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