1 //===- AsmWriterEmitter.cpp - Generate an assembly writer -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This tablegen backend is emits an assembly printer for the current target.
11 // Note that this is currently fairly skeletal, but will grow over time.
13 //===----------------------------------------------------------------------===//
15 #include "AsmWriterEmitter.h"
16 #include "CodeGenTarget.h"
22 static bool isIdentChar(char C) {
23 return (C >= 'a' && C <= 'z') ||
24 (C >= 'A' && C <= 'Z') ||
25 (C >= '0' && C <= '9') ||
30 struct AsmWriterOperand {
31 enum { isLiteralTextOperand, isMachineInstrOperand } OperandType;
33 /// Str - For isLiteralTextOperand, this IS the literal text. For
34 /// isMachineInstrOperand, this is the PrinterMethodName for the operand.
37 /// MiOpNo - For isMachineInstrOperand, this is the operand number of the
38 /// machine instruction.
41 /// OpVT - For isMachineInstrOperand, this is the value type for the
45 AsmWriterOperand(const std::string &LitStr)
46 : OperandType(isLiteralTextOperand), Str(LitStr) {}
48 AsmWriterOperand(const std::string &Printer, unsigned OpNo,
49 MVT::ValueType VT) : OperandType(isMachineInstrOperand),
50 Str(Printer), MIOpNo(OpNo), OpVT(VT){}
52 bool operator!=(const AsmWriterOperand &Other) const {
53 if (OperandType != Other.OperandType || Str != Other.Str) return true;
54 if (OperandType == isMachineInstrOperand)
55 return MIOpNo != Other.MIOpNo || OpVT != Other.OpVT;
58 bool operator==(const AsmWriterOperand &Other) const {
59 return !operator!=(Other);
61 void EmitCode(std::ostream &OS) const;
64 struct AsmWriterInst {
65 std::vector<AsmWriterOperand> Operands;
66 const CodeGenInstruction *CGI;
68 AsmWriterInst(const CodeGenInstruction &CGI, unsigned Variant);
70 /// MatchesAllButOneOp - If this instruction is exactly identical to the
71 /// specified instruction except for one differing operand, return the
72 /// differing operand number. Otherwise return ~0.
73 unsigned MatchesAllButOneOp(const AsmWriterInst &Other) const;
76 void AddLiteralString(const std::string &Str) {
77 // If the last operand was already a literal text string, append this to
78 // it, otherwise add a new operand.
79 if (!Operands.empty() &&
80 Operands.back().OperandType == AsmWriterOperand::isLiteralTextOperand)
81 Operands.back().Str.append(Str);
83 Operands.push_back(AsmWriterOperand(Str));
89 void AsmWriterOperand::EmitCode(std::ostream &OS) const {
90 if (OperandType == isLiteralTextOperand)
91 OS << "O << \"" << Str << "\"; ";
93 OS << Str << "(MI, " << MIOpNo << ", MVT::" << getEnumName(OpVT) << "); ";
97 /// ParseAsmString - Parse the specified Instruction's AsmString into this
100 AsmWriterInst::AsmWriterInst(const CodeGenInstruction &CGI, unsigned Variant) {
102 bool inVariant = false; // True if we are inside a {.|.|.} region.
104 const std::string &AsmString = CGI.AsmString;
105 std::string::size_type LastEmitted = 0;
106 while (LastEmitted != AsmString.size()) {
107 std::string::size_type DollarPos =
108 AsmString.find_first_of("${|}", LastEmitted);
109 if (DollarPos == std::string::npos) DollarPos = AsmString.size();
111 // Emit a constant string fragment.
112 if (DollarPos != LastEmitted) {
113 // TODO: this should eventually handle escaping.
114 AddLiteralString(std::string(AsmString.begin()+LastEmitted,
115 AsmString.begin()+DollarPos));
116 LastEmitted = DollarPos;
117 } else if (AsmString[DollarPos] == '{') {
119 throw "Nested variants found for instruction '" +
120 CGI.TheDef->getName() + "'!";
121 LastEmitted = DollarPos+1;
122 inVariant = true; // We are now inside of the variant!
123 for (unsigned i = 0; i != Variant; ++i) {
124 // Skip over all of the text for an irrelevant variant here. The
125 // next variant starts at |, or there may not be text for this
126 // variant if we see a }.
127 std::string::size_type NP =
128 AsmString.find_first_of("|}", LastEmitted);
129 if (NP == std::string::npos)
130 throw "Incomplete variant for instruction '" +
131 CGI.TheDef->getName() + "'!";
133 if (AsmString[NP] == '}') {
134 inVariant = false; // No text for this variant.
138 } else if (AsmString[DollarPos] == '|') {
140 throw "'|' character found outside of a variant in instruction '"
141 + CGI.TheDef->getName() + "'!";
142 // Move to the end of variant list.
