R600/SI: nuke SReg_1 v3

It's completely unnecessary and can be replace with proper
SReg_64 handling instead.

This actually fixes a piglit test on SI.

v2: use correct register class in addRegisterClass,
    set special classes as not allocatable
v3: revert setting special classes as not allocateable

This is a candidate for the stable branch.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@175355 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/AMDGPUISelLowering.h

@@ -135,16 +135,6 @@ enum {
 
 } // End namespace AMDGPUISD
 
-namespace SIISD {
-
-enum {
-  SI_FIRST = AMDGPUISD::LAST_AMDGPU_ISD_NUMBER,
-  VCC_AND,
-  VCC_BITCAST
-};
-
-} // End namespace SIISD
-
 } // End namespace llvm
 
 #endif // AMDGPUISELLOWERING_H

lib/Target/R600/SIISelLowering.cpp

@@ -31,8 +31,7 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
   addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
   addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
   addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
-  addRegisterClass(MVT::i1, &AMDGPU::SCCRegRegClass);
+  addRegisterClass(MVT::i1, &AMDGPU::SReg_64RegClass);
-  addRegisterClass(MVT::i1, &AMDGPU::VCCRegRegClass);
 
   addRegisterClass(MVT::v1i32, &AMDGPU::VReg_32RegClass);
   addRegisterClass(MVT::v2i32, &AMDGPU::VReg_64RegClass);
@@ -42,8 +41,6 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
 
   computeRegisterProperties();
 
-  setOperationAction(ISD::AND, MVT::i1, Custom);
-
   setOperationAction(ISD::ADD, MVT::i64, Legal);
   setOperationAction(ISD::ADD, MVT::i32, Legal);
 
@@ -202,7 +199,6 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::BRCOND: return LowerBRCOND(Op, DAG);
   case ISD::LOAD: return LowerLOAD(Op, DAG);
   case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
-  case ISD::AND: return Loweri1ContextSwitch(Op, DAG, ISD::AND);
   case ISD::INTRINSIC_WO_CHAIN: {
     unsigned IntrinsicID =
                          cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
@@ -219,30 +215,6 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   return SDValue();
 }
 
-/// \brief The function is for lowering i1 operations on the
-/// VCC register.
-///
-/// In the VALU context, VCC is a one bit register, but in the
-/// SALU context the VCC is a 64-bit register (1-bit per thread).  Since only
-/// the SALU can perform operations on the VCC register, we need to promote
-/// the operand types from i1 to i64 in order for tablegen to be able to match
-/// this operation to the correct SALU instruction.  We do this promotion by
-/// wrapping the operands in a CopyToReg node.
-///
-SDValue SITargetLowering::Loweri1ContextSwitch(SDValue Op,
-                                               SelectionDAG &DAG,
-                                               unsigned VCCNode) const {
-  DebugLoc DL = Op.getDebugLoc();
-
-  SDValue OpNode = DAG.getNode(VCCNode, DL, MVT::i64,
-                               DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i64,
-                                           Op.getOperand(0)),
-                               DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i64,
-                                           Op.getOperand(1)));
-
-  return DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i1, OpNode);
-}
-
 /// \brief Helper function for LowerBRCOND
 static SDNode *findUser(SDValue Value, unsigned Opcode) {
 
@@ -446,13 +418,3 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
   }
   return SDValue();
 }
-
-#define NODE_NAME_CASE(node) case SIISD::node: return #node;
-
-const char* SITargetLowering::getTargetNodeName(unsigned Opcode) const {
-  switch (Opcode) {
-  default: return AMDGPUTargetLowering::getTargetNodeName(Opcode);
-  NODE_NAME_CASE(VCC_AND)
-  NODE_NAME_CASE(VCC_BITCAST)
-  }
-}

lib/Target/R600/SIISelLowering.h

@@ -32,8 +32,6 @@ class SITargetLowering : public AMDGPUTargetLowering {
   void LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
               MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
 
-  SDValue Loweri1ContextSwitch(SDValue Op, SelectionDAG &DAG,
-                                           unsigned VCCNode) const;
   SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
@@ -45,7 +43,6 @@ public:
   virtual EVT getSetCCResultType(EVT VT) const;
   virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
   virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
-  virtual const char* getTargetNodeName(unsigned Opcode) const;
 };
 
 } // End namespace llvm

lib/Target/R600/SIInstrFormats.td

@@ -39,9 +39,6 @@ class SOP2_32 <bits<7> op, string opName, list<dag> pattern>
 class SOP2_64 <bits<7> op, string opName, list<dag> pattern>
   : SOP2 <op, (outs SReg_64:$dst), (ins SSrc_64:$src0, SSrc_64:$src1), opName, pattern>;
 
