1 | //===- PPC64.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 "SymbolTable.h" |
12 | #include "Symbols.h" |
13 | #include "SyntheticSections.h" |
14 | #include "Target.h" |
15 | #include "Thunks.h" |
16 | |
17 | using namespace llvm; |
18 | using namespace llvm::object; |
19 | using namespace llvm::support::endian; |
20 | using namespace llvm::ELF; |
21 | using namespace lld; |
22 | using namespace lld::elf; |
23 | |
24 | constexpr uint64_t ppc64TocOffset = 0x8000; |
25 | constexpr uint64_t dynamicThreadPointerOffset = 0x8000; |
26 | |
27 | namespace { |
28 | // The instruction encoding of bits 21-30 from the ISA for the Xform and Dform |
29 | // instructions that can be used as part of the initial exec TLS sequence. |
30 | enum XFormOpcd { |
31 | LBZX = 87, |
32 | LHZX = 279, |
33 | LWZX = 23, |
34 | LDX = 21, |
35 | STBX = 215, |
36 | STHX = 407, |
37 | STWX = 151, |
38 | STDX = 149, |
39 | LHAX = 343, |
40 | LWAX = 341, |
41 | LFSX = 535, |
42 | LFDX = 599, |
43 | STFSX = 663, |
44 | STFDX = 727, |
45 | ADD = 266, |
46 | }; |
47 | |
48 | enum DFormOpcd { |
49 | LBZ = 34, |
50 | LBZU = 35, |
51 | LHZ = 40, |
52 | LHZU = 41, |
53 | LHAU = 43, |
54 | LWZ = 32, |
55 | LWZU = 33, |
56 | LFSU = 49, |
57 | LFDU = 51, |
58 | STB = 38, |
59 | STBU = 39, |
60 | STH = 44, |
61 | STHU = 45, |
62 | STW = 36, |
63 | STWU = 37, |
64 | STFSU = 53, |
65 | STFDU = 55, |
66 | LHA = 42, |
67 | LFS = 48, |
68 | LFD = 50, |
69 | STFS = 52, |
70 | STFD = 54, |
71 | ADDI = 14 |
72 | }; |
73 | |
74 | enum DSFormOpcd { |
75 | LD = 58, |
76 | LWA = 58, |
77 | STD = 62 |
78 | }; |
79 | |
80 | constexpr uint32_t NOP = 0x60000000; |
81 | |
82 | enum class PPCLegacyInsn : uint32_t { |
83 | NOINSN = 0, |
84 | // Loads. |
85 | LBZ = 0x88000000, |
86 | LHZ = 0xa0000000, |
87 | LWZ = 0x80000000, |
88 | LHA = 0xa8000000, |
89 | LWA = 0xe8000002, |
90 | LD = 0xe8000000, |
91 | LFS = 0xC0000000, |
92 | LXSSP = 0xe4000003, |
93 | LFD = 0xc8000000, |
94 | LXSD = 0xe4000002, |
95 | LXV = 0xf4000001, |
96 | LXVP = 0x18000000, |
97 | |
98 | // Stores. |
99 | STB = 0x98000000, |
100 | STH = 0xb0000000, |
101 | STW = 0x90000000, |
102 | STD = 0xf8000000, |
103 | STFS = 0xd0000000, |
104 | STXSSP = 0xf4000003, |
105 | STFD = 0xd8000000, |
106 | STXSD = 0xf4000002, |
107 | STXV = 0xf4000005, |
108 | STXVP = 0x18000001 |
109 | }; |
110 | enum class PPCPrefixedInsn : uint64_t { |
111 | NOINSN = 0, |
112 | PREFIX_MLS = 0x0610000000000000, |
113 | PREFIX_8LS = 0x0410000000000000, |
114 | |
115 | // Loads. |
116 | PLBZ = PREFIX_MLS, |
117 | PLHZ = PREFIX_MLS, |
118 | PLWZ = PREFIX_MLS, |
119 | PLHA = PREFIX_MLS, |
120 | PLWA = PREFIX_8LS | 0xa4000000, |
121 | PLD = PREFIX_8LS | 0xe4000000, |
122 | PLFS = PREFIX_MLS, |
123 | PLXSSP = PREFIX_8LS | 0xac000000, |
124 | PLFD = PREFIX_MLS, |
125 | PLXSD = PREFIX_8LS | 0xa8000000, |
126 | PLXV = PREFIX_8LS | 0xc8000000, |
127 | PLXVP = PREFIX_8LS | 0xe8000000, |
128 | |
129 | // Stores. |
130 | PSTB = PREFIX_MLS, |
131 | PSTH = PREFIX_MLS, |
132 | PSTW = PREFIX_MLS, |
133 | PSTD = PREFIX_8LS | 0xf4000000, |
134 | PSTFS = PREFIX_MLS, |
135 | PSTXSSP = PREFIX_8LS | 0xbc000000, |
136 | PSTFD = PREFIX_MLS, |
137 | PSTXSD = PREFIX_8LS | 0xb8000000, |
138 | PSTXV = PREFIX_8LS | 0xd8000000, |
139 | PSTXVP = PREFIX_8LS | 0xf8000000 |
140 | }; |
141 | |
142 | static bool checkPPCLegacyInsn(uint32_t encoding) { |
143 | PPCLegacyInsn insn = static_cast<PPCLegacyInsn>(encoding); |
144 | if (insn == PPCLegacyInsn::NOINSN) |
145 | return false; |
146 | #define PCREL_OPT(Legacy, PCRel, InsnMask) \ |
147 | if (insn == PPCLegacyInsn::Legacy) \ |
148 | return true; |
149 | #include "PPCInsns.def" |
150 | #undef PCREL_OPT |
151 | return false; |
152 | } |
153 | |
154 | // Masks to apply to legacy instructions when converting them to prefixed, |
155 | // pc-relative versions. For the most part, the primary opcode is shared |
156 | // between the legacy instruction and the suffix of its prefixed version. |
157 | // However, there are some instances where that isn't the case (DS-Form and |
158 | // DQ-form instructions). |
159 | enum class LegacyToPrefixMask : uint64_t { |
160 | NOMASK = 0x0, |
161 | OPC_AND_RST = 0xffe00000, // Primary opc (0-5) and R[ST] (6-10). |
162 | ONLY_RST = 0x3e00000, // [RS]T (6-10). |
163 | ST_STX28_TO5 = |
164 | 0x8000000003e00000, // S/T (6-10) - The [S/T]X bit moves from 28 to 5. |
165 | }; |
166 | |
167 | class PPC64 final : public TargetInfo { |
168 | public: |
169 | PPC64(Ctx &); |
170 | int getTlsGdRelaxSkip(RelType type) const override; |
171 | uint32_t calcEFlags() const override; |
172 | RelExpr getRelExpr(RelType type, const Symbol &s, |
173 | const uint8_t *loc) const override; |
174 | RelType getDynRel(RelType type) const override; |
175 | int64_t getImplicitAddend(const uint8_t *buf, RelType type) const override; |
176 | void writePltHeader(uint8_t *buf) const override; |
177 | void writePlt(uint8_t *buf, const Symbol &sym, |
178 | uint64_t pltEntryAddr) const override; |
179 | void writeIplt(uint8_t *buf, const Symbol &sym, |
180 | uint64_t pltEntryAddr) const override; |
181 | void relocate(uint8_t *loc, const Relocation &rel, |
182 | uint64_t val) const override; |
183 | void writeGotHeader(uint8_t *buf) const override; |
184 | bool needsThunk(RelExpr expr, RelType type, const InputFile *file, |
185 | uint64_t branchAddr, const Symbol &s, |
186 | int64_t a) const override; |
187 | uint32_t getThunkSectionSpacing() const override; |
188 | bool inBranchRange(RelType type, uint64_t src, uint64_t dst) const override; |
189 | RelExpr adjustTlsExpr(RelType type, RelExpr expr) const override; |
190 | RelExpr adjustGotPcExpr(RelType type, int64_t addend, |
191 | const uint8_t *loc) const override; |
192 | void relaxGot(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
193 | void relocateAlloc(InputSectionBase &sec, uint8_t *buf) const override; |
194 | |
195 | bool adjustPrologueForCrossSplitStack(uint8_t *loc, uint8_t *end, |
196 | uint8_t stOther) const override; |
197 | |
198 | private: |
199 | void relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
200 | void relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
201 | void relaxTlsLdToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
202 | void relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
203 | }; |
204 | } // namespace |
205 | |
206 | uint64_t elf::getPPC64TocBase(Ctx &ctx) { |
207 | // The TOC consists of sections .got, .toc, .tocbss, .plt in that order. The |
208 | // TOC starts where the first of these sections starts. We always create a |
209 | // .got when we see a relocation that uses it, so for us the start is always |
210 | // the .got. |
211 | uint64_t tocVA = ctx.in.got->getVA(); |
212 | |
213 | // Per the ppc64-elf-linux ABI, The TOC base is TOC value plus 0x8000 |
214 | // thus permitting a full 64 Kbytes segment. Note that the glibc startup |
215 | // code (crt1.o) assumes that you can get from the TOC base to the |
216 | // start of the .toc section with only a single (signed) 16-bit relocation. |
217 | return tocVA + ppc64TocOffset; |
218 | } |
219 | |
220 | unsigned elf::getPPC64GlobalEntryToLocalEntryOffset(Ctx &ctx, uint8_t stOther) { |
221 | // The offset is encoded into the 3 most significant bits of the st_other |
222 | // field, with some special values described in section 3.4.1 of the ABI: |
223 | // 0 --> Zero offset between the GEP and LEP, and the function does NOT use |
224 | // the TOC pointer (r2). r2 will hold the same value on returning from |
225 | // the function as it did on entering the function. |
226 | // 1 --> Zero offset between the GEP and LEP, and r2 should be treated as a |
227 | // caller-saved register for all callers. |
228 | // 2-6 --> The binary logarithm of the offset eg: |
229 | // 2 --> 2^2 = 4 bytes --> 1 instruction. |
230 | // 6 --> 2^6 = 64 bytes --> 16 instructions. |
231 | // 7 --> Reserved. |
232 | uint8_t gepToLep = (stOther >> 5) & 7; |
233 | if (gepToLep < 2) |
234 | return 0; |
235 | |
236 | // The value encoded in the st_other bits is the |
237 | // log-base-2(offset). |
238 | if (gepToLep < 7) |
239 | return 1 << gepToLep; |
240 | |
241 | ErrAlways(ctx) |
242 | << "reserved value of 7 in the 3 most-significant-bits of st_other" ; |
243 | return 0; |
244 | } |
245 | |
246 | void elf::writePrefixedInst(Ctx &ctx, uint8_t *loc, uint64_t insn) { |
247 | insn = ctx.arg.isLE ? insn << 32 | insn >> 32 : insn; |
248 | write64(ctx, p: loc, v: insn); |
249 | } |
250 | |
251 | static bool addOptional(Ctx &ctx, StringRef name, uint64_t value, |
252 | std::vector<Defined *> &defined) { |
253 | Symbol *sym = ctx.symtab->find(name); |
254 | if (!sym || sym->isDefined()) |
255 | return false; |
256 | sym->resolve(ctx, other: Defined{ctx, ctx.internalFile, StringRef(), STB_GLOBAL, |
257 | STV_HIDDEN, STT_FUNC, value, |
258 | /*size=*/0, /*section=*/nullptr}); |
259 | defined.push_back(x: cast<Defined>(Val: sym)); |
260 | return true; |
261 | } |
