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 // Power10 138 ImmToIdxMap[PPC::PLBZ] = PPC::LBZX; ImmToIdxMap[PPC::PLBZ8] = PPC::LBZX8; 139 ImmToIdxMap[PPC::PLHZ] = PPC::LHZX; ImmToIdxMap[PPC::PLHZ8] = PPC::LHZX8; 140 ImmToIdxMap[PPC::PLHA] = PPC::LHAX; ImmToIdxMap[PPC::PLHA8] = PPC::LHAX8; 141 ImmToIdxMap[PPC::PLWZ] = PPC::LWZX; ImmToIdxMap[PPC::PLWZ8] = PPC::LWZX8; 142 ImmToIdxMap[PPC::PLWA] = PPC::LWAX; ImmToIdxMap[PPC::PLWA8] = PPC::LWAX; 143 ImmToIdxMap[PPC::PLD] = PPC::LDX; ImmToIdxMap[PPC::PSTD] = PPC::STDX; 144 145 ImmToIdxMap[PPC::PSTB] = PPC::STBX; ImmToIdxMap[PPC::PSTB8] = PPC::STBX8; 146 ImmToIdxMap[PPC::PSTH] = PPC::STHX; ImmToIdxMap[PPC::PSTH8] = PPC::STHX8; 147 ImmToIdxMap[PPC::PSTW] = PPC::STWX; ImmToIdxMap[PPC::PSTW8] = PPC::STWX8; 148 149 ImmToIdxMap[PPC::PLFS] = PPC::LFSX; ImmToIdxMap[PPC::PSTFS] = PPC::STFSX; 150 ImmToIdxMap[PPC::PLFD] = PPC::LFDX; ImmToIdxMap[PPC::PSTFD] = PPC::STFDX; 151 ImmToIdxMap[PPC::PLXSSP] = PPC::LXSSPX; ImmToIdxMap[PPC::PSTXSSP] = PPC::STXSSPX; 152 ImmToIdxMap[PPC::PLXSD] = PPC::LXSDX; ImmToIdxMap[PPC::PSTXSD] = PPC::STXSDX; 153 ImmToIdxMap[PPC::PLXV] = PPC::LXVX; ImmToIdxMap[PPC::PSTXV] = PPC::STXVX; 154 155 ImmToIdxMap[PPC::LXVP] = PPC::LXVPX; 156 ImmToIdxMap[PPC::STXVP] = PPC::STXVPX; 157 ImmToIdxMap[PPC::PLXVP] = PPC::LXVPX; 158 ImmToIdxMap[PPC::PSTXVP] = PPC::STXVPX; 159 } 160 161 /// getPointerRegClass - Return the register class to use to hold pointers. 162 /// This is used for addressing modes. 163 const TargetRegisterClass * 164 PPCRegisterInfo::getPointerRegClass(const MachineFunction &MF, unsigned Kind) 165 const { 166 // Note that PPCInstrInfo::FoldImmediate also directly uses this Kind value 167 // when it checks for ZERO folding. 168 if (Kind == 1) { 169 if (TM.isPPC64()) 170 return &PPC::G8RC_NOX0RegClass; 171 return &PPC::GPRC_NOR0RegClass; 172 } 173 174 if (TM.isPPC64()) 175 return &PPC::G8RCRegClass; 176 return &PPC::GPRCRegClass; 177 } 178 179 const MCPhysReg* 180 PPCRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 181 const PPCSubtarget &Subtarget = MF->getSubtarget<PPCSubtarget>(); 182 if (MF->getFunction().getCallingConv() == CallingConv::AnyReg) { 183 if (!TM.isPPC64() && Subtarget.isAIXABI()) 184 report_fatal_error("AnyReg unimplemented on 32-bit AIX."); 185 if (Subtarget.hasVSX()) { 186 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 187 return CSR_64_AllRegs_AIX_Dflt_VSX_SaveList; 188 return CSR_64_AllRegs_VSX_SaveList; 189 } 190 if (Subtarget.hasAltivec()) { 191 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 192 return CSR_64_AllRegs_AIX_Dflt_Altivec_SaveList; 193 return CSR_64_AllRegs_Altivec_SaveList; 194 } 195 return CSR_64_AllRegs_SaveList; 196 } 197 198 // On PPC64, we might need to save r2 (but only if it is not reserved). 199 // We do not need to treat R2 as callee-saved when using PC-Relative calls 200 // because any direct uses of R2 will cause it to be reserved. If the function 201 // is a leaf or the only uses of R2 are implicit uses for calls, the calls 202 // will use the @notoc relocation which will cause this function to set the 203 // st_other bit to 1, thereby communicating to its caller that it arbitrarily 204 // clobbers the TOC. 205 bool SaveR2 = MF->getRegInfo().isAllocatable(PPC::X2) && 206 !Subtarget.isUsingPCRelativeCalls(); 207 208 // Cold calling convention CSRs. 209 if (MF->getFunction().getCallingConv() == CallingConv::Cold) { 210 if (Subtarget.isAIXABI()) 211 report_fatal_error("Cold calling unimplemented on AIX."); 212 if (TM.isPPC64()) { 213 if (Subtarget.hasAltivec()) 214 return SaveR2 ? CSR_SVR64_ColdCC_R2_Altivec_SaveList 215 : CSR_SVR64_ColdCC_Altivec_SaveList; 216 return SaveR2 ? CSR_SVR64_ColdCC_R2_SaveList 217 : CSR_SVR64_ColdCC_SaveList; 218 } 219 // 32-bit targets. 220 if (Subtarget.hasAltivec()) 221 return CSR_SVR32_ColdCC_Altivec_SaveList; 222 else if (Subtarget.hasSPE()) 223 return CSR_SVR32_ColdCC_SPE_SaveList; 224 return CSR_SVR32_ColdCC_SaveList; 225 } 226 // Standard calling convention CSRs. 227 if (TM.isPPC64()) { 228 if (Subtarget.hasAltivec() && 229 (!Subtarget.isAIXABI() || TM.getAIXExtendedAltivecABI())) { 230 return SaveR2 ? CSR_PPC64_R2_Altivec_SaveList 231 : CSR_PPC64_Altivec_SaveList; 232 } 233 return SaveR2 ? CSR_PPC64_R2_SaveList : CSR_PPC64_SaveList; 234 } 235 // 32-bit targets. 236 if (Subtarget.isAIXABI()) { 237 if (Subtarget.hasAltivec()) 238 return TM.getAIXExtendedAltivecABI() ? CSR_AIX32_Altivec_SaveList 239 : CSR_AIX32_SaveList; 240 return CSR_AIX32_SaveList; 241 } 242 if (Subtarget.hasAltivec()) 243 return CSR_SVR432_Altivec_SaveList; 244 else if (Subtarget.hasSPE()) 245 return CSR_SVR432_SPE_SaveList; 246 return CSR_SVR432_SaveList; 247 } 248 249 const uint32_t * 250 PPCRegisterInfo::getCallPreservedMask(const MachineFunction &MF, 251 CallingConv::ID CC) const { 252 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 253 if (CC == CallingConv::AnyReg) { 254 if (Subtarget.hasVSX()) { 255 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 256 return CSR_64_AllRegs_AIX_Dflt_VSX_RegMask; 257 return CSR_64_AllRegs_VSX_RegMask; 258 } 259 if (Subtarget.hasAltivec()) { 260 if (Subtarget.isAIXABI() && !TM.getAIXExtendedAltivecABI()) 261 return CSR_64_AllRegs_AIX_Dflt_Altivec_RegMask; 262 return CSR_64_AllRegs_Altivec_RegMask; 263 } 264 return CSR_64_AllRegs_RegMask; 265 } 266 267 if (Subtarget.isAIXABI()) { 268 return TM.isPPC64() 269 ? ((Subtarget.hasAltivec() && TM.getAIXExtendedAltivecABI()) 270 ? CSR_PPC64_Altivec_RegMask 271 : CSR_PPC64_RegMask) 272 : ((Subtarget.hasAltivec() && TM.getAIXExtendedAltivecABI()) 273 ? CSR_AIX32_Altivec_RegMask 274 : CSR_AIX32_RegMask); 275 } 276 277 if (CC == CallingConv::Cold) { 278 return TM.isPPC64() ? (Subtarget.hasAltivec() ? CSR_SVR64_ColdCC_Altivec_RegMask 279 : CSR_SVR64_ColdCC_RegMask) 280 : (Subtarget.hasAltivec() ? CSR_SVR32_ColdCC_Altivec_RegMask 281 : (Subtarget.hasSPE() 282 ? CSR_SVR32_ColdCC_SPE_RegMask 283 : CSR_SVR32_ColdCC_RegMask)); 284 } 285 286 return TM.isPPC64() ? (Subtarget.hasAltivec() ? CSR_PPC64_Altivec_RegMask 287 : CSR_PPC64_RegMask) 288 : (Subtarget.hasAltivec() 289 ? CSR_SVR432_Altivec_RegMask 290 : (Subtarget.hasSPE() ? CSR_SVR432_SPE_RegMask 291 : CSR_SVR432_RegMask)); 292 } 293 294 const uint32_t* 295 PPCRegisterInfo::getNoPreservedMask() const { 296 return CSR_NoRegs_RegMask; 297 } 298 299 void PPCRegisterInfo::adjustStackMapLiveOutMask(uint32_t *Mask) const { 300 for (unsigned PseudoReg : {PPC::ZERO, PPC::ZERO8, PPC::RM}) 301 Mask[PseudoReg / 32] &= ~(1u << (PseudoReg % 32)); 302 } 303 304 BitVector PPCRegisterInfo::getReservedRegs(const MachineFunction &MF) const { 305 BitVector Reserved(getNumRegs()); 306 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 307 const PPCFrameLowering *TFI = getFrameLowering(MF); 308 309 // The ZERO register is not really a register, but the representation of r0 310 // when used in instructions that treat r0 as the constant 0. 311 markSuperRegs(Reserved, PPC::ZERO); 312 313 // The FP register is also not really a register, but is the representation 314 // of the frame pointer register used by ISD::FRAMEADDR. 315 markSuperRegs(Reserved, PPC::FP); 316 317 // The BP register is also not really a register, but is the representation 318 // of the base pointer register used by setjmp. 319 markSuperRegs(Reserved, PPC::BP); 320 321 // The counter registers must be reserved so that counter-based loops can 322 // be correctly formed (and the mtctr instructions are not DCE'd). 