[llvm] [AMDGPU] Analyze REG_SEQUENCE To Remove Redundant CMP Instructions (PR #167364)

[via llvm-commits]

================
@@ -11267,10 +11304,103 @@ bool SIInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
   if (!SrcReg || SrcReg.isPhysical())
     return false;
 
-  if (SrcReg2 && !getFoldableImm(SrcReg2, *MRI, CmpValue))
+  if (SrcReg2 && !getFoldableImm(SrcReg2, *MRI, CmpValue)) {
+    std::pair<MachineInstr *, unsigned> RegSequence2;
----------------
LU-JOHN wrote:

Did any testcases/benchmarks utilize this deeper analysis of SrcReg2?  If real code doesn't need this we should simplify it.

https://github.com/llvm/llvm-project/pull/167364