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 const uint64_t MaxNopLength;
74 X86AsmBackend(const Target &T, StringRef CPU)
75 : MCAsmBackend(), CPU(CPU), MaxNopLength(CPU == "slm" ? 7 : 15) {
76 HasNopl = CPU != "generic" && CPU != "i386" && CPU != "i486" &&
77 CPU != "i586" && CPU != "pentium" && CPU != "pentium-mmx" &&
78 CPU != "i686" && CPU != "k6" && CPU != "k6-2" && CPU != "k6-3" &&
79 CPU != "geode" && CPU != "winchip-c6" && CPU != "winchip2" &&
80 CPU != "c3" && CPU != "c3-2";
83 unsigned getNumFixupKinds() const override {
84 return X86::NumTargetFixupKinds;
87 const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override {
88 const static MCFixupKindInfo Infos[X86::NumTargetFixupKinds] = {
89 { "reloc_riprel_4byte", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel },
90 { "reloc_riprel_4byte_movq_load", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel},
91 { "reloc_signed_4byte", 0, 4 * 8, 0},
92 { "reloc_global_offset_table", 0, 4 * 8, 0}
95 if (Kind < FirstTargetFixupKind)
96 return MCAsmBackend::getFixupKindInfo(Kind);
98 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
100 return Infos[Kind - FirstTargetFixupKind];
103 void applyFixup(const MCFixup &Fixup, char *Data, unsigned DataSize,
104 uint64_t Value, bool IsPCRel) const override {
105 unsigned Size = 1 << getFixupKindLog2Size(Fixup.getKind());
107 assert(Fixup.getOffset() + Size <= DataSize &&
108 "Invalid fixup offset!");
110 // Check that uppper bits are either all zeros or all ones.
111 // Specifically ignore overflow/underflow as long as the leakage is
112 // limited to the lower bits. This is to remain compatible with
114 assert(isIntN(Size * 8 + 1, Value) &&
115 "Value does not fit in the Fixup field");
117 for (unsigned i = 0; i != Size; ++i)
118 Data[Fixup.getOffset() + i] = uint8_t(Value >> (i * 8));
121 bool mayNeedRelaxation(const MCInst &Inst) const override;
123 bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
124 const MCRelaxableFragment *DF,
125 const MCAsmLayout &Layout) const override;
127 void relaxInstruction(const MCInst &Inst, MCInst &Res) const override;
129 bool writeNopData(uint64_t Count, MCObjectWriter *OW) const override;
131 } // end anonymous namespace
133 static unsigned getRelaxedOpcodeBranch(unsigned Op) {
138 case X86::JAE_1: return X86::JAE_4;
139 case X86::JA_1: return X86::JA_4;
140 case X86::JBE_1: return X86::JBE_4;
141 case X86::JB_1: return X86::JB_4;
142 case X86::JE_1: return X86::JE_4;
143 case X86::JGE_1: return X86::JGE_4;
144 case X86::JG_1: return X86::JG_4;
145 case X86::JLE_1: return X86::JLE_4;
146 case X86::JL_1: return X86::JL_4;
147 case X86::JMP_1: return X86::JMP_4;
148 case X86::JNE_1: return X86::JNE_4;
149 case X86::JNO_1: return X86::JNO_4;
150 case X86::JNP_1: return X86::JNP_4;
151 case X86::JNS_1: return X86::JNS_4;
152 case X86::JO_1: return X86::JO_4;
153 case X86::JP_1: return X86::JP_4;
154 case X86::JS_1: return X86::JS_4;
158 static unsigned getRelaxedOpcodeArith(unsigned Op) {
164 case X86::IMUL16rri8: return X86::IMUL16rri;
165 case X86::IMUL16rmi8: return X86::IMUL16rmi;
166 case X86::IMUL32rri8: return X86::IMUL32rri;
167 case X86::IMUL32rmi8: return X86::IMUL32rmi;
168 case X86::IMUL64rri8: return X86::IMUL64rri32;
169 case X86::IMUL64rmi8: return X86::IMUL64rmi32;
172 case X86::AND16ri8: return X86::AND16ri;
173 case X86::AND16mi8: return X86::AND16mi;
