xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/MCTargetDesc/AArch64AsmBackend.cpp (revision 1db9f3b21e39176dd5b67cf8ac378633b172463e)
1 //===-- AArch64AsmBackend.cpp - AArch64 Assembler Backend -----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "MCTargetDesc/AArch64FixupKinds.h"
10 #include "MCTargetDesc/AArch64MCExpr.h"
11 #include "MCTargetDesc/AArch64MCTargetDesc.h"
12 #include "Utils/AArch64BaseInfo.h"
13 #include "llvm/BinaryFormat/MachO.h"
14 #include "llvm/MC/MCAsmBackend.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCContext.h"
17 #include "llvm/MC/MCDirectives.h"
18 #include "llvm/MC/MCELFObjectWriter.h"
19 #include "llvm/MC/MCFixupKindInfo.h"
20 #include "llvm/MC/MCObjectWriter.h"
21 #include "llvm/MC/MCRegisterInfo.h"
22 #include "llvm/MC/MCSectionELF.h"
23 #include "llvm/MC/MCSectionMachO.h"
24 #include "llvm/MC/MCSubtargetInfo.h"
25 #include "llvm/MC/MCTargetOptions.h"
26 #include "llvm/MC/MCValue.h"
27 #include "llvm/MC/TargetRegistry.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/TargetParser/Triple.h"
31 using namespace llvm;
32 
33 namespace {
34 
35 class AArch64AsmBackend : public MCAsmBackend {
36   static const unsigned PCRelFlagVal =
37       MCFixupKindInfo::FKF_IsAlignedDownTo32Bits | MCFixupKindInfo::FKF_IsPCRel;
38 protected:
39   Triple TheTriple;
40 
41 public:
42   AArch64AsmBackend(const Target &T, const Triple &TT, bool IsLittleEndian)
43       : MCAsmBackend(IsLittleEndian ? llvm::endianness::little
44                                     : llvm::endianness::big),
45         TheTriple(TT) {}
46 
47   unsigned getNumFixupKinds() const override {
48     return AArch64::NumTargetFixupKinds;
49   }
50 
51   std::optional<MCFixupKind> getFixupKind(StringRef Name) const override;
52 
53   const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override {
54     const static MCFixupKindInfo Infos[AArch64::NumTargetFixupKinds] = {
55         // This table *must* be in the order that the fixup_* kinds are defined
56         // in AArch64FixupKinds.h.
57         //
58         // Name                           Offset (bits) Size (bits)     Flags
59         {"fixup_aarch64_pcrel_adr_imm21", 0, 32, PCRelFlagVal},
60         {"fixup_aarch64_pcrel_adrp_imm21", 0, 32, PCRelFlagVal},
61         {"fixup_aarch64_add_imm12", 10, 12, 0},
62         {"fixup_aarch64_ldst_imm12_scale1", 10, 12, 0},
63         {"fixup_aarch64_ldst_imm12_scale2", 10, 12, 0},
64         {"fixup_aarch64_ldst_imm12_scale4", 10, 12, 0},
65         {"fixup_aarch64_ldst_imm12_scale8", 10, 12, 0},
66         {"fixup_aarch64_ldst_imm12_scale16", 10, 12, 0},
67         {"fixup_aarch64_ldr_pcrel_imm19", 5, 19, PCRelFlagVal},
68         {"fixup_aarch64_movw", 5, 16, 0},
69         {"fixup_aarch64_pcrel_branch14", 5, 14, PCRelFlagVal},
70         {"fixup_aarch64_pcrel_branch16", 5, 16, PCRelFlagVal},
71         {"fixup_aarch64_pcrel_branch19", 5, 19, PCRelFlagVal},
72         {"fixup_aarch64_pcrel_branch26", 0, 26, PCRelFlagVal},
73         {"fixup_aarch64_pcrel_call26", 0, 26, PCRelFlagVal}};
74 
75     // Fixup kinds from .reloc directive are like R_AARCH64_NONE. They do not
76     // require any extra processing.
77     if (Kind >= FirstLiteralRelocationKind)
78       return MCAsmBackend::getFixupKindInfo(FK_NONE);
79 
80     if (Kind < FirstTargetFixupKind)
81       return MCAsmBackend::getFixupKindInfo(Kind);
82 
83     assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
84            "Invalid kind!");
85     return Infos[Kind - FirstTargetFixupKind];
86   }
87 
88   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
89                   const MCValue &Target, MutableArrayRef<char> Data,
90                   uint64_t Value, bool IsResolved,
91                   const MCSubtargetInfo *STI) const override;
92 
93   bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
94                             const MCRelaxableFragment *DF,
95                             const MCAsmLayout &Layout) const override;
96   void relaxInstruction(MCInst &Inst,
97                         const MCSubtargetInfo &STI) const override;
98   bool writeNopData(raw_ostream &OS, uint64_t Count,
99                     const MCSubtargetInfo *STI) const override;
100 
101   unsigned getFixupKindContainereSizeInBytes(unsigned Kind) const;
102 
103   bool shouldForceRelocation(const MCAssembler &Asm, const MCFixup &Fixup,
104                              const MCValue &Target,
105                              const MCSubtargetInfo *STI) override;
106 };
107 
108 } // end anonymous namespace
109 
110 /// The number of bytes the fixup may change.
