1 //===- AsmWriterEmitter.cpp - Generate an assembly writer -----------------===//
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 tablegen backend emits an assembly printer for the current target.
10 // Note that this is currently fairly skeletal, but will grow over time.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "AsmWriterInst.h"
15 #include "CodeGenInstruction.h"
16 #include "CodeGenRegisters.h"
17 #include "CodeGenTarget.h"
18 #include "SequenceToOffsetTable.h"
19 #include "Types.h"
20 #include "llvm/ADT/ArrayRef.h"
21 #include "llvm/ADT/DenseMap.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallString.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/ADT/Twine.h"
28 #include "llvm/Support/Casting.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/Format.h"
31 #include "llvm/Support/FormatVariadic.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/TableGen/Error.h"
35 #include "llvm/TableGen/Record.h"
36 #include "llvm/TableGen/TableGenBackend.h"
37 #include <algorithm>
38 #include <cassert>
39 #include <cstddef>
40 #include <cstdint>
41 #include <deque>
42 #include <iterator>
43 #include <map>
44 #include <set>
45 #include <string>
46 #include <tuple>
47 #include <utility>
48 #include <vector>
49
50 using namespace llvm;
51
52 #define DEBUG_TYPE "asm-writer-emitter"
53
54 namespace {
55
56 class AsmWriterEmitter {
57 RecordKeeper &Records;
58 CodeGenTarget Target;
59 ArrayRef<const CodeGenInstruction *> NumberedInstructions;
60 std::vector<AsmWriterInst> Instructions;
61
62 public:
63 AsmWriterEmitter(RecordKeeper &R);
64
65 void run(raw_ostream &o);
66 private:
67 void EmitGetMnemonic(
68 raw_ostream &o,
69 std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
70 unsigned &BitsLeft, unsigned &AsmStrBits);
71 void EmitPrintInstruction(
72 raw_ostream &o,
73 std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
74 unsigned &BitsLeft, unsigned &AsmStrBits);
75 void EmitGetRegisterName(raw_ostream &o);
76 void EmitPrintAliasInstruction(raw_ostream &O);
77
78 void FindUniqueOperandCommands(std::vector<std::string> &UOC,
79 std::vector<std::vector<unsigned>> &InstIdxs,
80 std::vector<unsigned> &InstOpsUsed,
81 bool PassSubtarget) const;
82 };
83
84 } // end anonymous namespace
85
PrintCases(std::vector<std::pair<std::string,AsmWriterOperand>> & OpsToPrint,raw_ostream & O,bool PassSubtarget)86 static void PrintCases(std::vector<std::pair<std::string,
87 AsmWriterOperand>> &OpsToPrint, raw_ostream &O,
88 bool PassSubtarget) {
89 O << " case " << OpsToPrint.back().first << ":";
90 AsmWriterOperand TheOp = OpsToPrint.back().second;
91 OpsToPrint.pop_back();
92
93 // Check to see if any other operands are identical in this list, and if so,
94 // emit a case label for them.
95 for (unsigned i = OpsToPrint.size(); i != 0; --i)
96 if (OpsToPrint[i-1].second == TheOp) {
97 O << "\n case " << OpsToPrint[i-1].first << ":";
98 OpsToPrint.erase(OpsToPrint.begin()+i-1);
99 }
100
101 // Finally, emit the code.
102 O << "\n " << TheOp.getCode(PassSubtarget);
103 O << "\n break;\n";
104 }
105
106 /// EmitInstructions - Emit the last instruction in the vector and any other
107 /// instructions that are suitably similar to it.
EmitInstructions(std::vector<AsmWriterInst> & Insts,raw_ostream & O,bool PassSubtarget)108 static void EmitInstructions(std::vector<AsmWriterInst> &Insts,
109 raw_ostream &O, bool PassSubtarget) {
110 AsmWriterInst FirstInst = Insts.back();
111 Insts.pop_back();
112
113 std::vector<AsmWriterInst> SimilarInsts;
114 unsigned DifferingOperand = ~0;
115 for (unsigned i = Insts.size(); i != 0; --i) {
116 unsigned DiffOp = Insts[i-1].MatchesAllButOneOp(FirstInst);
117 if (DiffOp != ~1U) {
118 if (DifferingOperand == ~0U) // First match!
119 DifferingOperand = DiffOp;
120
121 // If this differs in the same operand as the rest of the instructions in
122 // this class, move it to the SimilarInsts list.
123 if (DifferingOperand == DiffOp || DiffOp == ~0U) {
124 SimilarInsts.push_back(Insts[i-1]);
125 Insts.erase(Insts.begin()+i-1);
126 }
127 }
128 }
129
130 O << " case " << FirstInst.CGI->Namespace << "::"
131 << FirstInst.CGI->TheDef->getName() << ":\n";
132 for (const AsmWriterInst &AWI : SimilarInsts)
133 O << " case " << AWI.CGI->Namespace << "::"
134 << AWI.CGI->TheDef->getName() << ":\n";
135 for (unsigned i = 0, e = FirstInst.Operands.size(); i != e; ++i) {
136 if (i != DifferingOperand) {
137 // If the operand is the same for all instructions, just print it.
138 O << " " << FirstInst.Operands[i].getCode(PassSubtarget);
139 } else {
140 // If this is the operand that varies between all of the instructions,
141 // emit a switch for just this operand now.
142 O << " switch (MI->getOpcode()) {\n";
143 O << " default: llvm_unreachable(\"Unexpected opcode.\");\n";
144 std::vector<std::pair<std::string, AsmWriterOperand>> OpsToPrint;
145 OpsToPrint.push_back(std::make_pair(FirstInst.CGI->Namespace.str() + "::" +
146 FirstInst.CGI->TheDef->getName().str(),
147 FirstInst.Operands[i]));
148
149 for (const AsmWriterInst &AWI : SimilarInsts) {
150 OpsToPrint.push_back(std::make_pair(AWI.CGI->Namespace.str()+"::" +
151 AWI.CGI->TheDef->getName().str(),
152 AWI.Operands[i]));
153 }
154 std::reverse(OpsToPrint.begin(), OpsToPrint.end());
155 while (!OpsToPrint.empty())
156 PrintCases(OpsToPrint, O, PassSubtarget);
157 O << " }";
158 }
159 O << "\n";
160 }
161 O << " break;\n";
162 }
163
164 void AsmWriterEmitter::
FindUniqueOperandCommands(std::vector<std::string> & UniqueOperandCommands,std::vector<std::vector<unsigned>> & InstIdxs,std::vector<unsigned> & InstOpsUsed,bool PassSubtarget) const165 FindUniqueOperandCommands(std::vector<std::string> &UniqueOperandCommands,
166 std::vector<std::vector<unsigned>> &InstIdxs,
167 std::vector<unsigned> &InstOpsUsed,
168 bool PassSubtarget) const {
169 // This vector parallels UniqueOperandCommands, keeping track of which
170 // instructions each case are used for. It is a comma separated string of
171 // enums.
172 std::vector<std::string> InstrsForCase;
173 InstrsForCase.resize(UniqueOperandCommands.size());
174 InstOpsUsed.assign(UniqueOperandCommands.size(), 0);
175
176 for (size_t i = 0, e = Instructions.size(); i != e; ++i) {
177 const AsmWriterInst &Inst = Instructions[i];
178 if (Inst.Operands.empty())
179 continue; // Instruction already done.
