Use %ull here.
[llvm/stm8.git] / lib / Target / X86 / AsmParser / X86AsmParser.cpp
blobc352bfcd8cceb8384772a4c894803f4ed4d1a3fe
1 //===-- X86AsmParser.cpp - Parse X86 assembly to MCInst instructions ------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
10 #include "llvm/Target/TargetAsmParser.h"
11 #include "X86.h"
12 #include "X86Subtarget.h"
13 #include "llvm/Target/TargetRegistry.h"
14 #include "llvm/Target/TargetAsmParser.h"
15 #include "llvm/MC/MCStreamer.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCInst.h"
18 #include "llvm/MC/MCParser/MCAsmLexer.h"
19 #include "llvm/MC/MCParser/MCAsmParser.h"
20 #include "llvm/MC/MCParser/MCParsedAsmOperand.h"
21 #include "llvm/ADT/SmallString.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/StringExtras.h"
24 #include "llvm/ADT/StringSwitch.h"
25 #include "llvm/ADT/Twine.h"
26 #include "llvm/Support/SourceMgr.h"
27 #include "llvm/Support/raw_ostream.h"
28 using namespace llvm;
30 namespace {
31 struct X86Operand;
33 class X86ATTAsmParser : public TargetAsmParser {
34 MCAsmParser &Parser;
35 TargetMachine &TM;
37 protected:
38 unsigned Is64Bit : 1;
40 private:
41 MCAsmParser &getParser() const { return Parser; }
43 MCAsmLexer &getLexer() const { return Parser.getLexer(); }
45 bool Error(SMLoc L, const Twine &Msg) { return Parser.Error(L, Msg); }
47 X86Operand *ParseOperand();
48 X86Operand *ParseMemOperand(unsigned SegReg, SMLoc StartLoc);
50 bool ParseDirectiveWord(unsigned Size, SMLoc L);
52 bool MatchAndEmitInstruction(SMLoc IDLoc,
53 SmallVectorImpl<MCParsedAsmOperand*> &Operands,
54 MCStreamer &Out);
56 /// isSrcOp - Returns true if operand is either (%rsi) or %ds:%(rsi)
57 /// in 64bit mode or (%edi) or %es:(%edi) in 32bit mode.
58 bool isSrcOp(X86Operand &Op);
60 /// isDstOp - Returns true if operand is either %es:(%rdi) in 64bit mode
61 /// or %es:(%edi) in 32bit mode.
62 bool isDstOp(X86Operand &Op);
64 /// @name Auto-generated Matcher Functions
65 /// {
67 #define GET_ASSEMBLER_HEADER
68 #include "X86GenAsmMatcher.inc"
70 /// }
72 public:
73 X86ATTAsmParser(const Target &T, MCAsmParser &parser, TargetMachine &TM)
74 : TargetAsmParser(T), Parser(parser), TM(TM) {
76 // Initialize the set of available features.
77 setAvailableFeatures(ComputeAvailableFeatures(
78 &TM.getSubtarget<X86Subtarget>()));
80 virtual bool ParseRegister(unsigned &RegNo, SMLoc &StartLoc, SMLoc &EndLoc);
82 virtual bool ParseInstruction(StringRef Name, SMLoc NameLoc,
83 SmallVectorImpl<MCParsedAsmOperand*> &Operands);
85 virtual bool ParseDirective(AsmToken DirectiveID);
88 class X86_32ATTAsmParser : public X86ATTAsmParser {
89 public:
90 X86_32ATTAsmParser(const Target &T, MCAsmParser &Parser, TargetMachine &TM)
91 : X86ATTAsmParser(T, Parser, TM) {
92 Is64Bit = false;
96 class X86_64ATTAsmParser : public X86ATTAsmParser {
97 public:
98 X86_64ATTAsmParser(const Target &T, MCAsmParser &Parser, TargetMachine &TM)
99 : X86ATTAsmParser(T, Parser, TM) {
100 Is64Bit = true;
104 } // end anonymous namespace
106 /// @name Auto-generated Match Functions
107 /// {
109 static unsigned MatchRegisterName(StringRef Name);
111 /// }
113 namespace {
115 /// X86Operand - Instances of this class represent a parsed X86 machine
116 /// instruction.
117 struct X86Operand : public MCParsedAsmOperand {
118 enum KindTy {
119 Token,
120 Register,
121 Immediate,
122 Memory
123 } Kind;
125 SMLoc StartLoc, EndLoc;
127 union {
128 struct {
129 const char *Data;
130 unsigned Length;
131 } Tok;
133 struct {
134 unsigned RegNo;
135 } Reg;
137 struct {
138 const MCExpr *Val;
139 } Imm;
141 struct {
142 unsigned SegReg;
143 const MCExpr *Disp;
144 unsigned BaseReg;
145 unsigned IndexReg;
146 unsigned Scale;
147 } Mem;
150 X86Operand(KindTy K, SMLoc Start, SMLoc End)
151 : Kind(K), StartLoc(Start), EndLoc(End) {}
153 /// getStartLoc - Get the location of the first token of this operand.
154 SMLoc getStartLoc() const { return StartLoc; }
155 /// getEndLoc - Get the location of the last token of this operand.
156 SMLoc getEndLoc() const { return EndLoc; }
158 virtual void dump(raw_ostream &OS) const {}
160 StringRef getToken() const {
161 assert(Kind == Token && "Invalid access!");
162 return StringRef(Tok.Data, Tok.Length);
164 void setTokenValue(StringRef Value) {
165 assert(Kind == Token && "Invalid access!");
166 Tok.Data = Value.data();
167 Tok.Length = Value.size();
170 unsigned getReg() const {
171 assert(Kind == Register && "Invalid access!");
172 return Reg.RegNo;
175 const MCExpr *getImm() const {
176 assert(Kind == Immediate && "Invalid access!");
177 return Imm.Val;
180 const MCExpr *getMemDisp() const {
181 assert(Kind == Memory && "Invalid access!");
182 return Mem.Disp;
184 unsigned getMemSegReg() const {
185 assert(Kind == Memory && "Invalid access!");
186 return Mem.SegReg;
188 unsigned getMemBaseReg() const {
189 assert(Kind == Memory && "Invalid access!");
190 return Mem.BaseReg;
192 unsigned getMemIndexReg() const {
193 assert(Kind == Memory && "Invalid access!");
194 return Mem.IndexReg;
196 unsigned getMemScale() const {
197 assert(Kind == Memory && "Invalid access!");
198 return Mem.Scale;
201 bool isToken() const {return Kind == Token; }
203 bool isImm() const { return Kind == Immediate; }
205 bool isImmSExti16i8() const {
206 if (!isImm())
207 return false;
209 // If this isn't a constant expr, just assume it fits and let relaxation
210 // handle it.
