1 //===-- PPCRegisterInfo.cpp - PowerPC Register Information ----------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains the PowerPC implementation of the TargetRegisterInfo 10 // class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "PPCRegisterInfo.h" 15 #include "PPCFrameLowering.h" 16 #include "PPCInstrBuilder.h" 17 #include "PPCMachineFunctionInfo.h" 18 #include "PPCSubtarget.h" 19 #include "PPCTargetMachine.h" 20 #include "llvm/ADT/BitVector.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/CodeGen/MachineFrameInfo.h" 24 #include "llvm/CodeGen/MachineFunction.h" 25 #include "llvm/CodeGen/MachineInstrBuilder.h" 26 #include "llvm/CodeGen/MachineModuleInfo.h" 27 #include "llvm/CodeGen/MachineRegisterInfo.h" 28 #include "llvm/CodeGen/RegisterScavenging.h" 29 #include "llvm/CodeGen/TargetFrameLowering.h" 30 #include "llvm/CodeGen/TargetInstrInfo.h" 31 #include "llvm/IR/CallingConv.h" 32 #include "llvm/IR/Constants.h" 33 #include "llvm/IR/Function.h" 34 #include "llvm/IR/Type.h" 35 #include "llvm/Support/CommandLine.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/ErrorHandling.h" 38 #include "llvm/Support/MathExtras.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include "llvm/Target/TargetMachine.h" 41 #include "llvm/Target/TargetOptions.h" 42 #include <cstdlib> 43 44 using namespace llvm; 45 46 #define DEBUG_TYPE "reginfo" 47 48 #define GET_REGINFO_TARGET_DESC 49 #include "PPCGenRegisterInfo.inc" 50 51 STATISTIC(InflateGPRC, "Number of gprc inputs for getLargestLegalClass"); 52 STATISTIC(InflateGP8RC, "Number of g8rc inputs for getLargestLegalClass"); 53 54 static cl::opt<bool> 55 EnableBasePointer("ppc-use-base-pointer", cl::Hidden, cl::init(true), 56 cl::desc("Enable use of a base pointer for complex stack frames")); 57 58 static cl::opt<bool> 59 AlwaysBasePointer("ppc-always-use-base-pointer", cl::Hidden, cl::init(false), 60 cl::desc("Force the use of a base pointer in every function")); 61 62 static cl::opt<bool> 63 EnableGPRToVecSpills("ppc-enable-gpr-to-vsr-spills", cl::Hidden, cl::init(false), 64 cl::desc("Enable spills from gpr to vsr rather than stack")); 65 66 static cl::opt<bool> 67 StackPtrConst("ppc-stack-ptr-caller-preserved", 68 cl::desc("Consider R1 caller preserved so stack saves of " 69 "caller preserved registers can be LICM candidates"), 70 cl::init(true), cl::Hidden); 71 72 static cl::opt<unsigned> 73 MaxCRBitSpillDist("ppc-max-crbit-spill-dist", 74 cl::desc("Maximum search distance for definition of CR bit " 75 "spill on ppc"), 76 cl::Hidden, cl::init(100)); 77 78 static unsigned offsetMinAlignForOpcode(unsigned OpC); 79 80 PPCRegisterInfo::PPCRegisterInfo(const PPCTargetMachine &TM) 81 : PPCGenRegisterInfo(TM.isPPC64() ? PPC::LR8 : PPC::LR, 82 TM.isPPC64() ? 0 : 1, 83 TM.isPPC64() ? 0 : 1), 84 TM(TM) { 85 ImmToIdxMap[PPC::LD] = PPC::LDX; ImmToIdxMap[PPC::STD] = PPC::STDX; 86 ImmToIdxMap[PPC::LBZ] = PPC::LBZX; ImmToIdxMap[PPC::STB] = PPC::STBX; 87 ImmToIdxMap[PPC::LHZ] = PPC::LHZX; ImmToIdxMap[PPC::LHA] = PPC::LHAX; 88 ImmToIdxMap[PPC::LWZ] = PPC::LWZX; ImmToIdxMap[PPC::LWA] = PPC::LWAX; 89 ImmToIdxMap[PPC::LFS] = PPC::LFSX; ImmToIdxMap[PPC::LFD] = PPC::LFDX; 90 ImmToIdxMap[PPC::STH] = PPC::STHX; ImmToIdxMap[PPC::STW] = PPC::STWX; 91 ImmToIdxMap[PPC::STFS] = PPC::STFSX; ImmToIdxMap[PPC::STFD] = PPC::STFDX; 92 ImmToIdxMap[PPC::ADDI] = PPC::ADD4; 93 ImmToIdxMap[PPC::LWA_32] = PPC::LWAX_32; 94 95 // 64-bit 96 ImmToIdxMap[PPC::LHA8] = PPC::LHAX8; ImmToIdxMap[PPC::LBZ8] = PPC::LBZX8; 97 ImmToIdxMap[PPC::LHZ8] = PPC::LHZX8; ImmToIdxMap[PPC::LWZ8] = PPC::LWZX8; 98 ImmToIdxMap[PPC::STB8] = PPC::STBX8; ImmToIdxMap[PPC::STH8] = PPC::STHX8; 99 ImmToIdxMap[PPC::STW8] = PPC::STWX8; ImmToIdxMap[PPC::STDU] = PPC::STDUX; 100 ImmToIdxMap[PPC::ADDI8] = PPC::ADD8; 101 102 // VSX 103 ImmToIdxMap[PPC::DFLOADf32] = PPC::LXSSPX; 104 ImmToIdxMap[PPC::DFLOADf64] = PPC::LXSDX; 105 ImmToIdxMap[PPC::SPILLTOVSR_LD] = PPC::SPILLTOVSR_LDX; 106 ImmToIdxMap[PPC::SPILLTOVSR_ST] = PPC::SPILLTOVSR_STX; 107 ImmToIdxMap[PPC::DFSTOREf32] = PPC::STXSSPX; 108 ImmToIdxMap[PPC::DFSTOREf64] = PPC::STXSDX; 109 ImmToIdxMap[PPC::LXV] = PPC::LXVX; 110 ImmToIdxMap[PPC::LXSD] = PPC::LXSDX; 111 ImmToIdxMap[PPC::LXSSP] = PPC::LXSSPX; 112 ImmToIdxMap[PPC::STXV] = PPC::STXVX; 113 ImmToIdxMap[PPC::STXSD] = PPC::STXSDX; 114 ImmToIdxMap[PPC::STXSSP] = PPC::STXSSPX; 115 116 // SPE 117 ImmToIdxMap[PPC::EVLDD] = PPC::EVLDDX; 118 ImmToIdxMap[PPC::EVSTDD] = PPC::EVSTDDX; 119 ImmToIdxMap[PPC::SPESTW] = PPC::SPESTWX; 120 ImmToIdxMap[PPC::SPELWZ] = PPC::SPELWZX; 121 } 122 123 /// getPointerRegClass - Return the register class to use to hold pointers. 124 /// This is used for addressing modes. 125 const TargetRegisterClass * 126 PPCRegisterInfo::getPointerRegClass(const MachineFunction &MF, unsigned Kind) 127 const { 128 // Note that PPCInstrInfo::FoldImmediate also directly uses this Kind value 129 // when it checks for ZERO folding. 130 if (Kind == 1) { 131 if (TM.isPPC64()) 132 return &PPC::G8RC_NOX0RegClass; 133 return &PPC::GPRC_NOR0RegClass; 134 } 135 136 if (TM.isPPC64()) 137 return &PPC::G8RCRegClass; 138 return &PPC::GPRCRegClass; 139 } 140 141 const MCPhysReg* 142 PPCRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 143 const PPCSubtarget &Subtarget = MF->getSubtarget<PPCSubtarget>(); 144 if (MF->getFunction().getCallingConv() == CallingConv::AnyReg) { 145 if (!TM.isPPC64() && Subtarget.isAIXABI()) 146 report_fatal_error("AnyReg unimplemented on 32-bit AIX."); 147 if (Subtarget.hasVSX()) 148 return CSR_64_AllRegs_VSX_SaveList; 149 if (Subtarget.hasAltivec()) 150 return CSR_64_AllRegs_Altivec_SaveList; 151 return CSR_64_AllRegs_SaveList; 152 } 153 154 // On PPC64, we might need to save r2 (but only if it is not reserved). 