-class SOP2_VCC <bits<7> op, string opName, list<dag> pattern>
-  : SOP2 <op, (outs SReg_1:$vcc), (ins SSrc_64:$src0, SSrc_64:$src1), opName, pattern>;
-
 class VOP1_Helper <bits<8> op, RegisterClass vrc, RegisterClass arc,
                    string opName, list<dag> pattern> : 
   VOP1 <
@@ -101,7 +98,7 @@ multiclass VOPC_Helper <bits<8> op, RegisterClass vrc, RegisterClass arc,
   def _e32 : VOPC <op, (ins arc:$src0, vrc:$src1), opName, pattern>;
   def _e64 : VOP3 <
     {0, op{7}, op{6}, op{5}, op{4}, op{3}, op{2}, op{1}, op{0}},
-    (outs SReg_1:$dst),
+    (outs SReg_64:$dst),
     (ins arc:$src0, vrc:$src1,
          InstFlag:$abs, InstFlag:$clamp,
          InstFlag:$omod, InstFlag:$neg),

lib/Target/R600/SIInstrInfo.td

@@ -7,37 +7,10 @@
 //
 //===----------------------------------------------------------------------===//
 
-//===----------------------------------------------------------------------===//
-// SI DAG Profiles
-//===----------------------------------------------------------------------===//
-def SDTVCCBinaryOp : SDTypeProfile<1, 2, [
-  SDTCisInt<0>, SDTCisInt<1>, SDTCisSameAs<1, 2>
-]>;
-
 //===----------------------------------------------------------------------===//
 // SI DAG Nodes
 //===----------------------------------------------------------------------===//
 
-// and operation on 64-bit wide vcc
-def SIsreg1_and : SDNode<"SIISD::VCC_AND", SDTVCCBinaryOp,
-  [SDNPCommutative, SDNPAssociative]
->;
-
-// Special bitcast node for sharing VCC register between VALU and SALU
-def SIsreg1_bitcast : SDNode<"SIISD::VCC_BITCAST",
-  SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>]>
->;
-
-// and operation on 64-bit wide vcc
-def SIvcc_and : SDNode<"SIISD::VCC_AND", SDTVCCBinaryOp,
-  [SDNPCommutative, SDNPAssociative]
->;
-
-// Special bitcast node for sharing VCC register between VALU and SALU
-def SIvcc_bitcast : SDNode<"SIISD::VCC_BITCAST",
-  SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>]>
->;
-
 // SMRD takes a 64bit memory address and can only add an 32bit offset
 def SIadd64bit32bit : SDNode<"ISD::ADD",
   SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>, SDTCisVT<0, i64>, SDTCisVT<2, i32>]>

lib/Target/R600/SIInstructions.td

@@ -781,15 +781,15 @@ def V_CNDMASK_B32_e32 : VOP2 <0x00000000, (outs VReg_32:$dst),
 }
 
 def V_CNDMASK_B32_e64 : VOP3 <0x00000100, (outs VReg_32:$dst),
-  (ins VReg_32:$src0, VReg_32:$src1, SReg_1:$src2, InstFlag:$abs, InstFlag:$clamp, InstFlag:$omod, InstFlag:$neg),
+  (ins VReg_32:$src0, VReg_32:$src1, SReg_64:$src2, InstFlag:$abs, InstFlag:$clamp, InstFlag:$omod, InstFlag:$neg),
   "V_CNDMASK_B32_e64",
-  [(set (i32 VReg_32:$dst), (select SReg_1:$src2, VReg_32:$src1, VReg_32:$src0))]
+  [(set (i32 VReg_32:$dst), (select (i1 SReg_64:$src2), VReg_32:$src1, VReg_32:$src0))]
 >;
 
 //f32 pattern for V_CNDMASK_B32_e64
 def : Pat <
-  (f32 (select SReg_1:$src2, VReg_32:$src1, VReg_32:$src0)),
+  (f32 (select (i1 SReg_64:$src2), VReg_32:$src1, VReg_32:$src0)),
-  (V_CNDMASK_B32_e64 VReg_32:$src0, VReg_32:$src1, SReg_1:$src2)
+  (V_CNDMASK_B32_e64 VReg_32:$src0, VReg_32:$src1, SReg_64:$src2)
 >;
 
 defm V_READLANE_B32 : VOP2_32 <0x00000001, "V_READLANE_B32", []>;
@@ -983,11 +983,14 @@ def : Pat <
 def S_AND_B32 : SOP2_32 <0x0000000e, "S_AND_B32", []>;
 
