1//===- PPC.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 "OutputSections.h"
10#include "Symbols.h"
11#include "SyntheticSections.h"
12#include "Target.h"
13#include "Thunks.h"
14#include "lld/Common/ErrorHandler.h"
15#include "llvm/Support/Endian.h"
16
17using namespace llvm;
18using namespace llvm::support::endian;
19using namespace llvm::ELF;
20using namespace lld;
21using namespace lld::elf;
22
23// Undefine the macro predefined by GCC powerpc32.
24#undef PPC
25
26namespace {
27class PPC final : public TargetInfo {
28public:
29 PPC();
30 RelExpr getRelExpr(RelType type, const Symbol &s,
31 const uint8_t *loc) const override;
32 RelType getDynRel(RelType type) const override;
33 int64_t getImplicitAddend(const uint8_t *buf, RelType type) const override;
34 void writeGotHeader(uint8_t *buf) const override;
35 void writePltHeader(uint8_t *buf) const override {
36 llvm_unreachable("should call writePPC32GlinkSection() instead");
37 }
38 void writePlt(uint8_t *buf, const Symbol &sym,
39 uint64_t pltEntryAddr) const override {
40 llvm_unreachable("should call writePPC32GlinkSection() instead");
41 }
42 void writeIplt(uint8_t *buf, const Symbol &sym,
43 uint64_t pltEntryAddr) const override;
44 void writeGotPlt(uint8_t *buf, const Symbol &s) const override;
45 bool needsThunk(RelExpr expr, RelType relocType, const InputFile *file,
46 uint64_t branchAddr, const Symbol &s,
47 int64_t a) const override;
48 uint32_t getThunkSectionSpacing() const override;
49 bool inBranchRange(RelType type, uint64_t src, uint64_t dst) const override;
50 void relocate(uint8_t *loc, const Relocation &rel,
51 uint64_t val) const override;
52 RelExpr adjustTlsExpr(RelType type, RelExpr expr) const override;
53 int getTlsGdRelaxSkip(RelType type) const override;
54 void relocateAlloc(InputSectionBase &sec, uint8_t *buf) const override;
55
56private:
57 void relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
58 void relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
59 void relaxTlsLdToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
60 void relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
61};
62} // namespace
63
64static uint16_t lo(uint32_t v) { return v; }
65static uint16_t ha(uint32_t v) { return (v + 0x8000) >> 16; }
66
67static uint32_t readFromHalf16(const uint8_t *loc) {
68 return read32(p: config->isLE ? loc : loc - 2);
69}
70
71static void writeFromHalf16(uint8_t *loc, uint32_t insn) {
72 write32(p: config->isLE ? loc : loc - 2, v: insn);
73}
74
75void elf::writePPC32GlinkSection(uint8_t *buf, size_t numEntries) {
76 // Create canonical PLT entries for non-PIE code. Compilers don't generate
77 // non-GOT-non-PLT relocations referencing external functions for -fpie/-fPIE.
78 uint32_t glink = in.plt->getVA(); // VA of .glink
79 if (!config->isPic) {
80 for (const Symbol *sym : cast<PPC32GlinkSection>(Val&: *in.plt).canonical_plts) {
81 writePPC32PltCallStub(buf, gotPltVA: sym->getGotPltVA(), file: nullptr, addend: 0);
82 buf += 16;
83 glink += 16;
84 }
85 }
86
87 // On PPC Secure PLT ABI, bl foo@plt jumps to a call stub, which loads an
88 // absolute address from a specific .plt slot (usually called .got.plt on
89 // other targets) and jumps there.
90 //
91 // a) With immediate binding (BIND_NOW), the .plt entry is resolved at load
92 // time. The .glink section is not used.
93 // b) With lazy binding, the .plt entry points to a `b PLTresolve`
94 // instruction in .glink, filled in by PPC::writeGotPlt().
95
96 // Write N `b PLTresolve` first.
