1 //===-- X86AsmBackend.cpp - X86 Assembler Backend -------------------------===//
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 #include "llvm/Target/TargetAsmBackend.h"
12 #include "X86FixupKinds.h"
13 #include "llvm/ADT/Twine.h"
14 #include "llvm/MC/ELFObjectWriter.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCObjectFormat.h"
18 #include "llvm/MC/MCObjectWriter.h"
19 #include "llvm/MC/MCSectionCOFF.h"
20 #include "llvm/MC/MCSectionELF.h"
21 #include "llvm/MC/MCSectionMachO.h"
22 #include "llvm/MC/MachObjectWriter.h"
23 #include "llvm/Support/ELF.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Target/TargetRegistry.h"
27 #include "llvm/Target/TargetAsmBackend.h"
31 static unsigned getFixupKindLog2Size(unsigned Kind) {
33 default: assert(0 && "invalid fixup kind!");
34 case X86::reloc_pcrel_1byte:
35 case FK_Data_1: return 0;
36 case X86::reloc_pcrel_2byte:
37 case FK_Data_2: return 1;
38 case X86::reloc_pcrel_4byte:
39 case X86::reloc_riprel_4byte:
40 case X86::reloc_riprel_4byte_movq_load:
41 case X86::reloc_signed_4byte:
42 case X86::reloc_global_offset_table:
43 case FK_Data_4: return 2;
44 case FK_Data_8: return 3;
49 class X86AsmBackend : public TargetAsmBackend {
51 X86AsmBackend(const Target &T)
52 : TargetAsmBackend(T) {}
54 void ApplyFixup(const MCFixup &Fixup, MCDataFragment &DF,
55 uint64_t Value) const {
56 unsigned Size = 1 << getFixupKindLog2Size(Fixup.getKind());
58 assert(Fixup.getOffset() + Size <= DF.getContents().size() &&
59 "Invalid fixup offset!");
60 for (unsigned i = 0; i != Size; ++i)
61 DF.getContents()[Fixup.getOffset() + i] = uint8_t(Value >> (i * 8));
64 bool MayNeedRelaxation(const MCInst &Inst) const;
66 void RelaxInstruction(const MCInst &Inst, MCInst &Res) const;
68 bool WriteNopData(uint64_t Count, MCObjectWriter *OW) const;
70 } // end anonymous namespace
72 static unsigned getRelaxedOpcode(unsigned Op) {
77 // This is used on i386 with things like addl $foo, %ebx
78 // FIXME: Should the other *i8 instructions be here too? If not, it might
79 // be better to just select X86::ADD32ri instead of X86::ADD32ri8.
80 case X86::ADD32ri8: return X86::ADD32ri;
82 case X86::JAE_1: return X86::JAE_4;
83 case X86::JA_1: return X86::JA_4;
84 case X86::JBE_1: return X86::JBE_4;
85 case X86::JB_1: return X86::JB_4;
86 case X86::JE_1: return X86::JE_4;
87 case X86::JGE_1: return X86::JGE_4;
88 case X86::JG_1: return X86::JG_4;
89 case X86::JLE_1: return X86::JLE_4;
90 case X86::JL_1: return X86::JL_4;
91 case X86::JMP_1: return X86::JMP_4;
92 case X86::JNE_1: return X86::JNE_4;
93 case X86::JNO_1: return X86::JNO_4;
94 case X86::JNP_1: return X86::JNP_4;
95 case X86::JNS_1: return X86::JNS_4;
96 case X86::JO_1: return X86::JO_4;
97 case X86::JP_1: return X86::JP_4;
98 case X86::JS_1: return X86::JS_4;
102 bool X86AsmBackend::MayNeedRelaxation(const MCInst &Inst) const {
103 // Check if this instruction is ever relaxable.
104 if (getRelaxedOpcode(Inst.getOpcode()) == Inst.getOpcode())
107 // If so, just assume it can be relaxed. Once we support relaxing more complex
108 // instructions we should check that the instruction actually has symbolic
109 // operands before doing this, but we need to be careful about things like
114 // FIXME: Can tblgen help at all here to verify there aren't other instructions
116 void X86AsmBackend::RelaxInstruction(const MCInst &Inst, MCInst &Res) const {
117 // The only relaxations X86 does is from a 1byte pcrel to a 4byte pcrel.
