case Instruction::FPTrunc:
case Instruction::FPExt:
case Instruction::GetElementPtr: {
+ // Bail out if we would create longer vector ops. We could allow creating
+ // longer vector ops, but that may result in more expensive codegen. We
+ // would also need to limit the transform to avoid undefined behavior for
+ // integer div/rem.
+ Type *ITy = I->getType();
+ if (ITy->isVectorTy() && Mask.size() > ITy->getVectorNumElements())
+ return false;
for (Value *Operand : I->operands()) {
if (!canEvaluateShuffled(Operand, Mask, Depth - 1))
return false;
if (isRHSID) return replaceInstUsesWith(SVI, RHS);
}
- if (isa<UndefValue>(RHS) && !SVI.increasesLength() &&
- canEvaluateShuffled(LHS, Mask)) {
+ if (isa<UndefValue>(RHS) && canEvaluateShuffled(LHS, Mask)) {
Value *V = evaluateInDifferentElementOrder(LHS, Mask);
return replaceInstUsesWith(SVI, V);
}
ret <3 x i32> %ext
}
-; TODO: Increasing length of insertelements (no math ops) is a good canonicalization.
+; Increasing length of insertelements (no math ops) is a good canonicalization.
define <3 x i8> @fold_inselts_with_widening_shuffle(i8 %x, i8 %y) {
; CHECK-LABEL: @fold_inselts_with_widening_shuffle(
-; CHECK-NEXT: [[INS0:%.*]] = insertelement <2 x i8> undef, i8 [[X:%.*]], i32 0
-; CHECK-NEXT: [[INS1:%.*]] = insertelement <2 x i8> [[INS0]], i8 [[Y:%.*]], i32 1
-; CHECK-NEXT: [[WIDEN:%.*]] = shufflevector <2 x i8> [[INS1]], <2 x i8> undef, <3 x i32> <i32 0, i32 1, i32 undef>
-; CHECK-NEXT: ret <3 x i8> [[WIDEN]]
+; CHECK-NEXT: [[TMP1:%.*]] = insertelement <3 x i8> undef, i8 [[X:%.*]], i32 0
+; CHECK-NEXT: [[TMP2:%.*]] = insertelement <3 x i8> [[TMP1]], i8 [[Y:%.*]], i32 1
+; CHECK-NEXT: ret <3 x i8> [[TMP2]]
;
%ins0 = insertelement <2 x i8> undef, i8 %x, i32 0
%ins1 = insertelement <2 x i8> %ins0, i8 %y, i32 1