| 1 | //===- llvm/CodeGen/DwarfExpression.cpp - Dwarf Debug Framework -----------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file contains support for writing dwarf debug info into asm files. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "DwarfExpression.h" |
| 14 | #include "DwarfCompileUnit.h" |
| 15 | #include "llvm/ADT/APInt.h" |
| 16 | #include "llvm/ADT/SmallBitVector.h" |
| 17 | #include "llvm/BinaryFormat/Dwarf.h" |
| 18 | #include "llvm/CodeGen/Register.h" |
| 19 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 20 | #include "llvm/IR/DataLayout.h" |
| 21 | #include "llvm/MC/MCAsmInfo.h" |
| 22 | #include "llvm/Support/ErrorHandling.h" |
| 23 | #include <algorithm> |
| 24 | |
| 25 | using namespace llvm; |
| 26 | |
| 27 | #define DEBUG_TYPE "dwarfdebug" |
| 28 | |
| 29 | void DwarfExpression::emitConstu(uint64_t Value) { |
| 30 | if (Value < 32) |
| 31 | emitOp(Op: dwarf::DW_OP_lit0 + Value); |
| 32 | else if (Value == std::numeric_limits<uint64_t>::max()) { |
| 33 | // Only do this for 64-bit values as the DWARF expression stack uses |
| 34 | // target-address-size values. |
| 35 | emitOp(Op: dwarf::DW_OP_lit0); |
| 36 | emitOp(Op: dwarf::DW_OP_not); |
| 37 | } else { |
| 38 | emitOp(Op: dwarf::DW_OP_constu); |
| 39 | emitUnsigned(Value); |
| 40 | } |
| 41 | } |
| 42 | |
| 43 | void DwarfExpression::addReg(int64_t DwarfReg, const char *) { |
| 44 | assert(DwarfReg >= 0 && "invalid negative dwarf register number" ); |
| 45 | assert((isUnknownLocation() || isRegisterLocation()) && |
| 46 | "location description already locked down" ); |
| 47 | LocationKind = Register; |
| 48 | if (DwarfReg < 32) { |
| 49 | emitOp(Op: dwarf::DW_OP_reg0 + DwarfReg, Comment); |
| 50 | } else { |
| 51 | emitOp(Op: dwarf::DW_OP_regx, Comment); |
| 52 | emitUnsigned(Value: DwarfReg); |
| 53 | } |
| 54 | } |
| 55 | |
| 56 | void DwarfExpression::addBReg(int64_t DwarfReg, int64_t Offset) { |
| 57 | assert(DwarfReg >= 0 && "invalid negative dwarf register number" ); |
| 58 | assert(!isRegisterLocation() && "location description already locked down" ); |
| 59 | if (DwarfReg < 32) { |
| 60 | emitOp(Op: dwarf::DW_OP_breg0 + DwarfReg); |
| 61 | } else { |
| 62 | emitOp(Op: dwarf::DW_OP_bregx); |
| 63 | emitUnsigned(Value: DwarfReg); |
| 64 | } |
| 65 | emitSigned(Value: Offset); |
| 66 | } |
| 67 | |
| 68 | void DwarfExpression::addFBReg(int64_t Offset) { |
| 69 | emitOp(Op: dwarf::DW_OP_fbreg); |
| 70 | emitSigned(Value: Offset); |
| 71 | } |
| 72 | |
| 73 | void DwarfExpression::addOpPiece(unsigned SizeInBits, unsigned OffsetInBits) { |
| 74 | if (!SizeInBits) |
| 75 | return; |
| 76 | |
| 77 | const unsigned SizeOfByte = 8; |
| 78 | if (OffsetInBits > 0 || SizeInBits % SizeOfByte) { |
| 79 | emitOp(Op: dwarf::DW_OP_bit_piece); |
| 80 | emitUnsigned(Value: SizeInBits); |
| 81 | emitUnsigned(Value: OffsetInBits); |
| 82 | } else { |
| 83 | emitOp(Op: dwarf::DW_OP_piece); |
| 84 | unsigned ByteSize = SizeInBits / SizeOfByte; |
| 85 | emitUnsigned(Value: ByteSize); |
| 86 | } |
| 87 | this->OffsetInBits += SizeInBits; |
| 88 | } |
| 89 | |
| 90 | void DwarfExpression::addShr(unsigned ShiftBy) { |
| 91 | emitConstu(Value: ShiftBy); |
| 92 | emitOp(Op: dwarf::DW_OP_shr); |
| 93 | } |
| 94 | |
| 95 | void DwarfExpression::addAnd(unsigned Mask) { |
| 96 | emitConstu(Value: Mask); |
| 97 | emitOp(Op: dwarf::DW_OP_and); |
| 98 | } |
| 99 | |
| 100 | bool DwarfExpression::addMachineReg(const TargetRegisterInfo &TRI, |
| 101 | llvm::Register MachineReg, |
| 102 | unsigned MaxSize) { |
| 103 | if (!MachineReg.isPhysical()) { |
| 104 | if (isFrameRegister(TRI, MachineReg)) { |
| 105 | DwarfRegs.push_back(Elt: Register::createRegister(RegNo: -1, Comment: nullptr)); |
| 106 | return true; |
| 107 | } |
| 108 | // Try getting dwarf register for virtual register anyway, eg. for NVPTX. |
| 109 | int64_t Reg = TRI.getDwarfRegNumForVirtReg(RegNum: MachineReg, isEH: false); |
| 110 | if (Reg > 0) { |
| 111 | DwarfRegs.push_back(Elt: Register::createRegister(RegNo: Reg, Comment: nullptr)); |
| 112 | return true; |
| 113 | } |
