1 //===-- AsmPrinterInlineAsm.cpp - AsmPrinter Inline Asm Handling ----------===//
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 inline assembler pieces of the AsmPrinter class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/SmallString.h"
14 #include "llvm/ADT/SmallVector.h"
15 #include "llvm/ADT/Twine.h"
16 #include "llvm/CodeGen/AsmPrinter.h"
17 #include "llvm/CodeGen/MachineBasicBlock.h"
18 #include "llvm/CodeGen/MachineFunction.h"
19 #include "llvm/CodeGen/MachineModuleInfo.h"
20 #include "llvm/CodeGen/TargetRegisterInfo.h"
21 #include "llvm/CodeGen/TargetSubtargetInfo.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DiagnosticInfo.h"
25 #include "llvm/IR/InlineAsm.h"
26 #include "llvm/IR/LLVMContext.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/MC/MCAsmInfo.h"
29 #include "llvm/MC/MCInstrInfo.h"
30 #include "llvm/MC/MCParser/MCAsmLexer.h"
31 #include "llvm/MC/MCParser/MCTargetAsmParser.h"
32 #include "llvm/MC/MCStreamer.h"
33 #include "llvm/MC/MCSymbol.h"
34 #include "llvm/MC/TargetRegistry.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MemoryBuffer.h"
37 #include "llvm/Support/SourceMgr.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetMachine.h"
40 using namespace llvm;
41 
42 #define DEBUG_TYPE "asm-printer"
43 
44 unsigned AsmPrinter::addInlineAsmDiagBuffer(StringRef AsmStr,
45                                             const MDNode *LocMDNode) const {
46   MCContext &Context = MMI->getContext();
47   Context.initInlineSourceManager();
48   SourceMgr &SrcMgr = *Context.getInlineSourceManager();
49   std::vector<const MDNode *> &LocInfos = Context.getLocInfos();
50 
51   std::unique_ptr<MemoryBuffer> Buffer;
52   // The inline asm source manager will outlive AsmStr, so make a copy of the
53   // string for SourceMgr to own.
54   Buffer = MemoryBuffer::getMemBufferCopy(AsmStr, "<inline asm>");
55 
56   // Tell SrcMgr about this buffer, it takes ownership of the buffer.
57   unsigned BufNum = SrcMgr.AddNewSourceBuffer(std::move(Buffer), SMLoc());
58 
59   // Store LocMDNode in DiagInfo, using BufNum as an identifier.
60   if (LocMDNode) {
61     LocInfos.resize(BufNum);
62     LocInfos[BufNum - 1] = LocMDNode;
63   }
64 
65   return BufNum;
66 }
67 
68 
69 /// EmitInlineAsm - Emit a blob of inline asm to the output streamer.
70 void AsmPrinter::emitInlineAsm(StringRef Str, const MCSubtargetInfo &STI,
71                                const MCTargetOptions &MCOptions,
72                                const MDNode *LocMDNode,
73                                InlineAsm::AsmDialect Dialect) const {
74   assert(!Str.empty() && "Can't emit empty inline asm block");
75 
76   // Remember if the buffer is nul terminated or not so we can avoid a copy.
77   bool isNullTerminated = Str.back() == 0;
78   if (isNullTerminated)
79     Str = Str.substr(0, Str.size()-1);
80 
81   // If the output streamer does not have mature MC support or the integrated
82   // assembler has been disabled or not required, just emit the blob textually.
83   // Otherwise parse the asm and emit it via MC support.
84   // This is useful in case the asm parser doesn't handle something but the
85   // system assembler does.
86   const MCAsmInfo *MCAI = TM.getMCAsmInfo();
87   assert(MCAI && "No MCAsmInfo");
88   if (!MCAI->useIntegratedAssembler() &&
89       !MCAI->parseInlineAsmUsingAsmParser() &&
90       !OutStreamer->isIntegratedAssemblerRequired()) {
91     emitInlineAsmStart();
92     OutStreamer->emitRawText(Str);
93     emitInlineAsmEnd(STI, nullptr);
94     return;
95   }
96 
97   unsigned BufNum = addInlineAsmDiagBuffer(Str, LocMDNode);
98   SourceMgr &SrcMgr = *MMI->getContext().getInlineSourceManager();
99   SrcMgr.setIncludeDirs(MCOptions.IASSearchPaths);
100 
101   std::unique_ptr<MCAsmParser> Parser(
102       createMCAsmParser(SrcMgr, OutContext, *OutStreamer, *MAI, BufNum));
103 
104   // Do not use assembler-level information for parsing inline assembly.
