1 | //===- lib/MC/MCWin64EH.cpp - MCWin64EH 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/MCWin64EH.h" |
10 | #include "llvm/ADT/Twine.h" |
11 | #include "llvm/MC/MCAssembler.h" |
12 | #include "llvm/MC/MCContext.h" |
13 | #include "llvm/MC/MCExpr.h" |
14 | #include "llvm/MC/MCObjectStreamer.h" |
15 | #include "llvm/MC/MCStreamer.h" |
16 | #include "llvm/MC/MCSymbol.h" |
17 | #include "llvm/MC/MCValue.h" |
18 | #include "llvm/Support/Win64EH.h" |
19 | |
20 | namespace llvm { |
21 | class MCSection; |
22 | |
23 | /// MCExpr that represents the epilog unwind code in an unwind table. |
24 | class MCUnwindV2EpilogTargetExpr final : public MCTargetExpr { |
25 | const MCSymbol *FunctionEnd; |
26 | const MCSymbol *UnwindV2Start; |
27 | const MCSymbol *EpilogEnd; |
28 | uint8_t EpilogSize; |
29 | SMLoc Loc; |
30 | |
31 | MCUnwindV2EpilogTargetExpr(const WinEH::FrameInfo &FrameInfo, |
32 | const WinEH::FrameInfo::Epilog &Epilog, |
33 | uint8_t EpilogSize_) |
34 | : FunctionEnd(FrameInfo.FuncletOrFuncEnd), |
35 | UnwindV2Start(Epilog.UnwindV2Start), EpilogEnd(Epilog.End), |
36 | EpilogSize(EpilogSize_), Loc(Epilog.Loc) {} |
37 | |
38 | public: |
39 | static MCUnwindV2EpilogTargetExpr * |
40 | create(const WinEH::FrameInfo &FrameInfo, |
41 | const WinEH::FrameInfo::Epilog &Epilog, uint8_t EpilogSize_, |
42 | MCContext &Ctx) { |
43 | return new (Ctx) MCUnwindV2EpilogTargetExpr(FrameInfo, Epilog, EpilogSize_); |
44 | } |
45 | |
46 | void printImpl(raw_ostream &OS, const MCAsmInfo *MAI) const override { |
47 | OS << ":epilog:" ; |
48 | UnwindV2Start->print(OS, MAI); |
49 | } |
50 | |
51 | bool evaluateAsRelocatableImpl(MCValue &Res, |
52 | const MCAssembler *Asm) const override; |
53 | |
54 | void visitUsedExpr(MCStreamer &Streamer) const override { |
55 | // Contains no sub-expressions. |
56 | } |
57 | |
58 | MCFragment *findAssociatedFragment() const override { |
59 | return UnwindV2Start->getFragment(); |
60 | } |
61 | }; |
62 | } |
63 | |
64 | using namespace llvm; |
65 | |
66 | // NOTE: All relocations generated here are 4-byte image-relative. |
67 | |
68 | static uint8_t CountOfUnwindCodes(std::vector<WinEH::Instruction> &Insns) { |
69 | uint8_t Count = 0; |
70 | for (const auto &I : Insns) { |
71 | switch (static_cast<Win64EH::UnwindOpcodes>(I.Operation)) { |
72 | default: |
73 | llvm_unreachable("Unsupported unwind code" ); |
74 | case Win64EH::UOP_PushNonVol: |
75 | case Win64EH::UOP_AllocSmall: |
76 | case Win64EH::UOP_SetFPReg: |
77 | case Win64EH::UOP_PushMachFrame: |
78 | Count += 1; |
79 | break; |
80 | case Win64EH::UOP_SaveNonVol: |
81 | case Win64EH::UOP_SaveXMM128: |
82 | Count += 2; |
83 | break; |
84 | case Win64EH::UOP_SaveNonVolBig: |
85 | case Win64EH::UOP_SaveXMM128Big: |
86 | Count += 3; |
87 | break; |
88 | case Win64EH::UOP_AllocLarge: |
89 | Count += (I.Offset > 512 * 1024 - 8) ? 3 : 2; |
90 | break; |
91 | } |
92 | } |
93 | return Count; |
94 | } |
95 | |
96 | static void EmitAbsDifference(MCStreamer &Streamer, const MCSymbol *LHS, |
97 | const MCSymbol *RHS) { |
98 | MCContext &Context = Streamer.getContext(); |
99 | const MCExpr *Diff = |
100 | MCBinaryExpr::createSub(LHS: MCSymbolRefExpr::create(Symbol: LHS, Ctx&: Context), |
101 | RHS: MCSymbolRefExpr::create(Symbol: RHS, Ctx&: Context), Ctx&: Context); |
102 | Streamer.emitValue(Value: Diff, Size: 1); |
103 | } |
104 | |
105 | static void EmitUnwindCode(MCStreamer &streamer, const MCSymbol *begin, |
106 | WinEH::Instruction &inst) { |
107 | uint8_t b2; |
108 | uint16_t w; |
109 | b2 = (inst.Operation & 0x0F); |
110 | switch (static_cast<Win64EH::UnwindOpcodes>(inst.Operation)) { |
111 | default: |
112 | llvm_unreachable("Unsupported unwind code" ); |
113 | case Win64EH::UOP_PushNonVol: |
114 | EmitAbsDifference(Streamer&: streamer, LHS: inst.Label, RHS: begin); |
115 | b2 |= (inst.Register & 0x0F) << 4; |
116 | streamer.emitInt8(Value: b2); |
117 | break; |
118 | case Win64EH::UOP_AllocLarge: |
119 | EmitAbsDifference(Streamer&: streamer, LHS: inst.Label, RHS: begin); |
120 | if (inst.Offset > 512 * 1024 - 8) { |
121 | b2 |= 0x10; |
122 | streamer.emitInt8(Value: b2); |
123 | w = inst.Offset & 0xFFF8; |
124 | streamer.emitInt16(Value: w); |
125 | w = inst.Offset >> 16; |
126 | } else { |
127 | streamer.emitInt8(Value: b2); |
128 | w = inst.Offset >> 3; |
129 | } |
130 | streamer.emitInt16(Value: w); |
131 | break; |
132 | case Win64EH::UOP_AllocSmall: |
133 | b2 |= (((inst.Offset - 8) >> 3) & 0x0F) << 4; |
134 | EmitAbsDifference(Streamer&: streamer, LHS: inst.Label, RHS: begin); |
135 | streamer.emitInt8(Value: b2); |
136 | break; |
137 | case Win64EH::UOP_SetFPReg: |
138 | EmitAbsDifference(Streamer&: streamer, LHS: inst.Label, RHS: begin); |
139 | streamer.emitInt8(Value: b2); |
140 | break; |
141 | case Win64EH::UOP_SaveNonVol: |
142 | case Win64EH::UOP_SaveXMM128: |
143 | b2 |= (inst.Register & 0x0F) << 4; |
144 | EmitAbsDifference(Streamer&: streamer, LHS: inst.Label, RHS: begin); |
145 | streamer.emitInt8(Value: b2); |
146 | w = inst.Offset >> 3; |
147 | if (inst.Operation == Win64EH::UOP_SaveXMM128) |
148 | w >>= 1; |
149 | streamer.emitInt16(Value: w); |
150 | break; |
151 | case Win64EH::UOP_SaveNonVolBig: |
152 | case Win64EH::UOP_SaveXMM128Big: |
153 | b2 |= (inst.Register & 0x0F) << 4; |
154 | EmitAbsDifference(Streamer&: streamer, LHS: inst.Label, RHS: begin); |
155 | streamer.emitInt8(Value: b2); |
156 | if (inst.Operation == Win64EH::UOP_SaveXMM128Big) |
157 | w = inst.Offset & 0xFFF0; |
158 | else |
159 | w = inst.Offset & 0xFFF8; |
160 | streamer.emitInt16(Value: w); |
161 | w = inst.Offset >> 16; |
162 | streamer.emitInt16(Value: w); |
163 | break; |
164 | case Win64EH::UOP_PushMachFrame: |
165 | if (inst.Offset == 1) |
166 | b2 |= 0x10; |
167 | EmitAbsDifference(Streamer&: streamer, LHS: inst.Label, RHS: begin); |
168 | streamer.emitInt8(Value: b2); |
169 | break; |
170 | } |
171 | } |
172 | |
173 | static void EmitSymbolRefWithOfs(MCStreamer &streamer, |
174 | const MCSymbol *Base, |
175 | int64_t Offset) { |
176 | MCContext &Context = streamer.getContext(); |
177 | const MCConstantExpr *OffExpr = MCConstantExpr::create(Value: Offset, Ctx&: Context); |
178 | const MCSymbolRefExpr *BaseRefRel = MCSymbolRefExpr::create(Symbol: Base, |
179 | specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, |
180 | Ctx&: Context); |
181 | streamer.emitValue(Value: MCBinaryExpr::createAdd(LHS: BaseRefRel, RHS: OffExpr, Ctx&: Context), Size: 4); |
182 | } |
183 | |
184 | static void EmitSymbolRefWithOfs(MCStreamer &streamer, |
185 | const MCSymbol *Base, |
186 | const MCSymbol *Other) { |
187 | MCContext &Context = streamer.getContext(); |
188 | const MCSymbolRefExpr *BaseRef = MCSymbolRefExpr::create(Symbol: Base, Ctx&: Context); |
189 | const MCSymbolRefExpr *OtherRef = MCSymbolRefExpr::create(Symbol: Other, Ctx&: Context); |
190 | const MCExpr *Ofs = MCBinaryExpr::createSub(LHS: OtherRef, RHS: BaseRef, Ctx&: Context); |
191 | const MCSymbolRefExpr *BaseRefRel = MCSymbolRefExpr::create(Symbol: Base, |
192 | specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, |
193 | Ctx&: Context); |
194 | streamer.emitValue(Value: MCBinaryExpr::createAdd(LHS: BaseRefRel, RHS: Ofs, Ctx&: Context), Size: 4); |
195 | } |
196 | |
197 | static void EmitRuntimeFunction(MCStreamer &streamer, |
198 | const WinEH::FrameInfo *info) { |
199 | MCContext &context = streamer.getContext(); |
200 | |
201 | streamer.emitValueToAlignment(Alignment: Align(4)); |
202 | EmitSymbolRefWithOfs(streamer, Base: info->Begin, Other: info->Begin); |
203 | EmitSymbolRefWithOfs(streamer, Base: info->Begin, Other: info->End); |
204 | streamer.emitValue(Value: MCSymbolRefExpr::create(Symbol: info->Symbol, |
205 | specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, |
206 | Ctx&: context), Size: 4); |
207 | } |
208 | |
209 | static std::optional<int64_t> |
210 | GetOptionalAbsDifference(const MCAssembler &Assembler, const MCSymbol *LHS, |
211 | const MCSymbol *RHS) { |
212 | MCContext &Context = Assembler.getContext(); |
213 | const MCExpr *Diff = |
214 | MCBinaryExpr::createSub(LHS: MCSymbolRefExpr::create(Symbol: LHS, Ctx&: Context), |
215 | RHS: MCSymbolRefExpr::create(Symbol: RHS, Ctx&: Context), Ctx&: Context); |
216 | // It should normally be possible to calculate the length of a function |
217 | // at this point, but it might not be possible in the presence of certain |
218 | // unusual constructs, like an inline asm with an alignment directive. |
219 | int64_t value; |
220 | if (!Diff->evaluateAsAbsolute(Res&: value, Asm: Assembler)) |
221 | return std::nullopt; |
222 | return value; |
223 | } |
224 | |
225 | static void EmitUnwindInfo(MCStreamer &streamer, WinEH::FrameInfo *info) { |
226 | // If this UNWIND_INFO already has a symbol, it's already been emitted. |
227 | if (info->Symbol) |
228 | return; |
229 | |
230 | MCContext &context = streamer.getContext(); |
231 | MCObjectStreamer *OS = (MCObjectStreamer *)(&streamer); |
232 | MCSymbol *Label = context.createTempSymbol(); |
233 | |
234 | streamer.emitValueToAlignment(Alignment: Align(4)); |
235 | streamer.emitLabel(Symbol: Label); |
236 | info->Symbol = Label; |
237 | |
238 | uint8_t numCodes = CountOfUnwindCodes(Insns&: info->Instructions); |
239 | bool LastEpilogIsAtEnd = false; |
240 | bool AddPaddingEpilogCode = false; |
241 | uint8_t EpilogSize = 0; |
242 | bool EnableUnwindV2 = (info->Version >= 2) && !info->EpilogMap.empty(); |
243 | if (EnableUnwindV2) { |
244 | auto &LastEpilog = info->EpilogMap.back().second; |
245 | |
246 | // Calculate the size of the epilogs. Note that we +1 to the size so that |
247 | // the terminator instruction is also included in the epilog (the Windows |
248 | // unwinder does a simple range check versus the current instruction pointer |
249 | // so, although there are terminators that are large than 1 byte, the |
250 | // starting address of the terminator instruction will always be considered |
251 | // inside the epilog). |
252 | auto MaybeSize = GetOptionalAbsDifference( |
253 | Assembler: OS->getAssembler(), LHS: LastEpilog.End, RHS: LastEpilog.UnwindV2Start); |
254 | if (!MaybeSize) { |
255 | context.reportError(L: LastEpilog.Loc, |
256 | Msg: "Failed to evaluate epilog size for Unwind v2" ); |
257 | return; |
258 | } |
259 | assert(*MaybeSize >= 0); |
260 | if (*MaybeSize >= (int64_t)UINT8_MAX) { |
261 | context.reportError(L: LastEpilog.Loc, |
262 | Msg: "Epilog size is too large for Unwind v2" ); |
263 | return; |
264 | } |
265 | EpilogSize = *MaybeSize + 1; |
266 | |
267 | // If the last epilog is at the end of the function, we can use a special |
268 | // encoding for it. Because of our +1 trick for the size, this will only |
269 | // work where that final terminator instruction is 1 byte long. |
270 | auto LastEpilogToFuncEnd = GetOptionalAbsDifference( |
271 | Assembler: OS->getAssembler(), LHS: info->FuncletOrFuncEnd, RHS: LastEpilog.UnwindV2Start); |
272 | LastEpilogIsAtEnd = (LastEpilogToFuncEnd == EpilogSize); |
273 | |
274 | // If we have an odd number of epilog codes, we need to add a padding code. |
275 | size_t numEpilogCodes = |
276 | info->EpilogMap.size() + (LastEpilogIsAtEnd ? 0 : 1); |
277 | if ((numEpilogCodes % 2) != 0) { |
278 | AddPaddingEpilogCode = true; |
279 | numEpilogCodes++; |
280 | } |
281 | |
282 | // Too many epilogs to handle. |
283 | if ((size_t)numCodes + numEpilogCodes > UINT8_MAX) { |
284 | context.reportError(L: info->FunctionLoc, |
285 | Msg: "Too many unwind codes with Unwind v2 enabled" ); |
286 | return; |
287 | } |
288 | |
289 | numCodes += numEpilogCodes; |
290 | } |
291 | |
292 | // Upper 3 bits are the version number. |
293 | uint8_t flags = info->Version; |
294 | if (info->ChainedParent) |
295 | flags |= Win64EH::UNW_ChainInfo << 3; |
296 | else { |
297 | if (info->HandlesUnwind) |
298 | flags |= Win64EH::UNW_TerminateHandler << 3; |
299 | if (info->HandlesExceptions) |
300 | flags |= Win64EH::UNW_ExceptionHandler << 3; |
301 | } |
302 | streamer.emitInt8(Value: flags); |
303 | |
304 | if (info->PrologEnd) |
305 | EmitAbsDifference(Streamer&: streamer, LHS: info->PrologEnd, RHS: info->Begin); |
306 | else |
307 | streamer.emitInt8(Value: 0); |
308 | |
309 | streamer.emitInt8(Value: numCodes); |
310 | |
311 | uint8_t frame = 0; |
312 | if (info->LastFrameInst >= 0) { |
313 | WinEH::Instruction &frameInst = info->Instructions[info->LastFrameInst]; |
314 | assert(frameInst.Operation == Win64EH::UOP_SetFPReg); |
315 | frame = (frameInst.Register & 0x0F) | (frameInst.Offset & 0xF0); |
316 | } |
317 | streamer.emitInt8(Value: frame); |
318 | |
319 | // Emit the epilog instructions. |
320 | if (EnableUnwindV2) { |
321 | MCDataFragment *DF = OS->getOrCreateDataFragment(); |
322 | |
323 | bool IsLast = true; |
324 | for (const auto &Epilog : llvm::reverse(C&: info->EpilogMap)) { |
325 | if (IsLast) { |
326 | IsLast = false; |
327 | uint8_t Flags = LastEpilogIsAtEnd ? 0x01 : 0; |
328 | streamer.emitInt8(Value: EpilogSize); |
329 | streamer.emitInt8(Value: (Flags << 4) | Win64EH::UOP_Epilog); |
330 | |
331 | if (LastEpilogIsAtEnd) |
332 | continue; |
333 | } |
334 | |
335 | // Each epilog is emitted as a fixup, since we can't measure the distance |
336 | // between the start of the epilog and the end of the function until |
337 | // layout has been completed. |
338 | auto *MCE = MCUnwindV2EpilogTargetExpr::create(FrameInfo: *info, Epilog: Epilog.second, |
339 | EpilogSize_: EpilogSize, Ctx&: context); |
340 | MCFixup Fixup = MCFixup::create(Offset: DF->getContents().size(), Value: MCE, Kind: FK_Data_2); |
341 | DF->addFixup(Fixup); |
342 | DF->appendContents(Num: 2, Elt: 0); |
343 | } |
344 | } |
345 | if (AddPaddingEpilogCode) |
346 | streamer.emitInt16(Value: Win64EH::UOP_Epilog << 8); |
347 | |
348 | // Emit unwind instructions (in reverse order). |
349 | uint8_t numInst = info->Instructions.size(); |