111 static unsigned getFixupKindNumBytes(unsigned Kind) {
112   switch (Kind) {
113   default:
114     llvm_unreachable("Unknown fixup kind!");
115 
116   case FK_Data_1:
117     return 1;
118 
119   case FK_Data_2:
120   case FK_SecRel_2:
121     return 2;
122 
123   case AArch64::fixup_aarch64_movw:
124   case AArch64::fixup_aarch64_pcrel_branch14:
125   case AArch64::fixup_aarch64_pcrel_branch16:
126   case AArch64::fixup_aarch64_add_imm12:
127   case AArch64::fixup_aarch64_ldst_imm12_scale1:
128   case AArch64::fixup_aarch64_ldst_imm12_scale2:
129   case AArch64::fixup_aarch64_ldst_imm12_scale4:
130   case AArch64::fixup_aarch64_ldst_imm12_scale8:
131   case AArch64::fixup_aarch64_ldst_imm12_scale16:
132   case AArch64::fixup_aarch64_ldr_pcrel_imm19:
133   case AArch64::fixup_aarch64_pcrel_branch19:
134     return 3;
135 
136   case AArch64::fixup_aarch64_pcrel_adr_imm21:
137   case AArch64::fixup_aarch64_pcrel_adrp_imm21:
138   case AArch64::fixup_aarch64_pcrel_branch26:
139   case AArch64::fixup_aarch64_pcrel_call26:
140   case FK_Data_4:
141   case FK_SecRel_4:
142     return 4;
143 
144   case FK_Data_8:
145     return 8;
146   }
147 }
148 
149 static unsigned AdrImmBits(unsigned Value) {
150   unsigned lo2 = Value & 0x3;
151   unsigned hi19 = (Value & 0x1ffffc) >> 2;
152   return (hi19 << 5) | (lo2 << 29);
153 }
154 
155 static uint64_t adjustFixupValue(const MCFixup &Fixup, const MCValue &Target,
156                                  uint64_t Value, MCContext &Ctx,
157                                  const Triple &TheTriple, bool IsResolved) {
158   int64_t SignedValue = static_cast<int64_t>(Value);
159   switch (Fixup.getTargetKind()) {
160   default:
161     llvm_unreachable("Unknown fixup kind!");
162   case AArch64::fixup_aarch64_pcrel_adr_imm21:
163     if (!isInt<21>(SignedValue))
164       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
165     return AdrImmBits(Value & 0x1fffffULL);
166   case AArch64::fixup_aarch64_pcrel_adrp_imm21:
167     assert(!IsResolved);
168     if (TheTriple.isOSBinFormatCOFF()) {
169       if (!isInt<21>(SignedValue))
170         Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
171       return AdrImmBits(Value & 0x1fffffULL);
172     }
173     return AdrImmBits((Value & 0x1fffff000ULL) >> 12);
174   case AArch64::fixup_aarch64_ldr_pcrel_imm19:
175   case AArch64::fixup_aarch64_pcrel_branch19:
176     // Signed 19-bit immediate which gets multiplied by 4
177     if (!isInt<21>(SignedValue))
178       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
179     if (Value & 0x3)
180       Ctx.reportError(Fixup.getLoc(), "fixup not sufficiently aligned");
181     // Low two bits are not encoded.
