[llvm] [AMDGPU] SILowerI1Copies: clear kill flags on COPY (PR #65883)
Carl Ritson via llvm-commits
llvm-commits at lists.llvm.org
Sun Sep 10 02:25:24 PDT 2023
https://github.com/perlfu updated https://github.com/llvm/llvm-project/pull/65883:
>From 530cd6e62da8c63ef64ba575dce76d76246f670d Mon Sep 17 00:00:00 2001
From: Carl Ritson <carl.ritson at amd.com>
Date: Sun, 10 Sep 2023 17:03:25 +0900
Subject: [PATCH 1/2] [AMDGPU] SILowerI1Copies: clear kill flags on COPY
Clear kill flags on COPY source as it will be reused.
---
llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp | 3 +
.../AMDGPU/lower-i1-copies-clear-kills.mir | 157 ++++++++++++++++++
2 files changed, 160 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir
diff --git a/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp b/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
index d4f0906f020ab5f..9e20adfeac884f1 100644
--- a/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
+++ b/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
@@ -708,6 +708,9 @@ bool SILowerI1Copies::lowerCopiesToI1() {
.addImm(0);
MI.getOperand(1).setReg(TmpReg);
SrcReg = TmpReg;
+ } else {
+ // SrcReg needs to be live beyond copy.
+ MI.clearKillInfo();
}
// Defs in a loop that are observed outside the loop must be transformed
diff --git a/llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir b/llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir
new file mode 100644
index 000000000000000..73537f5e2cdaed7
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir
@@ -0,0 +1,157 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 3
+# RUN: llc -run-pass=si-i1-copies -mtriple=amdgcn--amdpal -mcpu=gfx1030 -verify-machineinstrs -o - %s | FileCheck %s
+
+# Make sure that kill flag is clear on %23 to %0 copy when %23 is reused.
+
+---
+name: _amdgpu_cs_main
+tracksRegLiveness: true
+body: |
+ ; CHECK-LABEL: name: _amdgpu_cs_main
+ ; CHECK: bb.0:
+ ; CHECK-NEXT: successors: %bb.1(0x80000000)
+ ; CHECK-NEXT: liveins: $vgpr0, $vgpr1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+ ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+ ; CHECK-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+ ; CHECK-NEXT: [[V_CMP_NE_U32_e64_:%[0-9]+]]:sreg_32 = V_CMP_NE_U32_e64 [[COPY1]], [[S_MOV_B32_]], implicit $exec
+ ; CHECK-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sgpr_32 = S_MOV_B32 0
+ ; CHECK-NEXT: [[V_CMP_GT_F32_e64_:%[0-9]+]]:sreg_32 = nofpexcept V_CMP_GT_F32_e64 0, [[COPY]], 0, killed [[S_MOV_B32_1]], 0, implicit $mode, implicit $exec
+ ; CHECK-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32 = S_AND_B32 killed [[V_CMP_NE_U32_e64_]], killed [[V_CMP_GT_F32_e64_]], implicit-def dead $scc
+ ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[S_AND_B32_]]
+ ; CHECK-NEXT: [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 -1
+ ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[COPY2]]
+ ; CHECK-NEXT: [[S_XOR_B32_:%[0-9]+]]:sreg_32 = S_XOR_B32 [[COPY3]], killed [[S_MOV_B32_2]], implicit-def dead $scc
+ ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[S_XOR_B32_]]
+ ; CHECK-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]], implicit $exec
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.1:
+ ; CHECK-NEXT: successors: %bb.2(0x40000000), %bb.3(0x40000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[PHI:%[0-9]+]]:sreg_32 = PHI [[S_MOV_B32_]], %bb.0, %14, %bb.5
+ ; CHECK-NEXT: [[PHI1:%[0-9]+]]:sreg_32 = PHI [[S_MOV_B32_]], %bb.0, %16, %bb.5
+ ; CHECK-NEXT: [[PHI2:%[0-9]+]]:vgpr_32 = PHI [[COPY5]], %bb.0, %18, %bb.5
