ARMBaseRegisterInfo.cpp revision d0c38176690e9602a93a20a43f1bd084564a8116
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 42using namespace llvm; 43 44static cl::opt<bool> 45ForceAllBaseRegAlloc("arm-force-base-reg-alloc", cl::Hidden, cl::init(false), 46 cl::desc("Force use of virtual base registers for stack load/store")); 47static cl::opt<bool> 48EnableLocalStackAlloc("enable-local-stack-alloc", cl::init(true), cl::Hidden, 49 cl::desc("Enable pre-regalloc stack frame index allocation")); 50static cl::opt<bool> 51EnableBasePointer("arm-use-base-pointer", cl::Hidden, cl::init(true), 52 cl::desc("Enable use of a base pointer for complex stack frames")); 53 54ARMBaseRegisterInfo::ARMBaseRegisterInfo(const ARMBaseInstrInfo &tii, 55 const ARMSubtarget &sti) 56 : ARMGenRegisterInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP), 57 TII(tii), STI(sti), 58 FramePtr((STI.isTargetDarwin() || STI.isThumb()) ? ARM::R7 : ARM::R11), 59 BasePtr(ARM::R6) { 60} 61 62const unsigned* 63ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { 64 static const unsigned CalleeSavedRegs[] = { 65 ARM::LR, ARM::R11, ARM::R10, ARM::R9, ARM::R8, 66 ARM::R7, ARM::R6, ARM::R5, ARM::R4, 67 68 ARM::D15, ARM::D14, ARM::D13, ARM::D12, 69 ARM::D11, ARM::D10, ARM::D9, ARM::D8, 70 0 71 }; 72 73 static const unsigned DarwinCalleeSavedRegs[] = { 74 // Darwin ABI deviates from ARM standard ABI. R9 is not a callee-saved 75 // register. 76 ARM::LR, ARM::R7, ARM::R6, ARM::R5, ARM::R4, 77 ARM::R11, ARM::R10, ARM::R8, 78 79 ARM::D15, ARM::D14, ARM::D13, ARM::D12, 80 ARM::D11, ARM::D10, ARM::D9, ARM::D8, 81 0 82 }; 83 return STI.isTargetDarwin() ? DarwinCalleeSavedRegs : CalleeSavedRegs; 84} 85 86BitVector ARMBaseRegisterInfo:: 87getReservedRegs(const MachineFunction &MF) const { 88 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 89 90 // FIXME: avoid re-calculating this everytime. 91 BitVector Reserved(getNumRegs()); 92 Reserved.set(ARM::SP); 93 Reserved.set(ARM::PC); 94 Reserved.set(ARM::FPSCR); 95 if (TFI->hasFP(MF)) 96 Reserved.set(FramePtr); 97 if (hasBasePointer(MF)) 98 Reserved.set(BasePtr); 99 // Some targets reserve R9. 100 if (STI.isR9Reserved()) 101 Reserved.set(ARM::R9); 102 return Reserved; 103} 104 105bool ARMBaseRegisterInfo::isReservedReg(const MachineFunction &MF, 106 unsigned Reg) const { 107 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 108 109 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 && TFI->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 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 357 // Alternative register allocation orders when favoring even / odd registers 358 // of register pairs. 359 360 // No FP, R9 is available. 361 static const unsigned GPREven1[] = { 362 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, ARM::R10, 363 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, 364 ARM::R9, ARM::R11 365 }; 366 static const unsigned GPROdd1[] = { 367 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R9, ARM::R11, 368 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, 369 ARM::R8, ARM::R10 370 }; 371 372 // FP is R7, R9 is available. 373 static const unsigned GPREven2[] = { 374 ARM::R0, ARM::R2, ARM::R4, ARM::R8, ARM::R10, 375 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, 376 ARM::R9, ARM::R11 377 }; 378 static const unsigned GPROdd2[] = { 379 ARM::R1, ARM::R3, ARM::R5, ARM::R9, ARM::R11, 380 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, 381 ARM::R8, ARM::R10 382 }; 383 384 // FP is R11, R9 is available. 385 static const unsigned GPREven3[] = { 386 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R8, 387 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, 388 ARM::R9 389 }; 390 static const unsigned GPROdd3[] = { 391 ARM::R1, ARM::R3, ARM::R5, ARM::R6, ARM::R9, 392 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R7, 393 ARM::R8 394 }; 395 396 // No FP, R9 is not available. 397 static const unsigned GPREven4[] = { 398 ARM::R0, ARM::R2, ARM::R4, ARM::R6, ARM::R10, 399 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8, 400 ARM::R11 401 }; 402 static const unsigned GPROdd4[] = { 403 ARM::R1, ARM::R3, ARM::R5, ARM::R7, ARM::R11, 404 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8, 405 ARM::R10 406 }; 407 408 // FP is R7, R9 is not available. 409 static const unsigned GPREven5[] = { 410 ARM::R0, ARM::R2, ARM::R4, ARM::R10, 411 ARM::R1, ARM::R3, ARM::R12,ARM::LR, ARM::R5, ARM::R6, ARM::R8, 412 ARM::R11 413 }; 414 static const unsigned GPROdd5[] = { 415 ARM::R1, ARM::R3, ARM::R5, ARM::R11, 416 ARM::R0, ARM::R2, ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8, 417 ARM::R10 418 }; 419 420 // FP is R11, R9 is not available. 421 static const unsigned GPREven6[] = { 422 ARM::R0, ARM::R2, ARM::R4, ARM::R6, 423 ARM::R1, ARM::R3, ARM::R10,ARM::R12,ARM::LR, ARM::R5, ARM::R7, ARM::R8 424 }; 425 static const unsigned GPROdd6[] = { 426 ARM::R1, ARM::R3, ARM::R5, ARM::R7, 427 ARM::R0, ARM::R2, ARM::R10,ARM::R12,ARM::LR, ARM::R4, ARM::R6, ARM::R8 428 }; 429 430 431 if (HintType == ARMRI::RegPairEven) { 432 if (isPhysicalRegister(HintReg) && getRegisterPairEven(HintReg, MF) == 0) 433 // It's no longer possible to fulfill this hint. Return the default 434 // allocation order. 435 return std::make_pair(RC->allocation_order_begin(MF), 436 RC->allocation_order_end(MF)); 437 438 if (!TFI->hasFP(MF)) { 439 if (!STI.isR9Reserved()) 440 return std::make_pair(GPREven1, 441 GPREven1 + (sizeof(GPREven1)/sizeof(unsigned))); 442 else 443 return std::make_pair(GPREven4, 444 GPREven4 + (sizeof(GPREven4)/sizeof(unsigned))); 445 } else if (FramePtr == ARM::R7) { 446 if (!STI.isR9Reserved()) 447 return std::make_pair(GPREven2, 448 GPREven2 + (sizeof(GPREven2)/sizeof(unsigned))); 449 else 450 return std::make_pair(GPREven5, 451 GPREven5 + (sizeof(GPREven5)/sizeof(unsigned))); 452 } else { // FramePtr == ARM::R11 453 if (!STI.isR9Reserved()) 454 return std::make_pair(GPREven3, 455 GPREven3 + (sizeof(GPREven3)/sizeof(unsigned))); 456 else 457 return std::make_pair(GPREven6, 458 GPREven6 + (sizeof(GPREven6)/sizeof(unsigned))); 459 } 460 } else if (HintType == ARMRI::RegPairOdd) { 461 if (isPhysicalRegister(HintReg) && getRegisterPairOdd(HintReg, MF) == 0) 462 // It's no longer possible to fulfill this hint. Return the default 463 // allocation order. 