105   OutStreamer->setUseAssemblerInfoForParsing(false);
106 
107   // We create a new MCInstrInfo here since we might be at the module level
108   // and not have a MachineFunction to initialize the TargetInstrInfo from and
109   // we only need MCInstrInfo for asm parsing. We create one unconditionally
110   // because it's not subtarget dependent.
111   std::unique_ptr<MCInstrInfo> MII(TM.getTarget().createMCInstrInfo());
112   assert(MII && "Failed to create instruction info");
113   std::unique_ptr<MCTargetAsmParser> TAP(TM.getTarget().createMCAsmParser(
114       STI, *Parser, *MII, MCOptions));
115   if (!TAP)
116     report_fatal_error("Inline asm not supported by this streamer because"
117                        " we don't have an asm parser for this target\n");
118   Parser->setAssemblerDialect(Dialect);
119   Parser->setTargetParser(*TAP);
120   // Enable lexing Masm binary and hex integer literals in intel inline
121   // assembly.
122   if (Dialect == InlineAsm::AD_Intel)
123     Parser->getLexer().setLexMasmIntegers(true);
124 
125   emitInlineAsmStart();
126   // Don't implicitly switch to the text section before the asm.
127   (void)Parser->Run(/*NoInitialTextSection*/ true,
128                     /*NoFinalize*/ true);
129   emitInlineAsmEnd(STI, &TAP->getSTI());
130 }
131 
132 static void EmitInlineAsmStr(const char *AsmStr, const MachineInstr *MI,
133                              MachineModuleInfo *MMI, const MCAsmInfo *MAI,
134                              AsmPrinter *AP, uint64_t LocCookie,
135                              raw_ostream &OS) {
136   bool InputIsIntelDialect = MI->getInlineAsmDialect() == InlineAsm::AD_Intel;
137 
138   if (InputIsIntelDialect) {
139     // Switch to the inline assembly variant.
140     OS << "\t.intel_syntax\n\t";
141   }
142 
143   int CurVariant = -1; // The number of the {.|.|.} region we are in.
144   const char *LastEmitted = AsmStr; // One past the last character emitted.
145   unsigned NumOperands = MI->getNumOperands();
146 
147   int AsmPrinterVariant;
148   if (InputIsIntelDialect)
149     AsmPrinterVariant = 1; // X86MCAsmInfo.cpp's AsmWriterFlavorTy::Intel.
150   else
151     AsmPrinterVariant = MMI->getTarget().unqualifiedInlineAsmVariant();
152 
153   // FIXME: Should this happen for `asm inteldialect` as well?
154   if (!InputIsIntelDialect && MAI->getEmitGNUAsmStartIndentationMarker())
155     OS << '\t';
156 
157   while (*LastEmitted) {
158     switch (*LastEmitted) {
159     default: {
160       // Not a special case, emit the string section literally.
161       const char *LiteralEnd = LastEmitted+1;
162       while (*LiteralEnd && *LiteralEnd != '{' && *LiteralEnd != '|' &&
163              *LiteralEnd != '}' && *LiteralEnd != '$' && *LiteralEnd != '\n')
164         ++LiteralEnd;
165       if (CurVariant == -1 || CurVariant == AsmPrinterVariant)
166         OS.write(LastEmitted, LiteralEnd - LastEmitted);
167       LastEmitted = LiteralEnd;
168       break;
169     }
170     case '\n':
171       ++LastEmitted;   // Consume newline character.
172       OS << '\n';      // Indent code with newline.
173       break;
174     case '$': {
175       ++LastEmitted;   // Consume '$' character.
176       bool Done = true;
177 
178       // Handle escapes.
179       switch (*LastEmitted) {
180       default: Done = false; break;
181       case '$':     // $$ -> $
182         if (!InputIsIntelDialect)
183           if (CurVariant == -1 || CurVariant == AsmPrinterVariant)
184             OS << '$';
185         ++LastEmitted;  // Consume second '$' character.
186         break;
187       case '(':        // $( -> same as GCC's { character.
188         ++LastEmitted; // Consume '(' character.
189         if (CurVariant != -1)
190           report_fatal_error("Nested variants found in inline asm string: '" +
191                              Twine(AsmStr) + "'");
192         CurVariant = 0; // We're in the first variant now.
193         break;
194       case '|':
195         ++LastEmitted; // Consume '|' character.
196         if (CurVariant == -1)
197           OS << '|'; // This is gcc's behavior for | outside a variant.
198         else
199           ++CurVariant; // We're in the next variant.
200         break;
201       case ')':        // $) -> same as GCC's } char.
202         ++LastEmitted; // Consume ')' character.
