1//===-- llvm/lib/CodeGen/AsmPrinter/DebugHandlerBase.cpp -------*- C++ -*--===//
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// Common functionality for different debug information format backends.
10// LLVM currently supports DWARF and CodeView.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/CodeGen/DebugHandlerBase.h"
15#include "llvm/CodeGen/AsmPrinter.h"
16#include "llvm/CodeGen/MachineFunction.h"
17#include "llvm/CodeGen/MachineInstr.h"
18#include "llvm/CodeGen/MachineModuleInfo.h"
19#include "llvm/CodeGen/TargetSubtargetInfo.h"
20#include "llvm/IR/DebugInfo.h"
21#include "llvm/IR/Module.h"
22#include "llvm/MC/MCStreamer.h"
23#include "llvm/Support/CommandLine.h"
24
25using namespace llvm;
26
27#define DEBUG_TYPE "dwarfdebug"
28
29/// If true, we drop variable location ranges which exist entirely outside the
30/// variable's lexical scope instruction ranges.
31static cl::opt<bool> TrimVarLocs("trim-var-locs", cl::Hidden, cl::init(Val: true));
32
33std::optional<DbgVariableLocation>
34DbgVariableLocation::extractFromMachineInstruction(
35 const MachineInstr &Instruction) {
36 DbgVariableLocation Location;
37 // Variables calculated from multiple locations can't be represented here.
38 if (Instruction.getNumDebugOperands() != 1)
39 return std::nullopt;
40 if (!Instruction.getDebugOperand(Index: 0).isReg())
41 return std::nullopt;
42 Location.Register = Instruction.getDebugOperand(Index: 0).getReg().asMCReg();
43 Location.FragmentInfo.reset();
44 // We only handle expressions generated by DIExpression::appendOffset,
45 // which doesn't require a full stack machine.
46 int64_t Offset = 0;
47 const DIExpression *DIExpr = Instruction.getDebugExpression();
48 auto Op = DIExpr->expr_op_begin();
49 // We can handle a DBG_VALUE_LIST iff it has exactly one location operand that
50 // appears exactly once at the start of the expression.
51 if (Instruction.isDebugValueList()) {
52 if (Instruction.getNumDebugOperands() == 1 &&
53 Op->getOp() == dwarf::DW_OP_LLVM_arg)
54 ++Op;
55 else
56 return std::nullopt;
57 }
58 while (Op != DIExpr->expr_op_end()) {
59 switch (Op->getOp()) {
60 case dwarf::DW_OP_constu: {
61 int Value = Op->getArg(I: 0);
62 ++Op;
63 if (Op != DIExpr->expr_op_end()) {
64 switch (Op->getOp()) {
65 case dwarf::DW_OP_minus:
66 Offset -= Value;
67 break;
68 case dwarf::DW_OP_plus:
69 Offset += Value;
70 break;
71 default:
72 continue;
73 }
74 }
75 } break;
76 case dwarf::DW_OP_plus_uconst:
77 Offset += Op->getArg(I: 0);
78 break;
79 case dwarf::DW_OP_LLVM_fragment:
80 Location.FragmentInfo = {Op->getArg(I: 1), Op->getArg(I: 0)};
81 break;
82 case dwarf::DW_OP_deref:
83 Location.LoadChain.push_back(Elt: Offset);
84 Offset = 0;
85 break;
86 default:
87 return std::nullopt;
88 }
89 ++Op;
90 }
91
92 // Do one final implicit DW_OP_deref if this was an indirect DBG_VALUE
93 // instruction.
94 // FIXME: Replace these with DIExpression.
95 if (Instruction.isIndirectDebugValue())
96 Location.LoadChain.push_back(Elt: Offset);
97
98 return Location;
99}
100
101DebugHandlerBase::DebugHandlerBase(AsmPrinter *A) : Asm(A), MMI(Asm->MMI) {}
102
103DebugHandlerBase::~DebugHandlerBase() = default;
104
105void DebugHandlerBase::beginModule(Module *M) {
106 if (M->debug_compile_units().empty())
107 Asm = nullptr;
108}
109
110// Each LexicalScope has first instruction and last instruction to mark
111// beginning and end of a scope respectively. Create an inverse map that list
112// scopes starts (and ends) with an instruction. One instruction may start (or
113// end) multiple scopes. Ignore scopes that are not reachable.