464 return std::make_pair(RC->allocation_order_begin(MF), 465 RC->allocation_order_end(MF)); 466 467 if (!TFI->hasFP(MF)) { 468 if (!STI.isR9Reserved()) 469 return std::make_pair(GPROdd1, 470 GPROdd1 + (sizeof(GPROdd1)/sizeof(unsigned))); 471 else 472 return std::make_pair(GPROdd4, 473 GPROdd4 + (sizeof(GPROdd4)/sizeof(unsigned))); 474 } else if (FramePtr == ARM::R7) { 475 if (!STI.isR9Reserved()) 476 return std::make_pair(GPROdd2, 477 GPROdd2 + (sizeof(GPROdd2)/sizeof(unsigned))); 478 else 479 return std::make_pair(GPROdd5, 480 GPROdd5 + (sizeof(GPROdd5)/sizeof(unsigned))); 481 } else { // FramePtr == ARM::R11 482 if (!STI.isR9Reserved()) 483 return std::make_pair(GPROdd3, 484 GPROdd3 + (sizeof(GPROdd3)/sizeof(unsigned))); 485 else 486 return std::make_pair(GPROdd6, 487 GPROdd6 + (sizeof(GPROdd6)/sizeof(unsigned))); 488 } 489 } 490 return std::make_pair(RC->allocation_order_begin(MF), 491 RC->allocation_order_end(MF)); 492} 493 494/// ResolveRegAllocHint - Resolves the specified register allocation hint 495/// to a physical register. Returns the physical register if it is successful. 496unsigned 497ARMBaseRegisterInfo::ResolveRegAllocHint(unsigned Type, unsigned Reg, 498 const MachineFunction &MF) const { 499 if (Reg == 0 || !isPhysicalRegister(Reg)) 500 return 0; 501 if (Type == 0) 502 return Reg; 503 else if (Type == (unsigned)ARMRI::RegPairOdd) 504 // Odd register. 505 return getRegisterPairOdd(Reg, MF); 506 else if (Type == (unsigned)ARMRI::RegPairEven) 507 // Even register. 508 return getRegisterPairEven(Reg, MF); 509 return 0; 510} 511 512void 513ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg, 514 MachineFunction &MF) const { 515 MachineRegisterInfo *MRI = &MF.getRegInfo(); 516 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg); 517 if ((Hint.first == (unsigned)ARMRI::RegPairOdd || 518 Hint.first == (unsigned)ARMRI::RegPairEven) && 519 Hint.second && TargetRegisterInfo::isVirtualRegister(Hint.second)) { 520 // If 'Reg' is one of the even / odd register pair and it's now changed 521 // (e.g. coalesced) into a different register. The other register of the 522 // pair allocation hint must be updated to reflect the relationship 523 // change. 524 unsigned OtherReg = Hint.second; 525 Hint = MRI->getRegAllocationHint(OtherReg); 526 if (Hint.second == Reg) 527 // Make sure the pair has not already divorced. 528 MRI->setRegAllocationHint(OtherReg, Hint.first, NewReg); 529 } 530} 531 532bool ARMBaseRegisterInfo::hasBasePointer(const MachineFunction &MF) const { 533 const MachineFrameInfo *MFI = MF.getFrameInfo(); 534 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 535 536 if (!EnableBasePointer) 537 return false; 538 539 if (needsStackRealignment(MF) && MFI->hasVarSizedObjects()) 540 return true; 541 542 // Thumb has trouble with negative offsets from the FP. Thumb2 has a limited 543 // negative range for ldr/str (255), and thumb1 is positive offsets only. 544 // It's going to be better to use the SP or Base Pointer instead. When there 545 // are variable sized objects, we can't reference off of the SP, so we 546 // reserve a Base Pointer. 547 if (AFI->isThumbFunction() && MFI->hasVarSizedObjects()) { 548 // Conservatively estimate whether the negative offset from the frame 549 // pointer will be sufficient to reach. If a function has a smallish 550 // frame, it's less likely to have lots of spills and callee saved 551 // space, so it's all more likely to be within range of the frame pointer. 552 // If it's wrong, the scavenger will still enable access to work, it just 553 // won't be optimal. 554 if (AFI->isThumb2Function() && MFI->getLocalFrameSize() < 128) 555 return false; 556 return true; 557 } 558 559 return false; 560} 561 562bool ARMBaseRegisterInfo::canRealignStack(const MachineFunction &MF) const { 563 const MachineFrameInfo *MFI = MF.getFrameInfo(); 564 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 565 // We can't realign the stack if: 566 // 1. Dynamic stack realignment is explicitly disabled, 567 // 2. This is a Thumb1 function (it's not useful, so we don't bother), or 568 // 3. There are VLAs in the function and the base pointer is disabled. 569 return (RealignStack && !AFI->isThumb1OnlyFunction() && 570 (!MFI->hasVarSizedObjects() || EnableBasePointer)); 571} 572 573bool ARMBaseRegisterInfo:: 574needsStackRealignment(const MachineFunction &MF) const { 575 const MachineFrameInfo *MFI = MF.getFrameInfo(); 576 const Function *F = MF.getFunction(); 577 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 578 bool requiresRealignment = ((MFI->getLocalFrameMaxAlign() > StackAlign) || 579 F->hasFnAttr(Attribute::StackAlignment)); 580 581 return requiresRealignment && canRealignStack(MF); 582} 583 584bool ARMBaseRegisterInfo:: 585cannotEliminateFrame(const MachineFunction &MF) const { 586 const MachineFrameInfo *MFI = MF.getFrameInfo(); 587 if (DisableFramePointerElim(MF) && MFI->adjustsStack()) 588 return true; 589 return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken() 590 || needsStackRealignment(MF); 591} 592 593/// estimateStackSize - Estimate and return the size of the frame. 