97 for (size_t i = 0; i != numEntries; ++i)
98 write32(p: buf + 4 * i, v: 0x48000000 | 4 * (numEntries - i));
99 buf += 4 * numEntries;
100
101 // Then write PLTresolve(), which has two forms: PIC and non-PIC. PLTresolve()
102 // computes the PLT index (by computing the distance from the landing b to
103 // itself) and calls _dl_runtime_resolve() (in glibc).
104 uint32_t got = in.got->getVA();
105 const uint8_t *end = buf + 64;
106 if (config->isPic) {
107 uint32_t afterBcl = 4 * in.plt->getNumEntries() + 12;
108 uint32_t gotBcl = got + 4 - (glink + afterBcl);
109 write32(p: buf + 0, v: 0x3d6b0000 | ha(v: afterBcl)); // addis r11,r11,1f-glink@ha
110 write32(p: buf + 4, v: 0x7c0802a6); // mflr r0
111 write32(p: buf + 8, v: 0x429f0005); // bcl 20,30,.+4
112 write32(p: buf + 12, v: 0x396b0000 | lo(v: afterBcl)); // 1: addi r11,r11,1b-glink@l
113 write32(p: buf + 16, v: 0x7d8802a6); // mflr r12
114 write32(p: buf + 20, v: 0x7c0803a6); // mtlr r0
115 write32(p: buf + 24, v: 0x7d6c5850); // sub r11,r11,r12
116 write32(p: buf + 28, v: 0x3d8c0000 | ha(v: gotBcl)); // addis 12,12,GOT+4-1b@ha
117 if (ha(v: gotBcl) == ha(v: gotBcl + 4)) {
118 write32(p: buf + 32, v: 0x800c0000 | lo(v: gotBcl)); // lwz r0,r12,GOT+4-1b@l(r12)
119 write32(p: buf + 36,
120 v: 0x818c0000 | lo(v: gotBcl + 4)); // lwz r12,r12,GOT+8-1b@l(r12)
121 } else {
122 write32(p: buf + 32, v: 0x840c0000 | lo(v: gotBcl)); // lwzu r0,r12,GOT+4-1b@l(r12)
123 write32(p: buf + 36, v: 0x818c0000 | 4); // lwz r12,r12,4(r12)
124 }
125 write32(p: buf + 40, v: 0x7c0903a6); // mtctr 0
126 write32(p: buf + 44, v: 0x7c0b5a14); // add r0,11,11
127 write32(p: buf + 48, v: 0x7d605a14); // add r11,0,11
128 write32(p: buf + 52, v: 0x4e800420); // bctr
129 buf += 56;
130 } else {
131 write32(p: buf + 0, v: 0x3d800000 | ha(v: got + 4)); // lis r12,GOT+4@ha
132 write32(p: buf + 4, v: 0x3d6b0000 | ha(v: -glink)); // addis r11,r11,-glink@ha
133 if (ha(v: got + 4) == ha(v: got + 8))
134 write32(p: buf + 8, v: 0x800c0000 | lo(v: got + 4)); // lwz r0,GOT+4@l(r12)
135 else
136 write32(p: buf + 8, v: 0x840c0000 | lo(v: got + 4)); // lwzu r0,GOT+4@l(r12)
137 write32(p: buf + 12, v: 0x396b0000 | lo(v: -glink)); // addi r11,r11,-glink@l
138 write32(p: buf + 16, v: 0x7c0903a6); // mtctr r0
139 write32(p: buf + 20, v: 0x7c0b5a14); // add r0,r11,r11
140 if (ha(v: got + 4) == ha(v: got + 8))
141 write32(p: buf + 24, v: 0x818c0000 | lo(v: got + 8)); // lwz r12,GOT+8@l(r12)
142 else
143 write32(p: buf + 24, v: 0x818c0000 | 4); // lwz r12,4(r12)
144 write32(p: buf + 28, v: 0x7d605a14); // add r11,r0,r11
145 write32(p: buf + 32, v: 0x4e800420); // bctr
146 buf += 36;
147 }
148
149 // Pad with nop. They should not be executed.
