AMDGPU/SI: Remove VCCReg

[arsenm]

Update to trunk to fix assembler for vopc instructions

[• tstellarAMD]

lib/Target/AMDGPU/SIInstrInfo.td

1091–1093

Can you change this to something like:

class VOPC_Profile<ValueType vt0, ValueType vt1 = vt0> : ...

So that the VOP_I1_* class can inherit from this and don't need to duplicate the Asm32 string.

1100–1102

Can you rename this class to VOPC_I1_F32_I32 for consistency.

1107–1108

Same here, use VOPC prefix.

[arsenm]

Don't duplicate asm strings

[• tstellarAMD]

LGTM.

[arsenm]

r244380

lib/Target/AMDGPU/SIInstrFormats.td

 }
 
 class Enc64 {
   field bits<64> Inst;
   int Size = 8;
 }
 
 class VOPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVOPDst">;
-def VOPDstVCC : VOPDstOperand <VCCReg>;
 
 let Uses = [EXEC] in {
 
 class VOPAnyCommon <dag outs, dag ins, string asm, list<dag> pattern> :
     InstSI <outs, ins, asm, pattern> {
 
   let mayLoad = 0;
   let mayStore = 0;
   let hasSideEffects = 0;
   let UseNamedOperandTable = 1;
   let VALU = 1;
 }
 
 class VOPCCommon <dag ins, string asm, list<dag> pattern> :
-    VOPAnyCommon <(outs VOPDstVCC:$dst), ins, asm, pattern> {
+    VOPAnyCommon <(outs), ins, asm, pattern> {
 
-  let DisableEncoding = "$dst";
   let VOPC = 1;
   let Size = 4;
+  let Defs = [VCC];
 }
 
 class VOP1Common <dag outs, dag ins, string asm, list<dag> pattern> :
     VOPAnyCommon <outs, ins, asm, pattern> {
 
   let VOP1 = 1;
   let Size = 4;
 }

lib/Target/AMDGPU/SIInstrInfo.cpp

SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
   } else if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
     if (DestReg == AMDGPU::VCC) {
       if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
         BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), AMDGPU::VCC)
           .addReg(SrcReg, getKillRegState(KillSrc));
       } else {
         // FIXME: Hack until VReg_1 removed.
         assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
-        BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_I32_e32), AMDGPU::VCC)
+        BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_I32_e32))
           .addImm(0)
           .addReg(SrcReg, getKillRegState(KillSrc));
       }
 
       return;
     }
 
     assert(AMDGPU::SReg_64RegClass.contains(SrcReg));

lib/Target/AMDGPU/SIInstrInfo.td

 def VOP_F64_F64_I32 : VOPProfile <[f64, f64, i32, untyped]>;
 def VOP_I32_F32_F32 : VOPProfile <[i32, f32, f32, untyped]>;
 def VOP_I32_F32_I32 : VOPProfile <[i32, f32, i32, untyped]>;
 def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>;
 def VOP_I32_I32_I32_VCC : VOPProfile <[i32, i32, i32, untyped]> {
   let Src0RC32 = VCSrc_32;
 }
 
+// VOPC instructions are a special case because for the 32-bit
+// encoding, we want to display the implicit vcc write as if it were
+// an explicit $dst.
+class VOPC_Profile<ValueType vt> : VOPProfile <[i1, vt, vt, untyped]> {
+  let Asm32 = "_e32 vcc, $src0, $src1";
+}
+
+def VOPC_I1_F32_F32 : VOPC_Profile<f32>;
+def VOPC_I1_F64_F64 : VOPC_Profile<f64>;
+def VOPC_I1_I32_I32 : VOPC_Profile<i32>;
+def VOPC_I1_I64_I64 : VOPC_Profile<i64>;
+
 def VOP_I1_F32_I32 : VOPProfile <[i1, f32, i32, untyped]> {
+  let Asm32 = "_e32 vcc, $src0, $src1";
+
   let Ins64 = (ins InputModsNoDefault:$src0_modifiers, Src0RC64:$src0, Src1RC64:$src1);
   let Asm64 = "$dst, $src0_modifiers, $src1";
 }
 
 def VOP_I1_F64_I32 : VOPProfile <[i1, f64, i32, untyped]> {
+  let Asm32 = "_e32 vcc, $src0, $src1";
   let Ins64 = (ins InputModsNoDefault:$src0_modifiers, Src0RC64:$src0, Src1RC64:$src1);
   let Asm64 = "$dst, $src0_modifiers, $src1";
 }
 
