[llvm-branch-commits] [llvm] [AMDGPU] Handle FREEZE like COPY in SIFixSGPRCopies (PR #240474)

Mitch Briles via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Mon Nov 9 13:47:03 PST 2026


https://github.com/MitchBriles updated https://github.com/llvm/llvm-project/pull/240474

>From 7bde27c7c0028ba871fc57f6281020dc8f8abfd9 Mon Sep 17 00:00:00 2001
From: Mitch <mitchbriles at gmail.com>
Date: Wed, 28 Oct 2026 14:36:41 -0600
Subject: [PATCH 1/2] Precommit SIFixSGPRCopies test

Co-Authored-By: Claude Opus 5.5 <noreply at anthropic.com>
---
 .../CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir | 58 +++++++++++++++++++
 1 file changed, 58 insertions(+)
 create mode 100644 llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir

diff --git a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir
new file mode 100644
index 0000000000000..30f75d701b64d
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir
@@ -0,0 +1,58 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6
+# RUN: llc -mtriple=amdgpu11.00-amd-amdhsa -run-pass=si-fix-sgpr-copies -o - %s | FileCheck %s
+
+---
+name: freeze_vgpr_to_sgpr
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0, $vgpr1_vgpr2
+
+    ; CHECK-LABEL: name: freeze_vgpr_to_sgpr
+    ; CHECK: liveins: $vgpr0, $vgpr1_vgpr2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
+    ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:sreg_32 = FREEZE [[COPY]]
+    ; CHECK-NEXT: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[FREEZE]], [[FREEZE]], implicit-def dead $scc
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_ADD_I32_]]
+    ; CHECK-NEXT: GLOBAL_STORE_DWORD [[COPY1]], [[COPY2]], 0, 0, implicit $exec
+    ; CHECK-NEXT: S_ENDPGM 0
+    %0:vgpr_32 = COPY $vgpr0
+    %1:vreg_64 = COPY $vgpr1_vgpr2
+    %2:sreg_32 = FREEZE %0
+    %3:sreg_32 = S_ADD_I32 %2, %2, implicit-def dead $scc
+    %4:vgpr_32 = COPY %3
+    GLOBAL_STORE_DWORD %1, %4, 0, 0, implicit $exec
+    S_ENDPGM 0
+...
+---
+name: freeze_sgpr_moved_to_valu
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0, $sgpr0, $vgpr1_vgpr2
+
+    ; CHECK-LABEL: name: freeze_sgpr_moved_to_valu
+    ; CHECK: liveins: $vgpr0, $sgpr0, $vgpr1_vgpr2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+    ; CHECK-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]], [[COPY2]], 0, implicit $exec
+    ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:sreg_32 = FREEZE [[V_ADD_U32_e64_]]
+    ; CHECK-NEXT: [[S_MUL_I32_:%[0-9]+]]:sreg_32 = S_MUL_I32 [[FREEZE]], [[FREEZE]]
+    ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[S_MUL_I32_]]
+    ; CHECK-NEXT: GLOBAL_STORE_DWORD [[COPY1]], [[COPY3]], 0, 0, implicit $exec
+    ; CHECK-NEXT: S_ENDPGM 0
+    %0:vgpr_32 = COPY $vgpr0
+    %1:vreg_64 = COPY $vgpr1_vgpr2
+    %2:sreg_32 = COPY $sgpr0
+    %3:sreg_32 = COPY %0
+    %4:sreg_32 = S_ADD_I32 %3, %2, implicit-def dead $scc
+    %5:sreg_32 = FREEZE %4
+    %6:sreg_32 = S_MUL_I32 %5, %5
+    %7:vgpr_32 = COPY %6
+    GLOBAL_STORE_DWORD %1, %7, 0, 0, implicit $exec
+    S_ENDPGM 0
+...

>From ef0894fe5339c5f2ac97a23211d1aa8c64301031 Mon Sep 17 00:00:00 2001
From: Mitch <mitchbriles at gmail.com>
Date: Wed, 28 Oct 2026 14:37:02 -0600
Subject: [PATCH 2/2] [AMDGPU] Handle FREEZE like COPY in SIFixSGPRCopies

An SGPR FREEZE of a VGPR value cannot be lowered after register
allocation. We should treat FREEZE like the other single-input pseudos
(WQM and friends) when fixing SGPR copies and when moving instructions to
the VALU, so that its result is moved to a VGPR along with its users.

