ARMBaseRegisterInfo.cpp revision 391f228e7e00f62b79ad483b801f5f58f046b7ea
1//===- ARMBaseRegisterInfo.cpp - ARM Register Information -------*- C++ -*-===// 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 file contains the base ARM implementation of TargetRegisterInfo class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "ARM.h" 15#include "ARMAddressingModes.h" 16#include "ARMBaseInstrInfo.h" 17#include "ARMBaseRegisterInfo.h" 18#include "ARMInstrInfo.h" 19#include "ARMMachineFunctionInfo.h" 20#include "ARMSubtarget.h" 21#include "llvm/Constants.h" 22#include "llvm/DerivedTypes.h" 23#include "llvm/Function.h" 24#include "llvm/LLVMContext.h" 25#include "llvm/CodeGen/MachineConstantPool.h" 26#include "llvm/CodeGen/MachineFrameInfo.h" 27#include "llvm/CodeGen/MachineFunction.h" 28#include "llvm/CodeGen/MachineInstrBuilder.h" 29#include "llvm/CodeGen/MachineLocation.h" 30#include "llvm/CodeGen/MachineRegisterInfo.h" 31#include "llvm/CodeGen/RegisterScavenging.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/Support/ErrorHandling.h" 34#include "llvm/Support/raw_ostream.h" 35#include "llvm/Target/TargetFrameInfo.h" 36#include "llvm/Target/TargetMachine.h" 37#include "llvm/Target/TargetOptions.h" 38#include "llvm/ADT/BitVector.h" 39#include "llvm/ADT/SmallVector.h" 40#include "llvm/Support/CommandLine.h" 41 42namespace llvm { 43static cl::opt<bool> 44ForceAllBaseRegAlloc("arm-force-base-reg-alloc", cl::Hidden, cl::init(false), 45 cl::desc("Force use of virtual base registers for stack load/store")); 46static cl::opt<bool> 47EnableLocalStackAlloc("enable-local-stack-alloc", cl::init(true), cl::Hidden, 48 cl::desc("Enable pre-regalloc stack frame index allocation")); 49} 50 51using namespace llvm; 52 53static cl::opt<bool> 54EnableBasePointer("arm-use-base-pointer", cl::Hidden, cl::init(true), 55 cl::desc("Enable use of a base pointer for complex stack frames")); 56 57 58ARMBaseRegisterInfo::ARMBaseRegisterInfo(const ARMBaseInstrInfo &tii, 59 const ARMSubtarget &sti) 60 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP), 61 TII(tii), STI(sti), 62 FramePtr((STI.isTargetDarwin() || STI.isThumb()) ? ARM::R7 : ARM::R11), 63 BasePtr(ARM::R6) { 64} 65 66const unsigned* 67ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 68 static const unsigned CalleeSavedRegs[] = { 69 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8, 70 ARM::R7, ARM::R6, ARM::R5, ARM::R4, 71 72 ARM::D15, ARM::D14, ARM::D13, ARM::D12, 73 ARM::D11, ARM::D10, ARM::D9, ARM::D8, 74 0 75 }; 76 77 static const unsigned DarwinCalleeSavedRegs[] = { 78 // Darwin ABI deviates from ARM standard ABI. R9 is not a callee-saved 79 // register. 80 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4, 81 ARM::R11, ARM::R10, ARM::R8, 82 83 ARM::D15, ARM::D14, ARM::D13, ARM::D12, 84 ARM::D11, ARM::D10, ARM::D9, ARM::D8, 85 0 86 }; 87 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs; 88} 89 90BitVector ARMBaseRegisterInfo:: 91getReservedRegs(const MachineFunction &MF) const { 92 // FIXME: avoid re-calculating this everytime. 93 BitVector Reserved(getNumRegs()); 94 Reserved.set(ARM::SP); 95 Reserved.set(ARM::PC); 96 Reserved.set(ARM::FPSCR); 97 if (hasFP(MF)) 98 Reserved.set(FramePtr); 99 if (hasBasePointer(MF)) 100 Reserved.set(BasePtr); 101 // Some targets reserve R9. 102 if (STI.isR9Reserved()) 103 Reserved.set(ARM::R9); 104 return Reserved; 105} 106 107bool ARMBaseRegisterInfo::isReservedReg(const MachineFunction &MF, 108 unsigned Reg) const { 109 switch (Reg) { 110 default: break; 111 case ARM::SP: 112 case ARM::PC: 113 return true; 114 case ARM::R6: 115 if (hasBasePointer(MF)) 116 return true; 117 break; 118 case ARM::R7: 119 case ARM::R11: 120 if (FramePtr == Reg && hasFP(MF)) 121 return true; 122 break; 123 case ARM::R9: 124 return STI.isR9Reserved(); 125 } 126 127 return false; 128} 129 130const TargetRegisterClass * 131ARMBaseRegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A, 132 const TargetRegisterClass *B, 133 unsigned SubIdx) const { 134 switch (SubIdx) { 135 default: return 0; 136 case ARM::ssub_0: 137 case ARM::ssub_1: 138 case ARM::ssub_2: 139 case ARM::ssub_3: { 140 // S sub-registers. 141 if (A->getSize() == 8) { 142 if (B == &ARM::SPR_8RegClass) 143 return &ARM::DPR_8RegClass; 144 assert(B == &ARM::SPRRegClass && "Expecting SPR register class!"); 145 if (A == &ARM::DPR_8RegClass) 146 return A; 147 return &ARM::DPR_VFP2RegClass; 148 } 149 150 if (A->getSize() == 16) { 151 if (B == &ARM::SPR_8RegClass) 152 return &ARM::QPR_8RegClass; 153 return &ARM::QPR_VFP2RegClass; 154 } 155 156 if (A->getSize() == 32) { 157 if (B == &ARM::SPR_8RegClass) 158 return 0; // Do not allow coalescing! 159 return &ARM::QQPR_VFP2RegClass; 160 } 161 162 assert(A->getSize() == 64 && "Expecting a QQQQ register class!"); 163 return 0; // Do not allow coalescing! 164 } 165 case ARM::dsub_0: 166 case ARM::dsub_1: 167 case ARM::dsub_2: 168 case ARM::dsub_3: { 169 // D sub-registers. 170 if (A->getSize() == 16) { 171 if (B == &ARM::DPR_VFP2RegClass) 172 return &ARM::QPR_VFP2RegClass; 173 if (B == &ARM::DPR_8RegClass) 174 return 0; // Do not allow coalescing! 175 return A; 176 } 177 178 if (A->getSize() == 32) { 179 if (B == &ARM::DPR_VFP2RegClass) 180 return &ARM::QQPR_VFP2RegClass; 181 if (B == &ARM::DPR_8RegClass) 182 return 0; // Do not allow coalescing! 183 return A; 184 } 185 186 assert(A->getSize() == 64 && "Expecting a QQQQ register class!"); 187 if (B != &ARM::DPRRegClass) 188 return 0; // Do not allow coalescing! 189 return A; 190 } 191 case ARM::dsub_4: 192 case ARM::dsub_5: 193 case ARM::dsub_6: 194 case ARM::dsub_7: { 195 // D sub-registers of QQQQ registers. 196 if (A->getSize() == 64 && B == &ARM::DPRRegClass) 197 return A; 198 return 0; // Do not allow coalescing! 199 } 200 201 case ARM::qsub_0: 202 case ARM::qsub_1: { 203 // Q sub-registers. 204 if (A->getSize() == 32) { 205 if (B == &ARM::QPR_VFP2RegClass) 206 return &ARM::QQPR_VFP2RegClass; 207 if (B == &ARM::QPR_8RegClass) 208 return 0; // Do not allow coalescing! 209 return A; 210 } 211 212 assert(A->getSize() == 64 && "Expecting a QQQQ register class!"); 213 if (B == &ARM::QPRRegClass) 214 return A; 215 return 0; // Do not allow coalescing! 216 } 217 case ARM::qsub_2: 218 case ARM::qsub_3: { 219 // Q sub-registers of QQQQ registers. 220 if (A->getSize() == 64 && B == &ARM::QPRRegClass) 221 return A; 222 return 0; // Do not allow coalescing! 223 } 224 } 225 return 0; 226} 227 228bool 229ARMBaseRegisterInfo::canCombineSubRegIndices(const TargetRegisterClass *RC, 230 SmallVectorImpl<unsigned> &SubIndices, 231 unsigned &NewSubIdx) const { 232 233 unsigned Size = RC->getSize() * 8; 234 if (Size < 6) 235 return 0; 236 237 NewSubIdx = 0; // Whole register. 238 unsigned NumRegs = SubIndices.size(); 239 if (NumRegs == 8) { 240 // 8 D registers -> 1 QQQQ register. 241 return (Size == 512 && 242 SubIndices[0] == ARM::dsub_0 && 243 SubIndices[1] == ARM::dsub_1 && 244 SubIndices[2] == ARM::dsub_2 && 245 SubIndices[3] == ARM::dsub_3 && 246 SubIndices[4] == ARM::dsub_4 && 247 SubIndices[5] == ARM::dsub_5 && 248 SubIndices[6] == ARM::dsub_6 && 249 SubIndices[7] == ARM::dsub_7); 250 } else if (NumRegs == 4) { 251 if (SubIndices[0] == ARM::qsub_0) { 252 // 4 Q registers -> 1 QQQQ register. 253 return (Size == 512 && 254 SubIndices[1] == ARM::qsub_1 && 255 SubIndices[2] == ARM::qsub_2 && 256 SubIndices[3] == ARM::qsub_3); 257 } else if (SubIndices[0] == ARM::dsub_0) { 258 // 4 D registers -> 1 QQ register. 259 if (Size >= 256 && 260 SubIndices[1] == ARM::dsub_1 && 261 SubIndices[2] == ARM::dsub_2 && 262 SubIndices[3] == ARM::dsub_3) { 263 if (Size == 512) 264 NewSubIdx = ARM::qqsub_0; 265 return true; 266 } 267 } else if (SubIndices[0] == ARM::dsub_4) { 268 // 4 D registers -> 1 QQ register (2nd). 269 if (Size == 512 && 270 SubIndices[1] == ARM::dsub_5 && 271 SubIndices[2] == ARM::dsub_6 && 272 SubIndices[3] == ARM::dsub_7) { 273 NewSubIdx = ARM::qqsub_1; 274 return true; 275 } 276 } else if (SubIndices[0] == ARM::ssub_0) { 277 // 4 S registers -> 1 Q register. 278 if (Size >= 128 && 279 SubIndices[1] == ARM::ssub_1 && 280 SubIndices[2] == ARM::ssub_2 && 281 SubIndices[3] == ARM::ssub_3) { 282 if (Size >= 256) 283 NewSubIdx = ARM::qsub_0; 284 return true; 285 } 286 } 287 } else if (NumRegs == 2) { 288 if (SubIndices[0] == ARM::qsub_0) { 289 // 2 Q registers -> 1 QQ register. 290 if (Size >= 256 && SubIndices[1] == ARM::qsub_1) { 291 if (Size == 512) 292 NewSubIdx = ARM::qqsub_0; 293 return true; 294 } 295 } else if (SubIndices[0] == ARM::qsub_2) { 296 // 2 Q registers -> 1 QQ register (2nd). 