262 | |
263 | // If from is 14, write ${prefix}14: firstInsn; ${prefix}15: |
264 | // firstInsn+0x200008; ...; ${prefix}31: firstInsn+(31-14)*0x200008; $tail |
265 | // The labels are defined only if they exist in the symbol table. |
266 | static void writeSequence(Ctx &ctx, const char *prefix, int from, |
267 | uint32_t firstInsn, ArrayRef<uint32_t> tail) { |
268 | std::vector<Defined *> defined; |
269 | char name[16]; |
270 | int first; |
271 | const size_t size = 32 - from + tail.size(); |
272 | MutableArrayRef<uint32_t> buf(ctx.bAlloc.Allocate<uint32_t>(Num: size), size); |
273 | uint32_t *ptr = buf.data(); |
274 | for (int r = from; r < 32; ++r) { |
275 | format(Fmt: "%s%d" , Vals: prefix, Vals: r).snprint(Buffer: name, BufferSize: sizeof(name)); |
276 | if (addOptional(ctx, name, value: 4 * (r - from), defined) && defined.size() == 1) |
277 | first = r - from; |
278 | write32(ctx, p: ptr++, v: firstInsn + 0x200008 * (r - from)); |
279 | } |
280 | for (uint32_t insn : tail) |
281 | write32(ctx, p: ptr++, v: insn); |
282 | assert(ptr == &*buf.end()); |
283 | |
284 | if (defined.empty()) |
285 | return; |
286 | // The full section content has the extent of [begin, end). We drop unused |
287 | // instructions and write [first,end). |
288 | auto *sec = make<InputSection>( |
289 | args&: ctx.internalFile, args: ".text" , args: SHT_PROGBITS, args: SHF_ALLOC, /*addralign=*/args: 4, |
290 | /*entsize=*/args: 0, |
291 | args: ArrayRef(reinterpret_cast<uint8_t *>(buf.data() + first), |
292 | 4 * (buf.size() - first))); |
293 | ctx.inputSections.push_back(Elt: sec); |
294 | for (Defined *sym : defined) { |
295 | sym->section = sec; |
296 | sym->value -= 4 * first; |
297 | } |
298 | } |
299 | |
300 | // Implements some save and restore functions as described by ELF V2 ABI to be |
301 | // compatible with GCC. With GCC -Os, when the number of call-saved registers |
302 | // exceeds a certain threshold, GCC generates _savegpr0_* _restgpr0_* calls and |
303 | // expects the linker to define them. See |
304 | // https://sourceware.org/pipermail/binutils/2002-February/017444.html and |
305 | // https://sourceware.org/pipermail/binutils/2004-August/036765.html . This is |
306 | // weird because libgcc.a would be the natural place. The linker generation |
307 | // approach has the advantage that the linker can generate multiple copies to |
308 | // avoid long branch thunks. However, we don't consider the advantage |
309 | // significant enough to complicate our trunk implementation, so we take the |
310 | // simple approach and synthesize .text sections providing the implementation. |
311 | void elf::addPPC64SaveRestore(Ctx &ctx) { |
312 | constexpr uint32_t blr = 0x4e800020, mtlr_0 = 0x7c0803a6; |
313 | |
314 | // _restgpr0_14: ld 14, -144(1); _restgpr0_15: ld 15, -136(1); ... |
315 | // Tail: ld 0, 16(1); mtlr 0; blr |
316 | writeSequence(ctx, prefix: "_restgpr0_" , from: 14, firstInsn: 0xe9c1ff70, tail: {0xe8010010, mtlr_0, blr}); |
317 | // _restgpr1_14: ld 14, -144(12); _restgpr1_15: ld 15, -136(12); ... |
318 | // Tail: blr |
319 | writeSequence(ctx, prefix: "_restgpr1_" , from: 14, firstInsn: 0xe9ccff70, tail: {blr}); |
320 | // _savegpr0_14: std 14, -144(1); _savegpr0_15: std 15, -136(1); ... |
321 | // Tail: std 0, 16(1); blr |
322 | writeSequence(ctx, prefix: "_savegpr0_" , from: 14, firstInsn: 0xf9c1ff70, tail: {0xf8010010, blr}); |
323 | // _savegpr1_14: std 14, -144(12); _savegpr1_15: std 15, -136(12); ... |
324 | // Tail: blr |
325 | writeSequence(ctx, prefix: "_savegpr1_" , from: 14, firstInsn: 0xf9ccff70, tail: {blr}); |
326 | } |
327 | |
328 | // Find the R_PPC64_ADDR64 in .rela.toc with matching offset. |
329 | template <typename ELFT> |
330 | static std::pair<Defined *, int64_t> |
331 | getRelaTocSymAndAddend(InputSectionBase *tocSec, uint64_t offset) { |
332 | // .rela.toc contains exclusively R_PPC64_ADDR64 relocations sorted by |
333 | // r_offset: 0, 8, 16, etc. For a given Offset, Offset / 8 gives us the |
334 | // relocation index in most cases. |
335 | // |
336 | // In rare cases a TOC entry may store a constant that doesn't need an |
337 | // R_PPC64_ADDR64, the corresponding r_offset is therefore missing. Offset / 8 |
338 | // points to a relocation with larger r_offset. Do a linear probe then. |
339 | // Constants are extremely uncommon in .toc and the extra number of array |
340 | // accesses can be seen as a small constant. |
341 | ArrayRef<typename ELFT::Rela> relas = |
342 | tocSec->template relsOrRelas<ELFT>().relas; |
343 | if (relas.empty()) |
344 | return {}; |
345 | uint64_t index = std::min<uint64_t>(offset / 8, relas.size() - 1); |
346 | for (;;) { |
347 | if (relas[index].r_offset == offset) { |
348 | Symbol &sym = tocSec->file->getRelocTargetSym(relas[index]); |
349 | return {dyn_cast<Defined>(Val: &sym), getAddend<ELFT>(relas[index])}; |
350 | } |
351 | if (relas[index].r_offset < offset || index == 0) |
352 | break; |
353 | --index; |
354 | } |
355 | return {}; |
356 | } |
357 | |
358 | // When accessing a symbol defined in another translation unit, compilers |
359 | // reserve a .toc entry, allocate a local label and generate toc-indirect |
360 | // instructions: |
361 | // |
362 | // addis 3, 2, .LC0@toc@ha # R_PPC64_TOC16_HA |
363 | // ld 3, .LC0@toc@l(3) # R_PPC64_TOC16_LO_DS, load the address from a .toc entry |
364 | // ld/lwa 3, 0(3) # load the value from the address |
365 | // |
366 | // .section .toc,"aw",@progbits |
367 | // .LC0: .tc var[TC],var |
368 | // |
369 | // If var is defined, non-preemptable and addressable with a 32-bit signed |
370 | // offset from the toc base, the address of var can be computed by adding an |
371 | // offset to the toc base, saving a load. |
372 | // |
373 | // addis 3,2,var@toc@ha # this may be relaxed to a nop, |
374 | // addi 3,3,var@toc@l # then this becomes addi 3,2,var@toc |
375 | // ld/lwa 3, 0(3) # load the value from the address |
376 | // |
377 | // Returns true if the relaxation is performed. |
378 | static bool tryRelaxPPC64TocIndirection(Ctx &ctx, const Relocation &rel, |
379 | uint8_t *bufLoc) { |
380 | assert(ctx.arg.tocOptimize); |
381 | if (rel.addend < 0) |
382 | return false; |
383 | |
384 | // If the symbol is not the .toc section, this isn't a toc-indirection. |
385 | Defined *defSym = dyn_cast<Defined>(Val: rel.sym); |
386 | if (!defSym || !defSym->isSection() || defSym->section->name != ".toc" ) |
387 | return false; |
388 | |
389 | Defined *d; |
390 | int64_t addend; |
391 | auto *tocISB = cast<InputSectionBase>(Val: defSym->section); |
392 | std::tie(args&: d, args&: addend) = |
393 | ctx.arg.isLE ? getRelaTocSymAndAddend<ELF64LE>(tocSec: tocISB, offset: rel.addend) |
394 | : getRelaTocSymAndAddend<ELF64BE>(tocSec: tocISB, offset: rel.addend); |
395 | |
396 | // Only non-preemptable defined symbols can be relaxed. |
397 | if (!d || d->isPreemptible) |
398 | return false; |
399 | |
400 | // R_PPC64_ADDR64 should have created a canonical PLT for the non-preemptable |
401 | // ifunc and changed its type to STT_FUNC. |
402 | assert(!d->isGnuIFunc()); |
403 | |
404 | // Two instructions can materialize a 32-bit signed offset from the toc base. |
405 | uint64_t tocRelative = d->getVA(ctx, addend) - getPPC64TocBase(ctx); |
406 | if (!isInt<32>(x: tocRelative)) |
407 | return false; |
408 | |
409 | // Add PPC64TocOffset that will be subtracted by PPC64::relocate(). |
410 | static_cast<const PPC64 &>(*ctx.target) |
411 | .relaxGot(loc: bufLoc, rel, val: tocRelative + ppc64TocOffset); |
412 | return true; |
413 | } |
414 | |
415 | // Relocation masks following the #lo(value), #hi(value), #ha(value), |
416 | // #higher(value), #highera(value), #highest(value), and #highesta(value) |
417 | // macros defined in section 4.5.1. Relocation Types of the PPC-elf64abi |
418 | // document. |
419 | static uint16_t lo(uint64_t v) { return v; } |
420 | static uint16_t hi(uint64_t v) { return v >> 16; } |
421 | static uint64_t ha(uint64_t v) { return (v + 0x8000) >> 16; } |
422 | static uint16_t higher(uint64_t v) { return v >> 32; } |
423 | static uint16_t highera(uint64_t v) { return (v + 0x8000) >> 32; } |
424 | static uint16_t highest(uint64_t v) { return v >> 48; } |
425 | static uint16_t highesta(uint64_t v) { return (v + 0x8000) >> 48; } |
426 | |
427 | // Extracts the 'PO' field of an instruction encoding. |
428 | static uint8_t getPrimaryOpCode(uint32_t encoding) { return (encoding >> 26); } |
429 | |
430 | static bool isDQFormInstruction(uint32_t encoding) { |
431 | switch (getPrimaryOpCode(encoding)) { |
432 | default: |
433 | return false; |
434 | case 6: // Power10 paired loads/stores (lxvp, stxvp). |
435 | case 56: |
436 | // The only instruction with a primary opcode of 56 is `lq`. |
437 | return true; |
438 | case 61: |
439 | // There are both DS and DQ instruction forms with this primary opcode. |
440 | // Namely `lxv` and `stxv` are the DQ-forms that use it. |
441 | // The DS 'XO' bits being set to 01 is restricted to DQ form. |