155 // We do not need to treat R2 as callee-saved when using PC-Relative calls 156 // because any direct uses of R2 will cause it to be reserved. If the function 157 // is a leaf or the only uses of R2 are implicit uses for calls, the calls 158 // will use the @notoc relocation which will cause this function to set the 159 // st_other bit to 1, thereby communicating to its caller that it arbitrarily 160 // clobbers the TOC. 161 bool SaveR2 = MF->getRegInfo().isAllocatable(PPC::X2) && 162 !Subtarget.isUsingPCRelativeCalls(); 163 164 // Cold calling convention CSRs. 165 if (MF->getFunction().getCallingConv() == CallingConv::Cold) { 166 if (Subtarget.isAIXABI()) 167 report_fatal_error("Cold calling unimplemented on AIX."); 168 if (TM.isPPC64()) { 169 if (Subtarget.hasAltivec()) 170 return SaveR2 ? CSR_SVR64_ColdCC_R2_Altivec_SaveList 171 : CSR_SVR64_ColdCC_Altivec_SaveList; 172 return SaveR2 ? CSR_SVR64_ColdCC_R2_SaveList 173 : CSR_SVR64_ColdCC_SaveList; 174 } 175 // 32-bit targets. 176 if (Subtarget.hasAltivec()) 177 return CSR_SVR32_ColdCC_Altivec_SaveList; 178 else if (Subtarget.hasSPE()) 179 return CSR_SVR32_ColdCC_SPE_SaveList; 180 return CSR_SVR32_ColdCC_SaveList; 181 } 182 // Standard calling convention CSRs. 183 if (TM.isPPC64()) { 184 if (Subtarget.hasAltivec()) 185 return SaveR2 ? CSR_PPC64_R2_Altivec_SaveList 186 : CSR_PPC64_Altivec_SaveList; 187 return SaveR2 ? CSR_PPC64_R2_SaveList : CSR_PPC64_SaveList; 188 } 189 // 32-bit targets. 190 if (Subtarget.isAIXABI()) 191 return CSR_AIX32_SaveList; 192 if (Subtarget.hasAltivec()) 193 return CSR_SVR432_Altivec_SaveList; 194 else if (Subtarget.hasSPE()) 195 return CSR_SVR432_SPE_SaveList; 196 return CSR_SVR432_SaveList; 197 } 198 199 const uint32_t * 200 PPCRegisterInfo::getCallPreservedMask(const MachineFunction &MF, 201 CallingConv::ID CC) const { 202 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 203 if (CC == CallingConv::AnyReg) { 204 if (Subtarget.hasVSX()) 205 return CSR_64_AllRegs_VSX_RegMask; 206 if (Subtarget.hasAltivec()) 207 return CSR_64_AllRegs_Altivec_RegMask; 208 return CSR_64_AllRegs_RegMask; 209 } 210 211 if (Subtarget.isAIXABI()) { 212 assert(!Subtarget.hasAltivec() && "Altivec is not implemented on AIX yet."); 213 return TM.isPPC64() ? CSR_PPC64_RegMask : CSR_AIX32_RegMask; 214 } 215 216 if (CC == CallingConv::Cold) { 217 return TM.isPPC64() ? (Subtarget.hasAltivec() ? CSR_SVR64_ColdCC_Altivec_RegMask 218 : CSR_SVR64_ColdCC_RegMask) 219 : (Subtarget.hasAltivec() ? CSR_SVR32_ColdCC_Altivec_RegMask 220 : (Subtarget.hasSPE() 221 ? CSR_SVR32_ColdCC_SPE_RegMask 222 : CSR_SVR32_ColdCC_RegMask)); 223 } 224 225 return TM.isPPC64() ? (Subtarget.hasAltivec() ? CSR_PPC64_Altivec_RegMask 226 : CSR_PPC64_RegMask) 227 : (Subtarget.hasAltivec() 228 ? CSR_SVR432_Altivec_RegMask 229 : (Subtarget.hasSPE() ? CSR_SVR432_SPE_RegMask 230 : CSR_SVR432_RegMask)); 231 } 232 233 const uint32_t* 234 PPCRegisterInfo::getNoPreservedMask() const { 235 return CSR_NoRegs_RegMask; 236 } 237 238 void PPCRegisterInfo::adjustStackMapLiveOutMask(uint32_t *Mask) const { 239 for (unsigned PseudoReg : {PPC::ZERO, PPC::ZERO8, PPC::RM}) 240 Mask[PseudoReg / 32] &= ~(1u << (PseudoReg % 32)); 241 } 242 243 BitVector PPCRegisterInfo::getReservedRegs(const MachineFunction &MF) const { 244 BitVector Reserved(getNumRegs()); 245 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 246 const PPCFrameLowering *TFI = getFrameLowering(MF); 247 248 // The ZERO register is not really a register, but the representation of r0 249 // when used in instructions that treat r0 as the constant 0. 250 markSuperRegs(Reserved, PPC::ZERO); 251 252 // The FP register is also not really a register, but is the representation 253 // of the frame pointer register used by ISD::FRAMEADDR. 254 markSuperRegs(Reserved, PPC::FP); 255 256 // The BP register is also not really a register, but is the representation 257 // of the base pointer register used by setjmp. 258 markSuperRegs(Reserved, PPC::BP); 259 260 // The counter registers must be reserved so that counter-based loops can 261 // be correctly formed (and the mtctr instructions are not DCE'd). 262 markSuperRegs(Reserved, PPC::CTR); 263 markSuperRegs(Reserved, PPC::CTR8); 264 265 markSuperRegs(Reserved, PPC::R1); 266 markSuperRegs(Reserved, PPC::LR); 267 markSuperRegs(Reserved, PPC::LR8); 268 markSuperRegs(Reserved, PPC::RM); 269 270 markSuperRegs(Reserved, PPC::VRSAVE); 271 272 // The SVR4 ABI reserves r2 and r13 273 if (Subtarget.isSVR4ABI()) { 274 // We only reserve r2 if we need to use the TOC pointer. If we have no 275 // explicit uses of the TOC pointer (meaning we're a leaf function with 276 // no constant-pool loads, etc.) and we have no potential uses inside an 277 // inline asm block, then we can treat r2 has an ordinary callee-saved 278 // register. 279 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 280 if (!TM.isPPC64() || FuncInfo->usesTOCBasePtr() || MF.hasInlineAsm()) 281 markSuperRegs(Reserved, PPC::R2); // System-reserved register 282 markSuperRegs(Reserved, PPC::R13); // Small Data Area pointer register 283 } 284 285 // Always reserve r2 on AIX for now. 286 // TODO: Make r2 allocatable on AIX/XCOFF for some leaf functions. 287 if (Subtarget.isAIXABI()) 288 markSuperRegs(Reserved, PPC::R2); // System-reserved register 289 290 // On PPC64, r13 is the thread pointer. Never allocate this register. 291 if (TM.isPPC64()) 292 markSuperRegs(Reserved, PPC::R13); 293 294 if (TFI->needsFP(MF)) 295 markSuperRegs(Reserved, PPC::R31); 296 297 bool IsPositionIndependent = TM.isPositionIndependent(); 298 if (hasBasePointer(MF)) { 299 if (Subtarget.is32BitELFABI() && IsPositionIndependent) 300 markSuperRegs(Reserved, PPC::R29); 301 else 302 markSuperRegs(Reserved, PPC::R30); 303 } 304 305 if (Subtarget.is32BitELFABI() && IsPositionIndependent) 306 markSuperRegs(Reserved, PPC::R30); 307 308 // Reserve Altivec registers when Altivec is unavailable. 309 if (!Subtarget.hasAltivec()) 310 for (TargetRegisterClass::iterator I = PPC::VRRCRegClass.begin(), 311 IE = PPC::VRRCRegClass.end(); I != IE; ++I) 312 markSuperRegs(Reserved, *I); 313 314 assert(checkAllSuperRegsMarked(Reserved)); 315 return Reserved; 316 } 317 318 bool PPCRegisterInfo::requiresFrameIndexScavenging(const MachineFunction &MF) const { 319 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 320 const PPCInstrInfo *InstrInfo = Subtarget.getInstrInfo(); 321 const MachineFrameInfo &MFI = MF.getFrameInfo(); 322 const std::vector<CalleeSavedInfo> &Info = MFI.getCalleeSavedInfo(); 323 324 // If the callee saved info is invalid we have to default to true for safety. 