1 //===-- PPCISelLowering.h - PPC32 DAG Lowering Interface --------*- C++ -*-===// 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 // This file defines the interfaces that PPC uses to lower LLVM code into a 11 // selection DAG. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H 16 #define LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H 17 18 #include "PPC.h" 19 #include "PPCSubtarget.h" 20 #include "llvm/Target/TargetLowering.h" 21 #include "llvm/CodeGen/SelectionDAG.h" 22 23 namespace llvm { 24 namespace PPCISD { 25 enum NodeType { 26 // Start the numbering where the builtin ops and target ops leave off. 27 FIRST_NUMBER = ISD::BUILTIN_OP_END, 28 29 /// FSEL - Traditional three-operand fsel node. 30 /// 31 FSEL, 32 33 /// FCFID - The FCFID instruction, taking an f64 operand and producing 34 /// and f64 value containing the FP representation of the integer that 35 /// was temporarily in the f64 operand. 36 FCFID, 37 38 /// FCTI[D,W]Z - The FCTIDZ and FCTIWZ instructions, taking an f32 or f64 39 /// operand, producing an f64 value containing the integer representation 40 /// of that FP value. 41 FCTIDZ, FCTIWZ, 42 43 /// STFIWX - The STFIWX instruction. The first operand is an input token 44 /// chain, then an f64 value to store, then an address to store it to. 45 STFIWX, 46 47 // VMADDFP, VNMSUBFP - The VMADDFP and VNMSUBFP instructions, taking 48 // three v4f32 operands and producing a v4f32 result. 49 VMADDFP, VNMSUBFP, 50 51 /// VPERM - The PPC VPERM Instruction. 52 /// 53 VPERM, 54 55 /// Hi/Lo - These represent the high and low 16-bit parts of a global 56 /// address respectively. These nodes have two operands, the first of 57 /// which must be a TargetGlobalAddress, and the second of which must be a 58 /// Constant. Selected naively, these turn into 'lis G+C' and 'li G+C', 59 /// though these are usually folded into other nodes. 60 Hi, Lo, 61 62 TOC_ENTRY, 63 64 /// The following three target-specific nodes are used for calls through 65 /// function pointers in the 64-bit SVR4 ABI. 66 67 /// Restore the TOC from the TOC save area of the current stack frame. 68 /// This is basically a hard coded load instruction which additionally 69 /// takes/produces a flag. 70 TOC_RESTORE, 71 72 /// Like a regular LOAD but additionally taking/producing a flag. 73 LOAD, 74 75 /// LOAD into r2 (also taking/producing a flag). Like TOC_RESTORE, this is 76 /// a hard coded load instruction. 77 LOAD_TOC, 78 79 /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX) 80 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to 81 /// compute an allocation on the stack. 82 DYNALLOC, 83 84 /// GlobalBaseReg - On Darwin, this node represents the result of the mflr 85 /// at function entry, used for PIC code. 86 GlobalBaseReg, 87 88 /// These nodes represent the 32-bit PPC shifts that operate on 6-bit 89 /// shift amounts. These nodes are generated by the multi-precision shift 90 /// code. 91 SRL, SRA, SHL, 92 93 /// EXTSW_32 - This is the EXTSW instruction for use with "32-bit" 94 /// registers. 95 EXTSW_32, 96 97 /// CALL - A direct function call. 98 /// CALL_NOP_SVR4 is a call with the special NOP which follows 64-bit 99 /// SVR4 calls. 100 CALL_Darwin, CALL_SVR4, CALL_NOP_SVR4, 101 102 /// NOP - Special NOP which follows 64-bit SVR4 calls. 103 NOP, 104 105 /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a 106 /// MTCTR instruction. 107 MTCTR, 108 109 /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a 110 /// BCTRL instruction. 111 BCTRL_Darwin, BCTRL_SVR4, 112 113 /// Return with a flag operand, matched by 'blr' 114 RET_FLAG, 115 116 /// R32 = MFCR(CRREG, INFLAG) - Represents the MFCRpseud/MFOCRF 117 /// instructions. This copies the bits corresponding to the specified 118 /// CRREG into the resultant GPR. Bits corresponding to other CR regs 119 /// are undefined. 120 MFCR, 121 122 /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP* 123 /// instructions. For lack of better number, we use the opcode number 124 /// encoding for the OPC field to identify the compare. For example, 838 125 /// is VCMPGTSH. 126 VCMP, 127 128 /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the 129 /// altivec VCMP*o instructions. For lack of better number, we use the 130 /// opcode number encoding for the OPC field to identify the compare. For 131 /// example, 838 is VCMPGTSH. 132 VCMPo, 133 134 /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This 135 /// corresponds to the COND_BRANCH pseudo instruction. CRRC is the 136 /// condition register to branch on, OPC is the branch opcode to use (e.g. 137 /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is 138 /// an optional input flag argument. 139 COND_BRANCH, 140 141 // The following 5 instructions are used only as part of the 142 // long double-to-int conversion sequence. 