| 114 | return false; |
| 115 | } |
| 116 | |
| 117 | int64_t Reg = TRI.getDwarfRegNum(RegNum: MachineReg, isEH: false); |
| 118 | |
| 119 | // If this is a valid register number, emit it. |
| 120 | if (Reg >= 0) { |
| 121 | DwarfRegs.push_back(Elt: Register::createRegister(RegNo: Reg, Comment: nullptr)); |
| 122 | return true; |
| 123 | } |
| 124 | |
| 125 | // Walk up the super-register chain until we find a valid number. |
| 126 | // For example, EAX on x86_64 is a 32-bit fragment of RAX with offset 0. |
| 127 | for (MCPhysReg SR : TRI.superregs(Reg: MachineReg)) { |
| 128 | Reg = TRI.getDwarfRegNum(RegNum: SR, isEH: false); |
| 129 | if (Reg >= 0) { |
| 130 | unsigned Idx = TRI.getSubRegIndex(RegNo: SR, SubRegNo: MachineReg); |
| 131 | unsigned Size = TRI.getSubRegIdxSize(Idx); |
| 132 | unsigned RegOffset = TRI.getSubRegIdxOffset(Idx); |
| 133 | DwarfRegs.push_back(Elt: Register::createRegister(RegNo: Reg, Comment: "super-register" )); |
| 134 | // Use a DW_OP_bit_piece to describe the sub-register. |
| 135 | setSubRegisterPiece(SizeInBits: Size, OffsetInBits: RegOffset); |
| 136 | return true; |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | // Otherwise, attempt to find a covering set of sub-register numbers. |
| 141 | // For example, Q0 on ARM is a composition of D0+D1. |
| 142 | unsigned CurPos = 0; |
| 143 | // The size of the register in bits. |
| 144 | const TargetRegisterClass *RC = TRI.getMinimalPhysRegClass(Reg: MachineReg); |
| 145 | unsigned RegSize = TRI.getRegSizeInBits(RC: *RC); |
| 146 | // Keep track of the bits in the register we already emitted, so we |
| 147 | // can avoid emitting redundant aliasing subregs. Because this is |
| 148 | // just doing a greedy scan of all subregisters, it is possible that |
| 149 | // this doesn't find a combination of subregisters that fully cover |
| 150 | // the register (even though one may exist). |
| 151 | SmallBitVector Coverage(RegSize, false); |
| 152 | for (MCPhysReg SR : TRI.subregs(Reg: MachineReg)) { |
| 153 | unsigned Idx = TRI.getSubRegIndex(RegNo: MachineReg, SubRegNo: SR); |
| 154 | unsigned Size = TRI.getSubRegIdxSize(Idx); |
| 155 | unsigned Offset = TRI.getSubRegIdxOffset(Idx); |
| 156 | Reg = TRI.getDwarfRegNum(RegNum: SR, isEH: false); |
| 157 | if (Reg < 0) |
| 158 | continue; |
| 159 | |
| 160 | // Used to build the intersection between the bits we already |
| 161 | // emitted and the bits covered by this subregister. |
| 162 | SmallBitVector CurSubReg(RegSize, false); |
| 163 | CurSubReg.set(I: Offset, E: Offset + Size); |
| 164 | |
| 165 | // If this sub-register has a DWARF number and we haven't covered |
| 166 | // its range, and its range covers the value, emit a DWARF piece for it. |
| 167 | if (Offset < MaxSize && CurSubReg.test(RHS: Coverage)) { |
| 168 | // Emit a piece for any gap in the coverage. |
| 169 | if (Offset > CurPos) |
| 170 | DwarfRegs.push_back(Elt: Register::createSubRegister( |
| 171 | RegNo: -1, SizeInBits: Offset - CurPos, Comment: "no DWARF register encoding" )); |
| 172 | if (Offset == 0 && Size >= MaxSize) |
| 173 | DwarfRegs.push_back(Elt: Register::createRegister(RegNo: Reg, Comment: "sub-register" )); |
| 174 | else |
| 175 | DwarfRegs.push_back(Elt: Register::createSubRegister( |
| 176 | RegNo: Reg, SizeInBits: std::min<unsigned>(a: Size, b: MaxSize - Offset), Comment: "sub-register" )); |
| 177 | } |
| 178 | // Mark it as emitted. |
| 179 | Coverage.set(I: Offset, E: Offset + Size); |
| 180 | CurPos = Offset + Size; |
| 181 | } |
| 182 | // Failed to find any DWARF encoding. |
| 183 | if (CurPos == 0) |
| 184 | return false; |
| 185 | // Found a partial or complete DWARF encoding. |
| 186 | if (CurPos < RegSize) |
| 187 | DwarfRegs.push_back(Elt: Register::createSubRegister( |
| 188 | RegNo: -1, SizeInBits: RegSize - CurPos, Comment: "no DWARF register encoding" )); |
| 189 | return true; |
| 190 | } |
| 191 | |
| 192 | void DwarfExpression::addStackValue() { |
| 193 | if (DwarfVersion >= 4) |
| 194 | emitOp(Op: dwarf::DW_OP_stack_value); |
| 195 | } |
| 196 | |