442 | return (encoding & 3) == 0x1; |
443 | } |
444 | } |
445 | |
446 | static bool isDSFormInstruction(PPCLegacyInsn insn) { |
447 | switch (insn) { |
448 | default: |
449 | return false; |
450 | case PPCLegacyInsn::LWA: |
451 | case PPCLegacyInsn::LD: |
452 | case PPCLegacyInsn::LXSD: |
453 | case PPCLegacyInsn::LXSSP: |
454 | case PPCLegacyInsn::STD: |
455 | case PPCLegacyInsn::STXSD: |
456 | case PPCLegacyInsn::STXSSP: |
457 | return true; |
458 | } |
459 | } |
460 | |
461 | static PPCLegacyInsn getPPCLegacyInsn(uint32_t encoding) { |
462 | uint32_t opc = encoding & 0xfc000000; |
463 | |
464 | // If the primary opcode is shared between multiple instructions, we need to |
465 | // fix it up to match the actual instruction we are after. |
466 | if ((opc == 0xe4000000 || opc == 0xe8000000 || opc == 0xf4000000 || |
467 | opc == 0xf8000000) && |
468 | !isDQFormInstruction(encoding)) |
469 | opc = encoding & 0xfc000003; |
470 | else if (opc == 0xf4000000) |
471 | opc = encoding & 0xfc000007; |
472 | else if (opc == 0x18000000) |
473 | opc = encoding & 0xfc00000f; |
474 | |
475 | // If the value is not one of the enumerators in PPCLegacyInsn, we want to |
476 | // return PPCLegacyInsn::NOINSN. |
477 | if (!checkPPCLegacyInsn(encoding: opc)) |
478 | return PPCLegacyInsn::NOINSN; |
479 | return static_cast<PPCLegacyInsn>(opc); |
480 | } |
481 | |
482 | static PPCPrefixedInsn getPCRelativeForm(PPCLegacyInsn insn) { |
483 | switch (insn) { |
484 | #define PCREL_OPT(Legacy, PCRel, InsnMask) \ |
485 | case PPCLegacyInsn::Legacy: \ |
486 | return PPCPrefixedInsn::PCRel |
487 | #include "PPCInsns.def" |
488 | #undef PCREL_OPT |
489 | } |
490 | return PPCPrefixedInsn::NOINSN; |
491 | } |
492 | |
493 | static LegacyToPrefixMask getInsnMask(PPCLegacyInsn insn) { |
494 | switch (insn) { |
495 | #define PCREL_OPT(Legacy, PCRel, InsnMask) \ |
496 | case PPCLegacyInsn::Legacy: \ |
497 | return LegacyToPrefixMask::InsnMask |
498 | #include "PPCInsns.def" |
499 | #undef PCREL_OPT |
500 | } |
501 | return LegacyToPrefixMask::NOMASK; |
502 | } |
503 | static uint64_t getPCRelativeForm(uint32_t encoding) { |
504 | PPCLegacyInsn origInsn = getPPCLegacyInsn(encoding); |
505 | PPCPrefixedInsn pcrelInsn = getPCRelativeForm(insn: origInsn); |
506 | if (pcrelInsn == PPCPrefixedInsn::NOINSN) |
507 | return UINT64_C(-1); |
508 | LegacyToPrefixMask origInsnMask = getInsnMask(insn: origInsn); |
509 | uint64_t pcrelEncoding = |
510 | (uint64_t)pcrelInsn | (encoding & (uint64_t)origInsnMask); |
511 | |
512 | // If the mask requires moving bit 28 to bit 5, do that now. |
513 | if (origInsnMask == LegacyToPrefixMask::ST_STX28_TO5) |
514 | pcrelEncoding |= (encoding & 0x8) << 23; |
515 | return pcrelEncoding; |
516 | } |
517 | |
518 | static bool isInstructionUpdateForm(uint32_t encoding) { |
519 | switch (getPrimaryOpCode(encoding)) { |
520 | default: |
521 | return false; |
522 | case LBZU: |
523 | case LHAU: |
524 | case LHZU: |
525 | case LWZU: |
526 | case LFSU: |
527 | case LFDU: |
528 | case STBU: |
529 | case STHU: |
530 | case STWU: |
531 | case STFSU: |
532 | case STFDU: |
533 | return true; |
534 | // LWA has the same opcode as LD, and the DS bits is what differentiates |
535 | // between LD/LDU/LWA |
536 | case LD: |
537 | case STD: |
538 | return (encoding & 3) == 1; |
539 | } |
540 | } |
541 | |
542 | // Compute the total displacement between the prefixed instruction that gets |
543 | // to the start of the data and the load/store instruction that has the offset |
544 | // into the data structure. |
545 | // For example: |
546 | // paddi 3, 0, 1000, 1 |
547 | // lwz 3, 20(3) |
548 | // Should add up to 1020 for total displacement. |
549 | static int64_t getTotalDisp(uint64_t prefixedInsn, uint32_t accessInsn) { |
550 | int64_t disp34 = llvm::SignExtend64( |
551 | X: ((prefixedInsn & 0x3ffff00000000) >> 16) | (prefixedInsn & 0xffff), B: 34); |
552 | int32_t disp16 = llvm::SignExtend32(X: accessInsn & 0xffff, B: 16); |
553 | // For DS and DQ form instructions, we need to mask out the XO bits. |
554 | if (isDQFormInstruction(encoding: accessInsn)) |
555 | disp16 &= ~0xf; |
556 | else if (isDSFormInstruction(insn: getPPCLegacyInsn(encoding: accessInsn))) |
557 | disp16 &= ~0x3; |
558 | return disp34 + disp16; |
559 | } |
560 | |
561 | // There are a number of places when we either want to read or write an |
562 | // instruction when handling a half16 relocation type. On big-endian the buffer |
563 | // pointer is pointing into the middle of the word we want to extract, and on |
564 | // little-endian it is pointing to the start of the word. These 2 helpers are to |
565 | // simplify reading and writing in that context. |
566 | static void writeFromHalf16(Ctx &ctx, uint8_t *loc, uint32_t insn) { |
567 | write32(ctx, p: ctx.arg.isLE ? loc : loc - 2, v: insn); |
568 | } |
569 | |
570 | static uint32_t readFromHalf16(Ctx &ctx, const uint8_t *loc) { |
571 | return read32(ctx, p: ctx.arg.isLE ? loc : loc - 2); |
572 | } |
573 | |
574 | static uint64_t readPrefixedInst(Ctx &ctx, const uint8_t *loc) { |
575 | uint64_t fullInstr = read64(ctx, p: loc); |
576 | return ctx.arg.isLE ? (fullInstr << 32 | fullInstr >> 32) : fullInstr; |
577 | } |
578 | |
579 | PPC64::PPC64(Ctx &ctx) : TargetInfo(ctx) { |
580 | copyRel = R_PPC64_COPY; |
581 | gotRel = R_PPC64_GLOB_DAT; |
582 | pltRel = R_PPC64_JMP_SLOT; |
583 | relativeRel = R_PPC64_RELATIVE; |
584 | iRelativeRel = R_PPC64_IRELATIVE; |
585 | symbolicRel = R_PPC64_ADDR64; |
586 | pltHeaderSize = 60; |
587 | pltEntrySize = 4; |
588 | ipltEntrySize = 16; // PPC64PltCallStub::size |
589 | gotHeaderEntriesNum = 1; |
590 | gotPltHeaderEntriesNum = 2; |
591 | needsThunks = true; |
592 | |
593 | tlsModuleIndexRel = R_PPC64_DTPMOD64; |
594 | tlsOffsetRel = R_PPC64_DTPREL64; |
595 | |
596 | tlsGotRel = R_PPC64_TPREL64; |
597 | |
598 | needsMoreStackNonSplit = false; |
599 | |
600 | // We need 64K pages (at least under glibc/Linux, the loader won't |
601 | // set different permissions on a finer granularity than that). |
602 | defaultMaxPageSize = 65536; |
603 | |
604 | // The PPC64 ELF ABI v1 spec, says: |
605 | // |
606 | // It is normally desirable to put segments with different characteristics |
607 | // in separate 256 Mbyte portions of the address space, to give the |
608 | // operating system full paging flexibility in the 64-bit address space. |
609 | // |
610 | // And because the lowest non-zero 256M boundary is 0x10000000, PPC64 linkers |
611 | // use 0x10000000 as the starting address. |
612 | defaultImageBase = 0x10000000; |
613 | |
614 | write32(ctx, p: trapInstr.data(), v: 0x7fe00008); |
615 | } |
616 | |
617 | int PPC64::getTlsGdRelaxSkip(RelType type) const { |
618 | // A __tls_get_addr call instruction is marked with 2 relocations: |
619 | // |
620 | // R_PPC64_TLSGD / R_PPC64_TLSLD: marker relocation |
621 | // R_PPC64_REL24: __tls_get_addr |
622 | // |
623 | // After the relaxation we no longer call __tls_get_addr and should skip both |
624 | // relocations to not create a false dependence on __tls_get_addr being |
625 | // defined. |
626 | if (type == R_PPC64_TLSGD || type == R_PPC64_TLSLD) |
627 | return 2; |
628 | return 1; |
629 | } |
630 | |
631 | static uint32_t getEFlags(InputFile *file) { |
632 | if (file->ekind == ELF64BEKind) |
633 | return cast<ObjFile<ELF64BE>>(Val: file)->getObj().getHeader().e_flags; |
634 | return cast<ObjFile<ELF64LE>>(Val: file)->getObj().getHeader().e_flags; |
635 | } |
636 | |
637 | // This file implements v2 ABI. This function makes sure that all |
638 | // object files have v2 or an unspecified version as an ABI version. |
639 | uint32_t PPC64::calcEFlags() const { |
640 | for (InputFile *f : ctx.objectFiles) { |
641 | uint32_t flag = getEFlags(file: f); |
642 | if (flag == 1) |
643 | ErrAlways(ctx) << f << ": ABI version 1 is not supported" ; |
644 | else if (flag > 2) |
645 | ErrAlways(ctx) << f << ": unrecognized e_flags: " << flag; |
646 | } |
647 | return 2; |
648 | } |
649 | |
650 | void PPC64::relaxGot(uint8_t *loc, const Relocation &rel, uint64_t val) const { |
651 | switch (rel.type) { |
652 | case R_PPC64_TOC16_HA: |
653 | // Convert "addis reg, 2, .LC0@toc@h" to "addis reg, 2, var@toc@h" or "nop". |
654 | relocate(loc, rel, val); |
655 | break; |
656 | case R_PPC64_TOC16_LO_DS: { |
657 | // Convert "ld reg, .LC0@toc@l(reg)" to "addi reg, reg, var@toc@l" or |
658 | // "addi reg, 2, var@toc". |
659 | uint32_t insn = readFromHalf16(ctx, loc); |
660 | if (getPrimaryOpCode(encoding: insn) != LD) |
661 | ErrAlways(ctx) |
662 | << "expected a 'ld' for got-indirect to toc-relative relaxing" ; |
663 | writeFromHalf16(ctx, loc, insn: (insn & 0x03ffffff) | 0x38000000); |
664 | relocateNoSym(loc, type: R_PPC64_TOC16_LO, val); |
665 | break; |
666 | } |
667 | case R_PPC64_GOT_PCREL34: { |
668 | // Clear the first 8 bits of the prefix and the first 6 bits of the |
669 | // instruction (the primary opcode). |