114void DebugHandlerBase::identifyScopeMarkers() {
115 SmallVector<LexicalScope *, 4> WorkList;
116 WorkList.push_back(Elt: LScopes.getCurrentFunctionScope());
117 while (!WorkList.empty()) {
118 LexicalScope *S = WorkList.pop_back_val();
119
120 const SmallVectorImpl<LexicalScope *> &Children = S->getChildren();
121 if (!Children.empty())
122 WorkList.append(in_start: Children.begin(), in_end: Children.end());
123
124 if (S->isAbstractScope())
125 continue;
126
127 for (const InsnRange &R : S->getRanges()) {
128 assert(R.first && "InsnRange does not have first instruction!");
129 assert(R.second && "InsnRange does not have second instruction!");
130 requestLabelBeforeInsn(MI: R.first);
131 requestLabelAfterInsn(MI: R.second);
132 }
133 }
134}
135
136// Return Label preceding the instruction.
137MCSymbol *DebugHandlerBase::getLabelBeforeInsn(const MachineInstr *MI) {
138 MCSymbol *Label = LabelsBeforeInsn.lookup(Val: MI);
139 assert(Label && "Didn't insert label before instruction");
140 return Label;
141}
142
143// Return Label immediately following the instruction.
144MCSymbol *DebugHandlerBase::getLabelAfterInsn(const MachineInstr *MI) {
145 return LabelsAfterInsn.lookup(Val: MI);
146}
147
148/// If this type is derived from a base type then return base type size.
149uint64_t DebugHandlerBase::getBaseTypeSize(const DIType *Ty) {
150 assert(Ty);
151
152 unsigned Tag = Ty->getTag();
153
154 if (Tag != dwarf::DW_TAG_member && Tag != dwarf::DW_TAG_typedef &&
155 Tag != dwarf::DW_TAG_const_type && Tag != dwarf::DW_TAG_volatile_type &&
156 Tag != dwarf::DW_TAG_restrict_type && Tag != dwarf::DW_TAG_atomic_type &&
157 Tag != dwarf::DW_TAG_immutable_type &&
158 Tag != dwarf::DW_TAG_template_alias)
159 return Ty->getSizeInBits();
160
161 DIType *BaseType = nullptr;
162 if (const DIDerivedType *DDTy = dyn_cast<DIDerivedType>(Val: Ty))
163 BaseType = DDTy->getBaseType();
164 else if (const DISubrangeType *SRTy = dyn_cast<DISubrangeType>(Val: Ty))
165 BaseType = SRTy->getBaseType();
166
167 if (!BaseType)
168 return 0;
169
170 // If this is a derived type, go ahead and get the base type, unless it's a
171 // reference then it's just the size of the field. Pointer types have no need
172 // of this since they're a different type of qualification on the type.
173 if (BaseType->getTag() == dwarf::DW_TAG_reference_type ||
174 BaseType->getTag() == dwarf::DW_TAG_rvalue_reference_type)
175 return Ty->getSizeInBits();
176
177 return getBaseTypeSize(Ty: BaseType);
178}
179
180bool DebugHandlerBase::isUnsignedDIType(const DIType *Ty) {
181 if (isa<DIStringType>(Val: Ty)) {
182 // Some transformations (e.g. instcombine) may decide to turn a Fortran
183 // character object into an integer, and later ones (e.g. SROA) may
184 // further inject a constant integer in a llvm.dbg.value call to track
185 // the object's value. Here we trust the transformations are doing the
186 // right thing, and treat the constant as unsigned to preserve that value
187 // (i.e. avoid sign extension).
188 return true;
189 }
190
191 if (auto *SRTy = dyn_cast<DISubrangeType>(Val: Ty)) {
192 Ty = SRTy->getBaseType();
193 if (!Ty)
194 return false;
195 }
196
197 if (auto *CTy = dyn_cast<DICompositeType>(Val: Ty)) {
198 if (CTy->getTag() == dwarf::DW_TAG_enumeration_type) {
199 if (!(Ty = CTy->getBaseType()))
200 // FIXME: Enums without a fixed underlying type have unknown signedness
201 // here, leading to incorrectly emitted constants.
202 return false;
203 } else
204 // (Pieces of) aggregate types that get hacked apart by SROA may be
205 // represented by a constant. Encode them as unsigned bytes.
206 return true;
207 }
208
209 if (auto *DTy = dyn_cast<DIDerivedType>(Val: Ty)) {
210 dwarf::Tag T = (dwarf::Tag)Ty->getTag();
211 // Encode pointer constants as unsigned bytes. This is used at least for
212 // null pointer constant emission.
213 // FIXME: reference and rvalue_reference /probably/ shouldn't be allowed
214 // here, but accept them for now due to a bug in SROA producing bogus
215 // dbg.values.