| 197 | void DwarfExpression::addSignedConstant(int64_t Value) { |
| 198 | assert(isImplicitLocation() || isUnknownLocation()); |
| 199 | LocationKind = Implicit; |
| 200 | emitOp(Op: dwarf::DW_OP_consts); |
| 201 | emitSigned(Value); |
| 202 | } |
| 203 | |
| 204 | void DwarfExpression::addUnsignedConstant(uint64_t Value) { |
| 205 | assert(isImplicitLocation() || isUnknownLocation()); |
| 206 | LocationKind = Implicit; |
| 207 | emitConstu(Value); |
| 208 | } |
| 209 | |
| 210 | void DwarfExpression::addUnsignedConstant(const APInt &Value) { |
| 211 | assert(isImplicitLocation() || isUnknownLocation()); |
| 212 | LocationKind = Implicit; |
| 213 | |
| 214 | unsigned Size = Value.getBitWidth(); |
| 215 | const uint64_t *Data = Value.getRawData(); |
| 216 | |
| 217 | // Chop it up into 64-bit pieces, because that's the maximum that |
| 218 | // addUnsignedConstant takes. |
| 219 | unsigned Offset = 0; |
| 220 | while (Offset < Size) { |
| 221 | addUnsignedConstant(Value: *Data++); |
| 222 | if (Offset == 0 && Size <= 64) |
| 223 | break; |
| 224 | addStackValue(); |
| 225 | addOpPiece(SizeInBits: std::min(a: Size - Offset, b: 64u), OffsetInBits: Offset); |
| 226 | Offset += 64; |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | void DwarfExpression::addConstantFP(const APFloat &APF, const AsmPrinter &AP) { |
| 231 | assert(isImplicitLocation() || isUnknownLocation()); |
| 232 | APInt API = APF.bitcastToAPInt(); |
| 233 | int NumBytes = API.getBitWidth() / 8; |
| 234 | if (NumBytes == 4 /*float*/ || NumBytes == 8 /*double*/) { |
| 235 | // FIXME: Add support for `long double`. |
| 236 | emitOp(Op: dwarf::DW_OP_implicit_value); |
| 237 | emitUnsigned(Value: NumBytes /*Size of the block in bytes*/); |
| 238 | |
| 239 | // The loop below is emitting the value starting at least significant byte, |
| 240 | // so we need to perform a byte-swap to get the byte order correct in case |
| 241 | // of a big-endian target. |
| 242 | if (AP.getDataLayout().isBigEndian()) |
| 243 | API = API.byteSwap(); |
| 244 | |
| 245 | for (int i = 0; i < NumBytes; ++i) { |
| 246 | emitData1(Value: API.getZExtValue() & 0xFF); |
| 247 | API = API.lshr(shiftAmt: 8); |
| 248 | } |
| 249 | |
| 250 | return; |
| 251 | } |
| 252 | LLVM_DEBUG( |
| 253 | dbgs() << "Skipped DW_OP_implicit_value creation for ConstantFP of size: " |
| 254 | << API.getBitWidth() << " bits\n" ); |
| 255 | } |
| 256 | |
| 257 | bool DwarfExpression::addMachineRegExpression(const TargetRegisterInfo &TRI, |
| 258 | DIExpressionCursor &ExprCursor, |
| 259 | llvm::Register MachineReg, |
| 260 | unsigned FragmentOffsetInBits) { |
| 261 | auto Fragment = ExprCursor.getFragmentInfo(); |
| 262 | if (!addMachineReg(TRI, MachineReg, MaxSize: Fragment ? Fragment->SizeInBits : ~1U)) { |
| 263 | LocationKind = Unknown; |
| 264 | return false; |
| 265 | } |
| 266 | |
| 267 | bool HasComplexExpression = false; |
| 268 | auto Op = ExprCursor.peek(); |
| 269 | if (Op && Op->getOp() != dwarf::DW_OP_LLVM_fragment) |
| 270 | HasComplexExpression = true; |
| 271 | |
| 272 | // If the register can only be described by a complex expression (i.e., |
| 273 | // multiple subregisters) it doesn't safely compose with another complex |
| 274 | // expression. For example, it is not possible to apply a DW_OP_deref |
| 275 | // operation to multiple DW_OP_pieces, since composite location descriptions |
| 276 | // do not push anything on the DWARF stack. |
| 277 | // |
| 278 | // DW_OP_entry_value operations can only hold a DWARF expression or a |
| 279 | // register location description, so we can't emit a single entry value |
| 280 | // covering a composite location description. In the future we may want to |
| 281 | // emit entry value operations for each register location in the composite |
| 282 | // location, but until that is supported do not emit anything. |
| 283 | if ((HasComplexExpression || IsEmittingEntryValue) && DwarfRegs.size() > 1) { |
| 284 | if (IsEmittingEntryValue) |
| 285 | cancelEntryValue(); |
| 286 | DwarfRegs.clear(); |
| 287 | LocationKind = Unknown; |
| 288 | return false; |
| 289 | } |
| 290 | |
