1 //===-- X86AsmBackend.cpp - X86 Assembler Backend -------------------------===//
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 //===----------------------------------------------------------------------===//
9
10 #include "MCTargetDesc/X86BaseInfo.h"
11 #include "MCTargetDesc/X86FixupKinds.h"
12 #include "llvm/ADT/StringSwitch.h"
13 #include "llvm/MC/MCAsmBackend.h"
14 #include "llvm/MC/MCELFObjectWriter.h"
15 #include "llvm/MC/MCExpr.h"
16 #include "llvm/MC/MCFixupKindInfo.h"
17 #include "llvm/MC/MCInst.h"
18 #include "llvm/MC/MCMachObjectWriter.h"
19 #include "llvm/MC/MCObjectWriter.h"
20 #include "llvm/MC/MCRegisterInfo.h"
21 #include "llvm/MC/MCSectionCOFF.h"
22 #include "llvm/MC/MCSectionELF.h"
23 #include "llvm/MC/MCSectionMachO.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/ELF.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/MachO.h"
28 #include "llvm/Support/TargetRegistry.h"
29 #include "llvm/Support/raw_ostream.h"
30 using namespace llvm;
31
getFixupKindLog2Size(unsigned Kind)32 static unsigned getFixupKindLog2Size(unsigned Kind) {
33 switch (Kind) {
34 default:
35 llvm_unreachable("invalid fixup kind!");
36 case FK_PCRel_1:
37 case FK_SecRel_1:
38 case FK_Data_1:
39 return 0;
40 case FK_PCRel_2:
41 case FK_SecRel_2:
42 case FK_Data_2:
43 return 1;
44 case FK_PCRel_4:
45 case X86::reloc_riprel_4byte:
46 case X86::reloc_riprel_4byte_movq_load:
47 case X86::reloc_signed_4byte:
48 case X86::reloc_global_offset_table:
49 case FK_SecRel_4:
50 case FK_Data_4:
51 return 2;
52 case FK_PCRel_8:
53 case FK_SecRel_8:
54 case FK_Data_8:
55 case X86::reloc_global_offset_table8:
56 return 3;
57 }
58 }
59
60 namespace {
61
62 class X86ELFObjectWriter : public MCELFObjectTargetWriter {
63 public:
X86ELFObjectWriter(bool is64Bit,uint8_t OSABI,uint16_t EMachine,bool HasRelocationAddend,bool foobar)64 X86ELFObjectWriter(bool is64Bit, uint8_t OSABI, uint16_t EMachine,
65 bool HasRelocationAddend, bool foobar)
66 : MCELFObjectTargetWriter(is64Bit, OSABI, EMachine, HasRelocationAddend) {}
67 };
68
69 class X86AsmBackend : public MCAsmBackend {
70 const StringRef CPU;
71 bool HasNopl;
72 uint64_t MaxNopLength;
73 public:
X86AsmBackend(const Target & T,StringRef CPU)74 X86AsmBackend(const Target &T, StringRef CPU) : MCAsmBackend(), CPU(CPU) {
75 HasNopl = CPU != "generic" && CPU != "i386" && CPU != "i486" &&
76 CPU != "i586" && CPU != "pentium" && CPU != "pentium-mmx" &&
77 CPU != "i686" && CPU != "k6" && CPU != "k6-2" && CPU != "k6-3" &&
78 CPU != "geode" && CPU != "winchip-c6" && CPU != "winchip2" &&
79 CPU != "c3" && CPU != "c3-2";
80 // Max length of true long nop instruction is 15 bytes.
81 // Max length of long nop replacement instruction is 7 bytes.
82 // Taking into account SilverMont architecture features max length of nops
83 // is reduced for it to achieve better performance.
84 MaxNopLength = (!HasNopl || CPU == "slm") ? 7 : 15;
85 }
86
getNumFixupKinds() const87 unsigned getNumFixupKinds() const override {
88 return X86::NumTargetFixupKinds;
89 }
90
getFixupKindInfo(MCFixupKind Kind) const91 const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override {
92 const static MCFixupKindInfo Infos[X86::NumTargetFixupKinds] = {
93 { "reloc_riprel_4byte", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel },
94 { "reloc_riprel_4byte_movq_load", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel},
95 { "reloc_signed_4byte", 0, 4 * 8, 0},
96 { "reloc_global_offset_table", 0, 4 * 8, 0}
97 };
98
99 if (Kind < FirstTargetFixupKind)
100 return MCAsmBackend::getFixupKindInfo(Kind);
101
102 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
103 "Invalid kind!");
104 return Infos[Kind - FirstTargetFixupKind];
105 }
106
applyFixup(const MCFixup & Fixup,char * Data,unsigned DataSize,uint64_t Value,bool IsPCRel) const107 void applyFixup(const MCFixup &Fixup, char *Data, unsigned DataSize,
108 uint64_t Value, bool IsPCRel) const override {
109 unsigned Size = 1 << getFixupKindLog2Size(Fixup.getKind());
110
111 assert(Fixup.getOffset() + Size <= DataSize &&
112 "Invalid fixup offset!");
113
114 // Check that uppper bits are either all zeros or all ones.