323 markSuperRegs(Reserved, PPC::CTR); 324 markSuperRegs(Reserved, PPC::CTR8); 325 326 markSuperRegs(Reserved, PPC::R1); 327 markSuperRegs(Reserved, PPC::LR); 328 markSuperRegs(Reserved, PPC::LR8); 329 markSuperRegs(Reserved, PPC::RM); 330 331 markSuperRegs(Reserved, PPC::VRSAVE); 332 333 // The SVR4 ABI reserves r2 and r13 334 if (Subtarget.isSVR4ABI()) { 335 // We only reserve r2 if we need to use the TOC pointer. If we have no 336 // explicit uses of the TOC pointer (meaning we're a leaf function with 337 // no constant-pool loads, etc.) and we have no potential uses inside an 338 // inline asm block, then we can treat r2 has an ordinary callee-saved 339 // register. 340 const PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 341 if (!TM.isPPC64() || FuncInfo->usesTOCBasePtr() || MF.hasInlineAsm()) 342 markSuperRegs(Reserved, PPC::R2); // System-reserved register 343 markSuperRegs(Reserved, PPC::R13); // Small Data Area pointer register 344 } 345 346 // Always reserve r2 on AIX for now. 347 // TODO: Make r2 allocatable on AIX/XCOFF for some leaf functions. 348 if (Subtarget.isAIXABI()) 349 markSuperRegs(Reserved, PPC::R2); // System-reserved register 350 351 // On PPC64, r13 is the thread pointer. Never allocate this register. 352 if (TM.isPPC64()) 353 markSuperRegs(Reserved, PPC::R13); 354 355 if (TFI->needsFP(MF)) 356 markSuperRegs(Reserved, PPC::R31); 357 358 bool IsPositionIndependent = TM.isPositionIndependent(); 359 if (hasBasePointer(MF)) { 360 if (Subtarget.is32BitELFABI() && IsPositionIndependent) 361 markSuperRegs(Reserved, PPC::R29); 362 else 363 markSuperRegs(Reserved, PPC::R30); 364 } 365 366 if (Subtarget.is32BitELFABI() && IsPositionIndependent) 367 markSuperRegs(Reserved, PPC::R30); 368 369 // Reserve Altivec registers when Altivec is unavailable. 370 if (!Subtarget.hasAltivec()) 371 for (TargetRegisterClass::iterator I = PPC::VRRCRegClass.begin(), 372 IE = PPC::VRRCRegClass.end(); I != IE; ++I) 373 markSuperRegs(Reserved, *I); 374 375 if (Subtarget.isAIXABI() && Subtarget.hasAltivec() && 376 !TM.getAIXExtendedAltivecABI()) { 377 // In the AIX default Altivec ABI, vector registers VR20-VR31 are reserved 378 // and cannot be used. 379 for (auto Reg : CSR_Altivec_SaveList) { 380 if (Reg == 0) 381 break; 382 markSuperRegs(Reserved, Reg); 383 for (MCRegAliasIterator AS(Reg, this, true); AS.isValid(); ++AS) { 384 Reserved.set(*AS); 385 } 386 } 387 } 388 389 assert(checkAllSuperRegsMarked(Reserved)); 390 return Reserved; 391 } 392 393 bool PPCRegisterInfo::requiresFrameIndexScavenging(const MachineFunction &MF) const { 394 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 395 const PPCInstrInfo *InstrInfo = Subtarget.getInstrInfo(); 396 const MachineFrameInfo &MFI = MF.getFrameInfo(); 397 const std::vector<CalleeSavedInfo> &Info = MFI.getCalleeSavedInfo(); 398 399 LLVM_DEBUG(dbgs() << "requiresFrameIndexScavenging for " << MF.getName() 400 << ".\n"); 401 // If the callee saved info is invalid we have to default to true for safety. 402 if (!MFI.isCalleeSavedInfoValid()) { 403 LLVM_DEBUG(dbgs() << "TRUE - Invalid callee saved info.\n"); 404 return true; 405 } 406 407 // We will require the use of X-Forms because the frame is larger than what 408 // can be represented in signed 16 bits that fit in the immediate of a D-Form. 409 // If we need an X-Form then we need a register to store the address offset. 410 unsigned FrameSize = MFI.getStackSize(); 411 // Signed 16 bits means that the FrameSize cannot be more than 15 bits. 412 if (FrameSize & ~0x7FFF) { 413 LLVM_DEBUG(dbgs() << "TRUE - Frame size is too large for D-Form.\n"); 414 return true; 415 } 416 417 // The callee saved info is valid so it can be traversed. 418 // Checking for registers that need saving that do not have load or store 419 // forms where the address offset is an immediate. 420 for (unsigned i = 0; i < Info.size(); i++) { 421 // If the spill is to a register no scavenging is required. 422 if (Info[i].isSpilledToReg()) 423 continue; 424 425 int FrIdx = Info[i].getFrameIdx(); 426 Register Reg = Info[i].getReg(); 427 428 const TargetRegisterClass *RC = getMinimalPhysRegClass(Reg); 429 unsigned Opcode = InstrInfo->getStoreOpcodeForSpill(RC); 430 if (!MFI.isFixedObjectIndex(FrIdx)) { 431 // This is not a fixed object. If it requires alignment then we may still 432 // need to use the XForm. 433 if (offsetMinAlignForOpcode(Opcode) > 1) { 434 LLVM_DEBUG(dbgs() << "Memory Operand: " << InstrInfo->getName(Opcode) 435 << " for register " << printReg(Reg, this) << ".\n"); 436 LLVM_DEBUG(dbgs() << "TRUE - Not fixed frame object that requires " 437 << "alignment.\n"); 438 return true; 439 } 440 } 441 442 // This is eiher: 443 // 1) A fixed frame index object which we know are aligned so 444 // as long as we have a valid DForm/DSForm/DQForm (non XForm) we don't 445 // need to consider the alignment here. 446 // 2) A not fixed object but in that case we now know that the min required 447 // alignment is no more than 1 based on the previous check. 448 if (InstrInfo->isXFormMemOp(Opcode)) { 449 LLVM_DEBUG(dbgs() << "Memory Operand: " << InstrInfo->getName(Opcode) 450 << " for register " << printReg(Reg, this) << ".\n"); 451 LLVM_DEBUG(dbgs() << "TRUE - Memory operand is X-Form.\n"); 452 return true; 453 } 454 } 455 LLVM_DEBUG(dbgs() << "FALSE - Scavenging is not required.\n"); 456 return false; 457 } 458 459 bool PPCRegisterInfo::requiresVirtualBaseRegisters( 460 const MachineFunction &MF) const { 461 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 462 // Do not use virtual base registers when ROP protection is turned on. 463 // Virtual base registers break the layout of the local variable space and may 464 // push the ROP Hash location past the 512 byte range of the ROP store 465 // instruction. 466 return !Subtarget.hasROPProtect(); 467 } 468 469 bool PPCRegisterInfo::isCallerPreservedPhysReg(MCRegister PhysReg, 470 const MachineFunction &MF) const { 471 assert(Register::isPhysicalRegister(PhysReg)); 472 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 473 const MachineFrameInfo &MFI = MF.getFrameInfo(); 474 475 if (!Subtarget.is64BitELFABI() && !Subtarget.isAIXABI()) 476 return false; 477 if (PhysReg == Subtarget.getTOCPointerRegister()) 478 // X2/R2 is guaranteed to be preserved within a function if it is reserved. 479 // The reason it's reserved is that it's the TOC pointer (and the function 480 // uses the TOC). In functions where it isn't reserved (i.e. leaf functions 481 // with no TOC access), we can't claim that it is preserved. 482 return (getReservedRegs(MF).test(PhysReg)); 483 if (StackPtrConst && PhysReg == Subtarget.getStackPointerRegister() && 484 !MFI.hasVarSizedObjects() && !MFI.hasOpaqueSPAdjustment()) 485 // The value of the stack pointer does not change within a function after 486 // the prologue and before the epilogue if there are no dynamic allocations 487 // and no inline asm which clobbers X1/R1. 488 return true; 489 return false; 490 } 491 492 bool PPCRegisterInfo::getRegAllocationHints(Register VirtReg, 493 ArrayRef<MCPhysReg> Order, 494 SmallVectorImpl<MCPhysReg> &Hints, 495 const MachineFunction &MF, 496 const VirtRegMap *VRM, 497 const LiveRegMatrix *Matrix) const { 498 const MachineRegisterInfo *MRI = &MF.getRegInfo(); 499 500 // Call the base implementation first to set any hints based on the usual 501 // heuristics and decide what the return value should be. We want to return 502 // the same value returned by the base implementation. If the base 503 // implementation decides to return true and force the allocation then we 504 // will leave it as such. On the other hand if the base implementation 505 // decides to return false the following code will not force the allocation 506 // as we are just looking to provide a hint. 507 bool BaseImplRetVal = TargetRegisterInfo::getRegAllocationHints( 508 VirtReg, Order, Hints, MF, VRM, Matrix); 509 // We are interested in instructions that copy values to ACC/UACC. 510 // The copy into UACC will be simply a COPY to a subreg so we 511 // want to allocate the corresponding physical subreg for the source. 