[llvm] [InstCombine] Improve coverage of `foldSelectValueEquivalence` (PR #88298)

[via llvm-commits]

================
@@ -1308,11 +1310,52 @@ Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel,
       if (match(V, m_ImmConstant()))
         return replaceOperand(Sel, Swapped ? 2 : 1, V);
 
-      // If NewOp is a constant and OldOp is not replace iff NewOp doesn't
-      // contain and undef/poison elements.
-      if (match(NewOp, m_ImmConstant()) &&
-          isGuaranteedNotToBeUndefOrPoison(NewOp, SQ.AC, &Sel, &DT))
+      // We can't do any further replacement if NewOp may be undef/poison.
+      if (!isGuaranteedNotToBeUndefOrPoison(NewOp, SQ.AC, &Sel, &DT))
+        return nullptr;
+
+      // If NewOp is a constant, replace.
+      if (match(NewOp, m_ImmConstant()))
+        return replaceOperand(Sel, Swapped ? 2 : 1, V);
+
+      // If we simplified the TrueArm -> NewOp then replace.
+      // This handles things like `select`/`min`/`max`/`or`/`and`/etc...
+      if (NewOp == V)
----------------
goldsteinn wrote:

Hmm, maybe a case like `(select (icmp eq (or x, y), y), (or x, y), Q)`. Where we replace `y` with `(or x, y)` and then `(or x, y), x/y)` -> `(or x, y)`.

I think for it to inf loop through would need `NewOp == TrueArm`. I think we can probably safely just return null for that case.

Either way, ill run csmith for a bit.

What do you mean 128C server?

https://github.com/llvm/llvm-project/pull/88298