SIISelLowering.cpp revision 73a2c4b9db638cad83e412097ed3433649aab47b
1//===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// Most of the DAG lowering is handled in AMDGPUISelLowering.cpp. This file is 11// mostly EmitInstrWithCustomInserter(). 12// 13//===----------------------------------------------------------------------===// 14 15#include "SIISelLowering.h" 16#include "AMDIL.h" 17#include "AMDILIntrinsicInfo.h" 18#include "SIInstrInfo.h" 19#include "SIRegisterInfo.h" 20#include "llvm/CodeGen/MachineInstrBuilder.h" 21#include "llvm/CodeGen/MachineRegisterInfo.h" 22#include "llvm/CodeGen/SelectionDAG.h" 23 24using namespace llvm; 25 26SITargetLowering::SITargetLowering(TargetMachine &TM) : 27 AMDGPUTargetLowering(TM), 28 TII(static_cast<const SIInstrInfo*>(TM.getInstrInfo())) 29{ 30 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass); 31 addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass); 32 addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass); 33 addRegisterClass(MVT::i64, &AMDGPU::VReg_64RegClass); 34 addRegisterClass(MVT::i1, &AMDGPU::SCCRegRegClass); 35 addRegisterClass(MVT::i1, &AMDGPU::VCCRegRegClass); 36 37 addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass); 38 addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass); 39 40 computeRegisterProperties(); 41 42 setOperationAction(ISD::AND, MVT::i1, Custom); 43 44 setOperationAction(ISD::ADD, MVT::i64, Legal); 45 setOperationAction(ISD::ADD, MVT::i32, Legal); 46 47 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 48 49 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 50 51 // We need to custom lower loads from the USER_SGPR address space, so we can 52 // add the SGPRs as livein registers. 53 setOperationAction(ISD::LOAD, MVT::i32, Custom); 54 setOperationAction(ISD::LOAD, MVT::i64, Custom); 55 56 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 57 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 58 59 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand); 60 setTargetDAGCombine(ISD::SELECT_CC); 61 62 setTargetDAGCombine(ISD::SETCC); 63} 64 65MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter( 66 MachineInstr * MI, MachineBasicBlock * BB) const 67{ 68 const TargetInstrInfo * TII = getTargetMachine().getInstrInfo(); 69 MachineRegisterInfo & MRI = BB->getParent()->getRegInfo(); 70 MachineBasicBlock::iterator I = MI; 71 72 if (TII->get(MI->getOpcode()).TSFlags & SIInstrFlags::NEED_WAIT) { 73 AppendS_WAITCNT(MI, *BB, llvm::next(I)); 74 return BB; 75 } 76 77 switch (MI->getOpcode()) { 78 default: 79 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB); 80 81 case AMDGPU::CLAMP_SI: 82 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64)) 83 .addOperand(MI->getOperand(0)) 84 .addOperand(MI->getOperand(1)) 85 // VSRC1-2 are unused, but we still need to fill all the 86 // operand slots, so we just reuse the VSRC0 operand 87 .addOperand(MI->getOperand(1)) 88 .addOperand(MI->getOperand(1)) 89 .addImm(0) // ABS 90 .addImm(1) // CLAMP 91 .addImm(0) // OMOD 92 .addImm(0); // NEG 93 MI->eraseFromParent(); 94 break; 95 96 case AMDGPU::FABS_SI: 97 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64)) 98 .addOperand(MI->getOperand(0)) 99 .addOperand(MI->getOperand(1)) 100 // VSRC1-2 are unused, but we still need to fill all the 101 // operand slots, so we just reuse the VSRC0 operand 102 .addOperand(MI->getOperand(1)) 103 .addOperand(MI->getOperand(1)) 104 .addImm(1) // ABS 105 .addImm(0) // CLAMP 106 .addImm(0) // OMOD 107 .addImm(0); // NEG 