?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
AMDGPU: Mark test as XFAIL in expensive_checks builds
[llvm-project.git] / llvm / lib / ExecutionEngine / Orc / OrcABISupport.cpp
blobabf3931d94d1f4f603800a0b23db2df34949b035
1 //===------------- OrcABISupport.cpp - ABI specific support code ----------===//
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 "llvm/ExecutionEngine/Orc/OrcABISupport.h"
10 #include "llvm/Support/FormatVariadic.h"
11 #include "llvm/Support/raw_ostream.h"
12
13 #define DEBUG_TYPE "orc"
14
15 using namespace llvm;
16 using namespace llvm::orc;
17
18 template <typename ORCABI>
19 static bool stubAndPointerRangesOk(ExecutorAddr StubBlockAddr,
20 ExecutorAddr PointerBlockAddr,
21 unsigned NumStubs) {
22 constexpr unsigned MaxDisp = ORCABI::StubToPointerMaxDisplacement;
23 ExecutorAddr FirstStub = StubBlockAddr;
24 ExecutorAddr LastStub = FirstStub + ((NumStubs - 1) * ORCABI::StubSize);
25 ExecutorAddr FirstPointer = PointerBlockAddr;
26 ExecutorAddr LastPointer = FirstPointer + ((NumStubs - 1) * ORCABI::StubSize);
27
28 if (FirstStub < FirstPointer) {
29 if (LastStub >= FirstPointer)
30 return false; // Ranges overlap.
31 return (FirstPointer - FirstStub <= MaxDisp) &&
32 (LastPointer - LastStub <= MaxDisp); // out-of-range.
33 }
34
35 if (LastPointer >= FirstStub)
36 return false; // Ranges overlap.
37
38 return (FirstStub - FirstPointer <= MaxDisp) &&
39 (LastStub - LastPointer <= MaxDisp);
40 }
41
42 namespace llvm {
43 namespace orc {
44
45 void OrcAArch64::writeResolverCode(char *ResolverWorkingMem,
46 ExecutorAddr ResolverTargetAddress,
47 ExecutorAddr ReentryFnAddr,
48 ExecutorAddr ReentryCtxAddr) {
49
50 const uint32_t ResolverCode[] = {
51 // resolver_entry:
52 0xa9bf47fd, // 0x000: stp x29, x17, [sp, #-16]!
53 0x910003fd, // 0x004: mov x29, sp
54 0xa9bf73fb, // 0x008: stp x27, x28, [sp, #-16]!
55 0xa9bf6bf9, // 0x00c: stp x25, x26, [sp, #-16]!
56 0xa9bf63f7, // 0x010: stp x23, x24, [sp, #-16]!
57 0xa9bf5bf5, // 0x014: stp x21, x22, [sp, #-16]!
58 0xa9bf53f3, // 0x018: stp x19, x20, [sp, #-16]!
59 0xa9bf3fee, // 0x01c: stp x14, x15, [sp, #-16]!
60 0xa9bf37ec, // 0x020: stp x12, x13, [sp, #-16]!
61 0xa9bf2fea, // 0x024: stp x10, x11, [sp, #-16]!
62 0xa9bf27e8, // 0x028: stp x8, x9, [sp, #-16]!
63 0xa9bf1fe6, // 0x02c: stp x6, x7, [sp, #-16]!
64 0xa9bf17e4, // 0x030: stp x4, x5, [sp, #-16]!
65 0xa9bf0fe2, // 0x034: stp x2, x3, [sp, #-16]!
66 0xa9bf07e0, // 0x038: stp x0, x1, [sp, #-16]!
67 0xadbf7ffe, // 0x03c: stp q30, q31, [sp, #-32]!
68 0xadbf77fc, // 0x040: stp q28, q29, [sp, #-32]!
69 0xadbf6ffa, // 0x044: stp q26, q27, [sp, #-32]!
70 0xadbf67f8, // 0x048: stp q24, q25, [sp, #-32]!
71 0xadbf5ff6, // 0x04c: stp q22, q23, [sp, #-32]!
72 0xadbf57f4, // 0x050: stp q20, q21, [sp, #-32]!
73 0xadbf4ff2, // 0x054: stp q18, q19, [sp, #-32]!
74 0xadbf47f0, // 0x058: stp q16, q17, [sp, #-32]!
75 0xadbf3fee, // 0x05c: stp q14, q15, [sp, #-32]!
76 0xadbf37ec, // 0x060: stp q12, q13, [sp, #-32]!
77 0xadbf2fea, // 0x064: stp q10, q11, [sp, #-32]!
78 0xadbf27e8, // 0x068: stp q8, q9, [sp, #-32]!
79 0xadbf1fe6, // 0x06c: stp q6, q7, [sp, #-32]!
80 0xadbf17e4, // 0x070: stp q4, q5, [sp, #-32]!
81 0xadbf0fe2, // 0x074: stp q2, q3, [sp, #-32]!
82 0xadbf07e0, // 0x078: stp q0, q1, [sp, #-32]!
83 0x580004e0, // 0x07c: ldr x0, Lreentry_ctx_ptr
84 0xaa1e03e1, // 0x080: mov x1, x30
85 0xd1003021, // 0x084: sub x1, x1, #12
86 0x58000442, // 0x088: ldr x2, Lreentry_fn_ptr
87 0xd63f0040, // 0x08c: blr x2
88 0xaa0003f1, // 0x090: mov x17, x0
89 0xacc107e0, // 0x094: ldp q0, q1, [sp], #32
90 0xacc10fe2, // 0x098: ldp q2, q3, [sp], #32
91 0xacc117e4, // 0x09c: ldp q4, q5, [sp], #32
92 0xacc11fe6, // 0x0a0: ldp q6, q7, [sp], #32
93 0xacc127e8, // 0x0a4: ldp q8, q9, [sp], #32
94 0xacc12fea, // 0x0a8: ldp q10, q11, [sp], #32
95 0xacc137ec, // 0x0ac: ldp q12, q13, [sp], #32
96 0xacc13fee, // 0x0b0: ldp q14, q15, [sp], #32
97 0xacc147f0, // 0x0b4: ldp q16, q17, [sp], #32
98 0xacc14ff2, // 0x0b8: ldp q18, q19, [sp], #32
99 0xacc157f4, // 0x0bc: ldp q20, q21, [sp], #32
100 0xacc15ff6, // 0x0c0: ldp q22, q23, [sp], #32
101 0xacc167f8, // 0x0c4: ldp q24, q25, [sp], #32
102 0xacc16ffa, // 0x0c8: ldp q26, q27, [sp], #32
103 0xacc177fc, // 0x0cc: ldp q28, q29, [sp], #32
104 0xacc17ffe, // 0x0d0: ldp q30, q31, [sp], #32
105 0xa8c107e0, // 0x0d4: ldp x0, x1, [sp], #16
106 0xa8c10fe2, // 0x0d8: ldp x2, x3, [sp], #16
107 0xa8c117e4, // 0x0dc: ldp x4, x5, [sp], #16
108 0xa8c11fe6, // 0x0e0: ldp x6, x7, [sp], #16
109 0xa8c127e8, // 0x0e4: ldp x8, x9, [sp], #16
110 0xa8c12fea, // 0x0e8: ldp x10, x11, [sp], #16
111 0xa8c137ec, // 0x0ec: ldp x12, x13, [sp], #16
112 0xa8c13fee, // 0x0f0: ldp x14, x15, [sp], #16
113 0xa8c153f3, // 0x0f4: ldp x19, x20, [sp], #16
114 0xa8c15bf5, // 0x0f8: ldp x21, x22, [sp], #16
115 0xa8c163f7, // 0x0fc: ldp x23, x24, [sp], #16
116 0xa8c16bf9, // 0x100: ldp x25, x26, [sp], #16
117 0xa8c173fb, // 0x104: ldp x27, x28, [sp], #16
118 0xa8c17bfd, // 0x108: ldp x29, x30, [sp], #16
119 0xd65f0220, // 0x10c: ret x17
120 0x01234567, // 0x110: Lreentry_fn_ptr:
121 0xdeadbeef, // 0x114: .quad 0
122 0x98765432, // 0x118: Lreentry_ctx_ptr:
123 0xcafef00d // 0x11c: .quad 0
124 };
125
126 const unsigned ReentryFnAddrOffset = 0x110;
127 const unsigned ReentryCtxAddrOffset = 0x118;
128
129 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
130 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
131 sizeof(uint64_t));
132 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
133 sizeof(uint64_t));
134 }
135
136 void OrcAArch64::writeTrampolines(char *TrampolineBlockWorkingMem,
137 ExecutorAddr TrampolineBlockTargetAddress,
138 ExecutorAddr ResolverAddr,
139 unsigned NumTrampolines) {
140
141 unsigned OffsetToPtr = alignTo(NumTrampolines * TrampolineSize, 8);
142
143 memcpy(TrampolineBlockWorkingMem + OffsetToPtr, &ResolverAddr,
144 sizeof(uint64_t));
145
146 // OffsetToPtr is actually the offset from the PC for the 2nd instruction, so
147 // subtract 32-bits.
148 OffsetToPtr -= 4;
149
150 uint32_t *Trampolines =
151 reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
152
153 for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize) {
154 Trampolines[3 * I + 0] = 0xaa1e03f1; // mov x17, x30
155 Trampolines[3 * I + 1] = 0x58000010 | (OffsetToPtr << 3); // ldr x16, Lptr
156 Trampolines[3 * I + 2] = 0xd63f0200; // blr x16
157 }
158 }
159
160 void OrcAArch64::writeIndirectStubsBlock(
161 char *StubsBlockWorkingMem, ExecutorAddr StubsBlockTargetAddress,
162 ExecutorAddr PointersBlockTargetAddress, unsigned NumStubs) {
163 // Stub format is:
164 //
165 // .section __orc_stubs
166 // stub1:
167 // ldr x16, ptr1 ; PC-rel load of ptr1
168 // br x16 ; Jump to resolver
169 // stub2:
170 // ldr x16, ptr2 ; PC-rel load of ptr2
171 // br x16 ; Jump to resolver
172 //
173 // ...
174 //
175 // .section __orc_ptrs
176 // ptr1:
177 // .quad 0x0
178 // ptr2:
179 // .quad 0x0
180 //
181 // ...
182
183 static_assert(StubSize == PointerSize,
184 "Pointer and stub size must match for algorithm below");
185 assert(stubAndPointerRangesOk<OrcAArch64>(
186 StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
187 "PointersBlock is out of range");
188 uint64_t PtrDisplacement =
189 PointersBlockTargetAddress - StubsBlockTargetAddress;
190 assert((PtrDisplacement % 8 == 0) &&
191 "Displacement to pointer is not a multiple of 8");
192 uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlockWorkingMem);
193 uint64_t PtrOffsetField = ((PtrDisplacement >> 2) & 0x7ffff) << 5;
194
195 for (unsigned I = 0; I < NumStubs; ++I)
196 Stub[I] = 0xd61f020058000010 | PtrOffsetField;
197 }
198
199 void OrcX86_64_Base::writeTrampolines(char *TrampolineBlockWorkingMem,
200 ExecutorAddr TrampolineBlockTargetAddress,
201 ExecutorAddr ResolverAddr,
202 unsigned NumTrampolines) {
203
204 unsigned OffsetToPtr = NumTrampolines * TrampolineSize;
205
206 memcpy(TrampolineBlockWorkingMem + OffsetToPtr, &ResolverAddr,
207 sizeof(uint64_t));
208
209 uint64_t *Trampolines =
210 reinterpret_cast<uint64_t *>(TrampolineBlockWorkingMem);
211 uint64_t CallIndirPCRel = 0xf1c40000000015ff;
212
213 for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize)
214 Trampolines[I] = CallIndirPCRel | ((OffsetToPtr - 6) << 16);
215 }
216
217 void OrcX86_64_Base::writeIndirectStubsBlock(
218 char *StubsBlockWorkingMem, ExecutorAddr StubsBlockTargetAddress,
219 ExecutorAddr PointersBlockTargetAddress, unsigned NumStubs) {
220 // Stub format is:
221 //
222 // .section __orc_stubs
223 // stub1:
224 // jmpq *ptr1(%rip)
225 // .byte 0xC4 ; <- Invalid opcode padding.
226 // .byte 0xF1
227 // stub2:
228 // jmpq *ptr2(%rip)
229 //
230 // ...
231 //
232 // .section __orc_ptrs
233 // ptr1:
234 // .quad 0x0
235 // ptr2:
236 // .quad 0x0
237 //
238 // ...
239
240 // Populate the stubs page stubs and mark it executable.
241 static_assert(StubSize == PointerSize,
242 "Pointer and stub size must match for algorithm below");
243 assert(stubAndPointerRangesOk<OrcX86_64_Base>(
244 StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
245 "PointersBlock is out of range");
246 uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlockWorkingMem);
247 uint64_t PtrOffsetField =
248 (PointersBlockTargetAddress - StubsBlockTargetAddress - 6) << 16;
249 for (unsigned I = 0; I < NumStubs; ++I)
250 Stub[I] = 0xF1C40000000025ff | PtrOffsetField;
251 }
252
253 void OrcX86_64_SysV::writeResolverCode(char *ResolverWorkingMem,
254 ExecutorAddr ResolverTargetAddress,
255 ExecutorAddr ReentryFnAddr,
256 ExecutorAddr ReentryCtxAddr) {
257
258 LLVM_DEBUG({
259 dbgs() << "Writing resolver code to "
260 << formatv("{0:x16}", ResolverTargetAddress) << "\n";
261 });
262
263 const uint8_t ResolverCode[] = {
264 // resolver_entry:
265 0x55, // 0x00: pushq %rbp
266 0x48, 0x89, 0xe5, // 0x01: movq %rsp, %rbp
267 0x50, // 0x04: pushq %rax
268 0x53, // 0x05: pushq %rbx
269 0x51, // 0x06: pushq %rcx
270 0x52, // 0x07: pushq %rdx
271 0x56, // 0x08: pushq %rsi
272 0x57, // 0x09: pushq %rdi
273 0x41, 0x50, // 0x0a: pushq %r8
274 0x41, 0x51, // 0x0c: pushq %r9
275 0x41, 0x52, // 0x0e: pushq %r10
276 0x41, 0x53, // 0x10: pushq %r11
277 0x41, 0x54, // 0x12: pushq %r12
278 0x41, 0x55, // 0x14: pushq %r13
279 0x41, 0x56, // 0x16: pushq %r14
280 0x41, 0x57, // 0x18: pushq %r15
281 0x48, 0x81, 0xec, 0x08, 0x02, 0x00, 0x00, // 0x1a: subq 0x208, %rsp
282 0x48, 0x0f, 0xae, 0x04, 0x24, // 0x21: fxsave64 (%rsp)
283 0x48, 0xbf, // 0x26: movabsq <CBMgr>, %rdi
284
285 // 0x28: JIT re-entry ctx addr.
286 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
287
288 0x48, 0x8b, 0x75, 0x08, // 0x30: movq 8(%rbp), %rsi
289 0x48, 0x83, 0xee, 0x06, // 0x34: subq $6, %rsi
290 0x48, 0xb8, // 0x38: movabsq <REntry>, %rax
291
292 // 0x3a: JIT re-entry fn addr:
293 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
294
295 0xff, 0xd0, // 0x42: callq *%rax
296 0x48, 0x89, 0x45, 0x08, // 0x44: movq %rax, 8(%rbp)
297 0x48, 0x0f, 0xae, 0x0c, 0x24, // 0x48: fxrstor64 (%rsp)
298 0x48, 0x81, 0xc4, 0x08, 0x02, 0x00, 0x00, // 0x4d: addq 0x208, %rsp
299 0x41, 0x5f, // 0x54: popq %r15
300 0x41, 0x5e, // 0x56: popq %r14
301 0x41, 0x5d, // 0x58: popq %r13
302 0x41, 0x5c, // 0x5a: popq %r12
303 0x41, 0x5b, // 0x5c: popq %r11
304 0x41, 0x5a, // 0x5e: popq %r10
305 0x41, 0x59, // 0x60: popq %r9
306 0x41, 0x58, // 0x62: popq %r8
307 0x5f, // 0x64: popq %rdi
308 0x5e, // 0x65: popq %rsi
309 0x5a, // 0x66: popq %rdx
310 0x59, // 0x67: popq %rcx
311 0x5b, // 0x68: popq %rbx
312 0x58, // 0x69: popq %rax
313 0x5d, // 0x6a: popq %rbp
314 0xc3, // 0x6b: retq
315 };
316
317 const unsigned ReentryFnAddrOffset = 0x3a;
318 const unsigned ReentryCtxAddrOffset = 0x28;
319
320 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
321 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
322 sizeof(uint64_t));
323 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
324 sizeof(uint64_t));
325 }
326
327 void OrcX86_64_Win32::writeResolverCode(char *ResolverWorkingMem,
328 ExecutorAddr ResolverTargetAddress,
329 ExecutorAddr ReentryFnAddr,
330 ExecutorAddr ReentryCtxAddr) {
331
332 // resolverCode is similar to OrcX86_64 with differences specific to windows
333 // x64 calling convention: arguments go into rcx, rdx and come in reverse
334 // order, shadow space allocation on stack
335 const uint8_t ResolverCode[] = {
336 // resolver_entry:
337 0x55, // 0x00: pushq %rbp
338 0x48, 0x89, 0xe5, // 0x01: movq %rsp, %rbp
339 0x50, // 0x04: pushq %rax
340 0x53, // 0x05: pushq %rbx
341 0x51, // 0x06: pushq %rcx
342 0x52, // 0x07: pushq %rdx
343 0x56, // 0x08: pushq %rsi
344 0x57, // 0x09: pushq %rdi
345 0x41, 0x50, // 0x0a: pushq %r8
346 0x41, 0x51, // 0x0c: pushq %r9
347 0x41, 0x52, // 0x0e: pushq %r10
348 0x41, 0x53, // 0x10: pushq %r11
349 0x41, 0x54, // 0x12: pushq %r12
350 0x41, 0x55, // 0x14: pushq %r13
351 0x41, 0x56, // 0x16: pushq %r14
352 0x41, 0x57, // 0x18: pushq %r15
353 0x48, 0x81, 0xec, 0x08, 0x02, 0x00, 0x00, // 0x1a: subq 0x208, %rsp
354 0x48, 0x0f, 0xae, 0x04, 0x24, // 0x21: fxsave64 (%rsp)
355
356 0x48, 0xb9, // 0x26: movabsq <CBMgr>, %rcx
357 // 0x28: JIT re-entry ctx addr.
358 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
359
360 0x48, 0x8B, 0x55, 0x08, // 0x30: mov rdx, [rbp+0x8]
361 0x48, 0x83, 0xea, 0x06, // 0x34: sub rdx, 0x6
362
363 0x48, 0xb8, // 0x38: movabsq <REntry>, %rax
364 // 0x3a: JIT re-entry fn addr:
365 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
366
367 // 0x42: sub rsp, 0x20 (Allocate shadow space)
368 0x48, 0x83, 0xEC, 0x20,
369 0xff, 0xd0, // 0x46: callq *%rax
370
371 // 0x48: add rsp, 0x20 (Free shadow space)
372 0x48, 0x83, 0xC4, 0x20,
373
374 0x48, 0x89, 0x45, 0x08, // 0x4C: movq %rax, 8(%rbp)
375 0x48, 0x0f, 0xae, 0x0c, 0x24, // 0x50: fxrstor64 (%rsp)
376 0x48, 0x81, 0xc4, 0x08, 0x02, 0x00, 0x00, // 0x55: addq 0x208, %rsp
377 0x41, 0x5f, // 0x5C: popq %r15
378 0x41, 0x5e, // 0x5E: popq %r14
379 0x41, 0x5d, // 0x60: popq %r13
380 0x41, 0x5c, // 0x62: popq %r12
381 0x41, 0x5b, // 0x64: popq %r11
382 0x41, 0x5a, // 0x66: popq %r10
383 0x41, 0x59, // 0x68: popq %r9
384 0x41, 0x58, // 0x6a: popq %r8
385 0x5f, // 0x6c: popq %rdi
386 0x5e, // 0x6d: popq %rsi
387 0x5a, // 0x6e: popq %rdx
388 0x59, // 0x6f: popq %rcx
389 0x5b, // 0x70: popq %rbx
390 0x58, // 0x71: popq %rax
391 0x5d, // 0x72: popq %rbp
392 0xc3, // 0x73: retq
393 };
394
395 const unsigned ReentryFnAddrOffset = 0x3a;
396 const unsigned ReentryCtxAddrOffset = 0x28;
397
398 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
399 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
400 sizeof(uint64_t));
401 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
402 sizeof(uint64_t));
403 }
404
405 void OrcI386::writeResolverCode(char *ResolverWorkingMem,
406 ExecutorAddr ResolverTargetAddress,
407 ExecutorAddr ReentryFnAddr,
408 ExecutorAddr ReentryCtxAddr) {
409
410 assert((ReentryFnAddr.getValue() >> 32) == 0 && "ReentryFnAddr out of range");
411 assert((ReentryCtxAddr.getValue() >> 32) == 0 &&
412 "ReentryCtxAddr out of range");
413
414 const uint8_t ResolverCode[] = {
415 // resolver_entry:
416 0x55, // 0x00: pushl %ebp
417 0x89, 0xe5, // 0x01: movl %esp, %ebp
418 0x54, // 0x03: pushl %esp
419 0x83, 0xe4, 0xf0, // 0x04: andl $-0x10, %esp
420 0x50, // 0x07: pushl %eax
421 0x53, // 0x08: pushl %ebx
422 0x51, // 0x09: pushl %ecx
423 0x52, // 0x0a: pushl %edx
424 0x56, // 0x0b: pushl %esi
425 0x57, // 0x0c: pushl %edi
426 0x81, 0xec, 0x18, 0x02, 0x00, 0x00, // 0x0d: subl $0x218, %esp
427 0x0f, 0xae, 0x44, 0x24, 0x10, // 0x13: fxsave 0x10(%esp)
428 0x8b, 0x75, 0x04, // 0x18: movl 0x4(%ebp), %esi
429 0x83, 0xee, 0x05, // 0x1b: subl $0x5, %esi
430 0x89, 0x74, 0x24, 0x04, // 0x1e: movl %esi, 0x4(%esp)
431 0xc7, 0x04, 0x24, 0x00, 0x00, 0x00,
432 0x00, // 0x22: movl <cbmgr>, (%esp)
433 0xb8, 0x00, 0x00, 0x00, 0x00, // 0x29: movl <reentry>, %eax
434 0xff, 0xd0, // 0x2e: calll *%eax
435 0x89, 0x45, 0x04, // 0x30: movl %eax, 0x4(%ebp)
436 0x0f, 0xae, 0x4c, 0x24, 0x10, // 0x33: fxrstor 0x10(%esp)
437 0x81, 0xc4, 0x18, 0x02, 0x00, 0x00, // 0x38: addl $0x218, %esp
438 0x5f, // 0x3e: popl %edi
439 0x5e, // 0x3f: popl %esi
440 0x5a, // 0x40: popl %edx
441 0x59, // 0x41: popl %ecx
442 0x5b, // 0x42: popl %ebx
443 0x58, // 0x43: popl %eax
444 0x8b, 0x65, 0xfc, // 0x44: movl -0x4(%ebp), %esp
445 0x5d, // 0x48: popl %ebp
446 0xc3 // 0x49: retl
447 };
448
449 const unsigned ReentryFnAddrOffset = 0x2a;
450 const unsigned ReentryCtxAddrOffset = 0x25;
451
452 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
453 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
454 sizeof(uint32_t));
455 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
456 sizeof(uint32_t));
457 }
458
459 void OrcI386::writeTrampolines(char *TrampolineWorkingMem,
460 ExecutorAddr TrampolineBlockTargetAddress,
461 ExecutorAddr ResolverAddr,
462 unsigned NumTrampolines) {
463 assert((ResolverAddr.getValue() >> 32) == 0 && "ResolverAddr out of range");
464
465 uint64_t CallRelImm = 0xF1C4C400000000e8;
466 uint64_t ResolverRel = ResolverAddr - TrampolineBlockTargetAddress - 5;
467
468 uint64_t *Trampolines = reinterpret_cast<uint64_t *>(TrampolineWorkingMem);
469 for (unsigned I = 0; I < NumTrampolines; ++I, ResolverRel -= TrampolineSize)
470 Trampolines[I] = CallRelImm | (ResolverRel << 8);
471 }
472
473 void OrcI386::writeIndirectStubsBlock(char *StubsBlockWorkingMem,
474 ExecutorAddr StubsBlockTargetAddress,
475 ExecutorAddr PointersBlockTargetAddress,
476 unsigned NumStubs) {
477 assert((StubsBlockTargetAddress.getValue() >> 32) == 0 &&
478 "StubsBlockTargetAddress is out of range");
479 assert((PointersBlockTargetAddress.getValue() >> 32) == 0 &&
480 "PointersBlockTargetAddress is out of range");
481
482 // Stub format is:
483 //
484 // .section __orc_stubs
485 // stub1:
486 // jmpq *ptr1
487 // .byte 0xC4 ; <- Invalid opcode padding.
488 // .byte 0xF1
489 // stub2:
490 // jmpq *ptr2
491 //
492 // ...
493 //
494 // .section __orc_ptrs
495 // ptr1:
496 // .quad 0x0
497 // ptr2:
498 // .quad 0x0
499 //
500 // ...
501
502 assert(stubAndPointerRangesOk<OrcI386>(
503 StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
504 "PointersBlock is out of range");
505
506 uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlockWorkingMem);
507 uint64_t PtrAddr = PointersBlockTargetAddress.getValue();
508 for (unsigned I = 0; I < NumStubs; ++I, PtrAddr += 4)
509 Stub[I] = 0xF1C40000000025ff | (PtrAddr << 16);
510 }
511
512 void OrcMips32_Base::writeResolverCode(char *ResolverWorkingMem,
513 ExecutorAddr ResolverTargetAddress,
514 ExecutorAddr ReentryFnAddr,
515 ExecutorAddr ReentryCtxAddr,
516 bool isBigEndian) {
517
518 const uint32_t ResolverCode[] = {
519 // resolver_entry:
520 0x27bdff98, // 0x00: addiu $sp,$sp,-104
521 0xafa20000, // 0x04: sw $v0,0($sp)
522 0xafa30004, // 0x08: sw $v1,4($sp)
523 0xafa40008, // 0x0c: sw $a0,8($sp)
524 0xafa5000c, // 0x10: sw $a1,12($sp)
525 0xafa60010, // 0x14: sw $a2,16($sp)
526 0xafa70014, // 0x18: sw $a3,20($sp)
527 0xafb00018, // 0x1c: sw $s0,24($sp)
528 0xafb1001c, // 0x20: sw $s1,28($sp)
529 0xafb20020, // 0x24: sw $s2,32($sp)
530 0xafb30024, // 0x28: sw $s3,36($sp)
531 0xafb40028, // 0x2c: sw $s4,40($sp)
532 0xafb5002c, // 0x30: sw $s5,44($sp)
533 0xafb60030, // 0x34: sw $s6,48($sp)
534 0xafb70034, // 0x38: sw $s7,52($sp)
535 0xafa80038, // 0x3c: sw $t0,56($sp)
536 0xafa9003c, // 0x40: sw $t1,60($sp)
537 0xafaa0040, // 0x44: sw $t2,64($sp)
538 0xafab0044, // 0x48: sw $t3,68($sp)
539 0xafac0048, // 0x4c: sw $t4,72($sp)
540 0xafad004c, // 0x50: sw $t5,76($sp)
541 0xafae0050, // 0x54: sw $t6,80($sp)
542 0xafaf0054, // 0x58: sw $t7,84($sp)
543 0xafb80058, // 0x5c: sw $t8,88($sp)
544 0xafb9005c, // 0x60: sw $t9,92($sp)
545 0xafbe0060, // 0x64: sw $fp,96($sp)
546 0xafbf0064, // 0x68: sw $ra,100($sp)
547
548 // JIT re-entry ctx addr.
549 0x00000000, // 0x6c: lui $a0,ctx
550 0x00000000, // 0x70: addiu $a0,$a0,ctx
551
552 0x03e02825, // 0x74: move $a1, $ra
553 0x24a5ffec, // 0x78: addiu $a1,$a1,-20
554
555 // JIT re-entry fn addr:
556 0x00000000, // 0x7c: lui $t9,reentry
557 0x00000000, // 0x80: addiu $t9,$t9,reentry
558
559 0x0320f809, // 0x84: jalr $t9
560 0x00000000, // 0x88: nop
561 0x8fbf0064, // 0x8c: lw $ra,100($sp)
562 0x8fbe0060, // 0x90: lw $fp,96($sp)
563 0x8fb9005c, // 0x94: lw $t9,92($sp)
564 0x8fb80058, // 0x98: lw $t8,88($sp)
565 0x8faf0054, // 0x9c: lw $t7,84($sp)
566 0x8fae0050, // 0xa0: lw $t6,80($sp)
567 0x8fad004c, // 0xa4: lw $t5,76($sp)
568 0x8fac0048, // 0xa8: lw $t4,72($sp)
569 0x8fab0044, // 0xac: lw $t3,68($sp)
570 0x8faa0040, // 0xb0: lw $t2,64($sp)
571 0x8fa9003c, // 0xb4: lw $t1,60($sp)
572 0x8fa80038, // 0xb8: lw $t0,56($sp)
573 0x8fb70034, // 0xbc: lw $s7,52($sp)
574 0x8fb60030, // 0xc0: lw $s6,48($sp)
575 0x8fb5002c, // 0xc4: lw $s5,44($sp)
576 0x8fb40028, // 0xc8: lw $s4,40($sp)
577 0x8fb30024, // 0xcc: lw $s3,36($sp)
578 0x8fb20020, // 0xd0: lw $s2,32($sp)
579 0x8fb1001c, // 0xd4: lw $s1,28($sp)
580 0x8fb00018, // 0xd8: lw $s0,24($sp)
581 0x8fa70014, // 0xdc: lw $a3,20($sp)
582 0x8fa60010, // 0xe0: lw $a2,16($sp)
583 0x8fa5000c, // 0xe4: lw $a1,12($sp)
584 0x8fa40008, // 0xe8: lw $a0,8($sp)
585 0x27bd0068, // 0xec: addiu $sp,$sp,104
586 0x0300f825, // 0xf0: move $ra, $t8
587 0x03200008, // 0xf4: jr $t9
588 0x00000000, // 0xf8: move $t9, $v0/v1
589 };
590
591 const unsigned ReentryFnAddrOffset = 0x7c; // JIT re-entry fn addr lui
592 const unsigned ReentryCtxAddrOffset = 0x6c; // JIT re-entry context addr lui
593 const unsigned Offsett = 0xf8;
594
595 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
596
597 // Depending on endian return value will be in v0 or v1.
598 uint32_t MoveVxT9 = isBigEndian ? 0x0060c825 : 0x0040c825;
599 memcpy(ResolverWorkingMem + Offsett, &MoveVxT9, sizeof(MoveVxT9));
600
601 uint32_t ReentryCtxLUi =
602 0x3c040000 | (((ReentryCtxAddr.getValue() + 0x8000) >> 16) & 0xFFFF);
603 uint32_t ReentryCtxADDiu = 0x24840000 | (ReentryCtxAddr.getValue() & 0xFFFF);
604 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxLUi,
605 sizeof(ReentryCtxLUi));
606 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset + 4, &ReentryCtxADDiu,
607 sizeof(ReentryCtxADDiu));
608
609 uint32_t ReentryFnLUi =
610 0x3c190000 | (((ReentryFnAddr.getValue() + 0x8000) >> 16) & 0xFFFF);
611 uint32_t ReentryFnADDiu = 0x27390000 | (ReentryFnAddr.getValue() & 0xFFFF);
612 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnLUi,
613 sizeof(ReentryFnLUi));
614 memcpy(ResolverWorkingMem + ReentryFnAddrOffset + 4, &ReentryFnADDiu,
615 sizeof(ReentryFnADDiu));
616 }
617
618 void OrcMips32_Base::writeTrampolines(char *TrampolineBlockWorkingMem,
619 ExecutorAddr TrampolineBlockTargetAddress,
620 ExecutorAddr ResolverAddr,
621 unsigned NumTrampolines) {
622
623 assert((ResolverAddr.getValue() >> 32) == 0 && "ResolverAddr out of range");
624
625 uint32_t *Trampolines =
626 reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
627 uint32_t RHiAddr = ((ResolverAddr.getValue() + 0x8000) >> 16);
628
629 for (unsigned I = 0; I < NumTrampolines; ++I) {
630 // move $t8,$ra
631 // lui $t9,ResolverAddr
632 // addiu $t9,$t9,ResolverAddr
633 // jalr $t9
634 // nop
635 Trampolines[5 * I + 0] = 0x03e0c025;
636 Trampolines[5 * I + 1] = 0x3c190000 | (RHiAddr & 0xFFFF);
637 Trampolines[5 * I + 2] = 0x27390000 | (ResolverAddr.getValue() & 0xFFFF);
638 Trampolines[5 * I + 3] = 0x0320f809;
639 Trampolines[5 * I + 4] = 0x00000000;
640 }
641 }
642
643 void OrcMips32_Base::writeIndirectStubsBlock(
644 char *StubsBlockWorkingMem, ExecutorAddr StubsBlockTargetAddress,
645 ExecutorAddr PointersBlockTargetAddress, unsigned NumStubs) {
646 assert((StubsBlockTargetAddress.getValue() >> 32) == 0 &&
647 "InitialPtrVal is out of range");
648
649 // Stub format is:
650 //
651 // .section __orc_stubs
652 // stub1:
653 // lui $t9, ptr1
654 // lw $t9, %lo(ptr1)($t9)
655 // jr $t9
656 // stub2:
657 // lui $t9, ptr2
658 // lw $t9,%lo(ptr1)($t9)
659 // jr $t9
660 //
661 // ...
662 //
663 // .section __orc_ptrs
664 // ptr1:
665 // .word 0x0
666 // ptr2:
667 // .word 0x0
668 //
669 // i..
670
671 assert(stubAndPointerRangesOk<OrcMips32_Base>(
672 StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
673 "PointersBlock is out of range");
674
675 // Populate the stubs page stubs and mark it executable.
676 uint32_t *Stub = reinterpret_cast<uint32_t *>(StubsBlockWorkingMem);
677 uint64_t PtrAddr = PointersBlockTargetAddress.getValue();
678
679 for (unsigned I = 0; I < NumStubs; ++I) {
680 uint32_t HiAddr = ((PtrAddr + 0x8000) >> 16);
681 Stub[4 * I + 0] = 0x3c190000 | (HiAddr & 0xFFFF); // lui $t9,ptr1
682 Stub[4 * I + 1] = 0x8f390000 | (PtrAddr & 0xFFFF); // lw $t9,%lo(ptr1)($t9)
683 Stub[4 * I + 2] = 0x03200008; // jr $t9
684 Stub[4 * I + 3] = 0x00000000; // nop
685 PtrAddr += 4;
686 }
687 }
688
689 void OrcMips64::writeResolverCode(char *ResolverWorkingMem,
690 ExecutorAddr ResolverTargetAddress,
691 ExecutorAddr ReentryFnAddr,
692 ExecutorAddr ReentryCtxAddr) {
693
694 const uint32_t ResolverCode[] = {
695 //resolver_entry:
696 0x67bdff30, // 0x00: daddiu $sp,$sp,-208
697 0xffa20000, // 0x04: sd v0,0(sp)
698 0xffa30008, // 0x08: sd v1,8(sp)
699 0xffa40010, // 0x0c: sd a0,16(sp)
700 0xffa50018, // 0x10: sd a1,24(sp)
701 0xffa60020, // 0x14: sd a2,32(sp)
702 0xffa70028, // 0x18: sd a3,40(sp)
703 0xffa80030, // 0x1c: sd a4,48(sp)
704 0xffa90038, // 0x20: sd a5,56(sp)
705 0xffaa0040, // 0x24: sd a6,64(sp)
706 0xffab0048, // 0x28: sd a7,72(sp)
707 0xffac0050, // 0x2c: sd t0,80(sp)
708 0xffad0058, // 0x30: sd t1,88(sp)
709 0xffae0060, // 0x34: sd t2,96(sp)
710 0xffaf0068, // 0x38: sd t3,104(sp)
711 0xffb00070, // 0x3c: sd s0,112(sp)
712 0xffb10078, // 0x40: sd s1,120(sp)
713 0xffb20080, // 0x44: sd s2,128(sp)
714 0xffb30088, // 0x48: sd s3,136(sp)
715 0xffb40090, // 0x4c: sd s4,144(sp)
716 0xffb50098, // 0x50: sd s5,152(sp)
717 0xffb600a0, // 0x54: sd s6,160(sp)
718 0xffb700a8, // 0x58: sd s7,168(sp)
719 0xffb800b0, // 0x5c: sd t8,176(sp)
720 0xffb900b8, // 0x60: sd t9,184(sp)
721 0xffbe00c0, // 0x64: sd fp,192(sp)
722 0xffbf00c8, // 0x68: sd ra,200(sp)
723
724 // JIT re-entry ctx addr.
725 0x00000000, // 0x6c: lui $a0,heighest(ctx)
726 0x00000000, // 0x70: daddiu $a0,$a0,heigher(ctx)
727 0x00000000, // 0x74: dsll $a0,$a0,16
728 0x00000000, // 0x78: daddiu $a0,$a0,hi(ctx)
729 0x00000000, // 0x7c: dsll $a0,$a0,16
730 0x00000000, // 0x80: daddiu $a0,$a0,lo(ctx)
731
732 0x03e02825, // 0x84: move $a1, $ra
733 0x64a5ffdc, // 0x88: daddiu $a1,$a1,-36
734
735 // JIT re-entry fn addr:
736 0x00000000, // 0x8c: lui $t9,reentry
737 0x00000000, // 0x90: daddiu $t9,$t9,reentry
738 0x00000000, // 0x94: dsll $t9,$t9,
739 0x00000000, // 0x98: daddiu $t9,$t9,
740 0x00000000, // 0x9c: dsll $t9,$t9,
741 0x00000000, // 0xa0: daddiu $t9,$t9,
742 0x0320f809, // 0xa4: jalr $t9
743 0x00000000, // 0xa8: nop
744 0xdfbf00c8, // 0xac: ld ra, 200(sp)
745 0xdfbe00c0, // 0xb0: ld fp, 192(sp)
746 0xdfb900b8, // 0xb4: ld t9, 184(sp)
747 0xdfb800b0, // 0xb8: ld t8, 176(sp)
748 0xdfb700a8, // 0xbc: ld s7, 168(sp)
749 0xdfb600a0, // 0xc0: ld s6, 160(sp)
750 0xdfb50098, // 0xc4: ld s5, 152(sp)
751 0xdfb40090, // 0xc8: ld s4, 144(sp)
752 0xdfb30088, // 0xcc: ld s3, 136(sp)
753 0xdfb20080, // 0xd0: ld s2, 128(sp)
754 0xdfb10078, // 0xd4: ld s1, 120(sp)
755 0xdfb00070, // 0xd8: ld s0, 112(sp)
756 0xdfaf0068, // 0xdc: ld t3, 104(sp)
757 0xdfae0060, // 0xe0: ld t2, 96(sp)
758 0xdfad0058, // 0xe4: ld t1, 88(sp)
759 0xdfac0050, // 0xe8: ld t0, 80(sp)
760 0xdfab0048, // 0xec: ld a7, 72(sp)
761 0xdfaa0040, // 0xf0: ld a6, 64(sp)
762 0xdfa90038, // 0xf4: ld a5, 56(sp)
763 0xdfa80030, // 0xf8: ld a4, 48(sp)
764 0xdfa70028, // 0xfc: ld a3, 40(sp)
765 0xdfa60020, // 0x100: ld a2, 32(sp)
766 0xdfa50018, // 0x104: ld a1, 24(sp)
767 0xdfa40010, // 0x108: ld a0, 16(sp)
768 0xdfa30008, // 0x10c: ld v1, 8(sp)
769 0x67bd00d0, // 0x110: daddiu $sp,$sp,208
770 0x0300f825, // 0x114: move $ra, $t8
771 0x03200008, // 0x118: jr $t9
772 0x0040c825, // 0x11c: move $t9, $v0
773 };
774
775 const unsigned ReentryFnAddrOffset = 0x8c; // JIT re-entry fn addr lui
776 const unsigned ReentryCtxAddrOffset = 0x6c; // JIT re-entry ctx addr lui
777
778 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
779
780 uint32_t ReentryCtxLUi =
781 0x3c040000 |
782 (((ReentryCtxAddr.getValue() + 0x800080008000) >> 48) & 0xFFFF);
783 uint32_t ReentryCtxDADDiu =
784 0x64840000 | (((ReentryCtxAddr.getValue() + 0x80008000) >> 32) & 0xFFFF);
785 uint32_t ReentryCtxDSLL = 0x00042438;
786 uint32_t ReentryCtxDADDiu2 =
787 0x64840000 | ((((ReentryCtxAddr.getValue() + 0x8000) >> 16) & 0xFFFF));
788 uint32_t ReentryCtxDSLL2 = 0x00042438;
789 uint32_t ReentryCtxDADDiu3 =
790 0x64840000 | (ReentryCtxAddr.getValue() & 0xFFFF);
791
792 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxLUi,
793 sizeof(ReentryCtxLUi));
794 memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 4), &ReentryCtxDADDiu,
795 sizeof(ReentryCtxDADDiu));
796 memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 8), &ReentryCtxDSLL,
797 sizeof(ReentryCtxDSLL));
798 memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 12), &ReentryCtxDADDiu2,
799 sizeof(ReentryCtxDADDiu2));
800 memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 16), &ReentryCtxDSLL2,
801 sizeof(ReentryCtxDSLL2));
802 memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 20), &ReentryCtxDADDiu3,
803 sizeof(ReentryCtxDADDiu3));
804
805 uint32_t ReentryFnLUi =
806 0x3c190000 |
807 (((ReentryFnAddr.getValue() + 0x800080008000) >> 48) & 0xFFFF);
808
809 uint32_t ReentryFnDADDiu =
810 0x67390000 | (((ReentryFnAddr.getValue() + 0x80008000) >> 32) & 0xFFFF);
811
812 uint32_t ReentryFnDSLL = 0x0019cc38;
813
814 uint32_t ReentryFnDADDiu2 =
815 0x67390000 | (((ReentryFnAddr.getValue() + 0x8000) >> 16) & 0xFFFF);
816
817 uint32_t ReentryFnDSLL2 = 0x0019cc38;
818
819 uint32_t ReentryFnDADDiu3 = 0x67390000 | (ReentryFnAddr.getValue() & 0xFFFF);
820
821 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnLUi,
822 sizeof(ReentryFnLUi));
823 memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 4), &ReentryFnDADDiu,
824 sizeof(ReentryFnDADDiu));