512 // The copy into ACC will be a BUILD_UACC so we want to allocate 513 // the same number UACC for the source. 514 for (MachineInstr &Use : MRI->reg_nodbg_instructions(VirtReg)) { 515 const MachineOperand *ResultOp = nullptr; 516 Register ResultReg; 517 switch (Use.getOpcode()) { 518 case TargetOpcode::COPY: { 519 ResultOp = &Use.getOperand(0); 520 ResultReg = ResultOp->getReg(); 521 if (Register::isVirtualRegister(ResultReg) && 522 MRI->getRegClass(ResultReg)->contains(PPC::UACC0) && 523 VRM->hasPhys(ResultReg)) { 524 Register UACCPhys = VRM->getPhys(ResultReg); 525 Register HintReg = getSubReg(UACCPhys, ResultOp->getSubReg()); 526 // Ensure that the hint is a VSRp register. 527 if (HintReg >= PPC::VSRp0 && HintReg <= PPC::VSRp31) 528 Hints.push_back(HintReg); 529 } 530 break; 531 } 532 case PPC::BUILD_UACC: { 533 ResultOp = &Use.getOperand(0); 534 ResultReg = ResultOp->getReg(); 535 if (MRI->getRegClass(ResultReg)->contains(PPC::ACC0) && 536 VRM->hasPhys(ResultReg)) { 537 Register ACCPhys = VRM->getPhys(ResultReg); 538 assert((ACCPhys >= PPC::ACC0 && ACCPhys <= PPC::ACC7) && 539 "Expecting an ACC register for BUILD_UACC."); 540 Register HintReg = PPC::UACC0 + (ACCPhys - PPC::ACC0); 541 Hints.push_back(HintReg); 542 } 543 break; 544 } 545 } 546 } 547 return BaseImplRetVal; 548 } 549 550 unsigned PPCRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC, 551 MachineFunction &MF) const { 552 const PPCFrameLowering *TFI = getFrameLowering(MF); 553 const unsigned DefaultSafety = 1; 554 555 switch (RC->getID()) { 556 default: 557 return 0; 558 case PPC::G8RC_NOX0RegClassID: 559 case PPC::GPRC_NOR0RegClassID: 560 case PPC::SPERCRegClassID: 561 case PPC::G8RCRegClassID: 562 case PPC::GPRCRegClassID: { 563 unsigned FP = TFI->hasFP(MF) ? 1 : 0; 564 return 32 - FP - DefaultSafety; 565 } 566 case PPC::F4RCRegClassID: 567 case PPC::F8RCRegClassID: 568 case PPC::VSLRCRegClassID: 569 return 32 - DefaultSafety; 570 case PPC::VFRCRegClassID: 571 case PPC::VRRCRegClassID: { 572 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 573 // Vector registers VR20-VR31 are reserved and cannot be used in the default 574 // Altivec ABI on AIX. 575 if (!TM.getAIXExtendedAltivecABI() && Subtarget.isAIXABI()) 576 return 20 - DefaultSafety; 577 } 578 return 32 - DefaultSafety; 579 case PPC::VSFRCRegClassID: 580 case PPC::VSSRCRegClassID: 581 case PPC::VSRCRegClassID: { 582 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 583 if (!TM.getAIXExtendedAltivecABI() && Subtarget.isAIXABI()) 584 // Vector registers VR20-VR31 are reserved and cannot be used in the 585 // default Altivec ABI on AIX. 586 return 52 - DefaultSafety; 587 } 588 return 64 - DefaultSafety; 589 case PPC::CRRCRegClassID: 590 return 8 - DefaultSafety; 591 } 592 } 593 594 const TargetRegisterClass * 595 PPCRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC, 596 const MachineFunction &MF) const { 597 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 598 const auto *DefaultSuperclass = 599 TargetRegisterInfo::getLargestLegalSuperClass(RC, MF); 600 if (Subtarget.hasVSX()) { 601 // With VSX, we can inflate various sub-register classes to the full VSX 602 // register set. 603 604 // For Power9 we allow the user to enable GPR to vector spills. 605 // FIXME: Currently limited to spilling GP8RC. A follow on patch will add 606 // support to spill GPRC. 607 if (TM.isELFv2ABI() || Subtarget.isAIXABI()) { 608 if (Subtarget.hasP9Vector() && EnableGPRToVecSpills && 609 RC == &PPC::G8RCRegClass) { 610 InflateGP8RC++; 611 return &PPC::SPILLTOVSRRCRegClass; 612 } 613 if (RC == &PPC::GPRCRegClass && EnableGPRToVecSpills) 614 InflateGPRC++; 615 } 616 617 for (const auto *I = RC->getSuperClasses(); *I; ++I) { 618 if (getRegSizeInBits(**I) != getRegSizeInBits(*RC)) 619 continue; 620 621 switch ((*I)->getID()) { 622 case PPC::VSSRCRegClassID: 623 return Subtarget.hasP8Vector() ? *I : DefaultSuperclass; 624 case PPC::VSFRCRegClassID: 625 case PPC::VSRCRegClassID: 626 return *I; 627 case PPC::VSRpRCRegClassID: 628 return Subtarget.pairedVectorMemops() ? *I : DefaultSuperclass; 629 case PPC::ACCRCRegClassID: 630 case PPC::UACCRCRegClassID: 631 return Subtarget.hasMMA() ? *I : DefaultSuperclass; 632 } 633 } 634 } 635 636 return DefaultSuperclass; 637 } 638 639 //===----------------------------------------------------------------------===// 640 // Stack Frame Processing methods 641 //===----------------------------------------------------------------------===// 642 643 /// lowerDynamicAlloc - Generate the code for allocating an object in the 644 /// current frame. The sequence of code will be in the general form 645 /// 646 /// addi R0, SP, \#frameSize ; get the address of the previous frame 647 /// stwxu R0, SP, Rnegsize ; add and update the SP with the negated size 648 /// addi Rnew, SP, \#maxCalFrameSize ; get the top of the allocation 649 /// 650 void PPCRegisterInfo::lowerDynamicAlloc(MachineBasicBlock::iterator II) const { 651 // Get the instruction. 652 MachineInstr &MI = *II; 653 // Get the instruction's basic block. 654 MachineBasicBlock &MBB = *MI.getParent(); 655 // Get the basic block's function. 656 MachineFunction &MF = *MBB.getParent(); 657 // Get the frame info. 658 MachineFrameInfo &MFI = MF.getFrameInfo(); 659 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 660 // Get the instruction info. 661 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 662 // Determine whether 64-bit pointers are used. 663 bool LP64 = TM.isPPC64(); 664 DebugLoc dl = MI.getDebugLoc(); 665 666 // Get the maximum call stack size. 667 unsigned maxCallFrameSize = MFI.getMaxCallFrameSize(); 668 Align MaxAlign = MFI.getMaxAlign(); 669 assert(isAligned(MaxAlign, maxCallFrameSize) && 670 "Maximum call-frame size not sufficiently aligned"); 671 (void)MaxAlign; 672 673 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 674 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 675 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 676 bool KillNegSizeReg = MI.getOperand(1).isKill(); 677 Register NegSizeReg = MI.getOperand(1).getReg(); 678 679 prepareDynamicAlloca(II, NegSizeReg, KillNegSizeReg, Reg); 680 // Grow the stack and update the stack pointer link, then determine the 681 // address of new allocated space. 682 if (LP64) { 683 BuildMI(MBB, II, dl, TII.get(PPC::STDUX), PPC::X1) 684 .addReg(Reg, RegState::Kill) 685 .addReg(PPC::X1) 686 .addReg(NegSizeReg, getKillRegState(KillNegSizeReg)); 687 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), MI.getOperand(0).getReg()) 688 .addReg(PPC::X1) 689 .addImm(maxCallFrameSize); 690 } else { 691 BuildMI(MBB, II, dl, TII.get(PPC::STWUX), PPC::R1) 692 .addReg(Reg, RegState::Kill) 693 .addReg(PPC::R1) 694 .addReg(NegSizeReg, getKillRegState(KillNegSizeReg)); 695 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), MI.getOperand(0).getReg()) 696 .addReg(PPC::R1) 697 .addImm(maxCallFrameSize); 698 } 699 700 // Discard the DYNALLOC instruction. 701 MBB.erase(II); 702 } 703 704 /// To accomplish dynamic stack allocation, we have to calculate exact size 705 /// subtracted from the stack pointer according alignment information and get 706 /// previous frame pointer. 707 void PPCRegisterInfo::prepareDynamicAlloca(MachineBasicBlock::iterator II, 708 Register &NegSizeReg, 709 bool &KillNegSizeReg, 710 Register &FramePointer) const { 711 // Get the instruction. 712 MachineInstr &MI = *II; 713 // Get the instruction's basic block. 714 MachineBasicBlock &MBB = *MI.getParent(); 715 // Get the basic block's function. 