| 1 | //===- CodeGen/AsmPrinter/EHStreamer.cpp - Exception Directive Streamer ---===// |
| 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 | // This file contains support for writing exception info into assembly files. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "EHStreamer.h" |
| 14 | #include "llvm/ADT/SmallVector.h" |
| 15 | #include "llvm/ADT/Twine.h" |
| 16 | #include "llvm/BinaryFormat/Dwarf.h" |
| 17 | #include "llvm/CodeGen/AsmPrinter.h" |
| 18 | #include "llvm/CodeGen/MachineFunction.h" |
| 19 | #include "llvm/CodeGen/MachineInstr.h" |
| 20 | #include "llvm/CodeGen/MachineOperand.h" |
| 21 | #include "llvm/IR/Function.h" |
| 22 | #include "llvm/MC/MCAsmInfo.h" |
| 23 | #include "llvm/MC/MCContext.h" |
| 24 | #include "llvm/MC/MCStreamer.h" |
| 25 | #include "llvm/MC/MCSymbol.h" |
| 26 | #include "llvm/MC/MCTargetOptions.h" |
| 27 | #include "llvm/Support/Casting.h" |
| 28 | #include "llvm/Support/LEB128.h" |
| 29 | #include "llvm/Target/TargetLoweringObjectFile.h" |
| 30 | #include <algorithm> |
| 31 | #include <cassert> |
| 32 | #include <cstdint> |
| 33 | #include <vector> |
| 34 | |
| 35 | using namespace llvm; |
| 36 | |
| 37 | EHStreamer::EHStreamer(AsmPrinter *A) : Asm(A), MMI(Asm->MMI) {} |
| 38 | |
| 39 | EHStreamer::~EHStreamer() = default; |
| 40 | |
| 41 | /// How many leading type ids two landing pads have in common. |
| 42 | unsigned EHStreamer::sharedTypeIDs(const LandingPadInfo *L, |
| 43 | const LandingPadInfo *R) { |
| 44 | const std::vector<int> &LIds = L->TypeIds, &RIds = R->TypeIds; |
| 45 | return std::mismatch(first1: LIds.begin(), last1: LIds.end(), first2: RIds.begin(), last2: RIds.end()) |
| 46 | .first - |
| 47 | LIds.begin(); |
| 48 | } |
| 49 | |
| 50 | /// Compute the actions table and gather the first action index for each landing |
| 51 | /// pad site. |
| 52 | void EHStreamer::computeActionsTable( |
| 53 | const SmallVectorImpl<const LandingPadInfo *> &LandingPads, |
| 54 | SmallVectorImpl<ActionEntry> &Actions, |
| 55 | SmallVectorImpl<unsigned> &FirstActions) { |
| 56 | // The action table follows the call-site table in the LSDA. The individual |
| 57 | // records are of two types: |
| 58 | // |
| 59 | // * Catch clause |
| 60 | // * Exception specification |
| 61 | // |
| 62 | // The two record kinds have the same format, with only small differences. |
| 63 | // They are distinguished by the "switch value" field: Catch clauses |
| 64 | // (TypeInfos) have strictly positive switch values, and exception |
| 65 | // specifications (FilterIds) have strictly negative switch values. Value 0 |
| 66 | // indicates a catch-all clause. |
| 67 | // |
| 68 | // Negative type IDs index into FilterIds. Positive type IDs index into |
| 69 | // TypeInfos. The value written for a positive type ID is just the type ID |
| 70 | // itself. For a negative type ID, however, the value written is the |
| 71 | // (negative) byte offset of the corresponding FilterIds entry. The byte |
| 72 | // offset is usually equal to the type ID (because the FilterIds entries are |
| 73 | // written using a variable width encoding, which outputs one byte per entry |
| 74 | // as long as the value written is not too large) but can differ. This kind |
| 75 | // of complication does not occur for positive type IDs because type infos are |
| 76 | // output using a fixed width encoding. FilterOffsets[i] holds the byte |
| 77 | // offset corresponding to FilterIds[i]. |
| 78 | |
| 79 | const std::vector<unsigned> &FilterIds = Asm->MF->getFilterIds(); |
| 80 | SmallVector<int, 16> FilterOffsets; |
| 81 | FilterOffsets.reserve(N: FilterIds.size()); |
| 82 | int Offset = -1; |
| 83 | |
| 84 | for (unsigned FilterId : FilterIds) { |
| 85 | FilterOffsets.push_back(Elt: Offset); |
| 86 | Offset -= getULEB128Size(Value: FilterId); |
| 87 | } |
| 88 | |
| 89 | FirstActions.reserve(N: LandingPads.size()); |
| 90 | |
| 91 | int FirstAction = 0; |
| 92 | unsigned SizeActions = 0; // Total size of all action entries for a function |
| 93 | const LandingPadInfo *PrevLPI = nullptr; |
| 94 | |
| 95 | for (const LandingPadInfo *LPI : LandingPads) { |
| 96 | const std::vector<int> &TypeIds = LPI->TypeIds; |
| 97 | unsigned NumShared = PrevLPI ? sharedTypeIDs(L: LPI, R: PrevLPI) : 0; |
| 98 | unsigned SizeSiteActions = 0; // Total size of all entries for a landingpad |
| 99 | |
| 100 | if (NumShared < TypeIds.size()) { |
| 101 | // Size of one action entry (typeid + next action) |
| 102 | unsigned SizeActionEntry = 0; |
| 103 | unsigned PrevAction = (unsigned)-1; |
| 104 | |
| 105 | if (NumShared) { |
| 106 | unsigned SizePrevIds = PrevLPI->TypeIds.size(); |
| 107 | assert(Actions.size()); |
| 108 | PrevAction = Actions.size() - 1; |
| 109 | SizeActionEntry = getSLEB128Size(Value: Actions[PrevAction].NextAction) + |
| 110 | getSLEB128Size(Value: Actions[PrevAction].ValueForTypeID); |
| 111 | |
| 112 | for (unsigned j = NumShared; j != SizePrevIds; ++j) { |
| 113 | assert(PrevAction != (unsigned)-1 && "PrevAction is invalid!" ); |
