R600/SI: Keep 64-bit not on SALU

[Danil]

1 )

case AMDGPU::S_XOR_B64:
  splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);

There is a bug here. It is clearly if we create two more tests:

; SI-CHECK-LABEL: @vector_xor_i64
; SI-CHECK: V_XOR_B32_e32
; SI-CHECK: V_XOR_B32_e32
define void @vector_xor_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in0, i64 addrspace(1)* %in1) {
  %a = load i64 addrspace(1)* %in0
  %b = load i64 addrspace(1)* %in1
  %result = xor i64 %a, %b
  store i64 %result, i64 addrspace(1)* %out
  ret void
}

; SI-CHECK-LABEL: @scalar_xor_i64
; SI-CHECK: S_XOR_B64
define void @scalar_xor_i64(i64 addrspace(1)* %out, i64 %a, i64 %b) {
  %result = xor i64 %a, %b
  store i64 %result, i64 addrspace(1)* %out
  ret void
}

You should write:

case AMDGPU::S_XOR_B64:
  splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);

I believe you should place this tests into xor.ll (next to vector_xor_i32 and scalar_xor_i32) because they checks you code...

2 )

case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;

I believe you need not place here this string, because control flow never get it.
As example consider S_AND_B64, S_OR_B64 and S_XOR_B64 from:

case AMDGPU::S_AND_B64:
  splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
  Inst->eraseFromParent();
  continue;

case AMDGPU::S_OR_B64:
  splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
  Inst->eraseFromParent();
  continue;

case AMDGPU::S_XOR_B64:
  splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
  Inst->eraseFromParent();

3 )
using this patch I have a fail on CodeGen/R600/sub.ll:

 llvm\test\CodeGen\R600\sub.ll:48:6: error: expected string not found in input
;SI: S_XOR_B64
    ^
<stdin>:60:2: note: scanning from here
S_LOAD_DWORDX2 s[0:1], s[0:1], 0x9
    ^
<stdin>:83:2: note: possible intended match here
S_NOT_B64 s[4:5], s[4:5]
    ^

Have you?

[Danil]

2 )

I believe you need not place here this string, because control flow never get it.

I'm sorry I mean:
I believe you need not place here this string, because control flow never get it meaningfully.

[arsenm]

Fix regular 64-bit xos

[arsenm]

r210476

lib/Target/R600/SIInstrInfo.h

                          const TargetRegisterClass *RC,
                          const MachineOperand &Op) const;
 
-  void splitScalar64BitOp(SmallVectorImpl<MachineInstr *> & Worklist,
+  void splitScalar64BitUnaryOp(SmallVectorImpl<MachineInstr *> &Worklist,
                                MachineInstr *Inst, unsigned Opcode) const;
+
+  void splitScalar64BitBinaryOp(SmallVectorImpl<MachineInstr *> &Worklist,
+                                MachineInstr *Inst, unsigned Opcode) const;
 
   void addDescImplicitUseDef(const MCInstrDesc &Desc, MachineInstr *MI) const;
 

lib/Target/R600/SIInstrInfo.cpp

   case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
   case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
   case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
+  case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
   case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
   case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
   case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e32;

       continue;
     }
     case AMDGPU::S_AND_B64:
-      splitScalar64BitOp(Worklist, Inst, AMDGPU::S_AND_B32);
+      splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32);
       Inst->eraseFromParent();
       continue;
 
     case AMDGPU::S_OR_B64:
-      splitScalar64BitOp(Worklist, Inst, AMDGPU::S_OR_B32);
+      splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32);
       Inst->eraseFromParent();
       continue;
 
     case AMDGPU::S_XOR_B64:
-      splitScalar64BitOp(Worklist, Inst, AMDGPU::S_XOR_B32);
+      splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32);
       Inst->eraseFromParent();
       continue;
 
     case AMDGPU::S_NOT_B64:
-      splitScalar64BitOp(Worklist, Inst, AMDGPU::S_NOT_B32);
+      splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
       Inst->eraseFromParent();
       continue;
 

