1 //===------------- OrcABISupport.cpp - ABI specific support code ----------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9
10 #include "llvm/ExecutionEngine/Orc/OrcABISupport.h"
11 #include "llvm/ADT/Triple.h"
12 #include "llvm/Support/Process.h"
13
14 namespace llvm {
15 namespace orc {
16
writeResolverCode(uint8_t * ResolverMem,JITReentryFn ReentryFn,void * CallbackMgr)17 void OrcAArch64::writeResolverCode(uint8_t *ResolverMem, JITReentryFn ReentryFn,
18 void *CallbackMgr) {
19
20 const uint32_t ResolverCode[] = {
21 // resolver_entry:
22 0xa9bf47fd, // 0x000: stp x29, x17, [sp, #-16]!
23 0x910003fd, // 0x004: mov x29, sp
24 0xa9bf73fb, // 0x008: stp x27, x28, [sp, #-16]!
25 0xa9bf6bf9, // 0x00c: stp x25, x26, [sp, #-16]!
26 0xa9bf63f7, // 0x010: stp x23, x24, [sp, #-16]!
27 0xa9bf5bf5, // 0x014: stp x21, x22, [sp, #-16]!
28 0xa9bf53f3, // 0x018: stp x19, x20, [sp, #-16]!
29 0xa9bf3fee, // 0x01c: stp x14, x15, [sp, #-16]!
30 0xa9bf37ec, // 0x020: stp x12, x13, [sp, #-16]!
31 0xa9bf2fea, // 0x024: stp x10, x11, [sp, #-16]!
32 0xa9bf27e8, // 0x028: stp x8, x9, [sp, #-16]!
33 0xa9bf1fe6, // 0x02c: stp x6, x7, [sp, #-16]!
34 0xa9bf17e4, // 0x030: stp x4, x5, [sp, #-16]!
35 0xa9bf0fe2, // 0x034: stp x2, x3, [sp, #-16]!
36 0xa9bf07e0, // 0x038: stp x0, x1, [sp, #-16]!
37 0xadbf7ffe, // 0x03c: stp q30, q31, [sp, #-32]!
38 0xadbf77fc, // 0x040: stp q28, q29, [sp, #-32]!
39 0xadbf6ffa, // 0x044: stp q26, q27, [sp, #-32]!
40 0xadbf67f8, // 0x048: stp q24, q25, [sp, #-32]!
41 0xadbf5ff6, // 0x04c: stp q22, q23, [sp, #-32]!
42 0xadbf57f4, // 0x050: stp q20, q21, [sp, #-32]!
43 0xadbf4ff2, // 0x054: stp q18, q19, [sp, #-32]!
44 0xadbf47f0, // 0x058: stp q16, q17, [sp, #-32]!
45 0xadbf3fee, // 0x05c: stp q14, q15, [sp, #-32]!
46 0xadbf37ec, // 0x060: stp q12, q13, [sp, #-32]!
47 0xadbf2fea, // 0x064: stp q10, q11, [sp, #-32]!
48 0xadbf27e8, // 0x068: stp q8, q9, [sp, #-32]!
49 0xadbf1fe6, // 0x06c: stp q6, q7, [sp, #-32]!
50 0xadbf17e4, // 0x070: stp q4, q5, [sp, #-32]!
51 0xadbf0fe2, // 0x074: stp q2, q3, [sp, #-32]!
52 0xadbf07e0, // 0x078: stp q0, q1, [sp, #-32]!