| 114 | SizeActionEntry -= getSLEB128Size(Value: Actions[PrevAction].ValueForTypeID); |
| 115 | SizeActionEntry += -Actions[PrevAction].NextAction; |
| 116 | PrevAction = Actions[PrevAction].Previous; |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | // Compute the actions. |
| 121 | for (unsigned J = NumShared, M = TypeIds.size(); J != M; ++J) { |
| 122 | int TypeID = TypeIds[J]; |
| 123 | assert(-1 - TypeID < (int)FilterOffsets.size() && "Unknown filter id!" ); |
| 124 | int ValueForTypeID = |
| 125 | isFilterEHSelector(Selector: TypeID) ? FilterOffsets[-1 - TypeID] : TypeID; |
| 126 | unsigned SizeTypeID = getSLEB128Size(Value: ValueForTypeID); |
| 127 | |
| 128 | int NextAction = SizeActionEntry ? -(SizeActionEntry + SizeTypeID) : 0; |
| 129 | SizeActionEntry = SizeTypeID + getSLEB128Size(Value: NextAction); |
| 130 | SizeSiteActions += SizeActionEntry; |
| 131 | |
| 132 | ActionEntry Action = { .ValueForTypeID: ValueForTypeID, .NextAction: NextAction, .Previous: PrevAction }; |
| 133 | Actions.push_back(Elt: Action); |
| 134 | PrevAction = Actions.size() - 1; |
| 135 | } |
| 136 | |
| 137 | // Record the first action of the landing pad site. |
| 138 | FirstAction = SizeActions + SizeSiteActions - SizeActionEntry + 1; |
| 139 | } // else identical - re-use previous FirstAction |
| 140 | |
| 141 | // Information used when creating the call-site table. The action record |
| 142 | // field of the call site record is the offset of the first associated |
| 143 | // action record, relative to the start of the actions table. This value is |
| 144 | // biased by 1 (1 indicating the start of the actions table), and 0 |
| 145 | // indicates that there are no actions. |
| 146 | FirstActions.push_back(Elt: FirstAction); |
| 147 | |
| 148 | // Compute this sites contribution to size. |
| 149 | SizeActions += SizeSiteActions; |
| 150 | |
| 151 | PrevLPI = LPI; |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | /// Return `true' if this is a call to a function marked `nounwind'. Return |
| 156 | /// `false' otherwise. |
| 157 | bool EHStreamer::callToNoUnwindFunction(const MachineInstr *MI) { |
| 158 | assert(MI->isCall() && "This should be a call instruction!" ); |
| 159 | |
| 160 | bool MarkedNoUnwind = false; |
| 161 | bool SawFunc = false; |
| 162 | |
| 163 | for (const MachineOperand &MO : MI->operands()) { |
| 164 | if (!MO.isGlobal()) continue; |
| 165 | |
| 166 | const Function *F = dyn_cast<Function>(Val: MO.getGlobal()); |
| 167 | if (!F) continue; |
| 168 | |
| 169 | if (SawFunc) { |
| 170 | // Be conservative. If we have more than one function operand for this |
| 171 | // call, then we can't make the assumption that it's the callee and |
| 172 | // not a parameter to the call. |
| 173 | // |
| 174 | // FIXME: Determine if there's a way to say that `F' is the callee or |
| 175 | // parameter. |
| 176 | MarkedNoUnwind = false; |
| 177 | break; |
| 178 | } |
| 179 | |
| 180 | MarkedNoUnwind = F->doesNotThrow(); |
| 181 | SawFunc = true; |
| 182 | } |
| 183 | |
| 184 | return MarkedNoUnwind; |
| 185 | } |
| 186 | |
| 187 | void EHStreamer::computePadMap( |
| 188 | const SmallVectorImpl<const LandingPadInfo *> &LandingPads, |
| 189 | RangeMapType &PadMap) { |
| 190 | // Invokes and nounwind calls have entries in PadMap (due to being bracketed |
| 191 | // by try-range labels when lowered). Ordinary calls do not, so appropriate |
| 192 | // try-ranges for them need be deduced so we can put them in the LSDA. |
| 193 | for (unsigned i = 0, N = LandingPads.size(); i != N; ++i) { |
| 194 | const LandingPadInfo *LandingPad = LandingPads[i]; |
| 195 | for (unsigned j = 0, E = LandingPad->BeginLabels.size(); j != E; ++j) { |
| 196 | MCSymbol *BeginLabel = LandingPad->BeginLabels[j]; |
| 197 | MCSymbol *EndLabel = LandingPad->BeginLabels[j]; |
| 198 | // If we have deleted the code for a given invoke after registering it in |
| 199 | // the LandingPad label list, the associated symbols will not have been |
| 200 | // emitted. In that case, ignore this callsite entry. |
| 201 | if (!BeginLabel->isDefined() || !EndLabel->isDefined()) |
| 202 | continue; |
| 203 | assert(!PadMap.count(BeginLabel) && "Duplicate landing pad labels!" ); |
| 204 | PadRange P = { .PadIndex: i, .RangeIndex: j }; |
| 205 | PadMap[BeginLabel] = P; |
| 206 | } |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | /// Compute the call-site table. The entry for an invoke has a try-range |
| 211 | /// containing the call, a non-zero landing pad, and an appropriate action. The |
| 212 | /// entry for an ordinary call has a try-range containing the call and zero for |
| 213 | /// the landing pad and the action. Calls marked 'nounwind' have no entry and |
| 214 | /// must not be contained in the try-range of any entry - they form gaps in the |
| 215 | /// table. Entries must be ordered by try-range address. |
| 216 | /// |
| 217 | /// Call-sites are split into one or more call-site ranges associated with |
| 218 | /// different sections of the function. |
| 219 | /// |