594static unsigned estimateStackSize(MachineFunction &MF) { 595 const MachineFrameInfo *FFI = MF.getFrameInfo(); 596 int Offset = 0; 597 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) { 598 int FixedOff = -FFI->getObjectOffset(i); 599 if (FixedOff > Offset) Offset = FixedOff; 600 } 601 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) { 602 if (FFI->isDeadObjectIndex(i)) 603 continue; 604 Offset += FFI->getObjectSize(i); 605 unsigned Align = FFI->getObjectAlignment(i); 606 // Adjust to alignment boundary 607 Offset = (Offset+Align-1)/Align*Align; 608 } 609 return (unsigned)Offset; 610} 611 612/// estimateRSStackSizeLimit - Look at each instruction that references stack 613/// frames and return the stack size limit beyond which some of these 614/// instructions will require a scratch register during their expansion later. 615unsigned 616ARMBaseRegisterInfo::estimateRSStackSizeLimit(MachineFunction &MF) const { 617 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 618 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 619 unsigned Limit = (1 << 12) - 1; 620 for (MachineFunction::iterator BB = MF.begin(),E = MF.end(); BB != E; ++BB) { 621 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 622 I != E; ++I) { 623 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 624 if (!I->getOperand(i).isFI()) continue; 625 626 // When using ADDri to get the address of a stack object, 255 is the 627 // largest offset guaranteed to fit in the immediate offset. 628 if (I->getOpcode() == ARM::ADDri) { 629 Limit = std::min(Limit, (1U << 8) - 1); 630 break; 631 } 632 633 // Otherwise check the addressing mode. 634 switch (I->getDesc().TSFlags & ARMII::AddrModeMask) { 635 case ARMII::AddrMode3: 636 case ARMII::AddrModeT2_i8: 637 Limit = std::min(Limit, (1U << 8) - 1); 638 break; 639 case ARMII::AddrMode5: 640 case ARMII::AddrModeT2_i8s4: 641 Limit = std::min(Limit, ((1U << 8) - 1) * 4); 642 break; 643 case ARMII::AddrModeT2_i12: 644 // i12 supports only positive offset so these will be converted to 645 // i8 opcodes. See llvm::rewriteT2FrameIndex. 646 if (TFI->hasFP(MF) && AFI->hasStackFrame()) 647 Limit = std::min(Limit, (1U << 8) - 1); 648 break; 649 case ARMII::AddrMode4: 650 case ARMII::AddrMode6: 651 // Addressing modes 4 & 6 (load/store) instructions can't encode an 652 // immediate offset for stack references. 653 return 0; 654 default: 655 break; 656 } 657 break; // At most one FI per instruction 658 } 659 } 660 } 661 662 return Limit; 663} 664 665static unsigned GetFunctionSizeInBytes(const MachineFunction &MF, 666 const ARMBaseInstrInfo &TII) { 667 unsigned FnSize = 0; 668 for (MachineFunction::const_iterator MBBI = MF.begin(), E = MF.end(); 669 MBBI != E; ++MBBI) { 670 const MachineBasicBlock &MBB = *MBBI; 671 for (MachineBasicBlock::const_iterator I = MBB.begin(),E = MBB.end(); 672 I != E; ++I) 673 FnSize += TII.GetInstSizeInBytes(I); 674 } 675 return FnSize; 676} 677 678void 679ARMBaseRegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF, 680 RegScavenger *RS) const { 681 // This tells PEI to spill the FP as if it is any other callee-save register 682 // to take advantage the eliminateFrameIndex machinery. This also ensures it 683 // is spilled in the order specified by getCalleeSavedRegs() to make it easier 684 // to combine multiple loads / stores. 685 bool CanEliminateFrame = true; 686 bool CS1Spilled = false; 687 bool LRSpilled = false; 688 unsigned NumGPRSpills = 0; 689 SmallVector<unsigned, 4> UnspilledCS1GPRs; 690 SmallVector<unsigned, 4> UnspilledCS2GPRs; 691 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 692 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 693 MachineFrameInfo *MFI = MF.getFrameInfo(); 694 695 // Spill R4 if Thumb2 function requires stack realignment - it will be used as 696 // scratch register. 697 // FIXME: It will be better just to find spare register here. 698 if (needsStackRealignment(MF) && 699 AFI->isThumb2Function()) 700 MF.getRegInfo().setPhysRegUsed(ARM::R4); 701 702 // Spill LR if Thumb1 function uses variable length argument lists. 703 if (AFI->isThumb1OnlyFunction() && AFI->getVarArgsRegSaveSize() > 0) 704 MF.getRegInfo().setPhysRegUsed(ARM::LR); 705 706 // Spill the BasePtr if it's used. 707 if (hasBasePointer(MF)) 708 MF.getRegInfo().setPhysRegUsed(BasePtr); 709 710 // Don't spill FP if the frame can be eliminated. This is determined 711 // by scanning the callee-save registers to see if any is used. 712 const unsigned *CSRegs = getCalleeSavedRegs(); 713 for (unsigned i = 0; CSRegs[i]; ++i) { 714 unsigned Reg = CSRegs[i]; 715 bool Spilled = false; 716 if (MF.getRegInfo().isPhysRegUsed(Reg)) { 717 AFI->setCSRegisterIsSpilled(Reg); 718 Spilled = true; 719 CanEliminateFrame = false; 720 } else { 721 // Check alias registers too. 722 for (const unsigned *Aliases = getAliasSet(Reg); *Aliases; ++Aliases) { 723 if (MF.getRegInfo().isPhysRegUsed(*Aliases)) { 724 Spilled = true; 725 CanEliminateFrame = false; 726 } 727 } 728 } 729 730 if (!ARM::GPRRegisterClass->contains(Reg)) 731 continue; 732 733 if (Spilled) { 734 NumGPRSpills++; 735 736 if (!STI.isTargetDarwin()) { 737 if (Reg == ARM::LR) 738 LRSpilled = true; 739 CS1Spilled = true; 740 continue; 741 } 742 743 // Keep track if LR and any of R4, R5, R6, and R7 is spilled. 