115 // Specifically ignore overflow/underflow as long as the leakage is
116 // limited to the lower bits. This is to remain compatible with
117 // other assemblers.
118 assert(isIntN(Size * 8 + 1, Value) &&
119 "Value does not fit in the Fixup field");
120
121 for (unsigned i = 0; i != Size; ++i)
122 Data[Fixup.getOffset() + i] = uint8_t(Value >> (i * 8));
123 }
124
125 bool mayNeedRelaxation(const MCInst &Inst) const override;
126
127 bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
128 const MCRelaxableFragment *DF,
129 const MCAsmLayout &Layout) const override;
130
131 void relaxInstruction(const MCInst &Inst, MCInst &Res) const override;
132
133 bool writeNopData(uint64_t Count, MCObjectWriter *OW) const override;
134 };
135 } // end anonymous namespace
136
getRelaxedOpcodeBranch(unsigned Op)137 static unsigned getRelaxedOpcodeBranch(unsigned Op) {
138 switch (Op) {
139 default:
140 return Op;
141
142 case X86::JAE_1: return X86::JAE_4;
143 case X86::JA_1: return X86::JA_4;
144 case X86::JBE_1: return X86::JBE_4;
145 case X86::JB_1: return X86::JB_4;
146 case X86::JE_1: return X86::JE_4;
147 case X86::JGE_1: return X86::JGE_4;
148 case X86::JG_1: return X86::JG_4;
149 case X86::JLE_1: return X86::JLE_4;
150 case X86::JL_1: return X86::JL_4;
151 case X86::JMP_1: return X86::JMP_4;
152 case X86::JNE_1: return X86::JNE_4;
153 case X86::JNO_1: return X86::JNO_4;
154 case X86::JNP_1: return X86::JNP_4;
155 case X86::JNS_1: return X86::JNS_4;
156 case X86::JO_1: return X86::JO_4;
157 case X86::JP_1: return X86::JP_4;
158 case X86::JS_1: return X86::JS_4;
159 }
160 }
161
getRelaxedOpcodeArith(unsigned Op)162 static unsigned getRelaxedOpcodeArith(unsigned Op) {
163 switch (Op) {
164 default:
165 return Op;
166
167 // IMUL
168 case X86::IMUL16rri8: return X86::IMUL16rri;
169 case X86::IMUL16rmi8: return X86::IMUL16rmi;
170 case X86::IMUL32rri8: return X86::IMUL32rri;
171 case X86::IMUL32rmi8: return X86::IMUL32rmi;
172 case X86::IMUL64rri8: return X86::IMUL64rri32;
173 case X86::IMUL64rmi8: return X86::IMUL64rmi32;
174
175 // AND
176 case X86::AND16ri8: return X86::AND16ri;
177 case X86::AND16mi8: return X86::AND16mi;
178 case X86::AND32ri8: return X86::AND32ri;
179 case X86::AND32mi8: return X86::AND32mi;
180 case X86::AND64ri8: return X86::AND64ri32;
181 case X86::AND64mi8: return X86::AND64mi32;
182
183 // OR
184 case X86::OR16ri8: return X86::OR16ri;
185 case X86::OR16mi8: return X86::OR16mi;
186 case X86::OR32ri8: return X86::OR32ri;
187 case X86::OR32mi8: return X86::OR32mi;
188 case X86::OR64ri8: return X86::OR64ri32;
189 case X86::OR64mi8: return X86::OR64mi32;
190
191 // XOR
192 case X86::XOR16ri8: return X86::XOR16ri;
193 case X86::XOR16mi8: return X86::XOR16mi;
194 case X86::XOR32ri8: return X86::XOR32ri;
195 case X86::XOR32mi8: return X86::XOR32mi;
196 case X86::XOR64ri8: return X86::XOR64ri32;
197 case X86::XOR64mi8: return X86::XOR64mi32;
198
199 // ADD
200 case X86::ADD16ri8: return X86::ADD16ri;
201 case X86::ADD16mi8: return X86::ADD16mi;
202 case X86::ADD32ri8: return X86::ADD32ri;
203 case X86::ADD32mi8: return X86::ADD32mi;
204 case X86::ADD64ri8: return X86::ADD64ri32;
205 case X86::ADD64mi8: return X86::ADD64mi32;
206
207 // ADC
208 case X86::ADC16ri8: return X86::ADC16ri;
209 case X86::ADC16mi8: return X86::ADC16mi;
210 case X86::ADC32ri8: return X86::ADC32ri;
211 case X86::ADC32mi8: return X86::ADC32mi;
212 case X86::ADC64ri8: return X86::ADC64ri32;
213 case X86::ADC64mi8: return X86::ADC64mi32;
214
215 // SUB
216 case X86::SUB16ri8: return X86::SUB16ri;
217 case X86::SUB16mi8: return X86::SUB16mi;
218 case X86::SUB32ri8: return X86::SUB32ri;
219 case X86::SUB32mi8: return X86::SUB32mi;
220 case X86::SUB64ri8: return X86::SUB64ri32;
221 case X86::SUB64mi8: return X86::SUB64mi32;
222
223 // SBB