| 220 | /// - Without -basic-block-sections, all call-sites are grouped into one |
| 221 | /// call-site-range corresponding to the function section. |
| 222 | /// |
| 223 | /// - With -basic-block-sections, one call-site range is created for each |
| 224 | /// section, with its FragmentBeginLabel and FragmentEndLabel respectively |
| 225 | // set to the beginning and ending of the corresponding section and its |
| 226 | // ExceptionLabel set to the exception symbol dedicated for this section. |
| 227 | // Later, one LSDA header will be emitted for each call-site range with its |
| 228 | // call-sites following. The action table and type info table will be |
| 229 | // shared across all ranges. |
| 230 | void EHStreamer::computeCallSiteTable( |
| 231 | SmallVectorImpl<CallSiteEntry> &CallSites, |
| 232 | SmallVectorImpl<CallSiteRange> &CallSiteRanges, |
| 233 | const SmallVectorImpl<const LandingPadInfo *> &LandingPads, |
| 234 | const SmallVectorImpl<unsigned> &FirstActions) { |
| 235 | RangeMapType PadMap; |
| 236 | computePadMap(LandingPads, PadMap); |
| 237 | |
| 238 | // The end label of the previous invoke or nounwind try-range. |
| 239 | MCSymbol *LastLabel = Asm->getFunctionBegin(); |
| 240 | |
| 241 | // Whether there is a potentially throwing instruction (currently this means |
| 242 | // an ordinary call) between the end of the previous try-range and now. |
| 243 | bool SawPotentiallyThrowing = false; |
| 244 | |
| 245 | // Whether the last CallSite entry was for an invoke. |
| 246 | bool PreviousIsInvoke = false; |
| 247 | |
| 248 | bool IsSJLJ = Asm->MAI->getExceptionHandlingType() == ExceptionHandling::SjLj; |
| 249 | |
| 250 | // Visit all instructions in order of address. |
| 251 | for (const auto &MBB : *Asm->MF) { |
| 252 | if (&MBB == &Asm->MF->front() || MBB.isBeginSection()) { |
| 253 | // We start a call-site range upon function entry and at the beginning of |
| 254 | // every basic block section. |
| 255 | auto &Range = Asm->MBBSectionRanges[MBB.getSectionID()]; |
| 256 | CallSiteRanges.push_back(Elt: {.FragmentBeginLabel: Range.BeginLabel, .FragmentEndLabel: Range.EndLabel, |
| 257 | .ExceptionLabel: Asm->getMBBExceptionSym(MBB), |
| 258 | .CallSiteBeginIdx: CallSites.size()}); |
| 259 | PreviousIsInvoke = false; |
| 260 | SawPotentiallyThrowing = false; |
| 261 | LastLabel = nullptr; |
| 262 | } |
| 263 | |
| 264 | if (MBB.isEHPad()) |
| 265 | CallSiteRanges.back().IsLPRange = true; |
| 266 | |
| 267 | for (const auto &MI : MBB) { |
| 268 | if (!MI.isEHLabel()) { |
| 269 | if (MI.isCall()) |
| 270 | SawPotentiallyThrowing |= !callToNoUnwindFunction(MI: &MI); |
| 271 | continue; |
| 272 | } |
| 273 | |
| 274 | // End of the previous try-range? |
| 275 | MCSymbol *BeginLabel = MI.getOperand(i: 0).getMCSymbol(); |
| 276 | if (BeginLabel == LastLabel) |
| 277 | SawPotentiallyThrowing = false; |
| 278 | |
| 279 | // Beginning of a new try-range? |
| 280 | RangeMapType::const_iterator L = PadMap.find(Val: BeginLabel); |
| 281 | if (L == PadMap.end()) |
| 282 | // Nope, it was just some random label. |
| 283 | continue; |
| 284 | |
| 285 | const PadRange &P = L->second; |
| 286 | const LandingPadInfo *LandingPad = LandingPads[P.PadIndex]; |
| 287 | assert(BeginLabel == LandingPad->BeginLabels[P.RangeIndex] && |
| 288 | "Inconsistent landing pad map!" ); |
| 289 | |
| 290 | // For Dwarf and AIX exception handling (SjLj handling doesn't use this). |
| 291 | // If some instruction between the previous try-range and this one may |
| 292 | // throw, create a call-site entry with no landing pad for the region |
| 293 | // between the try-ranges. |
| 294 | if (SawPotentiallyThrowing && |
| 295 | (Asm->MAI->usesCFIForEH() || |
| 296 | Asm->MAI->getExceptionHandlingType() == ExceptionHandling::AIX)) { |
| 297 | CallSites.push_back(Elt: {.BeginLabel: LastLabel, .EndLabel: BeginLabel, .LPad: nullptr, .Action: 0}); |
| 298 | PreviousIsInvoke = false; |
| 299 | } |
| 300 | |
| 301 | LastLabel = LandingPad->EndLabels[P.RangeIndex]; |
| 302 | assert(BeginLabel && LastLabel && "Invalid landing pad!" ); |
| 303 | |
| 304 | if (!LandingPad->LandingPadLabel) { |
| 305 | // Create a gap. |
| 306 | PreviousIsInvoke = false; |
| 307 | } else { |
| 308 | // This try-range is for an invoke. |
| 309 | CallSiteEntry Site = { |
| 310 | .BeginLabel: BeginLabel, |
| 311 | .EndLabel: LastLabel, |
| 312 | .LPad: LandingPad, |
| 313 | .Action: FirstActions[P.PadIndex] |
| 314 | }; |
| 315 | |
| 316 | // Try to merge with the previous call-site. SJLJ doesn't do this |
| 317 | if (PreviousIsInvoke && !IsSJLJ) { |
| 318 | CallSiteEntry &Prev = CallSites.back(); |
| 319 | if (Site.LPad == Prev.LPad && Site.Action == Prev.Action) { |
| 320 | // Extend the range of the previous entry. |
| 321 | Prev.EndLabel = Site.EndLabel; |
| 322 | continue; |
| 323 | } |
| 324 | } |
| 325 | |
| 326 | // Otherwise, create a new call-site. |
| 327 | if (!IsSJLJ) |
| 328 | CallSites.push_back(Elt: Site); |
| 329 | else { |
| 330 | // SjLj EH must maintain the call sites in the order assigned |