325 if (!MFI.isCalleeSavedInfoValid()) 326 return true; 327 328 // We will require the use of X-Forms because the frame is larger than what 329 // can be represented in signed 16 bits that fit in the immediate of a D-Form. 330 // If we need an X-Form then we need a register to store the address offset. 331 unsigned FrameSize = MFI.getStackSize(); 332 // Signed 16 bits means that the FrameSize cannot be more than 15 bits. 333 if (FrameSize & ~0x7FFF) 334 return true; 335 336 // The callee saved info is valid so it can be traversed. 337 // Checking for registers that need saving that do not have load or store 338 // forms where the address offset is an immediate. 339 for (unsigned i = 0; i < Info.size(); i++) { 340 int FrIdx = Info[i].getFrameIdx(); 341 unsigned Reg = Info[i].getReg(); 342 343 const TargetRegisterClass *RC = getMinimalPhysRegClass(Reg); 344 unsigned Opcode = InstrInfo->getStoreOpcodeForSpill(RC); 345 if (!MFI.isFixedObjectIndex(FrIdx)) { 346 // This is not a fixed object. If it requires alignment then we may still 347 // need to use the XForm. 348 if (offsetMinAlignForOpcode(Opcode) > 1) 349 return true; 350 } 351 352 // This is eiher: 353 // 1) A fixed frame index object which we know are aligned so 354 // as long as we have a valid DForm/DSForm/DQForm (non XForm) we don't 355 // need to consider the alignment here. 356 // 2) A not fixed object but in that case we now know that the min required 357 // alignment is no more than 1 based on the previous check. 358 if (InstrInfo->isXFormMemOp(Opcode)) 359 return true; 360 } 361 return false; 362 } 363 364 bool PPCRegisterInfo::isCallerPreservedPhysReg(MCRegister PhysReg, 365 const MachineFunction &MF) const { 366 assert(Register::isPhysicalRegister(PhysReg)); 367 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 368 const MachineFrameInfo &MFI = MF.getFrameInfo(); 369 if (!TM.isPPC64()) 370 return false; 371 372 if (!Subtarget.isSVR4ABI()) 373 return false; 374 if (PhysReg == PPC::X2) 375 // X2 is guaranteed to be preserved within a function if it is reserved. 376 // The reason it's reserved is that it's the TOC pointer (and the function 377 // uses the TOC). In functions where it isn't reserved (i.e. leaf functions 378 // with no TOC access), we can't claim that it is preserved. 379 return (getReservedRegs(MF).test(PPC::X2)); 380 if (StackPtrConst && (PhysReg == PPC::X1) && !MFI.hasVarSizedObjects() 381 && !MFI.hasOpaqueSPAdjustment()) 382 // The value of the stack pointer does not change within a function after 383 // the prologue and before the epilogue if there are no dynamic allocations 384 // and no inline asm which clobbers X1. 385 return true; 386 return false; 387 } 388 389 unsigned PPCRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC, 390 MachineFunction &MF) const { 391 const PPCFrameLowering *TFI = getFrameLowering(MF); 392 const unsigned DefaultSafety = 1; 393 394 switch (RC->getID()) { 395 default: 396 return 0; 397 case PPC::G8RC_NOX0RegClassID: 398 case PPC::GPRC_NOR0RegClassID: 399 case PPC::SPERCRegClassID: 400 case PPC::G8RCRegClassID: 401 case PPC::GPRCRegClassID: { 402 unsigned FP = TFI->hasFP(MF) ? 1 : 0; 403 return 32 - FP - DefaultSafety; 404 } 405 case PPC::F8RCRegClassID: 406 case PPC::F4RCRegClassID: 407 case PPC::QFRCRegClassID: 408 case PPC::QSRCRegClassID: 409 case PPC::QBRCRegClassID: 410 case PPC::VRRCRegClassID: 411 case PPC::VFRCRegClassID: 412 case PPC::VSLRCRegClassID: 413 return 32 - DefaultSafety; 414 case PPC::VSRCRegClassID: 415 case PPC::VSFRCRegClassID: 416 case PPC::VSSRCRegClassID: 417 return 64 - DefaultSafety; 418 case PPC::CRRCRegClassID: 419 return 8 - DefaultSafety; 420 } 421 } 422 423 const TargetRegisterClass * 424 PPCRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC, 425 const MachineFunction &MF) const { 426 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 427 if (Subtarget.hasVSX()) { 428 // With VSX, we can inflate various sub-register classes to the full VSX 429 // register set. 430 431 // For Power9 we allow the user to enable GPR to vector spills. 432 // FIXME: Currently limited to spilling GP8RC. A follow on patch will add 433 // support to spill GPRC. 434 if (TM.isELFv2ABI()) { 435 if (Subtarget.hasP9Vector() && EnableGPRToVecSpills && 436 RC == &PPC::G8RCRegClass) { 437 InflateGP8RC++; 438 return &PPC::SPILLTOVSRRCRegClass; 439 } 440 if (RC == &PPC::GPRCRegClass && EnableGPRToVecSpills) 441 InflateGPRC++; 442 } 443 if (RC == &PPC::F8RCRegClass) 444 return &PPC::VSFRCRegClass; 445 else if (RC == &PPC::VRRCRegClass) 446 return &PPC::VSRCRegClass; 447 else if (RC == &PPC::F4RCRegClass && Subtarget.hasP8Vector()) 448 return &PPC::VSSRCRegClass; 449 } 450 451 return TargetRegisterInfo::getLargestLegalSuperClass(RC, MF); 452 } 453 454 //===----------------------------------------------------------------------===// 455 // Stack Frame Processing methods 456 //===----------------------------------------------------------------------===// 457 458 /// lowerDynamicAlloc - Generate the code for allocating an object in the 459 /// current frame. The sequence of code will be in the general form 460 /// 461 /// addi R0, SP, \#frameSize ; get the address of the previous frame 462 /// stwxu R0, SP, Rnegsize ; add and update the SP with the negated size 463 /// addi Rnew, SP, \#maxCalFrameSize ; get the top of the allocation 464 /// 465 void PPCRegisterInfo::lowerDynamicAlloc(MachineBasicBlock::iterator II) const { 466 // Get the instruction. 467 MachineInstr &MI = *II; 468 // Get the instruction's basic block. 469 MachineBasicBlock &MBB = *MI.getParent(); 470 // Get the basic block's function. 471 MachineFunction &MF = *MBB.getParent(); 472 // Get the frame info. 473 MachineFrameInfo &MFI = MF.getFrameInfo(); 474 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 475 // Get the instruction info. 476 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 477 // Determine whether 64-bit pointers are used. 478 bool LP64 = TM.isPPC64(); 479 DebugLoc dl = MI.getDebugLoc(); 480 481 // Get the maximum call stack size. 482 unsigned maxCallFrameSize = MFI.getMaxCallFrameSize(); 483 // Get the total frame size. 484 unsigned FrameSize = MFI.getStackSize(); 485 486 // Get stack alignments. 