744 switch (Reg) { 745 case ARM::LR: 746 LRSpilled = true; 747 // Fallthrough 748 case ARM::R4: 749 case ARM::R5: 750 case ARM::R6: 751 case ARM::R7: 752 CS1Spilled = true; 753 break; 754 default: 755 break; 756 } 757 } else { 758 if (!STI.isTargetDarwin()) { 759 UnspilledCS1GPRs.push_back(Reg); 760 continue; 761 } 762 763 switch (Reg) { 764 case ARM::R4: 765 case ARM::R5: 766 case ARM::R6: 767 case ARM::R7: 768 case ARM::LR: 769 UnspilledCS1GPRs.push_back(Reg); 770 break; 771 default: 772 UnspilledCS2GPRs.push_back(Reg); 773 break; 774 } 775 } 776 } 777 778 bool ForceLRSpill = false; 779 if (!LRSpilled && AFI->isThumb1OnlyFunction()) { 780 unsigned FnSize = GetFunctionSizeInBytes(MF, TII); 781 // Force LR to be spilled if the Thumb function size is > 2048. This enables 782 // use of BL to implement far jump. If it turns out that it's not needed 783 // then the branch fix up path will undo it. 784 if (FnSize >= (1 << 11)) { 785 CanEliminateFrame = false; 786 ForceLRSpill = true; 787 } 788 } 789 790 // If any of the stack slot references may be out of range of an immediate 791 // offset, make sure a register (or a spill slot) is available for the 792 // register scavenger. Note that if we're indexing off the frame pointer, the 793 // effective stack size is 4 bytes larger since the FP points to the stack 794 // slot of the previous FP. Also, if we have variable sized objects in the 795 // function, stack slot references will often be negative, and some of 796 // our instructions are positive-offset only, so conservatively consider 797 // that case to want a spill slot (or register) as well. Similarly, if 798 // the function adjusts the stack pointer during execution and the 799 // adjustments aren't already part of our stack size estimate, our offset 800 // calculations may be off, so be conservative. 801 // FIXME: We could add logic to be more precise about negative offsets 802 // and which instructions will need a scratch register for them. Is it 803 // worth the effort and added fragility? 804 bool BigStack = 805 (RS && 806 (estimateStackSize(MF) + ((TFI->hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >= 807 estimateRSStackSizeLimit(MF))) 808 || MFI->hasVarSizedObjects() 809 || (MFI->adjustsStack() && !TFI->canSimplifyCallFramePseudos(MF)); 810 811 bool ExtraCSSpill = false; 812 if (BigStack || !CanEliminateFrame || cannotEliminateFrame(MF)) { 813 AFI->setHasStackFrame(true); 814 815 // If LR is not spilled, but at least one of R4, R5, R6, and R7 is spilled. 816 // Spill LR as well so we can fold BX_RET to the registers restore (LDM). 817 if (!LRSpilled && CS1Spilled) { 818 MF.getRegInfo().setPhysRegUsed(ARM::LR); 819 AFI->setCSRegisterIsSpilled(ARM::LR); 820 NumGPRSpills++; 821 UnspilledCS1GPRs.erase(std::find(UnspilledCS1GPRs.begin(), 822 UnspilledCS1GPRs.end(), (unsigned)ARM::LR)); 823 ForceLRSpill = false; 824 ExtraCSSpill = true; 825 } 826 827 if (TFI->hasFP(MF)) { 828 MF.getRegInfo().setPhysRegUsed(FramePtr); 829 NumGPRSpills++; 830 } 831 832 // If stack and double are 8-byte aligned and we are spilling an odd number 833 // of GPRs, spill one extra callee save GPR so we won't have to pad between 834 // the integer and double callee save areas. 835 unsigned TargetAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 836 if (TargetAlign == 8 && (NumGPRSpills & 1)) { 837 if (CS1Spilled && !UnspilledCS1GPRs.empty()) { 838 for (unsigned i = 0, e = UnspilledCS1GPRs.size(); i != e; ++i) { 839 unsigned Reg = UnspilledCS1GPRs[i]; 840 // Don't spill high register if the function is thumb1 841 if (!AFI->isThumb1OnlyFunction() || 842 isARMLowRegister(Reg) || Reg == ARM::LR) { 843 MF.getRegInfo().setPhysRegUsed(Reg); 844 AFI->setCSRegisterIsSpilled(Reg); 845 if (!isReservedReg(MF, Reg)) 846 ExtraCSSpill = true; 847 break; 848 } 849 } 850 } else if (!UnspilledCS2GPRs.empty() && 851 !AFI->isThumb1OnlyFunction()) { 852 unsigned Reg = UnspilledCS2GPRs.front(); 853 MF.getRegInfo().setPhysRegUsed(Reg); 854 AFI->setCSRegisterIsSpilled(Reg); 855 if (!isReservedReg(MF, Reg)) 856 ExtraCSSpill = true; 857 } 858 } 859 860 // Estimate if we might need to scavenge a register at some point in order 861 // to materialize a stack offset. If so, either spill one additional 862 // callee-saved register or reserve a special spill slot to facilitate 863 // register scavenging. Thumb1 needs a spill slot for stack pointer 864 // adjustments also, even when the frame itself is small. 865 if (BigStack && !ExtraCSSpill) { 866 // If any non-reserved CS register isn't spilled, just spill one or two 867 // extra. That should take care of it! 868 unsigned NumExtras = TargetAlign / 4; 869 SmallVector<unsigned, 2> Extras; 870 while (NumExtras && !UnspilledCS1GPRs.empty()) { 871 unsigned Reg = UnspilledCS1GPRs.back(); 872 UnspilledCS1GPRs.pop_back(); 873 if (!isReservedReg(MF, Reg) && 874 (!AFI->isThumb1OnlyFunction() || isARMLowRegister(Reg) || 875 Reg == ARM::LR)) { 876 Extras.push_back(Reg); 877 NumExtras--; 878 } 879 } 880 // For non-Thumb1 functions, also check for hi-reg CS registers 881 if (!AFI->isThumb1OnlyFunction()) { 882 while (NumExtras && !UnspilledCS2GPRs.empty()) { 883 unsigned Reg = UnspilledCS2GPRs.back(); 884 UnspilledCS2GPRs.pop_back(); 885 if (!isReservedReg(MF, Reg)) { 886 Extras.push_back(Reg); 887 NumExtras--; 888 } 889 } 890 } 891 if (Extras.size() && NumExtras == 0) { 892 for (unsigned i = 0, e = Extras.size(); i != e; ++i) { 893 MF.getRegInfo().setPhysRegUsed(Extras[i]); 894 AFI->setCSRegisterIsSpilled(Extras[i]); 895 } 896 } else if (!AFI->isThumb1OnlyFunction()) { 897 // note: Thumb1 functions spill to R12, not the stack. Reserve a slot 898 // closest to SP or frame pointer. 