xref: /freebsd/contrib/llvm-project/lld/ELF/Arch/RISCV.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- RISCV.cpp ----------------------------------------------------------===//
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 "InputFiles.h"
10 #include "OutputSections.h"
11 #include "Symbols.h"
12 #include "SyntheticSections.h"
13 #include "Target.h"
14 #include "llvm/Support/ELFAttributes.h"
15 #include "llvm/Support/LEB128.h"
16 #include "llvm/Support/RISCVAttributeParser.h"
17 #include "llvm/Support/RISCVAttributes.h"
18 #include "llvm/Support/TimeProfiler.h"
19 #include "llvm/TargetParser/RISCVISAInfo.h"
20 
21 using namespace llvm;
22 using namespace llvm::object;
23 using namespace llvm::support::endian;
24 using namespace llvm::ELF;
25 using namespace lld;
26 using namespace lld::elf;
27 
28 namespace {
29 
30 class RISCV final : public TargetInfo {
31 public:
32   RISCV();
33   uint32_t calcEFlags() const override;
34   int64_t getImplicitAddend(const uint8_t *buf, RelType type) const override;
35   void writeGotHeader(uint8_t *buf) const override;
36   void writeGotPlt(uint8_t *buf, const Symbol &s) const override;
37   void writeIgotPlt(uint8_t *buf, const Symbol &s) const override;
38   void writePltHeader(uint8_t *buf) const override;
39   void writePlt(uint8_t *buf, const Symbol &sym,
40                 uint64_t pltEntryAddr) const override;
41   RelType getDynRel(RelType type) const override;
42   RelExpr getRelExpr(RelType type, const Symbol &s,
43                      const uint8_t *loc) const override;
44   void relocate(uint8_t *loc, const Relocation &rel,
45                 uint64_t val) const override;
46   void relocateAlloc(InputSectionBase &sec, uint8_t *buf) const override;
47   bool relaxOnce(int pass) const override;
48   void finalizeRelax(int passes) const override;
49 };
50 
51 } // end anonymous namespace
52 
53 // These are internal relocation numbers for GP relaxation. They aren't part
54 // of the psABI spec.
55 #define INTERNAL_R_RISCV_GPREL_I 256
56 #define INTERNAL_R_RISCV_GPREL_S 257
57 
58 const uint64_t dtpOffset = 0x800;
59 
60 namespace {
61 enum Op {
62   ADDI = 0x13,
63   AUIPC = 0x17,
64   JALR = 0x67,
65   LD = 0x3003,
66   LUI = 0x37,
67   LW = 0x2003,
68   SRLI = 0x5013,
69   SUB = 0x40000033,
70 };
71 
72 enum Reg {
73   X_RA = 1,
74   X_GP = 3,
75   X_TP = 4,
76   X_T0 = 5,
77   X_T1 = 6,
78   X_T2 = 7,
79   X_A0 = 10,
80   X_T3 = 28,
81 };
82 } // namespace
83 
hi20(uint32_t val)84 static uint32_t hi20(uint32_t val) { return (val + 0x800) >> 12; }
lo12(uint32_t val)85 static uint32_t lo12(uint32_t val) { return val & 4095; }
86 
itype(uint32_t op,uint32_t rd,uint32_t rs1,uint32_t imm)87 static uint32_t itype(uint32_t op, uint32_t rd, uint32_t rs1, uint32_t imm) {
88   return op | (rd << 7) | (rs1 << 15) | (imm << 20);
89 }
rtype(uint32_t op,uint32_t rd,uint32_t rs1,uint32_t rs2)90 static uint32_t rtype(uint32_t op, uint32_t rd, uint32_t rs1, uint32_t rs2) {
91   return op | (rd << 7) | (rs1 << 15) | (rs2 << 20);
92 }
utype(uint32_t op,uint32_t rd,uint32_t imm)93 static uint32_t utype(uint32_t op, uint32_t rd, uint32_t imm) {
94   return op | (rd << 7) | (imm << 12);
95 }
96 
97 // Extract bits v[begin:end], where range is inclusive, and begin must be < 63.
extractBits(uint64_t v,uint32_t begin,uint32_t end)98 static uint32_t extractBits(uint64_t v, uint32_t begin, uint32_t end) {
99   return (v & ((1ULL << (begin + 1)) - 1)) >> end;
100 }
101 
setLO12_I(uint32_t insn,uint32_t imm)102 static uint32_t setLO12_I(uint32_t insn, uint32_t imm) {
103   return (insn & 0xfffff) | (imm << 20);
104 }
setLO12_S(uint32_t insn,uint32_t imm)105 static uint32_t setLO12_S(uint32_t insn, uint32_t imm) {
106   return (insn & 0x1fff07f) | (extractBits(imm, 11, 5) << 25) |
107          (extractBits(imm, 4, 0) << 7);
108 }
109 
RISCV()110 RISCV::RISCV() {
111   copyRel = R_RISCV_COPY;
112   pltRel = R_RISCV_JUMP_SLOT;
113   relativeRel = R_RISCV_RELATIVE;
114   iRelativeRel = R_RISCV_IRELATIVE;
115   if (config->is64) {
116     symbolicRel = R_RISCV_64;
117     tlsModuleIndexRel = R_RISCV_TLS_DTPMOD64;
118     tlsOffsetRel = R_RISCV_TLS_DTPREL64;
119     tlsGotRel = R_RISCV_TLS_TPREL64;
120   } else {
121     symbolicRel = R_RISCV_32;
122     tlsModuleIndexRel = R_RISCV_TLS_DTPMOD32;
123     tlsOffsetRel = R_RISCV_TLS_DTPREL32;
124     tlsGotRel = R_RISCV_TLS_TPREL32;
125   }
126   gotRel = symbolicRel;
127   tlsDescRel = R_RISCV_TLSDESC;
128 
129   // .got[0] = _DYNAMIC
130   gotHeaderEntriesNum = 1;
131 
132   // .got.plt[0] = _dl_runtime_resolve, .got.plt[1] = link_map
133   gotPltHeaderEntriesNum = 2;
134 
135   pltHeaderSize = 32;
136   pltEntrySize = 16;
137   ipltEntrySize = 16;
138 }
139 
getEFlags(InputFile * f)140 static uint32_t getEFlags(InputFile *f) {
141   if (config->is64)
142     return cast<ObjFile<ELF64LE>>(f)->getObj().getHeader().e_flags;
143   return cast<ObjFile<ELF32LE>>(f)->getObj().getHeader().e_flags;
144 }
145 
calcEFlags() const146 uint32_t RISCV::calcEFlags() const {
147   // If there are only binary input files (from -b binary), use a
148   // value of 0 for the ELF header flags.
149   if (ctx.objectFiles.empty())
150     return 0;
151 
152   uint32_t target = getEFlags(ctx.objectFiles.front());
153 
154   for (InputFile *f : ctx.objectFiles) {
155     uint32_t eflags = getEFlags(f);
156     if (eflags & EF_RISCV_RVC)
157       target |= EF_RISCV_RVC;
158 
159     if ((eflags & EF_RISCV_FLOAT_ABI) != (target & EF_RISCV_FLOAT_ABI))
160       error(
161           toString(f) +
162           ": cannot link object files with different floating-point ABI from " +
163           toString(ctx.objectFiles[0]));
164 
165     if ((eflags & EF_RISCV_RVE) != (target & EF_RISCV_RVE))
166       error(toString(f) +
167             ": cannot link object files with different EF_RISCV_RVE");
168   }
169 
170   return target;
171 }
172 
getImplicitAddend(const uint8_t * buf,RelType type) const173 int64_t RISCV::getImplicitAddend(const uint8_t *buf, RelType type) const {
174   switch (type) {
175   default:
176     internalLinkerError(getErrorLocation(buf),
177                         "cannot read addend for relocation " + toString(type));
178     return 0;
179   case R_RISCV_32:
180   case R_RISCV_TLS_DTPMOD32:
181   case R_RISCV_TLS_DTPREL32:
182   case R_RISCV_TLS_TPREL32:
183     return SignExtend64<32>(read32le(buf));
184   case R_RISCV_64:
185   case R_RISCV_TLS_DTPMOD64:
186   case R_RISCV_TLS_DTPREL64:
187   case R_RISCV_TLS_TPREL64:
188     return read64le(buf);
189   case R_RISCV_RELATIVE:
190   case R_RISCV_IRELATIVE:
191     return config->is64 ? read64le(buf) : read32le(buf);
192   case R_RISCV_NONE:
193   case R_RISCV_JUMP_SLOT:
194     // These relocations are defined as not having an implicit addend.
