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 // Copies/moves of physical accumulators are expensive operations 79 // that should be avoided whenever possible. MMA instructions are 80 // meant to be used in performance-sensitive computational kernels. 81 // This option is provided, at least for the time being, to give the 82 // user a tool to detect this expensive operation and either rework 83 // their code or report a compiler bug if that turns out to be the 84 // cause. 85 #ifndef NDEBUG 86 static cl::opt<bool> 87 ReportAccMoves("ppc-report-acc-moves", 88 cl::desc("Emit information about accumulator register spills " 89 "and copies"), 90 cl::Hidden, cl::init(false)); 91 #endif 92 93 static unsigned offsetMinAlignForOpcode(unsigned OpC); 94 95 PPCRegisterInfo::PPCRegisterInfo(const PPCTargetMachine &TM) 96 : PPCGenRegisterInfo(TM.isPPC64() ? PPC::LR8 : PPC::LR, 97 TM.isPPC64() ? 0 : 1, 98 TM.isPPC64() ? 0 : 1), 99 TM(TM) { 100 ImmToIdxMap[PPC::LD] = PPC::LDX; ImmToIdxMap[PPC::STD] = PPC::STDX; 101 ImmToIdxMap[PPC::LBZ] = PPC::LBZX; ImmToIdxMap[PPC::STB] = PPC::STBX; 102 ImmToIdxMap[PPC::LHZ] = PPC::LHZX; ImmToIdxMap[PPC::LHA] = PPC::LHAX; 103 ImmToIdxMap[PPC::LWZ] = PPC::LWZX; ImmToIdxMap[PPC::LWA] = PPC::LWAX; 104 ImmToIdxMap[PPC::LFS] = PPC::LFSX; ImmToIdxMap[PPC::LFD] = PPC::LFDX; 105 ImmToIdxMap[PPC::STH] = PPC::STHX; ImmToIdxMap[PPC::STW] = PPC::STWX; 106 ImmToIdxMap[PPC::STFS] = PPC::STFSX; ImmToIdxMap[PPC::STFD] = PPC::STFDX; 107 ImmToIdxMap[PPC::ADDI] = PPC::ADD4; 108 ImmToIdxMap[PPC::LWA_32] = PPC::LWAX_32; 109 110 // 64-bit 111 ImmToIdxMap[PPC::LHA8] = PPC::LHAX8; ImmToIdxMap[PPC::LBZ8] = PPC::LBZX8; 112 ImmToIdxMap[PPC::LHZ8] = PPC::LHZX8; ImmToIdxMap[PPC::LWZ8] = PPC::LWZX8; 113 ImmToIdxMap[PPC::STB8] = PPC::STBX8; ImmToIdxMap[PPC::STH8] = PPC::STHX8; 114 ImmToIdxMap[PPC::STW8] = PPC::STWX8; ImmToIdxMap[PPC::STDU] = PPC::STDUX; 115 ImmToIdxMap[PPC::ADDI8] = PPC::ADD8; 116 117 // VSX 118 ImmToIdxMap[PPC::DFLOADf32] = PPC::LXSSPX; 119 ImmToIdxMap[PPC::DFLOADf64] = PPC::LXSDX; 120 ImmToIdxMap[PPC::SPILLTOVSR_LD] = PPC::SPILLTOVSR_LDX; 121 ImmToIdxMap[PPC::SPILLTOVSR_ST] = PPC::SPILLTOVSR_STX; 122 ImmToIdxMap[PPC::DFSTOREf32] = PPC::STXSSPX; 123 ImmToIdxMap[PPC::DFSTOREf64] = PPC::STXSDX; 124 ImmToIdxMap[PPC::LXV] = PPC::LXVX; 125 ImmToIdxMap[PPC::LXSD] = PPC::LXSDX; 126 ImmToIdxMap[PPC::LXSSP] = PPC::LXSSPX; 127 ImmToIdxMap[PPC::STXV] = PPC::STXVX; 128 ImmToIdxMap[PPC::STXSD] = PPC::STXSDX; 129 ImmToIdxMap[PPC::STXSSP] = PPC::STXSSPX; 130 131 // SPE 132 ImmToIdxMap[PPC::EVLDD] = PPC::EVLDDX; 133 ImmToIdxMap[PPC::EVSTDD] = PPC::EVSTDDX; 134 ImmToIdxMap[PPC::SPESTW] = PPC::SPESTWX; 135 ImmToIdxMap[PPC::SPELWZ] = PPC::SPELWZX; 136 } 137 138 /// getPointerRegClass - Return the register class to use to hold pointers. 139 /// This is used for addressing modes. 140 const TargetRegisterClass * 141 PPCRegisterInfo::getPointerRegClass(const MachineFunction &MF, unsigned Kind) 142 const { 143 // Note that PPCInstrInfo::FoldImmediate also directly uses this Kind value 144 // when it checks for ZERO folding. 145 if (Kind == 1) { 146 if (TM.isPPC64()) 147 return &PPC::G8RC_NOX0RegClass; 148 return &PPC::GPRC_NOR0RegClass; 149 } 150 151 if (TM.isPPC64()) 152 return &PPC::G8RCRegClass; 153 return &PPC::GPRCRegClass; 154 } 155 156 const MCPhysReg* 157 PPCRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 158 const PPCSubtarget &Subtarget = MF->getSubtarget<PPCSubtarget>(); 159 if (MF->getFunction().getCallingConv() == CallingConv::AnyReg) { 160 if (!TM.isPPC64() && Subtarget.isAIXABI()) 161 report_fatal_error("AnyReg unimplemented on 32-bit AIX."); 162 if (Subtarget.hasVSX()) { 163 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 164 return CSR_64_AllRegs_AIX_Dflt_VSX_SaveList; 165 return CSR_64_AllRegs_VSX_SaveList; 166 } 167 if (Subtarget.hasAltivec()) { 168 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 169 return CSR_64_AllRegs_AIX_Dflt_Altivec_SaveList; 170 return CSR_64_AllRegs_Altivec_SaveList; 171 } 172 return CSR_64_AllRegs_SaveList; 173 } 174 175 // On PPC64, we might need to save r2 (but only if it is not reserved). 176 // We do not need to treat R2 as callee-saved when using PC-Relative calls 177 // because any direct uses of R2 will cause it to be reserved. If the function 178 // is a leaf or the only uses of R2 are implicit uses for calls, the calls 179 // will use the @notoc relocation which will cause this function to set the 180 // st_other bit to 1, thereby communicating to its caller that it arbitrarily 181 // clobbers the TOC. 182 bool SaveR2 = MF->getRegInfo().isAllocatable(PPC::X2) && 183 !Subtarget.isUsingPCRelativeCalls(); 184 185 // Cold calling convention CSRs. 186 if (MF->getFunction().getCallingConv() == CallingConv::Cold) { 187 if (Subtarget.isAIXABI()) 188 report_fatal_error("Cold calling unimplemented on AIX."); 189 if (TM.isPPC64()) { 190 if (Subtarget.hasAltivec()) 191 return SaveR2 ? CSR_SVR64_ColdCC_R2_Altivec_SaveList 192 : CSR_SVR64_ColdCC_Altivec_SaveList; 193 return SaveR2 ? CSR_SVR64_ColdCC_R2_SaveList 194 : CSR_SVR64_ColdCC_SaveList; 195 } 196 // 32-bit targets. 197 if (Subtarget.hasAltivec()) 198 return CSR_SVR32_ColdCC_Altivec_SaveList; 199 else if (Subtarget.hasSPE()) 200 return CSR_SVR32_ColdCC_SPE_SaveList; 201 return CSR_SVR32_ColdCC_SaveList; 202 } 203 // Standard calling convention CSRs. 204 if (TM.isPPC64()) { 205 if (Subtarget.hasAltivec() && 206 (!Subtarget.isAIXABI() || TM.getAIXExtendedAltivecABI())) { 207 return SaveR2 ? CSR_PPC64_R2_Altivec_SaveList 208 : CSR_PPC64_Altivec_SaveList; 209 } 210 return SaveR2 ? CSR_PPC64_R2_SaveList : CSR_PPC64_SaveList; 211 } 212 // 32-bit targets. 213 if (Subtarget.isAIXABI()) { 214 if (Subtarget.hasAltivec()) 215 return TM.getAIXExtendedAltivecABI() ? CSR_AIX32_Altivec_SaveList 216 : CSR_AIX32_SaveList; 217 return CSR_AIX32_SaveList; 218 } 219 if (Subtarget.hasAltivec()) 220 return CSR_SVR432_Altivec_SaveList; 221 else if (Subtarget.hasSPE()) 222 return CSR_SVR432_SPE_SaveList; 223 return CSR_SVR432_SaveList; 224 } 225 226 const uint32_t * 227 PPCRegisterInfo::getCallPreservedMask(const MachineFunction &MF, 228 CallingConv::ID CC) const { 229 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 230 if (CC == CallingConv::AnyReg) { 231 if (Subtarget.hasVSX()) { 232 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 233 return CSR_64_AllRegs_AIX_Dflt_VSX_RegMask; 234 return CSR_64_AllRegs_VSX_RegMask; 235 } 236 if (Subtarget.hasAltivec()) { 237 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 238 return CSR_64_AllRegs_AIX_Dflt_Altivec_RegMask; 239 return CSR_64_AllRegs_Altivec_RegMask; 240 } 241 return CSR_64_AllRegs_RegMask; 242 } 243 244 if (Subtarget.isAIXABI()) { 245 return TM.isPPC64() 246 ? ((Subtarget.hasAltivec() && TM.getAIXExtendedAltivecABI()) 247 ? CSR_PPC64_Altivec_RegMask 248 : CSR_PPC64_RegMask) 249 : ((Subtarget.hasAltivec() && TM.getAIXExtendedAltivecABI()) 250 ? CSR_AIX32_Altivec_RegMask 251 : CSR_AIX32_RegMask); 252 } 253 254 if (CC == CallingConv::Cold) { 255 return TM.isPPC64() ? (Subtarget.hasAltivec() ? CSR_SVR64_ColdCC_Altivec_RegMask 256 : CSR_SVR64_ColdCC_RegMask) 257 : (Subtarget.hasAltivec() ? CSR_SVR32_ColdCC_Altivec_RegMask 258 : (Subtarget.hasSPE() 259 ? CSR_SVR32_ColdCC_SPE_RegMask 260 : CSR_SVR32_ColdCC_RegMask)); 261 } 262 263 return TM.isPPC64() ? (Subtarget.hasAltivec() ? CSR_PPC64_Altivec_RegMask 264 : CSR_PPC64_RegMask) 265 : (Subtarget.hasAltivec() 266 ? CSR_SVR432_Altivec_RegMask 267 : (Subtarget.hasSPE() ? CSR_SVR432_SPE_RegMask 268 : CSR_SVR432_RegMask)); 269 } 270 271 const uint32_t* 272 PPCRegisterInfo::getNoPreservedMask() const { 273 return CSR_NoRegs_RegMask; 274 } 275 276 void PPCRegisterInfo::adjustStackMapLiveOutMask(uint32_t *Mask) const { 277 for (unsigned PseudoReg : {PPC::ZERO, PPC::ZERO8, PPC::RM}) 278 Mask[PseudoReg / 32] &= ~(1u << (PseudoReg % 32)); 279 } 280 281 BitVector PPCRegisterInfo::getReservedRegs(const MachineFunction &MF) const { 282 BitVector Reserved(getNumRegs()); 283 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 284 const PPCFrameLowering *TFI = getFrameLowering(MF); 285 286 // The ZERO register is not really a register, but the representation of r0 287 // when used in instructions that treat r0 as the constant 0. 