| 291 | // Handle simple register locations. If we are supposed to emit |
| 292 | // a call site parameter expression and if that expression is just a register |
| 293 | // location, emit it with addBReg and offset 0, because we should emit a DWARF |
| 294 | // expression representing a value, rather than a location. |
| 295 | if ((!isParameterValue() && !isMemoryLocation() && !HasComplexExpression) || |
| 296 | isEntryValue()) { |
| 297 | auto FragmentInfo = ExprCursor.getFragmentInfo(); |
| 298 | unsigned RegSize = 0; |
| 299 | for (auto &Reg : DwarfRegs) { |
| 300 | RegSize += Reg.SubRegSize; |
| 301 | if (Reg.DwarfRegNo >= 0) |
| 302 | addReg(DwarfReg: Reg.DwarfRegNo, Comment: Reg.Comment); |
| 303 | if (FragmentInfo) |
| 304 | if (RegSize > FragmentInfo->SizeInBits) |
| 305 | // If the register is larger than the current fragment stop |
| 306 | // once the fragment is covered. |
| 307 | break; |
| 308 | addOpPiece(SizeInBits: Reg.SubRegSize); |
| 309 | } |
| 310 | |
| 311 | if (isEntryValue()) { |
| 312 | finalizeEntryValue(); |
| 313 | |
| 314 | if (!isIndirect() && !isParameterValue() && !HasComplexExpression && |
| 315 | DwarfVersion >= 4) |
| 316 | emitOp(Op: dwarf::DW_OP_stack_value); |
| 317 | } |
| 318 | |
| 319 | DwarfRegs.clear(); |
| 320 | // If we need to mask out a subregister, do it now, unless the next |
| 321 | // operation would emit an OpPiece anyway. |
| 322 | auto NextOp = ExprCursor.peek(); |
| 323 | if (SubRegisterSizeInBits && NextOp && |
| 324 | (NextOp->getOp() != dwarf::DW_OP_LLVM_fragment)) |
| 325 | maskSubRegister(); |
| 326 | return true; |
| 327 | } |
| 328 | |
| 329 | // Don't emit locations that cannot be expressed without DW_OP_stack_value. |
| 330 | if (DwarfVersion < 4) |
| 331 | if (any_of(Range&: ExprCursor, P: [](DIExpression::ExprOperand Op) -> bool { |
| 332 | return Op.getOp() == dwarf::DW_OP_stack_value; |
| 333 | })) { |
| 334 | DwarfRegs.clear(); |
| 335 | LocationKind = Unknown; |
| 336 | return false; |
| 337 | } |
| 338 | |
| 339 | // TODO: We should not give up here but the following code needs to be changed |
| 340 | // to deal with multiple (sub)registers first. |
| 341 | if (DwarfRegs.size() > 1) { |
| 342 | LLVM_DEBUG(dbgs() << "TODO: giving up on debug information due to " |
| 343 | "multi-register usage.\n" ); |
| 344 | DwarfRegs.clear(); |
| 345 | LocationKind = Unknown; |
| 346 | return false; |
| 347 | } |
| 348 | |
| 349 | auto Reg = DwarfRegs[0]; |
| 350 | bool FBReg = isFrameRegister(TRI, MachineReg); |
| 351 | int SignedOffset = 0; |
| 352 | assert(!Reg.isSubRegister() && "full register expected" ); |
| 353 | |
| 354 | // Pattern-match combinations for which more efficient representations exist. |
| 355 | // [Reg, DW_OP_plus_uconst, Offset] --> [DW_OP_breg, Offset]. |
| 356 | if (Op && (Op->getOp() == dwarf::DW_OP_plus_uconst)) { |
| 357 | uint64_t Offset = Op->getArg(I: 0); |
| 358 | uint64_t IntMax = static_cast<uint64_t>(std::numeric_limits<int>::max()); |
| 359 | if (Offset <= IntMax) { |
| 360 | SignedOffset = Offset; |
| 361 | ExprCursor.take(); |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | // [Reg, DW_OP_constu, Offset, DW_OP_plus] --> [DW_OP_breg, Offset] |
| 366 | // [Reg, DW_OP_constu, Offset, DW_OP_minus] --> [DW_OP_breg,-Offset] |
| 367 | // If Reg is a subregister we need to mask it out before subtracting. |
| 368 | if (Op && Op->getOp() == dwarf::DW_OP_constu) { |
| 369 | uint64_t Offset = Op->getArg(I: 0); |
| 370 | uint64_t IntMax = static_cast<uint64_t>(std::numeric_limits<int>::max()); |
| 371 | auto N = ExprCursor.peekNext(); |
| 372 | if (N && N->getOp() == dwarf::DW_OP_plus && Offset <= IntMax) { |
| 373 | SignedOffset = Offset; |
| 374 | ExprCursor.consume(N: 2); |
| 375 | } else if (N && N->getOp() == dwarf::DW_OP_minus && |
| 376 | !SubRegisterSizeInBits && Offset <= IntMax + 1) { |
| 377 | SignedOffset = -static_cast<int64_t>(Offset); |
| 378 | ExprCursor.consume(N: 2); |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | if (FBReg) |
| 383 | addFBReg(Offset: SignedOffset); |
| 384 | else |
