| 1 | //===- lib/MC/MCAssembler.cpp - Assembler Backend Implementation ----------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #include "llvm/MC/MCAssembler.h" |
| 10 | #include "llvm/ADT/ArrayRef.h" |
| 11 | #include "llvm/ADT/SmallString.h" |
| 12 | #include "llvm/ADT/SmallVector.h" |
| 13 | #include "llvm/ADT/Statistic.h" |
| 14 | #include "llvm/ADT/StringRef.h" |
| 15 | #include "llvm/ADT/Twine.h" |
| 16 | #include "llvm/MC/MCAsmBackend.h" |
| 17 | #include "llvm/MC/MCAsmInfo.h" |
| 18 | #include "llvm/MC/MCCodeEmitter.h" |
| 19 | #include "llvm/MC/MCCodeView.h" |
| 20 | #include "llvm/MC/MCContext.h" |
| 21 | #include "llvm/MC/MCDwarf.h" |
| 22 | #include "llvm/MC/MCExpr.h" |
| 23 | #include "llvm/MC/MCFixup.h" |
| 24 | #include "llvm/MC/MCFixupKindInfo.h" |
| 25 | #include "llvm/MC/MCInst.h" |
| 26 | #include "llvm/MC/MCObjectWriter.h" |
| 27 | #include "llvm/MC/MCSection.h" |
| 28 | #include "llvm/MC/MCSymbol.h" |
| 29 | #include "llvm/MC/MCValue.h" |
| 30 | #include "llvm/Support/Alignment.h" |
| 31 | #include "llvm/Support/Casting.h" |
| 32 | #include "llvm/Support/Debug.h" |
| 33 | #include "llvm/Support/EndianStream.h" |
| 34 | #include "llvm/Support/ErrorHandling.h" |
| 35 | #include "llvm/Support/LEB128.h" |
| 36 | #include "llvm/Support/raw_ostream.h" |
| 37 | #include <cassert> |
| 38 | #include <cstdint> |
| 39 | #include <tuple> |
| 40 | #include <utility> |
| 41 | |
| 42 | using namespace llvm; |
| 43 | |
| 44 | namespace llvm { |
| 45 | class MCSubtargetInfo; |
| 46 | } |
| 47 | |
| 48 | #define DEBUG_TYPE "assembler" |
| 49 | |
| 50 | namespace { |
| 51 | namespace stats { |
| 52 | |
| 53 | STATISTIC(EmittedFragments, "Number of emitted assembler fragments - total" ); |
| 54 | STATISTIC(EmittedRelaxableFragments, |
| 55 | "Number of emitted assembler fragments - relaxable" ); |
| 56 | STATISTIC(EmittedDataFragments, |
| 57 | "Number of emitted assembler fragments - data" ); |
| 58 | STATISTIC(EmittedAlignFragments, |
| 59 | "Number of emitted assembler fragments - align" ); |
| 60 | STATISTIC(EmittedFillFragments, |
| 61 | "Number of emitted assembler fragments - fill" ); |
| 62 | STATISTIC(EmittedNopsFragments, "Number of emitted assembler fragments - nops" ); |
| 63 | STATISTIC(EmittedOrgFragments, "Number of emitted assembler fragments - org" ); |
| 64 | STATISTIC(evaluateFixup, "Number of evaluated fixups" ); |
| 65 | STATISTIC(ObjectBytes, "Number of emitted object file bytes" ); |
| 66 | STATISTIC(RelaxationSteps, "Number of assembler layout and relaxation steps" ); |
| 67 | STATISTIC(RelaxedInstructions, "Number of relaxed instructions" ); |
| 68 | |
| 69 | } // end namespace stats |
| 70 | } // end anonymous namespace |
| 71 | |
| 72 | // FIXME FIXME FIXME: There are number of places in this file where we convert |
| 73 | // what is a 64-bit assembler value used for computation into a value in the |
| 74 | // object file, which may truncate it. We should detect that truncation where |
| 75 | // invalid and report errors back. |
| 76 | |
| 77 | /* *** */ |
| 78 | |
| 79 | MCAssembler::MCAssembler(MCContext &Context, |
| 80 | std::unique_ptr<MCAsmBackend> Backend, |
| 81 | std::unique_ptr<MCCodeEmitter> Emitter, |
| 82 | std::unique_ptr<MCObjectWriter> Writer) |
| 83 | : Context(Context), Backend(std::move(Backend)), |
| 84 | Emitter(std::move(Emitter)), Writer(std::move(Writer)) { |
| 85 | if (this->Backend) |
| 86 | this->Backend->setAssembler(this); |
| 87 | if (this->Writer) |
| 88 | this->Writer->setAssembler(this); |
| 89 | } |
| 90 | |
| 91 | void MCAssembler::reset() { |
| 92 | HasLayout = false; |
| 93 | HasFinalLayout = false; |
| 94 | RelaxAll = false; |
| 95 | Sections.clear(); |
| 96 | Symbols.clear(); |
| 97 | ThumbFuncs.clear(); |
| 98 | BundleAlignSize = 0; |
| 99 | |
| 100 | // reset objects owned by us |
| 101 | if (getBackendPtr()) |
| 102 | getBackendPtr()->reset(); |
| 103 | if (getEmitterPtr()) |
| 104 | getEmitterPtr()->reset(); |
| 105 | if (Writer) |
| 106 | Writer->reset(); |
| 107 | } |
| 108 | |
| 109 | bool MCAssembler::registerSection(MCSection &Section) { |
| 110 | if (Section.isRegistered()) |
| 111 | return false; |
| 112 | assert(Section.curFragList()->Head && "allocInitialFragment not called" ); |
| 113 | Sections.push_back(Elt: &Section); |
| 114 | Section.setIsRegistered(true); |
| 115 | return true; |
| 116 | } |
| 117 | |
| 118 | bool MCAssembler::isThumbFunc(const MCSymbol *Symbol) const { |
| 119 | if (ThumbFuncs.count(Ptr: Symbol)) |
| 120 | return true; |
| 121 | |
| 122 | if (!Symbol->isVariable()) |
| 123 | return false; |
| 124 | |
| 125 | const MCExpr *Expr = Symbol->getVariableValue(); |
| 126 | |
| 127 | MCValue V; |
| 128 | if (!Expr->evaluateAsRelocatable(Res&: V, Asm: nullptr)) |
| 129 | return false; |
| 130 | |
| 131 | if (V.getSubSym() || V.getSpecifier()) |
| 132 | return false; |
| 133 | |
| 134 | auto *Sym = V.getAddSym(); |
| 135 | if (!Sym || V.getSpecifier()) |
| 136 | return false; |
| 137 | |
| 138 | if (!isThumbFunc(Symbol: Sym)) |
| 139 | return false; |
| 140 | |
| 141 | ThumbFuncs.insert(Ptr: Symbol); // Cache it. |
| 142 | return true; |
| 143 | } |
| 144 | |
| 145 | bool MCAssembler::evaluateFixup(const MCFragment &F, MCFixup &Fixup, |
| 146 | MCValue &Target, uint64_t &Value, |
| 147 | bool RecordReloc, |
| 148 | MutableArrayRef<char> Contents) const { |
| 149 | ++stats::evaluateFixup; |
| 150 | |
| 151 | // FIXME: This code has some duplication with recordRelocation. We should |
| 152 | // probably merge the two into a single callback that tries to evaluate a |
| 153 | // fixup and records a relocation if one is needed. |
| 154 | |
| 155 | // On error claim to have completely evaluated the fixup, to prevent any |
| 156 | // further processing from being done. |
| 157 | const MCExpr *Expr = Fixup.getValue(); |
| 158 | Value = 0; |
| 159 | if (!Expr->evaluateAsRelocatable(Res&: Target, Asm: this)) { |
| 160 | reportError(L: Fixup.getLoc(), Msg: "expected relocatable expression" ); |
| 161 | return true; |
| 162 | } |
| 163 | |
| 164 | bool IsResolved = false; |
| 165 | unsigned FixupFlags = getBackend().getFixupKindInfo(Kind: Fixup.getKind()).Flags; |
| 166 | bool IsPCRel = FixupFlags & MCFixupKindInfo::FKF_IsPCRel; |
| 167 | if (FixupFlags & MCFixupKindInfo::FKF_IsTarget) { |
| 168 | IsResolved = getBackend().evaluateTargetFixup(Fixup, Target, Value); |
| 169 | } else { |
| 170 | const MCSymbol *Add = Target.getAddSym(); |
