1 files changed, 38 insertions, 96 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index a1e14d9..c8935f0 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -687,7 +687,8 @@ static void indirectCopyToAGPR(const SIInstrInfo &TII,
         if (!SafeToPropagate)
           break;
 
-        DefOp.setIsKill(false);
+        for (auto I = Def; I != MI; ++I)
+          I->clearRegisterKills(DefOp.getReg(), &RI);
       }
 
       MachineInstrBuilder Builder =
@@ -1625,41 +1626,6 @@ static unsigned getVGPRSpillSaveOpcode(unsigned Size) {
   }
 }
 
-static unsigned getAGPRSpillSaveOpcode(unsigned Size) {
-  switch (Size) {
-  case 4:
-    return AMDGPU::SI_SPILL_A32_SAVE;
-  case 8:
-    return AMDGPU::SI_SPILL_A64_SAVE;
-  case 12:
-    return AMDGPU::SI_SPILL_A96_SAVE;
-  case 16:
-    return AMDGPU::SI_SPILL_A128_SAVE;
-  case 20:
-    return AMDGPU::SI_SPILL_A160_SAVE;
-  case 24:
-    return AMDGPU::SI_SPILL_A192_SAVE;
-  case 28:
-    return AMDGPU::SI_SPILL_A224_SAVE;
-  case 32:
-    return AMDGPU::SI_SPILL_A256_SAVE;
-  case 36:
-    return AMDGPU::SI_SPILL_A288_SAVE;
-  case 40:
-    return AMDGPU::SI_SPILL_A320_SAVE;
-  case 44:
-    return AMDGPU::SI_SPILL_A352_SAVE;
-  case 48:
-    return AMDGPU::SI_SPILL_A384_SAVE;
-  case 64:
-    return AMDGPU::SI_SPILL_A512_SAVE;
-  case 128:
-    return AMDGPU::SI_SPILL_A1024_SAVE;
-  default:
-    llvm_unreachable("unknown register size");
-  }
-}
-
 static unsigned getAVSpillSaveOpcode(unsigned Size) {
   switch (Size) {
   case 4:
@@ -1707,22 +1673,20 @@ static unsigned getWWMRegSpillSaveOpcode(unsigned Size,
   return AMDGPU::SI_SPILL_WWM_V32_SAVE;
 }
 
-static unsigned getVectorRegSpillSaveOpcode(Register Reg,
-                                            const TargetRegisterClass *RC,
-                                            unsigned Size,
-                                            const SIRegisterInfo &TRI,
-                                            const SIMachineFunctionInfo &MFI) {
-  bool IsVectorSuperClass = TRI.isVectorSuperClass(RC);
+unsigned SIInstrInfo::getVectorRegSpillSaveOpcode(
+    Register Reg, const TargetRegisterClass *RC, unsigned Size,
+    const SIMachineFunctionInfo &MFI) const {
+  bool IsVectorSuperClass = RI.isVectorSuperClass(RC);
 
   // Choose the right opcode if spilling a WWM register.
   if (MFI.checkFlag(Reg, AMDGPU::VirtRegFlag::WWM_REG))
     return getWWMRegSpillSaveOpcode(Size, IsVectorSuperClass);
 
-  if (IsVectorSuperClass)
+  // TODO: Check if AGPRs are available
+  if (ST.hasMAIInsts())
     return getAVSpillSaveOpcode(Size);
 
-  return TRI.isAGPRClass(RC) ? getAGPRSpillSaveOpcode(Size)
-                             : getVGPRSpillSaveOpcode(Size);
+  return getVGPRSpillSaveOpcode(Size);
 }
 
 void SIInstrInfo::storeRegToStackSlot(
@@ -1770,8 +1734,8 @@ void SIInstrInfo::storeRegToStackSlot(
     return;
   }
 
-  unsigned Opcode = getVectorRegSpillSaveOpcode(VReg ? VReg : SrcReg, RC,
-                                                SpillSize, RI, *MFI);
+  unsigned Opcode =
+      getVectorRegSpillSaveOpcode(VReg ? VReg : SrcReg, RC, SpillSize, *MFI);
   MFI->setHasSpilledVGPRs();
 
   BuildMI(MBB, MI, DL, get(Opcode))
@@ -1854,41 +1818,6 @@ static unsigned getVGPRSpillRestoreOpcode(unsigned Size) {
   }
 }
 
-static unsigned getAGPRSpillRestoreOpcode(unsigned Size) {
-  switch (Size) {
-  case 4:
-    return AMDGPU::SI_SPILL_A32_RESTORE;
-  case 8:
-    return AMDGPU::SI_SPILL_A64_RESTORE;
-  case 12:
-    return AMDGPU::SI_SPILL_A96_RESTORE;
-  case 16:
-    return AMDGPU::SI_SPILL_A128_RESTORE;
-  case 20:
-    return AMDGPU::SI_SPILL_A160_RESTORE;
-  case 24:
-    return AMDGPU::SI_SPILL_A192_RESTORE;
-  case 28:
-    return AMDGPU::SI_SPILL_A224_RESTORE;
-  case 32:
-    return AMDGPU::SI_SPILL_A256_RESTORE;
-  case 36:
-    return AMDGPU::SI_SPILL_A288_RESTORE;
-  case 40:
-    return AMDGPU::SI_SPILL_A320_RESTORE;
-  case 44:
-    return AMDGPU::SI_SPILL_A352_RESTORE;
-  case 48:
-    return AMDGPU::SI_SPILL_A384_RESTORE;
-  case 64:
-    return AMDGPU::SI_SPILL_A512_RESTORE;
-  case 128:
-    return AMDGPU::SI_SPILL_A1024_RESTORE;
-  default:
-    llvm_unreachable("unknown register size");
-  }
-}
-
 static unsigned getAVSpillRestoreOpcode(unsigned Size) {
   switch (Size) {
   case 4:
@@ -1930,27 +1859,27 @@ static unsigned getWWMRegSpillRestoreOpcode(unsigned Size,
   if (Size != 4)
     llvm_unreachable("unknown wwm register spill size");
 