108 MI->eraseFromParent(); 109 break; 110 111 case AMDGPU::FNEG_SI: 112 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64)) 113 .addOperand(MI->getOperand(0)) 114 .addOperand(MI->getOperand(1)) 115 // VSRC1-2 are unused, but we still need to fill all the 116 // operand slots, so we just reuse the VSRC0 operand 117 .addOperand(MI->getOperand(1)) 118 .addOperand(MI->getOperand(1)) 119 .addImm(0) // ABS 120 .addImm(0) // CLAMP 121 .addImm(0) // OMOD 122 .addImm(1); // NEG 123 MI->eraseFromParent(); 124 break; 125 126 case AMDGPU::SI_INTERP: 127 LowerSI_INTERP(MI, *BB, I, MRI); 128 break; 129 case AMDGPU::SI_INTERP_CONST: 130 LowerSI_INTERP_CONST(MI, *BB, I); 131 break; 132 case AMDGPU::SI_KIL: 133 LowerSI_KIL(MI, *BB, I, MRI); 134 break; 135 case AMDGPU::SI_V_CNDLT: 136 LowerSI_V_CNDLT(MI, *BB, I, MRI); 137 break; 138 } 139 return BB; 140} 141 142void SITargetLowering::AppendS_WAITCNT(MachineInstr *MI, MachineBasicBlock &BB, 143 MachineBasicBlock::iterator I) const 144{ 145 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WAITCNT)) 146 .addImm(0); 147} 148 149void SITargetLowering::LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB, 150 MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const 151{ 152 unsigned tmp = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass); 153 MachineOperand dst = MI->getOperand(0); 154 MachineOperand iReg = MI->getOperand(1); 155 MachineOperand jReg = MI->getOperand(2); 156 MachineOperand attr_chan = MI->getOperand(3); 157 MachineOperand attr = MI->getOperand(4); 158 MachineOperand params = MI->getOperand(5); 159 160 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0) 161 .addOperand(params); 162 163 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P1_F32), tmp) 164 .addOperand(iReg) 165 .addOperand(attr_chan) 166 .addOperand(attr); 167 168 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P2_F32)) 169 .addOperand(dst) 170 .addReg(tmp) 171 .addOperand(jReg) 172 .addOperand(attr_chan) 173 .addOperand(attr); 174 175 MI->eraseFromParent(); 176} 177 178void SITargetLowering::LowerSI_INTERP_CONST(MachineInstr *MI, 179 MachineBasicBlock &BB, MachineBasicBlock::iterator I) const 180{ 181 MachineOperand dst = MI->getOperand(0); 182 MachineOperand attr_chan = MI->getOperand(1); 183 MachineOperand attr = MI->getOperand(2); 184 MachineOperand params = MI->getOperand(3); 185 186 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0) 187 .addOperand(params); 188 189 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_MOV_F32)) 190 .addOperand(dst) 191 .addOperand(attr_chan) 192 .addOperand(attr); 193 194 MI->eraseFromParent(); 195} 196 197void SITargetLowering::LowerSI_KIL(MachineInstr *MI, MachineBasicBlock &BB, 198 MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const 199{ 200 // Clear this pixel from the exec mask if the operand is negative 201 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMPX_LE_F32_e32), 202 AMDGPU::VCC) 203 .addReg(AMDGPU::SREG_LIT_0) 204 .addOperand(MI->getOperand(0)); 205 206 // If the exec mask is non-zero, skip the next two instructions 207 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_CBRANCH_EXECNZ)) 208 .addImm(3) 209 .addReg(AMDGPU::EXEC); 210 211 // Exec mask is zero: Export to NULL target... 