[InstCombine] canonicalize insertelement of scalar constant ahead of insertelement of variable

insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 -->
insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1

As noted in the code comment and seen in the test changes, the motivation is that by pulling
constant insertion up, we may be able to constant fold some insertelement instructions.

Differential Revision: https://reviews.llvm.org/D31196

llvm-svn: 298520

5 files changed, 49 insertions, 22 deletions

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index 10f03c4bdd8..c29d3276526 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -645,6 +645,36 @@ static Instruction *foldInsSequenceIntoBroadcast(InsertElementInst &InsElt) {
   return new ShuffleVectorInst(InsertFirst, UndefValue::get(VT), ZeroMask);
 }
 
+/// 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
+///
+/// This has the potential of eliminating the 2nd insertelement instruction
+/// via constant folding of the scalar constant into a vector constant.
+static Instruction *hoistInsEltConst(InsertElementInst &InsElt2,
+                                     InstCombiner::BuilderTy &Builder) {
+  auto *InsElt1 = dyn_cast<InsertElementInst>(InsElt2.getOperand(0));
+  if (!InsElt1 || !InsElt1->hasOneUse())
+    return nullptr;
+
+  Value *X, *Y;
+  Constant *ScalarC;
+  ConstantInt *IdxC1, *IdxC2;
+  if (match(InsElt1->getOperand(0), m_Value(X)) &&
+      match(InsElt1->getOperand(1), m_Value(Y)) && !isa<Constant>(Y) &&
+      match(InsElt1->getOperand(2), m_ConstantInt(IdxC1)) &&
+      match(InsElt2.getOperand(1), m_Constant(ScalarC)) &&
+      match(InsElt2.getOperand(2), m_ConstantInt(IdxC2)) && IdxC1 != IdxC2) {
+    Value *NewInsElt1 = Builder.CreateInsertElement(X, ScalarC, IdxC2);
+    return InsertElementInst::Create(NewInsElt1, Y, IdxC1);
+  }
+
+  return nullptr;
+}
+
 /// insertelt (shufflevector X, CVec, Mask|insertelt X, C1, CIndex1), C, CIndex
 /// --> shufflevector X, CVec', Mask'
 static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) {
@@ -806,6 +836,9 @@ Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
   if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE))
     return Shuf;
 
+  if (Instruction *NewInsElt = hoistInsEltConst(IE, *Builder))
+    return NewInsElt;
+
   // Turn a sequence of inserts that broadcasts a scalar into a single
   // insert + shufflevector.
   if (Instruction *Broadcast = foldInsSequenceIntoBroadcast(IE))
diff --git a/llvm/test/Transforms/InstCombine/bitcast-bigendian.ll b/llvm/test/Transforms/InstCombine/bitcast-bigendian.ll
index 1a91d11d8ae..a6b2ac85c9c 100644
--- a/llvm/test/Transforms/InstCombine/bitcast-bigendian.ll
+++ b/llvm/test/Transforms/InstCombine/bitcast-bigendian.ll
@@ -81,9 +81,8 @@ define <2 x float> @test5(float %A, float %B) {
 
 define <2 x float> @test6(float %A){
 ; CHECK-LABEL: @test6(
-; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x float> undef, float %A, i32 0
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x float> [[TMP1]], float 4.200000e+01, i32 1
-; CHECK-NEXT:    ret <2 x float> [[TMP2]]
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x float> <float undef, float 4.200000e+01>, float %A, i32 0
+; CHECK-NEXT:    ret <2 x float> [[TMP1]]
 ;
   %tmp23 = bitcast float %A to i32
   %tmp24 = zext i32 %tmp23 to i64
diff --git a/llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll b/llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll
index 4507deb7f02..29f774c5f62 100644
--- a/llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll
+++ b/llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll
@@ -86,11 +86,8 @@ define <8 x float> @widen_extract4(<8 x float> %ins, <2 x float> %ext) {
 
 define <8 x i16> @pr26015(<4 x i16> %t0) {
 ; CHECK-LABEL: @pr26015(
-; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i16> %t0, <4 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
-; CHECK-NEXT:    [[TMP2:%.*]] = extractelement <4 x i16> %t0, i32 2
-; CHECK-NEXT:    [[T2:%.*]] = insertelement <8 x i16> <i16 0, i16 0, i16 0, i16 undef, i16 0, i16 0, i16 undef, i16 undef>, i16 [[TMP2]], i32 3
-; CHECK-NEXT:    [[T3:%.*]] = insertelement <8 x i16> [[T2]], i16 0, i32 6
-; CHECK-NEXT:    [[T5:%.*]] = shufflevector <8 x i16> [[T3]], <8 x i16> [[TMP1]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 11>
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i16> %t0, <4 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
+; CHECK-NEXT:    [[T5:%.*]] = shufflevector <8 x i16> <i16 0, i16 0, i16 0, i16 undef, i16 0, i16 0, i16 0, i16 undef>, <8 x i16> [[TMP1]], <8 x i32> <i32 0, i32 1, i32 2, i32 10, i32 4, i32 5, i32 6, i32 11>
 ; CHECK-NEXT:    ret <8 x i16> [[T5]]
 ;
   %t1 = extractelement <4 x i16> %t0, i32 2
@@ -110,8 +107,7 @@ define <8 x i16> @pr25999(<4 x i16> %t0, i1 %b) {
 ; CHECK-NEXT:    br i1 %b, label %if, label %end
 ; CHECK:       if:
 ; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i16> %t0, <4 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
-; CHECK-NEXT:    [[T2:%.*]] = insertelement <8 x i16> <i16 0, i16 0, i16 0, i16 undef, i16 0, i16 0, i16 undef, i16 undef>, i16 [[T1]], i32 3
-; CHECK-NEXT:    [[T3:%.*]] = insertelement <8 x i16> [[T2]], i16 0, i32 6
+; CHECK-NEXT:    [[T3:%.*]] = insertelement <8 x i16> <i16 0, i16 0, i16 0, i16 undef, i16 0, i16 0, i16 0, i16 undef>, i16 [[T1]], i32 3
 ; CHECK-NEXT:    [[T5:%.*]] = shufflevector <8 x i16> [[T3]], <8 x i16> [[TMP1]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 11>
 ; CHECK-NEXT:    ret <8 x i16> [[T5]]
 ; CHECK:       end:
diff --git a/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll b/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll
index e29c9eef74a..5f27634da19 100644
--- a/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll
+++ b/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll
@@ -182,10 +182,9 @@ define <4 x float> @dead_shuffle_elt(<4 x float> %x, <2 x float> %y) nounwind {
 
