InstCombine: Fold more shuffles of shuffles.

Always fold a shuffle-of-shuffle into a single shuffle when there's only one
input vector in the first place. Continue to be more conservative when there's
multiple inputs.

rdar://13402653
PR15866

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@180802 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index de8a3acdbd855150dc49e727295d12a7f97c5bf0..56243059a618a57d904433933bd09a3b771845d0 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -614,11 +614,16 @@ Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
   // we are absolutely afraid of producing a shuffle mask not in the input
   // program, because the code gen may not be smart enough to turn a merged
   // shuffle into two specific shuffles: it may produce worse code.  As such,
-  // we only merge two shuffles if the result is either a splat or one of the
-  // input shuffle masks.  In this case, merging the shuffles just removes
-  // one instruction, which we know is safe.  This is good for things like
+  // we only merge two shuffles if the result is a splat, one of the input
+  // input shuffle masks, or if there's only one input to the shuffle.
+  // In this case, merging the shuffles just removes one instruction, which
+  // we know is safe.  This is good for things like
   // turning: (splat(splat)) -> splat, or
   // merge(V[0..n], V[n+1..2n]) -> V[0..2n]
+  //
+  // FIXME: This is almost certainly far, far too conservative. We should
+  // have a better model. Perhaps a TargetTransformInfo hook to ask whether
+  // a shuffle is considered OK?
   ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS);
   ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS);
   if (LHSShuffle)
@@ -743,8 +748,10 @@ Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
   }
 
   // If the result mask is equal to one of the original shuffle masks,
-  // or is a splat, do the replacement.
-  if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) {
+  // or is a splat, do the replacement. Similarly, if there is only one
+  // input vector, go ahead and do the folding.
+  if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask ||
+      isa<UndefValue>(RHS)) {
     SmallVector<Constant*, 16> Elts;
     Type *Int32Ty = Type::getInt32Ty(SVI.getContext());
     for (unsigned i = 0, e = newMask.size(); i != e; ++i) {

diff --git a/test/Transforms/BBVectorize/simple.ll b/test/Transforms/BBVectorize/simple.ll
index a447908d16cc7ff7d4a37221d79e77ab3622682d..a22ad15940cf3b156d29c451f2b6c4ffa19c5c50 100644 (file)
--- a/test/Transforms/BBVectorize/simple.ll
+++ b/test/Transforms/BBVectorize/simple.ll
@@ -139,11 +139,10 @@ define <8 x i8> @test6(<8 x i8> %A1, <8 x i8> %A2, <8 x i8> %B1, <8 x i8> %B2) {
 ; CHECK: %Z1 = add <16 x i8> %Y1, %X1.v.i1
         %Q1 = shufflevector <8 x i8> %Z1, <8 x i8> %Z2, <8 x i32> <i32 15, i32 8, i32 6, i32 1, i32 13, i32 10, i32 4, i32 3>
         %Q2 = shufflevector <8 x i8> %Z2, <8 x i8> %Z2, <8 x i32> <i32 6, i32 7, i32 0, i32 1, i32 2, i32 4, i32 4, i32 1>
-; CHECK: %Q1.v.i1 = shufflevector <16 x i8> %Z1, <16 x i8> undef, <16 x i32> <i32 8, i32 undef, i32 10, i32 undef, i32 undef, i32 13, i32 undef, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
-; CHECK: %Q1 = shufflevector <16 x i8> %Z1, <16 x i8> %Q1.v.i1, <16 x i32> <i32 23, i32 16, i32 6, i32 1, i32 21, i32 18, i32 4, i32 3, i32 14, i32 15, i32 8, i32 9, i32 10, i32 12, i32 12, i32 9>
-       %R  = mul <8 x i8> %Q1, %Q2
-; CHECK: %Q1.v.r1 = shufflevector <16 x i8> %Q1, <16 x i8> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
-; CHECK: %Q1.v.r2 = shufflevector <16 x i8> %Q1, <16 x i8> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+        %R  = mul <8 x i8> %Q1, %Q2
+; CHECK:  %Q1.v.i1 = shufflevector <16 x i8> %Z1, <16 x i8> undef, <16 x i32> <i32 8, i32 undef, i32 10, i32 undef, i32 undef, i32 13, i32 undef, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; CHECK:  %Q1.v.r1 = shufflevector <16 x i8> %Z1, <16 x i8> %Q1.v.i1, <8 x i32> <i32 23, i32 16, i32 6, i32 1, i32 21, i32 18, i32 4, i32 3>
+; CHECK:  %Q1.v.r2 = shufflevector <16 x i8> %Z1, <16 x i8> undef, <8 x i32> <i32 14, i32 15, i32 8, i32 9, i32 10, i32 12, i32 12, i32 9>
 ; CHECK: %R = mul <8 x i8> %Q1.v.r1, %Q1.v.r2
        ret <8 x i8> %R
 ; CHECK: ret <8 x i8> %R

diff --git a/test/Transforms/InstCombine/vec_shuffle.ll b/test/Transforms/InstCombine/vec_shuffle.ll
index 37d4d56e9131291bb847aec217f9c1b4db10530f..37c19f1558b930e5528e598eb12d312c3e5675bd 100644 (file)
--- a/test/Transforms/InstCombine/vec_shuffle.ll
+++ b/test/Transforms/InstCombine/vec_shuffle.ll
@@ -86,14 +86,14 @@ define <4 x i8> @test9(<16 x i8> %tmp6) nounwind {
 }
 
 ; Same as test9, but make sure that "undef" mask values are not confused with
-; mask values of 2*N, where N is the mask length.  These shuffles should not
-; be folded (because [8,9,4,8] may not be a mask supported by the target).
-define <4 x i8> @test9a(<16 x i8> %tmp6) nounwind {
+; mask values of 2*N, where N is the mask length of the result.  Make sure when
+; folding these shuffles that 'undef' mask values stay that way in the result
+; instead of getting mapped to the 2*N'th entry of the source.
+define <4 x i8> @test9a(<16 x i8> %in, <16 x i8> %in2) nounwind {
 ; CHECK: @test9a
-; CHECK-NEXT: shufflevector
-; CHECK-NEXT: shufflevector
+; CHECK-NEXT: shufflevector <16 x i8> %in, <16 x i8> %in2, <4 x i32> <i32 16, i32 9, i32 4, i32 undef>
 ; CHECK-NEXT: ret
-       %tmp7 = shufflevector <16 x i8> %tmp6, <16 x i8> undef, <4 x i32> < i32 undef, i32 9, i32 4, i32 8 >            ; <<4 x i8>> [#uses=1]
+       %tmp7 = shufflevector <16 x i8> %in, <16 x i8> %in2, <4 x i32> < i32 undef, i32 9, i32 4, i32 16 >              ; <<4 x i8>> [#uses=1]
        %tmp9 = shufflevector <4 x i8> %tmp7, <4 x i8> undef, <4 x i32> < i32 3, i32 1, i32 2, i32 0 >          ; <<4 x i8>> [#uses=1]
        ret <4 x i8> %tmp9
 }