| 385 | addBReg(DwarfReg: Reg.DwarfRegNo, Offset: SignedOffset); |
| 386 | DwarfRegs.clear(); |
| 387 | |
| 388 | // If we need to mask out a subregister, do it now, unless the next |
| 389 | // operation would emit an OpPiece anyway. |
| 390 | auto NextOp = ExprCursor.peek(); |
| 391 | if (SubRegisterSizeInBits && NextOp && |
| 392 | (NextOp->getOp() != dwarf::DW_OP_LLVM_fragment)) |
| 393 | maskSubRegister(); |
| 394 | |
| 395 | return true; |
| 396 | } |
| 397 | |
| 398 | void DwarfExpression::setEntryValueFlags(const MachineLocation &Loc) { |
| 399 | LocationFlags |= EntryValue; |
| 400 | if (Loc.isIndirect()) |
| 401 | LocationFlags |= Indirect; |
| 402 | } |
| 403 | |
| 404 | void DwarfExpression::setLocation(const MachineLocation &Loc, |
| 405 | const DIExpression *DIExpr) { |
| 406 | if (Loc.isIndirect()) |
| 407 | setMemoryLocationKind(); |
| 408 | |
| 409 | if (DIExpr->isEntryValue()) |
| 410 | setEntryValueFlags(Loc); |
| 411 | } |
| 412 | |
| 413 | void DwarfExpression::beginEntryValueExpression( |
| 414 | DIExpressionCursor &ExprCursor) { |
| 415 | auto Op = ExprCursor.take(); |
| 416 | (void)Op; |
| 417 | assert(Op && Op->getOp() == dwarf::DW_OP_LLVM_entry_value); |
| 418 | assert(!IsEmittingEntryValue && "Already emitting entry value?" ); |
| 419 | assert(Op->getArg(0) == 1 && |
| 420 | "Can currently only emit entry values covering a single operation" ); |
| 421 | |
| 422 | SavedLocationKind = LocationKind; |
| 423 | LocationKind = Register; |
| 424 | LocationFlags |= EntryValue; |
| 425 | IsEmittingEntryValue = true; |
| 426 | enableTemporaryBuffer(); |
| 427 | } |
| 428 | |
| 429 | void DwarfExpression::finalizeEntryValue() { |
| 430 | assert(IsEmittingEntryValue && "Entry value not open?" ); |
| 431 | disableTemporaryBuffer(); |
| 432 | |
| 433 | emitOp(Op: CU.getDwarf5OrGNULocationAtom(Loc: dwarf::DW_OP_entry_value)); |
| 434 | |
| 435 | // Emit the entry value's size operand. |
| 436 | unsigned Size = getTemporaryBufferSize(); |
| 437 | emitUnsigned(Value: Size); |
| 438 | |
| 439 | // Emit the entry value's DWARF block operand. |
| 440 | commitTemporaryBuffer(); |
| 441 | |
| 442 | LocationFlags &= ~EntryValue; |
| 443 | LocationKind = SavedLocationKind; |
| 444 | IsEmittingEntryValue = false; |
| 445 | } |
| 446 | |
| 447 | void DwarfExpression::cancelEntryValue() { |
| 448 | assert(IsEmittingEntryValue && "Entry value not open?" ); |
| 449 | disableTemporaryBuffer(); |
| 450 | |
| 451 | // The temporary buffer can't be emptied, so for now just assert that nothing |
| 452 | // has been emitted to it. |
| 453 | assert(getTemporaryBufferSize() == 0 && |
| 454 | "Began emitting entry value block before cancelling entry value" ); |
| 455 | |
| 456 | LocationKind = SavedLocationKind; |
| 457 | IsEmittingEntryValue = false; |
| 458 | } |
| 459 | |
| 460 | unsigned DwarfExpression::getOrCreateBaseType(unsigned BitSize, |
| 461 | dwarf::TypeKind Encoding) { |
| 462 | // Reuse the base_type if we already have one in this CU otherwise we |
| 463 | // create a new one. |
| 464 | unsigned I = 0, E = CU.ExprRefedBaseTypes.size(); |
| 465 | for (; I != E; ++I) |
| 466 | if (CU.ExprRefedBaseTypes[I].BitSize == BitSize && |
| 467 | CU.ExprRefedBaseTypes[I].Encoding == Encoding) |
| 468 | break; |
| 469 | |
| 470 | if (I == E) |
| 471 | CU.ExprRefedBaseTypes.emplace_back(args&: BitSize, args&: Encoding); |
| 472 | return I; |
| 473 | } |
| 474 | |
| 475 | /// Assuming a well-formed expression, match "DW_OP_deref* |
| 476 | /// DW_OP_LLVM_fragment?". |
| 477 | static bool isMemoryLocation(DIExpressionCursor ExprCursor) { |
| 478 | while (ExprCursor) { |
| 479 | auto Op = ExprCursor.take(); |
| 480 | switch (Op->getOp()) { |
| 481 | case dwarf::DW_OP_deref: |
| 482 | case dwarf::DW_OP_LLVM_fragment: |
| 483 | break; |
| 484 | default: |
| 485 | return false; |
| 486 | } |
| 487 | } |
| 488 | return true; |
| 489 | } |
| 490 | |
| 491 | void DwarfExpression::addExpression(DIExpressionCursor &&ExprCursor) { |
| 492 | addExpression(Expr: std::move(ExprCursor), |
| 493 | InsertArg: [](unsigned Idx, DIExpressionCursor &Cursor) -> bool { |
