maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/idot8u.ll

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

3167 lines
128 KiB
LLVM
Raw Normal View History

[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX7 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX8 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9 %s
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; RUN: llc -mtriple=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-DL %s
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; RUN: llc -mtriple=amdgcn -mcpu=gfx1011 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1012 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
define amdgpu_kernel void @udot8_acc32(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc32:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dword s6, s[6:7], 0x0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s20, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s7, s6, 28
; GFX7-NEXT: s_bfe_u32 s14, s6, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s6, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s6, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s6, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s6, 0x40008
; GFX7-NEXT: s_bfe_u32 s19, s6, 0x40004
; GFX7-NEXT: s_and_b32 s6, s6, 15
; GFX7-NEXT: s_lshr_b32 s5, s4, 28
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s13, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v0, s6
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s20
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v0, v1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s19
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s13, v1, v0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s12, v1, v0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s17
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s16
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s10, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s14
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s7
; GFX7-NEXT: v_mad_u32_u24 v0, s5, v1, v0
; GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc32:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s6, s[6:7], 0x0
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX8-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_lshr_b32 s7, s6, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s12, s6, 0x40018
; GFX8-NEXT: s_bfe_u32 s13, s6, 0x40014
; GFX8-NEXT: s_bfe_u32 s14, s6, 0x40010
; GFX8-NEXT: s_bfe_u32 s15, s6, 0x4000c
; GFX8-NEXT: s_bfe_u32 s16, s6, 0x40008
; GFX8-NEXT: s_bfe_u32 s17, s6, 0x40004
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_and_b32 s6, s6, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_lshr_b32 s3, s2, 28
; GFX8-NEXT: s_bfe_u32 s4, s2, 0x40018
; GFX8-NEXT: s_bfe_u32 s5, s2, 0x40014
; GFX8-NEXT: s_bfe_u32 s8, s2, 0x40010
; GFX8-NEXT: s_bfe_u32 s9, s2, 0x4000c
; GFX8-NEXT: s_bfe_u32 s10, s2, 0x40008
; GFX8-NEXT: s_bfe_u32 s11, s2, 0x40004
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_and_b32 s2, s2, 15
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v0, s2, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s17
; GFX8-NEXT: v_mad_u32_u24 v0, s11, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s16
; GFX8-NEXT: v_mad_u32_u24 v0, s10, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s15
; GFX8-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s14
; GFX8-NEXT: v_mad_u32_u24 v0, s8, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s13
; GFX8-NEXT: v_mad_u32_u24 v0, s5, v1, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s12
; GFX8-NEXT: v_mad_u32_u24 v0, s4, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s7
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc32:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s6, s[6:7], 0x0
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX9-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_lshr_b32 s7, s6, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s12, s6, 0x40018
; GFX9-NEXT: s_bfe_u32 s13, s6, 0x40014
; GFX9-NEXT: s_bfe_u32 s14, s6, 0x40010
; GFX9-NEXT: s_bfe_u32 s15, s6, 0x4000c
; GFX9-NEXT: s_bfe_u32 s16, s6, 0x40008
; GFX9-NEXT: s_bfe_u32 s17, s6, 0x40004
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_and_b32 s6, s6, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_lshr_b32 s3, s2, 28
; GFX9-NEXT: s_bfe_u32 s4, s2, 0x40018
; GFX9-NEXT: s_bfe_u32 s5, s2, 0x40014
; GFX9-NEXT: s_bfe_u32 s8, s2, 0x40010
; GFX9-NEXT: s_bfe_u32 s9, s2, 0x4000c
; GFX9-NEXT: s_bfe_u32 s10, s2, 0x40008
; GFX9-NEXT: s_bfe_u32 s11, s2, 0x40004
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_and_b32 s2, s2, 15
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s6
; GFX9-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v0, s2, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s17
; GFX9-NEXT: v_mad_u32_u24 v0, s11, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s16
; GFX9-NEXT: v_mad_u32_u24 v0, s10, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s15
; GFX9-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s14
; GFX9-NEXT: v_mad_u32_u24 v0, s8, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s13
; GFX9-NEXT: v_mad_u32_u24 v0, s5, v1, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s12
; GFX9-NEXT: v_mad_u32_u24 v0, s4, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s7
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s3, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: global_store_dword v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc32:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s2, s[6:7], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_load_dword s3, s[0:1], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s4, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s3
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_dot8_u32_u4 v2, s4, v0, v1
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc32:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] New SIInsertHardClauses pass Enable clausing of memory loads on gfx10 by adding a new pass to insert the s_clause instructions that mark the start of each hard clause. Differential Revision: https://reviews.llvm.org/D79792
2020-05-01 23:43:12 +08:00
; GFX10-DL-NEXT: s_clause 0x1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x34
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s6, s[4:5], 0x0
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s6
; GFX10-DL-NEXT: v_dot8_u32_u4 v2, s0, s1, v0
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s4
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s5
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i32 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%v1e0 = extractelement <8 x i4> %vec1, i64 0
%cv1e0 = zext i4 %v1e0 to i32
%v2e0 = extractelement <8 x i4> %vec2, i64 0
%cv2e0 = zext i4 %v2e0 to i32
%mul0 = mul nuw nsw i32 %cv1e0, %cv2e0
%v1e1 = extractelement <8 x i4> %vec1, i64 1
%cv1e1 = zext i4 %v1e1 to i32
%v2e1 = extractelement <8 x i4> %vec2, i64 1
%cv2e1 = zext i4 %v2e1 to i32
%mul1 = mul nuw nsw i32 %cv1e1, %cv2e1
%v1e2 = extractelement <8 x i4> %vec1, i64 2
%cv1e2 = zext i4 %v1e2 to i32
%v2e2 = extractelement <8 x i4> %vec2, i64 2
%cv2e2 = zext i4 %v2e2 to i32
%mul2 = mul nuw nsw i32 %cv1e2, %cv2e2
%v1e3 = extractelement <8 x i4> %vec1, i64 3
%cv1e3 = zext i4 %v1e3 to i32
%v2e3 = extractelement <8 x i4> %vec2, i64 3
%cv2e3 = zext i4 %v2e3 to i32
%mul3 = mul nuw nsw i32 %cv1e3, %cv2e3
%v1e4 = extractelement <8 x i4> %vec1, i64 4
%cv1e4 = zext i4 %v1e4 to i32
%v2e4 = extractelement <8 x i4> %vec2, i64 4
%cv2e4 = zext i4 %v2e4 to i32
%mul4 = mul nuw nsw i32 %cv1e4, %cv2e4
%v1e5 = extractelement <8 x i4> %vec1, i64 5
%cv1e5 = zext i4 %v1e5 to i32
%v2e5 = extractelement <8 x i4> %vec2, i64 5
%cv2e5 = zext i4 %v2e5 to i32
%mul5 = mul nuw nsw i32 %cv1e5, %cv2e5
%v1e6 = extractelement <8 x i4> %vec1, i64 6
%cv1e6 = zext i4 %v1e6 to i32
%v2e6 = extractelement <8 x i4> %vec2, i64 6
%cv2e6 = zext i4 %v2e6 to i32
%mul6 = mul nuw nsw i32 %cv1e6, %cv2e6
%v1e7 = extractelement <8 x i4> %vec1, i64 7
%cv1e7 = zext i4 %v1e7 to i32
%v2e7 = extractelement <8 x i4> %vec2, i64 7
%cv2e7 = zext i4 %v2e7 to i32
%mul7 = mul nuw nsw i32 %cv1e7, %cv2e7
%acc = load i32, i32 addrspace(1)* %dst, align 4
%add1 = add i32 %mul0, %acc
%add2 = add i32 %add1, %mul1
%add3 = add i32 %add2, %mul2
%add4 = add i32 %add3, %mul3
%add5 = add i32 %add4, %mul4
%add6 = add i32 %add5, %mul5
%add7 = add i32 %add6, %mul6
%add8 = add i32 %add7, %mul7
store i32 %add8, i32 addrspace(1)* %dst, align 4
ret void
}
; TODO: Remove the unnecessary instruction(that is zero-extending the
; 2nd MAD) to have the pattern-recognizer to kick in.
define amdgpu_kernel void @udot8_acc16(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc16:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_load_ushort v0, off, s[0:3], 0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s6, s4, 28
; GFX7-NEXT: s_bfe_u32 s14, s5, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s5, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s5, 0x40008
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40004
; GFX7-NEXT: s_lshr_b32 s13, s5, 28
; GFX7-NEXT: s_and_b32 s5, s5, 15
; GFX7-NEXT: s_bfe_u32 s7, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v1, s5
; GFX7-NEXT: v_mov_b32_e32 v2, s19
; GFX7-NEXT: v_mov_b32_e32 v3, s18
; GFX7-NEXT: v_mov_b32_e32 v4, s17
; GFX7-NEXT: v_mov_b32_e32 v5, s16
; GFX7-NEXT: v_mov_b32_e32 v6, s15
; GFX7-NEXT: v_mov_b32_e32 v7, s14
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v1, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s12, v2, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v3, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s10, v4, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v5, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v6, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s7, v7, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s13
; GFX7-NEXT: v_mad_u32_u24 v0, s6, v1, v0
; GFX7-NEXT: buffer_store_short v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: flat_load_ushort v2, v[0:1]
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX8-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX8-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX8-NEXT: s_bfe_u32 s13, s1, 0x4000c
; GFX8-NEXT: s_bfe_u32 s14, s1, 0x40008
; GFX8-NEXT: s_bfe_u32 s15, s1, 0x40004
; GFX8-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_and_b32 s1, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX8-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX8-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX8-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX8-NEXT: s_bfe_u32 s7, s0, 0x40008
; GFX8-NEXT: s_bfe_u32 s8, s0, 0x40004
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_and_b32 s0, s0, 15
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v4, s15
; GFX8-NEXT: v_mov_b32_e32 v5, s14
; GFX8-NEXT: v_mov_b32_e32 v6, s13
; GFX8-NEXT: v_mov_b32_e32 v7, s12
; GFX8-NEXT: v_mov_b32_e32 v8, s11
; GFX8-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s0, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s8, v4, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s6, v6, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX8-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: flat_store_short v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: global_load_ushort v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-NEXT: s_bfe_u32 s13, s1, 0x4000c
; GFX9-NEXT: s_bfe_u32 s14, s1, 0x40008
; GFX9-NEXT: s_bfe_u32 s15, s1, 0x40004
; GFX9-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_and_b32 s1, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX9-NEXT: s_bfe_u32 s7, s0, 0x40008
; GFX9-NEXT: s_bfe_u32 s8, s0, 0x40004
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_and_b32 s0, s0, 15
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v4, s15
; GFX9-NEXT: v_mov_b32_e32 v5, s14
; GFX9-NEXT: v_mov_b32_e32 v6, s13
; GFX9-NEXT: v_mov_b32_e32 v7, s12
; GFX9-NEXT: v_mov_b32_e32 v8, s11
; GFX9-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s0, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s8, v4, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s6, v6, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: global_store_short v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc16:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_load_ushort v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s13, s1, 0x4000c
; GFX9-DL-NEXT: s_bfe_u32 s14, s1, 0x40008
; GFX9-DL-NEXT: s_bfe_u32 s15, s1, 0x40004
; GFX9-DL-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_and_b32 s1, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX9-DL-NEXT: s_bfe_u32 s7, s0, 0x40008
; GFX9-DL-NEXT: s_bfe_u32 s8, s0, 0x40004
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_and_b32 s0, s0, 15
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v4, s15
; GFX9-DL-NEXT: v_mov_b32_e32 v5, s14
; GFX9-DL-NEXT: v_mov_b32_e32 v6, s13
; GFX9-DL-NEXT: v_mov_b32_e32 v7, s12
; GFX9-DL-NEXT: v_mov_b32_e32 v8, s11
; GFX9-DL-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s0, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s8, v4, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 0xffff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s6, v6, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: global_store_short v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc16:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX10-DL-NEXT: global_load_ushort v2, v[0:1], off
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40004
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40004
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40008
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v2, 0xffff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x4000c
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x4000c
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40010
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40010
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40014
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40014
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40018
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
; GFX10-DL-NEXT: s_lshr_b32 s1, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s0, s1, v2
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: global_store_short v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i16 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%v1e0 = extractelement <8 x i4> %vec1, i64 0
%cv1e0 = zext i4 %v1e0 to i16
%v2e0 = extractelement <8 x i4> %vec2, i64 0
%cv2e0 = zext i4 %v2e0 to i16
%mul0 = mul nuw nsw i16 %cv1e0, %cv2e0
%v1e1 = extractelement <8 x i4> %vec1, i64 1
%cv1e1 = zext i4 %v1e1 to i16
%v2e1 = extractelement <8 x i4> %vec2, i64 1
%cv2e1 = zext i4 %v2e1 to i16
%mul1 = mul nuw nsw i16 %cv1e1, %cv2e1
%v1e2 = extractelement <8 x i4> %vec1, i64 2
%cv1e2 = zext i4 %v1e2 to i16
%v2e2 = extractelement <8 x i4> %vec2, i64 2
%cv2e2 = zext i4 %v2e2 to i16
%mul2 = mul nuw nsw i16 %cv1e2, %cv2e2
%v1e3 = extractelement <8 x i4> %vec1, i64 3
%cv1e3 = zext i4 %v1e3 to i16
%v2e3 = extractelement <8 x i4> %vec2, i64 3
%cv2e3 = zext i4 %v2e3 to i16
%mul3 = mul nuw nsw i16 %cv1e3, %cv2e3
%v1e4 = extractelement <8 x i4> %vec1, i64 4
%cv1e4 = zext i4 %v1e4 to i16
%v2e4 = extractelement <8 x i4> %vec2, i64 4
%cv2e4 = zext i4 %v2e4 to i16
%mul4 = mul nuw nsw i16 %cv1e4, %cv2e4
%v1e5 = extractelement <8 x i4> %vec1, i64 5
%cv1e5 = zext i4 %v1e5 to i16
%v2e5 = extractelement <8 x i4> %vec2, i64 5
%cv2e5 = zext i4 %v2e5 to i16
%mul5 = mul nuw nsw i16 %cv1e5, %cv2e5
%v1e6 = extractelement <8 x i4> %vec1, i64 6
%cv1e6 = zext i4 %v1e6 to i16
%v2e6 = extractelement <8 x i4> %vec2, i64 6
%cv2e6 = zext i4 %v2e6 to i16
%mul6 = mul nuw nsw i16 %cv1e6, %cv2e6
%v1e7 = extractelement <8 x i4> %vec1, i64 7
%cv1e7 = zext i4 %v1e7 to i16
%v2e7 = extractelement <8 x i4> %vec2, i64 7
%cv2e7 = zext i4 %v2e7 to i16
%mul7 = mul nuw nsw i16 %cv1e7, %cv2e7
%acc = load i16, i16 addrspace(1)* %dst, align 4
%add1 = add i16 %mul0, %acc
%add2 = add i16 %add1, %mul1
%add3 = add i16 %add2, %mul2
%add4 = add i16 %add3, %mul3
%add5 = add i16 %add4, %mul4
%add6 = add i16 %add5, %mul5
%add7 = add i16 %add6, %mul6
%add8 = add i16 %add7, %mul7
store i16 %add8, i16 addrspace(1)* %dst, align 4
ret void
}
; TODO: Remove the unnecessary instruction(that is zero-extending the
; 2nd MAD) to have the pattern-recognizer to kick in.