203         if (CurVariant == -1)
204           OS << '}'; // This is gcc's behavior for } outside a variant.
205         else
206           CurVariant = -1;
207         break;
208       }
209       if (Done) break;
210 
211       bool HasCurlyBraces = false;
212       if (*LastEmitted == '{') {     // ${variable}
213         ++LastEmitted;               // Consume '{' character.
214         HasCurlyBraces = true;
215       }
216 
217       // If we have ${:foo}, then this is not a real operand reference, it is a
218       // "magic" string reference, just like in .td files.  Arrange to call
219       // PrintSpecial.
220       if (HasCurlyBraces && *LastEmitted == ':') {
221         ++LastEmitted;
222         const char *StrStart = LastEmitted;
223         const char *StrEnd = strchr(StrStart, '}');
224         if (!StrEnd)
225           report_fatal_error("Unterminated ${:foo} operand in inline asm"
226                              " string: '" + Twine(AsmStr) + "'");
227         if (CurVariant == -1 || CurVariant == AsmPrinterVariant)
228           AP->PrintSpecial(MI, OS, StringRef(StrStart, StrEnd - StrStart));
229         LastEmitted = StrEnd+1;
230         break;
231       }
232 
233       const char *IDStart = LastEmitted;
234       const char *IDEnd = IDStart;
235       while (isDigit(*IDEnd))
236         ++IDEnd;
237 
238       unsigned Val;
239       if (StringRef(IDStart, IDEnd-IDStart).getAsInteger(10, Val))
240         report_fatal_error("Bad $ operand number in inline asm string: '" +
241                            Twine(AsmStr) + "'");
242       LastEmitted = IDEnd;
243 
244       if (Val >= NumOperands - 1)
245         report_fatal_error("Invalid $ operand number in inline asm string: '" +
246                            Twine(AsmStr) + "'");
247 
248       char Modifier[2] = { 0, 0 };
249 
250       if (HasCurlyBraces) {
251         // If we have curly braces, check for a modifier character.  This
252         // supports syntax like ${0:u}, which correspond to "%u0" in GCC asm.
253         if (*LastEmitted == ':') {
254           ++LastEmitted;    // Consume ':' character.
255           if (*LastEmitted == 0)
256             report_fatal_error("Bad ${:} expression in inline asm string: '" +
257                                Twine(AsmStr) + "'");
258 
259           Modifier[0] = *LastEmitted;
260           ++LastEmitted;    // Consume modifier character.
261         }
262 
263         if (*LastEmitted != '}')
264           report_fatal_error("Bad ${} expression in inline asm string: '" +
265                              Twine(AsmStr) + "'");
266         ++LastEmitted;    // Consume '}' character.
267       }
268 
269       // Okay, we finally have a value number.  Ask the target to print this
270       // operand!
271       if (CurVariant == -1 || CurVariant == AsmPrinterVariant) {
272         unsigned OpNo = InlineAsm::MIOp_FirstOperand;
273 
274         bool Error = false;
275 
276         // Scan to find the machine operand number for the operand.
277         for (; Val; --Val) {
278           if (OpNo >= MI->getNumOperands())
279             break;
280           unsigned OpFlags = MI->getOperand(OpNo).getImm();
281           OpNo += InlineAsm::getNumOperandRegisters(OpFlags) + 1;
282         }
283 
284         // We may have a location metadata attached to the end of the
285         // instruction, and at no point should see metadata at any
286         // other point while processing. It's an error if so.
287         if (OpNo >= MI->getNumOperands() || MI->getOperand(OpNo).isMetadata()) {
288           Error = true;
289         } else {
290           unsigned OpFlags = MI->getOperand(OpNo).getImm();
291           ++OpNo; // Skip over the ID number.
292 
293           // FIXME: Shouldn't arch-independent output template handling go into
294           // PrintAsmOperand?
295           // Labels are target independent.
296           if (MI->getOperand(OpNo).isBlockAddress()) {
297             const BlockAddress *BA = MI->getOperand(OpNo).getBlockAddress();
298             MCSymbol *Sym = AP->GetBlockAddressSymbol(BA);
299             Sym->print(OS, AP->MAI);
300             MMI->getContext().registerInlineAsmLabel(Sym);
301           } else if (MI->getOperand(OpNo).isMBB()) {
302             const MCSymbol *Sym = MI->getOperand(OpNo).getMBB()->getSymbol();
303             Sym->print(OS, AP->MAI);
304           } else if (InlineAsm::isMemKind(OpFlags)) {
305             Error = AP->PrintAsmMemoryOperand(
306                 MI, OpNo, Modifier[0] ? Modifier : nullptr, OS);
307           } else {
308             Error = AP->PrintAsmOperand(MI, OpNo,
309                                         Modifier[0] ? Modifier : nullptr, OS);
310           }
311         }
312         if (Error) {
313           std::string msg;
314           raw_string_ostream Msg(msg);
315           Msg << "invalid operand in inline asm: '" << AsmStr << "'";
316           MMI->getModule()->getContext().emitError(LocCookie, Msg.str());
317         }
318       }
319       break;
320     }
321     }
322   }
323   if (InputIsIntelDialect)