53 0x580004e0, // 0x07c: ldr x0, Lcallbackmgr
54 0xaa1e03e1, // 0x080: mov x1, x30
55 0xd1003021, // 0x084: sub x1, x1, #12
56 0x58000442, // 0x088: ldr x2, Lreentry_fn_ptr
57 0xd63f0040, // 0x08c: blr x2
58 0xaa0003f1, // 0x090: mov x17, x0
59 0xacc107e0, // 0x094: ldp q0, q1, [sp], #32
60 0xacc10fe2, // 0x098: ldp q2, q3, [sp], #32
61 0xacc117e4, // 0x09c: ldp q4, q5, [sp], #32
62 0xacc11fe6, // 0x0a0: ldp q6, q7, [sp], #32
63 0xacc127e8, // 0x0a4: ldp q8, q9, [sp], #32
64 0xacc12fea, // 0x0a8: ldp q10, q11, [sp], #32
65 0xacc137ec, // 0x0ac: ldp q12, q13, [sp], #32
66 0xacc13fee, // 0x0b0: ldp q14, q15, [sp], #32
67 0xacc147f0, // 0x0b4: ldp q16, q17, [sp], #32
68 0xacc14ff2, // 0x0b8: ldp q18, q19, [sp], #32
69 0xacc157f4, // 0x0bc: ldp q20, q21, [sp], #32
70 0xacc15ff6, // 0x0c0: ldp q22, q23, [sp], #32
71 0xacc167f8, // 0x0c4: ldp q24, q25, [sp], #32
72 0xacc16ffa, // 0x0c8: ldp q26, q27, [sp], #32
73 0xacc177fc, // 0x0cc: ldp q28, q29, [sp], #32
74 0xacc17ffe, // 0x0d0: ldp q30, q31, [sp], #32
75 0xa8c107e0, // 0x0d4: ldp x0, x1, [sp], #16
76 0xa8c10fe2, // 0x0d8: ldp x2, x3, [sp], #16
77 0xa8c117e4, // 0x0dc: ldp x4, x5, [sp], #16
78 0xa8c11fe6, // 0x0e0: ldp x6, x7, [sp], #16
79 0xa8c127e8, // 0x0e4: ldp x8, x9, [sp], #16
80 0xa8c12fea, // 0x0e8: ldp x10, x11, [sp], #16
81 0xa8c137ec, // 0x0ec: ldp x12, x13, [sp], #16
82 0xa8c13fee, // 0x0f0: ldp x14, x15, [sp], #16
83 0xa8c153f3, // 0x0f4: ldp x19, x20, [sp], #16
84 0xa8c15bf5, // 0x0f8: ldp x21, x22, [sp], #16
85 0xa8c163f7, // 0x0fc: ldp x23, x24, [sp], #16
86 0xa8c16bf9, // 0x100: ldp x25, x26, [sp], #16
87 0xa8c173fb, // 0x104: ldp x27, x28, [sp], #16
88 0xa8c17bfd, // 0x108: ldp x29, x30, [sp], #16
89 0xd65f0220, // 0x10c: ret x17
90 0x01234567, // 0x110: Lreentry_fn_ptr:
91 0xdeadbeef, // 0x114: .quad 0
92 0x98765432, // 0x118: Lcallbackmgr:
93 0xcafef00d // 0x11c: .quad 0
94 };
95
96 const unsigned ReentryFnAddrOffset = 0x110;
97 const unsigned CallbackMgrAddrOffset = 0x118;
98
99 memcpy(ResolverMem, ResolverCode, sizeof(ResolverCode));
100 memcpy(ResolverMem + ReentryFnAddrOffset, &ReentryFn, sizeof(ReentryFn));
101 memcpy(ResolverMem + CallbackMgrAddrOffset, &CallbackMgr,
102 sizeof(CallbackMgr));
103 }
104
writeTrampolines(uint8_t * TrampolineMem,void * ResolverAddr,unsigned NumTrampolines)105 void OrcAArch64::writeTrampolines(uint8_t *TrampolineMem, void *ResolverAddr,
106 unsigned NumTrampolines) {
107
108 unsigned OffsetToPtr = alignTo(NumTrampolines * TrampolineSize, 8);
109
110 memcpy(TrampolineMem + OffsetToPtr, &ResolverAddr, sizeof(void *));
111
112 // OffsetToPtr is actually the offset from the PC for the 2nd instruction, so
113 // subtract 32-bits.
114 OffsetToPtr -= 4;
115
116 uint32_t *Trampolines = reinterpret_cast<uint32_t *>(TrampolineMem);
117
118 for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize) {
119 Trampolines[3 * I + 0] = 0xaa1e03f1; // mov x17, x30
120 Trampolines[3 * I + 1] = 0x58000010 | (OffsetToPtr << 3); // mov x16, Lptr
121 Trampolines[3 * I + 2] = 0xd63f0200; // blr x16
122 }
123
124 }
125
emitIndirectStubsBlock(IndirectStubsInfo & StubsInfo,unsigned MinStubs,void * InitialPtrVal)126 Error OrcAArch64::emitIndirectStubsBlock(IndirectStubsInfo &StubsInfo,
127 unsigned MinStubs,
128 void *InitialPtrVal) {
129 // Stub format is:
130 //
131 // .section __orc_stubs
132 // stub1:
133 // ldr x0, ptr1 ; PC-rel load of ptr1
134 // br x0 ; Jump to resolver
135 // stub2:
136 // ldr x0, ptr2 ; PC-rel load of ptr2
137 // br x0 ; Jump to resolver
138 //
139 // ...
