xref: /llvm-project/bolt/lib/Passes/AsmDump.cpp (revision 867faeec054abb4c035673189c1169fef45f54c8) 
1 //===- bolt/Passes/AsmDump.cpp - Dump BinaryFunction into assembly --------===//
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 implements the AsmDumpPass class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "bolt/Passes/AsmDump.h"
14 #include "llvm/CodeGen/AsmPrinter.h"
15 #include "llvm/MC/TargetRegistry.h"
16 #include "llvm/Support/FileSystem.h"
17 #include "llvm/Support/Path.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include <unordered_set>
20 
21 #define DEBUG_TYPE "asm-dump"
22 
23 using namespace llvm;
24 
25 namespace opts {
26 extern bool shouldPrint(const bolt::BinaryFunction &Function);
27 extern cl::OptionCategory BoltCategory;
28 extern cl::opt<unsigned> Verbosity;
29 
30 cl::opt<std::string> AsmDump("asm-dump",
31                              cl::desc("dump function into assembly"),
32                              cl::value_desc("dump folder"), cl::ValueOptional,
33                              cl::Hidden, cl::cat(BoltCategory));
34 } // end namespace opts
35 
36 namespace llvm {
37 namespace bolt {
38 
39 void dumpCFI(const BinaryFunction &BF, const MCInst &Instr, AsmPrinter &MAP) {
40   const MCCFIInstruction *CFIInstr = BF.getCFIFor(Instr);
41   switch (CFIInstr->getOperation()) {
42   // Skip unsupported CFI instructions.
43   case MCCFIInstruction::OpRememberState:
44   case MCCFIInstruction::OpRestoreState:
45     if (opts::Verbosity >= 2)
46       BF.getBinaryContext().errs()
47           << "BOLT-WARNING: AsmDump: skipping unsupported CFI instruction in "
48           << BF << ".\n";
49 
50     return;
51 
52   default:
53     // Emit regular CFI instructions.
54     MAP.emitCFIInstruction(*CFIInstr);
55   }
56 }
57 
58 void dumpTargetFunctionStub(raw_ostream &OS, const BinaryContext &BC,
59                             const MCSymbol *CalleeSymb,
60                             const BinarySection *&LastCS) {
61   const BinaryFunction *CalleeFunc = BC.getFunctionForSymbol(CalleeSymb);
62   if (!CalleeFunc || CalleeFunc->isPLTFunction())
63     return;
64 
65   if (CalleeFunc->getOriginSection() != LastCS) {
66     OS << ".section " << CalleeFunc->getOriginSectionName() << '\n';
67     LastCS = CalleeFunc->getOriginSection();
68   }
69   StringRef CalleeName = CalleeFunc->getOneName();
70   OS << ".set \"" << CalleeName << "\", 0\n";
71 }
72 
73 void dumpJumpTableSymbols(raw_ostream &OS, const JumpTable *JT, AsmPrinter &MAP,
74                           const BinarySection *&LastBS) {
75   if (&JT->getSection() != LastBS) {
76     OS << ".section " << JT->getSectionName() << '\n';
77     LastBS = &JT->getSection();
78   }
79   OS << "\"" << JT->getName() << "\":\n";
80   for (MCSymbol *JTEntry : JT->Entries)
81     MAP.OutStreamer->emitSymbolValue(JTEntry, JT->OutputEntrySize);
82   OS << '\n';
83 }
84 
85 void dumpBinaryDataSymbols(raw_ostream &OS, const BinaryData *BD,
86                            const BinarySection *&LastBS) {
87   if (BD->isJumpTable())
88     return;
89   if (&BD->getSection() != LastBS) {
90     OS << ".section " << BD->getSectionName() << '\n';
91     LastBS = &BD->getSection();
92   }
93   OS << "\"" << BD->getName() << "\": ";
94   OS << '\n';
95 }
96 
97 void dumpFunction(const BinaryFunction &BF) {
98   const BinaryContext &BC = BF.getBinaryContext();
99   if (!opts::shouldPrint(BF))
100     return;
101 
102   // Make sure the new directory exists, creating it if necessary.
103   if (!opts::AsmDump.empty()) {
104     if (std::error_code EC = sys::fs::create_directories(opts::AsmDump)) {
105       BC.errs() << "BOLT-ERROR: could not create directory '" << opts::AsmDump
106                 << "': " << EC.message() << '\n';
107       return;
108     }
109   }
110 
111   std::string PrintName = BF.getPrintName();
112   std::replace(PrintName.begin(), PrintName.end(), '/', '-');
113   std::string Filename =
114       opts::AsmDump.empty()
115           ? (PrintName + ".s")
116           : (opts::AsmDump + sys::path::get_separator() + PrintName + ".s")
117                 .str();
118   BC.outs() << "BOLT-INFO: Dumping function assembly to " << Filename << "\n";
119 
120   std::error_code EC;
121   raw_fd_ostream OS(Filename, EC, sys::fs::OF_None);
122   if (EC) {
123     BC.errs() << "BOLT-ERROR: " << EC.message() << ", unable to open "
124               << Filename << " for output.\n";
125     return;
126   }
127   OS.SetUnbuffered();
128 
129   // Create local MC context to isolate the effect of ephemeral assembly
130   // emission.