716 MachineFunction &MF = *MBB.getParent(); 717 // Get the frame info. 718 MachineFrameInfo &MFI = MF.getFrameInfo(); 719 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 720 // Get the instruction info. 721 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 722 // Determine whether 64-bit pointers are used. 723 bool LP64 = TM.isPPC64(); 724 DebugLoc dl = MI.getDebugLoc(); 725 // Get the total frame size. 726 unsigned FrameSize = MFI.getStackSize(); 727 728 // Get stack alignments. 729 const PPCFrameLowering *TFI = getFrameLowering(MF); 730 Align TargetAlign = TFI->getStackAlign(); 731 Align MaxAlign = MFI.getMaxAlign(); 732 733 // Determine the previous frame's address. If FrameSize can't be 734 // represented as 16 bits or we need special alignment, then we load the 735 // previous frame's address from 0(SP). Why not do an addis of the hi? 736 // Because R0 is our only safe tmp register and addi/addis treat R0 as zero. 737 // Constructing the constant and adding would take 3 instructions. 738 // Fortunately, a frame greater than 32K is rare. 739 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 740 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 741 742 if (MaxAlign < TargetAlign && isInt<16>(FrameSize)) { 743 if (LP64) 744 BuildMI(MBB, II, dl, TII.get(PPC::ADDI8), FramePointer) 745 .addReg(PPC::X31) 746 .addImm(FrameSize); 747 else 748 BuildMI(MBB, II, dl, TII.get(PPC::ADDI), FramePointer) 749 .addReg(PPC::R31) 750 .addImm(FrameSize); 751 } else if (LP64) { 752 BuildMI(MBB, II, dl, TII.get(PPC::LD), FramePointer) 753 .addImm(0) 754 .addReg(PPC::X1); 755 } else { 756 BuildMI(MBB, II, dl, TII.get(PPC::LWZ), FramePointer) 757 .addImm(0) 758 .addReg(PPC::R1); 759 } 760 // Determine the actual NegSizeReg according to alignment info. 761 if (LP64) { 762 if (MaxAlign > TargetAlign) { 763 unsigned UnalNegSizeReg = NegSizeReg; 764 NegSizeReg = MF.getRegInfo().createVirtualRegister(G8RC); 765 766 // Unfortunately, there is no andi, only andi., and we can't insert that 767 // here because we might clobber cr0 while it is live. 768 BuildMI(MBB, II, dl, TII.get(PPC::LI8), NegSizeReg) 769 .addImm(~(MaxAlign.value() - 1)); 770 771 unsigned NegSizeReg1 = NegSizeReg; 772 NegSizeReg = MF.getRegInfo().createVirtualRegister(G8RC); 773 BuildMI(MBB, II, dl, TII.get(PPC::AND8), NegSizeReg) 774 .addReg(UnalNegSizeReg, getKillRegState(KillNegSizeReg)) 775 .addReg(NegSizeReg1, RegState::Kill); 776 KillNegSizeReg = true; 777 } 778 } else { 779 if (MaxAlign > TargetAlign) { 780 unsigned UnalNegSizeReg = NegSizeReg; 781 NegSizeReg = MF.getRegInfo().createVirtualRegister(GPRC); 782 783 // Unfortunately, there is no andi, only andi., and we can't insert that 784 // here because we might clobber cr0 while it is live. 785 BuildMI(MBB, II, dl, TII.get(PPC::LI), NegSizeReg) 786 .addImm(~(MaxAlign.value() - 1)); 787 788 unsigned NegSizeReg1 = NegSizeReg; 789 NegSizeReg = MF.getRegInfo().createVirtualRegister(GPRC); 790 BuildMI(MBB, II, dl, TII.get(PPC::AND), NegSizeReg) 791 .addReg(UnalNegSizeReg, getKillRegState(KillNegSizeReg)) 792 .addReg(NegSizeReg1, RegState::Kill); 793 KillNegSizeReg = true; 794 } 795 } 796 } 797 798 void PPCRegisterInfo::lowerPrepareProbedAlloca( 799 MachineBasicBlock::iterator II) const { 800 MachineInstr &MI = *II; 801 // Get the instruction's basic block. 802 MachineBasicBlock &MBB = *MI.getParent(); 803 // Get the basic block's function. 804 MachineFunction &MF = *MBB.getParent(); 805 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 806 // Get the instruction info. 807 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 808 // Determine whether 64-bit pointers are used. 809 bool LP64 = TM.isPPC64(); 810 DebugLoc dl = MI.getDebugLoc(); 811 Register FramePointer = MI.getOperand(0).getReg(); 812 const Register ActualNegSizeReg = MI.getOperand(1).getReg(); 813 bool KillNegSizeReg = MI.getOperand(2).isKill(); 814 Register NegSizeReg = MI.getOperand(2).getReg(); 815 const MCInstrDesc &CopyInst = TII.get(LP64 ? PPC::OR8 : PPC::OR); 816 // RegAllocator might allocate FramePointer and NegSizeReg in the same phyreg. 817 if (FramePointer == NegSizeReg) { 818 assert(KillNegSizeReg && "FramePointer is a def and NegSizeReg is an use, " 819 "NegSizeReg should be killed"); 820 // FramePointer is clobbered earlier than the use of NegSizeReg in 821 // prepareDynamicAlloca, save NegSizeReg in ActualNegSizeReg to avoid 822 // misuse. 823 BuildMI(MBB, II, dl, CopyInst, ActualNegSizeReg) 824 .addReg(NegSizeReg) 825 .addReg(NegSizeReg); 826 NegSizeReg = ActualNegSizeReg; 827 KillNegSizeReg = false; 828 } 829 prepareDynamicAlloca(II, NegSizeReg, KillNegSizeReg, FramePointer); 830 // NegSizeReg might be updated in prepareDynamicAlloca if MaxAlign > 831 // TargetAlign. 832 if (NegSizeReg != ActualNegSizeReg) 833 BuildMI(MBB, II, dl, CopyInst, ActualNegSizeReg) 834 .addReg(NegSizeReg) 835 .addReg(NegSizeReg); 836 MBB.erase(II); 837 } 838 839 void PPCRegisterInfo::lowerDynamicAreaOffset( 840 MachineBasicBlock::iterator II) const { 841 // Get the instruction. 842 MachineInstr &MI = *II; 843 // Get the instruction's basic block. 844 MachineBasicBlock &MBB = *MI.getParent(); 845 // Get the basic block's function. 846 MachineFunction &MF = *MBB.getParent(); 847 // Get the frame info. 848 MachineFrameInfo &MFI = MF.getFrameInfo(); 849 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 850 // Get the instruction info. 851 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 852 853 unsigned maxCallFrameSize = MFI.getMaxCallFrameSize(); 854 bool is64Bit = TM.isPPC64(); 855 DebugLoc dl = MI.getDebugLoc(); 856 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LI8 : PPC::LI), 857 MI.getOperand(0).getReg()) 858 .addImm(maxCallFrameSize); 859 MBB.erase(II); 860 } 861 862 /// lowerCRSpilling - Generate the code for spilling a CR register. Instead of 863 /// reserving a whole register (R0), we scrounge for one here. This generates 864 /// code like this: 865 /// 866 /// mfcr rA ; Move the conditional register into GPR rA. 867 /// rlwinm rA, rA, SB, 0, 31 ; Shift the bits left so they are in CR0's slot. 868 /// stw rA, FI ; Store rA to the frame. 869 /// 870 void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II, 871 unsigned FrameIndex) const { 872 // Get the instruction. 873 MachineInstr &MI = *II; // ; SPILL_CR <SrcReg>, <offset> 874 // Get the instruction's basic block. 875 MachineBasicBlock &MBB = *MI.getParent(); 876 MachineFunction &MF = *MBB.getParent(); 877 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 878 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 879 DebugLoc dl = MI.getDebugLoc(); 880 881 bool LP64 = TM.isPPC64(); 882 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 883 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 884 885 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 886 Register SrcReg = MI.getOperand(0).getReg(); 887 888 // We need to store the CR in the low 4-bits of the saved value. First, issue 889 // an MFOCRF to save all of the CRBits and, if needed, kill the SrcReg. 890 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), Reg) 891 .addReg(SrcReg, getKillRegState(MI.getOperand(0).isKill())); 892 893 // If the saved register wasn't CR0, shift the bits left so that they are in 894 // CR0's slot. 895 if (SrcReg != PPC::CR0) { 896 Register Reg1 = Reg; 897 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 898 899 // rlwinm rA, rA, ShiftBits, 0, 31. 900 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 901 .addReg(Reg1, RegState::Kill) 902 .addImm(getEncodingValue(SrcReg) * 4) 903 .addImm(0) 904 .addImm(31); 905 } 906 907 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW)) 908 .addReg(Reg, RegState::Kill), 909 FrameIndex); 910 911 // Discard the pseudo instruction. 