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/GlobalISel/llvm.amdgcn.raw.buffer.stor...

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

309 lines
23 KiB
LLVM
Raw Normal View History

Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -stop-after=instruction-select -o - %s | FileCheck %s
; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx810 -stop-after=instruction-select -o - %s | FileCheck %s
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_f32(<4 x i32> inreg %rsrc, float %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_f32
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_X_OFFEN_exact [[COPY4]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom "TargetCustom7", align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.f32(float %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__voffset_4095__sgpr_soffset_f32(<4 x i32> inreg %rsrc, float %val, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__voffset_4095__sgpr_soffset_f32
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: BUFFER_STORE_FORMAT_X_OFFSET_exact [[COPY4]], [[REG_SEQUENCE]], [[COPY5]], 4095, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom "TargetCustom7" + 4095, align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.f32(float %val, <4 x i32> %rsrc, i32 4095, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32(<4 x i32> inreg %rsrc, <2 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XY_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom "TargetCustom7", align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.v2f32(<2 x float> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v3f32(<4 x i32> inreg %rsrc, <3 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v3f32
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_96 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY6]], %subreg.sub2
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XYZ_OFFEN_exact [[REG_SEQUENCE1]], [[COPY7]], [[REG_SEQUENCE]], [[COPY8]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 12 into custom "TargetCustom7", align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.v3f32(<3 x float> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f32(<4 x i32> inreg %rsrc, <4 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f32
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY8:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY9:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY6]], %subreg.sub2, [[COPY7]], %subreg.sub3
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XYZW_OFFEN_exact [[REG_SEQUENCE1]], [[COPY8]], [[REG_SEQUENCE]], [[COPY9]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16 into custom "TargetCustom7", align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f32(<4 x i32> %rsrc, <4 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f32
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr5
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr6
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr7
; CHECK: [[COPY8:%[0-9]+]]:vgpr_32 = COPY $vgpr8
; CHECK: [[COPY9:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY6]], %subreg.sub2, [[COPY7]], %subreg.sub3
; CHECK: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY10]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY10]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY10]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub1, implicit $exec
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[COPY11]], implicit $exec
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE4:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XYZW_OFFEN_exact [[REG_SEQUENCE1]], [[COPY8]], [[REG_SEQUENCE4]], [[COPY9]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16 into custom "TargetCustom7", align 1, addrspace 4)
AMDGPU/GlobalISel: Support wave32 waterfall loops llvm-svn: 373714
2019-10-04 16:35:35 +08:00
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_soffset4095(<4 x i32> inreg %rsrc, <2 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_soffset4095
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4095
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XY_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom "TargetCustom7", align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.v2f32(<2 x float> %val, <4 x i32> %rsrc, i32 %voffset, i32 4095, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_soffset4096(<4 x i32> inreg %rsrc, <2 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_soffset4096
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XY_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom "TargetCustom7", align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
call void @llvm.amdgcn.raw.buffer.store.format.v2f32(<2 x float> %val, <4 x i32> %rsrc, i32 %voffset, i32 4096, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_voffset_add_16(<4 x i32> inreg %rsrc, <2 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_voffset_add_16
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XY_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE]], [[COPY7]], 16, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom "TargetCustom7" + 16, align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
%voffset.add = add i32 %voffset, 16
call void @llvm.amdgcn.raw.buffer.store.format.v2f32(<2 x float> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_voffset_add_4095(<4 x i32> inreg %rsrc, <2 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_voffset_add_4095
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
Revert "Revert "[MIR] Target specific MIR formating and parsing"" There was an unguarded dereference of MF in a function that permitted nullptr. Fixed This reverts commit 71d64f72f934631aa2f12b9542c23f74f256f494.
2020-01-09 12:02:37 +08:00
; CHECK: BUFFER_STORE_FORMAT_XY_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE]], [[COPY7]], 4095, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom "TargetCustom7" + 4095, align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
%voffset.add = add i32 %voffset, 4095
call void @llvm.amdgcn.raw.buffer.store.format.v2f32(<2 x float> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
ret void
}
define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_voffset_add_4096(<4 x i32> inreg %rsrc, <2 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f32_voffset_add_4096
; CHECK: bb.1 (%ir-block.0):
; CHECK: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
AMDGPU: Relax 32-bit SGPR register class Mostly use SReg_32 instead of SReg_32_XM0 for arbitrary values. This will allow the register coalescer to do a better job eliminating copies to m0. For GlobalISel, as a terrible hack, use SGPR_32 for things that should use SCC until booleans are solved. llvm-svn: 375267
2019-10-19 02:26:37 +08:00
; CHECK: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
; CHECK: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr4
; CHECK: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr5