+ ; CHECK-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY [[COPY4]]
+ ; CHECK-NEXT: [[SI_IF:%[0-9]+]]:sreg_32 = SI_IF [[COPY6]], %bb.3, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: S_BRANCH %bb.2
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.2:
+ ; CHECK-NEXT: successors: %bb.3(0x80000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[V_OR_B32_e64_:%[0-9]+]]:vgpr_32 = V_OR_B32_e64 [[PHI1]], [[PHI2]], implicit $exec
+ ; CHECK-NEXT: [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 -1
+ ; CHECK-NEXT: [[S_OR_B32_:%[0-9]+]]:sreg_32 = S_OR_B32 [[S_AND_B32_]], $exec_lo, implicit-def $scc
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.3:
+ ; CHECK-NEXT: successors: %bb.4(0x40000000), %bb.5(0x40000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[PHI3:%[0-9]+]]:sreg_32 = PHI [[S_AND_B32_]], %bb.1, [[S_OR_B32_]], %bb.2
+ ; CHECK-NEXT: [[PHI4:%[0-9]+]]:vgpr_32 = PHI [[PHI2]], %bb.1, [[V_OR_B32_e64_]], %bb.2
+ ; CHECK-NEXT: SI_END_CF [[SI_IF]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: [[S_MOV_B32_4:%[0-9]+]]:sreg_32 = S_MOV_B32 -1
+ ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
+ ; CHECK-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY [[PHI3]]
+ ; CHECK-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[DEF]]
+ ; CHECK-NEXT: [[SI_IF1:%[0-9]+]]:sreg_32 = SI_IF [[COPY7]], %bb.5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: S_BRANCH %bb.4
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.4:
+ ; CHECK-NEXT: successors: %bb.5(0x80000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[S_MOV_B32_5:%[0-9]+]]:sreg_32 = S_MOV_B32 1
+ ; CHECK-NEXT: [[S_OR_B32_1:%[0-9]+]]:sreg_32 = S_OR_B32 [[PHI1]], killed [[S_MOV_B32_5]], implicit-def dead $scc
+ ; CHECK-NEXT: [[S_MOV_B32_6:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+ ; CHECK-NEXT: [[S_XOR_B32_1:%[0-9]+]]:sreg_32 = S_XOR_B32 $exec_lo, -1, implicit-def $scc
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.5:
+ ; CHECK-NEXT: successors: %bb.6(0x04000000), %bb.1(0x7c000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[PHI5:%[0-9]+]]:sreg_32 = PHI [[S_MOV_B32_4]], %bb.3, [[S_XOR_B32_1]], %bb.4
+ ; CHECK-NEXT: [[PHI6:%[0-9]+]]:vgpr_32 = PHI [[COPY8]], %bb.3, [[PHI4]], %bb.4
+ ; CHECK-NEXT: [[PHI7:%[0-9]+]]:sreg_32 = PHI [[DEF]], %bb.3, [[S_OR_B32_1]], %bb.4
+ ; CHECK-NEXT: SI_END_CF [[SI_IF1]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: [[COPY9:%[0-9]+]]:sreg_32 = COPY [[PHI5]]
+ ; CHECK-NEXT: [[SI_IF_BREAK:%[0-9]+]]:sreg_32 = SI_IF_BREAK [[COPY9]], [[PHI]], implicit-def dead $scc
+ ; CHECK-NEXT: SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: S_BRANCH %bb.6
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.6:
+ ; CHECK-NEXT: [[PHI8:%[0-9]+]]:sreg_32 = PHI [[SI_IF_BREAK]], %bb.5
+ ; CHECK-NEXT: SI_END_CF [[PHI8]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: S_ENDPGM 0
+ bb.0:
+ successors: %bb.1(0x80000000)
+ liveins: $vgpr0, $vgpr1
+
+ %17:vgpr_32 = COPY $vgpr1
+ %16:vgpr_32 = COPY $vgpr0
+ %19:sreg_32 = S_MOV_B32 0
+ %20:sreg_32 = V_CMP_NE_U32_e64 %16, %19, implicit $exec
+ %21:sgpr_32 = S_MOV_B32 0
+ %22:sreg_32 = nofpexcept V_CMP_GT_F32_e64 0, %17, 0, killed %21, 0, implicit $mode, implicit $exec
+ %23:sreg_32 = S_AND_B32 killed %20, killed %22, implicit-def dead $scc
+ %0:vreg_1 = COPY killed %23
+ %24:sreg_32 = S_MOV_B32 -1
+ %26:sreg_32 = COPY %0
+ %25:sreg_32 = S_XOR_B32 %26, killed %24, implicit-def dead $scc
+ %1:vreg_1 = COPY %25
+ %35:vgpr_32 = COPY %19, implicit $exec
+
+ bb.1:
+ successors: %bb.2(0x40000000), %bb.3(0x40000000)