150 for (; buf < end; buf += 4)
151 write32(p: buf, v: 0x60000000);
152}
153
154PPC::PPC() {
155 copyRel = R_PPC_COPY;
156 gotRel = R_PPC_GLOB_DAT;
157 pltRel = R_PPC_JMP_SLOT;
158 relativeRel = R_PPC_RELATIVE;
159 iRelativeRel = R_PPC_IRELATIVE;
160 symbolicRel = R_PPC_ADDR32;
161 gotHeaderEntriesNum = 3;
162 gotPltHeaderEntriesNum = 0;
163 pltHeaderSize = 0;
164 pltEntrySize = 4;
165 ipltEntrySize = 16;
166
167 needsThunks = true;
168
169 tlsModuleIndexRel = R_PPC_DTPMOD32;
170 tlsOffsetRel = R_PPC_DTPREL32;
171 tlsGotRel = R_PPC_TPREL32;
172
173 defaultMaxPageSize = 65536;
174 defaultImageBase = 0x10000000;
175
176 write32(p: trapInstr.data(), v: 0x7fe00008);
177}
178
179void PPC::writeIplt(uint8_t *buf, const Symbol &sym,
180 uint64_t /*pltEntryAddr*/) const {
181 // In -pie or -shared mode, assume r30 points to .got2+0x8000, and use a
182 // .got2.plt_pic32. thunk.
183 writePPC32PltCallStub(buf, gotPltVA: sym.getGotPltVA(), file: sym.file, addend: 0x8000);
184}
185
186void PPC::writeGotHeader(uint8_t *buf) const {
187 // _GLOBAL_OFFSET_TABLE_[0] = _DYNAMIC
188 // glibc stores _dl_runtime_resolve in _GLOBAL_OFFSET_TABLE_[1],
189 // link_map in _GLOBAL_OFFSET_TABLE_[2].
190 write32(p: buf, v: mainPart->dynamic->getVA());
191}
192
193void PPC::writeGotPlt(uint8_t *buf, const Symbol &s) const {
194 // Address of the symbol resolver stub in .glink .
195 write32(p: buf, v: in.plt->getVA() + in.plt->headerSize + 4 * s.getPltIdx());
196}
197
198bool PPC::needsThunk(RelExpr expr, RelType type, const InputFile *file,
199 uint64_t branchAddr, const Symbol &s, int64_t a) const {
200 if (type != R_PPC_LOCAL24PC && type != R_PPC_REL24 && type != R_PPC_PLTREL24)
201 return false;
202 if (s.isInPlt())
203 return true;
204 if (s.isUndefWeak())
205 return false;
206 return !PPC::inBranchRange(type, src: branchAddr, dst: s.getVA(addend: a));
207}
208
209uint32_t PPC::getThunkSectionSpacing() const { return 0x2000000; }
210
211bool PPC::inBranchRange(RelType type, uint64_t src, uint64_t dst) const {
212 uint64_t offset = dst - src;
213 if (type == R_PPC_LOCAL24PC || type == R_PPC_REL24 || type == R_PPC_PLTREL24)
214 return isInt<26>(x: offset);
215 llvm_unreachable("unsupported relocation type used in branch");
216}
217
218RelExpr PPC::getRelExpr(RelType type, const Symbol &s,
219 const uint8_t *loc) const {
220 switch (type) {
221 case R_PPC_NONE:
222 return R_NONE;
223 case R_PPC_ADDR16_HA:
224 case R_PPC_ADDR16_HI:
225 case R_PPC_ADDR16_LO:
226 case R_PPC_ADDR24:
227 case R_PPC_ADDR32:
228 return R_ABS;
229 case R_PPC_DTPREL16:
230 case R_PPC_DTPREL16_HA:
231 case R_PPC_DTPREL16_HI:
232 case R_PPC_DTPREL16_LO:
233 case R_PPC_DTPREL32:
234 return R_DTPREL;
235 case R_PPC_REL14:
236 case R_PPC_REL32:
237 case R_PPC_REL16_LO:
238 case R_PPC_REL16_HI:
239 case R_PPC_REL16_HA:
240 return R_PC;
241 case R_PPC_GOT16:
242 return R_GOT_OFF;
243 case R_PPC_LOCAL24PC:
244 case R_PPC_REL24:
245 return R_PLT_PC;
246 case R_PPC_PLTREL24:
247 return R_PPC32_PLTREL;