670 | uint64_t insn = readPrefixedInst(ctx, loc); |
671 | if ((insn & 0xfc000000) != 0xe4000000) |
672 | ErrAlways(ctx) |
673 | << "expected a 'pld' for got-indirect to pc-relative relaxing" ; |
674 | insn &= ~0xff000000fc000000; |
675 | |
676 | // Replace the cleared bits with the values for PADDI (0x600000038000000); |
677 | insn |= 0x600000038000000; |
678 | writePrefixedInst(ctx, loc, insn); |
679 | relocate(loc, rel, val); |
680 | break; |
681 | } |
682 | case R_PPC64_PCREL_OPT: { |
683 | // We can only relax this if the R_PPC64_GOT_PCREL34 at this offset can |
684 | // be relaxed. The eligibility for the relaxation needs to be determined |
685 | // on that relocation since this one does not relocate a symbol. |
686 | uint64_t insn = readPrefixedInst(ctx, loc); |
687 | uint32_t accessInsn = read32(ctx, p: loc + rel.addend); |
688 | uint64_t pcRelInsn = getPCRelativeForm(encoding: accessInsn); |
689 | |
690 | // This error is not necessary for correctness but is emitted for now |
691 | // to ensure we don't miss these opportunities in real code. It can be |
692 | // removed at a later date. |
693 | if (pcRelInsn == UINT64_C(-1)) { |
694 | Err(ctx) |
695 | << "unrecognized instruction for R_PPC64_PCREL_OPT relaxation: 0x" |
696 | << utohexstr(X: accessInsn, LowerCase: true); |
697 | break; |
698 | } |
699 | |
700 | int64_t totalDisp = getTotalDisp(prefixedInsn: insn, accessInsn); |
701 | if (!isInt<34>(x: totalDisp)) |
702 | break; // Displacement doesn't fit. |
703 | // Convert the PADDI to the prefixed version of accessInsn and convert |
704 | // accessInsn to a nop. |
705 | writePrefixedInst(ctx, loc, |
706 | insn: pcRelInsn | ((totalDisp & 0x3ffff0000) << 16) | |
707 | (totalDisp & 0xffff)); |
708 | write32(ctx, p: loc + rel.addend, v: NOP); // nop accessInsn. |
709 | break; |
710 | } |
711 | default: |
712 | llvm_unreachable("unexpected relocation type" ); |
713 | } |
714 | } |
715 | |
716 | void PPC64::relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, |
717 | uint64_t val) const { |
718 | // Reference: 3.7.4.2 of the 64-bit ELF V2 abi supplement. |
719 | // The general dynamic code sequence for a global `x` will look like: |
720 | // Instruction Relocation Symbol |
721 | // addis r3, r2, x@got@tlsgd@ha R_PPC64_GOT_TLSGD16_HA x |
722 | // addi r3, r3, x@got@tlsgd@l R_PPC64_GOT_TLSGD16_LO x |
723 | // bl __tls_get_addr(x@tlsgd) R_PPC64_TLSGD x |
724 | // R_PPC64_REL24 __tls_get_addr |
725 | // nop None None |
726 | |
727 | // Relaxing to local exec entails converting: |
728 | // addis r3, r2, x@got@tlsgd@ha into nop |
729 | // addi r3, r3, x@got@tlsgd@l into addis r3, r13, x@tprel@ha |
730 | // bl __tls_get_addr(x@tlsgd) into nop |
731 | // nop into addi r3, r3, x@tprel@l |
732 | |
733 | switch (rel.type) { |
734 | case R_PPC64_GOT_TLSGD16_HA: |
735 | writeFromHalf16(ctx, loc, insn: NOP); |
736 | break; |
737 | case R_PPC64_GOT_TLSGD16: |
738 | case R_PPC64_GOT_TLSGD16_LO: |
739 | writeFromHalf16(ctx, loc, insn: 0x3c6d0000); // addis r3, r13 |
740 | relocateNoSym(loc, type: R_PPC64_TPREL16_HA, val); |
741 | break; |
742 | case R_PPC64_GOT_TLSGD_PCREL34: |
743 | // Relax from paddi r3, 0, x@got@tlsgd@pcrel, 1 to |
744 | // paddi r3, r13, x@tprel, 0 |
745 | writePrefixedInst(ctx, loc, insn: 0x06000000386d0000); |
746 | relocateNoSym(loc, type: R_PPC64_TPREL34, val); |
747 | break; |
748 | case R_PPC64_TLSGD: { |
749 | // PC Relative Relaxation: |
750 | // Relax from bl __tls_get_addr@notoc(x@tlsgd) to |
751 | // nop |
752 | // TOC Relaxation: |
753 | // Relax from bl __tls_get_addr(x@tlsgd) |
754 | // nop |
755 | // to |
756 | // nop |
757 | // addi r3, r3, x@tprel@l |
758 | const uintptr_t locAsInt = reinterpret_cast<uintptr_t>(loc); |
759 | if (locAsInt % 4 == 0) { |
760 | write32(ctx, p: loc, v: NOP); // nop |
761 | write32(ctx, p: loc + 4, v: 0x38630000); // addi r3, r3 |
762 | // Since we are relocating a half16 type relocation and Loc + 4 points to |
763 | // the start of an instruction we need to advance the buffer by an extra |
764 | // 2 bytes on BE. |
765 | relocateNoSym(loc: loc + 4 + (ctx.arg.ekind == ELF64BEKind ? 2 : 0), |
766 | type: R_PPC64_TPREL16_LO, val); |
767 | } else if (locAsInt % 4 == 1) { |
768 | write32(ctx, p: loc - 1, v: NOP); |
769 | } else { |
770 | Err(ctx) << "R_PPC64_TLSGD has unexpected byte alignment" ; |
771 | } |
772 | break; |
773 | } |
774 | default: |
775 | llvm_unreachable("unsupported relocation for TLS GD to LE relaxation" ); |
776 | } |
777 | } |
778 | |
779 | void PPC64::relaxTlsLdToLe(uint8_t *loc, const Relocation &rel, |
780 | uint64_t val) const { |
781 | // Reference: 3.7.4.3 of the 64-bit ELF V2 abi supplement. |
782 | // The local dynamic code sequence for a global `x` will look like: |
783 | // Instruction Relocation Symbol |
784 | // addis r3, r2, x@got@tlsld@ha R_PPC64_GOT_TLSLD16_HA x |
785 | // addi r3, r3, x@got@tlsld@l R_PPC64_GOT_TLSLD16_LO x |
786 | // bl __tls_get_addr(x@tlsgd) R_PPC64_TLSLD x |
787 | // R_PPC64_REL24 __tls_get_addr |
788 | // nop None None |
789 | |
790 | // Relaxing to local exec entails converting: |
791 | // addis r3, r2, x@got@tlsld@ha into nop |
792 | // addi r3, r3, x@got@tlsld@l into addis r3, r13, 0 |
793 | // bl __tls_get_addr(x@tlsgd) into nop |
794 | // nop into addi r3, r3, 4096 |
795 | |
796 | switch (rel.type) { |
797 | case R_PPC64_GOT_TLSLD16_HA: |
798 | writeFromHalf16(ctx, loc, insn: NOP); |
799 | break; |
800 | case R_PPC64_GOT_TLSLD16_LO: |
801 | writeFromHalf16(ctx, loc, insn: 0x3c6d0000); // addis r3, r13, 0 |
802 | break; |
803 | case R_PPC64_GOT_TLSLD_PCREL34: |
804 | // Relax from paddi r3, 0, x1@got@tlsld@pcrel, 1 to |
805 | // paddi r3, r13, 0x1000, 0 |
806 | writePrefixedInst(ctx, loc, insn: 0x06000000386d1000); |
807 | break; |
808 | case R_PPC64_TLSLD: { |
809 | // PC Relative Relaxation: |
810 | // Relax from bl __tls_get_addr@notoc(x@tlsld) |
811 | // to |
812 | // nop |
813 | // TOC Relaxation: |
814 | // Relax from bl __tls_get_addr(x@tlsld) |
815 | // nop |
816 | // to |
817 | // nop |
818 | // addi r3, r3, 4096 |
819 | const uintptr_t locAsInt = reinterpret_cast<uintptr_t>(loc); |
820 | if (locAsInt % 4 == 0) { |
821 | write32(ctx, p: loc, v: NOP); |
822 | write32(ctx, p: loc + 4, v: 0x38631000); // addi r3, r3, 4096 |
823 | } else if (locAsInt % 4 == 1) { |
824 | write32(ctx, p: loc - 1, v: NOP); |
825 | } else { |
826 | Err(ctx) << "R_PPC64_TLSLD has unexpected byte alignment" ; |
827 | } |
828 | break; |
829 | } |
830 | case R_PPC64_DTPREL16: |
831 | case R_PPC64_DTPREL16_HA: |
832 | case R_PPC64_DTPREL16_HI: |
833 | case R_PPC64_DTPREL16_DS: |
834 | case R_PPC64_DTPREL16_LO: |
835 | case R_PPC64_DTPREL16_LO_DS: |
836 | case R_PPC64_DTPREL34: |
837 | relocate(loc, rel, val); |
838 | break; |
839 | default: |
840 | llvm_unreachable("unsupported relocation for TLS LD to LE relaxation" ); |
841 | } |
842 | } |
843 | |
844 | // Map X-Form instructions to their DS-Form counterparts, if applicable. |
845 | // The full encoding is returned here to distinguish between the different |
846 | // DS-Form instructions. |
847 | unsigned elf::getPPCDSFormOp(unsigned secondaryOp) { |
848 | switch (secondaryOp) { |
849 | case LWAX: |
850 | return (LWA << 26) | 0x2; |
851 | case LDX: |
852 | return LD << 26; |
853 | case STDX: |
854 | return STD << 26; |
855 | default: |
856 | return 0; |
857 | } |
858 | } |
859 | |
860 | unsigned elf::getPPCDFormOp(unsigned secondaryOp) { |
861 | switch (secondaryOp) { |
862 | case LBZX: |
863 | return LBZ << 26; |
864 | case LHZX: |
865 | return LHZ << 26; |
866 | case LWZX: |
867 | return LWZ << 26; |
868 | case STBX: |
869 | return STB << 26; |
870 | case STHX: |
871 | return STH << 26; |
872 | case STWX: |
873 | return STW << 26; |
874 | case LHAX: |
875 | return LHA << 26; |
876 | case LFSX: |
877 | return LFS << 26; |
878 | case LFDX: |
879 | return LFD << 26; |
880 | case STFSX: |
881 | return STFS << 26; |
882 | case STFDX: |
883 | return STFD << 26; |
884 | case ADD: |
885 | return ADDI << 26; |
886 | default: |
887 | return 0; |
888 | } |
889 | } |
890 | |
891 | void PPC64::relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, |
892 | uint64_t val) const { |
893 | // The initial exec code sequence for a global `x` will look like: |
894 | // Instruction Relocation Symbol |
895 | // addis r9, r2, x@got@tprel@ha R_PPC64_GOT_TPREL16_HA x |
896 | // ld r9, x@got@tprel@l(r9) R_PPC64_GOT_TPREL16_LO_DS x |
897 | // add r9, r9, x@tls R_PPC64_TLS x |
898 | |
899 | // Relaxing to local exec entails converting: |
900 | // addis r9, r2, x@got@tprel@ha into nop |
901 | // ld r9, x@got@tprel@l(r9) into addis r9, r13, x@tprel@ha |
902 | // add r9, r9, x@tls into addi r9, r9, x@tprel@l |
903 | |
904 | // x@tls R_PPC64_TLS is a relocation which does not compute anything, |
905 | // it is replaced with r13 (thread pointer). |
906 | |
907 | // The add instruction in the initial exec sequence has multiple variations |