297 if (Size == 512 && SubIndices[1] == ARM::qsub_3) { 298 NewSubIdx = ARM::qqsub_1; 299 return true; 300 } 301 } else if (SubIndices[0] == ARM::dsub_0) { 302 // 2 D registers -> 1 Q register. 303 if (Size >= 128 && SubIndices[1] == ARM::dsub_1) { 304 if (Size >= 256) 305 NewSubIdx = ARM::qsub_0; 306 return true; 307 } 308 } else if (SubIndices[0] == ARM::dsub_2) { 309 // 2 D registers -> 1 Q register (2nd). 310 if (Size >= 256 && SubIndices[1] == ARM::dsub_3) { 311 NewSubIdx = ARM::qsub_1; 312 return true; 313 } 314 } else if (SubIndices[0] == ARM::dsub_4) { 315 // 2 D registers -> 1 Q register (3rd). 316 if (Size == 512 && SubIndices[1] == ARM::dsub_5) { 317 NewSubIdx = ARM::qsub_2; 318 return true; 319 } 320 } else if (SubIndices[0] == ARM::dsub_6) { 321 // 2 D registers -> 1 Q register (3rd). 322 if (Size == 512 && SubIndices[1] == ARM::dsub_7) { 323 NewSubIdx = ARM::qsub_3; 324 return true; 325 } 326 } else if (SubIndices[0] == ARM::ssub_0) { 327 // 2 S registers -> 1 D register. 328 if (SubIndices[1] == ARM::ssub_1) { 329 if (Size >= 128) 330 NewSubIdx = ARM::dsub_0; 331 return true; 332 } 333 } else if (SubIndices[0] == ARM::ssub_2) { 334 // 2 S registers -> 1 D register (2nd). 335 if (Size >= 128 && SubIndices[1] == ARM::ssub_3) { 336 NewSubIdx = ARM::dsub_1; 337 return true; 338 } 339 } 340 } 341 return false; 342} 343 344 345const TargetRegisterClass * 346ARMBaseRegisterInfo::getPointerRegClass(unsigned Kind) const { 347 return ARM::GPRRegisterClass; 348} 349 350/// getAllocationOrder - Returns the register allocation order for a specified 351/// register class in the form of a pair of TargetRegisterClass iterators. 352std::pair<TargetRegisterClass::iterator,TargetRegisterClass::iterator> 353ARMBaseRegisterInfo::getAllocationOrder(const TargetRegisterClass *RC, 354 unsigned HintType, unsigned HintReg, 355 const MachineFunction &MF) const { 356 // Alternative register allocation orders when favoring even / odd registers 357 // of register pairs. 358 359 // No FP, R9 is available. 360 static const unsigned GPREven1[] = { 361 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, ARM::R10, 362 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, 363 ARM::R9, ARM::R11 364 }; 365 static const unsigned GPROdd1[] = { 366 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R9, ARM::R11, 367 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, 368 ARM::R8, ARM::R10 369 }; 370 371 // FP is R7, R9 is available. 372 static const unsigned GPREven2[] = { 373 ARM::R0, ARM::R2, ARM::R4, ARM::R8, ARM::R10, 374 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, 375 ARM::R9, ARM::R11 376 }; 377 static const unsigned GPROdd2[] = { 378 ARM::R1, ARM::R3, ARM::R5, ARM::R9, ARM::R11, 379 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, 380 ARM::R8, ARM::R10 381 }; 382 383 // FP is R11, R9 is available. 384 static const unsigned GPREven3[] = { 385 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, 386 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, 387 ARM::R9 388 }; 389 static const unsigned GPROdd3[] = { 390 ARM::R1, ARM::R3, ARM::R5, ARM::R6, ARM::R9, 391 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R7, 392 ARM::R8 393 }; 394 395 // No FP, R9 is not available. 396 static const unsigned GPREven4[] = { 397 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R10, 398 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8, 399 ARM::R11 400 }; 401 static const unsigned GPROdd4[] = { 402 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R11, 403 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8, 404 ARM::R10 405 }; 406 407 // FP is R7, R9 is not available. 408 static const unsigned GPREven5[] = { 409 ARM::R0, ARM::R2, ARM::R4, ARM::R10, 410 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, ARM::R8, 411 ARM::R11 412 }; 413 static const unsigned GPROdd5[] = { 414 ARM::R1, ARM::R3, ARM::R5, ARM::R11, 415 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8, 416 ARM::R10 417 }; 418 419 // FP is R11, R9 is not available. 420 static const unsigned GPREven6[] = { 421 ARM::R0, ARM::R2, ARM::R4, ARM::R6, 422 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8 423 }; 424 static const unsigned GPROdd6[] = { 425 ARM::R1, ARM::R3, ARM::R5, ARM::R7, 426 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8 427 }; 428 429 430 if (HintType == ARMRI::RegPairEven) { 431 if (isPhysicalRegister(HintReg) && getRegisterPairEven(HintReg, MF) == 0) 432 // It's no longer possible to fulfill this hint. Return the default 433 // allocation order. 434 return std::make_pair(RC->allocation_order_begin(MF), 435 RC->allocation_order_end(MF)); 436 437 if (!hasFP(MF)) { 438 if (!STI.isR9Reserved()) 439 return std::make_pair(GPREven1, 440 GPREven1 + (sizeof(GPREven1)/sizeof(unsigned))); 441 else 442 return std::make_pair(GPREven4, 443 GPREven4 + (sizeof(GPREven4)/sizeof(unsigned))); 444 } else if (FramePtr == ARM::R7) { 445 if (!STI.isR9Reserved()) 446 return std::make_pair(GPREven2, 447 GPREven2 + (sizeof(GPREven2)/sizeof(unsigned))); 448 else 449 return std::make_pair(GPREven5, 450 GPREven5 + (sizeof(GPREven5)/sizeof(unsigned))); 451 } else { // FramePtr == ARM::R11 452 if (!STI.isR9Reserved()) 453 return std::make_pair(GPREven3, 454 GPREven3 + (sizeof(GPREven3)/sizeof(unsigned))); 455 else 456 return std::make_pair(GPREven6, 457 GPREven6 + (sizeof(GPREven6)/sizeof(unsigned))); 458 } 459 } else if (HintType == ARMRI::RegPairOdd) { 460 if (isPhysicalRegister(HintReg) && getRegisterPairOdd(HintReg, MF) == 0) 461 // It's no longer possible to fulfill this hint. Return the default 462 // allocation order. 463 return std::make_pair(RC->allocation_order_begin(MF), 464 RC->allocation_order_end(MF)); 465 466 if (!hasFP(MF)) { 467 if (!STI.isR9Reserved()) 468 return std::make_pair(GPROdd1, 469 GPROdd1 + (sizeof(GPROdd1)/sizeof(unsigned))); 470 else 471 return std::make_pair(GPROdd4, 472 GPROdd4 + (sizeof(GPROdd4)/sizeof(unsigned))); 473 } else if (FramePtr == ARM::R7) { 474 if (!STI.isR9Reserved()) 475 return std::make_pair(GPROdd2, 476 GPROdd2 + (sizeof(GPROdd2)/sizeof(unsigned))); 477 else 478 return std::make_pair(GPROdd5, 479 GPROdd5 + (sizeof(GPROdd5)/sizeof(unsigned))); 480 } else { // FramePtr == ARM::R11 481 if (!STI.isR9Reserved()) 482 return std::make_pair(GPROdd3, 483 GPROdd3 + (sizeof(GPROdd3)/sizeof(unsigned))); 484 else 485 return std::make_pair(GPROdd6, 486 GPROdd6 + (sizeof(GPROdd6)/sizeof(unsigned))); 487 } 488 } 489 return std::make_pair(RC->allocation_order_begin(MF), 490 RC->allocation_order_end(MF)); 491} 492 493/// ResolveRegAllocHint - Resolves the specified register allocation hint 494/// to a physical register. Returns the physical register if it is successful. 495unsigned 496ARMBaseRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg, 497 const MachineFunction &MF) const { 498 if (Reg == 0 || !isPhysicalRegister(Reg)) 499 return 0; 500 if (Type == 0) 501 return Reg; 502 else if (Type == (unsigned)ARMRI::RegPairOdd) 503 // Odd register. 504 return getRegisterPairOdd(Reg, MF); 505 else if (Type == (unsigned)ARMRI::RegPairEven) 506 // Even register. 507 return getRegisterPairEven(Reg, MF); 508 return 0; 509} 510 511void 512ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg, 513 MachineFunction &MF) const { 514 MachineRegisterInfo *MRI = &MF.getRegInfo(); 515 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg); 516 if ((Hint.first == (unsigned)ARMRI::RegPairOdd || 517 Hint.first == (unsigned)ARMRI::RegPairEven) && 518 Hint.second && TargetRegisterInfo::isVirtualRegister(Hint.second)) { 519 // If 'Reg' is one of the even / odd register pair and it's now changed 520 // (e.g. coalesced) into a different register. The other register of the 521 // pair allocation hint must be updated to reflect the relationship 522 // change. 523 unsigned OtherReg = Hint.second; 524 Hint = MRI->getRegAllocationHint(OtherReg); 525 if (Hint.second == Reg) 526 // Make sure the pair has not already divorced. 527 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg); 528 } 529} 530 531/// hasFP - Return true if the specified function should have a dedicated frame 532/// pointer register. This is true if the function has variable sized allocas 533/// or if frame pointer elimination is disabled. 534/// 535bool ARMBaseRegisterInfo::hasFP(const MachineFunction &MF) const { 536 // Mac OS X requires FP not to be clobbered for backtracing purpose. 537 if (STI.isTargetDarwin()) 538 return true; 539 540 const MachineFrameInfo *MFI = MF.getFrameInfo(); 541 // Always eliminate non-leaf frame pointers. 