1 //===-- ARMAsmBackend.cpp - ARM 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/ARMMCTargetDesc.h"
11 #include "MCTargetDesc/ARMBaseInfo.h"
12 #include "MCTargetDesc/ARMFixupKinds.h"
13 #include "MCTargetDesc/ARMAddressingModes.h"
14 #include "llvm/ADT/Twine.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCDirectives.h"
17 #include "llvm/MC/MCELFObjectWriter.h"
18 #include "llvm/MC/MCExpr.h"
19 #include "llvm/MC/MCMachObjectWriter.h"
20 #include "llvm/MC/MCObjectWriter.h"
21 #include "llvm/MC/MCSectionELF.h"
22 #include "llvm/MC/MCSectionMachO.h"
23 #include "llvm/MC/MCAsmBackend.h"
24 #include "llvm/MC/MCSubtargetInfo.h"
25 #include "llvm/Object/MachOFormat.h"
26 #include "llvm/Support/ELF.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/raw_ostream.h"
32 class ARMELFObjectWriter : public MCELFObjectTargetWriter {
34 ARMELFObjectWriter(Triple::OSType OSType)
35 : MCELFObjectTargetWriter(/*Is64Bit*/ false, OSType, ELF::EM_ARM,
36 /*HasRelocationAddend*/ false) {}
39 class ARMAsmBackend : public MCAsmBackend {
40 const MCSubtargetInfo* STI;
41 bool isThumbMode; // Currently emitting Thumb code.
43 ARMAsmBackend(const Target &T, const StringRef TT)
44 : MCAsmBackend(), STI(ARM_MC::createARMMCSubtargetInfo(TT, "", "")),
45 isThumbMode(TT.startswith("thumb")) {}
51 unsigned getNumFixupKinds() const { return ARM::NumTargetFixupKinds; }
54 return (STI->getFeatureBits() & ARM::HasV6T2Ops) != 0;
57 const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const {
58 const static MCFixupKindInfo Infos[ARM::NumTargetFixupKinds] = {
59 // This table *must* be in the order that the fixup_* kinds are defined in
62 // Name Offset (bits) Size (bits) Flags
63 { "fixup_arm_ldst_pcrel_12", 0, 32, MCFixupKindInfo::FKF_IsPCRel },
64 { "fixup_t2_ldst_pcrel_12", 0, 32, MCFixupKindInfo::FKF_IsPCRel |
65 MCFixupKindInfo::FKF_IsAlignedDownTo32Bits},
66 { "fixup_arm_pcrel_10", 0, 32, MCFixupKindInfo::FKF_IsPCRel },
67 { "fixup_t2_pcrel_10", 0, 32, MCFixupKindInfo::FKF_IsPCRel |
68 MCFixupKindInfo::FKF_IsAlignedDownTo32Bits},
69 { "fixup_thumb_adr_pcrel_10",0, 8, MCFixupKindInfo::FKF_IsPCRel |
70 MCFixupKindInfo::FKF_IsAlignedDownTo32Bits},
71 { "fixup_arm_adr_pcrel_12", 0, 32, MCFixupKindInfo::FKF_IsPCRel },
72 { "fixup_t2_adr_pcrel_12", 0, 32, MCFixupKindInfo::FKF_IsPCRel |
73 MCFixupKindInfo::FKF_IsAlignedDownTo32Bits},
74 { "fixup_arm_condbranch", 0, 24, MCFixupKindInfo::FKF_IsPCRel },
75 { "fixup_arm_uncondbranch", 0, 24, MCFixupKindInfo::FKF_IsPCRel },
76 { "fixup_t2_condbranch", 0, 32, MCFixupKindInfo::FKF_IsPCRel },
77 { "fixup_t2_uncondbranch", 0, 32, MCFixupKindInfo::FKF_IsPCRel },
78 { "fixup_arm_thumb_br", 0, 16, MCFixupKindInfo::FKF_IsPCRel },
79 { "fixup_arm_thumb_bl", 0, 32, MCFixupKindInfo::FKF_IsPCRel },
80 { "fixup_arm_thumb_blx", 0, 32, MCFixupKindInfo::FKF_IsPCRel },
81 { "fixup_arm_thumb_cb", 0, 16, MCFixupKindInfo::FKF_IsPCRel },
82 { "fixup_arm_thumb_cp", 0, 8, MCFixupKindInfo::FKF_IsPCRel },
83 { "fixup_arm_thumb_bcc", 0, 8, MCFixupKindInfo::FKF_IsPCRel },
84 // movw / movt: 16-bits immediate but scattered into two chunks 0 - 12, 16 - 19.