140 //
141 // .section __orc_ptrs
142 // ptr1:
143 // .quad 0x0
144 // ptr2:
145 // .quad 0x0
146 //
147 // ...
148
149 const unsigned StubSize = IndirectStubsInfo::StubSize;
150
151 // Emit at least MinStubs, rounded up to fill the pages allocated.
152 unsigned PageSize = sys::Process::getPageSize();
153 unsigned NumPages = ((MinStubs * StubSize) + (PageSize - 1)) / PageSize;
154 unsigned NumStubs = (NumPages * PageSize) / StubSize;
155
156 // Allocate memory for stubs and pointers in one call.
157 std::error_code EC;
158 auto StubsMem = sys::OwningMemoryBlock(sys::Memory::allocateMappedMemory(
159 2 * NumPages * PageSize, nullptr,
160 sys::Memory::MF_READ | sys::Memory::MF_WRITE, EC));
161
162 if (EC)
163 return errorCodeToError(EC);
164
165 // Create separate MemoryBlocks representing the stubs and pointers.
166 sys::MemoryBlock StubsBlock(StubsMem.base(), NumPages * PageSize);
167 sys::MemoryBlock PtrsBlock(static_cast<char *>(StubsMem.base()) +
168 NumPages * PageSize,
169 NumPages * PageSize);
170
171 // Populate the stubs page stubs and mark it executable.
172 uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlock.base());
173 uint64_t PtrOffsetField = static_cast<uint64_t>(NumPages * PageSize)
174 << 3;
175
176 for (unsigned I = 0; I < NumStubs; ++I)
177 Stub[I] = 0xd61f020058000010 | PtrOffsetField;
178
179 if (auto EC = sys::Memory::protectMappedMemory(
180 StubsBlock, sys::Memory::MF_READ | sys::Memory::MF_EXEC))
181 return errorCodeToError(EC);
182
183 // Initialize all pointers to point at FailureAddress.
184 void **Ptr = reinterpret_cast<void **>(PtrsBlock.base());
185 for (unsigned I = 0; I < NumStubs; ++I)
186 Ptr[I] = InitialPtrVal;
187
188 StubsInfo = IndirectStubsInfo(NumStubs, std::move(StubsMem));
189
190 return Error::success();
191 }
192
writeTrampolines(uint8_t * TrampolineMem,void * ResolverAddr,unsigned NumTrampolines)193 void OrcX86_64_Base::writeTrampolines(uint8_t *TrampolineMem,
194 void *ResolverAddr,
195 unsigned NumTrampolines) {
196
197 unsigned OffsetToPtr = NumTrampolines * TrampolineSize;
198
199 memcpy(TrampolineMem + OffsetToPtr, &ResolverAddr, sizeof(void *));
200
201 uint64_t *Trampolines = reinterpret_cast<uint64_t *>(TrampolineMem);
202 uint64_t CallIndirPCRel = 0xf1c40000000015ff;
203
204 for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize)
205 Trampolines[I] = CallIndirPCRel | ((OffsetToPtr - 6) << 16);
206 }
207
emitIndirectStubsBlock(IndirectStubsInfo & StubsInfo,unsigned MinStubs,void * InitialPtrVal)208 Error OrcX86_64_Base::emitIndirectStubsBlock(IndirectStubsInfo &StubsInfo,
209 unsigned MinStubs,
210 void *InitialPtrVal) {
211 // Stub format is:
212 //
213 // .section __orc_stubs
214 // stub1:
215 // jmpq *ptr1(%rip)
216 // .byte 0xC4 ; <- Invalid opcode padding.
217 // .byte 0xF1
218 // stub2:
219 // jmpq *ptr2(%rip)
220 //
221 // ...
222 //
223 // .section __orc_ptrs
224 // ptr1:
225 // .quad 0x0
226 // ptr2:
227 // .quad 0x0
228 //
229 // ...