908 | // that need to be handled. If we are building an address it will use an add |
909 | // instruction, if we are accessing memory it will use any of the X-form |
910 | // indexed load or store instructions. |
911 | |
912 | unsigned offset = (ctx.arg.ekind == ELF64BEKind) ? 2 : 0; |
913 | switch (rel.type) { |
914 | case R_PPC64_GOT_TPREL16_HA: |
915 | write32(ctx, p: loc - offset, v: NOP); |
916 | break; |
917 | case R_PPC64_GOT_TPREL16_LO_DS: |
918 | case R_PPC64_GOT_TPREL16_DS: { |
919 | uint32_t regNo = read32(ctx, p: loc - offset) & 0x03e00000; // bits 6-10 |
920 | write32(ctx, p: loc - offset, v: 0x3c0d0000 | regNo); // addis RegNo, r13 |
921 | relocateNoSym(loc, type: R_PPC64_TPREL16_HA, val); |
922 | break; |
923 | } |
924 | case R_PPC64_GOT_TPREL_PCREL34: { |
925 | const uint64_t pldRT = readPrefixedInst(ctx, loc) & 0x0000000003e00000; |
926 | // paddi RT(from pld), r13, symbol@tprel, 0 |
927 | writePrefixedInst(ctx, loc, insn: 0x06000000380d0000 | pldRT); |
928 | relocateNoSym(loc, type: R_PPC64_TPREL34, val); |
929 | break; |
930 | } |
931 | case R_PPC64_TLS: { |
932 | const uintptr_t locAsInt = reinterpret_cast<uintptr_t>(loc); |
933 | if (locAsInt % 4 == 0) { |
934 | uint32_t primaryOp = getPrimaryOpCode(encoding: read32(ctx, p: loc)); |
935 | if (primaryOp != 31) |
936 | ErrAlways(ctx) << "unrecognized instruction for IE to LE R_PPC64_TLS" ; |
937 | uint32_t secondaryOp = (read32(ctx, p: loc) & 0x000007fe) >> 1; // bits 21-30 |
938 | uint32_t dFormOp = getPPCDFormOp(secondaryOp); |
939 | uint32_t finalReloc; |
940 | if (dFormOp == 0) { // Expecting a DS-Form instruction. |
941 | dFormOp = getPPCDSFormOp(secondaryOp); |
942 | if (dFormOp == 0) |
943 | ErrAlways(ctx) << "unrecognized instruction for IE to LE R_PPC64_TLS" ; |
944 | finalReloc = R_PPC64_TPREL16_LO_DS; |
945 | } else |
946 | finalReloc = R_PPC64_TPREL16_LO; |
947 | write32(ctx, p: loc, v: dFormOp | (read32(ctx, p: loc) & 0x03ff0000)); |
948 | relocateNoSym(loc: loc + offset, type: finalReloc, val); |
949 | } else if (locAsInt % 4 == 1) { |
950 | // If the offset is not 4 byte aligned then we have a PCRel type reloc. |
951 | // This version of the relocation is offset by one byte from the |
952 | // instruction it references. |
953 | uint32_t tlsInstr = read32(ctx, p: loc - 1); |
954 | uint32_t primaryOp = getPrimaryOpCode(encoding: tlsInstr); |
955 | if (primaryOp != 31) |
956 | Err(ctx) << "unrecognized instruction for IE to LE R_PPC64_TLS" ; |
957 | uint32_t secondaryOp = (tlsInstr & 0x000007FE) >> 1; // bits 21-30 |
958 | // The add is a special case and should be turned into a nop. The paddi |
959 | // that comes before it will already have computed the address of the |
960 | // symbol. |
961 | if (secondaryOp == 266) { |
962 | // Check if the add uses the same result register as the input register. |
963 | uint32_t rt = (tlsInstr & 0x03E00000) >> 21; // bits 6-10 |
964 | uint32_t ra = (tlsInstr & 0x001F0000) >> 16; // bits 11-15 |
965 | if (ra == rt) { |
966 | write32(ctx, p: loc - 1, v: NOP); |
967 | } else { |
968 | // mr rt, ra |
969 | write32(ctx, p: loc - 1, |
970 | v: 0x7C000378 | (rt << 16) | (ra << 21) | (ra << 11)); |
971 | } |
972 | } else { |
973 | uint32_t dFormOp = getPPCDFormOp(secondaryOp); |
974 | if (dFormOp == 0) { // Expecting a DS-Form instruction. |
975 | dFormOp = getPPCDSFormOp(secondaryOp); |
976 | if (dFormOp == 0) |
977 | Err(ctx) << "unrecognized instruction for IE to LE R_PPC64_TLS" ; |
978 | } |
979 | write32(ctx, p: loc - 1, v: (dFormOp | (tlsInstr & 0x03ff0000))); |
980 | } |
981 | } else { |
982 | Err(ctx) << "R_PPC64_TLS must be either 4 byte aligned or one byte " |
983 | "offset from 4 byte aligned" ; |
984 | } |
985 | break; |
986 | } |
987 | default: |
988 | llvm_unreachable("unknown relocation for IE to LE" ); |
989 | break; |
990 | } |
991 | } |
992 | |
993 | RelExpr PPC64::getRelExpr(RelType type, const Symbol &s, |
994 | const uint8_t *loc) const { |
995 | switch (type) { |
996 | case R_PPC64_NONE: |
997 | return R_NONE; |
998 | case R_PPC64_ADDR16: |
999 | case R_PPC64_ADDR16_DS: |
1000 | case R_PPC64_ADDR16_HA: |
1001 | case R_PPC64_ADDR16_HI: |
1002 | case R_PPC64_ADDR16_HIGH: |
1003 | case R_PPC64_ADDR16_HIGHER: |
1004 | case R_PPC64_ADDR16_HIGHERA: |
1005 | case R_PPC64_ADDR16_HIGHEST: |
1006 | case R_PPC64_ADDR16_HIGHESTA: |
1007 | case R_PPC64_ADDR16_LO: |
1008 | case R_PPC64_ADDR16_LO_DS: |
1009 | case R_PPC64_ADDR32: |
1010 | case R_PPC64_ADDR64: |
1011 | return R_ABS; |
1012 | case R_PPC64_GOT16: |
1013 | case R_PPC64_GOT16_DS: |
1014 | case R_PPC64_GOT16_HA: |
1015 | case R_PPC64_GOT16_HI: |
1016 | case R_PPC64_GOT16_LO: |
1017 | case R_PPC64_GOT16_LO_DS: |
1018 | return R_GOT_OFF; |
1019 | case R_PPC64_TOC16: |
1020 | case R_PPC64_TOC16_DS: |
1021 | case R_PPC64_TOC16_HI: |
1022 | case R_PPC64_TOC16_LO: |
1023 | return R_GOTREL; |
1024 | case R_PPC64_GOT_PCREL34: |
1025 | case R_PPC64_GOT_TPREL_PCREL34: |
1026 | case R_PPC64_PCREL_OPT: |
1027 | return R_GOT_PC; |
1028 | case R_PPC64_TOC16_HA: |
1029 | case R_PPC64_TOC16_LO_DS: |
1030 | return ctx.arg.tocOptimize ? RE_PPC64_RELAX_TOC : R_GOTREL; |
1031 | case R_PPC64_TOC: |
1032 | return RE_PPC64_TOCBASE; |
1033 | case R_PPC64_REL14: |
1034 | case R_PPC64_REL24: |
1035 | return RE_PPC64_CALL_PLT; |
1036 | case R_PPC64_REL24_NOTOC: |
1037 | return R_PLT_PC; |
1038 | case R_PPC64_REL16_LO: |
1039 | case R_PPC64_REL16_HA: |
1040 | case R_PPC64_REL16_HI: |
1041 | case R_PPC64_REL32: |
1042 | case R_PPC64_REL64: |
1043 | case R_PPC64_PCREL34: |
1044 | return R_PC; |
1045 | case R_PPC64_GOT_TLSGD16: |
1046 | case R_PPC64_GOT_TLSGD16_HA: |
1047 | case R_PPC64_GOT_TLSGD16_HI: |
1048 | case R_PPC64_GOT_TLSGD16_LO: |
1049 | return R_TLSGD_GOT; |
1050 | case R_PPC64_GOT_TLSGD_PCREL34: |
1051 | return R_TLSGD_PC; |
1052 | case R_PPC64_GOT_TLSLD16: |
1053 | case R_PPC64_GOT_TLSLD16_HA: |
1054 | case R_PPC64_GOT_TLSLD16_HI: |
1055 | case R_PPC64_GOT_TLSLD16_LO: |
1056 | return R_TLSLD_GOT; |
1057 | case R_PPC64_GOT_TLSLD_PCREL34: |
1058 | return R_TLSLD_PC; |
1059 | case R_PPC64_GOT_TPREL16_HA: |
1060 | case R_PPC64_GOT_TPREL16_LO_DS: |
1061 | case R_PPC64_GOT_TPREL16_DS: |
1062 | case R_PPC64_GOT_TPREL16_HI: |
1063 | return R_GOT_OFF; |
1064 | case R_PPC64_GOT_DTPREL16_HA: |
1065 | case R_PPC64_GOT_DTPREL16_LO_DS: |
1066 | case R_PPC64_GOT_DTPREL16_DS: |
1067 | case R_PPC64_GOT_DTPREL16_HI: |
1068 | return R_TLSLD_GOT_OFF; |
1069 | case R_PPC64_TPREL16: |
1070 | case R_PPC64_TPREL16_HA: |
1071 | case R_PPC64_TPREL16_LO: |
1072 | case R_PPC64_TPREL16_HI: |
1073 | case R_PPC64_TPREL16_DS: |
1074 | case R_PPC64_TPREL16_LO_DS: |
1075 | case R_PPC64_TPREL16_HIGHER: |
1076 | case R_PPC64_TPREL16_HIGHERA: |
1077 | case R_PPC64_TPREL16_HIGHEST: |
1078 | case R_PPC64_TPREL16_HIGHESTA: |
1079 | case R_PPC64_TPREL34: |
1080 | return R_TPREL; |
1081 | case R_PPC64_DTPREL16: |
1082 | case R_PPC64_DTPREL16_DS: |
1083 | case R_PPC64_DTPREL16_HA: |
1084 | case R_PPC64_DTPREL16_HI: |
1085 | case R_PPC64_DTPREL16_HIGHER: |
1086 | case R_PPC64_DTPREL16_HIGHERA: |
1087 | case R_PPC64_DTPREL16_HIGHEST: |
1088 | case R_PPC64_DTPREL16_HIGHESTA: |
1089 | case R_PPC64_DTPREL16_LO: |
1090 | case R_PPC64_DTPREL16_LO_DS: |
1091 | case R_PPC64_DTPREL64: |
1092 | case R_PPC64_DTPREL34: |
1093 | return R_DTPREL; |
1094 | case R_PPC64_TLSGD: |
1095 | return R_TLSDESC_CALL; |
1096 | case R_PPC64_TLSLD: |
1097 | return R_TLSLD_HINT; |
1098 | case R_PPC64_TLS: |
1099 | return R_TLSIE_HINT; |
1100 | default: |
1101 | Err(ctx) << getErrorLoc(ctx, loc) << "unknown relocation (" << type.v |
1102 | << ") against symbol " << &s; |
1103 | return R_NONE; |
1104 | } |
1105 | } |
1106 | |
1107 | RelType PPC64::getDynRel(RelType type) const { |
1108 | if (type == R_PPC64_ADDR64 || type == R_PPC64_TOC) |
1109 | return R_PPC64_ADDR64; |
1110 | return R_PPC64_NONE; |
1111 | } |
1112 | |
1113 | int64_t PPC64::getImplicitAddend(const uint8_t *buf, RelType type) const { |
1114 | switch (type) { |
1115 | case R_PPC64_NONE: |
1116 | case R_PPC64_GLOB_DAT: |
1117 | case R_PPC64_JMP_SLOT: |
1118 | return 0; |
1119 | case R_PPC64_REL32: |
1120 | return SignExtend64<32>(x: read32(ctx, p: buf)); |
1121 | case R_PPC64_ADDR64: |
1122 | case R_PPC64_REL64: |
1123 | case R_PPC64_RELATIVE: |
1124 | case R_PPC64_IRELATIVE: |
1125 | case R_PPC64_DTPMOD64: |
1126 | case R_PPC64_DTPREL64: |
1127 | case R_PPC64_TPREL64: |
1128 | return read64(ctx, p: buf); |
1129 | default: |
1130 | InternalErr(ctx, buf) << "cannot read addend for relocation " << type; |
1131 | return 0; |
1132 | } |
1133 | } |
1134 | |
1135 | void PPC64::(uint8_t *buf) const { |
1136 | write64(ctx, p: buf, v: getPPC64TocBase(ctx)); |
1137 | } |
1138 | |
1139 | void PPC64::(uint8_t *buf) const { |
1140 | // The generic resolver stub goes first. |
1141 | write32(ctx, p: buf + 0, v: 0x7c0802a6); // mflr r0 |
1142 | write32(ctx, p: buf + 4, v: 0x429f0005); // bcl 20,4*cr7+so,8 <_glink+0x8> |
1143 | write32(ctx, p: buf + 8, v: 0x7d6802a6); // mflr r11 |