180
181 std::string Command = " "+Inst.Operands[0].getCode(PassSubtarget)+"\n";
182
183 // Check to see if we already have 'Command' in UniqueOperandCommands.
184 // If not, add it.
185 auto I = llvm::find(UniqueOperandCommands, Command);
186 if (I != UniqueOperandCommands.end()) {
187 size_t idx = I - UniqueOperandCommands.begin();
188 InstrsForCase[idx] += ", ";
189 InstrsForCase[idx] += Inst.CGI->TheDef->getName();
190 InstIdxs[idx].push_back(i);
191 } else {
192 UniqueOperandCommands.push_back(std::move(Command));
193 InstrsForCase.push_back(std::string(Inst.CGI->TheDef->getName()));
194 InstIdxs.emplace_back();
195 InstIdxs.back().push_back(i);
196
197 // This command matches one operand so far.
198 InstOpsUsed.push_back(1);
199 }
200 }
201
202 // For each entry of UniqueOperandCommands, there is a set of instructions
203 // that uses it. If the next command of all instructions in the set are
204 // identical, fold it into the command.
205 for (size_t CommandIdx = 0, e = UniqueOperandCommands.size();
206 CommandIdx != e; ++CommandIdx) {
207
208 const auto &Idxs = InstIdxs[CommandIdx];
209
210 for (unsigned Op = 1; ; ++Op) {
211 // Find the first instruction in the set.
212 const AsmWriterInst &FirstInst = Instructions[Idxs.front()];
213 // If this instruction has no more operands, we isn't anything to merge
214 // into this command.
215 if (FirstInst.Operands.size() == Op)
216 break;
217
218 // Otherwise, scan to see if all of the other instructions in this command
219 // set share the operand.
220 if (any_of(drop_begin(Idxs), [&](unsigned Idx) {
221 const AsmWriterInst &OtherInst = Instructions[Idx];
222 return OtherInst.Operands.size() == Op ||
223 OtherInst.Operands[Op] != FirstInst.Operands[Op];
224 }))
225 break;
226
227 // Okay, everything in this command set has the same next operand. Add it
228 // to UniqueOperandCommands and remember that it was consumed.
229 std::string Command = " " +
230 FirstInst.Operands[Op].getCode(PassSubtarget) + "\n";
231
232 UniqueOperandCommands[CommandIdx] += Command;
233 InstOpsUsed[CommandIdx]++;
234 }
235 }
236
237 // Prepend some of the instructions each case is used for onto the case val.
238 for (unsigned i = 0, e = InstrsForCase.size(); i != e; ++i) {
239 std::string Instrs = InstrsForCase[i];
240 if (Instrs.size() > 70) {
241 Instrs.erase(Instrs.begin()+70, Instrs.end());
242 Instrs += "...";
243 }
244
245 if (!Instrs.empty())
246 UniqueOperandCommands[i] = " // " + Instrs + "\n" +
247 UniqueOperandCommands[i];
248 }
249 }
250
UnescapeString(std::string & Str)251 static void UnescapeString(std::string &Str) {
252 for (unsigned i = 0; i != Str.size(); ++i) {
253 if (Str[i] == '\\' && i != Str.size()-1) {
254 switch (Str[i+1]) {
255 default: continue; // Don't execute the code after the switch.
256 case 'a': Str[i] = '\a'; break;
257 case 'b': Str[i] = '\b'; break;
258 case 'e': Str[i] = 27; break;
259 case 'f': Str[i] = '\f'; break;
260 case 'n': Str[i] = '\n'; break;
261 case 'r': Str[i] = '\r'; break;
262 case 't': Str[i] = '\t'; break;
263 case 'v': Str[i] = '\v'; break;
264 case '"': Str[i] = '\"'; break;
265 case '\'': Str[i] = '\''; break;
266 case '\\': Str[i] = '\\'; break;
267 }
268 // Nuke the second character.
269 Str.erase(Str.begin()+i+1);
270 }
271 }
272 }
273
274 /// UnescapeAliasString - Supports literal braces in InstAlias asm string which
275 /// are escaped with '\\' to avoid being interpreted as variants. Braces must
276 /// be unescaped before c++ code is generated as (e.g.):
277 ///
278 /// AsmString = "foo \{$\x01\}";
279 ///
280 /// causes non-standard escape character warnings.
UnescapeAliasString(std::string & Str)281 static void UnescapeAliasString(std::string &Str) {
282 for (unsigned i = 0; i != Str.size(); ++i) {
283 if (Str[i] == '\\' && i != Str.size()-1) {
284 switch (Str[i+1]) {
285 default: continue; // Don't execute the code after the switch.
286 case '{': Str[i] = '{'; break;
287 case '}': Str[i] = '}'; break;
288 }
289 // Nuke the second character.
290 Str.erase(Str.begin()+i+1);
291 }
292 }
293 }
294
EmitGetMnemonic(raw_ostream & O,std::vector<std::vector<std::string>> & TableDrivenOperandPrinters,unsigned & BitsLeft,unsigned & AsmStrBits)295 void AsmWriterEmitter::EmitGetMnemonic(
296 raw_ostream &O,
297 std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
298 unsigned &BitsLeft, unsigned &AsmStrBits) {
299 Record *AsmWriter = Target.getAsmWriter();
300 StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
301 bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
302
303 O << "/// getMnemonic - This method is automatically generated by "
304 "tablegen\n"
305 "/// from the instruction set description.\n"
306 "std::pair<const char *, uint64_t> "
307 << Target.getName() << ClassName << "::getMnemonic(const MCInst *MI) {\n";
308
309 // Build an aggregate string, and build a table of offsets into it.
310 SequenceToOffsetTable<std::string> StringTable;
311
312 /// OpcodeInfo - This encodes the index of the string to use for the first
313 /// chunk of the output as well as indices used for operand printing.
314 std::vector<uint64_t> OpcodeInfo(NumberedInstructions.size());
315 const unsigned OpcodeInfoBits = 64;
316
317 // Add all strings to the string table upfront so it can generate an optimized
318 // representation.
319 for (AsmWriterInst &AWI : Instructions) {
320 if (AWI.Operands[0].OperandType ==
321 AsmWriterOperand::isLiteralTextOperand &&
322 !AWI.Operands[0].Str.empty()) {
323 std::string Str = AWI.Operands[0].Str;
324 UnescapeString(Str);
325 StringTable.add(Str);
326 }
327 }
328
329 StringTable.layout();
330
331 unsigned MaxStringIdx = 0;
332 for (AsmWriterInst &AWI : Instructions) {
333 unsigned Idx;
334 if (AWI.Operands[0].OperandType != AsmWriterOperand::isLiteralTextOperand ||
335 AWI.Operands[0].Str.empty()) {
336 // Something handled by the asmwriter printer, but with no leading string.
337 Idx = StringTable.get("");
338 } else {
339 std::string Str = AWI.Operands[0].Str;
340 UnescapeString(Str);
341 Idx = StringTable.get(Str);
342 MaxStringIdx = std::max(MaxStringIdx, Idx);
343
344 // Nuke the string from the operand list. It is now handled!