216 if (T == dwarf::DW_TAG_pointer_type ||
217 T == dwarf::DW_TAG_ptr_to_member_type ||
218 T == dwarf::DW_TAG_reference_type ||
219 T == dwarf::DW_TAG_rvalue_reference_type)
220 return true;
221 assert(T == dwarf::DW_TAG_typedef || T == dwarf::DW_TAG_const_type ||
222 T == dwarf::DW_TAG_volatile_type ||
223 T == dwarf::DW_TAG_restrict_type || T == dwarf::DW_TAG_atomic_type ||
224 T == dwarf::DW_TAG_immutable_type ||
225 T == dwarf::DW_TAG_template_alias);
226 assert(DTy->getBaseType() && "Expected valid base type");
227 return isUnsignedDIType(Ty: DTy->getBaseType());
228 }
229
230 auto *BTy = cast<DIBasicType>(Val: Ty);
231 unsigned Encoding = BTy->getEncoding();
232 assert((Encoding == dwarf::DW_ATE_unsigned ||
233 Encoding == dwarf::DW_ATE_unsigned_char ||
234 Encoding == dwarf::DW_ATE_signed ||
235 Encoding == dwarf::DW_ATE_signed_char ||
236 Encoding == dwarf::DW_ATE_float || Encoding == dwarf::DW_ATE_UTF ||
237 Encoding == dwarf::DW_ATE_boolean ||
238 Encoding == dwarf::DW_ATE_complex_float ||
239 Encoding == dwarf::DW_ATE_signed_fixed ||
240 Encoding == dwarf::DW_ATE_unsigned_fixed ||
241 (Ty->getTag() == dwarf::DW_TAG_unspecified_type &&
242 Ty->getName() == "decltype(nullptr)")) &&
243 "Unsupported encoding");
244 return Encoding == dwarf::DW_ATE_unsigned ||
245 Encoding == dwarf::DW_ATE_unsigned_char ||
246 Encoding == dwarf::DW_ATE_UTF || Encoding == dwarf::DW_ATE_boolean ||
247 Encoding == llvm::dwarf::DW_ATE_unsigned_fixed ||
248 Ty->getTag() == dwarf::DW_TAG_unspecified_type;
249}
250
251static bool hasDebugInfo(const MachineFunction *MF) {
252 auto *SP = MF->getFunction().getSubprogram();
253 if (!SP)
254 return false;
255 assert(SP->getUnit());
256 auto EK = SP->getUnit()->getEmissionKind();
257 if (EK == DICompileUnit::NoDebug)
258 return false;
259 return true;
260}
261
262void DebugHandlerBase::beginFunction(const MachineFunction *MF) {
263 PrevInstBB = nullptr;
264
265 if (!Asm || !hasDebugInfo(MF)) {
266 skippedNonDebugFunction();
267 return;
268 }
269
270 // Grab the lexical scopes for the function, if we don't have any of those
271 // then we're not going to be able to do anything.
272 LScopes.initialize(*MF);
273 if (LScopes.empty()) {
274 beginFunctionImpl(MF);
275 return;
276 }
277
278 // Make sure that each lexical scope will have a begin/end label.
279 identifyScopeMarkers();
280
281 // Calculate history for local variables.
282 assert(DbgValues.empty() && "DbgValues map wasn't cleaned!");
283 assert(DbgLabels.empty() && "DbgLabels map wasn't cleaned!");
284 calculateDbgEntityHistory(MF, TRI: Asm->MF->getSubtarget().getRegisterInfo(),
285 DbgValues, DbgLabels);
286 InstOrdering.initialize(MF: *MF);
287 if (TrimVarLocs)
288 DbgValues.trimLocationRanges(MF: *MF, LScopes, Ordering: InstOrdering);
289 LLVM_DEBUG(DbgValues.dump(MF->getName()));
290
291 // Request labels for the full history.
292 for (const auto &I : DbgValues) {
293 const auto &Entries = I.second;
294 if (Entries.empty())
295 continue;
296
297 auto IsDescribedByReg = [](const MachineInstr *MI) {
298 return any_of(Range: MI->debug_operands(),
299 P: [](auto &MO) { return MO.isReg() && MO.getReg(); });
300 };
301
302 // The first mention of a function argument gets the CurrentFnBegin label,
303 // so arguments are visible when breaking at function entry.
304 //
305 // We do not change the label for values that are described by registers,
306 // as that could place them above their defining instructions. We should
307 // ideally not change the labels for constant debug values either, since
308 // doing that violates the ranges that are calculated in the history map.
309 // However, we currently do not emit debug values for constant arguments
310 // directly at the start of the function, so this code is still useful.
311 const DILocalVariable *DIVar =
312 Entries.front().getInstr()->getDebugVariable();
313 if (DIVar->isParameter() &&
314 getDISubprogram(Scope: DIVar->getScope())->describes(F: &MF->getFunction())) {
315 if (!IsDescribedByReg(Entries.front().getInstr()))
316 LabelsBeforeInsn[Entries.front().getInstr()] = Asm->getFunctionBegin();
317 if (Entries.front().getInstr()->getDebugExpression()->isFragment()) {
318 // Mark all non-overlapping initial fragments.