542 return ((DisableFramePointerElim(MF) && MFI->hasCalls()) || 543 needsStackRealignment(MF) || 544 MFI->hasVarSizedObjects() || 545 MFI->isFrameAddressTaken()); 546} 547 548bool ARMBaseRegisterInfo::hasBasePointer(const MachineFunction &MF) const { 549 const MachineFrameInfo *MFI = MF.getFrameInfo(); 550 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 551 552 if (!EnableBasePointer) 553 return false; 554 555 if (needsStackRealignment(MF) && MFI->hasVarSizedObjects()) 556 return true; 557 558 // Thumb has trouble with negative offsets from the FP. Thumb2 has a limited 559 // negative range for ldr/str (255), and thumb1 is positive offsets only. 560 // It's going to be better to use the SP or Base Pointer instead. When there 561 // are variable sized objects, we can't reference off of the SP, so we 562 // reserve a Base Pointer. 563 if (AFI->isThumbFunction() && MFI->hasVarSizedObjects()) { 564 // Conservatively estimate whether the negative offset from the frame 565 // pointer will be sufficient to reach. If a function has a smallish 566 // frame, it's less likely to have lots of spills and callee saved 567 // space, so it's all more likely to be within range of the frame pointer. 568 // If it's wrong, the scavenger will still enable access to work, it just 569 // won't be optimal. 570 if (AFI->isThumb2Function() && MFI->getLocalFrameSize() < 128) 571 return false; 572 return true; 573 } 574 575 return false; 576} 577 578bool ARMBaseRegisterInfo::canRealignStack(const MachineFunction &MF) const { 579 const MachineFrameInfo *MFI = MF.getFrameInfo(); 580 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 581 // We can't realign the stack if: 582 // 1. Dynamic stack realignment is explicitly disabled, 583 // 2. This is a Thumb1 function (it's not useful, so we don't bother), or 584 // 3. There are VLAs in the function and the base pointer is disabled. 585 return (RealignStack && !AFI->isThumb1OnlyFunction() && 586 (!MFI->hasVarSizedObjects() || EnableBasePointer)); 587} 588 589bool ARMBaseRegisterInfo:: 590needsStackRealignment(const MachineFunction &MF) const { 591 const MachineFrameInfo *MFI = MF.getFrameInfo(); 592 const Function *F = MF.getFunction(); 593 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 594 bool requiresRealignment = ((MFI->getLocalFrameMaxAlign() > StackAlign) || 595 F->hasFnAttr(Attribute::StackAlignment)); 596 597 return requiresRealignment && canRealignStack(MF); 598} 599 600bool ARMBaseRegisterInfo:: 601cannotEliminateFrame(const MachineFunction &MF) const { 602 const MachineFrameInfo *MFI = MF.getFrameInfo(); 603 if (DisableFramePointerElim(MF) && MFI->adjustsStack()) 604 return true; 605 return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken() 606 || needsStackRealignment(MF); 607} 608 609/// estimateStackSize - Estimate and return the size of the frame. 610static unsigned estimateStackSize(MachineFunction &MF) { 611 const MachineFrameInfo *FFI = MF.getFrameInfo(); 612 int Offset = 0; 613 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) { 614 int FixedOff = -FFI->getObjectOffset(i); 615 if (FixedOff > Offset) Offset = FixedOff; 616 } 617 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) { 618 if (FFI->isDeadObjectIndex(i)) 619 continue; 620 Offset += FFI->getObjectSize(i); 621 unsigned Align = FFI->getObjectAlignment(i); 622 // Adjust to alignment boundary 623 Offset = (Offset+Align-1)/Align*Align; 624 } 625 return (unsigned)Offset; 626} 627 628/// estimateRSStackSizeLimit - Look at each instruction that references stack 629/// frames and return the stack size limit beyond which some of these 630/// instructions will require a scratch register during their expansion later. 631unsigned 632ARMBaseRegisterInfo::estimateRSStackSizeLimit(MachineFunction &MF) const { 633 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 634 unsigned Limit = (1 << 12) - 1; 635 for (MachineFunction::iterator BB = MF.begin(),E = MF.end(); BB != E; ++BB) { 636 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 637 I != E; ++I) { 638 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 639 if (!I->getOperand(i).isFI()) continue; 640 641 // When using ADDri to get the address of a stack object, 255 is the 642 // largest offset guaranteed to fit in the immediate offset. 643 if (I->getOpcode() == ARM::ADDri) { 644 Limit = std::min(Limit, (1U << 8) - 1); 645 break; 646 } 647 648 // Otherwise check the addressing mode. 649 switch (I->getDesc().TSFlags & ARMII::AddrModeMask) { 650 case ARMII::AddrMode3: 651 case ARMII::AddrModeT2_i8: 652 Limit = std::min(Limit, (1U << 8) - 1); 653 break; 654 case ARMII::AddrMode5: 655 case ARMII::AddrModeT2_i8s4: 656 Limit = std::min(Limit, ((1U << 8) - 1) * 4); 657 break; 658 case ARMII::AddrModeT2_i12: 659 // i12 supports only positive offset so these will be converted to 660 // i8 opcodes. See llvm::rewriteT2FrameIndex. 661 if (hasFP(MF) && AFI->hasStackFrame()) 662 Limit = std::min(Limit, (1U << 8) - 1); 663 break; 664 case ARMII::AddrMode4: 665 case ARMII::AddrMode6: 666 // Addressing modes 4 & 6 (load/store) instructions can't encode an 667 // immediate offset for stack references. 668 return 0; 669 default: 670 break; 671 } 672 break; // At most one FI per instruction 673 } 674 } 675 } 676 677 return Limit; 678} 679 680static unsigned GetFunctionSizeInBytes(const MachineFunction &MF, 681 const ARMBaseInstrInfo &TII) { 682 unsigned FnSize = 0; 683 for (MachineFunction::const_iterator MBBI = MF.begin(), E = MF.end(); 684 MBBI != E; ++MBBI) { 685 const MachineBasicBlock &MBB = *MBBI; 686 for (MachineBasicBlock::const_iterator I = MBB.begin(),E = MBB.end(); 687 I != E; ++I) 688 FnSize += TII.GetInstSizeInBytes(I); 689 } 690 return FnSize; 691} 692 693void 694ARMBaseRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, 695 RegScavenger *RS) const { 696 // This tells PEI to spill the FP as if it is any other callee-save register 697 // to take advantage the eliminateFrameIndex machinery. This also ensures it 698 // is spilled in the order specified by getCalleeSavedRegs() to make it easier 699 // to combine multiple loads / stores. 700 bool CanEliminateFrame = true; 701 bool CS1Spilled = false; 702 bool LRSpilled = false; 703 unsigned NumGPRSpills = 0; 704 SmallVector<unsigned, 4> UnspilledCS1GPRs; 705 SmallVector<unsigned, 4> UnspilledCS2GPRs; 706 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 707 MachineFrameInfo *MFI = MF.getFrameInfo(); 708 709 // Spill R4 if Thumb2 function requires stack realignment - it will be used as 710 // scratch register. 711 // FIXME: It will be better just to find spare register here. 712 if (needsStackRealignment(MF) && 713 AFI->isThumb2Function()) 714 MF.getRegInfo().setPhysRegUsed(ARM::R4); 715 716 // Spill LR if Thumb1 function uses variable length argument lists. 717 if (AFI->isThumb1OnlyFunction() && AFI->getVarArgsRegSaveSize() > 0) 718 MF.getRegInfo().setPhysRegUsed(ARM::LR); 719 720 // Spill the BasePtr if it's used. 721 if (hasBasePointer(MF)) 722 MF.getRegInfo().setPhysRegUsed(BasePtr); 723 724 // Don't spill FP if the frame can be eliminated. This is determined 725 // by scanning the callee-save registers to see if any is used. 726 const unsigned *CSRegs = getCalleeSavedRegs(); 727 for (unsigned i = 0; CSRegs[i]; ++i) { 728 unsigned Reg = CSRegs[i]; 729 bool Spilled = false; 730 if (MF.getRegInfo().isPhysRegUsed(Reg)) { 731 AFI->setCSRegisterIsSpilled(Reg); 732 Spilled = true; 733 CanEliminateFrame = false; 734 } else { 735 // Check alias registers too. 736 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) { 737 if (MF.getRegInfo().isPhysRegUsed(*Aliases)) { 738 Spilled = true; 739 CanEliminateFrame = false; 740 } 741 } 742 } 743 744 if (!ARM::GPRRegisterClass->contains(Reg)) 745 continue; 746 747 if (Spilled) { 748 NumGPRSpills++; 749 750 if (!STI.isTargetDarwin()) { 751 if (Reg == ARM::LR) 752 LRSpilled = true; 753 CS1Spilled = true; 754 continue; 755 } 756 757 // Keep track if LR and any of R4, R5, R6, and R7 is spilled. 