85 { "fixup_arm_movt_hi16", 0, 20, 0 },
86 { "fixup_arm_movw_lo16", 0, 20, 0 },
87 { "fixup_t2_movt_hi16", 0, 20, 0 },
88 { "fixup_t2_movw_lo16", 0, 20, 0 },
89 { "fixup_arm_movt_hi16_pcrel", 0, 20, MCFixupKindInfo::FKF_IsPCRel },
90 { "fixup_arm_movw_lo16_pcrel", 0, 20, MCFixupKindInfo::FKF_IsPCRel },
91 { "fixup_t2_movt_hi16_pcrel", 0, 20, MCFixupKindInfo::FKF_IsPCRel },
92 { "fixup_t2_movw_lo16_pcrel", 0, 20, MCFixupKindInfo::FKF_IsPCRel },
95 if (Kind < FirstTargetFixupKind)
96 return MCAsmBackend::getFixupKindInfo(Kind);
98 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
100 return Infos[Kind - FirstTargetFixupKind];
103 bool MayNeedRelaxation(const MCInst &Inst) const;
105 void RelaxInstruction(const MCInst &Inst, MCInst &Res) const;
107 bool WriteNopData(uint64_t Count, MCObjectWriter *OW) const;
109 void HandleAssemblerFlag(MCAssemblerFlag Flag) {
121 unsigned getPointerSize() const { return 4; }
122 bool isThumb() const { return isThumbMode; }
123 void setIsThumb(bool it) { isThumbMode = it; }
125 } // end anonymous namespace
127 static unsigned getRelaxedOpcode(unsigned Op) {
130 case ARM::tBcc: return ARM::t2Bcc;
134 bool ARMAsmBackend::MayNeedRelaxation(const MCInst &Inst) const {
135 if (getRelaxedOpcode(Inst.getOpcode()) != Inst.getOpcode())
140 void ARMAsmBackend::RelaxInstruction(const MCInst &Inst, MCInst &Res) const {
141 unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode());
143 // Sanity check w/ diagnostic if we get here w/ a bogus instruction.
144 if (RelaxedOp == Inst.getOpcode()) {
145 SmallString<256> Tmp;
146 raw_svector_ostream OS(Tmp);
147 Inst.dump_pretty(OS);
149 report_fatal_error("unexpected instruction to relax: " + OS.str());
152 // The instructions we're relaxing have (so far) the same operands.
153 // We just need to update to the proper opcode.
155 Res.setOpcode(RelaxedOp);
158 bool ARMAsmBackend::WriteNopData(uint64_t Count, MCObjectWriter *OW) const {
159 const uint16_t Thumb1_16bitNopEncoding = 0x46c0; // using MOV r8,r8
160 const uint16_t Thumb2_16bitNopEncoding = 0xbf00; // NOP
161 const uint32_t ARMv4_NopEncoding = 0xe1a0000; // using MOV r0,r0
162 const uint32_t ARMv6T2_NopEncoding = 0xe320f000; // NOP
164 const uint16_t nopEncoding = hasNOP() ? Thumb2_16bitNopEncoding
165 : Thumb1_16bitNopEncoding;
166 uint64_t NumNops = Count / 2;
167 for (uint64_t i = 0; i != NumNops; ++i)
168 OW->Write16(nopEncoding);
174 const uint32_t nopEncoding = hasNOP() ? ARMv6T2_NopEncoding
176 uint64_t NumNops = Count / 4;
177 for (uint64_t i = 0; i != NumNops; ++i)
178 OW->Write32(nopEncoding);
179 // FIXME: should this function return false when unable to write exactly
180 // 'Count' bytes with NOP encodings?
182 default: break; // No leftover bytes to write
183 case 1: OW->Write8(0); break;
184 case 2: OW->Write16(0); break;
185 case 3: OW->Write16(0); OW->Write8(0xa0); break;
191 static unsigned adjustFixupValue(unsigned Kind, uint64_t Value) {
194 llvm_unreachable("Unknown fixup kind!");
199 case ARM::fixup_arm_movt_hi16:
202 case ARM::fixup_arm_movw_lo16:
203 case ARM::fixup_arm_movt_hi16_pcrel:
204 case ARM::fixup_arm_movw_lo16_pcrel: {
205 unsigned Hi4 = (Value & 0xF000) >> 12;
206 unsigned Lo12 = Value & 0x0FFF;
207 // inst{19-16} = Hi4;
208 // inst{11-0} = Lo12;
209 Value = (Hi4 << 16) | (Lo12);
212 case ARM::fixup_t2_movt_hi16:
215 case ARM::fixup_t2_movw_lo16:
216 case ARM::fixup_t2_movt_hi16_pcrel: //FIXME: Shouldn't this be shifted like
217 // the other hi16 fixup?