324     OS << "\n\t.att_syntax";
325   OS << '\n' << (char)0;  // null terminate string.
326 }
327 
328 /// This method formats and emits the specified machine instruction that is an
329 /// inline asm.
330 void AsmPrinter::emitInlineAsm(const MachineInstr *MI) const {
331   assert(MI->isInlineAsm() && "printInlineAsm only works on inline asms");
332 
333   // Count the number of register definitions to find the asm string.
334   unsigned NumDefs = 0;
335   for (; MI->getOperand(NumDefs).isReg() && MI->getOperand(NumDefs).isDef();
336        ++NumDefs)
337     assert(NumDefs != MI->getNumOperands()-2 && "No asm string?");
338 
339   assert(MI->getOperand(NumDefs).isSymbol() && "No asm string?");
340 
341   // Disassemble the AsmStr, printing out the literal pieces, the operands, etc.
342   const char *AsmStr = MI->getOperand(NumDefs).getSymbolName();
343 
344   // If this asmstr is empty, just print the #APP/#NOAPP markers.
345   // These are useful to see where empty asm's wound up.
346   if (AsmStr[0] == 0) {
347     OutStreamer->emitRawComment(MAI->getInlineAsmStart());
348     OutStreamer->emitRawComment(MAI->getInlineAsmEnd());
349     return;
350   }
351 
352   // Emit the #APP start marker.  This has to happen even if verbose-asm isn't
353   // enabled, so we use emitRawComment.
354   OutStreamer->emitRawComment(MAI->getInlineAsmStart());
355 
356   // Get the !srcloc metadata node if we have it, and decode the loc cookie from
357   // it.
358   uint64_t LocCookie = 0;
359   const MDNode *LocMD = nullptr;
360   for (const MachineOperand &MO : llvm::reverse(MI->operands())) {
361     if (MO.isMetadata() && (LocMD = MO.getMetadata()) &&
362         LocMD->getNumOperands() != 0) {
363       if (const ConstantInt *CI =
364               mdconst::dyn_extract<ConstantInt>(LocMD->getOperand(0))) {
365         LocCookie = CI->getZExtValue();
366         break;
367       }
368     }
369   }
370 
371   // Emit the inline asm to a temporary string so we can emit it through
372   // EmitInlineAsm.
373   SmallString<256> StringData;
374   raw_svector_ostream OS(StringData);
375 
376   AsmPrinter *AP = const_cast<AsmPrinter*>(this);
377   EmitInlineAsmStr(AsmStr, MI, MMI, MAI, AP, LocCookie, OS);
378 
379   // Emit warnings if we use reserved registers on the clobber list, as
380   // that might lead to undefined behaviour.
381   SmallVector<Register, 8> RestrRegs;
382   const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
383   // Start with the first operand descriptor, and iterate over them.
384   for (unsigned I = InlineAsm::MIOp_FirstOperand, NumOps = MI->getNumOperands();
385        I < NumOps; ++I) {
386     const MachineOperand &MO = MI->getOperand(I);
387     if (!MO.isImm())
388       continue;
389     unsigned Flags = MO.getImm();
390     if (InlineAsm::getKind(Flags) == InlineAsm::Kind_Clobber) {
391       Register Reg = MI->getOperand(I + 1).getReg();
392       if (!TRI->isAsmClobberable(*MF, Reg))
393         RestrRegs.push_back(Reg);
394     }
395     // Skip to one before the next operand descriptor, if it exists.