912 MBB.erase(II); 913 } 914 915 void PPCRegisterInfo::lowerCRRestore(MachineBasicBlock::iterator II, 916 unsigned FrameIndex) const { 917 // Get the instruction. 918 MachineInstr &MI = *II; // ; <DestReg> = RESTORE_CR <offset> 919 // Get the instruction's basic block. 920 MachineBasicBlock &MBB = *MI.getParent(); 921 MachineFunction &MF = *MBB.getParent(); 922 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 923 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 924 DebugLoc dl = MI.getDebugLoc(); 925 926 bool LP64 = TM.isPPC64(); 927 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 928 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 929 930 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 931 Register DestReg = MI.getOperand(0).getReg(); 932 assert(MI.definesRegister(DestReg) && 933 "RESTORE_CR does not define its destination"); 934 935 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LWZ8 : PPC::LWZ), 936 Reg), FrameIndex); 937 938 // If the reloaded register isn't CR0, shift the bits right so that they are 939 // in the right CR's slot. 940 if (DestReg != PPC::CR0) { 941 Register Reg1 = Reg; 942 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 943 944 unsigned ShiftBits = getEncodingValue(DestReg)*4; 945 // rlwinm r11, r11, 32-ShiftBits, 0, 31. 946 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 947 .addReg(Reg1, RegState::Kill).addImm(32-ShiftBits).addImm(0) 948 .addImm(31); 949 } 950 951 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MTOCRF8 : PPC::MTOCRF), DestReg) 952 .addReg(Reg, RegState::Kill); 953 954 // Discard the pseudo instruction. 955 MBB.erase(II); 956 } 957 958 void PPCRegisterInfo::lowerCRBitSpilling(MachineBasicBlock::iterator II, 959 unsigned FrameIndex) const { 960 // Get the instruction. 961 MachineInstr &MI = *II; // ; SPILL_CRBIT <SrcReg>, <offset> 962 // Get the instruction's basic block. 963 MachineBasicBlock &MBB = *MI.getParent(); 964 MachineFunction &MF = *MBB.getParent(); 965 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 966 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 967 const TargetRegisterInfo* TRI = Subtarget.getRegisterInfo(); 968 DebugLoc dl = MI.getDebugLoc(); 969 970 bool LP64 = TM.isPPC64(); 971 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 972 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 973 974 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 975 Register SrcReg = MI.getOperand(0).getReg(); 976 977 // Search up the BB to find the definition of the CR bit. 978 MachineBasicBlock::reverse_iterator Ins = MI; 979 MachineBasicBlock::reverse_iterator Rend = MBB.rend(); 980 ++Ins; 981 unsigned CRBitSpillDistance = 0; 982 bool SeenUse = false; 983 for (; Ins != Rend; ++Ins) { 984 // Definition found. 985 if (Ins->modifiesRegister(SrcReg, TRI)) 986 break; 987 // Use found. 988 if (Ins->readsRegister(SrcReg, TRI)) 989 SeenUse = true; 990 // Unable to find CR bit definition within maximum search distance. 991 if (CRBitSpillDistance == MaxCRBitSpillDist) { 992 Ins = MI; 993 break; 994 } 995 // Skip debug instructions when counting CR bit spill distance. 996 if (!Ins->isDebugInstr()) 997 CRBitSpillDistance++; 998 } 999 1000 // Unable to find the definition of the CR bit in the MBB. 1001 if (Ins == MBB.rend()) 1002 Ins = MI; 1003 1004 bool SpillsKnownBit = false; 1005 // There is no need to extract the CR bit if its value is already known. 1006 switch (Ins->getOpcode()) { 1007 case PPC::CRUNSET: 1008 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LI8 : PPC::LI), Reg) 1009 .addImm(0); 1010 SpillsKnownBit = true; 1011 break; 1012 case PPC::CRSET: 1013 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LIS8 : PPC::LIS), Reg) 1014 .addImm(-32768); 1015 SpillsKnownBit = true; 1016 break; 1017 default: 1018 // On Power10, we can use SETNBC to spill all CR bits. SETNBC will set all 1019 // bits (specifically, it produces a -1 if the CR bit is set). Ultimately, 1020 // the bit that is of importance to us is bit 32 (bit 0 of a 32-bit 1021 // register), and SETNBC will set this. 1022 if (Subtarget.isISA3_1()) { 1023 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::SETNBC8 : PPC::SETNBC), Reg) 1024 .addReg(SrcReg, RegState::Undef); 1025 break; 1026 } 1027 1028 // On Power9, we can use SETB to extract the LT bit. This only works for 1029 // the LT bit since SETB produces -1/1/0 for LT/GT/<neither>. So the value 1030 // of the bit we care about (32-bit sign bit) will be set to the value of 1031 // the LT bit (regardless of the other bits in the CR field). 1032 if (Subtarget.isISA3_0()) { 1033 if (SrcReg == PPC::CR0LT || SrcReg == PPC::CR1LT || 1034 SrcReg == PPC::CR2LT || SrcReg == PPC::CR3LT || 1035 SrcReg == PPC::CR4LT || SrcReg == PPC::CR5LT || 1036 SrcReg == PPC::CR6LT || SrcReg == PPC::CR7LT) { 1037 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::SETB8 : PPC::SETB), Reg) 1038 .addReg(getCRFromCRBit(SrcReg), RegState::Undef); 1039 break; 1040 } 1041 } 1042 1043 // We need to move the CR field that contains the CR bit we are spilling. 1044 // The super register may not be explicitly defined (i.e. it can be defined 1045 // by a CR-logical that only defines the subreg) so we state that the CR 1046 // field is undef. Also, in order to preserve the kill flag on the CR bit, 1047 // we add it as an implicit use. 1048 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), Reg) 1049 .addReg(getCRFromCRBit(SrcReg), RegState::Undef) 1050 .addReg(SrcReg, 1051 RegState::Implicit | getKillRegState(MI.getOperand(0).isKill())); 1052 1053 // If the saved register wasn't CR0LT, shift the bits left so that the bit 1054 // to store is the first one. Mask all but that bit. 1055 Register Reg1 = Reg; 1056 Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 1057 1058 // rlwinm rA, rA, ShiftBits, 0, 0. 1059 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg) 1060 .addReg(Reg1, RegState::Kill) 1061 .addImm(getEncodingValue(SrcReg)) 1062 .addImm(0).addImm(0); 1063 } 1064 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW)) 1065 .addReg(Reg, RegState::Kill), 1066 FrameIndex); 1067 1068 bool KillsCRBit = MI.killsRegister(SrcReg, TRI); 1069 // Discard the pseudo instruction. 1070 MBB.erase(II); 1071 if (SpillsKnownBit && KillsCRBit && !SeenUse) { 1072 Ins->setDesc(TII.get(PPC::UNENCODED_NOP)); 1073 Ins->RemoveOperand(0); 1074 } 1075 } 1076 1077 void PPCRegisterInfo::lowerCRBitRestore(MachineBasicBlock::iterator II, 1078 unsigned FrameIndex) const { 1079 // Get the instruction. 1080 MachineInstr &MI = *II; // ; <DestReg> = RESTORE_CRBIT <offset> 1081 // Get the instruction's basic block. 1082 MachineBasicBlock &MBB = *MI.getParent(); 1083 MachineFunction &MF = *MBB.getParent(); 1084 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1085 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1086 DebugLoc dl = MI.getDebugLoc(); 1087 1088 bool LP64 = TM.isPPC64(); 1089 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 1090 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 1091 1092 Register Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 1093 Register DestReg = MI.getOperand(0).getReg(); 1094 assert(MI.definesRegister(DestReg) && 1095 "RESTORE_CRBIT does not define its destination"); 1096 1097 addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LWZ8 : PPC::LWZ), 1098 Reg), FrameIndex); 1099 1100 BuildMI(MBB, II, dl, TII.get(TargetOpcode::IMPLICIT_DEF), DestReg); 1101 1102 Register RegO = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC); 1103 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), RegO) 1104 .addReg(getCRFromCRBit(DestReg)); 1105 1106 unsigned ShiftBits = getEncodingValue(DestReg); 1107 // rlwimi r11, r10, 32-ShiftBits, ..., ... 