define amdgpu_kernel void @udot8_acc8(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc8:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_load_ubyte v0, off, s[0:3], 0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s6, s4, 28
; GFX7-NEXT: s_bfe_u32 s14, s5, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s5, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s5, 0x40008
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40004
; GFX7-NEXT: s_lshr_b32 s13, s5, 28
; GFX7-NEXT: s_and_b32 s5, s5, 15
; GFX7-NEXT: s_bfe_u32 s7, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v1, s5
; GFX7-NEXT: v_mov_b32_e32 v2, s19
; GFX7-NEXT: v_mov_b32_e32 v3, s18
; GFX7-NEXT: v_mov_b32_e32 v4, s17
; GFX7-NEXT: v_mov_b32_e32 v5, s16
; GFX7-NEXT: v_mov_b32_e32 v6, s15
; GFX7-NEXT: v_mov_b32_e32 v7, s14
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v1, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s12, v2, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v3, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s10, v4, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v5, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v6, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s7, v7, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s13
; GFX7-NEXT: v_mad_u32_u24 v0, s6, v1, v0
; GFX7-NEXT: buffer_store_byte v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc8:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: flat_load_ubyte v2, v[0:1]
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX8-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX8-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX8-NEXT: s_bfe_u32 s13, s1, 0x4000c
; GFX8-NEXT: s_bfe_u32 s14, s1, 0x40008
; GFX8-NEXT: s_bfe_u32 s15, s1, 0x40004
; GFX8-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_and_b32 s1, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX8-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX8-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX8-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX8-NEXT: s_bfe_u32 s7, s0, 0x40008
; GFX8-NEXT: s_bfe_u32 s8, s0, 0x40004
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_and_b32 s0, s0, 15
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v4, s15
; GFX8-NEXT: v_mov_b32_e32 v5, s14
; GFX8-NEXT: v_mov_b32_e32 v6, s13
; GFX8-NEXT: v_mov_b32_e32 v7, s12
; GFX8-NEXT: v_mov_b32_e32 v8, s11
; GFX8-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s0, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s8, v4, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 0xff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s6, v6, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX8-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc8:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-NEXT: s_bfe_u32 s13, s1, 0x4000c
; GFX9-NEXT: s_bfe_u32 s14, s1, 0x40008
; GFX9-NEXT: s_bfe_u32 s15, s1, 0x40004
; GFX9-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_and_b32 s1, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX9-NEXT: s_bfe_u32 s7, s0, 0x40008
; GFX9-NEXT: s_bfe_u32 s8, s0, 0x40004
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_and_b32 s0, s0, 15
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v4, s15
; GFX9-NEXT: v_mov_b32_e32 v5, s14
; GFX9-NEXT: v_mov_b32_e32 v6, s13
; GFX9-NEXT: v_mov_b32_e32 v7, s12
; GFX9-NEXT: v_mov_b32_e32 v8, s11
; GFX9-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s0, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s8, v4, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 0xff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s6, v6, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: global_store_byte v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc8:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s13, s1, 0x4000c
; GFX9-DL-NEXT: s_bfe_u32 s14, s1, 0x40008
; GFX9-DL-NEXT: s_bfe_u32 s15, s1, 0x40004
; GFX9-DL-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_and_b32 s1, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX9-DL-NEXT: s_bfe_u32 s7, s0, 0x40008
; GFX9-DL-NEXT: s_bfe_u32 s8, s0, 0x40004
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_and_b32 s0, s0, 15
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v4, s15
; GFX9-DL-NEXT: v_mov_b32_e32 v5, s14
; GFX9-DL-NEXT: v_mov_b32_e32 v6, s13
; GFX9-DL-NEXT: v_mov_b32_e32 v7, s12
; GFX9-DL-NEXT: v_mov_b32_e32 v8, s11
; GFX9-DL-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s0, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s8, v4, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 0xff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s6, v6, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc8:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX10-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40004
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40004
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40008
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v2, 0xff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x4000c
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x4000c
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40010
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40010
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40014
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40014
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40018
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
; GFX10-DL-NEXT: s_lshr_b32 s1, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s0, s1, v2
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i8 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%v1e0 = extractelement <8 x i4> %vec1, i64 0
%cv1e0 = zext i4 %v1e0 to i8
%v2e0 = extractelement <8 x i4> %vec2, i64 0
%cv2e0 = zext i4 %v2e0 to i8
%mul0 = mul nuw nsw i8 %cv1e0, %cv2e0
%v1e1 = extractelement <8 x i4> %vec1, i64 1
%cv1e1 = zext i4 %v1e1 to i8
%v2e1 = extractelement <8 x i4> %vec2, i64 1
%cv2e1 = zext i4 %v2e1 to i8
%mul1 = mul nuw nsw i8 %cv1e1, %cv2e1
%v1e2 = extractelement <8 x i4> %vec1, i64 2
%cv1e2 = zext i4 %v1e2 to i8
%v2e2 = extractelement <8 x i4> %vec2, i64 2
%cv2e2 = zext i4 %v2e2 to i8
%mul2 = mul nuw nsw i8 %cv1e2, %cv2e2
%v1e3 = extractelement <8 x i4> %vec1, i64 3
%cv1e3 = zext i4 %v1e3 to i8
%v2e3 = extractelement <8 x i4> %vec2, i64 3
%cv2e3 = zext i4 %v2e3 to i8
%mul3 = mul nuw nsw i8 %cv1e3, %cv2e3
%v1e4 = extractelement <8 x i4> %vec1, i64 4
%cv1e4 = zext i4 %v1e4 to i8
%v2e4 = extractelement <8 x i4> %vec2, i64 4
%cv2e4 = zext i4 %v2e4 to i8
%mul4 = mul nuw nsw i8 %cv1e4, %cv2e4
%v1e5 = extractelement <8 x i4> %vec1, i64 5
%cv1e5 = zext i4 %v1e5 to i8
%v2e5 = extractelement <8 x i4> %vec2, i64 5
%cv2e5 = zext i4 %v2e5 to i8
%mul5 = mul nuw nsw i8 %cv1e5, %cv2e5
%v1e6 = extractelement <8 x i4> %vec1, i64 6
%cv1e6 = zext i4 %v1e6 to i8
%v2e6 = extractelement <8 x i4> %vec2, i64 6
%cv2e6 = zext i4 %v2e6 to i8
%mul6 = mul nuw nsw i8 %cv1e6, %cv2e6
%v1e7 = extractelement <8 x i4> %vec1, i64 7
%cv1e7 = zext i4 %v1e7 to i8
%v2e7 = extractelement <8 x i4> %vec2, i64 7
%cv2e7 = zext i4 %v2e7 to i8
%mul7 = mul nuw nsw i8 %cv1e7, %cv2e7
%acc = load i8, i8 addrspace(1)* %dst, align 4
%add1 = add i8 %mul0, %acc
%add2 = add i8 %add1, %mul1
%add3 = add i8 %add2, %mul2
%add4 = add i8 %add3, %mul3
%add5 = add i8 %add4, %mul4
%add6 = add i8 %add5, %mul5
%add7 = add i8 %add6, %mul6
%add8 = add i8 %add7, %mul7
store i8 %add8, i8 addrspace(1)* %dst, align 4
ret void
}
; TODO: Remove the two unnecessary instructions(and+add after 2nd MAD)
; to have the pattern-recognizer to kick in.
define amdgpu_kernel void @udot8_acc4(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc4:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_load_ubyte v0, off, s[0:3], 0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s6, s4, 28
; GFX7-NEXT: s_bfe_u32 s14, s5, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s5, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s5, 0x40008
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40004
; GFX7-NEXT: s_lshr_b32 s13, s5, 28
; GFX7-NEXT: s_and_b32 s5, s5, 15
; GFX7-NEXT: s_bfe_u32 s7, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v1, s5
; GFX7-NEXT: v_mov_b32_e32 v2, s19
; GFX7-NEXT: v_mov_b32_e32 v3, s18
; GFX7-NEXT: v_mov_b32_e32 v4, s17
; GFX7-NEXT: v_mov_b32_e32 v5, s16
; GFX7-NEXT: v_mov_b32_e32 v6, s15
; GFX7-NEXT: v_mov_b32_e32 v7, s14
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v1, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s12, v2, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v3, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s10, v4, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v5, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v6, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s7, v7, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s13
; GFX7-NEXT: v_mad_u32_u24 v0, s6, v1, v0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_and_b32_e32 v0, 15, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_store_byte v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc4:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: flat_load_ubyte v2, v[0:1]
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_and_b32 s8, s0, 15
; GFX8-NEXT: s_and_b32 s15, s1, 15
; GFX8-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX8-NEXT: v_mov_b32_e32 v4, s15
; GFX8-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX8-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX8-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX8-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX8-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX8-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX8-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX8-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX8-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_bfe_u32 s0, s0, 0x4000c
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX8-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v7, s12
; GFX8-NEXT: v_mov_b32_e32 v8, s11
; GFX8-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX8-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 15, v2
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc4:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_and_b32 s8, s0, 15
; GFX9-NEXT: s_and_b32 s15, s1, 15
; GFX9-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX9-NEXT: v_mov_b32_e32 v4, s15
; GFX9-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX9-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX9-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s0, s0, 0x4000c
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX9-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v7, s12
; GFX9-NEXT: v_mov_b32_e32 v8, s11
; GFX9-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_add_u32_e32 v2, v2, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 15, v2
; GFX9-NEXT: global_store_byte v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc4:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_and_b32 s8, s0, 15
; GFX9-DL-NEXT: s_and_b32 s15, s1, 15
; GFX9-DL-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX9-DL-NEXT: v_mov_b32_e32 v4, s15
; GFX9-DL-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX9-DL-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX9-DL-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s0, s0, 0x4000c
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX9-DL-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v7, s12
; GFX9-DL-NEXT: v_mov_b32_e32 v8, s11
; GFX9-DL-NEXT: v_mov_b32_e32 v9, s10
[TargetLowering] Add SimplifyMultipleUseDemandedBits This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured. The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use. We do see a couple of regressions that need to be addressed: AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better). X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG. The code owners have confirmed its ok for these cases to fixed up in future patches. Differential Revision: https://reviews.llvm.org/D63281 llvm-svn: 366799
2019-07-23 20:39:08 +08:00
; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_add_u32_e32 v2, v2, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 15, v2
; GFX9-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc4:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX10-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s4, s1, 0x40008
; GFX10-DL-NEXT: s_bfe_u32 s5, s1, 0x4000c
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40004
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40004
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s0, 0x4000c
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mul_u32_u24_e64 v3, s3, s5
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s4, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40010
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40010
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v2, 15, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v2, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40014
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40014
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40018
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
; GFX10-DL-NEXT: s_lshr_b32 s1, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s0, s1, v2
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v2, 15, v2
; GFX10-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i4 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%v1e0 = extractelement <8 x i4> %vec1, i64 0
%v2e0 = extractelement <8 x i4> %vec2, i64 0
%mul0 = mul nuw nsw i4 %v1e0, %v2e0
%v1e1 = extractelement <8 x i4> %vec1, i64 1
%v2e1 = extractelement <8 x i4> %vec2, i64 1
%mul1 = mul nuw nsw i4 %v1e1, %v2e1
%v1e2 = extractelement <8 x i4> %vec1, i64 2
%v2e2 = extractelement <8 x i4> %vec2, i64 2
%mul2 = mul nuw nsw i4 %v1e2, %v2e2
%v1e3 = extractelement <8 x i4> %vec1, i64 3
%v2e3 = extractelement <8 x i4> %vec2, i64 3
%mul3 = mul nuw nsw i4 %v1e3, %v2e3
%v1e4 = extractelement <8 x i4> %vec1, i64 4
%v2e4 = extractelement <8 x i4> %vec2, i64 4
%mul4 = mul nuw nsw i4 %v1e4, %v2e4
%v1e5 = extractelement <8 x i4> %vec1, i64 5
%v2e5 = extractelement <8 x i4> %vec2, i64 5
%mul5 = mul nuw nsw i4 %v1e5, %v2e5
%v1e6 = extractelement <8 x i4> %vec1, i64 6
%v2e6 = extractelement <8 x i4> %vec2, i64 6
%mul6 = mul nuw nsw i4 %v1e6, %v2e6
%v1e7 = extractelement <8 x i4> %vec1, i64 7
%v2e7 = extractelement <8 x i4> %vec2, i64 7
%mul7 = mul nuw nsw i4 %v1e7, %v2e7
%acc = load i4, i4 addrspace(1)* %dst, align 4
%add1 = add i4 %mul0, %acc
%add2 = add i4 %add1, %mul1
%add3 = add i4 %add2, %mul2
%add4 = add i4 %add3, %mul3
%add5 = add i4 %add4, %mul4
%add6 = add i4 %add5, %mul5
%add7 = add i4 %add6, %mul6
%add8 = add i4 %add7, %mul7
store i4 %add8, i4 addrspace(1)* %dst, align 4
ret void
}
; TODO: Currently, permutation of udot8 is turned off due to a huge increase
; in the compile time.