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
; CHECK: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
AMDGPU: Rename add/sub with carry out instructions The hardware has created a real mess in the naming for add/sub, which have been renamed basically every generation. Switch the carry out pseudos to have the gfx9/gfx10 names. We were using the original SI/CI v_add_i32/v_sub_i32 names. Later targets reintroduced these names as carryless instructions with a saturating clamp bit, which we do not define. Do this rename so we can unambiguously add these missing instructions. The carry-in versions should also be renamed, but at least those had a consistent _u32 name to begin with. The 16-bit instructions were also renamed, but aren't ambiguous. This does regress assembler error message quality in some cases. In mismatched wave32/wave64 situations, this will switch from "unsupported instruction" to "invalid operand", with the error pointing at the wrong position. I couldn't quite follow how the assembler selects these, but the previous behavior seemed accidental to me. It looked like there was a partial attempt to handle this which was never completed (i.e. there is an AMDGPUOperand::isBoolReg but it isn't used for anything).
2020-07-14 21:18:36 +08:00
; CHECK: %13:vgpr_32, dead %17:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY6]], [[COPY8]], 0, implicit $exec
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: BUFFER_STORE_FORMAT_XY_OFFEN_exact [[REG_SEQUENCE1]], %13, [[REG_SEQUENCE]], [[COPY7]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom "TargetCustom7" + 4096, align 1, addrspace 4)
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: S_ENDPGM 0
%voffset.add = add i32 %voffset, 4096
call void @llvm.amdgcn.raw.buffer.store.format.v2f32(<2 x float> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
ret void
}
; Check what happens with offset add inside a waterfall loop
define amdgpu_ps void @raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f32_add_4096(<4 x i32> %rsrc, <4 x float> %val, i32 %voffset, i32 inreg %soffset) {
; CHECK-LABEL: name: raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f32_add_4096
; CHECK: bb.1 (%ir-block.0):
; CHECK: successors: %bb.2(0x80000000)
; CHECK: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
; CHECK: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
; CHECK: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
; CHECK: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
; CHECK: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr5
; CHECK: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr6
; CHECK: [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr7
; CHECK: [[COPY8:%[0-9]+]]:vgpr_32 = COPY $vgpr8
; CHECK: [[COPY9:%[0-9]+]]:sreg_32 = COPY $sgpr2
; CHECK: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
; CHECK: [[REG_SEQUENCE1:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY6]], %subreg.sub2, [[COPY7]], %subreg.sub3
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
; CHECK: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
AMDGPU: Rename add/sub with carry out instructions The hardware has created a real mess in the naming for add/sub, which have been renamed basically every generation. Switch the carry out pseudos to have the gfx9/gfx10 names. We were using the original SI/CI v_add_i32/v_sub_i32 names. Later targets reintroduced these names as carryless instructions with a saturating clamp bit, which we do not define. Do this rename so we can unambiguously add these missing instructions. The carry-in versions should also be renamed, but at least those had a consistent _u32 name to begin with. The 16-bit instructions were also renamed, but aren't ambiguous. This does regress assembler error message quality in some cases. In mismatched wave32/wave64 situations, this will switch from "unsupported instruction" to "invalid operand", with the error pointing at the wrong position. I couldn't quite follow how the assembler selects these, but the previous behavior seemed accidental to me. It looked like there was a partial attempt to handle this which was never completed (i.e. there is an AMDGPUOperand::isBoolReg but it isn't used for anything).
2020-07-14 21:18:36 +08:00
; CHECK: %15:vgpr_32, dead %35:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY8]], [[COPY10]], 0, implicit $exec
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
; CHECK: [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
; CHECK: bb.2:
; CHECK: successors: %bb.3(0x40000000), %bb.2(0x40000000)
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub1, implicit $exec
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY11]], implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY12]].sub0, implicit $exec
; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY12]].sub1, implicit $exec
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[COPY12]], implicit $exec
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
AMDGPU: Use SGPR_128 instead of SReg_128 for vregs SGPR_128 only includes the real allocatable SGPRs, and SReg_128 adds the additional non-allocatable TTMP registers. There's no point in allocating SReg_128 vregs. This shrinks the size of the classes regalloc needs to consider, which is usually good. llvm-svn: 374284
2019-10-10 15:11:33 +08:00
; CHECK: [[REG_SEQUENCE4:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
AMDGPU/GlobalISel: Move llvm.amdgcn.raw.buffer.store handling Treat this the same way as loads. There's less value to the intermediate nodes, but it's good to be consistent.
2020-01-14 09:39:09 +08:00
; CHECK: BUFFER_STORE_FORMAT_XYZW_OFFEN_exact [[REG_SEQUENCE1]], %15, [[REG_SEQUENCE4]], [[COPY9]], 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16 into custom "TargetCustom7" + 4096, align 1, addrspace 4)
AMDGPU/GlobalISel: Support wave32 waterfall loops llvm-svn: 373714
2019-10-04 16:35:35 +08:00
; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
Reapply r372285 "GlobalISel: Don't materialize immarg arguments to intrinsics" This reverts r372314, reapplying r372285 and the commits which depend on it (r372286-r372293, and r372296-r372297) This was missing one switch to getTargetConstant in an untested case. llvm-svn: 372338
2019-09-20 00:26:14 +08:00
; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
; CHECK: S_CBRANCH_EXECNZ %bb.2, implicit $exec
; CHECK: bb.3:
; CHECK: successors: %bb.4(0x80000000)
; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
; CHECK: bb.4:
; CHECK: S_ENDPGM 0
%voffset.add = add i32 %voffset, 4096
call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
ret void
}
declare void @llvm.amdgcn.raw.buffer.store.format.f32(float, <4 x i32>, i32, i32, i32 immarg)
declare void @llvm.amdgcn.raw.buffer.store.format.v2f32(<2 x float>, <4 x i32>, i32, i32, i32 immarg)
declare void @llvm.amdgcn.raw.buffer.store.format.v3f32(<3 x float>, <4 x i32>, i32, i32, i32 immarg)
declare void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float>, <4 x i32>, i32, i32, i32 immarg)