218 case ARM::fixup_t2_movw_lo16_pcrel: {
219 unsigned Hi4 = (Value & 0xF000) >> 12;
220 unsigned i = (Value & 0x800) >> 11;
221 unsigned Mid3 = (Value & 0x700) >> 8;
222 unsigned Lo8 = Value & 0x0FF;
223 // inst{19-16} = Hi4;
225 // inst{14-12} = Mid3;
227 Value = (Hi4 << 16) | (i << 26) | (Mid3 << 12) | (Lo8);
228 uint64_t swapped = (Value & 0xFFFF0000) >> 16;
229 swapped |= (Value & 0x0000FFFF) << 16;
232 case ARM::fixup_arm_ldst_pcrel_12:
233 // ARM PC-relative values are offset by 8.
236 case ARM::fixup_t2_ldst_pcrel_12: {
237 // Offset by 4, adjusted by two due to the half-word ordering of thumb.
240 if ((int64_t)Value < 0) {
244 assert ((Value < 4096) && "Out of range pc-relative fixup value!");
245 Value |= isAdd << 23;
247 // Same addressing mode as fixup_arm_pcrel_10,
248 // but with 16-bit halfwords swapped.
249 if (Kind == ARM::fixup_t2_ldst_pcrel_12) {
250 uint64_t swapped = (Value & 0xFFFF0000) >> 16;
251 swapped |= (Value & 0x0000FFFF) << 16;
257 case ARM::fixup_thumb_adr_pcrel_10:
258 return ((Value - 4) >> 2) & 0xff;
259 case ARM::fixup_arm_adr_pcrel_12: {
260 // ARM PC-relative values are offset by 8.
262 unsigned opc = 4; // bits {24-21}. Default to add: 0b0100
263 if ((int64_t)Value < 0) {
267 assert(ARM_AM::getSOImmVal(Value) != -1 &&
268 "Out of range pc-relative fixup value!");
269 // Encode the immediate and shift the opcode into place.
270 return ARM_AM::getSOImmVal(Value) | (opc << 21);
273 case ARM::fixup_t2_adr_pcrel_12: {
276 if ((int64_t)Value < 0) {
281 uint32_t out = (opc << 21);
282 out |= (Value & 0x800) << 15;
283 out |= (Value & 0x700) << 4;
284 out |= (Value & 0x0FF);
286 uint64_t swapped = (out & 0xFFFF0000) >> 16;
287 swapped |= (out & 0x0000FFFF) << 16;
291 case ARM::fixup_arm_condbranch:
292 case ARM::fixup_arm_uncondbranch:
293 // These values don't encode the low two bits since they're always zero.
294 // Offset by 8 just as above.
295 return 0xffffff & ((Value - 8) >> 2);
296 case ARM::fixup_t2_uncondbranch: {
298 Value >>= 1; // Low bit is not encoded.
301 bool I = Value & 0x800000;
302 bool J1 = Value & 0x400000;
303 bool J2 = Value & 0x200000;
307 out |= I << 26; // S bit
308 out |= !J1 << 13; // J1 bit
309 out |= !J2 << 11; // J2 bit
310 out |= (Value & 0x1FF800) << 5; // imm6 field
311 out |= (Value & 0x0007FF); // imm11 field
313 uint64_t swapped = (out & 0xFFFF0000) >> 16;
314 swapped |= (out & 0x0000FFFF) << 16;
317 case ARM::fixup_t2_condbranch: {
319 Value >>= 1; // Low bit is not encoded.
322 out |= (Value & 0x80000) << 7; // S bit
323 out |= (Value & 0x40000) >> 7; // J2 bit
324 out |= (Value & 0x20000) >> 4; // J1 bit
325 out |= (Value & 0x1F800) << 5; // imm6 field
326 out |= (Value & 0x007FF); // imm11 field
328 uint32_t swapped = (out & 0xFFFF0000) >> 16;
329 swapped |= (out & 0x0000FFFF) << 16;
332 case ARM::fixup_arm_thumb_bl: {
333 // The value doesn't encode the low bit (always zero) and is offset by
334 // four. The value is encoded into disjoint bit positions in the destination
335 // opcode. x = unchanged, I = immediate value bit, S = sign extension bit
337 // BL: xxxxxSIIIIIIIIII xxxxxIIIIIIIIIII
339 // Note that the halfwords are stored high first, low second; so we need
340 // to transpose the fixup value here to map properly.
