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/wwm-reserved.ll

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

196 lines
9.2 KiB
LLVM
Raw Normal View History

[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
; RUN: llc -O0 -march=amdgcn -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9,GFX9-O0 %s
; RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9,GFX9-O3 %s
[AMDGPU] Add some missing -LABEL checks
2020-06-19 22:33:48 +08:00
; GFX9-LABEL: {{^}}no_cfg:
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
define amdgpu_cs void @no_cfg(<4 x i32> inreg %tmp14) {
%tmp100 = call <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32> %tmp14, i32 0, i32 0, i32 0)
%tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
%tmp102 = extractelement <2 x i32> %tmp101, i32 0
%tmp103 = extractelement <2 x i32> %tmp101, i32 1
%tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
%tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
; GFX9: v_mov_b32_dpp v[[FIRST_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
; GFX9: v_add_u32_e32 v[[FIRST_ADD:[0-9]+]], v{{[0-9]+}}, v[[FIRST_MOV]]
; GFX9: v_mov_b32_e32 v[[FIRST:[0-9]+]], v[[FIRST_ADD]]
%tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
%tmp121 = add i32 %tmp105, %tmp120
%tmp122 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp121)
; GFX9: v_mov_b32_dpp v[[SECOND_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
; GFX9: v_add_u32_e32 v[[SECOND_ADD:[0-9]+]], v{{[0-9]+}}, v[[SECOND_MOV]]
; GFX9: v_mov_b32_e32 v[[SECOND:[0-9]+]], v[[SECOND_ADD]]
%tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
%tmp136 = add i32 %tmp107, %tmp135
%tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
; GFX9-O3: v_cmp_eq_u32_e32 vcc, v[[FIRST]], v[[SECOND]]
; GFX9-O0: v_cmp_eq_u32_e64 s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, v[[FIRST]], v[[SECOND]]
%tmp138 = icmp eq i32 %tmp122, %tmp137
%tmp139 = sext i1 %tmp138 to i32
%tmp140 = shl nsw i32 %tmp139, 1
%tmp141 = and i32 %tmp140, 2
%tmp145 = bitcast i32 %tmp141 to float
call void @llvm.amdgcn.raw.buffer.store.f32(float %tmp145, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
ret void
}
[AMDGPU] Add some missing -LABEL checks
2020-06-19 22:33:48 +08:00
; GFX9-LABEL: {{^}}cfg:
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
define amdgpu_cs void @cfg(<4 x i32> inreg %tmp14, i32 %arg) {
entry:
%tmp100 = call <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32> %tmp14, i32 0, i32 0, i32 0)
%tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
%tmp102 = extractelement <2 x i32> %tmp101, i32 0
%tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
; GFX9: v_mov_b32_dpp v[[FIRST_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
; GFX9: v_add_u32_e32 v[[FIRST_ADD:[0-9]+]], v{{[0-9]+}}, v[[FIRST_MOV]]
; GFX9: v_mov_b32_e32 v[[FIRST:[0-9]+]], v[[FIRST_ADD]]
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to 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-O0: buffer_store_dword v[[FIRST]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, 0 offset:[[FIRST_IMM_OFFSET:[0-9]+]]
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
%tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
%tmp121 = add i32 %tmp105, %tmp120
%tmp122 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp121)
%cond = icmp eq i32 %arg, 0
br i1 %cond, label %if, label %merge
if:
%tmp103 = extractelement <2 x i32> %tmp101, i32 1
%tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
; GFX9: v_mov_b32_dpp v[[SECOND_MOV:[0-9]+]], v{{[0-9]+}} row_bcast:31 row_mask:0xc bank_mask:0xf
; GFX9: v_add_u32_e32 v[[SECOND_ADD:[0-9]+]], v{{[0-9]+}}, v[[SECOND_MOV]]
; GFX9: v_mov_b32_e32 v[[SECOND:[0-9]+]], v[[SECOND_ADD]]
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to 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-O0: buffer_store_dword v[[SECOND]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, 0 offset:[[SECOND_IMM_OFFSET:[0-9]+]]
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
%tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
%tmp136 = add i32 %tmp107, %tmp135
%tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
br label %merge
merge:
%merge_value = phi i32 [ 0, %entry ], [%tmp137, %if ]
; GFX9-O3: v_cmp_eq_u32_e32 vcc, v[[FIRST]], v[[SECOND]]
[AMDGPU] Add Scratch Wave Offset to Scratch Buffer Descriptor in entry functions Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to 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-O0: buffer_load_dword v[[SECOND:[0-9]+]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, 0 offset:[[SECOND_IMM_OFFSET]]
; GFX9-O0: buffer_load_dword v[[FIRST:[0-9]+]], off, s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, 0 offset:[[FIRST_IMM_OFFSET]]
