[llvm] [AMDGPU] Refactor GFX11 VALU Mask Hazard Waitcnt Merging (PR #169213)

[Carl Ritson via llvm-commits]

================
@@ -437,9 +438,121 @@ class AMDGPUWaitSGPRHazards {
     return Changed;
   }
 
+  bool runWaitMerging(MachineFunction &MF) {
+    // Perform per-block merging of existing s_waitcnt_depctr instructions.
+    // Track set of SGPR writes before a given wait instruction, and search
+    // for reads of these SGPRs.
+    // Move the wait to just before the read to improve pipelining.
+    // If no related reads occur before subsequent wait then merged waits.
+    const unsigned ConstantMaskBits = AMDGPU::DepCtr::encodeFieldSaSdst(
+        AMDGPU::DepCtr::encodeFieldVaSdst(
+            AMDGPU::DepCtr::encodeFieldVaVcc(0, *ST), 0),
+        0);
+    const unsigned VccLoIdx = *sgprNumber(AMDGPU::VCC_LO, *TRI);
+    const unsigned VccHiIdx = *sgprNumber(AMDGPU::VCC_HI, *TRI);
+    bool Changed = false;
+    for (MachineBasicBlock &MBB : MF) {
+      std::bitset<128> WriteSet, PendingSALUWriteSet, PendingVALUWriteSet;
+      MachineInstr *PrevWait = nullptr;
+
+      auto CommitWrites = [&](unsigned Mask) {
+        if (!AMDGPU::DepCtr::decodeFieldSaSdst(Mask))
+          WriteSet |= PendingSALUWriteSet;
+        bool VccLoBit = WriteSet[VccLoIdx];
+        bool VccHiBit = WriteSet[VccHiIdx];
+        if (!AMDGPU::DepCtr::decodeFieldVaSdst(Mask)) {
+          WriteSet |= PendingVALUWriteSet;
+          WriteSet.set(VccLoIdx, VccLoBit);
+          WriteSet.set(VccLoIdx, VccHiBit);
+        }
+        if (!AMDGPU::DepCtr::decodeFieldVaVcc(Mask)) {
+          WriteSet.set(VccLoIdx, VccLoBit | PendingVALUWriteSet[VccLoIdx]);
+          WriteSet.set(VccHiIdx, VccHiBit | PendingVALUWriteSet[VccHiIdx]);
+        }
+        // Clear all pending writes
+        PendingSALUWriteSet.reset();
+        PendingVALUWriteSet.reset();
+      };
+
+      for (MachineBasicBlock::instr_iterator MI = MBB.instr_begin(),
+                                             E = MBB.instr_end();
+           MI != E; ++MI) {
+        if (MI->isMetaInstruction())
+          continue;
+
+        if (MI->getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && !MI->isBundled() &&
----------------
perlfu wrote:

Done

https://github.com/llvm/llvm-project/pull/169213