758 switch (Reg) { 759 case ARM::LR: 760 LRSpilled = true; 761 // Fallthrough 762 case ARM::R4: 763 case ARM::R5: 764 case ARM::R6: 765 case ARM::R7: 766 CS1Spilled = true; 767 break; 768 default: 769 break; 770 } 771 } else { 772 if (!STI.isTargetDarwin()) { 773 UnspilledCS1GPRs.push_back(Reg); 774 continue; 775 } 776 777 switch (Reg) { 778 case ARM::R4: 779 case ARM::R5: 780 case ARM::R6: 781 case ARM::R7: 782 case ARM::LR: 783 UnspilledCS1GPRs.push_back(Reg); 784 break; 785 default: 786 UnspilledCS2GPRs.push_back(Reg); 787 break; 788 } 789 } 790 } 791 792 bool ForceLRSpill = false; 793 if (!LRSpilled && AFI->isThumb1OnlyFunction()) { 794 unsigned FnSize = GetFunctionSizeInBytes(MF, TII); 795 // Force LR to be spilled if the Thumb function size is > 2048. This enables 796 // use of BL to implement far jump. If it turns out that it's not needed 797 // then the branch fix up path will undo it. 798 if (FnSize >= (1 << 11)) { 799 CanEliminateFrame = false; 800 ForceLRSpill = true; 801 } 802 } 803 804 // If any of the stack slot references may be out of range of an immediate 805 // offset, make sure a register (or a spill slot) is available for the 806 // register scavenger. Note that if we're indexing off the frame pointer, the 807 // effective stack size is 4 bytes larger since the FP points to the stack 808 // slot of the previous FP. Also, if we have variable sized objects in the 809 // function, stack slot references will often be negative, and some of 810 // our instructions are positive-offset only, so conservatively consider 811 // that case to want a spill slot (or register) as well. Similarly, if 812 // the function adjusts the stack pointer during execution and the 813 // adjustments aren't already part of our stack size estimate, our offset 814 // calculations may be off, so be conservative. 815 // FIXME: We could add logic to be more precise about negative offsets 816 // and which instructions will need a scratch register for them. Is it 817 // worth the effort and added fragility? 818 bool BigStack = 819 (RS && 820 (estimateStackSize(MF) + ((hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >= 821 estimateRSStackSizeLimit(MF))) 822 || MFI->hasVarSizedObjects() 823 || (MFI->adjustsStack() && !canSimplifyCallFramePseudos(MF)); 824 825 bool ExtraCSSpill = false; 826 if (BigStack || !CanEliminateFrame || cannotEliminateFrame(MF)) { 827 AFI->setHasStackFrame(true); 828 829 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled. 830 // Spill LR as well so we can fold BX_RET to the registers restore (LDM). 831 if (!LRSpilled && CS1Spilled) { 832 MF.getRegInfo().setPhysRegUsed(ARM::LR); 833 AFI->setCSRegisterIsSpilled(ARM::LR); 834 NumGPRSpills++; 835 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(), 836 UnspilledCS1GPRs.end(), (unsigned)ARM::LR)); 837 ForceLRSpill = false; 838 ExtraCSSpill = true; 839 } 840 841 if (hasFP(MF)) { 842 MF.getRegInfo().setPhysRegUsed(FramePtr); 843 NumGPRSpills++; 844 } 845 846 // If stack and double are 8-byte aligned and we are spilling an odd number 847 // of GPRs, spill one extra callee save GPR so we won't have to pad between 848 // the integer and double callee save areas. 849 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 850 if (TargetAlign == 8 && (NumGPRSpills & 1)) { 851 if (CS1Spilled && !UnspilledCS1GPRs.empty()) { 852 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) { 853 unsigned Reg = UnspilledCS1GPRs[i]; 854 // Don't spill high register if the function is thumb1 855 if (!AFI->isThumb1OnlyFunction() || 856 isARMLowRegister(Reg) || Reg == ARM::LR) { 857 MF.getRegInfo().setPhysRegUsed(Reg); 858 AFI->setCSRegisterIsSpilled(Reg); 859 if (!isReservedReg(MF, Reg)) 860 ExtraCSSpill = true; 861 break; 862 } 863 } 864 } else if (!UnspilledCS2GPRs.empty() && 865 !AFI->isThumb1OnlyFunction()) { 866 unsigned Reg = UnspilledCS2GPRs.front(); 867 MF.getRegInfo().setPhysRegUsed(Reg); 868 AFI->setCSRegisterIsSpilled(Reg); 869 if (!isReservedReg(MF, Reg)) 870 ExtraCSSpill = true; 871 } 872 } 873 874 // Estimate if we might need to scavenge a register at some point in order 875 // to materialize a stack offset. If so, either spill one additional 876 // callee-saved register or reserve a special spill slot to facilitate 877 // register scavenging. Thumb1 needs a spill slot for stack pointer 878 // adjustments also, even when the frame itself is small. 879 if (BigStack && !ExtraCSSpill) { 880 // If any non-reserved CS register isn't spilled, just spill one or two 881 // extra. That should take care of it! 882 unsigned NumExtras = TargetAlign / 4; 883 SmallVector<unsigned, 2> Extras; 884 while (NumExtras && !UnspilledCS1GPRs.empty()) { 885 unsigned Reg = UnspilledCS1GPRs.back(); 886 UnspilledCS1GPRs.pop_back(); 887 if (!isReservedReg(MF, Reg) && 888 (!AFI->isThumb1OnlyFunction() || isARMLowRegister(Reg) || 889 Reg == ARM::LR)) { 890 Extras.push_back(Reg); 891 NumExtras--; 892 } 893 } 894 // For non-Thumb1 functions, also check for hi-reg CS registers 895 if (!AFI->isThumb1OnlyFunction()) { 896 while (NumExtras && !UnspilledCS2GPRs.empty()) { 897 unsigned Reg = UnspilledCS2GPRs.back(); 898 UnspilledCS2GPRs.pop_back(); 899 if (!isReservedReg(MF, Reg)) { 900 Extras.push_back(Reg); 901 NumExtras--; 902 } 903 } 904 } 905 if (Extras.size() && NumExtras == 0) { 906 for (unsigned i = 0, e = Extras.size(); i != e; ++i) { 907 MF.getRegInfo().setPhysRegUsed(Extras[i]); 908 AFI->setCSRegisterIsSpilled(Extras[i]); 909 } 910 } else if (!AFI->isThumb1OnlyFunction()) { 911 // note: Thumb1 functions spill to R12, not the stack. Reserve a slot 912 // closest to SP or frame pointer. 913 const TargetRegisterClass *RC = ARM::GPRRegisterClass; 914 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(), 915 RC->getAlignment(), 916 false)); 917 } 918 } 919 } 920 921 if (ForceLRSpill) { 922 MF.getRegInfo().setPhysRegUsed(ARM::LR); 923 AFI->setCSRegisterIsSpilled(ARM::LR); 924 AFI->setLRIsSpilledForFarJump(true); 925 } 926} 927 928unsigned ARMBaseRegisterInfo::getRARegister() const { 929 return ARM::LR; 930} 931 932unsigned 933ARMBaseRegisterInfo::getFrameRegister(const MachineFunction &MF) const { 934 if (hasFP(MF)) 935 return FramePtr; 936 return ARM::SP; 937} 938 939// Provide a base+offset reference to an FI slot for debug info. It's the 940// same as what we use for resolving the code-gen references for now. 941// FIXME: This can go wrong when references are SP-relative and simple call 942// frames aren't used. 943int 944ARMBaseRegisterInfo::getFrameIndexReference(const MachineFunction &MF, int FI, 945 unsigned &FrameReg) const { 946 return ResolveFrameIndexReference(MF, FI, FrameReg, 0); 947} 948 949int 950ARMBaseRegisterInfo::ResolveFrameIndexReference(const MachineFunction &MF, 951 int FI, 952 unsigned &FrameReg, 953 int SPAdj) const { 954 const MachineFrameInfo *MFI = MF.getFrameInfo(); 955 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 956 int Offset = MFI->getObjectOffset(FI) + MFI->getStackSize(); 957 int FPOffset = Offset - AFI->getFramePtrSpillOffset(); 958 bool isFixed = MFI->isFixedObjectIndex(FI); 959 960 FrameReg = ARM::SP; 961 Offset += SPAdj; 962 if (AFI->isGPRCalleeSavedArea1Frame(FI)) 963 return Offset - AFI->getGPRCalleeSavedArea1Offset(); 964 else if (AFI->isGPRCalleeSavedArea2Frame(FI)) 965 return Offset - AFI->getGPRCalleeSavedArea2Offset(); 966 else if (AFI->isDPRCalleeSavedAreaFrame(FI)) 967 return Offset - AFI->getDPRCalleeSavedAreaOffset(); 968 969 // When dynamically realigning the stack, use the frame pointer for 970 // parameters, and the stack/base pointer for locals. 971 if (needsStackRealignment(MF)) { 972 assert (hasFP(MF) && "dynamic stack realignment without a FP!"); 973 if (isFixed) { 974 FrameReg = getFrameRegister(MF); 975 Offset = FPOffset; 976 } else if (MFI->hasVarSizedObjects()) { 977 assert(hasBasePointer(MF) && 978 "VLAs and dynamic stack alignment, but missing base pointer!"); 979 FrameReg = BasePtr; 980 } 981 return Offset; 982 } 983 984 // If there is a frame pointer, use it when we can. 985 if (hasFP(MF) && AFI->hasStackFrame()) { 986 // Use frame pointer to reference fixed objects. Use it for locals if 987 // there are VLAs (and thus the SP isn't reliable as a base). 988 if (isFixed || (MFI->hasVarSizedObjects() && !hasBasePointer(MF))) { 989 FrameReg = getFrameRegister(MF); 990 return FPOffset; 991 } else if (MFI->hasVarSizedObjects()) { 992 assert(hasBasePointer(MF) && "missing base pointer!"); 993 // Try to use the frame pointer if we can, else use the base pointer 994 // since it's available. This is handy for the emergency spill slot, in 995 // particular. 996 if (AFI->isThumb2Function()) { 997 if (FPOffset >= -255 && FPOffset < 0) { 998 FrameReg = getFrameRegister(MF); 999 return FPOffset; 1000 } 1001 } else 1002 FrameReg = BasePtr; 1003 } else if (AFI->isThumb2Function()) { 1004 // In Thumb2 mode, the negative offset is very limited. Try to avoid 1005 // out of range references. 