224 case X86::SBB16ri8: return X86::SBB16ri;
225 case X86::SBB16mi8: return X86::SBB16mi;
226 case X86::SBB32ri8: return X86::SBB32ri;
227 case X86::SBB32mi8: return X86::SBB32mi;
228 case X86::SBB64ri8: return X86::SBB64ri32;
229 case X86::SBB64mi8: return X86::SBB64mi32;
230
231 // CMP
232 case X86::CMP16ri8: return X86::CMP16ri;
233 case X86::CMP16mi8: return X86::CMP16mi;
234 case X86::CMP32ri8: return X86::CMP32ri;
235 case X86::CMP32mi8: return X86::CMP32mi;
236 case X86::CMP64ri8: return X86::CMP64ri32;
237 case X86::CMP64mi8: return X86::CMP64mi32;
238
239 // PUSH
240 case X86::PUSH32i8: return X86::PUSHi32;
241 case X86::PUSH16i8: return X86::PUSHi16;
242 case X86::PUSH64i8: return X86::PUSH64i32;
243 }
244 }
245
getRelaxedOpcode(unsigned Op)246 static unsigned getRelaxedOpcode(unsigned Op) {
247 unsigned R = getRelaxedOpcodeArith(Op);
248 if (R != Op)
249 return R;
250 return getRelaxedOpcodeBranch(Op);
251 }
252
mayNeedRelaxation(const MCInst & Inst) const253 bool X86AsmBackend::mayNeedRelaxation(const MCInst &Inst) const {
254 // Branches can always be relaxed.
255 if (getRelaxedOpcodeBranch(Inst.getOpcode()) != Inst.getOpcode())
256 return true;
257
258 // Check if this instruction is ever relaxable.
259 if (getRelaxedOpcodeArith(Inst.getOpcode()) == Inst.getOpcode())
260 return false;
261
262
263 // Check if the relaxable operand has an expression. For the current set of
264 // relaxable instructions, the relaxable operand is always the last operand.
265 unsigned RelaxableOp = Inst.getNumOperands() - 1;
266 if (Inst.getOperand(RelaxableOp).isExpr())
267 return true;
268
269 return false;
270 }
271
fixupNeedsRelaxation(const MCFixup & Fixup,uint64_t Value,const MCRelaxableFragment * DF,const MCAsmLayout & Layout) const272 bool X86AsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup,
273 uint64_t Value,
274 const MCRelaxableFragment *DF,
275 const MCAsmLayout &Layout) const {
276 // Relax if the value is too big for a (signed) i8.
277 return int64_t(Value) != int64_t(int8_t(Value));
278 }
279
280 // FIXME: Can tblgen help at all here to verify there aren't other instructions
281 // we can relax?
relaxInstruction(const MCInst & Inst,MCInst & Res) const282 void X86AsmBackend::relaxInstruction(const MCInst &Inst, MCInst &Res) const {
283 // The only relaxations X86 does is from a 1byte pcrel to a 4byte pcrel.
284 unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode());
285
286 if (RelaxedOp == Inst.getOpcode()) {
287 SmallString<256> Tmp;
288 raw_svector_ostream OS(Tmp);
289 Inst.dump_pretty(OS);
290 OS << "\n";
291 report_fatal_error("unexpected instruction to relax: " + OS.str());
292 }
293
294 Res = Inst;
295 Res.setOpcode(RelaxedOp);
296 }
297
298 /// \brief Write a sequence of optimal nops to the output, covering \p Count
299 /// bytes.
300 /// \return - true on success, false on failure
writeNopData(uint64_t Count,MCObjectWriter * OW) const301 bool X86AsmBackend::writeNopData(uint64_t Count, MCObjectWriter *OW) const {
302 static const uint8_t TrueNops[10][10] = {
303 // nop
304 {0x90},
305 // xchg %ax,%ax
306 {0x66, 0x90},
307 // nopl (%[re]ax)
308 {0x0f, 0x1f, 0x00},
309 // nopl 0(%[re]ax)
310 {0x0f, 0x1f, 0x40, 0x00},
311 // nopl 0(%[re]ax,%[re]ax,1)
312 {0x0f, 0x1f, 0x44, 0x00, 0x00},
313 // nopw 0(%[re]ax,%[re]ax,1)
314 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
315 // nopl 0L(%[re]ax)
316 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
317 // nopl 0L(%[re]ax,%[re]ax,1)
318 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
319 // nopw 0L(%[re]ax,%[re]ax,1)
320 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
321 // nopw %cs:0L(%[re]ax,%[re]ax,1)
322 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
323 };
324
325 // Alternative nop instructions for CPUs which don't support long nops.