 define <2 x float> @test_fptrunc(double %f) {
 ; CHECK-LABEL: @test_fptrunc(
-; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> undef, double %f, i32 0
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x double> [[TMP1]], double 0.000000e+00, i32 1
-; CHECK-NEXT:    [[TMP3:%.*]] = fptrunc <2 x double> [[TMP2]] to <2 x float>
-; CHECK-NEXT:    ret <2 x float> [[TMP3]]
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> <double undef, double 0.000000e+00>, double %f, i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = fptrunc <2 x double> [[TMP1]] to <2 x float>
+; CHECK-NEXT:    ret <2 x float> [[TMP2]]
 ;
   %tmp9 = insertelement <4 x double> undef, double %f, i32 0
   %tmp10 = insertelement <4 x double> %tmp9, double 0.000000e+00, i32 1
@@ -198,10 +197,9 @@ define <2 x float> @test_fptrunc(double %f) {
 
 define <2 x double> @test_fpext(float %f) {
 ; CHECK-LABEL: @test_fpext(
-; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x float> undef, float %f, i32 0
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x float> [[TMP1]], float 0.000000e+00, i32 1
-; CHECK-NEXT:    [[TMP3:%.*]] = fpext <2 x float> [[TMP2]] to <2 x double>
-; CHECK-NEXT:    ret <2 x double> [[TMP3]]
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x float> <float undef, float 0.000000e+00>, float %f, i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = fpext <2 x float> [[TMP1]] to <2 x double>
+; CHECK-NEXT:    ret <2 x double> [[TMP2]]
 ;
   %tmp9 = insertelement <4 x float> undef, float %f, i32 0
   %tmp10 = insertelement <4 x float> %tmp9, float 0.000000e+00, i32 1
@@ -223,8 +221,7 @@ define <4 x double> @test_shuffle(<4 x double> %f) {
 
 define <4 x float> @test_select(float %f, float %g) {
 ; CHECK-LABEL: @test_select(
-; CHECK-NEXT:    [[A0:%.*]] = insertelement <4 x float> undef, float %f, i32 0
-; CHECK-NEXT:    [[A3:%.*]] = insertelement <4 x float> [[A0]], float 3.000000e+00, i32 3
+; CHECK-NEXT:    [[A3:%.*]] = insertelement <4 x float> <float undef, float undef, float undef, float 3.000000e+00>, float %f, i32 0
 ; CHECK-NEXT:    [[RET:%.*]] = shufflevector <4 x float> [[A3]], <4 x float> <float undef, float 4.000000e+00, float 5.000000e+00, float undef>, <4 x i32> <i32 0, i32 5, i32 6, i32 3>
 ; CHECK-NEXT:    ret <4 x float> [[RET]]
 ;
diff --git a/llvm/test/Transforms/InstCombine/vector_insertelt_shuffle.ll b/llvm/test/Transforms/InstCombine/vector_insertelt_shuffle.ll
index 8db937a9a34..c358509d690 100644
--- a/llvm/test/Transforms/InstCombine/vector_insertelt_shuffle.ll
+++ b/llvm/test/Transforms/InstCombine/vector_insertelt_shuffle.ll
@@ -12,10 +12,12 @@ define <4 x float> @foo(<4 x float> %x) {
   ret <4 x float> %ins2
 }
 
+; Insert of a constant is canonicalized ahead of insert of a variable.
+
 define <4 x float> @bar(<4 x float> %x, float %a) {
 ; CHECK-LABEL: @bar(
-; CHECK-NEXT:    [[INS1:%.*]] = insertelement <4 x float> %x, float %a, i32 1
-; CHECK-NEXT:    [[INS2:%.*]] = insertelement <4 x float> [[INS1]], float 2.000000e+00, i32 2
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <4 x float> %x, float 2.000000e+00, i32 2
+; CHECK-NEXT:    [[INS2:%.*]] = insertelement <4 x float> [[TMP1]], float %a, i32 1
 ; CHECK-NEXT:    ret <4 x float> [[INS2]]
 ;
   %ins1 = insertelement<4 x float> %x, float %a, i32 1