230
231 const unsigned StubSize = IndirectStubsInfo::StubSize;
232
233 // Emit at least MinStubs, rounded up to fill the pages allocated.
234 unsigned PageSize = sys::Process::getPageSize();
235 unsigned NumPages = ((MinStubs * StubSize) + (PageSize - 1)) / PageSize;
236 unsigned NumStubs = (NumPages * PageSize) / StubSize;
237
238 // Allocate memory for stubs and pointers in one call.
239 std::error_code EC;
240 auto StubsMem = sys::OwningMemoryBlock(sys::Memory::allocateMappedMemory(
241 2 * NumPages * PageSize, nullptr,
242 sys::Memory::MF_READ | sys::Memory::MF_WRITE, EC));
243
244 if (EC)
245 return errorCodeToError(EC);
246
247 // Create separate MemoryBlocks representing the stubs and pointers.
248 sys::MemoryBlock StubsBlock(StubsMem.base(), NumPages * PageSize);
249 sys::MemoryBlock PtrsBlock(static_cast<char *>(StubsMem.base()) +
250 NumPages * PageSize,
251 NumPages * PageSize);
252
253 // Populate the stubs page stubs and mark it executable.
254 uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlock.base());
255 uint64_t PtrOffsetField = static_cast<uint64_t>(NumPages * PageSize - 6)
256 << 16;
257 for (unsigned I = 0; I < NumStubs; ++I)
258 Stub[I] = 0xF1C40000000025ff | PtrOffsetField;
259
260 if (auto EC = sys::Memory::protectMappedMemory(
261 StubsBlock, sys::Memory::MF_READ | sys::Memory::MF_EXEC))
262 return errorCodeToError(EC);
263
264 // Initialize all pointers to point at FailureAddress.
265 void **Ptr = reinterpret_cast<void **>(PtrsBlock.base());
266 for (unsigned I = 0; I < NumStubs; ++I)
267 Ptr[I] = InitialPtrVal;
268
269 StubsInfo = IndirectStubsInfo(NumStubs, std::move(StubsMem));
270
271 return Error::success();
272 }
273
writeResolverCode(uint8_t * ResolverMem,JITReentryFn ReentryFn,void * CallbackMgr)274 void OrcX86_64_SysV::writeResolverCode(uint8_t *ResolverMem,
275 JITReentryFn ReentryFn,
276 void *CallbackMgr) {
277
278 const uint8_t ResolverCode[] = {
279 // resolver_entry:
280 0x55, // 0x00: pushq %rbp
281 0x48, 0x89, 0xe5, // 0x01: movq %rsp, %rbp
282 0x50, // 0x04: pushq %rax
283 0x53, // 0x05: pushq %rbx
284 0x51, // 0x06: pushq %rcx
285 0x52, // 0x07: pushq %rdx
286 0x56, // 0x08: pushq %rsi
287 0x57, // 0x09: pushq %rdi
288 0x41, 0x50, // 0x0a: pushq %r8
289 0x41, 0x51, // 0x0c: pushq %r9
290 0x41, 0x52, // 0x0e: pushq %r10
291 0x41, 0x53, // 0x10: pushq %r11
292 0x41, 0x54, // 0x12: pushq %r12
293 0x41, 0x55, // 0x14: pushq %r13
294 0x41, 0x56, // 0x16: pushq %r14
295 0x41, 0x57, // 0x18: pushq %r15
296 0x48, 0x81, 0xec, 0x08, 0x02, 0x00, 0x00, // 0x1a: subq 0x208, %rsp
297 0x48, 0x0f, 0xae, 0x04, 0x24, // 0x21: fxsave64 (%rsp)
298 0x48, 0xbf, // 0x26: movabsq <CBMgr>, %rdi
299
300 // 0x28: Callback manager addr.