345 AWI.Operands.erase(AWI.Operands.begin());
346 }
347
348 // Bias offset by one since we want 0 as a sentinel.
349 OpcodeInfo[AWI.CGIIndex] = Idx+1;
350 }
351
352 // Figure out how many bits we used for the string index.
353 AsmStrBits = Log2_32_Ceil(MaxStringIdx + 2);
354
355 // To reduce code size, we compactify common instructions into a few bits
356 // in the opcode-indexed table.
357 BitsLeft = OpcodeInfoBits - AsmStrBits;
358
359 while (true) {
360 std::vector<std::string> UniqueOperandCommands;
361 std::vector<std::vector<unsigned>> InstIdxs;
362 std::vector<unsigned> NumInstOpsHandled;
363 FindUniqueOperandCommands(UniqueOperandCommands, InstIdxs,
364 NumInstOpsHandled, PassSubtarget);
365
366 // If we ran out of operands to print, we're done.
367 if (UniqueOperandCommands.empty()) break;
368
369 // Compute the number of bits we need to represent these cases, this is
370 // ceil(log2(numentries)).
371 unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
372
373 // If we don't have enough bits for this operand, don't include it.
374 if (NumBits > BitsLeft) {
375 LLVM_DEBUG(errs() << "Not enough bits to densely encode " << NumBits
376 << " more bits\n");
377 break;
378 }
379
380 // Otherwise, we can include this in the initial lookup table. Add it in.
381 for (size_t i = 0, e = InstIdxs.size(); i != e; ++i) {
382 unsigned NumOps = NumInstOpsHandled[i];
383 for (unsigned Idx : InstIdxs[i]) {
384 OpcodeInfo[Instructions[Idx].CGIIndex] |=
385 (uint64_t)i << (OpcodeInfoBits-BitsLeft);
386 // Remove the info about this operand from the instruction.
387 AsmWriterInst &Inst = Instructions[Idx];
388 if (!Inst.Operands.empty()) {
389 assert(NumOps <= Inst.Operands.size() &&
390 "Can't remove this many ops!");
391 Inst.Operands.erase(Inst.Operands.begin(),
392 Inst.Operands.begin()+NumOps);
393 }
394 }
395 }
396 BitsLeft -= NumBits;
397
398 // Remember the handlers for this set of operands.
399 TableDrivenOperandPrinters.push_back(std::move(UniqueOperandCommands));
400 }
401
402 // Emit the string table itself.
403 StringTable.emitStringLiteralDef(O, " static const char AsmStrs[]");
404
405 // Emit the lookup tables in pieces to minimize wasted bytes.
406 unsigned BytesNeeded = ((OpcodeInfoBits - BitsLeft) + 7) / 8;
407 unsigned Table = 0, Shift = 0;
408 SmallString<128> BitsString;
409 raw_svector_ostream BitsOS(BitsString);
410 // If the total bits is more than 32-bits we need to use a 64-bit type.
411 BitsOS << " uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
412 << "_t Bits = 0;\n";
413 while (BytesNeeded != 0) {
414 // Figure out how big this table section needs to be, but no bigger than 4.
415 unsigned TableSize = std::min(llvm::bit_floor(BytesNeeded), 4u);
416 BytesNeeded -= TableSize;
417 TableSize *= 8; // Convert to bits;
418 uint64_t Mask = (1ULL << TableSize) - 1;
419 O << " static const uint" << TableSize << "_t OpInfo" << Table
420 << "[] = {\n";
421 for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
422 O << " " << ((OpcodeInfo[i] >> Shift) & Mask) << "U,\t// "
423 << NumberedInstructions[i]->TheDef->getName() << "\n";
424 }
425 O << " };\n\n";
426 // Emit string to combine the individual table lookups.
427 BitsOS << " Bits |= ";
428 // If the total bits is more than 32-bits we need to use a 64-bit type.
429 if (BitsLeft < (OpcodeInfoBits - 32))
430 BitsOS << "(uint64_t)";
431 BitsOS << "OpInfo" << Table << "[MI->getOpcode()] << " << Shift << ";\n";
432 // Prepare the shift for the next iteration and increment the table count.
433 Shift += TableSize;
434 ++Table;
435 }
436
437 O << " // Emit the opcode for the instruction.\n";
438 O << BitsString;
439
440 // Return mnemonic string and bits.
441 O << " return {AsmStrs+(Bits & " << (1 << AsmStrBits) - 1
442 << ")-1, Bits};\n\n";
443
444 O << "}\n";
445 }
446
447 /// EmitPrintInstruction - Generate the code for the "printInstruction" method
448 /// implementation. Destroys all instances of AsmWriterInst information, by
449 /// clearing the Instructions vector.
EmitPrintInstruction(raw_ostream & O,std::vector<std::vector<std::string>> & TableDrivenOperandPrinters,unsigned & BitsLeft,unsigned & AsmStrBits)450 void AsmWriterEmitter::EmitPrintInstruction(
451 raw_ostream &O,
452 std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
453 unsigned &BitsLeft, unsigned &AsmStrBits) {
454 const unsigned OpcodeInfoBits = 64;
455 Record *AsmWriter = Target.getAsmWriter();
456 StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
457 bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
458
459 // This function has some huge switch statements that causing excessive
460 // compile time in LLVM profile instrumenation build. This print function
461 // usually is not frequently called in compilation. Here we disable the
462 // profile instrumenation for this function.
463 O << "/// printInstruction - This method is automatically generated by "
464 "tablegen\n"
465 "/// from the instruction set description.\n"
466 "LLVM_NO_PROFILE_INSTRUMENT_FUNCTION\n"
467 "void "
468 << Target.getName() << ClassName
469 << "::printInstruction(const MCInst *MI, uint64_t Address, "
470 << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
471 << "raw_ostream &O) {\n";
472
473 // Emit the initial tab character.
474 O << " O << \"\\t\";\n\n";
475
476 // Emit the starting string.
477 O << " auto MnemonicInfo = getMnemonic(MI);\n\n";
478 O << " O << MnemonicInfo.first;\n\n";
479
480 O << " uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
481 << "_t Bits = MnemonicInfo.second;\n"
482 << " assert(Bits != 0 && \"Cannot print this instruction.\");\n";
483
484 // Output the table driven operand information.
485 BitsLeft = OpcodeInfoBits-AsmStrBits;
486 for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
487 std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];
488
489 // Compute the number of bits we need to represent these cases, this is
490 // ceil(log2(numentries)).
491 unsigned NumBits = Log2_32_Ceil(Commands.size());
492 assert(NumBits <= BitsLeft && "consistency error");
493
494 // Emit code to extract this field from Bits.
495 O << "\n // Fragment " << i << " encoded into " << NumBits
496 << " bits for " << Commands.size() << " unique commands.\n";
497
498 if (Commands.size() == 2) {
499 // Emit two possibilitys with if/else.
500 O << " if ((Bits >> "
501 << (OpcodeInfoBits-BitsLeft) << ") & "
502 << ((1 << NumBits)-1) << ") {\n"
503 << Commands[1]
504 << " } else {\n"
505 << Commands[0]
506 << " }\n\n";
507 } else if (Commands.size() == 1) {
508 // Emit a single possibility.