| 171 | const MCSymbol *Sub = Target.getSubSym(); |
| 172 | Value = Target.getConstant(); |
| 173 | if (Add && Add->isDefined()) |
| 174 | Value += getSymbolOffset(S: *Add); |
| 175 | if (Sub && Sub->isDefined()) |
| 176 | Value -= getSymbolOffset(S: *Sub); |
| 177 | |
| 178 | bool ShouldAlignPC = |
| 179 | FixupFlags & MCFixupKindInfo::FKF_IsAlignedDownTo32Bits; |
| 180 | if (IsPCRel) { |
| 181 | uint64_t Offset = getFragmentOffset(F) + Fixup.getOffset(); |
| 182 | |
| 183 | // A number of ARM fixups in Thumb mode require that the effective PC |
| 184 | // address be determined as the 32-bit aligned version of the actual |
| 185 | // offset. |
| 186 | if (ShouldAlignPC) |
| 187 | Offset &= ~0x3; |
| 188 | Value -= Offset; |
| 189 | |
| 190 | if (Add && !Sub && !Add->isUndefined() && !Add->isAbsolute()) { |
| 191 | IsResolved = getWriter().isSymbolRefDifferenceFullyResolvedImpl( |
| 192 | SymA: *Add, FB: F, InSet: false, IsPCRel: true); |
| 193 | } |
| 194 | } else { |
| 195 | IsResolved = Target.isAbsolute(); |
| 196 | assert(!ShouldAlignPC && "FKF_IsAlignedDownTo32Bits must be PC-relative" ); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | if (!RecordReloc) |
| 201 | return IsResolved; |
| 202 | |
| 203 | if (IsResolved && mc::isRelocRelocation(FixupKind: Fixup.getKind())) |
| 204 | IsResolved = false; |
| 205 | if (IsPCRel) |
| 206 | Fixup.setPCRel(); |
| 207 | getBackend().applyFixup(F, Fixup, Target, Data: Contents, Value, IsResolved); |
| 208 | return true; |
| 209 | } |
| 210 | |
| 211 | uint64_t MCAssembler::computeFragmentSize(const MCFragment &F) const { |
| 212 | assert(getBackendPtr() && "Requires assembler backend" ); |
| 213 | switch (F.getKind()) { |
| 214 | case MCFragment::FT_Data: |
| 215 | return cast<MCDataFragment>(Val: F).getContents().size(); |
| 216 | case MCFragment::FT_Relaxable: |
| 217 | return cast<MCRelaxableFragment>(Val: F).getContents().size(); |
| 218 | case MCFragment::FT_Fill: { |
| 219 | auto &FF = cast<MCFillFragment>(Val: F); |
| 220 | int64_t NumValues = 0; |
| 221 | if (!FF.getNumValues().evaluateKnownAbsolute(Res&: NumValues, Asm: *this)) { |
| 222 | recordError(L: FF.getLoc(), Msg: "expected assembly-time absolute expression" ); |
| 223 | return 0; |
| 224 | } |
| 225 | int64_t Size = NumValues * FF.getValueSize(); |
| 226 | if (Size < 0) { |
| 227 | recordError(L: FF.getLoc(), Msg: "invalid number of bytes" ); |
| 228 | return 0; |
| 229 | } |
| 230 | return Size; |
| 231 | } |
| 232 | |
| 233 | case MCFragment::FT_Nops: |
| 234 | return cast<MCNopsFragment>(Val: F).getNumBytes(); |
| 235 | |
| 236 | case MCFragment::FT_LEB: |
| 237 | return cast<MCLEBFragment>(Val: F).getContents().size(); |
| 238 | |
| 239 | case MCFragment::FT_BoundaryAlign: |
| 240 | return cast<MCBoundaryAlignFragment>(Val: F).getSize(); |
| 241 | |
| 242 | case MCFragment::FT_SymbolId: |
| 243 | return 4; |
| 244 | |
| 245 | case MCFragment::FT_Align: { |
| 246 | const MCAlignFragment &AF = cast<MCAlignFragment>(Val: F); |
| 247 | unsigned Offset = getFragmentOffset(F: AF); |
| 248 | unsigned Size = offsetToAlignment(Value: Offset, Alignment: AF.getAlignment()); |
| 249 | |
| 250 | // Insert extra Nops for code alignment if the target define |
| 251 | // shouldInsertExtraNopBytesForCodeAlign target hook. |
| 252 | if (AF.getParent()->useCodeAlign() && AF.hasEmitNops() && |
| 253 | getBackend().shouldInsertExtraNopBytesForCodeAlign(AF, Size)) |
| 254 | return Size; |
| 255 | |
| 256 | // If we are padding with nops, force the padding to be larger than the |
| 257 | // minimum nop size. |
| 258 | if (Size > 0 && AF.hasEmitNops()) { |
| 259 | while (Size % getBackend().getMinimumNopSize()) |
| 260 | Size += AF.getAlignment().value(); |
| 261 | } |
| 262 | if (Size > AF.getMaxBytesToEmit()) |
| 263 | return 0; |
| 264 | return Size; |
| 265 | } |
| 266 | |
| 267 | case MCFragment::FT_Org: { |
| 268 | const MCOrgFragment &OF = cast<MCOrgFragment>(Val: F); |
| 269 | MCValue Value; |
| 270 | if (!OF.getOffset().evaluateAsValue(Res&: Value, Asm: *this)) { |
| 271 | recordError(L: OF.getLoc(), Msg: "expected assembly-time absolute expression" ); |
| 272 | return 0; |
| 273 | } |
| 274 | |
| 275 | uint64_t FragmentOffset = getFragmentOffset(F: OF); |
| 276 | int64_t TargetLocation = Value.getConstant(); |
| 277 | if (const auto *SA = Value.getAddSym()) { |
| 278 | uint64_t Val; |
| 279 | if (!getSymbolOffset(S: *SA, Val)) { |
| 280 | recordError(L: OF.getLoc(), Msg: "expected absolute expression" ); |
| 281 | return 0; |
| 282 | } |
| 283 | TargetLocation += Val; |
| 284 | } |
| 285 | int64_t Size = TargetLocation - FragmentOffset; |
| 286 | if (Size < 0 || Size >= 0x40000000) { |
| 287 | recordError(L: OF.getLoc(), Msg: "invalid .org offset '" + Twine(TargetLocation) + |
| 288 | "' (at offset '" + Twine(FragmentOffset) + |
| 289 | "')" ); |
| 290 | return 0; |
| 291 | } |
| 292 | return Size; |
| 293 | } |
| 294 | |
| 295 | case MCFragment::FT_Dwarf: |
| 296 | return cast<MCDwarfLineAddrFragment>(Val: F).getContents().size(); |
| 297 | case MCFragment::FT_DwarfFrame: |
| 298 | return cast<MCDwarfCallFrameFragment>(Val: F).getContents().size(); |
| 299 | case MCFragment::FT_CVInlineLines: |
| 300 | return cast<MCCVInlineLineTableFragment>(Val: F).getContents().size(); |
| 301 | case MCFragment::FT_CVDefRange: |
| 302 | return cast<MCCVDefRangeFragment>(Val: F).getContents().size(); |
| 303 | case MCFragment::FT_PseudoProbe: |
| 304 | return cast<MCPseudoProbeAddrFragment>(Val: F).getContents().size(); |
| 305 | } |
| 306 | |
| 307 | llvm_unreachable("invalid fragment kind" ); |
| 308 | } |
| 309 | |
| 310 | // Compute the amount of padding required before the fragment \p F to |
| 311 | // obey bundling restrictions, where \p FOffset is the fragment's offset in |
| 312 | // its section and \p FSize is the fragment's size. |
| 313 | static uint64_t computeBundlePadding(unsigned BundleSize, |
| 314 | const MCEncodedFragment *F, |
| 315 | uint64_t FOffset, uint64_t FSize) { |
| 316 | uint64_t OffsetInBundle = FOffset & (BundleSize - 1); |
| 317 | uint64_t EndOfFragment = OffsetInBundle + FSize; |
| 318 | |
| 319 | // There are two kinds of bundling restrictions: |
| 320 | // |
| 321 | // 1) For alignToBundleEnd(), add padding to ensure that the fragment will |
| 322 | // *end* on a bundle boundary. |
| 323 | // 2) Otherwise, check if the fragment would cross a bundle boundary. If it |
