[llvm-branch-commits] [llvm] [AMDGPU] Apply occupancy-aware register allocation anti-hints (PR #218074)

[Syadus Sefat via llvm-branch-commits]

================
@@ -4173,6 +4178,130 @@ bool SIRegisterInfo::getRegAllocationHints(Register VirtReg,
   }
 }
 
+bool SIRegisterInfo::shouldApplyAntiHints(
+    const MachineFunction &MF, unsigned NumAllocatedVGPRs,
+    unsigned &MaxVGPRsForCurrentOccupancy) const {
+
+  const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+  unsigned DynamicVGPRBlockSize = MFI->getDynamicVGPRBlockSize();
+  unsigned TargetOccupancy = MFI->getOccupancy();
+  unsigned CurrentOccupancy =
+      ST.getOccupancyWithNumVGPRs(NumAllocatedVGPRs, DynamicVGPRBlockSize);
+  MaxVGPRsForCurrentOccupancy =
+      ST.getMaxNumVGPRs(CurrentOccupancy, DynamicVGPRBlockSize);
+
+  LLVM_DEBUG(dbgs() << "anti-hints: " << NumAllocatedVGPRs
+                    << " VGPRs allocated, target occupancy " << TargetOccupancy
+                    << ", current occupancy " << CurrentOccupancy << '\n');
+
+  // If we are already at lowest occupancy, then there is no need to protect
+  // against occupancy regression.
+  if (CurrentOccupancy == 1)
+    return true;
+
+  // Set max VGPRs for target and current occupancy to early bail out if we are
+  // close to the limit.
+  unsigned MaxVGPRsCutOffForTargetOccupancy =
+      (ST.getMaxNumVGPRs(TargetOccupancy, DynamicVGPRBlockSize) * 80) / 100;
+  unsigned MaxVGPRsCutOffForCurrentOccupancy =
+      (MaxVGPRsForCurrentOccupancy * 95) / 100;
----------------
mssefat wrote:

I added the comments. For now, I would propose to keep them as internal heuristics rather than adding cli options in this PR. The values were hand-waved on a small benchmark but broader validation is still needed before exposing them as the user tuning controls. If further benchmarking indicates that users would benefit from adjusting them I can add hidden options in a follow-up PR.

https://github.com/llvm/llvm-project/pull/218074