350 | for (uint8_t c = 0; c < numInst; ++c) { |
351 | WinEH::Instruction inst = info->Instructions.back(); |
352 | info->Instructions.pop_back(); |
353 | EmitUnwindCode(streamer, begin: info->Begin, inst); |
354 | } |
355 | |
356 | // For alignment purposes, the instruction array will always have an even |
357 | // number of entries, with the final entry potentially unused (in which case |
358 | // the array will be one longer than indicated by the count of unwind codes |
359 | // field). |
360 | if (numCodes & 1) { |
361 | streamer.emitInt16(Value: 0); |
362 | } |
363 | |
364 | if (flags & (Win64EH::UNW_ChainInfo << 3)) |
365 | EmitRuntimeFunction(streamer, info: info->ChainedParent); |
366 | else if (flags & |
367 | ((Win64EH::UNW_TerminateHandler|Win64EH::UNW_ExceptionHandler) << 3)) |
368 | streamer.emitValue(Value: MCSymbolRefExpr::create(Symbol: info->ExceptionHandler, |
369 | specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, |
370 | Ctx&: context), Size: 4); |
371 | else if (numCodes == 0) { |
372 | // The minimum size of an UNWIND_INFO struct is 8 bytes. If we're not |
373 | // a chained unwind info, if there is no handler, and if there are fewer |
374 | // than 2 slots used in the unwind code array, we have to pad to 8 bytes. |
375 | streamer.emitInt32(Value: 0); |
376 | } |
377 | } |
378 | |
379 | bool MCUnwindV2EpilogTargetExpr::evaluateAsRelocatableImpl( |
380 | MCValue &Res, const MCAssembler *Asm) const { |
381 | // Calculate the offset to this epilog, and validate it's within the allowed |
382 | // range. |
383 | auto Offset = GetOptionalAbsDifference(Assembler: *Asm, LHS: FunctionEnd, RHS: UnwindV2Start); |
384 | if (!Offset) { |
385 | Asm->getContext().reportError( |
386 | L: Loc, Msg: "Failed to evaluate epilog offset for Unwind v2" ); |
387 | return false; |
388 | } |
389 | assert(*Offset > 0); |
390 | constexpr uint16_t MaxEpilogOffset = 0x0fff; |
391 | if (*Offset > MaxEpilogOffset) { |
392 | Asm->getContext().reportError(L: Loc, |
393 | Msg: "Epilog offset is too large for Unwind v2" ); |
394 | return false; |
395 | } |
396 | |
397 | // Sanity check that all epilogs are the same size. |
398 | auto Size = GetOptionalAbsDifference(Assembler: *Asm, LHS: EpilogEnd, RHS: UnwindV2Start); |
399 | if (Size != (EpilogSize - 1)) { |
400 | Asm->getContext().reportError( |
401 | L: Loc, |
402 | Msg: "Size of this epilog does not match size of last epilog in function" ); |
403 | return false; |
404 | } |
405 | |
406 | auto HighBits = *Offset >> 8; |
407 | Res = MCValue::get(Val: (HighBits << 12) | (Win64EH::UOP_Epilog << 8) | |
408 | (*Offset & 0xFF)); |
409 | return true; |
410 | } |
411 | |
412 | void llvm::Win64EH::UnwindEmitter::Emit(MCStreamer &Streamer) const { |
413 | // Emit the unwind info structs first. |
414 | for (const auto &CFI : Streamer.getWinFrameInfos()) { |
415 | MCSection *XData = Streamer.getAssociatedXDataSection(TextSec: CFI->TextSection); |
416 | Streamer.switchSection(Section: XData); |
417 | ::EmitUnwindInfo(streamer&: Streamer, info: CFI.get()); |
418 | } |
419 | |
420 | // Now emit RUNTIME_FUNCTION entries. |
421 | for (const auto &CFI : Streamer.getWinFrameInfos()) { |
422 | MCSection *PData = Streamer.getAssociatedPDataSection(TextSec: CFI->TextSection); |
423 | Streamer.switchSection(Section: PData); |
424 | EmitRuntimeFunction(streamer&: Streamer, info: CFI.get()); |
425 | } |
426 | } |
427 | |
428 | void llvm::Win64EH::UnwindEmitter::EmitUnwindInfo(MCStreamer &Streamer, |
429 | WinEH::FrameInfo *info, |
430 | bool HandlerData) const { |
431 | // Switch sections (the static function above is meant to be called from |
432 | // here and from Emit(). |
433 | MCSection *XData = Streamer.getAssociatedXDataSection(TextSec: info->TextSection); |
434 | Streamer.switchSection(Section: XData); |
435 | |
436 | ::EmitUnwindInfo(streamer&: Streamer, info); |
437 | } |
438 | |
439 | static const MCExpr *GetSubDivExpr(MCStreamer &Streamer, const MCSymbol *LHS, |
440 | const MCSymbol *RHS, int Div) { |
441 | MCContext &Context = Streamer.getContext(); |
442 | const MCExpr *Expr = |
443 | MCBinaryExpr::createSub(LHS: MCSymbolRefExpr::create(Symbol: LHS, Ctx&: Context), |
444 | RHS: MCSymbolRefExpr::create(Symbol: RHS, Ctx&: Context), Ctx&: Context); |
445 | if (Div != 1) |
446 | Expr = MCBinaryExpr::createDiv(LHS: Expr, RHS: MCConstantExpr::create(Value: Div, Ctx&: Context), |
447 | Ctx&: Context); |
448 | return Expr; |
449 | } |
450 | |
451 | static std::optional<int64_t> GetOptionalAbsDifference(MCStreamer &Streamer, |
452 | const MCSymbol *LHS, |
453 | const MCSymbol *RHS) { |
454 | MCObjectStreamer *OS = (MCObjectStreamer *)(&Streamer); |
455 | return GetOptionalAbsDifference(Assembler: OS->getAssembler(), LHS, RHS); |
456 | } |
457 | |
458 | static int64_t GetAbsDifference(MCStreamer &Streamer, const MCSymbol *LHS, |
459 | const MCSymbol *RHS) { |
460 | std::optional<int64_t> MaybeDiff = |
461 | GetOptionalAbsDifference(Streamer, LHS, RHS); |
462 | if (!MaybeDiff) |
463 | report_fatal_error(reason: "Failed to evaluate function length in SEH unwind info" ); |
464 | return *MaybeDiff; |
465 | } |
466 | |
467 | static void checkARM64Instructions(MCStreamer &Streamer, |
468 | ArrayRef<WinEH::Instruction> Insns, |
469 | const MCSymbol *Begin, const MCSymbol *End, |
470 | StringRef Name, StringRef Type) { |
471 | if (!End) |
472 | return; |
473 | std::optional<int64_t> MaybeDistance = |
474 | GetOptionalAbsDifference(Streamer, LHS: End, RHS: Begin); |
475 | if (!MaybeDistance) |
476 | return; |
477 | uint32_t Distance = (uint32_t)*MaybeDistance; |
478 | |
479 | for (const auto &I : Insns) { |
480 | switch (static_cast<Win64EH::UnwindOpcodes>(I.Operation)) { |
481 | default: |
482 | break; |
483 | case Win64EH::UOP_TrapFrame: |
484 | case Win64EH::UOP_PushMachFrame: |
485 | case Win64EH::UOP_Context: |
486 | case Win64EH::UOP_ECContext: |
487 | case Win64EH::UOP_ClearUnwoundToCall: |
488 | // Can't reason about these opcodes and how they map to actual |
489 | // instructions. |
490 | return; |
491 | } |
492 | } |
493 | // Exclude the end opcode which doesn't map to an instruction. |
494 | uint32_t InstructionBytes = 4 * (Insns.size() - 1); |
495 | if (Distance != InstructionBytes) { |
496 | Streamer.getContext().reportError( |
497 | L: SMLoc(), Msg: "Incorrect size for " + Name + " " + Type + ": " + |
498 | Twine(Distance) + |
499 | " bytes of instructions in range, but .seh directives " |
500 | "corresponding to " + |
501 | Twine(InstructionBytes) + " bytes\n" ); |
502 | } |
503 | } |
504 | |
505 | static uint32_t ARM64CountOfUnwindCodes(ArrayRef<WinEH::Instruction> Insns) { |
506 | uint32_t Count = 0; |
507 | for (const auto &I : Insns) { |
508 | switch (static_cast<Win64EH::UnwindOpcodes>(I.Operation)) { |
509 | default: |
510 | llvm_unreachable("Unsupported ARM64 unwind code" ); |
511 | case Win64EH::UOP_AllocSmall: |
512 | Count += 1; |
513 | break; |
514 | case Win64EH::UOP_AllocMedium: |
515 | Count += 2; |
516 | break; |
517 | case Win64EH::UOP_AllocLarge: |
518 | Count += 4; |
519 | break; |
520 | case Win64EH::UOP_SaveR19R20X: |
521 | Count += 1; |
522 | break; |
523 | case Win64EH::UOP_SaveFPLRX: |
524 | Count += 1; |
525 | break; |
526 | case Win64EH::UOP_SaveFPLR: |
527 | Count += 1; |
528 | break; |
529 | case Win64EH::UOP_SaveReg: |
530 | Count += 2; |
531 | break; |
532 | case Win64EH::UOP_SaveRegP: |
533 | Count += 2; |
534 | break; |
535 | case Win64EH::UOP_SaveRegPX: |
536 | Count += 2; |
537 | break; |
538 | case Win64EH::UOP_SaveRegX: |
539 | Count += 2; |
540 | break; |
541 | case Win64EH::UOP_SaveLRPair: |
542 | Count += 2; |
543 | break; |
544 | case Win64EH::UOP_SaveFReg: |
545 | Count += 2; |
546 | break; |
547 | case Win64EH::UOP_SaveFRegP: |
548 | Count += 2; |
549 | break; |
550 | case Win64EH::UOP_SaveFRegX: |
551 | Count += 2; |
552 | break; |
553 | case Win64EH::UOP_SaveFRegPX: |
554 | Count += 2; |
555 | break; |
556 | case Win64EH::UOP_SetFP: |
557 | Count += 1; |
558 | break; |
559 | case Win64EH::UOP_AddFP: |
560 | Count += 2; |
561 | break; |
562 | case Win64EH::UOP_Nop: |
563 | Count += 1; |
564 | break; |
565 | case Win64EH::UOP_End: |
566 | Count += 1; |
567 | break; |
568 | case Win64EH::UOP_SaveNext: |
569 | Count += 1; |
570 | break; |
571 | case Win64EH::UOP_TrapFrame: |
572 | Count += 1; |
573 | break; |
574 | case Win64EH::UOP_PushMachFrame: |
575 | Count += 1; |
576 | break; |
577 | case Win64EH::UOP_Context: |
578 | Count += 1; |
579 | break; |
580 | case Win64EH::UOP_ECContext: |
581 | Count += 1; |
582 | break; |
583 | case Win64EH::UOP_ClearUnwoundToCall: |
584 | Count += 1; |
585 | break; |
586 | case Win64EH::UOP_PACSignLR: |
587 | Count += 1; |
588 | break; |
589 | case Win64EH::UOP_AllocZ: |
590 | Count += 2; |
591 | break; |
592 | case Win64EH::UOP_SaveAnyRegI: |
593 | case Win64EH::UOP_SaveAnyRegIP: |
594 | case Win64EH::UOP_SaveAnyRegD: |
595 | case Win64EH::UOP_SaveAnyRegDP: |
596 | case Win64EH::UOP_SaveAnyRegQ: |
597 | case Win64EH::UOP_SaveAnyRegQP: |
598 | case Win64EH::UOP_SaveAnyRegIX: |
599 | case Win64EH::UOP_SaveAnyRegIPX: |
600 | case Win64EH::UOP_SaveAnyRegDX: |
601 | case Win64EH::UOP_SaveAnyRegDPX: |
602 | case Win64EH::UOP_SaveAnyRegQX: |
603 | case Win64EH::UOP_SaveAnyRegQPX: |
604 | case Win64EH::UOP_SaveZReg: |
605 | case Win64EH::UOP_SavePReg: |
606 | Count += 3; |
607 | break; |
608 | } |
609 | } |
610 | return Count; |
611 | } |
612 | |
613 | // Unwind opcode encodings and restrictions are documented at |
614 | // https://docs.microsoft.com/en-us/cpp/build/arm64-exception-handling |
615 | static void ARM64EmitUnwindCode(MCStreamer &streamer, |
616 | const WinEH::Instruction &inst) { |
617 | uint8_t b, reg; |
618 | switch (static_cast<Win64EH::UnwindOpcodes>(inst.Operation)) { |
619 | default: |
620 | llvm_unreachable("Unsupported ARM64 unwind code" ); |
621 | case Win64EH::UOP_AllocSmall: |
622 | b = (inst.Offset >> 4) & 0x1F; |
623 | streamer.emitInt8(Value: b); |
624 | break; |
625 | case Win64EH::UOP_AllocMedium: { |
626 | uint16_t hw = (inst.Offset >> 4) & 0x7FF; |
627 | b = 0xC0; |
628 | b |= (hw >> 8); |
629 | streamer.emitInt8(Value: b); |
630 | b = hw & 0xFF; |
631 | streamer.emitInt8(Value: b); |
632 | break; |
633 | } |
634 | case Win64EH::UOP_AllocLarge: { |
635 | uint32_t w; |
636 | b = 0xE0; |
637 | streamer.emitInt8(Value: b); |
638 | w = inst.Offset >> 4; |
639 | b = (w & 0x00FF0000) >> 16; |
640 | streamer.emitInt8(Value: b); |
641 | b = (w & 0x0000FF00) >> 8; |
642 | streamer.emitInt8(Value: b); |
643 | b = w & 0x000000FF; |
644 | streamer.emitInt8(Value: b); |
645 | break; |
646 | } |
647 | case Win64EH::UOP_SetFP: |
648 | b = 0xE1; |
649 | streamer.emitInt8(Value: b); |
650 | break; |
651 | case Win64EH::UOP_AddFP: |
652 | b = 0xE2; |
653 | streamer.emitInt8(Value: b); |
654 | b = (inst.Offset >> 3); |
655 | streamer.emitInt8(Value: b); |
656 | break; |
657 | case Win64EH::UOP_Nop: |
658 | b = 0xE3; |
659 | streamer.emitInt8(Value: b); |
660 | break; |
661 | case Win64EH::UOP_SaveR19R20X: |
662 | b = 0x20; |
663 | b |= (inst.Offset >> 3) & 0x1F; |
664 | streamer.emitInt8(Value: b); |
665 | break; |
666 | case Win64EH::UOP_SaveFPLRX: |
667 | b = 0x80; |
668 | b |= ((inst.Offset - 1) >> 3) & 0x3F; |
669 | streamer.emitInt8(Value: b); |
670 | break; |
671 | case Win64EH::UOP_SaveFPLR: |
672 | b = 0x40; |
673 | b |= (inst.Offset >> 3) & 0x3F; |
674 | streamer.emitInt8(Value: b); |
675 | break; |
676 | case Win64EH::UOP_SaveReg: |
677 | assert(inst.Register >= 19 && "Saved reg must be >= 19" ); |
678 | reg = inst.Register - 19; |
679 | b = 0xD0 | ((reg & 0xC) >> 2); |
680 | streamer.emitInt8(Value: b); |
681 | b = ((reg & 0x3) << 6) | (inst.Offset >> 3); |
682 | streamer.emitInt8(Value: b); |
683 | break; |
684 | case Win64EH::UOP_SaveRegX: |
685 | assert(inst.Register >= 19 && "Saved reg must be >= 19" ); |
686 | reg = inst.Register - 19; |
687 | b = 0xD4 | ((reg & 0x8) >> 3); |
688 | streamer.emitInt8(Value: b); |
689 | b = ((reg & 0x7) << 5) | ((inst.Offset >> 3) - 1); |
690 | streamer.emitInt8(Value: b); |
691 | break; |
692 | case Win64EH::UOP_SaveRegP: |
693 | assert(inst.Register >= 19 && "Saved registers must be >= 19" ); |
694 | reg = inst.Register - 19; |
695 | b = 0xC8 | ((reg & 0xC) >> 2); |
696 | streamer.emitInt8(Value: b); |
697 | b = ((reg & 0x3) << 6) | (inst.Offset >> 3); |
698 | streamer.emitInt8(Value: b); |
699 | break; |
700 | case Win64EH::UOP_SaveRegPX: |
701 | assert(inst.Register >= 19 && "Saved registers must be >= 19" ); |
702 | reg = inst.Register - 19; |
703 | b = 0xCC | ((reg & 0xC) >> 2); |
704 | streamer.emitInt8(Value: b); |
705 | b = ((reg & 0x3) << 6) | ((inst.Offset >> 3) - 1); |
706 | streamer.emitInt8(Value: b); |
707 | break; |
708 | case Win64EH::UOP_SaveLRPair: |
709 | assert(inst.Register >= 19 && "Saved reg must be >= 19" ); |
710 | reg = inst.Register - 19; |
711 | assert((reg % 2) == 0 && "Saved reg must be 19+2*X" ); |
712 | reg /= 2; |
713 | b = 0xD6 | ((reg & 0x7) >> 2); |
714 | streamer.emitInt8(Value: b); |
715 | b = ((reg & 0x3) << 6) | (inst.Offset >> 3); |
716 | streamer.emitInt8(Value: b); |
717 | break; |
718 | case Win64EH::UOP_SaveFReg: |
719 | assert(inst.Register >= 8 && "Saved dreg must be >= 8" ); |
720 | reg = inst.Register - 8; |
721 | b = 0xDC | ((reg & 0x4) >> 2); |
722 | streamer.emitInt8(Value: b); |
723 | b = ((reg & 0x3) << 6) | (inst.Offset >> 3); |
724 | streamer.emitInt8(Value: b); |
725 | break; |
726 | case Win64EH::UOP_SaveFRegX: |
727 | assert(inst.Register >= 8 && "Saved dreg must be >= 8" ); |
728 | reg = inst.Register - 8; |
729 | b = 0xDE; |
730 | streamer.emitInt8(Value: b); |
731 | b = ((reg & 0x7) << 5) | ((inst.Offset >> 3) - 1); |
732 | streamer.emitInt8(Value: b); |
733 | break; |
734 | case Win64EH::UOP_SaveFRegP: |
735 | assert(inst.Register >= 8 && "Saved dregs must be >= 8" ); |
736 | reg = inst.Register - 8; |