248 case R_PPC_GOT_TLSGD16:
249 return R_TLSGD_GOT;
250 case R_PPC_GOT_TLSLD16:
251 return R_TLSLD_GOT;
252 case R_PPC_GOT_TPREL16:
253 return R_GOT_OFF;
254 case R_PPC_TLS:
255 return R_TLSIE_HINT;
256 case R_PPC_TLSGD:
257 return R_TLSDESC_CALL;
258 case R_PPC_TLSLD:
259 return R_TLSLD_HINT;
260 case R_PPC_TPREL16:
261 case R_PPC_TPREL16_HA:
262 case R_PPC_TPREL16_LO:
263 case R_PPC_TPREL16_HI:
264 return R_TPREL;
265 default:
266 error(msg: getErrorLocation(loc) + "unknown relocation (" + Twine(type) +
267 ") against symbol " + toString(s));
268 return R_NONE;
269 }
270}
271
272RelType PPC::getDynRel(RelType type) const {
273 if (type == R_PPC_ADDR32)
274 return type;
275 return R_PPC_NONE;
276}
277
278int64_t PPC::getImplicitAddend(const uint8_t *buf, RelType type) const {
279 switch (type) {
280 case R_PPC_NONE:
281 case R_PPC_GLOB_DAT:
282 case R_PPC_JMP_SLOT:
283 return 0;
284 case R_PPC_ADDR32:
285 case R_PPC_REL32:
286 case R_PPC_RELATIVE:
287 case R_PPC_IRELATIVE:
288 case R_PPC_DTPMOD32:
289 case R_PPC_DTPREL32:
290 case R_PPC_TPREL32:
291 return SignExtend64<32>(x: read32(p: buf));
292 default:
293 internalLinkerError(loc: getErrorLocation(loc: buf),
294 msg: "cannot read addend for relocation " + toString(type));
295 return 0;
296 }
297}
298
299static std::pair<RelType, uint64_t> fromDTPREL(RelType type, uint64_t val) {
300 uint64_t dtpBiasedVal = val - 0x8000;
301 switch (type) {
302 case R_PPC_DTPREL16:
303 return {R_PPC64_ADDR16, dtpBiasedVal};
304 case R_PPC_DTPREL16_HA:
305 return {R_PPC_ADDR16_HA, dtpBiasedVal};
306 case R_PPC_DTPREL16_HI:
307 return {R_PPC_ADDR16_HI, dtpBiasedVal};
308 case R_PPC_DTPREL16_LO:
309 return {R_PPC_ADDR16_LO, dtpBiasedVal};
310 case R_PPC_DTPREL32:
311 return {R_PPC_ADDR32, dtpBiasedVal};
312 default:
313 return {type, val};
314 }
315}
316
317void PPC::relocate(uint8_t *loc, const Relocation &rel, uint64_t val) const {
318 RelType newType;
319 std::tie(args&: newType, args&: val) = fromDTPREL(type: rel.type, val);
320 switch (newType) {
321 case R_PPC_ADDR16:
322 checkIntUInt(loc, v: val, n: 16, rel);
323 write16(p: loc, v: val);
324 break;
325 case R_PPC_GOT16:
326 case R_PPC_GOT_TLSGD16:
327 case R_PPC_GOT_TLSLD16:
328 case R_PPC_GOT_TPREL16:
329 case R_PPC_TPREL16:
330 checkInt(loc, v: val, n: 16, rel);
331 write16(p: loc, v: val);
332 break;
333 case R_PPC_ADDR16_HA:
334 case R_PPC_DTPREL16_HA:
335 case R_PPC_GOT_TLSGD16_HA:
336 case R_PPC_GOT_TLSLD16_HA:
337 case R_PPC_GOT_TPREL16_HA:
338 case R_PPC_REL16_HA:
339 case R_PPC_TPREL16_HA:
340 write16(p: loc, v: ha(v: val));
341 break;
342 case R_PPC_ADDR16_HI:
343 case R_PPC_DTPREL16_HI:
344 case R_PPC_GOT_TLSGD16_HI:
345 case R_PPC_GOT_TLSLD16_HI:
346 case R_PPC_GOT_TPREL16_HI:
347 case R_PPC_REL16_HI:
348 case R_PPC_TPREL16_HI:
349 write16(p: loc, v: val >> 16);
350 break;
351 case R_PPC_ADDR16_LO:
352 case R_PPC_DTPREL16_LO:
353 case R_PPC_GOT_TLSGD16_LO:
354 case R_PPC_GOT_TLSLD16_LO:
355 case R_PPC_GOT_TPREL16_LO:
356 case R_PPC_REL16_LO:
357 case R_PPC_TPREL16_LO:
358 write16(p: loc, v: val);
359 break;
360 case R_PPC_ADDR32:
361 case R_PPC_REL32:
362 write32(p: loc, v: val);
363 break;
364 case R_PPC_REL14: {
365 uint32_t mask = 0x0000FFFC;
366 checkInt(loc, v: val, n: 16, rel);
367 checkAlignment(loc, v: val, n: 4, rel);
368 write32(p: loc, v: (read32(p: loc) & ~mask) | (val & mask));
369 break;
370 }
371 case R_PPC_ADDR24:
372 case R_PPC_REL24:
373 case R_PPC_LOCAL24PC:
374 case R_PPC_PLTREL24: {
375 uint32_t mask = 0x03FFFFFC;
376 checkInt(loc, v: val, n: 26, rel);
377 checkAlignment(loc, v: val, n: 4, rel);
378 write32(p: loc, v: (read32(p: loc) & ~mask) | (val & mask));
379 break;
380 }
381 default:
382 llvm_unreachable("unknown relocation");
383 }
384}
385
386RelExpr PPC::adjustTlsExpr(RelType type, RelExpr expr) const {
387 if (expr == R_RELAX_TLS_GD_TO_IE)
388 return R_RELAX_TLS_GD_TO_IE_GOT_OFF;
389 if (expr == R_RELAX_TLS_LD_TO_LE)
390 return R_RELAX_TLS_LD_TO_LE_ABS;
391 return expr;
392}
393
394int PPC::getTlsGdRelaxSkip(RelType type) const {
395 // A __tls_get_addr call instruction is marked with 2 relocations:
396 //
397 // R_PPC_TLSGD / R_PPC_TLSLD: marker relocation
398 // R_PPC_REL24: __tls_get_addr
399 //
400 // After the relaxation we no longer call __tls_get_addr and should skip both
401 // relocations to not create a false dependence on __tls_get_addr being
402 // defined.
403 if (type == R_PPC_TLSGD || type == R_PPC_TLSLD)
404 return 2;
405 return 1;
406}
407
408void PPC::relaxTlsGdToIe(uint8_t *loc, const Relocation &rel,
409 uint64_t val) const {
410 switch (rel.type) {
411 case R_PPC_GOT_TLSGD16: {
412 // addi rT, rA, x@got@tlsgd --> lwz rT, x@got@tprel(rA)
413 uint32_t insn = readFromHalf16(loc);
414 writeFromHalf16(loc, insn: 0x80000000 | (insn & 0x03ff0000));
415 relocateNoSym(loc, type: R_PPC_GOT_TPREL16, val);
416 break;
417 }
418 case R_PPC_TLSGD:
419 // bl __tls_get_addr(x@tldgd) --> add r3, r3, r2
420 write32(p: loc, v: 0x7c631214);
421 break;
422 default:
423 llvm_unreachable("unsupported relocation for TLS GD to IE relaxation");
424 }
425}
426
427void PPC::relaxTlsGdToLe(uint8_t *loc, const Relocation &rel,
428 uint64_t val) const {
429 switch (rel.type) {
430 case R_PPC_GOT_TLSGD16:
431 // addi r3, r31, x@got@tlsgd --> addis r3, r2, x@tprel@ha
432 writeFromHalf16(loc, insn: 0x3c620000 | ha(v: val));
433 break;
434 case R_PPC_TLSGD:
435 // bl __tls_get_addr(x@tldgd) --> add r3, r3, x@tprel@l
436 write32(p: loc, v: 0x38630000 | lo(v: val));
437 break;
438 default:
439 llvm_unreachable("unsupported relocation for TLS GD to LE relaxation");
440 }
441}
442
443void PPC::relaxTlsLdToLe(uint8_t *loc, const Relocation &rel,
444 uint64_t val) const {
445 switch (rel.type) {
446 case R_PPC_GOT_TLSLD16:
447 // addi r3, rA, x@got@tlsgd --> addis r3, r2, 0
448 writeFromHalf16(loc, insn: 0x3c620000);
449 break;
450 case R_PPC_TLSLD:
451 // r3+x@dtprel computes r3+x-0x8000, while we want it to compute r3+x@tprel
452 // = r3+x-0x7000, so add 4096 to r3.
