1 //===- MCExpr.cpp - Assembly Level Expression Implementation --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #define DEBUG_TYPE "mcexpr"
11 #include "llvm/MC/MCExpr.h"
12 #include "llvm/ADT/Statistic.h"
13 #include "llvm/ADT/StringSwitch.h"
14 #include "llvm/MC/MCAsmLayout.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCContext.h"
17 #include "llvm/MC/MCObjectWriter.h"
18 #include "llvm/MC/MCSymbol.h"
19 #include "llvm/MC/MCValue.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Support/ErrorHandling.h"
22 #include "llvm/Support/raw_ostream.h"
27 STATISTIC(MCExprEvaluate, "Number of MCExpr evaluations");
31 void MCExpr::print(raw_ostream &OS) const {
34 return cast<MCTargetExpr>(this)->PrintImpl(OS);
35 case MCExpr::Constant:
36 OS << cast<MCConstantExpr>(*this).getValue();
39 case MCExpr::SymbolRef: {
40 const MCSymbolRefExpr &SRE = cast<MCSymbolRefExpr>(*this);
41 const MCSymbol &Sym = SRE.getSymbol();
42 // Parenthesize names that start with $ so that they don't look like
44 bool UseParens = Sym.getName()[0] == '$';
46 if (SRE.getKind() == MCSymbolRefExpr::VK_PPC_DARWIN_HA16 ||
47 SRE.getKind() == MCSymbolRefExpr::VK_PPC_DARWIN_LO16) {
48 OS << MCSymbolRefExpr::getVariantKindName(SRE.getKind());
53 OS << '(' << Sym << ')';
57 if (SRE.getKind() == MCSymbolRefExpr::VK_ARM_PLT ||
58 SRE.getKind() == MCSymbolRefExpr::VK_ARM_TLSGD ||
59 SRE.getKind() == MCSymbolRefExpr::VK_ARM_GOT ||
60 SRE.getKind() == MCSymbolRefExpr::VK_ARM_GOTOFF ||
61 SRE.getKind() == MCSymbolRefExpr::VK_ARM_TPOFF ||
62 SRE.getKind() == MCSymbolRefExpr::VK_ARM_GOTTPOFF ||
63 SRE.getKind() == MCSymbolRefExpr::VK_ARM_TARGET1 ||
64 SRE.getKind() == MCSymbolRefExpr::VK_ARM_TARGET2)
65 OS << MCSymbolRefExpr::getVariantKindName(SRE.getKind());
66 else if (SRE.getKind() != MCSymbolRefExpr::VK_None &&
67 SRE.getKind() != MCSymbolRefExpr::VK_PPC_DARWIN_HA16 &&
68 SRE.getKind() != MCSymbolRefExpr::VK_PPC_DARWIN_LO16)
69 OS << '@' << MCSymbolRefExpr::getVariantKindName(SRE.getKind());
75 const MCUnaryExpr &UE = cast<MCUnaryExpr>(*this);
76 switch (UE.getOpcode()) {
77 case MCUnaryExpr::LNot: OS << '!'; break;
78 case MCUnaryExpr::Minus: OS << '-'; break;
79 case MCUnaryExpr::Not: OS << '~'; break;
80 case MCUnaryExpr::Plus: OS << '+'; break;
82 OS << *UE.getSubExpr();
86 case MCExpr::Binary: {
87 const MCBinaryExpr &BE = cast<MCBinaryExpr>(*this);
89 // Only print parens around the LHS if it is non-trivial.
90 if (isa<MCConstantExpr>(BE.getLHS()) || isa<MCSymbolRefExpr>(BE.getLHS())) {
93 OS << '(' << *BE.getLHS() << ')';
96 switch (BE.getOpcode()) {
97 case MCBinaryExpr::Add:
98 // Print "X-42" instead of "X+-42".