182     return (Value >> 2) & 0x7ffff;
183   case AArch64::fixup_aarch64_add_imm12:
184   case AArch64::fixup_aarch64_ldst_imm12_scale1:
185     if (TheTriple.isOSBinFormatCOFF() && !IsResolved)
186       Value &= 0xfff;
187     // Unsigned 12-bit immediate
188     if (!isUInt<12>(Value))
189       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
190     return Value;
191   case AArch64::fixup_aarch64_ldst_imm12_scale2:
192     if (TheTriple.isOSBinFormatCOFF() && !IsResolved)
193       Value &= 0xfff;
194     // Unsigned 12-bit immediate which gets multiplied by 2
195     if (!isUInt<13>(Value))
196       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
197     if (Value & 0x1)
198       Ctx.reportError(Fixup.getLoc(), "fixup must be 2-byte aligned");
199     return Value >> 1;
200   case AArch64::fixup_aarch64_ldst_imm12_scale4:
201     if (TheTriple.isOSBinFormatCOFF() && !IsResolved)
202       Value &= 0xfff;
203     // Unsigned 12-bit immediate which gets multiplied by 4
204     if (!isUInt<14>(Value))
205       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
206     if (Value & 0x3)
207       Ctx.reportError(Fixup.getLoc(), "fixup must be 4-byte aligned");
208     return Value >> 2;
209   case AArch64::fixup_aarch64_ldst_imm12_scale8:
210     if (TheTriple.isOSBinFormatCOFF() && !IsResolved)
211       Value &= 0xfff;
212     // Unsigned 12-bit immediate which gets multiplied by 8
213     if (!isUInt<15>(Value))
214       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
215     if (Value & 0x7)
216       Ctx.reportError(Fixup.getLoc(), "fixup must be 8-byte aligned");
217     return Value >> 3;
218   case AArch64::fixup_aarch64_ldst_imm12_scale16:
219     if (TheTriple.isOSBinFormatCOFF() && !IsResolved)
220       Value &= 0xfff;
221     // Unsigned 12-bit immediate which gets multiplied by 16
222     if (!isUInt<16>(Value))
223       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
224     if (Value & 0xf)
225       Ctx.reportError(Fixup.getLoc(), "fixup must be 16-byte aligned");
226     return Value >> 4;
227   case AArch64::fixup_aarch64_movw: {
228     AArch64MCExpr::VariantKind RefKind =
229         static_cast<AArch64MCExpr::VariantKind>(Target.getRefKind());
230     if (AArch64MCExpr::getSymbolLoc(RefKind) != AArch64MCExpr::VK_ABS &&
231         AArch64MCExpr::getSymbolLoc(RefKind) != AArch64MCExpr::VK_SABS) {
232       if (!RefKind) {
233         // The fixup is an expression
234         if (SignedValue > 0xFFFF || SignedValue < -0xFFFF)
235           Ctx.reportError(Fixup.getLoc(),
236                           "fixup value out of range [-0xFFFF, 0xFFFF]");
237 
238         // Invert the negative immediate because it will feed into a MOVN.
239         if (SignedValue < 0)
240           SignedValue = ~SignedValue;
241         Value = static_cast<uint64_t>(SignedValue);
242       } else
243         // VK_GOTTPREL, VK_TPREL, VK_DTPREL are movw fixups, but they can't
244         // ever be resolved in the assembler.
245         Ctx.reportError(Fixup.getLoc(),
246                         "relocation for a thread-local variable points to an "
247                         "absolute symbol");
248       return Value;
249     }
250 
251     if (!IsResolved) {
252       // FIXME: Figure out when this can actually happen, and verify our
253       // behavior.
254       Ctx.reportError(Fixup.getLoc(), "unresolved movw fixup not yet "
255                                       "implemented");
256       return Value;
257     }
258 
259     if (AArch64MCExpr::getSymbolLoc(RefKind) == AArch64MCExpr::VK_SABS) {
260       switch (AArch64MCExpr::getAddressFrag(RefKind)) {
261       case AArch64MCExpr::VK_G0:
262         break;
263       case AArch64MCExpr::VK_G1:
264         SignedValue = SignedValue >> 16;
265         break;
266       case AArch64MCExpr::VK_G2:
267         SignedValue = SignedValue >> 32;
268         break;
269       case AArch64MCExpr::VK_G3:
270         SignedValue = SignedValue >> 48;
271         break;
272       default:
273         llvm_unreachable("Variant kind doesn't correspond to fixup");
274       }
275 
276     } else {
277       switch (AArch64MCExpr::getAddressFrag(RefKind)) {
278       case AArch64MCExpr::VK_G0:
279         break;
280       case AArch64MCExpr::VK_G1:
281         Value = Value >> 16;
282         break;
283       case AArch64MCExpr::VK_G2:
284         Value = Value >> 32;
285         break;
286       case AArch64MCExpr::VK_G3:
287         Value = Value >> 48;
288         break;
289       default:
290         llvm_unreachable("Variant kind doesn't correspond to fixup");
291       }
292     }
293 
294     if (RefKind & AArch64MCExpr::VK_NC) {
295       Value &= 0xFFFF;
296     }
297     else if (AArch64MCExpr::getSymbolLoc(RefKind) == AArch64MCExpr::VK_SABS) {
298       if (SignedValue > 0xFFFF || SignedValue < -0xFFFF)
299         Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
300 
301       // Invert the negative immediate because it will feed into a MOVN.
302       if (SignedValue < 0)
303         SignedValue = ~SignedValue;
304       Value = static_cast<uint64_t>(SignedValue);
305     }
306     else if (Value > 0xFFFF) {
307       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
308     }
309     return Value;
310   }
311   case AArch64::fixup_aarch64_pcrel_branch14:
312     // Signed 16-bit immediate
313     if (!isInt<16>(SignedValue))
314       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
315     // Low two bits are not encoded (4-byte alignment assumed).
316     if (Value & 0x3)
317       Ctx.reportError(Fixup.getLoc(), "fixup not sufficiently aligned");
318     return (Value >> 2) & 0x3fff;
319   case AArch64::fixup_aarch64_pcrel_branch16:
320     // Unsigned PC-relative offset, so invert the negative immediate.