195     return 0;
196   case R_RISCV_TLSDESC:
197     return config->is64 ? read64le(buf + 8) : read32le(buf + 4);
198   }
199 }
200 
writeGotHeader(uint8_t * buf) const201 void RISCV::writeGotHeader(uint8_t *buf) const {
202   if (config->is64)
203     write64le(buf, mainPart->dynamic->getVA());
204   else
205     write32le(buf, mainPart->dynamic->getVA());
206 }
207 
writeGotPlt(uint8_t * buf,const Symbol & s) const208 void RISCV::writeGotPlt(uint8_t *buf, const Symbol &s) const {
209   if (config->is64)
210     write64le(buf, in.plt->getVA());
211   else
212     write32le(buf, in.plt->getVA());
213 }
214 
writeIgotPlt(uint8_t * buf,const Symbol & s) const215 void RISCV::writeIgotPlt(uint8_t *buf, const Symbol &s) const {
216   if (config->writeAddends) {
217     if (config->is64)
218       write64le(buf, s.getVA());
219     else
220       write32le(buf, s.getVA());
221   }
222 }
223 
writePltHeader(uint8_t * buf) const224 void RISCV::writePltHeader(uint8_t *buf) const {
225   // 1: auipc t2, %pcrel_hi(.got.plt)
226   // sub t1, t1, t3
227   // l[wd] t3, %pcrel_lo(1b)(t2); t3 = _dl_runtime_resolve
228   // addi t1, t1, -pltHeaderSize-12; t1 = &.plt[i] - &.plt[0]
229   // addi t0, t2, %pcrel_lo(1b)
230   // srli t1, t1, (rv64?1:2); t1 = &.got.plt[i] - &.got.plt[0]
231   // l[wd] t0, Wordsize(t0); t0 = link_map
232   // jr t3
233   uint32_t offset = in.gotPlt->getVA() - in.plt->getVA();
234   uint32_t load = config->is64 ? LD : LW;
235   write32le(buf + 0, utype(AUIPC, X_T2, hi20(offset)));
236   write32le(buf + 4, rtype(SUB, X_T1, X_T1, X_T3));
237   write32le(buf + 8, itype(load, X_T3, X_T2, lo12(offset)));
238   write32le(buf + 12, itype(ADDI, X_T1, X_T1, -target->pltHeaderSize - 12));
239   write32le(buf + 16, itype(ADDI, X_T0, X_T2, lo12(offset)));
240   write32le(buf + 20, itype(SRLI, X_T1, X_T1, config->is64 ? 1 : 2));
241   write32le(buf + 24, itype(load, X_T0, X_T0, config->wordsize));
242   write32le(buf + 28, itype(JALR, 0, X_T3, 0));
243 }
244 
writePlt(uint8_t * buf,const Symbol & sym,uint64_t pltEntryAddr) const245 void RISCV::writePlt(uint8_t *buf, const Symbol &sym,
246                      uint64_t pltEntryAddr) const {
247   // 1: auipc t3, %pcrel_hi(f@.got.plt)
248   // l[wd] t3, %pcrel_lo(1b)(t3)
249   // jalr t1, t3
250   // nop
251   uint32_t offset = sym.getGotPltVA() - pltEntryAddr;
252   write32le(buf + 0, utype(AUIPC, X_T3, hi20(offset)));
253   write32le(buf + 4, itype(config->is64 ? LD : LW, X_T3, X_T3, lo12(offset)));
254   write32le(buf + 8, itype(JALR, X_T1, X_T3, 0));
255   write32le(buf + 12, itype(ADDI, 0, 0, 0));
256 }
257 
getDynRel(RelType type) const258 RelType RISCV::getDynRel(RelType type) const {
259   return type == target->symbolicRel ? type
260                                      : static_cast<RelType>(R_RISCV_NONE);
261 }
262 
getRelExpr(const RelType type,const Symbol & s,const uint8_t * loc) const263 RelExpr RISCV::getRelExpr(const RelType type, const Symbol &s,
264                           const uint8_t *loc) const {
265   switch (type) {
266   case R_RISCV_NONE:
267     return R_NONE;
268   case R_RISCV_32:
269   case R_RISCV_64:
270   case R_RISCV_HI20:
271   case R_RISCV_LO12_I:
272   case R_RISCV_LO12_S:
273   case R_RISCV_RVC_LUI:
274     return R_ABS;
275   case R_RISCV_ADD8:
276   case R_RISCV_ADD16:
277   case R_RISCV_ADD32:
278   case R_RISCV_ADD64:
279   case R_RISCV_SET6:
280   case R_RISCV_SET8:
281   case R_RISCV_SET16:
282   case R_RISCV_SET32:
283   case R_RISCV_SUB6:
284   case R_RISCV_SUB8:
285   case R_RISCV_SUB16:
286   case R_RISCV_SUB32:
287   case R_RISCV_SUB64:
288     return R_RISCV_ADD;
289   case R_RISCV_JAL:
290   case R_RISCV_BRANCH:
291   case R_RISCV_PCREL_HI20:
292   case R_RISCV_RVC_BRANCH:
293   case R_RISCV_RVC_JUMP:
294   case R_RISCV_32_PCREL:
295     return R_PC;
296   case R_RISCV_CALL:
297   case R_RISCV_CALL_PLT:
298   case R_RISCV_PLT32:
299     return R_PLT_PC;
300   case R_RISCV_GOT_HI20:
301   case R_RISCV_GOT32_PCREL:
302     return R_GOT_PC;
303   case R_RISCV_PCREL_LO12_I:
304   case R_RISCV_PCREL_LO12_S:
305     return R_RISCV_PC_INDIRECT;
306   case R_RISCV_TLSDESC_HI20:
307   case R_RISCV_TLSDESC_LOAD_LO12:
308   case R_RISCV_TLSDESC_ADD_LO12:
309     return R_TLSDESC_PC;
310   case R_RISCV_TLSDESC_CALL:
311     return R_TLSDESC_CALL;
312   case R_RISCV_TLS_GD_HI20:
313     return R_TLSGD_PC;
314   case R_RISCV_TLS_GOT_HI20:
315     return R_GOT_PC;
316   case R_RISCV_TPREL_HI20:
317   case R_RISCV_TPREL_LO12_I:
318   case R_RISCV_TPREL_LO12_S:
319     return R_TPREL;
320   case R_RISCV_ALIGN:
321     return R_RELAX_HINT;
322   case R_RISCV_TPREL_ADD:
323   case R_RISCV_RELAX:
324     return config->relax ? R_RELAX_HINT : R_NONE;
325   case R_RISCV_SET_ULEB128:
326   case R_RISCV_SUB_ULEB128:
327     return R_RISCV_LEB128;
328   default:
329     error(getErrorLocation(loc) + "unknown relocation (" + Twine(type) +
330           ") against symbol " + toString(s));
331     return R_NONE;
332   }
333 }
334 
relocate(uint8_t * loc,const Relocation & rel,uint64_t val) const335 void RISCV::relocate(uint8_t *loc, const Relocation &rel, uint64_t val) const {
336   const unsigned bits = config->wordsize * 8;
337 
338   switch (rel.type) {
339   case R_RISCV_32:
340     write32le(loc, val);
341     return;
342   case R_RISCV_64:
343     write64le(loc, val);
344     return;
345 
346   case R_RISCV_RVC_BRANCH: {
347     checkInt(loc, val, 9, rel);
348     checkAlignment(loc, val, 2, rel);
349     uint16_t insn = read16le(loc) & 0xE383;
350     uint16_t imm8 = extractBits(val, 8, 8) << 12;
351     uint16_t imm4_3 = extractBits(val, 4, 3) << 10;
352     uint16_t imm7_6 = extractBits(val, 7, 6) << 5;
353     uint16_t imm2_1 = extractBits(val, 2, 1) << 3;
354     uint16_t imm5 = extractBits(val, 5, 5) << 2;