487 const PPCFrameLowering *TFI = getFrameLowering(MF); 488 Align TargetAlign = TFI->getStackAlign(); 489 Align MaxAlign = MFI.getMaxAlign(); 490 assert(isAligned(MaxAlign, maxCallFrameSize) && 491 "Maximum call-frame size not sufficiently aligned"); 492 493 // Determine the previous frame's address. If FrameSize can't be 494 // represented as 16 bits or we need special alignment, then we load the 495 // previous frame's address from 0(SP). Why not do an addis of the hi? 496 // Because R0 is our only safe tmp register and addi/addis treat R0 as zero. 497 // Constructing the constant and adding would take 3 instructions. 498 // Fortunately, a frame greater than 32K is rare. 499 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 500 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 501 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 502 503 if (MaxAlign < TargetAlign && isInt<16>(FrameSize)) { 504 if (LP64) 505 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), Reg) 506 .addReg(PPC::X31) 507 .addImm(FrameSize); 508 else 509 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), Reg) 510 .addReg(PPC::R31) 511 .addImm(FrameSize); 512 } else if (LP64) { 513 BuildMI(MBB, II, dl, TII.get(PPC::LD), Reg) 514 .addImm(0) 515 .addReg(PPC::X1); 516 } else { 517 BuildMI(MBB, II, dl, TII.get(PPC::LWZ), Reg) 518 .addImm(0) 519 .addReg(PPC::R1); 520 } 521 522 bool KillNegSizeReg = MI.getOperand(1).isKill(); 523 Register NegSizeReg = MI.getOperand(1).getReg(); 524 525 // Grow the stack and update the stack pointer link, then determine the 526 // address of new allocated space. 527 if (LP64) { 528 if (MaxAlign > TargetAlign) { 529 unsigned UnalNegSizeReg = NegSizeReg; 530 NegSizeReg = MF.getRegInfo().createVirtualRegister(G8RC); 531 532 // Unfortunately, there is no andi, only andi., and we can't insert that 533 // here because we might clobber cr0 while it is live. 534 BuildMI(MBB, II, dl, TII.get(PPC::LI8), NegSizeReg) 535 .addImm(~(MaxAlign.value() - 1)); 536 537 unsigned NegSizeReg1 = NegSizeReg; 538 NegSizeReg = MF.getRegInfo().createVirtualRegister(G8RC); 539 BuildMI(MBB, II, dl, TII.get(PPC::AND8), NegSizeReg) 540 .addReg(UnalNegSizeReg, getKillRegState(KillNegSizeReg)) 541 .addReg(NegSizeReg1, RegState::Kill); 542 KillNegSizeReg = true; 543 } 544 545 BuildMI(MBB, II, dl, TII.get(PPC::STDUX), PPC::X1) 546 .addReg(Reg, RegState::Kill) 547 .addReg(PPC::X1) 548 .addReg(NegSizeReg, getKillRegState(KillNegSizeReg)); 549 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), MI.getOperand(0).getReg()) 550 .addReg(PPC::X1) 551 .addImm(maxCallFrameSize); 552 } else { 553 if (MaxAlign > TargetAlign) { 554 unsigned UnalNegSizeReg = NegSizeReg; 555 NegSizeReg = MF.getRegInfo().createVirtualRegister(GPRC); 556 557 // Unfortunately, there is no andi, only andi., and we can't insert that 558 // here because we might clobber cr0 while it is live. 559 BuildMI(MBB, II, dl, TII.get(PPC::LI), NegSizeReg) 560 .addImm(~(MaxAlign.value() - 1)); 561 562 unsigned NegSizeReg1 = NegSizeReg; 563 NegSizeReg = MF.getRegInfo().createVirtualRegister(GPRC); 564 BuildMI(MBB, II, dl, TII.get(PPC::AND), NegSizeReg) 565 .addReg(UnalNegSizeReg, getKillRegState(KillNegSizeReg)) 566 .addReg(NegSizeReg1, RegState::Kill); 567 KillNegSizeReg = true; 568 } 569 570 BuildMI(MBB, II, dl, TII.get(PPC::STWUX), PPC::R1) 571 .addReg(Reg, RegState::Kill) 572 .addReg(PPC::R1) 573 .addReg(NegSizeReg, getKillRegState(KillNegSizeReg)); 574 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), MI.getOperand(0).getReg()) 575 .addReg(PPC::R1) 576 .addImm(maxCallFrameSize); 577 } 578 579 // Discard the DYNALLOC instruction. 580 MBB.erase(II); 581 } 582 583 void PPCRegisterInfo::lowerDynamicAreaOffset( 584 MachineBasicBlock::iterator II) const { 585 // Get the instruction. 586 MachineInstr &MI = *II; 587 // Get the instruction's basic block. 588 MachineBasicBlock &MBB = *MI.getParent(); 589 // Get the basic block's function. 590 MachineFunction &MF = *MBB.getParent(); 591 // Get the frame info. 592 MachineFrameInfo &MFI = MF.getFrameInfo(); 593 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 594 // Get the instruction info. 595 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 596 597 unsigned maxCallFrameSize = MFI.getMaxCallFrameSize(); 598 bool is64Bit = TM.isPPC64(); 599 DebugLoc dl = MI.getDebugLoc(); 600 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LI8 : PPC::LI), 601 MI.getOperand(0).getReg()) 602 .addImm(maxCallFrameSize); 603 MBB.erase(II); 604 } 605 606 /// lowerCRSpilling - Generate the code for spilling a CR register. Instead of 607 /// reserving a whole register (R0), we scrounge for one here. This generates 608 /// code like this: 609 /// 610 /// mfcr rA ; Move the conditional register into GPR rA. 611 /// rlwinm rA, rA, SB, 0, 31 ; Shift the bits left so they are in CR0's slot. 612 /// stw rA, FI ; Store rA to the frame. 613 /// 614 void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II, 615 unsigned FrameIndex) const { 616 // Get the instruction. 617 MachineInstr &MI = *II; // ; SPILL_CR <SrcReg>, <offset> 618 // Get the instruction's basic block. 619 MachineBasicBlock &MBB = *MI.getParent(); 620 MachineFunction &MF = *MBB.getParent(); 621 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 622 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 623 DebugLoc dl = MI.getDebugLoc(); 624 625 bool LP64 = TM.isPPC64(); 626 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 627 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 628 629 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 630 Register SrcReg = MI.getOperand(0).getReg(); 631 632 // We need to store the CR in the low 4-bits of the saved value. First, issue 633 // an MFOCRF to save all of the CRBits and, if needed, kill the SrcReg. 