mirrors/llvm-project
0
0
Fork 0
mirror of https://github.com/llvm/llvm-project.git synced 2025-04-14 17:06:38 +00:00
Code Issues Projects Releases Wiki Activity

[Matrix] Fix a miscompile due to an incorrect double-transpose fold (#135397)

Browse Source 
Transposes are only inverses of each other when they have matching
shapes.

rdar://145592582
This commit is contained in:
Jon Roelofs 2025-04-12 07:31:13 -07:00 committed by GitHub
parent d39d24cec1
commit e838b8b7e7
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 38 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
View File

@ -804,9 +804,10 @@ public:
m_Value(TA), m_ConstantInt(R), m_ConstantInt(C))))
return nullptr;
// Transpose of a transpose is a nop
// Transpose of a transpose is a nop when the shapes match.
Value *TATA;
if (match(TA, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(TATA)))) {
if (match(TA, m_Intrinsic<Intrinsic::matrix_transpose>(
m_Value(TATA), m_Specific(C), m_Specific(R)))) {
updateShapeAndReplaceAllUsesWith(I, TATA);
eraseFromParentAndMove(&I, II, BB);
eraseFromParentAndMove(TA, II, BB);

35
llvm/test/Transforms/LowerMatrixIntrinsics/transpose-fold.ll Normal file
View File

@ -0,0 +1,35 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s
; We can only fold matching transposes.
define void @reshape(ptr %in, ptr %out) {
; CHECK-LABEL: define void @reshape(
; CHECK-SAME: ptr [[IN:%.*]], ptr [[OUT:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*:]]
; CHECK-NEXT: [[COL_LOAD:%.*]] = load <4 x double>, ptr [[IN]], align 8
; CHECK-NEXT: [[SPLIT:%.*]] = shufflevector <4 x double> [[COL_LOAD]], <4 x double> poison, <2 x i32> <i32 0, i32 1>
; CHECK-NEXT: [[SPLIT1:%.*]] = shufflevector <4 x double> [[COL_LOAD]], <4 x double> poison, <2 x i32> <i32 2, i32 3>
; CHECK-NEXT: [[TMP0:%.*]] = extractelement <2 x double> [[SPLIT]], i64 0
; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x double> poison, double [[TMP0]], i64 0
; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[SPLIT1]], i64 0
; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x double> [[TMP1]], double [[TMP2]], i64 1
; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[SPLIT]], i64 1
; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x double> poison, double [[TMP4]], i64 0
; CHECK-NEXT: [[TMP6:%.*]] = extractelement <2 x double> [[SPLIT1]], i64 1
; CHECK-NEXT: [[TMP7:%.*]] = insertelement <2 x double> [[TMP5]], double [[TMP6]], i64 1
; CHECK-NEXT: store <2 x double> [[TMP3]], ptr [[OUT]], align 8
; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[OUT]], i64 2
; CHECK-NEXT: store <2 x double> [[TMP7]], ptr [[VEC_GEP]], align 8
; CHECK-NEXT: ret void
;
entry:
%0 = load <4 x double>, ptr %in, align 8
%1 = tail call <4 x double> @llvm.matrix.transpose.v4f64(<4 x double> %0, i32 4, i32 1)
%2 = tail call <4 x double> @llvm.matrix.transpose.v4f64(<4 x double> %1, i32 1, i32 4)
%3 = tail call <4 x double> @llvm.matrix.transpose.v4f64(<4 x double> %2, i32 2, i32 2)
%4 = tail call <4 x double> @llvm.matrix.transpose.v4f64(<4 x double> %3, i32 2, i32 2)
%5 = tail call <4 x double> @llvm.matrix.transpose.v4f64(<4 x double> %4, i32 2, i32 2)
store <4 x double> %5, ptr %out, align 8
ret void
}