288 markSuperRegs(Reserved, PPC::ZERO); 289 290 // The FP register is also not really a register, but is the representation 291 // of the frame pointer register used by ISD::FRAMEADDR. 292 markSuperRegs(Reserved, PPC::FP); 293 294 // The BP register is also not really a register, but is the representation 295 // of the base pointer register used by setjmp. 296 markSuperRegs(Reserved, PPC::BP); 297 298 // The counter registers must be reserved so that counter-based loops can 299 // be correctly formed (and the mtctr instructions are not DCE'd). 300 markSuperRegs(Reserved, PPC::CTR); 301 markSuperRegs(Reserved, PPC::CTR8); 302 303 markSuperRegs(Reserved, PPC::R1); 304 markSuperRegs(Reserved, PPC::LR); 305 markSuperRegs(Reserved, PPC::LR8); 306 markSuperRegs(Reserved, PPC::RM); 307 308 markSuperRegs(Reserved, PPC::VRSAVE); 309 310 // The SVR4 ABI reserves r2 and r13 311 if (Subtarget.isSVR4ABI()) { 312 // We only reserve r2 if we need to use the TOC pointer. If we have no 313 // explicit uses of the TOC pointer (meaning we're a leaf function with 314 // no constant-pool loads, etc.) and we have no potential uses inside an 315 // inline asm block, then we can treat r2 has an ordinary callee-saved 316 // register. 317 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 318 if (!TM.isPPC64() || FuncInfo->usesTOCBasePtr() || MF.hasInlineAsm()) 319 markSuperRegs(Reserved, PPC::R2); // System-reserved register 320 markSuperRegs(Reserved, PPC::R13); // Small Data Area pointer register 321 } 322 323 // Always reserve r2 on AIX for now. 324 // TODO: Make r2 allocatable on AIX/XCOFF for some leaf functions. 325 if (Subtarget.isAIXABI()) 326 markSuperRegs(Reserved, PPC::R2); // System-reserved register 327 328 // On PPC64, r13 is the thread pointer. Never allocate this register. 329 if (TM.isPPC64()) 330 markSuperRegs(Reserved, PPC::R13); 331 332 if (TFI->needsFP(MF)) 333 markSuperRegs(Reserved, PPC::R31); 334 335 bool IsPositionIndependent = TM.isPositionIndependent(); 336 if (hasBasePointer(MF)) { 337 if (Subtarget.is32BitELFABI() && IsPositionIndependent) 338 markSuperRegs(Reserved, PPC::R29); 339 else 340 markSuperRegs(Reserved, PPC::R30); 341 } 342 343 if (Subtarget.is32BitELFABI() && IsPositionIndependent) 344 markSuperRegs(Reserved, PPC::R30); 345 346 // Reserve Altivec registers when Altivec is unavailable. 347 if (!Subtarget.hasAltivec()) 348 for (TargetRegisterClass::iterator I = PPC::VRRCRegClass.begin(), 349 IE = PPC::VRRCRegClass.end(); I != IE; ++I) 350 markSuperRegs(Reserved, *I); 351 352 if (Subtarget.isAIXABI() && Subtarget.hasAltivec() && 353 !TM.getAIXExtendedAltivecABI()) { 354 // In the AIX default Altivec ABI, vector registers VR20-VR31 are reserved 355 // and cannot be used. 356 for (auto Reg : CSR_Altivec_SaveList) { 357 if (Reg == 0) 358 break; 359 markSuperRegs(Reserved, Reg); 360 for (MCRegAliasIterator AS(Reg, this, true); AS.isValid(); ++AS) { 361 Reserved.set(*AS); 362 } 363 } 364 } 365 366 assert(checkAllSuperRegsMarked(Reserved)); 367 return Reserved; 368 } 369 370 bool PPCRegisterInfo::requiresFrameIndexScavenging(const MachineFunction &MF) const { 371 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 372 const PPCInstrInfo *InstrInfo = Subtarget.getInstrInfo(); 373 const MachineFrameInfo &MFI = MF.getFrameInfo(); 374 const std::vector<CalleeSavedInfo> &Info = MFI.getCalleeSavedInfo(); 375 376 LLVM_DEBUG(dbgs() << "requiresFrameIndexScavenging for " << MF.getName() 377 << ".\n"); 378 // If the callee saved info is invalid we have to default to true for safety. 379 if (!MFI.isCalleeSavedInfoValid()) { 380 LLVM_DEBUG(dbgs() << "TRUE - Invalid callee saved info.\n"); 381 return true; 382 } 383 384 // We will require the use of X-Forms because the frame is larger than what 385 // can be represented in signed 16 bits that fit in the immediate of a D-Form. 386 // If we need an X-Form then we need a register to store the address offset. 387 unsigned FrameSize = MFI.getStackSize(); 388 // Signed 16 bits means that the FrameSize cannot be more than 15 bits. 389 if (FrameSize & ~0x7FFF) { 390 LLVM_DEBUG(dbgs() << "TRUE - Frame size is too large for D-Form.\n"); 391 return true; 392 } 393 394 // The callee saved info is valid so it can be traversed. 395 // Checking for registers that need saving that do not have load or store 396 // forms where the address offset is an immediate. 397 for (unsigned i = 0; i < Info.size(); i++) { 398 // If the spill is to a register no scavenging is required. 399 if (Info[i].isSpilledToReg()) 400 continue; 401 402 int FrIdx = Info[i].getFrameIdx(); 403 unsigned Reg = Info[i].getReg(); 404 405 const TargetRegisterClass *RC = getMinimalPhysRegClass(Reg); 406 unsigned Opcode = InstrInfo->getStoreOpcodeForSpill(RC); 407 if (!MFI.isFixedObjectIndex(FrIdx)) { 408 // This is not a fixed object. If it requires alignment then we may still 409 // need to use the XForm. 410 if (offsetMinAlignForOpcode(Opcode) > 1) { 411 LLVM_DEBUG(dbgs() << "Memory Operand: " << InstrInfo->getName(Opcode) 412 << " for register " << printReg(Reg, this) << ".\n"); 413 LLVM_DEBUG(dbgs() << "TRUE - Not fixed frame object that requires " 414 << "alignment.\n"); 415 return true; 416 } 417 } 418 419 // This is eiher: 420 // 1) A fixed frame index object which we know are aligned so 421 // as long as we have a valid DForm/DSForm/DQForm (non XForm) we don't 422 // need to consider the alignment here. 423 // 2) A not fixed object but in that case we now know that the min required 424 // alignment is no more than 1 based on the previous check. 425 if (InstrInfo->isXFormMemOp(Opcode)) { 426 LLVM_DEBUG(dbgs() << "Memory Operand: " << InstrInfo->getName(Opcode) 427 << " for register " << printReg(Reg, this) << ".\n"); 428 LLVM_DEBUG(dbgs() << "TRUE - Memory operand is X-Form.\n"); 429 return true; 430 } 431 } 432 LLVM_DEBUG(dbgs() << "FALSE - Scavenging is not required.\n"); 433 return false; 434 } 435 436 bool PPCRegisterInfo::requiresVirtualBaseRegisters( 437 const MachineFunction &MF) const { 438 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 439 // Do not use virtual base registers when ROP protection is turned on. 440 // Virtual base registers break the layout of the local variable space and may 441 // push the ROP Hash location past the 512 byte range of the ROP store 442 // instruction. 443 return !Subtarget.hasROPProtect(); 444 } 445 446 bool PPCRegisterInfo::isCallerPreservedPhysReg(MCRegister PhysReg, 447 const MachineFunction &MF) const { 448 assert(Register::isPhysicalRegister(PhysReg)); 449 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 450 const MachineFrameInfo &MFI = MF.getFrameInfo(); 451 452 if (!Subtarget.is64BitELFABI() && !Subtarget.isAIXABI()) 453 return false; 454 if (PhysReg == Subtarget.getTOCPointerRegister()) 455 // X2/R2 is guaranteed to be preserved within a function if it is reserved. 