634 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), Reg) 635 .addReg(SrcReg, getKillRegState(MI.getOperand(0).isKill())); 636 637 // If the saved register wasn't CR0, shift the bits left so that they are in 638 // CR0's slot. 639 if (SrcReg != PPC::CR0) { 640 Register Reg1 = Reg; 641 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 642 643 // rlwinm rA, rA, ShiftBits, 0, 31. 644 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 645 .addReg(Reg1, RegState::Kill) 646 .addImm(getEncodingValue(SrcReg) * 4) 647 .addImm(0) 648 .addImm(31); 649 } 650 651 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW)) 652 .addReg(Reg, RegState::Kill), 653 FrameIndex); 654 655 // Discard the pseudo instruction. 656 MBB.erase(II); 657 } 658 659 void PPCRegisterInfo::lowerCRRestore(MachineBasicBlock::iterator II, 660 unsigned FrameIndex) const { 661 // Get the instruction. 662 MachineInstr &MI = *II; // ; <DestReg> = RESTORE_CR <offset> 663 // Get the instruction's basic block. 664 MachineBasicBlock &MBB = *MI.getParent(); 665 MachineFunction &MF = *MBB.getParent(); 666 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 667 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 668 DebugLoc dl = MI.getDebugLoc(); 669 670 bool LP64 = TM.isPPC64(); 671 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 672 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 673 674 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 675 Register DestReg = MI.getOperand(0).getReg(); 676 assert(MI.definesRegister(DestReg) && 677 "RESTORE_CR does not define its destination"); 678 679 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LWZ8 : PPC::LWZ), 680 Reg), FrameIndex); 681 682 // If the reloaded register isn't CR0, shift the bits right so that they are 683 // in the right CR's slot. 684 if (DestReg != PPC::CR0) { 685 Register Reg1 = Reg; 686 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 687 688 unsigned ShiftBits = getEncodingValue(DestReg)*4; 689 // rlwinm r11, r11, 32-ShiftBits, 0, 31. 690 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 691 .addReg(Reg1, RegState::Kill).addImm(32-ShiftBits).addImm(0) 692 .addImm(31); 693 } 694 695 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MTOCRF8 : PPC::MTOCRF), DestReg) 696 .addReg(Reg, RegState::Kill); 697 698 // Discard the pseudo instruction. 699 MBB.erase(II); 700 } 701 702 void PPCRegisterInfo::lowerCRBitSpilling(MachineBasicBlock::iterator II, 703 unsigned FrameIndex) const { 704 // Get the instruction. 705 MachineInstr &MI = *II; // ; SPILL_CRBIT <SrcReg>, <offset> 706 // Get the instruction's basic block. 707 MachineBasicBlock &MBB = *MI.getParent(); 708 MachineFunction &MF = *MBB.getParent(); 709 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 710 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 711 const TargetRegisterInfo* TRI = Subtarget.getRegisterInfo(); 712 DebugLoc dl = MI.getDebugLoc(); 713 714 bool LP64 = TM.isPPC64(); 715 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 716 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 717 718 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 719 Register SrcReg = MI.getOperand(0).getReg(); 720 721 // Search up the BB to find the definition of the CR bit. 722 MachineBasicBlock::reverse_iterator Ins = MI; 723 MachineBasicBlock::reverse_iterator Rend = MBB.rend(); 724 ++Ins; 725 unsigned CRBitSpillDistance = 0; 726 bool SeenUse = false; 727 for (; Ins != Rend; ++Ins) { 728 // Definition found. 729 if (Ins->modifiesRegister(SrcReg, TRI)) 730 break; 731 // Use found. 732 if (Ins->readsRegister(SrcReg, TRI)) 733 SeenUse = true; 734 // Unable to find CR bit definition within maximum search distance. 735 if (CRBitSpillDistance == MaxCRBitSpillDist) { 736 Ins = MI; 737 break; 738 } 739 // Skip debug instructions when counting CR bit spill distance. 740 if (!Ins->isDebugInstr()) 741 CRBitSpillDistance++; 742 } 743 744 // Unable to find the definition of the CR bit in the MBB. 745 if (Ins == MBB.rend()) 746 Ins = MI; 747 748 bool SpillsKnownBit = false; 749 // There is no need to extract the CR bit if its value is already known. 750 switch (Ins->getOpcode()) { 751 case PPC::CRUNSET: 752 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LI8 : PPC::LI), Reg) 753 .addImm(0); 754 SpillsKnownBit = true; 755 break; 756 case PPC::CRSET: 757 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LIS8 : PPC::LIS), Reg) 758 .addImm(-32768); 759 SpillsKnownBit = true; 760 break; 761 default: 762 // On Power9, we can use SETB to extract the LT bit. This only works for 763 // the LT bit since SETB produces -1/1/0 for LT/GT/<neither>. So the value 764 // of the bit we care about (32-bit sign bit) will be set to the value of 765 // the LT bit (regardless of the other bits in the CR field). 766 if (Subtarget.isISA3_0()) { 767 if (SrcReg == PPC::CR0LT || SrcReg == PPC::CR1LT || 768 SrcReg == PPC::CR2LT || SrcReg == PPC::CR3LT || 769 SrcReg == PPC::CR4LT || SrcReg == PPC::CR5LT || 770 SrcReg == PPC::CR6LT || SrcReg == PPC::CR7LT) { 771 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::SETB8 : PPC::SETB), Reg) 772 .addReg(getCRFromCRBit(SrcReg), RegState::Undef); 773 break; 774 } 775 } 776 777 // We need to move the CR field that contains the CR bit we are spilling. 778 // The super register may not be explicitly defined (i.e. it can be defined 779 // by a CR-logical that only defines the subreg) so we state that the CR 780 // field is undef. Also, in order to preserve the kill flag on the CR bit, 781 // we add it as an implicit use. 782 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), Reg) 783 .addReg(getCRFromCRBit(SrcReg), RegState::Undef) 784 .addReg(SrcReg, 785 RegState::Implicit | getKillRegState(MI.getOperand(0).isKill())); 786 787 // If the saved register wasn't CR0LT, shift the bits left so that the bit 788 // to store is the first one. Mask all but that bit. 789 Register Reg1 = Reg; 790 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 791 792 // rlwinm rA, rA, ShiftBits, 0, 0. 793 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 794 .addReg(Reg1, RegState::Kill) 795 .addImm(getEncodingValue(SrcReg)) 796 .addImm(0).addImm(0); 797 } 798 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW)) 799 .addReg(Reg, RegState::Kill), 800 FrameIndex); 801 802 bool KillsCRBit = MI.killsRegister(SrcReg, TRI); 803 // Discard the pseudo instruction. 