[InstCombine] allow undef elements when forming splat from chain of insertelements

[spatel]

Summary

We allow forming a splat (broadcast) shuffle, but we were conservatively limiting that to cases where all elements of the vector are specified. It should be safe from a codegen perspective to allow undefined lanes of the vector because the expansion of a splat shuffle would become the chain of inserts again.

Forming splat shuffles can reduce IR and help enable further IR transforms. Motivating bugs:
https://bugs.llvm.org/show_bug.cgi?id=42174
https://bugs.llvm.org/show_bug.cgi?id=16739 (this one needs a lot of changes, and I'm still not sure how to fix it)

[RKSimon]

LGTM - cheers

llvm/trunk/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

if (CurrIE != &InsElt &&
         (!CurrIE->hasOneUse() && (NextIE != nullptr || !Idx->isZero())))
       return nullptr;
 
     ElementPresent[Idx->getZExtValue()] = true;
     FirstIE = CurrIE;
     CurrIE = NextIE;
   }
 
-  // Make sure we've seen an insert into every element.
-  if (llvm::any_of(ElementPresent, [](bool Present) { return !Present; }))
+  // If this is just a single insertelement (not a sequence), we are done.
+  if (FirstIE == &InsElt)
+    return nullptr;
+
+  // If we are not inserting into an undef vector, make sure we've seen an
+  // insert into every element.
+  // TODO: If the base vector is not undef, it might be better to create a splat
+  //       and then a select-shuffle (blend) with the base vector.
+  if (!isa<UndefValue>(FirstIE->getOperand(0)))
+    if (any_of(ElementPresent, [](bool Present) { return !Present; }))
       return nullptr;
 
   // Create the insert + shuffle.
   Type *Int32Ty = Type::getInt32Ty(InsElt.getContext());
   UndefValue *UndefVec = UndefValue::get(VecTy);
   Constant *Zero = ConstantInt::get(Int32Ty, 0);
   if (!cast<ConstantInt>(FirstIE->getOperand(2))->isZero())
     FirstIE = InsertElementInst::Create(UndefVec, SplatVal, Zero, "", &InsElt);
 
-  Constant *ZeroMask = ConstantVector::getSplat(NumElements, Zero);
-  return new ShuffleVectorInst(FirstIE, UndefVec, ZeroMask);
+  // Splat from element 0, but replace absent elements with undef in the mask.
+  SmallVector<Constant *, 16> Mask(NumElements, Zero);
+  for (unsigned i = 0; i != NumElements; ++i)
+    if (!ElementPresent[i])
+      Mask[i] = UndefValue::get(Int32Ty);
+
+  return new ShuffleVectorInst(FirstIE, UndefVec, ConstantVector::get(Mask));
 }
 
 /// If we have an insertelement instruction feeding into another insertelement
 /// and the 2nd is inserting a constant into the vector, canonicalize that
 /// constant insertion before the insertion of a variable:
 ///
 /// insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 -->
 /// insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1

llvm/trunk/test/Transforms/InstCombine/broadcast.ll

;
   %a1 = fadd <4 x float> %ins1, %ins1
   %ins2 = insertelement<4 x float> %ins1, float %v, i32 1
   %ins3 = insertelement<4 x float> %ins2, float %v, i32 2
   %ins4 = insertelement<4 x float> %ins3, float %v, i32 3
   %res = fadd <4 x float> %a1, %ins4
   ret <4 x float> %res
 }
 