| 331 | // to them by the SjLjPrepare pass. |
| 332 | unsigned SiteNo = Asm->MF->getCallSiteBeginLabel(BeginLabel); |
| 333 | if (CallSites.size() < SiteNo) |
| 334 | CallSites.resize(N: SiteNo); |
| 335 | CallSites[SiteNo - 1] = Site; |
| 336 | } |
| 337 | PreviousIsInvoke = true; |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | // We end the call-site range upon function exit and at the end of every |
| 342 | // basic block section. |
| 343 | if (&MBB == &Asm->MF->back() || MBB.isEndSection()) { |
| 344 | // If some instruction between the previous try-range and the end of the |
| 345 | // function may throw, create a call-site entry with no landing pad for |
| 346 | // the region following the try-range. |
| 347 | if (SawPotentiallyThrowing && !IsSJLJ) { |
| 348 | CallSiteEntry Site = {.BeginLabel: LastLabel, .EndLabel: CallSiteRanges.back().FragmentEndLabel, |
| 349 | .LPad: nullptr, .Action: 0}; |
| 350 | CallSites.push_back(Elt: Site); |
| 351 | SawPotentiallyThrowing = false; |
| 352 | } |
| 353 | CallSiteRanges.back().CallSiteEndIdx = CallSites.size(); |
| 354 | } |
| 355 | } |
| 356 | } |
| 357 | |
| 358 | /// Emit landing pads and actions. |
| 359 | /// |
| 360 | /// The general organization of the table is complex, but the basic concepts are |
| 361 | /// easy. First there is a header which describes the location and organization |
| 362 | /// of the three components that follow. |
| 363 | /// |
| 364 | /// 1. The landing pad site information describes the range of code covered by |
| 365 | /// the try. In our case it's an accumulation of the ranges covered by the |
| 366 | /// invokes in the try. There is also a reference to the landing pad that |
| 367 | /// handles the exception once processed. Finally an index into the actions |
| 368 | /// table. |
| 369 | /// 2. The action table, in our case, is composed of pairs of type IDs and next |
| 370 | /// action offset. Starting with the action index from the landing pad |
| 371 | /// site, each type ID is checked for a match to the current exception. If |
| 372 | /// it matches then the exception and type id are passed on to the landing |
| 373 | /// pad. Otherwise the next action is looked up. This chain is terminated |
| 374 | /// with a next action of zero. If no type id is found then the frame is |
| 375 | /// unwound and handling continues. |
| 376 | /// 3. Type ID table contains references to all the C++ typeinfo for all |
| 377 | /// catches in the function. This tables is reverse indexed base 1. |
| 378 | /// |
| 379 | /// Returns the starting symbol of an exception table. |
| 380 | MCSymbol *EHStreamer::emitExceptionTable() { |
| 381 | const MachineFunction *MF = Asm->MF; |
| 382 | const std::vector<const GlobalValue *> &TypeInfos = MF->getTypeInfos(); |
| 383 | const std::vector<unsigned> &FilterIds = MF->getFilterIds(); |
| 384 | const std::vector<LandingPadInfo> &PadInfos = MF->getLandingPads(); |
| 385 | |
| 386 | // Sort the landing pads in order of their type ids. This is used to fold |
| 387 | // duplicate actions. |
| 388 | SmallVector<const LandingPadInfo *, 64> LandingPads; |
| 389 | LandingPads.reserve(N: PadInfos.size()); |
| 390 | |
| 391 | for (const LandingPadInfo &LPI : PadInfos) { |
| 392 | // If a landing-pad has an associated label, but the label wasn't ever |
| 393 | // emitted, then skip it. (This can occur if the landingpad's MBB was |
| 394 | // deleted). |
| 395 | if (LPI.LandingPadLabel && !LPI.LandingPadLabel->isDefined()) |
| 396 | continue; |
| 397 | LandingPads.push_back(Elt: &LPI); |
| 398 | } |
| 399 | |
| 400 | // Order landing pads lexicographically by type id. |
| 401 | llvm::sort(C&: LandingPads, Comp: [](const LandingPadInfo *L, const LandingPadInfo *R) { |
| 402 | return L->TypeIds < R->TypeIds; |
| 403 | }); |
| 404 | |
| 405 | // Compute the actions table and gather the first action index for each |
| 406 | // landing pad site. |
| 407 | SmallVector<ActionEntry, 32> Actions; |
| 408 | SmallVector<unsigned, 64> FirstActions; |
| 409 | computeActionsTable(LandingPads, Actions, FirstActions); |
| 410 | |
| 411 | // Compute the call-site table and call-site ranges. Normally, there is only |
| 412 | // one call-site-range which covers the whole function. With |
| 413 | // -basic-block-sections, there is one call-site-range per basic block |
| 414 | // section. |
| 415 | SmallVector<CallSiteEntry, 64> CallSites; |
| 416 | SmallVector<CallSiteRange, 4> CallSiteRanges; |
| 417 | computeCallSiteTable(CallSites, CallSiteRanges, LandingPads, FirstActions); |
| 418 | |
| 419 | bool IsSJLJ = Asm->MAI->getExceptionHandlingType() == ExceptionHandling::SjLj; |
| 420 | bool IsWasm = Asm->MAI->getExceptionHandlingType() == ExceptionHandling::Wasm; |
| 421 | bool HasLEB128Directives = Asm->MAI->hasLEB128Directives(); |
| 422 | unsigned CallSiteEncoding = |
| 423 | IsSJLJ ? static_cast<unsigned>(dwarf::DW_EH_PE_udata4) : |
| 424 | Asm->getObjFileLowering().getCallSiteEncoding(); |
| 425 | bool HaveTTData = !TypeInfos.empty() || !FilterIds.empty(); |
| 426 | |