326 static const uint8_t AltNops[7][10] = {
327 // nop
328 {0x90},
329 // xchg %ax,%ax
330 {0x66, 0x90},
331 // lea 0x0(%esi),%esi
332 {0x8d, 0x76, 0x00},
333 // lea 0x0(%esi),%esi
334 {0x8d, 0x74, 0x26, 0x00},
335 // nop + lea 0x0(%esi),%esi
336 {0x90, 0x8d, 0x74, 0x26, 0x00},
337 // lea 0x0(%esi),%esi
338 {0x8d, 0xb6, 0x00, 0x00, 0x00, 0x00 },
339 // lea 0x0(%esi),%esi
340 {0x8d, 0xb4, 0x26, 0x00, 0x00, 0x00, 0x00},
341 };
342
343 // Select the right NOP table.
344 // FIXME: Can we get if CPU supports long nops from the subtarget somehow?
345 const uint8_t (*Nops)[10] = HasNopl ? TrueNops : AltNops;
346 assert(HasNopl || MaxNopLength <= 7);
347
348 // Emit as many largest nops as needed, then emit a nop of the remaining
349 // length.
350 do {
351 const uint8_t ThisNopLength = (uint8_t) std::min(Count, MaxNopLength);
352 const uint8_t Prefixes = ThisNopLength <= 10 ? 0 : ThisNopLength - 10;
353 for (uint8_t i = 0; i < Prefixes; i++)
354 OW->write8(0x66);
355 const uint8_t Rest = ThisNopLength - Prefixes;
356 for (uint8_t i = 0; i < Rest; i++)
357 OW->write8(Nops[Rest - 1][i]);
358 Count -= ThisNopLength;
359 } while (Count != 0);
360
361 return true;
362 }
363
364 /* *** */
365
366 namespace {
367
368 class ELFX86AsmBackend : public X86AsmBackend {
369 public:
370 uint8_t OSABI;
ELFX86AsmBackend(const Target & T,uint8_t OSABI,StringRef CPU)371 ELFX86AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
372 : X86AsmBackend(T, CPU), OSABI(OSABI) {}
373 };
374
375 class ELFX86_32AsmBackend : public ELFX86AsmBackend {
376 public:
ELFX86_32AsmBackend(const Target & T,uint8_t OSABI,StringRef CPU)377 ELFX86_32AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
378 : ELFX86AsmBackend(T, OSABI, CPU) {}
379
createObjectWriter(raw_pwrite_stream & OS) const380 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
381 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI, ELF::EM_386);
382 }
383 };
384
385 class ELFX86_X32AsmBackend : public ELFX86AsmBackend {
386 public:
ELFX86_X32AsmBackend(const Target & T,uint8_t OSABI,StringRef CPU)387 ELFX86_X32AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
388 : ELFX86AsmBackend(T, OSABI, CPU) {}
389
createObjectWriter(raw_pwrite_stream & OS) const390 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
391 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI,
392 ELF::EM_X86_64);
393 }
394 };
395
396 class ELFX86_IAMCUAsmBackend : public ELFX86AsmBackend {
397 public:
ELFX86_IAMCUAsmBackend(const Target & T,uint8_t OSABI,StringRef CPU)398 ELFX86_IAMCUAsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
399 : ELFX86AsmBackend(T, OSABI, CPU) {}
400
createObjectWriter(raw_pwrite_stream & OS) const401 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
402 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI,
403 ELF::EM_IAMCU);
404 }
405 };
406
407 class ELFX86_64AsmBackend : public ELFX86AsmBackend {
408 public:
ELFX86_64AsmBackend(const Target & T,uint8_t OSABI,StringRef CPU)409 ELFX86_64AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
410 : ELFX86AsmBackend(T, OSABI, CPU) {}
411
createObjectWriter(raw_pwrite_stream & OS) const412 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
413 return createX86ELFObjectWriter(OS, /*IsELF64*/ true, OSABI, ELF::EM_X86_64);
414 }
415 };
416
417 class WindowsX86AsmBackend : public X86AsmBackend {
418 bool Is64Bit;
419
420 public:
WindowsX86AsmBackend(const Target & T,bool is64Bit,StringRef CPU)421 WindowsX86AsmBackend(const Target &T, bool is64Bit, StringRef CPU)
422 : X86AsmBackend(T, CPU)
423 , Is64Bit(is64Bit) {
424 }
425
createObjectWriter(raw_pwrite_stream & OS) const426 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
427 return createX86WinCOFFObjectWriter(OS, Is64Bit);
428 }
429 };
430
431 namespace CU {
432
433 /// Compact unwind encoding values.
