AMDGPU/GlobalISel: Select exp with patterns

[arsenm]

Summary

This does produce slightly different code. Now a unique IMPLICIT_DEF
is emitted for each of the implicit_def operands, rather than reusing
the same one.

[kerbowa]

LGTM

[arsenm]

711a17afaff276f816aca5dc4a68fae4e17a2c12

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

MachineInstr *ICmp = BuildMI(*BB, &I, DL, TII.get(Opcode),
             .add(I.getOperand(3));
   RBI.constrainGenericRegister(ICmp->getOperand(0).getReg(),
                                *TRI.getBoolRC(), *MRI);
   bool Ret = constrainSelectedInstRegOperands(*ICmp, TII, TRI, RBI);
   I.eraseFromParent();
   return Ret;
 }
 
-static MachineInstr *
-buildEXP(const TargetInstrInfo &TII, MachineInstr *Insert, unsigned Tgt,
-         unsigned Reg0, unsigned Reg1, unsigned Reg2, unsigned Reg3,
-         unsigned VM, bool Compr, unsigned Enabled, bool Done) {
-  const DebugLoc &DL = Insert->getDebugLoc();
-  MachineBasicBlock &BB = *Insert->getParent();
-  unsigned Opcode = Done ? AMDGPU::EXP_DONE : AMDGPU::EXP;
-  return BuildMI(BB, Insert, DL, TII.get(Opcode))
-          .addImm(Tgt)
-          .addReg(Reg0)
-          .addReg(Reg1)
-          .addReg(Reg2)
-          .addReg(Reg3)
-          .addImm(VM)
-          .addImm(Compr)
-          .addImm(Enabled);
-}
-
 static bool isZero(Register Reg, MachineRegisterInfo &MRI) {
   int64_t C;
   if (mi_match(Reg, MRI, m_ICst(C)) && C == 0)
     return true;
 
   // FIXME: matcher should ignore copies
   return mi_match(Reg, MRI, m_Copy(m_ICst(C))) && C == 0;
 }
bool AMDGPUInstructionSelector::selectStoreIntrinsic(MachineInstr &MI,
   return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
 }
 
 bool AMDGPUInstructionSelector::selectG_INTRINSIC_W_SIDE_EFFECTS(
     MachineInstr &I) const {
   MachineBasicBlock *BB = I.getParent();
   unsigned IntrinsicID = I.getIntrinsicID();
   switch (IntrinsicID) {
-  case Intrinsic::amdgcn_exp: {
-    int64_t Tgt = I.getOperand(1).getImm();
-    int64_t Enabled = I.getOperand(2).getImm();
-    int64_t Done = I.getOperand(7).getImm();
-    int64_t VM = I.getOperand(8).getImm();
-
-    MachineInstr *Exp = buildEXP(TII, &I, Tgt, I.getOperand(3).getReg(),
-                                 I.getOperand(4).getReg(),
-                                 I.getOperand(5).getReg(),
-                                 I.getOperand(6).getReg(),
-                                 VM, false, Enabled, Done);
-
-    I.eraseFromParent();
-    return constrainSelectedInstRegOperands(*Exp, TII, TRI, RBI);
-  }
-  case Intrinsic::amdgcn_exp_compr: {
-    const DebugLoc &DL = I.getDebugLoc();
-    int64_t Tgt = I.getOperand(1).getImm();
-    int64_t Enabled = I.getOperand(2).getImm();
-    Register Reg0 = I.getOperand(3).getReg();
-    Register Reg1 = I.getOperand(4).getReg();
-    Register Undef = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
-    int64_t Done = I.getOperand(5).getImm();
-    int64_t VM = I.getOperand(6).getImm();
-
-    BuildMI(*BB, &I, DL, TII.get(AMDGPU::IMPLICIT_DEF), Undef);
-    MachineInstr *Exp = buildEXP(TII, &I, Tgt, Reg0, Reg1, Undef, Undef, VM,
-                                 true,  Enabled, Done);
-
-    I.eraseFromParent();
-    return constrainSelectedInstRegOperands(*Exp, TII, TRI, RBI);
-  }
   case Intrinsic::amdgcn_end_cf: {
     // FIXME: Manually selecting to avoid dealiing with the SReg_1 trick
     // SelectionDAG uses for wave32 vs wave64.
     BuildMI(*BB, &I, I.getDebugLoc(),
             TII.get(AMDGPU::SI_END_CF))
       .add(I.getOperand(1));
 
     Register Reg = I.getOperand(1).getReg();

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-amdgcn.exp.mir

body: |
   bb.0:
     liveins: $vgpr0
     %0:vgpr(s32) = COPY $vgpr0
 
     ; CHECK: EXP 1, %0, %0, %0, %0, 0, 0, 15, implicit $exec
     G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.exp),1, 15, %0:vgpr(s32), %0:vgpr(s32), %0:vgpr(s32), %0:vgpr(s32), 0, 0
 
     ; CHECK: EXP_DONE 1, %0, %0, %0, %0, 0, 0, 15, implicit $exec
-    G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.exp), 1, 15, %0:vgpr(s32), %0:vgpr(s32), %0:vgpr(s32), %0:vgpr(s32), 1, 0
+    G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.exp), 1, 15, %0:vgpr(s32), %0:vgpr(s32), %0:vgpr(s32), %0:vgpr(s32), -1, 0
 
     %5:vgpr(<2 x s16>) = G_BITCAST %0(s32)
 
     ; CHECK: [[UNDEF0:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
-    ; CHECK: EXP 1, %0, %0, [[UNDEF0]], [[UNDEF0]], 0, 1, 15, implicit $exec
+    ; CHECK: [[UNDEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; CHECK: EXP 1, %0, %0, [[UNDEF1]], [[UNDEF0]], 0, 1, 15, implicit $exec
     G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.exp.compr), 1, 15, %5:vgpr(<2 x s16>), %5:vgpr(<2 x s16>), 0, 0
 
-    ; CHECK: [[UNDEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
-    ; CHECK: EXP_DONE 1, %0, %0, [[UNDEF1]], [[UNDEF1]], 0, 1, 15, implicit $exec
-    G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.exp.compr), 1, 15, %5:vgpr(<2 x s16>), %5:vgpr(<2 x s16>), 1, 0
+    ; CHECK: [[UNDEF2:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; CHECK: [[UNDEF3:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; CHECK: EXP_DONE 1, %0, %0, [[UNDEF3]], [[UNDEF2]], 0, 1, 15, implicit $exec
+    G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.exp.compr), 1, 15, %5:vgpr(<2 x s16>), %5:vgpr(<2 x s16>), -1, 0