1108 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWIMI8 : PPC::RLWIMI), RegO) 1109 .addReg(RegO, RegState::Kill) 1110 .addReg(Reg, RegState::Kill) 1111 .addImm(ShiftBits ? 32 - ShiftBits : 0) 1112 .addImm(ShiftBits) 1113 .addImm(ShiftBits); 1114 1115 BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MTOCRF8 : PPC::MTOCRF), 1116 getCRFromCRBit(DestReg)) 1117 .addReg(RegO, RegState::Kill) 1118 // Make sure we have a use dependency all the way through this 1119 // sequence of instructions. We can't have the other bits in the CR 1120 // modified in between the mfocrf and the mtocrf. 1121 .addReg(getCRFromCRBit(DestReg), RegState::Implicit); 1122 1123 // Discard the pseudo instruction. 1124 MBB.erase(II); 1125 } 1126 1127 void PPCRegisterInfo::emitAccCopyInfo(MachineBasicBlock &MBB, 1128 MCRegister DestReg, MCRegister SrcReg) { 1129 #ifdef NDEBUG 1130 return; 1131 #else 1132 if (ReportAccMoves) { 1133 std::string Dest = PPC::ACCRCRegClass.contains(DestReg) ? "acc" : "uacc"; 1134 std::string Src = PPC::ACCRCRegClass.contains(SrcReg) ? "acc" : "uacc"; 1135 dbgs() << "Emitting copy from " << Src << " to " << Dest << ":\n"; 1136 MBB.dump(); 1137 } 1138 #endif 1139 } 1140 1141 static void emitAccSpillRestoreInfo(MachineBasicBlock &MBB, bool IsPrimed, 1142 bool IsRestore) { 1143 #ifdef NDEBUG 1144 return; 1145 #else 1146 if (ReportAccMoves) { 1147 dbgs() << "Emitting " << (IsPrimed ? "acc" : "uacc") << " register " 1148 << (IsRestore ? "restore" : "spill") << ":\n"; 1149 MBB.dump(); 1150 } 1151 #endif 1152 } 1153 1154 /// lowerACCSpilling - Generate the code for spilling the accumulator register. 1155 /// Similarly to other spills/reloads that use pseudo-ops, we do not actually 1156 /// eliminate the FrameIndex here nor compute the stack offset. We simply 1157 /// create a real instruction with an FI and rely on eliminateFrameIndex to 1158 /// handle the FI elimination. 1159 void PPCRegisterInfo::lowerACCSpilling(MachineBasicBlock::iterator II, 1160 unsigned FrameIndex) const { 1161 MachineInstr &MI = *II; // SPILL_ACC <SrcReg>, <offset> 1162 MachineBasicBlock &MBB = *MI.getParent(); 1163 MachineFunction &MF = *MBB.getParent(); 1164 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1165 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1166 DebugLoc DL = MI.getDebugLoc(); 1167 Register SrcReg = MI.getOperand(0).getReg(); 1168 bool IsKilled = MI.getOperand(0).isKill(); 1169 1170 bool IsPrimed = PPC::ACCRCRegClass.contains(SrcReg); 1171 Register Reg = 1172 PPC::VSRp0 + (SrcReg - (IsPrimed ? PPC::ACC0 : PPC::UACC0)) * 2; 1173 bool IsLittleEndian = Subtarget.isLittleEndian(); 1174 1175 emitAccSpillRestoreInfo(MBB, IsPrimed, false); 1176 1177 // De-prime the register being spilled, create two stores for the pair 1178 // subregisters accounting for endianness and then re-prime the register if 1179 // it isn't killed. This uses the Offset parameter to addFrameReference() to 1180 // adjust the offset of the store that is within the 64-byte stack slot. 1181 if (IsPrimed) 1182 BuildMI(MBB, II, DL, TII.get(PPC::XXMFACC), SrcReg).addReg(SrcReg); 1183 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STXVP)) 1184 .addReg(Reg, getKillRegState(IsKilled)), 1185 FrameIndex, IsLittleEndian ? 32 : 0); 1186 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STXVP)) 1187 .addReg(Reg + 1, getKillRegState(IsKilled)), 1188 FrameIndex, IsLittleEndian ? 0 : 32); 1189 if (IsPrimed && !IsKilled) 1190 BuildMI(MBB, II, DL, TII.get(PPC::XXMTACC), SrcReg).addReg(SrcReg); 1191 1192 // Discard the pseudo instruction. 1193 MBB.erase(II); 1194 } 1195 1196 /// lowerACCRestore - Generate the code to restore the accumulator register. 1197 void PPCRegisterInfo::lowerACCRestore(MachineBasicBlock::iterator II, 1198 unsigned FrameIndex) const { 1199 MachineInstr &MI = *II; // <DestReg> = RESTORE_ACC <offset> 1200 MachineBasicBlock &MBB = *MI.getParent(); 1201 MachineFunction &MF = *MBB.getParent(); 1202 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1203 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1204 DebugLoc DL = MI.getDebugLoc(); 1205 1206 Register DestReg = MI.getOperand(0).getReg(); 1207 assert(MI.definesRegister(DestReg) && 1208 "RESTORE_ACC does not define its destination"); 1209 1210 bool IsPrimed = PPC::ACCRCRegClass.contains(DestReg); 1211 Register Reg = 1212 PPC::VSRp0 + (DestReg - (IsPrimed ? PPC::ACC0 : PPC::UACC0)) * 2; 1213 bool IsLittleEndian = Subtarget.isLittleEndian(); 1214 1215 emitAccSpillRestoreInfo(MBB, IsPrimed, true); 1216 1217 // Create two loads for the pair subregisters accounting for endianness and 1218 // then prime the accumulator register being restored. 1219 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LXVP), Reg), 1220 FrameIndex, IsLittleEndian ? 32 : 0); 1221 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LXVP), Reg + 1), 1222 FrameIndex, IsLittleEndian ? 0 : 32); 1223 if (IsPrimed) 1224 BuildMI(MBB, II, DL, TII.get(PPC::XXMTACC), DestReg).addReg(DestReg); 1225 1226 // Discard the pseudo instruction. 1227 MBB.erase(II); 1228 } 1229 1230 /// lowerQuadwordSpilling - Generate code to spill paired general register. 1231 void PPCRegisterInfo::lowerQuadwordSpilling(MachineBasicBlock::iterator II, 1232 unsigned FrameIndex) const { 1233 MachineInstr &MI = *II; 1234 MachineBasicBlock &MBB = *MI.getParent(); 1235 MachineFunction &MF = *MBB.getParent(); 1236 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1237 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1238 DebugLoc DL = MI.getDebugLoc(); 1239 1240 Register SrcReg = MI.getOperand(0).getReg(); 1241 bool IsKilled = MI.getOperand(0).isKill(); 1242 1243 Register Reg = PPC::X0 + (SrcReg - PPC::G8p0) * 2; 1244 bool IsLittleEndian = Subtarget.isLittleEndian(); 1245 1246 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STD)) 1247 .addReg(Reg, getKillRegState(IsKilled)), 1248 FrameIndex, IsLittleEndian ? 8 : 0); 1249 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::STD)) 1250 .addReg(Reg + 1, getKillRegState(IsKilled)), 1251 FrameIndex, IsLittleEndian ? 0 : 8); 1252 1253 // Discard the pseudo instruction. 1254 MBB.erase(II); 1255 } 1256 1257 /// lowerQuadwordRestore - Generate code to restore paired general register. 1258 void PPCRegisterInfo::lowerQuadwordRestore(MachineBasicBlock::iterator II, 1259 unsigned FrameIndex) const { 1260 MachineInstr &MI = *II; 1261 MachineBasicBlock &MBB = *MI.getParent(); 1262 MachineFunction &MF = *MBB.getParent(); 1263 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1264 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1265 DebugLoc DL = MI.getDebugLoc(); 1266 1267 Register DestReg = MI.getOperand(0).getReg(); 1268 assert(MI.definesRegister(DestReg) && 1269 "RESTORE_QUADWORD does not define its destination"); 1270 1271 Register Reg = PPC::X0 + (DestReg - PPC::G8p0) * 2; 1272 bool IsLittleEndian = Subtarget.isLittleEndian(); 1273 1274 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LD), Reg), FrameIndex, 1275 IsLittleEndian ? 8 : 0); 1276 addFrameReference(BuildMI(MBB, II, DL, TII.get(PPC::LD), Reg + 1), FrameIndex, 1277 IsLittleEndian ? 0 : 8); 1278 1279 // Discard the pseudo instruction. 1280 MBB.erase(II); 1281 } 1282 1283 bool PPCRegisterInfo::hasReservedSpillSlot(const MachineFunction &MF, 1284 Register Reg, int &FrameIdx) const { 1285 // For the nonvolatile condition registers (CR2, CR3, CR4) return true to 1286 // prevent allocating an additional frame slot. 1287 // For 64-bit ELF and AIX, the CR save area is in the linkage area at SP+8, 1288 // for 32-bit AIX the CR save area is in the linkage area at SP+4. 