| 427 | // Type infos. |
| 428 | MCSection *LSDASection = Asm->getObjFileLowering().getSectionForLSDA( |
| 429 | MF->getFunction(), *Asm->CurrentFnSym, Asm->TM); |
| 430 | unsigned TTypeEncoding; |
| 431 | |
| 432 | if (!HaveTTData) { |
| 433 | // If there is no TypeInfo, then we just explicitly say that we're omitting |
| 434 | // that bit. |
| 435 | TTypeEncoding = dwarf::DW_EH_PE_omit; |
| 436 | } else { |
| 437 | // Okay, we have actual filters or typeinfos to emit. As such, we need to |
| 438 | // pick a type encoding for them. We're about to emit a list of pointers to |
| 439 | // typeinfo objects at the end of the LSDA. However, unless we're in static |
| 440 | // mode, this reference will require a relocation by the dynamic linker. |
| 441 | // |
| 442 | // Because of this, we have a couple of options: |
| 443 | // |
| 444 | // 1) If we are in -static mode, we can always use an absolute reference |
| 445 | // from the LSDA, because the static linker will resolve it. |
| 446 | // |
| 447 | // 2) Otherwise, if the LSDA section is writable, we can output the direct |
| 448 | // reference to the typeinfo and allow the dynamic linker to relocate |
| 449 | // it. Since it is in a writable section, the dynamic linker won't |
| 450 | // have a problem. |
| 451 | // |
| 452 | // 3) Finally, if we're in PIC mode and the LDSA section isn't writable, |
| 453 | // we need to use some form of indirection. For example, on Darwin, |
| 454 | // we can output a statically-relocatable reference to a dyld stub. The |
| 455 | // offset to the stub is constant, but the contents are in a section |
| 456 | // that is updated by the dynamic linker. This is easy enough, but we |
| 457 | // need to tell the personality function of the unwinder to indirect |
| 458 | // through the dyld stub. |
| 459 | // |
| 460 | // FIXME: When (3) is actually implemented, we'll have to emit the stubs |
| 461 | // somewhere. This predicate should be moved to a shared location that is |
| 462 | // in target-independent code. |
| 463 | // |
| 464 | TTypeEncoding = Asm->getObjFileLowering().getTTypeEncoding(); |
| 465 | } |
| 466 | |
| 467 | // Begin the exception table. |
| 468 | // Sometimes we want not to emit the data into separate section (e.g. ARM |
| 469 | // EHABI). In this case LSDASection will be NULL. |
| 470 | if (LSDASection) |
| 471 | Asm->OutStreamer->switchSection(Section: LSDASection); |
| 472 | Asm->emitAlignment(Alignment: Align(4)); |
| 473 | |
| 474 | // Emit the LSDA. |
| 475 | MCSymbol *GCCETSym = |
| 476 | Asm->OutContext.getOrCreateSymbol(Name: Twine("GCC_except_table" )+ |
| 477 | Twine(Asm->getFunctionNumber())); |
| 478 | Asm->OutStreamer->emitLabel(Symbol: GCCETSym); |
| 479 | MCSymbol *CstEndLabel = Asm->createTempSymbol( |
| 480 | Name: CallSiteRanges.size() > 1 ? "action_table_base" : "cst_end" ); |
| 481 | |
| 482 | MCSymbol *TTBaseLabel = nullptr; |
| 483 | if (HaveTTData) |
| 484 | TTBaseLabel = Asm->createTempSymbol(Name: "ttbase" ); |
| 485 | |
| 486 | const bool VerboseAsm = Asm->OutStreamer->isVerboseAsm(); |
| 487 | |
| 488 | // Helper for emitting references (offsets) for type table and the end of the |
| 489 | // call-site table (which marks the beginning of the action table). |
| 490 | // * For Itanium, these references will be emitted for every callsite range. |
| 491 | // * For SJLJ and Wasm, they will be emitted only once in the LSDA header. |
| 492 | auto EmitTypeTableRefAndCallSiteTableEndRef = [&]() { |
| 493 | Asm->emitEncodingByte(Val: TTypeEncoding, Desc: "@TType" ); |
| 494 | if (HaveTTData) { |
| 495 | // N.B.: There is a dependency loop between the size of the TTBase uleb128 |
| 496 | // here and the amount of padding before the aligned type table. The |
| 497 | // assembler must sometimes pad this uleb128 or insert extra padding |
| 498 | // before the type table. See PR35809 or GNU as bug 4029. |
| 499 | MCSymbol *TTBaseRefLabel = Asm->createTempSymbol(Name: "ttbaseref" ); |
| 500 | Asm->emitLabelDifferenceAsULEB128(Hi: TTBaseLabel, Lo: TTBaseRefLabel); |
| 501 | Asm->OutStreamer->emitLabel(Symbol: TTBaseRefLabel); |
| 502 | } |
| 503 | |
| 504 | // The Action table follows the call-site table. So we emit the |
| 505 | // label difference from here (start of the call-site table for SJLJ and |
| 506 | // Wasm, and start of a call-site range for Itanium) to the end of the |
| 507 | // whole call-site table (end of the last call-site range for Itanium). |
| 508 | MCSymbol *CstBeginLabel = Asm->createTempSymbol(Name: "cst_begin" ); |
| 509 | Asm->emitEncodingByte(Val: CallSiteEncoding, Desc: "Call site" ); |
| 510 | Asm->emitLabelDifferenceAsULEB128(Hi: CstEndLabel, Lo: CstBeginLabel); |
| 511 | Asm->OutStreamer->emitLabel(Symbol: CstBeginLabel); |
| 512 | }; |
| 513 | |
| 514 | // An alternative path to EmitTypeTableRefAndCallSiteTableEndRef. |
| 515 | // For some platforms, the system assembler does not accept the form of |
| 516 | // `.uleb128 label2 - label1`. In those situations, we would need to calculate |