321     SignedValue = -SignedValue;
322     Value = static_cast<uint64_t>(SignedValue);
323     // Check valid 18-bit unsigned range.
324     if (SignedValue < 0 || SignedValue > ((1 << 18) - 1))
325       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
326     // Low two bits are not encoded (4-byte alignment assumed).
327     if (Value & 0b11)
328       Ctx.reportError(Fixup.getLoc(), "fixup not sufficiently aligned");
329     return (Value >> 2) & 0xffff;
330   case AArch64::fixup_aarch64_pcrel_branch26:
331   case AArch64::fixup_aarch64_pcrel_call26:
332     if (TheTriple.isOSBinFormatCOFF() && !IsResolved && SignedValue != 0) {
333       // MSVC link.exe and lld do not support this relocation type
334       // with a non-zero offset
335       Ctx.reportError(Fixup.getLoc(),
336                       "cannot perform a PC-relative fixup with a non-zero "
337                       "symbol offset");
338     }
339     // Signed 28-bit immediate
340     if (!isInt<28>(SignedValue))
341       Ctx.reportError(Fixup.getLoc(), "fixup value out of range");
342     // Low two bits are not encoded (4-byte alignment assumed).
343     if (Value & 0x3)
344       Ctx.reportError(Fixup.getLoc(), "fixup not sufficiently aligned");
345     return (Value >> 2) & 0x3ffffff;
346   case FK_Data_1:
347   case FK_Data_2:
348   case FK_Data_4:
349   case FK_Data_8:
350   case FK_SecRel_2:
351   case FK_SecRel_4:
352     return Value;
353   }
354 }
355 
356 std::optional<MCFixupKind>
357 AArch64AsmBackend::getFixupKind(StringRef Name) const {
358   if (!TheTriple.isOSBinFormatELF())
359     return std::nullopt;
360 
361   unsigned Type = llvm::StringSwitch<unsigned>(Name)
362 #define ELF_RELOC(X, Y)  .Case(#X, Y)
363 #include "llvm/BinaryFormat/ELFRelocs/AArch64.def"
364 #undef ELF_RELOC
365                       .Case("BFD_RELOC_NONE", ELF::R_AARCH64_NONE)
366                       .Case("BFD_RELOC_16", ELF::R_AARCH64_ABS16)
367                       .Case("BFD_RELOC_32", ELF::R_AARCH64_ABS32)
368                       .Case("BFD_RELOC_64", ELF::R_AARCH64_ABS64)
369                       .Default(-1u);
370   if (Type == -1u)
371     return std::nullopt;
372   return static_cast<MCFixupKind>(FirstLiteralRelocationKind + Type);
373 }
374 
375 /// getFixupKindContainereSizeInBytes - The number of bytes of the
376 /// container involved in big endian or 0 if the item is little endian
377 unsigned AArch64AsmBackend::getFixupKindContainereSizeInBytes(unsigned Kind) const {
378   if (Endian == llvm::endianness::little)
379     return 0;
380 
381   switch (Kind) {
382   default:
383     llvm_unreachable("Unknown fixup kind!");
384 
385   case FK_Data_1:
386     return 1;
387   case FK_Data_2:
388     return 2;
389   case FK_Data_4:
390     return 4;
391   case FK_Data_8:
392     return 8;
393 
394   case AArch64::fixup_aarch64_movw:
395   case AArch64::fixup_aarch64_pcrel_branch14:
396   case AArch64::fixup_aarch64_pcrel_branch16:
397   case AArch64::fixup_aarch64_add_imm12:
398   case AArch64::fixup_aarch64_ldst_imm12_scale1:
399   case AArch64::fixup_aarch64_ldst_imm12_scale2:
400   case AArch64::fixup_aarch64_ldst_imm12_scale4:
401   case AArch64::fixup_aarch64_ldst_imm12_scale8:
402   case AArch64::fixup_aarch64_ldst_imm12_scale16:
403   case AArch64::fixup_aarch64_ldr_pcrel_imm19:
404   case AArch64::fixup_aarch64_pcrel_branch19:
405   case AArch64::fixup_aarch64_pcrel_adr_imm21:
406   case AArch64::fixup_aarch64_pcrel_adrp_imm21:
407   case AArch64::fixup_aarch64_pcrel_branch26:
408   case AArch64::fixup_aarch64_pcrel_call26:
409     // Instructions are always little endian
410     return 0;
411   }
412 }
413 
414 void AArch64AsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
415                                    const MCValue &Target,
416                                    MutableArrayRef<char> Data, uint64_t Value,
417                                    bool IsResolved,
418                                    const MCSubtargetInfo *STI) const {
419   if (Fixup.getTargetKind() == FK_Data_8 && TheTriple.isOSBinFormatELF()) {
420     auto RefKind = static_cast<AArch64MCExpr::VariantKind>(Target.getRefKind());
421     AArch64MCExpr::VariantKind SymLoc = AArch64MCExpr::getSymbolLoc(RefKind);
422     if (SymLoc == AArch64AuthMCExpr::VK_AUTH ||
423         SymLoc == AArch64AuthMCExpr::VK_AUTHADDR) {
424       assert(Value == 0);
425       const auto *Expr = cast<AArch64AuthMCExpr>(Fixup.getValue());
426       Value = (uint64_t(Expr->getDiscriminator()) << 32) |
427               (uint64_t(Expr->getKey()) << 60) |
428               (uint64_t(Expr->hasAddressDiversity()) << 63);
429     }
430   }
431 
432   if (!Value)
433     return; // Doesn't change encoding.