737 | b = 0xD8 | ((reg & 0x4) >> 2); |
738 | streamer.emitInt8(Value: b); |
739 | b = ((reg & 0x3) << 6) | (inst.Offset >> 3); |
740 | streamer.emitInt8(Value: b); |
741 | break; |
742 | case Win64EH::UOP_SaveFRegPX: |
743 | assert(inst.Register >= 8 && "Saved dregs must be >= 8" ); |
744 | reg = inst.Register - 8; |
745 | b = 0xDA | ((reg & 0x4) >> 2); |
746 | streamer.emitInt8(Value: b); |
747 | b = ((reg & 0x3) << 6) | ((inst.Offset >> 3) - 1); |
748 | streamer.emitInt8(Value: b); |
749 | break; |
750 | case Win64EH::UOP_End: |
751 | b = 0xE4; |
752 | streamer.emitInt8(Value: b); |
753 | break; |
754 | case Win64EH::UOP_SaveNext: |
755 | b = 0xE6; |
756 | streamer.emitInt8(Value: b); |
757 | break; |
758 | case Win64EH::UOP_TrapFrame: |
759 | b = 0xE8; |
760 | streamer.emitInt8(Value: b); |
761 | break; |
762 | case Win64EH::UOP_PushMachFrame: |
763 | b = 0xE9; |
764 | streamer.emitInt8(Value: b); |
765 | break; |
766 | case Win64EH::UOP_Context: |
767 | b = 0xEA; |
768 | streamer.emitInt8(Value: b); |
769 | break; |
770 | case Win64EH::UOP_ECContext: |
771 | b = 0xEB; |
772 | streamer.emitInt8(Value: b); |
773 | break; |
774 | case Win64EH::UOP_ClearUnwoundToCall: |
775 | b = 0xEC; |
776 | streamer.emitInt8(Value: b); |
777 | break; |
778 | case Win64EH::UOP_PACSignLR: |
779 | b = 0xFC; |
780 | streamer.emitInt8(Value: b); |
781 | break; |
782 | case Win64EH::UOP_SaveAnyRegI: |
783 | case Win64EH::UOP_SaveAnyRegIP: |
784 | case Win64EH::UOP_SaveAnyRegD: |
785 | case Win64EH::UOP_SaveAnyRegDP: |
786 | case Win64EH::UOP_SaveAnyRegQ: |
787 | case Win64EH::UOP_SaveAnyRegQP: |
788 | case Win64EH::UOP_SaveAnyRegIX: |
789 | case Win64EH::UOP_SaveAnyRegIPX: |
790 | case Win64EH::UOP_SaveAnyRegDX: |
791 | case Win64EH::UOP_SaveAnyRegDPX: |
792 | case Win64EH::UOP_SaveAnyRegQX: |
793 | case Win64EH::UOP_SaveAnyRegQPX: { |
794 | // This assumes the opcodes are listed in the enum in a particular order. |
795 | int Op = inst.Operation - Win64EH::UOP_SaveAnyRegI; |
796 | int Writeback = Op / 6; |
797 | int Paired = Op % 2; |
798 | int Mode = (Op / 2) % 3; |
799 | int Offset = inst.Offset >> 3; |
800 | if (Writeback || Paired || Mode == 2) |
801 | Offset >>= 1; |
802 | if (Writeback) |
803 | --Offset; |
804 | b = 0xE7; |
805 | streamer.emitInt8(Value: b); |
806 | assert(inst.Register < 32); |
807 | b = inst.Register | (Writeback << 5) | (Paired << 6); |
808 | streamer.emitInt8(Value: b); |
809 | b = Offset | (Mode << 6); |
810 | streamer.emitInt8(Value: b); |
811 | break; |
812 | } |
813 | case Win64EH::UOP_AllocZ: { |
814 | b = 0xDF; |
815 | streamer.emitInt8(Value: b); |
816 | b = inst.Offset; |
817 | streamer.emitInt8(Value: b); |
818 | break; |
819 | } |
820 | case Win64EH::UOP_SaveZReg: { |
821 | assert(inst.Register >= 8 && inst.Register <= 23); |
822 | assert(inst.Offset < 256); |
823 | b = 0xE7; |
824 | streamer.emitInt8(Value: b); |
825 | reg = inst.Register - 8; |
826 | b = ((inst.Offset & 0xC0) >> 1) | reg; |
827 | streamer.emitInt8(Value: b); |
828 | b = 0xC0 | (inst.Offset & 0x3F); |
829 | streamer.emitInt8(Value: b); |
830 | break; |
831 | } |
832 | case Win64EH::UOP_SavePReg: { |
833 | assert(inst.Register >= 4 && inst.Register <= 15); |
834 | assert(inst.Offset < 256); |
835 | b = 0xE7; |
836 | streamer.emitInt8(Value: b); |
837 | reg = inst.Register; |
838 | b = ((inst.Offset & 0xC0) >> 1) | 0x10 | reg; |
839 | streamer.emitInt8(Value: b); |
840 | b = 0xC0 | (inst.Offset & 0x3F); |
841 | streamer.emitInt8(Value: b); |
842 | break; |
843 | } |
844 | } |
845 | } |
846 | |
847 | // Returns the epilog symbol of an epilog with the exact same unwind code |
848 | // sequence, if it exists. Otherwise, returns nullptr. |
849 | // EpilogInstrs - Unwind codes for the current epilog. |
850 | // Epilogs - Epilogs that potentialy match the current epilog. |
851 | static MCSymbol* |
852 | FindMatchingEpilog(const std::vector<WinEH::Instruction>& EpilogInstrs, |
853 | const std::vector<MCSymbol *>& Epilogs, |
854 | const WinEH::FrameInfo *info) { |
855 | for (auto *EpilogStart : Epilogs) { |
856 | auto InstrsIter = info->EpilogMap.find(Key: EpilogStart); |
857 | assert(InstrsIter != info->EpilogMap.end() && |
858 | "Epilog not found in EpilogMap" ); |
859 | const auto &Instrs = InstrsIter->second.Instructions; |
860 | |
861 | if (Instrs.size() != EpilogInstrs.size()) |
862 | continue; |
863 | |
864 | bool Match = true; |
865 | for (unsigned i = 0; i < Instrs.size(); ++i) |
866 | if (Instrs[i] != EpilogInstrs[i]) { |
867 | Match = false; |
868 | break; |
869 | } |
870 | |
871 | if (Match) |
872 | return EpilogStart; |
873 | } |
874 | return nullptr; |
875 | } |
876 | |
877 | static void simplifyARM64Opcodes(std::vector<WinEH::Instruction> &Instructions, |
878 | bool Reverse) { |
879 | unsigned PrevOffset = -1; |
880 | unsigned PrevRegister = -1; |
881 | |
882 | auto VisitInstruction = [&](WinEH::Instruction &Inst) { |
883 | // Convert 2-byte opcodes into equivalent 1-byte ones. |
884 | if (Inst.Operation == Win64EH::UOP_SaveRegP && Inst.Register == 29) { |
885 | Inst.Operation = Win64EH::UOP_SaveFPLR; |
886 | Inst.Register = -1; |
887 | } else if (Inst.Operation == Win64EH::UOP_SaveRegPX && |
888 | Inst.Register == 29) { |
889 | Inst.Operation = Win64EH::UOP_SaveFPLRX; |
890 | Inst.Register = -1; |
891 | } else if (Inst.Operation == Win64EH::UOP_SaveRegPX && |
892 | Inst.Register == 19 && Inst.Offset <= 248) { |
893 | Inst.Operation = Win64EH::UOP_SaveR19R20X; |
894 | Inst.Register = -1; |
895 | } else if (Inst.Operation == Win64EH::UOP_AddFP && Inst.Offset == 0) { |
896 | Inst.Operation = Win64EH::UOP_SetFP; |
897 | } else if (Inst.Operation == Win64EH::UOP_SaveRegP && |
898 | Inst.Register == PrevRegister + 2 && |
899 | Inst.Offset == PrevOffset + 16) { |
900 | Inst.Operation = Win64EH::UOP_SaveNext; |
901 | Inst.Register = -1; |
902 | Inst.Offset = 0; |
903 | // Intentionally not creating UOP_SaveNext for float register pairs, |
904 | // as current versions of Windows (up to at least 20.04) is buggy |
905 | // regarding SaveNext for float pairs. |
906 | } |
907 | // Update info about the previous instruction, for detecting if |
908 | // the next one can be made a UOP_SaveNext |
909 | if (Inst.Operation == Win64EH::UOP_SaveR19R20X) { |
910 | PrevOffset = 0; |
911 | PrevRegister = 19; |
912 | } else if (Inst.Operation == Win64EH::UOP_SaveRegPX) { |
913 | PrevOffset = 0; |
914 | PrevRegister = Inst.Register; |
915 | } else if (Inst.Operation == Win64EH::UOP_SaveRegP) { |
916 | PrevOffset = Inst.Offset; |
917 | PrevRegister = Inst.Register; |
918 | } else if (Inst.Operation == Win64EH::UOP_SaveNext) { |
919 | PrevRegister += 2; |
920 | PrevOffset += 16; |
921 | } else { |
922 | PrevRegister = -1; |
923 | PrevOffset = -1; |
924 | } |
925 | }; |
926 | |
927 | // Iterate over instructions in a forward order (for prologues), |
928 | // backwards for epilogues (i.e. always reverse compared to how the |
929 | // opcodes are stored). |
930 | if (Reverse) { |
931 | for (auto It = Instructions.rbegin(); It != Instructions.rend(); It++) |
932 | VisitInstruction(*It); |
933 | } else { |
934 | for (WinEH::Instruction &Inst : Instructions) |
935 | VisitInstruction(Inst); |
936 | } |
937 | } |
938 | |
939 | // Check if an epilog exists as a subset of the end of a prolog (backwards). |
940 | static int |
941 | getARM64OffsetInProlog(const std::vector<WinEH::Instruction> &Prolog, |
942 | const std::vector<WinEH::Instruction> &Epilog) { |
943 | // Can't find an epilog as a subset if it is longer than the prolog. |
944 | if (Epilog.size() > Prolog.size()) |
945 | return -1; |
946 | |
947 | // Check that the epilog actually is a perfect match for the end (backwrds) |
948 | // of the prolog. |
949 | for (int I = Epilog.size() - 1; I >= 0; I--) { |
950 | if (Prolog[I] != Epilog[Epilog.size() - 1 - I]) |
951 | return -1; |
952 | } |
953 | |
954 | if (Epilog.size() == Prolog.size()) |
955 | return 0; |
956 | |
957 | // If the epilog was a subset of the prolog, find its offset. |
958 | return ARM64CountOfUnwindCodes(Insns: ArrayRef<WinEH::Instruction>( |
959 | &Prolog[Epilog.size()], Prolog.size() - Epilog.size())); |
960 | } |
961 | |
962 | static int checkARM64PackedEpilog(MCStreamer &streamer, WinEH::FrameInfo *info, |
963 | WinEH::FrameInfo::Segment *Seg, |
964 | int PrologCodeBytes) { |
965 | // Can only pack if there's one single epilog |
966 | if (Seg->Epilogs.size() != 1) |
967 | return -1; |
968 | |
969 | MCSymbol *Sym = Seg->Epilogs.begin()->first; |
970 | const std::vector<WinEH::Instruction> &Epilog = |
971 | info->EpilogMap[Sym].Instructions; |
972 | |
973 | // Check that the epilog actually is at the very end of the function, |
974 | // otherwise it can't be packed. |
975 | uint32_t DistanceFromEnd = |
976 | (uint32_t)(Seg->Offset + Seg->Length - Seg->Epilogs.begin()->second); |
977 | if (DistanceFromEnd / 4 != Epilog.size()) |
978 | return -1; |
979 | |
980 | int RetVal = -1; |
981 | // Even if we don't end up sharing opcodes with the prolog, we can still |
982 | // write the offset as a packed offset, if the single epilog is located at |
983 | // the end of the function and the offset (pointing after the prolog) fits |
984 | // as a packed offset. |
985 | if (PrologCodeBytes <= 31 && |
986 | PrologCodeBytes + ARM64CountOfUnwindCodes(Insns: Epilog) <= 124) |
987 | RetVal = PrologCodeBytes; |
988 | |
989 | int Offset = getARM64OffsetInProlog(Prolog: info->Instructions, Epilog); |
990 | if (Offset < 0) |
991 | return RetVal; |
992 | |
993 | // Check that the offset and prolog size fits in the first word; it's |
994 | // unclear whether the epilog count in the extension word can be taken |
995 | // as packed epilog offset. |
996 | if (Offset > 31 || PrologCodeBytes > 124) |
997 | return RetVal; |
998 | |
999 | // As we choose to express the epilog as part of the prolog, remove the |
1000 | // epilog from the map, so we don't try to emit its opcodes. |
1001 | info->EpilogMap.erase(Key: Sym); |
1002 | return Offset; |
1003 | } |
1004 | |
1005 | static bool tryARM64PackedUnwind(WinEH::FrameInfo *info, uint32_t FuncLength, |
1006 | int PackedEpilogOffset) { |
1007 | if (PackedEpilogOffset == 0) { |
1008 | // Fully symmetric prolog and epilog, should be ok for packed format. |
1009 | // For CR=3, the corresponding synthesized epilog actually lacks the |
1010 | // SetFP opcode, but unwinding should work just fine despite that |
1011 | // (if at the SetFP opcode, the unwinder considers it as part of the |
1012 | // function body and just unwinds the full prolog instead). |
1013 | } else if (PackedEpilogOffset == 1) { |
1014 | // One single case of differences between prolog and epilog is allowed: |
1015 | // The epilog can lack a single SetFP that is the last opcode in the |
1016 | // prolog, for the CR=3 case. |
1017 | if (info->Instructions.back().Operation != Win64EH::UOP_SetFP) |
1018 | return false; |
1019 | } else { |
1020 | // Too much difference between prolog and epilog. |
1021 | return false; |
1022 | } |
1023 | unsigned RegI = 0, RegF = 0; |
1024 | int Predecrement = 0; |
1025 | enum { |
1026 | Start, |
1027 | Start2, |
1028 | Start3, |
1029 | IntRegs, |
1030 | FloatRegs, |
1031 | InputArgs, |
1032 | StackAdjust, |
1033 | FrameRecord, |
1034 | End |
1035 | } Location = Start; |
1036 | bool StandaloneLR = false, FPLRPair = false; |
1037 | bool PAC = false; |
1038 | int StackOffset = 0; |
1039 | int Nops = 0; |
1040 | // Iterate over the prolog and check that all opcodes exactly match |
1041 | // the canonical order and form. A more lax check could verify that |
1042 | // all saved registers are in the expected locations, but not enforce |
1043 | // the order - that would work fine when unwinding from within |
1044 | // functions, but not be exactly right if unwinding happens within |
1045 | // prologs/epilogs. |
1046 | for (const WinEH::Instruction &Inst : info->Instructions) { |
1047 | switch (Inst.Operation) { |
1048 | case Win64EH::UOP_End: |
1049 | if (Location != Start) |
1050 | return false; |
1051 | Location = Start2; |
1052 | break; |
1053 | case Win64EH::UOP_PACSignLR: |
1054 | if (Location != Start2) |
1055 | return false; |
1056 | PAC = true; |
1057 | Location = Start3; |
1058 | break; |
1059 | case Win64EH::UOP_SaveR19R20X: |
1060 | if (Location != Start2 && Location != Start3) |
1061 | return false; |
1062 | Predecrement = Inst.Offset; |
1063 | RegI = 2; |
1064 | Location = IntRegs; |
1065 | break; |
1066 | case Win64EH::UOP_SaveRegX: |
1067 | if (Location != Start2 && Location != Start3) |
1068 | return false; |
1069 | Predecrement = Inst.Offset; |
1070 | if (Inst.Register == 19) |
1071 | RegI += 1; |
1072 | else if (Inst.Register == 30) |
1073 | StandaloneLR = true; |
1074 | else |
1075 | return false; |
1076 | // Odd register; can't be any further int registers. |
1077 | Location = FloatRegs; |
1078 | break; |
1079 | case Win64EH::UOP_SaveRegPX: |
1080 | // Can't have this in a canonical prologue. Either this has been |
1081 | // canonicalized into SaveR19R20X or SaveFPLRX, or it's an unsupported |
1082 | // register pair. |
1083 | // It can't be canonicalized into SaveR19R20X if the offset is |
1084 | // larger than 248 bytes, but even with the maximum case with |
1085 | // RegI=10/RegF=8/CR=1/H=1, we end up with SavSZ = 216, which should |
1086 | // fit into SaveR19R20X. |
1087 | // The unwinding opcodes can't describe the otherwise seemingly valid |
1088 | // case for RegI=1 CR=1, that would start with a |
1089 | // "stp x19, lr, [sp, #-...]!" as that fits neither SaveRegPX nor |
1090 | // SaveLRPair. |
1091 | return false; |
1092 | case Win64EH::UOP_SaveRegP: |
1093 | if (Location != IntRegs || Inst.Offset != 8 * RegI || |
1094 | Inst.Register != 19 + RegI) |
1095 | return false; |
1096 | RegI += 2; |
1097 | break; |
1098 | case Win64EH::UOP_SaveReg: |
1099 | if (Location != IntRegs || Inst.Offset != 8 * RegI) |
1100 | return false; |