define amdgpu_kernel void @udot8_CommutationInsideMAD(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_CommutationInsideMAD:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_load_ubyte v0, off, s[0:3], 0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s6, s4, 28
; GFX7-NEXT: s_bfe_u32 s14, s5, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s5, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s5, 0x40008
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40004
; GFX7-NEXT: s_lshr_b32 s13, s5, 28
; GFX7-NEXT: s_and_b32 s5, s5, 15
; GFX7-NEXT: s_bfe_u32 s7, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v1, s5
; GFX7-NEXT: v_mov_b32_e32 v2, s19
; GFX7-NEXT: v_mov_b32_e32 v3, s18
; GFX7-NEXT: v_mov_b32_e32 v4, s17
; GFX7-NEXT: v_mov_b32_e32 v5, s16
; GFX7-NEXT: v_mov_b32_e32 v6, s15
; GFX7-NEXT: v_mov_b32_e32 v7, s14
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v1, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s12, v2, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v3, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s10, v4, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v5, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v6, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s7, v7, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s13
; GFX7-NEXT: v_mad_u32_u24 v0, s6, v1, v0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_and_b32_e32 v0, 15, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_store_byte v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_CommutationInsideMAD:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: flat_load_ubyte v2, v[0:1]
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_and_b32 s8, s0, 15
; GFX8-NEXT: s_and_b32 s15, s1, 15
; GFX8-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX8-NEXT: v_mov_b32_e32 v4, s15
; GFX8-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX8-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX8-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX8-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX8-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX8-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX8-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX8-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX8-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_bfe_u32 s0, s0, 0x4000c
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX8-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v7, s12
; GFX8-NEXT: v_mov_b32_e32 v8, s11
; GFX8-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX8-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 15, v2
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_CommutationInsideMAD:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_and_b32 s8, s0, 15
; GFX9-NEXT: s_and_b32 s15, s1, 15
; GFX9-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX9-NEXT: v_mov_b32_e32 v4, s15
; GFX9-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX9-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX9-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s0, s0, 0x4000c
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX9-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v7, s12
; GFX9-NEXT: v_mov_b32_e32 v8, s11
; GFX9-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_add_u32_e32 v2, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 15, v2
; GFX9-NEXT: global_store_byte v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_CommutationInsideMAD:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_and_b32 s8, s0, 15
; GFX9-DL-NEXT: s_and_b32 s15, s1, 15
; GFX9-DL-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX9-DL-NEXT: v_mov_b32_e32 v4, s15
; GFX9-DL-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX9-DL-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX9-DL-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s0, s0, 0x4000c
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX9-DL-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v7, s12
; GFX9-DL-NEXT: v_mov_b32_e32 v8, s11
; GFX9-DL-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_add_u32_e32 v2, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 15, v2
; GFX9-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_CommutationInsideMAD:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX10-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s4, s0, 0x40008
; GFX10-DL-NEXT: s_bfe_u32 s5, s1, 0x40008
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40004
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40004
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x4000c
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x4000c
; GFX10-DL-NEXT: v_mul_u32_u24_e64 v3, s2, s3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s4, s5, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40010
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40010
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v3, 15, v3
; GFX10-DL-NEXT: v_and_b32_e32 v2, 15, v2
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40014
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40014
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40018
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
; GFX10-DL-NEXT: s_lshr_b32 s1, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s0, s1, v2
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v2, 15, v2
; GFX10-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i4 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%v1e0 = extractelement <8 x i4> %vec1, i64 0
%v2e0 = extractelement <8 x i4> %vec2, i64 0
%mul0 = mul nuw nsw i4 %v1e0, %v2e0
%v1e1 = extractelement <8 x i4> %vec1, i64 1
%v2e1 = extractelement <8 x i4> %vec2, i64 1
%mul1 = mul nuw nsw i4 %v1e1, %v2e1
%v1e2 = extractelement <8 x i4> %vec1, i64 2
%v2e2 = extractelement <8 x i4> %vec2, i64 2
%mul2 = mul nuw nsw i4 %v1e2, %v2e2
%v1e3 = extractelement <8 x i4> %vec1, i64 3
%v2e3 = extractelement <8 x i4> %vec2, i64 3
%mul3 = mul nuw nsw i4 %v1e3, %v2e3
%v1e4 = extractelement <8 x i4> %vec1, i64 4
%v2e4 = extractelement <8 x i4> %vec2, i64 4
%mul4 = mul nuw nsw i4 %v1e4, %v2e4
%v1e5 = extractelement <8 x i4> %vec1, i64 5
%v2e5 = extractelement <8 x i4> %vec2, i64 5
%mul5 = mul nuw nsw i4 %v1e5, %v2e5
%v1e6 = extractelement <8 x i4> %vec1, i64 6
%v2e6 = extractelement <8 x i4> %vec2, i64 6
%mul6 = mul nuw nsw i4 %v1e6, %v2e6
%v1e7 = extractelement <8 x i4> %vec1, i64 7
%v2e7 = extractelement <8 x i4> %vec2, i64 7
%mul7 = mul nuw nsw i4 %v1e7, %v2e7
%acc = load i4, i4 addrspace(1)* %dst, align 4
%add1 = add i4 %mul0, %acc
%add2 = add i4 %mul1, %add1
%add3 = add i4 %mul2, %add2
%add4 = add i4 %mul3, %add3
%add5 = add i4 %mul4, %add4
%add6 = add i4 %mul5, %add5
%add7 = add i4 %mul6, %add6
%add8 = add i4 %mul7, %add7
store i4 %add8, i4 addrspace(1)* %dst, align 4
ret void
}
define amdgpu_kernel void @udot8_multiuses_mul1(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_multiuses_mul1:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dword s6, s[6:7], 0x0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s20, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_bfe_u32 s19, s6, 0x40004
; GFX7-NEXT: s_lshr_b32 s7, s6, 28
; GFX7-NEXT: s_bfe_u32 s14, s6, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s6, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s6, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s6, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s6, 0x40008
; GFX7-NEXT: s_and_b32 s6, s6, 15
; GFX7-NEXT: s_lshr_b32 s5, s4, 28
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s13, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v0, s6
; GFX7-NEXT: v_mov_b32_e32 v1, s20
; GFX7-NEXT: v_mad_u32_u24 v1, s4, v0, v1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v2, s19
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mad_u32_u24 v1, s13, v2, v1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v2, s18
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mad_u32_u24 v1, s12, v2, v1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v2, s17
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v1, s11, v2, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v2, s16
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v1, s10, v2, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v2, s15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v1, s9, v2, v1
; GFX7-NEXT: v_mov_b32_e32 v2, s14
; GFX7-NEXT: v_mad_u32_u24 v1, s8, v2, v1
; GFX7-NEXT: v_mov_b32_e32 v2, s7
; GFX7-NEXT: v_mad_u32_u24 v1, s5, v2, v1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v1, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_multiuses_mul1:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s6, s[6:7], 0x0
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX8-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s17, s6, 0x40004
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_lshr_b32 s7, s6, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s12, s6, 0x40018
; GFX8-NEXT: s_bfe_u32 s13, s6, 0x40014
; GFX8-NEXT: s_bfe_u32 s14, s6, 0x40010
; GFX8-NEXT: s_bfe_u32 s15, s6, 0x4000c
; GFX8-NEXT: s_bfe_u32 s16, s6, 0x40008
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_and_b32 s6, s6, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_lshr_b32 s3, s2, 28
; GFX8-NEXT: s_bfe_u32 s4, s2, 0x40018
; GFX8-NEXT: s_bfe_u32 s5, s2, 0x40014
; GFX8-NEXT: s_bfe_u32 s8, s2, 0x40010
; GFX8-NEXT: s_bfe_u32 s9, s2, 0x4000c
; GFX8-NEXT: s_bfe_u32 s10, s2, 0x40008
; GFX8-NEXT: s_bfe_u32 s11, s2, 0x40004
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_and_b32 s2, s2, 15
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v0, s6
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mad_u32_u24 v1, s2, v0, v1
; GFX8-NEXT: v_mov_b32_e32 v2, s17
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v0, s2, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mad_u32_u24 v1, s11, v2, v1
; GFX8-NEXT: v_mov_b32_e32 v2, s16
; GFX8-NEXT: v_mad_u32_u24 v1, s10, v2, v1
; GFX8-NEXT: v_mov_b32_e32 v2, s15
; GFX8-NEXT: v_mad_u32_u24 v1, s9, v2, v1
; GFX8-NEXT: v_mov_b32_e32 v2, s14
; GFX8-NEXT: v_mad_u32_u24 v1, s8, v2, v1
; GFX8-NEXT: v_mov_b32_e32 v2, s13
; GFX8-NEXT: v_mad_u32_u24 v1, s5, v2, v1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v2, s12
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mad_u32_u24 v1, s4, v2, v1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v2, s7
; GFX8-NEXT: v_mad_u32_u24 v1, s3, v2, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_multiuses_mul1:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s6, s[6:7], 0x0
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX9-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s17, s6, 0x40004
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_lshr_b32 s7, s6, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s12, s6, 0x40018
; GFX9-NEXT: s_bfe_u32 s13, s6, 0x40014
; GFX9-NEXT: s_bfe_u32 s14, s6, 0x40010
; GFX9-NEXT: s_bfe_u32 s15, s6, 0x4000c
; GFX9-NEXT: s_bfe_u32 s16, s6, 0x40008
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_and_b32 s6, s6, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_lshr_b32 s3, s2, 28
; GFX9-NEXT: s_bfe_u32 s4, s2, 0x40018
; GFX9-NEXT: s_bfe_u32 s5, s2, 0x40014
; GFX9-NEXT: s_bfe_u32 s8, s2, 0x40010
; GFX9-NEXT: s_bfe_u32 s9, s2, 0x4000c
; GFX9-NEXT: s_bfe_u32 s10, s2, 0x40008
; GFX9-NEXT: s_bfe_u32 s11, s2, 0x40004
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_and_b32 s2, s2, 15
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s6
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mad_u32_u24 v1, s2, v0, v1
; GFX9-NEXT: v_mov_b32_e32 v2, s17
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v0, s2, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mad_u32_u24 v1, s11, v2, v1
; GFX9-NEXT: v_mov_b32_e32 v2, s16
; GFX9-NEXT: v_mad_u32_u24 v1, s10, v2, v1
; GFX9-NEXT: v_mov_b32_e32 v2, s15
; GFX9-NEXT: v_mad_u32_u24 v1, s9, v2, v1
; GFX9-NEXT: v_mov_b32_e32 v2, s14
; GFX9-NEXT: v_mad_u32_u24 v1, s8, v2, v1
; GFX9-NEXT: v_mov_b32_e32 v2, s13
; GFX9-NEXT: v_mad_u32_u24 v1, s5, v2, v1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v2, s12
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mad_u32_u24 v1, s4, v2, v1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v2, s7
; GFX9-NEXT: v_mad_u32_u24 v1, s3, v2, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_add_u32_e32 v2, v0, v1
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: global_store_dword v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_multiuses_mul1:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s6, s[6:7], 0x0
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX9-DL-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s17, s6, 0x40004