804 MBB.erase(II); 805 if (SpillsKnownBit && KillsCRBit && !SeenUse) { 806 Ins->setDesc(TII.get(PPC::UNENCODED_NOP)); 807 Ins->RemoveOperand(0); 808 } 809 } 810 811 void PPCRegisterInfo::lowerCRBitRestore(MachineBasicBlock::iterator II, 812 unsigned FrameIndex) const { 813 // Get the instruction. 814 MachineInstr &MI = *II; // ; <DestReg> = RESTORE_CRBIT <offset> 815 // Get the instruction's basic block. 816 MachineBasicBlock &MBB = *MI.getParent(); 817 MachineFunction &MF = *MBB.getParent(); 818 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 819 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 820 DebugLoc dl = MI.getDebugLoc(); 821 822 bool LP64 = TM.isPPC64(); 823 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 824 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 825 826 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 827 Register DestReg = MI.getOperand(0).getReg(); 828 assert(MI.definesRegister(DestReg) && 829 "RESTORE_CRBIT does not define its destination"); 830 831 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LWZ8 : PPC::LWZ), 832 Reg), FrameIndex); 833 834 BuildMI(MBB, II, dl, TII.get(TargetOpcode::IMPLICIT_DEF), DestReg); 835 836 Register RegO = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 837 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), RegO) 838 .addReg(getCRFromCRBit(DestReg)); 839 840 unsigned ShiftBits = getEncodingValue(DestReg); 841 // rlwimi r11, r10, 32-ShiftBits, ..., ... 842 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWIMI8 : PPC::RLWIMI), RegO) 843 .addReg(RegO, RegState::Kill) 844 .addReg(Reg, RegState::Kill) 845 .addImm(ShiftBits ? 32 - ShiftBits : 0) 846 .addImm(ShiftBits) 847 .addImm(ShiftBits); 848 849 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MTOCRF8 : PPC::MTOCRF), 850 getCRFromCRBit(DestReg)) 851 .addReg(RegO, RegState::Kill) 852 // Make sure we have a use dependency all the way through this 853 // sequence of instructions. We can't have the other bits in the CR 854 // modified in between the mfocrf and the mtocrf. 855 .addReg(getCRFromCRBit(DestReg), RegState::Implicit); 856 857 // Discard the pseudo instruction. 858 MBB.erase(II); 859 } 860 861 void PPCRegisterInfo::lowerVRSAVESpilling(MachineBasicBlock::iterator II, 862 unsigned FrameIndex) const { 863 // Get the instruction. 864 MachineInstr &MI = *II; // ; SPILL_VRSAVE <SrcReg>, <offset> 865 // Get the instruction's basic block. 866 MachineBasicBlock &MBB = *MI.getParent(); 867 MachineFunction &MF = *MBB.getParent(); 868 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 869 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 870 DebugLoc dl = MI.getDebugLoc(); 871 872 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 873 Register Reg = MF.getRegInfo().createVirtualRegister(GPRC); 874 Register SrcReg = MI.getOperand(0).getReg(); 875 876 BuildMI(MBB, II, dl, TII.get(PPC::MFVRSAVEv), Reg) 877 .addReg(SrcReg, getKillRegState(MI.getOperand(0).isKill())); 878 879 addFrameReference( 880 BuildMI(MBB, II, dl, TII.get(PPC::STW)).addReg(Reg, RegState::Kill), 881 FrameIndex); 882 883 // Discard the pseudo instruction. 884 MBB.erase(II); 885 } 886 887 void PPCRegisterInfo::lowerVRSAVERestore(MachineBasicBlock::iterator II, 888 unsigned FrameIndex) const { 889 // Get the instruction. 890 MachineInstr &MI = *II; // ; <DestReg> = RESTORE_VRSAVE <offset> 891 // Get the instruction's basic block. 892 MachineBasicBlock &MBB = *MI.getParent(); 893 MachineFunction &MF = *MBB.getParent(); 894 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 895 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 896 DebugLoc dl = MI.getDebugLoc(); 897 898 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 899 Register Reg = MF.getRegInfo().createVirtualRegister(GPRC); 900 Register DestReg = MI.getOperand(0).getReg(); 901 assert(MI.definesRegister(DestReg) && 902 "RESTORE_VRSAVE does not define its destination"); 903 904 addFrameReference(BuildMI(MBB, II, dl, TII.get(PPC::LWZ), 905 Reg), FrameIndex); 906 907 BuildMI(MBB, II, dl, TII.get(PPC::MTVRSAVEv), DestReg) 908 .addReg(Reg, RegState::Kill); 909 910 // Discard the pseudo instruction. 911 MBB.erase(II); 912 } 913 914 bool PPCRegisterInfo::hasReservedSpillSlot(const MachineFunction &MF, 915 Register Reg, int &FrameIdx) const { 916 // For the nonvolatile condition registers (CR2, CR3, CR4) return true to 917 // prevent allocating an additional frame slot. 918 // For 64-bit ELF and AIX, the CR save area is in the linkage area at SP+8, 919 // for 32-bit AIX the CR save area is in the linkage area at SP+4. 920 // We have created a FrameIndex to that spill slot to keep the CalleSaveInfos 921 // valid. 922 // For 32-bit ELF, we have previously created the stack slot if needed, so 923 // return its FrameIdx. 924 if (PPC::CR2 <= Reg && Reg <= PPC::CR4) { 925 FrameIdx = MF.getInfo<PPCFunctionInfo>()->getCRSpillFrameIndex(); 926 return true; 927 } 928 return false; 929 } 930 931 // If the offset must be a multiple of some value, return what that value is. 932 static unsigned offsetMinAlignForOpcode(unsigned OpC) { 933 switch (OpC) { 934 default: 935 return 1; 936 case PPC::LWA: 937 case PPC::LWA_32: 938 case PPC::LD: 939 case PPC::LDU: 940 case PPC::STD: 941 case PPC::STDU: 942 case PPC::DFLOADf32: 943 case PPC::DFLOADf64: 944 case PPC::DFSTOREf32: 945 case PPC::DFSTOREf64: 946 case PPC::LXSD: 947 case PPC::LXSSP: 948 case PPC::STXSD: 949 case PPC::STXSSP: 950 return 4; 951 case PPC::EVLDD: 952 case PPC::EVSTDD: 953 return 8; 954 case PPC::LXV: 955 case PPC::STXV: 956 return 16; 957 } 958 } 959 960 // If the offset must be a multiple of some value, return what that value is. 961 static unsigned offsetMinAlign(const MachineInstr &MI) { 962 unsigned OpC = MI.getOpcode(); 963 return offsetMinAlignForOpcode(OpC); 964 } 965 966 // Return the OffsetOperandNo given the FIOperandNum (and the instruction). 967 static unsigned getOffsetONFromFION(const MachineInstr &MI, 968 unsigned FIOperandNum) { 969 // Take into account whether it's an add or mem instruction 970 unsigned OffsetOperandNo = (FIOperandNum == 2) ? 