1289 // We have created a FrameIndex to that spill slot to keep the CalleSaveInfos 1290 // valid. 1291 // For 32-bit ELF, we have previously created the stack slot if needed, so 1292 // return its FrameIdx. 1293 if (PPC::CR2 <= Reg && Reg <= PPC::CR4) { 1294 FrameIdx = MF.getInfo<PPCFunctionInfo>()->getCRSpillFrameIndex(); 1295 return true; 1296 } 1297 return false; 1298 } 1299 1300 // If the offset must be a multiple of some value, return what that value is. 1301 static unsigned offsetMinAlignForOpcode(unsigned OpC) { 1302 switch (OpC) { 1303 default: 1304 return 1; 1305 case PPC::LWA: 1306 case PPC::LWA_32: 1307 case PPC::LD: 1308 case PPC::LDU: 1309 case PPC::STD: 1310 case PPC::STDU: 1311 case PPC::DFLOADf32: 1312 case PPC::DFLOADf64: 1313 case PPC::DFSTOREf32: 1314 case PPC::DFSTOREf64: 1315 case PPC::LXSD: 1316 case PPC::LXSSP: 1317 case PPC::STXSD: 1318 case PPC::STXSSP: 1319 case PPC::STQ: 1320 return 4; 1321 case PPC::EVLDD: 1322 case PPC::EVSTDD: 1323 return 8; 1324 case PPC::LXV: 1325 case PPC::STXV: 1326 case PPC::LQ: 1327 case PPC::LXVP: 1328 case PPC::STXVP: 1329 return 16; 1330 } 1331 } 1332 1333 // If the offset must be a multiple of some value, return what that value is. 1334 static unsigned offsetMinAlign(const MachineInstr &MI) { 1335 unsigned OpC = MI.getOpcode(); 1336 return offsetMinAlignForOpcode(OpC); 1337 } 1338 1339 // Return the OffsetOperandNo given the FIOperandNum (and the instruction). 1340 static unsigned getOffsetONFromFION(const MachineInstr &MI, 1341 unsigned FIOperandNum) { 1342 // Take into account whether it's an add or mem instruction 1343 unsigned OffsetOperandNo = (FIOperandNum == 2) ? 1 : 2; 1344 if (MI.isInlineAsm()) 1345 OffsetOperandNo = FIOperandNum - 1; 1346 else if (MI.getOpcode() == TargetOpcode::STACKMAP || 1347 MI.getOpcode() == TargetOpcode::PATCHPOINT) 1348 OffsetOperandNo = FIOperandNum + 1; 1349 1350 return OffsetOperandNo; 1351 } 1352 1353 void 1354 PPCRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 1355 int SPAdj, unsigned FIOperandNum, 1356 RegScavenger *RS) const { 1357 assert(SPAdj == 0 && "Unexpected"); 1358 1359 // Get the instruction. 1360 MachineInstr &MI = *II; 1361 // Get the instruction's basic block. 1362 MachineBasicBlock &MBB = *MI.getParent(); 1363 // Get the basic block's function. 1364 MachineFunction &MF = *MBB.getParent(); 1365 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1366 // Get the instruction info. 1367 const PPCInstrInfo &TII = *Subtarget.getInstrInfo(); 1368 // Get the frame info. 1369 MachineFrameInfo &MFI = MF.getFrameInfo(); 1370 DebugLoc dl = MI.getDebugLoc(); 1371 1372 unsigned OffsetOperandNo = getOffsetONFromFION(MI, FIOperandNum); 1373 1374 // Get the frame index. 1375 int FrameIndex = MI.getOperand(FIOperandNum).getIndex(); 1376 1377 // Get the frame pointer save index. Users of this index are primarily 1378 // DYNALLOC instructions. 1379 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); 1380 int FPSI = FI->getFramePointerSaveIndex(); 1381 // Get the instruction opcode. 1382 unsigned OpC = MI.getOpcode(); 1383 1384 if ((OpC == PPC::DYNAREAOFFSET || OpC == PPC::DYNAREAOFFSET8)) { 1385 lowerDynamicAreaOffset(II); 1386 return; 1387 } 1388 1389 // Special case for dynamic alloca. 1390 if (FPSI && FrameIndex == FPSI && 1391 (OpC == PPC::DYNALLOC || OpC == PPC::DYNALLOC8)) { 1392 lowerDynamicAlloc(II); 1393 return; 1394 } 1395 1396 if (FPSI && FrameIndex == FPSI && 1397 (OpC == PPC::PREPARE_PROBED_ALLOCA_64 || 1398 OpC == PPC::PREPARE_PROBED_ALLOCA_32 || 1399 OpC == PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_64 || 1400 OpC == PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_32)) { 1401 lowerPrepareProbedAlloca(II); 1402 return; 1403 } 1404 1405 // Special case for pseudo-ops SPILL_CR and RESTORE_CR, etc. 1406 if (OpC == PPC::SPILL_CR) { 1407 lowerCRSpilling(II, FrameIndex); 1408 return; 1409 } else if (OpC == PPC::RESTORE_CR) { 1410 lowerCRRestore(II, FrameIndex); 1411 return; 1412 } else if (OpC == PPC::SPILL_CRBIT) { 1413 lowerCRBitSpilling(II, FrameIndex); 1414 return; 1415 } else if (OpC == PPC::RESTORE_CRBIT) { 1416 lowerCRBitRestore(II, FrameIndex); 1417 return; 1418 } else if (OpC == PPC::SPILL_ACC || OpC == PPC::SPILL_UACC) { 1419 lowerACCSpilling(II, FrameIndex); 1420 return; 1421 } else if (OpC == PPC::RESTORE_ACC || OpC == PPC::RESTORE_UACC) { 1422 lowerACCRestore(II, FrameIndex); 1423 return; 1424 } else if (OpC == PPC::SPILL_QUADWORD) { 1425 lowerQuadwordSpilling(II, FrameIndex); 1426 return; 1427 } else if (OpC == PPC::RESTORE_QUADWORD) { 1428 lowerQuadwordRestore(II, FrameIndex); 1429 return; 1430 } 1431 1432 // Replace the FrameIndex with base register with GPR1 (SP) or GPR31 (FP). 1433 MI.getOperand(FIOperandNum).ChangeToRegister( 1434 FrameIndex < 0 ? getBaseRegister(MF) : getFrameRegister(MF), false); 1435 1436 // If the instruction is not present in ImmToIdxMap, then it has no immediate 1437 // form (and must be r+r). 1438 bool noImmForm = !MI.isInlineAsm() && OpC != TargetOpcode::STACKMAP && 1439 OpC != TargetOpcode::PATCHPOINT && !ImmToIdxMap.count(OpC); 1440 1441 // Now add the frame object offset to the offset from r1. 1442 int Offset = MFI.getObjectOffset(FrameIndex); 1443 Offset += MI.getOperand(OffsetOperandNo).getImm(); 1444 1445 // If we're not using a Frame Pointer that has been set to the value of the 1446 // SP before having the stack size subtracted from it, then add the stack size 1447 // to Offset to get the correct offset. 1448 // Naked functions have stack size 0, although getStackSize may not reflect 1449 // that because we didn't call all the pieces that compute it for naked 1450 // functions. 1451 if (!MF.getFunction().hasFnAttribute(Attribute::Naked)) { 1452 if (!(hasBasePointer(MF) && FrameIndex < 0)) 1453 Offset += MFI.getStackSize(); 1454 } 1455 1456 // If we encounter an LXVP/STXVP with an offset that doesn't fit, we can 1457 // transform it to the prefixed version so we don't have to use the XForm. 1458 if ((OpC == PPC::LXVP || OpC == PPC::STXVP) && 1459 (!isInt<16>(Offset) || (Offset % offsetMinAlign(MI)) != 0) && 1460 Subtarget.hasPrefixInstrs()) { 1461 unsigned NewOpc = OpC == PPC::LXVP ? PPC::PLXVP : PPC::PSTXVP; 1462 MI.setDesc(TII.get(NewOpc)); 1463 OpC = NewOpc; 1464 } 1465 1466 // If we can, encode the offset directly into the instruction. If this is a 1467 // normal PPC "ri" instruction, any 16-bit value can be safely encoded. If 1468 // this is a PPC64 "ix" instruction, only a 16-bit value with the low two bits 1469 // clear can be encoded. This is extremely uncommon, because normally you 1470 // only "std" to a stack slot that is at least 4-byte aligned, but it can 1471 // happen in invalid code. 1472 assert(OpC != PPC::DBG_VALUE && 1473 "This should be handled in a target-independent way"); 1474 // FIXME: This should be factored out to a separate function as prefixed 1475 // instructions add a number of opcodes for which we can use 34-bit imm. 1476 bool OffsetFitsMnemonic = (OpC == PPC::EVSTDD || OpC == PPC::EVLDD) ? 