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s7, s6, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s12, s6, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s13, s6, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s14, s6, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s15, s6, 0x4000c
; GFX9-DL-NEXT: s_bfe_u32 s16, s6, 0x40008
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_and_b32 s6, s6, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s3, s2, 28
; GFX9-DL-NEXT: s_bfe_u32 s4, s2, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s5, s2, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s8, s2, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s9, s2, 0x4000c
; GFX9-DL-NEXT: s_bfe_u32 s10, s2, 0x40008
; GFX9-DL-NEXT: s_bfe_u32 s11, s2, 0x40004
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_and_b32 s2, s2, 15
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s6
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s2, v0, v1
; GFX9-DL-NEXT: v_mov_b32_e32 v2, s17
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v0, s2, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s11, v2, v1
; GFX9-DL-NEXT: v_mov_b32_e32 v2, s16
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s10, v2, v1
; GFX9-DL-NEXT: v_mov_b32_e32 v2, s15
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s9, v2, v1
; GFX9-DL-NEXT: v_mov_b32_e32 v2, s14
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s8, v2, v1
; GFX9-DL-NEXT: v_mov_b32_e32 v2, s13
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s5, v2, v1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v2, s12
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s4, v2, v1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v2, s7
; GFX9-DL-NEXT: v_mad_u32_u24 v1, s3, v2, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: v_add_u32_e32 v2, v0, v1
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_multiuses_mul1:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] New SIInsertHardClauses pass Enable clausing of memory loads on gfx10 by adding a new pass to insert the s_clause instructions that mark the start of each hard clause. Differential Revision: https://reviews.llvm.org/D79792
2020-05-01 23:43:12 +08:00
; GFX10-DL-NEXT: s_clause 0x1
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x34
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_load_dword s6, s[4:5], 0x0
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s6
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
; GFX10-DL-NEXT: s_bfe_u32 s6, s0, 0x40004
; GFX10-DL-NEXT: s_bfe_u32 s7, s1, 0x40004
; GFX10-DL-NEXT: v_mad_u32_u24 v0, s2, s3, v0
; GFX10-DL-NEXT: v_mad_u32_u24 v1, s6, s7, v0
; GFX10-DL-NEXT: s_bfe_u32 s6, s0, 0x40008
; GFX10-DL-NEXT: s_bfe_u32 s7, s1, 0x40008
; GFX10-DL-NEXT: v_mad_u32_u24 v0, s2, s3, v0
; GFX10-DL-NEXT: v_mad_u32_u24 v1, s6, s7, v1
; GFX10-DL-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX10-DL-NEXT: s_bfe_u32 s7, s1, 0x4000c
; GFX10-DL-NEXT: v_mad_u32_u24 v1, s6, s7, v1
; GFX10-DL-NEXT: s_bfe_u32 s6, s0, 0x40010
; GFX10-DL-NEXT: s_bfe_u32 s7, s1, 0x40010
; GFX10-DL-NEXT: v_mad_u32_u24 v1, s6, s7, v1
; GFX10-DL-NEXT: s_bfe_u32 s6, s0, 0x40014
; GFX10-DL-NEXT: s_bfe_u32 s7, s1, 0x40014
; GFX10-DL-NEXT: v_mad_u32_u24 v1, s6, s7, v1
; GFX10-DL-NEXT: s_bfe_u32 s6, s0, 0x40018
; GFX10-DL-NEXT: s_bfe_u32 s7, s1, 0x40018
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
; GFX10-DL-NEXT: s_lshr_b32 s1, s1, 28
; GFX10-DL-NEXT: v_mad_u32_u24 v1, s6, s7, v1
; GFX10-DL-NEXT: v_mad_u32_u24 v1, s0, s1, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v0, v1
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s4
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s5
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i32 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%v1e0 = extractelement <8 x i4> %vec1, i64 0
%cv1e0 = zext i4 %v1e0 to i32
%v2e0 = extractelement <8 x i4> %vec2, i64 0
%cv2e0 = zext i4 %v2e0 to i32
%mul0 = mul nuw nsw i32 %cv1e0, %cv2e0
%v1e1 = extractelement <8 x i4> %vec1, i64 1
%cv1e1 = zext i4 %v1e1 to i32
%v2e1 = extractelement <8 x i4> %vec2, i64 1
%cv2e1 = zext i4 %v2e1 to i32
%mul1 = mul nuw nsw i32 %cv1e1, %cv2e1
%v1e2 = extractelement <8 x i4> %vec1, i64 2
%cv1e2 = zext i4 %v1e2 to i32
%v2e2 = extractelement <8 x i4> %vec2, i64 2
%cv2e2 = zext i4 %v2e2 to i32
%mul2 = mul nuw nsw i32 %cv1e2, %cv2e2
%v1e3 = extractelement <8 x i4> %vec1, i64 3
%cv1e3 = zext i4 %v1e3 to i32
%v2e3 = extractelement <8 x i4> %vec2, i64 3
%cv2e3 = zext i4 %v2e3 to i32
%mul3 = mul nuw nsw i32 %cv1e3, %cv2e3
%v1e4 = extractelement <8 x i4> %vec1, i64 4
%cv1e4 = zext i4 %v1e4 to i32
%v2e4 = extractelement <8 x i4> %vec2, i64 4
%cv2e4 = zext i4 %v2e4 to i32
%mul4 = mul nuw nsw i32 %cv1e4, %cv2e4
%v1e5 = extractelement <8 x i4> %vec1, i64 5
%cv1e5 = zext i4 %v1e5 to i32
%v2e5 = extractelement <8 x i4> %vec2, i64 5
%cv2e5 = zext i4 %v2e5 to i32
%mul5 = mul nuw nsw i32 %cv1e5, %cv2e5
%v1e6 = extractelement <8 x i4> %vec1, i64 6
%cv1e6 = zext i4 %v1e6 to i32
%v2e6 = extractelement <8 x i4> %vec2, i64 6
%cv2e6 = zext i4 %v2e6 to i32
%mul6 = mul nuw nsw i32 %cv1e6, %cv2e6
%v1e7 = extractelement <8 x i4> %vec1, i64 7
%cv1e7 = zext i4 %v1e7 to i32
%v2e7 = extractelement <8 x i4> %vec2, i64 7
%cv2e7 = zext i4 %v2e7 to i32
%mul7 = mul nuw nsw i32 %cv1e7, %cv2e7
%acc = load i32, i32 addrspace(1)* %dst, align 4
%add1 = add i32 %mul0, %acc
%add = add i32 %mul0, %add1
%add2 = add i32 %add1, %mul1
%add3 = add i32 %add2, %mul2
%add4 = add i32 %add3, %mul3
%add5 = add i32 %add4, %mul4
%add6 = add i32 %add5, %mul5
%add7 = add i32 %add6, %mul6
%add8 = add i32 %add7, %mul7
%res = add i32 %add, %add8
store i32 %res, i32 addrspace(1)* %dst, align 4
ret void
}
define amdgpu_kernel void @udot8_acc32_vecMul(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc32_vecMul:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dword s6, s[6:7], 0x0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s20, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s7, s6, 28
; GFX7-NEXT: s_bfe_u32 s14, s6, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s6, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s6, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s6, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s6, 0x40008
; GFX7-NEXT: s_bfe_u32 s19, s6, 0x40004
; GFX7-NEXT: s_and_b32 s6, s6, 15
; GFX7-NEXT: s_lshr_b32 s5, s4, 28
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s13, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v0, s6
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s20
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v0, v1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s19
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s13, v1, v0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s12, v1, v0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s17
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s16
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s10, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s14
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s7
; GFX7-NEXT: v_mad_u32_u24 v0, s5, v1, v0
; GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc32_vecMul:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s6, s[6:7], 0x0
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX8-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_lshr_b32 s7, s6, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s12, s6, 0x40018
; GFX8-NEXT: s_bfe_u32 s13, s6, 0x40014
; GFX8-NEXT: s_bfe_u32 s14, s6, 0x40010
; GFX8-NEXT: s_bfe_u32 s15, s6, 0x4000c
; GFX8-NEXT: s_bfe_u32 s16, s6, 0x40008
; GFX8-NEXT: s_bfe_u32 s17, s6, 0x40004
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_and_b32 s6, s6, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_lshr_b32 s3, s2, 28
; GFX8-NEXT: s_bfe_u32 s4, s2, 0x40018
; GFX8-NEXT: s_bfe_u32 s5, s2, 0x40014
; GFX8-NEXT: s_bfe_u32 s8, s2, 0x40010
; GFX8-NEXT: s_bfe_u32 s9, s2, 0x4000c
; GFX8-NEXT: s_bfe_u32 s10, s2, 0x40008
; GFX8-NEXT: s_bfe_u32 s11, s2, 0x40004
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_and_b32 s2, s2, 15
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v0, s6
; GFX8-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v0, s2, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s17
; GFX8-NEXT: v_mad_u32_u24 v0, s11, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s16
; GFX8-NEXT: v_mad_u32_u24 v0, s10, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s15
; GFX8-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s14
; GFX8-NEXT: v_mad_u32_u24 v0, s8, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s13
; GFX8-NEXT: v_mad_u32_u24 v0, s5, v1, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s12
; GFX8-NEXT: v_mad_u32_u24 v0, s4, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s7
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc32_vecMul:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s6, s[6:7], 0x0
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX9-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_lshr_b32 s7, s6, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s12, s6, 0x40018
; GFX9-NEXT: s_bfe_u32 s13, s6, 0x40014
; GFX9-NEXT: s_bfe_u32 s14, s6, 0x40010
; GFX9-NEXT: s_bfe_u32 s15, s6, 0x4000c
; GFX9-NEXT: s_bfe_u32 s16, s6, 0x40008
; GFX9-NEXT: s_bfe_u32 s17, s6, 0x40004
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_and_b32 s6, s6, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_lshr_b32 s3, s2, 28
; GFX9-NEXT: s_bfe_u32 s4, s2, 0x40018
; GFX9-NEXT: s_bfe_u32 s5, s2, 0x40014
; GFX9-NEXT: s_bfe_u32 s8, s2, 0x40010
; GFX9-NEXT: s_bfe_u32 s9, s2, 0x4000c
; GFX9-NEXT: s_bfe_u32 s10, s2, 0x40008
; GFX9-NEXT: s_bfe_u32 s11, s2, 0x40004
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_and_b32 s2, s2, 15
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s6
; GFX9-NEXT: v_mov_b32_e32 v1, s18
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v0, s2, v0, v1
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s17
; GFX9-NEXT: v_mad_u32_u24 v0, s11, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s16
; GFX9-NEXT: v_mad_u32_u24 v0, s10, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s15
; GFX9-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s14
; GFX9-NEXT: v_mad_u32_u24 v0, s8, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s13
; GFX9-NEXT: v_mad_u32_u24 v0, s5, v1, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s12
; GFX9-NEXT: v_mad_u32_u24 v0, s4, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s7
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s3, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: global_store_dword v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc32_vecMul:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s2, s[6:7], 0x0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_load_dword s3, s[0:1], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s4, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s3
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_dot8_u32_u4 v2, s4, v0, v1
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc32_vecMul:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] New SIInsertHardClauses pass Enable clausing of memory loads on gfx10 by adding a new pass to insert the s_clause instructions that mark the start of each hard clause. Differential Revision: https://reviews.llvm.org/D79792
2020-05-01 23:43:12 +08:00
; GFX10-DL-NEXT: s_clause 0x1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x34
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s6, s[4:5], 0x0
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s6
; GFX10-DL-NEXT: v_dot8_u32_u4 v2, s0, s1, v0
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s4
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s5
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i32 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%cvec1 = zext <8 x i4> %vec1 to <8 x i32>
%cvec2 = zext <8 x i4> %vec2 to <8 x i32>
%mul = mul <8 x i32> %cvec1, %cvec2
%mul0 = extractelement <8 x i32> %mul, i64 0
%mul1 = extractelement <8 x i32> %mul, i64 1
%mul2 = extractelement <8 x i32> %mul, i64 2
%mul3 = extractelement <8 x i32> %mul, i64 3
%mul4 = extractelement <8 x i32> %mul, i64 4
%mul5 = extractelement <8 x i32> %mul, i64 5
%mul6 = extractelement <8 x i32> %mul, i64 6
%mul7 = extractelement <8 x i32> %mul, i64 7
%acc = load i32, i32 addrspace(1)* %dst, align 4
%add1 = add i32 %mul0, %acc
%add2 = add i32 %add1, %mul1
%add3 = add i32 %add2, %mul2
%add4 = add i32 %add3, %mul3
%add5 = add i32 %add4, %mul4
%add6 = add i32 %add5, %mul5
%add7 = add i32 %add6, %mul6
%add8 = add i32 %add7, %mul7
store i32 %add8, i32 addrspace(1)* %dst, align 4
ret void
}
; TODO: Clean up the code(by default pk_mad_I16 should be generated), then
; support the pattern.