1006 if (FPOffset >= -255 && FPOffset < 0) { 1007 FrameReg = getFrameRegister(MF); 1008 return FPOffset; 1009 } 1010 } else if (Offset > (FPOffset < 0 ? -FPOffset : FPOffset)) { 1011 // Otherwise, use SP or FP, whichever is closer to the stack slot. 1012 FrameReg = getFrameRegister(MF); 1013 return FPOffset; 1014 } 1015 } 1016 // Use the base pointer if we have one. 1017 if (hasBasePointer(MF)) 1018 FrameReg = BasePtr; 1019 return Offset; 1020} 1021 1022int 1023ARMBaseRegisterInfo::getFrameIndexOffset(const MachineFunction &MF, 1024 int FI) const { 1025 unsigned FrameReg; 1026 return getFrameIndexReference(MF, FI, FrameReg); 1027} 1028 1029unsigned ARMBaseRegisterInfo::getEHExceptionRegister() const { 1030 llvm_unreachable("What is the exception register"); 1031 return 0; 1032} 1033 1034unsigned ARMBaseRegisterInfo::getEHHandlerRegister() const { 1035 llvm_unreachable("What is the exception handler register"); 1036 return 0; 1037} 1038 1039int ARMBaseRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const { 1040 return ARMGenRegisterInfo::getDwarfRegNumFull(RegNum, 0); 1041} 1042 1043unsigned ARMBaseRegisterInfo::getRegisterPairEven(unsigned Reg, 1044 const MachineFunction &MF) const { 1045 switch (Reg) { 1046 default: break; 1047 // Return 0 if either register of the pair is a special register. 1048 // So no R12, etc. 1049 case ARM::R1: 1050 return ARM::R0; 1051 case ARM::R3: 1052 return ARM::R2; 1053 case ARM::R5: 1054 return ARM::R4; 1055 case ARM::R7: 1056 return (isReservedReg(MF, ARM::R7) || isReservedReg(MF, ARM::R6)) 1057 ? 0 : ARM::R6; 1058 case ARM::R9: 1059 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R8; 1060 case ARM::R11: 1061 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R10; 1062 1063 case ARM::S1: 1064 return ARM::S0; 1065 case ARM::S3: 1066 return ARM::S2; 1067 case ARM::S5: 1068 return ARM::S4; 1069 case ARM::S7: 1070 return ARM::S6; 1071 case ARM::S9: 1072 return ARM::S8; 1073 case ARM::S11: 1074 return ARM::S10; 1075 case ARM::S13: 1076 return ARM::S12; 1077 case ARM::S15: 1078 return ARM::S14; 1079 case ARM::S17: 1080 return ARM::S16; 1081 case ARM::S19: 1082 return ARM::S18; 1083 case ARM::S21: 1084 return ARM::S20; 1085 case ARM::S23: 1086 return ARM::S22; 1087 case ARM::S25: 1088 return ARM::S24; 1089 case ARM::S27: 1090 return ARM::S26; 1091 case ARM::S29: 1092 return ARM::S28; 1093 case ARM::S31: 1094 return ARM::S30; 1095 1096 case ARM::D1: 1097 return ARM::D0; 1098 case ARM::D3: 1099 return ARM::D2; 1100 case ARM::D5: 1101 return ARM::D4; 1102 case ARM::D7: 1103 return ARM::D6; 1104 case ARM::D9: 1105 return ARM::D8; 1106 case ARM::D11: 1107 return ARM::D10; 1108 case ARM::D13: 1109 return ARM::D12; 1110 case ARM::D15: 1111 return ARM::D14; 1112 case ARM::D17: 1113 return ARM::D16; 1114 case ARM::D19: 1115 return ARM::D18; 1116 case ARM::D21: 1117 return ARM::D20; 1118 case ARM::D23: 1119 return ARM::D22; 1120 case ARM::D25: 1121 return ARM::D24; 1122 case ARM::D27: 1123 return ARM::D26; 1124 case ARM::D29: 1125 return ARM::D28; 1126 case ARM::D31: 1127 return ARM::D30; 1128 } 1129 1130 return 0; 1131} 1132 1133unsigned ARMBaseRegisterInfo::getRegisterPairOdd(unsigned Reg, 1134 const MachineFunction &MF) const { 1135 switch (Reg) { 1136 default: break; 1137 // Return 0 if either register of the pair is a special register. 1138 // So no R12, etc. 1139 case ARM::R0: 1140 return ARM::R1; 1141 case ARM::R2: 1142 return ARM::R3; 1143 case ARM::R4: 1144 return ARM::R5; 1145 case ARM::R6: 1146 return (isReservedReg(MF, ARM::R7) || isReservedReg(MF, ARM::R6)) 1147 ? 0 : ARM::R7; 1148 case ARM::R8: 1149 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R9; 1150 case ARM::R10: 1151 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R11; 1152 1153 case ARM::S0: 1154 return ARM::S1; 1155 case ARM::S2: 1156 return ARM::S3; 1157 case ARM::S4: 1158 return ARM::S5; 1159 case ARM::S6: 1160 return ARM::S7; 1161 case ARM::S8: 1162 return ARM::S9; 1163 case ARM::S10: 1164 return ARM::S11; 1165 case ARM::S12: 1166 return ARM::S13; 1167 case ARM::S14: 1168 return ARM::S15; 1169 case ARM::S16: 1170 return ARM::S17; 1171 case ARM::S18: 1172 return ARM::S19; 1173 case ARM::S20: 1174 return ARM::S21; 1175 case ARM::S22: 1176 return ARM::S23; 1177 case ARM::S24: 1178 return ARM::S25; 1179 case ARM::S26: 1180 return ARM::S27; 1181 case ARM::S28: 1182 return ARM::S29; 1183 case ARM::S30: 1184 return ARM::S31; 1185 1186 case ARM::D0: 1187 return ARM::D1; 1188 case ARM::D2: 1189 return ARM::D3; 1190 case ARM::D4: 1191 return ARM::D5; 1192 case ARM::D6: 1193 return ARM::D7; 1194 case ARM::D8: 1195 return ARM::D9; 1196 case ARM::D10: 1197 return ARM::D11; 1198 case ARM::D12: 1199 return ARM::D13; 1200 case ARM::D14: 1201 return ARM::D15; 1202 case ARM::D16: 1203 return ARM::D17; 1204 case ARM::D18: 1205 return ARM::D19; 1206 case ARM::D20: 1207 return ARM::D21; 1208 case ARM::D22: 1209 return ARM::D23; 1210 case ARM::D24: 1211 return ARM::D25; 1212 case ARM::D26: 1213 return ARM::D27; 1214 case ARM::D28: 1215 return ARM::D29; 1216 case ARM::D30: 1217 return ARM::D31; 1218 } 1219 1220 return 0; 1221} 1222 1223/// emitLoadConstPool - Emits a load from constpool to materialize the 1224/// specified immediate. 1225void ARMBaseRegisterInfo:: 1226emitLoadConstPool(MachineBasicBlock &MBB, 1227 MachineBasicBlock::iterator &MBBI, 1228 DebugLoc dl, 1229 unsigned DestReg, unsigned SubIdx, int Val, 1230 ARMCC::CondCodes Pred, 1231 unsigned PredReg) const { 1232 MachineFunction &MF = *MBB.getParent(); 1233 MachineConstantPool *ConstantPool = MF.getConstantPool(); 1234 const Constant *C = 1235 ConstantInt::get(Type::getInt32Ty(MF.getFunction()->getContext()), Val); 1236 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4); 1237 1238 BuildMI(MBB, MBBI, dl, TII.get(ARM::LDRcp)) 1239 .addReg(DestReg, getDefRegState(true), SubIdx) 1240 .addConstantPoolIndex(Idx) 1241 .addImm(0).addImm(Pred).addReg(PredReg); 1242} 1243 1244bool ARMBaseRegisterInfo:: 1245requiresRegisterScavenging(const MachineFunction &MF) const { 1246 return true; 1247} 1248 1249bool ARMBaseRegisterInfo:: 1250requiresFrameIndexScavenging(const MachineFunction &MF) const { 1251 return true; 1252} 1253 1254bool ARMBaseRegisterInfo:: 1255requiresVirtualBaseRegisters(const MachineFunction &MF) const { 1256 return EnableLocalStackAlloc; 1257} 1258 1259// hasReservedCallFrame - Under normal circumstances, when a frame pointer is 1260// not required, we reserve argument space for call sites in the function 1261// immediately on entry to the current function. This eliminates the need for 1262// add/sub sp brackets around call sites. Returns true if the call frame is 1263// included as part of the stack frame. 1264bool ARMBaseRegisterInfo:: 1265hasReservedCallFrame(const MachineFunction &MF) const { 1266 const MachineFrameInfo *FFI = MF.getFrameInfo(); 1267 unsigned CFSize = FFI->getMaxCallFrameSize(); 1268 // It's not always a good idea to include the call frame as part of the 1269 // stack frame. ARM (especially Thumb) has small immediate offset to 1270 // address the stack frame. So a large call frame can cause poor codegen 1271 // and may even makes it impossible to scavenge a register. 1272 if (CFSize >= ((1 << 12) - 1) / 2) // Half of imm12 1273 return false; 1274 1275 return !MF.getFrameInfo()->hasVarSizedObjects(); 1276} 1277 1278// canSimplifyCallFramePseudos - If there is a reserved call frame, the 1279// call frame pseudos can be simplified. Unlike most targets, having a FP 1280// is not sufficient here since we still may reference some objects via SP 1281// even when FP is available in Thumb2 mode. 1282bool ARMBaseRegisterInfo:: 1283canSimplifyCallFramePseudos(const MachineFunction &MF) const { 1284 return hasReservedCallFrame(MF) || MF.getFrameInfo()->hasVarSizedObjects(); 1285} 1286 1287static void 1288emitSPUpdate(bool isARM, 1289 MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, 1290 DebugLoc dl, const ARMBaseInstrInfo &TII, 1291 int NumBytes, 1292 ARMCC::CondCodes Pred = ARMCC::AL, unsigned PredReg = 0) { 1293 if (isARM) 1294 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes, 1295 Pred, PredReg, TII); 1296 else 1297 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes, 1298 Pred, PredReg, TII); 1299} 1300 1301 1302void ARMBaseRegisterInfo:: 1303eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, 1304 MachineBasicBlock::iterator I) const { 1305 if (!hasReservedCallFrame(MF)) { 1306 // If we have alloca, convert as follows: 1307 // ADJCALLSTACKDOWN -> sub, sp, sp, amount 1308 // ADJCALLSTACKUP -> add, sp, sp, amount 1309 MachineInstr *Old = I; 1310 DebugLoc dl = Old->getDebugLoc(); 1311 unsigned Amount = Old->getOperand(0).getImm(); 1312 if (Amount != 0) { 1313 // We need to keep the stack aligned properly. To do this, we round the 1314 // amount of space needed for the outgoing arguments up to the next 1315 // alignment boundary. 