341 unsigned isNeg = (int64_t(Value - 4) < 0) ? 1 : 0;
343 Value = 0x3fffff & ((Value - 4) >> 1);
344 Binary = (Value & 0x7ff) << 16; // Low imm11 value.
345 Binary |= (Value & 0x1ffc00) >> 11; // High imm10 value.
346 Binary |= isNeg << 10; // Sign bit.
349 case ARM::fixup_arm_thumb_blx: {
350 // The value doesn't encode the low two bits (always zero) and is offset by
351 // four (see fixup_arm_thumb_cp). The value is encoded into disjoint bit
352 // positions in the destination opcode. x = unchanged, I = immediate value
353 // bit, S = sign extension bit, 0 = zero.
355 // BLX: xxxxxSIIIIIIIIII xxxxxIIIIIIIIII0
357 // Note that the halfwords are stored high first, low second; so we need
358 // to transpose the fixup value here to map properly.
359 unsigned isNeg = (int64_t(Value-4) < 0) ? 1 : 0;
361 Value = 0xfffff & ((Value - 2) >> 2);
362 Binary = (Value & 0x3ff) << 17; // Low imm10L value.
363 Binary |= (Value & 0xffc00) >> 10; // High imm10H value.
364 Binary |= isNeg << 10; // Sign bit.
367 case ARM::fixup_arm_thumb_cp:
368 // Offset by 4, and don't encode the low two bits. Two bytes of that
369 // 'off by 4' is implicitly handled by the half-word ordering of the
370 // Thumb encoding, so we only need to adjust by 2 here.
371 return ((Value - 2) >> 2) & 0xff;
372 case ARM::fixup_arm_thumb_cb: {
373 // Offset by 4 and don't encode the lower bit, which is always 0.
374 uint32_t Binary = (Value - 4) >> 1;
375 return ((Binary & 0x20) << 4) | ((Binary & 0x1f) << 3);
377 case ARM::fixup_arm_thumb_br:
378 // Offset by 4 and don't encode the lower bit, which is always 0.
379 return ((Value - 4) >> 1) & 0x7ff;
380 case ARM::fixup_arm_thumb_bcc:
381 // Offset by 4 and don't encode the lower bit, which is always 0.
382 return ((Value - 4) >> 1) & 0xff;
383 case ARM::fixup_arm_pcrel_10:
384 Value = Value - 4; // ARM fixups offset by an additional word and don't
385 // need to adjust for the half-word ordering.
387 case ARM::fixup_t2_pcrel_10: {
388 // Offset by 4, adjusted by two due to the half-word ordering of thumb.
391 if ((int64_t)Value < 0) {
395 // These values don't encode the low two bits since they're always zero.
397 assert ((Value < 256) && "Out of range pc-relative fixup value!");
398 Value |= isAdd << 23;
400 // Same addressing mode as fixup_arm_pcrel_10,
401 // but with 16-bit halfwords swapped.
402 if (Kind == ARM::fixup_t2_pcrel_10) {
403 uint32_t swapped = (Value & 0xFFFF0000) >> 16;
404 swapped |= (Value & 0x0000FFFF) << 16;
415 // FIXME: This should be in a separate file.
416 // ELF is an ELF of course...
417 class ELFARMAsmBackend : public ARMAsmBackend {
419 Triple::OSType OSType;
420 ELFARMAsmBackend(const Target &T, const StringRef TT,
421 Triple::OSType _OSType)
422 : ARMAsmBackend(T, TT), OSType(_OSType) { }
424 void ApplyFixup(const MCFixup &Fixup, char *Data, unsigned DataSize,
425 uint64_t Value) const;
427 MCObjectWriter *createObjectWriter(raw_ostream &OS) const {
428 return createELFObjectWriter(new ARMELFObjectWriter(OSType), OS,
429 /*IsLittleEndian*/ true);
433 // FIXME: Raise this to share code between Darwin and ELF.
434 void ELFARMAsmBackend::ApplyFixup(const MCFixup &Fixup, char *Data,
435 unsigned DataSize, uint64_t Value) const {
436 unsigned NumBytes = 4; // FIXME: 2 for Thumb
437 Value = adjustFixupValue(Fixup.getKind(), Value);
438 if (!Value) return; // Doesn't change encoding.