355     insn |= imm8 | imm4_3 | imm7_6 | imm2_1 | imm5;
356 
357     write16le(loc, insn);
358     return;
359   }
360 
361   case R_RISCV_RVC_JUMP: {
362     checkInt(loc, val, 12, rel);
363     checkAlignment(loc, val, 2, rel);
364     uint16_t insn = read16le(loc) & 0xE003;
365     uint16_t imm11 = extractBits(val, 11, 11) << 12;
366     uint16_t imm4 = extractBits(val, 4, 4) << 11;
367     uint16_t imm9_8 = extractBits(val, 9, 8) << 9;
368     uint16_t imm10 = extractBits(val, 10, 10) << 8;
369     uint16_t imm6 = extractBits(val, 6, 6) << 7;
370     uint16_t imm7 = extractBits(val, 7, 7) << 6;
371     uint16_t imm3_1 = extractBits(val, 3, 1) << 3;
372     uint16_t imm5 = extractBits(val, 5, 5) << 2;
373     insn |= imm11 | imm4 | imm9_8 | imm10 | imm6 | imm7 | imm3_1 | imm5;
374 
375     write16le(loc, insn);
376     return;
377   }
378 
379   case R_RISCV_RVC_LUI: {
380     int64_t imm = SignExtend64(val + 0x800, bits) >> 12;
381     checkInt(loc, imm, 6, rel);
382     if (imm == 0) { // `c.lui rd, 0` is illegal, convert to `c.li rd, 0`
383       write16le(loc, (read16le(loc) & 0x0F83) | 0x4000);
384     } else {
385       uint16_t imm17 = extractBits(val + 0x800, 17, 17) << 12;
386       uint16_t imm16_12 = extractBits(val + 0x800, 16, 12) << 2;
387       write16le(loc, (read16le(loc) & 0xEF83) | imm17 | imm16_12);
388     }
389     return;
390   }
391 
392   case R_RISCV_JAL: {
393     checkInt(loc, val, 21, rel);
394     checkAlignment(loc, val, 2, rel);
395 
396     uint32_t insn = read32le(loc) & 0xFFF;
397     uint32_t imm20 = extractBits(val, 20, 20) << 31;
398     uint32_t imm10_1 = extractBits(val, 10, 1) << 21;
399     uint32_t imm11 = extractBits(val, 11, 11) << 20;
400     uint32_t imm19_12 = extractBits(val, 19, 12) << 12;
401     insn |= imm20 | imm10_1 | imm11 | imm19_12;
402 
403     write32le(loc, insn);
404     return;
405   }
406 
407   case R_RISCV_BRANCH: {
408     checkInt(loc, val, 13, rel);
409     checkAlignment(loc, val, 2, rel);
410 
411     uint32_t insn = read32le(loc) & 0x1FFF07F;
412     uint32_t imm12 = extractBits(val, 12, 12) << 31;
413     uint32_t imm10_5 = extractBits(val, 10, 5) << 25;
414     uint32_t imm4_1 = extractBits(val, 4, 1) << 8;
415     uint32_t imm11 = extractBits(val, 11, 11) << 7;
416     insn |= imm12 | imm10_5 | imm4_1 | imm11;
417 
418     write32le(loc, insn);
419     return;
420   }
421 
422   // auipc + jalr pair
423   case R_RISCV_CALL:
424   case R_RISCV_CALL_PLT: {
425     int64_t hi = SignExtend64(val + 0x800, bits) >> 12;
426     checkInt(loc, hi, 20, rel);
427     if (isInt<20>(hi)) {
428       relocateNoSym(loc, R_RISCV_PCREL_HI20, val);
429       relocateNoSym(loc + 4, R_RISCV_PCREL_LO12_I, val);
430     }
431     return;
432   }
433 
434   case R_RISCV_GOT_HI20:
435   case R_RISCV_PCREL_HI20:
436   case R_RISCV_TLSDESC_HI20:
437   case R_RISCV_TLS_GD_HI20:
438   case R_RISCV_TLS_GOT_HI20:
439   case R_RISCV_TPREL_HI20:
440   case R_RISCV_HI20: {
441     uint64_t hi = val + 0x800;
442     checkInt(loc, SignExtend64(hi, bits) >> 12, 20, rel);
443     write32le(loc, (read32le(loc) & 0xFFF) | (hi & 0xFFFFF000));
444     return;
445   }
446 
447   case R_RISCV_PCREL_LO12_I:
448   case R_RISCV_TLSDESC_LOAD_LO12:
449   case R_RISCV_TLSDESC_ADD_LO12:
450   case R_RISCV_TPREL_LO12_I:
451   case R_RISCV_LO12_I: {
452     uint64_t hi = (val + 0x800) >> 12;
453     uint64_t lo = val - (hi << 12);
454     write32le(loc, setLO12_I(read32le(loc), lo & 0xfff));
455     return;
456   }
457 
458   case R_RISCV_PCREL_LO12_S:
459   case R_RISCV_TPREL_LO12_S:
460   case R_RISCV_LO12_S: {
461     uint64_t hi = (val + 0x800) >> 12;
462     uint64_t lo = val - (hi << 12);
463     write32le(loc, setLO12_S(read32le(loc), lo));
464     return;
465   }
466 
467   case INTERNAL_R_RISCV_GPREL_I:
468   case INTERNAL_R_RISCV_GPREL_S: {
469     Defined *gp = ElfSym::riscvGlobalPointer;
470     int64_t displace = SignExtend64(val - gp->getVA(), bits);
471     checkInt(loc, displace, 12, rel);
472     uint32_t insn = (read32le(loc) & ~(31 << 15)) | (X_GP << 15);
473     if (rel.type == INTERNAL_R_RISCV_GPREL_I)
474       insn = setLO12_I(insn, displace);
475     else
476       insn = setLO12_S(insn, displace);
477     write32le(loc, insn);
478     return;
479   }
480 
481   case R_RISCV_ADD8:
482     *loc += val;
483     return;
484   case R_RISCV_ADD16:
485     write16le(loc, read16le(loc) + val);
486     return;
487   case R_RISCV_ADD32:
488     write32le(loc, read32le(loc) + val);
489     return;
490   case R_RISCV_ADD64:
491     write64le(loc, read64le(loc) + val);
492     return;
493   case R_RISCV_SUB6:
494     *loc = (*loc & 0xc0) | (((*loc & 0x3f) - val) & 0x3f);
495     return;
496   case R_RISCV_SUB8:
497     *loc -= val;
498     return;
499   case R_RISCV_SUB16:
500     write16le(loc, read16le(loc) - val);
501     return;
502   case R_RISCV_SUB32:
503     write32le(loc, read32le(loc) - val);
504     return;
505   case R_RISCV_SUB64:
506     write64le(loc, read64le(loc) - val);
507     return;
508   case R_RISCV_SET6:
509     *loc = (*loc & 0xc0) | (val & 0x3f);
510     return;
511   case R_RISCV_SET8:
512     *loc = val;
513     return;
514   case R_RISCV_SET16:
515     write16le(loc, val);
516     return;
517   case R_RISCV_SET32:
518   case R_RISCV_32_PCREL:
519   case R_RISCV_PLT32:
520   case R_RISCV_GOT32_PCREL:
521     checkInt(loc, val, 32, rel);
522     write32le(loc, val);
523     return;
524 
525   case R_RISCV_TLS_DTPREL32:
526     write32le(loc, val - dtpOffset);
527     break;
528   case R_RISCV_TLS_DTPREL64:
529     write64le(loc, val - dtpOffset);
530     break;
531 
532   case R_RISCV_RELAX:
533     return;
534   case R_RISCV_TLSDESC:
535     // The addend is stored in the second word.