174 case X86::AND32ri8: return X86::AND32ri;
175 case X86::AND32mi8: return X86::AND32mi;
176 case X86::AND64ri8: return X86::AND64ri32;
177 case X86::AND64mi8: return X86::AND64mi32;
180 case X86::OR16ri8: return X86::OR16ri;
181 case X86::OR16mi8: return X86::OR16mi;
182 case X86::OR32ri8: return X86::OR32ri;
183 case X86::OR32mi8: return X86::OR32mi;
184 case X86::OR64ri8: return X86::OR64ri32;
185 case X86::OR64mi8: return X86::OR64mi32;
188 case X86::XOR16ri8: return X86::XOR16ri;
189 case X86::XOR16mi8: return X86::XOR16mi;
190 case X86::XOR32ri8: return X86::XOR32ri;
191 case X86::XOR32mi8: return X86::XOR32mi;
192 case X86::XOR64ri8: return X86::XOR64ri32;
193 case X86::XOR64mi8: return X86::XOR64mi32;
196 case X86::ADD16ri8: return X86::ADD16ri;
197 case X86::ADD16mi8: return X86::ADD16mi;
198 case X86::ADD32ri8: return X86::ADD32ri;
199 case X86::ADD32mi8: return X86::ADD32mi;
200 case X86::ADD64ri8: return X86::ADD64ri32;
201 case X86::ADD64mi8: return X86::ADD64mi32;
204 case X86::ADC16ri8: return X86::ADC16ri;
205 case X86::ADC16mi8: return X86::ADC16mi;
206 case X86::ADC32ri8: return X86::ADC32ri;
207 case X86::ADC32mi8: return X86::ADC32mi;
208 case X86::ADC64ri8: return X86::ADC64ri32;
209 case X86::ADC64mi8: return X86::ADC64mi32;
212 case X86::SUB16ri8: return X86::SUB16ri;
213 case X86::SUB16mi8: return X86::SUB16mi;
214 case X86::SUB32ri8: return X86::SUB32ri;
215 case X86::SUB32mi8: return X86::SUB32mi;
216 case X86::SUB64ri8: return X86::SUB64ri32;
217 case X86::SUB64mi8: return X86::SUB64mi32;
220 case X86::SBB16ri8: return X86::SBB16ri;
221 case X86::SBB16mi8: return X86::SBB16mi;
222 case X86::SBB32ri8: return X86::SBB32ri;
223 case X86::SBB32mi8: return X86::SBB32mi;
224 case X86::SBB64ri8: return X86::SBB64ri32;
225 case X86::SBB64mi8: return X86::SBB64mi32;
228 case X86::CMP16ri8: return X86::CMP16ri;
229 case X86::CMP16mi8: return X86::CMP16mi;
230 case X86::CMP32ri8: return X86::CMP32ri;
231 case X86::CMP32mi8: return X86::CMP32mi;
232 case X86::CMP64ri8: return X86::CMP64ri32;
233 case X86::CMP64mi8: return X86::CMP64mi32;
236 case X86::PUSH32i8: return X86::PUSHi32;
237 case X86::PUSH16i8: return X86::PUSHi16;
238 case X86::PUSH64i8: return X86::PUSH64i32;
242 static unsigned getRelaxedOpcode(unsigned Op) {
243 unsigned R = getRelaxedOpcodeArith(Op);
246 return getRelaxedOpcodeBranch(Op);
249 bool X86AsmBackend::mayNeedRelaxation(const MCInst &Inst) const {
250 // Branches can always be relaxed.
251 if (getRelaxedOpcodeBranch(Inst.getOpcode()) != Inst.getOpcode())
254 // Check if this instruction is ever relaxable.
255 if (getRelaxedOpcodeArith(Inst.getOpcode()) == Inst.getOpcode())
259 // Check if the relaxable operand has an expression. For the current set of
260 // relaxable instructions, the relaxable operand is always the last operand.
261 unsigned RelaxableOp = Inst.getNumOperands() - 1;
262 if (Inst.getOperand(RelaxableOp).isExpr())
268 bool X86AsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup,
270 const MCRelaxableFragment *DF,
271 const MCAsmLayout &Layout) const {
272 // Relax if the value is too big for a (signed) i8.
273 return int64_t(Value) != int64_t(int8_t(Value));
276 // FIXME: Can tblgen help at all here to verify there aren't other instructions
278 void X86AsmBackend::relaxInstruction(const MCInst &Inst, MCInst &Res) const {
279 // The only relaxations X86 does is from a 1byte pcrel to a 4byte pcrel.