 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
 def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>;
 def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>;
 def VOP_CNDMASK : VOPProfile <[i32, i32, i32, untyped]> {

 }
 
 multiclass VOPC_m <vopc op, dag outs, dag ins, string asm, list<dag> pattern,
                    string opName, bit DefExec, string revOpName = ""> {
   def "" : VOPC_Pseudo <outs, ins, pattern, opName>;
 
   def _si : VOPC<op.SI, ins, asm, []>,
             SIMCInstr <opName#"_e32", SISubtarget.SI> {
-    let Defs = !if(DefExec, [EXEC], []);
+    let Defs = !if(DefExec, [VCC, EXEC], [VCC]);
     let hasSideEffects = DefExec;
   }
 
   def _vi : VOPC<op.VI, ins, asm, []>,
             SIMCInstr <opName#"_e32", SISubtarget.VI> {
-    let Defs = !if(DefExec, [EXEC], []);
+    let Defs = !if(DefExec, [VCC, EXEC], [VCC]);
     let hasSideEffects = DefExec;
   }
 }
 
 multiclass VOPC_Helper <vopc op, string opName,
                         dag ins32, string asm32, list<dag> pat32,
                         dag out64, dag ins64, string asm64, list<dag> pat64,
                         bit HasMods, bit DefExec, string revOp> {
!if(P.HasModifiers,
       [(set i1:$dst,
           (AMDGPUfp_class (P.Src0VT (VOP3Mods0Clamp0OMod P.Src0VT:$src0, i32:$src0_modifiers)), P.Src1VT:$src1))],
       [(set i1:$dst, (AMDGPUfp_class P.Src0VT:$src0, P.Src1VT:$src1))]),
   P.HasModifiers, DefExec, opName
 >;
 
 
 multiclass VOPC_F32 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
-  VOPCInst <op, opName, VOP_F32_F32_F32, cond, revOp>;
+  VOPCInst <op, opName, VOPC_I1_F32_F32, cond, revOp>;
 
 multiclass VOPC_F64 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
-  VOPCInst <op, opName, VOP_F64_F64_F64, cond, revOp>;
+  VOPCInst <op, opName, VOPC_I1_F64_F64, cond, revOp>;
 
 multiclass VOPC_I32 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
-  VOPCInst <op, opName, VOP_I32_I32_I32, cond, revOp>;
+  VOPCInst <op, opName, VOPC_I1_I32_I32, cond, revOp>;
 
 multiclass VOPC_I64 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
-  VOPCInst <op, opName, VOP_I64_I64_I64, cond, revOp>;
+  VOPCInst <op, opName, VOPC_I1_I64_I64, cond, revOp>;
 
 
 multiclass VOPCX <vopc op, string opName, VOPProfile P,
                   PatLeaf cond = COND_NULL,
                   string revOp = "">
   : VOPCInst <op, opName, P, cond, revOp, 1>;
 
 multiclass VOPCX_F32 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOP_F32_F32_F32, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_F32_F32, COND_NULL, revOp>;
 
 multiclass VOPCX_F64 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOP_F64_F64_F64, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_F64_F64, COND_NULL, revOp>;
 
 multiclass VOPCX_I32 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOP_I32_I32_I32, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_I32_I32, COND_NULL, revOp>;
 
 multiclass VOPCX_I64 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOP_I64_I64_I64, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_I64_I64, COND_NULL, revOp>;
 
