[llvm] [InstCombine] Missing optimization: fold mul (select a, b), (select b, a) to mul a, b (PR #74953)

[via llvm-commits]

================
@@ -4217,3 +4220,25 @@ InstCombinerImpl::transformCallThroughTrampoline(CallBase &Call,
   Call.setCalledFunction(FTy, NestF);
   return &Call;
 }
+
+// op(select(%v, %x, %y), select(%v, %y, %x)) --> op(%x, %y)
+Instruction *
+InstCombinerImpl::foldCommutativeIntrinsicOverSelects(IntrinsicInst &II) {
+  if (!II.isCommutative()) {
+    return nullptr;
+  }
+
+  Value *A, *B, *C, *D, *E, *F;
+  bool LHSIsSelect =
+      match(II.getOperand(0), m_Select(m_Value(A), m_Value(B), m_Value(C)));
+  bool RHSIsSelect =
+      match(II.getOperand(1), m_Select(m_Value(D), m_Value(E), m_Value(F)));
+
----------------
goldsteinn wrote:

Maybe use `m_Specific` here instead of manually checking in condition below.

https://github.com/llvm/llvm-project/pull/74953