1144 | write32(ctx, p: buf + 12, v: 0x7c0803a6); // mtlr r0 |
1145 | write32(ctx, p: buf + 16, v: 0x7d8b6050); // subf r12, r11, r12 |
1146 | write32(ctx, p: buf + 20, v: 0x380cffcc); // subi r0,r12,52 |
1147 | write32(ctx, p: buf + 24, v: 0x7800f082); // srdi r0,r0,62,2 |
1148 | write32(ctx, p: buf + 28, v: 0xe98b002c); // ld r12,44(r11) |
1149 | write32(ctx, p: buf + 32, v: 0x7d6c5a14); // add r11,r12,r11 |
1150 | write32(ctx, p: buf + 36, v: 0xe98b0000); // ld r12,0(r11) |
1151 | write32(ctx, p: buf + 40, v: 0xe96b0008); // ld r11,8(r11) |
1152 | write32(ctx, p: buf + 44, v: 0x7d8903a6); // mtctr r12 |
1153 | write32(ctx, p: buf + 48, v: 0x4e800420); // bctr |
1154 | |
1155 | // The 'bcl' instruction will set the link register to the address of the |
1156 | // following instruction ('mflr r11'). Here we store the offset from that |
1157 | // instruction to the first entry in the GotPlt section. |
1158 | int64_t gotPltOffset = ctx.in.gotPlt->getVA() - (ctx.in.plt->getVA() + 8); |
1159 | write64(ctx, p: buf + 52, v: gotPltOffset); |
1160 | } |
1161 | |
1162 | void PPC64::writePlt(uint8_t *buf, const Symbol &sym, |
1163 | uint64_t /*pltEntryAddr*/) const { |
1164 | int32_t offset = pltHeaderSize + sym.getPltIdx(ctx) * pltEntrySize; |
1165 | // bl __glink_PLTresolve |
1166 | write32(ctx, p: buf, v: 0x48000000 | ((-offset) & 0x03fffffc)); |
1167 | } |
1168 | |
1169 | void PPC64::writeIplt(uint8_t *buf, const Symbol &sym, |
1170 | uint64_t /*pltEntryAddr*/) const { |
1171 | writePPC64LoadAndBranch(ctx, buf, |
1172 | offset: sym.getGotPltVA(ctx) - getPPC64TocBase(ctx)); |
1173 | } |
1174 | |
1175 | static std::pair<RelType, uint64_t> toAddr16Rel(RelType type, uint64_t val) { |
1176 | // Relocations relative to the toc-base need to be adjusted by the Toc offset. |
1177 | uint64_t tocBiasedVal = val - ppc64TocOffset; |
1178 | // Relocations relative to dtv[dtpmod] need to be adjusted by the DTP offset. |
1179 | uint64_t dtpBiasedVal = val - dynamicThreadPointerOffset; |
1180 | |
1181 | switch (type) { |
1182 | // TOC biased relocation. |
1183 | case R_PPC64_GOT16: |
1184 | case R_PPC64_GOT_TLSGD16: |
1185 | case R_PPC64_GOT_TLSLD16: |
1186 | case R_PPC64_TOC16: |
1187 | return {R_PPC64_ADDR16, tocBiasedVal}; |
1188 | case R_PPC64_GOT16_DS: |
1189 | case R_PPC64_TOC16_DS: |
1190 | case R_PPC64_GOT_TPREL16_DS: |
1191 | case R_PPC64_GOT_DTPREL16_DS: |
1192 | return {R_PPC64_ADDR16_DS, tocBiasedVal}; |
1193 | case R_PPC64_GOT16_HA: |
1194 | case R_PPC64_GOT_TLSGD16_HA: |
1195 | case R_PPC64_GOT_TLSLD16_HA: |
1196 | case R_PPC64_GOT_TPREL16_HA: |
1197 | case R_PPC64_GOT_DTPREL16_HA: |
1198 | case R_PPC64_TOC16_HA: |
1199 | return {R_PPC64_ADDR16_HA, tocBiasedVal}; |
1200 | case R_PPC64_GOT16_HI: |
1201 | case R_PPC64_GOT_TLSGD16_HI: |
1202 | case R_PPC64_GOT_TLSLD16_HI: |
1203 | case R_PPC64_GOT_TPREL16_HI: |
1204 | case R_PPC64_GOT_DTPREL16_HI: |
1205 | case R_PPC64_TOC16_HI: |
1206 | return {R_PPC64_ADDR16_HI, tocBiasedVal}; |
1207 | case R_PPC64_GOT16_LO: |
1208 | case R_PPC64_GOT_TLSGD16_LO: |
1209 | case R_PPC64_GOT_TLSLD16_LO: |
1210 | case R_PPC64_TOC16_LO: |
1211 | return {R_PPC64_ADDR16_LO, tocBiasedVal}; |
1212 | case R_PPC64_GOT16_LO_DS: |
1213 | case R_PPC64_TOC16_LO_DS: |
1214 | case R_PPC64_GOT_TPREL16_LO_DS: |
1215 | case R_PPC64_GOT_DTPREL16_LO_DS: |
1216 | return {R_PPC64_ADDR16_LO_DS, tocBiasedVal}; |
1217 | |
1218 | // Dynamic Thread pointer biased relocation types. |
1219 | case R_PPC64_DTPREL16: |
1220 | return {R_PPC64_ADDR16, dtpBiasedVal}; |
1221 | case R_PPC64_DTPREL16_DS: |
1222 | return {R_PPC64_ADDR16_DS, dtpBiasedVal}; |
1223 | case R_PPC64_DTPREL16_HA: |
1224 | return {R_PPC64_ADDR16_HA, dtpBiasedVal}; |
1225 | case R_PPC64_DTPREL16_HI: |
1226 | return {R_PPC64_ADDR16_HI, dtpBiasedVal}; |
1227 | case R_PPC64_DTPREL16_HIGHER: |
1228 | return {R_PPC64_ADDR16_HIGHER, dtpBiasedVal}; |
1229 | case R_PPC64_DTPREL16_HIGHERA: |
1230 | return {R_PPC64_ADDR16_HIGHERA, dtpBiasedVal}; |
1231 | case R_PPC64_DTPREL16_HIGHEST: |
1232 | return {R_PPC64_ADDR16_HIGHEST, dtpBiasedVal}; |
1233 | case R_PPC64_DTPREL16_HIGHESTA: |
1234 | return {R_PPC64_ADDR16_HIGHESTA, dtpBiasedVal}; |
1235 | case R_PPC64_DTPREL16_LO: |
1236 | return {R_PPC64_ADDR16_LO, dtpBiasedVal}; |
1237 | case R_PPC64_DTPREL16_LO_DS: |
1238 | return {R_PPC64_ADDR16_LO_DS, dtpBiasedVal}; |
1239 | case R_PPC64_DTPREL64: |
1240 | return {R_PPC64_ADDR64, dtpBiasedVal}; |
1241 | |
1242 | default: |
1243 | return {type, val}; |
1244 | } |
1245 | } |
1246 | |
1247 | static bool isTocOptType(RelType type) { |
1248 | switch (type) { |
1249 | case R_PPC64_GOT16_HA: |
1250 | case R_PPC64_GOT16_LO_DS: |
1251 | case R_PPC64_TOC16_HA: |
1252 | case R_PPC64_TOC16_LO_DS: |
1253 | case R_PPC64_TOC16_LO: |
1254 | return true; |
1255 | default: |
1256 | return false; |
1257 | } |
1258 | } |
1259 | |
1260 | void PPC64::relocate(uint8_t *loc, const Relocation &rel, uint64_t val) const { |
1261 | RelType type = rel.type; |
1262 | bool shouldTocOptimize = isTocOptType(type); |
1263 | // For dynamic thread pointer relative, toc-relative, and got-indirect |
1264 | // relocations, proceed in terms of the corresponding ADDR16 relocation type. |
1265 | std::tie(args&: type, args&: val) = toAddr16Rel(type, val); |
1266 | |
1267 | switch (type) { |
1268 | case R_PPC64_ADDR14: { |
1269 | checkAlignment(ctx, loc, v: val, n: 4, rel); |
1270 | // Preserve the AA/LK bits in the branch instruction |
1271 | uint8_t aalk = loc[3]; |
1272 | write16(ctx, p: loc + 2, v: (aalk & 3) | (val & 0xfffc)); |
1273 | break; |
1274 | } |
1275 | case R_PPC64_ADDR16: |
1276 | checkIntUInt(ctx, loc, v: val, n: 16, rel); |
1277 | write16(ctx, p: loc, v: val); |
1278 | break; |
1279 | case R_PPC64_ADDR32: |
1280 | checkIntUInt(ctx, loc, v: val, n: 32, rel); |
1281 | write32(ctx, p: loc, v: val); |
1282 | break; |
1283 | case R_PPC64_ADDR16_DS: |
1284 | case R_PPC64_TPREL16_DS: { |
1285 | checkInt(ctx, loc, v: val, n: 16, rel); |
1286 | // DQ-form instructions use bits 28-31 as part of the instruction encoding |
1287 | // DS-form instructions only use bits 30-31. |
1288 | uint16_t mask = isDQFormInstruction(encoding: readFromHalf16(ctx, loc)) ? 0xf : 0x3; |
1289 | checkAlignment(ctx, loc, v: lo(v: val), n: mask + 1, rel); |
1290 | write16(ctx, p: loc, v: (read16(ctx, p: loc) & mask) | lo(v: val)); |
1291 | } break; |
1292 | case R_PPC64_ADDR16_HA: |
1293 | case R_PPC64_REL16_HA: |
1294 | case R_PPC64_TPREL16_HA: |
1295 | if (ctx.arg.tocOptimize && shouldTocOptimize && ha(v: val) == 0) |
1296 | writeFromHalf16(ctx, loc, insn: NOP); |
1297 | else { |
1298 | checkInt(ctx, loc, v: val + 0x8000, n: 32, rel); |
1299 | write16(ctx, p: loc, v: ha(v: val)); |
1300 | } |
1301 | break; |
1302 | case R_PPC64_ADDR16_HI: |
1303 | case R_PPC64_REL16_HI: |
1304 | case R_PPC64_TPREL16_HI: |
1305 | checkInt(ctx, loc, v: val, n: 32, rel); |
1306 | write16(ctx, p: loc, v: hi(v: val)); |
1307 | break; |
1308 | case R_PPC64_ADDR16_HIGH: |
1309 | write16(ctx, p: loc, v: hi(v: val)); |
1310 | break; |
1311 | case R_PPC64_ADDR16_HIGHER: |
1312 | case R_PPC64_TPREL16_HIGHER: |
1313 | write16(ctx, p: loc, v: higher(v: val)); |
1314 | break; |
1315 | case R_PPC64_ADDR16_HIGHERA: |
1316 | case R_PPC64_TPREL16_HIGHERA: |
1317 | write16(ctx, p: loc, v: highera(v: val)); |
1318 | break; |
1319 | case R_PPC64_ADDR16_HIGHEST: |
1320 | case R_PPC64_TPREL16_HIGHEST: |
1321 | write16(ctx, p: loc, v: highest(v: val)); |
1322 | break; |
1323 | case R_PPC64_ADDR16_HIGHESTA: |
1324 | case R_PPC64_TPREL16_HIGHESTA: |
1325 | write16(ctx, p: loc, v: highesta(v: val)); |
1326 | break; |
1327 | case R_PPC64_ADDR16_LO: |
1328 | case R_PPC64_REL16_LO: |
1329 | case R_PPC64_TPREL16_LO: |
1330 | // When the high-adjusted part of a toc relocation evaluates to 0, it is |
1331 | // changed into a nop. The lo part then needs to be updated to use the |
1332 | // toc-pointer register r2, as the base register. |
1333 | if (ctx.arg.tocOptimize && shouldTocOptimize && ha(v: val) == 0) { |
1334 | uint32_t insn = readFromHalf16(ctx, loc); |
1335 | if (isInstructionUpdateForm(encoding: insn)) |
1336 | Err(ctx) << getErrorLoc(ctx, loc) |
1337 | << "can't toc-optimize an update instruction: 0x" |
1338 | << utohexstr(X: insn, LowerCase: true); |
1339 | writeFromHalf16(ctx, loc, insn: (insn & 0xffe00000) | 0x00020000 | lo(v: val)); |
1340 | } else { |
1341 | write16(ctx, p: loc, v: lo(v: val)); |
1342 | } |
1343 | break; |
1344 | case R_PPC64_ADDR16_LO_DS: |
1345 | case R_PPC64_TPREL16_LO_DS: { |
1346 | // DQ-form instructions use bits 28-31 as part of the instruction encoding |
1347 | // DS-form instructions only use bits 30-31. |
1348 | uint32_t insn = readFromHalf16(ctx, loc); |
1349 | uint16_t mask = isDQFormInstruction(encoding: insn) ? 0xf : 0x3; |
1350 | checkAlignment(ctx, loc, v: lo(v: val), n: mask + 1, rel); |
1351 | if (ctx.arg.tocOptimize && shouldTocOptimize && ha(v: val) == 0) { |
1352 | // When the high-adjusted part of a toc relocation evaluates to 0, it is |
