[llvm] [X86] Handle BSF/BSR "zero-input pass through" behaviour (PR #123623)

[Simon Pilgrim via llvm-commits]

================
@@ -5391,9 +5392,9 @@ bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
         //      ...                 // EFLAGS not changed
         //      testl %eax, %eax    // <-- can be removed
         if (IsCmpZero) {
-          NewCC = isUseDefConvertible(Inst);
-          if (NewCC != X86::COND_INVALID && Inst.getOperand(1).isReg() &&
-              Inst.getOperand(1).getReg() == SrcReg) {
+          std::tie(NewCC, OpNo) = isUseDefConvertible(Inst);
+          if (NewCC != X86::COND_INVALID && Inst.getOperand(OpNo).isReg() &&
+              Inst.getOperand(OpNo).getReg() == SrcReg) {
----------------
RKSimon wrote:

This is at the MI level, and the passthrough/fallback operand is tied to the destination reg (like any x86 binop) - it'd be weird to have a commutation and I expect could lead to further problems.

https://github.com/llvm/llvm-project/pull/123623