434   unsigned Kind = Fixup.getKind();
435   if (Kind >= FirstLiteralRelocationKind)
436     return;
437   unsigned NumBytes = getFixupKindNumBytes(Kind);
438   MCFixupKindInfo Info = getFixupKindInfo(Fixup.getKind());
439   MCContext &Ctx = Asm.getContext();
440   int64_t SignedValue = static_cast<int64_t>(Value);
441   // Apply any target-specific value adjustments.
442   Value = adjustFixupValue(Fixup, Target, Value, Ctx, TheTriple, IsResolved);
443 
444   // Shift the value into position.
445   Value <<= Info.TargetOffset;
446 
447   unsigned Offset = Fixup.getOffset();
448   assert(Offset + NumBytes <= Data.size() && "Invalid fixup offset!");
449 
450   // Used to point to big endian bytes.
451   unsigned FulleSizeInBytes = getFixupKindContainereSizeInBytes(Fixup.getKind());
452 
453   // For each byte of the fragment that the fixup touches, mask in the
454   // bits from the fixup value.
455   if (FulleSizeInBytes == 0) {
456     // Handle as little-endian
457     for (unsigned i = 0; i != NumBytes; ++i) {
458       Data[Offset + i] |= uint8_t((Value >> (i * 8)) & 0xff);
459     }
460   } else {
461     // Handle as big-endian
462     assert((Offset + FulleSizeInBytes) <= Data.size() && "Invalid fixup size!");
463     assert(NumBytes <= FulleSizeInBytes && "Invalid fixup size!");
464     for (unsigned i = 0; i != NumBytes; ++i) {
465       unsigned Idx = FulleSizeInBytes - 1 - i;
466       Data[Offset + Idx] |= uint8_t((Value >> (i * 8)) & 0xff);
467     }
468   }
469 
470   // FIXME: getFixupKindInfo() and getFixupKindNumBytes() could be fixed to
471   // handle this more cleanly. This may affect the output of -show-mc-encoding.
472   AArch64MCExpr::VariantKind RefKind =
473       static_cast<AArch64MCExpr::VariantKind>(Target.getRefKind());
474   if (AArch64MCExpr::getSymbolLoc(RefKind) == AArch64MCExpr::VK_SABS ||
475       (!RefKind && Fixup.getTargetKind() == AArch64::fixup_aarch64_movw)) {
476     // If the immediate is negative, generate MOVN else MOVZ.
477     // (Bit 30 = 0) ==> MOVN, (Bit 30 = 1) ==> MOVZ.
478     if (SignedValue < 0)
479       Data[Offset + 3] &= ~(1 << 6);
480     else
481       Data[Offset + 3] |= (1 << 6);
482   }
483 }
484 
485 bool AArch64AsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup,
486                                              uint64_t Value,
487                                              const MCRelaxableFragment *DF,
488                                              const MCAsmLayout &Layout) const {
489   // FIXME:  This isn't correct for AArch64. Just moving the "generic" logic
490   // into the targets for now.
491   //
492   // Relax if the value is too big for a (signed) i8.
493   return int64_t(Value) != int64_t(int8_t(Value));
494 }
495 
496 void AArch64AsmBackend::relaxInstruction(MCInst &Inst,
497                                          const MCSubtargetInfo &STI) const {
498   llvm_unreachable("AArch64AsmBackend::relaxInstruction() unimplemented");
499 }
500 
501 bool AArch64AsmBackend::writeNopData(raw_ostream &OS, uint64_t Count,
502                                      const MCSubtargetInfo *STI) const {
503   // If the count is not 4-byte aligned, we must be writing data into the text
504   // section (otherwise we have unaligned instructions, and thus have far
505   // bigger problems), so just write zeros instead.
506   OS.write_zeros(Count % 4);
507 
508   // We are properly aligned, so write NOPs as requested.