456 // The reason it's reserved is that it's the TOC pointer (and the function 457 // uses the TOC). In functions where it isn't reserved (i.e. leaf functions 458 // with no TOC access), we can't claim that it is preserved. 459 return (getReservedRegs(MF).test(PhysReg)); 460 if (StackPtrConst && PhysReg == Subtarget.getStackPointerRegister() && 461 !MFI.hasVarSizedObjects() && !MFI.hasOpaqueSPAdjustment()) 462 // The value of the stack pointer does not change within a function after 463 // the prologue and before the epilogue if there are no dynamic allocations 464 // and no inline asm which clobbers X1/R1. 465 return true; 466 return false; 467 } 468 469 unsigned PPCRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC, 470 MachineFunction &MF) const { 471 const PPCFrameLowering *TFI = getFrameLowering(MF); 472 const unsigned DefaultSafety = 1; 473 474 switch (RC->getID()) { 475 default: 476 return 0; 477 case PPC::G8RC_NOX0RegClassID: 478 case PPC::GPRC_NOR0RegClassID: 479 case PPC::SPERCRegClassID: 480 case PPC::G8RCRegClassID: 481 case PPC::GPRCRegClassID: { 482 unsigned FP = TFI->hasFP(MF) ? 1 : 0; 483 return 32 - FP - DefaultSafety; 484 } 485 case PPC::F4RCRegClassID: 486 case PPC::F8RCRegClassID: 487 case PPC::VSLRCRegClassID: 488 return 32 - DefaultSafety; 489 case PPC::VFRCRegClassID: 490 case PPC::VRRCRegClassID: { 491 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 492 // Vector registers VR20-VR31 are reserved and cannot be used in the default 493 // Altivec ABI on AIX. 494 if (!TM.getAIXExtendedAltivecABI() && Subtarget.isAIXABI()) 495 return 20 - DefaultSafety; 496 } 497 return 32 - DefaultSafety; 498 case PPC::VSFRCRegClassID: 499 case PPC::VSSRCRegClassID: 500 case PPC::VSRCRegClassID: { 501 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 502 if (!TM.getAIXExtendedAltivecABI() && Subtarget.isAIXABI()) 503 // Vector registers VR20-VR31 are reserved and cannot be used in the 504 // default Altivec ABI on AIX. 505 return 52 - DefaultSafety; 506 } 507 return 64 - DefaultSafety; 508 case PPC::CRRCRegClassID: 509 return 8 - DefaultSafety; 510 } 511 } 512 513 const TargetRegisterClass * 514 PPCRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC, 515 const MachineFunction &MF) const { 516 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 517 const auto *DefaultSuperclass = 518 TargetRegisterInfo::getLargestLegalSuperClass(RC, MF); 519 if (Subtarget.hasVSX()) { 520 // With VSX, we can inflate various sub-register classes to the full VSX 521 // register set. 522 523 // For Power9 we allow the user to enable GPR to vector spills. 524 // FIXME: Currently limited to spilling GP8RC. A follow on patch will add 525 // support to spill GPRC. 526 if (TM.isELFv2ABI() || Subtarget.isAIXABI()) { 527 if (Subtarget.hasP9Vector() && EnableGPRToVecSpills && 528 RC == &PPC::G8RCRegClass) { 529 InflateGP8RC++; 530 return &PPC::SPILLTOVSRRCRegClass; 531 } 532 if (RC == &PPC::GPRCRegClass && EnableGPRToVecSpills) 533 InflateGPRC++; 534 } 535 536 for (const auto *I = RC->getSuperClasses(); *I; ++I) { 537 if (getRegSizeInBits(**I) != getRegSizeInBits(*RC)) 538 continue; 539 540 switch ((*I)->getID()) { 541 case PPC::VSSRCRegClassID: 542 return Subtarget.hasP8Vector() ? *I : DefaultSuperclass; 543 case PPC::VSFRCRegClassID: 544 case PPC::VSRCRegClassID: 545 return *I; 546 case PPC::VSRpRCRegClassID: 547 return Subtarget.pairedVectorMemops() ? *I : DefaultSuperclass; 548 case PPC::ACCRCRegClassID: 549 case PPC::UACCRCRegClassID: 550 return Subtarget.hasMMA() ? *I : DefaultSuperclass; 551 } 552 } 553 } 554 555 return DefaultSuperclass; 556 } 557 558 //===----------------------------------------------------------------------===// 559 // Stack Frame Processing methods 560 //===----------------------------------------------------------------------===// 561 562 /// lowerDynamicAlloc - Generate the code for allocating an object in the 563 /// current frame. The sequence of code will be in the general form 564 /// 565 /// addi R0, SP, \#frameSize ; get the address of the previous frame 566 /// stwxu R0, SP, Rnegsize ; add and update the SP with the negated size 567 /// addi Rnew, SP, \#maxCalFrameSize ; get the top of the allocation 568 /// 569 void PPCRegisterInfo::lowerDynamicAlloc(MachineBasicBlock::iterator II) const { 570 // Get the instruction. 571 MachineInstr &MI = *II; 572 // Get the instruction's basic block. 573 MachineBasicBlock &MBB = *MI.getParent(); 574 // Get the basic block's function. 575 MachineFunction &MF = *MBB.getParent(); 576 // Get the frame info. 577 MachineFrameInfo &MFI = MF.getFrameInfo(); 578 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 579 // Get the instruction info. 580 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 581 // Determine whether 64-bit pointers are used. 582 bool LP64 = TM.isPPC64(); 583 DebugLoc dl = MI.getDebugLoc(); 584 585 // Get the maximum call stack size. 586 unsigned maxCallFrameSize = MFI.getMaxCallFrameSize(); 587 Align MaxAlign = MFI.getMaxAlign(); 588 assert(isAligned(MaxAlign, maxCallFrameSize) && 589 "Maximum call-frame size not sufficiently aligned"); 590 (void)MaxAlign; 591 592 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 593 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 594 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 595 bool KillNegSizeReg = MI.getOperand(1).isKill(); 596 Register NegSizeReg = MI.getOperand(1).getReg(); 597 598 prepareDynamicAlloca(II, NegSizeReg, KillNegSizeReg, Reg); 599 // Grow the stack and update the stack pointer link, then determine the 600 // address of new allocated space. 601 if (LP64) { 602 BuildMI(MBB, II, dl, TII.get(PPC::STDUX), PPC::X1) 603 .addReg(Reg, RegState::Kill) 604 .addReg(PPC::X1) 605 .addReg(NegSizeReg, getKillRegState(KillNegSizeReg)); 606 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), MI.getOperand(0).getReg()) 607 .addReg(PPC::X1) 608 .addImm(maxCallFrameSize); 609 } else { 610 BuildMI(MBB, II, dl, TII.get(PPC::STWUX), PPC::R1) 611 .addReg(Reg, RegState::Kill) 612 .addReg(PPC::R1) 613 .addReg(NegSizeReg, getKillRegState(KillNegSizeReg)); 614 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), MI.getOperand(0).getReg()) 615 .addReg(PPC::R1) 616 .addImm(maxCallFrameSize); 617 } 618 619 // Discard the DYNALLOC instruction. 620 MBB.erase(II); 621 } 622 623 /// To accomplish dynamic stack allocation, we have to calculate exact size 624 /// subtracted from the stack pointer according alignment information and get 625 /// previous frame pointer. 626 void PPCRegisterInfo::prepareDynamicAlloca(MachineBasicBlock::iterator II, 627 Register &NegSizeReg, 628 bool &KillNegSizeReg, 629 Register &FramePointer) const { 630 // Get the instruction. 631 MachineInstr &MI = *II; 632 // Get the instruction's basic block. 633 MachineBasicBlock &MBB = *MI.getParent(); 634 // Get the basic block's function. 635 MachineFunction &MF = *MBB.getParent(); 636 // Get the frame info. 637 MachineFrameInfo &MFI = MF.getFrameInfo(); 638 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 639 // Get the instruction info. 640 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 641 // Determine whether 64-bit pointers are used. 642 bool LP64 = TM.isPPC64(); 643 DebugLoc dl = MI.getDebugLoc(); 644 // Get the total frame size. 645 unsigned FrameSize = MFI.getStackSize(); 646 647 // Get stack alignments. 