+
+ %2:sreg_32 = PHI %19, %bb.0, %14, %bb.5
+ %3:sreg_32 = PHI %19, %bb.0, %12, %bb.5
+ %4:vgpr_32 = PHI %35, %bb.0, %11, %bb.5
+ %27:sreg_32 = COPY %1
+ %5:sreg_32 = SI_IF %27, %bb.3, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ S_BRANCH %bb.2
+
+ bb.2:
+ successors: %bb.3(0x80000000)
+
+ %6:vgpr_32 = V_OR_B32_e64 %3, %4, implicit $exec
+ %28:sreg_32 = S_MOV_B32 -1
+ %36:vreg_1 = COPY %28, implicit $exec
+
+ bb.3:
+ successors: %bb.4(0x40000000), %bb.5(0x40000000)
+
+ %7:vgpr_32 = PHI %4, %bb.1, %6, %bb.2
+ %8:vreg_1 = PHI %0, %bb.1, %36, %bb.2
+ SI_END_CF %5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %30:sreg_32 = S_MOV_B32 -1
+ %29:sreg_32 = IMPLICIT_DEF
+ %31:sreg_32 = COPY %8
+ %37:vgpr_32 = COPY %29
+ %38:vreg_1 = COPY %30, implicit $exec
+ %9:sreg_32 = SI_IF %31, %bb.5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ S_BRANCH %bb.4
+
+ bb.4:
+ successors: %bb.5(0x80000000)
+
+ %33:sreg_32 = S_MOV_B32 1
+ %10:sreg_32 = S_OR_B32 %3, killed %33, implicit-def dead $scc
+ %32:sreg_32 = S_MOV_B32 0
+ %39:vreg_1 = COPY %32, implicit $exec
+
+ bb.5:
+ successors: %bb.6(0x04000000), %bb.1(0x7c000000)
+
+ %11:vgpr_32 = PHI %37, %bb.3, %7, %bb.4
+ %12:sreg_32 = PHI %29, %bb.3, %10, %bb.4
+ %13:vreg_1 = PHI %38, %bb.3, %39, %bb.4
+ SI_END_CF %9, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %34:sreg_32 = COPY %13
+ %14:sreg_32 = SI_IF_BREAK %34, %2, implicit-def dead $scc
+ SI_LOOP %14, %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ S_BRANCH %bb.6
+
+ bb.6:
+ %15:sreg_32 = PHI %14, %bb.5
+ SI_END_CF %15, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ S_ENDPGM 0
+
+...
>From 8ceed89eaf573e33f3c01ad3a353c5a35c68c243 Mon Sep 17 00:00:00 2001
From: Carl Ritson <carl.ritson at amd.com>
Date: Sun, 10 Sep 2023 18:21:11 +0900
Subject: [PATCH 2/2] PR #65883 - Address Reviewer Comments
- Change from clearKillInfo to getOperand(1).setIsKill(false)
- Regenerate test with reduced register numbers
---
llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp | 2 +-
.../AMDGPU/lower-i1-copies-clear-kills.mir | 84 +++++++++----------
2 files changed, 43 insertions(+), 43 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp b/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
index 9e20adfeac884f1..a11f02e7db3504b 100644
--- a/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
+++ b/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
@@ -710,7 +710,7 @@ bool SILowerI1Copies::lowerCopiesToI1() {
SrcReg = TmpReg;
} else {
// SrcReg needs to be live beyond copy.
- MI.clearKillInfo();
+ MI.getOperand(1).setIsKill(false);
}
// Defs in a loop that are observed outside the loop must be transformed
diff --git a/llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir b/llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir
index 73537f5e2cdaed7..c5e2ba5d8c7cba3 100644
--- a/llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir
+++ b/llvm/test/CodeGen/AMDGPU/lower-i1-copies-clear-kills.mir
@@ -84,74 +84,74 @@ body: |
successors: %bb.1(0x80000000)
liveins: $vgpr0, $vgpr1
- %17:vgpr_32 = COPY $vgpr1
- %16:vgpr_32 = COPY $vgpr0
- %19:sreg_32 = S_MOV_B32 0
- %20:sreg_32 = V_CMP_NE_U32_e64 %16, %19, implicit $exec
- %21:sgpr_32 = S_MOV_B32 0
- %22:sreg_32 = nofpexcept V_CMP_GT_F32_e64 0, %17, 0, killed %21, 0, implicit $mode, implicit $exec
- %23:sreg_32 = S_AND_B32 killed %20, killed %22, implicit-def dead $scc
- %0:vreg_1 = COPY killed %23
- %24:sreg_32 = S_MOV_B32 -1
- %26:sreg_32 = COPY %0
- %25:sreg_32 = S_XOR_B32 %26, killed %24, implicit-def dead $scc
- %1:vreg_1 = COPY %25