118 unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode());
120 if (RelaxedOp == Inst.getOpcode()) {
121 SmallString<256> Tmp;
122 raw_svector_ostream OS(Tmp);
123 Inst.dump_pretty(OS);
125 report_fatal_error("unexpected instruction to relax: " + OS.str());
129 Res.setOpcode(RelaxedOp);
132 /// WriteNopData - Write optimal nops to the output file for the \arg Count
133 /// bytes. This returns the number of bytes written. It may return 0 if
134 /// the \arg Count is more than the maximum optimal nops.
136 /// FIXME this is X86 32-bit specific and should move to a better place.
137 bool X86AsmBackend::WriteNopData(uint64_t Count, MCObjectWriter *OW) const {
138 static const uint8_t Nops[16][16] = {
146 {0x0f, 0x1f, 0x40, 0x00},
147 // nopl 0(%[re]ax,%[re]ax,1)
148 {0x0f, 0x1f, 0x44, 0x00, 0x00},
149 // nopw 0(%[re]ax,%[re]ax,1)
150 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
152 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
153 // nopl 0L(%[re]ax,%[re]ax,1)
154 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
155 // nopw 0L(%[re]ax,%[re]ax,1)
156 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
157 // nopw %cs:0L(%[re]ax,%[re]ax,1)
158 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
159 // nopl 0(%[re]ax,%[re]ax,1)
160 // nopw 0(%[re]ax,%[re]ax,1)
161 {0x0f, 0x1f, 0x44, 0x00, 0x00,
162 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
163 // nopw 0(%[re]ax,%[re]ax,1)
164 // nopw 0(%[re]ax,%[re]ax,1)
165 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
166 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
167 // nopw 0(%[re]ax,%[re]ax,1)
168 // nopl 0L(%[re]ax) */
169 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
170 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
173 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
174 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
176 // nopl 0L(%[re]ax,%[re]ax,1)
177 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
178 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}
181 // Write an optimal sequence for the first 15 bytes.
182 uint64_t OptimalCount = (Count < 16) ? Count : 15;
183 for (uint64_t i = 0, e = OptimalCount; i != e; i++)
184 OW->Write8(Nops[OptimalCount - 1][i]);
186 // Finish with single byte nops.
187 for (uint64_t i = OptimalCount, e = Count; i != e; ++i)
196 class ELFX86AsmBackend : public X86AsmBackend {
197 MCELFObjectFormat Format;
200 Triple::OSType OSType;
201 ELFX86AsmBackend(const Target &T, Triple::OSType _OSType)
202 : X86AsmBackend(T), OSType(_OSType) {
203 HasScatteredSymbols = true;
204 HasReliableSymbolDifference = true;
207 virtual const MCObjectFormat &getObjectFormat() const {
211 virtual bool doesSectionRequireSymbols(const MCSection &Section) const {
212 const MCSectionELF &ES = static_cast<const MCSectionELF&>(Section);
213 return ES.getFlags() & MCSectionELF::SHF_MERGE;
216 bool isVirtualSection(const MCSection &Section) const {
217 const MCSectionELF &SE = static_cast<const MCSectionELF&>(Section);
218 return SE.getType() == MCSectionELF::SHT_NOBITS;
222 class ELFX86_32AsmBackend : public ELFX86AsmBackend {
224 ELFX86_32AsmBackend(const Target &T, Triple::OSType OSType)
225 : ELFX86AsmBackend(T, OSType) {}
227 unsigned getPointerSize() const {
231 MCObjectWriter *createObjectWriter(raw_ostream &OS) const {
232 return new ELFObjectWriter(OS, /*Is64Bit=*/false,
234 /*IsLittleEndian=*/true,
235 /*HasRelocationAddend=*/false);
239 class ELFX86_64AsmBackend : public ELFX86AsmBackend {
241 ELFX86_64AsmBackend(const Target &T, Triple::OSType OSType)
242 : ELFX86AsmBackend(T, OSType) {}
244 unsigned getPointerSize() const {
248 MCObjectWriter *createObjectWriter(raw_ostream &OS) const {
249 return new ELFObjectWriter(OS, /*Is64Bit=*/true,
250 OSType, ELF::EM_X86_64,
251 /*IsLittleEndian=*/true,