Co-Authored-By: Claude Opus 5.5 <noreply at anthropic.com>
---
 llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp          |  1 +
 llvm/lib/Target/AMDGPU/SIInstrInfo.cpp              |  5 ++++-
 llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir | 12 ++++++------
 3 files changed, 11 insertions(+), 7 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp b/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
index c7d9b534a9225..51cf2e4eab798 100644
--- a/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
@@ -682,6 +682,7 @@ bool SIFixSGPRCopies::run(MachineFunction &MF) {
       case AMDGPU::STRICT_WQM:
       case AMDGPU::SOFT_WQM:
       case AMDGPU::STRICT_WWM:
+      case AMDGPU::FREEZE:
       case AMDGPU::INSERT_SUBREG:
       case AMDGPU::PHI:
       case AMDGPU::REG_SEQUENCE: {
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 5023e9d83efb2..2a3b7a66b5330 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -6369,6 +6369,7 @@ unsigned SIInstrInfo::getVALUOp(unsigned Opc) const {
   case AMDGPU::PHI: return AMDGPU::PHI;
   case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
   case AMDGPU::WQM: return AMDGPU::WQM;
+  case AMDGPU::FREEZE: return AMDGPU::FREEZE;
   case AMDGPU::SOFT_WQM: return AMDGPU::SOFT_WQM;
   case AMDGPU::STRICT_WWM: return AMDGPU::STRICT_WWM;
   case AMDGPU::STRICT_WQM: return AMDGPU::STRICT_WQM;
@@ -9969,6 +9970,7 @@ void SIInstrInfo::addUsersToMoveToVALUWorklist(
     case AMDGPU::SOFT_WQM:
     case AMDGPU::STRICT_WWM:
     case AMDGPU::STRICT_WQM:
+    case AMDGPU::FREEZE:
     case AMDGPU::REG_SEQUENCE:
     case AMDGPU::PHI:
     case AMDGPU::INSERT_SUBREG:
@@ -10204,7 +10206,8 @@ const TargetRegisterClass *SIInstrInfo::getDestEquivalentVGPRClass(
   case AMDGPU::WQM:
   case AMDGPU::SOFT_WQM:
   case AMDGPU::STRICT_WWM:
-  case AMDGPU::STRICT_WQM: {
+  case AMDGPU::STRICT_WQM:
+  case AMDGPU::FREEZE: {
     const TargetRegisterClass *SrcRC = getOpRegClass(Inst, 1);
     if (RI.isAGPRClass(SrcRC)) {
       if (RI.isAGPRClass(NewDstRC))
diff --git a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir
index 30f75d701b64d..820204911fa24 100644
--- a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir
+++ b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-freeze.mir
@@ -13,9 +13,9 @@ body: |
     ; CHECK-NEXT: {{  $}}
     ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
-    ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:sreg_32 = FREEZE [[COPY]]
-    ; CHECK-NEXT: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[FREEZE]], [[FREEZE]], implicit-def dead $scc
-    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_ADD_I32_]]
+    ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:vgpr_32 = FREEZE [[COPY]]
+    ; CHECK-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[FREEZE]], [[FREEZE]], 0, implicit $exec
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_U32_e64_]]
     ; CHECK-NEXT: GLOBAL_STORE_DWORD [[COPY1]], [[COPY2]], 0, 0, implicit $exec
     ; CHECK-NEXT: S_ENDPGM 0
     %0:vgpr_32 = COPY $vgpr0
@@ -40,9 +40,9 @@ body: |
     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
     ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
     ; CHECK-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]], [[COPY2]], 0, implicit $exec
-    ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:sreg_32 = FREEZE [[V_ADD_U32_e64_]]
-    ; CHECK-NEXT: [[S_MUL_I32_:%[0-9]+]]:sreg_32 = S_MUL_I32 [[FREEZE]], [[FREEZE]]
-    ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[S_MUL_I32_]]
+    ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:vgpr_32 = FREEZE [[V_ADD_U32_e64_]]
+    ; CHECK-NEXT: [[V_MUL_LO_U32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_LO_U32_e64 [[FREEZE]], [[FREEZE]], implicit $exec
+    ; CHECK-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[V_MUL_LO_U32_e64_]]
     ; CHECK-NEXT: GLOBAL_STORE_DWORD [[COPY1]], [[COPY3]], 0, 0, implicit $exec
     ; CHECK-NEXT: S_ENDPGM 0
     %0:vgpr_32 = COPY $vgpr0



More information about the llvm-branch-commits mailing list