1 : 2; 971 if (MI.isInlineAsm()) 972 OffsetOperandNo = FIOperandNum - 1; 973 else if (MI.getOpcode() == TargetOpcode::STACKMAP || 974 MI.getOpcode() == TargetOpcode::PATCHPOINT) 975 OffsetOperandNo = FIOperandNum + 1; 976 977 return OffsetOperandNo; 978 } 979 980 void 981 PPCRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 982 int SPAdj, unsigned FIOperandNum, 983 RegScavenger *RS) const { 984 assert(SPAdj == 0 && "Unexpected"); 985 986 // Get the instruction. 987 MachineInstr &MI = *II; 988 // Get the instruction's basic block. 989 MachineBasicBlock &MBB = *MI.getParent(); 990 // Get the basic block's function. 991 MachineFunction &MF = *MBB.getParent(); 992 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 993 // Get the instruction info. 994 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 995 // Get the frame info. 996 MachineFrameInfo &MFI = MF.getFrameInfo(); 997 DebugLoc dl = MI.getDebugLoc(); 998 999 unsigned OffsetOperandNo = getOffsetONFromFION(MI, FIOperandNum); 1000 1001 // Get the frame index. 1002 int FrameIndex = MI.getOperand(FIOperandNum).getIndex(); 1003 1004 // Get the frame pointer save index. Users of this index are primarily 1005 // DYNALLOC instructions. 1006 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 1007 int FPSI = FI->getFramePointerSaveIndex(); 1008 // Get the instruction opcode. 1009 unsigned OpC = MI.getOpcode(); 1010 1011 if ((OpC == PPC::DYNAREAOFFSET || OpC == PPC::DYNAREAOFFSET8)) { 1012 lowerDynamicAreaOffset(II); 1013 return; 1014 } 1015 1016 // Special case for dynamic alloca. 1017 if (FPSI && FrameIndex == FPSI && 1018 (OpC == PPC::DYNALLOC || OpC == PPC::DYNALLOC8)) { 1019 lowerDynamicAlloc(II); 1020 return; 1021 } 1022 1023 // Special case for pseudo-ops SPILL_CR and RESTORE_CR, etc. 1024 if (OpC == PPC::SPILL_CR) { 1025 lowerCRSpilling(II, FrameIndex); 1026 return; 1027 } else if (OpC == PPC::RESTORE_CR) { 1028 lowerCRRestore(II, FrameIndex); 1029 return; 1030 } else if (OpC == PPC::SPILL_CRBIT) { 1031 lowerCRBitSpilling(II, FrameIndex); 1032 return; 1033 } else if (OpC == PPC::RESTORE_CRBIT) { 1034 lowerCRBitRestore(II, FrameIndex); 1035 return; 1036 } else if (OpC == PPC::SPILL_VRSAVE) { 1037 lowerVRSAVESpilling(II, FrameIndex); 1038 return; 1039 } else if (OpC == PPC::RESTORE_VRSAVE) { 1040 lowerVRSAVERestore(II, FrameIndex); 1041 return; 1042 } 1043 1044 // Replace the FrameIndex with base register with GPR1 (SP) or GPR31 (FP). 1045 MI.getOperand(FIOperandNum).ChangeToRegister( 1046 FrameIndex < 0 ? getBaseRegister(MF) : getFrameRegister(MF), false); 1047 1048 // If the instruction is not present in ImmToIdxMap, then it has no immediate 1049 // form (and must be r+r). 1050 bool noImmForm = !MI.isInlineAsm() && OpC != TargetOpcode::STACKMAP && 1051 OpC != TargetOpcode::PATCHPOINT && !ImmToIdxMap.count(OpC); 1052 1053 // Now add the frame object offset to the offset from r1. 1054 int Offset = MFI.getObjectOffset(FrameIndex); 1055 Offset += MI.getOperand(OffsetOperandNo).getImm(); 1056 1057 // If we're not using a Frame Pointer that has been set to the value of the 1058 // SP before having the stack size subtracted from it, then add the stack size 1059 // to Offset to get the correct offset. 1060 // Naked functions have stack size 0, although getStackSize may not reflect 1061 // that because we didn't call all the pieces that compute it for naked 1062 // functions. 1063 if (!MF.getFunction().hasFnAttribute(Attribute::Naked)) { 1064 if (!(hasBasePointer(MF) && FrameIndex < 0)) 1065 Offset += MFI.getStackSize(); 1066 } 1067 1068 // If we can, encode the offset directly into the instruction. If this is a 1069 // normal PPC "ri" instruction, any 16-bit value can be safely encoded. If 1070 // this is a PPC64 "ix" instruction, only a 16-bit value with the low two bits 1071 // clear can be encoded. This is extremely uncommon, because normally you 1072 // only "std" to a stack slot that is at least 4-byte aligned, but it can 1073 // happen in invalid code. 1074 assert(OpC != PPC::DBG_VALUE && 1075 "This should be handled in a target-independent way"); 1076 bool OffsetFitsMnemonic = (OpC == PPC::EVSTDD || OpC == PPC::EVLDD) ? 1077 isUInt<8>(Offset) : 1078 isInt<16>(Offset); 1079 if (!noImmForm && ((OffsetFitsMnemonic && 1080 ((Offset % offsetMinAlign(MI)) == 0)) || 1081 OpC == TargetOpcode::STACKMAP || 1082 OpC == TargetOpcode::PATCHPOINT)) { 1083 MI.getOperand(OffsetOperandNo).ChangeToImmediate(Offset); 1084 return; 1085 } 1086 1087 // The offset doesn't fit into a single register, scavenge one to build the 1088 // offset in. 1089 1090 bool is64Bit = TM.isPPC64(); 1091 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 1092 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 1093 const TargetRegisterClass *RC = is64Bit ? G8RC : GPRC; 1094 Register SRegHi = MF.getRegInfo().createVirtualRegister(RC), 1095 SReg = MF.getRegInfo().createVirtualRegister(RC); 1096 1097 // Insert a set of rA with the full offset value before the ld, st, or add 1098 if (isInt<16>(Offset)) 1099 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LI8 : PPC::LI), SReg) 1100 .addImm(Offset); 1101 else { 1102 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LIS8 : PPC::LIS), SRegHi) 1103 .addImm(Offset >> 16); 1104 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::ORI8 : PPC::ORI), SReg) 1105 .addReg(SRegHi, RegState::Kill) 1106 .addImm(Offset); 1107 } 1108 1109 // Convert into indexed form of the instruction: 1110 // 1111 // sth 0:rA, 1:imm 2:(rB) ==> sthx 0:rA, 2:rB, 1:r0 1112 // addi 0:rA 1:rB, 2, imm ==> add 0:rA, 1:rB, 2:r0 1113 unsigned OperandBase; 1114 1115 if (noImmForm) 1116 OperandBase = 1; 1117 else if (OpC != TargetOpcode::INLINEASM && 1118 OpC != TargetOpcode::INLINEASM_BR) { 1119 assert(ImmToIdxMap.count(OpC) && 1120 "No indexed form of load or store available!"); 1121 unsigned NewOpcode = ImmToIdxMap.find(OpC)->second; 1122 MI.setDesc(TII.get(NewOpcode)); 1123 OperandBase = 1; 1124 } else { 1125 OperandBase = OffsetOperandNo; 1126 } 1127 1128 Register StackReg = MI.getOperand(FIOperandNum).getReg(); 1129 