| 324 | // would, add padding until the end of the bundle so that the fragment |
| 325 | // will start in a new one. |
| 326 | if (F->alignToBundleEnd()) { |
| 327 | // Three possibilities here: |
| 328 | // |
| 329 | // A) The fragment just happens to end at a bundle boundary, so we're good. |
| 330 | // B) The fragment ends before the current bundle boundary: pad it just |
| 331 | // enough to reach the boundary. |
| 332 | // C) The fragment ends after the current bundle boundary: pad it until it |
| 333 | // reaches the end of the next bundle boundary. |
| 334 | // |
| 335 | // Note: this code could be made shorter with some modulo trickery, but it's |
| 336 | // intentionally kept in its more explicit form for simplicity. |
| 337 | if (EndOfFragment == BundleSize) |
| 338 | return 0; |
| 339 | else if (EndOfFragment < BundleSize) |
| 340 | return BundleSize - EndOfFragment; |
| 341 | else { // EndOfFragment > BundleSize |
| 342 | return 2 * BundleSize - EndOfFragment; |
| 343 | } |
| 344 | } else if (OffsetInBundle > 0 && EndOfFragment > BundleSize) |
| 345 | return BundleSize - OffsetInBundle; |
| 346 | else |
| 347 | return 0; |
| 348 | } |
| 349 | |
| 350 | void MCAssembler::layoutBundle(MCFragment *Prev, MCFragment *F) const { |
| 351 | // If bundling is enabled and this fragment has instructions in it, it has to |
| 352 | // obey the bundling restrictions. With padding, we'll have: |
| 353 | // |
| 354 | // |
| 355 | // BundlePadding |
| 356 | // ||| |
| 357 | // ------------------------------------- |
| 358 | // Prev |##########| F | |
| 359 | // ------------------------------------- |
| 360 | // ^ |
| 361 | // | |
| 362 | // F->Offset |
| 363 | // |
| 364 | // The fragment's offset will point to after the padding, and its computed |
| 365 | // size won't include the padding. |
| 366 | // |
| 367 | // ".align N" is an example of a directive that introduces multiple |
| 368 | // fragments. We could add a special case to handle ".align N" by emitting |
| 369 | // within-fragment padding (which would produce less padding when N is less |
| 370 | // than the bundle size), but for now we don't. |
| 371 | // |
| 372 | assert(isa<MCEncodedFragment>(F) && |
| 373 | "Only MCEncodedFragment implementations have instructions" ); |
| 374 | MCEncodedFragment *EF = cast<MCEncodedFragment>(Val: F); |
| 375 | uint64_t FSize = computeFragmentSize(F: *EF); |
| 376 | |
| 377 | if (FSize > getBundleAlignSize()) |
| 378 | report_fatal_error(reason: "Fragment can't be larger than a bundle size" ); |
| 379 | |
| 380 | uint64_t RequiredBundlePadding = |
| 381 | computeBundlePadding(BundleSize: getBundleAlignSize(), F: EF, FOffset: EF->Offset, FSize); |
| 382 | if (RequiredBundlePadding > UINT8_MAX) |
| 383 | report_fatal_error(reason: "Padding cannot exceed 255 bytes" ); |
| 384 | EF->setBundlePadding(static_cast<uint8_t>(RequiredBundlePadding)); |
| 385 | EF->Offset += RequiredBundlePadding; |
| 386 | if (auto *DF = dyn_cast_or_null<MCDataFragment>(Val: Prev)) |
| 387 | if (DF->getContents().empty()) |
| 388 | DF->Offset = EF->Offset; |
| 389 | } |
| 390 | |
| 391 | // Simple getSymbolOffset helper for the non-variable case. |
| 392 | static bool getLabelOffset(const MCAssembler &Asm, const MCSymbol &S, |
| 393 | bool ReportError, uint64_t &Val) { |
| 394 | if (!S.getFragment()) { |
| 395 | if (ReportError) |
| 396 | reportFatalUsageError(reason: "cannot evaluate undefined symbol '" + S.getName() + |
| 397 | "'" ); |
| 398 | return false; |
| 399 | } |
| 400 | Val = Asm.getFragmentOffset(F: *S.getFragment()) + S.getOffset(); |
| 401 | return true; |
| 402 | } |
| 403 | |
| 404 | static bool getSymbolOffsetImpl(const MCAssembler &Asm, const MCSymbol &S, |
| 405 | bool ReportError, uint64_t &Val) { |
| 406 | if (!S.isVariable()) |
| 407 | return getLabelOffset(Asm, S, ReportError, Val); |
| 408 | |
| 409 | // If SD is a variable, evaluate it. |
| 410 | MCValue Target; |
| 411 | if (!S.getVariableValue()->evaluateAsValue(Res&: Target, Asm)) |
| 412 | reportFatalUsageError(reason: "cannot evaluate equated symbol '" + S.getName() + |
| 413 | "'" ); |
| 414 | |
| 415 | uint64_t Offset = Target.getConstant(); |
| 416 | |
| 417 | const MCSymbol *A = Target.getAddSym(); |
| 418 | if (A) { |
| 419 | uint64_t ValA; |
| 420 | // FIXME: On most platforms, `Target`'s component symbols are labels from |
| 421 | // having been simplified during evaluation, but on Mach-O they can be |
| 422 | // variables due to PR19203. This, and the line below for `B` can be |
| 423 | // restored to call `getLabelOffset` when PR19203 is fixed. |
| 424 | if (!getSymbolOffsetImpl(Asm, S: *A, ReportError, Val&: ValA)) |
| 425 | return false; |
| 426 | Offset += ValA; |
| 427 | } |
| 428 | |
| 429 | const MCSymbol *B = Target.getSubSym(); |
| 430 | if (B) { |
| 431 | uint64_t ValB; |
| 432 | if (!getSymbolOffsetImpl(Asm, S: *B, ReportError, Val&: ValB)) |
| 433 | return false; |
| 434 | Offset -= ValB; |
| 435 | } |
| 436 | |
| 437 | Val = Offset; |
| 438 | return true; |
| 439 | } |
| 440 | |
| 441 | bool MCAssembler::getSymbolOffset(const MCSymbol &S, uint64_t &Val) const { |
| 442 | return getSymbolOffsetImpl(Asm: *this, S, ReportError: false, Val); |
| 443 | } |
| 444 | |
| 445 | uint64_t MCAssembler::getSymbolOffset(const MCSymbol &S) const { |
| 446 | uint64_t Val; |
| 447 | getSymbolOffsetImpl(Asm: *this, S, ReportError: true, Val); |
| 448 | return Val; |
| 449 | } |
| 450 | |
| 451 | const MCSymbol *MCAssembler::getBaseSymbol(const MCSymbol &Symbol) const { |
| 452 | assert(HasLayout); |
| 453 | if (!Symbol.isVariable()) |
| 454 | return &Symbol; |
| 455 | |
| 456 | const MCExpr *Expr = Symbol.getVariableValue(); |
| 457 | MCValue Value; |
| 458 | if (!Expr->evaluateAsValue(Res&: Value, Asm: *this)) { |
| 459 | reportError(L: Expr->getLoc(), Msg: "expression could not be evaluated" ); |
| 460 | return nullptr; |
| 461 | } |
| 462 | |
| 463 | const MCSymbol *SymB = Value.getSubSym(); |
| 464 | if (SymB) { |
| 465 | reportError(L: Expr->getLoc(), |
| 466 | Msg: Twine("symbol '" ) + SymB->getName() + |
| 467 | "' could not be evaluated in a subtraction expression" ); |
| 468 | return nullptr; |
| 469 | } |
| 470 | |
| 471 | const MCSymbol *A = Value.getAddSym(); |
| 472 | if (!A) |
| 473 | return nullptr; |
| 474 | |
| 475 | const MCSymbol &ASym = *A; |
| 476 | if (ASym.isCommon()) { |
| 477 | reportError(L: Expr->getLoc(), Msg: "Common symbol '" + ASym.getName() + |
| 478 | "' cannot be used in assignment expr" ); |