509 O << Commands[0] << "\n\n";
510 } else {
511 O << " switch ((Bits >> "
512 << (OpcodeInfoBits-BitsLeft) << ") & "
513 << ((1 << NumBits)-1) << ") {\n"
514 << " default: llvm_unreachable(\"Invalid command number.\");\n";
515
516 // Print out all the cases.
517 for (unsigned j = 0, e = Commands.size(); j != e; ++j) {
518 O << " case " << j << ":\n";
519 O << Commands[j];
520 O << " break;\n";
521 }
522 O << " }\n\n";
523 }
524 BitsLeft -= NumBits;
525 }
526
527 // Okay, delete instructions with no operand info left.
528 llvm::erase_if(Instructions,
529 [](AsmWriterInst &Inst) { return Inst.Operands.empty(); });
530
531 // Because this is a vector, we want to emit from the end. Reverse all of the
532 // elements in the vector.
533 std::reverse(Instructions.begin(), Instructions.end());
534
535
536 // Now that we've emitted all of the operand info that fit into 64 bits, emit
537 // information for those instructions that are left. This is a less dense
538 // encoding, but we expect the main 64-bit table to handle the majority of
539 // instructions.
540 if (!Instructions.empty()) {
541 // Find the opcode # of inline asm.
542 O << " switch (MI->getOpcode()) {\n";
543 O << " default: llvm_unreachable(\"Unexpected opcode.\");\n";
544 while (!Instructions.empty())
545 EmitInstructions(Instructions, O, PassSubtarget);
546
547 O << " }\n";
548 }
549
550 O << "}\n";
551 }
552
553 static void
emitRegisterNameString(raw_ostream & O,StringRef AltName,const std::deque<CodeGenRegister> & Registers)554 emitRegisterNameString(raw_ostream &O, StringRef AltName,
555 const std::deque<CodeGenRegister> &Registers) {
556 SequenceToOffsetTable<std::string> StringTable;
557 SmallVector<std::string, 4> AsmNames(Registers.size());
558 unsigned i = 0;
559 for (const auto &Reg : Registers) {
560 std::string &AsmName = AsmNames[i++];
561
562 // "NoRegAltName" is special. We don't need to do a lookup for that,
563 // as it's just a reference to the default register name.
564 if (AltName == "" || AltName == "NoRegAltName") {
565 AsmName = std::string(Reg.TheDef->getValueAsString("AsmName"));
566 if (AsmName.empty())
567 AsmName = std::string(Reg.getName());
568 } else {
569 // Make sure the register has an alternate name for this index.
570 std::vector<Record*> AltNameList =
571 Reg.TheDef->getValueAsListOfDefs("RegAltNameIndices");
572 unsigned Idx = 0, e;
573 for (e = AltNameList.size();
574 Idx < e && (AltNameList[Idx]->getName() != AltName);
575 ++Idx)
576 ;
577 // If the register has an alternate name for this index, use it.
578 // Otherwise, leave it empty as an error flag.
579 if (Idx < e) {
580 std::vector<StringRef> AltNames =
581 Reg.TheDef->getValueAsListOfStrings("AltNames");
582 if (AltNames.size() <= Idx)
583 PrintFatalError(Reg.TheDef->getLoc(),
584 "Register definition missing alt name for '" +
585 AltName + "'.");
586 AsmName = std::string(AltNames[Idx]);
587 }
588 }
589 StringTable.add(AsmName);
590 }
591
592 StringTable.layout();
593 StringTable.emitStringLiteralDef(O, Twine(" static const char AsmStrs") +
594 AltName + "[]");
595
596 O << " static const " << getMinimalTypeForRange(StringTable.size() - 1, 32)
597 << " RegAsmOffset" << AltName << "[] = {";
598 for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
599 if ((i % 14) == 0)
600 O << "\n ";
601 O << StringTable.get(AsmNames[i]) << ", ";
602 }
603 O << "\n };\n"
604 << "\n";
605 }
606
EmitGetRegisterName(raw_ostream & O)607 void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) {
608 Record *AsmWriter = Target.getAsmWriter();
609 StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
610 const auto &Registers = Target.getRegBank().getRegisters();
611 const std::vector<Record*> &AltNameIndices = Target.getRegAltNameIndices();
612 bool hasAltNames = AltNameIndices.size() > 1;
613 StringRef Namespace = Registers.front().TheDef->getValueAsString("Namespace");
614
615 O <<
616 "\n\n/// getRegisterName - This method is automatically generated by tblgen\n"
617 "/// from the register set description. This returns the assembler name\n"
618 "/// for the specified register.\n"
619 "const char *" << Target.getName() << ClassName << "::";
620 if (hasAltNames)
621 O << "\ngetRegisterName(MCRegister Reg, unsigned AltIdx) {\n";
622 else
623 O << "getRegisterName(MCRegister Reg) {\n";
624 O << " unsigned RegNo = Reg.id();\n"
625 << " assert(RegNo && RegNo < " << (Registers.size() + 1)
626 << " && \"Invalid register number!\");\n"
627 << "\n";
628
629 if (hasAltNames) {
630 for (const Record *R : AltNameIndices)
631 emitRegisterNameString(O, R->getName(), Registers);
632 } else
633 emitRegisterNameString(O, "", Registers);
634
635 if (hasAltNames) {
636 O << " switch(AltIdx) {\n"
637 << " default: llvm_unreachable(\"Invalid register alt name index!\");\n";
638 for (const Record *R : AltNameIndices) {
639 StringRef AltName = R->getName();
640 O << " case ";
641 if (!Namespace.empty())
642 O << Namespace << "::";
643 O << AltName << ":\n";
644 if (R->isValueUnset("FallbackRegAltNameIndex"))
645 O << " assert(*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
646 << "[RegNo-1]) &&\n"
647 << " \"Invalid alt name index for register!\");\n";
648 else {
649 O << " if (!*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
650 << "[RegNo-1]))\n"
651 << " return getRegisterName(RegNo, ";
652 if (!Namespace.empty())
653 O << Namespace << "::";
654 O << R->getValueAsDef("FallbackRegAltNameIndex")->getName() << ");\n";
655 }
656 O << " return AsmStrs" << AltName << "+RegAsmOffset" << AltName
657 << "[RegNo-1];\n";
658 }
659 O << " }\n";
660 } else {
661 O << " assert (*(AsmStrs+RegAsmOffset[RegNo-1]) &&\n"
662 << " \"Invalid alt name index for register!\");\n"
663 << " return AsmStrs+RegAsmOffset[RegNo-1];\n";
664 }
665 O << "}\n";
666 }
667
668 namespace {
669
670 // IAPrinter - Holds information about an InstAlias. Two InstAliases match if
671 // they both have the same conditionals. In which case, we cannot print out the
672 // alias for that pattern.