440 unsigned Offset = Fixup.getOffset();
442 // For each byte of the fragment that the fixup touches, mask in the bits from
443 // the fixup value. The Value has been "split up" into the appropriate
445 for (unsigned i = 0; i != NumBytes; ++i)
446 Data[Offset + i] |= uint8_t((Value >> (i * 8)) & 0xff);
449 // FIXME: This should be in a separate file.
450 class DarwinARMAsmBackend : public ARMAsmBackend {
452 const object::mach::CPUSubtypeARM Subtype;
453 DarwinARMAsmBackend(const Target &T, const StringRef TT,
454 object::mach::CPUSubtypeARM st)
455 : ARMAsmBackend(T, TT), Subtype(st) { }
457 MCObjectWriter *createObjectWriter(raw_ostream &OS) const {
458 return createARMMachObjectWriter(OS, /*Is64Bit=*/false,
459 object::mach::CTM_ARM,
463 void ApplyFixup(const MCFixup &Fixup, char *Data, unsigned DataSize,
464 uint64_t Value) const;
466 virtual bool doesSectionRequireSymbols(const MCSection &Section) const {
471 /// getFixupKindNumBytes - The number of bytes the fixup may change.
472 static unsigned getFixupKindNumBytes(unsigned Kind) {
475 llvm_unreachable("Unknown fixup kind!");
478 case ARM::fixup_arm_thumb_bcc:
479 case ARM::fixup_arm_thumb_cp:
480 case ARM::fixup_thumb_adr_pcrel_10:
484 case ARM::fixup_arm_thumb_br:
485 case ARM::fixup_arm_thumb_cb:
488 case ARM::fixup_arm_ldst_pcrel_12:
489 case ARM::fixup_arm_pcrel_10:
490 case ARM::fixup_arm_adr_pcrel_12:
491 case ARM::fixup_arm_condbranch:
492 case ARM::fixup_arm_uncondbranch:
496 case ARM::fixup_t2_ldst_pcrel_12:
497 case ARM::fixup_t2_condbranch:
498 case ARM::fixup_t2_uncondbranch:
499 case ARM::fixup_t2_pcrel_10:
500 case ARM::fixup_t2_adr_pcrel_12:
501 case ARM::fixup_arm_thumb_bl:
502 case ARM::fixup_arm_thumb_blx:
503 case ARM::fixup_arm_movt_hi16:
504 case ARM::fixup_arm_movw_lo16:
505 case ARM::fixup_arm_movt_hi16_pcrel:
506 case ARM::fixup_arm_movw_lo16_pcrel:
507 case ARM::fixup_t2_movt_hi16:
508 case ARM::fixup_t2_movw_lo16:
509 case ARM::fixup_t2_movt_hi16_pcrel:
510 case ARM::fixup_t2_movw_lo16_pcrel:
515 void DarwinARMAsmBackend::ApplyFixup(const MCFixup &Fixup, char *Data,
516 unsigned DataSize, uint64_t Value) const {
517 unsigned NumBytes = getFixupKindNumBytes(Fixup.getKind());
518 Value = adjustFixupValue(Fixup.getKind(), Value);
519 if (!Value) return; // Doesn't change encoding.
521 unsigned Offset = Fixup.getOffset();
522 assert(Offset + NumBytes <= DataSize && "Invalid fixup offset!");
524 // For each byte of the fragment that the fixup touches, mask in the
525 // bits from the fixup value.
526 for (unsigned i = 0; i != NumBytes; ++i)
527 Data[Offset + i] |= uint8_t((Value >> (i * 8)) & 0xff);
530 } // end anonymous namespace
532 MCAsmBackend *llvm::createARMAsmBackend(const Target &T, StringRef TT) {
533 Triple TheTriple(TT);
535 if (TheTriple.isOSDarwin()) {
536 if (TheTriple.getArchName() == "armv4t" ||
537 TheTriple.getArchName() == "thumbv4t")
538 return new DarwinARMAsmBackend(T, TT, object::mach::CSARM_V4T);
539 else if (TheTriple.getArchName() == "armv5e" ||
540 TheTriple.getArchName() == "thumbv5e")
541 return new DarwinARMAsmBackend(T, TT, object::mach::CSARM_V5TEJ);
542 else if (TheTriple.getArchName() == "armv6" ||
543 TheTriple.getArchName() == "thumbv6")
544 return new DarwinARMAsmBackend(T, TT, object::mach::CSARM_V6);
545 return new DarwinARMAsmBackend(T, TT, object::mach::CSARM_V7);
548 if (TheTriple.isOSWindows())
549 assert(0 && "Windows not supported on ARM");
551 return new ELFARMAsmBackend(T, TT, Triple(TT).getOS());