252 /*HasRelocationAddend=*/true);
256 class WindowsX86AsmBackend : public X86AsmBackend {
258 MCCOFFObjectFormat Format;
261 WindowsX86AsmBackend(const Target &T, bool is64Bit)
264 HasScatteredSymbols = true;
267 virtual const MCObjectFormat &getObjectFormat() const {
271 unsigned getPointerSize() const {
278 MCObjectWriter *createObjectWriter(raw_ostream &OS) const {
279 return createWinCOFFObjectWriter(OS, Is64Bit);
282 bool isVirtualSection(const MCSection &Section) const {
283 const MCSectionCOFF &SE = static_cast<const MCSectionCOFF&>(Section);
284 return SE.getCharacteristics() & COFF::IMAGE_SCN_CNT_UNINITIALIZED_DATA;
288 class DarwinX86AsmBackend : public X86AsmBackend {
289 MCMachOObjectFormat Format;
292 DarwinX86AsmBackend(const Target &T)
294 HasScatteredSymbols = true;
297 virtual const MCObjectFormat &getObjectFormat() const {
301 bool isVirtualSection(const MCSection &Section) const {
302 const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section);
303 return (SMO.getType() == MCSectionMachO::S_ZEROFILL ||
304 SMO.getType() == MCSectionMachO::S_GB_ZEROFILL ||
305 SMO.getType() == MCSectionMachO::S_THREAD_LOCAL_ZEROFILL);
309 class DarwinX86_32AsmBackend : public DarwinX86AsmBackend {
311 DarwinX86_32AsmBackend(const Target &T)
312 : DarwinX86AsmBackend(T) {}
314 unsigned getPointerSize() const {
318 MCObjectWriter *createObjectWriter(raw_ostream &OS) const {
319 return new MachObjectWriter(OS, /*Is64Bit=*/false);
323 class DarwinX86_64AsmBackend : public DarwinX86AsmBackend {
325 DarwinX86_64AsmBackend(const Target &T)
326 : DarwinX86AsmBackend(T) {
327 HasReliableSymbolDifference = true;
330 unsigned getPointerSize() const {
334 MCObjectWriter *createObjectWriter(raw_ostream &OS) const {
335 return new MachObjectWriter(OS, /*Is64Bit=*/true);
338 virtual bool doesSectionRequireSymbols(const MCSection &Section) const {
339 // Temporary labels in the string literals sections require symbols. The
340 // issue is that the x86_64 relocation format does not allow symbol +
341 // offset, and so the linker does not have enough information to resolve the
342 // access to the appropriate atom unless an external relocation is used. For
343 // non-cstring sections, we expect the compiler to use a non-temporary label
344 // for anything that could have an addend pointing outside the symbol.
346 // See <rdar://problem/4765733>.
347 const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section);
348 return SMO.getType() == MCSectionMachO::S_CSTRING_LITERALS;
351 virtual bool isSectionAtomizable(const MCSection &Section) const {
352 const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section);
353 // Fixed sized data sections are uniqued, they cannot be diced into atoms.
354 switch (SMO.getType()) {
358 case MCSectionMachO::S_4BYTE_LITERALS:
359 case MCSectionMachO::S_8BYTE_LITERALS:
360 case MCSectionMachO::S_16BYTE_LITERALS:
361 case MCSectionMachO::S_LITERAL_POINTERS:
362 case MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS:
363 case MCSectionMachO::S_LAZY_SYMBOL_POINTERS:
364 case MCSectionMachO::S_MOD_INIT_FUNC_POINTERS:
365 case MCSectionMachO::S_MOD_TERM_FUNC_POINTERS:
366 case MCSectionMachO::S_INTERPOSING:
372 } // end anonymous namespace
374 TargetAsmBackend *llvm::createX86_32AsmBackend(const Target &T,
375 const std::string &TT) {
376 switch (Triple(TT).getOS()) {
378 return new DarwinX86_32AsmBackend(T);
379 case Triple::MinGW32:
382 return new WindowsX86AsmBackend(T, false);
384 return new ELFX86_32AsmBackend(T, Triple(TT).getOS());
388 TargetAsmBackend *llvm::createX86_64AsmBackend(const Target &T,
389 const std::string &TT) {
390 switch (Triple(TT).getOS()) {
392 return new DarwinX86_64AsmBackend(T);
393 case Triple::MinGW64:
396 return new WindowsX86AsmBackend(T, true);
398 return new ELFX86_64AsmBackend(T, Triple(TT).getOS());