[llvm] [AMDGPU]: Rewrite mbcnt_lo/mbcnt_hi to work item ID where applicable (PR #160496)

[Teja Alaghari via llvm-commits]

================
@@ -1312,9 +1317,122 @@ GCNTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
     break;
   }
   case Intrinsic::amdgcn_mbcnt_hi: {
-    // exec_hi is all 0, so this is just a copy.
-    if (ST->isWave32())
+    // exec_hi is all 0, so this is just a copy on wave32.
+    if (ST && ST->isWave32())
       return IC.replaceInstUsesWith(II, II.getArgOperand(1));
+
+    // Pattern: mbcnt.hi(~0, mbcnt.lo(~0, 0))
+    if (auto *HiArg1 = dyn_cast<CallInst>(II.getArgOperand(1))) {
+      Function *CalledF = HiArg1->getCalledFunction();
+      bool IsMbcntLo = false;
+      if (CalledF) {
+        // Fast-path: if this is a declared intrinsic, check the intrinsic ID.
+        if (CalledF->getIntrinsicID() == Intrinsic::amdgcn_mbcnt_lo) {
+          IsMbcntLo = true;
+        } else {
+          // Fallback: accept a declared function with the canonical name, but
----------------
TejaX-Alaghari wrote:

Initially, I had sketched to handle explicit calls to declared `@llvm.amdgcn.mbcnt.*` functions (not intrinsics). So call sites that are declared functions (not recognized IntrinsicInsts) are also handled.

But now I recognize its totally redundant! So, removed it now. 

https://github.com/llvm/llvm-project/pull/160496