280 unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode());
282 if (RelaxedOp == Inst.getOpcode()) {
283 SmallString<256> Tmp;
284 raw_svector_ostream OS(Tmp);
285 Inst.dump_pretty(OS);
287 report_fatal_error("unexpected instruction to relax: " + OS.str());
291 Res.setOpcode(RelaxedOp);
294 /// \brief Write a sequence of optimal nops to the output, covering \p Count
296 /// \return - true on success, false on failure
297 bool X86AsmBackend::writeNopData(uint64_t Count, MCObjectWriter *OW) const {
298 static const uint8_t Nops[10][10] = {
306 {0x0f, 0x1f, 0x40, 0x00},
307 // nopl 0(%[re]ax,%[re]ax,1)
308 {0x0f, 0x1f, 0x44, 0x00, 0x00},
309 // nopw 0(%[re]ax,%[re]ax,1)
310 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
312 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
313 // nopl 0L(%[re]ax,%[re]ax,1)
314 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
315 // nopw 0L(%[re]ax,%[re]ax,1)
316 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
317 // nopw %cs:0L(%[re]ax,%[re]ax,1)
318 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
321 // This CPU doesn't support long nops. If needed add more.
322 // FIXME: Can we get this from the subtarget somehow?
323 // FIXME: We could generated something better than plain 0x90.
325 for (uint64_t i = 0; i < Count; ++i)
330 // 15 is the longest single nop instruction. Emit as many 15-byte nops as
331 // needed, then emit a nop of the remaining length.
333 const uint8_t ThisNopLength = (uint8_t) std::min(Count, MaxNopLength);
334 const uint8_t Prefixes = ThisNopLength <= 10 ? 0 : ThisNopLength - 10;
335 for (uint8_t i = 0; i < Prefixes; i++)
337 const uint8_t Rest = ThisNopLength - Prefixes;
338 for (uint8_t i = 0; i < Rest; i++)
339 OW->write8(Nops[Rest - 1][i]);
340 Count -= ThisNopLength;
341 } while (Count != 0);
350 class ELFX86AsmBackend : public X86AsmBackend {
353 ELFX86AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
354 : X86AsmBackend(T, CPU), OSABI(OSABI) {}
357 class ELFX86_32AsmBackend : public ELFX86AsmBackend {
359 ELFX86_32AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
360 : ELFX86AsmBackend(T, OSABI, CPU) {}
362 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
363 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI, ELF::EM_386);
367 class ELFX86_X32AsmBackend : public ELFX86AsmBackend {
369 ELFX86_X32AsmBackend(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,
378 class ELFX86_IAMCUAsmBackend : public ELFX86AsmBackend {
380 ELFX86_IAMCUAsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
381 : ELFX86AsmBackend(T, OSABI, CPU) {}
383 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
384 return createX86ELFObjectWriter(OS, /*IsELF64*/ false, OSABI,
389 class ELFX86_64AsmBackend : public ELFX86AsmBackend {
391 ELFX86_64AsmBackend(const Target &T, uint8_t OSABI, StringRef CPU)
392 : ELFX86AsmBackend(T, OSABI, CPU) {}
394 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
395 return createX86ELFObjectWriter(OS, /*IsELF64*/ true, OSABI, ELF::EM_X86_64);
399 class WindowsX86AsmBackend : public X86AsmBackend {
403 WindowsX86AsmBackend(const Target &T, bool is64Bit, StringRef CPU)
404 : X86AsmBackend(T, CPU)
408 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
409 return createX86WinCOFFObjectWriter(OS, Is64Bit);
415 /// Compact unwind encoding values.
416 enum CompactUnwindEncodings {
417 /// [RE]BP based frame where [RE]BP is pused on the stack immediately after
418 /// the return address, then [RE]SP is moved to [RE]BP.
419 UNWIND_MODE_BP_FRAME = 0x01000000,
421 /// A frameless function with a small constant stack size.
422 UNWIND_MODE_STACK_IMMD = 0x02000000,
424 /// A frameless function with a large constant stack size.
425 UNWIND_MODE_STACK_IND = 0x03000000,
427 /// No compact unwind encoding is available.
428 UNWIND_MODE_DWARF = 0x04000000,
430 /// Mask for encoding the frame registers.
431 UNWIND_BP_FRAME_REGISTERS = 0x00007FFF,
433 /// Mask for encoding the frameless registers.
434 UNWIND_FRAMELESS_STACK_REG_PERMUTATION = 0x000003FF
437 } // end CU namespace
439 class DarwinX86AsmBackend : public X86AsmBackend {
440 const MCRegisterInfo &MRI;
442 /// \brief Number of registers that can be saved in a compact unwind encoding.