define amdgpu_kernel void @udot8_acc16_vecMul(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc16_vecMul:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_load_ushort v0, off, s[0:3], 0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s17, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mov_b32_e32 v4, s17
; GFX7-NEXT: s_bfe_u32 s14, s5, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s5, 0x40010
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: s_bfe_u32 s18, s5, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s13, s5, 28
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: s_and_b32 s5, s5, 15
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mov_b32_e32 v2, s19
; GFX7-NEXT: v_mul_u32_u24_e32 v2, s12, v2
; GFX7-NEXT: v_mul_u32_u24_e32 v4, s10, v4
; GFX7-NEXT: s_lshr_b32 s6, s4, 28
; GFX7-NEXT: s_bfe_u32 s7, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40010
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mov_b32_e32 v3, s18
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: s_and_b32 s4, s4, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s5
; GFX7-NEXT: v_mul_u32_u24_e32 v1, s4, v1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mul_u32_u24_e32 v3, s11, v3
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: v_or_b32_e32 v3, v3, v4
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX7-NEXT: v_or_b32_e32 v1, v1, v2
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: v_alignbit_b32 v2, v3, v2, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mov_b32_e32 v5, s16
; GFX7-NEXT: v_mov_b32_e32 v6, s15
; GFX7-NEXT: v_mov_b32_e32 v7, s14
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt vmcnt(0)
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v0, v1
[DAG][AMDGPU][X86] Add SimplifyMultipleUseDemandedBits handling for SIGN/ZERO_EXTEND + SIGN/ZERO_EXTEND_VECTOR_INREG Peek through multiple use ops like we already do for ANY_EXTEND/ANY_EXTEND_VECTOR_INREG Differential Revision: https://reviews.llvm.org/D84863
2020-07-30 00:33:56 +08:00
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v2, v0
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v3, v0
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v4, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v5, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v6, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s7, v7, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s13
; GFX7-NEXT: v_mad_u32_u24 v0, s6, v1, v0
; GFX7-NEXT: buffer_store_short v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc16_vecMul:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: flat_load_ushort v2, v[0:1]
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX8-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX8-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX8-NEXT: s_bfe_u32 s13, s1, 0x4000c
; GFX8-NEXT: s_bfe_u32 s14, s1, 0x40008
; GFX8-NEXT: s_bfe_u32 s15, s1, 0x40004
; GFX8-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_and_b32 s1, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX8-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX8-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX8-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX8-NEXT: s_bfe_u32 s7, s0, 0x40008
; GFX8-NEXT: s_bfe_u32 s8, s0, 0x40004
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_and_b32 s0, s0, 15
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v4, s15
; GFX8-NEXT: v_mov_b32_e32 v5, s14
; GFX8-NEXT: v_mov_b32_e32 v6, s13
; GFX8-NEXT: v_mov_b32_e32 v7, s12
; GFX8-NEXT: v_mov_b32_e32 v8, s11
; GFX8-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s0, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s8, v4, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s6, v6, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX8-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: flat_store_short v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc16_vecMul:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_load_dword s6, s[6:7], 0x0
; GFX9-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s7, s6, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_lshr_b32 s12, s6, 28
; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s12
; GFX9-NEXT: s_bfe_u32 s3, s2, 0x40018
; GFX9-NEXT: s_lshr_b32 s4, s2, 28
; GFX9-NEXT: s_bfe_u32 s13, s6, 0x40010
; GFX9-NEXT: s_bfe_u32 s14, s6, 0x40014
; GFX9-NEXT: s_pack_ll_b32_b16 s3, s3, s4
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s7
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_pk_mul_lo_u16 v2, s3, v0
; GFX9-NEXT: s_pack_ll_b32_b16 s3, s13, s14
; GFX9-NEXT: s_bfe_u32 s5, s2, 0x40010
; GFX9-NEXT: s_bfe_u32 s8, s2, 0x40014
; GFX9-NEXT: s_bfe_u32 s15, s6, 0x40008
; GFX9-NEXT: s_bfe_u32 s16, s6, 0x4000c
; GFX9-NEXT: s_and_b32 s17, s6, 15
; GFX9-NEXT: v_mov_b32_e32 v0, s3
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s5, s8
; GFX9-NEXT: s_bfe_u32 s9, s2, 0x40008
; GFX9-NEXT: s_bfe_u32 s10, s2, 0x4000c
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s6, s6, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_pack_ll_b32_b16 s3, s15, s16
; GFX9-NEXT: v_pk_mul_lo_u16 v3, s4, v0
; GFX9-NEXT: s_and_b32 s11, s2, 15
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s2, s2, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s3
; GFX9-NEXT: s_pack_ll_b32_b16 s4, s9, s10
; GFX9-NEXT: s_pack_ll_b32_b16 s3, s17, s6
; GFX9-NEXT: v_pk_mul_lo_u16 v4, s4, v0
; GFX9-NEXT: s_pack_ll_b32_b16 s2, s11, s2
; GFX9-NEXT: v_mov_b32_e32 v0, s3
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_pk_mul_lo_u16 v5, s2, v0
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: global_load_ushort v6, v[0:1], off
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_add_u32_e32 v6, v5, v6
; GFX9-NEXT: v_add_u32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT: v_add_u32_sdwa v5, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:BYTE_0
; GFX9-NEXT: v_add_u32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT: v_add_u32_e32 v4, v4, v3
; GFX9-NEXT: v_add_u32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT: v_add_u32_e32 v3, v3, v2
; GFX9-NEXT: v_add_u32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: global_store_short v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc16_vecMul:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_load_dword s6, s[6:7], 0x0
; GFX9-DL-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s7, s6, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s12, s6, 28
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s7, s7, s12
; GFX9-DL-NEXT: s_bfe_u32 s3, s2, 0x40018
; GFX9-DL-NEXT: s_lshr_b32 s4, s2, 28
; GFX9-DL-NEXT: s_bfe_u32 s13, s6, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s14, s6, 0x40014
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s3, s3, s4
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s7
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_pk_mul_lo_u16 v2, s3, v0
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s3, s13, s14
; GFX9-DL-NEXT: s_bfe_u32 s5, s2, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s8, s2, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s15, s6, 0x40008
; GFX9-DL-NEXT: s_bfe_u32 s16, s6, 0x4000c
; GFX9-DL-NEXT: s_and_b32 s17, s6, 15
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s3
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s4, s5, s8
; GFX9-DL-NEXT: s_bfe_u32 s9, s2, 0x40008
; GFX9-DL-NEXT: s_bfe_u32 s10, s2, 0x4000c
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s6, s6, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s3, s15, s16
; GFX9-DL-NEXT: v_pk_mul_lo_u16 v3, s4, v0
; GFX9-DL-NEXT: s_and_b32 s11, s2, 15
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s2, s2, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s3
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s4, s9, s10
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s3, s17, s6
; GFX9-DL-NEXT: v_pk_mul_lo_u16 v4, s4, v0
; GFX9-DL-NEXT: s_pack_ll_b32_b16 s2, s11, s2
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s3
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_pk_mul_lo_u16 v5, s2, v0
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_load_ushort v6, v[0:1], off
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_add_u32_e32 v6, v5, v6
; GFX9-DL-NEXT: v_add_u32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-DL-NEXT: v_add_u32_sdwa v5, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:BYTE_0
; GFX9-DL-NEXT: v_add_u32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-DL-NEXT: v_add_u32_e32 v4, v4, v3
; GFX9-DL-NEXT: v_add_u32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-DL-NEXT: v_add_u32_e32 v3, v3, v2
; GFX9-DL-NEXT: v_add_u32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: global_store_short v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc16_vecMul:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX10-DL-NEXT: global_load_ushort v2, v[0:1], off
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s5, s0, 0x40004
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
; GFX10-DL-NEXT: s_bfe_u32 s4, s1, 0x40004
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s2, s2, s5
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s3, s3, s4
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s4, s1, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_pk_mul_lo_u16 v3, s2, s3
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s5, s1, 0x4000c
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40008
; GFX10-DL-NEXT: s_bfe_u32 s3, s0, 0x4000c
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s4, s4, s5
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s2, s2, s3
; GFX10-DL-NEXT: s_bfe_u32 s3, s0, 0x40014
; GFX10-DL-NEXT: v_pk_mul_lo_u16 v4, s2, s4
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40010
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s4, s1, 0x40010
; GFX10-DL-NEXT: s_bfe_u32 s5, s1, 0x40014
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s2, s2, s3
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s4, s4, s5
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40018
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s1, s1, 28
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s1, s3, s1
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v3, v2
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_pk_mul_lo_u16 v3, s2, s4
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40018
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:BYTE_0
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_pack_ll_b32_b16 s0, s2, s0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-DL-NEXT: v_pk_mul_lo_u16 v4, s0, s1
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v2, v3
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v2, v4
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-DL-NEXT: global_store_short v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i16 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%cvec1 = zext <8 x i4> %vec1 to <8 x i16>
%cvec2 = zext <8 x i4> %vec2 to <8 x i16>
%mul = mul <8 x i16> %cvec1, %cvec2
%mul0 = extractelement <8 x i16> %mul, i64 0
%mul1 = extractelement <8 x i16> %mul, i64 1
%mul2 = extractelement <8 x i16> %mul, i64 2
%mul3 = extractelement <8 x i16> %mul, i64 3
%mul4 = extractelement <8 x i16> %mul, i64 4
%mul5 = extractelement <8 x i16> %mul, i64 5
%mul6 = extractelement <8 x i16> %mul, i64 6
%mul7 = extractelement <8 x i16> %mul, i64 7
%acc = load i16, i16 addrspace(1)* %dst, align 4
%add1 = add i16 %mul0, %acc
%add2 = add i16 %add1, %mul1
%add3 = add i16 %add2, %mul2
%add4 = add i16 %add3, %mul3
%add5 = add i16 %add4, %mul4
%add6 = add i16 %add5, %mul5
%add7 = add i16 %add6, %mul6
%add8 = add i16 %add7, %mul7
store i16 %add8, i16 addrspace(1)* %dst, align 4
ret void
}
; TODO: Cleanup the code to generate MAD; pattern should be recognized then.
define amdgpu_kernel void @udot8_acc8_vecMul(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc8_vecMul:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_load_ubyte v0, off, s[0:3], 0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_bfe_u32 s6, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s13, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40004
; GFX7-NEXT: s_lshr_b32 s17, s5, 28
; GFX7-NEXT: v_mov_b32_e32 v8, s13
; GFX7-NEXT: s_bfe_u32 s14, s5, 0x40008
; GFX7-NEXT: s_and_b32 s16, s5, 15
; GFX7-NEXT: s_bfe_u32 s18, s5, 0x40018
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40004
; GFX7-NEXT: v_mov_b32_e32 v6, s15
; GFX7-NEXT: s_lshr_b32 s10, s4, 28
; GFX7-NEXT: v_mov_b32_e32 v4, s17
; GFX7-NEXT: v_mul_u32_u24_e32 v4, s10, v4
; GFX7-NEXT: v_mul_u32_u24_e32 v6, s8, v6
; GFX7-NEXT: v_mul_u32_u24_e32 v8, s6, v8
; GFX7-NEXT: s_bfe_u32 s5, s5, 0x40010
; GFX7-NEXT: s_bfe_u32 s7, s4, 0x40008
; GFX7-NEXT: v_mov_b32_e32 v7, s14
; GFX7-NEXT: s_and_b32 s9, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v5, s16
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x40018
; GFX7-NEXT: v_mov_b32_e32 v3, s18
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40014
; GFX7-NEXT: v_mov_b32_e32 v2, s19
; GFX7-NEXT: v_mul_u32_u24_e32 v2, s12, v2
; GFX7-NEXT: s_bfe_u32 s4, s4, 0x40010
; GFX7-NEXT: v_mov_b32_e32 v1, s5
; GFX7-NEXT: v_mul_u32_u24_e32 v3, s11, v3
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 8, v4
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mul_u32_u24_e32 v5, s9, v5
; GFX7-NEXT: v_mul_u32_u24_e32 v7, s7, v7
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 8, v6
[TargetLowering] Add SimplifyMultipleUseDemandedBits This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured. The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use. We do see a couple of regressions that need to be addressed: AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better). X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG. The code owners have confirmed its ok for these cases to fixed up in future patches. Differential Revision: https://reviews.llvm.org/D63281 llvm-svn: 366799
2019-07-23 20:39:08 +08:00
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 8, v8
; GFX7-NEXT: v_or_b32_e32 v3, v3, v4
; GFX7-NEXT: v_or_b32_e32 v4, v5, v6
; GFX7-NEXT: v_or_b32_e32 v5, v7, v8
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mul_u32_u24_e32 v9, s4, v1
[TargetLowering] Add SimplifyMultipleUseDemandedBits This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured. The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use. We do see a couple of regressions that need to be addressed: AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better). X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG. The code owners have confirmed its ok for these cases to fixed up in future patches. Differential Revision: https://reviews.llvm.org/D63281 llvm-svn: 366799
2019-07-23 20:39:08 +08:00
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 8, v2
; GFX7-NEXT: v_or_b32_e32 v2, v9, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX7-NEXT: v_or_b32_e32 v2, v2, v3
; GFX7-NEXT: v_or_b32_e32 v3, v4, v5
; GFX7-NEXT: v_alignbit_b32 v4, v2, v3, 8
; GFX7-NEXT: v_alignbit_b32 v5, v2, v3, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 24, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 8, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 24, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt vmcnt(0)
[TargetLowering] Add SimplifyMultipleUseDemandedBits This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured. The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use. We do see a couple of regressions that need to be addressed: AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better). X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG. The code owners have confirmed its ok for these cases to fixed up in future patches. Differential Revision: https://reviews.llvm.org/D63281 llvm-svn: 366799
2019-07-23 20:39:08 +08:00
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v0, v3
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v4, v0
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v5, v0
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v6, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v1, v0
[TargetLowering] Add SimplifyMultipleUseDemandedBits This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured. The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use. We do see a couple of regressions that need to be addressed: AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better). X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG. The code owners have confirmed its ok for these cases to fixed up in future patches. Differential Revision: https://reviews.llvm.org/D63281 llvm-svn: 366799
2019-07-23 20:39:08 +08:00
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v0, v7
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v0, v8
; GFX7-NEXT: v_add_i32_e32 v0, vcc, v0, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_store_byte v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc8_vecMul:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: flat_load_ubyte v2, v[0:1]
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: s_load_dword s1, s[4:5], 0x0
; GFX8-NEXT: s_load_dword s2, s[6:7], 0x0
; GFX8-NEXT: s_mov_b32 s0, 0xffff
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s7, s1, 0x40004
; GFX8-NEXT: s_bfe_u32 s9, s1, 0x4000c
; GFX8-NEXT: s_bfe_u32 s14, s2, 0x40004
; GFX8-NEXT: s_and_b32 s15, s2, 15
; GFX8-NEXT: s_bfe_u32 s16, s2, 0x4000c
; GFX8-NEXT: s_bfe_u32 s3, s1, 0x40014
; GFX8-NEXT: s_lshr_b32 s5, s1, 28
; GFX8-NEXT: s_bfe_u32 s10, s2, 0x40014
; GFX8-NEXT: s_bfe_u32 s11, s2, 0x40010
; GFX8-NEXT: s_lshr_b32 s12, s2, 28
; GFX8-NEXT: s_bfe_u32 s13, s2, 0x40018
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: s_bfe_u32 s2, s2, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_and_b32 s8, s1, 15