1101 | if (Inst.Register == 19 + RegI) |
1102 | RegI += 1; |
1103 | else if (Inst.Register == 30) |
1104 | StandaloneLR = true; |
1105 | else |
1106 | return false; |
1107 | // Odd register; can't be any further int registers. |
1108 | Location = FloatRegs; |
1109 | break; |
1110 | case Win64EH::UOP_SaveLRPair: |
1111 | if (Location != IntRegs || Inst.Offset != 8 * RegI || |
1112 | Inst.Register != 19 + RegI) |
1113 | return false; |
1114 | RegI += 1; |
1115 | StandaloneLR = true; |
1116 | Location = FloatRegs; |
1117 | break; |
1118 | case Win64EH::UOP_SaveFRegX: |
1119 | // Packed unwind can't handle prologs that only save one single |
1120 | // float register. |
1121 | return false; |
1122 | case Win64EH::UOP_SaveFReg: |
1123 | if (Location != FloatRegs || RegF == 0 || Inst.Register != 8 + RegF || |
1124 | Inst.Offset != 8 * (RegI + (StandaloneLR ? 1 : 0) + RegF)) |
1125 | return false; |
1126 | RegF += 1; |
1127 | Location = InputArgs; |
1128 | break; |
1129 | case Win64EH::UOP_SaveFRegPX: |
1130 | if ((Location != Start2 && Location != Start3) || Inst.Register != 8) |
1131 | return false; |
1132 | Predecrement = Inst.Offset; |
1133 | RegF = 2; |
1134 | Location = FloatRegs; |
1135 | break; |
1136 | case Win64EH::UOP_SaveFRegP: |
1137 | if ((Location != IntRegs && Location != FloatRegs) || |
1138 | Inst.Register != 8 + RegF || |
1139 | Inst.Offset != 8 * (RegI + (StandaloneLR ? 1 : 0) + RegF)) |
1140 | return false; |
1141 | RegF += 2; |
1142 | Location = FloatRegs; |
1143 | break; |
1144 | case Win64EH::UOP_SaveNext: |
1145 | if (Location == IntRegs) |
1146 | RegI += 2; |
1147 | else if (Location == FloatRegs) |
1148 | RegF += 2; |
1149 | else |
1150 | return false; |
1151 | break; |
1152 | case Win64EH::UOP_Nop: |
1153 | if (Location != IntRegs && Location != FloatRegs && Location != InputArgs) |
1154 | return false; |
1155 | Location = InputArgs; |
1156 | Nops++; |
1157 | break; |
1158 | case Win64EH::UOP_AllocSmall: |
1159 | case Win64EH::UOP_AllocMedium: |
1160 | if (Location != Start2 && Location != Start3 && Location != IntRegs && |
1161 | Location != FloatRegs && Location != InputArgs && |
1162 | Location != StackAdjust) |
1163 | return false; |
1164 | // Can have either a single decrement, or a pair of decrements with |
1165 | // 4080 and another decrement. |
1166 | if (StackOffset == 0) |
1167 | StackOffset = Inst.Offset; |
1168 | else if (StackOffset != 4080) |
1169 | return false; |
1170 | else |
1171 | StackOffset += Inst.Offset; |
1172 | Location = StackAdjust; |
1173 | break; |
1174 | case Win64EH::UOP_SaveFPLRX: |
1175 | // Not allowing FPLRX after StackAdjust; if a StackAdjust is used, it |
1176 | // should be followed by a FPLR instead. |
1177 | if (Location != Start2 && Location != Start3 && Location != IntRegs && |
1178 | Location != FloatRegs && Location != InputArgs) |
1179 | return false; |
1180 | StackOffset = Inst.Offset; |
1181 | Location = FrameRecord; |
1182 | FPLRPair = true; |
1183 | break; |
1184 | case Win64EH::UOP_SaveFPLR: |
1185 | // This can only follow after a StackAdjust |
1186 | if (Location != StackAdjust || Inst.Offset != 0) |
1187 | return false; |
1188 | Location = FrameRecord; |
1189 | FPLRPair = true; |
1190 | break; |
1191 | case Win64EH::UOP_SetFP: |
1192 | if (Location != FrameRecord) |
1193 | return false; |
1194 | Location = End; |
1195 | break; |
1196 | case Win64EH::UOP_SaveAnyRegI: |
1197 | case Win64EH::UOP_SaveAnyRegIP: |
1198 | case Win64EH::UOP_SaveAnyRegD: |
1199 | case Win64EH::UOP_SaveAnyRegDP: |
1200 | case Win64EH::UOP_SaveAnyRegQ: |
1201 | case Win64EH::UOP_SaveAnyRegQP: |
1202 | case Win64EH::UOP_SaveAnyRegIX: |
1203 | case Win64EH::UOP_SaveAnyRegIPX: |
1204 | case Win64EH::UOP_SaveAnyRegDX: |
1205 | case Win64EH::UOP_SaveAnyRegDPX: |
1206 | case Win64EH::UOP_SaveAnyRegQX: |
1207 | case Win64EH::UOP_SaveAnyRegQPX: |
1208 | // These are never canonical; they don't show up with the usual Arm64 |
1209 | // calling convention. |
1210 | return false; |
1211 | case Win64EH::UOP_AllocLarge: |
1212 | // Allocations this large can't be represented in packed unwind (and |
1213 | // usually don't fit the canonical form anyway because we need to use |
1214 | // __chkstk to allocate the stack space). |
1215 | return false; |
1216 | case Win64EH::UOP_AddFP: |
1217 | // "add x29, sp, #N" doesn't show up in the canonical pattern (except for |
1218 | // N=0, which is UOP_SetFP). |
1219 | return false; |
1220 | case Win64EH::UOP_AllocZ: |
1221 | case Win64EH::UOP_SaveZReg: |
1222 | case Win64EH::UOP_SavePReg: |
1223 | // Canonical prologues don't support spilling SVE registers. |
1224 | return false; |
1225 | case Win64EH::UOP_TrapFrame: |
1226 | case Win64EH::UOP_Context: |
1227 | case Win64EH::UOP_ECContext: |
1228 | case Win64EH::UOP_ClearUnwoundToCall: |
1229 | case Win64EH::UOP_PushMachFrame: |
1230 | // These are special opcodes that aren't normally generated. |
1231 | return false; |
1232 | default: |
1233 | report_fatal_error(reason: "Unknown Arm64 unwind opcode" ); |
1234 | } |
1235 | } |
1236 | if (RegI > 10 || RegF > 8) |
1237 | return false; |
1238 | if (StandaloneLR && FPLRPair) |
1239 | return false; |
1240 | if (FPLRPair && Location != End) |
1241 | return false; |
1242 | if (Nops != 0 && Nops != 4) |
1243 | return false; |
1244 | if (PAC && !FPLRPair) |
1245 | return false; |
1246 | int H = Nops == 4; |
1247 | // There's an inconsistency regarding packed unwind info with homed |
1248 | // parameters; according to the documentation, the epilog shouldn't have |
1249 | // the same corresponding nops (and thus, to set the H bit, we should |
1250 | // require an epilog which isn't exactly symmetrical - we shouldn't accept |
1251 | // an exact mirrored epilog for those cases), but in practice, |
1252 | // RtlVirtualUnwind behaves as if it does expect the epilogue to contain |
1253 | // the same nops. See https://github.com/llvm/llvm-project/issues/54879. |
1254 | // To play it safe, don't produce packed unwind info with homed parameters. |
1255 | if (H) |
1256 | return false; |
1257 | int IntSZ = 8 * RegI; |
1258 | if (StandaloneLR) |
1259 | IntSZ += 8; |
1260 | int FpSZ = 8 * RegF; // RegF not yet decremented |
1261 | int SavSZ = (IntSZ + FpSZ + 8 * 8 * H + 0xF) & ~0xF; |
1262 | if (Predecrement != SavSZ) |
1263 | return false; |
1264 | if (FPLRPair && StackOffset < 16) |
1265 | return false; |
1266 | if (StackOffset % 16) |
1267 | return false; |
1268 | uint32_t FrameSize = (StackOffset + SavSZ) / 16; |
1269 | if (FrameSize > 0x1FF) |
1270 | return false; |
1271 | assert(RegF != 1 && "One single float reg not allowed" ); |
1272 | if (RegF > 0) |
1273 | RegF--; // Convert from actual number of registers, to value stored |
1274 | assert(FuncLength <= 0x7FF && "FuncLength should have been checked earlier" ); |
1275 | int Flag = 0x01; // Function segments not supported yet |
1276 | int CR = PAC ? 2 : FPLRPair ? 3 : StandaloneLR ? 1 : 0; |
1277 | info->PackedInfo |= Flag << 0; |
1278 | info->PackedInfo |= (FuncLength & 0x7FF) << 2; |
1279 | info->PackedInfo |= (RegF & 0x7) << 13; |
1280 | info->PackedInfo |= (RegI & 0xF) << 16; |
1281 | info->PackedInfo |= (H & 0x1) << 20; |
1282 | info->PackedInfo |= (CR & 0x3) << 21; |
1283 | info->PackedInfo |= (FrameSize & 0x1FF) << 23; |
1284 | return true; |
1285 | } |
1286 | |
1287 | static void ARM64ProcessEpilogs(WinEH::FrameInfo *info, |
1288 | WinEH::FrameInfo::Segment *Seg, |
1289 | uint32_t &TotalCodeBytes, |
1290 | MapVector<MCSymbol *, uint32_t> &EpilogInfo) { |
1291 | |
1292 | std::vector<MCSymbol *> EpilogStarts; |
1293 | for (auto &I : Seg->Epilogs) |
1294 | EpilogStarts.push_back(x: I.first); |
1295 | |
1296 | // Epilogs processed so far. |
1297 | std::vector<MCSymbol *> AddedEpilogs; |
1298 | for (auto *S : EpilogStarts) { |
1299 | MCSymbol *EpilogStart = S; |
1300 | auto &EpilogInstrs = info->EpilogMap[S].Instructions; |
1301 | uint32_t CodeBytes = ARM64CountOfUnwindCodes(Insns: EpilogInstrs); |
1302 | |
1303 | MCSymbol* MatchingEpilog = |
1304 | FindMatchingEpilog(EpilogInstrs, Epilogs: AddedEpilogs, info); |
1305 | int PrologOffset; |
1306 | if (MatchingEpilog) { |
1307 | assert(EpilogInfo.contains(MatchingEpilog) && |
1308 | "Duplicate epilog not found" ); |
1309 | EpilogInfo[EpilogStart] = EpilogInfo.lookup(Key: MatchingEpilog); |
1310 | // Clear the unwind codes in the EpilogMap, so that they don't get output |
1311 | // in ARM64EmitUnwindInfoForSegment(). |
1312 | EpilogInstrs.clear(); |
1313 | } else if ((PrologOffset = getARM64OffsetInProlog(Prolog: info->Instructions, |
1314 | Epilog: EpilogInstrs)) >= 0) { |
1315 | EpilogInfo[EpilogStart] = PrologOffset; |
1316 | // If the segment doesn't have a prolog, an end_c will be emitted before |
1317 | // prolog opcodes. So epilog start index in opcodes array is moved by 1. |
1318 | if (!Seg->HasProlog) |
1319 | EpilogInfo[EpilogStart] += 1; |
1320 | // Clear the unwind codes in the EpilogMap, so that they don't get output |
1321 | // in ARM64EmitUnwindInfoForSegment(). |
1322 | EpilogInstrs.clear(); |
1323 | } else { |
1324 | EpilogInfo[EpilogStart] = TotalCodeBytes; |
1325 | TotalCodeBytes += CodeBytes; |
1326 | AddedEpilogs.push_back(x: EpilogStart); |
1327 | } |
1328 | } |
1329 | } |
1330 | |
1331 | static void ARM64FindSegmentsInFunction(MCStreamer &streamer, |
1332 | WinEH::FrameInfo *info, |
1333 | int64_t RawFuncLength) { |
1334 | if (info->PrologEnd) |
1335 | checkARM64Instructions(Streamer&: streamer, Insns: info->Instructions, Begin: info->Begin, |
1336 | End: info->PrologEnd, Name: info->Function->getName(), |
1337 | Type: "prologue" ); |
1338 | struct EpilogStartEnd { |
1339 | MCSymbol *Start; |
1340 | int64_t Offset; |
1341 | int64_t End; |
1342 | }; |
1343 | // Record Start and End of each epilog. |
1344 | SmallVector<struct EpilogStartEnd, 4> Epilogs; |
1345 | for (auto &I : info->EpilogMap) { |
1346 | MCSymbol *Start = I.first; |
1347 | auto &Instrs = I.second.Instructions; |
1348 | int64_t Offset = GetAbsDifference(Streamer&: streamer, LHS: Start, RHS: info->Begin); |
1349 | checkARM64Instructions(Streamer&: streamer, Insns: Instrs, Begin: Start, End: I.second.End, |
1350 | Name: info->Function->getName(), Type: "epilogue" ); |
1351 | assert((Epilogs.size() == 0 || Offset >= Epilogs.back().End) && |
1352 | "Epilogs should be monotonically ordered" ); |
1353 | // Exclue the end opcode from Instrs.size() when calculating the end of the |
1354 | // epilog. |
1355 | Epilogs.push_back(Elt: {.Start: Start, .Offset: Offset, .End: Offset + (int64_t)(Instrs.size() - 1) * 4}); |
1356 | } |
1357 | |
1358 | unsigned E = 0; |
1359 | int64_t SegLimit = 0xFFFFC; |
1360 | int64_t SegOffset = 0; |
1361 | |
1362 | if (RawFuncLength > SegLimit) { |
1363 | |
1364 | int64_t RemainingLength = RawFuncLength; |
1365 | |
1366 | while (RemainingLength > SegLimit) { |
1367 | // Try divide the function into segments, requirements: |
1368 | // 1. Segment length <= 0xFFFFC; |
1369 | // 2. Each Prologue or Epilogue must be fully within a segment. |
1370 | int64_t SegLength = SegLimit; |
1371 | int64_t SegEnd = SegOffset + SegLength; |
1372 | // Keep record on symbols and offsets of epilogs in this segment. |
1373 | MapVector<MCSymbol *, int64_t> EpilogsInSegment; |
1374 | |
1375 | while (E < Epilogs.size() && Epilogs[E].End < SegEnd) { |
1376 | // Epilogs within current segment. |
1377 | EpilogsInSegment[Epilogs[E].Start] = Epilogs[E].Offset; |
1378 | ++E; |
1379 | } |
1380 | |
1381 | // At this point, we have: |
1382 | // 1. Put all epilogs in segments already. No action needed here; or |
1383 | // 2. Found an epilog that will cross segments boundry. We need to |
1384 | // move back current segment's end boundry, so the epilog is entirely |
1385 | // in the next segment; or |
1386 | // 3. Left at least one epilog that is entirely after this segment. |
1387 | // It'll be handled by the next iteration, or the last segment. |
1388 | if (E < Epilogs.size() && Epilogs[E].Offset <= SegEnd) |
1389 | // Move back current Segment's end boundry. |
1390 | SegLength = Epilogs[E].Offset - SegOffset; |
1391 | |
1392 | auto Seg = WinEH::FrameInfo::Segment( |
1393 | SegOffset, SegLength, /* HasProlog */!SegOffset); |
1394 | Seg.Epilogs = std::move(EpilogsInSegment); |
1395 | info->Segments.push_back(x: Seg); |
1396 | |
1397 | SegOffset += SegLength; |
1398 | RemainingLength -= SegLength; |
1399 | } |
1400 | } |
1401 | |
1402 | // Add the last segment when RawFuncLength > 0xFFFFC, |
1403 | // or the only segment otherwise. |
1404 | auto LastSeg = |
1405 | WinEH::FrameInfo::Segment(SegOffset, RawFuncLength - SegOffset, |
1406 | /* HasProlog */!SegOffset); |
1407 | for (; E < Epilogs.size(); ++E) |
1408 | LastSeg.Epilogs[Epilogs[E].Start] = Epilogs[E].Offset; |
1409 | info->Segments.push_back(x: LastSeg); |
1410 | } |
1411 | |
1412 | static void ARM64EmitUnwindInfoForSegment(MCStreamer &streamer, |
1413 | WinEH::FrameInfo *info, |
1414 | WinEH::FrameInfo::Segment &Seg, |
1415 | bool TryPacked = true) { |
1416 | MCContext &context = streamer.getContext(); |
1417 | MCSymbol *Label = context.createTempSymbol(); |
1418 | |
1419 | streamer.emitValueToAlignment(Alignment: Align(4)); |
1420 | streamer.emitLabel(Symbol: Label); |
1421 | Seg.Symbol = Label; |
1422 | // Use the 1st segemnt's label as function's. |
1423 | if (Seg.Offset == 0) |
1424 | info->Symbol = Label; |
1425 | |
1426 | bool HasProlog = Seg.HasProlog; |
1427 | bool HasEpilogs = (Seg.Epilogs.size() != 0); |
1428 | |
1429 | uint32_t SegLength = (uint32_t)Seg.Length / 4; |
1430 | uint32_t PrologCodeBytes = info->PrologCodeBytes; |
1431 | |
1432 | int PackedEpilogOffset = HasEpilogs ? |
1433 | checkARM64PackedEpilog(streamer, info, Seg: &Seg, PrologCodeBytes) : -1; |
1434 | |
1435 | // TODO: |
1436 | // 1. Enable packed unwind info (.pdata only) for multi-segment functions. |
1437 | // 2. Emit packed unwind info (.pdata only) for segments that have neithor |