536     if (config->is64)
537       write64le(loc + 8, val);
538     else
539       write32le(loc + 4, val);
540     break;
541   default:
542     llvm_unreachable("unknown relocation");
543   }
544 }
545 
relaxable(ArrayRef<Relocation> relocs,size_t i)546 static bool relaxable(ArrayRef<Relocation> relocs, size_t i) {
547   return i + 1 != relocs.size() && relocs[i + 1].type == R_RISCV_RELAX;
548 }
549 
tlsdescToIe(uint8_t * loc,const Relocation & rel,uint64_t val)550 static void tlsdescToIe(uint8_t *loc, const Relocation &rel, uint64_t val) {
551   switch (rel.type) {
552   case R_RISCV_TLSDESC_HI20:
553   case R_RISCV_TLSDESC_LOAD_LO12:
554     write32le(loc, 0x00000013); // nop
555     break;
556   case R_RISCV_TLSDESC_ADD_LO12:
557     write32le(loc, utype(AUIPC, X_A0, hi20(val))); // auipc a0,<hi20>
558     break;
559   case R_RISCV_TLSDESC_CALL:
560     if (config->is64)
561       write32le(loc, itype(LD, X_A0, X_A0, lo12(val))); // ld a0,<lo12>(a0)
562     else
563       write32le(loc, itype(LW, X_A0, X_A0, lo12(val))); // lw a0,<lo12>(a0)
564     break;
565   default:
566     llvm_unreachable("unsupported relocation for TLSDESC to IE");
567   }
568 }
569 
tlsdescToLe(uint8_t * loc,const Relocation & rel,uint64_t val)570 static void tlsdescToLe(uint8_t *loc, const Relocation &rel, uint64_t val) {
571   switch (rel.type) {
572   case R_RISCV_TLSDESC_HI20:
573   case R_RISCV_TLSDESC_LOAD_LO12:
574     write32le(loc, 0x00000013); // nop
575     return;
576   case R_RISCV_TLSDESC_ADD_LO12:
577     if (isInt<12>(val))
578       write32le(loc, 0x00000013); // nop
579     else
580       write32le(loc, utype(LUI, X_A0, hi20(val))); // lui a0,<hi20>
581     return;
582   case R_RISCV_TLSDESC_CALL:
583     if (isInt<12>(val))
584       write32le(loc, itype(ADDI, X_A0, 0, val)); // addi a0,zero,<lo12>
585     else
586       write32le(loc, itype(ADDI, X_A0, X_A0, lo12(val))); // addi a0,a0,<lo12>
587     return;
588   default:
589     llvm_unreachable("unsupported relocation for TLSDESC to LE");
590   }
591 }
592 
relocateAlloc(InputSectionBase & sec,uint8_t * buf) const593 void RISCV::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const {
594   uint64_t secAddr = sec.getOutputSection()->addr;
595   if (auto *s = dyn_cast<InputSection>(&sec))
596     secAddr += s->outSecOff;
597   else if (auto *ehIn = dyn_cast<EhInputSection>(&sec))
598     secAddr += ehIn->getParent()->outSecOff;
599   uint64_t tlsdescVal = 0;
600   bool tlsdescRelax = false, isToLe = false;
601   const ArrayRef<Relocation> relocs = sec.relocs();
602   for (size_t i = 0, size = relocs.size(); i != size; ++i) {
603     const Relocation &rel = relocs[i];
604     uint8_t *loc = buf + rel.offset;
605     uint64_t val =
606         sec.getRelocTargetVA(sec.file, rel.type, rel.addend,
607                              secAddr + rel.offset, *rel.sym, rel.expr);
608 
609     switch (rel.expr) {
610     case R_RELAX_HINT:
611       continue;
612     case R_TLSDESC_PC:
613       // For R_RISCV_TLSDESC_HI20, store &got(sym)-PC to be used by the
614       // following two instructions L[DW] and ADDI.
615       if (rel.type == R_RISCV_TLSDESC_HI20)
616         tlsdescVal = val;
617       else
618         val = tlsdescVal;
619       break;
620     case R_RELAX_TLS_GD_TO_IE:
621       // Only R_RISCV_TLSDESC_HI20 reaches here. tlsdescVal will be finalized
622       // after we see R_RISCV_TLSDESC_ADD_LO12 in the R_RELAX_TLS_GD_TO_LE case.
623       // The net effect is that tlsdescVal will be smaller than `val` to take
624       // into account of NOP instructions (in the absence of R_RISCV_RELAX)
625       // before AUIPC.
626       tlsdescVal = val + rel.offset;
627       isToLe = false;
628       tlsdescRelax = relaxable(relocs, i);
629       if (!tlsdescRelax)
630         tlsdescToIe(loc, rel, val);
631       continue;
632     case R_RELAX_TLS_GD_TO_LE:
633       // See the comment in handleTlsRelocation. For TLSDESC=>IE,
634       // R_RISCV_TLSDESC_{LOAD_LO12,ADD_LO12,CALL} also reach here. If isToLe is
635       // false, this is actually TLSDESC=>IE optimization.
636       if (rel.type == R_RISCV_TLSDESC_HI20) {
637         tlsdescVal = val;
638         isToLe = true;
639         tlsdescRelax = relaxable(relocs, i);
640       } else {
641         if (!isToLe && rel.type == R_RISCV_TLSDESC_ADD_LO12)
642           tlsdescVal -= rel.offset;
643         val = tlsdescVal;
644       }
645       // When NOP conversion is eligible and relaxation applies, don't write a
646       // NOP in case an unrelated instruction follows the current instruction.
647       if (tlsdescRelax &&
648           (rel.type == R_RISCV_TLSDESC_HI20 ||
649            rel.type == R_RISCV_TLSDESC_LOAD_LO12 ||
650            (rel.type == R_RISCV_TLSDESC_ADD_LO12 && isToLe && !hi20(val))))
651         continue;
652       if (isToLe)
653         tlsdescToLe(loc, rel, val);
654       else
655         tlsdescToIe(loc, rel, val);
656       continue;
657     case R_RISCV_LEB128:
658       if (i + 1 < size) {
659         const Relocation &rel1 = relocs[i + 1];
660         if (rel.type == R_RISCV_SET_ULEB128 &&
661             rel1.type == R_RISCV_SUB_ULEB128 && rel.offset == rel1.offset) {
662           auto val = rel.sym->getVA(rel.addend) - rel1.sym->getVA(rel1.addend);
663           if (overwriteULEB128(loc, val) >= 0x80)
664             errorOrWarn(sec.getLocation(rel.offset) + ": ULEB128 value " +
665                         Twine(val) + " exceeds available space; references '" +
666                         lld::toString(*rel.sym) + "'");
667           ++i;
668           continue;
669         }
670       }
671       errorOrWarn(sec.getLocation(rel.offset) +
672                   ": R_RISCV_SET_ULEB128 not paired with R_RISCV_SUB_SET128");
673       return;
674     default:
675       break;
676     }
677     relocate(loc, rel, val);
678   }
679 }
680 
initSymbolAnchors()681 void elf::initSymbolAnchors() {
682   SmallVector<InputSection *, 0> storage;
683   for (OutputSection *osec : outputSections) {
684     if (!(osec->flags & SHF_EXECINSTR))
685       continue;
686     for (InputSection *sec : getInputSections(*osec, storage)) {
687       sec->relaxAux = make<RelaxAux>();
688       if (sec->relocs().size()) {
689         sec->relaxAux->relocDeltas =
690             std::make_unique<uint32_t[]>(sec->relocs().size());
691         sec->relaxAux->relocTypes =
692             std::make_unique<RelType[]>(sec->relocs().size());
693       }
694     }
695   }
696   // Store anchors (st_value and st_value+st_size) for symbols relative to text
697   // sections.