509   Count /= 4;
510   for (uint64_t i = 0; i != Count; ++i)
511     OS.write("\x1f\x20\x03\xd5", 4);
512   return true;
513 }
514 
515 bool AArch64AsmBackend::shouldForceRelocation(const MCAssembler &Asm,
516                                               const MCFixup &Fixup,
517                                               const MCValue &Target,
518                                               const MCSubtargetInfo *STI) {
519   unsigned Kind = Fixup.getKind();
520   if (Kind >= FirstLiteralRelocationKind)
521     return true;
522 
523   // The ADRP instruction adds some multiple of 0x1000 to the current PC &
524   // ~0xfff. This means that the required offset to reach a symbol can vary by
525   // up to one step depending on where the ADRP is in memory. For example:
526   //
527   //     ADRP x0, there
528   //  there:
529   //
530   // If the ADRP occurs at address 0xffc then "there" will be at 0x1000 and
531   // we'll need that as an offset. At any other address "there" will be in the
532   // same page as the ADRP and the instruction should encode 0x0. Assuming the
533   // section isn't 0x1000-aligned, we therefore need to delegate this decision
534   // to the linker -- a relocation!
535   if (Kind == AArch64::fixup_aarch64_pcrel_adrp_imm21)
536     return true;
537 
538   return false;
539 }
540 
541 namespace {
542 
543 namespace CU {
544 
545 /// Compact unwind encoding values.
546 enum CompactUnwindEncodings {
547   /// A "frameless" leaf function, where no non-volatile registers are
548   /// saved. The return remains in LR throughout the function.
549   UNWIND_ARM64_MODE_FRAMELESS = 0x02000000,
550 
551   /// No compact unwind encoding available. Instead the low 23-bits of
552   /// the compact unwind encoding is the offset of the DWARF FDE in the
553   /// __eh_frame section. This mode is never used in object files. It is only
554   /// generated by the linker in final linked images, which have only DWARF info
555   /// for a function.
556   UNWIND_ARM64_MODE_DWARF = 0x03000000,
557 
558   /// This is a standard arm64 prologue where FP/LR are immediately
559   /// pushed on the stack, then SP is copied to FP. If there are any
560   /// non-volatile register saved, they are copied into the stack fame in pairs
561   /// in a contiguous ranger right below the saved FP/LR pair. Any subset of the
562   /// five X pairs and four D pairs can be saved, but the memory layout must be
563   /// in register number order.
564   UNWIND_ARM64_MODE_FRAME = 0x04000000,
565 
566   /// Frame register pair encodings.
567   UNWIND_ARM64_FRAME_X19_X20_PAIR = 0x00000001,
568   UNWIND_ARM64_FRAME_X21_X22_PAIR = 0x00000002,
569   UNWIND_ARM64_FRAME_X23_X24_PAIR = 0x00000004,
570   UNWIND_ARM64_FRAME_X25_X26_PAIR = 0x00000008,
571   UNWIND_ARM64_FRAME_X27_X28_PAIR = 0x00000010,
572   UNWIND_ARM64_FRAME_D8_D9_PAIR = 0x00000100,
573   UNWIND_ARM64_FRAME_D10_D11_PAIR = 0x00000200,
574   UNWIND_ARM64_FRAME_D12_D13_PAIR = 0x00000400,
575   UNWIND_ARM64_FRAME_D14_D15_PAIR = 0x00000800
576 };
577 
578 } // end CU namespace
579 
580 // FIXME: This should be in a separate file.
581 class DarwinAArch64AsmBackend : public AArch64AsmBackend {
582   const MCRegisterInfo &MRI;
583 
584   /// Encode compact unwind stack adjustment for frameless functions.
585   /// See UNWIND_ARM64_FRAMELESS_STACK_SIZE_MASK in compact_unwind_encoding.h.
586   /// The stack size always needs to be 16 byte aligned.
587   uint32_t encodeStackAdjustment(uint32_t StackSize) const {
588     return (StackSize / 16) << 12;
589   }
590 
591 public:
592   DarwinAArch64AsmBackend(const Target &T, const Triple &TT,
593                           const MCRegisterInfo &MRI)
594       : AArch64AsmBackend(T, TT, /*IsLittleEndian*/ true), MRI(MRI) {}
595 
596   std::unique_ptr<MCObjectTargetWriter>
597   createObjectTargetWriter() const override {
598     uint32_t CPUType = cantFail(MachO::getCPUType(TheTriple));
599     uint32_t CPUSubType = cantFail(MachO::getCPUSubType(TheTriple));
600     return createAArch64MachObjectWriter(CPUType, CPUSubType,
601                                          TheTriple.isArch32Bit());
602   }
603 
604   /// Generate the compact unwind encoding from the CFI directives.
605   uint32_t generateCompactUnwindEncoding(const MCDwarfFrameInfo *FI,
606                                          const MCContext *Ctxt) const override {
607     ArrayRef<MCCFIInstruction> Instrs = FI->Instructions;
608     if (Instrs.empty())
609       return CU::UNWIND_ARM64_MODE_FRAMELESS;
610     if (!isDarwinCanonicalPersonality(FI->Personality) &&
611         !Ctxt->emitCompactUnwindNonCanonical())
612       return CU::UNWIND_ARM64_MODE_DWARF;
613 
614     bool HasFP = false;
615     unsigned StackSize = 0;
616 
617     uint32_t CompactUnwindEncoding = 0;
618     int CurOffset = 0;
619     for (size_t i = 0, e = Instrs.size(); i != e; ++i) {
620       const MCCFIInstruction &Inst = Instrs[i];
621 
622       switch (Inst.getOperation()) {
623       default:
624         // Cannot handle this directive:  bail out.