1477 isUInt<8>(Offset) : 1478 isInt<16>(Offset); 1479 if (TII.isPrefixed(MI.getOpcode())) 1480 OffsetFitsMnemonic = isInt<34>(Offset); 1481 if (!noImmForm && ((OffsetFitsMnemonic && 1482 ((Offset % offsetMinAlign(MI)) == 0)) || 1483 OpC == TargetOpcode::STACKMAP || 1484 OpC == TargetOpcode::PATCHPOINT)) { 1485 MI.getOperand(OffsetOperandNo).ChangeToImmediate(Offset); 1486 return; 1487 } 1488 1489 // The offset doesn't fit into a single register, scavenge one to build the 1490 // offset in. 1491 1492 bool is64Bit = TM.isPPC64(); 1493 const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; 1494 const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; 1495 const TargetRegisterClass *RC = is64Bit ? G8RC : GPRC; 1496 Register SRegHi = MF.getRegInfo().createVirtualRegister(RC), 1497 SReg = MF.getRegInfo().createVirtualRegister(RC); 1498 1499 // Insert a set of rA with the full offset value before the ld, st, or add 1500 if (isInt<16>(Offset)) 1501 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LI8 : PPC::LI), SReg) 1502 .addImm(Offset); 1503 else { 1504 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::LIS8 : PPC::LIS), SRegHi) 1505 .addImm(Offset >> 16); 1506 BuildMI(MBB, II, dl, TII.get(is64Bit ? PPC::ORI8 : PPC::ORI), SReg) 1507 .addReg(SRegHi, RegState::Kill) 1508 .addImm(Offset); 1509 } 1510 1511 // Convert into indexed form of the instruction: 1512 // 1513 // sth 0:rA, 1:imm 2:(rB) ==> sthx 0:rA, 2:rB, 1:r0 1514 // addi 0:rA 1:rB, 2, imm ==> add 0:rA, 1:rB, 2:r0 1515 unsigned OperandBase; 1516 1517 if (noImmForm) 1518 OperandBase = 1; 1519 else if (OpC != TargetOpcode::INLINEASM && 1520 OpC != TargetOpcode::INLINEASM_BR) { 1521 assert(ImmToIdxMap.count(OpC) && 1522 "No indexed form of load or store available!"); 1523 unsigned NewOpcode = ImmToIdxMap.find(OpC)->second; 1524 MI.setDesc(TII.get(NewOpcode)); 1525 OperandBase = 1; 1526 } else { 1527 OperandBase = OffsetOperandNo; 1528 } 1529 1530 Register StackReg = MI.getOperand(FIOperandNum).getReg(); 1531 MI.getOperand(OperandBase).ChangeToRegister(StackReg, false); 1532 MI.getOperand(OperandBase + 1).ChangeToRegister(SReg, false, false, true); 1533 } 1534 1535 Register PPCRegisterInfo::getFrameRegister(const MachineFunction &MF) const { 1536 const PPCFrameLowering *TFI = getFrameLowering(MF); 1537 1538 if (!TM.isPPC64()) 1539 return TFI->hasFP(MF) ? PPC::R31 : PPC::R1; 1540 else 1541 return TFI->hasFP(MF) ? PPC::X31 : PPC::X1; 1542 } 1543 1544 Register PPCRegisterInfo::getBaseRegister(const MachineFunction &MF) const { 1545 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1546 if (!hasBasePointer(MF)) 1547 return getFrameRegister(MF); 1548 1549 if (TM.isPPC64()) 1550 return PPC::X30; 1551 1552 if (Subtarget.isSVR4ABI() && TM.isPositionIndependent()) 1553 return PPC::R29; 1554 1555 return PPC::R30; 1556 } 1557 1558 bool PPCRegisterInfo::hasBasePointer(const MachineFunction &MF) const { 1559 if (!EnableBasePointer) 1560 return false; 1561 if (AlwaysBasePointer) 1562 return true; 1563 1564 // If we need to realign the stack, then the stack pointer can no longer 1565 // serve as an offset into the caller's stack space. As a result, we need a 1566 // base pointer. 1567 return hasStackRealignment(MF); 1568 } 1569 1570 /// Returns true if the instruction's frame index 1571 /// reference would be better served by a base register other than FP 1572 /// or SP. Used by LocalStackFrameAllocation to determine which frame index 1573 /// references it should create new base registers for. 1574 bool PPCRegisterInfo:: 1575 needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const { 1576 assert(Offset < 0 && "Local offset must be negative"); 1577 1578 // It's the load/store FI references that cause issues, as it can be difficult 1579 // to materialize the offset if it won't fit in the literal field. Estimate 1580 // based on the size of the local frame and some conservative assumptions 1581 // about the rest of the stack frame (note, this is pre-regalloc, so 1582 // we don't know everything for certain yet) whether this offset is likely 1583 // to be out of range of the immediate. Return true if so. 1584 1585 // We only generate virtual base registers for loads and stores that have 1586 // an r+i form. Return false for everything else. 1587 unsigned OpC = MI->getOpcode(); 1588 if (!ImmToIdxMap.count(OpC)) 1589 return false; 1590 1591 // Don't generate a new virtual base register just to add zero to it. 1592 if ((OpC == PPC::ADDI || OpC == PPC::ADDI8) && 1593 MI->getOperand(2).getImm() == 0) 1594 return false; 1595 1596 MachineBasicBlock &MBB = *MI->getParent(); 1597 MachineFunction &MF = *MBB.getParent(); 1598 const PPCFrameLowering *TFI = getFrameLowering(MF); 1599 unsigned StackEst = TFI->determineFrameLayout(MF, true); 1600 1601 // If we likely don't need a stack frame, then we probably don't need a 1602 // virtual base register either. 1603 if (!StackEst) 1604 return false; 1605 1606 // Estimate an offset from the stack pointer. 1607 // The incoming offset is relating to the SP at the start of the function, 1608 // but when we access the local it'll be relative to the SP after local 1609 // allocation, so adjust our SP-relative offset by that allocation size. 1610 Offset += StackEst; 1611 1612 // The frame pointer will point to the end of the stack, so estimate the 1613 // offset as the difference between the object offset and the FP location. 1614 return !isFrameOffsetLegal(MI, getBaseRegister(MF), Offset); 1615 } 1616 1617 /// Insert defining instruction(s) for BaseReg to 1618 /// be a pointer to FrameIdx at the beginning of the basic block. 1619 Register PPCRegisterInfo::materializeFrameBaseRegister(MachineBasicBlock *MBB, 1620 int FrameIdx, 1621 int64_t Offset) const { 1622 unsigned ADDriOpc = TM.isPPC64() ? PPC::ADDI8 : PPC::ADDI; 1623 1624 MachineBasicBlock::iterator Ins = MBB->begin(); 1625 DebugLoc DL; // Defaults to "unknown" 1626 if (Ins != MBB->end()) 1627 DL = Ins->getDebugLoc(); 1628 1629 const MachineFunction &MF = *MBB->getParent(); 1630 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1631 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1632 const MCInstrDesc &MCID = TII.get(ADDriOpc); 1633 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 1634 const TargetRegisterClass *RC = getPointerRegClass(MF); 1635 Register BaseReg = MRI.createVirtualRegister(RC); 1636 MRI.constrainRegClass(BaseReg, TII.getRegClass(MCID, 0, this, MF)); 1637 1638 BuildMI(*MBB, Ins, DL, MCID, BaseReg) 1639 .addFrameIndex(FrameIdx).addImm(Offset); 1640 1641 return BaseReg; 1642 } 1643 1644 void PPCRegisterInfo::resolveFrameIndex(MachineInstr &MI, Register BaseReg, 1645 int64_t Offset) const { 1646 unsigned FIOperandNum = 0; 1647 while (!MI.getOperand(FIOperandNum).isFI()) { 1648 ++FIOperandNum; 1649 assert(FIOperandNum < MI.getNumOperands() && 1650 "Instr doesn't have FrameIndex operand!"); 1651 } 1652 1653 MI.getOperand(FIOperandNum).ChangeToRegister(BaseReg, false); 1654 unsigned OffsetOperandNo = getOffsetONFromFION(MI, FIOperandNum); 1655 Offset += MI.getOperand(OffsetOperandNo).getImm(); 1656 MI.getOperand(OffsetOperandNo).ChangeToImmediate(Offset); 1657 1658 MachineBasicBlock &MBB = *MI.getParent(); 1659 MachineFunction &MF = *MBB.getParent(); 1660 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); 1661 const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); 1662 const MCInstrDesc &MCID = MI.getDesc(); 1663 MachineRegisterInfo &MRI = MF.getRegInfo(); 1664 MRI.constrainRegClass(BaseReg, 1665 TII.getRegClass(MCID, FIOperandNum, this, MF)); 1666 } 1667 1668 bool PPCRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, 1669 Register BaseReg, 1670 int64_t Offset) const { 1671 unsigned FIOperandNum = 0; 1672 while (!MI->getOperand(FIOperandNum).isFI()) { 1673 ++FIOperandNum; 1674 assert(FIOperandNum < MI->getNumOperands() && 1675 "Instr doesn't have FrameIndex operand!"); 1676 } 1677 1678 unsigned OffsetOperandNo = getOffsetONFromFION(*MI, FIOperandNum); 1679 Offset += MI->getOperand(OffsetOperandNo).getImm(); 1680 1681 return MI->getOpcode() == PPC::DBG_VALUE || // DBG_VALUE is always Reg+Imm 1682 MI->getOpcode() == TargetOpcode::STACKMAP || 1683 MI->getOpcode() == TargetOpcode::PATCHPOINT || 1684 (isInt<16>(Offset) && (Offset % offsetMinAlign(*MI)) == 0); 1685 } 1686