| 517 | // the size between label1 and label2 manually. |
| 518 | // In this case, we would need to calculate the LSDA size and the call |
| 519 | // site table size. |
| 520 | auto EmitTypeTableOffsetAndCallSiteTableOffset = [&]() { |
| 521 | assert(CallSiteEncoding == dwarf::DW_EH_PE_udata4 && !HasLEB128Directives && |
| 522 | "Targets supporting .uleb128 do not need to take this path." ); |
| 523 | if (CallSiteRanges.size() > 1) |
| 524 | report_fatal_error( |
| 525 | reason: "-fbasic-block-sections is not yet supported on " |
| 526 | "platforms that do not have general LEB128 directive support." ); |
| 527 | |
| 528 | uint64_t CallSiteTableSize = 0; |
| 529 | const CallSiteRange &CSRange = CallSiteRanges.back(); |
| 530 | for (size_t CallSiteIdx = CSRange.CallSiteBeginIdx; |
| 531 | CallSiteIdx < CSRange.CallSiteEndIdx; ++CallSiteIdx) { |
| 532 | const CallSiteEntry &S = CallSites[CallSiteIdx]; |
| 533 | // Each call site entry consists of 3 udata4 fields (12 bytes) and |
| 534 | // 1 ULEB128 field. |
| 535 | CallSiteTableSize += 12 + getULEB128Size(Value: S.Action); |
| 536 | assert(isUInt<32>(CallSiteTableSize) && "CallSiteTableSize overflows." ); |
| 537 | } |
| 538 | |
| 539 | Asm->emitEncodingByte(Val: TTypeEncoding, Desc: "@TType" ); |
| 540 | if (HaveTTData) { |
| 541 | const unsigned ByteSizeOfCallSiteOffset = |
| 542 | getULEB128Size(Value: CallSiteTableSize); |
| 543 | uint64_t ActionTableSize = 0; |
| 544 | for (const ActionEntry &Action : Actions) { |
| 545 | // Each action entry consists of two SLEB128 fields. |
| 546 | ActionTableSize += getSLEB128Size(Value: Action.ValueForTypeID) + |
| 547 | getSLEB128Size(Value: Action.NextAction); |
| 548 | assert(isUInt<32>(ActionTableSize) && "ActionTableSize overflows." ); |
| 549 | } |
| 550 | |
| 551 | const unsigned TypeInfoSize = |
| 552 | Asm->GetSizeOfEncodedValue(Encoding: TTypeEncoding) * MF->getTypeInfos().size(); |
| 553 | |
| 554 | const uint64_t LSDASizeBeforeAlign = |
| 555 | 1 // Call site encoding byte. |
| 556 | + ByteSizeOfCallSiteOffset // ULEB128 encoding of CallSiteTableSize. |
| 557 | + CallSiteTableSize // Call site table content. |
| 558 | + ActionTableSize; // Action table content. |
| 559 | |
| 560 | const uint64_t LSDASizeWithoutAlign = LSDASizeBeforeAlign + TypeInfoSize; |
| 561 | const unsigned ByteSizeOfLSDAWithoutAlign = |
| 562 | getULEB128Size(Value: LSDASizeWithoutAlign); |
| 563 | const uint64_t DisplacementBeforeAlign = |
| 564 | 2 // LPStartEncoding and TypeTableEncoding. |
| 565 | + ByteSizeOfLSDAWithoutAlign + LSDASizeBeforeAlign; |
| 566 | |
| 567 | // The type info area starts with 4 byte alignment. |
| 568 | const unsigned NeedAlignVal = (4 - DisplacementBeforeAlign % 4) % 4; |
| 569 | uint64_t LSDASizeWithAlign = LSDASizeWithoutAlign + NeedAlignVal; |
| 570 | const unsigned ByteSizeOfLSDAWithAlign = |
| 571 | getULEB128Size(Value: LSDASizeWithAlign); |
| 572 | |
| 573 | // The LSDASizeWithAlign could use 1 byte less padding for alignment |
| 574 | // when the data we use to represent the LSDA Size "needs" to be 1 byte |
| 575 | // larger than the one previously calculated without alignment. |
| 576 | if (ByteSizeOfLSDAWithAlign > ByteSizeOfLSDAWithoutAlign) |
| 577 | LSDASizeWithAlign -= 1; |
| 578 | |
| 579 | Asm->OutStreamer->emitULEB128IntValue(Value: LSDASizeWithAlign, |
| 580 | PadTo: ByteSizeOfLSDAWithAlign); |
| 581 | } |
| 582 | |
| 583 | Asm->emitEncodingByte(Val: CallSiteEncoding, Desc: "Call site" ); |
| 584 | Asm->OutStreamer->emitULEB128IntValue(Value: CallSiteTableSize); |
| 585 | }; |
| 586 | |
| 587 | // SjLj / Wasm Exception handling |
| 588 | if (IsSJLJ || IsWasm) { |
| 589 | Asm->OutStreamer->emitLabel(Symbol: Asm->getMBBExceptionSym(MBB: Asm->MF->front())); |
| 590 | |
| 591 | // emit the LSDA header. |
| 592 | Asm->emitEncodingByte(Val: dwarf::DW_EH_PE_omit, Desc: "@LPStart" ); |
| 593 | EmitTypeTableRefAndCallSiteTableEndRef(); |
| 594 | |
| 595 | unsigned idx = 0; |
| 596 | for (SmallVectorImpl<CallSiteEntry>::const_iterator |
| 597 | I = CallSites.begin(), E = CallSites.end(); I != E; ++I, ++idx) { |
| 598 | const CallSiteEntry &S = *I; |
| 599 | |
| 600 | // Index of the call site entry. |
| 601 | if (VerboseAsm) { |
| 602 | Asm->OutStreamer->AddComment(T: ">> Call Site " + Twine(idx) + " <<" ); |
| 603 | Asm->OutStreamer->AddComment(T: " On exception at call site " +Twine(idx)); |
| 604 | } |
| 605 | Asm->emitULEB128(Value: idx); |
| 606 | |
| 607 | // Offset of the first associated action record, relative to the start of |
| 608 | // the action table. This value is biased by 1 (1 indicates the start of |
| 609 | // the action table), and 0 indicates that there are no actions. |
| 610 | if (VerboseAsm) { |
| 611 | if (S.Action == 0) |
| 612 | Asm->OutStreamer->AddComment(T: " Action: cleanup" ); |
| 613 | else |
| 614 | Asm->OutStreamer->AddComment(T: " Action: " + |
| 615 | Twine((S.Action - 1) / 2 + 1)); |
| 616 | } |
| 617 | Asm->emitULEB128(Value: S.Action); |
| 618 | } |
| 619 | Asm->OutStreamer->emitLabel(Symbol: CstEndLabel); |