211 const MCConstantExpr *CE = dyn_cast<MCConstantExpr>(getImm());
212 if (!CE)
213 return true;
215 // Otherwise, check the value is in a range that makes sense for this
216 // extension.
217 uint64_t Value = CE->getValue();
218 return (( Value <= 0x000000000000007FULL)||
219 (0x000000000000FF80ULL <= Value && Value <= 0x000000000000FFFFULL)||
220 (0xFFFFFFFFFFFFFF80ULL <= Value && Value <= 0xFFFFFFFFFFFFFFFFULL));
222 bool isImmSExti32i8() const {
223 if (!isImm())
224 return false;
226 // If this isn't a constant expr, just assume it fits and let relaxation
227 // handle it.
228 const MCConstantExpr *CE = dyn_cast<MCConstantExpr>(getImm());
229 if (!CE)
230 return true;
232 // Otherwise, check the value is in a range that makes sense for this
233 // extension.
234 uint64_t Value = CE->getValue();
235 return (( Value <= 0x000000000000007FULL)||
236 (0x00000000FFFFFF80ULL <= Value && Value <= 0x00000000FFFFFFFFULL)||
237 (0xFFFFFFFFFFFFFF80ULL <= Value && Value <= 0xFFFFFFFFFFFFFFFFULL));
239 bool isImmSExti64i8() const {
240 if (!isImm())
241 return false;
243 // If this isn't a constant expr, just assume it fits and let relaxation
244 // handle it.
245 const MCConstantExpr *CE = dyn_cast<MCConstantExpr>(getImm());
246 if (!CE)
247 return true;
249 // Otherwise, check the value is in a range that makes sense for this
250 // extension.
251 uint64_t Value = CE->getValue();
252 return (( Value <= 0x000000000000007FULL)||
253 (0xFFFFFFFFFFFFFF80ULL <= Value && Value <= 0xFFFFFFFFFFFFFFFFULL));
255 bool isImmSExti64i32() const {
256 if (!isImm())
257 return false;
259 // If this isn't a constant expr, just assume it fits and let relaxation
260 // handle it.
261 const MCConstantExpr *CE = dyn_cast<MCConstantExpr>(getImm());
262 if (!CE)
263 return true;
265 // Otherwise, check the value is in a range that makes sense for this
266 // extension.
267 uint64_t Value = CE->getValue();
268 return (( Value <= 0x000000007FFFFFFFULL)||
269 (0xFFFFFFFF80000000ULL <= Value && Value <= 0xFFFFFFFFFFFFFFFFULL));
272 bool isMem() const { return Kind == Memory; }
274 bool isAbsMem() const {
275 return Kind == Memory && !getMemSegReg() && !getMemBaseReg() &&
276 !getMemIndexReg() && getMemScale() == 1;
279 bool isReg() const { return Kind == Register; }
281 void addExpr(MCInst &Inst, const MCExpr *Expr) const {
282 // Add as immediates when possible.
283 if (const MCConstantExpr *CE = dyn_cast<MCConstantExpr>(Expr))
284 Inst.addOperand(MCOperand::CreateImm(CE->getValue()));
285 else
286 Inst.addOperand(MCOperand::CreateExpr(Expr));
289 void addRegOperands(MCInst &Inst, unsigned N) const {
290 assert(N == 1 && "Invalid number of operands!");
291 Inst.addOperand(MCOperand::CreateReg(getReg()));
294 void addImmOperands(MCInst &Inst, unsigned N) const {
295 assert(N == 1 && "Invalid number of operands!");
296 addExpr(Inst, getImm());
299 void addMemOperands(MCInst &Inst, unsigned N) const {
300 assert((N == 5) && "Invalid number of operands!");
301 Inst.addOperand(MCOperand::CreateReg(getMemBaseReg()));
302 Inst.addOperand(MCOperand::CreateImm(getMemScale()));
303 Inst.addOperand(MCOperand::CreateReg(getMemIndexReg()));
304 addExpr(Inst, getMemDisp());
305 Inst.addOperand(MCOperand::CreateReg(getMemSegReg()));
308 void addAbsMemOperands(MCInst &Inst, unsigned N) const {
309 assert((N == 1) && "Invalid number of operands!");
310 Inst.addOperand(MCOperand::CreateExpr(getMemDisp()));
313 static X86Operand *CreateToken(StringRef Str, SMLoc Loc) {
314 X86Operand *Res = new X86Operand(Token, Loc, Loc);
315 Res->Tok.Data = Str.data();
316 Res->Tok.Length = Str.size();
317 return Res;
320 static X86Operand *CreateReg(unsigned RegNo, SMLoc StartLoc, SMLoc EndLoc) {
321 X86Operand *Res = new X86Operand(Register, StartLoc, EndLoc);
322 Res->Reg.RegNo = RegNo;
323 return Res;
326 static X86Operand *CreateImm(const MCExpr *Val, SMLoc StartLoc, SMLoc EndLoc){
327 X86Operand *Res = new X86Operand(Immediate, StartLoc, EndLoc);
328 Res->Imm.Val = Val;
329 return Res;
332 /// Create an absolute memory operand.
333 static X86Operand *CreateMem(const MCExpr *Disp, SMLoc StartLoc,
334 SMLoc EndLoc) {
335 X86Operand *Res = new X86Operand(Memory, StartLoc, EndLoc);
336 Res->Mem.SegReg = 0;
337 Res->Mem.Disp = Disp;
338 Res->Mem.BaseReg = 0;
339 Res->Mem.IndexReg = 0;
340 Res->Mem.Scale = 1;
341 return Res;
344 /// Create a generalized memory operand.