648 const PPCFrameLowering *TFI = getFrameLowering(MF); 649 Align TargetAlign = TFI->getStackAlign(); 650 Align MaxAlign = MFI.getMaxAlign(); 651 652 // Determine the previous frame's address. If FrameSize can't be 653 // represented as 16 bits or we need special alignment, then we load the 654 // previous frame's address from 0(SP). Why not do an addis of the hi? 655 // Because R0 is our only safe tmp register and addi/addis treat R0 as zero. 656 // Constructing the constant and adding would take 3 instructions. 657 // Fortunately, a frame greater than 32K is rare. 658 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 659 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 660 661 if (MaxAlign < TargetAlign && isInt<16>(FrameSize)) { 662 if (LP64) 663 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), FramePointer) 664 .addReg(PPC::X31) 665 .addImm(FrameSize); 666 else 667 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), FramePointer) 668 .addReg(PPC::R31) 669 .addImm(FrameSize); 670 } else if (LP64) { 671 BuildMI(MBB, II, dl, TII.get(PPC::LD), FramePointer) 672 .addImm(0) 673 .addReg(PPC::X1); 674 } else { 675 BuildMI(MBB, II, dl, TII.get(PPC::LWZ), FramePointer) 676 .addImm(0) 677 .addReg(PPC::R1); 678 } 679 // Determine the actual NegSizeReg according to alignment info. 680 if (LP64) { 681 if (MaxAlign > TargetAlign) { 682 unsigned UnalNegSizeReg = NegSizeReg; 683 NegSizeReg = MF.getRegInfo().createVirtualRegister(G8RC); 684 685 // Unfortunately, there is no andi, only andi., and we can't insert that 686 // here because we might clobber cr0 while it is live. 687 BuildMI(MBB, II, dl, TII.get(PPC::LI8), NegSizeReg) 688 .addImm(~(MaxAlign.value() - 1)); 689 690 unsigned NegSizeReg1 = NegSizeReg; 691 NegSizeReg = MF.getRegInfo().createVirtualRegister(G8RC); 692 BuildMI(MBB, II, dl, TII.get(PPC::AND8), NegSizeReg) 693 .addReg(UnalNegSizeReg, getKillRegState(KillNegSizeReg)) 694 .addReg(NegSizeReg1, RegState::Kill); 695 KillNegSizeReg = true; 696 } 697 } else { 698 if (MaxAlign > TargetAlign) { 699 unsigned UnalNegSizeReg = NegSizeReg; 700 NegSizeReg = MF.getRegInfo().createVirtualRegister(GPRC); 701 702 // Unfortunately, there is no andi, only andi., and we can't insert that 703 // here because we might clobber cr0 while it is live. 704 BuildMI(MBB, II, dl, TII.get(PPC::LI), NegSizeReg) 705 .addImm(~(MaxAlign.value() - 1)); 706 707 unsigned NegSizeReg1 = NegSizeReg; 708 NegSizeReg = MF.getRegInfo().createVirtualRegister(GPRC); 709 BuildMI(MBB, II, dl, TII.get(PPC::AND), NegSizeReg) 710 .addReg(UnalNegSizeReg, getKillRegState(KillNegSizeReg)) 711 .addReg(NegSizeReg1, RegState::Kill); 712 KillNegSizeReg = true; 713 } 714 } 715 } 716 717 void PPCRegisterInfo::lowerPrepareProbedAlloca( 718 MachineBasicBlock::iterator II) const { 719 MachineInstr &MI = *II; 720 // Get the instruction's basic block. 721 MachineBasicBlock &MBB = *MI.getParent(); 722 // Get the basic block's function. 723 MachineFunction &MF = *MBB.getParent(); 724 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 725 // Get the instruction info. 726 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 727 // Determine whether 64-bit pointers are used. 728 bool LP64 = TM.isPPC64(); 729 DebugLoc dl = MI.getDebugLoc(); 730 Register FramePointer = MI.getOperand(0).getReg(); 731 const Register ActualNegSizeReg = MI.getOperand(1).getReg(); 732 bool KillNegSizeReg = MI.getOperand(2).isKill(); 733 Register NegSizeReg = MI.getOperand(2).getReg(); 734 const MCInstrDesc &CopyInst = TII.get(LP64 ? PPC::OR8 : PPC::OR); 735 // RegAllocator might allocate FramePointer and NegSizeReg in the same phyreg. 736 if (FramePointer == NegSizeReg) { 737 assert(KillNegSizeReg && "FramePointer is a def and NegSizeReg is an use, " 738 "NegSizeReg should be killed"); 739 // FramePointer is clobbered earlier than the use of NegSizeReg in 740 // prepareDynamicAlloca, save NegSizeReg in ActualNegSizeReg to avoid 741 // misuse. 742 BuildMI(MBB, II, dl, CopyInst, ActualNegSizeReg) 743 .addReg(NegSizeReg) 744 .addReg(NegSizeReg); 745 NegSizeReg = ActualNegSizeReg; 746 KillNegSizeReg = false; 747 } 748 prepareDynamicAlloca(II, NegSizeReg, KillNegSizeReg, FramePointer); 749 // NegSizeReg might be updated in prepareDynamicAlloca if MaxAlign > 750 // TargetAlign. 751 if (NegSizeReg != ActualNegSizeReg) 752 BuildMI(MBB, II, dl, CopyInst, ActualNegSizeReg) 753 .addReg(NegSizeReg) 754 .addReg(NegSizeReg); 755 MBB.erase(II); 756 } 757 758 void PPCRegisterInfo::lowerDynamicAreaOffset( 759 MachineBasicBlock::iterator II) const { 760 // Get the instruction. 761 MachineInstr &MI = *II; 762 // Get the instruction's basic block. 763 MachineBasicBlock &MBB = *MI.getParent(); 764 // Get the basic block's function. 765 MachineFunction &MF = *MBB.getParent(); 766 // Get the frame info. 767 MachineFrameInfo &MFI = MF.getFrameInfo(); 768 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 769 // Get the instruction info. 770 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 771 772 unsigned maxCallFrameSize = MFI.getMaxCallFrameSize(); 773 bool is64Bit = TM.isPPC64(); 774 DebugLoc dl = MI.getDebugLoc(); 775 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LI8 : PPC::LI), 776 MI.getOperand(0).getReg()) 777 .addImm(maxCallFrameSize); 778 MBB.erase(II); 779 } 780 781 /// lowerCRSpilling - Generate the code for spilling a CR register. Instead of 782 /// reserving a whole register (R0), we scrounge for one here. This generates 783 /// code like this: 784 /// 785 /// mfcr rA ; Move the conditional register into GPR rA. 786 /// rlwinm rA, rA, SB, 0, 31 ; Shift the bits left so they are in CR0's slot. 787 /// stw rA, FI ; Store rA to the frame. 788 /// 789 void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II, 790 unsigned FrameIndex) const { 791 // Get the instruction. 792 MachineInstr &MI = *II; // ; SPILL_CR <SrcReg>, <offset> 793 // Get the instruction's basic block. 794 MachineBasicBlock &MBB = *MI.getParent(); 795 MachineFunction &MF = *MBB.getParent(); 796 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 797 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 798 DebugLoc dl = MI.getDebugLoc(); 799 800 bool LP64 = TM.isPPC64(); 801 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 802 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 803 804 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 805 Register SrcReg = MI.getOperand(0).getReg(); 806 807 // We need to store the CR in the low 4-bits of the saved value. First, issue 808 // an MFOCRF to save all of the CRBits and, if needed, kill the SrcReg. 809 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), Reg) 810 .addReg(SrcReg, getKillRegState(MI.getOperand(0).isKill())); 811 812 // If the saved register wasn't CR0, shift the bits left so that they are in 813 // CR0's slot. 814 if (SrcReg != PPC::CR0) { 815 Register Reg1 = Reg; 816 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 817 818 // rlwinm rA, rA, ShiftBits, 0, 31. 819 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 820 .addReg(Reg1, RegState::Kill) 821 .addImm(getEncodingValue(SrcReg) * 4) 822 .addImm(0) 823 .addImm(31); 824 } 825 826 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW)) 827 .addReg(Reg, RegState::Kill), 828 FrameIndex); 829 830 // Discard the pseudo instruction. 