1438 | // prolog nor epilog. |
1439 | if (info->Segments.size() == 1 && PackedEpilogOffset >= 0 && |
1440 | uint32_t(PackedEpilogOffset) < PrologCodeBytes && |
1441 | !info->HandlesExceptions && SegLength <= 0x7ff && TryPacked) { |
1442 | // Matching prolog/epilog and no exception handlers; check if the |
1443 | // prolog matches the patterns that can be described by the packed |
1444 | // format. |
1445 | |
1446 | // info->Symbol was already set even if we didn't actually write any |
1447 | // unwind info there. Keep using that as indicator that this unwind |
1448 | // info has been generated already. |
1449 | if (tryARM64PackedUnwind(info, FuncLength: SegLength, PackedEpilogOffset)) |
1450 | return; |
1451 | } |
1452 | |
1453 | // If the prolog is not in this segment, we need to emit an end_c, which takes |
1454 | // 1 byte, before prolog unwind ops. |
1455 | if (!HasProlog) { |
1456 | PrologCodeBytes += 1; |
1457 | if (PackedEpilogOffset >= 0) |
1458 | PackedEpilogOffset += 1; |
1459 | // If a segment has neither prolog nor epilog, "With full .xdata record, |
1460 | // Epilog Count = 1. Epilog Start Index points to end_c." |
1461 | // https://docs.microsoft.com/en-us/cpp/build/arm64-exception-handling#function-fragments |
1462 | // TODO: We can remove this if testing shows zero epilog scope is ok with |
1463 | // MS unwinder. |
1464 | if (!HasEpilogs) |
1465 | // Pack the fake epilog into phantom prolog. |
1466 | PackedEpilogOffset = 0; |
1467 | } |
1468 | |
1469 | uint32_t TotalCodeBytes = PrologCodeBytes; |
1470 | |
1471 | // Process epilogs. |
1472 | MapVector<MCSymbol *, uint32_t> EpilogInfo; |
1473 | ARM64ProcessEpilogs(info, Seg: &Seg, TotalCodeBytes, EpilogInfo); |
1474 | |
1475 | // Code Words, Epilog count, E, X, Vers, Function Length |
1476 | uint32_t row1 = 0x0; |
1477 | uint32_t CodeWords = TotalCodeBytes / 4; |
1478 | uint32_t CodeWordsMod = TotalCodeBytes % 4; |
1479 | if (CodeWordsMod) |
1480 | CodeWords++; |
1481 | uint32_t EpilogCount = |
1482 | PackedEpilogOffset >= 0 ? PackedEpilogOffset : Seg.Epilogs.size(); |
1483 | bool ExtensionWord = EpilogCount > 31 || TotalCodeBytes > 124; |
1484 | if (!ExtensionWord) { |
1485 | row1 |= (EpilogCount & 0x1F) << 22; |
1486 | row1 |= (CodeWords & 0x1F) << 27; |
1487 | } |
1488 | if (info->HandlesExceptions) // X |
1489 | row1 |= 1 << 20; |
1490 | if (PackedEpilogOffset >= 0) // E |
1491 | row1 |= 1 << 21; |
1492 | row1 |= SegLength & 0x3FFFF; |
1493 | streamer.emitInt32(Value: row1); |
1494 | |
1495 | // Extended Code Words, Extended Epilog Count |
1496 | if (ExtensionWord) { |
1497 | // FIXME: We should be able to split unwind info into multiple sections. |
1498 | if (CodeWords > 0xFF || EpilogCount > 0xFFFF) |
1499 | report_fatal_error( |
1500 | reason: "SEH unwind data splitting is only implemented for large functions, " |
1501 | "cases of too many code words or too many epilogs will be done " |
1502 | "later" ); |
1503 | uint32_t row2 = 0x0; |
1504 | row2 |= (CodeWords & 0xFF) << 16; |
1505 | row2 |= (EpilogCount & 0xFFFF); |
1506 | streamer.emitInt32(Value: row2); |
1507 | } |
1508 | |
1509 | if (PackedEpilogOffset < 0) { |
1510 | // Epilog Start Index, Epilog Start Offset |
1511 | for (auto &I : EpilogInfo) { |
1512 | MCSymbol *EpilogStart = I.first; |
1513 | uint32_t EpilogIndex = I.second; |
1514 | // Epilog offset within the Segment. |
1515 | uint32_t EpilogOffset = (uint32_t)(Seg.Epilogs[EpilogStart] - Seg.Offset); |
1516 | if (EpilogOffset) |
1517 | EpilogOffset /= 4; |
1518 | uint32_t row3 = EpilogOffset; |
1519 | row3 |= (EpilogIndex & 0x3FF) << 22; |
1520 | streamer.emitInt32(Value: row3); |
1521 | } |
1522 | } |
1523 | |
1524 | // Note that even for segments that have no prolog, we still need to emit |
1525 | // prolog unwinding opcodes so that the unwinder knows how to unwind from |
1526 | // such a segment. |
1527 | // The end_c opcode at the start indicates to the unwinder that the actual |
1528 | // prolog is outside of the current segment, and the unwinder shouldn't try |
1529 | // to check for unwinding from a partial prolog. |
1530 | if (!HasProlog) |
1531 | // Emit an end_c. |
1532 | streamer.emitInt8(Value: (uint8_t)0xE5); |
1533 | |
1534 | // Emit prolog unwind instructions (in reverse order). |
1535 | for (auto Inst : llvm::reverse(C&: info->Instructions)) |
1536 | ARM64EmitUnwindCode(streamer, inst: Inst); |
1537 | |
1538 | // Emit epilog unwind instructions |
1539 | for (auto &I : Seg.Epilogs) { |
1540 | auto &EpilogInstrs = info->EpilogMap[I.first].Instructions; |
1541 | for (const WinEH::Instruction &inst : EpilogInstrs) |
1542 | ARM64EmitUnwindCode(streamer, inst); |
1543 | } |
1544 | |
1545 | int32_t BytesMod = CodeWords * 4 - TotalCodeBytes; |
1546 | assert(BytesMod >= 0); |
1547 | for (int i = 0; i < BytesMod; i++) |
1548 | streamer.emitInt8(Value: 0xE3); |
1549 | |
1550 | if (info->HandlesExceptions) |
1551 | streamer.emitValue( |
1552 | Value: MCSymbolRefExpr::create(Symbol: info->ExceptionHandler, |
1553 | specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, Ctx&: context), |
1554 | Size: 4); |
1555 | } |
1556 | |
1557 | // Populate the .xdata section. The format of .xdata on ARM64 is documented at |
1558 | // https://docs.microsoft.com/en-us/cpp/build/arm64-exception-handling |
1559 | static void ARM64EmitUnwindInfo(MCStreamer &streamer, WinEH::FrameInfo *info, |
1560 | bool TryPacked = true) { |
1561 | // If this UNWIND_INFO already has a symbol, it's already been emitted. |
1562 | if (info->Symbol) |
1563 | return; |
1564 | // If there's no unwind info here (not even a terminating UOP_End), the |
1565 | // unwind info is considered bogus and skipped. If this was done in |
1566 | // response to an explicit .seh_handlerdata, the associated trailing |
1567 | // handler data is left orphaned in the xdata section. |
1568 | if (info->empty()) { |
1569 | info->EmitAttempted = true; |
1570 | return; |
1571 | } |
1572 | if (info->EmitAttempted) { |
1573 | // If we tried to emit unwind info before (due to an explicit |
1574 | // .seh_handlerdata directive), but skipped it (because there was no |
1575 | // valid information to emit at the time), and it later got valid unwind |
1576 | // opcodes, we can't emit it here, because the trailing handler data |
1577 | // was already emitted elsewhere in the xdata section. |
1578 | streamer.getContext().reportError( |
1579 | L: SMLoc(), Msg: "Earlier .seh_handlerdata for " + info->Function->getName() + |
1580 | " skipped due to no unwind info at the time " |
1581 | "(.seh_handlerdata too early?), but the function later " |
1582 | "did get unwind info that can't be emitted" ); |
1583 | return; |
1584 | } |
1585 | |
1586 | simplifyARM64Opcodes(Instructions&: info->Instructions, Reverse: false); |
1587 | for (auto &I : info->EpilogMap) |
1588 | simplifyARM64Opcodes(Instructions&: I.second.Instructions, Reverse: true); |
1589 | |
1590 | int64_t RawFuncLength; |
1591 | if (!info->FuncletOrFuncEnd) { |
1592 | report_fatal_error(reason: "FuncletOrFuncEnd not set" ); |
1593 | } else { |
1594 | // FIXME: GetAbsDifference tries to compute the length of the function |
1595 | // immediately, before the whole file is emitted, but in general |
1596 | // that's impossible: the size in bytes of certain assembler directives |
1597 | // like .align and .fill is not known until the whole file is parsed and |
1598 | // relaxations are applied. Currently, GetAbsDifference fails with a fatal |
1599 | // error in that case. (We mostly don't hit this because inline assembly |
1600 | // specifying those directives is rare, and we don't normally try to |
1601 | // align loops on AArch64.) |
1602 | // |
1603 | // There are two potential approaches to delaying the computation. One, |
1604 | // we could emit something like ".word (endfunc-beginfunc)/4+0x10800000", |
1605 | // as long as we have some conservative estimate we could use to prove |
1606 | // that we don't need to split the unwind data. Emitting the constant |
1607 | // is straightforward, but there's no existing code for estimating the |
1608 | // size of the function. |
1609 | // |
1610 | // The other approach would be to use a dedicated, relaxable fragment, |
1611 | // which could grow to accommodate splitting the unwind data if |
1612 | // necessary. This is more straightforward, since it automatically works |
1613 | // without any new infrastructure, and it's consistent with how we handle |
1614 | // relaxation in other contexts. But it would require some refactoring |
1615 | // to move parts of the pdata/xdata emission into the implementation of |
1616 | // a fragment. We could probably continue to encode the unwind codes |
1617 | // here, but we'd have to emit the pdata, the xdata header, and the |
1618 | // epilogue scopes later, since they depend on whether the we need to |
1619 | // split the unwind data. |
1620 | // |
1621 | // If this is fixed, remove code in AArch64ISelLowering.cpp that |
1622 | // disables loop alignment on Windows. |
1623 | RawFuncLength = GetAbsDifference(Streamer&: streamer, LHS: info->FuncletOrFuncEnd, |
1624 | RHS: info->Begin); |
1625 | } |
1626 | |
1627 | ARM64FindSegmentsInFunction(streamer, info, RawFuncLength); |
1628 | |
1629 | info->PrologCodeBytes = ARM64CountOfUnwindCodes(Insns: info->Instructions); |
1630 | for (auto &S : info->Segments) |
1631 | ARM64EmitUnwindInfoForSegment(streamer, info, Seg&: S, TryPacked); |
1632 | |
1633 | // Clear prolog instructions after unwind info is emitted for all segments. |
1634 | info->Instructions.clear(); |
1635 | } |
1636 | |
1637 | static uint32_t ARMCountOfUnwindCodes(ArrayRef<WinEH::Instruction> Insns) { |
1638 | uint32_t Count = 0; |
1639 | for (const auto &I : Insns) { |
1640 | switch (static_cast<Win64EH::UnwindOpcodes>(I.Operation)) { |
1641 | default: |
1642 | llvm_unreachable("Unsupported ARM unwind code" ); |
1643 | case Win64EH::UOP_AllocSmall: |
1644 | Count += 1; |
1645 | break; |
1646 | case Win64EH::UOP_AllocLarge: |
1647 | Count += 3; |
1648 | break; |
1649 | case Win64EH::UOP_AllocHuge: |
1650 | Count += 4; |
1651 | break; |
1652 | case Win64EH::UOP_WideAllocMedium: |
1653 | Count += 2; |
1654 | break; |
1655 | case Win64EH::UOP_WideAllocLarge: |
1656 | Count += 3; |
1657 | break; |
1658 | case Win64EH::UOP_WideAllocHuge: |
1659 | Count += 4; |
1660 | break; |
1661 | case Win64EH::UOP_WideSaveRegMask: |
1662 | Count += 2; |
1663 | break; |
1664 | case Win64EH::UOP_SaveSP: |
1665 | Count += 1; |
1666 | break; |
1667 | case Win64EH::UOP_SaveRegsR4R7LR: |
1668 | Count += 1; |
1669 | break; |
1670 | case Win64EH::UOP_WideSaveRegsR4R11LR: |
1671 | Count += 1; |
1672 | break; |
1673 | case Win64EH::UOP_SaveFRegD8D15: |
1674 | Count += 1; |
1675 | break; |
1676 | case Win64EH::UOP_SaveRegMask: |
1677 | Count += 2; |
1678 | break; |
1679 | case Win64EH::UOP_SaveLR: |
1680 | Count += 2; |
1681 | break; |
1682 | case Win64EH::UOP_SaveFRegD0D15: |
1683 | Count += 2; |
1684 | break; |
1685 | case Win64EH::UOP_SaveFRegD16D31: |
1686 | Count += 2; |
1687 | break; |
1688 | case Win64EH::UOP_Nop: |
1689 | case Win64EH::UOP_WideNop: |
1690 | case Win64EH::UOP_End: |
1691 | case Win64EH::UOP_EndNop: |
1692 | case Win64EH::UOP_WideEndNop: |
1693 | Count += 1; |
1694 | break; |
1695 | case Win64EH::UOP_Custom: { |
1696 | int J; |
1697 | for (J = 3; J > 0; J--) |
1698 | if (I.Offset & (0xffu << (8 * J))) |
1699 | break; |
1700 | Count += J + 1; |
1701 | break; |
1702 | } |
1703 | } |
1704 | } |
1705 | return Count; |
1706 | } |
1707 | |
1708 | static uint32_t ARMCountOfInstructionBytes(ArrayRef<WinEH::Instruction> Insns, |
1709 | bool *HasCustom = nullptr) { |
1710 | uint32_t Count = 0; |
1711 | for (const auto &I : Insns) { |
1712 | switch (static_cast<Win64EH::UnwindOpcodes>(I.Operation)) { |
1713 | default: |
1714 | llvm_unreachable("Unsupported ARM unwind code" ); |
1715 | case Win64EH::UOP_AllocSmall: |
1716 | case Win64EH::UOP_AllocLarge: |
1717 | case Win64EH::UOP_AllocHuge: |
1718 | Count += 2; |
1719 | break; |
1720 | case Win64EH::UOP_WideAllocMedium: |
1721 | case Win64EH::UOP_WideAllocLarge: |
1722 | case Win64EH::UOP_WideAllocHuge: |
1723 | Count += 4; |
1724 | break; |
1725 | case Win64EH::UOP_WideSaveRegMask: |
1726 | case Win64EH::UOP_WideSaveRegsR4R11LR: |
1727 | Count += 4; |
1728 | break; |
1729 | case Win64EH::UOP_SaveSP: |
1730 | Count += 2; |
1731 | break; |
1732 | case Win64EH::UOP_SaveRegMask: |
1733 | case Win64EH::UOP_SaveRegsR4R7LR: |
1734 | Count += 2; |
1735 | break; |
1736 | case Win64EH::UOP_SaveFRegD8D15: |
1737 | case Win64EH::UOP_SaveFRegD0D15: |
1738 | case Win64EH::UOP_SaveFRegD16D31: |
1739 | Count += 4; |
1740 | break; |
1741 | case Win64EH::UOP_SaveLR: |
1742 | Count += 4; |
1743 | break; |
1744 | case Win64EH::UOP_Nop: |
1745 | case Win64EH::UOP_EndNop: |
1746 | Count += 2; |
1747 | break; |
1748 | case Win64EH::UOP_WideNop: |
1749 | case Win64EH::UOP_WideEndNop: |
1750 | Count += 4; |
1751 | break; |
1752 | case Win64EH::UOP_End: |
1753 | // This doesn't map to any instruction |
1754 | break; |
1755 | case Win64EH::UOP_Custom: |
1756 | // We can't reason about what instructions this maps to; return a |
1757 | // phony number to make sure we don't accidentally do epilog packing. |
1758 | Count += 1000; |
1759 | if (HasCustom) |
1760 | *HasCustom = true; |
1761 | break; |
1762 | } |
1763 | } |
1764 | return Count; |
1765 | } |
1766 | |
1767 | static void checkARMInstructions(MCStreamer &Streamer, |
1768 | ArrayRef<WinEH::Instruction> Insns, |
1769 | const MCSymbol *Begin, const MCSymbol *End, |
1770 | StringRef Name, StringRef Type) { |
1771 | if (!End) |
1772 | return; |
1773 | std::optional<int64_t> MaybeDistance = |
1774 | GetOptionalAbsDifference(Streamer, LHS: End, RHS: Begin); |
1775 | if (!MaybeDistance) |
1776 | return; |
1777 | uint32_t Distance = (uint32_t)*MaybeDistance; |
1778 | bool HasCustom = false; |