-define <4 x float> @bad1(float %arg) {
-; CHECK-LABEL: @bad1(
-; CHECK-NEXT:    [[T4:%.*]] = insertelement <4 x float> undef, float [[ARG:%.*]], i32 1
-; CHECK-NEXT:    [[T5:%.*]] = insertelement <4 x float> [[T4]], float [[ARG]], i32 2
-; CHECK-NEXT:    [[T6:%.*]] = insertelement <4 x float> [[T5]], float [[ARG]], i32 3
+; The insert is changed to allow the canonical shuffle-splat pattern from element 0.
+
+define <4 x float> @splat_undef1(float %arg) {
+; CHECK-LABEL: @splat_undef1(
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <4 x float> undef, float [[ARG:%.*]], i32 0
+; CHECK-NEXT:    [[T6:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> undef, <4 x i32> <i32 undef, i32 0, i32 0, i32 0>
 ; CHECK-NEXT:    ret <4 x float> [[T6]]
 ;
   %t = insertelement <4 x float> undef, float %arg, i32 1
   %t4 = insertelement <4 x float> %t, float %arg, i32 1
   %t5 = insertelement <4 x float> %t4, float %arg, i32 2
   %t6 = insertelement <4 x float> %t5, float %arg, i32 3
   ret <4 x float> %t6
 }
 
-define <4 x float> @bad2(float %arg) {
-; CHECK-LABEL: @bad2(
+; Re-uses the existing first insertelement.
+
+define <4 x float> @splat_undef2(float %arg) {
+; CHECK-LABEL: @splat_undef2(
 ; CHECK-NEXT:    [[T:%.*]] = insertelement <4 x float> undef, float [[ARG:%.*]], i32 0
-; CHECK-NEXT:    [[T5:%.*]] = insertelement <4 x float> [[T]], float [[ARG]], i32 2
-; CHECK-NEXT:    [[T6:%.*]] = insertelement <4 x float> [[T5]], float [[ARG]], i32 3
+; CHECK-NEXT:    [[T6:%.*]] = shufflevector <4 x float> [[T]], <4 x float> undef, <4 x i32> <i32 0, i32 undef, i32 0, i32 0>
 ; CHECK-NEXT:    ret <4 x float> [[T6]]
 ;
   %t = insertelement <4 x float> undef, float %arg, i32 0
   %t5 = insertelement <4 x float> %t, float %arg, i32 2
   %t6 = insertelement <4 x float> %t5, float %arg, i32 3
   ret <4 x float> %t6
 }
 

 ; CHECK-LABEL: @bad4(
 ; CHECK-NEXT:    [[T:%.*]] = insertelement <1 x float> undef, float [[ARG:%.*]], i32 0
 ; CHECK-NEXT:    ret <1 x float> [[T]]
 ;
   %t = insertelement <1 x float> undef, float %arg, i32 0
   ret <1 x float> %t
 }
 
-define <4 x float> @bad5(float %arg) {
-; CHECK-LABEL: @bad5(
+; Multiple undef elements are ok.
+; TODO: Multiple uses triggers the transform at %t4, but we could form another splat from %t6 and simplify?
+
+define <4 x float> @splat_undef3(float %arg) {
+; CHECK-LABEL: @splat_undef3(
 ; CHECK-NEXT:    [[T:%.*]] = insertelement <4 x float> undef, float [[ARG:%.*]], i32 0
-; CHECK-NEXT:    [[T4:%.*]] = insertelement <4 x float> [[T]], float [[ARG]], i32 1
+; CHECK-NEXT:    [[T4:%.*]] = shufflevector <4 x float> [[T]], <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 undef, i32 undef>
 ; CHECK-NEXT:    [[T5:%.*]] = insertelement <4 x float> [[T4]], float [[ARG]], i32 2
 ; CHECK-NEXT:    [[T6:%.*]] = insertelement <4 x float> [[T5]], float [[ARG]], i32 3
 ; CHECK-NEXT:    [[T7:%.*]] = fadd <4 x float> [[T6]], [[T4]]
 ; CHECK-NEXT:    ret <4 x float> [[T7]]
 ;
   %t = insertelement <4 x float> undef, float %arg, i32 0
   %t4 = insertelement <4 x float> %t, float %arg, i32 1
   %t5 = insertelement <4 x float> %t4, float %arg, i32 2