899 const TargetRegisterClass *RC = ARM::GPRRegisterClass; 900 RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(), 901 RC->getAlignment(), 902 false)); 903 } 904 } 905 } 906 907 if (ForceLRSpill) { 908 MF.getRegInfo().setPhysRegUsed(ARM::LR); 909 AFI->setCSRegisterIsSpilled(ARM::LR); 910 AFI->setLRIsSpilledForFarJump(true); 911 } 912} 913 914unsigned ARMBaseRegisterInfo::getRARegister() const { 915 return ARM::LR; 916} 917 918unsigned 919ARMBaseRegisterInfo::getFrameRegister(const MachineFunction &MF) const { 920 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 921 922 if (TFI->hasFP(MF)) 923 return FramePtr; 924 return ARM::SP; 925} 926 927// Provide a base+offset reference to an FI slot for debug info. It's the 928// same as what we use for resolving the code-gen references for now. 929// FIXME: This can go wrong when references are SP-relative and simple call 930// frames aren't used. 931int 932ARMBaseRegisterInfo::getFrameIndexReference(const MachineFunction &MF, int FI, 933 unsigned &FrameReg) const { 934 return ResolveFrameIndexReference(MF, FI, FrameReg, 0); 935} 936 937int 938ARMBaseRegisterInfo::ResolveFrameIndexReference(const MachineFunction &MF, 939 int FI, 940 unsigned &FrameReg, 941 int SPAdj) const { 942 const MachineFrameInfo *MFI = MF.getFrameInfo(); 943 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 944 const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 945 int Offset = MFI->getObjectOffset(FI) + MFI->getStackSize(); 946 int FPOffset = Offset - AFI->getFramePtrSpillOffset(); 947 bool isFixed = MFI->isFixedObjectIndex(FI); 948 949 FrameReg = ARM::SP; 950 Offset += SPAdj; 951 if (AFI->isGPRCalleeSavedArea1Frame(FI)) 952 return Offset - AFI->getGPRCalleeSavedArea1Offset(); 953 else if (AFI->isGPRCalleeSavedArea2Frame(FI)) 954 return Offset - AFI->getGPRCalleeSavedArea2Offset(); 955 else if (AFI->isDPRCalleeSavedAreaFrame(FI)) 956 return Offset - AFI->getDPRCalleeSavedAreaOffset(); 957 958 // When dynamically realigning the stack, use the frame pointer for 959 // parameters, and the stack/base pointer for locals. 960 if (needsStackRealignment(MF)) { 961 assert (TFI->hasFP(MF) && "dynamic stack realignment without a FP!"); 962 if (isFixed) { 963 FrameReg = getFrameRegister(MF); 964 Offset = FPOffset; 965 } else if (MFI->hasVarSizedObjects()) { 966 assert(hasBasePointer(MF) && 967 "VLAs and dynamic stack alignment, but missing base pointer!"); 968 FrameReg = BasePtr; 969 } 970 return Offset; 971 } 972 973 // If there is a frame pointer, use it when we can. 974 if (TFI->hasFP(MF) && AFI->hasStackFrame()) { 975 // Use frame pointer to reference fixed objects. Use it for locals if 976 // there are VLAs (and thus the SP isn't reliable as a base). 977 if (isFixed || (MFI->hasVarSizedObjects() && !hasBasePointer(MF))) { 978 FrameReg = getFrameRegister(MF); 979 return FPOffset; 980 } else if (MFI->hasVarSizedObjects()) { 981 assert(hasBasePointer(MF) && "missing base pointer!"); 982 // Try to use the frame pointer if we can, else use the base pointer 983 // since it's available. This is handy for the emergency spill slot, in 984 // particular. 985 if (AFI->isThumb2Function()) { 986 if (FPOffset >= -255 && FPOffset < 0) { 987 FrameReg = getFrameRegister(MF); 988 return FPOffset; 989 } 990 } else 991 FrameReg = BasePtr; 992 } else if (AFI->isThumb2Function()) { 993 // In Thumb2 mode, the negative offset is very limited. Try to avoid 994 // out of range references. 995 if (FPOffset >= -255 && FPOffset < 0) { 996 FrameReg = getFrameRegister(MF); 997 return FPOffset; 998 } 999 } else if (Offset > (FPOffset < 0 ? -FPOffset : FPOffset)) { 1000 // Otherwise, use SP or FP, whichever is closer to the stack slot. 1001 FrameReg = getFrameRegister(MF); 1002 return FPOffset; 1003 } 1004 } 1005 // Use the base pointer if we have one. 1006 if (hasBasePointer(MF)) 1007 FrameReg = BasePtr; 1008 return Offset; 1009} 1010 1011int 1012ARMBaseRegisterInfo::getFrameIndexOffset(const MachineFunction &MF, 1013 int FI) const { 1014 unsigned FrameReg; 1015 return getFrameIndexReference(MF, FI, FrameReg); 1016} 1017 1018unsigned ARMBaseRegisterInfo::getEHExceptionRegister() const { 1019 llvm_unreachable("What is the exception register"); 1020 return 0; 1021} 1022 1023unsigned ARMBaseRegisterInfo::getEHHandlerRegister() const { 1024 llvm_unreachable("What is the exception handler register"); 1025 return 0; 1026} 1027 1028int ARMBaseRegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const { 1029 return ARMGenRegisterInfo::getDwarfRegNumFull(RegNum, 0); 1030} 1031 1032unsigned ARMBaseRegisterInfo::getRegisterPairEven(unsigned Reg, 1033 const MachineFunction &MF) const { 1034 switch (Reg) { 1035 default: break; 1036 // Return 0 if either register of the pair is a special register. 1037 // So no R12, etc. 1038 case ARM::R1: 1039 return ARM::R0; 1040 case ARM::R3: 1041 return ARM::R2; 1042 case ARM::R5: 1043 return ARM::R4; 1044 case ARM::R7: 1045 return (isReservedReg(MF, ARM::R7) || isReservedReg(MF, ARM::R6)) 1046 ? 0 : ARM::R6; 1047 case ARM::R9: 1048 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R8; 1049 case ARM::R11: 1050 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R10; 1051 1052 case ARM::S1: 1053 return ARM::S0; 1054 case ARM::S3: 1055 return ARM::S2; 1056 case ARM::S5: 1057 return ARM::S4; 1058 case ARM::S7: 1059 return ARM::S6; 1060 case ARM::S9: 1061 return ARM::S8; 1062 case ARM::S11: 1063 return ARM::S10; 1064 case ARM::S13: 1065 return ARM::S12; 1066 case ARM::S15: 1067 return ARM::S14; 1068 case ARM::S17: 1069 return ARM::S16; 1070 case ARM::S19: 1071 return ARM::S18; 1072 case ARM::S21: 1073 return ARM::S20; 1074 case ARM::S23: 1075 return ARM::S22; 1076 case ARM::S25: 1077 return ARM::S24; 1078 case ARM::S27: 1079 return ARM::S26; 1080 case ARM::S29: 1081 return ARM::S28; 1082 case ARM::S31: 1083 return ARM::S30; 1084 1085 case ARM::D1: 1086 return ARM::D0; 1087 case ARM::D3: 1088 return ARM::D2; 1089 case ARM::D5: 1090 return ARM::D4; 1091 case ARM::D7: 1092 return ARM::D6; 1093 case ARM::D9: 1094 return ARM::D8; 1095 case ARM::D11: 1096 return ARM::D10; 1097 case ARM::D13: 1098 return ARM::D12; 1099 case ARM::D15: 1100 return ARM::D14; 1101 case ARM::D17: 1102 return ARM::D16; 1103 case ARM::D19: 1104 return ARM::D18; 1105 case ARM::D21: 1106 return ARM::D20; 1107 case ARM::D23: 1108 return ARM::D22; 1109 case ARM::D25: 1110 return ARM::D24; 1111 case ARM::D27: 1112 return ARM::D26; 1113 case ARM::D29: 1114 return ARM::D28; 1115 case ARM::D31: 1116 return ARM::D30; 1117 } 1118 1119 return 0; 1120} 1121 1122unsigned ARMBaseRegisterInfo::getRegisterPairOdd(unsigned Reg, 1123 const MachineFunction &MF) const { 1124 switch (Reg) { 1125 default: break; 1126 // Return 0 if either register of the pair is a special register. 