345 static X86Operand *CreateMem(unsigned SegReg, const MCExpr *Disp,
346 unsigned BaseReg, unsigned IndexReg,
347 unsigned Scale, SMLoc StartLoc, SMLoc EndLoc) {
348 // We should never just have a displacement, that should be parsed as an
349 // absolute memory operand.
350 assert((SegReg || BaseReg || IndexReg) && "Invalid memory operand!");
352 // The scale should always be one of {1,2,4,8}.
353 assert(((Scale == 1 || Scale == 2 || Scale == 4 || Scale == 8)) &&
354 "Invalid scale!");
355 X86Operand *Res = new X86Operand(Memory, StartLoc, EndLoc);
356 Res->Mem.SegReg = SegReg;
357 Res->Mem.Disp = Disp;
358 Res->Mem.BaseReg = BaseReg;
359 Res->Mem.IndexReg = IndexReg;
360 Res->Mem.Scale = Scale;
361 return Res;
365 } // end anonymous namespace.
367 bool X86ATTAsmParser::isSrcOp(X86Operand &Op) {
368 unsigned basereg = Is64Bit ? X86::RSI : X86::ESI;
370 return (Op.isMem() &&
371 (Op.Mem.SegReg == 0 || Op.Mem.SegReg == X86::DS) &&
372 isa<MCConstantExpr>(Op.Mem.Disp) &&
373 cast<MCConstantExpr>(Op.Mem.Disp)->getValue() == 0 &&
374 Op.Mem.BaseReg == basereg && Op.Mem.IndexReg == 0);
377 bool X86ATTAsmParser::isDstOp(X86Operand &Op) {
378 unsigned basereg = Is64Bit ? X86::RDI : X86::EDI;
380 return Op.isMem() && Op.Mem.SegReg == X86::ES &&
381 isa<MCConstantExpr>(Op.Mem.Disp) &&
382 cast<MCConstantExpr>(Op.Mem.Disp)->getValue() == 0 &&
383 Op.Mem.BaseReg == basereg && Op.Mem.IndexReg == 0;
386 bool X86ATTAsmParser::ParseRegister(unsigned &RegNo,
387 SMLoc &StartLoc, SMLoc &EndLoc) {
388 RegNo = 0;
389 const AsmToken &TokPercent = Parser.getTok();
390 assert(TokPercent.is(AsmToken::Percent) && "Invalid token kind!");
391 StartLoc = TokPercent.getLoc();
392 Parser.Lex(); // Eat percent token.
394 const AsmToken &Tok = Parser.getTok();
395 if (Tok.isNot(AsmToken::Identifier))
396 return Error(Tok.getLoc(), "invalid register name");
398 // FIXME: Validate register for the current architecture; we have to do
399 // validation later, so maybe there is no need for this here.
400 RegNo = MatchRegisterName(Tok.getString());
402 // If the match failed, try the register name as lowercase.
403 if (RegNo == 0)
404 RegNo = MatchRegisterName(LowercaseString(Tok.getString()));
406 // FIXME: This should be done using Requires<In32BitMode> and
407 // Requires<In64BitMode> so "eiz" usage in 64-bit instructions
408 // can be also checked.
409 if (RegNo == X86::RIZ && !Is64Bit)
410 return Error(Tok.getLoc(), "riz register in 64-bit mode only");
412 // Parse "%st" as "%st(0)" and "%st(1)", which is multiple tokens.
413 if (RegNo == 0 && (Tok.getString() == "st" || Tok.getString() == "ST")) {
414 RegNo = X86::ST0;
415 EndLoc = Tok.getLoc();
416 Parser.Lex(); // Eat 'st'
418 // Check to see if we have '(4)' after %st.
419 if (getLexer().isNot(AsmToken::LParen))
420 return false;
421 // Lex the paren.
422 getParser().Lex();
424 const AsmToken &IntTok = Parser.getTok();
425 if (IntTok.isNot(AsmToken::Integer))
426 return Error(IntTok.getLoc(), "expected stack index");
427 switch (IntTok.getIntVal()) {
428 case 0: RegNo = X86::ST0; break;
429 case 1: RegNo = X86::ST1; break;
430 case 2: RegNo = X86::ST2; break;
431 case 3: RegNo = X86::ST3; break;
432 case 4: RegNo = X86::ST4; break;
433 case 5: RegNo = X86::ST5; break;
434 case 6: RegNo = X86::ST6; break;
435 case 7: RegNo = X86::ST7; break;
436 default: return Error(IntTok.getLoc(), "invalid stack index");
439 if (getParser().Lex().isNot(AsmToken::RParen))
440 return Error(Parser.getTok().getLoc(), "expected ')'");
442 EndLoc = Tok.getLoc();
443 Parser.Lex(); // Eat ')'
444 return false;
447 // If this is "db[0-7]", match it as an alias
448 // for dr[0-7].
449 if (RegNo == 0 && Tok.getString().size() == 3 &&
450 Tok.getString().startswith("db")) {
451 switch (Tok.getString()[2]) {
452 case '0': RegNo = X86::DR0; break;
453 case '1': RegNo = X86::DR1; break;
454 case '2': RegNo = X86::DR2; break;
455 case '3': RegNo = X86::DR3; break;
456 case '4': RegNo = X86::DR4; break;
457 case '5': RegNo = X86::DR5; break;
458 case '6': RegNo = X86::DR6; break;
459 case '7': RegNo = X86::DR7; break;
462 if (RegNo != 0) {
463 EndLoc = Tok.getLoc();
464 Parser.Lex(); // Eat it.
465 return false;
469 if (RegNo == 0)
470 return Error(Tok.getLoc(), "invalid register name");
472 EndLoc = Tok.getLoc();
473 Parser.Lex(); // Eat identifier token.
474 return false;
477 X86Operand *X86ATTAsmParser::ParseOperand() {
478 switch (getLexer().getKind()) {
479 default:
480 // Parse a memory operand with no segment register.
