1//===-- SystemZFrameLowering.cpp - Frame lowering for SystemZ -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9 10#include "SystemZFrameLowering.h" 11#include "SystemZCallingConv.h" 12#include "SystemZInstrBuilder.h" 13#include "SystemZInstrInfo.h" 14#include "SystemZMachineFunctionInfo.h" 15#include "SystemZRegisterInfo.h" 16#include "SystemZSubtarget.h" 17#include "llvm/CodeGen/MachineModuleInfo.h" 18#include "llvm/CodeGen/MachineRegisterInfo.h" 19#include "llvm/CodeGen/RegisterScavenging.h" 20#include "llvm/IR/Function.h" 21 22using namespace llvm; 23 24namespace { 25// The ABI-defined register save slots, relative to the incoming stack 26// pointer. 27static const TargetFrameLowering::SpillSlot SpillOffsetTable[] = { 28 { SystemZ::R2D, 0x10 }, 29 { SystemZ::R3D, 0x18 }, 30 { SystemZ::R4D, 0x20 }, 31 { SystemZ::R5D, 0x28 }, 32 { SystemZ::R6D, 0x30 }, 33 { SystemZ::R7D, 0x38 }, 34 { SystemZ::R8D, 0x40 }, 35 { SystemZ::R9D, 0x48 }, 36 { SystemZ::R10D, 0x50 }, 37 { SystemZ::R11D, 0x58 }, 38 { SystemZ::R12D, 0x60 }, 39 { SystemZ::R13D, 0x68 }, 40 { SystemZ::R14D, 0x70 }, 41 { SystemZ::R15D, 0x78 }, 42 { SystemZ::F0D, 0x80 }, 43 { SystemZ::F2D, 0x88 }, 44 { SystemZ::F4D, 0x90 }, 45 { SystemZ::F6D, 0x98 } 46}; 47} // end anonymous namespace 48 49SystemZFrameLowering::SystemZFrameLowering() 50 : TargetFrameLowering(TargetFrameLowering::StackGrowsDown, 8, 51 -SystemZMC::CallFrameSize, 8, 52 false /* StackRealignable */) { 53 // Create a mapping from register number to save slot offset. 54 RegSpillOffsets.grow(SystemZ::NUM_TARGET_REGS); 55 for (unsigned I = 0, E = array_lengthof(SpillOffsetTable); I != E; ++I) 56 RegSpillOffsets[SpillOffsetTable[I].Reg] = SpillOffsetTable[I].Offset; 57} 58 59const TargetFrameLowering::SpillSlot * 60SystemZFrameLowering::getCalleeSavedSpillSlots(unsigned &NumEntries) const { 61 NumEntries = array_lengthof(SpillOffsetTable); 62 return SpillOffsetTable; 63} 64 65void SystemZFrameLowering::determineCalleeSaves(MachineFunction &MF, 66 BitVector &SavedRegs, 67 RegScavenger *RS) const { 68 TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS); 69 70 MachineFrameInfo *MFFrame = MF.getFrameInfo(); 71 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 72 bool HasFP = hasFP(MF); 73 SystemZMachineFunctionInfo *MFI = MF.getInfo<SystemZMachineFunctionInfo>(); 74 bool IsVarArg = MF.getFunction()->isVarArg(); 75 76 // va_start stores incoming FPR varargs in the normal way, but delegates 77 // the saving of incoming GPR varargs to spillCalleeSavedRegisters(). 78 // Record these pending uses, which typically include the call-saved 79 // argument register R6D. 80 if (IsVarArg) 81 for (unsigned I = MFI->getVarArgsFirstGPR(); I < SystemZ::NumArgGPRs; ++I) 82 SavedRegs.set(SystemZ::ArgGPRs[I]); 83 84 // If the function requires a frame pointer, record that the hard 85 // frame pointer will be clobbered. 86 if (HasFP) 87 SavedRegs.set(SystemZ::R11D); 88 89 // If the function calls other functions, record that the return 90 // address register will be clobbered. 