443 enum { CU_NUM_SAVED_REGS = 6 };
445 mutable unsigned SavedRegs[CU_NUM_SAVED_REGS];
448 unsigned OffsetSize; ///< Offset of a "push" instruction.
449 unsigned MoveInstrSize; ///< Size of a "move" instruction.
450 unsigned StackDivide; ///< Amount to adjust stack size by.
452 /// \brief Size of a "push" instruction for the given register.
453 unsigned PushInstrSize(unsigned Reg) const {
473 /// \brief Implementation of algorithm to generate the compact unwind encoding
474 /// for the CFI instructions.
476 generateCompactUnwindEncodingImpl(ArrayRef<MCCFIInstruction> Instrs) const {
477 if (Instrs.empty()) return 0;
479 // Reset the saved registers.
480 unsigned SavedRegIdx = 0;
481 memset(SavedRegs, 0, sizeof(SavedRegs));
485 // Encode that we are using EBP/RBP as the frame pointer.
486 uint32_t CompactUnwindEncoding = 0;
488 unsigned SubtractInstrIdx = Is64Bit ? 3 : 2;
489 unsigned InstrOffset = 0;
490 unsigned StackAdjust = 0;
491 unsigned StackSize = 0;
492 unsigned PrevStackSize = 0;
493 unsigned NumDefCFAOffsets = 0;
495 for (unsigned i = 0, e = Instrs.size(); i != e; ++i) {
496 const MCCFIInstruction &Inst = Instrs[i];
498 switch (Inst.getOperation()) {
500 // Any other CFI directives indicate a frame that we aren't prepared
501 // to represent via compact unwind, so just bail out.
503 case MCCFIInstruction::OpDefCfaRegister: {
504 // Defines a frame pointer. E.g.
508 // .cfi_def_cfa_register %rbp
511 assert(MRI.getLLVMRegNum(Inst.getRegister(), true) ==
512 (Is64Bit ? X86::RBP : X86::EBP) && "Invalid frame pointer!");
515 memset(SavedRegs, 0, sizeof(SavedRegs));
518 InstrOffset += MoveInstrSize;
521 case MCCFIInstruction::OpDefCfaOffset: {
522 // Defines a new offset for the CFA. E.g.
528 // .cfi_def_cfa_offset 16
534 // .cfi_def_cfa_offset 80
536 PrevStackSize = StackSize;
537 StackSize = std::abs(Inst.getOffset()) / StackDivide;
541 case MCCFIInstruction::OpOffset: {
542 // Defines a "push" of a callee-saved register. E.g.
550 // .cfi_offset %rbx, -40
551 // .cfi_offset %r14, -32
552 // .cfi_offset %r15, -24
554 if (SavedRegIdx == CU_NUM_SAVED_REGS)
555 // If there are too many saved registers, we cannot use a compact
557 return CU::UNWIND_MODE_DWARF;
559 unsigned Reg = MRI.getLLVMRegNum(Inst.getRegister(), true);
560 SavedRegs[SavedRegIdx++] = Reg;
561 StackAdjust += OffsetSize;
562 InstrOffset += PushInstrSize(Reg);
568 StackAdjust /= StackDivide;
571 if ((StackAdjust & 0xFF) != StackAdjust)
572 // Offset was too big for a compact unwind encoding.
573 return CU::UNWIND_MODE_DWARF;
575 // Get the encoding of the saved registers when we have a frame pointer.
576 uint32_t RegEnc = encodeCompactUnwindRegistersWithFrame();
577 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF;
579 CompactUnwindEncoding |= CU::UNWIND_MODE_BP_FRAME;
580 CompactUnwindEncoding |= (StackAdjust & 0xFF) << 16;
581 CompactUnwindEncoding |= RegEnc & CU::UNWIND_BP_FRAME_REGISTERS;
583 // If the amount of the stack allocation is the size of a register, then
584 // we "push" the RAX/EAX register onto the stack instead of adjusting the
585 // stack pointer with a SUB instruction. We don't support the push of the
586 // RAX/EAX register with compact unwind. So we check for that situation
588 if ((NumDefCFAOffsets == SavedRegIdx + 1 &&
589 StackSize - PrevStackSize == 1) ||
590 (Instrs.size() == 1 && NumDefCFAOffsets == 1 && StackSize == 2))
591 return CU::UNWIND_MODE_DWARF;
593 SubtractInstrIdx += InstrOffset;
596 if ((StackSize & 0xFF) == StackSize) {
597 // Frameless stack with a small stack size.
598 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IMMD;
600 // Encode the stack size.