99 if (const MCConstantExpr *RHSC = dyn_cast<MCConstantExpr>(BE.getRHS())) {
100 if (RHSC->getValue() < 0) {
101 OS << RHSC->getValue();
108 case MCBinaryExpr::And: OS << '&'; break;
109 case MCBinaryExpr::Div: OS << '/'; break;
110 case MCBinaryExpr::EQ: OS << "=="; break;
111 case MCBinaryExpr::GT: OS << '>'; break;
112 case MCBinaryExpr::GTE: OS << ">="; break;
113 case MCBinaryExpr::LAnd: OS << "&&"; break;
114 case MCBinaryExpr::LOr: OS << "||"; break;
115 case MCBinaryExpr::LT: OS << '<'; break;
116 case MCBinaryExpr::LTE: OS << "<="; break;
117 case MCBinaryExpr::Mod: OS << '%'; break;
118 case MCBinaryExpr::Mul: OS << '*'; break;
119 case MCBinaryExpr::NE: OS << "!="; break;
120 case MCBinaryExpr::Or: OS << '|'; break;
121 case MCBinaryExpr::Shl: OS << "<<"; break;
122 case MCBinaryExpr::Shr: OS << ">>"; break;
123 case MCBinaryExpr::Sub: OS << '-'; break;
124 case MCBinaryExpr::Xor: OS << '^'; break;
127 // Only print parens around the LHS if it is non-trivial.
128 if (isa<MCConstantExpr>(BE.getRHS()) || isa<MCSymbolRefExpr>(BE.getRHS())) {
131 OS << '(' << *BE.getRHS() << ')';
137 llvm_unreachable("Invalid expression kind!");
140 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
141 void MCExpr::dump() const {
149 const MCBinaryExpr *MCBinaryExpr::Create(Opcode Opc, const MCExpr *LHS,
150 const MCExpr *RHS, MCContext &Ctx) {
151 return new (Ctx) MCBinaryExpr(Opc, LHS, RHS);
154 const MCUnaryExpr *MCUnaryExpr::Create(Opcode Opc, const MCExpr *Expr,
156 return new (Ctx) MCUnaryExpr(Opc, Expr);
159 const MCConstantExpr *MCConstantExpr::Create(int64_t Value, MCContext &Ctx) {
160 return new (Ctx) MCConstantExpr(Value);
165 const MCSymbolRefExpr *MCSymbolRefExpr::Create(const MCSymbol *Sym,
168 return new (Ctx) MCSymbolRefExpr(Sym, Kind);
171 const MCSymbolRefExpr *MCSymbolRefExpr::Create(StringRef Name, VariantKind Kind,
173 return Create(Ctx.GetOrCreateSymbol(Name), Kind, Ctx);
176 StringRef MCSymbolRefExpr::getVariantKindName(VariantKind Kind) {
178 case VK_Invalid: return "<<invalid>>";
179 case VK_None: return "<<none>>";
181 case VK_GOT: return "GOT";
182 case VK_GOTOFF: return "GOTOFF";
183 case VK_GOTPCREL: return "GOTPCREL";
184 case VK_GOTTPOFF: return "GOTTPOFF";
185 case VK_INDNTPOFF: return "INDNTPOFF";
186 case VK_NTPOFF: return "NTPOFF";
187 case VK_GOTNTPOFF: return "GOTNTPOFF";
188 case VK_PLT: return "PLT";
189 case VK_TLSGD: return "TLSGD";
190 case VK_TLSLD: return "TLSLD";
191 case VK_TLSLDM: return "TLSLDM";
192 case VK_TPOFF: return "TPOFF";
193 case VK_DTPOFF: return "DTPOFF";
194 case VK_TLVP: return "TLVP";
195 case VK_SECREL: return "SECREL";
196 case VK_ARM_PLT: return "(PLT)";
197 case VK_ARM_GOT: return "(GOT)";
198 case VK_ARM_GOTOFF: return "(GOTOFF)";
199 case VK_ARM_TPOFF: return "(tpoff)";
200 case VK_ARM_GOTTPOFF: return "(gottpoff)";
201 case VK_ARM_TLSGD: return "(tlsgd)";
202 case VK_ARM_TARGET1: return "(target1)";
203 case VK_ARM_TARGET2: return "(target2)";