143 std::string::size_type NP = AsmString.find('}', LastEmitted);
144 if (NP == std::string::npos)
145 throw "Incomplete variant for instruction '" +
146 CGI.TheDef->getName() + "'!";
149 } else if (AsmString[DollarPos] == '}') {
151 throw "'}' character found outside of a variant in instruction '"
152 + CGI.TheDef->getName() + "'!";
153 LastEmitted = DollarPos+1;
155 } else if (DollarPos+1 != AsmString.size() &&
156 AsmString[DollarPos+1] == '$') {
157 AddLiteralString("$"); // "$$" -> $
158 LastEmitted = DollarPos+2;
160 // Get the name of the variable.
161 std::string::size_type VarEnd = DollarPos+1;
163 // handle ${foo}bar as $foo by detecting whether the character following
164 // the dollar sign is a curly brace. If so, advance VarEnd and DollarPos
165 // so the variable name does not contain the leading curly brace.
166 bool hasCurlyBraces = false;
167 if (VarEnd < AsmString.size() && '{' == AsmString[VarEnd]) {
168 hasCurlyBraces = true;
173 while (VarEnd < AsmString.size() && isIdentChar(AsmString[VarEnd]))
175 std::string VarName(AsmString.begin()+DollarPos+1,
176 AsmString.begin()+VarEnd);
178 // In order to avoid starting the next string at the terminating curly
179 // brace, advance the end position past it if we found an opening curly
181 if (hasCurlyBraces) {
182 if (VarEnd >= AsmString.size())
183 throw "Reached end of string before terminating curly brace in '"
184 + CGI.TheDef->getName() + "'";
185 if (AsmString[VarEnd] != '}')
186 throw "Variant name beginning with '{' did not end with '}' in '"
187 + CGI.TheDef->getName() + "'";
191 throw "Stray '$' in '" + CGI.TheDef->getName() +
192 "' asm string, maybe you want $$?";
194 unsigned OpNo = CGI.getOperandNamed(VarName);
195 CodeGenInstruction::OperandInfo OpInfo = CGI.OperandList[OpNo];
197 // If this is a two-address instruction and we are not accessing the
198 // 0th operand, remove an operand.
199 unsigned MIOp = OpInfo.MIOperandNo;
200 if (CGI.isTwoAddress && MIOp != 0) {
202 throw "Should refer to operand #0 instead of #1 for two-address"
203 " instruction '" + CGI.TheDef->getName() + "'!";
207 Operands.push_back(AsmWriterOperand(OpInfo.PrinterMethodName,
209 LastEmitted = VarEnd;
213 AddLiteralString("\\n");
216 /// MatchesAllButOneOp - If this instruction is exactly identical to the
217 /// specified instruction except for one differing operand, return the differing
218 /// operand number. If more than one operand mismatches, return ~1, otherwise
219 /// if the instructions are identical return ~0.
220 unsigned AsmWriterInst::MatchesAllButOneOp(const AsmWriterInst &Other)const{
221 if (Operands.size() != Other.Operands.size()) return ~1;
223 unsigned MismatchOperand = ~0U;
224 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
225 if (Operands[i] != Other.Operands[i])
226 if (MismatchOperand != ~0U) // Already have one mismatch?
231 return MismatchOperand;
234 static void PrintCases(std::vector<std::pair<std::string,
235 AsmWriterOperand> > &OpsToPrint, std::ostream &O) {
236 O << " case " << OpsToPrint.back().first << ": ";
237 AsmWriterOperand TheOp = OpsToPrint.back().second;
238 OpsToPrint.pop_back();
240 // Check to see if any other operands are identical in this list, and if so,
241 // emit a case label for them.
242 for (unsigned i = OpsToPrint.size(); i != 0; --i)
243 if (OpsToPrint[i-1].second == TheOp) {
244 O << "\n case " << OpsToPrint[i-1].first << ": ";
245 OpsToPrint.erase(OpsToPrint.begin()+i-1);
248 // Finally, emit the code.
254 /// EmitInstructions - Emit the last instruction in the vector and any other
255 /// instructions that are suitably similar to it.
256 static void EmitInstructions(std::vector<AsmWriterInst> &Insts,
258 AsmWriterInst FirstInst = Insts.back();
261 std::vector<AsmWriterInst> SimilarInsts;
262 unsigned DifferingOperand = ~0;
263 for (unsigned i = Insts.size(); i != 0; --i) {
264 unsigned DiffOp = Insts[i-1].MatchesAllButOneOp(FirstInst);
266 if (DifferingOperand == ~0U) // First match!
267 DifferingOperand = DiffOp;
269 // If this differs in the same operand as the rest of the instructions in
270 // this class, move it to the SimilarInsts list.