| 479 | return nullptr; |
| 480 | } |
| 481 | |
| 482 | return &ASym; |
| 483 | } |
| 484 | |
| 485 | uint64_t MCAssembler::getSectionAddressSize(const MCSection &Sec) const { |
| 486 | assert(HasLayout); |
| 487 | // The size is the last fragment's end offset. |
| 488 | const MCFragment &F = *Sec.curFragList()->Tail; |
| 489 | return getFragmentOffset(F) + computeFragmentSize(F); |
| 490 | } |
| 491 | |
| 492 | uint64_t MCAssembler::getSectionFileSize(const MCSection &Sec) const { |
| 493 | // Virtual sections have no file size. |
| 494 | if (Sec.isVirtualSection()) |
| 495 | return 0; |
| 496 | return getSectionAddressSize(Sec); |
| 497 | } |
| 498 | |
| 499 | bool MCAssembler::registerSymbol(const MCSymbol &Symbol) { |
| 500 | bool Changed = !Symbol.isRegistered(); |
| 501 | if (Changed) { |
| 502 | Symbol.setIsRegistered(true); |
| 503 | Symbols.push_back(Elt: &Symbol); |
| 504 | } |
| 505 | return Changed; |
| 506 | } |
| 507 | |
| 508 | void MCAssembler::writeFragmentPadding(raw_ostream &OS, |
| 509 | const MCEncodedFragment &EF, |
| 510 | uint64_t FSize) const { |
| 511 | assert(getBackendPtr() && "Expected assembler backend" ); |
| 512 | // Should NOP padding be written out before this fragment? |
| 513 | unsigned BundlePadding = EF.getBundlePadding(); |
| 514 | if (BundlePadding > 0) { |
| 515 | assert(isBundlingEnabled() && |
| 516 | "Writing bundle padding with disabled bundling" ); |
| 517 | assert(EF.hasInstructions() && |
| 518 | "Writing bundle padding for a fragment without instructions" ); |
| 519 | |
| 520 | unsigned TotalLength = BundlePadding + static_cast<unsigned>(FSize); |
| 521 | const MCSubtargetInfo *STI = EF.getSubtargetInfo(); |
| 522 | if (EF.alignToBundleEnd() && TotalLength > getBundleAlignSize()) { |
| 523 | // If the padding itself crosses a bundle boundary, it must be emitted |
| 524 | // in 2 pieces, since even nop instructions must not cross boundaries. |
| 525 | // v--------------v <- BundleAlignSize |
| 526 | // v---------v <- BundlePadding |
| 527 | // ---------------------------- |
| 528 | // | Prev |####|####| F | |
| 529 | // ---------------------------- |
| 530 | // ^-------------------^ <- TotalLength |
| 531 | unsigned DistanceToBoundary = TotalLength - getBundleAlignSize(); |
| 532 | if (!getBackend().writeNopData(OS, Count: DistanceToBoundary, STI)) |
| 533 | report_fatal_error(reason: "unable to write NOP sequence of " + |
| 534 | Twine(DistanceToBoundary) + " bytes" ); |
| 535 | BundlePadding -= DistanceToBoundary; |
| 536 | } |
| 537 | if (!getBackend().writeNopData(OS, Count: BundlePadding, STI)) |
| 538 | report_fatal_error(reason: "unable to write NOP sequence of " + |
| 539 | Twine(BundlePadding) + " bytes" ); |
| 540 | } |
| 541 | } |
| 542 | |
| 543 | /// Write the fragment \p F to the output file. |
| 544 | static void writeFragment(raw_ostream &OS, const MCAssembler &Asm, |
| 545 | const MCFragment &F) { |
| 546 | // FIXME: Embed in fragments instead? |
| 547 | uint64_t FragmentSize = Asm.computeFragmentSize(F); |
| 548 | |
| 549 | llvm::endianness Endian = Asm.getBackend().Endian; |
| 550 | |
| 551 | if (const MCEncodedFragment *EF = dyn_cast<MCEncodedFragment>(Val: &F)) |
| 552 | Asm.writeFragmentPadding(OS, EF: *EF, FSize: FragmentSize); |
| 553 | |
| 554 | // This variable (and its dummy usage) is to participate in the assert at |
| 555 | // the end of the function. |
| 556 | uint64_t Start = OS.tell(); |
| 557 | (void) Start; |
| 558 | |
| 559 | ++stats::EmittedFragments; |
| 560 | |
| 561 | switch (F.getKind()) { |
| 562 | case MCFragment::FT_Align: { |
| 563 | ++stats::EmittedAlignFragments; |
| 564 | const MCAlignFragment &AF = cast<MCAlignFragment>(Val: F); |
| 565 | assert(AF.getValueSize() && "Invalid virtual align in concrete fragment!" ); |
| 566 | |
| 567 | uint64_t Count = FragmentSize / AF.getValueSize(); |
| 568 | |
| 569 | // FIXME: This error shouldn't actually occur (the front end should emit |
| 570 | // multiple .align directives to enforce the semantics it wants), but is |
| 571 | // severe enough that we want to report it. How to handle this? |
| 572 | if (Count * AF.getValueSize() != FragmentSize) |
| 573 | report_fatal_error(reason: "undefined .align directive, value size '" + |
| 574 | Twine(AF.getValueSize()) + |
| 575 | "' is not a divisor of padding size '" + |
| 576 | Twine(FragmentSize) + "'" ); |
| 577 | |
| 578 | // See if we are aligning with nops, and if so do that first to try to fill |
| 579 | // the Count bytes. Then if that did not fill any bytes or there are any |
| 580 | // bytes left to fill use the Value and ValueSize to fill the rest. |
| 581 | // If we are aligning with nops, ask that target to emit the right data. |
| 582 | if (AF.hasEmitNops()) { |
| 583 | if (!Asm.getBackend().writeNopData(OS, Count, STI: AF.getSubtargetInfo())) |
| 584 | report_fatal_error(reason: "unable to write nop sequence of " + |
| 585 | Twine(Count) + " bytes" ); |
| 586 | break; |
| 587 | } |
| 588 | |
| 589 | // Otherwise, write out in multiples of the value size. |
| 590 | for (uint64_t i = 0; i != Count; ++i) { |
| 591 | switch (AF.getValueSize()) { |
| 592 | default: llvm_unreachable("Invalid size!" ); |
| 593 | case 1: OS << char(AF.getValue()); break; |
| 594 | case 2: |
| 595 | support::endian::write<uint16_t>(os&: OS, value: AF.getValue(), endian: Endian); |
| 596 | break; |
| 597 | case 4: |
| 598 | support::endian::write<uint32_t>(os&: OS, value: AF.getValue(), endian: Endian); |
| 599 | break; |
| 600 | case 8: |
| 601 | support::endian::write<uint64_t>(os&: OS, value: AF.getValue(), endian: Endian); |
| 602 | break; |
| 603 | } |
| 604 | } |
| 605 | break; |
| 606 | } |
| 607 | |
| 608 | case MCFragment::FT_Data: |
| 609 | ++stats::EmittedDataFragments; |
| 610 | OS << StringRef(cast<MCDataFragment>(Val: F).getContents().data(), |
| 611 | cast<MCDataFragment>(Val: F).getContents().size()); |
| 612 | break; |
| 613 | |
| 614 | case MCFragment::FT_Relaxable: |
| 615 | ++stats::EmittedRelaxableFragments; |
| 616 | OS << StringRef(cast<MCRelaxableFragment>(Val: F).getContents().data(), |
| 617 | cast<MCRelaxableFragment>(Val: F).getContents().size()); |
| 618 | break; |
| 619 | |
| 620 | case MCFragment::FT_Fill: { |
| 621 | ++stats::EmittedFillFragments; |
| 622 | const MCFillFragment &FF = cast<MCFillFragment>(Val: F); |
| 623 | uint64_t V = FF.getValue(); |
| 624 | unsigned VSize = FF.getValueSize(); |