204 case VK_PPC_TOC: return "tocbase";
205 case VK_PPC_TOC_ENTRY: return "toc";
206 case VK_PPC_DARWIN_HA16: return "ha16";
207 case VK_PPC_DARWIN_LO16: return "lo16";
208 case VK_PPC_GAS_HA16: return "ha";
209 case VK_PPC_GAS_LO16: return "l";
210 case VK_PPC_TPREL16_HA: return "tprel@ha";
211 case VK_PPC_TPREL16_LO: return "tprel@l";
212 case VK_Mips_GPREL: return "GPREL";
213 case VK_Mips_GOT_CALL: return "GOT_CALL";
214 case VK_Mips_GOT16: return "GOT16";
215 case VK_Mips_GOT: return "GOT";
216 case VK_Mips_ABS_HI: return "ABS_HI";
217 case VK_Mips_ABS_LO: return "ABS_LO";
218 case VK_Mips_TLSGD: return "TLSGD";
219 case VK_Mips_TLSLDM: return "TLSLDM";
220 case VK_Mips_DTPREL_HI: return "DTPREL_HI";
221 case VK_Mips_DTPREL_LO: return "DTPREL_LO";
222 case VK_Mips_GOTTPREL: return "GOTTPREL";
223 case VK_Mips_TPREL_HI: return "TPREL_HI";
224 case VK_Mips_TPREL_LO: return "TPREL_LO";
225 case VK_Mips_GPOFF_HI: return "GPOFF_HI";
226 case VK_Mips_GPOFF_LO: return "GPOFF_LO";
227 case VK_Mips_GOT_DISP: return "GOT_DISP";
228 case VK_Mips_GOT_PAGE: return "GOT_PAGE";
229 case VK_Mips_GOT_OFST: return "GOT_OFST";
230 case VK_Mips_HIGHER: return "HIGHER";
231 case VK_Mips_HIGHEST: return "HIGHEST";
233 llvm_unreachable("Invalid variant kind");
236 MCSymbolRefExpr::VariantKind
237 MCSymbolRefExpr::getVariantKindForName(StringRef Name) {
238 return StringSwitch<VariantKind>(Name)
241 .Case("GOTOFF", VK_GOTOFF)
242 .Case("gotoff", VK_GOTOFF)
243 .Case("GOTPCREL", VK_GOTPCREL)
244 .Case("gotpcrel", VK_GOTPCREL)
245 .Case("GOTTPOFF", VK_GOTTPOFF)
246 .Case("gottpoff", VK_GOTTPOFF)
247 .Case("INDNTPOFF", VK_INDNTPOFF)
248 .Case("indntpoff", VK_INDNTPOFF)
249 .Case("NTPOFF", VK_NTPOFF)
250 .Case("ntpoff", VK_NTPOFF)
251 .Case("GOTNTPOFF", VK_GOTNTPOFF)
252 .Case("gotntpoff", VK_GOTNTPOFF)
255 .Case("TLSGD", VK_TLSGD)
256 .Case("tlsgd", VK_TLSGD)
257 .Case("TLSLD", VK_TLSLD)
258 .Case("tlsld", VK_TLSLD)
259 .Case("TLSLDM", VK_TLSLDM)
260 .Case("tlsldm", VK_TLSLDM)
261 .Case("TPOFF", VK_TPOFF)
262 .Case("tpoff", VK_TPOFF)
263 .Case("DTPOFF", VK_DTPOFF)
264 .Case("dtpoff", VK_DTPOFF)
265 .Case("TLVP", VK_TLVP)
266 .Case("tlvp", VK_TLVP)
267 .Default(VK_Invalid);
272 void MCTargetExpr::anchor() {}
276 bool MCExpr::EvaluateAsAbsolute(int64_t &Res) const {
277 return EvaluateAsAbsolute(Res, 0, 0, 0);
280 bool MCExpr::EvaluateAsAbsolute(int64_t &Res,
281 const MCAsmLayout &Layout) const {
282 return EvaluateAsAbsolute(Res, &Layout.getAssembler(), &Layout, 0);
285 bool MCExpr::EvaluateAsAbsolute(int64_t &Res,
286 const MCAsmLayout &Layout,
287 const SectionAddrMap &Addrs) const {
288 return EvaluateAsAbsolute(Res, &Layout.getAssembler(), &Layout, &Addrs);
291 bool MCExpr::EvaluateAsAbsolute(int64_t &Res, const MCAssembler &Asm) const {
292 return EvaluateAsAbsolute(Res, &Asm, 0, 0);
295 bool MCExpr::EvaluateAsAbsolute(int64_t &Res, const MCAssembler *Asm,
296 const MCAsmLayout *Layout,
297 const SectionAddrMap *Addrs) const {
300 // Fast path constants.