831 MBB.erase(II); 832 } 833 834 void PPCRegisterInfo::lowerCRRestore(MachineBasicBlock::iterator II, 835 unsigned FrameIndex) const { 836 // Get the instruction. 837 MachineInstr &MI = *II; // ; <DestReg> = RESTORE_CR <offset> 838 // Get the instruction's basic block. 839 MachineBasicBlock &MBB = *MI.getParent(); 840 MachineFunction &MF = *MBB.getParent(); 841 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 842 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 843 DebugLoc dl = MI.getDebugLoc(); 844 845 bool LP64 = TM.isPPC64(); 846 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 847 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 848 849 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 850 Register DestReg = MI.getOperand(0).getReg(); 851 assert(MI.definesRegister(DestReg) && 852 "RESTORE_CR does not define its destination"); 853 854 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LWZ8 : PPC::LWZ), 855 Reg), FrameIndex); 856 857 // If the reloaded register isn't CR0, shift the bits right so that they are 858 // in the right CR's slot. 859 if (DestReg != PPC::CR0) { 860 Register Reg1 = Reg; 861 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 862 863 unsigned ShiftBits = getEncodingValue(DestReg)*4; 864 // rlwinm r11, r11, 32-ShiftBits, 0, 31. 865 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 866 .addReg(Reg1, RegState::Kill).addImm(32-ShiftBits).addImm(0) 867 .addImm(31); 868 } 869 870 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MTOCRF8 : PPC::MTOCRF), DestReg) 871 .addReg(Reg, RegState::Kill); 872 873 // Discard the pseudo instruction. 874 MBB.erase(II); 875 } 876 877 void PPCRegisterInfo::lowerCRBitSpilling(MachineBasicBlock::iterator II, 878 unsigned FrameIndex) const { 879 // Get the instruction. 880 MachineInstr &MI = *II; // ; SPILL_CRBIT <SrcReg>, <offset> 881 // Get the instruction's basic block. 882 MachineBasicBlock &MBB = *MI.getParent(); 883 MachineFunction &MF = *MBB.getParent(); 884 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 885 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 886 const TargetRegisterInfo* TRI = Subtarget.getRegisterInfo(); 887 DebugLoc dl = MI.getDebugLoc(); 888 889 bool LP64 = TM.isPPC64(); 890 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 891 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 892 893 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 894 Register SrcReg = MI.getOperand(0).getReg(); 895 896 // Search up the BB to find the definition of the CR bit. 897 MachineBasicBlock::reverse_iterator Ins = MI; 898 MachineBasicBlock::reverse_iterator Rend = MBB.rend(); 899 ++Ins; 900 unsigned CRBitSpillDistance = 0; 901 bool SeenUse = false; 902 for (; Ins != Rend; ++Ins) { 903 // Definition found. 904 if (Ins->modifiesRegister(SrcReg, TRI)) 905 break; 906 // Use found. 907 if (Ins->readsRegister(SrcReg, TRI)) 908 SeenUse = true; 909 // Unable to find CR bit definition within maximum search distance. 910 if (CRBitSpillDistance == MaxCRBitSpillDist) { 911 Ins = MI; 912 break; 913 } 914 // Skip debug instructions when counting CR bit spill distance. 915 if (!Ins->isDebugInstr()) 916 CRBitSpillDistance++; 917 } 918 919 // Unable to find the definition of the CR bit in the MBB. 920 if (Ins == MBB.rend()) 921 Ins = MI; 922 923 bool SpillsKnownBit = false; 924 // There is no need to extract the CR bit if its value is already known. 925 switch (Ins->getOpcode()) { 926 case PPC::CRUNSET: 927 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LI8 : PPC::LI), Reg) 928 .addImm(0); 929 SpillsKnownBit = true; 930 break; 931 case PPC::CRSET: 932 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LIS8 : PPC::LIS), Reg) 933 .addImm(-32768); 934 SpillsKnownBit = true; 935 break; 936 default: 937 // On Power10, we can use SETNBC to spill all CR bits. SETNBC will set all 938 // bits (specifically, it produces a -1 if the CR bit is set). Ultimately, 939 // the bit that is of importance to us is bit 32 (bit 0 of a 32-bit 940 // register), and SETNBC will set this. 941 if (Subtarget.isISA3_1()) { 942 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::SETNBC8 : PPC::SETNBC), Reg) 943 .addReg(SrcReg, RegState::Undef); 944 break; 945 } 946 947 // On Power9, we can use SETB to extract the LT bit. This only works for 948 // the LT bit since SETB produces -1/1/0 for LT/GT/<neither>. So the value 949 // of the bit we care about (32-bit sign bit) will be set to the value of 950 // the LT bit (regardless of the other bits in the CR field). 951 if (Subtarget.isISA3_0()) { 952 if (SrcReg == PPC::CR0LT || SrcReg == PPC::CR1LT || 953 SrcReg == PPC::CR2LT || SrcReg == PPC::CR3LT || 954 SrcReg == PPC::CR4LT || SrcReg == PPC::CR5LT || 955 SrcReg == PPC::CR6LT || SrcReg == PPC::CR7LT) { 956 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::SETB8 : PPC::SETB), Reg) 957 .addReg(getCRFromCRBit(SrcReg), RegState::Undef); 958 break; 959 } 960 } 961 962 // We need to move the CR field that contains the CR bit we are spilling. 963 // The super register may not be explicitly defined (i.e. it can be defined 964 // by a CR-logical that only defines the subreg) so we state that the CR 965 // field is undef. Also, in order to preserve the kill flag on the CR bit, 966 // we add it as an implicit use. 967 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), Reg) 968 .addReg(getCRFromCRBit(SrcReg), RegState::Undef) 969 .addReg(SrcReg, 970 RegState::Implicit | getKillRegState(MI.getOperand(0).isKill())); 971 972 // If the saved register wasn't CR0LT, shift the bits left so that the bit 973 // to store is the first one. Mask all but that bit. 974 Register Reg1 = Reg; 975 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 976 977 // rlwinm rA, rA, ShiftBits, 0, 0. 978 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 979 .addReg(Reg1, RegState::Kill) 980 .addImm(getEncodingValue(SrcReg)) 981 .addImm(0).addImm(0); 982 } 983 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW)) 984 .addReg(Reg, RegState::Kill), 985 FrameIndex); 986 987 bool KillsCRBit = MI.killsRegister(SrcReg, TRI); 988 // Discard the pseudo instruction. 989 MBB.erase(II); 990 if (SpillsKnownBit && KillsCRBit && !SeenUse) { 991 Ins->setDesc(TII.get(PPC::UNENCODED_NOP)); 992 Ins->RemoveOperand(0); 993 } 994 } 995 996 void PPCRegisterInfo::lowerCRBitRestore(MachineBasicBlock::iterator II, 997 unsigned FrameIndex) const { 998 // Get the instruction. 999 MachineInstr &MI = *II; // ; <DestReg> = RESTORE_CRBIT <offset> 1000 // Get the instruction's basic block. 1001 MachineBasicBlock &MBB = *MI.getParent(); 1002 MachineFunction &MF = *MBB.getParent(); 1003 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1004 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1005 DebugLoc dl = MI.getDebugLoc(); 1006 1007 bool LP64 = TM.isPPC64(); 1008 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 1009 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 1010 1011 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 1012 Register DestReg = MI.getOperand(0).getReg(); 1013 assert(MI.definesRegister(DestReg) && 1014 "RESTORE_CRBIT does not define its destination"); 1015 1016 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LWZ8 : PPC::LWZ), 1017 Reg), FrameIndex); 1018 1019 BuildMI(MBB, II, dl, TII.get(TargetOpcode::IMPLICIT_DEF), DestReg); 1020 1021 Register RegO = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 1022 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), RegO) 1023 .addReg(getCRFromCRBit(DestReg)); 1024 1025 unsigned ShiftBits = getEncodingValue(DestReg); 1026 // rlwimi r11, r10, 32-ShiftBits, ..., ... 