| 625 | const unsigned MaxChunkSize = 16; |
| 626 | char Data[MaxChunkSize]; |
| 627 | assert(0 < VSize && VSize <= MaxChunkSize && "Illegal fragment fill size" ); |
| 628 | // Duplicate V into Data as byte vector to reduce number of |
| 629 | // writes done. As such, do endian conversion here. |
| 630 | for (unsigned I = 0; I != VSize; ++I) { |
| 631 | unsigned index = Endian == llvm::endianness::little ? I : (VSize - I - 1); |
| 632 | Data[I] = uint8_t(V >> (index * 8)); |
| 633 | } |
| 634 | for (unsigned I = VSize; I < MaxChunkSize; ++I) |
| 635 | Data[I] = Data[I - VSize]; |
| 636 | |
| 637 | // Set to largest multiple of VSize in Data. |
| 638 | const unsigned NumPerChunk = MaxChunkSize / VSize; |
| 639 | // Set ChunkSize to largest multiple of VSize in Data |
| 640 | const unsigned ChunkSize = VSize * NumPerChunk; |
| 641 | |
| 642 | // Do copies by chunk. |
| 643 | StringRef Ref(Data, ChunkSize); |
| 644 | for (uint64_t I = 0, E = FragmentSize / ChunkSize; I != E; ++I) |
| 645 | OS << Ref; |
| 646 | |
| 647 | // do remainder if needed. |
| 648 | unsigned TrailingCount = FragmentSize % ChunkSize; |
| 649 | if (TrailingCount) |
| 650 | OS.write(Ptr: Data, Size: TrailingCount); |
| 651 | break; |
| 652 | } |
| 653 | |
| 654 | case MCFragment::FT_Nops: { |
| 655 | ++stats::EmittedNopsFragments; |
| 656 | const MCNopsFragment &NF = cast<MCNopsFragment>(Val: F); |
| 657 | |
| 658 | int64_t NumBytes = NF.getNumBytes(); |
| 659 | int64_t ControlledNopLength = NF.getControlledNopLength(); |
| 660 | int64_t MaximumNopLength = |
| 661 | Asm.getBackend().getMaximumNopSize(STI: *NF.getSubtargetInfo()); |
| 662 | |
| 663 | assert(NumBytes > 0 && "Expected positive NOPs fragment size" ); |
| 664 | assert(ControlledNopLength >= 0 && "Expected non-negative NOP size" ); |
| 665 | |
| 666 | if (ControlledNopLength > MaximumNopLength) { |
| 667 | Asm.reportError(L: NF.getLoc(), Msg: "illegal NOP size " + |
| 668 | std::to_string(val: ControlledNopLength) + |
| 669 | ". (expected within [0, " + |
| 670 | std::to_string(val: MaximumNopLength) + "])" ); |
| 671 | // Clamp the NOP length as reportError does not stop the execution |
| 672 | // immediately. |
| 673 | ControlledNopLength = MaximumNopLength; |
| 674 | } |
| 675 | |
| 676 | // Use maximum value if the size of each NOP is not specified |
| 677 | if (!ControlledNopLength) |
| 678 | ControlledNopLength = MaximumNopLength; |
| 679 | |
| 680 | while (NumBytes) { |
| 681 | uint64_t NumBytesToEmit = |
| 682 | (uint64_t)std::min(a: NumBytes, b: ControlledNopLength); |
| 683 | assert(NumBytesToEmit && "try to emit empty NOP instruction" ); |
| 684 | if (!Asm.getBackend().writeNopData(OS, Count: NumBytesToEmit, |
| 685 | STI: NF.getSubtargetInfo())) { |
| 686 | report_fatal_error(reason: "unable to write nop sequence of the remaining " + |
| 687 | Twine(NumBytesToEmit) + " bytes" ); |
| 688 | break; |
| 689 | } |
| 690 | NumBytes -= NumBytesToEmit; |
| 691 | } |
| 692 | break; |
| 693 | } |
| 694 | |
| 695 | case MCFragment::FT_LEB: { |
| 696 | const MCLEBFragment &LF = cast<MCLEBFragment>(Val: F); |
| 697 | OS << StringRef(LF.getContents().data(), LF.getContents().size()); |
| 698 | break; |
| 699 | } |
| 700 | |
| 701 | case MCFragment::FT_BoundaryAlign: { |
| 702 | const MCBoundaryAlignFragment &BF = cast<MCBoundaryAlignFragment>(Val: F); |
| 703 | if (!Asm.getBackend().writeNopData(OS, Count: FragmentSize, STI: BF.getSubtargetInfo())) |
| 704 | report_fatal_error(reason: "unable to write nop sequence of " + |
| 705 | Twine(FragmentSize) + " bytes" ); |
| 706 | break; |
| 707 | } |
| 708 | |
| 709 | case MCFragment::FT_SymbolId: { |
| 710 | const MCSymbolIdFragment &SF = cast<MCSymbolIdFragment>(Val: F); |
| 711 | support::endian::write<uint32_t>(os&: OS, value: SF.getSymbol()->getIndex(), endian: Endian); |
| 712 | break; |
| 713 | } |
| 714 | |
| 715 | case MCFragment::FT_Org: { |
| 716 | ++stats::EmittedOrgFragments; |
| 717 | const MCOrgFragment &OF = cast<MCOrgFragment>(Val: F); |
| 718 | |
| 719 | for (uint64_t i = 0, e = FragmentSize; i != e; ++i) |
| 720 | OS << char(OF.getValue()); |
| 721 | |
| 722 | break; |
| 723 | } |
| 724 | |
| 725 | case MCFragment::FT_Dwarf: { |
| 726 | const MCDwarfLineAddrFragment &OF = cast<MCDwarfLineAddrFragment>(Val: F); |
| 727 | OS << StringRef(OF.getContents().data(), OF.getContents().size()); |
| 728 | break; |
| 729 | } |
| 730 | case MCFragment::FT_DwarfFrame: { |
| 731 | const MCDwarfCallFrameFragment &CF = cast<MCDwarfCallFrameFragment>(Val: F); |
| 732 | OS << StringRef(CF.getContents().data(), CF.getContents().size()); |
| 733 | break; |
| 734 | } |
| 735 | case MCFragment::FT_CVInlineLines: { |
| 736 | const auto &OF = cast<MCCVInlineLineTableFragment>(Val: F); |
| 737 | OS << StringRef(OF.getContents().data(), OF.getContents().size()); |
| 738 | break; |
| 739 | } |
| 740 | case MCFragment::FT_CVDefRange: { |
| 741 | const auto &DRF = cast<MCCVDefRangeFragment>(Val: F); |
| 742 | OS << StringRef(DRF.getContents().data(), DRF.getContents().size()); |
| 743 | break; |
| 744 | } |
| 745 | case MCFragment::FT_PseudoProbe: { |
| 746 | const MCPseudoProbeAddrFragment &PF = cast<MCPseudoProbeAddrFragment>(Val: F); |
| 747 | OS << StringRef(PF.getContents().data(), PF.getContents().size()); |
| 748 | break; |
| 749 | } |
| 750 | } |
| 751 | |
| 752 | assert(OS.tell() - Start == FragmentSize && |
| 753 | "The stream should advance by fragment size" ); |
| 754 | } |
| 755 | |
| 756 | void MCAssembler::writeSectionData(raw_ostream &OS, |
| 757 | const MCSection *Sec) const { |
| 758 | assert(getBackendPtr() && "Expected assembler backend" ); |
| 759 | |
| 760 | // Ignore virtual sections. |
| 761 | if (Sec->isVirtualSection()) { |
| 762 | assert(getSectionFileSize(*Sec) == 0 && "Invalid size for section!" ); |
| 763 | |
| 764 | // Check that contents are only things legal inside a virtual section. |
| 765 | for (const MCFragment &F : *Sec) { |
| 766 | switch (F.getKind()) { |
| 767 | default: llvm_unreachable("Invalid fragment in virtual section!" ); |
| 768 | case MCFragment::FT_Data: { |
| 769 | // Check that we aren't trying to write a non-zero contents (or fixups) |
| 770 | // into a virtual section. This is to support clients which use standard |
| 771 | // directives to fill the contents of virtual sections. |
| 772 | const MCDataFragment &DF = cast<MCDataFragment>(Val: F); |
| 773 | if (DF.getFixups().size()) |
| 774 | reportError(L: SMLoc(), Msg: Sec->getVirtualSectionKind() + " section '" + |