| 620 | } else { |
| 621 | // Itanium LSDA exception handling |
| 622 | |
| 623 | // The call-site table is a list of all call sites that may throw an |
| 624 | // exception (including C++ 'throw' statements) in the procedure |
| 625 | // fragment. It immediately follows the LSDA header. Each entry indicates, |
| 626 | // for a given call, the first corresponding action record and corresponding |
| 627 | // landing pad. |
| 628 | // |
| 629 | // The table begins with the number of bytes, stored as an LEB128 |
| 630 | // compressed, unsigned integer. The records immediately follow the record |
| 631 | // count. They are sorted in increasing call-site address. Each record |
| 632 | // indicates: |
| 633 | // |
| 634 | // * The position of the call-site. |
| 635 | // * The position of the landing pad. |
| 636 | // * The first action record for that call site. |
| 637 | // |
| 638 | // A missing entry in the call-site table indicates that a call is not |
| 639 | // supposed to throw. |
| 640 | |
| 641 | assert(CallSiteRanges.size() != 0 && "No call-site ranges!" ); |
| 642 | |
| 643 | // There should be only one call-site range which includes all the landing |
| 644 | // pads. Find that call-site range here. |
| 645 | const CallSiteRange *LandingPadRange = nullptr; |
| 646 | for (const CallSiteRange &CSRange : CallSiteRanges) { |
| 647 | if (CSRange.IsLPRange) { |
| 648 | assert(LandingPadRange == nullptr && |
| 649 | "All landing pads must be in a single callsite range." ); |
| 650 | LandingPadRange = &CSRange; |
| 651 | } |
| 652 | } |
| 653 | |
| 654 | // The call-site table is split into its call-site ranges, each being |
| 655 | // emitted as: |
| 656 | // [ LPStartEncoding | LPStart ] |
| 657 | // [ TypeTableEncoding | TypeTableOffset ] |
| 658 | // [ CallSiteEncoding | CallSiteTableEndOffset ] |
| 659 | // cst_begin -> { call-site entries contained in this range } |
| 660 | // |
| 661 | // and is followed by the next call-site range. |
| 662 | // |
| 663 | // For each call-site range, CallSiteTableEndOffset is computed as the |
| 664 | // difference between cst_begin of that range and the last call-site-table's |
| 665 | // end label. This offset is used to find the action table. |
| 666 | |
| 667 | unsigned Entry = 0; |
| 668 | for (const CallSiteRange &CSRange : CallSiteRanges) { |
| 669 | if (CSRange.CallSiteBeginIdx != 0) { |
| 670 | // Align the call-site range for all ranges except the first. The |
| 671 | // first range is already aligned due to the exception table alignment. |
| 672 | Asm->emitAlignment(Alignment: Align(4)); |
| 673 | } |
| 674 | Asm->OutStreamer->emitLabel(Symbol: CSRange.ExceptionLabel); |
| 675 | |
| 676 | // Emit the LSDA header. |
| 677 | // LPStart is omitted if either we have a single call-site range (in which |
| 678 | // case the function entry is treated as @LPStart) or if this function has |
| 679 | // no landing pads (in which case @LPStart is undefined). |
| 680 | if (CallSiteRanges.size() == 1 || LandingPadRange == nullptr) { |
| 681 | Asm->emitEncodingByte(Val: dwarf::DW_EH_PE_omit, Desc: "@LPStart" ); |
| 682 | } else if (!Asm->isPositionIndependent()) { |
| 683 | // For more than one call-site ranges, LPStart must be explicitly |
| 684 | // specified. |
| 685 | // For non-PIC we can simply use the absolute value. |
| 686 | Asm->emitEncodingByte(Val: dwarf::DW_EH_PE_absptr, Desc: "@LPStart" ); |
| 687 | Asm->OutStreamer->emitSymbolValue(Sym: LandingPadRange->FragmentBeginLabel, |
| 688 | Size: Asm->MAI->getCodePointerSize()); |
| 689 | } else { |
| 690 | // For PIC mode, we Emit a PC-relative address for LPStart. |
| 691 | Asm->emitEncodingByte(Val: dwarf::DW_EH_PE_pcrel, Desc: "@LPStart" ); |
| 692 | MCContext &Context = Asm->OutStreamer->getContext(); |
| 693 | MCSymbol *Dot = Context.createTempSymbol(); |
| 694 | Asm->OutStreamer->emitLabel(Symbol: Dot); |
| 695 | Asm->OutStreamer->emitValue( |
| 696 | Value: MCBinaryExpr::createSub( |
| 697 | LHS: MCSymbolRefExpr::create(Symbol: LandingPadRange->FragmentBeginLabel, |
| 698 | Ctx&: Context), |
| 699 | RHS: MCSymbolRefExpr::create(Symbol: Dot, Ctx&: Context), Ctx&: Context), |
| 700 | Size: Asm->MAI->getCodePointerSize()); |
| 701 | } |
| 702 | |
| 703 | if (HasLEB128Directives) |
| 704 | EmitTypeTableRefAndCallSiteTableEndRef(); |
| 705 | else |
| 706 | EmitTypeTableOffsetAndCallSiteTableOffset(); |
| 707 | |
| 708 | for (size_t CallSiteIdx = CSRange.CallSiteBeginIdx; |
| 709 | CallSiteIdx != CSRange.CallSiteEndIdx; ++CallSiteIdx) { |
| 710 | const CallSiteEntry &S = CallSites[CallSiteIdx]; |
| 711 | |
| 712 | MCSymbol *EHFuncBeginSym = CSRange.FragmentBeginLabel; |
| 713 | MCSymbol *EHFuncEndSym = CSRange.FragmentEndLabel; |
| 714 | |
| 715 | MCSymbol *BeginLabel = S.BeginLabel; |
| 716 | if (!BeginLabel) |
| 717 | BeginLabel = EHFuncBeginSym; |
| 718 | MCSymbol *EndLabel = S.EndLabel; |
| 719 | if (!EndLabel) |
| 720 | EndLabel = EHFuncEndSym; |
| 721 | |
| 722 | // Offset of the call site relative to the start of the procedure. |
| 723 | if (VerboseAsm) |