 def S_AND_B64 : SOP2_64 <0x0000000f, "S_AND_B64",
-  [(set SReg_64:$dst, (and SSrc_64:$src0, SSrc_64:$src1))]
+  [(set SReg_64:$dst, (i64 (and SSrc_64:$src0, SSrc_64:$src1)))]
 >;
-def S_AND_VCC : SOP2_VCC <0x0000000f, "S_AND_B64",
+
-  [(set SReg_1:$vcc, (SIvcc_and SSrc_64:$src0, SSrc_64:$src1))]
+def : Pat <
+  (i1 (and SSrc_64:$src0, SSrc_64:$src1)),
+  (S_AND_B64 SSrc_64:$src0, SSrc_64:$src1)
 >;
+
 def S_OR_B32 : SOP2_32 <0x00000010, "S_OR_B32", []>;
 def S_OR_B64 : SOP2_64 <0x00000011, "S_OR_B64", []>;
 def S_XOR_B32 : SOP2_32 <0x00000012, "S_XOR_B32", []>;
@@ -1069,9 +1072,9 @@ let isBranch = 1, isTerminator = 1 in {
 
 def SI_IF : InstSI <
   (outs SReg_64:$dst),
-  (ins SReg_1:$vcc, brtarget:$target),
+  (ins SReg_64:$vcc, brtarget:$target),
   "SI_IF",
-  [(set SReg_64:$dst, (int_SI_if SReg_1:$vcc, bb:$target))]
+  [(set SReg_64:$dst, (int_SI_if SReg_64:$vcc, bb:$target))]
 >;
 
 def SI_ELSE : InstSI <
@@ -1101,9 +1104,9 @@ def SI_BREAK : InstSI <
 
 def SI_IF_BREAK : InstSI <
   (outs SReg_64:$dst),
-  (ins SReg_1:$vcc, SReg_64:$src),
+  (ins SReg_64:$vcc, SReg_64:$src),
   "SI_IF_BREAK",
-  [(set SReg_64:$dst, (int_SI_if_break SReg_1:$vcc, SReg_64:$src))]
+  [(set SReg_64:$dst, (int_SI_if_break SReg_64:$vcc, SReg_64:$src))]
 >;
 
 def SI_ELSE_BREAK : InstSI <
@@ -1260,30 +1263,15 @@ def : BitConvert <i32, f32, VReg_32>;
 def : BitConvert <f32, i32, SReg_32>;
 def : BitConvert <f32, i32, VReg_32>;
 
-def : Pat <
-  (i64 (SIsreg1_bitcast SReg_1:$vcc)),
-  (S_MOV_B64 (COPY_TO_REGCLASS SReg_1:$vcc, SReg_64))
->;
-
-def : Pat <
-  (i1 (SIsreg1_bitcast SReg_64:$vcc)),
-  (COPY_TO_REGCLASS SReg_64:$vcc, SReg_1)
->;
-
-def : Pat <
-  (i64 (SIvcc_bitcast VCCReg:$vcc)),
-  (S_MOV_B64 (COPY_TO_REGCLASS VCCReg:$vcc, SReg_64))
->;
-
-def : Pat <
-  (i1 (SIvcc_bitcast SReg_64:$vcc)),
-  (COPY_TO_REGCLASS SReg_64:$vcc, VCCReg)
->;
-
 /********** ================== **********/
 /********** Immediate Patterns **********/
 /********** ================== **********/
 
+def : Pat <
+  (i1 imm:$imm),
+  (S_MOV_B64 imm:$imm)
+>;
+
 def : Pat <
   (i32 imm:$imm),
   (V_MOV_B32_e32 imm:$imm)

lib/Target/R600/SIRegisterInfo.td

@@ -137,9 +137,7 @@ def SReg_32 : RegisterClass<"AMDGPU", [f32, i32], 32,
     (add SGPR_32, M0, EXEC_LO, EXEC_HI)
 >;
 
-def SReg_64 : RegisterClass<"AMDGPU", [i64], 64, (add SGPR_64, VCC, EXEC)>;
+def SReg_64 : RegisterClass<"AMDGPU", [i1, i64], 64, (add SGPR_64, VCC, EXEC)>;
-
-def SReg_1 : RegisterClass<"AMDGPU", [i1], 1, (add VCC, SGPR_64, EXEC)>;
 
 def SReg_128 : RegisterClass<"AMDGPU", [v4f32, v4i32], 128, (add SGPR_128)>;
 
@@ -178,7 +176,7 @@ def VReg_512 : RegisterClass<"AMDGPU", [v16i32], 512, (add VGPR_512)>;
 // [SV]Src_* operands can have either an immediate or an register
 def SSrc_32 : RegisterClass<"AMDGPU", [i32, f32], 32, (add SReg_32)>;
 
-def SSrc_64 : RegisterClass<"AMDGPU", [i64], 64, (add SReg_64)>;
+def SSrc_64 : RegisterClass<"AMDGPU", [i1, i64], 64, (add SReg_64)>;
 
 def VSrc_32 : RegisterClass<"AMDGPU", [i32, f32], 32, (add VReg_32, SReg_32)>;