[SLP] limit vectorization of Constant subclasses (PR33958)

This is a fix for:
https://bugs.llvm.org/show_bug.cgi?id=33958

It seems universally true that we would not want to transform this kind of
sequence on any target, but if that's not correct, then we could view this
as a target-specific cost model problem. We could also white-list ConstantInt,
ConstantFP, etc. rather than blacklist Global and ConstantExpr.

Differential Revision: https://reviews.llvm.org/D67362

llvm-svn: 371931

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 7fbcb23..c18972c 100644 (file)
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -194,10 +194,13 @@ static bool allSameBlock(ArrayRef<Value *> VL) {
   return true;
 }
 
-/// \returns True if all of the values in \p VL are constants.
+/// \returns True if all of the values in \p VL are constants (but not
+/// globals/constant expressions).
 static bool allConstant(ArrayRef<Value *> VL) {
+  // Constant expressions and globals can't be vectorized like normal integer/FP
+  // constants.
   for (Value *i : VL)
-    if (!isa<Constant>(i))
+    if (!isa<Constant>(i) || isa<ConstantExpr>(i) || isa<GlobalValue>(i))
       return false;
   return true;
 }

diff --git a/llvm/test/Transforms/SLPVectorizer/X86/consecutive-access.ll b/llvm/test/Transforms/SLPVectorizer/X86/consecutive-access.ll
index b77c010..56a7063 100644 (file)
--- a/llvm/test/Transforms/SLPVectorizer/X86/consecutive-access.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/consecutive-access.ll
@@ -549,14 +549,17 @@ for.body:                                         ; preds = %entry, %for.body
   br i1 %cmp, label %for.body, label %for.cond.cleanup
 }
 
+; Globals/constant expressions are not normal constants.
+; They should not be treated as the usual vectorization candidates.
+
 @g1 = external global i32, align 4
 @g2 = external global i32, align 4
 
 define void @PR33958(i32** nocapture %p) {
 ; CHECK-LABEL: @PR33958(
-; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i32*, i32** [[P:%.*]], i64 1
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32** [[P]] to <2 x i32*>*
-; CHECK-NEXT:    store <2 x i32*> <i32* @g1, i32* @g2>, <2 x i32*>* [[TMP1]], align 8
+; CHECK-NEXT:    store i32* @g1, i32** [[P:%.*]], align 8
+; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i32*, i32** [[P]], i64 1
+; CHECK-NEXT:    store i32* @g2, i32** [[ARRAYIDX1]], align 8
 ; CHECK-NEXT:    ret void
 ;
   store i32* @g1, i32** %p, align 8
@@ -567,9 +570,9 @@ define void @PR33958(i32** nocapture %p) {
 
 define void @store_constant_expression(i64* %p) {
 ; CHECK-LABEL: @store_constant_expression(
-; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i64, i64* [[P:%.*]], i64 1
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i64* [[P]] to <2 x i64>*
-; CHECK-NEXT:    store <2 x i64> <i64 ptrtoint (i32* @g1 to i64), i64 ptrtoint (i32* @g2 to i64)>, <2 x i64>* [[TMP1]], align 8
+; CHECK-NEXT:    store i64 ptrtoint (i32* @g1 to i64), i64* [[P:%.*]], align 8
+; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i64, i64* [[P]], i64 1
+; CHECK-NEXT:    store i64 ptrtoint (i32* @g2 to i64), i64* [[ARRAYIDX1]], align 8
 ; CHECK-NEXT:    ret void
 ;
   store i64 ptrtoint (i32* @g1 to i64), i64* %p, align 8