143 144 /// OUTFLAG = MFFS F8RC - This moves the FPSCR (not modelled) into the 145 /// register. 146 MFFS, 147 148 /// OUTFLAG = MTFSB0 INFLAG - This clears a bit in the FPSCR. 149 MTFSB0, 150 151 /// OUTFLAG = MTFSB1 INFLAG - This sets a bit in the FPSCR. 152 MTFSB1, 153 154 /// F8RC, OUTFLAG = FADDRTZ F8RC, F8RC, INFLAG - This is an FADD done with 155 /// rounding towards zero. It has flags added so it won't move past the 156 /// FPSCR-setting instructions. 157 FADDRTZ, 158 159 /// MTFSF = F8RC, INFLAG - This moves the register into the FPSCR. 160 MTFSF, 161 162 /// LARX = This corresponds to PPC l{w|d}arx instrcution: load and 163 /// reserve indexed. This is used to implement atomic operations. 164 LARX, 165 166 /// STCX = This corresponds to PPC stcx. instrcution: store conditional 167 /// indexed. This is used to implement atomic operations. 168 STCX, 169 170 /// TC_RETURN - A tail call return. 171 /// operand #0 chain 172 /// operand #1 callee (register or absolute) 173 /// operand #2 stack adjustment 174 /// operand #3 optional in flag 175 TC_RETURN, 176 177 /// STD_32 - This is the STD instruction for use with "32-bit" registers. 178 STD_32 = ISD::FIRST_TARGET_MEMORY_OPCODE, 179 180 /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a 181 /// byte-swapping store instruction. It byte-swaps the low "Type" bits of 182 /// the GPRC input, then stores it through Ptr. Type can be either i16 or 183 /// i32. 184 STBRX, 185 186 /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a 187 /// byte-swapping load instruction. It loads "Type" bits, byte swaps it, 188 /// then puts it in the bottom bits of the GPRC. TYPE can be either i16 189 /// or i32. 190 LBRX 191 }; 192 } 193 194 /// Define some predicates that are used for node matching. 195 namespace PPC { 196 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a 197 /// VPKUHUM instruction. 198 bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary); 199 200 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a 201 /// VPKUWUM instruction. 202 bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary); 203 204 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for 205 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes). 206 bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, 207 bool isUnary); 208 209 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for 210 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes). 211 bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, 212 bool isUnary); 213 214 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift 215 /// amount, otherwise return -1. 216 int isVSLDOIShuffleMask(SDNode *N, bool isUnary); 217 218 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand 219 /// specifies a splat of a single element that is suitable for input to 220 /// VSPLTB/VSPLTH/VSPLTW. 221 bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize); 222 223 /// isAllNegativeZeroVector - Returns true if all elements of build_vector 224 /// are -0.0. 225 bool isAllNegativeZeroVector(SDNode *N); 226 227 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the 228 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask. 229 unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize); 230 231 /// get_VSPLTI_elt - If this is a build_vector of constants which can be 232 /// formed by using a vspltis[bhw] instruction of the specified element 233 /// size, return the constant being splatted. The ByteSize field indicates 234 /// the number of bytes of each element [124] -> [bhw]. 235 SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG); 236 } 237 238 class PPCTargetLowering : public TargetLowering { 239 const PPCSubtarget &PPCSubTarget; 240 241 public: 242 explicit PPCTargetLowering(PPCTargetMachine &TM); 243 244 /// getTargetNodeName() - This method returns the name of a target specific 245 /// DAG node. 246 virtual const char *getTargetNodeName(unsigned Opcode) const; 247 getShiftAmountTy(EVT LHSTy)248 virtual MVT getShiftAmountTy(EVT LHSTy) const { return MVT::i32; } 249 250 /// getSetCCResultType - Return the ISD::SETCC ValueType 251 virtual EVT getSetCCResultType(EVT VT) const; 252 253 /// getPreIndexedAddressParts - returns true by value, base pointer and 254 /// offset pointer and addressing mode by reference if the node's address 255 /// can be legally represented as pre-indexed load / store address. 256 virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, 257 SDValue &Offset, 258 ISD::MemIndexedMode &AM, 259 SelectionDAG &DAG) const; 260 261 /// SelectAddressRegReg - Given the specified addressed, check to see if it 262 /// can be represented as an indexed [r+r] operation. Returns false if it 263 /// can be more efficiently represented with [r+imm]. 