1353 | // changed into a nop. The lo part then needs to be updated to use the toc |
1354 | // pointer register r2, as the base register. |
1355 | if (isInstructionUpdateForm(encoding: insn)) |
1356 | Err(ctx) << getErrorLoc(ctx, loc) |
1357 | << "can't toc-optimize an update instruction: 0x" |
1358 | << utohexstr(X: insn, LowerCase: true); |
1359 | insn &= 0xffe00000 | mask; |
1360 | writeFromHalf16(ctx, loc, insn: insn | 0x00020000 | lo(v: val)); |
1361 | } else { |
1362 | write16(ctx, p: loc, v: (read16(ctx, p: loc) & mask) | lo(v: val)); |
1363 | } |
1364 | } break; |
1365 | case R_PPC64_TPREL16: |
1366 | checkInt(ctx, loc, v: val, n: 16, rel); |
1367 | write16(ctx, p: loc, v: val); |
1368 | break; |
1369 | case R_PPC64_REL32: |
1370 | checkInt(ctx, loc, v: val, n: 32, rel); |
1371 | write32(ctx, p: loc, v: val); |
1372 | break; |
1373 | case R_PPC64_ADDR64: |
1374 | case R_PPC64_REL64: |
1375 | case R_PPC64_TOC: |
1376 | write64(ctx, p: loc, v: val); |
1377 | break; |
1378 | case R_PPC64_REL14: { |
1379 | uint32_t mask = 0x0000FFFC; |
1380 | checkInt(ctx, loc, v: val, n: 16, rel); |
1381 | checkAlignment(ctx, loc, v: val, n: 4, rel); |
1382 | write32(ctx, p: loc, v: (read32(ctx, p: loc) & ~mask) | (val & mask)); |
1383 | break; |
1384 | } |
1385 | case R_PPC64_REL24: |
1386 | case R_PPC64_REL24_NOTOC: { |
1387 | uint32_t mask = 0x03FFFFFC; |
1388 | checkInt(ctx, loc, v: val, n: 26, rel); |
1389 | checkAlignment(ctx, loc, v: val, n: 4, rel); |
1390 | write32(ctx, p: loc, v: (read32(ctx, p: loc) & ~mask) | (val & mask)); |
1391 | break; |
1392 | } |
1393 | case R_PPC64_DTPREL64: |
1394 | write64(ctx, p: loc, v: val - dynamicThreadPointerOffset); |
1395 | break; |
1396 | case R_PPC64_DTPREL34: |
1397 | // The Dynamic Thread Vector actually points 0x8000 bytes past the start |
1398 | // of the TLS block. Therefore, in the case of R_PPC64_DTPREL34 we first |
1399 | // need to subtract that value then fallthrough to the general case. |
1400 | val -= dynamicThreadPointerOffset; |
1401 | [[fallthrough]]; |
1402 | case R_PPC64_PCREL34: |
1403 | case R_PPC64_GOT_PCREL34: |
1404 | case R_PPC64_GOT_TLSGD_PCREL34: |
1405 | case R_PPC64_GOT_TLSLD_PCREL34: |
1406 | case R_PPC64_GOT_TPREL_PCREL34: |
1407 | case R_PPC64_TPREL34: { |
1408 | const uint64_t si0Mask = 0x00000003ffff0000; |
1409 | const uint64_t si1Mask = 0x000000000000ffff; |
1410 | const uint64_t fullMask = 0x0003ffff0000ffff; |
1411 | checkInt(ctx, loc, v: val, n: 34, rel); |
1412 | |
1413 | uint64_t instr = readPrefixedInst(ctx, loc) & ~fullMask; |
1414 | writePrefixedInst(ctx, loc, |
1415 | insn: instr | ((val & si0Mask) << 16) | (val & si1Mask)); |
1416 | break; |
1417 | } |
1418 | // If we encounter a PCREL_OPT relocation that we won't optimize. |
1419 | case R_PPC64_PCREL_OPT: |
1420 | break; |
1421 | default: |
1422 | llvm_unreachable("unknown relocation" ); |
1423 | } |
1424 | } |
1425 | |
1426 | bool PPC64::needsThunk(RelExpr expr, RelType type, const InputFile *file, |
1427 | uint64_t branchAddr, const Symbol &s, int64_t a) const { |
1428 | if (type != R_PPC64_REL14 && type != R_PPC64_REL24 && |
1429 | type != R_PPC64_REL24_NOTOC) |
1430 | return false; |
1431 | |
1432 | // If a function is in the Plt it needs to be called with a call-stub. |
1433 | if (s.isInPlt(ctx)) |
1434 | return true; |
1435 | |
1436 | // This check looks at the st_other bits of the callee with relocation |
1437 | // R_PPC64_REL14 or R_PPC64_REL24. If the value is 1, then the callee |
1438 | // clobbers the TOC and we need an R2 save stub. |
1439 | if (type != R_PPC64_REL24_NOTOC && (s.stOther >> 5) == 1) |
1440 | return true; |
1441 | |
1442 | if (type == R_PPC64_REL24_NOTOC && (s.stOther >> 5) > 1) |
1443 | return true; |
1444 | |
1445 | // An undefined weak symbol not in a PLT does not need a thunk. If it is |
1446 | // hidden, its binding has been converted to local, so we just check |
1447 | // isUndefined() here. A undefined non-weak symbol has been errored. |
1448 | if (s.isUndefined()) |
1449 | return false; |
1450 | |
1451 | // If the offset exceeds the range of the branch type then it will need |
1452 | // a range-extending thunk. |
1453 | // See the comment in getRelocTargetVA() about RE_PPC64_CALL. |
1454 | return !inBranchRange( |
1455 | type, src: branchAddr, |
1456 | dst: s.getVA(ctx, addend: a) + getPPC64GlobalEntryToLocalEntryOffset(ctx, stOther: s.stOther)); |
1457 | } |
1458 | |
1459 | uint32_t PPC64::getThunkSectionSpacing() const { |
1460 | // See comment in Arch/ARM.cpp for a more detailed explanation of |
1461 | // getThunkSectionSpacing(). For PPC64 we pick the constant here based on |
1462 | // R_PPC64_REL24, which is used by unconditional branch instructions. |
1463 | // 0x2000000 = (1 << 24-1) * 4 |
1464 | return 0x2000000; |
1465 | } |
1466 | |
1467 | bool PPC64::inBranchRange(RelType type, uint64_t src, uint64_t dst) const { |
1468 | int64_t offset = dst - src; |
1469 | if (type == R_PPC64_REL14) |
1470 | return isInt<16>(x: offset); |
1471 | if (type == R_PPC64_REL24 || type == R_PPC64_REL24_NOTOC) |
1472 | return isInt<26>(x: offset); |
1473 | llvm_unreachable("unsupported relocation type used in branch" ); |
1474 | } |
1475 | |
1476 | RelExpr PPC64::adjustTlsExpr(RelType type, RelExpr expr) const { |
1477 | if (type != R_PPC64_GOT_TLSGD_PCREL34 && expr == R_RELAX_TLS_GD_TO_IE) |
1478 | return R_RELAX_TLS_GD_TO_IE_GOT_OFF; |
1479 | if (expr == R_RELAX_TLS_LD_TO_LE) |
1480 | return R_RELAX_TLS_LD_TO_LE_ABS; |
1481 | return expr; |
1482 | } |
1483 | |
1484 | RelExpr PPC64::adjustGotPcExpr(RelType type, int64_t addend, |
1485 | const uint8_t *loc) const { |
1486 | if ((type == R_PPC64_GOT_PCREL34 || type == R_PPC64_PCREL_OPT) && |
1487 | ctx.arg.pcRelOptimize) { |
1488 | // It only makes sense to optimize pld since paddi means that the address |
1489 | // of the object in the GOT is required rather than the object itself. |
1490 | if ((readPrefixedInst(ctx, loc) & 0xfc000000) == 0xe4000000) |
1491 | return RE_PPC64_RELAX_GOT_PC; |
1492 | } |
1493 | return R_GOT_PC; |
1494 | } |
1495 | |
1496 | // Reference: 3.7.4.1 of the 64-bit ELF V2 abi supplement. |
1497 | // The general dynamic code sequence for a global `x` uses 4 instructions. |
1498 | // Instruction Relocation Symbol |
1499 | // addis r3, r2, x@got@tlsgd@ha R_PPC64_GOT_TLSGD16_HA x |
1500 | // addi r3, r3, x@got@tlsgd@l R_PPC64_GOT_TLSGD16_LO x |
1501 | // bl __tls_get_addr(x@tlsgd) R_PPC64_TLSGD x |
1502 | // R_PPC64_REL24 __tls_get_addr |
1503 | // nop None None |
1504 | // |
1505 | // Relaxing to initial-exec entails: |
1506 | // 1) Convert the addis/addi pair that builds the address of the tls_index |
1507 | // struct for 'x' to an addis/ld pair that loads an offset from a got-entry. |
1508 | // 2) Convert the call to __tls_get_addr to a nop. |
1509 | // 3) Convert the nop following the call to an add of the loaded offset to the |
1510 | // thread pointer. |
1511 | // Since the nop must directly follow the call, the R_PPC64_TLSGD relocation is |
1512 | // used as the relaxation hint for both steps 2 and 3. |
1513 | void PPC64::relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, |
1514 | uint64_t val) const { |
1515 | switch (rel.type) { |
1516 | case R_PPC64_GOT_TLSGD16_HA: |
1517 | // This is relaxed from addis rT, r2, sym@got@tlsgd@ha to |
1518 | // addis rT, r2, sym@got@tprel@ha. |
1519 | relocateNoSym(loc, type: R_PPC64_GOT_TPREL16_HA, val); |
1520 | return; |
1521 | case R_PPC64_GOT_TLSGD16: |
1522 | case R_PPC64_GOT_TLSGD16_LO: { |
1523 | // Relax from addi r3, rA, sym@got@tlsgd@l to |
1524 | // ld r3, sym@got@tprel@l(rA) |
1525 | uint32_t ra = (readFromHalf16(ctx, loc) & (0x1f << 16)); |
1526 | writeFromHalf16(ctx, loc, insn: 0xe8600000 | ra); |
1527 | relocateNoSym(loc, type: R_PPC64_GOT_TPREL16_LO_DS, val); |
1528 | return; |
1529 | } |
1530 | case R_PPC64_GOT_TLSGD_PCREL34: { |
1531 | // Relax from paddi r3, 0, sym@got@tlsgd@pcrel, 1 to |
1532 | // pld r3, sym@got@tprel@pcrel |
1533 | writePrefixedInst(ctx, loc, insn: 0x04100000e4600000); |
1534 | relocateNoSym(loc, type: R_PPC64_GOT_TPREL_PCREL34, val); |
1535 | return; |
1536 | } |
1537 | case R_PPC64_TLSGD: { |
1538 | // PC Relative Relaxation: |
1539 | // Relax from bl __tls_get_addr@notoc(x@tlsgd) to |
1540 | // nop |
1541 | // TOC Relaxation: |
1542 | // Relax from bl __tls_get_addr(x@tlsgd) |
1543 | // nop |
1544 | // to |
1545 | // nop |
1546 | // add r3, r3, r13 |
1547 | const uintptr_t locAsInt = reinterpret_cast<uintptr_t>(loc); |
1548 | if (locAsInt % 4 == 0) { |
1549 | write32(ctx, p: loc, v: NOP); // bl __tls_get_addr(sym@tlsgd) --> nop |
1550 | write32(ctx, p: loc + 4, v: 0x7c636a14); // nop --> add r3, r3, r13 |
1551 | } else if (locAsInt % 4 == 1) { |
1552 | // bl __tls_get_addr(sym@tlsgd) --> add r3, r3, r13 |
1553 | write32(ctx, p: loc - 1, v: 0x7c636a14); |
1554 | } else { |
1555 | Err(ctx) << "R_PPC64_TLSGD has unexpected byte alignment" ; |
