[llvm] r343406 - [InstCombine] allow lengthening of insertelement to eliminate shuffles

[Sanjay Patel via llvm-commits]

Author: spatel
Date: Sun Sep 30 06:50:42 2018
New Revision: 343406

URL: http://llvm.org/viewvc/llvm-project?rev=343406&view=rev
Log:
[InstCombine] allow lengthening of insertelement to eliminate shuffles

As noted in post-commit comments for D52548, the limitation on 
increasing vector length can be applied by opcode.
As a first step, this patch only allows insertelement to be
widened because that has no logical downsides for IR and has 
little risk of pessimizing codegen.

This may cause PR39132 to go into hiding during a full compile,
but that bug is not fixed.

Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
    llvm/trunk/test/Transforms/InstCombine/vec_shuffle.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineVectorOps.cpp?rev=343406&r1=343405&r2=343406&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineVectorOps.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineVectorOps.cpp Sun Sep 30 06:50:42 2018
@@ -918,6 +918,13 @@ static bool canEvaluateShuffled(Value *V
     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;
@@ -1464,8 +1471,7 @@ Instruction *InstCombiner::visitShuffleV
     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);
   }

Modified: llvm/trunk/test/Transforms/InstCombine/vec_shuffle.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/vec_shuffle.ll?rev=343406&r1=343405&r2=343406&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/vec_shuffle.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/vec_shuffle.ll Sun Sep 30 06:50:42 2018
@@ -218,14 +218,13 @@ define <3 x i32> @div_wider(i32 %y, i32
   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