673 class IAPrinter {
674 std::map<StringRef, std::pair<int, int>> OpMap;
675
676 std::vector<std::string> Conds;
677
678 std::string Result;
679 std::string AsmString;
680
681 unsigned NumMIOps;
682
683 public:
IAPrinter(std::string R,std::string AS,unsigned NumMIOps)684 IAPrinter(std::string R, std::string AS, unsigned NumMIOps)
685 : Result(std::move(R)), AsmString(std::move(AS)), NumMIOps(NumMIOps) {}
686
addCond(std::string C)687 void addCond(std::string C) { Conds.push_back(std::move(C)); }
getConds() const688 ArrayRef<std::string> getConds() const { return Conds; }
getCondCount() const689 size_t getCondCount() const { return Conds.size(); }
690
addOperand(StringRef Op,int OpIdx,int PrintMethodIdx=-1)691 void addOperand(StringRef Op, int OpIdx, int PrintMethodIdx = -1) {
692 assert(OpIdx >= 0 && OpIdx < 0xFE && "Idx out of range");
693 assert(PrintMethodIdx >= -1 && PrintMethodIdx < 0xFF &&
694 "Idx out of range");
695 OpMap[Op] = std::make_pair(OpIdx, PrintMethodIdx);
696 }
697
getNumMIOps()698 unsigned getNumMIOps() { return NumMIOps; }
699
getResult()700 StringRef getResult() { return Result; }
701
isOpMapped(StringRef Op)702 bool isOpMapped(StringRef Op) { return OpMap.find(Op) != OpMap.end(); }
getOpIndex(StringRef Op)703 int getOpIndex(StringRef Op) { return OpMap[Op].first; }
getOpData(StringRef Op)704 std::pair<int, int> &getOpData(StringRef Op) { return OpMap[Op]; }
705
parseName(StringRef::iterator Start,StringRef::iterator End)706 std::pair<StringRef, StringRef::iterator> parseName(StringRef::iterator Start,
707 StringRef::iterator End) {
708 StringRef::iterator I = Start;
709 StringRef::iterator Next;
710 if (*I == '{') {
711 // ${some_name}
712 Start = ++I;
713 while (I != End && *I != '}')
714 ++I;
715 Next = I;
716 // eat the final '}'
717 if (Next != End)
718 ++Next;
719 } else {
720 // $name, just eat the usual suspects.
721 while (I != End && (isAlnum(*I) || *I == '_'))
722 ++I;
723 Next = I;
724 }
725
726 return std::make_pair(StringRef(Start, I - Start), Next);
727 }
728
formatAliasString(uint32_t & UnescapedSize)729 std::string formatAliasString(uint32_t &UnescapedSize) {
730 // Directly mangle mapped operands into the string. Each operand is
731 // identified by a '$' sign followed by a byte identifying the number of the
732 // operand. We add one to the index to avoid zero bytes.
733 StringRef ASM(AsmString);
734 std::string OutString;
735 raw_string_ostream OS(OutString);
736 for (StringRef::iterator I = ASM.begin(), E = ASM.end(); I != E;) {
737 OS << *I;
738 ++UnescapedSize;
739 if (*I == '$') {
740 StringRef Name;
741 std::tie(Name, I) = parseName(++I, E);
742 assert(isOpMapped(Name) && "Unmapped operand!");
743
744 int OpIndex, PrintIndex;
745 std::tie(OpIndex, PrintIndex) = getOpData(Name);
746 if (PrintIndex == -1) {
747 // Can use the default printOperand route.
748 OS << format("\\x%02X", (unsigned char)OpIndex + 1);
749 ++UnescapedSize;
750 } else {
751 // 3 bytes if a PrintMethod is needed: 0xFF, the MCInst operand
752 // number, and which of our pre-detected Methods to call.
753 OS << format("\\xFF\\x%02X\\x%02X", OpIndex + 1, PrintIndex + 1);
754 UnescapedSize += 3;
755 }
756 } else {
757 ++I;
758 }
759 }
760 return OutString;
761 }
762
operator ==(const IAPrinter & RHS) const763 bool operator==(const IAPrinter &RHS) const {
764 if (NumMIOps != RHS.NumMIOps)
765 return false;
766 if (Conds.size() != RHS.Conds.size())
767 return false;
768
769 unsigned Idx = 0;
770 for (const auto &str : Conds)
771 if (str != RHS.Conds[Idx++])
772 return false;
773
774 return true;
775 }
776 };
777
778 } // end anonymous namespace
779
CountNumOperands(StringRef AsmString,unsigned Variant)780 static unsigned CountNumOperands(StringRef AsmString, unsigned Variant) {
781 return AsmString.count(' ') + AsmString.count('\t');
782 }
783
784 namespace {
785
786 struct AliasPriorityComparator {
787 typedef std::pair<CodeGenInstAlias, int> ValueType;
operator ()__anon2ca7038a0511::AliasPriorityComparator788 bool operator()(const ValueType &LHS, const ValueType &RHS) const {
789 if (LHS.second == RHS.second) {
790 // We don't actually care about the order, but for consistency it
791 // shouldn't depend on pointer comparisons.
792 return LessRecordByID()(LHS.first.TheDef, RHS.first.TheDef);
793 }
794
795 // Aliases with larger priorities should be considered first.
796 return LHS.second > RHS.second;
797 }
798 };
799
800 } // end anonymous namespace
801
EmitPrintAliasInstruction(raw_ostream & O)802 void AsmWriterEmitter::EmitPrintAliasInstruction(raw_ostream &O) {
803 Record *AsmWriter = Target.getAsmWriter();
804
805 O << "\n#ifdef PRINT_ALIAS_INSTR\n";
806 O << "#undef PRINT_ALIAS_INSTR\n\n";
807
808 //////////////////////////////
809 // Gather information about aliases we need to print
810 //////////////////////////////
811
812 // Emit the method that prints the alias instruction.
813 StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
814 unsigned Variant = AsmWriter->getValueAsInt("Variant");
815 bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
816
817 std::vector<Record*> AllInstAliases =
818 Records.getAllDerivedDefinitions("InstAlias");
819
820 // Create a map from the qualified name to a list of potential matches.
821 typedef std::set<std::pair<CodeGenInstAlias, int>, AliasPriorityComparator>
822 AliasWithPriority;
823 std::map<std::string, AliasWithPriority> AliasMap;
824 for (Record *R : AllInstAliases) {
825 int Priority = R->getValueAsInt("EmitPriority");
826 if (Priority < 1)
827 continue; // Aliases with priority 0 are never emitted.
828
829 const DagInit *DI = R->getValueAsDag("ResultInst");
830 AliasMap[getQualifiedName(DI->getOperatorAsDef(R->getLoc()))].insert(
831 std::make_pair(CodeGenInstAlias(R, Target), Priority));
832 }
833
834 // A map of which conditions need to be met for each instruction operand
835 // before it can be matched to the mnemonic.
836 std::map<std::string, std::vector<IAPrinter>> IAPrinterMap;
837
838 std::vector<std::pair<std::string, bool>> PrintMethods;
839
840 // A list of MCOperandPredicates for all operands in use, and the reverse map
841 std::vector<const Record*> MCOpPredicates;
842 DenseMap<const Record*, unsigned> MCOpPredicateMap;
843
844 for (auto &Aliases : AliasMap) {
845 // Collection of instruction alias rules. May contain ambiguous rules.