- %35:vgpr_32 = COPY %19, implicit $exec
+ %0:vgpr_32 = COPY $vgpr1
+ %1:vgpr_32 = COPY $vgpr0
+ %2:sreg_32 = S_MOV_B32 0
+ %3:sreg_32 = V_CMP_NE_U32_e64 %1, %2, implicit $exec
+ %4:sgpr_32 = S_MOV_B32 0
+ %5:sreg_32 = nofpexcept V_CMP_GT_F32_e64 0, %0, 0, killed %4, 0, implicit $mode, implicit $exec
+ %6:sreg_32 = S_AND_B32 killed %3, killed %5, implicit-def dead $scc
+ %7:vreg_1 = COPY killed %6
+ %8:sreg_32 = S_MOV_B32 -1
+ %9:sreg_32 = COPY %7
+ %10:sreg_32 = S_XOR_B32 %9, killed %8, implicit-def dead $scc
+ %11:vreg_1 = COPY %10
+ %12:vgpr_32 = COPY %2, implicit $exec
bb.1:
successors: %bb.2(0x40000000), %bb.3(0x40000000)
- %2:sreg_32 = PHI %19, %bb.0, %14, %bb.5
- %3:sreg_32 = PHI %19, %bb.0, %12, %bb.5
- %4:vgpr_32 = PHI %35, %bb.0, %11, %bb.5
- %27:sreg_32 = COPY %1
- %5:sreg_32 = SI_IF %27, %bb.3, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %13:sreg_32 = PHI %2, %bb.0, %14, %bb.5
+ %15:sreg_32 = PHI %2, %bb.0, %16, %bb.5
+ %17:vgpr_32 = PHI %12, %bb.0, %18, %bb.5
+ %19:sreg_32 = COPY %11
+ %20:sreg_32 = SI_IF %19, %bb.3, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
S_BRANCH %bb.2
bb.2:
successors: %bb.3(0x80000000)
- %6:vgpr_32 = V_OR_B32_e64 %3, %4, implicit $exec
- %28:sreg_32 = S_MOV_B32 -1
- %36:vreg_1 = COPY %28, implicit $exec
+ %21:vgpr_32 = V_OR_B32_e64 %15, %17, implicit $exec
+ %22:sreg_32 = S_MOV_B32 -1
+ %23:vreg_1 = COPY %22, implicit $exec
bb.3:
successors: %bb.4(0x40000000), %bb.5(0x40000000)
- %7:vgpr_32 = PHI %4, %bb.1, %6, %bb.2
- %8:vreg_1 = PHI %0, %bb.1, %36, %bb.2
- SI_END_CF %5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
- %30:sreg_32 = S_MOV_B32 -1
- %29:sreg_32 = IMPLICIT_DEF
- %31:sreg_32 = COPY %8
- %37:vgpr_32 = COPY %29
- %38:vreg_1 = COPY %30, implicit $exec
- %9:sreg_32 = SI_IF %31, %bb.5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %24:vgpr_32 = PHI %17, %bb.1, %21, %bb.2
+ %25:vreg_1 = PHI %7, %bb.1, %23, %bb.2
+ SI_END_CF %20, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %26:sreg_32 = S_MOV_B32 -1
+ %27:sreg_32 = IMPLICIT_DEF
+ %28:sreg_32 = COPY %25
+ %29:vgpr_32 = COPY %27
+ %30:vreg_1 = COPY %26, implicit $exec
+ %31:sreg_32 = SI_IF %28, %bb.5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
S_BRANCH %bb.4
bb.4:
successors: %bb.5(0x80000000)
- %33:sreg_32 = S_MOV_B32 1
- %10:sreg_32 = S_OR_B32 %3, killed %33, implicit-def dead $scc
- %32:sreg_32 = S_MOV_B32 0
- %39:vreg_1 = COPY %32, implicit $exec
+ %32:sreg_32 = S_MOV_B32 1
+ %33:sreg_32 = S_OR_B32 %15, killed %32, implicit-def dead $scc
+ %34:sreg_32 = S_MOV_B32 0
+ %35:vreg_1 = COPY %34, implicit $exec
bb.5:
successors: %bb.6(0x04000000), %bb.1(0x7c000000)
- %11:vgpr_32 = PHI %37, %bb.3, %7, %bb.4
- %12:sreg_32 = PHI %29, %bb.3, %10, %bb.4
- %13:vreg_1 = PHI %38, %bb.3, %39, %bb.4
- SI_END_CF %9, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
- %34:sreg_32 = COPY %13
- %14:sreg_32 = SI_IF_BREAK %34, %2, implicit-def dead $scc
+ %18:vgpr_32 = PHI %29, %bb.3, %24, %bb.4
+ %16:sreg_32 = PHI %27, %bb.3, %33, %bb.4
+ %36:vreg_1 = PHI %30, %bb.3, %35, %bb.4
+ SI_END_CF %31, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %37:sreg_32 = COPY %36
+ %14:sreg_32 = SI_IF_BREAK %37, %13, implicit-def dead $scc
SI_LOOP %14, %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
S_BRANCH %bb.6
bb.6:
- %15:sreg_32 = PHI %14, %bb.5
- SI_END_CF %15, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %38:sreg_32 = PHI %14, %bb.5
+ SI_END_CF %38, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
S_ENDPGM 0
...
More information about the llvm-commits
mailing list