| 494 | llvm_unreachable("unhandled opcode found in expression" ); |
| 495 | }); |
| 496 | } |
| 497 | |
| 498 | bool DwarfExpression::addExpression( |
| 499 | DIExpressionCursor &&ExprCursor, |
| 500 | llvm::function_ref<bool(unsigned, DIExpressionCursor &)> InsertArg) { |
| 501 | // Entry values can currently only cover the initial register location, |
| 502 | // and not any other parts of the following DWARF expression. |
| 503 | assert(!IsEmittingEntryValue && "Can't emit entry value around expression" ); |
| 504 | |
| 505 | std::optional<DIExpression::ExprOperand> PrevConvertOp; |
| 506 | |
| 507 | while (ExprCursor) { |
| 508 | auto Op = ExprCursor.take(); |
| 509 | uint64_t OpNum = Op->getOp(); |
| 510 | |
| 511 | if (OpNum >= dwarf::DW_OP_reg0 && OpNum <= dwarf::DW_OP_reg31) { |
| 512 | emitOp(Op: OpNum); |
| 513 | continue; |
| 514 | } else if (OpNum >= dwarf::DW_OP_breg0 && OpNum <= dwarf::DW_OP_breg31) { |
| 515 | addBReg(DwarfReg: OpNum - dwarf::DW_OP_breg0, Offset: Op->getArg(I: 0)); |
| 516 | continue; |
| 517 | } |
| 518 | |
| 519 | switch (OpNum) { |
| 520 | case dwarf::DW_OP_LLVM_arg: |
| 521 | if (!InsertArg(Op->getArg(I: 0), ExprCursor)) { |
| 522 | LocationKind = Unknown; |
| 523 | return false; |
| 524 | } |
| 525 | break; |
| 526 | case dwarf::DW_OP_LLVM_fragment: { |
| 527 | unsigned SizeInBits = Op->getArg(I: 1); |
| 528 | unsigned FragmentOffset = Op->getArg(I: 0); |
| 529 | // The fragment offset must have already been adjusted by emitting an |
| 530 | // empty DW_OP_piece / DW_OP_bit_piece before we emitted the base |
| 531 | // location. |
| 532 | assert(OffsetInBits >= FragmentOffset && "fragment offset not added?" ); |
| 533 | assert(SizeInBits >= OffsetInBits - FragmentOffset && "size underflow" ); |
| 534 | |
| 535 | // If addMachineReg already emitted DW_OP_piece operations to represent |
| 536 | // a super-register by splicing together sub-registers, subtract the size |
| 537 | // of the pieces that was already emitted. |
| 538 | SizeInBits -= OffsetInBits - FragmentOffset; |
| 539 | |
| 540 | // If addMachineReg requested a DW_OP_bit_piece to stencil out a |
| 541 | // sub-register that is smaller than the current fragment's size, use it. |
| 542 | if (SubRegisterSizeInBits) |
| 543 | SizeInBits = std::min<unsigned>(a: SizeInBits, b: SubRegisterSizeInBits); |
| 544 | |
| 545 | // Emit a DW_OP_stack_value for implicit location descriptions. |
| 546 | if (isImplicitLocation()) |
| 547 | addStackValue(); |
| 548 | |
| 549 | // Emit the DW_OP_piece. |
| 550 | addOpPiece(SizeInBits, OffsetInBits: SubRegisterOffsetInBits); |
| 551 | setSubRegisterPiece(SizeInBits: 0, OffsetInBits: 0); |
| 552 | // Reset the location description kind. |
| 553 | LocationKind = Unknown; |
| 554 | return true; |
| 555 | } |
| 556 | case dwarf::DW_OP_LLVM_extract_bits_sext: |
| 557 | case dwarf::DW_OP_LLVM_extract_bits_zext: { |
| 558 | unsigned SizeInBits = Op->getArg(I: 1); |
| 559 | unsigned BitOffset = Op->getArg(I: 0); |
| 560 | |
| 561 | // If we have a memory location then dereference to get the value, though |
| 562 | // we have to make sure we don't dereference any bytes past the end of the |
| 563 | // object. |
| 564 | if (isMemoryLocation()) { |
| 565 | emitOp(Op: dwarf::DW_OP_deref_size); |
| 566 | emitUnsigned(Value: alignTo(Value: BitOffset + SizeInBits, Align: 8) / 8); |
| 567 | } |
| 568 | |
| 569 | // Extract the bits by a shift left (to shift out the bits after what we |
| 570 | // want to extract) followed by shift right (to shift the bits to position |
| 571 | // 0 and also sign/zero extend). These operations are done in the DWARF |
| 572 | // "generic type" whose size is the size of a pointer. |
| 573 | unsigned PtrSizeInBytes = CU.getAsmPrinter()->MAI->getCodePointerSize(); |
| 574 | unsigned LeftShift = PtrSizeInBytes * 8 - (SizeInBits + BitOffset); |
| 575 | unsigned RightShift = LeftShift + BitOffset; |
| 576 | if (LeftShift) { |
| 577 | emitOp(Op: dwarf::DW_OP_constu); |
| 578 | emitUnsigned(Value: LeftShift); |
| 579 | emitOp(Op: dwarf::DW_OP_shl); |
| 580 | } |