846 std::vector<IAPrinter> IAPs;
847
848 for (auto &Alias : Aliases.second) {
849 const CodeGenInstAlias &CGA = Alias.first;
850 unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
851 std::string FlatInstAsmString =
852 CodeGenInstruction::FlattenAsmStringVariants(CGA.ResultInst->AsmString,
853 Variant);
854 unsigned NumResultOps = CountNumOperands(FlatInstAsmString, Variant);
855
856 std::string FlatAliasAsmString =
857 CodeGenInstruction::FlattenAsmStringVariants(CGA.AsmString, Variant);
858 UnescapeAliasString(FlatAliasAsmString);
859
860 // Don't emit the alias if it has more operands than what it's aliasing.
861 if (NumResultOps < CountNumOperands(FlatAliasAsmString, Variant))
862 continue;
863
864 StringRef Namespace = Target.getName();
865 unsigned NumMIOps = 0;
866 for (auto &ResultInstOpnd : CGA.ResultInst->Operands)
867 NumMIOps += ResultInstOpnd.MINumOperands;
868
869 IAPrinter IAP(CGA.Result->getAsString(), FlatAliasAsmString, NumMIOps);
870
871 unsigned MIOpNum = 0;
872 for (unsigned i = 0, e = LastOpNo; i != e; ++i) {
873 // Skip over tied operands as they're not part of an alias declaration.
874 auto &Operands = CGA.ResultInst->Operands;
875 while (true) {
876 unsigned OpNum = Operands.getSubOperandNumber(MIOpNum).first;
877 if (Operands[OpNum].MINumOperands == 1 &&
878 Operands[OpNum].getTiedRegister() != -1) {
879 // Tied operands of different RegisterClass should be explicit within
880 // an instruction's syntax and so cannot be skipped.
881 int TiedOpNum = Operands[OpNum].getTiedRegister();
882 if (Operands[OpNum].Rec->getName() ==
883 Operands[TiedOpNum].Rec->getName()) {
884 ++MIOpNum;
885 continue;
886 }
887 }
888 break;
889 }
890
891 // Ignore unchecked result operands.
892 while (IAP.getCondCount() < MIOpNum)
893 IAP.addCond("AliasPatternCond::K_Ignore, 0");
894
895 const CodeGenInstAlias::ResultOperand &RO = CGA.ResultOperands[i];
896
897 switch (RO.Kind) {
898 case CodeGenInstAlias::ResultOperand::K_Record: {
899 const Record *Rec = RO.getRecord();
900 StringRef ROName = RO.getName();
901 int PrintMethodIdx = -1;
902
903 // These two may have a PrintMethod, which we want to record (if it's
904 // the first time we've seen it) and provide an index for the aliasing
905 // code to use.
906 if (Rec->isSubClassOf("RegisterOperand") ||
907 Rec->isSubClassOf("Operand")) {
908 StringRef PrintMethod = Rec->getValueAsString("PrintMethod");
909 bool IsPCRel =
910 Rec->getValueAsString("OperandType") == "OPERAND_PCREL";
911 if (PrintMethod != "" && PrintMethod != "printOperand") {
912 PrintMethodIdx = llvm::find_if(PrintMethods,
913 [&](auto &X) {
914 return X.first == PrintMethod;
915 }) -
916 PrintMethods.begin();
917 if (static_cast<unsigned>(PrintMethodIdx) == PrintMethods.size())
918 PrintMethods.emplace_back(std::string(PrintMethod), IsPCRel);
919 }
920 }
921
922 if (Rec->isSubClassOf("RegisterOperand"))
923 Rec = Rec->getValueAsDef("RegClass");
924 if (Rec->isSubClassOf("RegisterClass")) {
925 if (!IAP.isOpMapped(ROName)) {
926 IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
927 Record *R = CGA.ResultOperands[i].getRecord();
928 if (R->isSubClassOf("RegisterOperand"))
929 R = R->getValueAsDef("RegClass");
930 IAP.addCond(std::string(
931 formatv("AliasPatternCond::K_RegClass, {0}::{1}RegClassID",
932 Namespace, R->getName())));
933 } else {
934 IAP.addCond(std::string(formatv(
935 "AliasPatternCond::K_TiedReg, {0}", IAP.getOpIndex(ROName))));
936 }
937 } else {
938 // Assume all printable operands are desired for now. This can be
939 // overridden in the InstAlias instantiation if necessary.
940 IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
941
942 // There might be an additional predicate on the MCOperand
943 unsigned Entry = MCOpPredicateMap[Rec];
944 if (!Entry) {
945 if (!Rec->isValueUnset("MCOperandPredicate")) {
946 MCOpPredicates.push_back(Rec);
947 Entry = MCOpPredicates.size();
948 MCOpPredicateMap[Rec] = Entry;
949 } else
950 break; // No conditions on this operand at all
951 }
952 IAP.addCond(
953 std::string(formatv("AliasPatternCond::K_Custom, {0}", Entry)));
954 }
955 break;
956 }
957 case CodeGenInstAlias::ResultOperand::K_Imm: {
958 // Just because the alias has an immediate result, doesn't mean the
959 // MCInst will. An MCExpr could be present, for example.
960 auto Imm = CGA.ResultOperands[i].getImm();
961 int32_t Imm32 = int32_t(Imm);
962 if (Imm != Imm32)
963 PrintFatalError("Matching an alias with an immediate out of the "
964 "range of int32_t is not supported");
965 IAP.addCond(std::string(
966 formatv("AliasPatternCond::K_Imm, uint32_t({0})", Imm32)));
967 break;
968 }
969 case CodeGenInstAlias::ResultOperand::K_Reg:
970 if (!CGA.ResultOperands[i].getRegister()) {
971 IAP.addCond(std::string(formatv(
972 "AliasPatternCond::K_Reg, {0}::NoRegister", Namespace)));
973 break;
974 }
975
976 StringRef Reg = CGA.ResultOperands[i].getRegister()->getName();
977 IAP.addCond(std::string(
978 formatv("AliasPatternCond::K_Reg, {0}::{1}", Namespace, Reg)));
979 break;
980 }
981
982 MIOpNum += RO.getMINumOperands();
983 }
984
985 std::vector<Record *> ReqFeatures;
986 if (PassSubtarget) {
987 // We only consider ReqFeatures predicates if PassSubtarget
988 std::vector<Record *> RF =
989 CGA.TheDef->getValueAsListOfDefs("Predicates");
990 copy_if(RF, std::back_inserter(ReqFeatures), [](Record *R) {
991 return R->getValueAsBit("AssemblerMatcherPredicate");
992 });
993 }
994
995 for (Record *const R : ReqFeatures) {
996 const DagInit *D = R->getValueAsDag("AssemblerCondDag");
997 std::string CombineType = D->getOperator()->getAsString();
998 if (CombineType != "any_of" && CombineType != "all_of")
999 PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1000 if (D->getNumArgs() == 0)
1001 PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1002 bool IsOr = CombineType == "any_of";
1003 // Change (any_of FeatureAll, (any_of ...)) to (any_of FeatureAll, ...).