601 CompactUnwindEncoding |= (StackSize & 0xFF) << 16;
603 if ((StackAdjust & 0x7) != StackAdjust)
604 // The extra stack adjustments are too big for us to handle.
605 return CU::UNWIND_MODE_DWARF;
607 // Frameless stack with an offset too large for us to encode compactly.
608 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IND;
610 // Encode the offset to the nnnnnn value in the 'subl $nnnnnn, ESP'
612 CompactUnwindEncoding |= (SubtractInstrIdx & 0xFF) << 16;
614 // Encode any extra stack stack adjustments (done via push
616 CompactUnwindEncoding |= (StackAdjust & 0x7) << 13;
619 // Encode the number of registers saved. (Reverse the list first.)
620 std::reverse(&SavedRegs[0], &SavedRegs[SavedRegIdx]);
621 CompactUnwindEncoding |= (SavedRegIdx & 0x7) << 10;
623 // Get the encoding of the saved registers when we don't have a frame
625 uint32_t RegEnc = encodeCompactUnwindRegistersWithoutFrame(SavedRegIdx);
626 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF;
628 // Encode the register encoding.
629 CompactUnwindEncoding |=
630 RegEnc & CU::UNWIND_FRAMELESS_STACK_REG_PERMUTATION;
633 return CompactUnwindEncoding;
637 /// \brief Get the compact unwind number for a given register. The number
638 /// corresponds to the enum lists in compact_unwind_encoding.h.
639 int getCompactUnwindRegNum(unsigned Reg) const {
640 static const MCPhysReg CU32BitRegs[7] = {
641 X86::EBX, X86::ECX, X86::EDX, X86::EDI, X86::ESI, X86::EBP, 0
643 static const MCPhysReg CU64BitRegs[] = {
644 X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0
646 const MCPhysReg *CURegs = Is64Bit ? CU64BitRegs : CU32BitRegs;
647 for (int Idx = 1; *CURegs; ++CURegs, ++Idx)
654 /// \brief Return the registers encoded for a compact encoding with a frame
656 uint32_t encodeCompactUnwindRegistersWithFrame() const {
657 // Encode the registers in the order they were saved --- 3-bits per
658 // register. The list of saved registers is assumed to be in reverse
659 // order. The registers are numbered from 1 to CU_NUM_SAVED_REGS.
661 for (int i = 0, Idx = 0; i != CU_NUM_SAVED_REGS; ++i) {
662 unsigned Reg = SavedRegs[i];
665 int CURegNum = getCompactUnwindRegNum(Reg);
666 if (CURegNum == -1) return ~0U;
668 // Encode the 3-bit register number in order, skipping over 3-bits for
670 RegEnc |= (CURegNum & 0x7) << (Idx++ * 3);
673 assert((RegEnc & 0x3FFFF) == RegEnc &&
674 "Invalid compact register encoding!");
678 /// \brief Create the permutation encoding used with frameless stacks. It is
679 /// passed the number of registers to be saved and an array of the registers
681 uint32_t encodeCompactUnwindRegistersWithoutFrame(unsigned RegCount) const {
682 // The saved registers are numbered from 1 to 6. In order to encode the
683 // order in which they were saved, we re-number them according to their
684 // place in the register order. The re-numbering is relative to the last
685 // re-numbered register. E.g., if we have registers {6, 2, 4, 5} saved in
695 for (unsigned i = 0; i < RegCount; ++i) {
696 int CUReg = getCompactUnwindRegNum(SavedRegs[i]);
697 if (CUReg == -1) return ~0U;
698 SavedRegs[i] = CUReg;
702 std::reverse(&SavedRegs[0], &SavedRegs[CU_NUM_SAVED_REGS]);
704 uint32_t RenumRegs[CU_NUM_SAVED_REGS];
705 for (unsigned i = CU_NUM_SAVED_REGS - RegCount; i < CU_NUM_SAVED_REGS; ++i){
706 unsigned Countless = 0;
707 for (unsigned j = CU_NUM_SAVED_REGS - RegCount; j < i; ++j)
708 if (SavedRegs[j] < SavedRegs[i])
711 RenumRegs[i] = SavedRegs[i] - Countless - 1;
714 // Take the renumbered values and encode them into a 10-bit number.