481 return ParseMemOperand(0, Parser.getTok().getLoc());
482 case AsmToken::Percent: {
483 // Read the register.
484 unsigned RegNo;
485 SMLoc Start, End;
486 if (ParseRegister(RegNo, Start, End)) return 0;
487 if (RegNo == X86::EIZ || RegNo == X86::RIZ) {
488 Error(Start, "eiz and riz can only be used as index registers");
489 return 0;
492 // If this is a segment register followed by a ':', then this is the start
493 // of a memory reference, otherwise this is a normal register reference.
494 if (getLexer().isNot(AsmToken::Colon))
495 return X86Operand::CreateReg(RegNo, Start, End);
498 getParser().Lex(); // Eat the colon.
499 return ParseMemOperand(RegNo, Start);
501 case AsmToken::Dollar: {
502 // $42 -> immediate.
503 SMLoc Start = Parser.getTok().getLoc(), End;
504 Parser.Lex();
505 const MCExpr *Val;
506 if (getParser().ParseExpression(Val, End))
507 return 0;
508 return X86Operand::CreateImm(Val, Start, End);
513 /// ParseMemOperand: segment: disp(basereg, indexreg, scale). The '%ds:' prefix
514 /// has already been parsed if present.
515 X86Operand *X86ATTAsmParser::ParseMemOperand(unsigned SegReg, SMLoc MemStart) {
517 // We have to disambiguate a parenthesized expression "(4+5)" from the start
518 // of a memory operand with a missing displacement "(%ebx)" or "(,%eax)". The
519 // only way to do this without lookahead is to eat the '(' and see what is
520 // after it.
521 const MCExpr *Disp = MCConstantExpr::Create(0, getParser().getContext());
522 if (getLexer().isNot(AsmToken::LParen)) {
523 SMLoc ExprEnd;
524 if (getParser().ParseExpression(Disp, ExprEnd)) return 0;
526 // After parsing the base expression we could either have a parenthesized
527 // memory address or not. If not, return now. If so, eat the (.
528 if (getLexer().isNot(AsmToken::LParen)) {
529 // Unless we have a segment register, treat this as an immediate.
530 if (SegReg == 0)
531 return X86Operand::CreateMem(Disp, MemStart, ExprEnd);
532 return X86Operand::CreateMem(SegReg, Disp, 0, 0, 1, MemStart, ExprEnd);
535 // Eat the '('.
536 Parser.Lex();
537 } else {
538 // Okay, we have a '('. We don't know if this is an expression or not, but
539 // so we have to eat the ( to see beyond it.
540 SMLoc LParenLoc = Parser.getTok().getLoc();
541 Parser.Lex(); // Eat the '('.
543 if (getLexer().is(AsmToken::Percent) || getLexer().is(AsmToken::Comma)) {
544 // Nothing to do here, fall into the code below with the '(' part of the
545 // memory operand consumed.
546 } else {
547 SMLoc ExprEnd;
549 // It must be an parenthesized expression, parse it now.
550 if (getParser().ParseParenExpression(Disp, ExprEnd))
551 return 0;
553 // After parsing the base expression we could either have a parenthesized
554 // memory address or not. If not, return now. If so, eat the (.
555 if (getLexer().isNot(AsmToken::LParen)) {
556 // Unless we have a segment register, treat this as an immediate.
557 if (SegReg == 0)
558 return X86Operand::CreateMem(Disp, LParenLoc, ExprEnd);
559 return X86Operand::CreateMem(SegReg, Disp, 0, 0, 1, MemStart, ExprEnd);
562 // Eat the '('.
563 Parser.Lex();
567 // If we reached here, then we just ate the ( of the memory operand. Process
568 // the rest of the memory operand.
569 unsigned BaseReg = 0, IndexReg = 0, Scale = 1;
571 if (getLexer().is(AsmToken::Percent)) {
572 SMLoc L;
573 if (ParseRegister(BaseReg, L, L)) return 0;
574 if (BaseReg == X86::EIZ || BaseReg == X86::RIZ) {
575 Error(L, "eiz and riz can only be used as index registers");
576 return 0;
580 if (getLexer().is(AsmToken::Comma)) {
581 Parser.Lex(); // Eat the comma.
583 // Following the comma we should have either an index register, or a scale
584 // value. We don't support the later form, but we want to parse it
585 // correctly.
587 // Not that even though it would be completely consistent to support syntax
588 // like "1(%eax,,1)", the assembler doesn't. Use "eiz" or "riz" for this.
589 if (getLexer().is(AsmToken::Percent)) {
590 SMLoc L;
591 if (ParseRegister(IndexReg, L, L)) return 0;
593 if (getLexer().isNot(AsmToken::RParen)) {
594 // Parse the scale amount:
595 // ::= ',' [scale-expression]
596 if (getLexer().isNot(AsmToken::Comma)) {
597 Error(Parser.getTok().getLoc(),
598 "expected comma in scale expression");
599 return 0;
601 Parser.Lex(); // Eat the comma.
603 if (getLexer().isNot(AsmToken::RParen)) {
604 SMLoc Loc = Parser.getTok().getLoc();
606 int64_t ScaleVal;
607 if (getParser().ParseAbsoluteExpression(ScaleVal))
608 return 0;
610 // Validate the scale amount.
611 if (ScaleVal != 1 && ScaleVal != 2 && ScaleVal != 4 && ScaleVal != 8){
612 Error(Loc, "scale factor in address must be 1, 2, 4 or 8");
613 return 0;
615 Scale = (unsigned)ScaleVal;
618 } else if (getLexer().isNot(AsmToken::RParen)) {
619 // A scale amount without an index is ignored.
620 // index.
621 SMLoc Loc = Parser.getTok().getLoc();
623 int64_t Value;
624 if (getParser().ParseAbsoluteExpression(Value))
625 return 0;
627 if (Value != 1)
628 Warning(Loc, "scale factor without index register is ignored");
629 Scale = 1;
633 // Ok, we've eaten the memory operand, verify we have a ')' and eat it too.