1027 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWIMI8 : PPC::RLWIMI), RegO) 1028 .addReg(RegO, RegState::Kill) 1029 .addReg(Reg, RegState::Kill) 1030 .addImm(ShiftBits ? 32 - ShiftBits : 0) 1031 .addImm(ShiftBits) 1032 .addImm(ShiftBits); 1033 1034 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MTOCRF8 : PPC::MTOCRF), 1035 getCRFromCRBit(DestReg)) 1036 .addReg(RegO, RegState::Kill) 1037 // Make sure we have a use dependency all the way through this 1038 // sequence of instructions. We can't have the other bits in the CR 1039 // modified in between the mfocrf and the mtocrf. 1040 .addReg(getCRFromCRBit(DestReg), RegState::Implicit); 1041 1042 // Discard the pseudo instruction. 1043 MBB.erase(II); 1044 } 1045 1046 void PPCRegisterInfo::emitAccCopyInfo(MachineBasicBlock &MBB, 1047 MCRegister DestReg, MCRegister SrcReg) { 1048 #ifdef NDEBUG 1049 return; 1050 #else 1051 if (ReportAccMoves) { 1052 std::string Dest = PPC::ACCRCRegClass.contains(DestReg) ? "acc" : "uacc"; 1053 std::string Src = PPC::ACCRCRegClass.contains(SrcReg) ? "acc" : "uacc"; 1054 dbgs() << "Emitting copy from " << Src << " to " << Dest << ":\n"; 1055 MBB.dump(); 1056 } 1057 #endif 1058 } 1059 1060 static void emitAccSpillRestoreInfo(MachineBasicBlock &MBB, bool IsPrimed, 1061 bool IsRestore) { 1062 #ifdef NDEBUG 1063 return; 1064 #else 1065 if (ReportAccMoves) { 1066 dbgs() << "Emitting " << (IsPrimed ? "acc" : "uacc") << " register " 1067 << (IsRestore ? "restore" : "spill") << ":\n"; 1068 MBB.dump(); 1069 } 1070 #endif 1071 } 1072 1073 /// lowerACCSpilling - Generate the code for spilling the accumulator register. 1074 /// Similarly to other spills/reloads that use pseudo-ops, we do not actually 1075 /// eliminate the FrameIndex here nor compute the stack offset. We simply 1076 /// create a real instruction with an FI and rely on eliminateFrameIndex to 1077 /// handle the FI elimination. 1078 void PPCRegisterInfo::lowerACCSpilling(MachineBasicBlock::iterator II, 1079 unsigned FrameIndex) const { 1080 MachineInstr &MI = *II; // SPILL_ACC <SrcReg>, <offset> 1081 MachineBasicBlock &MBB = *MI.getParent(); 1082 MachineFunction &MF = *MBB.getParent(); 1083 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1084 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1085 DebugLoc DL = MI.getDebugLoc(); 1086 Register SrcReg = MI.getOperand(0).getReg(); 1087 bool IsKilled = MI.getOperand(0).isKill(); 1088 1089 bool IsPrimed = PPC::ACCRCRegClass.contains(SrcReg); 1090 Register Reg = 1091 PPC::VSRp0 + (SrcReg - (IsPrimed ? PPC::ACC0 : PPC::UACC0)) * 2; 1092 bool IsLittleEndian = Subtarget.isLittleEndian(); 1093 1094 emitAccSpillRestoreInfo(MBB, IsPrimed, false); 1095 1096 // De-prime the register being spilled, create two stores for the pair 1097 // subregisters accounting for endianness and then re-prime the register if 1098 // it isn't killed. This uses the Offset parameter to addFrameReference() to 1099 // adjust the offset of the store that is within the 64-byte stack slot. 1100 if (IsPrimed) 1101 BuildMI(MBB, II, DL, TII.get(PPC::XXMFACC), SrcReg).addReg(SrcReg); 1102 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STXVP)) 1103 .addReg(Reg, getKillRegState(IsKilled)), 1104 FrameIndex, IsLittleEndian ? 32 : 0); 1105 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STXVP)) 1106 .addReg(Reg + 1, getKillRegState(IsKilled)), 1107 FrameIndex, IsLittleEndian ? 0 : 32); 1108 if (IsPrimed && !IsKilled) 1109 BuildMI(MBB, II, DL, TII.get(PPC::XXMTACC), SrcReg).addReg(SrcReg); 1110 1111 // Discard the pseudo instruction. 1112 MBB.erase(II); 1113 } 1114 1115 /// lowerACCRestore - Generate the code to restore the accumulator register. 1116 void PPCRegisterInfo::lowerACCRestore(MachineBasicBlock::iterator II, 1117 unsigned FrameIndex) const { 1118 MachineInstr &MI = *II; // <DestReg> = RESTORE_ACC <offset> 1119 MachineBasicBlock &MBB = *MI.getParent(); 1120 MachineFunction &MF = *MBB.getParent(); 1121 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1122 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1123 DebugLoc DL = MI.getDebugLoc(); 1124 1125 Register DestReg = MI.getOperand(0).getReg(); 1126 assert(MI.definesRegister(DestReg) && 1127 "RESTORE_ACC does not define its destination"); 1128 1129 bool IsPrimed = PPC::ACCRCRegClass.contains(DestReg); 1130 Register Reg = 1131 PPC::VSRp0 + (DestReg - (IsPrimed ? PPC::ACC0 : PPC::UACC0)) * 2; 1132 bool IsLittleEndian = Subtarget.isLittleEndian(); 1133 1134 emitAccSpillRestoreInfo(MBB, IsPrimed, true); 1135 1136 // Create two loads for the pair subregisters accounting for endianness and 1137 // then prime the accumulator register being restored. 1138 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LXVP), Reg), 1139 FrameIndex, IsLittleEndian ? 32 : 0); 1140 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LXVP), Reg + 1), 1141 FrameIndex, IsLittleEndian ? 0 : 32); 1142 if (IsPrimed) 1143 BuildMI(MBB, II, DL, TII.get(PPC::XXMTACC), DestReg).addReg(DestReg); 1144 1145 // Discard the pseudo instruction. 1146 MBB.erase(II); 1147 } 1148 1149 /// lowerQuadwordSpilling - Generate code to spill paired general register. 1150 void PPCRegisterInfo::lowerQuadwordSpilling(MachineBasicBlock::iterator II, 1151 unsigned FrameIndex) const { 1152 MachineInstr &MI = *II; 1153 MachineBasicBlock &MBB = *MI.getParent(); 1154 MachineFunction &MF = *MBB.getParent(); 1155 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1156 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1157 DebugLoc DL = MI.getDebugLoc(); 1158 1159 Register SrcReg = MI.getOperand(0).getReg(); 1160 bool IsKilled = MI.getOperand(0).isKill(); 1161 1162 Register Reg = PPC::X0 + (SrcReg - PPC::G8p0) * 2; 1163 bool IsLittleEndian = Subtarget.isLittleEndian(); 1164 1165 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STD)) 1166 .addReg(Reg, getKillRegState(IsKilled)), 1167 FrameIndex, IsLittleEndian ? 8 : 0); 1168 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STD)) 1169 .addReg(Reg + 1, getKillRegState(IsKilled)), 1170 FrameIndex, IsLittleEndian ? 0 : 8); 1171 1172 // Discard the pseudo instruction. 1173 MBB.erase(II); 1174 } 1175 1176 /// lowerQuadwordRestore - Generate code to restore paired general register. 1177 void PPCRegisterInfo::lowerQuadwordRestore(MachineBasicBlock::iterator II, 1178 unsigned FrameIndex) const { 1179 MachineInstr &MI = *II; 1180 MachineBasicBlock &MBB = *MI.getParent(); 1181 MachineFunction &MF = *MBB.getParent(); 1182 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1183 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1184 DebugLoc DL = MI.getDebugLoc(); 1185 1186 Register DestReg = MI.getOperand(0).getReg(); 1187 assert(MI.definesRegister(DestReg) && 1188 "RESTORE_QUADWORD does not define its destination"); 1189 1190 Register Reg = PPC::X0 + (DestReg - PPC::G8p0) * 2; 1191 bool IsLittleEndian = Subtarget.isLittleEndian(); 1192 1193 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LD), Reg), FrameIndex, 1194 IsLittleEndian ? 8 : 0); 1195 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LD), Reg + 1), FrameIndex, 1196 IsLittleEndian ? 0 : 8); 1197 1198 // Discard the pseudo instruction. 