[PATCH] D128071: [SLP]Fix a crash when insert subvector is out of range.

[Alexey Bataev via Phabricator via llvm-commits]

ABataev updated this revision to Diff 438026.
ABataev added a comment.

Adjust OffsetBeg for better shuffle mask estimation.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D128071/new/

https://reviews.llvm.org/D128071

Files:
  llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
  llvm/test/Transforms/SLPVectorizer/AArch64/buildvector-vectorize.ll


Index: llvm/test/Transforms/SLPVectorizer/AArch64/buildvector-vectorize.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/SLPVectorizer/AArch64/buildvector-vectorize.ll
@@ -0,0 +1,21 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -slp-vectorizer -S -mtriple=aarch64 < %s | FileCheck %s
+
+define void @test(ptr %p) {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[INC:%.*]] = getelementptr inbounds i16, ptr [[P:%.*]], i64 1
+; CHECK-NEXT:    [[TMP0:%.*]] = load i16, ptr [[INC]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = load i16, ptr [[P]], align 2
+; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <8 x i16> zeroinitializer, i16 [[TMP0]], i32 5
+; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <8 x i16> [[TMP2]], i16 [[TMP1]], i32 7
+; CHECK-NEXT:    ret void
+;
+entry:
+  %inc = getelementptr inbounds i16, ptr %p, i64 1
+  %0 = load i16, ptr %inc, align 4
+  %1 = load i16, ptr %p, align 2
+  %2 = insertelement <8 x i16> zeroinitializer, i16 %0, i32 5
+  %3 = insertelement <8 x i16> %2, i16 %1, i32 7
+  ret void
+}
Index: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
===================================================================
--- llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -6134,6 +6134,16 @@
           PowerOf2Ceil(OffsetEnd - OffsetBeg + 1),
           ((OffsetEnd - OffsetBeg + VecScalarsSz) / VecScalarsSz) *
               VecScalarsSz);
+      bool IsWholeSubvector =
+          OffsetBeg == Offset && ((OffsetEnd + 1) % VecScalarsSz == 0);
+      // Check if we can safely insert a subvector. If it is not possible, just
+      // generate a whole-sized vector and shuffle the source vector and the new
+      // subvector.
+      if (OffsetBeg + InsertVecSz > VecSz) {
+        // Align OffsetBeg to generate correct mask.
+        OffsetBeg = alignDown(OffsetBeg, VecSz, Offset);
+        InsertVecSz = VecSz;
+      }
 
       APInt DemandedElts = APInt::getZero(NumElts);
       // TODO: Add support for Instruction::InsertValue.
@@ -6177,7 +6187,7 @@
       // TODO: Implement the analysis of the FirstInsert->getOperand(0)
       // subvector of ActualVecTy.
       if (!isUndefVector(FirstInsert->getOperand(0)) && NumScalars != NumElts &&
-          (Offset != OffsetBeg || (OffsetEnd + 1) % VecScalarsSz != 0)) {
+          !IsWholeSubvector) {
         if (InsertVecSz != VecSz) {
           auto *ActualVecTy =
               FixedVectorType::get(SrcVecTy->getElementType(), VecSz);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D128071.438026.patch
Type: text/x-patch
Size: 2601 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220617/91aed682/attachment.bin>