634 if (getLexer().isNot(AsmToken::RParen)) {
635 Error(Parser.getTok().getLoc(), "unexpected token in memory operand");
636 return 0;
638 SMLoc MemEnd = Parser.getTok().getLoc();
639 Parser.Lex(); // Eat the ')'.
641 return X86Operand::CreateMem(SegReg, Disp, BaseReg, IndexReg, Scale,
642 MemStart, MemEnd);
645 bool X86ATTAsmParser::
646 ParseInstruction(StringRef Name, SMLoc NameLoc,
647 SmallVectorImpl<MCParsedAsmOperand*> &Operands) {
648 StringRef PatchedName = Name;
650 // FIXME: Hack to recognize setneb as setne.
651 if (PatchedName.startswith("set") && PatchedName.endswith("b") &&
652 PatchedName != "setb" && PatchedName != "setnb")
653 PatchedName = PatchedName.substr(0, Name.size()-1);
655 // FIXME: Hack to recognize cmp<comparison code>{ss,sd,ps,pd}.
656 const MCExpr *ExtraImmOp = 0;
657 if ((PatchedName.startswith("cmp") || PatchedName.startswith("vcmp")) &&
658 (PatchedName.endswith("ss") || PatchedName.endswith("sd") ||
659 PatchedName.endswith("ps") || PatchedName.endswith("pd"))) {
660 bool IsVCMP = PatchedName.startswith("vcmp");
661 unsigned SSECCIdx = IsVCMP ? 4 : 3;
662 unsigned SSEComparisonCode = StringSwitch<unsigned>(
663 PatchedName.slice(SSECCIdx, PatchedName.size() - 2))
664 .Case("eq", 0)
665 .Case("lt", 1)
666 .Case("le", 2)
667 .Case("unord", 3)
668 .Case("neq", 4)
669 .Case("nlt", 5)
670 .Case("nle", 6)
671 .Case("ord", 7)
672 .Case("eq_uq", 8)
673 .Case("nge", 9)
674 .Case("ngt", 0x0A)
675 .Case("false", 0x0B)
676 .Case("neq_oq", 0x0C)
677 .Case("ge", 0x0D)
678 .Case("gt", 0x0E)
679 .Case("true", 0x0F)
680 .Case("eq_os", 0x10)
681 .Case("lt_oq", 0x11)
682 .Case("le_oq", 0x12)
683 .Case("unord_s", 0x13)
684 .Case("neq_us", 0x14)
685 .Case("nlt_uq", 0x15)
686 .Case("nle_uq", 0x16)
687 .Case("ord_s", 0x17)
688 .Case("eq_us", 0x18)
689 .Case("nge_uq", 0x19)
690 .Case("ngt_uq", 0x1A)
691 .Case("false_os", 0x1B)
692 .Case("neq_os", 0x1C)
693 .Case("ge_oq", 0x1D)
694 .Case("gt_oq", 0x1E)
695 .Case("true_us", 0x1F)
696 .Default(~0U);
697 if (SSEComparisonCode != ~0U) {
698 ExtraImmOp = MCConstantExpr::Create(SSEComparisonCode,
699 getParser().getContext());
700 if (PatchedName.endswith("ss")) {
701 PatchedName = IsVCMP ? "vcmpss" : "cmpss";
702 } else if (PatchedName.endswith("sd")) {
703 PatchedName = IsVCMP ? "vcmpsd" : "cmpsd";
704 } else if (PatchedName.endswith("ps")) {
705 PatchedName = IsVCMP ? "vcmpps" : "cmpps";
706 } else {
707 assert(PatchedName.endswith("pd") && "Unexpected mnemonic!");
708 PatchedName = IsVCMP ? "vcmppd" : "cmppd";
713 // FIXME: Hack to recognize vpclmul<src1_quadword, src2_quadword>dq
714 if (PatchedName.startswith("vpclmul")) {
715 unsigned CLMULQuadWordSelect = StringSwitch<unsigned>(
716 PatchedName.slice(7, PatchedName.size() - 2))
717 .Case("lqlq", 0x00) // src1[63:0], src2[63:0]
718 .Case("hqlq", 0x01) // src1[127:64], src2[63:0]
719 .Case("lqhq", 0x10) // src1[63:0], src2[127:64]
720 .Case("hqhq", 0x11) // src1[127:64], src2[127:64]
721 .Default(~0U);
722 if (CLMULQuadWordSelect != ~0U) {
723 ExtraImmOp = MCConstantExpr::Create(CLMULQuadWordSelect,
724 getParser().getContext());
725 assert(PatchedName.endswith("dq") && "Unexpected mnemonic!");
726 PatchedName = "vpclmulqdq";
730 Operands.push_back(X86Operand::CreateToken(PatchedName, NameLoc));
732 if (ExtraImmOp)
733 Operands.push_back(X86Operand::CreateImm(ExtraImmOp, NameLoc, NameLoc));
736 // Determine whether this is an instruction prefix.
737 bool isPrefix =
738 Name == "lock" || Name == "rep" ||
739 Name == "repe" || Name == "repz" ||
740 Name == "repne" || Name == "repnz" ||
741 Name == "rex64" || Name == "data16";
744 // This does the actual operand parsing. Don't parse any more if we have a
745 // prefix juxtaposed with an operation like "lock incl 4(%rax)", because we
746 // just want to parse the "lock" as the first instruction and the "incl" as
747 // the next one.
748 if (getLexer().isNot(AsmToken::EndOfStatement) && !isPrefix) {
750 // Parse '*' modifier.
751 if (getLexer().is(AsmToken::Star)) {
752 SMLoc Loc = Parser.getTok().getLoc();
753 Operands.push_back(X86Operand::CreateToken("*", Loc));
754 Parser.Lex(); // Eat the star.
757 // Read the first operand.
758 if (X86Operand *Op = ParseOperand())
759 Operands.push_back(Op);
760 else {
761 Parser.EatToEndOfStatement();
762 return true;
765 while (getLexer().is(AsmToken::Comma)) {
766 Parser.Lex(); // Eat the comma.
768 // Parse and remember the operand.
