[AMDGPU] Support GFX12 VDSDIR instructions WAITVMSRC operand in GCNHa…

…zardRecognizer (#77628)

Modify GCNHazardRecognizer::fixLdsDirectVMEMHazard() so the waitvsrc
operand
in gfx12 DS_PARAM_LOAD or DS_DIRECT_LOAD instructions is set
appropriately
depending on whether a hazard is found or not, rather than inserting an
S_WAITCNT_DEPCTR instruction if a hazard needs to be mitigated.

Co-authored-by: Stephen Thomas <Stephen.Thomas@amd.com>

llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp

@@ -1450,20 +1450,27 @@ bool GCNHazardRecognizer::fixLdsDirectVMEMHazard(MachineInstr *MI) {
       return false;
     return I.readsRegister(VDSTReg, &TRI) || I.modifiesRegister(VDSTReg, &TRI);
   };
-  auto IsExpiredFn = [](const MachineInstr &I, int) {
+  bool LdsdirCanWait = ST.hasLdsWaitVMSRC();
+  auto IsExpiredFn = [this, LdsdirCanWait](const MachineInstr &I, int) {
     return SIInstrInfo::isVALU(I) || SIInstrInfo::isEXP(I) ||
            (I.getOpcode() == AMDGPU::S_WAITCNT && !I.getOperand(0).getImm()) ||
            (I.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR &&
-            AMDGPU::DepCtr::decodeFieldVmVsrc(I.getOperand(0).getImm()) == 0);
+            AMDGPU::DepCtr::decodeFieldVmVsrc(I.getOperand(0).getImm()) == 0) ||
+           (LdsdirCanWait && SIInstrInfo::isLDSDIR(I) &&
+            !TII.getNamedOperand(I, AMDGPU::OpName::waitvsrc)->getImm());
   };
 
   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
       std::numeric_limits<int>::max())
     return false;
 
-  BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
-          TII.get(AMDGPU::S_WAITCNT_DEPCTR))
-      .addImm(AMDGPU::DepCtr::encodeFieldVmVsrc(0));
+  if (LdsdirCanWait) {
+    TII.getNamedOperand(*MI, AMDGPU::OpName::waitvsrc)->setImm(0);
+  } else {
+    BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
+            TII.get(AMDGPU::S_WAITCNT_DEPCTR))
+        .addImm(AMDGPU::DepCtr::encodeFieldVmVsrc(0));
+  }
 
   return true;
 }

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -1128,6 +1128,8 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
 
   bool hasLdsDirect() const { return getGeneration() >= GFX11; }
 
+  bool hasLdsWaitVMSRC() const { return getGeneration() >= GFX12; }
+
   bool hasVALUPartialForwardingHazard() const {
     return getGeneration() >= GFX11;
   }