264 bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index, 265 SelectionDAG &DAG) const; 266 267 /// SelectAddressRegImm - Returns true if the address N can be represented 268 /// by a base register plus a signed 16-bit displacement [r+imm], and if it 269 /// is not better represented as reg+reg. 270 bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base, 271 SelectionDAG &DAG) const; 272 273 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be 274 /// represented as an indexed [r+r] operation. 275 bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index, 276 SelectionDAG &DAG) const; 277 278 /// SelectAddressRegImmShift - Returns true if the address N can be 279 /// represented by a base register plus a signed 14-bit displacement 280 /// [r+imm*4]. Suitable for use by STD and friends. 281 bool SelectAddressRegImmShift(SDValue N, SDValue &Disp, SDValue &Base, 282 SelectionDAG &DAG) const; 283 284 Sched::Preference getSchedulingPreference(SDNode *N) const; 285 286 /// LowerOperation - Provide custom lowering hooks for some operations. 287 /// 288 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 289 290 /// ReplaceNodeResults - Replace the results of node with an illegal result 291 /// type with new values built out of custom code. 292 /// 293 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 294 SelectionDAG &DAG) const; 295 296 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 297 298 virtual void computeMaskedBitsForTargetNode(const SDValue Op, 299 APInt &KnownZero, 300 APInt &KnownOne, 301 const SelectionDAG &DAG, 302 unsigned Depth = 0) const; 303 304 virtual MachineBasicBlock * 305 EmitInstrWithCustomInserter(MachineInstr *MI, 306 MachineBasicBlock *MBB) const; 307 MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI, 308 MachineBasicBlock *MBB, bool is64Bit, 309 unsigned BinOpcode) const; 310 MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr *MI, 311 MachineBasicBlock *MBB, 312 bool is8bit, unsigned Opcode) const; 313 314 ConstraintType getConstraintType(const std::string &Constraint) const; 315 316 /// Examine constraint string and operand type and determine a weight value. 317 /// The operand object must already have been set up with the operand type. 318 ConstraintWeight getSingleConstraintMatchWeight( 319 AsmOperandInfo &info, const char *constraint) const; 320 321 std::pair<unsigned, const TargetRegisterClass*> 322 getRegForInlineAsmConstraint(const std::string &Constraint, 323 EVT VT) const; 324 325 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 326 /// function arguments in the caller parameter area. This is the actual 327 /// alignment, not its logarithm. 328 unsigned getByValTypeAlignment(Type *Ty) const; 329 330 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 331 /// vector. If it is invalid, don't add anything to Ops. 332 virtual void LowerAsmOperandForConstraint(SDValue Op, 333 std::string &Constraint, 334 std::vector<SDValue> &Ops, 335 SelectionDAG &DAG) const; 336 337 /// isLegalAddressingMode - Return true if the addressing mode represented 338 /// by AM is legal for this target, for a load/store of the specified type. 339 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const; 340 341 /// isLegalAddressImmediate - Return true if the integer value can be used 342 /// as the offset of the target addressing mode for load / store of the 343 /// given type. 344 virtual bool isLegalAddressImmediate(int64_t V, Type *Ty) const; 345 346 /// isLegalAddressImmediate - Return true if the GlobalValue can be used as 347 /// the offset of the target addressing mode. 348 virtual bool isLegalAddressImmediate(GlobalValue *GV) const; 349 350 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const; 351 352 /// getOptimalMemOpType - Returns the target specific optimal type for load 353 /// and store operations as a result of memset, memcpy, and memmove 354 /// lowering. If DstAlign is zero that means it's safe to destination 355 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it 356 /// means there isn't a need to check it against alignment requirement, 357 /// probably because the source does not need to be loaded. If 358 /// 'IsZeroVal' is true, that means it's safe to return a 359 /// non-scalar-integer type, e.g. empty string source, constant, or loaded 360 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is 361 /// constant so it does not need to be loaded. 362 /// It returns EVT::Other if the type should be determined using generic 363 /// target-independent logic. 