1004 if (IsOr && D->getNumArgs() == 2 && isa<DagInit>(D->getArg(1))) {
1005 DagInit *RHS = dyn_cast<DagInit>(D->getArg(1));
1006 SmallVector<Init *> Args{D->getArg(0)};
1007 SmallVector<StringInit *> ArgNames{D->getArgName(0)};
1008 for (unsigned i = 0, e = RHS->getNumArgs(); i != e; ++i) {
1009 Args.push_back(RHS->getArg(i));
1010 ArgNames.push_back(RHS->getArgName(i));
1011 }
1012 D = DagInit::get(D->getOperator(), nullptr, Args, ArgNames);
1013 }
1014
1015 for (auto *Arg : D->getArgs()) {
1016 bool IsNeg = false;
1017 if (auto *NotArg = dyn_cast<DagInit>(Arg)) {
1018 if (NotArg->getOperator()->getAsString() != "not" ||
1019 NotArg->getNumArgs() != 1)
1020 PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1021 Arg = NotArg->getArg(0);
1022 IsNeg = true;
1023 }
1024 if (!isa<DefInit>(Arg) ||
1025 !cast<DefInit>(Arg)->getDef()->isSubClassOf("SubtargetFeature"))
1026 PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1027
1028 IAP.addCond(std::string(formatv(
1029 "AliasPatternCond::K_{0}{1}Feature, {2}::{3}", IsOr ? "Or" : "",
1030 IsNeg ? "Neg" : "", Namespace, Arg->getAsString())));
1031 }
1032 // If an AssemblerPredicate with ors is used, note end of list should
1033 // these be combined.
1034 if (IsOr)
1035 IAP.addCond("AliasPatternCond::K_EndOrFeatures, 0");
1036 }
1037
1038 IAPrinterMap[Aliases.first].push_back(std::move(IAP));
1039 }
1040 }
1041
1042 //////////////////////////////
1043 // Write out the printAliasInstr function
1044 //////////////////////////////
1045
1046 std::string Header;
1047 raw_string_ostream HeaderO(Header);
1048
1049 HeaderO << "bool " << Target.getName() << ClassName
1050 << "::printAliasInstr(const MCInst"
1051 << " *MI, uint64_t Address, "
1052 << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
1053 << "raw_ostream &OS) {\n";
1054
1055 std::string PatternsForOpcode;
1056 raw_string_ostream OpcodeO(PatternsForOpcode);
1057
1058 unsigned PatternCount = 0;
1059 std::string Patterns;
1060 raw_string_ostream PatternO(Patterns);
1061
1062 unsigned CondCount = 0;
1063 std::string Conds;
1064 raw_string_ostream CondO(Conds);
1065
1066 // All flattened alias strings.
1067 std::map<std::string, uint32_t> AsmStringOffsets;
1068 std::vector<std::pair<uint32_t, std::string>> AsmStrings;
1069 size_t AsmStringsSize = 0;
1070
1071 // Iterate over the opcodes in enum order so they are sorted by opcode for
1072 // binary search.
1073 for (const CodeGenInstruction *Inst : NumberedInstructions) {
1074 auto It = IAPrinterMap.find(getQualifiedName(Inst->TheDef));
1075 if (It == IAPrinterMap.end())
1076 continue;
1077 std::vector<IAPrinter> &IAPs = It->second;
1078 std::vector<IAPrinter*> UniqueIAPs;
1079
1080 // Remove any ambiguous alias rules.
1081 for (auto &LHS : IAPs) {
1082 bool IsDup = false;
1083 for (const auto &RHS : IAPs) {
1084 if (&LHS != &RHS && LHS == RHS) {
1085 IsDup = true;
1086 break;
1087 }
1088 }
1089
1090 if (!IsDup)
1091 UniqueIAPs.push_back(&LHS);
1092 }
1093
1094 if (UniqueIAPs.empty()) continue;
1095
1096 unsigned PatternStart = PatternCount;
1097
1098 // Insert the pattern start and opcode in the pattern list for debugging.
1099 PatternO << formatv(" // {0} - {1}\n", It->first, PatternStart);
1100
1101 for (IAPrinter *IAP : UniqueIAPs) {
1102 // Start each condition list with a comment of the resulting pattern that
1103 // we're trying to match.
1104 unsigned CondStart = CondCount;
1105 CondO << formatv(" // {0} - {1}\n", IAP->getResult(), CondStart);
1106 for (const auto &Cond : IAP->getConds())
1107 CondO << " {" << Cond << "},\n";
1108 CondCount += IAP->getCondCount();
1109
1110 // After operands have been examined, re-encode the alias string with
1111 // escapes indicating how operands should be printed.
1112 uint32_t UnescapedSize = 0;
1113 std::string EncodedAsmString = IAP->formatAliasString(UnescapedSize);
1114 auto Insertion =
1115 AsmStringOffsets.insert({EncodedAsmString, AsmStringsSize});
1116 if (Insertion.second) {
1117 // If the string is new, add it to the vector.
1118 AsmStrings.push_back({AsmStringsSize, EncodedAsmString});
1119 AsmStringsSize += UnescapedSize + 1;
1120 }
1121 unsigned AsmStrOffset = Insertion.first->second;
1122
1123 PatternO << formatv(" {{{0}, {1}, {2}, {3} },\n", AsmStrOffset,
1124 CondStart, IAP->getNumMIOps(), IAP->getCondCount());
1125 ++PatternCount;
1126 }
1127
1128 OpcodeO << formatv(" {{{0}, {1}, {2} },\n", It->first, PatternStart,
1129 PatternCount - PatternStart);
1130 }
1131
1132 if (OpcodeO.str().empty()) {
1133 O << HeaderO.str();
1134 O << " return false;\n";
1135 O << "}\n\n";
1136 O << "#endif // PRINT_ALIAS_INSTR\n";
1137 return;
1138 }
1139
1140 // Forward declare the validation method if needed.
1141 if (!MCOpPredicates.empty())
1142 O << "static bool " << Target.getName() << ClassName
1143 << "ValidateMCOperand(const MCOperand &MCOp,\n"
1144 << " const MCSubtargetInfo &STI,\n"
1145 << " unsigned PredicateIndex);\n";
1146
1147 O << HeaderO.str();
1148 O.indent(2) << "static const PatternsForOpcode OpToPatterns[] = {\n";
1149 O << OpcodeO.str();
1150 O.indent(2) << "};\n\n";
1151 O.indent(2) << "static const AliasPattern Patterns[] = {\n";
1152 O << PatternO.str();
1153 O.indent(2) << "};\n\n";
1154 O.indent(2) << "static const AliasPatternCond Conds[] = {\n";
1155 O << CondO.str();
1156 O.indent(2) << "};\n\n";
1157 O.indent(2) << "static const char AsmStrings[] =\n";
1158 for (const auto &P : AsmStrings) {
1159 O.indent(4) << "/* " << P.first << " */ \"" << P.second << "\\0\"\n";
1160 }
1161
1162 O.indent(2) << ";\n\n";
1163
1164 // Assert that the opcode table is sorted. Use a static local constructor to
1165 // ensure that the check only happens once on first run.
