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

[Juan Manuel Martinez Caamaño via llvm-commits]

================
@@ -2090,6 +2097,146 @@ 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;
+
+  // Only optimize the pattern mbcnt.lo(~0, 0) which counts active lanes with
+  // lower IDs.
+  if (!match(&I,
+             m_Intrinsic<Intrinsic::amdgcn_mbcnt_lo>(m_AllOnes(), m_Zero())))
+    return false;
+
+  unsigned Wave = ST.getWavefrontSize();
+
+  if (auto MaybeX = ST.getReqdWorkGroupSize(F, 0)) {
+    unsigned XLen = *MaybeX;
+
+    // When XLen == wave_size, each work group contains exactly one wave, so
+    // mbcnt.lo(~0, 0) directly equals the workitem ID within the group.
+    if (XLen == Wave) {
+      Function *WorkitemIdFn = Intrinsic::getOrInsertDeclaration(
+          I.getModule(), Intrinsic::amdgcn_workitem_id_x);
+      CallInst *NewCall = CallInst::Create(WorkitemIdFn, I.getName());
+      ReplaceInstWithInst(&I, NewCall);
+      ST.makeLIDRangeMetadata(NewCall);
+      return true;
+    }
+    // When work group evenly splits into waves and wave size is power-of-2,
+    // we can compute lane ID within wave using bit masking:
+    // lane_id = workitem.id.x & (wave_size - 1).
+    if (ST.hasWavefrontsEvenlySplittingXDim(F, /*RequiresUniformYZ=*/true)) {
+      if (isPowerOf2_32(Wave)) {
----------------
jmmartinez wrote:

Sorry for the mix, I was talking about the `isPowerOf2_32(Wave)`

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