MI.getOperand(OperandBase).ChangeToRegister(StackReg, false); 1130 MI.getOperand(OperandBase + 1).ChangeToRegister(SReg, false, false, true); 1131 } 1132 1133 Register PPCRegisterInfo::getFrameRegister(const MachineFunction &MF) const { 1134 const PPCFrameLowering *TFI = getFrameLowering(MF); 1135 1136 if (!TM.isPPC64()) 1137 return TFI->hasFP(MF) ? PPC::R31 : PPC::R1; 1138 else 1139 return TFI->hasFP(MF) ? PPC::X31 : PPC::X1; 1140 } 1141 1142 Register PPCRegisterInfo::getBaseRegister(const MachineFunction &MF) const { 1143 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1144 if (!hasBasePointer(MF)) 1145 return getFrameRegister(MF); 1146 1147 if (TM.isPPC64()) 1148 return PPC::X30; 1149 1150 if (Subtarget.isSVR4ABI() && TM.isPositionIndependent()) 1151 return PPC::R29; 1152 1153 return PPC::R30; 1154 } 1155 1156 bool PPCRegisterInfo::hasBasePointer(const MachineFunction &MF) const { 1157 if (!EnableBasePointer) 1158 return false; 1159 if (AlwaysBasePointer) 1160 return true; 1161 1162 // If we need to realign the stack, then the stack pointer can no longer 1163 // serve as an offset into the caller's stack space. As a result, we need a 1164 // base pointer. 1165 return needsStackRealignment(MF); 1166 } 1167 1168 /// Returns true if the instruction's frame index 1169 /// reference would be better served by a base register other than FP 1170 /// or SP. Used by LocalStackFrameAllocation to determine which frame index 1171 /// references it should create new base registers for. 1172 bool PPCRegisterInfo:: 1173 needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const { 1174 assert(Offset < 0 && "Local offset must be negative"); 1175 1176 // It's the load/store FI references that cause issues, as it can be difficult 1177 // to materialize the offset if it won't fit in the literal field. Estimate 1178 // based on the size of the local frame and some conservative assumptions 1179 // about the rest of the stack frame (note, this is pre-regalloc, so 1180 // we don't know everything for certain yet) whether this offset is likely 1181 // to be out of range of the immediate. Return true if so. 1182 1183 // We only generate virtual base registers for loads and stores that have 1184 // an r+i form. Return false for everything else. 1185 unsigned OpC = MI->getOpcode(); 1186 if (!ImmToIdxMap.count(OpC)) 1187 return false; 1188 1189 // Don't generate a new virtual base register just to add zero to it. 1190 if ((OpC == PPC::ADDI || OpC == PPC::ADDI8) && 1191 MI->getOperand(2).getImm() == 0) 1192 return false; 1193 1194 MachineBasicBlock &MBB = *MI->getParent(); 1195 MachineFunction &MF = *MBB.getParent(); 1196 const PPCFrameLowering *TFI = getFrameLowering(MF); 1197 unsigned StackEst = TFI->determineFrameLayout(MF, true); 1198 1199 // If we likely don't need a stack frame, then we probably don't need a 1200 // virtual base register either. 1201 if (!StackEst) 1202 return false; 1203 1204 // Estimate an offset from the stack pointer. 1205 // The incoming offset is relating to the SP at the start of the function, 1206 // but when we access the local it'll be relative to the SP after local 1207 // allocation, so adjust our SP-relative offset by that allocation size. 1208 Offset += StackEst; 1209 1210 // The frame pointer will point to the end of the stack, so estimate the 1211 // offset as the difference between the object offset and the FP location. 1212 return !isFrameOffsetLegal(MI, getBaseRegister(MF), Offset); 1213 } 1214 1215 /// Insert defining instruction(s) for BaseReg to 1216 /// be a pointer to FrameIdx at the beginning of the basic block. 1217 void PPCRegisterInfo::materializeFrameBaseRegister(MachineBasicBlock *MBB, 1218 Register BaseReg, 1219 int FrameIdx, 1220 int64_t Offset) const { 1221 unsigned ADDriOpc = TM.isPPC64() ? PPC::ADDI8 : PPC::ADDI; 1222 1223 MachineBasicBlock::iterator Ins = MBB->begin(); 1224 DebugLoc DL; // Defaults to "unknown" 1225 if (Ins != MBB->end()) 1226 DL = Ins->getDebugLoc(); 1227 1228 const MachineFunction &MF = *MBB->getParent(); 1229 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1230 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1231 const MCInstrDesc &MCID = TII.get(ADDriOpc); 1232 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 1233 MRI.constrainRegClass(BaseReg, TII.getRegClass(MCID, 0, this, MF)); 1234 1235 BuildMI(*MBB, Ins, DL, MCID, BaseReg) 1236 .addFrameIndex(FrameIdx).addImm(Offset); 1237 } 1238 1239 void PPCRegisterInfo::resolveFrameIndex(MachineInstr &MI, Register BaseReg, 1240 int64_t Offset) const { 1241 unsigned FIOperandNum = 0; 1242 while (!MI.getOperand(FIOperandNum).isFI()) { 1243 ++FIOperandNum; 1244 assert(FIOperandNum < MI.getNumOperands() && 1245 "Instr doesn't have FrameIndex operand!"); 1246 } 1247 1248 MI.getOperand(FIOperandNum).ChangeToRegister(BaseReg, false); 1249 unsigned OffsetOperandNo = getOffsetONFromFION(MI, FIOperandNum); 1250 Offset += MI.getOperand(OffsetOperandNo).getImm(); 1251 MI.getOperand(OffsetOperandNo).ChangeToImmediate(Offset); 1252 1253 MachineBasicBlock &MBB = *MI.getParent(); 1254 MachineFunction &MF = *MBB.getParent(); 1255 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1256 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1257 const MCInstrDesc &MCID = MI.getDesc(); 1258 MachineRegisterInfo &MRI = MF.getRegInfo(); 1259 MRI.constrainRegClass(BaseReg, 1260 TII.getRegClass(MCID, FIOperandNum, this, MF)); 1261 } 1262 1263 bool PPCRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, 1264 Register BaseReg, 1265 int64_t Offset) const { 1266 unsigned FIOperandNum = 0; 1267 while (!MI->getOperand(FIOperandNum).isFI()) { 1268 ++FIOperandNum; 1269 assert(FIOperandNum < MI->getNumOperands() && 1270 "Instr doesn't have FrameIndex operand!"); 1271 } 1272 1273 unsigned OffsetOperandNo = getOffsetONFromFION(*MI, FIOperandNum); 1274 Offset += MI->getOperand(OffsetOperandNo).getImm(); 1275 1276 return MI->getOpcode() == PPC::DBG_VALUE || // DBG_VALUE is always Reg+Imm 1277 MI->getOpcode() == TargetOpcode::STACKMAP || 1278 MI->getOpcode() == TargetOpcode::PATCHPOINT || 1279 (isInt<16>(Offset) && (Offset % offsetMinAlign(*MI)) == 0); 1280 } 1281