434 enum CompactUnwindEncodings {
435 /// [RE]BP based frame where [RE]BP is pused on the stack immediately after
436 /// the return address, then [RE]SP is moved to [RE]BP.
437 UNWIND_MODE_BP_FRAME = 0x01000000,
438
439 /// A frameless function with a small constant stack size.
440 UNWIND_MODE_STACK_IMMD = 0x02000000,
441
442 /// A frameless function with a large constant stack size.
443 UNWIND_MODE_STACK_IND = 0x03000000,
444
445 /// No compact unwind encoding is available.
446 UNWIND_MODE_DWARF = 0x04000000,
447
448 /// Mask for encoding the frame registers.
449 UNWIND_BP_FRAME_REGISTERS = 0x00007FFF,
450
451 /// Mask for encoding the frameless registers.
452 UNWIND_FRAMELESS_STACK_REG_PERMUTATION = 0x000003FF
453 };
454
455 } // end CU namespace
456
457 class DarwinX86AsmBackend : public X86AsmBackend {
458 const MCRegisterInfo &MRI;
459
460 /// \brief Number of registers that can be saved in a compact unwind encoding.
461 enum { CU_NUM_SAVED_REGS = 6 };
462
463 mutable unsigned SavedRegs[CU_NUM_SAVED_REGS];
464 bool Is64Bit;
465
466 unsigned OffsetSize; ///< Offset of a "push" instruction.
467 unsigned MoveInstrSize; ///< Size of a "move" instruction.
468 unsigned StackDivide; ///< Amount to adjust stack size by.
469 protected:
470 /// \brief Size of a "push" instruction for the given register.
PushInstrSize(unsigned Reg) const471 unsigned PushInstrSize(unsigned Reg) const {
472 switch (Reg) {
473 case X86::EBX:
474 case X86::ECX:
475 case X86::EDX:
476 case X86::EDI:
477 case X86::ESI:
478 case X86::EBP:
479 case X86::RBX:
480 case X86::RBP:
481 return 1;
482 case X86::R12:
483 case X86::R13:
484 case X86::R14:
485 case X86::R15:
486 return 2;
487 }
488 return 1;
489 }
490
491 /// \brief Implementation of algorithm to generate the compact unwind encoding
492 /// for the CFI instructions.
493 uint32_t
generateCompactUnwindEncodingImpl(ArrayRef<MCCFIInstruction> Instrs) const494 generateCompactUnwindEncodingImpl(ArrayRef<MCCFIInstruction> Instrs) const {
495 if (Instrs.empty()) return 0;
496
497 // Reset the saved registers.
498 unsigned SavedRegIdx = 0;
499 memset(SavedRegs, 0, sizeof(SavedRegs));
500
501 bool HasFP = false;
502
503 // Encode that we are using EBP/RBP as the frame pointer.
504 uint32_t CompactUnwindEncoding = 0;
505
506 unsigned SubtractInstrIdx = Is64Bit ? 3 : 2;
507 unsigned InstrOffset = 0;
508 unsigned StackAdjust = 0;
509 unsigned StackSize = 0;
510 unsigned PrevStackSize = 0;
511 unsigned NumDefCFAOffsets = 0;
512
513 for (unsigned i = 0, e = Instrs.size(); i != e; ++i) {
514 const MCCFIInstruction &Inst = Instrs[i];
515
516 switch (Inst.getOperation()) {
517 default:
518 // Any other CFI directives indicate a frame that we aren't prepared
519 // to represent via compact unwind, so just bail out.
520 return 0;
521 case MCCFIInstruction::OpDefCfaRegister: {
522 // Defines a frame pointer. E.g.
523 //
524 // movq %rsp, %rbp
525 // L0:
526 // .cfi_def_cfa_register %rbp
527 //
528 HasFP = true;
529 assert(MRI.getLLVMRegNum(Inst.getRegister(), true) ==
530 (Is64Bit ? X86::RBP : X86::EBP) && "Invalid frame pointer!");
531
532 // Reset the counts.
533 memset(SavedRegs, 0, sizeof(SavedRegs));
534 StackAdjust = 0;
535 SavedRegIdx = 0;
536 InstrOffset += MoveInstrSize;
537 break;
538 }
539 case MCCFIInstruction::OpDefCfaOffset: {
540 // Defines a new offset for the CFA. E.g.
541 //
542 // With frame:
543 //
544 // pushq %rbp
545 // L0:
546 // .cfi_def_cfa_offset 16
547 //
548 // Without frame:
549 //
550 // subq $72, %rsp
551 // L0:
552 // .cfi_def_cfa_offset 80
553 //
554 PrevStackSize = StackSize;
555 StackSize = std::abs(Inst.getOffset()) / StackDivide;
556 ++NumDefCFAOffsets;
557 break;
558 }
559 case MCCFIInstruction::OpOffset: {
560 // Defines a "push" of a callee-saved register. E.g.
561 //
562 // pushq %r15
563 // pushq %r14
564 // pushq %rbx
565 // L0:
566 // subq $120, %rsp
567 // L1:
568 // .cfi_offset %rbx, -40
569 // .cfi_offset %r14, -32
570 // .cfi_offset %r15, -24
571 //
572 if (SavedRegIdx == CU_NUM_SAVED_REGS)
573 // If there are too many saved registers, we cannot use a compact
574 // unwind encoding.
575 return CU::UNWIND_MODE_DWARF;
576
577 unsigned Reg = MRI.getLLVMRegNum(Inst.getRegister(), true);
578 SavedRegs[SavedRegIdx++] = Reg;
579 StackAdjust += OffsetSize;
580 InstrOffset += PushInstrSize(Reg);
581 break;
582 }
583 }
584 }
585
586 StackAdjust /= StackDivide;
587
588 if (HasFP) {
589 if ((StackAdjust & 0xFF) != StackAdjust)
590 // Offset was too big for a compact unwind encoding.
591 return CU::UNWIND_MODE_DWARF;
592
593 // Get the encoding of the saved registers when we have a frame pointer.
594 uint32_t RegEnc = encodeCompactUnwindRegistersWithFrame();
595 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF;
596
597 CompactUnwindEncoding |= CU::UNWIND_MODE_BP_FRAME;
598 CompactUnwindEncoding |= (StackAdjust & 0xFF) << 16;
599 CompactUnwindEncoding |= RegEnc & CU::UNWIND_BP_FRAME_REGISTERS;
600 } else {
601 // If the amount of the stack allocation is the size of a register, then
602 // we "push" the RAX/EAX register onto the stack instead of adjusting the
603 // stack pointer with a SUB instruction. We don't support the push of the
604 // RAX/EAX register with compact unwind. So we check for that situation
605 // here.
606 if ((NumDefCFAOffsets == SavedRegIdx + 1 &&
607 StackSize - PrevStackSize == 1) ||
608 (Instrs.size() == 1 && NumDefCFAOffsets == 1 && StackSize == 2))
609 return CU::UNWIND_MODE_DWARF;
610
611 SubtractInstrIdx += InstrOffset;
612 ++StackAdjust;
613
614 if ((StackSize & 0xFF) == StackSize) {
615 // Frameless stack with a small stack size.
616 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IMMD;
617
618 // Encode the stack size.
619 CompactUnwindEncoding |= (StackSize & 0xFF) << 16;
620 } else {
621 if ((StackAdjust & 0x7) != StackAdjust)
622 // The extra stack adjustments are too big for us to handle.
623 return CU::UNWIND_MODE_DWARF;
624
625 // Frameless stack with an offset too large for us to encode compactly.
626 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IND;
627
628 // Encode the offset to the nnnnnn value in the 'subl $nnnnnn, ESP'
629 // instruction.
630 CompactUnwindEncoding |= (SubtractInstrIdx & 0xFF) << 16;
631
632 // Encode any extra stack stack adjustments (done via push
633 // instructions).
634 CompactUnwindEncoding |= (StackAdjust & 0x7) << 13;
635 }
636
637 // Encode the number of registers saved. (Reverse the list first.)
638 std::reverse(&SavedRegs[0], &SavedRegs[SavedRegIdx]);
639 CompactUnwindEncoding |= (SavedRegIdx & 0x7) << 10;
640
641 // Get the encoding of the saved registers when we don't have a frame
642 // pointer.
643 uint32_t RegEnc = encodeCompactUnwindRegistersWithoutFrame(SavedRegIdx);
644 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF;
645
646 // Encode the register encoding.
647 CompactUnwindEncoding |=
648 RegEnc & CU::UNWIND_FRAMELESS_STACK_REG_PERMUTATION;
649 }
650
651 return CompactUnwindEncoding;
652 }
653
654 private:
655 /// \brief Get the compact unwind number for a given register. The number
656 /// corresponds to the enum lists in compact_unwind_encoding.h.
getCompactUnwindRegNum(unsigned Reg) const657 int getCompactUnwindRegNum(unsigned Reg) const {
658 static const MCPhysReg CU32BitRegs[7] = {
659 X86::EBX, X86::ECX, X86::EDX, X86::EDI, X86::ESI, X86::EBP, 0
660 };
661 static const MCPhysReg CU64BitRegs[] = {
662 X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0
663 };
664 const MCPhysReg *CURegs = Is64Bit ? CU64BitRegs : CU32BitRegs;
665 for (int Idx = 1; *CURegs; ++CURegs, ++Idx)
666 if (*CURegs == Reg)
667 return Idx;
668
669 return -1;
670 }
671
672 /// \brief Return the registers encoded for a compact encoding with a frame
673 /// pointer.
encodeCompactUnwindRegistersWithFrame() const674 uint32_t encodeCompactUnwindRegistersWithFrame() const {
675 // Encode the registers in the order they were saved --- 3-bits per
676 // register. The list of saved registers is assumed to be in reverse
677 // order. The registers are numbered from 1 to CU_NUM_SAVED_REGS.
678 uint32_t RegEnc = 0;
679 for (int i = 0, Idx = 0; i != CU_NUM_SAVED_REGS; ++i) {
680 unsigned Reg = SavedRegs[i];
681 if (Reg == 0) break;
682
683 int CURegNum = getCompactUnwindRegNum(Reg);
684 if (CURegNum == -1) return ~0U;
685
686 // Encode the 3-bit register number in order, skipping over 3-bits for
687 // each register.
688 RegEnc |= (CURegNum & 0x7) << (Idx++ * 3);
689 }
690
691 assert((RegEnc & 0x3FFFF) == RegEnc &&
692 "Invalid compact register encoding!");
693 return RegEnc;
694 }
695
696 /// \brief Create the permutation encoding used with frameless stacks. It is
697 /// passed the number of registers to be saved and an array of the registers
698 /// saved.
encodeCompactUnwindRegistersWithoutFrame(unsigned RegCount) const699 uint32_t encodeCompactUnwindRegistersWithoutFrame(unsigned RegCount) const {
700 // The saved registers are numbered from 1 to 6. In order to encode the
701 // order in which they were saved, we re-number them according to their
702 // place in the register order. The re-numbering is relative to the last
703 // re-numbered register. E.g., if we have registers {6, 2, 4, 5} saved in
704 // that order:
705 //
706 // Orig Re-Num
707 // ---- ------
708 // 6 6
709 // 2 2
710 // 4 3
711 // 5 3
712 //
713 for (unsigned i = 0; i < RegCount; ++i) {
714 int CUReg = getCompactUnwindRegNum(SavedRegs[i]);
715 if (CUReg == -1) return ~0U;
716 SavedRegs[i] = CUReg;
717 }
718
719 // Reverse the list.
720 std::reverse(&SavedRegs[0], &SavedRegs[CU_NUM_SAVED_REGS]);
721
722 uint32_t RenumRegs[CU_NUM_SAVED_REGS];
723 for (unsigned i = CU_NUM_SAVED_REGS - RegCount; i < CU_NUM_SAVED_REGS; ++i){
724 unsigned Countless = 0;
725 for (unsigned j = CU_NUM_SAVED_REGS - RegCount; j < i; ++j)
726 if (SavedRegs[j] < SavedRegs[i])
727 ++Countless;
728
729 RenumRegs[i] = SavedRegs[i] - Countless - 1;
730 }
731
732 // Take the renumbered values and encode them into a 10-bit number.
733 uint32_t permutationEncoding = 0;
734 switch (RegCount) {
735 case 6:
736 permutationEncoding |= 120 * RenumRegs[0] + 24 * RenumRegs[1]
737 + 6 * RenumRegs[2] + 2 * RenumRegs[3]
738 + RenumRegs[4];
739 break;
740 case 5:
741 permutationEncoding |= 120 * RenumRegs[1] + 24 * RenumRegs[2]
742 + 6 * RenumRegs[3] + 2 * RenumRegs[4]
743 + RenumRegs[5];
744 break;
745 case 4:
746 permutationEncoding |= 60 * RenumRegs[2] + 12 * RenumRegs[3]
747 + 3 * RenumRegs[4] + RenumRegs[5];
748 break;
749 case 3:
750 permutationEncoding |= 20 * RenumRegs[3] + 4 * RenumRegs[4]
751 + RenumRegs[5];
752 break;
753 case 2:
754 permutationEncoding |= 5 * RenumRegs[4] + RenumRegs[5];
755 break;
756 case 1:
757 permutationEncoding |= RenumRegs[5];
758 break;
759 }
760
761 assert((permutationEncoding & 0x3FF) == permutationEncoding &&
762 "Invalid compact register encoding!");
763 return permutationEncoding;
764 }
765
766 public:
DarwinX86AsmBackend(const Target & T,const MCRegisterInfo & MRI,StringRef CPU,bool Is64Bit)767 DarwinX86AsmBackend(const Target &T, const MCRegisterInfo &MRI, StringRef CPU,
768 bool Is64Bit)
769 : X86AsmBackend(T, CPU), MRI(MRI), Is64Bit(Is64Bit) {
770 memset(SavedRegs, 0, sizeof(SavedRegs));
771 OffsetSize = Is64Bit ? 8 : 4;
772 MoveInstrSize = Is64Bit ? 3 : 2;
773 StackDivide = Is64Bit ? 8 : 4;
774 }
775 };
776
777 class DarwinX86_32AsmBackend : public DarwinX86AsmBackend {
778 public:
DarwinX86_32AsmBackend(const Target & T,const MCRegisterInfo & MRI,StringRef CPU)779 DarwinX86_32AsmBackend(const Target &T, const MCRegisterInfo &MRI,
780 StringRef CPU)
781 : DarwinX86AsmBackend(T, MRI, CPU, false) {}
782
createObjectWriter(raw_pwrite_stream & OS) const783 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
784 return createX86MachObjectWriter(OS, /*Is64Bit=*/false,
785 MachO::CPU_TYPE_I386,
786 MachO::CPU_SUBTYPE_I386_ALL);
787 }
788
789 /// \brief Generate the compact unwind encoding for the CFI instructions.
generateCompactUnwindEncoding(ArrayRef<MCCFIInstruction> Instrs) const790 uint32_t generateCompactUnwindEncoding(
791 ArrayRef<MCCFIInstruction> Instrs) const override {
792 return generateCompactUnwindEncodingImpl(Instrs);
793 }
794 };
795
796 class DarwinX86_64AsmBackend : public DarwinX86AsmBackend {
797 const MachO::CPUSubTypeX86 Subtype;
798 public:
DarwinX86_64AsmBackend(const Target & T,const MCRegisterInfo & MRI,StringRef CPU,MachO::CPUSubTypeX86 st)799 DarwinX86_64AsmBackend(const Target &T, const MCRegisterInfo &MRI,
800 StringRef CPU, MachO::CPUSubTypeX86 st)
801 : DarwinX86AsmBackend(T, MRI, CPU, true), Subtype(st) {}
802
createObjectWriter(raw_pwrite_stream & OS) const803 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
804 return createX86MachObjectWriter(OS, /*Is64Bit=*/true,
805 MachO::CPU_TYPE_X86_64, Subtype);
806 }
807
808 /// \brief Generate the compact unwind encoding for the CFI instructions.
generateCompactUnwindEncoding(ArrayRef<MCCFIInstruction> Instrs) const809 uint32_t generateCompactUnwindEncoding(
810 ArrayRef<MCCFIInstruction> Instrs) const override {
811 return generateCompactUnwindEncodingImpl(Instrs);
812 }
813 };
814
815 } // end anonymous namespace
816
createX86_32AsmBackend(const Target & T,const MCRegisterInfo & MRI,const Triple & TheTriple,StringRef CPU)817 MCAsmBackend *llvm::createX86_32AsmBackend(const Target &T,
818 const MCRegisterInfo &MRI,
819 const Triple &TheTriple,
820 StringRef CPU) {
821 if (TheTriple.isOSBinFormatMachO())
822 return new DarwinX86_32AsmBackend(T, MRI, CPU);
823
824 if (TheTriple.isOSWindows() && !TheTriple.isOSBinFormatELF())
825 return new WindowsX86AsmBackend(T, false, CPU);
826
827 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
828
829 if (TheTriple.isOSIAMCU())
830 return new ELFX86_IAMCUAsmBackend(T, OSABI, CPU);
831
832 return new ELFX86_32AsmBackend(T, OSABI, CPU);
833 }
834
createX86_64AsmBackend(const Target & T,const MCRegisterInfo & MRI,const Triple & TheTriple,StringRef CPU)835 MCAsmBackend *llvm::createX86_64AsmBackend(const Target &T,
836 const MCRegisterInfo &MRI,
837 const Triple &TheTriple,
838 StringRef CPU) {
839 if (TheTriple.isOSBinFormatMachO()) {
840 MachO::CPUSubTypeX86 CS =
841 StringSwitch<MachO::CPUSubTypeX86>(TheTriple.getArchName())
842 .Case("x86_64h", MachO::CPU_SUBTYPE_X86_64_H)
843 .Default(MachO::CPU_SUBTYPE_X86_64_ALL);
844 return new DarwinX86_64AsmBackend(T, MRI, CPU, CS);
845 }
846
847 if (TheTriple.isOSWindows() && !TheTriple.isOSBinFormatELF())
848 return new WindowsX86AsmBackend(T, true, CPU);
849
850 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
851
852 if (TheTriple.getEnvironment() == Triple::GNUX32)
853 return new ELFX86_X32AsmBackend(T, OSABI, CPU);
854 return new ELFX86_64AsmBackend(T, OSABI, CPU);
855 }
856