301 if (const MCConstantExpr *CE = dyn_cast<MCConstantExpr>(this)) {
302 Res = CE->getValue();
306 // FIXME: The use if InSet = Addrs is a hack. Setting InSet causes us
307 // absolutize differences across sections and that is what the MachO writer
310 EvaluateAsRelocatableImpl(Value, Asm, Layout, Addrs, /*InSet*/ Addrs);
312 // Record the current value.
313 Res = Value.getConstant();
315 return IsRelocatable && Value.isAbsolute();
318 /// \brief Helper method for \see EvaluateSymbolAdd().
319 static void AttemptToFoldSymbolOffsetDifference(const MCAssembler *Asm,
320 const MCAsmLayout *Layout,
321 const SectionAddrMap *Addrs,
323 const MCSymbolRefExpr *&A,
324 const MCSymbolRefExpr *&B,
329 const MCSymbol &SA = A->getSymbol();
330 const MCSymbol &SB = B->getSymbol();
332 if (SA.isUndefined() || SB.isUndefined())
335 if (!Asm->getWriter().IsSymbolRefDifferenceFullyResolved(*Asm, A, B, InSet))
338 MCSymbolData &AD = Asm->getSymbolData(SA);
339 MCSymbolData &BD = Asm->getSymbolData(SB);
341 if (AD.getFragment() == BD.getFragment()) {
342 Addend += (AD.getOffset() - BD.getOffset());
344 // Pointers to Thumb symbols need to have their low-bit set to allow
346 if (Asm->isThumbFunc(&SA))
349 // Clear the symbol expr pointers to indicate we have folded these
358 const MCSectionData &SecA = *AD.getFragment()->getParent();
359 const MCSectionData &SecB = *BD.getFragment()->getParent();
361 if ((&SecA != &SecB) && !Addrs)
365 Addend += (Layout->getSymbolOffset(&Asm->getSymbolData(A->getSymbol())) -
366 Layout->getSymbolOffset(&Asm->getSymbolData(B->getSymbol())));
367 if (Addrs && (&SecA != &SecB))
368 Addend += (Addrs->lookup(&SecA) - Addrs->lookup(&SecB));
370 // Pointers to Thumb symbols need to have their low-bit set to allow
372 if (Asm->isThumbFunc(&SA))
375 // Clear the symbol expr pointers to indicate we have folded these
380 /// \brief Evaluate the result of an add between (conceptually) two MCValues.
382 /// This routine conceptually attempts to construct an MCValue:
383 /// Result = (Result_A - Result_B + Result_Cst)
384 /// from two MCValue's LHS and RHS where
385 /// Result = LHS + RHS
387 /// Result = (LHS_A - LHS_B + LHS_Cst) + (RHS_A - RHS_B + RHS_Cst).
389 /// This routine attempts to aggresively fold the operands such that the result
390 /// is representable in an MCValue, but may not always succeed.
392 /// \returns True on success, false if the result is not representable in an
395 /// NOTE: It is really important to have both the Asm and Layout arguments.
396 /// They might look redundant, but this function can be used before layout
397 /// is done (see the object streamer for example) and having the Asm argument
398 /// lets us avoid relaxations early.
399 static bool EvaluateSymbolicAdd(const MCAssembler *Asm,
400 const MCAsmLayout *Layout,
401 const SectionAddrMap *Addrs,
403 const MCValue &LHS,const MCSymbolRefExpr *RHS_A,
404 const MCSymbolRefExpr *RHS_B, int64_t RHS_Cst,
406 // FIXME: This routine (and other evaluation parts) are *incredibly* sloppy
407 // about dealing with modifiers. This will ultimately bite us, one day.