91 if (MFFrame->hasCalls()) 92 SavedRegs.set(SystemZ::R14D); 93 94 // If we are saving GPRs other than the stack pointer, we might as well 95 // save and restore the stack pointer at the same time, via STMG and LMG. 96 // This allows the deallocation to be done by the LMG, rather than needing 97 // a separate %r15 addition. 98 const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(&MF); 99 for (unsigned I = 0; CSRegs[I]; ++I) { 100 unsigned Reg = CSRegs[I]; 101 if (SystemZ::GR64BitRegClass.contains(Reg) && SavedRegs.test(Reg)) { 102 SavedRegs.set(SystemZ::R15D); 103 break; 104 } 105 } 106} 107 108// Add GPR64 to the save instruction being built by MIB, which is in basic 109// block MBB. IsImplicit says whether this is an explicit operand to the 110// instruction, or an implicit one that comes between the explicit start 111// and end registers. 112static void addSavedGPR(MachineBasicBlock &MBB, MachineInstrBuilder &MIB, 113 unsigned GPR64, bool IsImplicit) { 114 const TargetRegisterInfo *RI = 115 MBB.getParent()->getSubtarget().getRegisterInfo(); 116 unsigned GPR32 = RI->getSubReg(GPR64, SystemZ::subreg_l32); 117 bool IsLive = MBB.isLiveIn(GPR64) || MBB.isLiveIn(GPR32); 118 if (!IsLive || !IsImplicit) { 119 MIB.addReg(GPR64, getImplRegState(IsImplicit) | getKillRegState(!IsLive)); 120 if (!IsLive) 121 MBB.addLiveIn(GPR64); 122 } 123} 124 125bool SystemZFrameLowering:: 126spillCalleeSavedRegisters(MachineBasicBlock &MBB, 127 MachineBasicBlock::iterator MBBI, 128 const std::vector<CalleeSavedInfo> &CSI, 129 const TargetRegisterInfo *TRI) const { 130 if (CSI.empty()) 131 return false; 132 133 MachineFunction &MF = *MBB.getParent(); 134 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 135 SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); 136 bool IsVarArg = MF.getFunction()->isVarArg(); 137 DebugLoc DL; 138 139 // Scan the call-saved GPRs and find the bounds of the register spill area. 140 unsigned LowGPR = 0; 141 unsigned HighGPR = SystemZ::R15D; 142 unsigned StartOffset = -1U; 143 for (unsigned I = 0, E = CSI.size(); I != E; ++I) { 144 unsigned Reg = CSI[I].getReg(); 145 if (SystemZ::GR64BitRegClass.contains(Reg)) { 146 unsigned Offset = RegSpillOffsets[Reg]; 147 assert(Offset && "Unexpected GPR save"); 148 if (StartOffset > Offset) { 149 LowGPR = Reg; 150 StartOffset = Offset; 151 } 152 } 153 } 154 155 // Save the range of call-saved registers, for use by the epilogue inserter. 156 ZFI->setLowSavedGPR(LowGPR); 157 ZFI->setHighSavedGPR(HighGPR); 158 159 // Include the GPR varargs, if any. R6D is call-saved, so would 160 // be included by the loop above, but we also need to handle the 161 // call-clobbered argument registers. 162 if (IsVarArg) { 163 unsigned FirstGPR = ZFI->getVarArgsFirstGPR(); 164 if (FirstGPR < SystemZ::NumArgGPRs) { 165 unsigned Reg = SystemZ::ArgGPRs[FirstGPR]; 166 unsigned Offset = RegSpillOffsets[Reg]; 167 if (StartOffset > Offset) { 168 LowGPR = Reg; StartOffset = Offset; 169 } 170 } 171 } 172 173 // Save GPRs 174 if (LowGPR) { 175 assert(LowGPR != HighGPR && "Should be saving %r15 and something else"); 176 177 // Build an STMG instruction. 178 MachineInstrBuilder MIB = BuildMI(MBB, MBBI, DL, TII->get(SystemZ::STMG)); 179 180 // Add the explicit register operands. 181 addSavedGPR(MBB, MIB, LowGPR, false); 182 addSavedGPR(MBB, MIB, HighGPR, false); 183 184 // Add the address. 185 MIB.addReg(SystemZ::R15D).addImm(StartOffset); 186 187 // Make sure all call-saved GPRs are included as operands and are 188 // marked as live on entry. 189 for (unsigned I = 0, E = CSI.size(); I != E; ++I) { 190 unsigned Reg = CSI[I].getReg(); 191 if (SystemZ::GR64BitRegClass.contains(Reg)) 192 addSavedGPR(MBB, MIB, Reg, true); 193 } 194 195 // ...likewise GPR varargs. 196 if (IsVarArg) 197 for (unsigned I = ZFI->getVarArgsFirstGPR(); I < SystemZ::NumArgGPRs; ++I) 198 addSavedGPR(MBB, MIB, SystemZ::ArgGPRs[I], true); 199 } 200 201 // Save FPRs in the normal TargetInstrInfo way. 202 for (unsigned I = 0, E = CSI.size(); I != E; ++I) { 203 unsigned Reg = CSI[I].getReg(); 204 if (SystemZ::FP64BitRegClass.contains(Reg)) { 205 MBB.addLiveIn(Reg); 206 TII->storeRegToStackSlot(MBB, MBBI, Reg, true, CSI[I].getFrameIdx(), 207 &SystemZ::FP64BitRegClass, TRI); 208 } 209 } 210 211 return true; 212} 213 214bool SystemZFrameLowering:: 215restoreCalleeSavedRegisters(MachineBasicBlock &MBB, 216 MachineBasicBlock::iterator MBBI, 217 const std::vector<CalleeSavedInfo> &CSI, 218 const TargetRegisterInfo *TRI) const { 219 if (CSI.empty()) 220 return false; 221 222 MachineFunction &MF = *MBB.getParent(); 223 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 224 SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); 225 bool HasFP = hasFP(MF); 226 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 227 228 // Restore FPRs in the normal TargetInstrInfo way. 229 for (unsigned I = 0, E = CSI.size(); I != E; ++I) { 230 unsigned Reg = CSI[I].getReg(); 231 if (SystemZ::FP64BitRegClass.contains(Reg)) 232 TII->loadRegFromStackSlot(MBB, MBBI, Reg, CSI[I].getFrameIdx(), 233 &SystemZ::FP64BitRegClass, TRI); 234 } 235 236 // Restore call-saved GPRs (but not call-clobbered varargs, which at 237 // this point might hold return values). 238 unsigned LowGPR = ZFI->getLowSavedGPR(); 239 unsigned HighGPR = ZFI->getHighSavedGPR(); 240 unsigned StartOffset = RegSpillOffsets[LowGPR]; 241 if (LowGPR) { 242 // If we saved any of %r2-%r5 as varargs, we should also be saving 243 // and restoring %r6. If we're saving %r6 or above, we should be 244 // restoring it too. 245 assert(LowGPR != HighGPR && "Should be loading %r15 and something else"); 246 247 // Build an LMG instruction. 248 MachineInstrBuilder MIB = BuildMI(MBB, MBBI, DL, TII->get(SystemZ::LMG)); 249 250 // Add the explicit register operands. 251 MIB.addReg(LowGPR, RegState::Define); 252 MIB.addReg(HighGPR, RegState::Define); 253 254 // Add the address. 255 MIB.addReg(HasFP ? SystemZ::R11D : SystemZ::R15D); 256 MIB.addImm(StartOffset); 257 258 // Do a second scan adding regs as being defined by instruction 259 for (unsigned I = 0, E = CSI.size(); I != E; ++I) { 260 unsigned Reg = CSI[I].getReg(); 261 if (Reg != LowGPR && Reg != HighGPR) 262 MIB.addReg(Reg, RegState::ImplicitDefine); 263 } 264 } 265 266 return true; 267} 268 269void SystemZFrameLowering:: 270processFunctionBeforeFrameFinalized(MachineFunction &MF, 271 RegScavenger *RS) const { 272 MachineFrameInfo *MFFrame = MF.getFrameInfo(); 273 uint64_t MaxReach = (MFFrame->estimateStackSize(MF) + 274 SystemZMC::CallFrameSize * 2); 275 if (!isUInt<12>(MaxReach)) { 276 // We may need register scavenging slots if some parts of the frame 277 // are outside the reach of an unsigned 12-bit displacement. 278 // Create 2 for the case where both addresses in an MVC are 279 // out of range. 280 RS->addScavengingFrameIndex(MFFrame->CreateStackObject(8, 8, false)); 281 RS->addScavengingFrameIndex(MFFrame->CreateStackObject(8, 8, false)); 282 } 283} 284 285// Emit instructions before MBBI (in MBB) to add NumBytes to Reg. 286static void emitIncrement(MachineBasicBlock &MBB, 287 MachineBasicBlock::iterator &MBBI, 288 const DebugLoc &DL, 289 unsigned Reg, int64_t NumBytes, 290 const TargetInstrInfo *TII) { 291 while (NumBytes) { 292 unsigned Opcode; 293 int64_t ThisVal = NumBytes; 294 if (isInt<16>(NumBytes)) 295 Opcode = SystemZ::AGHI; 296 else { 297 Opcode = SystemZ::AGFI; 298 // Make sure we maintain 8-byte stack alignment. 299 int64_t MinVal = -uint64_t(1) << 31; 300 int64_t MaxVal = (int64_t(1) << 31) - 8; 301 if (ThisVal < MinVal) 302 ThisVal = MinVal; 303 else if (ThisVal > MaxVal) 304 ThisVal = MaxVal; 305 } 306 MachineInstr *MI = BuildMI(MBB, MBBI, DL, TII->get(Opcode), Reg) 307 .addReg(Reg).addImm(ThisVal); 308 // The CC implicit def is dead. 309 MI->getOperand(3).setIsDead(); 310 NumBytes -= ThisVal; 311 } 312} 313 314void SystemZFrameLowering::emitPrologue(MachineFunction &MF, 315 MachineBasicBlock &MBB) const { 316 assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported"); 317 MachineFrameInfo *MFFrame = MF.getFrameInfo(); 318 auto *ZII = 319 static_cast<const SystemZInstrInfo *>(MF.getSubtarget().getInstrInfo()); 320 SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); 321 MachineBasicBlock::iterator MBBI = MBB.begin(); 322 MachineModuleInfo &MMI = MF.getMMI(); 323 const MCRegisterInfo *MRI = MMI.getContext().getRegisterInfo(); 324 const std::vector<CalleeSavedInfo> &CSI = MFFrame->getCalleeSavedInfo(); 325 bool HasFP = hasFP(MF); 326 327 // Debug location must be unknown since the first debug location is used 328 // to determine the end of the prologue. 329 DebugLoc DL; 330 331 // The current offset of the stack pointer from the CFA. 332 int64_t SPOffsetFromCFA = -SystemZMC::CFAOffsetFromInitialSP; 333 334 if (ZFI->getLowSavedGPR()) { 335 // Skip over the GPR saves. 336 if (MBBI != MBB.end() && MBBI->getOpcode() == SystemZ::STMG) 337 ++MBBI; 338 else 339 llvm_unreachable("Couldn't skip over GPR saves"); 340 341 // Add CFI for the GPR saves. 342 for (auto &Save : CSI) { 343 unsigned Reg = Save.getReg(); 344 if (SystemZ::GR64BitRegClass.contains(Reg)) { 345 int64_t Offset = SPOffsetFromCFA + RegSpillOffsets[Reg]; 346 unsigned CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset( 347 nullptr, MRI->getDwarfRegNum(Reg, true), Offset)); 348 BuildMI(MBB, MBBI, DL, ZII->get(TargetOpcode::CFI_INSTRUCTION)) 349 .addCFIIndex(CFIIndex); 350 } 351 } 352 } 353 354 uint64_t StackSize = getAllocatedStackSize(MF); 355 if (StackSize) { 356 // Allocate StackSize bytes. 357 int64_t Delta = -int64_t(StackSize); 358 emitIncrement(MBB, MBBI, DL, SystemZ::R15D, Delta, ZII); 359 360 // Add CFI for the allocation. 361 unsigned CFIIndex = MMI.addFrameInst( 362 MCCFIInstruction::createDefCfaOffset(nullptr, SPOffsetFromCFA + Delta)); 363 BuildMI(MBB, MBBI, DL, ZII->get(TargetOpcode::CFI_INSTRUCTION)) 364 .addCFIIndex(CFIIndex); 365 SPOffsetFromCFA += Delta; 366 } 367 368 if (HasFP) { 369 // Copy the base of the frame to R11. 370 BuildMI(MBB, MBBI, DL, ZII->get(SystemZ::LGR), SystemZ::R11D) 371 .addReg(SystemZ::R15D); 372 373 // Add CFI for the new frame location. 374 unsigned HardFP = MRI->getDwarfRegNum(SystemZ::R11D, true); 375 unsigned CFIIndex = MMI.addFrameInst( 376 MCCFIInstruction::createDefCfaRegister(nullptr, HardFP)); 377 BuildMI(MBB, MBBI, DL, ZII->get(TargetOpcode::CFI_INSTRUCTION)) 378 .addCFIIndex(CFIIndex); 379 380 // Mark the FramePtr as live at the beginning of every block except 381 // the entry block. (We'll have marked R11 as live on entry when 382 // saving the GPRs.) 383 for (auto I = std::next(MF.begin()), E = MF.end(); I != E; ++I) 384 I->addLiveIn(SystemZ::R11D); 385 } 386 387 // Skip over the FPR saves. 388 SmallVector<unsigned, 8> CFIIndexes; 389 for (auto &Save : CSI) { 390 unsigned Reg = Save.getReg(); 391 if (SystemZ::FP64BitRegClass.contains(Reg)) { 392 if (MBBI != MBB.end() && 393 (MBBI->getOpcode() == SystemZ::STD || 394 MBBI->getOpcode() == SystemZ::STDY)) 395 ++MBBI; 396 else 397 llvm_unreachable("Couldn't skip over FPR save"); 398 399 // Add CFI for the this save. 400 unsigned DwarfReg = MRI->getDwarfRegNum(Reg, true); 401 unsigned IgnoredFrameReg; 402 int64_t Offset = 403 getFrameIndexReference(MF, Save.getFrameIdx(), IgnoredFrameReg); 404 405 unsigned CFIIndex = MMI.addFrameInst(MCCFIInstruction::createOffset( 406 nullptr, DwarfReg, SPOffsetFromCFA + Offset)); 407 CFIIndexes.push_back(CFIIndex); 408 } 409 } 410 // Complete the CFI for the FPR saves, modelling them as taking effect 411 // after the last save. 412 for (auto CFIIndex : CFIIndexes) { 413 BuildMI(MBB, MBBI, DL, ZII->get(TargetOpcode::CFI_INSTRUCTION)) 414 .addCFIIndex(CFIIndex); 415 } 416} 417 418void SystemZFrameLowering::emitEpilogue(MachineFunction &MF, 419 MachineBasicBlock &MBB) const { 420 MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr(); 421 auto *ZII = 422 static_cast<const SystemZInstrInfo *>(MF.getSubtarget().getInstrInfo()); 423 SystemZMachineFunctionInfo *ZFI = MF.getInfo<SystemZMachineFunctionInfo>(); 424 425 // Skip the return instruction. 426 assert(MBBI->isReturn() && "Can only insert epilogue into returning blocks"); 427 428 uint64_t StackSize = getAllocatedStackSize(MF); 429 if (ZFI->getLowSavedGPR()) { 430 --MBBI; 431 unsigned Opcode = MBBI->getOpcode(); 432 if (Opcode != SystemZ::LMG) 433 llvm_unreachable("Expected to see callee-save register restore code"); 434 435 unsigned AddrOpNo = 2; 436 DebugLoc DL = MBBI->getDebugLoc(); 437 uint64_t Offset = StackSize + MBBI->getOperand(AddrOpNo + 1).getImm(); 438 unsigned NewOpcode = ZII->getOpcodeForOffset(Opcode, Offset); 439 440 // If the offset is too large, use the largest stack-aligned offset 441 // and add the rest to the base register (the stack or frame pointer). 