| 581 | emitOp(Op: dwarf::DW_OP_constu); |
| 582 | emitUnsigned(Value: RightShift); |
| 583 | emitOp(Op: OpNum == dwarf::DW_OP_LLVM_extract_bits_sext ? dwarf::DW_OP_shra |
| 584 | : dwarf::DW_OP_shr); |
| 585 | |
| 586 | // The value is now at the top of the stack, so set the location to |
| 587 | // implicit so that we get a stack_value at the end. |
| 588 | LocationKind = Implicit; |
| 589 | break; |
| 590 | } |
| 591 | case dwarf::DW_OP_plus_uconst: |
| 592 | assert(!isRegisterLocation()); |
| 593 | emitOp(Op: dwarf::DW_OP_plus_uconst); |
| 594 | emitUnsigned(Value: Op->getArg(I: 0)); |
| 595 | break; |
| 596 | case dwarf::DW_OP_plus: |
| 597 | case dwarf::DW_OP_minus: |
| 598 | case dwarf::DW_OP_mul: |
| 599 | case dwarf::DW_OP_div: |
| 600 | case dwarf::DW_OP_mod: |
| 601 | case dwarf::DW_OP_or: |
| 602 | case dwarf::DW_OP_and: |
| 603 | case dwarf::DW_OP_xor: |
| 604 | case dwarf::DW_OP_shl: |
| 605 | case dwarf::DW_OP_shr: |
| 606 | case dwarf::DW_OP_shra: |
| 607 | case dwarf::DW_OP_lit0: |
| 608 | case dwarf::DW_OP_not: |
| 609 | case dwarf::DW_OP_dup: |
| 610 | case dwarf::DW_OP_push_object_address: |
| 611 | case dwarf::DW_OP_over: |
| 612 | case dwarf::DW_OP_eq: |
| 613 | case dwarf::DW_OP_ne: |
| 614 | case dwarf::DW_OP_gt: |
| 615 | case dwarf::DW_OP_ge: |
| 616 | case dwarf::DW_OP_lt: |
| 617 | case dwarf::DW_OP_le: |
| 618 | emitOp(Op: OpNum); |
| 619 | break; |
| 620 | case dwarf::DW_OP_deref: |
| 621 | assert(!isRegisterLocation()); |
| 622 | if (!isMemoryLocation() && ::isMemoryLocation(ExprCursor)) |
| 623 | // Turning this into a memory location description makes the deref |
| 624 | // implicit. |
| 625 | LocationKind = Memory; |
| 626 | else |
| 627 | emitOp(Op: dwarf::DW_OP_deref); |
| 628 | break; |
| 629 | case dwarf::DW_OP_constu: |
| 630 | assert(!isRegisterLocation()); |
| 631 | emitConstu(Value: Op->getArg(I: 0)); |
| 632 | break; |
| 633 | case dwarf::DW_OP_consts: |
| 634 | assert(!isRegisterLocation()); |
| 635 | emitOp(Op: dwarf::DW_OP_consts); |
| 636 | emitSigned(Value: Op->getArg(I: 0)); |
| 637 | break; |
| 638 | case dwarf::DW_OP_LLVM_convert: { |
| 639 | unsigned BitSize = Op->getArg(I: 0); |
| 640 | dwarf::TypeKind Encoding = static_cast<dwarf::TypeKind>(Op->getArg(I: 1)); |
| 641 | if (DwarfVersion >= 5 && CU.getDwarfDebug().useOpConvert()) { |
| 642 | emitOp(Op: dwarf::DW_OP_convert); |
| 643 | // If targeting a location-list; simply emit the index into the raw |
| 644 | // byte stream as ULEB128, DwarfDebug::emitDebugLocEntry has been |
| 645 | // fitted with means to extract it later. |
| 646 | // If targeting a inlined DW_AT_location; insert a DIEBaseTypeRef |
| 647 | // (containing the index and a resolve mechanism during emit) into the |
| 648 | // DIE value list. |
| 649 | emitBaseTypeRef(Idx: getOrCreateBaseType(BitSize, Encoding)); |
| 650 | } else { |
| 651 | if (PrevConvertOp && PrevConvertOp->getArg(I: 0) < BitSize) { |
| 652 | if (Encoding == dwarf::DW_ATE_signed) |
| 653 | emitLegacySExt(FromBits: PrevConvertOp->getArg(I: 0)); |
| 654 | else if (Encoding == dwarf::DW_ATE_unsigned) |
| 655 | emitLegacyZExt(FromBits: PrevConvertOp->getArg(I: 0)); |
| 656 | PrevConvertOp = std::nullopt; |
| 657 | } else { |
| 658 | PrevConvertOp = Op; |
| 659 | } |
| 660 | } |
| 661 | break; |
| 662 | } |
| 663 | case dwarf::DW_OP_stack_value: |
| 664 | LocationKind = Implicit; |
| 665 | break; |
| 666 | case dwarf::DW_OP_swap: |
| 667 | assert(!isRegisterLocation()); |
| 668 | emitOp(Op: dwarf::DW_OP_swap); |
| 669 | break; |
| 670 | case dwarf::DW_OP_xderef: |
| 671 | assert(!isRegisterLocation()); |
| 672 | emitOp(Op: dwarf::DW_OP_xderef); |
| 673 | break; |
| 674 | case dwarf::DW_OP_deref_size: |
| 675 | emitOp(Op: dwarf::DW_OP_deref_size); |
| 676 | emitData1(Value: Op->getArg(I: 0)); |
| 677 | break; |
| 678 | case dwarf::DW_OP_LLVM_tag_offset: |
| 679 | TagOffset = Op->getArg(I: 0); |
| 680 | break; |
| 681 | case dwarf::DW_OP_regx: |
| 682 | emitOp(Op: dwarf::DW_OP_regx); |
| 683 | emitUnsigned(Value: Op->getArg(I: 0)); |
| 684 | break; |
| 685 | case dwarf::DW_OP_bregx: |