| 724 | Asm->OutStreamer->AddComment(T: ">> Call Site " + Twine(++Entry) + |
| 725 | " <<" ); |
| 726 | Asm->emitCallSiteOffset(Hi: BeginLabel, Lo: EHFuncBeginSym, Encoding: CallSiteEncoding); |
| 727 | if (VerboseAsm) |
| 728 | Asm->OutStreamer->AddComment(T: Twine(" Call between " ) + |
| 729 | BeginLabel->getName() + " and " + |
| 730 | EndLabel->getName()); |
| 731 | Asm->emitCallSiteOffset(Hi: EndLabel, Lo: BeginLabel, Encoding: CallSiteEncoding); |
| 732 | |
| 733 | // Offset of the landing pad relative to the start of the landing pad |
| 734 | // fragment. |
| 735 | if (!S.LPad) { |
| 736 | if (VerboseAsm) |
| 737 | Asm->OutStreamer->AddComment(T: " has no landing pad" ); |
| 738 | Asm->emitCallSiteValue(Value: 0, Encoding: CallSiteEncoding); |
| 739 | } else { |
| 740 | if (VerboseAsm) |
| 741 | Asm->OutStreamer->AddComment(T: Twine(" jumps to " ) + |
| 742 | S.LPad->LandingPadLabel->getName()); |
| 743 | Asm->emitCallSiteOffset(Hi: S.LPad->LandingPadLabel, |
| 744 | Lo: LandingPadRange->FragmentBeginLabel, |
| 745 | Encoding: CallSiteEncoding); |
| 746 | } |
| 747 | |
| 748 | // Offset of the first associated action record, relative to the start |
| 749 | // of the action table. This value is biased by 1 (1 indicates the start |
| 750 | // of the action table), and 0 indicates that there are no actions. |
| 751 | if (VerboseAsm) { |
| 752 | if (S.Action == 0) |
| 753 | Asm->OutStreamer->AddComment(T: " On action: cleanup" ); |
| 754 | else |
| 755 | Asm->OutStreamer->AddComment(T: " On action: " + |
| 756 | Twine((S.Action - 1) / 2 + 1)); |
| 757 | } |
| 758 | Asm->emitULEB128(Value: S.Action); |
| 759 | } |
| 760 | } |
| 761 | Asm->OutStreamer->emitLabel(Symbol: CstEndLabel); |
| 762 | } |
| 763 | |
| 764 | // Emit the Action Table. |
| 765 | int Entry = 0; |
| 766 | for (const ActionEntry &Action : Actions) { |
| 767 | if (VerboseAsm) { |
| 768 | // Emit comments that decode the action table. |
| 769 | Asm->OutStreamer->AddComment(T: ">> Action Record " + Twine(++Entry) + " <<" ); |
| 770 | } |
| 771 | |
| 772 | // Type Filter |
| 773 | // |
| 774 | // Used by the runtime to match the type of the thrown exception to the |
| 775 | // type of the catch clauses or the types in the exception specification. |
| 776 | if (VerboseAsm) { |
| 777 | if (Action.ValueForTypeID > 0) |
| 778 | Asm->OutStreamer->AddComment(T: " Catch TypeInfo " + |
| 779 | Twine(Action.ValueForTypeID)); |
| 780 | else if (Action.ValueForTypeID < 0) |
| 781 | Asm->OutStreamer->AddComment(T: " Filter TypeInfo " + |
| 782 | Twine(Action.ValueForTypeID)); |
| 783 | else |
| 784 | Asm->OutStreamer->AddComment(T: " Cleanup" ); |
| 785 | } |
| 786 | Asm->emitSLEB128(Value: Action.ValueForTypeID); |
| 787 | |
| 788 | // Action Record |
| 789 | if (VerboseAsm) { |
| 790 | if (Action.Previous == unsigned(-1)) { |
| 791 | Asm->OutStreamer->AddComment(T: " No further actions" ); |
| 792 | } else { |
| 793 | Asm->OutStreamer->AddComment(T: " Continue to action " + |
| 794 | Twine(Action.Previous + 1)); |
| 795 | } |
| 796 | } |
| 797 | Asm->emitSLEB128(Value: Action.NextAction); |
| 798 | } |
| 799 | |
| 800 | if (HaveTTData) { |
| 801 | Asm->emitAlignment(Alignment: Align(4)); |
| 802 | emitTypeInfos(TTypeEncoding, TTBaseLabel); |
| 803 | } |
| 804 | |
| 805 | Asm->emitAlignment(Alignment: Align(4)); |
| 806 | return GCCETSym; |
| 807 | } |
| 808 | |
| 809 | void EHStreamer::emitTypeInfos(unsigned TTypeEncoding, MCSymbol *TTBaseLabel) { |
| 810 | const MachineFunction *MF = Asm->MF; |
| 811 | const std::vector<const GlobalValue *> &TypeInfos = MF->getTypeInfos(); |
| 812 | const std::vector<unsigned> &FilterIds = MF->getFilterIds(); |
| 813 | |
| 814 | const bool VerboseAsm = Asm->OutStreamer->isVerboseAsm(); |
| 815 | |
| 816 | int Entry = 0; |
| 817 | // Emit the Catch TypeInfos. |
| 818 | if (VerboseAsm && !TypeInfos.empty()) { |
| 819 | Asm->OutStreamer->AddComment(T: ">> Catch TypeInfos <<" ); |
| 820 | Asm->OutStreamer->addBlankLine(); |
| 821 | Entry = TypeInfos.size(); |
| 822 | } |
| 823 | |
| 824 | for (const GlobalValue *GV : llvm::reverse(C: TypeInfos)) { |
| 825 | if (VerboseAsm) |
| 826 | Asm->OutStreamer->AddComment(T: "TypeInfo " + Twine(Entry--)); |
| 827 | Asm->emitTTypeReference(GV, Encoding: TTypeEncoding); |
| 828 | } |
| 829 | |
| 830 | Asm->OutStreamer->emitLabel(Symbol: TTBaseLabel); |
| 831 | |
| 832 | // Emit the Exception Specifications. |
| 833 | if (VerboseAsm && !FilterIds.empty()) { |
| 834 | Asm->OutStreamer->AddComment(T: ">> Filter TypeInfos <<" ); |
| 835 | Asm->OutStreamer->addBlankLine(); |
| 836 | Entry = 0; |
| 837 | } |
| 838 | for (std::vector<unsigned>::const_iterator |
| 839 | I = FilterIds.begin(), E = FilterIds.end(); I < E; ++I) { |
| 840 | unsigned TypeID = *I; |
| 841 | if (VerboseAsm) { |
| 842 | --Entry; |
| 843 | if (isFilterEHSelector(Selector: TypeID)) |
| 844 | Asm->OutStreamer->AddComment(T: "FilterInfo " + Twine(Entry)); |
| 845 | } |
| 846 | |
| 847 | Asm->emitULEB128(Value: TypeID); |
| 848 | } |
| 849 | } |
| 850 | |