Index: include/llvm/IR/InstrTypes.h =================================================================== --- include/llvm/IR/InstrTypes.h +++ include/llvm/IR/InstrTypes.h @@ -358,6 +358,10 @@ /// isExact - Determine whether the exact flag is set. bool isExact() const; + /// Convenience method to copy wrapping, exact, and fast-math flag values + /// from V to this instruction. + void copyFlags(const Value *V); + // Methods for support type inquiry through isa, cast, and dyn_cast: static inline bool classof(const Instruction *I) { return I->isBinaryOp(); Index: lib/IR/Instructions.cpp =================================================================== --- lib/IR/Instructions.cpp +++ lib/IR/Instructions.cpp @@ -2030,6 +2030,24 @@ return cast(this)->isExact(); } +void BinaryOperator::copyFlags(const Value *V) { + // Copy the wrapping flags. + const OverflowingBinaryOperator *OB = dyn_cast(V); + if (OB) { + setHasNoSignedWrap(OB->hasNoSignedWrap()); + setHasNoUnsignedWrap(OB->hasNoUnsignedWrap()); + } + // Copy the exact flag. + const PossiblyExactOperator *PE = dyn_cast(V); + if (PE) + setIsExact(PE->isExact()); + + // Copy the fast-math flags. + const FPMathOperator *FP = dyn_cast(V); + if (FP) + setFastMathFlags(FP->getFastMathFlags()); +} + //===----------------------------------------------------------------------===// // FPMathOperator Class //===----------------------------------------------------------------------===// Index: lib/Transforms/Vectorize/LoopVectorize.cpp =================================================================== --- lib/Transforms/Vectorize/LoopVectorize.cpp +++ lib/Transforms/Vectorize/LoopVectorize.cpp @@ -3181,19 +3181,9 @@ for (unsigned Part = 0; Part < UF; ++Part) { Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]); - // Update the NSW, NUW and Exact flags. Notice: V can be an Undef. - BinaryOperator *VecOp = dyn_cast(V); - if (VecOp && isa(BinOp)) { - VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); - VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); - } - if (VecOp && isa(VecOp)) - VecOp->setIsExact(BinOp->isExact()); - - // Copy the fast-math flags. - if (VecOp && isa(V)) - VecOp->setFastMathFlags(it->getFastMathFlags()); - + if (BinaryOperator *VecOp = dyn_cast(V)) + VecOp->copyFlags(BinOp); + Entry[Part] = V; } Index: test/Transforms/LoopVectorize/exact.ll =================================================================== --- test/Transforms/LoopVectorize/exact.ll +++ test/Transforms/LoopVectorize/exact.ll @@ -0,0 +1,24 @@ +; RUN: opt < %s -loop-vectorize -force-vector-width=4 -S | FileCheck %s + +;CHECK-LABEL: @lshr_exact( +;CHECK: lshr exact <4 x i32> +target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" +target triple = "x86_64-apple-macosx10.9.0" + +define void @lshr_exact(i32* %x) { +entry: + br label %for.body + +for.body: + %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ] + %arrayidx = getelementptr inbounds i32* %x, i64 %iv + %0 = load i32* %arrayidx, align 4 + %conv1 = lshr exact i32 %0, 1 + store i32 %conv1, i32* %arrayidx, align 4 + %iv.next = add nuw nsw i64 %iv, 1 + %exitcond = icmp eq i64 %iv.next, 256 + br i1 %exitcond, label %for.end, label %for.body + +for.end: + ret void +}