769 if (X86Operand *Op = ParseOperand())
770 Operands.push_back(Op);
771 else {
772 Parser.EatToEndOfStatement();
773 return true;
777 if (getLexer().isNot(AsmToken::EndOfStatement)) {
778 SMLoc Loc = getLexer().getLoc();
779 Parser.EatToEndOfStatement();
780 return Error(Loc, "unexpected token in argument list");
784 if (getLexer().is(AsmToken::EndOfStatement))
785 Parser.Lex(); // Consume the EndOfStatement
786 else if (isPrefix && getLexer().is(AsmToken::Slash))
787 Parser.Lex(); // Consume the prefix separator Slash
789 // This is a terrible hack to handle "out[bwl]? %al, (%dx)" ->
790 // "outb %al, %dx". Out doesn't take a memory form, but this is a widely
791 // documented form in various unofficial manuals, so a lot of code uses it.
792 if ((Name == "outb" || Name == "outw" || Name == "outl" || Name == "out") &&
793 Operands.size() == 3) {
794 X86Operand &Op = *(X86Operand*)Operands.back();
795 if (Op.isMem() && Op.Mem.SegReg == 0 &&
796 isa<MCConstantExpr>(Op.Mem.Disp) &&
797 cast<MCConstantExpr>(Op.Mem.Disp)->getValue() == 0 &&
798 Op.Mem.BaseReg == MatchRegisterName("dx") && Op.Mem.IndexReg == 0) {
799 SMLoc Loc = Op.getEndLoc();
800 Operands.back() = X86Operand::CreateReg(Op.Mem.BaseReg, Loc, Loc);
801 delete &Op;
804 // Same hack for "in[bwl]? (%dx), %al" -> "inb %dx, %al".
805 if ((Name == "inb" || Name == "inw" || Name == "inl" || Name == "in") &&
806 Operands.size() == 3) {
807 X86Operand &Op = *(X86Operand*)Operands.begin()[1];
808 if (Op.isMem() && Op.Mem.SegReg == 0 &&
809 isa<MCConstantExpr>(Op.Mem.Disp) &&
810 cast<MCConstantExpr>(Op.Mem.Disp)->getValue() == 0 &&
811 Op.Mem.BaseReg == MatchRegisterName("dx") && Op.Mem.IndexReg == 0) {
812 SMLoc Loc = Op.getEndLoc();
813 Operands.begin()[1] = X86Operand::CreateReg(Op.Mem.BaseReg, Loc, Loc);
814 delete &Op;
817 // Transform "ins[bwl] %dx, %es:(%edi)" into "ins[bwl]"
818 if (Name.startswith("ins") && Operands.size() == 3 &&
819 (Name == "insb" || Name == "insw" || Name == "insl")) {
820 X86Operand &Op = *(X86Operand*)Operands.begin()[1];
821 X86Operand &Op2 = *(X86Operand*)Operands.begin()[2];
822 if (Op.isReg() && Op.getReg() == X86::DX && isDstOp(Op2)) {
823 Operands.pop_back();
824 Operands.pop_back();
825 delete &Op;
826 delete &Op2;
830 // Transform "outs[bwl] %ds:(%esi), %dx" into "out[bwl]"
831 if (Name.startswith("outs") && Operands.size() == 3 &&
832 (Name == "outsb" || Name == "outsw" || Name == "outsl")) {
833 X86Operand &Op = *(X86Operand*)Operands.begin()[1];
834 X86Operand &Op2 = *(X86Operand*)Operands.begin()[2];
835 if (isSrcOp(Op) && Op2.isReg() && Op2.getReg() == X86::DX) {
836 Operands.pop_back();
837 Operands.pop_back();
838 delete &Op;
839 delete &Op2;
843 // Transform "movs[bwl] %ds:(%esi), %es:(%edi)" into "movs[bwl]"
844 if (Name.startswith("movs") && Operands.size() == 3 &&
845 (Name == "movsb" || Name == "movsw" || Name == "movsl" ||
846 (Is64Bit && Name == "movsq"))) {
847 X86Operand &Op = *(X86Operand*)Operands.begin()[1];
848 X86Operand &Op2 = *(X86Operand*)Operands.begin()[2];
849 if (isSrcOp(Op) && isDstOp(Op2)) {
850 Operands.pop_back();
851 Operands.pop_back();
852 delete &Op;
853 delete &Op2;
856 // Transform "lods[bwl] %ds:(%esi),{%al,%ax,%eax,%rax}" into "lods[bwl]"
857 if (Name.startswith("lods") && Operands.size() == 3 &&
858 (Name == "lods" || Name == "lodsb" || Name == "lodsw" ||
859 Name == "lodsl" || (Is64Bit && Name == "lodsq"))) {
860 X86Operand *Op1 = static_cast<X86Operand*>(Operands[1]);
861 X86Operand *Op2 = static_cast<X86Operand*>(Operands[2]);
862 if (isSrcOp(*Op1) && Op2->isReg()) {
863 const char *ins;
864 unsigned reg = Op2->getReg();
865 bool isLods = Name == "lods";
866 if (reg == X86::AL && (isLods || Name == "lodsb"))
867 ins = "lodsb";
868 else if (reg == X86::AX && (isLods || Name == "lodsw"))
869 ins = "lodsw";
870 else if (reg == X86::EAX && (isLods || Name == "lodsl"))
871 ins = "lodsl";
872 else if (reg == X86::RAX && (isLods || Name == "lodsq"))
873 ins = "lodsq";
874 else
875 ins = NULL;
876 if (ins != NULL) {
877 Operands.pop_back();
878 Operands.pop_back();
879 delete Op1;
880 delete Op2;
881 if (Name != ins)
882 static_cast<X86Operand*>(Operands[0])->setTokenValue(ins);
886 // Transform "stos[bwl] {%al,%ax,%eax,%rax},%es:(%edi)" into "stos[bwl]"
887 if (Name.startswith("stos") && Operands.size() == 3 &&
888 (Name == "stos" || Name == "stosb" || Name == "stosw" ||
889 Name == "stosl" || (Is64Bit && Name == "stosq"))) {
890 X86Operand *Op1 = static_cast<X86Operand*>(Operands[1]);
891 X86Operand *Op2 = static_cast<X86Operand*>(Operands[2]);
892 if (isDstOp(*Op2) && Op1->isReg()) {
893 const char *ins;
894 unsigned reg = Op1->getReg();
895 bool isStos = Name == "stos";
896 if (reg == X86::AL && (isStos || Name == "stosb"))
897 ins = "stosb";
898 else if (reg == X86::AX && (isStos || Name == "stosw"))
899 ins = "stosw";
900 else if (reg == X86::EAX && (isStos || Name == "stosl"))
901 ins = "stosl";
902 else if (reg == X86::RAX && (isStos || Name == "stosq"))
903 ins = "stosq";
904 else
905 ins = NULL;
906 if (ins != NULL) {
907 Operands.pop_back();
908 Operands.pop_back();
909 delete Op1;
910 delete Op2;
911 if (Name != ins)
912 static_cast<X86Operand*>(Operands[0])->setTokenValue(ins);
917 // FIXME: Hack to handle recognize s{hr,ar,hl} $1, <op>. Canonicalize to
918 // "shift <op>".