1779 | uint32_t InstructionBytes = ARMCountOfInstructionBytes(Insns, HasCustom: &HasCustom); |
1780 | if (HasCustom) |
1781 | return; |
1782 | if (Distance != InstructionBytes) { |
1783 | Streamer.getContext().reportError( |
1784 | L: SMLoc(), Msg: "Incorrect size for " + Name + " " + Type + ": " + |
1785 | Twine(Distance) + |
1786 | " bytes of instructions in range, but .seh directives " |
1787 | "corresponding to " + |
1788 | Twine(InstructionBytes) + " bytes\n" ); |
1789 | } |
1790 | } |
1791 | |
1792 | static bool isARMTerminator(const WinEH::Instruction &inst) { |
1793 | switch (static_cast<Win64EH::UnwindOpcodes>(inst.Operation)) { |
1794 | case Win64EH::UOP_End: |
1795 | case Win64EH::UOP_EndNop: |
1796 | case Win64EH::UOP_WideEndNop: |
1797 | return true; |
1798 | default: |
1799 | return false; |
1800 | } |
1801 | } |
1802 | |
1803 | // Unwind opcode encodings and restrictions are documented at |
1804 | // https://docs.microsoft.com/en-us/cpp/build/arm-exception-handling |
1805 | static void ARMEmitUnwindCode(MCStreamer &streamer, |
1806 | const WinEH::Instruction &inst) { |
1807 | uint32_t w, lr; |
1808 | int i; |
1809 | switch (static_cast<Win64EH::UnwindOpcodes>(inst.Operation)) { |
1810 | default: |
1811 | llvm_unreachable("Unsupported ARM unwind code" ); |
1812 | case Win64EH::UOP_AllocSmall: |
1813 | assert((inst.Offset & 3) == 0); |
1814 | assert(inst.Offset / 4 <= 0x7f); |
1815 | streamer.emitInt8(Value: inst.Offset / 4); |
1816 | break; |
1817 | case Win64EH::UOP_WideSaveRegMask: |
1818 | assert((inst.Register & ~0x5fff) == 0); |
1819 | lr = (inst.Register >> 14) & 1; |
1820 | w = 0x8000 | (inst.Register & 0x1fff) | (lr << 13); |
1821 | streamer.emitInt8(Value: (w >> 8) & 0xff); |
1822 | streamer.emitInt8(Value: (w >> 0) & 0xff); |
1823 | break; |
1824 | case Win64EH::UOP_SaveSP: |
1825 | assert(inst.Register <= 0x0f); |
1826 | streamer.emitInt8(Value: 0xc0 | inst.Register); |
1827 | break; |
1828 | case Win64EH::UOP_SaveRegsR4R7LR: |
1829 | assert(inst.Register >= 4 && inst.Register <= 7); |
1830 | assert(inst.Offset <= 1); |
1831 | streamer.emitInt8(Value: 0xd0 | (inst.Register - 4) | (inst.Offset << 2)); |
1832 | break; |
1833 | case Win64EH::UOP_WideSaveRegsR4R11LR: |
1834 | assert(inst.Register >= 8 && inst.Register <= 11); |
1835 | assert(inst.Offset <= 1); |
1836 | streamer.emitInt8(Value: 0xd8 | (inst.Register - 8) | (inst.Offset << 2)); |
1837 | break; |
1838 | case Win64EH::UOP_SaveFRegD8D15: |
1839 | assert(inst.Register >= 8 && inst.Register <= 15); |
1840 | streamer.emitInt8(Value: 0xe0 | (inst.Register - 8)); |
1841 | break; |
1842 | case Win64EH::UOP_WideAllocMedium: |
1843 | assert((inst.Offset & 3) == 0); |
1844 | assert(inst.Offset / 4 <= 0x3ff); |
1845 | w = 0xe800 | (inst.Offset / 4); |
1846 | streamer.emitInt8(Value: (w >> 8) & 0xff); |
1847 | streamer.emitInt8(Value: (w >> 0) & 0xff); |
1848 | break; |
1849 | case Win64EH::UOP_SaveRegMask: |
1850 | assert((inst.Register & ~0x40ff) == 0); |
1851 | lr = (inst.Register >> 14) & 1; |
1852 | w = 0xec00 | (inst.Register & 0x0ff) | (lr << 8); |
1853 | streamer.emitInt8(Value: (w >> 8) & 0xff); |
1854 | streamer.emitInt8(Value: (w >> 0) & 0xff); |
1855 | break; |
1856 | case Win64EH::UOP_SaveLR: |
1857 | assert((inst.Offset & 3) == 0); |
1858 | assert(inst.Offset / 4 <= 0x0f); |
1859 | streamer.emitInt8(Value: 0xef); |
1860 | streamer.emitInt8(Value: inst.Offset / 4); |
1861 | break; |
1862 | case Win64EH::UOP_SaveFRegD0D15: |
1863 | assert(inst.Register <= 15); |
1864 | assert(inst.Offset <= 15); |
1865 | assert(inst.Register <= inst.Offset); |
1866 | streamer.emitInt8(Value: 0xf5); |
1867 | streamer.emitInt8(Value: (inst.Register << 4) | inst.Offset); |
1868 | break; |
1869 | case Win64EH::UOP_SaveFRegD16D31: |
1870 | assert(inst.Register >= 16 && inst.Register <= 31); |
1871 | assert(inst.Offset >= 16 && inst.Offset <= 31); |
1872 | assert(inst.Register <= inst.Offset); |
1873 | streamer.emitInt8(Value: 0xf6); |
1874 | streamer.emitInt8(Value: ((inst.Register - 16) << 4) | (inst.Offset - 16)); |
1875 | break; |
1876 | case Win64EH::UOP_AllocLarge: |
1877 | assert((inst.Offset & 3) == 0); |
1878 | assert(inst.Offset / 4 <= 0xffff); |
1879 | w = inst.Offset / 4; |
1880 | streamer.emitInt8(Value: 0xf7); |
1881 | streamer.emitInt8(Value: (w >> 8) & 0xff); |
1882 | streamer.emitInt8(Value: (w >> 0) & 0xff); |
1883 | break; |
1884 | case Win64EH::UOP_AllocHuge: |
1885 | assert((inst.Offset & 3) == 0); |
1886 | assert(inst.Offset / 4 <= 0xffffff); |
1887 | w = inst.Offset / 4; |
1888 | streamer.emitInt8(Value: 0xf8); |
1889 | streamer.emitInt8(Value: (w >> 16) & 0xff); |
1890 | streamer.emitInt8(Value: (w >> 8) & 0xff); |
1891 | streamer.emitInt8(Value: (w >> 0) & 0xff); |
1892 | break; |
1893 | case Win64EH::UOP_WideAllocLarge: |
1894 | assert((inst.Offset & 3) == 0); |
1895 | assert(inst.Offset / 4 <= 0xffff); |
1896 | w = inst.Offset / 4; |
1897 | streamer.emitInt8(Value: 0xf9); |
1898 | streamer.emitInt8(Value: (w >> 8) & 0xff); |
1899 | streamer.emitInt8(Value: (w >> 0) & 0xff); |
1900 | break; |
1901 | case Win64EH::UOP_WideAllocHuge: |
1902 | assert((inst.Offset & 3) == 0); |
1903 | assert(inst.Offset / 4 <= 0xffffff); |
1904 | w = inst.Offset / 4; |
1905 | streamer.emitInt8(Value: 0xfa); |
1906 | streamer.emitInt8(Value: (w >> 16) & 0xff); |
1907 | streamer.emitInt8(Value: (w >> 8) & 0xff); |
1908 | streamer.emitInt8(Value: (w >> 0) & 0xff); |
1909 | break; |
1910 | case Win64EH::UOP_Nop: |
1911 | streamer.emitInt8(Value: 0xfb); |
1912 | break; |
1913 | case Win64EH::UOP_WideNop: |
1914 | streamer.emitInt8(Value: 0xfc); |
1915 | break; |
1916 | case Win64EH::UOP_EndNop: |
1917 | streamer.emitInt8(Value: 0xfd); |
1918 | break; |
1919 | case Win64EH::UOP_WideEndNop: |
1920 | streamer.emitInt8(Value: 0xfe); |
1921 | break; |
1922 | case Win64EH::UOP_End: |
1923 | streamer.emitInt8(Value: 0xff); |
1924 | break; |
1925 | case Win64EH::UOP_Custom: |
1926 | for (i = 3; i > 0; i--) |
1927 | if (inst.Offset & (0xffu << (8 * i))) |
1928 | break; |
1929 | for (; i >= 0; i--) |
1930 | streamer.emitInt8(Value: (inst.Offset >> (8 * i)) & 0xff); |
1931 | break; |
1932 | } |
1933 | } |
1934 | |
1935 | // Check if an epilog exists as a subset of the end of a prolog (backwards). |
1936 | // An epilog may end with one out of three different end opcodes; if this |
1937 | // is the first epilog that shares opcodes with the prolog, we can tolerate |
1938 | // that this opcode differs (and the caller will update the prolog to use |
1939 | // the same end opcode as the epilog). If another epilog already shares |
1940 | // opcodes with the prolog, the ending opcode must be a strict match. |
1941 | static int getARMOffsetInProlog(const std::vector<WinEH::Instruction> &Prolog, |
1942 | const std::vector<WinEH::Instruction> &Epilog, |
1943 | bool CanTweakProlog) { |
1944 | // Can't find an epilog as a subset if it is longer than the prolog. |
1945 | if (Epilog.size() > Prolog.size()) |
1946 | return -1; |
1947 | |
1948 | // Check that the epilog actually is a perfect match for the end (backwrds) |
1949 | // of the prolog. |
1950 | // If we can adjust the prolog afterwards, don't check that the end opcodes |
1951 | // match. |
1952 | int EndIdx = CanTweakProlog ? 1 : 0; |
1953 | for (int I = Epilog.size() - 1; I >= EndIdx; I--) { |
1954 | // TODO: Could also allow minor mismatches, e.g. "add sp, #16" vs |
1955 | // "push {r0-r3}". |
1956 | if (Prolog[I] != Epilog[Epilog.size() - 1 - I]) |
1957 | return -1; |
1958 | } |
1959 | |
1960 | if (CanTweakProlog) { |
1961 | // Check that both prolog and epilog end with an expected end opcode. |
1962 | if (Prolog.front().Operation != Win64EH::UOP_End) |
1963 | return -1; |
1964 | if (Epilog.back().Operation != Win64EH::UOP_End && |
1965 | Epilog.back().Operation != Win64EH::UOP_EndNop && |
1966 | Epilog.back().Operation != Win64EH::UOP_WideEndNop) |
1967 | return -1; |
1968 | } |
1969 | |
1970 | // If the epilog was a subset of the prolog, find its offset. |
1971 | if (Epilog.size() == Prolog.size()) |
1972 | return 0; |
1973 | return ARMCountOfUnwindCodes(Insns: ArrayRef<WinEH::Instruction>( |
1974 | &Prolog[Epilog.size()], Prolog.size() - Epilog.size())); |
1975 | } |
1976 | |
1977 | static int checkARMPackedEpilog(MCStreamer &streamer, WinEH::FrameInfo *info, |
1978 | int PrologCodeBytes) { |
1979 | // Can only pack if there's one single epilog |
1980 | if (info->EpilogMap.size() != 1) |
1981 | return -1; |
1982 | |
1983 | const WinEH::FrameInfo::Epilog &EpilogInfo = info->EpilogMap.begin()->second; |
1984 | // Can only pack if the epilog is unconditional |
1985 | if (EpilogInfo.Condition != 0xe) // ARMCC::AL |
1986 | return -1; |
1987 | |
1988 | const std::vector<WinEH::Instruction> &Epilog = EpilogInfo.Instructions; |
1989 | // Make sure we have at least the trailing end opcode |
1990 | if (info->Instructions.empty() || Epilog.empty()) |
1991 | return -1; |
1992 | |
1993 | // Check that the epilog actually is at the very end of the function, |
1994 | // otherwise it can't be packed. |
1995 | std::optional<int64_t> MaybeDistance = GetOptionalAbsDifference( |
1996 | Streamer&: streamer, LHS: info->FuncletOrFuncEnd, RHS: info->EpilogMap.begin()->first); |
1997 | if (!MaybeDistance) |
1998 | return -1; |
1999 | uint32_t DistanceFromEnd = (uint32_t)*MaybeDistance; |
2000 | uint32_t InstructionBytes = ARMCountOfInstructionBytes(Insns: Epilog); |
2001 | if (DistanceFromEnd != InstructionBytes) |
2002 | return -1; |
2003 | |
2004 | int RetVal = -1; |
2005 | // Even if we don't end up sharing opcodes with the prolog, we can still |
2006 | // write the offset as a packed offset, if the single epilog is located at |
2007 | // the end of the function and the offset (pointing after the prolog) fits |
2008 | // as a packed offset. |
2009 | if (PrologCodeBytes <= 31 && |
2010 | PrologCodeBytes + ARMCountOfUnwindCodes(Insns: Epilog) <= 63) |
2011 | RetVal = PrologCodeBytes; |
2012 | |
2013 | int Offset = |
2014 | getARMOffsetInProlog(Prolog: info->Instructions, Epilog, /*CanTweakProlog=*/true); |
2015 | if (Offset < 0) |
2016 | return RetVal; |
2017 | |
2018 | // Check that the offset and prolog size fits in the first word; it's |
2019 | // unclear whether the epilog count in the extension word can be taken |
2020 | // as packed epilog offset. |
2021 | if (Offset > 31 || PrologCodeBytes > 63) |
2022 | return RetVal; |
2023 | |
2024 | // Replace the regular end opcode of the prolog with the one from the |
2025 | // epilog. |
2026 | info->Instructions.front() = Epilog.back(); |
2027 | |
2028 | // As we choose to express the epilog as part of the prolog, remove the |
2029 | // epilog from the map, so we don't try to emit its opcodes. |
2030 | info->EpilogMap.clear(); |
2031 | return Offset; |
2032 | } |
2033 | |
2034 | static bool parseRegMask(unsigned Mask, bool &HasLR, bool &HasR11, |
2035 | unsigned &Folded, int &IntRegs) { |
2036 | if (Mask & (1 << 14)) { |
2037 | HasLR = true; |
2038 | Mask &= ~(1 << 14); |
2039 | } |
2040 | if (Mask & (1 << 11)) { |
2041 | HasR11 = true; |
2042 | Mask &= ~(1 << 11); |
2043 | } |
2044 | Folded = 0; |
2045 | IntRegs = -1; |
2046 | if (!Mask) |
2047 | return true; |
2048 | int First = 0; |
2049 | // Shift right until we have the bits at the bottom |
2050 | while ((Mask & 1) == 0) { |
2051 | First++; |
2052 | Mask >>= 1; |
2053 | } |
2054 | if ((Mask & (Mask + 1)) != 0) |
2055 | return false; // Not a consecutive series of bits? Can't be packed. |
2056 | // Count the bits |
2057 | int N = 0; |
2058 | while (Mask & (1 << N)) |
2059 | N++; |
2060 | if (First < 4) { |
2061 | if (First + N < 4) |
2062 | return false; |
2063 | Folded = 4 - First; |
2064 | N -= Folded; |
2065 | First = 4; |
2066 | } |
2067 | if (First > 4) |
2068 | return false; // Can't be packed |
2069 | if (N >= 1) |
2070 | IntRegs = N - 1; |
2071 | return true; |
2072 | } |
2073 | |
2074 | static bool tryARMPackedUnwind(MCStreamer &streamer, WinEH::FrameInfo *info, |
2075 | uint32_t FuncLength) { |
2076 | int Step = 0; |
2077 | bool Homing = false; |
2078 | bool HasR11 = false; |
2079 | bool HasChain = false; |
2080 | bool HasLR = false; |
2081 | int IntRegs = -1; // r4 - r(4+N) |
2082 | int FloatRegs = -1; // d8 - d(8+N) |
2083 | unsigned PF = 0; // Number of extra pushed registers |
2084 | unsigned StackAdjust = 0; |
2085 | // Iterate over the prolog and check that all opcodes exactly match |
2086 | // the canonical order and form. |
2087 | for (const WinEH::Instruction &Inst : info->Instructions) { |
2088 | switch (Inst.Operation) { |
2089 | default: |
2090 | llvm_unreachable("Unsupported ARM unwind code" ); |
2091 | case Win64EH::UOP_Custom: |
2092 | case Win64EH::UOP_AllocLarge: |
2093 | case Win64EH::UOP_AllocHuge: |
2094 | case Win64EH::UOP_WideAllocLarge: |
2095 | case Win64EH::UOP_WideAllocHuge: |
2096 | case Win64EH::UOP_SaveFRegD0D15: |
2097 | case Win64EH::UOP_SaveFRegD16D31: |
2098 | // Can't be packed |
2099 | return false; |
2100 | case Win64EH::UOP_SaveSP: |
2101 | // Can't be packed; we can't rely on restoring sp from r11 when |
2102 | // unwinding a packed prologue. |
2103 | return false; |
2104 | case Win64EH::UOP_SaveLR: |
2105 | // Can't be present in a packed prologue |
2106 | return false; |
2107 | |
2108 | case Win64EH::UOP_End: |
2109 | case Win64EH::UOP_EndNop: |
2110 | case Win64EH::UOP_WideEndNop: |
2111 | if (Step != 0) |
2112 | return false; |
2113 | Step = 1; |
2114 | break; |
2115 | |
2116 | case Win64EH::UOP_SaveRegsR4R7LR: |
2117 | case Win64EH::UOP_WideSaveRegsR4R11LR: |
2118 | // push {r4-r11,lr} |
2119 | if (Step != 1 && Step != 2) |
2120 | return false; |
2121 | assert(Inst.Register >= 4 && Inst.Register <= 11); // r4-rX |
2122 | assert(Inst.Offset <= 1); // Lr |
2123 | IntRegs = Inst.Register - 4; |
2124 | if (Inst.Register == 11) { |
2125 | HasR11 = true; |
2126 | IntRegs--; |
2127 | } |