   return &AMDGPU::VReg_32RegClass;
 }
 
-void SIInstrInfo::splitScalar64BitOp(SmallVectorImpl<MachineInstr *> &Worklist,
-                                     MachineInstr *Inst,
-                                     unsigned Opcode) const {
+void SIInstrInfo::splitScalar64BitUnaryOp(
+  SmallVectorImpl<MachineInstr *> &Worklist,
+  MachineInstr *Inst,
+  unsigned Opcode) const {
+  MachineBasicBlock &MBB = *Inst->getParent();
+  MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
+
+  MachineOperand &Dest = Inst->getOperand(0);
+  MachineOperand &Src0 = Inst->getOperand(1);
+  DebugLoc DL = Inst->getDebugLoc();
+
+  MachineBasicBlock::iterator MII = Inst;
+
+  const MCInstrDesc &InstDesc = get(Opcode);
+  const TargetRegisterClass *Src0RC = Src0.isReg() ?
+    MRI.getRegClass(Src0.getReg()) :
+    &AMDGPU::SGPR_32RegClass;
+
+  const TargetRegisterClass *Src0SubRC = RI.getSubRegClass(Src0RC, AMDGPU::sub0);
+
+  MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
+                                                       AMDGPU::sub0, Src0SubRC);
+
+  const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
+  const TargetRegisterClass *DestSubRC = RI.getSubRegClass(DestRC, AMDGPU::sub0);
+
+  unsigned DestSub0 = MRI.createVirtualRegister(DestRC);
+  MachineInstr *LoHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub0)
+    .addOperand(SrcReg0Sub0);
+
+  MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
+                                                       AMDGPU::sub1, Src0SubRC);
+
+  unsigned DestSub1 = MRI.createVirtualRegister(DestSubRC);
+  MachineInstr *HiHalf = BuildMI(MBB, MII, DL, InstDesc, DestSub1)
+    .addOperand(SrcReg0Sub1);
+
+  unsigned FullDestReg = MRI.createVirtualRegister(DestRC);
+  BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
+    .addReg(DestSub0)
+    .addImm(AMDGPU::sub0)
+    .addReg(DestSub1)
+    .addImm(AMDGPU::sub1);
+
+  MRI.replaceRegWith(Dest.getReg(), FullDestReg);
+
+  // Try to legalize the operands in case we need to swap the order to keep it
+  // valid.
+  Worklist.push_back(LoHalf);
+  Worklist.push_back(HiHalf);
+}
+
+void SIInstrInfo::splitScalar64BitBinaryOp(
+  SmallVectorImpl<MachineInstr *> &Worklist,
+  MachineInstr *Inst,
+  unsigned Opcode) const {
   MachineBasicBlock &MBB = *Inst->getParent();
   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
 

lib/Target/R600/SIInstructions.td

   [(set i32:$dst, (not i32:$src0))]
 >;
 
-def S_NOT_B64 : SOP1_64 <0x00000008, "S_NOT_B64", []>;
+def S_NOT_B64 : SOP1_64 <0x00000008, "S_NOT_B64",
+  [(set i64:$dst, (not i64:$src0))]
+>;
 def S_WQM_B32 : SOP1_32 <0x00000009, "S_WQM_B32", []>;
 def S_WQM_B64 : SOP1_64 <0x0000000a, "S_WQM_B64", []>;
 def S_BREV_B32 : SOP1_32 <0x0000000b, "S_BREV_B32", []>;

test/CodeGen/R600/sub.ll

 ;EG-DAG: SUB_INT
 ;EG-DAG: SUB_INT
 
-;SI: S_XOR_B64
+;SI: S_NOT_B64
 ;SI-DAG: S_ADD_I32
 ;SI-DAG: S_ADDC_U32
 ;SI-DAG: S_ADD_I32

test/CodeGen/R600/xor.ll

   store i32 %result, i32 addrspace(1)* %out
   ret void
 }
+
+; SI-CHECK-LABEL: @vector_xor_i64
+; SI-CHECK: V_XOR_B32_e32
+; SI-CHECK: V_XOR_B32_e32
+; SI-CHECK: S_ENDPGM
+define void @vector_xor_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in0, i64 addrspace(1)* %in1) {
+  %a = load i64 addrspace(1)* %in0
+  %b = load i64 addrspace(1)* %in1
+  %result = xor i64 %a, %b
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-CHECK-LABEL: @scalar_xor_i64
+; SI-CHECK: S_XOR_B64
+; SI-CHECK: S_ENDPGM
+define void @scalar_xor_i64(i64 addrspace(1)* %out, i64 %a, i64 %b) {
+  %result = xor i64 %a, %b
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-CHECK-LABEL: @scalar_not_i64
+; SI-CHECK: S_NOT_B64
+define void @scalar_not_i64(i64 addrspace(1)* %out, i64 %a) {
+  %result = xor i64 %a, -1
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-CHECK-LABEL: @vector_not_i64
+; SI-CHECK: V_NOT_B32
+; SI-CHECK: V_NOT_B32
+define void @vector_not_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in0, i64 addrspace(1)* %in1) {
+  %a = load i64 addrspace(1)* %in0
+  %b = load i64 addrspace(1)* %in1
+  %result = xor i64 %a, -1
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}