919 if ((Name.startswith("shr") || Name.startswith("sar") ||
920 Name.startswith("shl") || Name.startswith("sal") ||
921 Name.startswith("rcl") || Name.startswith("rcr") ||
922 Name.startswith("rol") || Name.startswith("ror")) &&
923 Operands.size() == 3) {
924 X86Operand *Op1 = static_cast<X86Operand*>(Operands[1]);
925 if (Op1->isImm() && isa<MCConstantExpr>(Op1->getImm()) &&
926 cast<MCConstantExpr>(Op1->getImm())->getValue() == 1) {
927 delete Operands[1];
928 Operands.erase(Operands.begin() + 1);
932 // Transforms "int $3" into "int3" as a size optimization. We can't write an
933 // instalias with an immediate operand yet.
934 if (Name == "int" && Operands.size() == 2) {
935 X86Operand *Op1 = static_cast<X86Operand*>(Operands[1]);
936 if (Op1->isImm() && isa<MCConstantExpr>(Op1->getImm()) &&
937 cast<MCConstantExpr>(Op1->getImm())->getValue() == 3) {
938 delete Operands[1];
939 Operands.erase(Operands.begin() + 1);
940 static_cast<X86Operand*>(Operands[0])->setTokenValue("int3");
944 return false;
947 bool X86ATTAsmParser::
948 MatchAndEmitInstruction(SMLoc IDLoc,
949 SmallVectorImpl<MCParsedAsmOperand*> &Operands,
950 MCStreamer &Out) {
951 assert(!Operands.empty() && "Unexpect empty operand list!");
952 X86Operand *Op = static_cast<X86Operand*>(Operands[0]);
953 assert(Op->isToken() && "Leading operand should always be a mnemonic!");
955 // First, handle aliases that expand to multiple instructions.
956 // FIXME: This should be replaced with a real .td file alias mechanism.
957 // Also, MatchInstructionImpl should do actually *do* the EmitInstruction
958 // call.
959 if (Op->getToken() == "fstsw" || Op->getToken() == "fstcw" ||
960 Op->getToken() == "fstsww" || Op->getToken() == "fstcww" ||
961 Op->getToken() == "finit" || Op->getToken() == "fsave" ||
962 Op->getToken() == "fstenv" || Op->getToken() == "fclex") {
963 MCInst Inst;
964 Inst.setOpcode(X86::WAIT);
965 Out.EmitInstruction(Inst);
967 const char *Repl =
968 StringSwitch<const char*>(Op->getToken())
969 .Case("finit", "fninit")
970 .Case("fsave", "fnsave")
971 .Case("fstcw", "fnstcw")
972 .Case("fstcww", "fnstcw")
973 .Case("fstenv", "fnstenv")
974 .Case("fstsw", "fnstsw")
975 .Case("fstsww", "fnstsw")
976 .Case("fclex", "fnclex")
977 .Default(0);
978 assert(Repl && "Unknown wait-prefixed instruction");
979 delete Operands[0];
980 Operands[0] = X86Operand::CreateToken(Repl, IDLoc);
983 bool WasOriginallyInvalidOperand = false;
984 unsigned OrigErrorInfo;
985 MCInst Inst;
987 // First, try a direct match.
988 switch (MatchInstructionImpl(Operands, Inst, OrigErrorInfo)) {
989 case Match_Success:
990 Out.EmitInstruction(Inst);
991 return false;
992 case Match_MissingFeature:
993 Error(IDLoc, "instruction requires a CPU feature not currently enabled");
994 return true;
995 case Match_ConversionFail:
996 return Error(IDLoc, "unable to convert operands to instruction");
997 case Match_InvalidOperand:
998 WasOriginallyInvalidOperand = true;
999 break;
1000 case Match_MnemonicFail:
1001 break;
1004 // FIXME: Ideally, we would only attempt suffix matches for things which are
1005 // valid prefixes, and we could just infer the right unambiguous
1006 // type. However, that requires substantially more matcher support than the
1007 // following hack.
1009 // Change the operand to point to a temporary token.
1010 StringRef Base = Op->getToken();
1011 SmallString<16> Tmp;
1012 Tmp += Base;
1013 Tmp += ' ';
1014 Op->setTokenValue(Tmp.str());
1016 // If this instruction starts with an 'f', then it is a floating point stack
1017 // instruction. These come in up to three forms for 32-bit, 64-bit, and
1018 // 80-bit floating point, which use the suffixes s,l,t respectively.
1020 // Otherwise, we assume that this may be an integer instruction, which comes
1021 // in 8/16/32/64-bit forms using the b,w,l,q suffixes respectively.
1022 const char *Suffixes = Base[0] != 'f' ? "bwlq" : "slt\0";
1024 // Check for the various suffix matches.