1127 // So no R12, etc. 1128 case ARM::R0: 1129 return ARM::R1; 1130 case ARM::R2: 1131 return ARM::R3; 1132 case ARM::R4: 1133 return ARM::R5; 1134 case ARM::R6: 1135 return (isReservedReg(MF, ARM::R7) || isReservedReg(MF, ARM::R6)) 1136 ? 0 : ARM::R7; 1137 case ARM::R8: 1138 return isReservedReg(MF, ARM::R9) ? 0 :ARM::R9; 1139 case ARM::R10: 1140 return isReservedReg(MF, ARM::R11) ? 0 : ARM::R11; 1141 1142 case ARM::S0: 1143 return ARM::S1; 1144 case ARM::S2: 1145 return ARM::S3; 1146 case ARM::S4: 1147 return ARM::S5; 1148 case ARM::S6: 1149 return ARM::S7; 1150 case ARM::S8: 1151 return ARM::S9; 1152 case ARM::S10: 1153 return ARM::S11; 1154 case ARM::S12: 1155 return ARM::S13; 1156 case ARM::S14: 1157 return ARM::S15; 1158 case ARM::S16: 1159 return ARM::S17; 1160 case ARM::S18: 1161 return ARM::S19; 1162 case ARM::S20: 1163 return ARM::S21; 1164 case ARM::S22: 1165 return ARM::S23; 1166 case ARM::S24: 1167 return ARM::S25; 1168 case ARM::S26: 1169 return ARM::S27; 1170 case ARM::S28: 1171 return ARM::S29; 1172 case ARM::S30: 1173 return ARM::S31; 1174 1175 case ARM::D0: 1176 return ARM::D1; 1177 case ARM::D2: 1178 return ARM::D3; 1179 case ARM::D4: 1180 return ARM::D5; 1181 case ARM::D6: 1182 return ARM::D7; 1183 case ARM::D8: 1184 return ARM::D9; 1185 case ARM::D10: 1186 return ARM::D11; 1187 case ARM::D12: 1188 return ARM::D13; 1189 case ARM::D14: 1190 return ARM::D15; 1191 case ARM::D16: 1192 return ARM::D17; 1193 case ARM::D18: 1194 return ARM::D19; 1195 case ARM::D20: 1196 return ARM::D21; 1197 case ARM::D22: 1198 return ARM::D23; 1199 case ARM::D24: 1200 return ARM::D25; 1201 case ARM::D26: 1202 return ARM::D27; 1203 case ARM::D28: 1204 return ARM::D29; 1205 case ARM::D30: 1206 return ARM::D31; 1207 } 1208 1209 return 0; 1210} 1211 1212/// emitLoadConstPool - Emits a load from constpool to materialize the 1213/// specified immediate. 1214void ARMBaseRegisterInfo:: 1215emitLoadConstPool(MachineBasicBlock &MBB, 1216 MachineBasicBlock::iterator &MBBI, 1217 DebugLoc dl, 1218 unsigned DestReg, unsigned SubIdx, int Val, 1219 ARMCC::CondCodes Pred, 1220 unsigned PredReg) const { 1221 MachineFunction &MF = *MBB.getParent(); 1222 MachineConstantPool *ConstantPool = MF.getConstantPool(); 1223 const Constant *C = 1224 ConstantInt::get(Type::getInt32Ty(MF.getFunction()->getContext()), Val); 1225 unsigned Idx = ConstantPool->getConstantPoolIndex(C, 4); 1226 1227 BuildMI(MBB, MBBI, dl, TII.get(ARM::LDRcp)) 1228 .addReg(DestReg, getDefRegState(true), SubIdx) 1229 .addConstantPoolIndex(Idx) 1230 .addImm(0).addImm(Pred).addReg(PredReg); 1231} 1232 1233bool ARMBaseRegisterInfo:: 1234requiresRegisterScavenging(const MachineFunction &MF) const { 1235 return true; 1236} 1237 1238bool ARMBaseRegisterInfo:: 1239requiresFrameIndexScavenging(const MachineFunction &MF) const { 1240 return true; 1241} 1242 1243bool ARMBaseRegisterInfo:: 1244requiresVirtualBaseRegisters(const MachineFunction &MF) const { 1245 return EnableLocalStackAlloc; 1246} 1247 1248static void 1249emitSPUpdate(bool isARM, 1250 MachineBasicBlock &MBB, MachineBasicBlock::iterator &MBBI, 1251 DebugLoc dl, const ARMBaseInstrInfo &TII, 1252 int NumBytes, 1253 ARMCC::CondCodes Pred = ARMCC::AL, unsigned PredReg = 0) { 1254 if (isARM) 1255 emitARMRegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes, 1256 Pred, PredReg, TII); 1257 else 1258 emitT2RegPlusImmediate(MBB, MBBI, dl, ARM::SP, ARM::SP, NumBytes, 1259 Pred, PredReg, TII); 1260} 1261 1262 1263void ARMBaseRegisterInfo:: 1264eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, 1265 MachineBasicBlock::iterator I) const { 1266 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 1267 if (!TFI->hasReservedCallFrame(MF)) { 1268 // If we have alloca, convert as follows: 1269 // ADJCALLSTACKDOWN -> sub, sp, sp, amount 1270 // ADJCALLSTACKUP -> add, sp, sp, amount 1271 MachineInstr *Old = I; 1272 DebugLoc dl = Old->getDebugLoc(); 1273 unsigned Amount = Old->getOperand(0).getImm(); 1274 if (Amount != 0) { 1275 // We need to keep the stack aligned properly. To do this, we round the 1276 // amount of space needed for the outgoing arguments up to the next 1277 // alignment boundary. 1278 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment(); 1279 Amount = (Amount+Align-1)/Align*Align; 1280 1281 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1282 assert(!AFI->isThumb1OnlyFunction() && 1283 "This eliminateCallFramePseudoInstr does not support Thumb1!"); 1284 bool isARM = !AFI->isThumbFunction(); 1285 1286 // Replace the pseudo instruction with a new instruction... 1287 unsigned Opc = Old->getOpcode(); 1288 int PIdx = Old->findFirstPredOperandIdx(); 1289 ARMCC::CondCodes Pred = (PIdx == -1) 1290 ? ARMCC::AL : (ARMCC::CondCodes)Old->getOperand(PIdx).getImm(); 1291 if (Opc == ARM::ADJCALLSTACKDOWN || Opc == ARM::tADJCALLSTACKDOWN) { 1292 // Note: PredReg is operand 2 for ADJCALLSTACKDOWN. 1293 unsigned PredReg = Old->getOperand(2).getReg(); 1294 emitSPUpdate(isARM, MBB, I, dl, TII, -Amount, Pred, PredReg); 1295 } else { 1296 // Note: PredReg is operand 3 for ADJCALLSTACKUP. 