212 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::EXP)) 213 .addImm(0) 214 .addImm(0x09) // V_008DFC_SQ_EXP_NULL 215 .addImm(0) 216 .addImm(1) 217 .addImm(1) 218 .addReg(AMDGPU::SREG_LIT_0) 219 .addReg(AMDGPU::SREG_LIT_0) 220 .addReg(AMDGPU::SREG_LIT_0) 221 .addReg(AMDGPU::SREG_LIT_0); 222 223 // ... and terminate wavefront 224 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_ENDPGM)); 225 226 MI->eraseFromParent(); 227} 228 229void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB, 230 MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const 231{ 232 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMP_LT_F32_e32), 233 AMDGPU::VCC) 234 .addOperand(MI->getOperand(1)) 235 .addReg(AMDGPU::SREG_LIT_0); 236 237 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CNDMASK_B32)) 238 .addOperand(MI->getOperand(0)) 239 .addReg(AMDGPU::VCC) 240 .addOperand(MI->getOperand(2)) 241 .addOperand(MI->getOperand(3)); 242 243 MI->eraseFromParent(); 244} 245 246EVT SITargetLowering::getSetCCResultType(EVT VT) const 247{ 248 return MVT::i1; 249} 250 251//===----------------------------------------------------------------------===// 252// Custom DAG Lowering Operations 253//===----------------------------------------------------------------------===// 254 255SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const 256{ 257 switch (Op.getOpcode()) { 258 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG); 259 case ISD::BR_CC: return LowerBR_CC(Op, DAG); 260 case ISD::LOAD: return LowerLOAD(Op, DAG); 261 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 262 case ISD::AND: return Loweri1ContextSwitch(Op, DAG, ISD::AND); 263 case ISD::INTRINSIC_WO_CHAIN: { 264 unsigned IntrinsicID = 265 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 266 EVT VT = Op.getValueType(); 267 switch (IntrinsicID) { 268 case AMDGPUIntrinsic::SI_vs_load_buffer_index: 269 return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass, 270 AMDGPU::VGPR0, VT); 271 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG); 272 } 273 break; 274 } 275 } 276 return SDValue(); 277} 278 279/// Loweri1ContextSwitch - The function is for lowering i1 operations on the 280/// VCC register. In the VALU context, VCC is a one bit register, but in the 281/// SALU context the VCC is a 64-bit register (1-bit per thread). Since only 282/// the SALU can perform operations on the VCC register, we need to promote 283/// the operand types from i1 to i64 in order for tablegen to be able to match 284/// this operation to the correct SALU instruction. We do this promotion by 285/// wrapping the operands in a CopyToReg node. 286/// 287SDValue SITargetLowering::Loweri1ContextSwitch(SDValue Op, 288 SelectionDAG &DAG, 289 unsigned VCCNode) const 290{ 291 DebugLoc DL = Op.getDebugLoc(); 292 293 SDValue OpNode = DAG.getNode(VCCNode, DL, MVT::i64, 294 DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i64, 295 Op.getOperand(0)), 296 DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i64, 297 Op.getOperand(1))); 298 299 return DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i1, OpNode); 300} 301 302SDValue SITargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const 303{ 304 SDValue Chain = Op.getOperand(0); 305 SDValue CC = Op.getOperand(1); 306 SDValue LHS = Op.getOperand(2); 307 SDValue RHS = Op.getOperand(3); 308 SDValue JumpT = Op.getOperand(4); 309 SDValue CmpValue; 310 SDValue Result; 311 CmpValue = DAG.getNode( 312 ISD::SETCC, 313 Op.getDebugLoc(), 314 MVT::i1, 315 LHS, RHS, 316 CC); 317 318 Result = DAG.getNode( 319 AMDGPUISD::BRANCH_COND, 320 CmpValue.getDebugLoc(), 321 MVT::Other, Chain, 322 JumpT, CmpValue); 323 return Result; 324} 325 326SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const 327{ 328 EVT VT = Op.getValueType(); 329 LoadSDNode *Ptr = dyn_cast<LoadSDNode>(Op); 330 331 assert(Ptr); 332 333 unsigned AddrSpace = Ptr->getPointerInfo().getAddrSpace(); 334 335 // We only need to lower USER_SGPR address space loads 336 if (AddrSpace != AMDGPUAS::USER_SGPR_ADDRESS) { 337 return SDValue(); 338 } 339 340 // Loads from the USER_SGPR address space can only have constant value 341 // pointers. 