698   //
699   // For a defined symbol foo, we may have `d->file != file` with --wrap=foo.
700   // We should process foo, as the defining object file's symbol table may not
701   // contain foo after redirectSymbols changed the foo entry to __wrap_foo. To
702   // avoid adding a Defined that is undefined in one object file, use
703   // `!d->scriptDefined` to exclude symbols that are definitely not wrapped.
704   //
705   // `relaxAux->anchors` may contain duplicate symbols, but that is fine.
706   for (InputFile *file : ctx.objectFiles)
707     for (Symbol *sym : file->getSymbols()) {
708       auto *d = dyn_cast<Defined>(sym);
709       if (!d || (d->file != file && !d->scriptDefined))
710         continue;
711       if (auto *sec = dyn_cast_or_null<InputSection>(d->section))
712         if (sec->flags & SHF_EXECINSTR && sec->relaxAux) {
713           // If sec is discarded, relaxAux will be nullptr.
714           sec->relaxAux->anchors.push_back({d->value, d, false});
715           sec->relaxAux->anchors.push_back({d->value + d->size, d, true});
716         }
717     }
718   // Sort anchors by offset so that we can find the closest relocation
719   // efficiently. For a zero size symbol, ensure that its start anchor precedes
720   // its end anchor. For two symbols with anchors at the same offset, their
721   // order does not matter.
722   for (OutputSection *osec : outputSections) {
723     if (!(osec->flags & SHF_EXECINSTR))
724       continue;
725     for (InputSection *sec : getInputSections(*osec, storage)) {
726       llvm::sort(sec->relaxAux->anchors, [](auto &a, auto &b) {
727         return std::make_pair(a.offset, a.end) <
728                std::make_pair(b.offset, b.end);
729       });
730     }
731   }
732 }
733 
734 // Relax R_RISCV_CALL/R_RISCV_CALL_PLT auipc+jalr to c.j, c.jal, or jal.
relaxCall(const InputSection & sec,size_t i,uint64_t loc,Relocation & r,uint32_t & remove)735 static void relaxCall(const InputSection &sec, size_t i, uint64_t loc,
736                       Relocation &r, uint32_t &remove) {
737   const bool rvc = getEFlags(sec.file) & EF_RISCV_RVC;
738   const Symbol &sym = *r.sym;
739   const uint64_t insnPair = read64le(sec.content().data() + r.offset);
740   const uint32_t rd = extractBits(insnPair, 32 + 11, 32 + 7);
741   const uint64_t dest =
742       (r.expr == R_PLT_PC ? sym.getPltVA() : sym.getVA()) + r.addend;
743   const int64_t displace = dest - loc;
744 
745   if (rvc && isInt<12>(displace) && rd == 0) {
746     sec.relaxAux->relocTypes[i] = R_RISCV_RVC_JUMP;
747     sec.relaxAux->writes.push_back(0xa001); // c.j
748     remove = 6;
749   } else if (rvc && isInt<12>(displace) && rd == X_RA &&
750              !config->is64) { // RV32C only
751     sec.relaxAux->relocTypes[i] = R_RISCV_RVC_JUMP;
752     sec.relaxAux->writes.push_back(0x2001); // c.jal
753     remove = 6;
754   } else if (isInt<21>(displace)) {
755     sec.relaxAux->relocTypes[i] = R_RISCV_JAL;
756     sec.relaxAux->writes.push_back(0x6f | rd << 7); // jal
757     remove = 4;
758   }
759 }
760 
761 // Relax local-exec TLS when hi20 is zero.
relaxTlsLe(const InputSection & sec,size_t i,uint64_t loc,Relocation & r,uint32_t & remove)762 static void relaxTlsLe(const InputSection &sec, size_t i, uint64_t loc,
763                        Relocation &r, uint32_t &remove) {
764   uint64_t val = r.sym->getVA(r.addend);
765   if (hi20(val) != 0)
766     return;
767   uint32_t insn = read32le(sec.content().data() + r.offset);
768   switch (r.type) {
769   case R_RISCV_TPREL_HI20:
770   case R_RISCV_TPREL_ADD:
771     // Remove lui rd, %tprel_hi(x) and add rd, rd, tp, %tprel_add(x).
772     sec.relaxAux->relocTypes[i] = R_RISCV_RELAX;
773     remove = 4;
774     break;
775   case R_RISCV_TPREL_LO12_I:
776     // addi rd, rd, %tprel_lo(x) => addi rd, tp, st_value(x)
777     sec.relaxAux->relocTypes[i] = R_RISCV_32;
778     insn = (insn & ~(31 << 15)) | (X_TP << 15);
779     sec.relaxAux->writes.push_back(setLO12_I(insn, val));
780     break;
781   case R_RISCV_TPREL_LO12_S:
782     // sw rs, %tprel_lo(x)(rd) => sw rs, st_value(x)(rd)
783     sec.relaxAux->relocTypes[i] = R_RISCV_32;
784     insn = (insn & ~(31 << 15)) | (X_TP << 15);
785     sec.relaxAux->writes.push_back(setLO12_S(insn, val));
786     break;
787   }
788 }
789 
relaxHi20Lo12(const InputSection & sec,size_t i,uint64_t loc,Relocation & r,uint32_t & remove)790 static void relaxHi20Lo12(const InputSection &sec, size_t i, uint64_t loc,
791                           Relocation &r, uint32_t &remove) {
792   const Defined *gp = ElfSym::riscvGlobalPointer;
793   if (!gp)
794     return;
795 
796   if (!isInt<12>(r.sym->getVA(r.addend) - gp->getVA()))
797     return;
798 
799   switch (r.type) {
800   case R_RISCV_HI20:
801     // Remove lui rd, %hi20(x).
802     sec.relaxAux->relocTypes[i] = R_RISCV_RELAX;
803     remove = 4;
804     break;
805   case R_RISCV_LO12_I:
806     sec.relaxAux->relocTypes[i] = INTERNAL_R_RISCV_GPREL_I;
807     break;
808   case R_RISCV_LO12_S:
809     sec.relaxAux->relocTypes[i] = INTERNAL_R_RISCV_GPREL_S;
810     break;
811   }
812 }
813 
relax(InputSection & sec)814 static bool relax(InputSection &sec) {
815   const uint64_t secAddr = sec.getVA();
816   const MutableArrayRef<Relocation> relocs = sec.relocs();
817   auto &aux = *sec.relaxAux;
818   bool changed = false;
819   ArrayRef<SymbolAnchor> sa = ArrayRef(aux.anchors);
820   uint64_t delta = 0;
821   bool tlsdescRelax = false, toLeShortForm = false;
822 
823   std::fill_n(aux.relocTypes.get(), relocs.size(), R_RISCV_NONE);
824   aux.writes.clear();
825   for (auto [i, r] : llvm::enumerate(relocs)) {
826     const uint64_t loc = secAddr + r.offset - delta;
827     uint32_t &cur = aux.relocDeltas[i], remove = 0;
828     switch (r.type) {
829     case R_RISCV_ALIGN: {
830       const uint64_t nextLoc = loc + r.addend;
831       const uint64_t align = PowerOf2Ceil(r.addend + 2);
832       // All bytes beyond the alignment boundary should be removed.