1297 unsigned PredReg = Old->getOperand(3).getReg(); 1298 assert(Opc == ARM::ADJCALLSTACKUP || Opc == ARM::tADJCALLSTACKUP); 1299 emitSPUpdate(isARM, MBB, I, dl, TII, Amount, Pred, PredReg); 1300 } 1301 } 1302 } 1303 MBB.erase(I); 1304} 1305 1306int64_t ARMBaseRegisterInfo:: 1307getFrameIndexInstrOffset(const MachineInstr *MI, int Idx) const { 1308 const TargetInstrDesc &Desc = MI->getDesc(); 1309 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); 1310 int64_t InstrOffs = 0;; 1311 int Scale = 1; 1312 unsigned ImmIdx = 0; 1313 switch (AddrMode) { 1314 case ARMII::AddrModeT2_i8: 1315 case ARMII::AddrModeT2_i12: 1316 case ARMII::AddrMode_i12: 1317 InstrOffs = MI->getOperand(Idx+1).getImm(); 1318 Scale = 1; 1319 break; 1320 case ARMII::AddrMode5: { 1321 // VFP address mode. 1322 const MachineOperand &OffOp = MI->getOperand(Idx+1); 1323 InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm()); 1324 if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub) 1325 InstrOffs = -InstrOffs; 1326 Scale = 4; 1327 break; 1328 } 1329 case ARMII::AddrMode2: { 1330 ImmIdx = Idx+2; 1331 InstrOffs = ARM_AM::getAM2Offset(MI->getOperand(ImmIdx).getImm()); 1332 if (ARM_AM::getAM2Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub) 1333 InstrOffs = -InstrOffs; 1334 break; 1335 } 1336 case ARMII::AddrMode3: { 1337 ImmIdx = Idx+2; 1338 InstrOffs = ARM_AM::getAM3Offset(MI->getOperand(ImmIdx).getImm()); 1339 if (ARM_AM::getAM3Op(MI->getOperand(ImmIdx).getImm()) == ARM_AM::sub) 1340 InstrOffs = -InstrOffs; 1341 break; 1342 } 1343 case ARMII::AddrModeT1_s: { 1344 ImmIdx = Idx+1; 1345 InstrOffs = MI->getOperand(ImmIdx).getImm(); 1346 Scale = 4; 1347 break; 1348 } 1349 default: 1350 llvm_unreachable("Unsupported addressing mode!"); 1351 break; 1352 } 1353 1354 return InstrOffs * Scale; 1355} 1356 1357/// needsFrameBaseReg - Returns true if the instruction's frame index 1358/// reference would be better served by a base register other than FP 1359/// or SP. Used by LocalStackFrameAllocation to determine which frame index 1360/// references it should create new base registers for. 1361bool ARMBaseRegisterInfo:: 1362needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const { 1363 for (unsigned i = 0; !MI->getOperand(i).isFI(); ++i) { 1364 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!"); 1365 } 1366 1367 // It's the load/store FI references that cause issues, as it can be difficult 1368 // to materialize the offset if it won't fit in the literal field. Estimate 1369 // based on the size of the local frame and some conservative assumptions 1370 // about the rest of the stack frame (note, this is pre-regalloc, so 1371 // we don't know everything for certain yet) whether this offset is likely 1372 // to be out of range of the immediate. Return true if so. 1373 1374 // We only generate virtual base registers for loads and stores, so 1375 // return false for everything else. 1376 unsigned Opc = MI->getOpcode(); 1377 switch (Opc) { 1378 case ARM::LDRi12: case ARM::LDRH: case ARM::LDRBi12: 1379 case ARM::STRi12: case ARM::STRH: case ARM::STRBi12: 1380 case ARM::t2LDRi12: case ARM::t2LDRi8: 1381 case ARM::t2STRi12: case ARM::t2STRi8: 1382 case ARM::VLDRS: case ARM::VLDRD: 1383 case ARM::VSTRS: case ARM::VSTRD: 1384 case ARM::tSTRspi: case ARM::tLDRspi: 1385 if (ForceAllBaseRegAlloc) 1386 return true; 1387 break; 1388 default: 1389 return false; 1390 } 1391 1392 // Without a virtual base register, if the function has variable sized 1393 // objects, all fixed-size local references will be via the frame pointer, 1394 // Approximate the offset and see if it's legal for the instruction. 1395 // Note that the incoming offset is based on the SP value at function entry, 1396 // so it'll be negative. 1397 MachineFunction &MF = *MI->getParent()->getParent(); 1398 const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo(); 1399 MachineFrameInfo *MFI = MF.getFrameInfo(); 1400 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1401 1402 // Estimate an offset from the frame pointer. 1403 // Conservatively assume all callee-saved registers get pushed. R4-R6 1404 // will be earlier than the FP, so we ignore those. 1405 // R7, LR 1406 int64_t FPOffset = Offset - 8; 1407 // ARM and Thumb2 functions also need to consider R8-R11 and D8-D15 1408 if (!AFI->isThumbFunction() || !AFI->isThumb1OnlyFunction()) 1409 FPOffset -= 80; 1410 // Estimate an offset from the stack pointer. 1411 // The incoming offset is relating to the SP at the start of the function, 1412 // but when we access the local it'll be relative to the SP after local 1413 // allocation, so adjust our SP-relative offset by that allocation size. 1414 Offset = -Offset; 1415 Offset += MFI->getLocalFrameSize(); 1416 // Assume that we'll have at least some spill slots allocated. 1417 // FIXME: This is a total SWAG number. We should run some statistics 1418 // and pick a real one. 1419 Offset += 128; // 128 bytes of spill slots 1420 1421 // If there is a frame pointer, try using it. 1422 // The FP is only available if there is no dynamic realignment. We 1423 // don't know for sure yet whether we'll need that, so we guess based 1424 // on whether there are any local variables that would trigger it. 1425 unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment(); 1426 if (TFI->hasFP(MF) && 1427 !((MFI->getLocalFrameMaxAlign() > StackAlign) && canRealignStack(MF))) { 1428 if (isFrameOffsetLegal(MI, FPOffset)) 1429 return false; 1430 } 1431 // If we can reference via the stack pointer, try that. 1432 // FIXME: This (and the code that resolves the references) can be improved 1433 // to only disallow SP relative references in the live range of 1434 // the VLA(s). In practice, it's unclear how much difference that 1435 // would make, but it may be worth doing. 1436 if (!MFI->hasVarSizedObjects() && isFrameOffsetLegal(MI, Offset)) 1437 return false; 1438 1439 // The offset likely isn't legal, we want to allocate a virtual base register. 