[AMDGPU] w/a for s_setreg_b32 gfx1250 hazard with MODE register (#153879)

Author: rampitec

llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp

@@ -1206,6 +1206,8 @@ void GCNHazardRecognizer::fixHazards(MachineInstr *MI) {
     fixDsAtomicAsyncBarrierArriveB64(MI);
   if (ST.hasScratchBaseForwardingHazard())
     fixScratchBaseForwardingHazard(MI);
+  if (ST.setRegModeNeedsVNOPs())
+    fixSetRegMode(MI);
 }
 
 static bool isVCmpXWritesExec(const SIInstrInfo &TII, const SIRegisterInfo &TRI,
@@ -3546,3 +3548,13 @@ bool GCNHazardRecognizer::fixScratchBaseForwardingHazard(MachineInstr *MI) {
           AMDGPU::DepCtr::encodeFieldSaSdst(0), 0));
   return true;
 }
+
+bool GCNHazardRecognizer::fixSetRegMode(MachineInstr *MI) {
+  if (!isSSetReg(MI->getOpcode()) ||
+      MI->getOperand(1).getImm() != AMDGPU::Hwreg::ID_MODE)
+    return false;
+
+  BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII.get(AMDGPU::V_NOP_e32));
+  BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII.get(AMDGPU::V_NOP_e32));
+  return true;
+}

llvm/lib/Target/AMDGPU/GCNHazardRecognizer.h

@@ -113,6 +113,7 @@ class GCNHazardRecognizer final : public ScheduleHazardRecognizer {
   bool fixGetRegWaitIdle(MachineInstr *MI);
   bool fixDsAtomicAsyncBarrierArriveB64(MachineInstr *MI);
   bool fixScratchBaseForwardingHazard(MachineInstr *MI);
+  bool fixSetRegMode(MachineInstr *MI);
 
   int checkMAIHazards(MachineInstr *MI);
   int checkMAIHazards908(MachineInstr *MI);

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -1345,6 +1345,10 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
 
   bool hasVALUReadSGPRHazard() const { return GFX12Insts && !GFX1250Insts; }
 
+  bool setRegModeNeedsVNOPs() const {
+    return GFX1250Insts && getGeneration() == GFX12;
+  }
+
   /// Return if operations acting on VGPR tuples require even alignment.
   bool needsAlignedVGPRs() const { return GFX90AInsts || GFX1250Insts; }
 

llvm/test/CodeGen/AMDGPU/hazards-gfx1250.mir

@@ -493,3 +493,57 @@ body: |
     liveins: $vgpr0
     $vgpr0 = V_ADD_U32_e32 $src_flat_scratch_base_lo, $vgpr0, implicit $exec
 ...
+
+---
+name: s_setreg_b32_hwreg_mode
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $sgpr0
+    ; GCN-LABEL: name: s_setreg_b32_hwreg_mode
+    ; GCN: liveins: $sgpr0
+    ; GCN-NEXT: {{  $}}
+    ; GCN-NEXT: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: S_SETREG_B32 $sgpr0, 1, implicit-def $mode, implicit $mode
+    S_SETREG_B32 $sgpr0, 1, implicit-def $mode, implicit $mode
+...
+
+---
+name: s_setreg_b32_mode
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $sgpr0
+    ; GCN-LABEL: name: s_setreg_b32_mode
+    ; GCN: liveins: $sgpr0
+    ; GCN-NEXT: {{  $}}
+    ; GCN-NEXT: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: S_SETREG_B32_mode $sgpr0, 1, implicit-def $mode, implicit $mode
+    S_SETREG_B32_mode $sgpr0, 1, implicit-def $mode, implicit $mode
+...
+
+---
+name: s_setreg_imm32_b32_hwreg_mode
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: s_setreg_imm32_b32_hwreg_mode
+    ; GCN: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: S_SETREG_IMM32_B32 1, 1, implicit-def $mode, implicit $mode
+    S_SETREG_IMM32_B32 1, 1, implicit-def $mode, implicit $mode
+...
+
+---
+name: s_setreg_imm32_b32_mode
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GCN-LABEL: name: s_setreg_imm32_b32_mode
+    ; GCN: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: V_NOP_e32 implicit $exec
+    ; GCN-NEXT: S_SETREG_IMM32_B32_mode 1, 1, implicit-def $mode, implicit $mode
+    S_SETREG_IMM32_B32_mode 1, 1, implicit-def $mode, implicit $mode
+...