2128 | if (Inst.Offset) |
2129 | HasLR = true; |
2130 | Step = 3; |
2131 | break; |
2132 | |
2133 | case Win64EH::UOP_SaveRegMask: |
2134 | if (Step == 1 && Inst.Register == 0x0f) { |
2135 | // push {r0-r3} |
2136 | Homing = true; |
2137 | Step = 2; |
2138 | break; |
2139 | } |
2140 | [[fallthrough]]; |
2141 | case Win64EH::UOP_WideSaveRegMask: |
2142 | if (Step != 1 && Step != 2) |
2143 | return false; |
2144 | // push {r4-r9,r11,lr} |
2145 | // push {r11,lr} |
2146 | // push {r1-r5} |
2147 | if (!parseRegMask(Mask: Inst.Register, HasLR, HasR11, Folded&: PF, IntRegs)) |
2148 | return false; |
2149 | Step = 3; |
2150 | break; |
2151 | |
2152 | case Win64EH::UOP_Nop: |
2153 | // mov r11, sp |
2154 | if (Step != 3 || !HasR11 || IntRegs >= 0 || PF > 0) |
2155 | return false; |
2156 | HasChain = true; |
2157 | Step = 4; |
2158 | break; |
2159 | case Win64EH::UOP_WideNop: |
2160 | // add.w r11, sp, #xx |
2161 | if (Step != 3 || !HasR11 || (IntRegs < 0 && PF == 0)) |
2162 | return false; |
2163 | HasChain = true; |
2164 | Step = 4; |
2165 | break; |
2166 | |
2167 | case Win64EH::UOP_SaveFRegD8D15: |
2168 | if (Step != 1 && Step != 2 && Step != 3 && Step != 4) |
2169 | return false; |
2170 | assert(Inst.Register >= 8 && Inst.Register <= 15); |
2171 | if (Inst.Register == 15) |
2172 | return false; // Can't pack this case, R==7 means no IntRegs |
2173 | if (IntRegs >= 0) |
2174 | return false; |
2175 | FloatRegs = Inst.Register - 8; |
2176 | Step = 5; |
2177 | break; |
2178 | |
2179 | case Win64EH::UOP_AllocSmall: |
2180 | case Win64EH::UOP_WideAllocMedium: |
2181 | if (Step != 1 && Step != 2 && Step != 3 && Step != 4 && Step != 5) |
2182 | return false; |
2183 | if (PF > 0) // Can't have both folded and explicit stack allocation |
2184 | return false; |
2185 | if (Inst.Offset / 4 >= 0x3f4) |
2186 | return false; |
2187 | StackAdjust = Inst.Offset / 4; |
2188 | Step = 6; |
2189 | break; |
2190 | } |
2191 | } |
2192 | if (HasR11 && !HasChain) { |
2193 | if (IntRegs + 4 == 10) { |
2194 | // r11 stored, but not chaining; can be packed if already saving r4-r10 |
2195 | // and we can fit r11 into this range. |
2196 | IntRegs++; |
2197 | HasR11 = false; |
2198 | } else |
2199 | return false; |
2200 | } |
2201 | if (HasChain && !HasLR) |
2202 | return false; |
2203 | |
2204 | // Packed uneind info can't express multiple epilogues. |
2205 | if (info->EpilogMap.size() > 1) |
2206 | return false; |
2207 | |
2208 | unsigned EF = 0; |
2209 | int Ret = 0; |
2210 | if (info->EpilogMap.size() == 0) { |
2211 | Ret = 3; // No epilogue |
2212 | } else { |
2213 | // As the prologue and epilogue aren't exact mirrors of each other, |
2214 | // we have to check the epilogue too and see if it matches what we've |
2215 | // concluded from the prologue. |
2216 | const WinEH::FrameInfo::Epilog &EpilogInfo = |
2217 | info->EpilogMap.begin()->second; |
2218 | if (EpilogInfo.Condition != 0xe) // ARMCC::AL |
2219 | return false; |
2220 | const std::vector<WinEH::Instruction> &Epilog = EpilogInfo.Instructions; |
2221 | std::optional<int64_t> MaybeDistance = GetOptionalAbsDifference( |
2222 | Streamer&: streamer, LHS: info->FuncletOrFuncEnd, RHS: info->EpilogMap.begin()->first); |
2223 | if (!MaybeDistance) |
2224 | return false; |
2225 | uint32_t DistanceFromEnd = (uint32_t)*MaybeDistance; |
2226 | uint32_t InstructionBytes = ARMCountOfInstructionBytes(Insns: Epilog); |
2227 | if (DistanceFromEnd != InstructionBytes) |
2228 | return false; |
2229 | |
2230 | bool GotStackAdjust = false; |
2231 | bool GotFloatRegs = false; |
2232 | bool GotIntRegs = false; |
2233 | bool GotHomingRestore = false; |
2234 | bool GotLRRestore = false; |
2235 | bool NeedsReturn = false; |
2236 | bool GotReturn = false; |
2237 | |
2238 | Step = 6; |
2239 | for (const WinEH::Instruction &Inst : Epilog) { |
2240 | switch (Inst.Operation) { |
2241 | default: |
2242 | llvm_unreachable("Unsupported ARM unwind code" ); |
2243 | case Win64EH::UOP_Custom: |
2244 | case Win64EH::UOP_AllocLarge: |
2245 | case Win64EH::UOP_AllocHuge: |
2246 | case Win64EH::UOP_WideAllocLarge: |
2247 | case Win64EH::UOP_WideAllocHuge: |
2248 | case Win64EH::UOP_SaveFRegD0D15: |
2249 | case Win64EH::UOP_SaveFRegD16D31: |
2250 | case Win64EH::UOP_SaveSP: |
2251 | case Win64EH::UOP_Nop: |
2252 | case Win64EH::UOP_WideNop: |
2253 | // Can't be packed in an epilogue |
2254 | return false; |
2255 | |
2256 | case Win64EH::UOP_AllocSmall: |
2257 | case Win64EH::UOP_WideAllocMedium: |
2258 | if (Inst.Offset / 4 >= 0x3f4) |
2259 | return false; |
2260 | if (Step == 6) { |
2261 | if (Homing && FloatRegs < 0 && IntRegs < 0 && StackAdjust == 0 && |
2262 | PF == 0 && Inst.Offset == 16) { |
2263 | GotHomingRestore = true; |
2264 | Step = 10; |
2265 | } else { |
2266 | if (StackAdjust > 0) { |
2267 | // Got stack adjust in prologue too; must match. |
2268 | if (StackAdjust != Inst.Offset / 4) |
2269 | return false; |
2270 | GotStackAdjust = true; |
2271 | } else if (PF == Inst.Offset / 4) { |
2272 | // Folded prologue, non-folded epilogue |
2273 | StackAdjust = Inst.Offset / 4; |
2274 | GotStackAdjust = true; |
2275 | } else { |
2276 | // StackAdjust == 0 in prologue, mismatch |
2277 | return false; |
2278 | } |
2279 | Step = 7; |
2280 | } |
2281 | } else if (Step == 7 || Step == 8 || Step == 9) { |
2282 | if (!Homing || Inst.Offset != 16) |
2283 | return false; |
2284 | GotHomingRestore = true; |
2285 | Step = 10; |
2286 | } else |
2287 | return false; |
2288 | break; |
2289 | |
2290 | case Win64EH::UOP_SaveFRegD8D15: |
2291 | if (Step != 6 && Step != 7) |
2292 | return false; |
2293 | assert(Inst.Register >= 8 && Inst.Register <= 15); |
2294 | if (FloatRegs != (int)(Inst.Register - 8)) |
2295 | return false; |
2296 | GotFloatRegs = true; |
2297 | Step = 8; |
2298 | break; |
2299 | |
2300 | case Win64EH::UOP_SaveRegsR4R7LR: |
2301 | case Win64EH::UOP_WideSaveRegsR4R11LR: { |
2302 | // push {r4-r11,lr} |
2303 | if (Step != 6 && Step != 7 && Step != 8) |
2304 | return false; |
2305 | assert(Inst.Register >= 4 && Inst.Register <= 11); // r4-rX |
2306 | assert(Inst.Offset <= 1); // Lr |
2307 | if (Homing && HasLR) { |
2308 | // If homing and LR is backed up, we can either restore LR here |
2309 | // and return with Ret == 1 or 2, or return with SaveLR below |
2310 | if (Inst.Offset) { |
2311 | GotLRRestore = true; |
2312 | NeedsReturn = true; |
2313 | } else { |
2314 | // Expecting a separate SaveLR below |
2315 | } |
2316 | } else { |
2317 | if (HasLR != (Inst.Offset == 1)) |
2318 | return false; |
2319 | } |
2320 | GotLRRestore = Inst.Offset == 1; |
2321 | if (IntRegs < 0) // This opcode must include r4 |
2322 | return false; |
2323 | int Expected = IntRegs; |
2324 | if (HasChain) { |
2325 | // Can't express r11 here unless IntRegs describe r4-r10 |
2326 | if (IntRegs != 6) |
2327 | return false; |
2328 | Expected++; |
2329 | } |
2330 | if (Expected != (int)(Inst.Register - 4)) |
2331 | return false; |
2332 | GotIntRegs = true; |
2333 | Step = 9; |
2334 | break; |
2335 | } |
2336 | |
2337 | case Win64EH::UOP_SaveRegMask: |
2338 | case Win64EH::UOP_WideSaveRegMask: { |
2339 | if (Step != 6 && Step != 7 && Step != 8) |
2340 | return false; |
2341 | // push {r4-r9,r11,lr} |
2342 | // push {r11,lr} |
2343 | // push {r1-r5} |
2344 | bool CurHasLR = false, CurHasR11 = false; |
2345 | int Regs; |
2346 | if (!parseRegMask(Mask: Inst.Register, HasLR&: CurHasLR, HasR11&: CurHasR11, Folded&: EF, IntRegs&: Regs)) |
2347 | return false; |
2348 | if (EF > 0) { |
2349 | if (EF != PF && EF != StackAdjust) |
2350 | return false; |
2351 | } |
2352 | if (Homing && HasLR) { |
2353 | // If homing and LR is backed up, we can either restore LR here |
2354 | // and return with Ret == 1 or 2, or return with SaveLR below |
2355 | if (CurHasLR) { |
2356 | GotLRRestore = true; |
2357 | NeedsReturn = true; |
2358 | } else { |
2359 | // Expecting a separate SaveLR below |
2360 | } |
2361 | } else { |
2362 | if (CurHasLR != HasLR) |
2363 | return false; |
2364 | GotLRRestore = CurHasLR; |
2365 | } |
2366 | int Expected = IntRegs; |
2367 | if (HasChain) { |
2368 | // If we have chaining, the mask must have included r11. |
2369 | if (!CurHasR11) |
2370 | return false; |
2371 | } else if (Expected == 7) { |
2372 | // If we don't have chaining, the mask could still include r11, |
2373 | // expressed as part of IntRegs Instead. |
2374 | Expected--; |
2375 | if (!CurHasR11) |
2376 | return false; |
2377 | } else { |
2378 | // Neither HasChain nor r11 included in IntRegs, must not have r11 |
2379 | // here either. |
2380 | if (CurHasR11) |
2381 | return false; |
2382 | } |
2383 | if (Expected != Regs) |
2384 | return false; |
2385 | GotIntRegs = true; |
2386 | Step = 9; |
2387 | break; |
2388 | } |
2389 | |
2390 | case Win64EH::UOP_SaveLR: |
2391 | if (Step != 6 && Step != 7 && Step != 8 && Step != 9) |
2392 | return false; |
2393 | if (!Homing || Inst.Offset != 20 || GotLRRestore) |
2394 | return false; |
2395 | GotLRRestore = true; |
2396 | GotHomingRestore = true; |
2397 | Step = 10; |
2398 | break; |
2399 | |
2400 | case Win64EH::UOP_EndNop: |
2401 | case Win64EH::UOP_WideEndNop: |
2402 | GotReturn = true; |
2403 | Ret = (Inst.Operation == Win64EH::UOP_EndNop) ? 1 : 2; |
2404 | [[fallthrough]]; |
2405 | case Win64EH::UOP_End: |
2406 | if (Step != 6 && Step != 7 && Step != 8 && Step != 9 && Step != 10) |
2407 | return false; |
2408 | Step = 11; |
2409 | break; |
2410 | } |
2411 | } |
2412 | |
2413 | if (Step != 11) |
2414 | return false; |
2415 | if (StackAdjust > 0 && !GotStackAdjust && EF == 0) |
2416 | return false; |
2417 | if (FloatRegs >= 0 && !GotFloatRegs) |
2418 | return false; |
2419 | if (IntRegs >= 0 && !GotIntRegs) |
2420 | return false; |
2421 | if (Homing && !GotHomingRestore) |
2422 | return false; |
2423 | if (HasLR && !GotLRRestore) |
2424 | return false; |
2425 | if (NeedsReturn && !GotReturn) |
2426 | return false; |
2427 | } |
2428 | |
2429 | assert(PF == 0 || EF == 0 || |
2430 | StackAdjust == 0); // Can't have adjust in all three |
2431 | if (PF > 0 || EF > 0) { |
2432 | StackAdjust = PF > 0 ? (PF - 1) : (EF - 1); |
2433 | assert(StackAdjust <= 3); |
2434 | StackAdjust |= 0x3f0; |
2435 | if (PF > 0) |
2436 | StackAdjust |= 1 << 2; |
2437 | if (EF > 0) |
2438 | StackAdjust |= 1 << 3; |
2439 | } |
2440 | |
2441 | assert(FuncLength <= 0x7FF && "FuncLength should have been checked earlier" ); |
2442 | int Flag = info->Fragment ? 0x02 : 0x01; |
2443 | int H = Homing ? 1 : 0; |
2444 | int L = HasLR ? 1 : 0; |
2445 | int C = HasChain ? 1 : 0; |
2446 | assert(IntRegs < 0 || FloatRegs < 0); |
2447 | unsigned Reg, R; |
2448 | if (IntRegs >= 0) { |
2449 | Reg = IntRegs; |
2450 | assert(Reg <= 7); |
2451 | R = 0; |
2452 | } else if (FloatRegs >= 0) { |
2453 | Reg = FloatRegs; |
2454 | assert(Reg < 7); |
2455 | R = 1; |
2456 | } else { |
2457 | // No int or float regs stored (except possibly R11,LR) |
2458 | Reg = 7; |
2459 | R = 1; |
2460 | } |
2461 | info->PackedInfo |= Flag << 0; |
2462 | info->PackedInfo |= (FuncLength & 0x7FF) << 2; |
2463 | info->PackedInfo |= (Ret & 0x3) << 13; |
2464 | info->PackedInfo |= H << 15; |
2465 | info->PackedInfo |= Reg << 16; |
2466 | info->PackedInfo |= R << 19; |
2467 | info->PackedInfo |= L << 20; |
2468 | info->PackedInfo |= C << 21; |
2469 | assert(StackAdjust <= 0x3ff); |
2470 | info->PackedInfo |= StackAdjust << 22; |
2471 | return true; |
2472 | } |
2473 | |
2474 | // Populate the .xdata section. The format of .xdata on ARM is documented at |
2475 | // https://docs.microsoft.com/en-us/cpp/build/arm-exception-handling |
2476 | static void ARMEmitUnwindInfo(MCStreamer &streamer, WinEH::FrameInfo *info, |
2477 | bool TryPacked = true) { |
2478 | // If this UNWIND_INFO already has a symbol, it's already been emitted. |
2479 | if (info->Symbol) |
2480 | return; |
2481 | // If there's no unwind info here (not even a terminating UOP_End), the |
2482 | // unwind info is considered bogus and skipped. If this was done in |
2483 | // response to an explicit .seh_handlerdata, the associated trailing |
2484 | // handler data is left orphaned in the xdata section. |
2485 | if (info->empty()) { |
2486 | info->EmitAttempted = true; |
2487 | return; |
2488 | } |
2489 | if (info->EmitAttempted) { |
2490 | // If we tried to emit unwind info before (due to an explicit |
2491 | // .seh_handlerdata directive), but skipped it (because there was no |
2492 | // valid information to emit at the time), and it later got valid unwind |
2493 | // opcodes, we can't emit it here, because the trailing handler data |
2494 | // was already emitted elsewhere in the xdata section. |
2495 | streamer.getContext().reportError( |
2496 | L: SMLoc(), Msg: "Earlier .seh_handlerdata for " + info->Function->getName() + |
2497 | " skipped due to no unwind info at the time " |
2498 | "(.seh_handlerdata too early?), but the function later " |
2499 | "did get unwind info that can't be emitted" ); |
2500 | return; |
2501 | } |
2502 | |
2503 | MCContext &context = streamer.getContext(); |
2504 | MCSymbol *Label = context.createTempSymbol(); |
2505 | |
2506 | streamer.emitValueToAlignment(Alignment: Align(4)); |
2507 | streamer.emitLabel(Symbol: Label); |
2508 | info->Symbol = Label; |
2509 | |
2510 | if (!info->PrologEnd) |
2511 | streamer.getContext().reportError(L: SMLoc(), Msg: "Prologue in " + |
2512 | info->Function->getName() + |
2513 | " not correctly terminated" ); |
2514 | |
2515 | if (info->PrologEnd && !info->Fragment) |
2516 | checkARMInstructions(Streamer&: streamer, Insns: info->Instructions, Begin: info->Begin, |
2517 | End: info->PrologEnd, Name: info->Function->getName(), |
2518 | Type: "prologue" ); |
2519 | for (auto &I : info->EpilogMap) { |
2520 | MCSymbol *EpilogStart = I.first; |
2521 | auto &Epilog = I.second; |
2522 | checkARMInstructions(Streamer&: streamer, Insns: Epilog.Instructions, Begin: EpilogStart, End: Epilog.End, |