271 if (DifferingOperand == DiffOp || DiffOp == ~0U) {
272 SimilarInsts.push_back(Insts[i-1]);
273 Insts.erase(Insts.begin()+i-1);
278 std::string Namespace = FirstInst.CGI->Namespace;
280 O << " case " << Namespace << "::"
281 << FirstInst.CGI->TheDef->getName() << ":\n";
282 for (unsigned i = 0, e = SimilarInsts.size(); i != e; ++i)
283 O << " case " << Namespace << "::"
284 << SimilarInsts[i].CGI->TheDef->getName() << ":\n";
285 for (unsigned i = 0, e = FirstInst.Operands.size(); i != e; ++i) {
286 if (i != DifferingOperand) {
287 // If the operand is the same for all instructions, just print it.
289 FirstInst.Operands[i].EmitCode(O);
291 // If this is the operand that varies between all of the instructions,
292 // emit a switch for just this operand now.
293 O << " switch (MI->getOpcode()) {\n";
294 std::vector<std::pair<std::string, AsmWriterOperand> > OpsToPrint;
295 OpsToPrint.push_back(std::make_pair(Namespace+"::"+
296 FirstInst.CGI->TheDef->getName(),
297 FirstInst.Operands[i]));
299 for (unsigned si = 0, e = SimilarInsts.size(); si != e; ++si) {
300 AsmWriterInst &AWI = SimilarInsts[si];
301 OpsToPrint.push_back(std::make_pair(Namespace+"::"+
302 AWI.CGI->TheDef->getName(),
305 std::reverse(OpsToPrint.begin(), OpsToPrint.end());
306 while (!OpsToPrint.empty())
307 PrintCases(OpsToPrint, O);
316 void AsmWriterEmitter::run(std::ostream &O) {
317 EmitSourceFileHeader("Assembly Writer Source Fragment", O);
319 CodeGenTarget Target;
320 Record *AsmWriter = Target.getAsmWriter();
321 std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
322 unsigned Variant = AsmWriter->getValueAsInt("Variant");
325 "/// printInstruction - This method is automatically generated by tablegen\n"
326 "/// from the instruction set description. This method returns true if the\n"
327 "/// machine instruction was sufficiently described to print it, otherwise\n"
328 "/// it returns false.\n"
329 "bool " << Target.getName() << ClassName
330 << "::printInstruction(const MachineInstr *MI) {\n";
332 std::string Namespace = Target.inst_begin()->second.Namespace;
334 std::vector<AsmWriterInst> Instructions;
336 for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
337 E = Target.inst_end(); I != E; ++I)
338 if (!I->second.AsmString.empty())
339 Instructions.push_back(AsmWriterInst(I->second, Variant));
341 // If all of the instructions start with a constant string (a very very common
342 // occurance), emit all of the constant strings as a big table lookup instead
343 // of requiring a switch for them.
344 bool AllStartWithString = true;
346 for (unsigned i = 0, e = Instructions.size(); i != e; ++i)
347 if (Instructions[i].Operands.empty() ||
348 Instructions[i].Operands[0].OperandType !=
349 AsmWriterOperand::isLiteralTextOperand) {
350 AllStartWithString = false;
354 if (AllStartWithString) {
355 // Compute the CodeGenInstruction -> AsmWriterInst mapping. Note that not
356 // all machine instructions are necessarily being printed, so there may be
357 // target instructions not in this map.
358 std::map<const CodeGenInstruction*, AsmWriterInst*> CGIAWIMap;
359 for (unsigned i = 0, e = Instructions.size(); i != e; ++i)
360 CGIAWIMap.insert(std::make_pair(Instructions[i].CGI, &Instructions[i]));
362 // Emit a table of constant strings.
363 std::vector<const CodeGenInstruction*> NumberedInstructions;
364 Target.getInstructionsByEnumValue(NumberedInstructions);
366 O << " static const char * const OpStrs[] = {\n";
367 for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
368 AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions[i]];
370 // Something not handled by the asmwriter printer.
373 O << " \"" << AWI->Operands[0].Str << "\",\t// ";
374 // Nuke the string from the operand list. It is now handled!
375 AWI->Operands.erase(AWI->Operands.begin());
377 O << NumberedInstructions[i]->TheDef->getName() << "\n";
380 << " // Emit the opcode for the instruction.\n"
381 << " if (const char *AsmStr = OpStrs[MI->getOpcode()])\n"
382 << " O << AsmStr;\n\n";
385 // Because this is a vector we want to emit from the end. Reverse all of the
386 // elements in the vector.
387 std::reverse(Instructions.begin(), Instructions.end());
389 O << " switch (MI->getOpcode()) {\n"
390 " default: return false;\n";
392 while (!Instructions.empty())
393 EmitInstructions(Instructions, O);