1316 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment(); 1317 Amount = (Amount+Align-1)/Align*Align; 1318 1319 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1320 assert(!AFI->isThumb1OnlyFunction() && 1321 "This eliminateCallFramePseudoInstr does not support Thumb1!"); 1322 bool isARM = !AFI->isThumbFunction(); 1323 1324 // Replace the pseudo instruction with a new instruction... 1325 unsigned Opc = Old->getOpcode(); 1326 int PIdx = Old->findFirstPredOperandIdx(); 1327 ARMCC::CondCodes Pred = (PIdx == -1) 1328 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(PIdx).getImm(); 1329 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) { 1330 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN. 1331 unsigned PredReg = Old->getOperand(2).getReg(); 1332 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, Pred, PredReg); 1333 } else { 1334 // Note: PredReg is operand 3 for ADJCALLSTACKUP. 1335 unsigned PredReg = Old->getOperand(3).getReg(); 1336 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP); 1337 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, Pred, PredReg); 1338 } 1339 } 1340 } 1341 MBB.erase(I); 1342} 1343 1344int64_t ARMBaseRegisterInfo:: 1345getFrameIndexInstrOffset(const MachineInstr *MI, int Idx) const { 1346 const TargetInstrDesc &Desc = MI->getDesc(); 1347 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); 1348 int64_t InstrOffs = 0;; 1349 int Scale = 1; 1350 unsigned ImmIdx = 0; 1351 switch (AddrMode) { 1352 case ARMII::AddrModeT2_i8: 1353 case ARMII::AddrModeT2_i12: 1354 case ARMII::AddrMode_i12: 1355 InstrOffs = MI->getOperand(Idx+1).getImm(); 1356 Scale = 1; 1357 break; 1358 case ARMII::AddrMode5: { 1359 // VFP address mode. 1360 const MachineOperand &OffOp = MI->getOperand(Idx+1); 1361 InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm()); 1362 if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub) 1363 InstrOffs = -InstrOffs; 1364 Scale = 4; 1365 break; 1366 } 1367 case ARMII::AddrMode2: { 1368 ImmIdx = Idx+2; 1369 InstrOffs = ARM_AM::getAM2Offset(MI->getOperand(ImmIdx).getImm()); 1370 if (ARM_AM::getAM2Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub) 1371 InstrOffs = -InstrOffs; 1372 break; 1373 } 1374 case ARMII::AddrMode3: { 1375 ImmIdx = Idx+2; 1376 InstrOffs = ARM_AM::getAM3Offset(MI->getOperand(ImmIdx).getImm()); 1377 if (ARM_AM::getAM3Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub) 1378 InstrOffs = -InstrOffs; 1379 break; 1380 } 1381 case ARMII::AddrModeT1_s: { 1382 ImmIdx = Idx+1; 1383 InstrOffs = MI->getOperand(ImmIdx).getImm(); 1384 Scale = 4; 1385 break; 1386 } 1387 default: 1388 llvm_unreachable("Unsupported addressing mode!"); 1389 break; 1390 } 1391 1392 return InstrOffs * Scale; 1393} 1394 1395/// needsFrameBaseReg - Returns true if the instruction's frame index 1396/// reference would be better served by a base register other than FP 1397/// or SP. Used by LocalStackFrameAllocation to determine which frame index 1398/// references it should create new base registers for. 1399bool ARMBaseRegisterInfo:: 1400needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const { 1401 for (unsigned i = 0; !MI->getOperand(i).isFI(); ++i) { 1402 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!"); 1403 } 1404 1405 // It's the load/store FI references that cause issues, as it can be difficult 1406 // to materialize the offset if it won't fit in the literal field. Estimate 1407 // based on the size of the local frame and some conservative assumptions 1408 // about the rest of the stack frame (note, this is pre-regalloc, so 1409 // we don't know everything for certain yet) whether this offset is likely 1410 // to be out of range of the immediate. Return true if so. 1411 1412 // We only generate virtual base registers for loads and stores, so 1413 // return false for everything else. 1414 unsigned Opc = MI->getOpcode(); 1415 switch (Opc) { 1416 case ARM::LDRi12: case ARM::LDRH: case ARM::LDRBi12: 1417 case ARM::STRi12: case ARM::STRH: case ARM::STRBi12: 1418 case ARM::t2LDRi12: case ARM::t2LDRi8: 1419 case ARM::t2STRi12: case ARM::t2STRi8: 1420 case ARM::VLDRS: case ARM::VLDRD: 1421 case ARM::VSTRS: case ARM::VSTRD: 1422 case ARM::tSTRspi: case ARM::tLDRspi: 1423 if (ForceAllBaseRegAlloc) 1424 return true; 1425 break; 1426 default: 1427 return false; 1428 } 1429 1430 // Without a virtual base register, if the function has variable sized 1431 // objects, all fixed-size local references will be via the frame pointer, 1432 // Approximate the offset and see if it's legal for the instruction. 1433 // Note that the incoming offset is based on the SP value at function entry, 1434 // so it'll be negative. 1435 MachineFunction &MF = *MI->getParent()->getParent(); 1436 MachineFrameInfo *MFI = MF.getFrameInfo(); 1437 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1438 1439 // Estimate an offset from the frame pointer. 1440 // Conservatively assume all callee-saved registers get pushed. R4-R6 1441 // will be earlier than the FP, so we ignore those. 1442 // R7, LR 1443 int64_t FPOffset = Offset - 8; 1444 // ARM and Thumb2 functions also need to consider R8-R11 and D8-D15 1445 if (!AFI->isThumbFunction() || !AFI->isThumb1OnlyFunction()) 1446 FPOffset -= 80; 1447 // Estimate an offset from the stack pointer. 1448 // The incoming offset is relating to the SP at the start of the function, 1449 // but when we access the local it'll be relative to the SP after local 1450 // allocation, so adjust our SP-relative offset by that allocation size. 1451 Offset = -Offset; 1452 Offset += MFI->getLocalFrameSize(); 1453 // Assume that we'll have at least some spill slots allocated. 1454 // FIXME: This is a total SWAG number. We should run some statistics 1455 // and pick a real one. 1456 Offset += 128; // 128 bytes of spill slots 1457 1458 // If there is a frame pointer, try using it. 1459 // The FP is only available if there is no dynamic realignment. We 1460 // don't know for sure yet whether we'll need that, so we guess based 1461 // on whether there are any local variables that would trigger it. 1462 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 1463 if (hasFP(MF) && 1464 !((MFI->getLocalFrameMaxAlign() > StackAlign) && canRealignStack(MF))) { 1465 if (isFrameOffsetLegal(MI, FPOffset)) 1466 return false; 1467 } 1468 // If we can reference via the stack pointer, try that. 1469 // FIXME: This (and the code that resolves the references) can be improved 1470 // to only disallow SP relative references in the live range of 1471 // the VLA(s). In practice, it's unclear how much difference that 1472 // would make, but it may be worth doing. 1473 if (!MFI->hasVarSizedObjects() && isFrameOffsetLegal(MI, Offset)) 1474 return false; 1475 1476 // The offset likely isn't legal, we want to allocate a virtual base register. 1477 return true; 1478} 1479 1480/// materializeFrameBaseRegister - Insert defining instruction(s) for 1481/// BaseReg to be a pointer to FrameIdx before insertion point I. 1482void ARMBaseRegisterInfo:: 1483materializeFrameBaseRegister(MachineBasicBlock::iterator I, unsigned BaseReg, 1484 int FrameIdx, int64_t Offset) const { 1485 ARMFunctionInfo *AFI = 1486 I->getParent()->getParent()->getInfo<ARMFunctionInfo>(); 1487 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : 1488 (AFI->isThumb1OnlyFunction() ? ARM::tADDrSPi : ARM::t2ADDri); 1489 1490 MachineInstrBuilder MIB = 1491 BuildMI(*I->getParent(), I, I->getDebugLoc(), TII.get(ADDriOpc), BaseReg) 1492 .addFrameIndex(FrameIdx).addImm(Offset); 1493 if (!AFI->isThumb1OnlyFunction()) 1494 AddDefaultCC(AddDefaultPred(MIB)); 1495} 1496 1497void 1498ARMBaseRegisterInfo::resolveFrameIndex(MachineBasicBlock::iterator I, 1499 unsigned BaseReg, int64_t Offset) const { 1500 MachineInstr &MI = *I; 1501 MachineBasicBlock &MBB = *MI.getParent(); 1502 MachineFunction &MF = *MBB.getParent(); 1503 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1504 int Off = Offset; // ARM doesn't need the general 64-bit offsets 1505 unsigned i = 0; 1506 1507 assert(!AFI->isThumb1OnlyFunction() && 1508 "This resolveFrameIndex does not support Thumb1!"); 1509 1510 while (!MI.getOperand(i).isFI()) { 1511 ++i; 1512 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); 1513 } 1514 bool Done = false; 1515 if (!AFI->isThumbFunction()) 1516 Done = rewriteARMFrameIndex(MI, i, BaseReg, Off, TII); 1517 else { 1518 assert(AFI->isThumb2Function()); 1519 Done = rewriteT2FrameIndex(MI, i, BaseReg, Off, TII); 1520 } 1521 assert (Done && "Unable to resolve frame index!"); 1522} 1523 1524bool ARMBaseRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, 1525 int64_t Offset) const { 1526 const TargetInstrDesc &Desc = MI->getDesc(); 1527 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); 1528 unsigned i = 0; 1529 1530 while (!MI->getOperand(i).isFI()) { 1531 ++i; 1532 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!"); 1533 } 1534 1535 // AddrMode4 and AddrMode6 cannot handle any offset. 