396     I += InlineAsm::getNumOperandRegisters(Flags);
397   }
398 
399   if (!RestrRegs.empty()) {
400     std::string Msg = "inline asm clobber list contains reserved registers: ";
401     ListSeparator LS;
402     for (const Register RR : RestrRegs) {
403       Msg += LS;
404       Msg += TRI->getRegAsmName(RR);
405     }
406     const char *Note =
407         "Reserved registers on the clobber list may not be "
408         "preserved across the asm statement, and clobbering them may "
409         "lead to undefined behaviour.";
410     MMI->getModule()->getContext().diagnose(DiagnosticInfoInlineAsm(
411         LocCookie, Msg, DiagnosticSeverity::DS_Warning));
412     MMI->getModule()->getContext().diagnose(
413         DiagnosticInfoInlineAsm(LocCookie, Note, DiagnosticSeverity::DS_Note));
414   }
415 
416   emitInlineAsm(OS.str(), getSubtargetInfo(), TM.Options.MCOptions, LocMD,
417                 MI->getInlineAsmDialect());
418 
419   // Emit the #NOAPP end marker.  This has to happen even if verbose-asm isn't
420   // enabled, so we use emitRawComment.
421   OutStreamer->emitRawComment(MAI->getInlineAsmEnd());
422 }
423 
424 /// PrintSpecial - Print information related to the specified machine instr
425 /// that is independent of the operand, and may be independent of the instr
426 /// itself.  This can be useful for portably encoding the comment character
427 /// or other bits of target-specific knowledge into the asmstrings.  The
428 /// syntax used is ${:comment}.  Targets can override this to add support
429 /// for their own strange codes.
430 void AsmPrinter::PrintSpecial(const MachineInstr *MI, raw_ostream &OS,
431                               StringRef Code) const {
432   if (Code == "private") {
433     const DataLayout &DL = MF->getDataLayout();
434     OS << DL.getPrivateGlobalPrefix();
435   } else if (Code == "comment") {
436     OS << MAI->getCommentString();
437   } else if (Code == "uid") {
438     // Comparing the address of MI isn't sufficient, because machineinstrs may
439     // be allocated to the same address across functions.
440 
441     // If this is a new LastFn instruction, bump the counter.
442     if (LastMI != MI || LastFn != getFunctionNumber()) {
443       ++Counter;
444       LastMI = MI;
445       LastFn = getFunctionNumber();
446     }
447     OS << Counter;
448   } else {
449     std::string msg;
450     raw_string_ostream Msg(msg);
451     Msg << "Unknown special formatter '" << Code
452          << "' for machine instr: " << *MI;
453     report_fatal_error(Twine(Msg.str()));
454   }
455 }
456 
457 void AsmPrinter::PrintSymbolOperand(const MachineOperand &MO, raw_ostream &OS) {
458   assert(MO.isGlobal() && "caller should check MO.isGlobal");
459   getSymbolPreferLocal(*MO.getGlobal())->print(OS, MAI);
460   printOffset(MO.getOffset(), OS);
461 }
462 
463 /// PrintAsmOperand - Print the specified operand of MI, an INLINEASM
464 /// instruction, using the specified assembler variant.  Targets should
465 /// override this to format as appropriate for machine specific ExtraCodes
466 /// or when the arch-independent handling would be too complex otherwise.
467 bool AsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
468                                  const char *ExtraCode, raw_ostream &O) {
469   // Does this asm operand have a single letter operand modifier?
470   if (ExtraCode && ExtraCode[0]) {
471     if (ExtraCode[1] != 0) return true; // Unknown modifier.
472 
473     // https://gcc.gnu.org/onlinedocs/gccint/Output-Template.html
474     const MachineOperand &MO = MI->getOperand(OpNo);
475     switch (ExtraCode[0]) {
476     default:
477       return true;  // Unknown modifier.
478     case 'a': // Print as memory address.
479       if (MO.isReg()) {
480         PrintAsmMemoryOperand(MI, OpNo, nullptr, O);
481         return false;
482       }
483       LLVM_FALLTHROUGH; // GCC allows '%a' to behave like '%c' with immediates.
484     case 'c': // Substitute immediate value without immediate syntax
485       if (MO.isImm()) {
486         O << MO.getImm();
487         return false;
488       }
489       if (MO.isGlobal()) {
490         PrintSymbolOperand(MO, O);
491         return false;
492       }
493       return true;
494     case 'n':  // Negate the immediate constant.
495       if (!MO.isImm())
496         return true;
497       O << -MO.getImm();
498       return false;
499     case 's':  // The GCC deprecated s modifier
500       if (!MO.isImm())
501         return true;
502       O << ((32 - MO.getImm()) & 31);
503       return false;
504     }
505   }
506   return true;
507 }
508 
509 bool AsmPrinter::PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo,
510                                        const char *ExtraCode, raw_ostream &O) {
511   // Target doesn't support this yet!
512   return true;
513 }
514 
515 void AsmPrinter::emitInlineAsmStart() const {}
516 
517 void AsmPrinter::emitInlineAsmEnd(const MCSubtargetInfo &StartInfo,
518                                   const MCSubtargetInfo *EndInfo) const {}
519