364 virtual EVT 365 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign, 366 bool IsZeroVal, bool MemcpyStrSrc, 367 MachineFunction &MF) const; 368 369 private: 370 SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const; 371 SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const; 372 373 bool 374 IsEligibleForTailCallOptimization(SDValue Callee, 375 CallingConv::ID CalleeCC, 376 bool isVarArg, 377 const SmallVectorImpl<ISD::InputArg> &Ins, 378 SelectionDAG& DAG) const; 379 380 SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG, 381 int SPDiff, 382 SDValue Chain, 383 SDValue &LROpOut, 384 SDValue &FPOpOut, 385 bool isDarwinABI, 386 DebugLoc dl) const; 387 388 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 389 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 390 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; 391 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; 392 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const; 393 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const; 394 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const; 395 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const; 396 SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const; 397 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG, 398 const PPCSubtarget &Subtarget) const; 399 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG, 400 const PPCSubtarget &Subtarget) const; 401 SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG, 402 const PPCSubtarget &Subtarget) const; 403 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG, 404 const PPCSubtarget &Subtarget) const; 405 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; 406 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG, DebugLoc dl) const; 407 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 408 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const; 409 SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const; 410 SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const; 411 SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const; 412 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const; 413 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const; 414 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; 415 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const; 416 SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const; 417 418 SDValue LowerCallResult(SDValue Chain, SDValue InFlag, 419 CallingConv::ID CallConv, bool isVarArg, 420 const SmallVectorImpl<ISD::InputArg> &Ins, 421 DebugLoc dl, SelectionDAG &DAG, 422 SmallVectorImpl<SDValue> &InVals) const; 423 SDValue FinishCall(CallingConv::ID CallConv, DebugLoc dl, bool isTailCall, 424 bool isVarArg, 425 SelectionDAG &DAG, 426 SmallVector<std::pair<unsigned, SDValue>, 8> 427 &RegsToPass, 428 SDValue InFlag, SDValue Chain, 429 SDValue &Callee, 430 int SPDiff, unsigned NumBytes, 431 const SmallVectorImpl<ISD::InputArg> &Ins, 432 SmallVectorImpl<SDValue> &InVals) const; 433 434 virtual SDValue 435 LowerFormalArguments(SDValue Chain, 436 CallingConv::ID CallConv, bool isVarArg, 437 const SmallVectorImpl<ISD::InputArg> &Ins, 438 DebugLoc dl, SelectionDAG &DAG, 439 SmallVectorImpl<SDValue> &InVals) const; 440 441 virtual SDValue 442 LowerCall(SDValue Chain, SDValue Callee, CallingConv::ID CallConv, 443 bool isVarArg, bool doesNotRet, bool &isTailCall, 444 const SmallVectorImpl<ISD::OutputArg> &Outs, 445 const SmallVectorImpl<SDValue> &OutVals, 446 const SmallVectorImpl<ISD::InputArg> &Ins, 447 DebugLoc dl, SelectionDAG &DAG, 448 SmallVectorImpl<SDValue> &InVals) const; 449 450 virtual bool 451 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF, 452 bool isVarArg, 453 const SmallVectorImpl<ISD::OutputArg> &Outs, 454 LLVMContext &Context) const; 455 456 virtual SDValue 457 LowerReturn(SDValue Chain, 458 CallingConv::ID CallConv, bool isVarArg, 459 const SmallVectorImpl<ISD::OutputArg> &Outs, 460 const SmallVectorImpl<SDValue> &OutVals, 461 DebugLoc dl, SelectionDAG &DAG) const; 462 463 SDValue 464 LowerFormalArguments_Darwin(SDValue Chain, 465 CallingConv::ID CallConv, bool isVarArg, 466 const SmallVectorImpl<ISD::InputArg> &Ins, 467 DebugLoc dl, SelectionDAG &DAG, 468 SmallVectorImpl<SDValue> &InVals) const; 469 SDValue 470 LowerFormalArguments_SVR4(SDValue Chain, 471 CallingConv::ID CallConv, bool isVarArg, 472 const SmallVectorImpl<ISD::InputArg> &Ins, 473 DebugLoc dl, SelectionDAG &DAG, 474 SmallVectorImpl<SDValue> &InVals) const; 475 476 SDValue 477 LowerCall_Darwin(SDValue Chain, SDValue Callee, CallingConv::ID CallConv, 478 bool isVarArg, bool isTailCall, 479 const SmallVectorImpl<ISD::OutputArg> &Outs, 480 const SmallVectorImpl<SDValue> &OutVals, 481 const SmallVectorImpl<ISD::InputArg> &Ins, 482 DebugLoc dl, SelectionDAG &DAG, 483 SmallVectorImpl<SDValue> &InVals) const; 484 SDValue 485 LowerCall_SVR4(SDValue Chain, SDValue Callee, CallingConv::ID CallConv, 486 bool isVarArg, bool isTailCall, 487 const SmallVectorImpl<ISD::OutputArg> &Outs, 488 const SmallVectorImpl<SDValue> &OutVals, 489 const SmallVectorImpl<ISD::InputArg> &Ins, 490 DebugLoc dl, SelectionDAG &DAG, 491 SmallVectorImpl<SDValue> &InVals) const; 492 }; 493 } 494 495 #endif // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H 496