625         return CU::UNWIND_ARM64_MODE_DWARF;
626       case MCCFIInstruction::OpDefCfa: {
627         // Defines a frame pointer.
628         unsigned XReg =
629             getXRegFromWReg(*MRI.getLLVMRegNum(Inst.getRegister(), true));
630 
631         // Other CFA registers than FP are not supported by compact unwind.
632         // Fallback on DWARF.
633         // FIXME: When opt-remarks are supported in MC, add a remark to notify
634         // the user.
635         if (XReg != AArch64::FP)
636           return CU::UNWIND_ARM64_MODE_DWARF;
637 
638         if (i + 2 >= e)
639           return CU::UNWIND_ARM64_MODE_DWARF;
640 
641         const MCCFIInstruction &LRPush = Instrs[++i];
642         if (LRPush.getOperation() != MCCFIInstruction::OpOffset)
643           return CU::UNWIND_ARM64_MODE_DWARF;
644         const MCCFIInstruction &FPPush = Instrs[++i];
645         if (FPPush.getOperation() != MCCFIInstruction::OpOffset)
646           return CU::UNWIND_ARM64_MODE_DWARF;
647 
648         if (FPPush.getOffset() + 8 != LRPush.getOffset())
649           return CU::UNWIND_ARM64_MODE_DWARF;
650         CurOffset = FPPush.getOffset();
651 
652         unsigned LRReg = *MRI.getLLVMRegNum(LRPush.getRegister(), true);
653         unsigned FPReg = *MRI.getLLVMRegNum(FPPush.getRegister(), true);
654 
655         LRReg = getXRegFromWReg(LRReg);
656         FPReg = getXRegFromWReg(FPReg);
657 
658         if (LRReg != AArch64::LR || FPReg != AArch64::FP)
659           return CU::UNWIND_ARM64_MODE_DWARF;
660 
661         // Indicate that the function has a frame.
662         CompactUnwindEncoding |= CU::UNWIND_ARM64_MODE_FRAME;
663         HasFP = true;
664         break;
665       }
666       case MCCFIInstruction::OpDefCfaOffset: {
667         if (StackSize != 0)
668           return CU::UNWIND_ARM64_MODE_DWARF;
669         StackSize = std::abs(Inst.getOffset());
670         break;
671       }
672       case MCCFIInstruction::OpOffset: {
673         // Registers are saved in pairs. We expect there to be two consecutive
674         // `.cfi_offset' instructions with the appropriate registers specified.
675         unsigned Reg1 = *MRI.getLLVMRegNum(Inst.getRegister(), true);
676         if (i + 1 == e)
677           return CU::UNWIND_ARM64_MODE_DWARF;
678 
679         if (CurOffset != 0 && Inst.getOffset() != CurOffset - 8)
680           return CU::UNWIND_ARM64_MODE_DWARF;
681         CurOffset = Inst.getOffset();
682 
683         const MCCFIInstruction &Inst2 = Instrs[++i];
684         if (Inst2.getOperation() != MCCFIInstruction::OpOffset)
685           return CU::UNWIND_ARM64_MODE_DWARF;
686         unsigned Reg2 = *MRI.getLLVMRegNum(Inst2.getRegister(), true);
687 
688         if (Inst2.getOffset() != CurOffset - 8)
689           return CU::UNWIND_ARM64_MODE_DWARF;
690         CurOffset = Inst2.getOffset();
691 
692         // N.B. The encodings must be in register number order, and the X
693         // registers before the D registers.