301 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
302
303 0x48, 0x8b, 0x75, 0x08, // 0x30: movq 8(%rbp), %rsi
304 0x48, 0x83, 0xee, 0x06, // 0x34: subq $6, %rsi
305 0x48, 0xb8, // 0x38: movabsq <REntry>, %rax
306
307 // 0x3a: JIT re-entry fn addr:
308 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
309
310 0xff, 0xd0, // 0x42: callq *%rax
311 0x48, 0x89, 0x45, 0x08, // 0x44: movq %rax, 8(%rbp)
312 0x48, 0x0f, 0xae, 0x0c, 0x24, // 0x48: fxrstor64 (%rsp)
313 0x48, 0x81, 0xc4, 0x08, 0x02, 0x00, 0x00, // 0x4d: addq 0x208, %rsp
314 0x41, 0x5f, // 0x54: popq %r15
315 0x41, 0x5e, // 0x56: popq %r14
316 0x41, 0x5d, // 0x58: popq %r13
317 0x41, 0x5c, // 0x5a: popq %r12
318 0x41, 0x5b, // 0x5c: popq %r11
319 0x41, 0x5a, // 0x5e: popq %r10
320 0x41, 0x59, // 0x60: popq %r9
321 0x41, 0x58, // 0x62: popq %r8
322 0x5f, // 0x64: popq %rdi
323 0x5e, // 0x65: popq %rsi
324 0x5a, // 0x66: popq %rdx
325 0x59, // 0x67: popq %rcx
326 0x5b, // 0x68: popq %rbx
327 0x58, // 0x69: popq %rax
328 0x5d, // 0x6a: popq %rbp
329 0xc3, // 0x6b: retq
330 };
331
332 const unsigned ReentryFnAddrOffset = 0x3a;
333 const unsigned CallbackMgrAddrOffset = 0x28;
334
335 memcpy(ResolverMem, ResolverCode, sizeof(ResolverCode));
336 memcpy(ResolverMem + ReentryFnAddrOffset, &ReentryFn, sizeof(ReentryFn));
337 memcpy(ResolverMem + CallbackMgrAddrOffset, &CallbackMgr,
338 sizeof(CallbackMgr));
339 }
340
writeResolverCode(uint8_t * ResolverMem,JITReentryFn ReentryFn,void * CallbackMgr)341 void OrcX86_64_Win32::writeResolverCode(uint8_t *ResolverMem,
342 JITReentryFn ReentryFn,
343 void *CallbackMgr) {
344
345 // resolverCode is similar to OrcX86_64 with differences specific to windows x64 calling convention:
346 // arguments go into rcx, rdx and come in reverse order, shadow space allocation on stack
347 const uint8_t ResolverCode[] = {
348 // resolver_entry:
349 0x55, // 0x00: pushq %rbp
350 0x48, 0x89, 0xe5, // 0x01: movq %rsp, %rbp
351 0x50, // 0x04: pushq %rax
352 0x53, // 0x05: pushq %rbx
353 0x51, // 0x06: pushq %rcx
354 0x52, // 0x07: pushq %rdx
355 0x56, // 0x08: pushq %rsi
356 0x57, // 0x09: pushq %rdi
357 0x41, 0x50, // 0x0a: pushq %r8
358 0x41, 0x51, // 0x0c: pushq %r9
359 0x41, 0x52, // 0x0e: pushq %r10
360 0x41, 0x53, // 0x10: pushq %r11
361 0x41, 0x54, // 0x12: pushq %r12
362 0x41, 0x55, // 0x14: pushq %r13
363 0x41, 0x56, // 0x16: pushq %r14
364 0x41, 0x57, // 0x18: pushq %r15
365 0x48, 0x81, 0xec, 0x08, 0x02, 0x00, 0x00, // 0x1a: subq 0x208, %rsp
366 0x48, 0x0f, 0xae, 0x04, 0x24, // 0x21: fxsave64 (%rsp)
367
368 0x48, 0xb9, // 0x26: movabsq <CBMgr>, %rcx
369 // 0x28: Callback manager addr.