| 686 | emitOp(Op: dwarf::DW_OP_bregx); |
| 687 | emitUnsigned(Value: Op->getArg(I: 0)); |
| 688 | emitSigned(Value: Op->getArg(I: 1)); |
| 689 | break; |
| 690 | default: |
| 691 | llvm_unreachable("unhandled opcode found in expression" ); |
| 692 | } |
| 693 | } |
| 694 | |
| 695 | if (isImplicitLocation() && !isParameterValue()) |
| 696 | // Turn this into an implicit location description. |
| 697 | addStackValue(); |
| 698 | |
| 699 | return true; |
| 700 | } |
| 701 | |
| 702 | /// add masking operations to stencil out a subregister. |
| 703 | void DwarfExpression::maskSubRegister() { |
| 704 | assert(SubRegisterSizeInBits && "no subregister was registered" ); |
| 705 | if (SubRegisterOffsetInBits > 0) |
| 706 | addShr(ShiftBy: SubRegisterOffsetInBits); |
| 707 | uint64_t Mask = (1ULL << (uint64_t)SubRegisterSizeInBits) - 1ULL; |
| 708 | addAnd(Mask); |
| 709 | } |
| 710 | |
| 711 | void DwarfExpression::finalize() { |
| 712 | assert(DwarfRegs.size() == 0 && "dwarf registers not emitted" ); |
| 713 | // Emit any outstanding DW_OP_piece operations to mask out subregisters. |
| 714 | if (SubRegisterSizeInBits == 0) |
| 715 | return; |
| 716 | // Don't emit a DW_OP_piece for a subregister at offset 0. |
| 717 | if (SubRegisterOffsetInBits == 0) |
| 718 | return; |
| 719 | addOpPiece(SizeInBits: SubRegisterSizeInBits, OffsetInBits: SubRegisterOffsetInBits); |
| 720 | } |
| 721 | |
| 722 | void DwarfExpression::addFragmentOffset(const DIExpression *Expr) { |
| 723 | if (!Expr || !Expr->isFragment()) |
| 724 | return; |
| 725 | |
| 726 | uint64_t FragmentOffset = Expr->getFragmentInfo()->OffsetInBits; |
| 727 | assert(FragmentOffset >= OffsetInBits && |
| 728 | "overlapping or duplicate fragments" ); |
| 729 | if (FragmentOffset > OffsetInBits) |
| 730 | addOpPiece(SizeInBits: FragmentOffset - OffsetInBits); |
| 731 | OffsetInBits = FragmentOffset; |
| 732 | } |
| 733 | |
| 734 | void DwarfExpression::emitLegacySExt(unsigned FromBits) { |
| 735 | // (((X >> (FromBits - 1)) * (~0)) << FromBits) | X |
| 736 | emitOp(Op: dwarf::DW_OP_dup); |
| 737 | emitOp(Op: dwarf::DW_OP_constu); |
| 738 | emitUnsigned(Value: FromBits - 1); |
| 739 | emitOp(Op: dwarf::DW_OP_shr); |
| 740 | emitOp(Op: dwarf::DW_OP_lit0); |
| 741 | emitOp(Op: dwarf::DW_OP_not); |
| 742 | emitOp(Op: dwarf::DW_OP_mul); |
| 743 | emitOp(Op: dwarf::DW_OP_constu); |
| 744 | emitUnsigned(Value: FromBits); |
| 745 | emitOp(Op: dwarf::DW_OP_shl); |
| 746 | emitOp(Op: dwarf::DW_OP_or); |
| 747 | } |
| 748 | |
| 749 | void DwarfExpression::emitLegacyZExt(unsigned FromBits) { |
| 750 | // Heuristic to decide the most efficient encoding. |
| 751 | // A ULEB can encode 7 1-bits per byte. |
| 752 | if (FromBits / 7 < 1+1+1+1+1) { |
| 753 | // (X & (1 << FromBits - 1)) |
| 754 | emitOp(Op: dwarf::DW_OP_constu); |
| 755 | emitUnsigned(Value: (1ULL << FromBits) - 1); |
| 756 | } else { |
| 757 | // Note that the DWARF 4 stack consists of pointer-sized elements, |
| 758 | // so technically it doesn't make sense to shift left more than 64 |
| 759 | // bits. We leave that for the consumer to decide though. LLDB for |
| 760 | // example uses APInt for the stack elements and can still deal |
| 761 | // with this. |
| 762 | emitOp(Op: dwarf::DW_OP_lit1); |
| 763 | emitOp(Op: dwarf::DW_OP_constu); |
| 764 | emitUnsigned(Value: FromBits); |
| 765 | emitOp(Op: dwarf::DW_OP_shl); |
| 766 | emitOp(Op: dwarf::DW_OP_lit1); |
| 767 | emitOp(Op: dwarf::DW_OP_minus); |
| 768 | } |
| 769 | emitOp(Op: dwarf::DW_OP_and); |
| 770 | } |
| 771 | |
| 772 | void DwarfExpression::addWasmLocation(unsigned Index, uint64_t Offset) { |
| 773 | emitOp(Op: dwarf::DW_OP_WASM_location); |
| 774 | emitUnsigned(Value: Index == 4/*TI_LOCAL_INDIRECT*/ ? 0/*TI_LOCAL*/ : Index); |
| 775 | emitUnsigned(Value: Offset); |
| 776 | if (Index == 4 /*TI_LOCAL_INDIRECT*/) { |
| 777 | assert(LocationKind == Unknown); |
| 778 | LocationKind = Memory; |
| 779 | } else { |
| 780 | assert(LocationKind == Implicit || LocationKind == Unknown); |
| 781 | LocationKind = Implicit; |
| 782 | } |
| 783 | } |
| 784 | |