342 ConstantSDNode *BasePtr = dyn_cast<ConstantSDNode>(Ptr->getBasePtr()); 343 assert(BasePtr); 344 345 unsigned TypeDwordWidth = VT.getSizeInBits() / 32; 346 const TargetRegisterClass * dstClass; 347 switch (TypeDwordWidth) { 348 default: 349 assert(!"USER_SGPR value size not implemented"); 350 return SDValue(); 351 case 1: 352 dstClass = &AMDGPU::SReg_32RegClass; 353 break; 354 case 2: 355 dstClass = &AMDGPU::SReg_64RegClass; 356 break; 357 } 358 uint64_t Index = BasePtr->getZExtValue(); 359 assert(Index % TypeDwordWidth == 0 && "USER_SGPR not properly aligned"); 360 unsigned SGPRIndex = Index / TypeDwordWidth; 361 unsigned Reg = dstClass->getRegister(SGPRIndex); 362 363 DAG.ReplaceAllUsesOfValueWith(Op, CreateLiveInRegister(DAG, dstClass, Reg, 364 VT)); 365 return SDValue(); 366} 367 368SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const 369{ 370 SDValue LHS = Op.getOperand(0); 371 SDValue RHS = Op.getOperand(1); 372 SDValue True = Op.getOperand(2); 373 SDValue False = Op.getOperand(3); 374 SDValue CC = Op.getOperand(4); 375 EVT VT = Op.getValueType(); 376 DebugLoc DL = Op.getDebugLoc(); 377 378 SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC); 379 return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False); 380} 381 382//===----------------------------------------------------------------------===// 383// Custom DAG optimizations 384//===----------------------------------------------------------------------===// 385 386SDValue SITargetLowering::PerformDAGCombine(SDNode *N, 387 DAGCombinerInfo &DCI) const { 388 SelectionDAG &DAG = DCI.DAG; 389 DebugLoc DL = N->getDebugLoc(); 390 EVT VT = N->getValueType(0); 391 392 switch (N->getOpcode()) { 393 default: break; 394 case ISD::SELECT_CC: { 395 N->dump(); 396 ConstantSDNode *True, *False; 397 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc) 398 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2))) 399 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3))) 400 && True->isAllOnesValue() 401 && False->isNullValue() 402 && VT == MVT::i1) { 403 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0), 404 N->getOperand(1), N->getOperand(4)); 405 406 } 407 break; 408 } 409 case ISD::SETCC: { 410 SDValue Arg0 = N->getOperand(0); 411 SDValue Arg1 = N->getOperand(1); 412 SDValue CC = N->getOperand(2); 413 ConstantSDNode * C = NULL; 414 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get(); 415 416 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne) 417 if (VT == MVT::i1 418 && Arg0.getOpcode() == ISD::SIGN_EXTEND 419 && Arg0.getOperand(0).getValueType() == MVT::i1 420 && (C = dyn_cast<ConstantSDNode>(Arg1)) 421 && C->isNullValue() 422 && CCOp == ISD::SETNE) { 423 return SimplifySetCC(VT, Arg0.getOperand(0), 424 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL); 425 } 426 break; 427 } 428 } 429 return SDValue(); 430} 431 432#define NODE_NAME_CASE(node) case SIISD::node: return #node; 433 434const char* SITargetLowering::getTargetNodeName(unsigned Opcode) const 435{ 436 switch (Opcode) { 437 default: return AMDGPUTargetLowering::getTargetNodeName(Opcode); 438 NODE_NAME_CASE(VCC_AND) 439 NODE_NAME_CASE(VCC_BITCAST) 440 } 441} 442