1440 return true; 1441} 1442 1443/// materializeFrameBaseRegister - Insert defining instruction(s) for 1444/// BaseReg to be a pointer to FrameIdx before insertion point I. 1445void ARMBaseRegisterInfo:: 1446materializeFrameBaseRegister(MachineBasicBlock::iterator I, unsigned BaseReg, 1447 int FrameIdx, int64_t Offset) const { 1448 ARMFunctionInfo *AFI = 1449 I->getParent()->getParent()->getInfo<ARMFunctionInfo>(); 1450 unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : 1451 (AFI->isThumb1OnlyFunction() ? ARM::tADDrSPi : ARM::t2ADDri); 1452 1453 MachineInstrBuilder MIB = 1454 BuildMI(*I->getParent(), I, I->getDebugLoc(), TII.get(ADDriOpc), BaseReg) 1455 .addFrameIndex(FrameIdx).addImm(Offset); 1456 if (!AFI->isThumb1OnlyFunction()) 1457 AddDefaultCC(AddDefaultPred(MIB)); 1458} 1459 1460void 1461ARMBaseRegisterInfo::resolveFrameIndex(MachineBasicBlock::iterator I, 1462 unsigned BaseReg, int64_t Offset) const { 1463 MachineInstr &MI = *I; 1464 MachineBasicBlock &MBB = *MI.getParent(); 1465 MachineFunction &MF = *MBB.getParent(); 1466 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1467 int Off = Offset; // ARM doesn't need the general 64-bit offsets 1468 unsigned i = 0; 1469 1470 assert(!AFI->isThumb1OnlyFunction() && 1471 "This resolveFrameIndex does not support Thumb1!"); 1472 1473 while (!MI.getOperand(i).isFI()) { 1474 ++i; 1475 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); 1476 } 1477 bool Done = false; 1478 if (!AFI->isThumbFunction()) 1479 Done = rewriteARMFrameIndex(MI, i, BaseReg, Off, TII); 1480 else { 1481 assert(AFI->isThumb2Function()); 1482 Done = rewriteT2FrameIndex(MI, i, BaseReg, Off, TII); 1483 } 1484 assert (Done && "Unable to resolve frame index!"); 1485} 1486 1487bool ARMBaseRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, 1488 int64_t Offset) const { 1489 const TargetInstrDesc &Desc = MI->getDesc(); 1490 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); 1491 unsigned i = 0; 1492 1493 while (!MI->getOperand(i).isFI()) { 1494 ++i; 1495 assert(i < MI->getNumOperands() &&"Instr doesn't have FrameIndex operand!"); 1496 } 1497 1498 // AddrMode4 and AddrMode6 cannot handle any offset. 1499 if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6) 1500 return Offset == 0; 1501 1502 unsigned NumBits = 0; 1503 unsigned Scale = 1; 1504 bool isSigned = true; 1505 switch (AddrMode) { 1506 case ARMII::AddrModeT2_i8: 1507 case ARMII::AddrModeT2_i12: 1508 // i8 supports only negative, and i12 supports only positive, so 1509 // based on Offset sign, consider the appropriate instruction 1510 Scale = 1; 1511 if (Offset < 0) { 1512 NumBits = 8; 1513 Offset = -Offset; 1514 } else { 1515 NumBits = 12; 1516 } 1517 break; 1518 case ARMII::AddrMode5: 1519 // VFP address mode. 1520 NumBits = 8; 1521 Scale = 4; 1522 break; 1523 case ARMII::AddrMode_i12: 1524 case ARMII::AddrMode2: 1525 NumBits = 12; 1526 break; 1527 case ARMII::AddrMode3: 1528 NumBits = 8; 1529 break; 1530 case ARMII::AddrModeT1_s: 1531 NumBits = 5; 1532 Scale = 4; 1533 isSigned = false; 1534 break; 1535 default: 1536 llvm_unreachable("Unsupported addressing mode!"); 1537 break; 1538 } 1539 1540 Offset += getFrameIndexInstrOffset(MI, i); 1541 // Make sure the offset is encodable for instructions that scale the 1542 // immediate. 1543 if ((Offset & (Scale-1)) != 0) 1544 return false; 1545 1546 if (isSigned && Offset < 0) 1547 Offset = -Offset; 1548 1549 unsigned Mask = (1 << NumBits) - 1; 1550 if ((unsigned)Offset <= Mask * Scale) 1551 return true; 1552 1553 return false; 1554} 1555 1556void 1557ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, 1558 int SPAdj, RegScavenger *RS) const { 1559 unsigned i = 0; 1560 MachineInstr &MI = *II; 1561 MachineBasicBlock &MBB = *MI.getParent(); 1562 MachineFunction &MF = *MBB.getParent(); 1563 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1564 assert(!AFI->isThumb1OnlyFunction() && 1565 "This eliminateFrameIndex does not support Thumb1!"); 1566 1567 while (!MI.getOperand(i).isFI()) { 1568 ++i; 1569 assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!"); 1570 } 1571 1572 int FrameIndex = MI.getOperand(i).getIndex(); 1573 unsigned FrameReg; 1574 1575 int Offset = ResolveFrameIndexReference(MF, FrameIndex, FrameReg, SPAdj); 1576 1577 // Special handling of dbg_value instructions. 1578 if (MI.isDebugValue()) { 1579 MI.getOperand(i). ChangeToRegister(FrameReg, false /*isDef*/); 1580 MI.getOperand(i+1).ChangeToImmediate(Offset); 1581 return; 1582 } 1583 1584 // Modify MI as necessary to handle as much of 'Offset' as possible 1585 bool Done = false; 1586 if (!AFI->isThumbFunction()) 1587 Done = rewriteARMFrameIndex(MI, i, FrameReg, Offset, TII); 1588 else { 1589 assert(AFI->isThumb2Function()); 1590 Done = rewriteT2FrameIndex(MI, i, FrameReg, Offset, TII); 1591 } 1592 if (Done) 1593 return; 1594 1595 // If we get here, the immediate doesn't fit into the instruction. We folded 1596 // as much as possible above, handle the rest, providing a register that is 1597 // SP+LargeImm. 1598 assert((Offset || 1599 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 || 1600 (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) && 1601 "This code isn't needed if offset already handled!"); 1602 1603 unsigned ScratchReg = 0; 1604 int PIdx = MI.findFirstPredOperandIdx(); 1605 ARMCC::CondCodes Pred = (PIdx == -1) 1606 ? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm(); 1607 unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg(); 1608 if (Offset == 0) 1609 // Must be addrmode4/6. 1610 MI.getOperand(i).ChangeToRegister(FrameReg, false, false, false); 1611 else { 1612 ScratchReg = MF.getRegInfo().createVirtualRegister(ARM::GPRRegisterClass); 1613 if (!AFI->isThumbFunction()) 1614 emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg, 1615 Offset, Pred, PredReg, TII); 1616 else { 1617 assert(AFI->isThumb2Function()); 1618 emitT2RegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg, 1619 Offset, Pred, PredReg, TII); 1620 } 1621 MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true); 1622 } 1623} 1624 1625#include "ARMGenRegisterInfo.inc" 1626