1536 if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6) 1537 return Offset == 0; 1538 1539 unsigned NumBits = 0; 1540 unsigned Scale = 1; 1541 bool isSigned = true; 1542 switch (AddrMode) { 1543 case ARMII::AddrModeT2_i8: 1544 case ARMII::AddrModeT2_i12: 1545 // i8 supports only negative, and i12 supports only positive, so 1546 // based on Offset sign, consider the appropriate instruction 1547 Scale = 1; 1548 if (Offset < 0) { 1549 NumBits = 8; 1550 Offset = -Offset; 1551 } else { 1552 NumBits = 12; 1553 } 1554 break; 1555 case ARMII::AddrMode5: 1556 // VFP address mode. 1557 NumBits = 8; 1558 Scale = 4; 1559 break; 1560 case ARMII::AddrMode_i12: 1561 case ARMII::AddrMode2: 1562 NumBits = 12; 1563 break; 1564 case ARMII::AddrMode3: 1565 NumBits = 8; 1566 break; 1567 case ARMII::AddrModeT1_s: 1568 NumBits = 5; 1569 Scale = 4; 1570 isSigned = false; 1571 break; 1572 default: 1573 llvm_unreachable("Unsupported addressing mode!"); 1574 break; 1575 } 1576 1577 Offset += getFrameIndexInstrOffset(MI, i); 1578 // Make sure the offset is encodable for instructions that scale the 1579 // immediate. 1580 if ((Offset & (Scale-1)) != 0) 1581 return false; 1582 1583 if (isSigned && Offset < 0) 1584 Offset = -Offset; 1585 1586 unsigned Mask = (1 << NumBits) - 1; 1587 if ((unsigned)Offset <= Mask * Scale) 1588 return true; 1589 1590 return false; 1591} 1592 1593void 1594ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 1595 int SPAdj, RegScavenger *RS) const { 1596 unsigned i = 0; 1597 MachineInstr &MI = *II; 1598 MachineBasicBlock &MBB = *MI.getParent(); 1599 MachineFunction &MF = *MBB.getParent(); 1600 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1601 assert(!AFI->isThumb1OnlyFunction() && 1602 "This eliminateFrameIndex does not support Thumb1!"); 1603 1604 while (!MI.getOperand(i).isFI()) { 1605 ++i; 1606 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); 1607 } 1608 1609 int FrameIndex = MI.getOperand(i).getIndex(); 1610 unsigned FrameReg; 1611 1612 int Offset = ResolveFrameIndexReference(MF, FrameIndex, FrameReg, SPAdj); 1613 1614 // Special handling of dbg_value instructions. 1615 if (MI.isDebugValue()) { 1616 MI.getOperand(i). ChangeToRegister(FrameReg, false /*isDef*/); 1617 MI.getOperand(i+1).ChangeToImmediate(Offset); 1618 return; 1619 } 1620 1621 // Modify MI as necessary to handle as much of 'Offset' as possible 1622 bool Done = false; 1623 if (!AFI->isThumbFunction()) 1624 Done = rewriteARMFrameIndex(MI, i, FrameReg, Offset, TII); 1625 else { 1626 assert(AFI->isThumb2Function()); 1627 Done = rewriteT2FrameIndex(MI, i, FrameReg, Offset, TII); 1628 } 1629 if (Done) 1630 return; 1631 1632 // If we get here, the immediate doesn't fit into the instruction. We folded 1633 // as much as possible above, handle the rest, providing a register that is 1634 // SP+LargeImm. 1635 assert((Offset || 1636 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 || 1637 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) && 1638 "This code isn't needed if offset already handled!"); 1639 1640 unsigned ScratchReg = 0; 1641 int PIdx = MI.findFirstPredOperandIdx(); 1642 ARMCC::CondCodes Pred = (PIdx == -1) 1643 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm(); 1644 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg(); 1645 if (Offset == 0) 1646 // Must be addrmode4/6. 1647 MI.getOperand(i).ChangeToRegister(FrameReg, false, false, false); 1648 else { 1649 ScratchReg = MF.getRegInfo().createVirtualRegister(ARM::GPRRegisterClass); 1650 if (!AFI->isThumbFunction()) 1651 emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg, 1652 Offset, Pred, PredReg, TII); 1653 else { 1654 assert(AFI->isThumb2Function()); 1655 emitT2RegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg, 1656 Offset, Pred, PredReg, TII); 1657 } 1658 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true); 1659 } 1660} 1661 1662/// Move iterator past the next bunch of callee save load / store ops for 1663/// the particular spill area (1: integer area 1, 2: integer area 2, 1664/// 3: fp area, 0: don't care). 1665static void movePastCSLoadStoreOps(MachineBasicBlock &MBB, 1666 MachineBasicBlock::iterator &MBBI, 1667 int Opc1, int Opc2, unsigned Area, 1668 const ARMSubtarget &STI) { 1669 while (MBBI != MBB.end() && 1670 ((MBBI->getOpcode() == Opc1) || (MBBI->getOpcode() == Opc2))) { 1671 1672 if (Area == 3) { 1673 bool Done = false; 1674 unsigned Category = 0; 1675 switch (MBBI->getOperand(0).getReg()) { 1676 case ARM::R4: case ARM::R5: case ARM::R6: case ARM::R7: 1677 case ARM::LR: 1678 Category = 1; 1679 break; 1680 case ARM::R8: case ARM::R9: case ARM::R10: case ARM::R11: 1681 Category = STI.isTargetDarwin() ? 2 : 1; 1682 break; 1683 case ARM::D8: case ARM::D9: case ARM::D10: case ARM::D11: 1684 case ARM::D12: case ARM::D13: case ARM::D14: case ARM::D15: 1685 Category = 3; 1686 break; 1687 default: 1688 Done = true; 1689 break; 1690 } 1691 if (Done || Category != Area) 1692 break; 1693 } 1694 1695 ++MBBI; 1696 } 1697} 1698 1699void ARMBaseRegisterInfo:: 1700emitPrologue(MachineFunction &MF) const { 1701 MachineBasicBlock &MBB = MF.front(); 1702 MachineBasicBlock::iterator MBBI = MBB.begin(); 1703 MachineFrameInfo *MFI = MF.getFrameInfo(); 1704 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1705 assert(!AFI->isThumb1OnlyFunction() && 1706 "This emitPrologue does not support Thumb1!"); 1707 bool isARM = !AFI->isThumbFunction(); 1708 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize(); 1709 unsigned NumBytes = MFI->getStackSize(); 1710 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 1711 DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1712 1713 // Determine the sizes of each callee-save spill areas and record which frame 1714 // belongs to which callee-save spill areas. 1715 unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0; 1716 int FramePtrSpillFI = 0; 1717 1718 // Allocate the vararg register save area. This is not counted in NumBytes. 1719 if (VARegSaveSize) 1720 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -VARegSaveSize); 1721 1722 if (!AFI->hasStackFrame()) { 1723 if (NumBytes != 0) 1724 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes); 1725 return; 1726 } 1727 1728 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 1729 unsigned Reg = CSI[i].getReg(); 1730 int FI = CSI[i].getFrameIdx(); 1731 switch (Reg) { 1732 case ARM::R4: 1733 case ARM::R5: 1734 case ARM::R6: 1735 case ARM::R7: 1736 case ARM::LR: 1737 if (Reg == FramePtr) 1738 FramePtrSpillFI = FI; 1739 AFI->addGPRCalleeSavedArea1Frame(FI); 1740 GPRCS1Size += 4; 1741 break; 1742 case ARM::R8: 1743 case ARM::R9: 1744 case ARM::R10: 1745 case ARM::R11: 1746 if (Reg == FramePtr) 1747 FramePtrSpillFI = FI; 1748 if (STI.isTargetDarwin()) { 1749 AFI->addGPRCalleeSavedArea2Frame(FI); 1750 GPRCS2Size += 4; 1751 } else { 1752 AFI->addGPRCalleeSavedArea1Frame(FI); 1753 GPRCS1Size += 4; 1754 } 1755 break; 1756 default: 1757 AFI->addDPRCalleeSavedAreaFrame(FI); 1758 DPRCSSize += 8; 1759 } 1760 } 1761 1762 movePastCSLoadStoreOps(MBB, MBBI, ARM::tPUSH, 0, 1, STI); 1763 1764 // Set FP to point to the stack slot that contains the previous FP. 1765 // For Darwin, FP is R7, which has now been stored in spill area 1. 1766 // Otherwise, if this is not Darwin, all the callee-saved registers go 1767 // into spill area 1, including the FP in R11. In either case, it is 1768 // now safe to emit this assignment. 1769 bool HasFP = hasFP(MF); 1770 if (HasFP) { 1771 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : ARM::t2ADDri; 1772 MachineInstrBuilder MIB = 1773 BuildMI(MBB, MBBI, dl, TII.get(ADDriOpc), FramePtr) 1774 .addFrameIndex(FramePtrSpillFI).addImm(0); 1775 AddDefaultCC(AddDefaultPred(MIB)); 1776 } 1777 1778 // Build the new SUBri to adjust SP for integer callee-save spill area 2. 