408 const MCSymbolRefExpr *LHS_A = LHS.getSymA();
409 const MCSymbolRefExpr *LHS_B = LHS.getSymB();
410 int64_t LHS_Cst = LHS.getConstant();
412 // Fold the result constant immediately.
413 int64_t Result_Cst = LHS_Cst + RHS_Cst;
415 assert((!Layout || Asm) &&
416 "Must have an assembler object if layout is given!");
418 // If we have a layout, we can fold resolved differences.
420 // First, fold out any differences which are fully resolved. By
421 // reassociating terms in
422 // Result = (LHS_A - LHS_B + LHS_Cst) + (RHS_A - RHS_B + RHS_Cst).
423 // we have the four possible differences:
428 // Since we are attempting to be as aggressive as possible about folding, we
429 // attempt to evaluate each possible alternative.
430 AttemptToFoldSymbolOffsetDifference(Asm, Layout, Addrs, InSet, LHS_A, LHS_B,
432 AttemptToFoldSymbolOffsetDifference(Asm, Layout, Addrs, InSet, LHS_A, RHS_B,
434 AttemptToFoldSymbolOffsetDifference(Asm, Layout, Addrs, InSet, RHS_A, LHS_B,
436 AttemptToFoldSymbolOffsetDifference(Asm, Layout, Addrs, InSet, RHS_A, RHS_B,
440 // We can't represent the addition or subtraction of two symbols.
441 if ((LHS_A && RHS_A) || (LHS_B && RHS_B))
444 // At this point, we have at most one additive symbol and one subtractive
445 // symbol -- find them.
446 const MCSymbolRefExpr *A = LHS_A ? LHS_A : RHS_A;
447 const MCSymbolRefExpr *B = LHS_B ? LHS_B : RHS_B;
449 // If we have a negated symbol, then we must have also have a non-negated
450 // symbol in order to encode the expression.
454 Res = MCValue::get(A, B, Result_Cst);
458 bool MCExpr::EvaluateAsRelocatable(MCValue &Res,
459 const MCAsmLayout &Layout) const {
460 return EvaluateAsRelocatableImpl(Res, &Layout.getAssembler(), &Layout,
464 bool MCExpr::EvaluateAsRelocatableImpl(MCValue &Res,
465 const MCAssembler *Asm,
466 const MCAsmLayout *Layout,
467 const SectionAddrMap *Addrs,
469 ++stats::MCExprEvaluate;
473 return cast<MCTargetExpr>(this)->EvaluateAsRelocatableImpl(Res, Layout);
476 Res = MCValue::get(cast<MCConstantExpr>(this)->getValue());
480 const MCSymbolRefExpr *SRE = cast<MCSymbolRefExpr>(this);
481 const MCSymbol &Sym = SRE->getSymbol();
483 // Evaluate recursively if this is a variable.
484 if (Sym.isVariable() && SRE->getKind() == MCSymbolRefExpr::VK_None) {
485 bool Ret = Sym.getVariableValue()->EvaluateAsRelocatableImpl(Res, Asm,
489 // If we failed to simplify this to a constant, let the target
491 if (Ret && !Res.getSymA() && !Res.getSymB())
495 Res = MCValue::get(SRE, 0, 0);
500 const MCUnaryExpr *AUE = cast<MCUnaryExpr>(this);
503 if (!AUE->getSubExpr()->EvaluateAsRelocatableImpl(Value, Asm, Layout,
507 switch (AUE->getOpcode()) {
508 case MCUnaryExpr::LNot:
509 if (!Value.isAbsolute())
511 Res = MCValue::get(!Value.getConstant());
513 case MCUnaryExpr::Minus:
514 /// -(a - b + const) ==> (b - a - const)
515 if (Value.getSymA() && !Value.getSymB())
517 Res = MCValue::get(Value.getSymB(), Value.getSymA(),
518 -Value.getConstant());
520 case MCUnaryExpr::Not:
521 if (!Value.isAbsolute())
523 Res = MCValue::get(~Value.getConstant());
525 case MCUnaryExpr::Plus:
534 const MCBinaryExpr *ABE = cast<MCBinaryExpr>(this);
535 MCValue LHSValue, RHSValue;
537 if (!ABE->getLHS()->EvaluateAsRelocatableImpl(LHSValue, Asm, Layout,
539 !ABE->getRHS()->EvaluateAsRelocatableImpl(RHSValue, Asm, Layout,
543 // We only support a few operations on non-constant expressions, handle
545 if (!LHSValue.isAbsolute() || !RHSValue.isAbsolute()) {
546 switch (ABE->getOpcode()) {
549 case MCBinaryExpr::Sub:
550 // Negate RHS and add.
