[InstCombine][x86] try harder to convert blendv intrinsic to generic IR (PR38814)

Missing optimizations with blendv are shown in:
https://bugs.llvm.org/show_bug.cgi?id=38814

If this works, it's an easier and more powerful solution than adding pattern matching 
for a few special cases in the backend. The potential danger with this transform in IR
is that the condition value can get separated from the select, and the backend might 
not be able to make a blendv out of it again. I don't think that's too likely, but 
I've kept this patch minimal with a 'TODO', so we can test that theory in the wild 
before expanding the transform.

Differential Revision: https://reviews.llvm.org/D52059

llvm-svn: 342324

1 files changed, 15 insertions, 7 deletions

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 4755e8509b3..bff46597266 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -2929,16 +2929,10 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
   case Intrinsic::x86_avx_blendv_ps_256:
   case Intrinsic::x86_avx_blendv_pd_256:
   case Intrinsic::x86_avx2_pblendvb: {
-    // Convert blendv* to vector selects if the mask is constant.
-    // This optimization is convoluted because the intrinsic is defined as
-    // getting a vector of floats or doubles for the ps and pd versions.
-    // FIXME: That should be changed.
-
+    // fold (blend A, A, Mask) -> A
     Value *Op0 = II->getArgOperand(0);
     Value *Op1 = II->getArgOperand(1);
     Value *Mask = II->getArgOperand(2);
-
-    // fold (blend A, A, Mask) -> A
     if (Op0 == Op1)
       return replaceInstUsesWith(CI, Op0);
 
@@ -2951,6 +2945,20 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
       Constant *NewSelector = getNegativeIsTrueBoolVec(ConstantMask);
       return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
     }
+
+    // Convert to a vector select if we can bypass casts and find a boolean
+    // vector condition value.
+    Value *BoolVec;
+    if (match(peekThroughBitcast(Mask), m_SExt(m_Value(BoolVec)))) {
+      auto *VTy = dyn_cast<VectorType>(BoolVec->getType());
+      if (VTy && VTy->getScalarSizeInBits() == 1 &&
+          VTy->getVectorNumElements() == II->getType()->getVectorNumElements())
+        return SelectInst::Create(BoolVec, Op1, Op0);
+      // TODO: If we can find a boolean vector condition with less elements,
+      //       then we can form a vector select by bitcasting Op0/Op1 to a
+      //       vector type with wider elements and bitcasting the result.
+    }
+
     break;
   }