694 
695         // X19/X20 pair = 0x00000001,
696         // X21/X22 pair = 0x00000002,
697         // X23/X24 pair = 0x00000004,
698         // X25/X26 pair = 0x00000008,
699         // X27/X28 pair = 0x00000010
700         Reg1 = getXRegFromWReg(Reg1);
701         Reg2 = getXRegFromWReg(Reg2);
702 
703         if (Reg1 == AArch64::X19 && Reg2 == AArch64::X20 &&
704             (CompactUnwindEncoding & 0xF1E) == 0)
705           CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_X19_X20_PAIR;
706         else if (Reg1 == AArch64::X21 && Reg2 == AArch64::X22 &&
707                  (CompactUnwindEncoding & 0xF1C) == 0)
708           CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_X21_X22_PAIR;
709         else if (Reg1 == AArch64::X23 && Reg2 == AArch64::X24 &&
710                  (CompactUnwindEncoding & 0xF18) == 0)
711           CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_X23_X24_PAIR;
712         else if (Reg1 == AArch64::X25 && Reg2 == AArch64::X26 &&
713                  (CompactUnwindEncoding & 0xF10) == 0)
714           CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_X25_X26_PAIR;
715         else if (Reg1 == AArch64::X27 && Reg2 == AArch64::X28 &&
716                  (CompactUnwindEncoding & 0xF00) == 0)
717           CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_X27_X28_PAIR;
718         else {
719           Reg1 = getDRegFromBReg(Reg1);
720           Reg2 = getDRegFromBReg(Reg2);
721 
722           // D8/D9 pair   = 0x00000100,
723           // D10/D11 pair = 0x00000200,
724           // D12/D13 pair = 0x00000400,
725           // D14/D15 pair = 0x00000800
726           if (Reg1 == AArch64::D8 && Reg2 == AArch64::D9 &&
727               (CompactUnwindEncoding & 0xE00) == 0)
728             CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_D8_D9_PAIR;
729           else if (Reg1 == AArch64::D10 && Reg2 == AArch64::D11 &&
730                    (CompactUnwindEncoding & 0xC00) == 0)
731             CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_D10_D11_PAIR;
732           else if (Reg1 == AArch64::D12 && Reg2 == AArch64::D13 &&
733                    (CompactUnwindEncoding & 0x800) == 0)
734             CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_D12_D13_PAIR;
735           else if (Reg1 == AArch64::D14 && Reg2 == AArch64::D15)
736             CompactUnwindEncoding |= CU::UNWIND_ARM64_FRAME_D14_D15_PAIR;
737           else
738             // A pair was pushed which we cannot handle.
739             return CU::UNWIND_ARM64_MODE_DWARF;
740         }
741 
742         break;
743       }
744       }
745     }
746 
747     if (!HasFP) {
748       // With compact unwind info we can only represent stack adjustments of up
749       // to 65520 bytes.
750       if (StackSize > 65520)
751         return CU::UNWIND_ARM64_MODE_DWARF;
752 
753       CompactUnwindEncoding |= CU::UNWIND_ARM64_MODE_FRAMELESS;
754       CompactUnwindEncoding |= encodeStackAdjustment(StackSize);
755     }
756 
757     return CompactUnwindEncoding;
758   }
759 };
760 
761 } // end anonymous namespace
762 
763 namespace {
764 
765 class ELFAArch64AsmBackend : public AArch64AsmBackend {
766 public:
767   uint8_t OSABI;
768   bool IsILP32;
769 
770   ELFAArch64AsmBackend(const Target &T, const Triple &TT, uint8_t OSABI,
771                        bool IsLittleEndian, bool IsILP32)
772       : AArch64AsmBackend(T, TT, IsLittleEndian), OSABI(OSABI),
773         IsILP32(IsILP32) {}
774 
775   std::unique_ptr<MCObjectTargetWriter>
776   createObjectTargetWriter() const override {
777     return createAArch64ELFObjectWriter(OSABI, IsILP32);
778   }
779 };
780 
781 }
782 
783 namespace {
784 class COFFAArch64AsmBackend : public AArch64AsmBackend {
785 public:
786   COFFAArch64AsmBackend(const Target &T, const Triple &TheTriple)
787       : AArch64AsmBackend(T, TheTriple, /*IsLittleEndian*/ true) {}
788 
789   std::unique_ptr<MCObjectTargetWriter>
790   createObjectTargetWriter() const override {
791     return createAArch64WinCOFFObjectWriter(TheTriple);
792   }
793 };
794 }
795 
796 MCAsmBackend *llvm::createAArch64leAsmBackend(const Target &T,
797                                               const MCSubtargetInfo &STI,
798                                               const MCRegisterInfo &MRI,
799                                               const MCTargetOptions &Options) {
800   const Triple &TheTriple = STI.getTargetTriple();
801   if (TheTriple.isOSBinFormatMachO()) {
802     return new DarwinAArch64AsmBackend(T, TheTriple, MRI);
803   }
804 
805   if (TheTriple.isOSBinFormatCOFF())
806     return new COFFAArch64AsmBackend(T, TheTriple);
807 
808   assert(TheTriple.isOSBinFormatELF() && "Invalid target");
809 
810   uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
811   bool IsILP32 = STI.getTargetTriple().getEnvironment() == Triple::GNUILP32;
812   return new ELFAArch64AsmBackend(T, TheTriple, OSABI, /*IsLittleEndian=*/true,
813                                   IsILP32);
814 }
815 
816 MCAsmBackend *llvm::createAArch64beAsmBackend(const Target &T,
817                                               const MCSubtargetInfo &STI,
818                                               const MCRegisterInfo &MRI,
819                                               const MCTargetOptions &Options) {
820   const Triple &TheTriple = STI.getTargetTriple();
821   assert(TheTriple.isOSBinFormatELF() &&
822          "Big endian is only supported for ELF targets!");
823   uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS());
824   bool IsILP32 = STI.getTargetTriple().getEnvironment() == Triple::GNUILP32;
825   return new ELFAArch64AsmBackend(T, TheTriple, OSABI, /*IsLittleEndian=*/false,
826                                   IsILP32);
827 }
828