1199 MBB.erase(II); 1200 } 1201 1202 bool PPCRegisterInfo::hasReservedSpillSlot(const MachineFunction &MF, 1203 Register Reg, int &FrameIdx) const { 1204 // For the nonvolatile condition registers (CR2, CR3, CR4) return true to 1205 // prevent allocating an additional frame slot. 1206 // For 64-bit ELF and AIX, the CR save area is in the linkage area at SP+8, 1207 // for 32-bit AIX the CR save area is in the linkage area at SP+4. 1208 // We have created a FrameIndex to that spill slot to keep the CalleSaveInfos 1209 // valid. 1210 // For 32-bit ELF, we have previously created the stack slot if needed, so 1211 // return its FrameIdx. 1212 if (PPC::CR2 <= Reg && Reg <= PPC::CR4) { 1213 FrameIdx = MF.getInfo<PPCFunctionInfo>()->getCRSpillFrameIndex(); 1214 return true; 1215 } 1216 return false; 1217 } 1218 1219 // If the offset must be a multiple of some value, return what that value is. 1220 static unsigned offsetMinAlignForOpcode(unsigned OpC) { 1221 switch (OpC) { 1222 default: 1223 return 1; 1224 case PPC::LWA: 1225 case PPC::LWA_32: 1226 case PPC::LD: 1227 case PPC::LDU: 1228 case PPC::STD: 1229 case PPC::STDU: 1230 case PPC::DFLOADf32: 1231 case PPC::DFLOADf64: 1232 case PPC::DFSTOREf32: 1233 case PPC::DFSTOREf64: 1234 case PPC::LXSD: 1235 case PPC::LXSSP: 1236 case PPC::STXSD: 1237 case PPC::STXSSP: 1238 case PPC::STQ: 1239 return 4; 1240 case PPC::EVLDD: 1241 case PPC::EVSTDD: 1242 return 8; 1243 case PPC::LXV: 1244 case PPC::STXV: 1245 case PPC::LQ: 1246 return 16; 1247 } 1248 } 1249 1250 // If the offset must be a multiple of some value, return what that value is. 1251 static unsigned offsetMinAlign(const MachineInstr &MI) { 1252 unsigned OpC = MI.getOpcode(); 1253 return offsetMinAlignForOpcode(OpC); 1254 } 1255 1256 // Return the OffsetOperandNo given the FIOperandNum (and the instruction). 1257 static unsigned getOffsetONFromFION(const MachineInstr &MI, 1258 unsigned FIOperandNum) { 1259 // Take into account whether it's an add or mem instruction 1260 unsigned OffsetOperandNo = (FIOperandNum == 2) ? 1 : 2; 1261 if (MI.isInlineAsm()) 1262 OffsetOperandNo = FIOperandNum - 1; 1263 else if (MI.getOpcode() == TargetOpcode::STACKMAP || 1264 MI.getOpcode() == TargetOpcode::PATCHPOINT) 1265 OffsetOperandNo = FIOperandNum + 1; 1266 1267 return OffsetOperandNo; 1268 } 1269 1270 void 1271 PPCRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 1272 int SPAdj, unsigned FIOperandNum, 1273 RegScavenger *RS) const { 1274 assert(SPAdj == 0 && "Unexpected"); 1275 1276 // Get the instruction. 1277 MachineInstr &MI = *II; 1278 // Get the instruction's basic block. 1279 MachineBasicBlock &MBB = *MI.getParent(); 1280 // Get the basic block's function. 1281 MachineFunction &MF = *MBB.getParent(); 1282 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1283 // Get the instruction info. 1284 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1285 // Get the frame info. 1286 MachineFrameInfo &MFI = MF.getFrameInfo(); 1287 DebugLoc dl = MI.getDebugLoc(); 1288 1289 unsigned OffsetOperandNo = getOffsetONFromFION(MI, FIOperandNum); 1290 1291 // Get the frame index. 1292 int FrameIndex = MI.getOperand(FIOperandNum).getIndex(); 1293 1294 // Get the frame pointer save index. Users of this index are primarily 1295 // DYNALLOC instructions. 1296 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 1297 int FPSI = FI->getFramePointerSaveIndex(); 1298 // Get the instruction opcode. 1299 unsigned OpC = MI.getOpcode(); 1300 1301 if ((OpC == PPC::DYNAREAOFFSET || OpC == PPC::DYNAREAOFFSET8)) { 1302 lowerDynamicAreaOffset(II); 1303 return; 1304 } 1305 1306 // Special case for dynamic alloca. 1307 if (FPSI && FrameIndex == FPSI && 1308 (OpC == PPC::DYNALLOC || OpC == PPC::DYNALLOC8)) { 1309 lowerDynamicAlloc(II); 1310 return; 1311 } 1312 1313 if (FPSI && FrameIndex == FPSI && 1314 (OpC == PPC::PREPARE_PROBED_ALLOCA_64 || 1315 OpC == PPC::PREPARE_PROBED_ALLOCA_32 || 1316 OpC == PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_64 || 1317 OpC == PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_32)) { 1318 lowerPrepareProbedAlloca(II); 1319 return; 1320 } 1321 1322 // Special case for pseudo-ops SPILL_CR and RESTORE_CR, etc. 1323 if (OpC == PPC::SPILL_CR) { 1324 lowerCRSpilling(II, FrameIndex); 1325 return; 1326 } else if (OpC == PPC::RESTORE_CR) { 1327 lowerCRRestore(II, FrameIndex); 1328 return; 1329 } else if (OpC == PPC::SPILL_CRBIT) { 1330 lowerCRBitSpilling(II, FrameIndex); 1331 return; 1332 } else if (OpC == PPC::RESTORE_CRBIT) { 1333 lowerCRBitRestore(II, FrameIndex); 1334 return; 1335 } else if (OpC == PPC::SPILL_ACC || OpC == PPC::SPILL_UACC) { 1336 lowerACCSpilling(II, FrameIndex); 1337 return; 1338 } else if (OpC == PPC::RESTORE_ACC || OpC == PPC::RESTORE_UACC) { 1339 lowerACCRestore(II, FrameIndex); 1340 return; 1341 } else if (OpC == PPC::SPILL_QUADWORD) { 1342 lowerQuadwordSpilling(II, FrameIndex); 1343 return; 1344 } else if (OpC == PPC::RESTORE_QUADWORD) { 1345 lowerQuadwordRestore(II, FrameIndex); 1346 return; 1347 } 1348 1349 // Replace the FrameIndex with base register with GPR1 (SP) or GPR31 (FP). 1350 MI.getOperand(FIOperandNum).ChangeToRegister( 1351 FrameIndex < 0 ? getBaseRegister(MF) : getFrameRegister(MF), false); 1352 1353 // If the instruction is not present in ImmToIdxMap, then it has no immediate 1354 // form (and must be r+r). 1355 bool noImmForm = !MI.isInlineAsm() && OpC != TargetOpcode::STACKMAP && 1356 OpC != TargetOpcode::PATCHPOINT && !ImmToIdxMap.count(OpC); 1357 1358 // Now add the frame object offset to the offset from r1. 1359 int Offset = MFI.getObjectOffset(FrameIndex); 1360 Offset += MI.getOperand(OffsetOperandNo).getImm(); 1361 1362 // If we're not using a Frame Pointer that has been set to the value of the 1363 // SP before having the stack size subtracted from it, then add the stack size 1364 // to Offset to get the correct offset. 1365 // Naked functions have stack size 0, although getStackSize may not reflect 1366 // that because we didn't call all the pieces that compute it for naked 1367 // functions. 1368 if (!MF.getFunction().hasFnAttribute(Attribute::Naked)) { 1369 if (!(hasBasePointer(MF) && FrameIndex < 0)) 1370 Offset += MFI.getStackSize(); 1371 } 1372 1373 // If we can, encode the offset directly into the instruction. If this is a 1374 // normal PPC "ri" instruction, any 16-bit value can be safely encoded. If 1375 // this is a PPC64 "ix" instruction, only a 16-bit value with the low two bits 1376 // clear can be encoded. This is extremely uncommon, because normally you 1377 // only "std" to a stack slot that is at least 4-byte aligned, but it can 1378 // happen in invalid code. 1379 assert(OpC != PPC::DBG_VALUE && 1380 "This should be handled in a target-independent way"); 1381 bool OffsetFitsMnemonic = (OpC == PPC::EVSTDD || OpC == PPC::EVLDD) ? 