453 // bl __tls_get_addr(x@tlsld) --> addi r3, r3, 4096
454 write32(p: loc, v: 0x38631000);
455 break;
456 case R_PPC_DTPREL16:
457 case R_PPC_DTPREL16_HA:
458 case R_PPC_DTPREL16_HI:
459 case R_PPC_DTPREL16_LO:
460 relocate(loc, rel, val);
461 break;
462 default:
463 llvm_unreachable("unsupported relocation for TLS LD to LE relaxation");
464 }
465}
466
467void PPC::relaxTlsIeToLe(uint8_t *loc, const Relocation &rel,
468 uint64_t val) const {
469 switch (rel.type) {
470 case R_PPC_GOT_TPREL16: {
471 // lwz rT, x@got@tprel(rA) --> addis rT, r2, x@tprel@ha
472 uint32_t rt = readFromHalf16(loc) & 0x03e00000;
473 writeFromHalf16(loc, insn: 0x3c020000 | rt | ha(v: val));
474 break;
475 }
476 case R_PPC_TLS: {
477 uint32_t insn = read32(p: loc);
478 if (insn >> 26 != 31)
479 error(msg: "unrecognized instruction for IE to LE R_PPC_TLS");
480 // addi rT, rT, x@tls --> addi rT, rT, x@tprel@l
481 unsigned secondaryOp = (read32(p: loc) & 0x000007fe) >> 1;
482 uint32_t dFormOp = getPPCDFormOp(secondaryOp);
483 if (dFormOp == 0) { // Expecting a DS-Form instruction.
484 dFormOp = getPPCDSFormOp(secondaryOp);
485 if (dFormOp == 0)
486 error(msg: "unrecognized instruction for IE to LE R_PPC_TLS");
487 }
488 write32(p: loc, v: (dFormOp | (insn & 0x03ff0000) | lo(v: val)));
489 break;
490 }
491 default:
492 llvm_unreachable("unsupported relocation for TLS IE to LE relaxation");
493 }
494}
495
496void PPC::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const {
497 uint64_t secAddr = sec.getOutputSection()->addr;
498 if (auto *s = dyn_cast<InputSection>(Val: &sec))
499 secAddr += s->outSecOff;
500 for (const Relocation &rel : sec.relocs()) {
501 uint8_t *loc = buf + rel.offset;
502 const uint64_t val = SignExtend64(
503 X: sec.getRelocTargetVA(File: sec.file, Type: rel.type, A: rel.addend,
504 P: secAddr + rel.offset, Sym: *rel.sym, Expr: rel.expr),
505 B: 32);
506 switch (rel.expr) {
507 case R_RELAX_TLS_GD_TO_IE_GOT_OFF:
508 relaxTlsGdToIe(loc, rel, val);
509 break;
510 case R_RELAX_TLS_GD_TO_LE:
511 relaxTlsGdToLe(loc, rel, val);
512 break;
513 case R_RELAX_TLS_LD_TO_LE_ABS:
514 relaxTlsLdToLe(loc, rel, val);
515 break;
516 case R_RELAX_TLS_IE_TO_LE:
517 relaxTlsIeToLe(loc, rel, val);
518 break;
519 default:
520 relocate(loc, rel, val);
521 break;
522 }
523 }
524}
525
526TargetInfo *elf::getPPCTargetInfo() {
527 static PPC target;
528 return &target;
529}
530