370 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
371
372 0x48, 0x8B, 0x55, 0x08, // 0x30: mov rdx, [rbp+0x8]
373 0x48, 0x83, 0xea, 0x06, // 0x34: sub rdx, 0x6
374
375 0x48, 0xb8, // 0x38: movabsq <REntry>, %rax
376 // 0x3a: JIT re-entry fn addr:
377 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
378
379 // 0x42: sub rsp, 0x20 (Allocate shadow space)
380 0x48, 0x83, 0xEC, 0x20,
381 0xff, 0xd0, // 0x46: callq *%rax
382
383 // 0x48: add rsp, 0x20 (Free shadow space)
384 0x48, 0x83, 0xC4, 0x20,
385
386 0x48, 0x89, 0x45, 0x08, // 0x4C: movq %rax, 8(%rbp)
387 0x48, 0x0f, 0xae, 0x0c, 0x24, // 0x50: fxrstor64 (%rsp)
388 0x48, 0x81, 0xc4, 0x08, 0x02, 0x00, 0x00, // 0x55: addq 0x208, %rsp
389 0x41, 0x5f, // 0x5C: popq %r15
390 0x41, 0x5e, // 0x5E: popq %r14
391 0x41, 0x5d, // 0x60: popq %r13
392 0x41, 0x5c, // 0x62: popq %r12
393 0x41, 0x5b, // 0x64: popq %r11
394 0x41, 0x5a, // 0x66: popq %r10
395 0x41, 0x59, // 0x68: popq %r9
396 0x41, 0x58, // 0x6a: popq %r8
397 0x5f, // 0x6c: popq %rdi
398 0x5e, // 0x6d: popq %rsi
399 0x5a, // 0x6e: popq %rdx
400 0x59, // 0x6f: popq %rcx
401 0x5b, // 0x70: popq %rbx
402 0x58, // 0x71: popq %rax
403 0x5d, // 0x72: popq %rbp
404 0xc3, // 0x73: retq
405 };
406
407
408 const unsigned ReentryFnAddrOffset = 0x3a;
409 const unsigned CallbackMgrAddrOffset = 0x28;
410
411 memcpy(ResolverMem, ResolverCode, sizeof(ResolverCode));
412 memcpy(ResolverMem + ReentryFnAddrOffset, &ReentryFn, sizeof(ReentryFn));
413 memcpy(ResolverMem + CallbackMgrAddrOffset, &CallbackMgr,
414 sizeof(CallbackMgr));
415 }
416
writeResolverCode(uint8_t * ResolverMem,JITReentryFn ReentryFn,void * CallbackMgr)417 void OrcI386::writeResolverCode(uint8_t *ResolverMem, JITReentryFn ReentryFn,
418 void *CallbackMgr) {
419
420 const uint8_t ResolverCode[] = {
421 // resolver_entry:
422 0x55, // 0x00: pushl %ebp
423 0x89, 0xe5, // 0x01: movl %esp, %ebp
424 0x54, // 0x03: pushl %esp
425 0x83, 0xe4, 0xf0, // 0x04: andl $-0x10, %esp
426 0x50, // 0x07: pushl %eax
427 0x53, // 0x08: pushl %ebx
428 0x51, // 0x09: pushl %ecx
429 0x52, // 0x0a: pushl %edx
430 0x56, // 0x0b: pushl %esi
431 0x57, // 0x0c: pushl %edi
432 0x81, 0xec, 0x18, 0x02, 0x00, 0x00, // 0x0d: subl $0x218, %esp
433 0x0f, 0xae, 0x44, 0x24, 0x10, // 0x13: fxsave 0x10(%esp)
434 0x8b, 0x75, 0x04, // 0x18: movl 0x4(%ebp), %esi
435 0x83, 0xee, 0x05, // 0x1b: subl $0x5, %esi
436 0x89, 0x74, 0x24, 0x04, // 0x1e: movl %esi, 0x4(%esp)
437 0xc7, 0x04, 0x24, 0x00, 0x00, 0x00,
438 0x00, // 0x22: movl <cbmgr>, (%esp)
439 0xb8, 0x00, 0x00, 0x00, 0x00, // 0x29: movl <reentry>, %eax
440 0xff, 0xd0, // 0x2e: calll *%eax
441 0x89, 0x45, 0x04, // 0x30: movl %eax, 0x4(%ebp)
442 0x0f, 0xae, 0x4c, 0x24, 0x10, // 0x33: fxrstor 0x10(%esp)
443 0x81, 0xc4, 0x18, 0x02, 0x00, 0x00, // 0x38: addl $0x218, %esp
444 0x5f, // 0x3e: popl %edi
445 0x5e, // 0x3f: popl %esi
446 0x5a, // 0x40: popl %edx
447 0x59, // 0x41: popl %ecx
448 0x5b, // 0x42: popl %ebx
449 0x58, // 0x43: popl %eax
450 0x8b, 0x65, 0xfc, // 0x44: movl -0x4(%ebp), %esp
451 0x5d, // 0x48: popl %ebp
452 0xc3 // 0x49: retl
453 };
454
455 const unsigned ReentryFnAddrOffset = 0x2a;
456 const unsigned CallbackMgrAddrOffset = 0x25;
