[llvm] 76f0ff8 - [SLP]Add an extra check to avoid infinite vectorization attempts

[Alexey Bataev via llvm-commits]

Author: Alexey Bataev
Date: 2024-11-25T14:27:44-08:00
New Revision: 76f0ff8210d56a050d2679926a2fdddd3a8c16d6

URL: https://github.com/llvm/llvm-project/commit/76f0ff8210d56a050d2679926a2fdddd3a8c16d6
DIFF: https://github.com/llvm/llvm-project/commit/76f0ff8210d56a050d2679926a2fdddd3a8c16d6.diff

LOG: [SLP]Add an extra check to avoid infinite vectorization attempts

Added extra check for the cost of the buildvector if the -slp-threshold
option is used. Prevents infinite vectorization attempts.

Added: 
    

Modified: 
    llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
    llvm/test/Transforms/SLPVectorizer/X86/buildvector-schedule-for-subvector.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 34b80c1e360815..115cbd4d2ce5e4 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -12034,7 +12034,14 @@ bool BoUpSLP::isTreeTinyAndNotFullyVectorizable(bool ForReduction) const {
   if (VectorizableTree.back()->isGather() &&
       VectorizableTree.back()->isAltShuffle() &&
       VectorizableTree.back()->getVectorFactor() > 2 &&
-      allSameBlock(VectorizableTree.back()->Scalars))
+      allSameBlock(VectorizableTree.back()->Scalars) &&
+      !VectorizableTree.back()->Scalars.front()->getType()->isVectorTy() &&
+      TTI->getScalarizationOverhead(
+          getWidenedType(VectorizableTree.back()->Scalars.front()->getType(),
+                         VectorizableTree.back()->getVectorFactor()),
+          APInt::getAllOnes(VectorizableTree.back()->getVectorFactor()),
+          /*Insert=*/true, /*Extract=*/false,
+          TTI::TCK_RecipThroughput) > -SLPCostThreshold)
     return false;
 
   // Otherwise, we can't vectorize the tree. It is both tiny and not fully

diff  --git a/llvm/test/Transforms/SLPVectorizer/X86/buildvector-schedule-for-subvector.ll b/llvm/test/Transforms/SLPVectorizer/X86/buildvector-schedule-for-subvector.ll
index a0b390011faa6f..3bf13b76a93327 100644
--- a/llvm/test/Transforms/SLPVectorizer/X86/buildvector-schedule-for-subvector.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/buildvector-schedule-for-subvector.ll
@@ -7,7 +7,7 @@ define void @test() {
 ; CHECK-NEXT:    [[ADD:%.*]] = add i32 1, 0
 ; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <4 x i32> <i32 0, i32 0, i32 0, i32 poison>, i32 [[ADD]], i32 3
 ; CHECK-NEXT:    [[TMP1:%.*]] = icmp samesign ult <4 x i32> [[TMP0]], zeroinitializer
-; CHECK-NEXT:    [[ICMP:%.*]] = icmp samesign ult i32 0, 0
+; CHECK-NEXT:    [[ICMP:%.*]] = extractelement <4 x i1> [[TMP1]], i32 2
 ; CHECK-NEXT:    [[SELECT:%.*]] = select i1 [[ICMP]], i32 0, i32 0
 ; CHECK-NEXT:    [[ZEXT:%.*]] = zext i32 [[SELECT]] to i64
 ; CHECK-NEXT:    [[GETELEMENTPTR:%.*]] = getelementptr ptr addrspace(1), ptr addrspace(1) null, i64 [[ZEXT]]
@@ -16,8 +16,6 @@ define void @test() {
 ; CHECK-NEXT:    [[CALL:%.*]] = call i32 null(<2 x double> zeroinitializer)
 ; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <4 x i32> <i32 0, i32 0, i32 0, i32 poison>, i32 [[CALL]], i32 3
 ; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq <4 x i32> [[TMP2]], zeroinitializer
-; CHECK-NEXT:    [[TMP4:%.*]] = call <8 x i1> @llvm.vector.insert.v8i1.v4i1(<8 x i1> poison, <4 x i1> [[TMP3]], i64 0)
-; CHECK-NEXT:    [[TMP5:%.*]] = call <8 x i1> @llvm.vector.insert.v8i1.v4i1(<8 x i1> [[TMP4]], <4 x i1> [[TMP1]], i64 4)
 ; CHECK-NEXT:    ret void
 ;
 bb: