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 "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"
32 static unsigned getFixupKindLog2Size(unsigned Kind) {
35 llvm_unreachable("invalid fixup kind!");
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:
55 case X86::reloc_global_offset_table8:
62 class X86ELFObjectWriter : public MCELFObjectTargetWriter {
64 X86ELFObjectWriter(bool is64Bit, uint8_t OSABI, uint16_t EMachine,
65 bool HasRelocationAddend, bool foobar)
66 : MCELFObjectTargetWriter(is64Bit, OSABI, EMachine, HasRelocationAddend) {}
69 class X86AsmBackend : public MCAsmBackend {
72 uint64_t MaxNopLength;
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;
87 unsigned getNumFixupKinds() const override {
88 return X86::NumTargetFixupKinds;
91 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}
99 if (Kind < FirstTargetFixupKind)
100 return MCAsmBackend::getFixupKindInfo(Kind);
102 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
104 return Infos[Kind - FirstTargetFixupKind];
107 void applyFixup(const MCFixup &Fixup, char *Data, unsigned DataSize,
108 uint64_t Value, bool IsPCRel) const override {
109 unsigned Size = 1 << getFixupKindLog2Size(Fixup.getKind());
111 assert(Fixup.getOffset() + Size <= DataSize &&
112 "Invalid fixup offset!");
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
118 assert(isIntN(Size * 8 + 1, Value) &&
119 "Value does not fit in the Fixup field");
121 for (unsigned i = 0; i != Size; ++i)
122 Data[Fixup.getOffset() + i] = uint8_t(Value >> (i * 8));
125 bool mayNeedRelaxation(const MCInst &Inst) const override;
127 bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
128 const MCRelaxableFragment *DF,
129 const MCAsmLayout &Layout) const override;
131 void relaxInstruction(const MCInst &Inst, MCInst &Res) const override;
133 bool writeNopData(uint64_t Count, MCObjectWriter *OW) const override;
135 } // end anonymous namespace
137 static unsigned getRelaxedOpcodeBranch(unsigned Op) {
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;
162 static unsigned getRelaxedOpcodeArith(unsigned Op) {
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;
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;
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;
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;
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;
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;
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;
224 case X86::CMP16ri8: return X86::CMP16ri;
225 case X86::CMP16mi8: return X86::CMP16mi;
226 case X86::CMP32ri8: return X86::CMP32ri;
227 case X86::CMP32mi8: return X86::CMP32mi;
228 case X86::CMP64ri8: return X86::CMP64ri32;
229 case X86::CMP64mi8: return X86::CMP64mi32;
232 case X86::PUSH32i8: return X86::PUSHi32;
233 case X86::PUSH16i8: return X86::PUSHi16;
234 case X86::PUSH64i8: return X86::PUSH64i32;
238 static unsigned getRelaxedOpcode(unsigned Op) {
239 unsigned R = getRelaxedOpcodeArith(Op);
242 return getRelaxedOpcodeBranch(Op);
245 bool X86AsmBackend::mayNeedRelaxation(const MCInst &Inst) const {
246 // Branches can always be relaxed.
247 if (getRelaxedOpcodeBranch(Inst.getOpcode()) != Inst.getOpcode())
250 // Check if this instruction is ever relaxable.
251 if (getRelaxedOpcodeArith(Inst.getOpcode()) == Inst.getOpcode())
255 // Check if the relaxable operand has an expression. For the current set of
256 // relaxable instructions, the relaxable operand is always the last operand.
257 unsigned RelaxableOp = Inst.getNumOperands() - 1;
258 if (Inst.getOperand(RelaxableOp).isExpr())
264 bool X86AsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup,
266 const MCRelaxableFragment *DF,
267 const MCAsmLayout &Layout) const {
268 // Relax if the value is too big for a (signed) i8.
269 return int64_t(Value) != int64_t(int8_t(Value));
272 // FIXME: Can tblgen help at all here to verify there aren't other instructions
274 void X86AsmBackend::relaxInstruction(const MCInst &Inst, MCInst &Res) const {
275 // The only relaxations X86 does is from a 1byte pcrel to a 4byte pcrel.
276 unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode());
278 if (RelaxedOp == Inst.getOpcode()) {
279 SmallString<256> Tmp;
280 raw_svector_ostream OS(Tmp);
281 Inst.dump_pretty(OS);
283 report_fatal_error("unexpected instruction to relax: " + OS.str());
287 Res.setOpcode(RelaxedOp);
290 /// \brief Write a sequence of optimal nops to the output, covering \p Count
292 /// \return - true on success, false on failure
293 bool X86AsmBackend::writeNopData(uint64_t Count, MCObjectWriter *OW) const {
294 static const uint8_t TrueNops[10][10] = {
302 {0x0f, 0x1f, 0x40, 0x00},
303 // nopl 0(%[re]ax,%[re]ax,1)
304 {0x0f, 0x1f, 0x44, 0x00, 0x00},
305 // nopw 0(%[re]ax,%[re]ax,1)
306 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
308 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
309 // nopl 0L(%[re]ax,%[re]ax,1)
310 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
311 // nopw 0L(%[re]ax,%[re]ax,1)
312 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
313 // nopw %cs:0L(%[re]ax,%[re]ax,1)
314 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
317 // Alternative nop instructions for CPUs which don't support long nops.
318 static const uint8_t AltNops[7][10] = {
323 // lea 0x0(%esi),%esi
325 // lea 0x0(%esi),%esi
326 {0x8d, 0x74, 0x26, 0x00},
327 // nop + lea 0x0(%esi),%esi
328 {0x90, 0x8d, 0x74, 0x26, 0x00},
329 // lea 0x0(%esi),%esi
330 {0x8d, 0xb6, 0x00, 0x00, 0x00, 0x00 },
331 // lea 0x0(%esi),%esi
332 {0x8d, 0xb4, 0x26, 0x00, 0x00, 0x00, 0x00},
335 // Select the right NOP table.
336 // FIXME: Can we get if CPU supports long nops from the subtarget somehow?
337 const uint8_t (*Nops)[10] = HasNopl ? TrueNops : AltNops;
338 assert(HasNopl || MaxNopLength <= 7);
340 // Emit as many largest nops as needed, then emit a nop of the remaining
343 const uint8_t ThisNopLength = (uint8_t) std::min(Count, MaxNopLength);
344 const uint8_t Prefixes = ThisNopLength <= 10 ? 0 : ThisNopLength - 10;
345 for (uint8_t i = 0; i < Prefixes; i++)
347 const uint8_t Rest = ThisNopLength - Prefixes;
348 for (uint8_t i = 0; i < Rest; i++)
349 OW->write8(Nops[Rest - 1][i]);
350 Count -= ThisNopLength;
351 } while (Count != 0);
360 class ELFX86AsmBackend : public X86AsmBackend {
363 ELFX86AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
364 : X86AsmBackend(T, CPU), OSABI(OSABI) {}
367 class ELFX86_32AsmBackend : public ELFX86AsmBackend {
369 ELFX86_32AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
370 : ELFX86AsmBackend(T, OSABI, CPU) {}
372 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
373 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI, ELF::EM_386);
377 class ELFX86_X32AsmBackend : public ELFX86AsmBackend {
379 ELFX86_X32AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
380 : ELFX86AsmBackend(T, OSABI, CPU) {}
382 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
383 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI,
388 class ELFX86_IAMCUAsmBackend : public ELFX86AsmBackend {
390 ELFX86_IAMCUAsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
391 : ELFX86AsmBackend(T, OSABI, CPU) {}
393 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
394 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI,
399 class ELFX86_64AsmBackend : public ELFX86AsmBackend {
401 ELFX86_64AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
402 : ELFX86AsmBackend(T, OSABI, CPU) {}
404 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
405 return createX86ELFObjectWriter(OS, /*IsELF64*/ true, OSABI, ELF::EM_X86_64);
409 class WindowsX86AsmBackend : public X86AsmBackend {
413 WindowsX86AsmBackend(const Target &T, bool is64Bit, StringRef CPU)
414 : X86AsmBackend(T, CPU)
418 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
419 return createX86WinCOFFObjectWriter(OS, Is64Bit);
425 /// Compact unwind encoding values.
426 enum CompactUnwindEncodings {
427 /// [RE]BP based frame where [RE]BP is pused on the stack immediately after
428 /// the return address, then [RE]SP is moved to [RE]BP.
429 UNWIND_MODE_BP_FRAME = 0x01000000,
431 /// A frameless function with a small constant stack size.
432 UNWIND_MODE_STACK_IMMD = 0x02000000,
434 /// A frameless function with a large constant stack size.
435 UNWIND_MODE_STACK_IND = 0x03000000,
437 /// No compact unwind encoding is available.
438 UNWIND_MODE_DWARF = 0x04000000,
440 /// Mask for encoding the frame registers.
441 UNWIND_BP_FRAME_REGISTERS = 0x00007FFF,
443 /// Mask for encoding the frameless registers.
444 UNWIND_FRAMELESS_STACK_REG_PERMUTATION = 0x000003FF
447 } // end CU namespace
449 class DarwinX86AsmBackend : public X86AsmBackend {
450 const MCRegisterInfo &MRI;
452 /// \brief Number of registers that can be saved in a compact unwind encoding.
453 enum { CU_NUM_SAVED_REGS = 6 };
455 mutable unsigned SavedRegs[CU_NUM_SAVED_REGS];
458 unsigned OffsetSize; ///< Offset of a "push" instruction.
459 unsigned MoveInstrSize; ///< Size of a "move" instruction.
460 unsigned StackDivide; ///< Amount to adjust stack size by.
462 /// \brief Size of a "push" instruction for the given register.
463 unsigned PushInstrSize(unsigned Reg) const {
483 /// \brief Implementation of algorithm to generate the compact unwind encoding
484 /// for the CFI instructions.
486 generateCompactUnwindEncodingImpl(ArrayRef<MCCFIInstruction> Instrs) const {
487 if (Instrs.empty()) return 0;
489 // Reset the saved registers.
490 unsigned SavedRegIdx = 0;
491 memset(SavedRegs, 0, sizeof(SavedRegs));
495 // Encode that we are using EBP/RBP as the frame pointer.
496 uint32_t CompactUnwindEncoding = 0;
498 unsigned SubtractInstrIdx = Is64Bit ? 3 : 2;
499 unsigned InstrOffset = 0;
500 unsigned StackAdjust = 0;
501 unsigned StackSize = 0;
502 unsigned PrevStackSize = 0;
503 unsigned NumDefCFAOffsets = 0;
505 for (unsigned i = 0, e = Instrs.size(); i != e; ++i) {
506 const MCCFIInstruction &Inst = Instrs[i];
508 switch (Inst.getOperation()) {
510 // Any other CFI directives indicate a frame that we aren't prepared
511 // to represent via compact unwind, so just bail out.
513 case MCCFIInstruction::OpDefCfaRegister: {
514 // Defines a frame pointer. E.g.
518 // .cfi_def_cfa_register %rbp
521 assert(MRI.getLLVMRegNum(Inst.getRegister(), true) ==
522 (Is64Bit ? X86::RBP : X86::EBP) && "Invalid frame pointer!");
525 memset(SavedRegs, 0, sizeof(SavedRegs));
528 InstrOffset += MoveInstrSize;
531 case MCCFIInstruction::OpDefCfaOffset: {
532 // Defines a new offset for the CFA. E.g.
538 // .cfi_def_cfa_offset 16
544 // .cfi_def_cfa_offset 80
546 PrevStackSize = StackSize;
547 StackSize = std::abs(Inst.getOffset()) / StackDivide;
551 case MCCFIInstruction::OpOffset: {
552 // Defines a "push" of a callee-saved register. E.g.
560 // .cfi_offset %rbx, -40
561 // .cfi_offset %r14, -32
562 // .cfi_offset %r15, -24
564 if (SavedRegIdx == CU_NUM_SAVED_REGS)
565 // If there are too many saved registers, we cannot use a compact
567 return CU::UNWIND_MODE_DWARF;
569 unsigned Reg = MRI.getLLVMRegNum(Inst.getRegister(), true);
570 SavedRegs[SavedRegIdx++] = Reg;
571 StackAdjust += OffsetSize;
572 InstrOffset += PushInstrSize(Reg);
578 StackAdjust /= StackDivide;
581 if ((StackAdjust & 0xFF) != StackAdjust)
582 // Offset was too big for a compact unwind encoding.
583 return CU::UNWIND_MODE_DWARF;
585 // Get the encoding of the saved registers when we have a frame pointer.
586 uint32_t RegEnc = encodeCompactUnwindRegistersWithFrame();
587 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF;
589 CompactUnwindEncoding |= CU::UNWIND_MODE_BP_FRAME;
590 CompactUnwindEncoding |= (StackAdjust & 0xFF) << 16;
591 CompactUnwindEncoding |= RegEnc & CU::UNWIND_BP_FRAME_REGISTERS;
593 // If the amount of the stack allocation is the size of a register, then
594 // we "push" the RAX/EAX register onto the stack instead of adjusting the
595 // stack pointer with a SUB instruction. We don't support the push of the
596 // RAX/EAX register with compact unwind. So we check for that situation
598 if ((NumDefCFAOffsets == SavedRegIdx + 1 &&
599 StackSize - PrevStackSize == 1) ||
600 (Instrs.size() == 1 && NumDefCFAOffsets == 1 && StackSize == 2))
601 return CU::UNWIND_MODE_DWARF;
603 SubtractInstrIdx += InstrOffset;
606 if ((StackSize & 0xFF) == StackSize) {
607 // Frameless stack with a small stack size.
608 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IMMD;
610 // Encode the stack size.
611 CompactUnwindEncoding |= (StackSize & 0xFF) << 16;
613 if ((StackAdjust & 0x7) != StackAdjust)
614 // The extra stack adjustments are too big for us to handle.
615 return CU::UNWIND_MODE_DWARF;
617 // Frameless stack with an offset too large for us to encode compactly.
618 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IND;
620 // Encode the offset to the nnnnnn value in the 'subl $nnnnnn, ESP'
622 CompactUnwindEncoding |= (SubtractInstrIdx & 0xFF) << 16;
624 // Encode any extra stack stack adjustments (done via push
626 CompactUnwindEncoding |= (StackAdjust & 0x7) << 13;
629 // Encode the number of registers saved. (Reverse the list first.)
630 std::reverse(&SavedRegs[0], &SavedRegs[SavedRegIdx]);
631 CompactUnwindEncoding |= (SavedRegIdx & 0x7) << 10;
633 // Get the encoding of the saved registers when we don't have a frame
635 uint32_t RegEnc = encodeCompactUnwindRegistersWithoutFrame(SavedRegIdx);
636 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF;
638 // Encode the register encoding.
639 CompactUnwindEncoding |=
640 RegEnc & CU::UNWIND_FRAMELESS_STACK_REG_PERMUTATION;
643 return CompactUnwindEncoding;
647 /// \brief Get the compact unwind number for a given register. The number
648 /// corresponds to the enum lists in compact_unwind_encoding.h.
649 int getCompactUnwindRegNum(unsigned Reg) const {
650 static const MCPhysReg CU32BitRegs[7] = {
651 X86::EBX, X86::ECX, X86::EDX, X86::EDI, X86::ESI, X86::EBP, 0
653 static const MCPhysReg CU64BitRegs[] = {
654 X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0
656 const MCPhysReg *CURegs = Is64Bit ? CU64BitRegs : CU32BitRegs;
657 for (int Idx = 1; *CURegs; ++CURegs, ++Idx)
664 /// \brief Return the registers encoded for a compact encoding with a frame
666 uint32_t encodeCompactUnwindRegistersWithFrame() const {
667 // Encode the registers in the order they were saved --- 3-bits per
668 // register. The list of saved registers is assumed to be in reverse
669 // order. The registers are numbered from 1 to CU_NUM_SAVED_REGS.
671 for (int i = 0, Idx = 0; i != CU_NUM_SAVED_REGS; ++i) {
672 unsigned Reg = SavedRegs[i];
675 int CURegNum = getCompactUnwindRegNum(Reg);
676 if (CURegNum == -1) return ~0U;
678 // Encode the 3-bit register number in order, skipping over 3-bits for
680 RegEnc |= (CURegNum & 0x7) << (Idx++ * 3);
683 assert((RegEnc & 0x3FFFF) == RegEnc &&
684 "Invalid compact register encoding!");
688 /// \brief Create the permutation encoding used with frameless stacks. It is
689 /// passed the number of registers to be saved and an array of the registers
691 uint32_t encodeCompactUnwindRegistersWithoutFrame(unsigned RegCount) const {
692 // The saved registers are numbered from 1 to 6. In order to encode the
693 // order in which they were saved, we re-number them according to their
694 // place in the register order. The re-numbering is relative to the last
695 // re-numbered register. E.g., if we have registers {6, 2, 4, 5} saved in
705 for (unsigned i = 0; i < RegCount; ++i) {
706 int CUReg = getCompactUnwindRegNum(SavedRegs[i]);
707 if (CUReg == -1) return ~0U;
708 SavedRegs[i] = CUReg;
712 std::reverse(&SavedRegs[0], &SavedRegs[CU_NUM_SAVED_REGS]);
714 uint32_t RenumRegs[CU_NUM_SAVED_REGS];
715 for (unsigned i = CU_NUM_SAVED_REGS - RegCount; i < CU_NUM_SAVED_REGS; ++i){
716 unsigned Countless = 0;
717 for (unsigned j = CU_NUM_SAVED_REGS - RegCount; j < i; ++j)
718 if (SavedRegs[j] < SavedRegs[i])
721 RenumRegs[i] = SavedRegs[i] - Countless - 1;
724 // Take the renumbered values and encode them into a 10-bit number.
725 uint32_t permutationEncoding = 0;
728 permutationEncoding |= 120 * RenumRegs[0] + 24 * RenumRegs[1]
729 + 6 * RenumRegs[2] + 2 * RenumRegs[3]
733 permutationEncoding |= 120 * RenumRegs[1] + 24 * RenumRegs[2]
734 + 6 * RenumRegs[3] + 2 * RenumRegs[4]
738 permutationEncoding |= 60 * RenumRegs[2] + 12 * RenumRegs[3]
739 + 3 * RenumRegs[4] + RenumRegs[5];
742 permutationEncoding |= 20 * RenumRegs[3] + 4 * RenumRegs[4]
746 permutationEncoding |= 5 * RenumRegs[4] + RenumRegs[5];
749 permutationEncoding |= RenumRegs[5];
753 assert((permutationEncoding & 0x3FF) == permutationEncoding &&
754 "Invalid compact register encoding!");
755 return permutationEncoding;
759 DarwinX86AsmBackend(const Target &T, const MCRegisterInfo &MRI, StringRef CPU,
761 : X86AsmBackend(T, CPU), MRI(MRI), Is64Bit(Is64Bit) {
762 memset(SavedRegs, 0, sizeof(SavedRegs));
763 OffsetSize = Is64Bit ? 8 : 4;
764 MoveInstrSize = Is64Bit ? 3 : 2;
765 StackDivide = Is64Bit ? 8 : 4;
769 class DarwinX86_32AsmBackend : public DarwinX86AsmBackend {
771 DarwinX86_32AsmBackend(const Target &T, const MCRegisterInfo &MRI,
773 : DarwinX86AsmBackend(T, MRI, CPU, false) {}
775 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
776 return createX86MachObjectWriter(OS, /*Is64Bit=*/false,
777 MachO::CPU_TYPE_I386,
778 MachO::CPU_SUBTYPE_I386_ALL);
781 /// \brief Generate the compact unwind encoding for the CFI instructions.
782 uint32_t generateCompactUnwindEncoding(
783 ArrayRef<MCCFIInstruction> Instrs) const override {
784 return generateCompactUnwindEncodingImpl(Instrs);
788 class DarwinX86_64AsmBackend : public DarwinX86AsmBackend {
789 const MachO::CPUSubTypeX86 Subtype;
791 DarwinX86_64AsmBackend(const Target &T, const MCRegisterInfo &MRI,
792 StringRef CPU, MachO::CPUSubTypeX86 st)
793 : DarwinX86AsmBackend(T, MRI, CPU, true), Subtype(st) {}
795 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
796 return createX86MachObjectWriter(OS, /*Is64Bit=*/true,
797 MachO::CPU_TYPE_X86_64, Subtype);
800 /// \brief Generate the compact unwind encoding for the CFI instructions.
801 uint32_t generateCompactUnwindEncoding(
802 ArrayRef<MCCFIInstruction> Instrs) const override {
803 return generateCompactUnwindEncodingImpl(Instrs);
807 } // end anonymous namespace
809 MCAsmBackend *llvm::createX86_32AsmBackend(const Target &T,
810 const MCRegisterInfo &MRI,
811 const Triple &TheTriple,
813 if (TheTriple.isOSBinFormatMachO())
814 return new DarwinX86_32AsmBackend(T, MRI, CPU);
816 if (TheTriple.isOSWindows() && !TheTriple.isOSBinFormatELF())
817 return new WindowsX86AsmBackend(T, false, CPU);
819 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
821 if (TheTriple.isOSIAMCU())
822 return new ELFX86_IAMCUAsmBackend(T, OSABI, CPU);
824 return new ELFX86_32AsmBackend(T, OSABI, CPU);
827 MCAsmBackend *llvm::createX86_64AsmBackend(const Target &T,
828 const MCRegisterInfo &MRI,
829 const Triple &TheTriple,
831 if (TheTriple.isOSBinFormatMachO()) {
832 MachO::CPUSubTypeX86 CS =
833 StringSwitch<MachO::CPUSubTypeX86>(TheTriple.getArchName())
834 .Case("x86_64h", MachO::CPU_SUBTYPE_X86_64_H)
835 .Default(MachO::CPU_SUBTYPE_X86_64_ALL);
836 return new DarwinX86_64AsmBackend(T, MRI, CPU, CS);
839 if (TheTriple.isOSWindows() && !TheTriple.isOSBinFormatELF())
840 return new WindowsX86AsmBackend(T, true, CPU);
842 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
844 if (TheTriple.getEnvironment() == Triple::GNUX32)
845 return new ELFX86_X32AsmBackend(T, OSABI, CPU);
846 return new ELFX86_64AsmBackend(T, OSABI, CPU);