-  if (IsVectorSuperClass)
+  if (IsVectorSuperClass) // TODO: Always use this if there are AGPRs
     return AMDGPU::SI_SPILL_WWM_AV32_RESTORE;
 
   return AMDGPU::SI_SPILL_WWM_V32_RESTORE;
 }
 
-static unsigned
-getVectorRegSpillRestoreOpcode(Register Reg, const TargetRegisterClass *RC,
-                               unsigned Size, const SIRegisterInfo &TRI,
-                               const SIMachineFunctionInfo &MFI) {
-  bool IsVectorSuperClass = TRI.isVectorSuperClass(RC);
+unsigned SIInstrInfo::getVectorRegSpillRestoreOpcode(
+    Register Reg, const TargetRegisterClass *RC, unsigned Size,
+    const SIMachineFunctionInfo &MFI) const {
+  bool IsVectorSuperClass = RI.isVectorSuperClass(RC);
 
   // Choose the right opcode if restoring a WWM register.
   if (MFI.checkFlag(Reg, AMDGPU::VirtRegFlag::WWM_REG))
     return getWWMRegSpillRestoreOpcode(Size, IsVectorSuperClass);
 
-  if (IsVectorSuperClass)
+  // TODO: Check if AGPRs are available
+  if (ST.hasMAIInsts())
     return getAVSpillRestoreOpcode(Size);
 
-  return TRI.isAGPRClass(RC) ? getAGPRSpillRestoreOpcode(Size)
-                             : getVGPRSpillRestoreOpcode(Size);
+  assert(!RI.isAGPRClass(RC));
+  return getVGPRSpillRestoreOpcode(Size);
 }
 
 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
@@ -1998,7 +1927,7 @@ void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
   }
 
   unsigned Opcode = getVectorRegSpillRestoreOpcode(VReg ? VReg : DestReg, RC,
-                                                   SpillSize, RI, *MFI);
+                                                   SpillSize, *MFI);
   BuildMI(MBB, MI, DL, get(Opcode), DestReg)
       .addFrameIndex(FrameIndex)           // vaddr
       .addReg(MFI->getStackPtrOffsetReg()) // scratch_offset
@@ -6460,7 +6389,7 @@ bool SIInstrInfo::moveFlatAddrToVGPR(MachineInstr &Inst) const {
   if (OldSAddrIdx < 0)
     return false;
 
-  assert(isSegmentSpecificFLAT(Inst));
+  assert(isSegmentSpecificFLAT(Inst) || (isFLAT(Inst) && ST.hasFlatGVSMode()));
 
   int NewOpc = AMDGPU::getGlobalVaddrOp(Opc);
   if (NewOpc < 0)
@@ -6484,7 +6413,7 @@ bool SIInstrInfo::moveFlatAddrToVGPR(MachineInstr &Inst) const {
   if (OldVAddrIdx >= 0) {
     MachineOperand &VAddr = Inst.getOperand(OldVAddrIdx);
     VAddrDef = MRI.getUniqueVRegDef(VAddr.getReg());
-    if (!VAddrDef || VAddrDef->getOpcode() != AMDGPU::V_MOV_B32_e32 ||
+    if (!VAddrDef || !VAddrDef->isMoveImmediate() ||
         !VAddrDef->getOperand(1).isImm() ||
         VAddrDef->getOperand(1).getImm() != 0)
       return false;
@@ -6537,7 +6466,7 @@ bool SIInstrInfo::moveFlatAddrToVGPR(MachineInstr &Inst) const {
 // FIXME: Remove this when SelectionDAG is obsoleted.
 void SIInstrInfo::legalizeOperandsFLAT(MachineRegisterInfo &MRI,
                                        MachineInstr &MI) const {
-  if (!isSegmentSpecificFLAT(MI))
+  if (!isSegmentSpecificFLAT(MI) && !ST.hasFlatGVSMode())
     return;
 
   // Fixup SGPR operands in VGPRs. We only select these when the DAG divergence
@@ -10466,10 +10395,23 @@ bool SIInstrInfo::isGlobalMemoryObject(const MachineInstr *MI) const {
   return TargetInstrInfo::isGlobalMemoryObject(MI);
 }
 
+bool SIInstrInfo::isXDLWMMA(const MachineInstr &MI) const {
+  if (!isWMMA(MI) && !isSWMMAC(MI))
+    return false;
+
+  if (AMDGPU::isGFX1250(ST))
+    return AMDGPU::getWMMAIsXDL(MI.getOpcode());
+
+  return true;
+}
+
 bool SIInstrInfo::isXDL(const MachineInstr &MI) const {
   unsigned Opcode = MI.getOpcode();
 
-  if (!SIInstrInfo::isMAI(MI) || isDGEMM(Opcode) ||
+  if (AMDGPU::isGFX12Plus(ST))
+    return isDOT(MI) || isXDLWMMA(MI);
+
+  if (!isMAI(MI) || isDGEMM(Opcode) ||
       Opcode == AMDGPU::V_ACCVGPR_WRITE_B32_e64 ||
       Opcode == AMDGPU::V_ACCVGPR_READ_B32_e64)
     return false;