1382 isUInt<8>(Offset) : 1383 isInt<16>(Offset); 1384 if (!noImmForm && ((OffsetFitsMnemonic && 1385 ((Offset % offsetMinAlign(MI)) == 0)) || 1386 OpC == TargetOpcode::STACKMAP || 1387 OpC == TargetOpcode::PATCHPOINT)) { 1388 MI.getOperand(OffsetOperandNo).ChangeToImmediate(Offset); 1389 return; 1390 } 1391 1392 // The offset doesn't fit into a single register, scavenge one to build the 1393 // offset in. 1394 1395 bool is64Bit = TM.isPPC64(); 1396 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 1397 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 1398 const TargetRegisterClass *RC = is64Bit ? G8RC : GPRC; 1399 Register SRegHi = MF.getRegInfo().createVirtualRegister(RC), 1400 SReg = MF.getRegInfo().createVirtualRegister(RC); 1401 1402 // Insert a set of rA with the full offset value before the ld, st, or add 1403 if (isInt<16>(Offset)) 1404 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LI8 : PPC::LI), SReg) 1405 .addImm(Offset); 1406 else { 1407 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LIS8 : PPC::LIS), SRegHi) 1408 .addImm(Offset >> 16); 1409 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::ORI8 : PPC::ORI), SReg) 1410 .addReg(SRegHi, RegState::Kill) 1411 .addImm(Offset); 1412 } 1413 1414 // Convert into indexed form of the instruction: 1415 // 1416 // sth 0:rA, 1:imm 2:(rB) ==> sthx 0:rA, 2:rB, 1:r0 1417 // addi 0:rA 1:rB, 2, imm ==> add 0:rA, 1:rB, 2:r0 1418 unsigned OperandBase; 1419 1420 if (noImmForm) 1421 OperandBase = 1; 1422 else if (OpC != TargetOpcode::INLINEASM && 1423 OpC != TargetOpcode::INLINEASM_BR) { 1424 assert(ImmToIdxMap.count(OpC) && 1425 "No indexed form of load or store available!"); 1426 unsigned NewOpcode = ImmToIdxMap.find(OpC)->second; 1427 MI.setDesc(TII.get(NewOpcode)); 1428 OperandBase = 1; 1429 } else { 1430 OperandBase = OffsetOperandNo; 1431 } 1432 1433 Register StackReg = MI.getOperand(FIOperandNum).getReg(); 1434 MI.getOperand(OperandBase).ChangeToRegister(StackReg, false); 1435 MI.getOperand(OperandBase + 1).ChangeToRegister(SReg, false, false, true); 1436 } 1437 1438 Register PPCRegisterInfo::getFrameRegister(const MachineFunction &MF) const { 1439 const PPCFrameLowering *TFI = getFrameLowering(MF); 1440 1441 if (!TM.isPPC64()) 1442 return TFI->hasFP(MF) ? PPC::R31 : PPC::R1; 1443 else 1444 return TFI->hasFP(MF) ? PPC::X31 : PPC::X1; 1445 } 1446 1447 Register PPCRegisterInfo::getBaseRegister(const MachineFunction &MF) const { 1448 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1449 if (!hasBasePointer(MF)) 1450 return getFrameRegister(MF); 1451 1452 if (TM.isPPC64()) 1453 return PPC::X30; 1454 1455 if (Subtarget.isSVR4ABI() && TM.isPositionIndependent()) 1456 return PPC::R29; 1457 1458 return PPC::R30; 1459 } 1460 1461 bool PPCRegisterInfo::hasBasePointer(const MachineFunction &MF) const { 1462 if (!EnableBasePointer) 1463 return false; 1464 if (AlwaysBasePointer) 1465 return true; 1466 1467 // If we need to realign the stack, then the stack pointer can no longer 1468 // serve as an offset into the caller's stack space. As a result, we need a 1469 // base pointer. 1470 return hasStackRealignment(MF); 1471 } 1472 1473 /// Returns true if the instruction's frame index 1474 /// reference would be better served by a base register other than FP 1475 /// or SP. Used by LocalStackFrameAllocation to determine which frame index 1476 /// references it should create new base registers for. 1477 bool PPCRegisterInfo:: 1478 needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const { 1479 assert(Offset < 0 && "Local offset must be negative"); 1480 1481 // It's the load/store FI references that cause issues, as it can be difficult 1482 // to materialize the offset if it won't fit in the literal field. Estimate 1483 // based on the size of the local frame and some conservative assumptions 1484 // about the rest of the stack frame (note, this is pre-regalloc, so 1485 // we don't know everything for certain yet) whether this offset is likely 1486 // to be out of range of the immediate. Return true if so. 1487 1488 // We only generate virtual base registers for loads and stores that have 1489 // an r+i form. Return false for everything else. 1490 unsigned OpC = MI->getOpcode(); 1491 if (!ImmToIdxMap.count(OpC)) 1492 return false; 1493 1494 // Don't generate a new virtual base register just to add zero to it. 1495 if ((OpC == PPC::ADDI || OpC == PPC::ADDI8) && 1496 MI->getOperand(2).getImm() == 0) 1497 return false; 1498 1499 MachineBasicBlock &MBB = *MI->getParent(); 1500 MachineFunction &MF = *MBB.getParent(); 1501 const PPCFrameLowering *TFI = getFrameLowering(MF); 1502 unsigned StackEst = TFI->determineFrameLayout(MF, true); 1503 1504 // If we likely don't need a stack frame, then we probably don't need a 1505 // virtual base register either. 1506 if (!StackEst) 1507 return false; 1508 1509 // Estimate an offset from the stack pointer. 1510 // The incoming offset is relating to the SP at the start of the function, 1511 // but when we access the local it'll be relative to the SP after local 1512 // allocation, so adjust our SP-relative offset by that allocation size. 1513 Offset += StackEst; 1514 1515 // The frame pointer will point to the end of the stack, so estimate the 1516 // offset as the difference between the object offset and the FP location. 1517 return !isFrameOffsetLegal(MI, getBaseRegister(MF), Offset); 1518 } 1519 1520 /// Insert defining instruction(s) for BaseReg to 1521 /// be a pointer to FrameIdx at the beginning of the basic block. 1522 Register PPCRegisterInfo::materializeFrameBaseRegister(MachineBasicBlock *MBB, 1523 int FrameIdx, 1524 int64_t Offset) const { 1525 unsigned ADDriOpc = TM.isPPC64() ? PPC::ADDI8 : PPC::ADDI; 1526 1527 MachineBasicBlock::iterator Ins = MBB->begin(); 1528 DebugLoc DL; // Defaults to "unknown" 1529 if (Ins != MBB->end()) 1530 DL = Ins->getDebugLoc(); 1531 1532 const MachineFunction &MF = *MBB->getParent(); 1533 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1534 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1535 const MCInstrDesc &MCID = TII.get(ADDriOpc); 1536 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 1537 const TargetRegisterClass *RC = getPointerRegClass(MF); 1538 Register BaseReg = MRI.createVirtualRegister(RC); 1539 MRI.constrainRegClass(BaseReg, TII.getRegClass(MCID, 0, this, MF)); 1540 1541 BuildMI(*MBB, Ins, DL, MCID, BaseReg) 1542 .addFrameIndex(FrameIdx).addImm(Offset); 1543 1544 return BaseReg; 1545 } 1546 1547 void PPCRegisterInfo::resolveFrameIndex(MachineInstr &MI, Register BaseReg, 1548 int64_t Offset) const { 1549 unsigned FIOperandNum = 0; 1550 while (!MI.getOperand(FIOperandNum).isFI()) { 1551 ++FIOperandNum; 1552 assert(FIOperandNum < MI.getNumOperands() && 1553 "Instr doesn't have FrameIndex operand!"); 1554 } 1555 1556 MI.getOperand(FIOperandNum).ChangeToRegister(BaseReg, false); 1557 unsigned OffsetOperandNo = getOffsetONFromFION(MI, FIOperandNum); 1558 Offset += MI.getOperand(OffsetOperandNo).getImm(); 1559 MI.getOperand(OffsetOperandNo).ChangeToImmediate(Offset); 1560 1561 MachineBasicBlock &MBB = *MI.getParent(); 1562 MachineFunction &MF = *MBB.getParent(); 1563 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1564 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1565 const MCInstrDesc &MCID = MI.getDesc(); 1566 MachineRegisterInfo &MRI = MF.getRegInfo(); 1567 MRI.constrainRegClass(BaseReg, 1568 TII.getRegClass(MCID, FIOperandNum, this, MF)); 1569 } 1570 1571 bool PPCRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, 1572 Register BaseReg, 1573 int64_t Offset) const { 1574 unsigned FIOperandNum = 0; 1575 while (!MI->getOperand(FIOperandNum).isFI()) { 1576 ++FIOperandNum; 1577 assert(FIOperandNum < MI->getNumOperands() && 1578 "Instr doesn't have FrameIndex operand!"); 1579 } 1580 1581 unsigned OffsetOperandNo = getOffsetONFromFION(*MI, FIOperandNum); 1582 Offset += MI->getOperand(OffsetOperandNo).getImm(); 1583 1584 return MI->getOpcode() == PPC::DBG_VALUE || // DBG_VALUE is always Reg+Imm 1585 MI->getOpcode() == TargetOpcode::STACKMAP || 1586 MI->getOpcode() == TargetOpcode::PATCHPOINT || 1587 (isInt<16>(Offset) && (Offset % offsetMinAlign(*MI)) == 0); 1588 } 1589