; GFX8-NEXT: v_mov_b32_e32 v4, s16
; GFX8-NEXT: v_mov_b32_e32 v5, s9
; GFX8-NEXT: v_mov_b32_e32 v6, s15
; GFX8-NEXT: v_mov_b32_e32 v7, s14
; GFX8-NEXT: v_mov_b32_e32 v8, s7
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mul_u32_u24_sdwa v4, v5, v4 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mul_u32_u24_e32 v5, s8, v6
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mul_u32_u24_sdwa v6, v8, v7 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s4, s1, 0x40010
; GFX8-NEXT: s_bfe_u32 s6, s1, 0x40018
; GFX8-NEXT: v_mov_b32_e32 v9, s13
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: s_bfe_u32 s1, s1, 0x40008
; GFX8-NEXT: v_mov_b32_e32 v3, s2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v10, s12
; GFX8-NEXT: v_mov_b32_e32 v11, s5
; GFX8-NEXT: v_mov_b32_e32 v12, s11
; GFX8-NEXT: v_mov_b32_e32 v13, s10
; GFX8-NEXT: v_mov_b32_e32 v14, s3
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_mul_u32_u24_e32 v3, s1, v3
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_or_b32_e32 v5, v5, v6
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mul_u32_u24_e32 v7, s6, v9
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_mul_u32_u24_sdwa v8, v11, v10 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mul_u32_u24_e32 v9, s4, v12
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_mul_u32_u24_sdwa v10, v14, v13 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT: v_and_b32_e32 v5, s0, v5
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_or_b32_e32 v9, v9, v10
; GFX8-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT: v_and_b32_e32 v4, s0, v9
; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
; GFX8-NEXT: v_or_b32_e32 v6, v4, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 8, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v8, 8, v6
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt vmcnt(0)
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v5
[TargetLowering] Add SimplifyMultipleUseDemandedBits This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured. The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use. We do see a couple of regressions that need to be addressed: AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better). X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG. The code owners have confirmed its ok for these cases to fixed up in future patches. Differential Revision: https://reviews.llvm.org/D63281 llvm-svn: 366799
2019-07-23 20:39:08 +08:00
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v7, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_add_u32_sdwa v2, vcc, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_0
; GFX8-NEXT: v_add_u32_sdwa v2, vcc, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v4
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v8, v2
; GFX8-NEXT: v_add_u32_sdwa v2, vcc, v6, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX8-NEXT: v_add_u32_sdwa v2, vcc, v6, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc8_vecMul:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: global_load_ubyte v2, v[0:1], off
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_load_dword s1, s[4:5], 0x0
; GFX9-NEXT: s_load_dword s2, s[6:7], 0x0
; GFX9-NEXT: s_mov_b32 s0, 0xffff
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_bfe_u32 s3, s1, 0x40010
; GFX9-NEXT: s_bfe_u32 s10, s2, 0x40010
; GFX9-NEXT: s_bfe_u32 s11, s2, 0x40014
; GFX9-NEXT: s_bfe_u32 s12, s2, 0x40018
; GFX9-NEXT: s_lshr_b32 s13, s2, 28
; GFX9-NEXT: s_and_b32 s14, s2, 15
; GFX9-NEXT: s_bfe_u32 s15, s2, 0x40004
; GFX9-NEXT: s_bfe_u32 s16, s2, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v3, s10
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_bfe_u32 s2, s2, 0x4000c
; GFX9-NEXT: s_bfe_u32 s4, s1, 0x40014
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v4, s11
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_bfe_u32 s5, s1, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v5, s12
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_lshr_b32 s6, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v6, s13
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_and_b32 s7, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v7, s14
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_bfe_u32 s8, s1, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v8, s15
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_bfe_u32 s9, s1, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v9, s16
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: s_bfe_u32 s1, s1, 0x4000c
; GFX9-NEXT: v_mov_b32_e32 v10, s2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mul_lo_u16_e32 v3, s3, v3
; GFX9-NEXT: v_mul_lo_u16_sdwa v4, s4, v4 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT: v_mul_lo_u16_e32 v5, s5, v5
; GFX9-NEXT: v_mul_lo_u16_sdwa v6, s6, v6 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT: v_mul_lo_u16_e32 v7, s7, v7
; GFX9-NEXT: v_mul_lo_u16_sdwa v8, s8, v8 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_or_b32_e32 v3, v3, v4
; GFX9-NEXT: v_or_b32_sdwa v4, v5, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-NEXT: v_or_b32_e32 v5, v7, v8
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mul_lo_u16_e32 v9, s9, v9
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: v_mul_lo_u16_sdwa v10, s1, v10 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT: v_and_b32_e32 v5, s0, v5
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-NEXT: v_or_b32_sdwa v6, v9, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT: v_or_b32_e32 v6, v5, v6
; GFX9-NEXT: v_lshrrev_b32_e32 v7, 8, v6
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-NEXT: v_and_b32_e32 v3, s0, v3
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-NEXT: v_or_b32_e32 v4, v3, v4
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt vmcnt(0)
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-NEXT: v_add_u32_e32 v2, v5, v2
; GFX9-NEXT: v_add_u32_e32 v2, v2, v7
; GFX9-NEXT: v_add_u32_sdwa v2, v2, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_2
; GFX9-NEXT: v_add_u32_sdwa v2, v2, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_3
; GFX9-NEXT: v_add_u32_e32 v2, v2, v3
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_lshrrev_b32_e32 v3, 8, v4
; GFX9-NEXT: v_add_u32_e32 v2, v2, v3
; GFX9-NEXT: v_add_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT: v_add_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_3
; GFX9-NEXT: global_store_byte v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc8_vecMul:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_load_ubyte v2, v[0:1], off
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_load_dword s1, s[4:5], 0x0
; GFX9-DL-NEXT: s_load_dword s2, s[6:7], 0x0
; GFX9-DL-NEXT: s_mov_b32 s0, 0xffff
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s3, s1, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s10, s2, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s11, s2, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s12, s2, 0x40018
; GFX9-DL-NEXT: s_lshr_b32 s13, s2, 28
; GFX9-DL-NEXT: s_and_b32 s14, s2, 15
; GFX9-DL-NEXT: s_bfe_u32 s15, s2, 0x40004
; GFX9-DL-NEXT: s_bfe_u32 s16, s2, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s10
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s2, s2, 0x4000c
; GFX9-DL-NEXT: s_bfe_u32 s4, s1, 0x40014
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v4, s11
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s5, s1, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v5, s12
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s6, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v6, s13
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_and_b32 s7, s1, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v7, s14
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s8, s1, 0x40004
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v8, s15
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s9, s1, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v9, s16
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s1, s1, 0x4000c
; GFX9-DL-NEXT: v_mov_b32_e32 v10, s2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mul_lo_u16_e32 v3, s3, v3
; GFX9-DL-NEXT: v_mul_lo_u16_sdwa v4, s4, v4 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-DL-NEXT: v_mul_lo_u16_e32 v5, s5, v5
; GFX9-DL-NEXT: v_mul_lo_u16_sdwa v6, s6, v6 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-DL-NEXT: v_mul_lo_u16_e32 v7, s7, v7
; GFX9-DL-NEXT: v_mul_lo_u16_sdwa v8, s8, v8 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_or_b32_e32 v3, v3, v4
; GFX9-DL-NEXT: v_or_b32_sdwa v4, v5, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_or_b32_e32 v5, v7, v8
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mul_lo_u16_e32 v9, s9, v9
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: v_mul_lo_u16_sdwa v10, s1, v10 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-DL-NEXT: v_and_b32_e32 v5, s0, v5
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_or_b32_sdwa v6, v9, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-DL-NEXT: v_or_b32_e32 v6, v5, v6
; GFX9-DL-NEXT: v_lshrrev_b32_e32 v7, 8, v6
[MachineScheduler] Fix the TopDepth/BotHeightReduce latency heuristics tryLatency compares two sched candidates. For the top zone it prefers the one with lesser depth, but only if that depth is greater than the total latency of the instructions we've already scheduled -- otherwise its latency would be hidden and there would be no stall. Unfortunately it only tests the depth of one of the candidates. This can lead to situations where the TopDepthReduce heuristic does not kick in, but a lower priority heuristic chooses the other candidate, whose depth *is* greater than the already scheduled latency, which causes a stall. The fix is to apply the heuristic if the depth of *either* candidate is greater than the already scheduled latency. All this also applies to the BotHeightReduce heuristic in the bottom zone. Differential Revision: https://reviews.llvm.org/D72392
2020-01-07 23:43:46 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v3, s0, v3
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_or_b32_e32 v4, v3, v4
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_add_u32_e32 v2, v5, v2
; GFX9-DL-NEXT: v_add_u32_e32 v2, v2, v7
; GFX9-DL-NEXT: v_add_u32_sdwa v2, v2, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_2
; GFX9-DL-NEXT: v_add_u32_sdwa v2, v2, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_3
; GFX9-DL-NEXT: v_add_u32_e32 v2, v2, v3
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_lshrrev_b32_e32 v3, 8, v4
; GFX9-DL-NEXT: v_add_u32_e32 v2, v2, v3
; GFX9-DL-NEXT: v_add_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-DL-NEXT: v_add_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_3
; GFX9-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc8_vecMul:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX10-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s0, 0x40004
; GFX10-DL-NEXT: s_bfe_u32 s5, s1, 0x40004
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v3, s3, s5
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
; GFX10-DL-NEXT: s_bfe_u32 s6, s0, 0x4000c
; GFX10-DL-NEXT: s_bfe_u32 s5, s1, 0x4000c
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v4, s2, s3
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v5, s6, s5
AMDGPU: Fix i16 arithmetic pattern redundancy There were 2 problems here. First, these patterns were duplicated to handle the inverted shift operands instead of using the commuted PatFrags. Second, the point of the zext folding patterns don't apply to the non-0ing high subtargets. They should be skipped instead of inserting the extension. The zeroing high code would be emitted when necessary anyway. This was also emitting unnecessary zexts in cases where the high bits were undefined. llvm-svn: 374092
2019-10-09 01:36:38 +08:00
; GFX10-DL-NEXT: v_lshlrev_b16_e64 v3, 8, v3
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s4, s0, 0x40008
; GFX10-DL-NEXT: s_bfe_u32 s2, s1, 0x40008
; GFX10-DL-NEXT: s_mov_b32 s3, 0xffff
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v6, s4, s2
AMDGPU: Fix i16 arithmetic pattern redundancy There were 2 problems here. First, these patterns were duplicated to handle the inverted shift operands instead of using the commuted PatFrags. Second, the point of the zext folding patterns don't apply to the non-0ing high subtargets. They should be skipped instead of inserting the extension. The zeroing high code would be emitted when necessary anyway. This was also emitting unnecessary zexts in cases where the high bits were undefined. llvm-svn: 374092
2019-10-09 01:36:38 +08:00
; GFX10-DL-NEXT: v_or_b32_e32 v3, v4, v3
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_lshlrev_b16_e64 v5, 8, v5
; GFX10-DL-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX10-DL-NEXT: s_bfe_u32 s6, s1, 0x40014
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40010
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v3, s3, v3
AMDGPU: Fix i16 arithmetic pattern redundancy There were 2 problems here. First, these patterns were duplicated to handle the inverted shift operands instead of using the commuted PatFrags. Second, the point of the zext folding patterns don't apply to the non-0ing high subtargets. They should be skipped instead of inserting the extension. The zeroing high code would be emitted when necessary anyway. This was also emitting unnecessary zexts in cases where the high bits were undefined. llvm-svn: 374092
2019-10-09 01:36:38 +08:00
; GFX10-DL-NEXT: v_or_b32_sdwa v4, v6, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v5, s4, s6
; GFX10-DL-NEXT: s_bfe_u32 s5, s0, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s7, s1, 0x40010
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
AMDGPU: Fix i16 arithmetic pattern redundancy There were 2 problems here. First, these patterns were duplicated to handle the inverted shift operands instead of using the commuted PatFrags. Second, the point of the zext folding patterns don't apply to the non-0ing high subtargets. They should be skipped instead of inserting the extension. The zeroing high code would be emitted when necessary anyway. This was also emitting unnecessary zexts in cases where the high bits were undefined. llvm-svn: 374092
2019-10-09 01:36:38 +08:00
; GFX10-DL-NEXT: v_or_b32_e32 v4, v3, v4
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s4, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v6, s2, s7
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v7, s0, s4
; GFX10-DL-NEXT: v_lshlrev_b16_e64 v5, 8, v5
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_lshrrev_b32_e32 v8, 8, v4
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s0, s1, 0x40018
; GFX10-DL-NEXT: v_mul_lo_u16_e64 v10, s5, s0
; GFX10-DL-NEXT: v_or_b32_e32 v5, v6, v5
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_lshlrev_b16_e64 v7, 8, v7
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v5, s3, v5
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
AMDGPU: Fix i16 arithmetic pattern redundancy There were 2 problems here. First, these patterns were duplicated to handle the inverted shift operands instead of using the commuted PatFrags. Second, the point of the zext folding patterns don't apply to the non-0ing high subtargets. They should be skipped instead of inserting the extension. The zeroing high code would be emitted when necessary anyway. This was also emitting unnecessary zexts in cases where the high bits were undefined. llvm-svn: 374092
2019-10-09 01:36:38 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_or_b32_sdwa v3, v10, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v2, v8
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_or_b32_e32 v3, v5, v3
AMDGPU: Fix i16 arithmetic pattern redundancy There were 2 problems here. First, these patterns were duplicated to handle the inverted shift operands instead of using the commuted PatFrags. Second, the point of the zext folding patterns don't apply to the non-0ing high subtargets. They should be skipped instead of inserting the extension. The zeroing high code would be emitted when necessary anyway. This was also emitting unnecessary zexts in cases where the high bits were undefined. llvm-svn: 374092
2019-10-09 01:36:38 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_2
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_3
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_lshrrev_b32_e32 v4, 8, v3
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v2, v5
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v2, v4
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-DL-NEXT: v_add_nc_u32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_3
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i8 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%cvec1 = zext <8 x i4> %vec1 to <8 x i8>
%cvec2 = zext <8 x i4> %vec2 to <8 x i8>
%mul = mul <8 x i8> %cvec1, %cvec2
%mul0 = extractelement <8 x i8> %mul, i64 0
%mul1 = extractelement <8 x i8> %mul, i64 1
%mul2 = extractelement <8 x i8> %mul, i64 2
%mul3 = extractelement <8 x i8> %mul, i64 3
%mul4 = extractelement <8 x i8> %mul, i64 4
%mul5 = extractelement <8 x i8> %mul, i64 5
%mul6 = extractelement <8 x i8> %mul, i64 6
%mul7 = extractelement <8 x i8> %mul, i64 7
%acc = load i8, i8 addrspace(1)* %dst, align 4
%add1 = add i8 %mul0, %acc
%add2 = add i8 %add1, %mul1
%add3 = add i8 %add2, %mul2
%add4 = add i8 %add3, %mul3
%add5 = add i8 %add4, %mul4
%add6 = add i8 %add5, %mul5
%add7 = add i8 %add6, %mul6
%add8 = add i8 %add7, %mul7
store i8 %add8, i8 addrspace(1)* %dst, align 4
ret void
}
; TODO: Once the adictional "and+add" are removed, the pattern will be recognized.
