AMDGPU/SI: Handle div_fmas hazard in GCNHazardRecognizer

[arsenm]

LGTM

lib/Target/AMDGPU/GCNHazardRecognizer.cpp

281

const

[arsenm]

I think the mir tests are supposed to go in test/CodeGen/MIR/AMDGPU

lib/Target/AMDGPU/GCNHazardRecognizer.h

class GCNHazardRecognizer final : public ScheduleHazardRecognizer {
   int getWaitStatesSinceDef(unsigned Reg,
                             function_ref<bool(MachineInstr *)> IsHazardDef =
                                 [](MachineInstr *) { return true; });
 
   int checkSMEMSoftClauseHazards(MachineInstr *SMEM);
   int checkSMRDHazards(MachineInstr *SMRD);
   int checkVMEMHazards(MachineInstr* VMEM);
   int checkDPPHazards(MachineInstr *DPP);
+  int checkDivFMasHazards(MachineInstr *DivFMas);
 public:
   GCNHazardRecognizer(const MachineFunction &MF);
   // We can only issue one instruction per cycle.
   bool atIssueLimit() const override { return true; }
   void EmitInstruction(SUnit *SU) override;
   void EmitInstruction(MachineInstr *MI) override;
   HazardType getHazardType(SUnit *SU, int Stalls) override;
   void EmitNoop() override;

lib/Target/AMDGPU/GCNHazardRecognizer.cpp

 void GCNHazardRecognizer::EmitInstruction(SUnit *SU) {
   EmitInstruction(SU->getInstr());
 }
 
 void GCNHazardRecognizer::EmitInstruction(MachineInstr *MI) {
   CurrCycleInstr = MI;
 }
 
+static bool isDivFMas(unsigned Opcode) {
+  return Opcode == AMDGPU::V_DIV_FMAS_F32 || Opcode == AMDGPU::V_DIV_FMAS_F64;
+}
+
 ScheduleHazardRecognizer::HazardType
 GCNHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
   MachineInstr *MI = SU->getInstr();
 
   if (SIInstrInfo::isSMRD(*MI) && checkSMRDHazards(MI) > 0)
     return NoopHazard;
 
   if (SIInstrInfo::isVMEM(*MI) && checkVMEMHazards(MI) > 0)
     return NoopHazard;
 
   if (SIInstrInfo::isDPP(*MI) && checkDPPHazards(MI) > 0)
     return NoopHazard;
 
+  if (isDivFMas(MI->getOpcode()) && checkDivFMasHazards(MI) > 0)
+    return NoopHazard;
+
   return NoHazard;
 }
 
 unsigned GCNHazardRecognizer::PreEmitNoops(SUnit *SU) {
   return PreEmitNoops(SU->getInstr());
 }
 
 unsigned GCNHazardRecognizer::PreEmitNoops(MachineInstr *MI) {
   if (SIInstrInfo::isSMRD(*MI))
     return std::max(0, checkSMRDHazards(MI));
 
   if (SIInstrInfo::isVMEM(*MI))
     return std::max(0, checkVMEMHazards(MI));
 
   if (SIInstrInfo::isDPP(*MI))
     return std::max(0, checkDPPHazards(MI));
 
+  if (isDivFMas(MI->getOpcode()))
+    return std::max(0, checkDivFMasHazards(MI));
+
   return 0;
 }
 
 void GCNHazardRecognizer::EmitNoop() {
   EmittedInstrs.push_front(nullptr);
 }
 
 void GCNHazardRecognizer::AdvanceCycle() {
if (!Use.isReg() || !TRI->isVGPR(MF.getRegInfo(), Use.getReg()))
       continue;
     int WaitStatesNeededForUse =
         DppVgprWaitStates - getWaitStatesSinceDef(Use.getReg());
     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
   }
 
   return WaitStatesNeeded;
 }
+
+int GCNHazardRecognizer::checkDivFMasHazards(MachineInstr *DivFMas) {
+  const SIInstrInfo *TII = ST.getInstrInfo();
+
+  // v_div_fmas requires 4 wait states after a write to vcc from a VALU
+  // instruction.
+  int DivFMasWaitStates = 4;
+  auto IsHazardDefFn = [TII] (MachineInstr *MI) { return TII->isVALU(*MI); };
+  int WaitStatesNeeded = getWaitStatesSinceDef(AMDGPU::VCC, IsHazardDefFn);
+
+  return DivFMasWaitStates - WaitStatesNeeded;
+}

test/CodeGen/AMDGPU/inserted-wait-states.mir

+# RUN: llc -march=amdgcn -run-pass post-RA-hazard-rec  %s -o - | FileCheck %s
+
+# CHECK-LABEL: bb.0:
+# CHECK: S_MOV_B64
+# CHECK-NOT: S_NOP
+# CHECK: V_DIV_FMAS
+
+# CHECK-LABEL: bb.1:
+# CHECK: V_CMP_EQ_I32
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: V_DIV_FMAS_F32
+
+# CHECK-LABEL: bb.2:
+# CHECK: V_CMP_EQ_I32
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: V_DIV_FMAS_F32
+
+# CHECK-LABEL: bb.3:
+# CHECK: V_DIV_SCALE_F32
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: S_NOP
+# CHECK: V_DIV_FMAS_F32
+--- |
+  define void @test0() { ret void }
+...
+---
+name: test0
+
+body: |
+  bb.0:
+    successors: %bb.1
+    %vcc = S_MOV_B64 0
+    %vgpr0 = V_DIV_FMAS_F32 0, %vgpr1, 0, %vgpr2, 0, %vgpr3, 0, 0, implicit %vcc, implicit %exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    successors: %bb.2
+    implicit %vcc = V_CMP_EQ_I32_e32 %vgpr1, %vgpr2, implicit %exec
+    %vgpr0 = V_DIV_FMAS_F32 0, %vgpr1, 0, %vgpr2, 0, %vgpr3, 0, 0, implicit %vcc, implicit %exec
+    S_BRANCH %bb.2
+
+  bb.2:
+    successors: %bb.3
+    %vcc = V_CMP_EQ_I32_e64 %vgpr1, %vgpr2, implicit %exec
+    %vgpr0 = V_DIV_FMAS_F32 0, %vgpr1, 0, %vgpr2, 0, %vgpr3, 0, 0, implicit %vcc, implicit %exec
+    S_BRANCH %bb.3
+
+  bb.3:
+    %vgpr4, %vcc = V_DIV_SCALE_F32 0, %vgpr1, 0, %vgpr1, 0, %vgpr3, 0, 0, implicit %exec
+    %vgpr0 = V_DIV_FMAS_F32 0, %vgpr1, 0, %vgpr2, 0, %vgpr3, 0, 0, implicit %vcc, implicit %exec
+    S_ENDPGM
+...