1779 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS2Size); 1780 1781 // Build the new SUBri to adjust SP for FP callee-save spill area. 1782 movePastCSLoadStoreOps(MBB, MBBI, ARM::tPUSH, 0, 2, STI); 1783 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -DPRCSSize); 1784 1785 // Determine starting offsets of spill areas. 1786 unsigned DPRCSOffset = NumBytes - (GPRCS1Size + GPRCS2Size + DPRCSSize); 1787 unsigned GPRCS2Offset = DPRCSOffset + DPRCSSize; 1788 unsigned GPRCS1Offset = GPRCS2Offset + GPRCS2Size; 1789 if (HasFP) 1790 AFI->setFramePtrSpillOffset(MFI->getObjectOffset(FramePtrSpillFI) + 1791 NumBytes); 1792 AFI->setGPRCalleeSavedArea1Offset(GPRCS1Offset); 1793 AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset); 1794 AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset); 1795 1796 movePastCSLoadStoreOps(MBB, MBBI, ARM::VSTRD, 0, 3, STI); 1797 NumBytes = DPRCSOffset; 1798 if (NumBytes) { 1799 // Adjust SP after all the callee-save spills. 1800 emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes); 1801 if (HasFP) 1802 AFI->setShouldRestoreSPFromFP(true); 1803 } 1804 1805 if (STI.isTargetELF() && hasFP(MF)) { 1806 MFI->setOffsetAdjustment(MFI->getOffsetAdjustment() - 1807 AFI->getFramePtrSpillOffset()); 1808 AFI->setShouldRestoreSPFromFP(true); 1809 } 1810 1811 AFI->setGPRCalleeSavedArea1Size(GPRCS1Size); 1812 AFI->setGPRCalleeSavedArea2Size(GPRCS2Size); 1813 AFI->setDPRCalleeSavedAreaSize(DPRCSSize); 1814 1815 // If we need dynamic stack realignment, do it here. Be paranoid and make 1816 // sure if we also have VLAs, we have a base pointer for frame access. 1817 if (needsStackRealignment(MF)) { 1818 unsigned MaxAlign = MFI->getMaxAlignment(); 1819 assert (!AFI->isThumb1OnlyFunction()); 1820 if (!AFI->isThumbFunction()) { 1821 // Emit bic sp, sp, MaxAlign 1822 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, 1823 TII.get(ARM::BICri), ARM::SP) 1824 .addReg(ARM::SP, RegState::Kill) 1825 .addImm(MaxAlign-1))); 1826 } else { 1827 // We cannot use sp as source/dest register here, thus we're emitting the 1828 // following sequence: 1829 // mov r4, sp 1830 // bic r4, r4, MaxAlign 1831 // mov sp, r4 1832 // FIXME: It will be better just to find spare register here. 1833 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2tgpr), ARM::R4) 1834 .addReg(ARM::SP, RegState::Kill); 1835 AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, 1836 TII.get(ARM::t2BICri), ARM::R4) 1837 .addReg(ARM::R4, RegState::Kill) 1838 .addImm(MaxAlign-1))); 1839 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVtgpr2gpr), ARM::SP) 1840 .addReg(ARM::R4, RegState::Kill); 1841 } 1842 1843 AFI->setShouldRestoreSPFromFP(true); 1844 } 1845 1846 // If we need a base pointer, set it up here. It's whatever the value 1847 // of the stack pointer is at this point. Any variable size objects 1848 // will be allocated after this, so we can still use the base pointer 1849 // to reference locals. 1850 if (hasBasePointer(MF)) { 1851 if (isARM) 1852 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), BasePtr) 1853 .addReg(ARM::SP) 1854 .addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); 1855 else 1856 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2gpr), BasePtr) 1857 .addReg(ARM::SP); 1858 } 1859 1860 // If the frame has variable sized objects then the epilogue must restore 1861 // the sp from fp. 1862 if (!AFI->shouldRestoreSPFromFP() && MFI->hasVarSizedObjects()) 1863 AFI->setShouldRestoreSPFromFP(true); 1864} 1865 1866static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) { 1867 for (unsigned i = 0; CSRegs[i]; ++i) 1868 if (Reg == CSRegs[i]) 1869 return true; 1870 return false; 1871} 1872 1873static bool isCSRestore(MachineInstr *MI, 1874 const ARMBaseInstrInfo &TII, 1875 const unsigned *CSRegs) { 1876 1877 // Integer spill area is handled with pop. 1878 if (MI->getOpcode() == ARM::tRestore || 1879 MI->getOpcode() == ARM::tPOP) { 1880 // The first two operands are predicates. The last two are 1881 // imp-def and imp-use of SP. Check everything in between. 1882 for (int i = 2, e = MI->getNumOperands() - 2; i != e; ++i) 1883 if (!isCalleeSavedRegister(MI->getOperand(i).getReg(), CSRegs)) 1884 return false; 1885 return true; 1886 } 1887 1888 // Or if this is a fp reg spill. 1889 if (MI->getOpcode() == (int)ARM::VLDRD && 1890 MI->getOperand(1).isFI() && 1891 isCalleeSavedRegister(MI->getOperand(0).getReg(), CSRegs)) 1892 return true; 1893 1894 return false; 1895} 1896 1897void ARMBaseRegisterInfo:: 1898emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const { 1899 MachineBasicBlock::iterator MBBI = prior(MBB.end()); 1900 assert(MBBI->getDesc().isReturn() && 1901 "Can only insert epilog into returning blocks"); 1902 unsigned RetOpcode = MBBI->getOpcode(); 1903 DebugLoc dl = MBBI->getDebugLoc(); 1904 MachineFrameInfo *MFI = MF.getFrameInfo(); 1905 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1906 assert(!AFI->isThumb1OnlyFunction() && 1907 "This emitEpilogue does not support Thumb1!"); 1908 bool isARM = !AFI->isThumbFunction(); 1909 1910 unsigned VARegSaveSize = AFI->getVarArgsRegSaveSize(); 1911 int NumBytes = (int)MFI->getStackSize(); 1912 1913 if (!AFI->hasStackFrame()) { 1914 if (NumBytes != 0) 1915 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes); 1916 } else { 1917 // Unwind MBBI to point to first LDR / VLDRD. 1918 const unsigned *CSRegs = getCalleeSavedRegs(); 1919 if (MBBI != MBB.begin()) { 1920 do 1921 --MBBI; 1922 while (MBBI != MBB.begin() && isCSRestore(MBBI, TII, CSRegs)); 1923 if (!isCSRestore(MBBI, TII, CSRegs)) 1924 ++MBBI; 1925 } 1926 1927 // Move SP to start of FP callee save spill area. 1928 NumBytes -= (AFI->getGPRCalleeSavedArea1Size() + 1929 AFI->getGPRCalleeSavedArea2Size() + 1930 AFI->getDPRCalleeSavedAreaSize()); 1931 1932 // Reset SP based on frame pointer only if the stack frame extends beyond 1933 // frame pointer stack slot or target is ELF and the function has FP. 1934 if (AFI->shouldRestoreSPFromFP()) { 1935 NumBytes = AFI->getFramePtrSpillOffset() - NumBytes; 1936 if (NumBytes) { 1937 if (isARM) 1938 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes, 1939 ARMCC::AL, 0, TII); 1940 else 1941 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, FramePtr, -NumBytes, 1942 ARMCC::AL, 0, TII); 1943 } else { 1944 // Thumb2 or ARM. 1945 if (isARM) 1946 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), ARM::SP) 1947 .addReg(FramePtr).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); 1948 else 1949 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2gpr), ARM::SP) 1950 .addReg(FramePtr); 1951 } 1952 } else if (NumBytes) 1953 emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes); 1954 1955 // Move SP to start of integer callee save spill area 2. 1956 movePastCSLoadStoreOps(MBB, MBBI, ARM::VLDRD, 0, 3, STI); 1957 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getDPRCalleeSavedAreaSize()); 1958 1959 // Move SP to start of integer callee save spill area 1. 1960 movePastCSLoadStoreOps(MBB, MBBI, ARM::tPOP, 0, 2, STI); 1961 emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getGPRCalleeSavedArea2Size()); 1962 } 1963 1964 if (RetOpcode == ARM::TCRETURNdi || RetOpcode == ARM::TCRETURNdiND || 1965 RetOpcode == ARM::TCRETURNri || RetOpcode == ARM::TCRETURNriND) { 1966 // Tail call return: adjust the stack pointer and jump to callee. 1967 MBBI = prior(MBB.end()); 1968 MachineOperand &JumpTarget = MBBI->getOperand(0); 1969 1970 // Jump to label or value in register. 1971 if (RetOpcode == ARM::TCRETURNdi) { 1972 BuildMI(MBB, MBBI, dl, 1973 TII.get(STI.isThumb() ? ARM::TAILJMPdt : ARM::TAILJMPd)). 1974 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(), 1975 JumpTarget.getTargetFlags()); 1976 } else if (RetOpcode == ARM::TCRETURNdiND) { 1977 BuildMI(MBB, MBBI, dl, 1978 TII.get(STI.isThumb() ? ARM::TAILJMPdNDt : ARM::TAILJMPdND)). 1979 addGlobalAddress(JumpTarget.getGlobal(), JumpTarget.getOffset(), 1980 JumpTarget.getTargetFlags()); 1981 } else if (RetOpcode == ARM::TCRETURNri) { 1982 BuildMI(MBB, MBBI, dl, TII.get(ARM::TAILJMPr)). 1983 addReg(JumpTarget.getReg(), RegState::Kill); 1984 } else if (RetOpcode == ARM::TCRETURNriND) { 1985 BuildMI(MBB, MBBI, dl, TII.get(ARM::TAILJMPrND)). 1986 addReg(JumpTarget.getReg(), RegState::Kill); 1987 } 1988 1989 MachineInstr *NewMI = prior(MBBI); 1990 for (unsigned i = 1, e = MBBI->getNumOperands(); i != e; ++i) 1991 NewMI->addOperand(MBBI->getOperand(i)); 1992 1993 // Delete the pseudo instruction TCRETURN. 1994 MBB.erase(MBBI); 1995 } 1996 1997 if (VARegSaveSize) 1998 emitSPUpdate(isARM, MBB, MBBI, dl, TII, VARegSaveSize); 1999} 2000 2001#include "ARMGenRegisterInfo.inc" 2002