1166 O << "#ifndef NDEBUG\n";
1167 O.indent(2) << "static struct SortCheck {\n";
1168 O.indent(2) << " SortCheck(ArrayRef<PatternsForOpcode> OpToPatterns) {\n";
1169 O.indent(2) << " assert(std::is_sorted(\n";
1170 O.indent(2) << " OpToPatterns.begin(), OpToPatterns.end(),\n";
1171 O.indent(2) << " [](const PatternsForOpcode &L, const "
1172 "PatternsForOpcode &R) {\n";
1173 O.indent(2) << " return L.Opcode < R.Opcode;\n";
1174 O.indent(2) << " }) &&\n";
1175 O.indent(2) << " \"tablegen failed to sort opcode patterns\");\n";
1176 O.indent(2) << " }\n";
1177 O.indent(2) << "} sortCheckVar(OpToPatterns);\n";
1178 O << "#endif\n\n";
1179
1180 O.indent(2) << "AliasMatchingData M {\n";
1181 O.indent(2) << " ArrayRef(OpToPatterns),\n";
1182 O.indent(2) << " ArrayRef(Patterns),\n";
1183 O.indent(2) << " ArrayRef(Conds),\n";
1184 O.indent(2) << " StringRef(AsmStrings, std::size(AsmStrings)),\n";
1185 if (MCOpPredicates.empty())
1186 O.indent(2) << " nullptr,\n";
1187 else
1188 O.indent(2) << " &" << Target.getName() << ClassName << "ValidateMCOperand,\n";
1189 O.indent(2) << "};\n";
1190
1191 O.indent(2) << "const char *AsmString = matchAliasPatterns(MI, "
1192 << (PassSubtarget ? "&STI" : "nullptr") << ", M);\n";
1193 O.indent(2) << "if (!AsmString) return false;\n\n";
1194
1195 // Code that prints the alias, replacing the operands with the ones from the
1196 // MCInst.
1197 O << " unsigned I = 0;\n";
1198 O << " while (AsmString[I] != ' ' && AsmString[I] != '\\t' &&\n";
1199 O << " AsmString[I] != '$' && AsmString[I] != '\\0')\n";
1200 O << " ++I;\n";
1201 O << " OS << '\\t' << StringRef(AsmString, I);\n";
1202
1203 O << " if (AsmString[I] != '\\0') {\n";
1204 O << " if (AsmString[I] == ' ' || AsmString[I] == '\\t') {\n";
1205 O << " OS << '\\t';\n";
1206 O << " ++I;\n";
1207 O << " }\n";
1208 O << " do {\n";
1209 O << " if (AsmString[I] == '$') {\n";
1210 O << " ++I;\n";
1211 O << " if (AsmString[I] == (char)0xff) {\n";
1212 O << " ++I;\n";
1213 O << " int OpIdx = AsmString[I++] - 1;\n";
1214 O << " int PrintMethodIdx = AsmString[I++] - 1;\n";
1215 O << " printCustomAliasOperand(MI, Address, OpIdx, PrintMethodIdx, ";
1216 O << (PassSubtarget ? "STI, " : "");
1217 O << "OS);\n";
1218 O << " } else\n";
1219 O << " printOperand(MI, unsigned(AsmString[I++]) - 1, ";
1220 O << (PassSubtarget ? "STI, " : "");
1221 O << "OS);\n";
1222 O << " } else {\n";
1223 O << " OS << AsmString[I++];\n";
1224 O << " }\n";
1225 O << " } while (AsmString[I] != '\\0');\n";
1226 O << " }\n\n";
1227
1228 O << " return true;\n";
1229 O << "}\n\n";
1230
1231 //////////////////////////////
1232 // Write out the printCustomAliasOperand function
1233 //////////////////////////////
1234
1235 O << "void " << Target.getName() << ClassName << "::"
1236 << "printCustomAliasOperand(\n"
1237 << " const MCInst *MI, uint64_t Address, unsigned OpIdx,\n"
1238 << " unsigned PrintMethodIdx,\n"
1239 << (PassSubtarget ? " const MCSubtargetInfo &STI,\n" : "")
1240 << " raw_ostream &OS) {\n";
1241 if (PrintMethods.empty())
1242 O << " llvm_unreachable(\"Unknown PrintMethod kind\");\n";
1243 else {
1244 O << " switch (PrintMethodIdx) {\n"
1245 << " default:\n"
1246 << " llvm_unreachable(\"Unknown PrintMethod kind\");\n"
1247 << " break;\n";
1248
1249 for (unsigned i = 0; i < PrintMethods.size(); ++i) {
1250 O << " case " << i << ":\n"
1251 << " " << PrintMethods[i].first << "(MI, "
1252 << (PrintMethods[i].second ? "Address, " : "") << "OpIdx, "
1253 << (PassSubtarget ? "STI, " : "") << "OS);\n"
1254 << " break;\n";
1255 }
1256 O << " }\n";
1257 }
1258 O << "}\n\n";
1259
1260 if (!MCOpPredicates.empty()) {
1261 O << "static bool " << Target.getName() << ClassName
1262 << "ValidateMCOperand(const MCOperand &MCOp,\n"
1263 << " const MCSubtargetInfo &STI,\n"
1264 << " unsigned PredicateIndex) {\n"
1265 << " switch (PredicateIndex) {\n"
1266 << " default:\n"
1267 << " llvm_unreachable(\"Unknown MCOperandPredicate kind\");\n"
1268 << " break;\n";
1269
1270 for (unsigned i = 0; i < MCOpPredicates.size(); ++i) {
1271 StringRef MCOpPred = MCOpPredicates[i]->getValueAsString("MCOperandPredicate");
1272 O << " case " << i + 1 << ": {\n"
1273 << MCOpPred.data() << "\n"
1274 << " }\n";
1275 }
1276 O << " }\n"
1277 << "}\n\n";
1278 }
1279
1280 O << "#endif // PRINT_ALIAS_INSTR\n";
1281 }
1282
AsmWriterEmitter(RecordKeeper & R)1283 AsmWriterEmitter::AsmWriterEmitter(RecordKeeper &R) : Records(R), Target(R) {
1284 Record *AsmWriter = Target.getAsmWriter();
1285 unsigned Variant = AsmWriter->getValueAsInt("Variant");
1286
1287 // Get the instruction numbering.
1288 NumberedInstructions = Target.getInstructionsByEnumValue();
1289
1290 for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
1291 const CodeGenInstruction *I = NumberedInstructions[i];
1292 if (!I->AsmString.empty() && I->TheDef->getName() != "PHI")
1293 Instructions.emplace_back(*I, i, Variant);
1294 }
1295 }
1296
run(raw_ostream & O)1297 void AsmWriterEmitter::run(raw_ostream &O) {
1298 std::vector<std::vector<std::string>> TableDrivenOperandPrinters;
1299 unsigned BitsLeft = 0;
1300 unsigned AsmStrBits = 0;
1301 EmitGetMnemonic(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1302 EmitPrintInstruction(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1303 EmitGetRegisterName(O);
1304 EmitPrintAliasInstruction(O);
1305 }
1306
1307 namespace llvm {
1308
EmitAsmWriter(RecordKeeper & RK,raw_ostream & OS)1309 void EmitAsmWriter(RecordKeeper &RK, raw_ostream &OS) {
1310 emitSourceFileHeader("Assembly Writer Source Fragment", OS);
1311 AsmWriterEmitter(RK).run(OS);
1312 }
1313
1314 } // end namespace llvm
1315