319 for (const auto *I = Entries.begin(); I != Entries.end(); ++I) {
320 if (!I->isDbgValue())
321 continue;
322 const DIExpression *Fragment = I->getInstr()->getDebugExpression();
323 if (std::any_of(first: Entries.begin(), last: I,
324 pred: [&](DbgValueHistoryMap::Entry Pred) {
325 return Pred.isDbgValue() &&
326 Fragment->fragmentsOverlap(
327 Other: Pred.getInstr()->getDebugExpression());
328 }))
329 break;
330 // The code that generates location lists for DWARF assumes that the
331 // entries' start labels are monotonically increasing, and since we
332 // don't change the label for fragments that are described by
333 // registers, we must bail out when encountering such a fragment.
334 if (IsDescribedByReg(I->getInstr()))
335 break;
336 LabelsBeforeInsn[I->getInstr()] = Asm->getFunctionBegin();
337 }
338 }
339 }
340
341 for (const auto &Entry : Entries) {
342 if (Entry.isDbgValue())
343 requestLabelBeforeInsn(MI: Entry.getInstr());
344 else
345 requestLabelAfterInsn(MI: Entry.getInstr());
346 }
347 }
348
349 // Ensure there is a symbol before DBG_LABEL.
350 for (const auto &I : DbgLabels) {
351 const MachineInstr *MI = I.second;
352 requestLabelBeforeInsn(MI);
353 }
354
355 PrevInstLoc = DebugLoc();
356 PrevLabel = Asm->getFunctionBegin();
357 beginFunctionImpl(MF);
358}
359
360void DebugHandlerBase::beginInstruction(const MachineInstr *MI) {
361 if (!Asm || !Asm->hasDebugInfo())
362 return;
363
364 assert(CurMI == nullptr);
365 CurMI = MI;
366
367 // Insert labels where requested.
368 DenseMap<const MachineInstr *, MCSymbol *>::iterator I =
369 LabelsBeforeInsn.find(Val: MI);
370
371 // No label needed.
372 if (I == LabelsBeforeInsn.end())
373 return;
374
375 // Label already assigned.
376 if (I->second)
377 return;
378
379 if (!PrevLabel) {
380 PrevLabel = MMI->getContext().createTempSymbol();
381 Asm->OutStreamer->emitLabel(Symbol: PrevLabel);
382 }
383 I->second = PrevLabel;
384}
385
386void DebugHandlerBase::endInstruction() {
387 if (!Asm || !Asm->hasDebugInfo())
388 return;
389
390 assert(CurMI != nullptr);
391 // Don't create a new label after DBG_VALUE and other instructions that don't
392 // generate code.
393 if (!CurMI->isMetaInstruction()) {
394 PrevLabel = nullptr;
395 PrevInstBB = CurMI->getParent();
396 }
397
398 DenseMap<const MachineInstr *, MCSymbol *>::iterator I =
399 LabelsAfterInsn.find(Val: CurMI);
400
401 // No label needed or label already assigned.
402 if (I == LabelsAfterInsn.end() || I->second) {
403 CurMI = nullptr;
404 return;
405 }
406
407 // We need a label after this instruction. With basic block sections, just
408 // use the end symbol of the section if this is the last instruction of the
409 // section. This reduces the need for an additional label and also helps
410 // merging ranges.
411 if (CurMI->getParent()->isEndSection() && CurMI->getNextNode() == nullptr) {
412 PrevLabel = CurMI->getParent()->getEndSymbol();
413 } else if (!PrevLabel) {
414 PrevLabel = MMI->getContext().createTempSymbol();
415 Asm->OutStreamer->emitLabel(Symbol: PrevLabel);
416 }
417 I->second = PrevLabel;
418 CurMI = nullptr;
419}
420
421void DebugHandlerBase::endFunction(const MachineFunction *MF) {
422 if (Asm && hasDebugInfo(MF))
423 endFunctionImpl(MF);
424 DbgValues.clear();
425 DbgLabels.clear();
426 LabelsBeforeInsn.clear();
427 LabelsAfterInsn.clear();
428 InstOrdering.clear();
429}
430
431void DebugHandlerBase::beginBasicBlockSection(const MachineBasicBlock &MBB) {
432 EpilogBeginBlock = nullptr;
433 if (!MBB.isEntryBlock())
434 PrevLabel = MBB.getSymbol();
435}
436
437void DebugHandlerBase::endBasicBlockSection(const MachineBasicBlock &MBB) {
438 PrevLabel = nullptr;
439}
440