define amdgpu_kernel void @udot8_acc4_vecMul(<8 x i4> addrspace(1)* %src1,
; GFX7-LABEL: udot8_acc4_vecMul:
; GFX7: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_load_ubyte v0, off, s[0:3], 0
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: s_lshr_b32 s6, s4, 28
; GFX7-NEXT: s_bfe_u32 s14, s5, 0x40018
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s16, s5, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s18, s5, 0x40008
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40004
; GFX7-NEXT: s_lshr_b32 s13, s5, 28
; GFX7-NEXT: s_and_b32 s5, s5, 15
; GFX7-NEXT: s_bfe_u32 s7, s4, 0x40018
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s9, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s11, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x40004
; GFX7-NEXT: s_and_b32 s4, s4, 15
; GFX7-NEXT: v_mov_b32_e32 v1, s5
; GFX7-NEXT: v_mov_b32_e32 v2, s19
; GFX7-NEXT: v_mov_b32_e32 v3, s18
; GFX7-NEXT: v_mov_b32_e32 v4, s17
; GFX7-NEXT: v_mov_b32_e32 v5, s16
; GFX7-NEXT: v_mov_b32_e32 v6, s15
; GFX7-NEXT: v_mov_b32_e32 v7, s14
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: v_mad_u32_u24 v0, s4, v1, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s12, v2, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v3, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s10, v4, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v5, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s8, v6, v0
; GFX7-NEXT: v_mad_u32_u24 v0, s7, v7, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s13
; GFX7-NEXT: v_mad_u32_u24 v0, s6, v1, v0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: v_and_b32_e32 v0, 15, v0
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX7-NEXT: buffer_store_byte v0, off, s[0:3], 0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_acc4_vecMul:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: flat_load_ubyte v2, v[0:1]
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_and_b32 s8, s0, 15
; GFX8-NEXT: s_and_b32 s15, s1, 15
; GFX8-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX8-NEXT: v_mov_b32_e32 v4, s15
; GFX8-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX8-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX8-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX8-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX8-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX8-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX8-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX8-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX8-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: s_bfe_u32 s0, s0, 0x4000c
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX8-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mov_b32_e32 v7, s12
; GFX8-NEXT: v_mov_b32_e32 v8, s11
; GFX8-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v3, v2
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX8-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX8-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX8-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX8-NEXT: v_and_b32_e32 v2, 15, v2
; GFX8-NEXT: flat_store_byte v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_acc4_vecMul:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_and_b32 s8, s0, 15
; GFX9-NEXT: s_and_b32 s15, s1, 15
; GFX9-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX9-NEXT: v_mov_b32_e32 v4, s15
; GFX9-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX9-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX9-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: s_bfe_u32 s0, s0, 0x4000c
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX9-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mov_b32_e32 v7, s12
; GFX9-NEXT: v_mov_b32_e32 v8, s11
; GFX9-NEXT: v_mov_b32_e32 v9, s10
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_add_u32_e32 v2, v2, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
; GFX9-NEXT: v_and_b32_e32 v2, 15, v2
; GFX9-NEXT: global_store_byte v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-LABEL: udot8_acc4_vecMul:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s0, s[4:5], 0x0
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_load_dword s1, s[6:7], 0x0
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_and_b32 s8, s0, 15
; GFX9-DL-NEXT: s_and_b32 s15, s1, 15
; GFX9-DL-NEXT: s_bfe_u32 s14, s1, 0x40004
; GFX9-DL-NEXT: v_mov_b32_e32 v4, s15
; GFX9-DL-NEXT: s_bfe_u32 s10, s1, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s11, s1, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s12, s1, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s13, s1, 0x40008
; GFX9-DL-NEXT: s_lshr_b32 s9, s1, 28
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s1, s1, 0x4000c
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s7, s0, 0x40004
; GFX9-DL-NEXT: v_mov_b32_e32 v5, s14
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_lshr_b32 s2, s0, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s3, s0, 0x40018
; GFX9-DL-NEXT: s_bfe_u32 s4, s0, 0x40014
; GFX9-DL-NEXT: s_bfe_u32 s5, s0, 0x40010
; GFX9-DL-NEXT: s_bfe_u32 s6, s0, 0x40008
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: s_bfe_u32 s0, s0, 0x4000c
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s1
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v6, s13
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mul_u32_u24_e32 v3, s0, v3
; GFX9-DL-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v7, s12
; GFX9-DL-NEXT: v_mov_b32_e32 v8, s11
; GFX9-DL-NEXT: v_mov_b32_e32 v9, s10
[TargetLowering] Add SimplifyMultipleUseDemandedBits This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured. The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use. We do see a couple of regressions that need to be addressed: AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better). X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG. The code owners have confirmed its ok for these cases to fixed up in future patches. Differential Revision: https://reviews.llvm.org/D63281 llvm-svn: 366799
2019-07-23 20:39:08 +08:00
; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s8, v4, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s7, v5, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s6, v6, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 15, v2
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_add_u32_e32 v2, v2, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s5, v7, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s4, v8, v2
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s3, v9, v2
; GFX9-DL-NEXT: v_mov_b32_e32 v3, s9
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX9-DL-NEXT: v_mad_u32_u24 v2, s2, v3, v2
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
; GFX9-DL-NEXT: v_and_b32_e32 v2, 15, v2
; GFX9-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_acc4_vecMul:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] Bundle loads before post-RA scheduler We are relying on atrificial DAG edges inserted by the MemOpClusterMutation to keep loads and stores together in the post-RA scheduler. This does not work all the time since it allows to schedule a completely independent instruction in the middle of the cluster. Removed the DAG mutation and added pass to bundle already clustered instructions. These bundles are unpacked before the memory legalizer because it does not work with bundles but also because it allows to insert waitcounts in the middle of a store cluster. Removing artificial edges also allows a more relaxed scheduling. Differential Revision: https://reviews.llvm.org/D72737
2020-01-14 06:54:17 +08:00
; GFX10-DL-NEXT: global_load_ubyte v2, v[0:1], off
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_and_b32 s2, s0, 15
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_and_b32 s3, s1, 15
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s4, s1, 0x40008
; GFX10-DL-NEXT: s_bfe_u32 s5, s1, 0x4000c
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40004
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40004
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40008
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s0, 0x4000c
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_mul_u32_u24_e64 v3, s3, s5
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s4, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40010
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40010
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v3, 15, v3
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v2, 15, v2
[SelectionDAG] Check for any recursion depth greater than or equal to limit instead of just equal the limit. If anything called the recursive isKnownNeverNaN/computeKnownBits/ComputeNumSignBits/SimplifyDemandedBits/SimplifyMultipleUseDemandedBits with an incorrect depth then we could continue to recurse if we'd already exceeded the depth limit. This replaces the limit check (Depth == 6) with a (Depth >= 6) to make sure that we don't circumvent it. This causes a couple of regressions as a mixture of calls (SimplifyMultipleUseDemandedBits + combineX86ShufflesRecursively) were calling with depths that were already over the limit. I've fixed SimplifyMultipleUseDemandedBits to not do this. combineX86ShufflesRecursively is trickier as we get a lot of regressions if we reduce its own limit from 8 to 6 (it also starts at Depth == 1 instead of Depth == 0 like the others....) - I'll see what I can do in future patches. llvm-svn: 367171
2019-07-27 20:48:46 +08:00
; GFX10-DL-NEXT: v_add_nc_u32_e32 v2, v2, v3
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] More accurate gfx10 latencies Differential Revision: https://reviews.llvm.org/D81012
2020-03-11 19:33:30 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40014
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40014
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s2, s0, 0x40018
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: s_bfe_u32 s3, s1, 0x40018
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_lshr_b32 s0, s0, 28
; GFX10-DL-NEXT: s_lshr_b32 s1, s1, 28
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
2020-01-22 06:27:57 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s2, s3, v2
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: v_mad_u32_u24 v2, s0, s1, v2
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: v_and_b32_e32 v2, 15, v2
; GFX10-DL-NEXT: global_store_byte v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Match udot8 pattern Summary: D.u32 = S0.u4[0] * S1.u4[0] + S0.u4[1] * S1.u4[1] + S0.u4[2] * S1.u4[2] + S0.u4[3] * S1.u4[3] + S0.u4[4] * S1.u4[4] + S0.u4[5] * S1.u4[5] + S0.u4[6] * S1.u4[6] + S0.u4[7] * S1.u4[7] + S2.u32 Author: FarhanaAleen Reviewed By: arsenm, nhaehnle Differential Revision: https://reviews.llvm.org/D51947 llvm-svn: 342497
2018-09-19 00:59:48 +08:00
<8 x i4> addrspace(1)* %src2,
i4 addrspace(1)* nocapture %dst) {
entry:
%vec1 = load <8 x i4>, <8 x i4> addrspace(1)* %src1
%vec2 = load <8 x i4>, <8 x i4> addrspace(1)* %src2
%mul = mul <8 x i4> %vec1, %vec2
%mul0 = extractelement <8 x i4> %mul, i64 0
%mul1 = extractelement <8 x i4> %mul, i64 1
%mul2 = extractelement <8 x i4> %mul, i64 2
%mul3 = extractelement <8 x i4> %mul, i64 3
%mul4 = extractelement <8 x i4> %mul, i64 4
%mul5 = extractelement <8 x i4> %mul, i64 5
%mul6 = extractelement <8 x i4> %mul, i64 6
%mul7 = extractelement <8 x i4> %mul, i64 7
%acc = load i4, i4 addrspace(1)* %dst, align 4
%add1 = add i4 %mul0, %acc
