Index: lib/Transforms/InstCombine/InstCombineVectorOps.cpp =================================================================== --- lib/Transforms/InstCombine/InstCombineVectorOps.cpp +++ lib/Transforms/InstCombine/InstCombineVectorOps.cpp @@ -383,24 +383,16 @@ auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType), ConstantVector::get(ExtendMask)); - // Insert the new shuffle after the vector operand of the extract is defined - // or at the start of the basic block, so any subsequent extracts can use it. - bool ReplaceAllExtUsers; - if (auto *ExtVecOpInst = dyn_cast(ExtVecOp)) { - WideVec->insertAfter(ExtVecOpInst); - ReplaceAllExtUsers = true; - } else { - // TODO: Insert at start of function, so it's always safe to replace all? - IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt()); - ReplaceAllExtUsers = false; - } + // Insert at the start of the basic block, so any subsequent extracts in this + // basic block can use it. + // TODO: Insert at start of function, so it's safe to replace all users? + IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt()); // Replace extracts from the original narrow vector with extracts from the new // wide vector. for (User *U : ExtVecOp->users()) { ExtractElementInst *OldExt = dyn_cast(U); - if (!OldExt || - (!ReplaceAllExtUsers && OldExt->getParent() != WideVec->getParent())) + if (!OldExt || OldExt->getParent() != WideVec->getParent()) continue; auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1)); NewExt->insertAfter(WideVec); Index: test/Transforms/InstCombine/insert-extract-shuffle.ll =================================================================== --- test/Transforms/InstCombine/insert-extract-shuffle.ll +++ test/Transforms/InstCombine/insert-extract-shuffle.ll @@ -125,3 +125,29 @@ ret <8 x i16> %t6 } +; The widening shuffle must be inserted at a valid point (after the PHIs). + +define <4 x double> @pr25999_phis(i1 %c, i1 %c2, <2 x double> %a, <4 x double> %b) { +; CHECK-LABEL: @pr25999_phis( +; CHECK: %tmp1 = phi <2 x double> [ %a, %bb1 ], [ %r, %bb2 ] +; CHECK-NEXT: %tmp2 = phi <4 x double> [ %b, %bb1 ], [ zeroinitializer, %bb2 ] +; CHECK-NEXT: %[[WIDEVEC:.*]] = shufflevector <2 x double> %tmp1, <2 x double> undef, <4 x i32> +; CHECK-NEXT: %tmp4 = shufflevector <4 x double> %tmp2, <4 x double> %[[WIDEVEC]], <4 x i32> +; CHECK-NEXT: ret <4 x double> %tmp4 +bb1: + br i1 %c, label %bb2, label %bb3 + +bb2: + %r = call <2 x double> @dummy(<2 x double> %a) + br label %bb3 + +bb3: + %tmp1 = phi <2 x double> [ %a, %bb1 ], [ %r, %bb2 ] + %tmp2 = phi <4 x double> [ %b, %bb1 ], [ zeroinitializer, %bb2 ] + %tmp3 = extractelement <2 x double> %tmp1, i32 0 + %tmp4 = insertelement <4 x double> %tmp2, double %tmp3, i32 2 + ret <4 x double> %tmp4 +} + +declare <2 x double> @dummy(<2 x double>) +