[llvm] ed87f0a - [AMDGPU] Visit all PHIs in each call to optimizeLiveType (#147522)

[via llvm-commits]

Author: Pierre van Houtryve
Date: 2025-07-10T09:29:48+02:00
New Revision: ed87f0afba24d56b509dfc21298b2700d886912b

URL: https://github.com/llvm/llvm-project/commit/ed87f0afba24d56b509dfc21298b2700d886912b
DIFF: https://github.com/llvm/llvm-project/commit/ed87f0afba24d56b509dfc21298b2700d886912b.diff

LOG: [AMDGPU] Visit all PHIs in each call to optimizeLiveType (#147522)

Make the Visited set a local variable, otherwise we can reject a PHI
(those that do not have a zeroinitializer constant) but mark it as
visited,
and the rest of the function thinks the PHI is ok when it isn't.
This is a bit crude but it's the only fix that consistently worked in my
testing.

Fixes SWDEV-541767

Added: 
    llvm/test/CodeGen/AMDGPU/amdgpu-late-codegenprepare-crash-splat.ll

Modified: 
    llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp
index df976cf3f7fdb..523c66c72273c 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp
@@ -79,8 +79,6 @@ class LiveRegOptimizer {
 
   /// The scalar type to convert to
   Type *const ConvertToScalar;
-  /// The set of visited Instructions
-  SmallPtrSet<Instruction *, 4> Visited;
   /// Map of Value -> Converted Value
   ValueToValueMap ValMap;
   /// Map of containing conversions from Optimal Type -> Original Type per BB.
@@ -288,6 +286,7 @@ bool LiveRegOptimizer::optimizeLiveType(
   SmallPtrSet<PHINode *, 4> PhiNodes;
   SmallPtrSet<Instruction *, 4> Defs;
   SmallPtrSet<Instruction *, 4> Uses;
+  SmallPtrSet<Instruction *, 4> Visited;
 
   Worklist.push_back(cast<Instruction>(I));
   while (!Worklist.empty()) {

diff  --git a/llvm/test/CodeGen/AMDGPU/amdgpu-late-codegenprepare-crash-splat.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-late-codegenprepare-crash-splat.ll
new file mode 100644
index 0000000000000..b8464c37a5dc2
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-late-codegenprepare-crash-splat.ll
@@ -0,0 +1,72 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=amdgpu-late-codegenprepare %s  | FileCheck %s
+
+; This crashed because the PHI with a splat was rejected, but then we marked the PHI
+; as visited and tried to convert one of its user afterwards.
+
+define amdgpu_kernel void @widget(ptr %arg, ptr %arg1, ptr %arg2) {
+; CHECK-LABEL: define amdgpu_kernel void @widget(
+; CHECK-SAME: ptr [[ARG:%.*]], ptr [[ARG1:%.*]], ptr [[ARG2:%.*]]) {
+; CHECK-NEXT:  [[BB:.*]]:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARG]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = trunc i32 [[TMP0]] to i1
+; CHECK-NEXT:    [[ARG1_LOAD:%.*]] = load <4 x i8>, ptr [[ARG1]], align 4
+; CHECK-NEXT:    [[ARG2_LOAD:%.*]] = load i64, ptr [[ARG2]], align 4
+; CHECK-NEXT:    br label %[[BB_1:.*]]
+; CHECK:       [[BB_1]]:
+; CHECK-NEXT:    [[PHI:%.*]] = phi ptr [ null, %[[BB]] ], [ [[ARG1]], %[[BB_6:.*]] ]
+; CHECK-NEXT:    [[PHI4:%.*]] = phi <4 x i8> [ splat (i8 1), %[[BB]] ], [ [[PHI15:%.*]], %[[BB_6]] ]
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_2:.*]], label %[[BB_6]]
+; CHECK:       [[BB_2]]:
+; CHECK-NEXT:    [[PHI7:%.*]] = phi <4 x i8> [ [[PHI13:%.*]], %[[BB_5:.*]] ], [ [[PHI4]], %[[BB_1]] ]
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_4:.*]], label %[[BB_5]]
+; CHECK:       [[BB_3:.*]]:
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_4]], label %[[BB_EXIT:.*]]
+; CHECK:       [[BB_4]]:
+; CHECK-NEXT:    [[PHI11:%.*]] = phi <4 x i8> [ [[PHI7]], %[[BB_3]] ], [ zeroinitializer, %[[BB_2]] ]
+; CHECK-NEXT:    store <4 x i8> [[PHI11]], ptr [[PHI]], align 1
+; CHECK-NEXT:    br label %[[BB_5]]
+; CHECK:       [[BB_5]]:
+; CHECK-NEXT:    [[PHI13]] = phi <4 x i8> [ zeroinitializer, %[[BB_4]] ], [ [[PHI7]], %[[BB_2]] ]
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[BB_2]], label %[[BB_6]]
+; CHECK:       [[BB_6]]:
+; CHECK-NEXT:    [[PHI15]] = phi <4 x i8> [ [[ARG1_LOAD]], %[[BB_1]] ], [ zeroinitializer, %[[BB_5]] ]
+; CHECK-NEXT:    br label %[[BB_1]]
+; CHECK:       [[BB_EXIT]]:
+; CHECK-NEXT:    ret void
+;
+bb:
+  %ld = load i32, ptr %arg, align 4
+  %ld.trunc = trunc i32 %ld to i1
+  %arg1.load = load <4 x i8>, ptr %arg1, align 4
+  %arg2.load = load i64, ptr %arg2, align 4
+  br label %bb.1
+
+bb.1:
+  %phi = phi ptr [ null, %bb ], [ %arg1, %bb.6 ]
+  %phi4 = phi <4 x i8> [ splat (i8 1), %bb ], [ %phi15, %bb.6 ]
+  br i1 %ld.trunc, label %bb.2, label %bb.6
+
+bb.2:
+  %phi7 = phi <4 x i8> [ %phi13, %bb.5 ], [ %phi4, %bb.1 ]
+  br i1 %ld.trunc, label %bb.4, label %bb.5
+
+bb.3:
+  br i1 %ld.trunc, label %bb.4, label %bb.exit
+
+bb.4:
+  %phi11 = phi <4 x i8> [ %phi7, %bb.3 ], [ zeroinitializer, %bb.2 ]
+  store <4 x i8> %phi11, ptr %phi, align 1
+  br label %bb.5
+
+bb.5:
+  %phi13 = phi <4 x i8> [ zeroinitializer, %bb.4 ], [ %phi7, %bb.2 ]
+  br i1 %ld.trunc, label %bb.2, label %bb.6
+
+bb.6:
+  %phi15 = phi <4 x i8> [ %arg1.load, %bb.1 ], [ zeroinitializer, %bb.5 ]
+  br label %bb.1
+
+bb.exit:
+  ret void
+}