%add2 = add i4 %add1, %mul1
%add3 = add i4 %add2, %mul2
%add4 = add i4 %add3, %mul3
%add5 = add i4 %add4, %mul4
%add6 = add i4 %add5, %mul5
%add7 = add i4 %add6, %mul6
%add8 = add i4 %add7, %mul7
store i4 %add8, i4 addrspace(1)* %dst, align 4
[AMDGPU] Match signed dot4/8 pattern. Summary: This patch matches signed dot4 and dot8 pattern. Author: FarhanaAleen Reviewed By: msearles Differential Revision: https://reviews.llvm.org/D52520 llvm-svn: 343798
2018-10-05 00:57:37 +08:00
ret void
}
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
define amdgpu_kernel void @udot8_variant1(i32 addrspace(1)* %v1addr,
; GFX7-LABEL: udot8_variant1:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; GFX7-NEXT: s_mov_b32 s3, 0xf000
; GFX7-NEXT: s_mov_b32 s2, -1
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
; GFX7-NEXT: s_load_dword s4, s[4:5], 0x0
; GFX7-NEXT: s_load_dword s5, s[6:7], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: s_load_dword s20, s[0:1], 0x0
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: s_and_b32 s6, s4, 15
; GFX7-NEXT: s_and_b32 s7, s5, 15
; GFX7-NEXT: s_bfe_u32 s8, s4, 0x40004
; GFX7-NEXT: s_bfe_u32 s10, s4, 0x40008
; GFX7-NEXT: s_bfe_u32 s12, s4, 0x4000c
; GFX7-NEXT: s_bfe_u32 s14, s4, 0x40010
; GFX7-NEXT: s_bfe_u32 s16, s4, 0x40014
; GFX7-NEXT: s_bfe_u32 s18, s4, 0x40018
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX7-NEXT: s_lshr_b32 s4, s4, 28
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v0, s6
; GFX7-NEXT: v_mov_b32_e32 v1, s20
; GFX7-NEXT: v_mad_u32_u24 v0, s7, v0, v1
; GFX7-NEXT: s_bfe_u32 s9, s5, 0x40004
; GFX7-NEXT: s_bfe_u32 s11, s5, 0x40008
; GFX7-NEXT: s_bfe_u32 s13, s5, 0x4000c
; GFX7-NEXT: s_bfe_u32 s15, s5, 0x40010
; GFX7-NEXT: s_bfe_u32 s17, s5, 0x40014
; GFX7-NEXT: s_bfe_u32 s19, s5, 0x40018
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX7-NEXT: s_lshr_b32 s5, s5, 28
; GFX7-NEXT: v_mov_b32_e32 v1, s4
; GFX7-NEXT: v_mad_u32_u24 v0, s5, v1, v0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX7-NEXT: v_mov_b32_e32 v1, s8
; GFX7-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s10
; GFX7-NEXT: v_mad_u32_u24 v0, s11, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s12
; GFX7-NEXT: v_mad_u32_u24 v0, s13, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s14
; GFX7-NEXT: v_mad_u32_u24 v0, s15, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s16
; GFX7-NEXT: v_mad_u32_u24 v0, s17, v1, v0
; GFX7-NEXT: v_mov_b32_e32 v1, s18
; GFX7-NEXT: v_mad_u32_u24 v0, s19, v1, v0
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX7-NEXT: s_endpgm
;
; GFX8-LABEL: udot8_variant1:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_load_dword s2, s[4:5], 0x0
; GFX8-NEXT: s_load_dword s3, s[6:7], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: s_and_b32 s4, s2, 15
; GFX8-NEXT: s_and_b32 s5, s3, 15
; GFX8-NEXT: s_bfe_u32 s6, s2, 0x40004
; GFX8-NEXT: s_bfe_u32 s8, s2, 0x40008
; GFX8-NEXT: s_bfe_u32 s10, s2, 0x4000c
; GFX8-NEXT: s_bfe_u32 s12, s2, 0x40010
; GFX8-NEXT: s_bfe_u32 s14, s2, 0x40014
; GFX8-NEXT: s_bfe_u32 s16, s2, 0x40018
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX8-NEXT: s_lshr_b32 s2, s2, 28
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v0, s4
; GFX8-NEXT: v_mov_b32_e32 v1, s18
; GFX8-NEXT: v_mad_u32_u24 v0, s5, v0, v1
; GFX8-NEXT: s_bfe_u32 s7, s3, 0x40004
; GFX8-NEXT: s_bfe_u32 s9, s3, 0x40008
; GFX8-NEXT: s_bfe_u32 s11, s3, 0x4000c
; GFX8-NEXT: s_bfe_u32 s13, s3, 0x40010
; GFX8-NEXT: s_bfe_u32 s15, s3, 0x40014
; GFX8-NEXT: s_bfe_u32 s17, s3, 0x40018
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX8-NEXT: s_lshr_b32 s3, s3, 28
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX8-NEXT: v_mov_b32_e32 v1, s2
; GFX8-NEXT: v_mad_u32_u24 v0, s3, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s6
; GFX8-NEXT: v_mad_u32_u24 v0, s7, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s8
; GFX8-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s10
; GFX8-NEXT: v_mad_u32_u24 v0, s11, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s12
; GFX8-NEXT: v_mad_u32_u24 v0, s13, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s14
; GFX8-NEXT: v_mad_u32_u24 v0, s15, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v1, s16
; GFX8-NEXT: v_mad_u32_u24 v2, s17, v1, v0
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX8-NEXT: flat_store_dword v[0:1], v2
; GFX8-NEXT: s_endpgm
;
; GFX9-LABEL: udot8_variant1:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_load_dword s2, s[4:5], 0x0
; GFX9-NEXT: s_load_dword s3, s[6:7], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_load_dword s18, s[0:1], 0x0
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: s_and_b32 s4, s2, 15
; GFX9-NEXT: s_and_b32 s5, s3, 15
; GFX9-NEXT: s_bfe_u32 s6, s2, 0x40004
; GFX9-NEXT: s_bfe_u32 s8, s2, 0x40008
; GFX9-NEXT: s_bfe_u32 s10, s2, 0x4000c
; GFX9-NEXT: s_bfe_u32 s12, s2, 0x40010
; GFX9-NEXT: s_bfe_u32 s14, s2, 0x40014
; GFX9-NEXT: s_bfe_u32 s16, s2, 0x40018
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX9-NEXT: s_lshr_b32 s2, s2, 28
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: v_mov_b32_e32 v1, s18
; GFX9-NEXT: v_mad_u32_u24 v0, s5, v0, v1
; GFX9-NEXT: s_bfe_u32 s7, s3, 0x40004
; GFX9-NEXT: s_bfe_u32 s9, s3, 0x40008
; GFX9-NEXT: s_bfe_u32 s11, s3, 0x4000c
; GFX9-NEXT: s_bfe_u32 s13, s3, 0x40010
; GFX9-NEXT: s_bfe_u32 s15, s3, 0x40014
; GFX9-NEXT: s_bfe_u32 s17, s3, 0x40018
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX9-NEXT: s_lshr_b32 s3, s3, 28
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-NEXT: v_mov_b32_e32 v1, s2
; GFX9-NEXT: v_mad_u32_u24 v0, s3, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s6
; GFX9-NEXT: v_mad_u32_u24 v0, s7, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s8
; GFX9-NEXT: v_mad_u32_u24 v0, s9, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s10
; GFX9-NEXT: v_mad_u32_u24 v0, s11, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s12
; GFX9-NEXT: v_mad_u32_u24 v0, s13, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s14
; GFX9-NEXT: v_mad_u32_u24 v0, s15, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v1, s16
; GFX9-NEXT: v_mad_u32_u24 v2, s17, v1, v0
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX9-NEXT: global_store_dword v[0:1], v2, off
; GFX9-NEXT: s_endpgm
;
; GFX9-DL-LABEL: udot8_variant1:
; GFX9-DL: ; %bb.0: ; %entry
; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: s_load_dword s2, s[4:5], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX9-DL-NEXT: s_load_dword s3, s[0:1], 0x0
; GFX9-DL-NEXT: s_load_dword s4, s[6:7], 0x0
; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s2
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s3
; GFX9-DL-NEXT: v_dot8_u32_u4 v2, s4, v0, v1
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
; GFX9-DL-NEXT: v_mov_b32_e32 v0, s0
; GFX9-DL-NEXT: v_mov_b32_e32 v1, s1
; GFX9-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX9-DL-NEXT: s_endpgm
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
;
; GFX10-DL-LABEL: udot8_variant1:
; GFX10-DL: ; %bb.0: ; %entry
[AMDGPU] New SIInsertHardClauses pass Enable clausing of memory loads on gfx10 by adding a new pass to insert the s_clause instructions that mark the start of each hard clause. Differential Revision: https://reviews.llvm.org/D79792
2020-05-01 23:43:12 +08:00
; GFX10-DL-NEXT: s_clause 0x1
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x34
; GFX10-DL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: ; implicit-def: $vcc_hi
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: s_load_dword s6, s[4:5], 0x0
; GFX10-DL-NEXT: s_load_dword s0, s[0:1], 0x0
; GFX10-DL-NEXT: s_load_dword s1, s[2:3], 0x0
[AMDGPU] Fix typo in SIInstrInfo::memOpsHaveSameBasePtr Summary: The typo has been present since memOpsHaveSameBasePtr was introduced in r313208. It caused SIInstrInfo::shouldClusterMemOps to cluster more mem ops than it was supposed to. Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D71616
2019-12-18 00:09:02 +08:00
; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
[AMDGPU] Remove dubious logic in bidirectional list scheduler Summary: pickNodeBidirectional tried to compare the best top candidate and the best bottom candidate by examining TopCand.Reason and BotCand.Reason. This is unsound because, after calling pickNodeFromQueue, Cand.Reason does not reflect the most important reason why Cand was chosen. Rather it reflects the most recent reason why it beat some other potential candidate, which could have been for some low priority tie breaker reason. I have seen this cause problems where TopCand is a good candidate, but because TopCand.Reason is ORDER (which is very low priority) it is repeatedly ignored in favour of a mediocre BotCand. This is not how bidirectional scheduling is supposed to work. To fix this I changed the code to always compare TopCand and BotCand directly, like the generic implementation of pickNodeBidirectional does. This removes some uncommented AMDGPU-specific logic; if this logic turns out to be important then perhaps it could be moved into an override of tryCandidate instead. Graphics shader benchmarking on gfx10 shows a lot more positive than negative effects from this change. Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D68338
2019-10-07 22:33:59 +08:00
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s6
; GFX10-DL-NEXT: v_dot8_u32_u4 v2, s1, s0, v0
; GFX10-DL-NEXT: v_mov_b32_e32 v0, s4
; GFX10-DL-NEXT: v_mov_b32_e32 v1, s5
[AMDGPU] gfx10 tests. NFC. llvm-svn: 363946
2019-06-21 00:29:40 +08:00
; GFX10-DL-NEXT: global_store_dword v[0:1], v2, off
; GFX10-DL-NEXT: s_endpgm
[AMDGPU] Handle the idot8 pattern generated by FE. Summary: Different variants of idot8 codegen dag patterns are not generated by llvm-tablegen due to a huge increase in the compile time. Support the pattern that clang FE generates after reordering the additions in integer-dot8 source language pattern. Author: FarhanaAleen Reviewed By: arsenm Differential Revision: https://reviews.llvm.org/D53937 llvm-svn: 345902
2018-11-02 06:48:19 +08:00
i32 addrspace(1)* %v2addr,
i32 addrspace(1)* %dst) {
entry:
%v1 = load i32, i32 addrspace(1)* %v1addr, align 4
%v2 = load i32, i32 addrspace(1)* %v2addr, align 4
%and = and i32 %v1, 15
%and1 = and i32 %v2, 15
%mul1 = mul nuw nsw i32 %and1, %and
%shr = lshr i32 %v1, 4
%and2 = and i32 %shr, 15
%shr3 = lshr i32 %v2, 4
%and4 = and i32 %shr3, 15
%mul2 = mul nuw nsw i32 %and4, %and2
%shr6 = lshr i32 %v1, 8
%and7 = and i32 %shr6, 15
%shr8 = lshr i32 %v2, 8
%and9 = and i32 %shr8, 15
%mul3 = mul nuw nsw i32 %and9, %and7
%shr12 = lshr i32 %v1, 12
%and13 = and i32 %shr12, 15
%shr14 = lshr i32 %v2, 12
%and15 = and i32 %shr14, 15
%mul4 = mul nuw nsw i32 %and15, %and13
%shr18 = lshr i32 %v1, 16
%and19 = and i32 %shr18, 15
%shr20 = lshr i32 %v2, 16
%and21 = and i32 %shr20, 15
%mul5 = mul nuw nsw i32 %and21, %and19
%shr24 = lshr i32 %v1, 20
%and25 = and i32 %shr24, 15
%shr26 = lshr i32 %v2, 20
%and27 = and i32 %shr26, 15
%mul6 = mul nuw nsw i32 %and27, %and25
%shr30 = lshr i32 %v1, 24
%and31 = and i32 %shr30, 15
%shr32 = lshr i32 %v2, 24
%and33 = and i32 %shr32, 15
%mul7 = mul nuw nsw i32 %and33, %and31
%shr36 = lshr i32 %v1, 28
%shr37 = lshr i32 %v2, 28
%mul8 = mul nuw nsw i32 %shr37, %shr36
%acc = load i32, i32 addrspace(1)* %dst, align 4
%add1 = add i32 %mul1, %acc
%add2 = add i32 %add1, %mul8
%add3 = add i32 %add2, %mul2
%add4 = add i32 %add3, %mul3
%add5 = add i32 %add4, %mul4
%add6 = add i32 %add5, %mul5
%add7 = add i32 %add6, %mul6
%add8 = add i32 %add7, %mul7
store i32 %add8, i32 addrspace(1)* %dst, align 4
ret void
}