AMDGPU/GlobalISel: Implement select() for G_IMPLICIT_DEF

[nhaehnle]

Is the default implementation really unable to do this? That seems a bit silly...

[tstellar]

We also need to constrain the register class when we select G_IMPLICIT_DEF.

[tstellar]

In D46150#1082433, @nhaehnle wrote:

Is the default implementation really unable to do this? That seems a bit silly...

The default implementation doesn't know how to constrain register classes for generic instructions, which is why targets have to select these generic opcodes. The original version of my patch was not doing that, but it is now.

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPUInstructionSelector.h

private:
   /// tblgen-erated 'select' implementation.
   bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
 
   MachineOperand getSubOperand64(MachineOperand &MO, unsigned SubIdx) const;
   bool selectCOPY(MachineInstr &I) const;
   bool selectG_CONSTANT(MachineInstr &I) const;
   bool selectG_ADD(MachineInstr &I) const;
   bool selectG_GEP(MachineInstr &I) const;
+  bool selectG_IMPLICIT_DEF(MachineInstr &I) const;
   bool selectG_INTRINSIC(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
   bool selectG_INTRINSIC_W_SIDE_EFFECTS(MachineInstr &I,
                                         CodeGenCoverage &CoverageInfo) const;
   bool hasVgprParts(ArrayRef<GEPInfo> AddrInfo) const;
   void getAddrModeInfo(const MachineInstr &Load, const MachineRegisterInfo &MRI,
                        SmallVectorImpl<GEPInfo> &AddrInfo) const;
   bool selectSMRD(MachineInstr &I, ArrayRef<GEPInfo> AddrInfo) const;
   bool selectG_LOAD(MachineInstr &I) const;

lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

bool AMDGPUInstructionSelector::selectG_ADD(MachineInstr &I) const {
   I.eraseFromParent();
   return true;
 }
 
 bool AMDGPUInstructionSelector::selectG_GEP(MachineInstr &I) const {
   return selectG_ADD(I);
 }
 
+bool AMDGPUInstructionSelector::selectG_IMPLICIT_DEF(MachineInstr &I) const {
+  MachineBasicBlock *BB = I.getParent();
+  MachineFunction *MF = BB->getParent();
+  MachineRegisterInfo &MRI = MF->getRegInfo();
+  const MachineOperand &MO = I.getOperand(0);
+  const TargetRegisterClass *RC =
+      TRI.getConstrainedRegClassForOperand(MO, MRI);
+  if (RC)
+    RBI.constrainGenericRegister(MO.getReg(), *RC, MRI);
+  I.setDesc(TII.get(TargetOpcode::IMPLICIT_DEF));
+  return true;
+}
+
 bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I,
                                           CodeGenCoverage &CoverageInfo) const {
   unsigned IntrinsicID =  I.getOperand(1).getIntrinsicID();
 
   switch (IntrinsicID) {
   default:
     break;
   case Intrinsic::amdgcn_cvt_pkrtz:
case TargetOpcode::G_ADD:
     return selectG_ADD(I);
   case TargetOpcode::G_BITCAST:
     return selectCOPY(I);
   case TargetOpcode::G_CONSTANT:
   case TargetOpcode::G_FCONSTANT:
     return selectG_CONSTANT(I);
   case TargetOpcode::G_GEP:
     return selectG_GEP(I);
+  case TargetOpcode::G_IMPLICIT_DEF:
+    return selectG_IMPLICIT_DEF(I);
   case TargetOpcode::G_INTRINSIC:
     return selectG_INTRINSIC(I, CoverageInfo);
   case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
     return selectG_INTRINSIC_W_SIDE_EFFECTS(I, CoverageInfo);
   case TargetOpcode::G_LOAD:
     return selectG_LOAD(I);
   case TargetOpcode::G_STORE:
     return selectG_STORE(I);

test/CodeGen/AMDGPU/GlobalISel/inst-select-implicit-def.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -run-pass=instruction-select -verify-machineinstrs -global-isel %s -o - | FileCheck %s -check-prefixes=GCN
+
+--- |
+  define amdgpu_kernel void @implicit_def(i32 addrspace(1)* %global0) {ret void}
+...
+---
+
+name:            implicit_def
+legalized:       true
+regBankSelected: true
+
+
+body: |
+  bb.0:
+    liveins: $vgpr3_vgpr4
+    ; GCN-LABEL: name: implicit_def
+    ; GCN: [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
+    ; GCN: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN: FLAT_STORE_DWORD [[COPY]], [[DEF]], 0, 0, 0, implicit $exec, implicit $flat_scr
+    %0:vgpr(s64) = COPY $vgpr3_vgpr4
+    %1:vgpr(s32) = G_IMPLICIT_DEF
+    G_STORE %1, %0 :: (store 4 into %ir.global0)
+...
+---