833       remove = nextLoc - ((loc + align - 1) & -align);
834       // If we can't satisfy this alignment, we've found a bad input.
835       if (LLVM_UNLIKELY(static_cast<int32_t>(remove) < 0)) {
836         errorOrWarn(getErrorLocation((const uint8_t*)loc) +
837                     "insufficient padding bytes for " + lld::toString(r.type) +
838                     ": " + Twine(r.addend) + " bytes available "
839                     "for requested alignment of " + Twine(align) + " bytes");
840         remove = 0;
841       }
842       break;
843     }
844     case R_RISCV_CALL:
845     case R_RISCV_CALL_PLT:
846       if (relaxable(relocs, i))
847         relaxCall(sec, i, loc, r, remove);
848       break;
849     case R_RISCV_TPREL_HI20:
850     case R_RISCV_TPREL_ADD:
851     case R_RISCV_TPREL_LO12_I:
852     case R_RISCV_TPREL_LO12_S:
853       if (relaxable(relocs, i))
854         relaxTlsLe(sec, i, loc, r, remove);
855       break;
856     case R_RISCV_HI20:
857     case R_RISCV_LO12_I:
858     case R_RISCV_LO12_S:
859       if (relaxable(relocs, i))
860         relaxHi20Lo12(sec, i, loc, r, remove);
861       break;
862     case R_RISCV_TLSDESC_HI20:
863       // For TLSDESC=>LE, we can use the short form if hi20 is zero.
864       tlsdescRelax = relaxable(relocs, i);
865       toLeShortForm = tlsdescRelax && r.expr == R_RELAX_TLS_GD_TO_LE &&
866                       !hi20(r.sym->getVA(r.addend));
867       [[fallthrough]];
868     case R_RISCV_TLSDESC_LOAD_LO12:
869       // For TLSDESC=>LE/IE, AUIPC and L[DW] are removed if relaxable.
870       if (tlsdescRelax && r.expr != R_TLSDESC_PC)
871         remove = 4;
872       break;
873     case R_RISCV_TLSDESC_ADD_LO12:
874       if (toLeShortForm)
875         remove = 4;
876       break;
877     }
878 
879     // For all anchors whose offsets are <= r.offset, they are preceded by
880     // the previous relocation whose `relocDeltas` value equals `delta`.
881     // Decrease their st_value and update their st_size.
882     for (; sa.size() && sa[0].offset <= r.offset; sa = sa.slice(1)) {
883       if (sa[0].end)
884         sa[0].d->size = sa[0].offset - delta - sa[0].d->value;
885       else
886         sa[0].d->value = sa[0].offset - delta;
887     }
888     delta += remove;
889     if (delta != cur) {
890       cur = delta;
891       changed = true;
892     }
893   }
894 
895   for (const SymbolAnchor &a : sa) {
896     if (a.end)
897       a.d->size = a.offset - delta - a.d->value;
898     else
899       a.d->value = a.offset - delta;
900   }
901   // Inform assignAddresses that the size has changed.
902   if (!isUInt<32>(delta))
903     fatal("section size decrease is too large: " + Twine(delta));
904   sec.bytesDropped = delta;
905   return changed;
906 }
907 
908 // When relaxing just R_RISCV_ALIGN, relocDeltas is usually changed only once in
909 // the absence of a linker script. For call and load/store R_RISCV_RELAX, code
910 // shrinkage may reduce displacement and make more relocations eligible for
911 // relaxation. Code shrinkage may increase displacement to a call/load/store
912 // target at a higher fixed address, invalidating an earlier relaxation. Any
913 // change in section sizes can have cascading effect and require another
914 // relaxation pass.
relaxOnce(int pass) const915 bool RISCV::relaxOnce(int pass) const {
916   llvm::TimeTraceScope timeScope("RISC-V relaxOnce");
917   if (config->relocatable)
918     return false;
919 
920   if (pass == 0)
921     initSymbolAnchors();
922 
923   SmallVector<InputSection *, 0> storage;
924   bool changed = false;
925   for (OutputSection *osec : outputSections) {
926     if (!(osec->flags & SHF_EXECINSTR))
927       continue;
928     for (InputSection *sec : getInputSections(*osec, storage))
929       changed |= relax(*sec);
930   }
931   return changed;
932 }
933 
finalizeRelax(int passes) const934 void RISCV::finalizeRelax(int passes) const {
935   llvm::TimeTraceScope timeScope("Finalize RISC-V relaxation");
936   log("relaxation passes: " + Twine(passes));
937   SmallVector<InputSection *, 0> storage;
938   for (OutputSection *osec : outputSections) {
939     if (!(osec->flags & SHF_EXECINSTR))
940       continue;
941     for (InputSection *sec : getInputSections(*osec, storage)) {
942       RelaxAux &aux = *sec->relaxAux;
943       if (!aux.relocDeltas)
944         continue;
945 
946       MutableArrayRef<Relocation> rels = sec->relocs();
947       ArrayRef<uint8_t> old = sec->content();
948       size_t newSize = old.size() - aux.relocDeltas[rels.size() - 1];
949       size_t writesIdx = 0;
950       uint8_t *p = context().bAlloc.Allocate<uint8_t>(newSize);
951       uint64_t offset = 0;
952       int64_t delta = 0;
953       sec->content_ = p;
954       sec->size = newSize;
955       sec->bytesDropped = 0;
956 
957       // Update section content: remove NOPs for R_RISCV_ALIGN and rewrite
958       // instructions for relaxed relocations.
959       for (size_t i = 0, e = rels.size(); i != e; ++i) {
960         uint32_t remove = aux.relocDeltas[i] - delta;
961         delta = aux.relocDeltas[i];
962         if (remove == 0 && aux.relocTypes[i] == R_RISCV_NONE)
963           continue;
964 
965         // Copy from last location to the current relocated location.
966         const Relocation &r = rels[i];
967         uint64_t size = r.offset - offset;
968         memcpy(p, old.data() + offset, size);
969         p += size;
970 
971         // For R_RISCV_ALIGN, we will place `offset` in a location (among NOPs)
972         // to satisfy the alignment requirement. If both `remove` and r.addend
973         // are multiples of 4, it is as if we have skipped some NOPs. Otherwise
974         // we are in the middle of a 4-byte NOP, and we need to rewrite the NOP
975         // sequence.
976         int64_t skip = 0;
977         if (r.type == R_RISCV_ALIGN) {
978           if (remove % 4 || r.addend % 4) {
979             skip = r.addend - remove;
980             int64_t j = 0;
981             for (; j + 4 <= skip; j += 4)
982               write32le(p + j, 0x00000013); // nop
983             if (j != skip) {
984               assert(j + 2 == skip);
985               write16le(p + j, 0x0001); // c.nop
986             }
987           }
988         } else if (RelType newType = aux.relocTypes[i]) {
989           switch (newType) {
990           case INTERNAL_R_RISCV_GPREL_I:
991           case INTERNAL_R_RISCV_GPREL_S:
992             break;
993           case R_RISCV_RELAX:
994             // Used by relaxTlsLe to indicate the relocation is ignored.
