1 //===-- PrologEpilogInserter.cpp - Insert Prolog/Epilog code in function --===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass is responsible for finalizing the functions frame layout, saving 11 // callee saved registers, and for emitting prolog & epilog code for the 12 // function. 13 // 14 // This pass must be run after register allocation. After this pass is 15 // executed, it is illegal to construct MO_FrameIndex operands. 16 // 17 // This pass provides an optional shrink wrapping variant of prolog/epilog 18 // insertion, enabled via --shrink-wrap. See ShrinkWrapping.cpp. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #define DEBUG_TYPE "pei" 23 #include "PrologEpilogInserter.h" 24 #include "llvm/InlineAsm.h" 25 #include "llvm/CodeGen/MachineDominators.h" 26 #include "llvm/CodeGen/MachineLoopInfo.h" 27 #include "llvm/CodeGen/MachineInstr.h" 28 #include "llvm/CodeGen/MachineFrameInfo.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/RegisterScavenging.h" 31 #include "llvm/Target/TargetMachine.h" 32 #include "llvm/Target/TargetOptions.h" 33 #include "llvm/Target/TargetRegisterInfo.h" 34 #include "llvm/Target/TargetFrameLowering.h" 35 #include "llvm/Target/TargetInstrInfo.h" 36 #include "llvm/Support/CommandLine.h" 37 #include "llvm/Support/Compiler.h" 38 #include "llvm/Support/Debug.h" 39 #include "llvm/ADT/IndexedMap.h" 40 #include "llvm/ADT/SmallSet.h" 41 #include "llvm/ADT/Statistic.h" 42 #include "llvm/ADT/STLExtras.h" 43 #include <climits> 44 45 using namespace llvm; 46 47 char PEI::ID = 0; 48 49 INITIALIZE_PASS_BEGIN(PEI, "prologepilog", 50 "Prologue/Epilogue Insertion", false, false) 51 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 52 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 53 INITIALIZE_PASS_END(PEI, "prologepilog", 54 "Prologue/Epilogue Insertion", false, false) 55 56 STATISTIC(NumVirtualFrameRegs, "Number of virtual frame regs encountered"); 57 STATISTIC(NumScavengedRegs, "Number of frame index regs scavenged"); 58 STATISTIC(NumBytesStackSpace, 59 "Number of bytes used for stack in all functions"); 60 61 /// createPrologEpilogCodeInserter - This function returns a pass that inserts 62 /// prolog and epilog code, and eliminates abstract frame references. 63 /// 64 FunctionPass *llvm::createPrologEpilogCodeInserter() { return new PEI(); } 65 66 /// runOnMachineFunction - Insert prolog/epilog code and replace abstract 67 /// frame indexes with appropriate references. 68 /// 69 bool PEI::runOnMachineFunction(MachineFunction &Fn) { 70 const Function* F = Fn.getFunction(); 71 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 72 const TargetFrameLowering *TFI = Fn.getTarget().getFrameLowering(); 73 74 RS = TRI->requiresRegisterScavenging(Fn) ? new RegScavenger() : NULL; 75 FrameIndexVirtualScavenging = TRI->requiresFrameIndexScavenging(Fn); 76 77 // Calculate the MaxCallFrameSize and AdjustsStack variables for the 78 // function's frame information. Also eliminates call frame pseudo 79 // instructions. 80 calculateCallsInformation(Fn); 81 82 // Allow the target machine to make some adjustments to the function 83 // e.g. UsedPhysRegs before calculateCalleeSavedRegisters. 84 TFI->processFunctionBeforeCalleeSavedScan(Fn, RS); 85 86 // Scan the function for modified callee saved registers and insert spill code 87 // for any callee saved registers that are modified. 88 calculateCalleeSavedRegisters(Fn); 89 90 // Determine placement of CSR spill/restore code: 91 // - With shrink wrapping, place spills and restores to tightly 92 // enclose regions in the Machine CFG of the function where 93 // they are used. 94 // - Without shink wrapping (default), place all spills in the 95 // entry block, all restores in return blocks. 96 placeCSRSpillsAndRestores(Fn); 97 98 // Add the code to save and restore the callee saved registers 99 if (!F->hasFnAttr(Attribute::Naked)) 100 insertCSRSpillsAndRestores(Fn); 101 102 // Allow the target machine to make final modifications to the function 103 // before the frame layout is finalized. 104 TFI->processFunctionBeforeFrameFinalized(Fn); 105 106 // Calculate actual frame offsets for all abstract stack objects... 107 calculateFrameObjectOffsets(Fn); 108 109 // Add prolog and epilog code to the function. This function is required 110 // to align the stack frame as necessary for any stack variables or 111 // called functions. Because of this, calculateCalleeSavedRegisters() 112 // must be called before this function in order to set the AdjustsStack 113 // and MaxCallFrameSize variables. 114 if (!F->hasFnAttr(Attribute::Naked)) 115 insertPrologEpilogCode(Fn); 116 117 // Replace all MO_FrameIndex operands with physical register references 118 // and actual offsets. 119 // 120 replaceFrameIndices(Fn); 121 122 // If register scavenging is needed, as we've enabled doing it as a 123 // post-pass, scavenge the virtual registers that frame index elimiation 124 // inserted. 125 if (TRI->requiresRegisterScavenging(Fn) && FrameIndexVirtualScavenging) 126 scavengeFrameVirtualRegs(Fn); 127 128 delete RS; 129 clearAllSets(); 130 return true; 131 } 132 133 #if 0 134 void PEI::getAnalysisUsage(AnalysisUsage &AU) const { 135 AU.setPreservesCFG(); 136 if (ShrinkWrapping || ShrinkWrapFunc != "") { 137 AU.addRequired<MachineLoopInfo>(); 138 AU.addRequired<MachineDominatorTree>(); 139 } 140 AU.addPreserved<MachineLoopInfo>(); 141 AU.addPreserved<MachineDominatorTree>(); 142 MachineFunctionPass::getAnalysisUsage(AU); 143 } 144 #endif 145 146 /// calculateCallsInformation - Calculate the MaxCallFrameSize and AdjustsStack 147 /// variables for the function's frame information and eliminate call frame 148 /// pseudo instructions. 149 void PEI::calculateCallsInformation(MachineFunction &Fn) { 150 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 151 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 152 const TargetFrameLowering *TFI = Fn.getTarget().getFrameLowering(); 153 MachineFrameInfo *MFI = Fn.getFrameInfo(); 154 155 unsigned MaxCallFrameSize = 0; 156 bool AdjustsStack = MFI->adjustsStack(); 157 158 // Get the function call frame set-up and tear-down instruction opcode 159 int FrameSetupOpcode = TII.getCallFrameSetupOpcode(); 160 int FrameDestroyOpcode = TII.getCallFrameDestroyOpcode(); 161 162 // Early exit for targets which have no call frame setup/destroy pseudo 163 // instructions. 164 if (FrameSetupOpcode == -1 && FrameDestroyOpcode == -1) 165 return; 166 167 std::vector<MachineBasicBlock::iterator> FrameSDOps; 168 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) 169 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) 170 if (I->getOpcode() == FrameSetupOpcode || 171 I->getOpcode() == FrameDestroyOpcode) { 172 assert(I->getNumOperands() >= 1 && "Call Frame Setup/Destroy Pseudo" 173 " instructions should have a single immediate argument!"); 174 unsigned Size = I->getOperand(0).getImm(); 175 if (Size > MaxCallFrameSize) MaxCallFrameSize = Size; 176 AdjustsStack = true; 177 FrameSDOps.push_back(I); 178 } else if (I->isInlineAsm()) { 179 // Some inline asm's need a stack frame, as indicated by operand 1. 180 unsigned ExtraInfo = I->getOperand(InlineAsm::MIOp_ExtraInfo).getImm(); 181 if (ExtraInfo & InlineAsm::Extra_IsAlignStack) 182 AdjustsStack = true; 183 } 184 185 MFI->setAdjustsStack(AdjustsStack); 186 MFI->setMaxCallFrameSize(MaxCallFrameSize); 187 188 for (std::vector<MachineBasicBlock::iterator>::iterator 189 i = FrameSDOps.begin(), e = FrameSDOps.end(); i != e; ++i) { 190 MachineBasicBlock::iterator I = *i; 191 192 // If call frames are not being included as part of the stack frame, and 193 // the target doesn't indicate otherwise, remove the call frame pseudos 194 // here. The sub/add sp instruction pairs are still inserted, but we don't 195 // need to track the SP adjustment for frame index elimination. 196 if (TFI->canSimplifyCallFramePseudos(Fn)) 197 RegInfo->eliminateCallFramePseudoInstr(Fn, *I->getParent(), I); 198 } 199 } 200 201 202 /// calculateCalleeSavedRegisters - Scan the function for modified callee saved 203 /// registers. 204 void PEI::calculateCalleeSavedRegisters(MachineFunction &Fn) { 205 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 206 const TargetFrameLowering *TFI = Fn.getTarget().getFrameLowering(); 207 MachineFrameInfo *MFI = Fn.getFrameInfo(); 208 209 // Get the callee saved register list... 210 const unsigned *CSRegs = RegInfo->getCalleeSavedRegs(&Fn); 211 212 // These are used to keep track the callee-save area. Initialize them. 213 MinCSFrameIndex = INT_MAX; 214 MaxCSFrameIndex = 0; 215 216 // Early exit for targets which have no callee saved registers. 217 if (CSRegs == 0 || CSRegs[0] == 0) 218 return; 219 220 // In Naked functions we aren't going to save any registers. 221 if (Fn.getFunction()->hasFnAttr(Attribute::Naked)) 222 return; 223 224 std::vector<CalleeSavedInfo> CSI; 225 for (unsigned i = 0; CSRegs[i]; ++i) { 226 unsigned Reg = CSRegs[i]; 227 if (Fn.getRegInfo().isPhysRegUsed(Reg)) { 228 // If the reg is modified, save it! 229 CSI.push_back(CalleeSavedInfo(Reg)); 230 } else { 231 for (const unsigned *AliasSet = RegInfo->getAliasSet(Reg); 232 *AliasSet; ++AliasSet) { // Check alias registers too. 233 if (Fn.getRegInfo().isPhysRegUsed(*AliasSet)) { 234 CSI.push_back(CalleeSavedInfo(Reg)); 235 break; 236 } 237 } 238 } 239 } 240 241 if (CSI.empty()) 242 return; // Early exit if no callee saved registers are modified! 243 244 unsigned NumFixedSpillSlots; 245 const TargetFrameLowering::SpillSlot *FixedSpillSlots = 246 TFI->getCalleeSavedSpillSlots(NumFixedSpillSlots); 247 248 // Now that we know which registers need to be saved and restored, allocate 249 // stack slots for them. 250 for (std::vector<CalleeSavedInfo>::iterator 251 I = CSI.begin(), E = CSI.end(); I != E; ++I) { 252 unsigned Reg = I->getReg(); 253 const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg); 254 255 int FrameIdx; 256 if (RegInfo->hasReservedSpillSlot(Fn, Reg, FrameIdx)) { 257 I->setFrameIdx(FrameIdx); 258 continue; 259 } 260 261 // Check to see if this physreg must be spilled to a particular stack slot 262 // on this target. 263 const TargetFrameLowering::SpillSlot *FixedSlot = FixedSpillSlots; 264 while (FixedSlot != FixedSpillSlots+NumFixedSpillSlots && 265 FixedSlot->Reg != Reg) 266 ++FixedSlot; 267 268 if (FixedSlot == FixedSpillSlots + NumFixedSpillSlots) { 269 // Nope, just spill it anywhere convenient. 270 unsigned Align = RC->getAlignment(); 271 unsigned StackAlign = TFI->getStackAlignment(); 272 273 // We may not be able to satisfy the desired alignment specification of 274 // the TargetRegisterClass if the stack alignment is smaller. Use the 275 // min. 276 Align = std::min(Align, StackAlign); 277 FrameIdx = MFI->CreateStackObject(RC->getSize(), Align, true); 278 if ((unsigned)FrameIdx < MinCSFrameIndex) MinCSFrameIndex = FrameIdx; 279 if ((unsigned)FrameIdx > MaxCSFrameIndex) MaxCSFrameIndex = FrameIdx; 280 } else { 281 // Spill it to the stack where we must. 282 FrameIdx = MFI->CreateFixedObject(RC->getSize(), FixedSlot->Offset, true); 283 } 284 285 I->setFrameIdx(FrameIdx); 286 } 287 288 MFI->setCalleeSavedInfo(CSI); 289 } 290 291 /// insertCSRSpillsAndRestores - Insert spill and restore code for 292 /// callee saved registers used in the function, handling shrink wrapping. 293 /// 294 void PEI::insertCSRSpillsAndRestores(MachineFunction &Fn) { 295 // Get callee saved register information. 296 MachineFrameInfo *MFI = Fn.getFrameInfo(); 297 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 298 299 MFI->setCalleeSavedInfoValid(true); 300 301 // Early exit if no callee saved registers are modified! 302 if (CSI.empty()) 303 return; 304 305 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 306 const TargetFrameLowering *TFI = Fn.getTarget().getFrameLowering(); 307 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 308 MachineBasicBlock::iterator I; 309 310 if (! ShrinkWrapThisFunction) { 311 // Spill using target interface. 312 I = EntryBlock->begin(); 313 if (!TFI->spillCalleeSavedRegisters(*EntryBlock, I, CSI, TRI)) { 314 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 315 // Add the callee-saved register as live-in. 316 // It's killed at the spill. 317 EntryBlock->addLiveIn(CSI[i].getReg()); 318 319 // Insert the spill to the stack frame. 320 unsigned Reg = CSI[i].getReg(); 321 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 322 TII.storeRegToStackSlot(*EntryBlock, I, Reg, true, 323 CSI[i].getFrameIdx(), RC, TRI); 324 } 325 } 326 327 // Restore using target interface. 328 for (unsigned ri = 0, re = ReturnBlocks.size(); ri != re; ++ri) { 329 MachineBasicBlock* MBB = ReturnBlocks[ri]; 330 I = MBB->end(); --I; 331 332 // Skip over all terminator instructions, which are part of the return 333 // sequence. 334 MachineBasicBlock::iterator I2 = I; 335 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator()) 336 I = I2; 337 338 bool AtStart = I == MBB->begin(); 339 MachineBasicBlock::iterator BeforeI = I; 340 if (!AtStart) 341 --BeforeI; 342 343 // Restore all registers immediately before the return and any 344 // terminators that precede it. 345 if (!TFI->restoreCalleeSavedRegisters(*MBB, I, CSI, TRI)) { 346 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 347 unsigned Reg = CSI[i].getReg(); 348 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 349 TII.loadRegFromStackSlot(*MBB, I, Reg, 350 CSI[i].getFrameIdx(), 351 RC, TRI); 352 assert(I != MBB->begin() && 353 "loadRegFromStackSlot didn't insert any code!"); 354 // Insert in reverse order. loadRegFromStackSlot can insert 355 // multiple instructions. 356 if (AtStart) 357 I = MBB->begin(); 358 else { 359 I = BeforeI; 360 ++I; 361 } 362 } 363 } 364 } 365 return; 366 } 367 368 // Insert spills. 369 std::vector<CalleeSavedInfo> blockCSI; 370 for (CSRegBlockMap::iterator BI = CSRSave.begin(), 371 BE = CSRSave.end(); BI != BE; ++BI) { 372 MachineBasicBlock* MBB = BI->first; 373 CSRegSet save = BI->second; 374 375 if (save.empty()) 376 continue; 377 378 blockCSI.clear(); 379 for (CSRegSet::iterator RI = save.begin(), 380 RE = save.end(); RI != RE; ++RI) { 381 blockCSI.push_back(CSI[*RI]); 382 } 383 assert(blockCSI.size() > 0 && 384 "Could not collect callee saved register info"); 385 386 I = MBB->begin(); 387 388 // When shrink wrapping, use stack slot stores/loads. 389 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 390 // Add the callee-saved register as live-in. 391 // It's killed at the spill. 392 MBB->addLiveIn(blockCSI[i].getReg()); 393 394 // Insert the spill to the stack frame. 395 unsigned Reg = blockCSI[i].getReg(); 396 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 397 TII.storeRegToStackSlot(*MBB, I, Reg, 398 true, 399 blockCSI[i].getFrameIdx(), 400 RC, TRI); 401 } 402 } 403 404 for (CSRegBlockMap::iterator BI = CSRRestore.begin(), 405 BE = CSRRestore.end(); BI != BE; ++BI) { 406 MachineBasicBlock* MBB = BI->first; 407 CSRegSet restore = BI->second; 408 409 if (restore.empty()) 410 continue; 411 412 blockCSI.clear(); 413 for (CSRegSet::iterator RI = restore.begin(), 414 RE = restore.end(); RI != RE; ++RI) { 415 blockCSI.push_back(CSI[*RI]); 416 } 417 assert(blockCSI.size() > 0 && 418 "Could not find callee saved register info"); 419 420 // If MBB is empty and needs restores, insert at the _beginning_. 421 if (MBB->empty()) { 422 I = MBB->begin(); 423 } else { 424 I = MBB->end(); 425 --I; 426 427 // Skip over all terminator instructions, which are part of the 428 // return sequence. 429 if (! I->getDesc().isTerminator()) { 430 ++I; 431 } else { 432 MachineBasicBlock::iterator I2 = I; 433 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator()) 434 I = I2; 435 } 436 } 437 438 bool AtStart = I == MBB->begin(); 439 MachineBasicBlock::iterator BeforeI = I; 440 if (!AtStart) 441 --BeforeI; 442 443 // Restore all registers immediately before the return and any 444 // terminators that precede it. 445 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 446 unsigned Reg = blockCSI[i].getReg(); 447 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 448 TII.loadRegFromStackSlot(*MBB, I, Reg, 449 blockCSI[i].getFrameIdx(), 450 RC, TRI); 451 assert(I != MBB->begin() && 452 "loadRegFromStackSlot didn't insert any code!"); 453 // Insert in reverse order. loadRegFromStackSlot can insert 454 // multiple instructions. 455 if (AtStart) 456 I = MBB->begin(); 457 else { 458 I = BeforeI; 459 ++I; 460 } 461 } 462 } 463 } 464 465 /// AdjustStackOffset - Helper function used to adjust the stack frame offset. 466 static inline void 467 AdjustStackOffset(MachineFrameInfo *MFI, int FrameIdx, 468 bool StackGrowsDown, int64_t &Offset, 469 unsigned &MaxAlign) { 470 // If the stack grows down, add the object size to find the lowest address. 471 if (StackGrowsDown) 472 Offset += MFI->getObjectSize(FrameIdx); 473 474 unsigned Align = MFI->getObjectAlignment(FrameIdx); 475 476 // If the alignment of this object is greater than that of the stack, then 477 // increase the stack alignment to match. 478 MaxAlign = std::max(MaxAlign, Align); 479 480 // Adjust to alignment boundary. 481 Offset = (Offset + Align - 1) / Align * Align; 482 483 if (StackGrowsDown) { 484 DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << -Offset << "]\n"); 485 MFI->setObjectOffset(FrameIdx, -Offset); // Set the computed offset 486 } else { 487 DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << Offset << "]\n"); 488 MFI->setObjectOffset(FrameIdx, Offset); 489 Offset += MFI->getObjectSize(FrameIdx); 490 } 491 } 492 493 /// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the 494 /// abstract stack objects. 495 /// 496 void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { 497 const TargetFrameLowering &TFI = *Fn.getTarget().getFrameLowering(); 498 499 bool StackGrowsDown = 500 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 501 502 // Loop over all of the stack objects, assigning sequential addresses... 503 MachineFrameInfo *MFI = Fn.getFrameInfo(); 504 505 // Start at the beginning of the local area. 506 // The Offset is the distance from the stack top in the direction 507 // of stack growth -- so it's always nonnegative. 508 int LocalAreaOffset = TFI.getOffsetOfLocalArea(); 509 if (StackGrowsDown) 510 LocalAreaOffset = -LocalAreaOffset; 511 assert(LocalAreaOffset >= 0 512 && "Local area offset should be in direction of stack growth"); 513 int64_t Offset = LocalAreaOffset; 514 515 // If there are fixed sized objects that are preallocated in the local area, 516 // non-fixed objects can't be allocated right at the start of local area. 517 // We currently don't support filling in holes in between fixed sized 518 // objects, so we adjust 'Offset' to point to the end of last fixed sized 519 // preallocated object. 520 for (int i = MFI->getObjectIndexBegin(); i != 0; ++i) { 521 int64_t FixedOff; 522 if (StackGrowsDown) { 523 // The maximum distance from the stack pointer is at lower address of 524 // the object -- which is given by offset. For down growing stack 525 // the offset is negative, so we negate the offset to get the distance. 526 FixedOff = -MFI->getObjectOffset(i); 527 } else { 528 // The maximum distance from the start pointer is at the upper 529 // address of the object. 530 FixedOff = MFI->getObjectOffset(i) + MFI->getObjectSize(i); 531 } 532 if (FixedOff > Offset) Offset = FixedOff; 533 } 534 535 // First assign frame offsets to stack objects that are used to spill 536 // callee saved registers. 537 if (StackGrowsDown) { 538 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) { 539 // If the stack grows down, we need to add the size to find the lowest 540 // address of the object. 541 Offset += MFI->getObjectSize(i); 542 543 unsigned Align = MFI->getObjectAlignment(i); 544 // Adjust to alignment boundary 545 Offset = (Offset+Align-1)/Align*Align; 546 547 MFI->setObjectOffset(i, -Offset); // Set the computed offset 548 } 549 } else { 550 int MaxCSFI = MaxCSFrameIndex, MinCSFI = MinCSFrameIndex; 551 for (int i = MaxCSFI; i >= MinCSFI ; --i) { 552 unsigned Align = MFI->getObjectAlignment(i); 553 // Adjust to alignment boundary 554 Offset = (Offset+Align-1)/Align*Align; 555 556 MFI->setObjectOffset(i, Offset); 557 Offset += MFI->getObjectSize(i); 558 } 559 } 560 561 unsigned MaxAlign = MFI->getMaxAlignment(); 562 563 // Make sure the special register scavenging spill slot is closest to the 564 // frame pointer if a frame pointer is required. 565 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 566 if (RS && TFI.hasFP(Fn) && RegInfo->useFPForScavengingIndex(Fn) && 567 !RegInfo->needsStackRealignment(Fn)) { 568 int SFI = RS->getScavengingFrameIndex(); 569 if (SFI >= 0) 570 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 571 } 572 573 // FIXME: Once this is working, then enable flag will change to a target 574 // check for whether the frame is large enough to want to use virtual 575 // frame index registers. Functions which don't want/need this optimization 576 // will continue to use the existing code path. 577 if (MFI->getUseLocalStackAllocationBlock()) { 578 unsigned Align = MFI->getLocalFrameMaxAlign(); 579 580 // Adjust to alignment boundary. 581 Offset = (Offset + Align - 1) / Align * Align; 582 583 DEBUG(dbgs() << "Local frame base offset: " << Offset << "\n"); 584 585 // Resolve offsets for objects in the local block. 586 for (unsigned i = 0, e = MFI->getLocalFrameObjectCount(); i != e; ++i) { 587 std::pair<int, int64_t> Entry = MFI->getLocalFrameObjectMap(i); 588 int64_t FIOffset = (StackGrowsDown ? -Offset : Offset) + Entry.second; 589 DEBUG(dbgs() << "alloc FI(" << Entry.first << ") at SP[" << 590 FIOffset << "]\n"); 591 MFI->setObjectOffset(Entry.first, FIOffset); 592 } 593 // Allocate the local block 594 Offset += MFI->getLocalFrameSize(); 595 596 MaxAlign = std::max(Align, MaxAlign); 597 } 598 599 // Make sure that the stack protector comes before the local variables on the 600 // stack. 601 SmallSet<int, 16> LargeStackObjs; 602 if (MFI->getStackProtectorIndex() >= 0) { 603 AdjustStackOffset(MFI, MFI->getStackProtectorIndex(), StackGrowsDown, 604 Offset, MaxAlign); 605 606 // Assign large stack objects first. 607 for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 608 if (MFI->isObjectPreAllocated(i) && 609 MFI->getUseLocalStackAllocationBlock()) 610 continue; 611 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex) 612 continue; 613 if (RS && (int)i == RS->getScavengingFrameIndex()) 614 continue; 615 if (MFI->isDeadObjectIndex(i)) 616 continue; 617 if (MFI->getStackProtectorIndex() == (int)i) 618 continue; 619 if (!MFI->MayNeedStackProtector(i)) 620 continue; 621 622 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign); 623 LargeStackObjs.insert(i); 624 } 625 } 626 627 // Then assign frame offsets to stack objects that are not used to spill 628 // callee saved registers. 629 for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 630 if (MFI->isObjectPreAllocated(i) && 631 MFI->getUseLocalStackAllocationBlock()) 632 continue; 633 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex) 634 continue; 635 if (RS && (int)i == RS->getScavengingFrameIndex()) 636 continue; 637 if (MFI->isDeadObjectIndex(i)) 638 continue; 639 if (MFI->getStackProtectorIndex() == (int)i) 640 continue; 641 if (LargeStackObjs.count(i)) 642 continue; 643 644 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign); 645 } 646 647 // Make sure the special register scavenging spill slot is closest to the 648 // stack pointer. 