715 uint32_t permutationEncoding = 0;
718 permutationEncoding |= 120 * RenumRegs[0] + 24 * RenumRegs[1]
719 + 6 * RenumRegs[2] + 2 * RenumRegs[3]
723 permutationEncoding |= 120 * RenumRegs[1] + 24 * RenumRegs[2]
724 + 6 * RenumRegs[3] + 2 * RenumRegs[4]
728 permutationEncoding |= 60 * RenumRegs[2] + 12 * RenumRegs[3]
729 + 3 * RenumRegs[4] + RenumRegs[5];
732 permutationEncoding |= 20 * RenumRegs[3] + 4 * RenumRegs[4]
736 permutationEncoding |= 5 * RenumRegs[4] + RenumRegs[5];
739 permutationEncoding |= RenumRegs[5];
743 assert((permutationEncoding & 0x3FF) == permutationEncoding &&
744 "Invalid compact register encoding!");
745 return permutationEncoding;
749 DarwinX86AsmBackend(const Target &T, const MCRegisterInfo &MRI, StringRef CPU,
751 : X86AsmBackend(T, CPU), MRI(MRI), Is64Bit(Is64Bit) {
752 memset(SavedRegs, 0, sizeof(SavedRegs));
753 OffsetSize = Is64Bit ? 8 : 4;
754 MoveInstrSize = Is64Bit ? 3 : 2;
755 StackDivide = Is64Bit ? 8 : 4;
759 class DarwinX86_32AsmBackend : public DarwinX86AsmBackend {
761 DarwinX86_32AsmBackend(const Target &T, const MCRegisterInfo &MRI,
763 : DarwinX86AsmBackend(T, MRI, CPU, false) {}
765 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
766 return createX86MachObjectWriter(OS, /*Is64Bit=*/false,
767 MachO::CPU_TYPE_I386,
768 MachO::CPU_SUBTYPE_I386_ALL);
771 /// \brief Generate the compact unwind encoding for the CFI instructions.
772 uint32_t generateCompactUnwindEncoding(
773 ArrayRef<MCCFIInstruction> Instrs) const override {
774 return generateCompactUnwindEncodingImpl(Instrs);
778 class DarwinX86_64AsmBackend : public DarwinX86AsmBackend {
779 const MachO::CPUSubTypeX86 Subtype;
781 DarwinX86_64AsmBackend(const Target &T, const MCRegisterInfo &MRI,
782 StringRef CPU, MachO::CPUSubTypeX86 st)
783 : DarwinX86AsmBackend(T, MRI, CPU, true), Subtype(st) {}
785 MCObjectWriter *createObjectWriter(raw_pwrite_stream &OS) const override {
786 return createX86MachObjectWriter(OS, /*Is64Bit=*/true,
787 MachO::CPU_TYPE_X86_64, Subtype);
790 /// \brief Generate the compact unwind encoding for the CFI instructions.
791 uint32_t generateCompactUnwindEncoding(
792 ArrayRef<MCCFIInstruction> Instrs) const override {
793 return generateCompactUnwindEncodingImpl(Instrs);
797 } // end anonymous namespace
799 MCAsmBackend *llvm::createX86_32AsmBackend(const Target &T,
800 const MCRegisterInfo &MRI,
801 const Triple &TheTriple,
803 if (TheTriple.isOSBinFormatMachO())
804 return new DarwinX86_32AsmBackend(T, MRI, CPU);
806 if (TheTriple.isOSWindows() && !TheTriple.isOSBinFormatELF())
807 return new WindowsX86AsmBackend(T, false, CPU);
809 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
811 if (TheTriple.isOSIAMCU())
812 return new ELFX86_IAMCUAsmBackend(T, OSABI, CPU);
814 return new ELFX86_32AsmBackend(T, OSABI, CPU);
817 MCAsmBackend *llvm::createX86_64AsmBackend(const Target &T,
818 const MCRegisterInfo &MRI,
819 const Triple &TheTriple,
821 if (TheTriple.isOSBinFormatMachO()) {
822 MachO::CPUSubTypeX86 CS =
823 StringSwitch<MachO::CPUSubTypeX86>(TheTriple.getArchName())
824 .Case("x86_64h", MachO::CPU_SUBTYPE_X86_64_H)
825 .Default(MachO::CPU_SUBTYPE_X86_64_ALL);
826 return new DarwinX86_64AsmBackend(T, MRI, CPU, CS);
829 if (TheTriple.isOSWindows() && !TheTriple.isOSBinFormatELF())
830 return new WindowsX86AsmBackend(T, true, CPU);
832 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
834 if (TheTriple.getEnvironment() == Triple::GNUX32)
835 return new ELFX86_X32AsmBackend(T, OSABI, CPU);
836 return new ELFX86_64AsmBackend(T, OSABI, CPU);