457
458 memcpy(ResolverMem, ResolverCode, sizeof(ResolverCode));
459 memcpy(ResolverMem + ReentryFnAddrOffset, &ReentryFn, sizeof(ReentryFn));
460 memcpy(ResolverMem + CallbackMgrAddrOffset, &CallbackMgr,
461 sizeof(CallbackMgr));
462 }
463
writeTrampolines(uint8_t * TrampolineMem,void * ResolverAddr,unsigned NumTrampolines)464 void OrcI386::writeTrampolines(uint8_t *TrampolineMem, void *ResolverAddr,
465 unsigned NumTrampolines) {
466
467 uint64_t CallRelImm = 0xF1C4C400000000e8;
468 uint64_t Resolver = reinterpret_cast<uint64_t>(ResolverAddr);
469 uint64_t ResolverRel =
470 Resolver - reinterpret_cast<uint64_t>(TrampolineMem) - 5;
471
472 uint64_t *Trampolines = reinterpret_cast<uint64_t *>(TrampolineMem);
473 for (unsigned I = 0; I < NumTrampolines; ++I, ResolverRel -= TrampolineSize)
474 Trampolines[I] = CallRelImm | (ResolverRel << 8);
475 }
476
emitIndirectStubsBlock(IndirectStubsInfo & StubsInfo,unsigned MinStubs,void * InitialPtrVal)477 Error OrcI386::emitIndirectStubsBlock(IndirectStubsInfo &StubsInfo,
478 unsigned MinStubs, void *InitialPtrVal) {
479 // Stub format is:
480 //
481 // .section __orc_stubs
482 // stub1:
483 // jmpq *ptr1
484 // .byte 0xC4 ; <- Invalid opcode padding.
485 // .byte 0xF1
486 // stub2:
487 // jmpq *ptr2
488 //
489 // ...
490 //
491 // .section __orc_ptrs
492 // ptr1:
493 // .quad 0x0
494 // ptr2:
495 // .quad 0x0
496 //
497 // ...
498
499 const unsigned StubSize = IndirectStubsInfo::StubSize;
500
501 // Emit at least MinStubs, rounded up to fill the pages allocated.
502 unsigned PageSize = sys::Process::getPageSize();
503 unsigned NumPages = ((MinStubs * StubSize) + (PageSize - 1)) / PageSize;
504 unsigned NumStubs = (NumPages * PageSize) / StubSize;
505
506 // Allocate memory for stubs and pointers in one call.
507 std::error_code EC;
508 auto StubsMem = sys::OwningMemoryBlock(sys::Memory::allocateMappedMemory(
509 2 * NumPages * PageSize, nullptr,
510 sys::Memory::MF_READ | sys::Memory::MF_WRITE, EC));
511
512 if (EC)
513 return errorCodeToError(EC);
514
515 // Create separate MemoryBlocks representing the stubs and pointers.
516 sys::MemoryBlock StubsBlock(StubsMem.base(), NumPages * PageSize);
517 sys::MemoryBlock PtrsBlock(static_cast<char *>(StubsMem.base()) +
518 NumPages * PageSize,
519 NumPages * PageSize);
520
521 // Populate the stubs page stubs and mark it executable.
522 uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlock.base());
523 uint64_t PtrAddr = reinterpret_cast<uint64_t>(PtrsBlock.base());
524 for (unsigned I = 0; I < NumStubs; ++I, PtrAddr += 4)
525 Stub[I] = 0xF1C40000000025ff | (PtrAddr << 16);
526
527 if (auto EC = sys::Memory::protectMappedMemory(
528 StubsBlock, sys::Memory::MF_READ | sys::Memory::MF_EXEC))
529 return errorCodeToError(EC);
530
531 // Initialize all pointers to point at FailureAddress.
532 void **Ptr = reinterpret_cast<void **>(PtrsBlock.base());
533 for (unsigned I = 0; I < NumStubs; ++I)
534 Ptr[I] = InitialPtrVal;
535
536 StubsInfo = IndirectStubsInfo(NumStubs, std::move(StubsMem));
537
538 return Error::success();
539 }
540
541 } // End namespace orc.
542 } // End namespace llvm.
543