2523 | Name: info->Function->getName(), Type: "epilogue" ); |
2524 | if (Epilog.Instructions.empty() || |
2525 | !isARMTerminator(inst: Epilog.Instructions.back())) |
2526 | streamer.getContext().reportError( |
2527 | L: SMLoc(), Msg: "Epilogue in " + info->Function->getName() + |
2528 | " not correctly terminated" ); |
2529 | } |
2530 | |
2531 | std::optional<int64_t> RawFuncLength; |
2532 | const MCExpr *FuncLengthExpr = nullptr; |
2533 | if (!info->FuncletOrFuncEnd) { |
2534 | report_fatal_error(reason: "FuncletOrFuncEnd not set" ); |
2535 | } else { |
2536 | // As the size of many thumb2 instructions isn't known until later, |
2537 | // we can't always rely on being able to calculate the absolute |
2538 | // length of the function here. If we can't calculate it, defer it |
2539 | // to a relocation. |
2540 | // |
2541 | // In such a case, we won't know if the function is too long so that |
2542 | // the unwind info would need to be split (but this isn't implemented |
2543 | // anyway). |
2544 | RawFuncLength = |
2545 | GetOptionalAbsDifference(Streamer&: streamer, LHS: info->FuncletOrFuncEnd, RHS: info->Begin); |
2546 | if (!RawFuncLength) |
2547 | FuncLengthExpr = |
2548 | GetSubDivExpr(Streamer&: streamer, LHS: info->FuncletOrFuncEnd, RHS: info->Begin, Div: 2); |
2549 | } |
2550 | uint32_t FuncLength = 0; |
2551 | if (RawFuncLength) |
2552 | FuncLength = (uint32_t)*RawFuncLength / 2; |
2553 | if (FuncLength > 0x3FFFF) |
2554 | report_fatal_error(reason: "SEH unwind data splitting not yet implemented" ); |
2555 | uint32_t PrologCodeBytes = ARMCountOfUnwindCodes(Insns: info->Instructions); |
2556 | uint32_t TotalCodeBytes = PrologCodeBytes; |
2557 | |
2558 | if (!info->HandlesExceptions && RawFuncLength && FuncLength <= 0x7ff && |
2559 | TryPacked) { |
2560 | // No exception handlers; check if the prolog and epilog matches the |
2561 | // patterns that can be described by the packed format. If we don't |
2562 | // know the exact function length yet, we can't do this. |
2563 | |
2564 | // info->Symbol was already set even if we didn't actually write any |
2565 | // unwind info there. Keep using that as indicator that this unwind |
2566 | // info has been generated already. |
2567 | |
2568 | if (tryARMPackedUnwind(streamer, info, FuncLength)) |
2569 | return; |
2570 | } |
2571 | |
2572 | int PackedEpilogOffset = |
2573 | checkARMPackedEpilog(streamer, info, PrologCodeBytes); |
2574 | |
2575 | // Process epilogs. |
2576 | MapVector<MCSymbol *, uint32_t> EpilogInfo; |
2577 | // Epilogs processed so far. |
2578 | std::vector<MCSymbol *> AddedEpilogs; |
2579 | |
2580 | bool CanTweakProlog = true; |
2581 | for (auto &I : info->EpilogMap) { |
2582 | MCSymbol *EpilogStart = I.first; |
2583 | auto &EpilogInstrs = I.second.Instructions; |
2584 | uint32_t CodeBytes = ARMCountOfUnwindCodes(Insns: EpilogInstrs); |
2585 | |
2586 | MCSymbol *MatchingEpilog = |
2587 | FindMatchingEpilog(EpilogInstrs, Epilogs: AddedEpilogs, info); |
2588 | int PrologOffset; |
2589 | if (MatchingEpilog) { |
2590 | assert(EpilogInfo.contains(MatchingEpilog) && |
2591 | "Duplicate epilog not found" ); |
2592 | EpilogInfo[EpilogStart] = EpilogInfo.lookup(Key: MatchingEpilog); |
2593 | // Clear the unwind codes in the EpilogMap, so that they don't get output |
2594 | // in the logic below. |
2595 | EpilogInstrs.clear(); |
2596 | } else if ((PrologOffset = getARMOffsetInProlog( |
2597 | Prolog: info->Instructions, Epilog: EpilogInstrs, CanTweakProlog)) >= 0) { |
2598 | if (CanTweakProlog) { |
2599 | // Replace the regular end opcode of the prolog with the one from the |
2600 | // epilog. |
2601 | info->Instructions.front() = EpilogInstrs.back(); |
2602 | // Later epilogs need a strict match for the end opcode. |
2603 | CanTweakProlog = false; |
2604 | } |
2605 | EpilogInfo[EpilogStart] = PrologOffset; |
2606 | // Clear the unwind codes in the EpilogMap, so that they don't get output |
2607 | // in the logic below. |
2608 | EpilogInstrs.clear(); |
2609 | } else { |
2610 | EpilogInfo[EpilogStart] = TotalCodeBytes; |
2611 | TotalCodeBytes += CodeBytes; |
2612 | AddedEpilogs.push_back(x: EpilogStart); |
2613 | } |
2614 | } |
2615 | |
2616 | // Code Words, Epilog count, F, E, X, Vers, Function Length |
2617 | uint32_t row1 = 0x0; |
2618 | uint32_t CodeWords = TotalCodeBytes / 4; |
2619 | uint32_t CodeWordsMod = TotalCodeBytes % 4; |
2620 | if (CodeWordsMod) |
2621 | CodeWords++; |
2622 | uint32_t EpilogCount = |
2623 | PackedEpilogOffset >= 0 ? PackedEpilogOffset : info->EpilogMap.size(); |
2624 | bool ExtensionWord = EpilogCount > 31 || CodeWords > 15; |
2625 | if (!ExtensionWord) { |
2626 | row1 |= (EpilogCount & 0x1F) << 23; |
2627 | row1 |= (CodeWords & 0x0F) << 28; |
2628 | } |
2629 | if (info->HandlesExceptions) // X |
2630 | row1 |= 1 << 20; |
2631 | if (PackedEpilogOffset >= 0) // E |
2632 | row1 |= 1 << 21; |
2633 | if (info->Fragment) // F |
2634 | row1 |= 1 << 22; |
2635 | row1 |= FuncLength & 0x3FFFF; |
2636 | if (RawFuncLength) |
2637 | streamer.emitInt32(Value: row1); |
2638 | else |
2639 | streamer.emitValue( |
2640 | Value: MCBinaryExpr::createOr(LHS: FuncLengthExpr, |
2641 | RHS: MCConstantExpr::create(Value: row1, Ctx&: context), Ctx&: context), |
2642 | Size: 4); |
2643 | |
2644 | // Extended Code Words, Extended Epilog Count |
2645 | if (ExtensionWord) { |
2646 | // FIXME: We should be able to split unwind info into multiple sections. |
2647 | if (CodeWords > 0xFF || EpilogCount > 0xFFFF) |
2648 | report_fatal_error(reason: "SEH unwind data splitting not yet implemented" ); |
2649 | uint32_t row2 = 0x0; |
2650 | row2 |= (CodeWords & 0xFF) << 16; |
2651 | row2 |= (EpilogCount & 0xFFFF); |
2652 | streamer.emitInt32(Value: row2); |
2653 | } |
2654 | |
2655 | if (PackedEpilogOffset < 0) { |
2656 | // Epilog Start Index, Epilog Start Offset |
2657 | for (auto &I : EpilogInfo) { |
2658 | MCSymbol *EpilogStart = I.first; |
2659 | uint32_t EpilogIndex = I.second; |
2660 | |
2661 | std::optional<int64_t> MaybeEpilogOffset = |
2662 | GetOptionalAbsDifference(Streamer&: streamer, LHS: EpilogStart, RHS: info->Begin); |
2663 | const MCExpr *OffsetExpr = nullptr; |
2664 | uint32_t EpilogOffset = 0; |
2665 | if (MaybeEpilogOffset) |
2666 | EpilogOffset = *MaybeEpilogOffset / 2; |
2667 | else |
2668 | OffsetExpr = GetSubDivExpr(Streamer&: streamer, LHS: EpilogStart, RHS: info->Begin, Div: 2); |
2669 | |
2670 | assert(info->EpilogMap.contains(EpilogStart)); |
2671 | unsigned Condition = info->EpilogMap[EpilogStart].Condition; |
2672 | assert(Condition <= 0xf); |
2673 | |
2674 | uint32_t row3 = EpilogOffset; |
2675 | row3 |= Condition << 20; |
2676 | row3 |= (EpilogIndex & 0x3FF) << 24; |
2677 | if (MaybeEpilogOffset) |
2678 | streamer.emitInt32(Value: row3); |
2679 | else |
2680 | streamer.emitValue( |
2681 | Value: MCBinaryExpr::createOr( |
2682 | LHS: OffsetExpr, RHS: MCConstantExpr::create(Value: row3, Ctx&: context), Ctx&: context), |
2683 | Size: 4); |
2684 | } |
2685 | } |
2686 | |
2687 | // Emit prolog unwind instructions (in reverse order). |
2688 | uint8_t numInst = info->Instructions.size(); |
2689 | for (uint8_t c = 0; c < numInst; ++c) { |
2690 | WinEH::Instruction inst = info->Instructions.back(); |
2691 | info->Instructions.pop_back(); |
2692 | ARMEmitUnwindCode(streamer, inst); |
2693 | } |
2694 | |
2695 | // Emit epilog unwind instructions |
2696 | for (auto &I : info->EpilogMap) { |
2697 | auto &EpilogInstrs = I.second.Instructions; |
2698 | for (const WinEH::Instruction &inst : EpilogInstrs) |
2699 | ARMEmitUnwindCode(streamer, inst); |
2700 | } |
2701 | |
2702 | int32_t BytesMod = CodeWords * 4 - TotalCodeBytes; |
2703 | assert(BytesMod >= 0); |
2704 | for (int i = 0; i < BytesMod; i++) |
2705 | streamer.emitInt8(Value: 0xFB); |
2706 | |
2707 | if (info->HandlesExceptions) |
2708 | streamer.emitValue( |
2709 | Value: MCSymbolRefExpr::create(Symbol: info->ExceptionHandler, |
2710 | specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, Ctx&: context), |
2711 | Size: 4); |
2712 | } |
2713 | |
2714 | static void ARM64EmitRuntimeFunction(MCStreamer &streamer, |
2715 | const WinEH::FrameInfo *info) { |
2716 | MCContext &context = streamer.getContext(); |
2717 | |
2718 | streamer.emitValueToAlignment(Alignment: Align(4)); |
2719 | for (const auto &S : info->Segments) { |
2720 | EmitSymbolRefWithOfs(streamer, Base: info->Begin, Offset: S.Offset); |
2721 | if (info->PackedInfo) |
2722 | streamer.emitInt32(Value: info->PackedInfo); |
2723 | else |
2724 | streamer.emitValue( |
2725 | Value: MCSymbolRefExpr::create(Symbol: S.Symbol, specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, |
2726 | Ctx&: context), |
2727 | Size: 4); |
2728 | } |
2729 | } |
2730 | |
2731 | |
2732 | static void ARMEmitRuntimeFunction(MCStreamer &streamer, |
2733 | const WinEH::FrameInfo *info) { |
2734 | MCContext &context = streamer.getContext(); |
2735 | |
2736 | streamer.emitValueToAlignment(Alignment: Align(4)); |
2737 | EmitSymbolRefWithOfs(streamer, Base: info->Begin, Other: info->Begin); |
2738 | if (info->PackedInfo) |
2739 | streamer.emitInt32(Value: info->PackedInfo); |
2740 | else |
2741 | streamer.emitValue( |
2742 | Value: MCSymbolRefExpr::create(Symbol: info->Symbol, specifier: MCSymbolRefExpr::VK_COFF_IMGREL32, |
2743 | Ctx&: context), |
2744 | Size: 4); |
2745 | } |
2746 | |
2747 | void llvm::Win64EH::ARM64UnwindEmitter::Emit(MCStreamer &Streamer) const { |
2748 | // Emit the unwind info structs first. |
2749 | for (const auto &CFI : Streamer.getWinFrameInfos()) { |
2750 | WinEH::FrameInfo *Info = CFI.get(); |
2751 | if (Info->empty()) |
2752 | continue; |
2753 | MCSection *XData = Streamer.getAssociatedXDataSection(TextSec: CFI->TextSection); |
2754 | Streamer.switchSection(Section: XData); |
2755 | ARM64EmitUnwindInfo(streamer&: Streamer, info: Info); |
2756 | } |
2757 | |
2758 | // Now emit RUNTIME_FUNCTION entries. |
2759 | for (const auto &CFI : Streamer.getWinFrameInfos()) { |
2760 | WinEH::FrameInfo *Info = CFI.get(); |
2761 | // ARM64EmitUnwindInfo above clears the info struct, so we can't check |
2762 | // empty here. But if a Symbol is set, we should create the corresponding |
2763 | // pdata entry. |
2764 | if (!Info->Symbol) |
2765 | continue; |
2766 | MCSection *PData = Streamer.getAssociatedPDataSection(TextSec: CFI->TextSection); |
2767 | Streamer.switchSection(Section: PData); |
2768 | ARM64EmitRuntimeFunction(streamer&: Streamer, info: Info); |
2769 | } |
2770 | } |
2771 | |
2772 | void llvm::Win64EH::ARM64UnwindEmitter::EmitUnwindInfo(MCStreamer &Streamer, |
2773 | WinEH::FrameInfo *info, |
2774 | bool HandlerData) const { |
2775 | // Called if there's an .seh_handlerdata directive before the end of the |
2776 | // function. This forces writing the xdata record already here - and |
2777 | // in this case, the function isn't actually ended already, but the xdata |
2778 | // record needs to know the function length. In these cases, if the funclet |
2779 | // end hasn't been marked yet, the xdata function length won't cover the |
2780 | // whole function, only up to this point. |
2781 | if (!info->FuncletOrFuncEnd) { |
2782 | Streamer.switchSection(Section: info->TextSection); |
2783 | info->FuncletOrFuncEnd = Streamer.emitCFILabel(); |
2784 | } |
2785 | // Switch sections (the static function above is meant to be called from |
2786 | // here and from Emit(). |
2787 | MCSection *XData = Streamer.getAssociatedXDataSection(TextSec: info->TextSection); |
2788 | Streamer.switchSection(Section: XData); |
2789 | ARM64EmitUnwindInfo(streamer&: Streamer, info, /* TryPacked = */ !HandlerData); |
2790 | } |
2791 | |
2792 | void llvm::Win64EH::ARMUnwindEmitter::Emit(MCStreamer &Streamer) const { |
2793 | // Emit the unwind info structs first. |
2794 | for (const auto &CFI : Streamer.getWinFrameInfos()) { |
2795 | WinEH::FrameInfo *Info = CFI.get(); |
2796 | if (Info->empty()) |
2797 | continue; |
2798 | MCSection *XData = Streamer.getAssociatedXDataSection(TextSec: CFI->TextSection); |
2799 | Streamer.switchSection(Section: XData); |
2800 | ARMEmitUnwindInfo(streamer&: Streamer, info: Info); |
2801 | } |
2802 | |
2803 | // Now emit RUNTIME_FUNCTION entries. |
2804 | for (const auto &CFI : Streamer.getWinFrameInfos()) { |
2805 | WinEH::FrameInfo *Info = CFI.get(); |
2806 | // ARMEmitUnwindInfo above clears the info struct, so we can't check |
2807 | // empty here. But if a Symbol is set, we should create the corresponding |
2808 | // pdata entry. |
2809 | if (!Info->Symbol) |
2810 | continue; |
2811 | MCSection *PData = Streamer.getAssociatedPDataSection(TextSec: CFI->TextSection); |
2812 | Streamer.switchSection(Section: PData); |
2813 | ARMEmitRuntimeFunction(streamer&: Streamer, info: Info); |
2814 | } |
2815 | } |
2816 | |
2817 | void llvm::Win64EH::ARMUnwindEmitter::EmitUnwindInfo(MCStreamer &Streamer, |
2818 | WinEH::FrameInfo *info, |
2819 | bool HandlerData) const { |
2820 | // Called if there's an .seh_handlerdata directive before the end of the |
2821 | // function. This forces writing the xdata record already here - and |
2822 | // in this case, the function isn't actually ended already, but the xdata |
2823 | // record needs to know the function length. In these cases, if the funclet |
2824 | // end hasn't been marked yet, the xdata function length won't cover the |
2825 | // whole function, only up to this point. |
2826 | if (!info->FuncletOrFuncEnd) { |
2827 | Streamer.switchSection(Section: info->TextSection); |
2828 | info->FuncletOrFuncEnd = Streamer.emitCFILabel(); |
2829 | } |
2830 | // Switch sections (the static function above is meant to be called from |
2831 | // here and from Emit(). |
2832 | MCSection *XData = Streamer.getAssociatedXDataSection(TextSec: info->TextSection); |
2833 | Streamer.switchSection(Section: XData); |
2834 | ARMEmitUnwindInfo(streamer&: Streamer, info, /* TryPacked = */ !HandlerData); |
2835 | } |
2836 | |