131   BinaryContext::IndependentCodeEmitter MCEInstance =
132       BC.createIndependentMCCodeEmitter();
133   MCContext *LocalCtx = MCEInstance.LocalCtx.get();
134   std::unique_ptr<MCAsmBackend> MAB(
135       BC.TheTarget->createMCAsmBackend(*BC.STI, *BC.MRI, MCTargetOptions()));
136   int AsmPrinterVariant = BC.AsmInfo->getAssemblerDialect();
137   MCInstPrinter *InstructionPrinter(BC.TheTarget->createMCInstPrinter(
138       *BC.TheTriple, AsmPrinterVariant, *BC.AsmInfo, *BC.MII, *BC.MRI));
139   auto FOut = std::make_unique<formatted_raw_ostream>(OS);
140   FOut->SetUnbuffered();
141   std::unique_ptr<MCStreamer> AsmStreamer(
142       createAsmStreamer(*LocalCtx, std::move(FOut), InstructionPrinter,
143                         std::move(MCEInstance.MCE), std::move(MAB)));
144   AsmStreamer->initSections(true, *BC.STI);
145   std::unique_ptr<TargetMachine> TM(BC.TheTarget->createTargetMachine(
146       BC.TripleName, "", "", TargetOptions(), std::nullopt));
147   std::unique_ptr<AsmPrinter> MAP(
148       BC.TheTarget->createAsmPrinter(*TM, std::move(AsmStreamer)));
149 
150   StringRef FunctionName = BF.getOneName();
151   OS << "  .globl " << FunctionName << '\n';
152   OS << "  .type " << FunctionName << ", %function\n";
153   OS << FunctionName << ":\n";
154 
155   // FDATA for the entry point
156   if (uint64_t EntryExecCount = BF.getKnownExecutionCount())
157     OS << "# FDATA: 0 [unknown] 0 "
158        << "1 " << FunctionName << " 0 "
159        << "0 " << EntryExecCount << '\n';
160 
161   // Binary data references from the function.
162   std::unordered_set<const BinaryData *> BDReferences;
163   // Function references from the function (to avoid constructing call graph).
164   std::unordered_set<const MCSymbol *> CallReferences;
165 
166   MAP->OutStreamer->emitCFIStartProc(/*IsSimple=*/false);
167   for (const BinaryBasicBlock *BB : BF.getLayout().blocks()) {
168     OS << BB->getName() << ": \n";
169 
170     const std::string BranchLabel = Twine(BB->getName(), "_br").str();
171     const MCInst *LastInst = BB->getLastNonPseudoInstr();
172 
173     for (const MCInst &Instr : *BB) {
174       // Dump pseudo instructions (CFI)
175       if (BC.MIB->isPseudo(Instr)) {
176         if (BC.MIB->isCFI(Instr))
177           dumpCFI(BF, Instr, *MAP.get());
178         continue;
179       }
180 
181       // Analyze symbol references (data, functions) from the instruction.
182       bool IsCall = BC.MIB->isCall(Instr);
183       for (const MCOperand &Operand : MCPlus::primeOperands(Instr)) {
184         if (Operand.isExpr() &&
185             Operand.getExpr()->getKind() == MCExpr::SymbolRef) {
186           std::pair<const MCSymbol *, uint64_t> TSI =
187               BC.MIB->getTargetSymbolInfo(Operand.getExpr());
188           const MCSymbol *Symbol = TSI.first;
189           if (IsCall)
190             CallReferences.insert(Symbol);
191           else if (const BinaryData *BD =
192                        BC.getBinaryDataByName(Symbol->getName()))
193             BDReferences.insert(BD);
194         }
195       }
196 
197       if (&Instr == LastInst && (BB->succ_size() || IsCall))
198         OS << BranchLabel << ":\n";
199 
200       BC.InstPrinter->printInst(&Instr, 0, "", *BC.STI, OS);
201       OS << '\n';
202     }
203 
204     // Dump profile data in FDATA format (as parsed by link_fdata).
205     for (const BinaryBasicBlock *Succ : BB->successors()) {
206       const BinaryBasicBlock::BinaryBranchInfo BI = BB->getBranchInfo(*Succ);
207       if (!BI.MispredictedCount && !BI.Count)
208         continue;
209 
210       OS << "# FDATA: 1 " << FunctionName << " #" << BranchLabel << "# "
211          << "1 " << FunctionName << " #" << Succ->getName() << "# "
212          << BI.MispredictedCount << " " << BI.Count << '\n';
213     }
214 
215     OS << '\n';
216   }
217   MAP->OutStreamer->emitCFIEndProc();
218 
219   OS << ".size " << FunctionName << ", .-" << FunctionName << '\n';
220 
221   const BinarySection *LastSection = BF.getOriginSection();
222   // Print stubs for all target functions.
223   for (const MCSymbol *CalleeSymb : CallReferences)
224     dumpTargetFunctionStub(OS, BC, CalleeSymb, LastSection);
225 
226   OS << "# Jump tables\n";
227   // Print all jump tables.
228   for (auto &JTI : BF.jumpTables())
229     dumpJumpTableSymbols(OS, JTI.second, *MAP.get(), LastSection);
230 
231   OS << "# BinaryData\n";
232   // Print data references.
233   for (const BinaryData *BD : BDReferences)
234     dumpBinaryDataSymbols(OS, BD, LastSection);
235 }
236 
237 Error AsmDumpPass::runOnFunctions(BinaryContext &BC) {
238   for (const auto &BFIt : BC.getBinaryFunctions())
239     dumpFunction(BFIt.second);
240   return Error::success();
241 }
242 
243 } // namespace bolt
244 } // namespace llvm
245