551 return EvaluateSymbolicAdd(Asm, Layout, Addrs, InSet, LHSValue,
552 RHSValue.getSymB(), RHSValue.getSymA(),
553 -RHSValue.getConstant(),
556 case MCBinaryExpr::Add:
557 return EvaluateSymbolicAdd(Asm, Layout, Addrs, InSet, LHSValue,
558 RHSValue.getSymA(), RHSValue.getSymB(),
559 RHSValue.getConstant(),
564 // FIXME: We need target hooks for the evaluation. It may be limited in
565 // width, and gas defines the result of comparisons and right shifts
566 // differently from Apple as.
567 int64_t LHS = LHSValue.getConstant(), RHS = RHSValue.getConstant();
569 switch (ABE->getOpcode()) {
570 case MCBinaryExpr::Add: Result = LHS + RHS; break;
571 case MCBinaryExpr::And: Result = LHS & RHS; break;
572 case MCBinaryExpr::Div: Result = LHS / RHS; break;
573 case MCBinaryExpr::EQ: Result = LHS == RHS; break;
574 case MCBinaryExpr::GT: Result = LHS > RHS; break;
575 case MCBinaryExpr::GTE: Result = LHS >= RHS; break;
576 case MCBinaryExpr::LAnd: Result = LHS && RHS; break;
577 case MCBinaryExpr::LOr: Result = LHS || RHS; break;
578 case MCBinaryExpr::LT: Result = LHS < RHS; break;
579 case MCBinaryExpr::LTE: Result = LHS <= RHS; break;
580 case MCBinaryExpr::Mod: Result = LHS % RHS; break;
581 case MCBinaryExpr::Mul: Result = LHS * RHS; break;
582 case MCBinaryExpr::NE: Result = LHS != RHS; break;
583 case MCBinaryExpr::Or: Result = LHS | RHS; break;
584 case MCBinaryExpr::Shl: Result = LHS << RHS; break;
585 case MCBinaryExpr::Shr: Result = LHS >> RHS; break;
586 case MCBinaryExpr::Sub: Result = LHS - RHS; break;
587 case MCBinaryExpr::Xor: Result = LHS ^ RHS; break;
590 Res = MCValue::get(Result);
595 llvm_unreachable("Invalid assembly expression kind!");
598 const MCSection *MCExpr::FindAssociatedSection() const {
601 // We never look through target specific expressions.
602 return cast<MCTargetExpr>(this)->FindAssociatedSection();
605 return MCSymbol::AbsolutePseudoSection;
608 const MCSymbolRefExpr *SRE = cast<MCSymbolRefExpr>(this);
609 const MCSymbol &Sym = SRE->getSymbol();
612 return &Sym.getSection();
618 return cast<MCUnaryExpr>(this)->getSubExpr()->FindAssociatedSection();
621 const MCBinaryExpr *BE = cast<MCBinaryExpr>(this);
622 const MCSection *LHS_S = BE->getLHS()->FindAssociatedSection();
623 const MCSection *RHS_S = BE->getRHS()->FindAssociatedSection();
625 // If either section is absolute, return the other.
626 if (LHS_S == MCSymbol::AbsolutePseudoSection)
628 if (RHS_S == MCSymbol::AbsolutePseudoSection)
631 // Otherwise, return the first non-null section.
632 return LHS_S ? LHS_S : RHS_S;
636 llvm_unreachable("Invalid assembly expression kind!");