| 775 | Sec->getName() + "' cannot have fixups" ); |
| 776 | for (char C : DF.getContents()) |
| 777 | if (C) { |
| 778 | reportError(L: SMLoc(), Msg: Sec->getVirtualSectionKind() + " section '" + |
| 779 | Sec->getName() + |
| 780 | "' cannot have non-zero initializers" ); |
| 781 | break; |
| 782 | } |
| 783 | break; |
| 784 | } |
| 785 | case MCFragment::FT_Align: |
| 786 | // Check that we aren't trying to write a non-zero value into a virtual |
| 787 | // section. |
| 788 | assert((cast<MCAlignFragment>(F).getValueSize() == 0 || |
| 789 | cast<MCAlignFragment>(F).getValue() == 0) && |
| 790 | "Invalid align in virtual section!" ); |
| 791 | break; |
| 792 | case MCFragment::FT_Fill: |
| 793 | assert((cast<MCFillFragment>(F).getValue() == 0) && |
| 794 | "Invalid fill in virtual section!" ); |
| 795 | break; |
| 796 | case MCFragment::FT_Org: |
| 797 | break; |
| 798 | } |
| 799 | } |
| 800 | |
| 801 | return; |
| 802 | } |
| 803 | |
| 804 | uint64_t Start = OS.tell(); |
| 805 | (void)Start; |
| 806 | |
| 807 | for (const MCFragment &F : *Sec) |
| 808 | writeFragment(OS, Asm: *this, F); |
| 809 | |
| 810 | flushPendingErrors(); |
| 811 | assert(getContext().hadError() || |
| 812 | OS.tell() - Start == getSectionAddressSize(*Sec)); |
| 813 | } |
| 814 | |
| 815 | void MCAssembler::layout() { |
| 816 | assert(getBackendPtr() && "Expected assembler backend" ); |
| 817 | DEBUG_WITH_TYPE("mc-dump-pre" , { |
| 818 | errs() << "assembler backend - pre-layout\n--\n" ; |
| 819 | dump(); |
| 820 | }); |
| 821 | |
| 822 | // Assign section ordinals. |
| 823 | unsigned SectionIndex = 0; |
| 824 | for (MCSection &Sec : *this) { |
| 825 | Sec.setOrdinal(SectionIndex++); |
| 826 | |
| 827 | // Chain together fragments from all subsections. |
| 828 | if (Sec.Subsections.size() > 1) { |
| 829 | MCDataFragment Dummy; |
| 830 | MCFragment *Tail = &Dummy; |
| 831 | for (auto &[_, List] : Sec.Subsections) { |
| 832 | assert(List.Head); |
| 833 | Tail->Next = List.Head; |
| 834 | Tail = List.Tail; |
| 835 | } |
| 836 | Sec.Subsections.clear(); |
| 837 | Sec.Subsections.push_back(Elt: {0u, {.Head: Dummy.getNext(), .Tail: Tail}}); |
| 838 | Sec.CurFragList = &Sec.Subsections[0].second; |
| 839 | |
| 840 | unsigned FragmentIndex = 0; |
| 841 | for (MCFragment &Frag : Sec) |
| 842 | Frag.setLayoutOrder(FragmentIndex++); |
| 843 | } |
| 844 | } |
| 845 | |
| 846 | // Layout until everything fits. |
| 847 | this->HasLayout = true; |
| 848 | for (MCSection &Sec : *this) |
| 849 | layoutSection(Sec); |
| 850 | while (relaxOnce()) |
| 851 | if (getContext().hadError()) |
| 852 | return; |
| 853 | |
| 854 | // Some targets might want to adjust fragment offsets. If so, perform another |
| 855 | // layout iteration. |
| 856 | if (getBackend().finishLayout(Asm: *this)) |
| 857 | for (MCSection &Sec : *this) |
| 858 | layoutSection(Sec); |
| 859 | |
| 860 | flushPendingErrors(); |
| 861 | |
| 862 | DEBUG_WITH_TYPE("mc-dump" , { |
| 863 | errs() << "assembler backend - final-layout\n--\n" ; |
| 864 | dump(); }); |
| 865 | |
| 866 | // Allow the object writer a chance to perform post-layout binding (for |
| 867 | // example, to set the index fields in the symbol data). |
| 868 | getWriter().executePostLayoutBinding(); |
| 869 | |
| 870 | // Fragment sizes are finalized. For RISC-V linker relaxation, this flag |
| 871 | // helps check whether a PC-relative fixup is fully resolved. |
| 872 | this->HasFinalLayout = true; |
| 873 | |
| 874 | // Evaluate and apply the fixups, generating relocation entries as necessary. |
| 875 | for (MCSection &Sec : *this) { |
| 876 | for (MCFragment &Frag : Sec) { |
| 877 | // Process fragments with fixups here. |
| 878 | if (auto *F = dyn_cast<MCEncodedFragment>(Val: &Frag)) { |
| 879 | auto Contents = F->getContents(); |
| 880 | for (MCFixup &Fixup : F->getFixups()) { |
| 881 | uint64_t FixedValue; |
| 882 | MCValue Target; |
| 883 | evaluateFixup(F: Frag, Fixup, Target, Value&: FixedValue, |
| 884 | /*RecordReloc=*/true, Contents); |
| 885 | } |
| 886 | } else if (auto *AF = dyn_cast<MCAlignFragment>(Val: &Frag)) { |
| 887 | // For RISC-V linker relaxation, an alignment relocation might be |
| 888 | // needed. |
| 889 | if (AF->hasEmitNops()) |
| 890 | getBackend().shouldInsertFixupForCodeAlign(Asm&: *this, AF&: *AF); |
| 891 | } |
| 892 | } |
| 893 | } |
| 894 | } |
| 895 | |
| 896 | void MCAssembler::Finish() { |
| 897 | layout(); |
| 898 | |
| 899 | // Write the object file. |
| 900 | stats::ObjectBytes += getWriter().writeObject(); |
| 901 | |
| 902 | HasLayout = false; |
| 903 | assert(PendingErrors.empty()); |
| 904 | } |
| 905 | |
| 906 | bool MCAssembler::fixupNeedsRelaxation(const MCRelaxableFragment &F, |
| 907 | const MCFixup &Fixup) const { |
| 908 | assert(getBackendPtr() && "Expected assembler backend" ); |
| 909 | MCValue Target; |
| 910 | uint64_t Value; |
| 911 | bool Resolved = evaluateFixup(F, Fixup&: const_cast<MCFixup &>(Fixup), Target, Value, |
| 912 | /*RecordReloc=*/false, Contents: {}); |
| 913 | return getBackend().fixupNeedsRelaxationAdvanced(Fixup, Target, Value, |
| 914 | Resolved); |
| 915 | } |
| 916 | |
| 917 | bool MCAssembler::fragmentNeedsRelaxation(const MCRelaxableFragment &F) const { |
| 918 | assert(getBackendPtr() && "Expected assembler backend" ); |
| 919 | // If this inst doesn't ever need relaxation, ignore it. This occurs when we |
| 920 | // are intentionally pushing out inst fragments, or because we relaxed a |
| 921 | // previous instruction to one that doesn't need relaxation. |
| 922 | if (!getBackend().mayNeedRelaxation(Inst: F.getInst(), STI: *F.getSubtargetInfo())) |
| 923 | return false; |
| 924 | |
| 925 | for (const MCFixup &Fixup : F.getFixups()) |
| 926 | if (fixupNeedsRelaxation(F, Fixup)) |
| 927 | return true; |
| 928 | |
| 929 | return false; |
| 930 | } |
| 931 | |
| 932 | bool MCAssembler::relaxInstruction(MCRelaxableFragment &F) { |
| 933 | assert(getEmitterPtr() && |
| 934 | "Expected CodeEmitter defined for relaxInstruction" ); |
| 935 | if (!fragmentNeedsRelaxation(F)) |
| 936 | return false; |
| 937 | |
| 938 | ++stats::RelaxedInstructions; |
| 939 | |
| 940 | // FIXME-PERF: We could immediately lower out instructions if we can tell |
| 941 | // they are fully resolved, to avoid retesting on later passes. |
| 942 | |
| 943 | // Relax the fragment. |
| 944 | |
| 945 | MCInst Relaxed = F.getInst(); |