649 if (RS && (!TFI.hasFP(Fn) || RegInfo->needsStackRealignment(Fn) || 650 !RegInfo->useFPForScavengingIndex(Fn))) { 651 int SFI = RS->getScavengingFrameIndex(); 652 if (SFI >= 0) 653 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 654 } 655 656 if (!TFI.targetHandlesStackFrameRounding()) { 657 // If we have reserved argument space for call sites in the function 658 // immediately on entry to the current function, count it as part of the 659 // overall stack size. 660 if (MFI->adjustsStack() && TFI.hasReservedCallFrame(Fn)) 661 Offset += MFI->getMaxCallFrameSize(); 662 663 // Round up the size to a multiple of the alignment. If the function has 664 // any calls or alloca's, align to the target's StackAlignment value to 665 // ensure that the callee's frame or the alloca data is suitably aligned; 666 // otherwise, for leaf functions, align to the TransientStackAlignment 667 // value. 668 unsigned StackAlign; 669 if (MFI->adjustsStack() || MFI->hasVarSizedObjects() || 670 (RegInfo->needsStackRealignment(Fn) && MFI->getObjectIndexEnd() != 0)) 671 StackAlign = TFI.getStackAlignment(); 672 else 673 StackAlign = TFI.getTransientStackAlignment(); 674 675 // If the frame pointer is eliminated, all frame offsets will be relative to 676 // SP not FP. Align to MaxAlign so this works. 677 StackAlign = std::max(StackAlign, MaxAlign); 678 unsigned AlignMask = StackAlign - 1; 679 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask); 680 } 681 682 // Update frame info to pretend that this is part of the stack... 683 int64_t StackSize = Offset - LocalAreaOffset; 684 MFI->setStackSize(StackSize); 685 NumBytesStackSpace += StackSize; 686 } 687 688 /// insertPrologEpilogCode - Scan the function for modified callee saved 689 /// registers, insert spill code for these callee saved registers, then add 690 /// prolog and epilog code to the function. 691 /// 692 void PEI::insertPrologEpilogCode(MachineFunction &Fn) { 693 const TargetFrameLowering &TFI = *Fn.getTarget().getFrameLowering(); 694 695 // Add prologue to the function... 696 TFI.emitPrologue(Fn); 697 698 // Add epilogue to restore the callee-save registers in each exiting block 699 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) { 700 // If last instruction is a return instruction, add an epilogue 701 if (!I->empty() && I->back().getDesc().isReturn()) 702 TFI.emitEpilogue(Fn, *I); 703 } 704 705 // Emit additional code that is required to support segmented stacks, if 706 // we've been asked for it. This, when linked with a runtime with support 707 // for segmented stacks (libgcc is one), will result in allocating stack 708 // space in small chunks instead of one large contiguous block. 709 if (EnableSegmentedStacks) 710 TFI.adjustForSegmentedStacks(Fn); 711 } 712 713 /// replaceFrameIndices - Replace all MO_FrameIndex operands with physical 714 /// register references and actual offsets. 715 /// 716 void PEI::replaceFrameIndices(MachineFunction &Fn) { 717 if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do? 718 719 const TargetMachine &TM = Fn.getTarget(); 720 assert(TM.getRegisterInfo() && "TM::getRegisterInfo() must be implemented!"); 721 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 722 const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); 723 const TargetFrameLowering *TFI = TM.getFrameLowering(); 724 bool StackGrowsDown = 725 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 726 int FrameSetupOpcode = TII.getCallFrameSetupOpcode(); 727 int FrameDestroyOpcode = TII.getCallFrameDestroyOpcode(); 728 729 for (MachineFunction::iterator BB = Fn.begin(), 730 E = Fn.end(); BB != E; ++BB) { 731 #ifndef NDEBUG 732 int SPAdjCount = 0; // frame setup / destroy count. 733 #endif 734 int SPAdj = 0; // SP offset due to call frame setup / destroy. 735 if (RS && !FrameIndexVirtualScavenging) RS->enterBasicBlock(BB); 736 737 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 738 739 if (I->getOpcode() == FrameSetupOpcode || 740 I->getOpcode() == FrameDestroyOpcode) { 741 #ifndef NDEBUG 742 // Track whether we see even pairs of them 743 SPAdjCount += I->getOpcode() == FrameSetupOpcode ? 1 : -1; 744 #endif 745 // Remember how much SP has been adjusted to create the call 746 // frame. 