442 if (!NewOpcode) { 443 uint64_t NumBytes = Offset - 0x7fff8; 444 emitIncrement(MBB, MBBI, DL, MBBI->getOperand(AddrOpNo).getReg(), 445 NumBytes, ZII); 446 Offset -= NumBytes; 447 NewOpcode = ZII->getOpcodeForOffset(Opcode, Offset); 448 assert(NewOpcode && "No restore instruction available"); 449 } 450 451 MBBI->setDesc(ZII->get(NewOpcode)); 452 MBBI->getOperand(AddrOpNo + 1).ChangeToImmediate(Offset); 453 } else if (StackSize) { 454 DebugLoc DL = MBBI->getDebugLoc(); 455 emitIncrement(MBB, MBBI, DL, SystemZ::R15D, StackSize, ZII); 456 } 457} 458 459bool SystemZFrameLowering::hasFP(const MachineFunction &MF) const { 460 return (MF.getTarget().Options.DisableFramePointerElim(MF) || 461 MF.getFrameInfo()->hasVarSizedObjects() || 462 MF.getInfo<SystemZMachineFunctionInfo>()->getManipulatesSP()); 463} 464 465int SystemZFrameLowering::getFrameIndexReference(const MachineFunction &MF, 466 int FI, 467 unsigned &FrameReg) const { 468 const MachineFrameInfo *MFFrame = MF.getFrameInfo(); 469 const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo(); 470 471 // Fill in FrameReg output argument. 472 FrameReg = RI->getFrameRegister(MF); 473 474 // Start with the offset of FI from the top of the caller-allocated frame 475 // (i.e. the top of the 160 bytes allocated by the caller). This initial 476 // offset is therefore negative. 477 int64_t Offset = (MFFrame->getObjectOffset(FI) + 478 MFFrame->getOffsetAdjustment()); 479 480 // Make the offset relative to the incoming stack pointer. 481 Offset -= getOffsetOfLocalArea(); 482 483 // Make the offset relative to the bottom of the frame. 484 Offset += getAllocatedStackSize(MF); 485 486 return Offset; 487} 488 489uint64_t SystemZFrameLowering:: 490getAllocatedStackSize(const MachineFunction &MF) const { 491 const MachineFrameInfo *MFFrame = MF.getFrameInfo(); 492 493 // Start with the size of the local variables and spill slots. 494 uint64_t StackSize = MFFrame->getStackSize(); 495 496 // We need to allocate the ABI-defined 160-byte base area whenever 497 // we allocate stack space for our own use and whenever we call another 498 // function. 499 if (StackSize || MFFrame->hasVarSizedObjects() || MFFrame->hasCalls()) 500 StackSize += SystemZMC::CallFrameSize; 501 502 return StackSize; 503} 504 505bool 506SystemZFrameLowering::hasReservedCallFrame(const MachineFunction &MF) const { 507 // The ABI requires us to allocate 160 bytes of stack space for the callee, 508 // with any outgoing stack arguments being placed above that. It seems 509 // better to make that area a permanent feature of the frame even if 510 // we're using a frame pointer. 511 return true; 512} 513 514void SystemZFrameLowering:: 515eliminateCallFramePseudoInstr(MachineFunction &MF, 516 MachineBasicBlock &MBB, 517 MachineBasicBlock::iterator MI) const { 518 switch (MI->getOpcode()) { 519 case SystemZ::ADJCALLSTACKDOWN: 520 case SystemZ::ADJCALLSTACKUP: 521 assert(hasReservedCallFrame(MF) && 522 "ADJSTACKDOWN and ADJSTACKUP should be no-ops"); 523 MBB.erase(MI); 524 break; 525 526 default: 527 llvm_unreachable("Unexpected call frame instruction"); 528 } 529} 530