1025 Tmp[Base.size()] = Suffixes[0];
1026 unsigned ErrorInfoIgnore;
1027 MatchResultTy Match1, Match2, Match3, Match4;
1029 Match1 = MatchInstructionImpl(Operands, Inst, ErrorInfoIgnore);
1030 Tmp[Base.size()] = Suffixes[1];
1031 Match2 = MatchInstructionImpl(Operands, Inst, ErrorInfoIgnore);
1032 Tmp[Base.size()] = Suffixes[2];
1033 Match3 = MatchInstructionImpl(Operands, Inst, ErrorInfoIgnore);
1034 Tmp[Base.size()] = Suffixes[3];
1035 Match4 = MatchInstructionImpl(Operands, Inst, ErrorInfoIgnore);
1037 // Restore the old token.
1038 Op->setTokenValue(Base);
1040 // If exactly one matched, then we treat that as a successful match (and the
1041 // instruction will already have been filled in correctly, since the failing
1042 // matches won't have modified it).
1043 unsigned NumSuccessfulMatches =
1044 (Match1 == Match_Success) + (Match2 == Match_Success) +
1045 (Match3 == Match_Success) + (Match4 == Match_Success);
1046 if (NumSuccessfulMatches == 1) {
1047 Out.EmitInstruction(Inst);
1048 return false;
1051 // Otherwise, the match failed, try to produce a decent error message.
1053 // If we had multiple suffix matches, then identify this as an ambiguous
1054 // match.
1055 if (NumSuccessfulMatches > 1) {
1056 char MatchChars[4];
1057 unsigned NumMatches = 0;
1058 if (Match1 == Match_Success) MatchChars[NumMatches++] = Suffixes[0];
1059 if (Match2 == Match_Success) MatchChars[NumMatches++] = Suffixes[1];
1060 if (Match3 == Match_Success) MatchChars[NumMatches++] = Suffixes[2];
1061 if (Match4 == Match_Success) MatchChars[NumMatches++] = Suffixes[3];
1063 SmallString<126> Msg;
1064 raw_svector_ostream OS(Msg);
1065 OS << "ambiguous instructions require an explicit suffix (could be ";
1066 for (unsigned i = 0; i != NumMatches; ++i) {
1067 if (i != 0)
1068 OS << ", ";
1069 if (i + 1 == NumMatches)
1070 OS << "or ";
1071 OS << "'" << Base << MatchChars[i] << "'";
1073 OS << ")";
1074 Error(IDLoc, OS.str());
1075 return true;
1078 // Okay, we know that none of the variants matched successfully.
1080 // If all of the instructions reported an invalid mnemonic, then the original
1081 // mnemonic was invalid.
1082 if ((Match1 == Match_MnemonicFail) && (Match2 == Match_MnemonicFail) &&
1083 (Match3 == Match_MnemonicFail) && (Match4 == Match_MnemonicFail)) {
1084 if (!WasOriginallyInvalidOperand) {
1085 Error(IDLoc, "invalid instruction mnemonic '" + Base + "'");
1086 return true;
1089 // Recover location info for the operand if we know which was the problem.
1090 SMLoc ErrorLoc = IDLoc;
1091 if (OrigErrorInfo != ~0U) {
1092 if (OrigErrorInfo >= Operands.size())
1093 return Error(IDLoc, "too few operands for instruction");
1095 ErrorLoc = ((X86Operand*)Operands[OrigErrorInfo])->getStartLoc();
1096 if (ErrorLoc == SMLoc()) ErrorLoc = IDLoc;
1099 return Error(ErrorLoc, "invalid operand for instruction");
1102 // If one instruction matched with a missing feature, report this as a
1103 // missing feature.
1104 if ((Match1 == Match_MissingFeature) + (Match2 == Match_MissingFeature) +
1105 (Match3 == Match_MissingFeature) + (Match4 == Match_MissingFeature) == 1){
1106 Error(IDLoc, "instruction requires a CPU feature not currently enabled");
1107 return true;
1110 // If one instruction matched with an invalid operand, report this as an
1111 // operand failure.
1112 if ((Match1 == Match_InvalidOperand) + (Match2 == Match_InvalidOperand) +
1113 (Match3 == Match_InvalidOperand) + (Match4 == Match_InvalidOperand) == 1){
1114 Error(IDLoc, "invalid operand for instruction");
1115 return true;
1118 // If all of these were an outright failure, report it in a useless way.
1119 // FIXME: We should give nicer diagnostics about the exact failure.
1120 Error(IDLoc, "unknown use of instruction mnemonic without a size suffix");
1121 return true;
1125 bool X86ATTAsmParser::ParseDirective(AsmToken DirectiveID) {
1126 StringRef IDVal = DirectiveID.getIdentifier();
1127 if (IDVal == ".word")
1128 return ParseDirectiveWord(2, DirectiveID.getLoc());
1129 return true;
1132 /// ParseDirectiveWord
1133 /// ::= .word [ expression (, expression)* ]
1134 bool X86ATTAsmParser::ParseDirectiveWord(unsigned Size, SMLoc L) {
1135 if (getLexer().isNot(AsmToken::EndOfStatement)) {
1136 for (;;) {
1137 const MCExpr *Value;
1138 if (getParser().ParseExpression(Value))
1139 return true;
1141 getParser().getStreamer().EmitValue(Value, Size, 0 /*addrspace*/);
1143 if (getLexer().is(AsmToken::EndOfStatement))
1144 break;
1146 // FIXME: Improve diagnostic.
1147 if (getLexer().isNot(AsmToken::Comma))
1148 return Error(L, "unexpected token in directive");
1149 Parser.Lex();
1153 Parser.Lex();
1154 return false;
1160 extern "C" void LLVMInitializeX86AsmLexer();
1162 // Force static initialization.
1163 extern "C" void LLVMInitializeX86AsmParser() {
1164 RegisterAsmParser<X86_32ATTAsmParser> X(TheX86_32Target);
1165 RegisterAsmParser<X86_64ATTAsmParser> Y(TheX86_64Target);
1166 LLVMInitializeX86AsmLexer();
1169 #define GET_REGISTER_MATCHER
1170 #define GET_MATCHER_IMPLEMENTATION
1171 #include "X86GenAsmMatcher.inc"