| 946 | getBackend().relaxInstruction(Inst&: Relaxed, STI: *F.getSubtargetInfo()); |
| 947 | |
| 948 | // Encode the new instruction. |
| 949 | F.setInst(Relaxed); |
| 950 | SmallVector<char, 16> Data; |
| 951 | SmallVector<MCFixup, 1> Fixups; |
| 952 | getEmitter().encodeInstruction(Inst: Relaxed, CB&: Data, Fixups, STI: *F.getSubtargetInfo()); |
| 953 | F.setContents(Data); |
| 954 | F.setFixups(Fixups); |
| 955 | return true; |
| 956 | } |
| 957 | |
| 958 | bool MCAssembler::relaxLEB(MCLEBFragment &LF) { |
| 959 | const unsigned OldSize = static_cast<unsigned>(LF.getContents().size()); |
| 960 | unsigned PadTo = OldSize; |
| 961 | int64_t Value; |
| 962 | LF.clearFixups(); |
| 963 | // Use evaluateKnownAbsolute for Mach-O as a hack: .subsections_via_symbols |
| 964 | // requires that .uleb128 A-B is foldable where A and B reside in different |
| 965 | // fragments. This is used by __gcc_except_table. |
| 966 | bool Abs = getWriter().getSubsectionsViaSymbols() |
| 967 | ? LF.getValue().evaluateKnownAbsolute(Res&: Value, Asm: *this) |
| 968 | : LF.getValue().evaluateAsAbsolute(Res&: Value, Asm: *this); |
| 969 | if (!Abs) { |
| 970 | bool Relaxed, UseZeroPad; |
| 971 | std::tie(args&: Relaxed, args&: UseZeroPad) = getBackend().relaxLEB128(LF, Value); |
| 972 | if (!Relaxed) { |
| 973 | reportError(L: LF.getValue().getLoc(), |
| 974 | Msg: Twine(LF.isSigned() ? ".s" : ".u" ) + |
| 975 | "leb128 expression is not absolute" ); |
| 976 | LF.setValue(MCConstantExpr::create(Value: 0, Ctx&: Context)); |
| 977 | } |
| 978 | uint8_t Tmp[10]; // maximum size: ceil(64/7) |
| 979 | PadTo = std::max(a: PadTo, b: encodeULEB128(Value: uint64_t(Value), p: Tmp)); |
| 980 | if (UseZeroPad) |
| 981 | Value = 0; |
| 982 | } |
| 983 | uint8_t Data[16]; |
| 984 | size_t Size = 0; |
| 985 | // The compiler can generate EH table assembly that is impossible to assemble |
| 986 | // without either adding padding to an LEB fragment or adding extra padding |
| 987 | // to a later alignment fragment. To accommodate such tables, relaxation can |
| 988 | // only increase an LEB fragment size here, not decrease it. See PR35809. |
| 989 | if (LF.isSigned()) |
| 990 | Size = encodeSLEB128(Value, p: Data, PadTo); |
| 991 | else |
| 992 | Size = encodeULEB128(Value, p: Data, PadTo); |
| 993 | LF.setContents({reinterpret_cast<char *>(Data), Size}); |
| 994 | return OldSize != Size; |
| 995 | } |
| 996 | |
| 997 | /// Check if the branch crosses the boundary. |
| 998 | /// |
| 999 | /// \param StartAddr start address of the fused/unfused branch. |
| 1000 | /// \param Size size of the fused/unfused branch. |
| 1001 | /// \param BoundaryAlignment alignment requirement of the branch. |
| 1002 | /// \returns true if the branch cross the boundary. |
| 1003 | static bool mayCrossBoundary(uint64_t StartAddr, uint64_t Size, |
| 1004 | Align BoundaryAlignment) { |
| 1005 | uint64_t EndAddr = StartAddr + Size; |
| 1006 | return (StartAddr >> Log2(A: BoundaryAlignment)) != |
| 1007 | ((EndAddr - 1) >> Log2(A: BoundaryAlignment)); |
| 1008 | } |
| 1009 | |
| 1010 | /// Check if the branch is against the boundary. |
| 1011 | /// |
| 1012 | /// \param StartAddr start address of the fused/unfused branch. |
| 1013 | /// \param Size size of the fused/unfused branch. |
| 1014 | /// \param BoundaryAlignment alignment requirement of the branch. |
| 1015 | /// \returns true if the branch is against the boundary. |
| 1016 | static bool isAgainstBoundary(uint64_t StartAddr, uint64_t Size, |
| 1017 | Align BoundaryAlignment) { |
| 1018 | uint64_t EndAddr = StartAddr + Size; |
| 1019 | return (EndAddr & (BoundaryAlignment.value() - 1)) == 0; |
| 1020 | } |
| 1021 | |
| 1022 | /// Check if the branch needs padding. |
| 1023 | /// |
| 1024 | /// \param StartAddr start address of the fused/unfused branch. |
| 1025 | /// \param Size size of the fused/unfused branch. |
| 1026 | /// \param BoundaryAlignment alignment requirement of the branch. |
| 1027 | /// \returns true if the branch needs padding. |
| 1028 | static bool needPadding(uint64_t StartAddr, uint64_t Size, |
| 1029 | Align BoundaryAlignment) { |
| 1030 | return mayCrossBoundary(StartAddr, Size, BoundaryAlignment) || |
| 1031 | isAgainstBoundary(StartAddr, Size, BoundaryAlignment); |
| 1032 | } |
| 1033 | |
| 1034 | bool MCAssembler::relaxBoundaryAlign(MCBoundaryAlignFragment &BF) { |
| 1035 | // BoundaryAlignFragment that doesn't need to align any fragment should not be |
| 1036 | // relaxed. |
| 1037 | if (!BF.getLastFragment()) |
| 1038 | return false; |
| 1039 | |
| 1040 | uint64_t AlignedOffset = getFragmentOffset(F: BF); |
| 1041 | uint64_t AlignedSize = 0; |
| 1042 | for (const MCFragment *F = BF.getNext();; F = F->getNext()) { |
| 1043 | AlignedSize += computeFragmentSize(F: *F); |
| 1044 | if (F == BF.getLastFragment()) |
| 1045 | break; |
| 1046 | } |
| 1047 | |
| 1048 | Align BoundaryAlignment = BF.getAlignment(); |
| 1049 | uint64_t NewSize = needPadding(StartAddr: AlignedOffset, Size: AlignedSize, BoundaryAlignment) |
| 1050 | ? offsetToAlignment(Value: AlignedOffset, Alignment: BoundaryAlignment) |
| 1051 | : 0U; |
| 1052 | if (NewSize == BF.getSize()) |
| 1053 | return false; |
| 1054 | BF.setSize(NewSize); |
| 1055 | return true; |
| 1056 | } |
| 1057 | |
| 1058 | bool MCAssembler::relaxDwarfLineAddr(MCDwarfLineAddrFragment &DF) { |
| 1059 | bool WasRelaxed; |
| 1060 | if (getBackend().relaxDwarfLineAddr(DF, WasRelaxed)) |
| 1061 | return WasRelaxed; |
| 1062 | |
| 1063 | MCContext &Context = getContext(); |
| 1064 | auto OldSize = DF.getContents().size(); |
| 1065 | int64_t AddrDelta; |
| 1066 | bool Abs = DF.getAddrDelta().evaluateKnownAbsolute(Res&: AddrDelta, Asm: *this); |
| 1067 | assert(Abs && "We created a line delta with an invalid expression" ); |
| 1068 | (void)Abs; |
| 1069 | int64_t LineDelta; |
| 1070 | LineDelta = DF.getLineDelta(); |
| 1071 | SmallVector<char, 8> Data; |
| 1072 | |
| 1073 | MCDwarfLineAddr::encode(Context, Params: getDWARFLinetableParams(), LineDelta, |
| 1074 | AddrDelta, OS&: Data); |
| 1075 | DF.setContents(Data); |
| 1076 | DF.clearFixups(); |
| 1077 | return OldSize != Data.size(); |
| 1078 | } |
| 1079 | |
| 1080 | bool MCAssembler::relaxDwarfCallFrameFragment(MCDwarfCallFrameFragment &DF) { |
| 1081 | bool WasRelaxed; |
| 1082 | if (getBackend().relaxDwarfCFA(DF, WasRelaxed)) |
| 1083 | return WasRelaxed; |
| 1084 | |
| 1085 | MCContext &Context = getContext(); |
| 1086 | int64_t Value; |
