[llvm] [InstCombine][X86] Fold blendv(x, y, shuffle(bitcast(sext(m)))) -> select(shuffle(m),x,y) (PR #96882)

[via llvm-commits]

================
@@ -2800,6 +2800,23 @@ X86TTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
       return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
     }
 
+    // Peek through a one-use shuffle - VectorCombine should have simplified
+    // this for cases where we're splitting wider vectors to use blendv
+    // intrinsics.
+    Value *MaskSrc = nullptr;
+    ArrayRef<int> ShuffleMask;
+    if (match(Mask, PatternMatch::m_OneUse(PatternMatch::m_Shuffle(
+                        PatternMatch::m_Value(MaskSrc), PatternMatch::m_Undef(),
+                        PatternMatch::m_Mask(ShuffleMask))))) {
----------------
goldsteinn wrote:

nit: We should probably just start using `PatternMatch` namespace at the top of the function. That can be a later NFC though.

https://github.com/llvm/llvm-project/pull/96882