 multiclass VOP3_Helper <vop3 op, string opName, dag outs, dag ins, string asm,
                         list<dag> pat, int NumSrcArgs, bit HasMods> : VOP3_m <
     op, outs, ins, opName#" "#asm, pat, opName, NumSrcArgs, HasMods
 >;
 
 multiclass VOPC_CLASS_F32 <vopc op, string opName> :
   VOPCClassInst <op, opName, VOP_I1_F32_I32, 0>;

lib/Target/AMDGPU/SILowerControlFlow.cpp

#endif
   // Clear this thread from the exec mask if the operand is negative
   if ((Op.isImm())) {
     // Constant operand: Set exec mask to 0 or do nothing
     if (Op.getImm() & 0x80000000) {
       BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
               .addImm(0);
     }
   } else {
-    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMPX_LE_F32_e32), AMDGPU::VCC)
+    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMPX_LE_F32_e32))
            .addImm(0)
            .addOperand(Op);
   }
 
   MI.eraseFromParent();
 }
 
 void SILowerControlFlowPass::LoadM0(MachineInstr &MI, MachineInstr *MovRel, int Offset) {
BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32),
             AMDGPU::VCC_LO)
             .addReg(Idx);
 
     // Move index from VCC into M0
     BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
             .addReg(AMDGPU::VCC_LO);
 
     // Compare the just read M0 value to all possible Idx values
-    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMP_EQ_U32_e32), AMDGPU::VCC)
-            .addReg(AMDGPU::M0)
-            .addReg(Idx);
+    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMP_EQ_U32_e32))
 
     // Update EXEC, save the original EXEC value to VCC
     BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), AMDGPU::VCC)
             .addReg(AMDGPU::VCC);
 
     if (Offset) {
       BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ADD_I32), AMDGPU::M0)
               .addReg(AMDGPU::M0)

lib/Target/AMDGPU/SIRegisterInfo.td

 //  Register classes used as source and destination
 //===----------------------------------------------------------------------===//
 
 class RegImmMatcher<string name> : AsmOperandClass {
   let Name = name;
   let RenderMethod = "addRegOrImmOperands";
 }
 
-def VCCReg : RegisterClass<"AMDGPU", [i64, i1], 64, (add VCC)>;
-
 // Register class for all scalar registers (SGPRs + Special Registers)
 def SReg_32 : RegisterClass<"AMDGPU", [i32, f32], 32,
   (add SGPR_32, M0, VCC_LO, VCC_HI, EXEC_LO, EXEC_HI, FLAT_SCR_LO, FLAT_SCR_HI)
 >;
 
 def SGPR_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64], 64, (add SGPR_64Regs)>;
 
 def SReg_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64, i1], 64,

lib/Target/AMDGPU/SIShrinkInstructions.cpp

for (I = MBB.begin(); I != MBB.end(); I = Next) {
       }
 
       // We can shrink this instruction
       DEBUG(dbgs() << "Shrinking "; MI.dump(); dbgs() << '\n';);
 
       MachineInstrBuilder Inst32 =
           BuildMI(MBB, I, MI.getDebugLoc(), TII->get(Op32));
 
+      // Add the dst operand if the 32-bit encoding also has an explicit $dst.
+      // For VOPC instructions, this is replaced by an implicit def of vcc.
+      int Op32DstIdx = AMDGPU::getNamedOperandIdx(Op32, AMDGPU::OpName::dst);
+      if (Op32DstIdx != -1) {
         // dst
         Inst32.addOperand(MI.getOperand(0));
+      } else {
+        assert(MI.getOperand(0).getReg() == AMDGPU::VCC &&
+               "Unexpected case");
+      }
+
 
       Inst32.addOperand(*TII->getNamedOperand(MI, AMDGPU::OpName::src0));
 
       const MachineOperand *Src1 =
           TII->getNamedOperand(MI, AMDGPU::OpName::src1);
       if (Src1)
         Inst32.addOperand(*Src1);