| 1087 | bool Abs = DF.getAddrDelta().evaluateAsAbsolute(Res&: Value, Asm: *this); |
| 1088 | if (!Abs) { |
| 1089 | reportError(L: DF.getAddrDelta().getLoc(), |
| 1090 | Msg: "invalid CFI advance_loc expression" ); |
| 1091 | DF.setAddrDelta(MCConstantExpr::create(Value: 0, Ctx&: Context)); |
| 1092 | return false; |
| 1093 | } |
| 1094 | |
| 1095 | auto OldSize = DF.getContents().size(); |
| 1096 | SmallVector<char, 8> Data; |
| 1097 | MCDwarfFrameEmitter::encodeAdvanceLoc(Context, AddrDelta: Value, OS&: Data); |
| 1098 | DF.setContents(Data); |
| 1099 | DF.clearFixups(); |
| 1100 | return OldSize != Data.size(); |
| 1101 | } |
| 1102 | |
| 1103 | bool MCAssembler::relaxCVInlineLineTable(MCCVInlineLineTableFragment &F) { |
| 1104 | unsigned OldSize = F.getContents().size(); |
| 1105 | getContext().getCVContext().encodeInlineLineTable(Asm: *this, F); |
| 1106 | return OldSize != F.getContents().size(); |
| 1107 | } |
| 1108 | |
| 1109 | bool MCAssembler::relaxCVDefRange(MCCVDefRangeFragment &F) { |
| 1110 | unsigned OldSize = F.getContents().size(); |
| 1111 | getContext().getCVContext().encodeDefRange(Asm: *this, F); |
| 1112 | return OldSize != F.getContents().size(); |
| 1113 | } |
| 1114 | |
| 1115 | bool MCAssembler::relaxFill(MCFillFragment &F) { |
| 1116 | uint64_t Size = computeFragmentSize(F); |
| 1117 | if (F.getSize() == Size) |
| 1118 | return false; |
| 1119 | F.setSize(Size); |
| 1120 | return true; |
| 1121 | } |
| 1122 | |
| 1123 | bool MCAssembler::relaxPseudoProbeAddr(MCPseudoProbeAddrFragment &PF) { |
| 1124 | uint64_t OldSize = PF.getContents().size(); |
| 1125 | int64_t AddrDelta; |
| 1126 | bool Abs = PF.getAddrDelta().evaluateKnownAbsolute(Res&: AddrDelta, Asm: *this); |
| 1127 | assert(Abs && "We created a pseudo probe with an invalid expression" ); |
| 1128 | (void)Abs; |
| 1129 | SmallVector<char, 8> Data; |
| 1130 | raw_svector_ostream OSE(Data); |
| 1131 | |
| 1132 | // AddrDelta is a signed integer |
| 1133 | encodeSLEB128(Value: AddrDelta, OS&: OSE, PadTo: OldSize); |
| 1134 | PF.setContents(Data); |
| 1135 | PF.clearFixups(); |
| 1136 | return OldSize != Data.size(); |
| 1137 | } |
| 1138 | |
| 1139 | bool MCAssembler::relaxFragment(MCFragment &F) { |
| 1140 | switch(F.getKind()) { |
| 1141 | default: |
| 1142 | return false; |
| 1143 | case MCFragment::FT_Relaxable: |
| 1144 | assert(!getRelaxAll() && |
| 1145 | "Did not expect a MCRelaxableFragment in RelaxAll mode" ); |
| 1146 | return relaxInstruction(F&: cast<MCRelaxableFragment>(Val&: F)); |
| 1147 | case MCFragment::FT_Dwarf: |
| 1148 | return relaxDwarfLineAddr(DF&: cast<MCDwarfLineAddrFragment>(Val&: F)); |
| 1149 | case MCFragment::FT_DwarfFrame: |
| 1150 | return relaxDwarfCallFrameFragment(DF&: cast<MCDwarfCallFrameFragment>(Val&: F)); |
| 1151 | case MCFragment::FT_LEB: |
| 1152 | return relaxLEB(LF&: cast<MCLEBFragment>(Val&: F)); |
| 1153 | case MCFragment::FT_BoundaryAlign: |
| 1154 | return relaxBoundaryAlign(BF&: cast<MCBoundaryAlignFragment>(Val&: F)); |
| 1155 | case MCFragment::FT_CVInlineLines: |
| 1156 | return relaxCVInlineLineTable(F&: cast<MCCVInlineLineTableFragment>(Val&: F)); |
| 1157 | case MCFragment::FT_CVDefRange: |
| 1158 | return relaxCVDefRange(F&: cast<MCCVDefRangeFragment>(Val&: F)); |
| 1159 | case MCFragment::FT_Fill: |
| 1160 | return relaxFill(F&: cast<MCFillFragment>(Val&: F)); |
| 1161 | case MCFragment::FT_PseudoProbe: |
| 1162 | return relaxPseudoProbeAddr(PF&: cast<MCPseudoProbeAddrFragment>(Val&: F)); |
| 1163 | } |
| 1164 | } |
| 1165 | |
| 1166 | void MCAssembler::layoutSection(MCSection &Sec) { |
| 1167 | MCFragment *Prev = nullptr; |
| 1168 | uint64_t Offset = 0; |
| 1169 | for (MCFragment &F : Sec) { |
| 1170 | F.Offset = Offset; |
| 1171 | if (LLVM_UNLIKELY(isBundlingEnabled())) { |
| 1172 | if (F.hasInstructions()) { |
| 1173 | layoutBundle(Prev, F: &F); |
| 1174 | Offset = F.Offset; |
| 1175 | } |
| 1176 | Prev = &F; |
| 1177 | } |
| 1178 | Offset += computeFragmentSize(F); |
| 1179 | } |
| 1180 | } |
| 1181 | |
| 1182 | bool MCAssembler::relaxOnce() { |
| 1183 | ++stats::RelaxationSteps; |
| 1184 | PendingErrors.clear(); |
| 1185 | |
| 1186 | // Size of fragments in one section can depend on the size of fragments in |
| 1187 | // another. If any fragment has changed size, we have to re-layout (and |
| 1188 | // as a result possibly further relax) all sections. |
| 1189 | bool ChangedAny = false; |
| 1190 | for (MCSection &Sec : *this) { |
| 1191 | // Assume each iteration finalizes at least one extra fragment. If the |
| 1192 | // layout does not converge after N+1 iterations, bail out. |
| 1193 | auto MaxIter = Sec.curFragList()->Tail->getLayoutOrder() + 1; |
| 1194 | for (;;) { |
| 1195 | bool Changed = false; |
| 1196 | for (MCFragment &F : Sec) |
| 1197 | if (relaxFragment(F)) |
| 1198 | Changed = true; |
| 1199 | |
| 1200 | ChangedAny |= Changed; |
| 1201 | if (!Changed || --MaxIter == 0) |
| 1202 | break; |
| 1203 | layoutSection(Sec); |
| 1204 | } |
| 1205 | } |
| 1206 | return ChangedAny; |
| 1207 | } |
| 1208 | |
| 1209 | void MCAssembler::reportError(SMLoc L, const Twine &Msg) const { |
| 1210 | getContext().reportError(L, Msg); |
| 1211 | } |
| 1212 | |
| 1213 | void MCAssembler::recordError(SMLoc Loc, const Twine &Msg) const { |
| 1214 | PendingErrors.emplace_back(Args&: Loc, Args: Msg.str()); |
| 1215 | } |
| 1216 | |
| 1217 | void MCAssembler::flushPendingErrors() const { |
| 1218 | for (auto &Err : PendingErrors) |
| 1219 | reportError(L: Err.first, Msg: Err.second); |
| 1220 | PendingErrors.clear(); |
| 1221 | } |
| 1222 | |
| 1223 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
| 1224 | LLVM_DUMP_METHOD void MCAssembler::dump() const{ |
| 1225 | raw_ostream &OS = errs(); |
| 1226 | DenseMap<const MCFragment *, SmallVector<const MCSymbol *, 0>> FragToSyms; |
| 1227 | // Scan symbols and build a map of fragments to their corresponding symbols. |
| 1228 | // For variable symbols, we don't want to call their getFragment, which might |
| 1229 | // modify `Fragment`. |
| 1230 | for (const MCSymbol &Sym : symbols()) |
| 1231 | if (!Sym.isVariable()) |
| 1232 | if (auto *F = Sym.getFragment()) |
| 1233 | FragToSyms.try_emplace(F).first->second.push_back(&Sym); |
| 1234 | |
| 1235 | OS << "Sections:[" ; |
| 1236 | for (const MCSection &Sec : *this) { |
| 1237 | OS << '\n'; |
| 1238 | Sec.dump(&FragToSyms); |
| 1239 | } |
| 1240 | OS << "\n]\n" ; |
| 1241 | } |
| 1242 | #endif |
| 1243 | |