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
; GFX9-O0: v_cmp_eq_u32_e64 s{{\[}}{{[0-9]+}}:{{[0-9]+}}{{\]}}, v[[FIRST]], v[[SECOND]]
%tmp138 = icmp eq i32 %tmp122, %merge_value
%tmp139 = sext i1 %tmp138 to i32
%tmp140 = shl nsw i32 %tmp139, 1
%tmp141 = and i32 %tmp140, 2
%tmp145 = bitcast i32 %tmp141 to float
call void @llvm.amdgcn.raw.buffer.store.f32(float %tmp145, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
ret void
}
[AMDGPU] Add some missing -LABEL checks
2020-06-19 22:33:48 +08:00
; GFX9-LABEL: {{^}}called:
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
define i32 @called(i32 %a) noinline {
; GFX9: v_add_u32_e32 v1, v0, v0
%add = add i32 %a, %a
[AMDGPU] gfx1010: prefer V_MUL_LO_U32 over V_MUL_LO_I32 GFX10 deprecates v_mul_lo_i32 instruction, so choose u32 form for all targets. Differential Revision: https://reviews.llvm.org/D61525 llvm-svn: 360094
2019-05-07 06:27:05 +08:00
; GFX9: v_mul_lo_u32 v0, v1, v0
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
%mul = mul i32 %add, %a
; GFX9: v_sub_u32_e32 v0, v0, v1
%sub = sub i32 %mul, %add
ret i32 %sub
}
[AMDGPU] Add some missing -LABEL checks
2020-06-19 22:33:48 +08:00
; GFX9-LABEL: {{^}}call:
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
define amdgpu_kernel void @call(<4 x i32> inreg %tmp14, i32 inreg %arg) {
RegAllocFast: Record internal state based on register units Record internal state based on register units. This is often more efficient as there are typically fewer register units to update compared to iterating over all the aliases of a register. Original patch by Matthias Braun, but I've been rebasing and fixing it for almost 2 years and fixed a few bugs causing intermediate failures to make this patch independent of the changes in https://reviews.llvm.org/D52010.
2019-03-20 01:50:22 +08:00
; GFX9-O0: v_mov_b32_e32 v0, s0
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
; GFX9-O3: v_mov_b32_e32 v2, s0
; GFX9-NEXT: s_not_b64 exec, exec
RegAllocFast: Record internal state based on register units Record internal state based on register units. This is often more efficient as there are typically fewer register units to update compared to iterating over all the aliases of a register. Original patch by Matthias Braun, but I've been rebasing and fixing it for almost 2 years and fixed a few bugs causing intermediate failures to make this patch independent of the changes in https://reviews.llvm.org/D52010.
2019-03-20 01:50:22 +08:00
; GFX9-O0-NEXT: v_mov_b32_e32 v0, s1
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_not_b64 exec, exec
%tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %arg, i32 0)
; GFX9-O0: buffer_store_dword v0
; GFX9-O3: v_mov_b32_e32 v0, v2
; GFX9: s_swappc_b64
%tmp134 = call i32 @called(i32 %tmp107)
; GFX9-O0: buffer_load_dword v1
; GFX9-O3: v_mov_b32_e32 v1, v0
; GFX9-O0: v_add_u32_e32 v0, v0, v1
; GFX9-O3: v_add_u32_e32 v1, v1, v2
%tmp136 = add i32 %tmp134, %tmp107
%tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
; GFX9-O0: buffer_store_dword v2
; GFX9-O3: buffer_store_dword v0
call void @llvm.amdgcn.raw.buffer.store.i32(i32 %tmp137, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
ret void
}
[AMDGPU] Add some missing -LABEL checks
2020-06-19 22:33:48 +08:00
; GFX9-LABEL: {{^}}called_i64:
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
define i64 @called_i64(i64 %a) noinline {
%add = add i64 %a, %a
%mul = mul i64 %add, %a
%sub = sub i64 %mul, %add
ret i64 %sub
}
[AMDGPU] Add some missing -LABEL checks
2020-06-19 22:33:48 +08:00
; GFX9-LABEL: {{^}}call_i64:
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
define amdgpu_kernel void @call_i64(<4 x i32> inreg %tmp14, i64 inreg %arg) {
; GFX9-O0: v_mov_b32_e32 v0, s0
; GFX9-O0: v_mov_b32_e32 v1, s1
; GFX9-O3: v_mov_b32_e32 v7, s1
; GFX9-O3: v_mov_b32_e32 v6, s0
; GFX9-NEXT: s_not_b64 exec, exec
; GFX9-O0-NEXT: v_mov_b32_e32 v0, s2
; GFX9-O0-NEXT: v_mov_b32_e32 v1, s3
; GFX9-O3-NEXT: v_mov_b32_e32 v6, 0
; GFX9-O3-NEXT: v_mov_b32_e32 v7, 0
; GFX9-NEXT: s_not_b64 exec, exec
%tmp107 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %arg, i64 0)
; GFX9-O0: buffer_store_dword v0
; GFX9-O0: buffer_store_dword v1
; GFX9: s_swappc_b64
%tmp134 = call i64 @called_i64(i64 %tmp107)
RegAllocFast: Record internal state based on register units Record internal state based on register units. This is often more efficient as there are typically fewer register units to update compared to iterating over all the aliases of a register. Original patch by Matthias Braun, but I've been rebasing and fixing it for almost 2 years and fixed a few bugs causing intermediate failures to make this patch independent of the changes in https://reviews.llvm.org/D52010.