995             break;
996           case R_RISCV_RVC_JUMP:
997             skip = 2;
998             write16le(p, aux.writes[writesIdx++]);
999             break;
1000           case R_RISCV_JAL:
1001             skip = 4;
1002             write32le(p, aux.writes[writesIdx++]);
1003             break;
1004           case R_RISCV_32:
1005             // Used by relaxTlsLe to write a uint32_t then suppress the handling
1006             // in relocateAlloc.
1007             skip = 4;
1008             write32le(p, aux.writes[writesIdx++]);
1009             aux.relocTypes[i] = R_RISCV_NONE;
1010             break;
1011           default:
1012             llvm_unreachable("unsupported type");
1013           }
1014         }
1015 
1016         p += skip;
1017         offset = r.offset + skip + remove;
1018       }
1019       memcpy(p, old.data() + offset, old.size() - offset);
1020 
1021       // Subtract the previous relocDeltas value from the relocation offset.
1022       // For a pair of R_RISCV_CALL/R_RISCV_RELAX with the same offset, decrease
1023       // their r_offset by the same delta.
1024       delta = 0;
1025       for (size_t i = 0, e = rels.size(); i != e;) {
1026         uint64_t cur = rels[i].offset;
1027         do {
1028           rels[i].offset -= delta;
1029           if (aux.relocTypes[i] != R_RISCV_NONE)
1030             rels[i].type = aux.relocTypes[i];
1031         } while (++i != e && rels[i].offset == cur);
1032         delta = aux.relocDeltas[i - 1];
1033       }
1034     }
1035   }
1036 }
1037 
1038 namespace {
1039 // Representation of the merged .riscv.attributes input sections. The psABI
1040 // specifies merge policy for attributes. E.g. if we link an object without an
1041 // extension with an object with the extension, the output Tag_RISCV_arch shall
1042 // contain the extension. Some tools like objdump parse .riscv.attributes and
1043 // disabling some instructions if the first Tag_RISCV_arch does not contain an
1044 // extension.
1045 class RISCVAttributesSection final : public SyntheticSection {
1046 public:
RISCVAttributesSection()1047   RISCVAttributesSection()
1048       : SyntheticSection(0, SHT_RISCV_ATTRIBUTES, 1, ".riscv.attributes") {}
1049 
getSize() const1050   size_t getSize() const override { return size; }
1051   void writeTo(uint8_t *buf) override;
1052 
1053   static constexpr StringRef vendor = "riscv";
1054   DenseMap<unsigned, unsigned> intAttr;
1055   DenseMap<unsigned, StringRef> strAttr;
1056   size_t size = 0;
1057 };
1058 } // namespace
1059 
mergeArch(RISCVISAUtils::OrderedExtensionMap & mergedExts,unsigned & mergedXlen,const InputSectionBase * sec,StringRef s)1060 static void mergeArch(RISCVISAUtils::OrderedExtensionMap &mergedExts,
1061                       unsigned &mergedXlen, const InputSectionBase *sec,
1062                       StringRef s) {
1063   auto maybeInfo = RISCVISAInfo::parseNormalizedArchString(s);
1064   if (!maybeInfo) {
1065     errorOrWarn(toString(sec) + ": " + s + ": " +
1066                 llvm::toString(maybeInfo.takeError()));
1067     return;
1068   }
1069 
1070   // Merge extensions.
1071   RISCVISAInfo &info = **maybeInfo;
1072   if (mergedExts.empty()) {
1073     mergedExts = info.getExtensions();
1074     mergedXlen = info.getXLen();
1075   } else {
1076     for (const auto &ext : info.getExtensions()) {
1077       auto p = mergedExts.insert(ext);
1078       if (!p.second) {
1079         if (std::tie(p.first->second.Major, p.first->second.Minor) <
1080             std::tie(ext.second.Major, ext.second.Minor))
1081           p.first->second = ext.second;
1082       }
1083     }
1084   }
1085 }
1086 
mergeAtomic(DenseMap<unsigned,unsigned>::iterator it,const InputSectionBase * oldSection,const InputSectionBase * newSection,RISCVAttrs::RISCVAtomicAbiTag oldTag,RISCVAttrs::RISCVAtomicAbiTag newTag)1087 static void mergeAtomic(DenseMap<unsigned, unsigned>::iterator it,
1088                         const InputSectionBase *oldSection,
1089                         const InputSectionBase *newSection,
1090                         RISCVAttrs::RISCVAtomicAbiTag oldTag,
1091                         RISCVAttrs::RISCVAtomicAbiTag newTag) {
1092   using RISCVAttrs::RISCVAtomicAbiTag;
1093   // Same tags stay the same, and UNKNOWN is compatible with anything
1094   if (oldTag == newTag || newTag == RISCVAtomicAbiTag::UNKNOWN)
1095     return;
1096 
1097   auto reportAbiError = [&]() {
1098     errorOrWarn("atomic abi mismatch for " + oldSection->name + "\n>>> " +
1099                 toString(oldSection) +
1100                 ": atomic_abi=" + Twine(static_cast<unsigned>(oldTag)) +
1101                 "\n>>> " + toString(newSection) +
1102                 ": atomic_abi=" + Twine(static_cast<unsigned>(newTag)));
1103   };
1104 
1105   auto reportUnknownAbiError = [](const InputSectionBase *section,
1106                                   RISCVAtomicAbiTag tag) {
1107     switch (tag) {
1108     case RISCVAtomicAbiTag::UNKNOWN:
1109     case RISCVAtomicAbiTag::A6C:
1110     case RISCVAtomicAbiTag::A6S:
1111     case RISCVAtomicAbiTag::A7:
1112       return;
1113     };
1114     errorOrWarn("unknown atomic abi for " + section->name + "\n>>> " +
1115                 toString(section) +
1116                 ": atomic_abi=" + Twine(static_cast<unsigned>(tag)));
1117   };
1118   switch (oldTag) {
1119   case RISCVAtomicAbiTag::UNKNOWN:
1120     it->getSecond() = static_cast<unsigned>(newTag);
1121     return;
1122   case RISCVAtomicAbiTag::A6C:
1123     switch (newTag) {
1124     case RISCVAtomicAbiTag::A6S:
1125       it->getSecond() = static_cast<unsigned>(RISCVAtomicAbiTag::A6C);
1126       return;
1127     case RISCVAtomicAbiTag::A7:
1128       reportAbiError();
1129       return;
1130     case RISCVAttrs::RISCVAtomicAbiTag::UNKNOWN:
1131     case RISCVAttrs::RISCVAtomicAbiTag::A6C:
1132       return;
1133     };
1134     break;
1135 
1136   case RISCVAtomicAbiTag::A6S:
1137     switch (newTag) {
1138     case RISCVAtomicAbiTag::A6C:
1139       it->getSecond() = static_cast<unsigned>(RISCVAtomicAbiTag::A6C);
1140       return;
1141     case RISCVAtomicAbiTag::A7:
1142       it->getSecond() = static_cast<unsigned>(RISCVAtomicAbiTag::A7);
1143       return;
1144     case RISCVAttrs::RISCVAtomicAbiTag::UNKNOWN:
1145     case RISCVAttrs::RISCVAtomicAbiTag::A6S:
1146       return;
1147     };
1148     break;
1149 
1150   case RISCVAtomicAbiTag::A7:
1151     switch (newTag) {
1152     case RISCVAtomicAbiTag::A6S:
1153       it->getSecond() = static_cast<unsigned>(RISCVAtomicAbiTag::A7);
1154       return;
1155     case RISCVAtomicAbiTag::A6C:
1156       reportAbiError();
1157       return;
1158     case RISCVAttrs::RISCVAtomicAbiTag::UNKNOWN:
1159     case RISCVAttrs::RISCVAtomicAbiTag::A7:
1160       return;
1161     };
1162     break;
1163   };
1164 
1165   // If we get here, then we have an invalid tag, so report it.