1556 | } |
1557 | return; |
1558 | } |
1559 | default: |
1560 | llvm_unreachable("unsupported relocation for TLS GD to IE relaxation" ); |
1561 | } |
1562 | } |
1563 | |
1564 | void PPC64::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const { |
1565 | uint64_t secAddr = sec.getOutputSection()->addr; |
1566 | if (auto *s = dyn_cast<InputSection>(Val: &sec)) |
1567 | secAddr += s->outSecOff; |
1568 | else if (auto *ehIn = dyn_cast<EhInputSection>(Val: &sec)) |
1569 | secAddr += ehIn->getParent()->outSecOff; |
1570 | uint64_t lastPPCRelaxedRelocOff = -1; |
1571 | for (const Relocation &rel : sec.relocs()) { |
1572 | uint8_t *loc = buf + rel.offset; |
1573 | const uint64_t val = sec.getRelocTargetVA(ctx, r: rel, p: secAddr + rel.offset); |
1574 | switch (rel.expr) { |
1575 | case RE_PPC64_RELAX_GOT_PC: { |
1576 | // The R_PPC64_PCREL_OPT relocation must appear immediately after |
1577 | // R_PPC64_GOT_PCREL34 in the relocations table at the same offset. |
1578 | // We can only relax R_PPC64_PCREL_OPT if we have also relaxed |
1579 | // the associated R_PPC64_GOT_PCREL34 since only the latter has an |
1580 | // associated symbol. So save the offset when relaxing R_PPC64_GOT_PCREL34 |
1581 | // and only relax the other if the saved offset matches. |
1582 | if (rel.type == R_PPC64_GOT_PCREL34) |
1583 | lastPPCRelaxedRelocOff = rel.offset; |
1584 | if (rel.type == R_PPC64_PCREL_OPT && rel.offset != lastPPCRelaxedRelocOff) |
1585 | break; |
1586 | relaxGot(loc, rel, val); |
1587 | break; |
1588 | } |
1589 | case RE_PPC64_RELAX_TOC: |
1590 | // rel.sym refers to the STT_SECTION symbol associated to the .toc input |
1591 | // section. If an R_PPC64_TOC16_LO (.toc + addend) references the TOC |
1592 | // entry, there may be R_PPC64_TOC16_HA not paired with |
1593 | // R_PPC64_TOC16_LO_DS. Don't relax. This loses some relaxation |
1594 | // opportunities but is safe. |
1595 | if (ctx.ppc64noTocRelax.count(V: {rel.sym, rel.addend}) || |
1596 | !tryRelaxPPC64TocIndirection(ctx, rel, bufLoc: loc)) |
1597 | relocate(loc, rel, val); |
1598 | break; |
1599 | case RE_PPC64_CALL: |
1600 | // If this is a call to __tls_get_addr, it may be part of a TLS |
1601 | // sequence that has been relaxed and turned into a nop. In this |
1602 | // case, we don't want to handle it as a call. |
1603 | if (read32(ctx, p: loc) == 0x60000000) // nop |
1604 | break; |
1605 | |
1606 | // Patch a nop (0x60000000) to a ld. |
1607 | if (rel.sym->needsTocRestore()) { |
1608 | // gcc/gfortran 5.4, 6.3 and earlier versions do not add nop for |
1609 | // recursive calls even if the function is preemptible. This is not |
1610 | // wrong in the common case where the function is not preempted at |
1611 | // runtime. Just ignore. |
1612 | if ((rel.offset + 8 > sec.content().size() || |
1613 | read32(ctx, p: loc + 4) != 0x60000000) && |
1614 | rel.sym->file != sec.file) { |
1615 | // Use substr(6) to remove the "__plt_" prefix. |
1616 | Err(ctx) << getErrorLoc(ctx, loc) << "call to " |
1617 | << toStr(ctx, *rel.sym).substr(pos: 6) |
1618 | << " lacks nop, can't restore toc" ; |
1619 | break; |
1620 | } |
1621 | write32(ctx, p: loc + 4, v: 0xe8410018); // ld %r2, 24(%r1) |
1622 | } |
1623 | relocate(loc, rel, val); |
1624 | break; |
1625 | case R_RELAX_TLS_GD_TO_IE: |
1626 | case R_RELAX_TLS_GD_TO_IE_GOT_OFF: |
1627 | relaxTlsGdToIe(loc, rel, val); |
1628 | break; |
1629 | case R_RELAX_TLS_GD_TO_LE: |
1630 | relaxTlsGdToLe(loc, rel, val); |
1631 | break; |
1632 | case R_RELAX_TLS_LD_TO_LE_ABS: |
1633 | relaxTlsLdToLe(loc, rel, val); |
1634 | break; |
1635 | case R_RELAX_TLS_IE_TO_LE: |
1636 | relaxTlsIeToLe(loc, rel, val); |
1637 | break; |
1638 | default: |
1639 | relocate(loc, rel, val); |
1640 | break; |
1641 | } |
1642 | } |
1643 | } |
1644 | |
1645 | // The prologue for a split-stack function is expected to look roughly |
1646 | // like this: |
1647 | // .Lglobal_entry_point: |
1648 | // # TOC pointer initialization. |
1649 | // ... |
1650 | // .Llocal_entry_point: |
1651 | // # load the __private_ss member of the threads tcbhead. |
1652 | // ld r0,-0x7000-64(r13) |
1653 | // # subtract the functions stack size from the stack pointer. |
1654 | // addis r12, r1, ha(-stack-frame size) |
1655 | // addi r12, r12, l(-stack-frame size) |
1656 | // # compare needed to actual and branch to allocate_more_stack if more |
1657 | // # space is needed, otherwise fallthrough to 'normal' function body. |
1658 | // cmpld cr7,r12,r0 |
1659 | // blt- cr7, .Lallocate_more_stack |
1660 | // |
1661 | // -) The allocate_more_stack block might be placed after the split-stack |
1662 | // prologue and the `blt-` replaced with a `bge+ .Lnormal_func_body` |
1663 | // instead. |
1664 | // -) If either the addis or addi is not needed due to the stack size being |
1665 | // smaller then 32K or a multiple of 64K they will be replaced with a nop, |
1666 | // but there will always be 2 instructions the linker can overwrite for the |
1667 | // adjusted stack size. |
1668 | // |
1669 | // The linkers job here is to increase the stack size used in the addis/addi |
1670 | // pair by split-stack-size-adjust. |
1671 | // addis r12, r1, ha(-stack-frame size - split-stack-adjust-size) |
1672 | // addi r12, r12, l(-stack-frame size - split-stack-adjust-size) |
1673 | bool PPC64::adjustPrologueForCrossSplitStack(uint8_t *loc, uint8_t *end, |
1674 | uint8_t stOther) const { |
1675 | // If the caller has a global entry point adjust the buffer past it. The start |
1676 | // of the split-stack prologue will be at the local entry point. |
1677 | loc += getPPC64GlobalEntryToLocalEntryOffset(ctx, stOther); |
1678 | |
1679 | // At the very least we expect to see a load of some split-stack data from the |
1680 | // tcb, and 2 instructions that calculate the ending stack address this |
1681 | // function will require. If there is not enough room for at least 3 |
1682 | // instructions it can't be a split-stack prologue. |
1683 | if (loc + 12 >= end) |
1684 | return false; |
1685 | |
1686 | // First instruction must be `ld r0, -0x7000-64(r13)` |
1687 | if (read32(ctx, p: loc) != 0xe80d8fc0) |
1688 | return false; |
1689 | |
1690 | int16_t hiImm = 0; |
1691 | int16_t loImm = 0; |
1692 | // First instruction can be either an addis if the frame size is larger then |
1693 | // 32K, or an addi if the size is less then 32K. |
1694 | int32_t firstInstr = read32(ctx, p: loc + 4); |
1695 | if (getPrimaryOpCode(encoding: firstInstr) == 15) { |
1696 | hiImm = firstInstr & 0xFFFF; |
1697 | } else if (getPrimaryOpCode(encoding: firstInstr) == 14) { |
1698 | loImm = firstInstr & 0xFFFF; |
1699 | } else { |
1700 | return false; |
1701 | } |
1702 | |
1703 | // Second instruction is either an addi or a nop. If the first instruction was |
1704 | // an addi then LoImm is set and the second instruction must be a nop. |
1705 | uint32_t secondInstr = read32(ctx, p: loc + 8); |
1706 | if (!loImm && getPrimaryOpCode(encoding: secondInstr) == 14) { |
1707 | loImm = secondInstr & 0xFFFF; |
1708 | } else if (secondInstr != NOP) { |
1709 | return false; |
1710 | } |
1711 | |
1712 | // The register operands of the first instruction should be the stack-pointer |
1713 | // (r1) as the input (RA) and r12 as the output (RT). If the second |
1714 | // instruction is not a nop, then it should use r12 as both input and output. |
1715 | auto checkRegOperands = [](uint32_t instr, uint8_t expectedRT, |
1716 | uint8_t expectedRA) { |
1717 | return ((instr & 0x3E00000) >> 21 == expectedRT) && |
1718 | ((instr & 0x1F0000) >> 16 == expectedRA); |
1719 | }; |
1720 | if (!checkRegOperands(firstInstr, 12, 1)) |
1721 | return false; |
1722 | if (secondInstr != NOP && !checkRegOperands(secondInstr, 12, 12)) |
1723 | return false; |
1724 | |
1725 | int32_t stackFrameSize = (hiImm * 65536) + loImm; |
1726 | // Check that the adjusted size doesn't overflow what we can represent with 2 |
1727 | // instructions. |
1728 | if (stackFrameSize < ctx.arg.splitStackAdjustSize + INT32_MIN) { |
1729 | Err(ctx) << getErrorLoc(ctx, loc) |
1730 | << "split-stack prologue adjustment overflows" ; |
1731 | return false; |
1732 | } |
1733 | |
1734 | int32_t adjustedStackFrameSize = |
1735 | stackFrameSize - ctx.arg.splitStackAdjustSize; |
1736 | |
1737 | loImm = adjustedStackFrameSize & 0xFFFF; |
1738 | hiImm = (adjustedStackFrameSize + 0x8000) >> 16; |
1739 | if (hiImm) { |
1740 | write32(ctx, p: loc + 4, v: 0x3d810000 | (uint16_t)hiImm); |
1741 | // If the low immediate is zero the second instruction will be a nop. |
1742 | secondInstr = loImm ? 0x398C0000 | (uint16_t)loImm : NOP; |
1743 | write32(ctx, p: loc + 8, v: secondInstr); |
1744 | } else { |
1745 | // addi r12, r1, imm |
1746 | write32(ctx, p: loc + 4, v: (0x39810000) | (uint16_t)loImm); |
1747 | write32(ctx, p: loc + 8, v: NOP); |
1748 | } |
1749 | |
1750 | return true; |
1751 | } |
1752 | |
1753 | void elf::setPPC64TargetInfo(Ctx &ctx) { ctx.target.reset(p: new PPC64(ctx)); } |
1754 | |