2019-03-20 01:50:22 +08:00
; GFX9-O0: buffer_load_dword v4
; GFX9-O0: buffer_load_dword v5
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
%tmp136 = add i64 %tmp134, %tmp107
%tmp137 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp136)
%tmp138 = bitcast i64 %tmp137 to <2 x i32>
call void @llvm.amdgcn.raw.buffer.store.v2i32(<2 x i32> %tmp138, <4 x i32> %tmp14, i32 4, i32 0, i32 0)
ret void
}
[AMDGPU] Add some missing -LABEL checks
2020-06-19 22:33:48 +08:00
; GFX9-LABEL: {{^}}_amdgpu_cs_main:
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
define amdgpu_cs void @_amdgpu_cs_main(<4 x i32> inreg %desc, i32 %index) {
%tmp17 = shl i32 %index, 5
; GFX9: buffer_load_dwordx4
%tmp18 = tail call <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32> %desc, i32 %tmp17, i32 0)
%.i0.upto1.bc = bitcast <4 x i32> %tmp18 to <2 x i64>
%tmp19 = or i32 %tmp17, 16
; GFX9: buffer_load_dwordx2
%tmp20 = tail call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %desc, i32 %tmp19, i32 0)
%.i0.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 0
%tmp22 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i0.upto1.extract, i64 9223372036854775807)
%tmp97 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp22)
%.i1.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 1
%tmp99 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i1.upto1.extract, i64 9223372036854775807)
%tmp174 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp99)
%.i25 = bitcast <2 x i32> %tmp20 to i64
%tmp176 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i25, i64 9223372036854775807)
%tmp251 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp176)
%.cast = bitcast i64 %tmp97 to <2 x float>
%.cast6 = bitcast i64 %tmp174 to <2 x float>
%.cast7 = bitcast i64 %tmp251 to <2 x float>
%tmp254 = shufflevector <2 x float> %.cast, <2 x float> %.cast6, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
; GFX9: buffer_store_dwordx4
tail call void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float> %tmp254, <4 x i32> %desc, i32 %tmp17, i32 0, i32 0)
; GFX9: buffer_store_dwordx2
tail call void @llvm.amdgcn.raw.buffer.store.v2f32(<2 x float> %.cast7, <4 x i32> %desc, i32 %tmp19, i32 0, i32 0)
ret void
}
declare i32 @llvm.amdgcn.wwm.i32(i32)
declare i64 @llvm.amdgcn.wwm.i64(i64)
declare i32 @llvm.amdgcn.set.inactive.i32(i32, i32)
declare i64 @llvm.amdgcn.set.inactive.i64(i64, i64)
declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32, i32, i32, i1)
declare <2 x float> @llvm.amdgcn.raw.buffer.load.v2f32(<4 x i32>, i32, i32, i32)
declare void @llvm.amdgcn.raw.buffer.store.f32(float, <4 x i32>, i32, i32, i32)
declare void @llvm.amdgcn.raw.buffer.store.i32(i32, <4 x i32>, i32, i32, i32)
declare void @llvm.amdgcn.raw.buffer.store.v2i32(<2 x i32>, <4 x i32>, i32, i32, i32)
declare void @llvm.amdgcn.raw.buffer.store.v2f32(<2 x float>, <4 x i32>, i32, i32, i32)
AMDGPU: Decompose all values to 32-bit pieces for calling conventions This is the more natural lowering, and presents more opportunities to reduce 64-bit ops to 32-bit. This should also help avoid issues graphics shaders have had with 64-bit values, and simplify argument lowering in globalisel. llvm-svn: 366578
2019-07-19 21:57:44 +08:00
declare void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float>, <4 x i32>, i32, i32, i32)
[AMDGPU] Pre-allocate WWM registers to reduce VGPR pressure. This change incorporates an effort by Connor Abbot to change how we deal with WWM operations potentially trashing valid values in inactive lanes. Previously, the SIFixWWMLiveness pass would work out which registers were being trashed within WWM regions, and ensure that the register allocator did not have any values it was depending on resident in those registers if the WWM section would trash them. This worked perfectly well, but would cause sometimes severe register pressure when the WWM section resided before divergent control flow (or at least that is where I mostly observed it). This fix instead runs through the WWM sections and pre allocates some registers for WWM. It then reserves these registers so that the register allocator cannot use them. This results in a significant register saving on some WWM shaders I'm working with (130 -> 104 VGPRs, with just this change!). Differential Revision: https://reviews.llvm.org/D59295 llvm-svn: 357400
2019-04-01 23:19:52 +08:00
declare <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32>, i32, i32)
declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)