1166   // Putting these checks at the end allows us to only do these checks when we
1167   // need to, since this is expected to be a rare occurrence.
1168   reportUnknownAbiError(oldSection, oldTag);
1169   reportUnknownAbiError(newSection, newTag);
1170 }
1171 
1172 static RISCVAttributesSection *
mergeAttributesSection(const SmallVector<InputSectionBase *,0> & sections)1173 mergeAttributesSection(const SmallVector<InputSectionBase *, 0> &sections) {
1174   using RISCVAttrs::RISCVAtomicAbiTag;
1175   RISCVISAUtils::OrderedExtensionMap exts;
1176   const InputSectionBase *firstStackAlign = nullptr;
1177   const InputSectionBase *firstAtomicAbi = nullptr;
1178   unsigned firstStackAlignValue = 0, xlen = 0;
1179   bool hasArch = false;
1180 
1181   in.riscvAttributes = std::make_unique<RISCVAttributesSection>();
1182   auto &merged = static_cast<RISCVAttributesSection &>(*in.riscvAttributes);
1183 
1184   // Collect all tags values from attributes section.
1185   const auto &attributesTags = RISCVAttrs::getRISCVAttributeTags();
1186   for (const InputSectionBase *sec : sections) {
1187     RISCVAttributeParser parser;
1188     if (Error e = parser.parse(sec->content(), llvm::endianness::little))
1189       warn(toString(sec) + ": " + llvm::toString(std::move(e)));
1190     for (const auto &tag : attributesTags) {
1191       switch (RISCVAttrs::AttrType(tag.attr)) {
1192         // Integer attributes.
1193       case RISCVAttrs::STACK_ALIGN:
1194         if (auto i = parser.getAttributeValue(tag.attr)) {
1195           auto r = merged.intAttr.try_emplace(tag.attr, *i);
1196           if (r.second) {
1197             firstStackAlign = sec;
1198             firstStackAlignValue = *i;
1199           } else if (r.first->second != *i) {
1200             errorOrWarn(toString(sec) + " has stack_align=" + Twine(*i) +
1201                         " but " + toString(firstStackAlign) +
1202                         " has stack_align=" + Twine(firstStackAlignValue));
1203           }
1204         }
1205         continue;
1206       case RISCVAttrs::UNALIGNED_ACCESS:
1207         if (auto i = parser.getAttributeValue(tag.attr))
1208           merged.intAttr[tag.attr] |= *i;
1209         continue;
1210 
1211         // String attributes.
1212       case RISCVAttrs::ARCH:
1213         if (auto s = parser.getAttributeString(tag.attr)) {
1214           hasArch = true;
1215           mergeArch(exts, xlen, sec, *s);
1216         }
1217         continue;
1218 
1219         // Attributes which use the default handling.
1220       case RISCVAttrs::PRIV_SPEC:
1221       case RISCVAttrs::PRIV_SPEC_MINOR:
1222       case RISCVAttrs::PRIV_SPEC_REVISION:
1223         break;
1224 
1225       case RISCVAttrs::AttrType::ATOMIC_ABI:
1226         if (auto i = parser.getAttributeValue(tag.attr)) {
1227           auto r = merged.intAttr.try_emplace(tag.attr, *i);
1228           if (r.second)
1229             firstAtomicAbi = sec;
1230           else
1231             mergeAtomic(r.first, firstAtomicAbi, sec,
1232                         static_cast<RISCVAtomicAbiTag>(r.first->getSecond()),
1233                         static_cast<RISCVAtomicAbiTag>(*i));
1234         }
1235         continue;
1236       }
1237 
1238       // Fallback for deprecated priv_spec* and other unknown attributes: retain
1239       // the attribute if all input sections agree on the value. GNU ld uses 0
1240       // and empty strings as default values which are not dumped to the output.
1241       // TODO Adjust after resolution to
1242       // https://github.com/riscv-non-isa/riscv-elf-psabi-doc/issues/352
1243       if (tag.attr % 2 == 0) {
1244         if (auto i = parser.getAttributeValue(tag.attr)) {
1245           auto r = merged.intAttr.try_emplace(tag.attr, *i);
1246           if (!r.second && r.first->second != *i)
1247             r.first->second = 0;
1248         }
1249       } else if (auto s = parser.getAttributeString(tag.attr)) {
1250         auto r = merged.strAttr.try_emplace(tag.attr, *s);
1251         if (!r.second && r.first->second != *s)
1252           r.first->second = {};
1253       }
1254     }
1255   }
1256 
1257   if (hasArch && xlen != 0) {
1258     if (auto result = RISCVISAInfo::createFromExtMap(xlen, exts)) {
1259       merged.strAttr.try_emplace(RISCVAttrs::ARCH,
1260                                  saver().save((*result)->toString()));
1261     } else {
1262       errorOrWarn(llvm::toString(result.takeError()));
1263     }
1264   }
1265 
1266   // The total size of headers: format-version [ <section-length> "vendor-name"
1267   // [ <file-tag> <size>.
1268   size_t size = 5 + merged.vendor.size() + 1 + 5;
1269   for (auto &attr : merged.intAttr)
1270     if (attr.second != 0)
1271       size += getULEB128Size(attr.first) + getULEB128Size(attr.second);
1272   for (auto &attr : merged.strAttr)
1273     if (!attr.second.empty())
1274       size += getULEB128Size(attr.first) + attr.second.size() + 1;
1275   merged.size = size;
1276   return &merged;
1277 }
1278 
writeTo(uint8_t * buf)1279 void RISCVAttributesSection::writeTo(uint8_t *buf) {
1280   const size_t size = getSize();
1281   uint8_t *const end = buf + size;
1282   *buf = ELFAttrs::Format_Version;
1283   write32(buf + 1, size - 1);
1284   buf += 5;
1285 
1286   memcpy(buf, vendor.data(), vendor.size());
1287   buf += vendor.size() + 1;
1288 
1289   *buf = ELFAttrs::File;
1290   write32(buf + 1, end - buf);
1291   buf += 5;
1292 
1293   for (auto &attr : intAttr) {
1294     if (attr.second == 0)
1295       continue;
1296     buf += encodeULEB128(attr.first, buf);
1297     buf += encodeULEB128(attr.second, buf);
1298   }
1299   for (auto &attr : strAttr) {
1300     if (attr.second.empty())
1301       continue;
1302     buf += encodeULEB128(attr.first, buf);
1303     memcpy(buf, attr.second.data(), attr.second.size());
1304     buf += attr.second.size() + 1;
1305   }
1306 }
1307 
mergeRISCVAttributesSections()1308 void elf::mergeRISCVAttributesSections() {
1309   // Find the first input SHT_RISCV_ATTRIBUTES; return if not found.
1310   size_t place =
1311       llvm::find_if(ctx.inputSections,
1312                     [](auto *s) { return s->type == SHT_RISCV_ATTRIBUTES; }) -
1313       ctx.inputSections.begin();
1314   if (place == ctx.inputSections.size())
1315     return;
1316 
1317   // Extract all SHT_RISCV_ATTRIBUTES sections into `sections`.
1318   SmallVector<InputSectionBase *, 0> sections;
1319   llvm::erase_if(ctx.inputSections, [&](InputSectionBase *s) {
1320     if (s->type != SHT_RISCV_ATTRIBUTES)
1321       return false;
1322     sections.push_back(s);
1323     return true;
1324   });
1325 
1326   // Add the merged section.
1327   ctx.inputSections.insert(ctx.inputSections.begin() + place,
1328                            mergeAttributesSection(sections));
1329 }
1330 
getRISCVTargetInfo()1331 TargetInfo *elf::getRISCVTargetInfo() {
1332   static RISCV target;
1333   return &target;
1334 }
1335