[PATCH] D93192: [SLP] Fix vector element size for the store chains

[Anton Afanasyev via Phabricator via llvm-commits]

anton-afanasyev created this revision.
anton-afanasyev added reviewers: ABataev, RKSimon, dtemirbulatov.
Herald added a subscriber: hiraditya.
anton-afanasyev requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Vector element size could be different for different store chains.
This patch prevents wrong computation of maximum number of elements
for that case.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D93192

Files:
  llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
  llvm/test/Transforms/SLPVectorizer/X86/combined-stores-chains.ll


Index: llvm/test/Transforms/SLPVectorizer/X86/combined-stores-chains.ll
===================================================================
--- llvm/test/Transforms/SLPVectorizer/X86/combined-stores-chains.ll
+++ llvm/test/Transforms/SLPVectorizer/X86/combined-stores-chains.ll
@@ -23,28 +23,21 @@
 ; CHECK-NEXT:    [[T252:%.*]] = getelementptr inbounds i64, i64* [[T02]], i64 9
 ; CHECK-NEXT:    [[T292:%.*]] = getelementptr inbounds i64, i64* [[T02]], i64 10
 ; CHECK-NEXT:    [[T322:%.*]] = getelementptr inbounds i64, i64* [[T02]], i64 11
-; CHECK-NEXT:    [[T19:%.*]] = load i32, i32* [[T14]], align 4
-; CHECK-NEXT:    [[T23:%.*]] = load i32, i32* [[T18]], align 4
-; CHECK-NEXT:    [[T27:%.*]] = load i32, i32* [[T22]], align 4
-; CHECK-NEXT:    [[T30:%.*]] = load i32, i32* [[T26]], align 4
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i64* [[T142]] to <2 x i64>*
-; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, <2 x i64>* [[TMP1]], align 8
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i64* [[T222]] to <2 x i64>*
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[T14]] to <4 x i32>*
+; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i64* [[T142]] to <2 x i64>*
 ; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i64>, <2 x i64>* [[TMP3]], align 8
-; CHECK-NEXT:    [[T20:%.*]] = add nsw i32 [[T19]], 4
-; CHECK-NEXT:    [[T24:%.*]] = add nsw i32 [[T23]], 4
-; CHECK-NEXT:    [[T28:%.*]] = add nsw i32 [[T27]], 6
-; CHECK-NEXT:    [[T31:%.*]] = add nsw i32 [[T30]], 7
-; CHECK-NEXT:    [[TMP5:%.*]] = add nsw <2 x i64> [[TMP2]], <i64 4, i64 4>
-; CHECK-NEXT:    [[TMP6:%.*]] = add nsw <2 x i64> [[TMP4]], <i64 6, i64 7>
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast i64* [[T212]] to <2 x i64>*
-; CHECK-NEXT:    store <2 x i64> [[TMP5]], <2 x i64>* [[TMP7]], align 8
-; CHECK-NEXT:    [[TMP8:%.*]] = bitcast i64* [[T292]] to <2 x i64>*
-; CHECK-NEXT:    store <2 x i64> [[TMP6]], <2 x i64>* [[TMP8]], align 8
-; CHECK-NEXT:    store i32 [[T20]], i32* [[T21]], align 4
-; CHECK-NEXT:    store i32 [[T24]], i32* [[T25]], align 4
-; CHECK-NEXT:    store i32 [[T28]], i32* [[T29]], align 4
-; CHECK-NEXT:    store i32 [[T31]], i32* [[T32]], align 4
+; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i64* [[T222]] to <2 x i64>*
+; CHECK-NEXT:    [[TMP6:%.*]] = load <2 x i64>, <2 x i64>* [[TMP5]], align 8
+; CHECK-NEXT:    [[TMP7:%.*]] = add nsw <4 x i32> [[TMP2]], <i32 4, i32 4, i32 6, i32 7>
+; CHECK-NEXT:    [[TMP8:%.*]] = add nsw <2 x i64> [[TMP4]], <i64 4, i64 4>
+; CHECK-NEXT:    [[TMP9:%.*]] = add nsw <2 x i64> [[TMP6]], <i64 6, i64 7>
+; CHECK-NEXT:    [[TMP10:%.*]] = bitcast i64* [[T212]] to <2 x i64>*
+; CHECK-NEXT:    store <2 x i64> [[TMP8]], <2 x i64>* [[TMP10]], align 8
+; CHECK-NEXT:    [[TMP11:%.*]] = bitcast i64* [[T292]] to <2 x i64>*
+; CHECK-NEXT:    store <2 x i64> [[TMP9]], <2 x i64>* [[TMP11]], align 8
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[T21]] to <4 x i32>*
+; CHECK-NEXT:    store <4 x i32> [[TMP7]], <4 x i32>* [[TMP12]], align 4
 ; CHECK-NEXT:    ret void
 ;
   %t0 = bitcast i8* %v0 to i32*
Index: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
===================================================================
--- llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -6076,7 +6076,7 @@
 
     // If a vector register can't hold 1 element, we are done.
     unsigned MaxVecRegSize = R.getMaxVecRegSize();
-    unsigned EltSize = R.getVectorElementSize(Stores[0]);
+    unsigned EltSize = R.getVectorElementSize(Operands[0]);
     if (MaxVecRegSize % EltSize != 0)
       continue;
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D93192.311487.patch
Type: text/x-patch
Size: 3628 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201214/18093065/attachment.bin>