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

[Matt Arsenault via llvm-commits]

================
@@ -2090,6 +2097,139 @@ INITIALIZE_PASS_DEPENDENCY(UniformityInfoWrapperPass)
 INITIALIZE_PASS_END(AMDGPUCodeGenPrepare, DEBUG_TYPE, "AMDGPU IR optimizations",
                     false, false)
 
+/// Optimize mbcnt.lo calls on wave32 architectures for lane ID computation.
+bool AMDGPUCodeGenPrepareImpl::visitMbcntLo(IntrinsicInst &I) const {
+  // Abort if wave size is not known at compile time.
+  if (!ST.isWaveSizeKnown())
+    return false;
+
+  // This optimization only applies to wave32 targets where mbcnt.lo operates on
+  // the full execution mask.
+  if (!ST.isWave32())
+    return false;
----------------
arsenm wrote:

The ST.isWaveSizeKnown check above is redundant with this 

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