[ConstantFold][SVE] Skip scalable vectors in ConstantFoldInsertElementInstruction.

[huihuiz]

Summary

Should not constant fold insertelement instruction for scalable vector type.

[huihuiz]

current upstream crash with "void llvm::Value::doRAUW(llvm::Value *, llvm::Value::ReplaceMetadataUses): Assertion `New->getType() == getType() && "replaceAllUses of value with new value of different type!"' failed."

opt -S -instcombine test.ll

test.ll

define <vscale x 4 x i32> @f(i32 %val) {
  %i = insertelement <vscale x 4 x i32> undef, i32 1, i32 0
  ret <vscale x 4 x i32> %i
}

[efriedma]

instcombine probably isn't the best place to stick a test for ConstantFold.cpp. Not sure where we're sticking tests for ConstantFold.cpp these days, though. Maybe test/Analysis/ConstantFolding/ is appropriate. (The directory name is actually referring to different code, but it's closer.)

llvm/lib/IR/ConstantFold.cpp

840

We could handle the special case of inserting zero into a zero vector. Not important, though.

llvm/test/Transforms/InstCombine/insertelement.ll

71

Tests involving non-constant operands don't really seem related.

[huihuiz]

Addressed review feedback.

[efriedma]

LGTM. But please wait a couple days before you merge in case anyone wants to comment on the API usage.

[spatel]

llvm/lib/IR/ConstantFold.cpp

836–837

It would be clearer to say:

// Do not iterate on scalable vector. 
// The number of elements is unknown at compile-time.

llvm/test/Analysis/ConstantFolding/insertelement.ll

2 ↗

(On Diff #232009)

This should RUN with "-constprop" rather than "-instcombine".

[huihuiz]

Thank you Sanjay!

[spatel]

LGTM

llvm/lib/IR/ConstantFold.cpp

Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
                                                      Constant *Elt,
                                                      Constant *Idx) {
   if (isa<UndefValue>(Idx))
     return UndefValue::get(Val->getType());
 
   ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
   if (!CIdx) return nullptr;
 
+  // Should not iterate over scalable vector, as num element is known at
+  // runtime.
+  VectorType *ValTy = cast<VectorType>(Val->getType());
+  if (ValTy->isScalable())
+    return nullptr;
+
   unsigned NumElts = Val->getType()->getVectorNumElements();
   if (CIdx->uge(NumElts))
     return UndefValue::get(Val->getType());
 
   SmallVector<Constant*, 16> Result;
   Result.reserve(NumElts);
   auto *Ty = Type::getInt32Ty(Val->getContext());
   uint64_t IdxVal = CIdx->getZExtValue();

llvm/test/Transforms/InstCombine/insertelement.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -instcombine -S | FileCheck %s
+
+
+define <4 x i32> @insertelement_fixedlength_constant() {
+; CHECK-LABEL: @insertelement_fixedlength_constant(
+; CHECK-NEXT:    ret <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+;
+  %i = insertelement <4 x i32> undef, i32 1, i32 0
+  ret <4 x i32> %i
+}
+
+define <4 x i32> @insertelement_fixedlength_val1(<4 x i32> %val) {
+; CHECK-LABEL: @insertelement_fixedlength_val1(
+; CHECK-NEXT:    [[I:%.*]] = insertelement <4 x i32> [[VAL:%.*]], i32 1, i32 0
+; CHECK-NEXT:    ret <4 x i32> [[I]]
+;
+  %i = insertelement <4 x i32> %val, i32 1, i32 0
+  ret <4 x i32> %i
+}
+
+define <4 x i32> @insertelement_fixedlength_val2(i32 %val) {
+; CHECK-LABEL: @insertelement_fixedlength_val2(
+; CHECK-NEXT:    [[I:%.*]] = insertelement <4 x i32> undef, i32 [[VAL:%.*]], i32 0
+; CHECK-NEXT:    ret <4 x i32> [[I]]
+;
+  %i = insertelement <4 x i32> undef, i32 %val, i32 0
+  ret <4 x i32> %i
+}
+
+define <4 x i32> @insertelement_fixedlength_val3(i32 %val) {
+; CHECK-LABEL: @insertelement_fixedlength_val3(
+; CHECK-NEXT:    [[I:%.*]] = insertelement <4 x i32> undef, i32 1, i32 [[VAL:%.*]]
+; CHECK-NEXT:    ret <4 x i32> [[I]]
+;
+  %i = insertelement <4 x i32> undef, i32 1, i32 %val
+  ret <4 x i32> %i
+}
+
+define <vscale x 4 x i32> @insertelement_scalable_constant() {
+; CHECK-LABEL: @insertelement_scalable_constant(
+; CHECK-NEXT:    ret <vscale x 4 x i32> insertelement (<vscale x 4 x i32> undef, i32 1, i32 0)
+;
+  %i = insertelement <vscale x 4 x i32> undef, i32 1, i32 0
+  ret <vscale x 4 x i32> %i
+}
+
+define <vscale x 4 x i32> @insertelement_scalable_val1(<vscale x 4 x i32> %val) {
+; CHECK-LABEL: @insertelement_scalable_val1(
+; CHECK-NEXT:    [[I:%.*]] = insertelement <vscale x 4 x i32> [[VAL:%.*]], i32 1, i32 0
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[I]]
+;
+  %i = insertelement <vscale x 4 x i32> %val, i32 1, i32 0
+  ret <vscale x 4 x i32> %i
+}
+
+define <vscale x 4 x i32> @insertelement_scalable_val2(i32 %val) {
+; CHECK-LABEL: @insertelement_scalable_val2(
+; CHECK-NEXT:    [[I:%.*]] = insertelement <vscale x 4 x i32> undef, i32 [[VAL:%.*]], i32 0
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[I]]
+;
+  %i = insertelement <vscale x 4 x i32> undef, i32 %val, i32 0
+  ret <vscale x 4 x i32> %i
+}
+
+define <vscale x 4 x i32> @insertelement_scalable_val3(i32 %val) {
+; CHECK-LABEL: @insertelement_scalable_val3(
+; CHECK-NEXT:    [[I:%.*]] = insertelement <vscale x 4 x i32> undef, i32 1, i32 [[VAL:%.*]]
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[I]]
+;
+  %i = insertelement <vscale x 4 x i32> undef, i32 1, i32 %val
+  ret <vscale x 4 x i32> %i
+}