747 int Size = I->getOperand(0).getImm(); 748 749 if ((!StackGrowsDown && I->getOpcode() == FrameSetupOpcode) || 750 (StackGrowsDown && I->getOpcode() == FrameDestroyOpcode)) 751 Size = -Size; 752 753 SPAdj += Size; 754 755 MachineBasicBlock::iterator PrevI = BB->end(); 756 if (I != BB->begin()) PrevI = prior(I); 757 TRI.eliminateCallFramePseudoInstr(Fn, *BB, I); 758 759 // Visit the instructions created by eliminateCallFramePseudoInstr(). 760 if (PrevI == BB->end()) 761 I = BB->begin(); // The replaced instr was the first in the block. 762 else 763 I = llvm::next(PrevI); 764 continue; 765 } 766 767 MachineInstr *MI = I; 768 bool DoIncr = true; 769 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 770 if (MI->getOperand(i).isFI()) { 771 // Some instructions (e.g. inline asm instructions) can have 772 // multiple frame indices and/or cause eliminateFrameIndex 773 // to insert more than one instruction. We need the register 774 // scavenger to go through all of these instructions so that 775 // it can update its register information. We keep the 776 // iterator at the point before insertion so that we can 777 // revisit them in full. 778 bool AtBeginning = (I == BB->begin()); 779 if (!AtBeginning) --I; 780 781 // If this instruction has a FrameIndex operand, we need to 782 // use that target machine register info object to eliminate 783 // it. 784 TRI.eliminateFrameIndex(MI, SPAdj, 785 FrameIndexVirtualScavenging ? NULL : RS); 786 787 // Reset the iterator if we were at the beginning of the BB. 788 if (AtBeginning) { 789 I = BB->begin(); 790 DoIncr = false; 791 } 792 793 MI = 0; 794 break; 795 } 796 797 if (DoIncr && I != BB->end()) ++I; 798 799 // Update register states. 800 if (RS && !FrameIndexVirtualScavenging && MI) RS->forward(MI); 801 } 802 803 // If we have evenly matched pairs of frame setup / destroy instructions, 804 // make sure the adjustments come out to zero. If we don't have matched 805 // pairs, we can't be sure the missing bit isn't in another basic block 806 // due to a custom inserter playing tricks, so just asserting SPAdj==0 807 // isn't sufficient. See tMOVCC on Thumb1, for example. 808 assert((SPAdjCount || SPAdj == 0) && 809 "Unbalanced call frame setup / destroy pairs?"); 810 } 811 } 812 813 /// scavengeFrameVirtualRegs - Replace all frame index virtual registers 814 /// with physical registers. Use the register scavenger to find an 815 /// appropriate register to use. 816 void PEI::scavengeFrameVirtualRegs(MachineFunction &Fn) { 817 // Run through the instructions and find any virtual registers. 818 for (MachineFunction::iterator BB = Fn.begin(), 819 E = Fn.end(); BB != E; ++BB) { 820 RS->enterBasicBlock(BB); 821 822 unsigned VirtReg = 0; 823 unsigned ScratchReg = 0; 824 int SPAdj = 0; 825 826 // The instruction stream may change in the loop, so check BB->end() 827 // directly. 828 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 829 MachineInstr *MI = I; 830 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 831 if (MI->getOperand(i).isReg()) { 832 MachineOperand &MO = MI->getOperand(i); 833 unsigned Reg = MO.getReg(); 834 if (Reg == 0) 835 continue; 836 if (!TargetRegisterInfo::isVirtualRegister(Reg)) 837 continue; 838 839 ++NumVirtualFrameRegs; 840 841 // Have we already allocated a scratch register for this virtual? 842 if (Reg != VirtReg) { 843 // When we first encounter a new virtual register, it 844 // must be a definition. 845 assert(MI->getOperand(i).isDef() && 846 "frame index virtual missing def!"); 847 // Scavenge a new scratch register 848 VirtReg = Reg; 849 const TargetRegisterClass *RC = Fn.getRegInfo().getRegClass(Reg); 850 ScratchReg = RS->scavengeRegister(RC, I, SPAdj); 851 ++NumScavengedRegs; 852 } 853 // Replace this reference to the virtual register with the 854 // scratch register. 855 assert (ScratchReg && "Missing scratch register!"); 856 MI->getOperand(i).setReg(ScratchReg); 857 858 } 859 } 860 RS->forward(I); 861 ++I; 862 } 863 } 864 } 865