825 memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 8), &ReentryFnDSLL,
826 sizeof(ReentryFnDSLL));
827 memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 12), &ReentryFnDADDiu2,
828 sizeof(ReentryFnDADDiu2));
829 memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 16), &ReentryFnDSLL2,
830 sizeof(ReentryFnDSLL2));
831 memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 20), &ReentryFnDADDiu3,
832 sizeof(ReentryFnDADDiu3));
833 }
834
835 void OrcMips64::writeTrampolines(char *TrampolineBlockWorkingMem,
836 ExecutorAddr TrampolineBlockTargetAddress,
837 ExecutorAddr ResolverAddr,
838 unsigned NumTrampolines) {
839
840 uint32_t *Trampolines =
841 reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
842
843 uint64_t HeighestAddr = ((ResolverAddr.getValue() + 0x800080008000) >> 48);
844 uint64_t HeigherAddr = ((ResolverAddr.getValue() + 0x80008000) >> 32);
845 uint64_t HiAddr = ((ResolverAddr.getValue() + 0x8000) >> 16);
846
847 for (unsigned I = 0; I < NumTrampolines; ++I) {
848 Trampolines[10 * I + 0] = 0x03e0c025; // move $t8,$ra
849 Trampolines[10 * I + 1] = 0x3c190000 | (HeighestAddr & 0xFFFF); // lui $t9,resolveAddr
850 Trampolines[10 * I + 2] = 0x67390000 | (HeigherAddr & 0xFFFF); // daddiu $t9,$t9,%higher(resolveAddr)
851 Trampolines[10 * I + 3] = 0x0019cc38; // dsll $t9,$t9,16
852 Trampolines[10 * I + 4] = 0x67390000 | (HiAddr & 0xFFFF); // daddiu $t9,$t9,%hi(ptr)
853 Trampolines[10 * I + 5] = 0x0019cc38; // dsll $t9,$t9,16
854 Trampolines[10 * I + 6] = 0x67390000 | (ResolverAddr.getValue() &
855 0xFFFF); // daddiu $t9,$t9,%lo(ptr)
856 Trampolines[10 * I + 7] = 0x0320f809; // jalr $t9
857 Trampolines[10 * I + 8] = 0x00000000; // nop
858 Trampolines[10 * I + 9] = 0x00000000; // nop
859 }
860 }
861
862 void OrcMips64::writeIndirectStubsBlock(char *StubsBlockWorkingMem,
863 ExecutorAddr StubsBlockTargetAddress,
864 ExecutorAddr PointersBlockTargetAddress,
865 unsigned NumStubs) {
866 // Stub format is:
867 //
868 // .section __orc_stubs
869 // stub1:
870 // lui $t9,ptr1
871 // dsll $t9,$t9,16
872 // daddiu $t9,$t9,%hi(ptr)
873 // dsll $t9,$t9,16
874 // ld $t9,%lo(ptr)
875 // jr $t9
876 // stub2:
877 // lui $t9,ptr1
878 // dsll $t9,$t9,16
879 // daddiu $t9,$t9,%hi(ptr)
880 // dsll $t9,$t9,16
881 // ld $t9,%lo(ptr)
882 // jr $t9
883 //
884 // ...
885 //
886 // .section __orc_ptrs
887 // ptr1:
888 // .dword 0x0
889 // ptr2:
890 // .dword 0x0
891 //
892 // ...
893
894 assert(stubAndPointerRangesOk<OrcMips64>(
895 StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
896 "PointersBlock is out of range");
897
898 // Populate the stubs page stubs and mark it executable.
899 uint32_t *Stub = reinterpret_cast<uint32_t *>(StubsBlockWorkingMem);
900 uint64_t PtrAddr = PointersBlockTargetAddress.getValue();
901
902 for (unsigned I = 0; I < NumStubs; ++I, PtrAddr += 8) {
903 uint64_t HeighestAddr = ((PtrAddr + 0x800080008000) >> 48);
904 uint64_t HeigherAddr = ((PtrAddr + 0x80008000) >> 32);
905 uint64_t HiAddr = ((PtrAddr + 0x8000) >> 16);
906 Stub[8 * I + 0] = 0x3c190000 | (HeighestAddr & 0xFFFF); // lui $t9,ptr1
907 Stub[8 * I + 1] = 0x67390000 | (HeigherAddr & 0xFFFF); // daddiu $t9,$t9,%higher(ptr)
908 Stub[8 * I + 2] = 0x0019cc38; // dsll $t9,$t9,16
909 Stub[8 * I + 3] = 0x67390000 | (HiAddr & 0xFFFF); // daddiu $t9,$t9,%hi(ptr)
910 Stub[8 * I + 4] = 0x0019cc38; // dsll $t9,$t9,16
911 Stub[8 * I + 5] = 0xdf390000 | (PtrAddr & 0xFFFF); // ld $t9,%lo(ptr)
912 Stub[8 * I + 6] = 0x03200008; // jr $t9
913 Stub[8 * I + 7] = 0x00000000; // nop
914 }
915 }
916
917 void OrcRiscv64::writeResolverCode(char *ResolverWorkingMem,
918 ExecutorAddr ResolverTargetAddress,
919 ExecutorAddr ReentryFnAddr,
920 ExecutorAddr ReentryCtxAddr) {
921
922 const uint32_t ResolverCode[] = {
923 0xef810113, // 0x00: addi sp,sp,-264
924 0x00813023, // 0x04: sd s0,0(sp)
925 0x00913423, // 0x08: sd s1,8(sp)
926 0x01213823, // 0x0c: sd s2,16(sp)
927 0x01313c23, // 0x10: sd s3,24(sp)
928 0x03413023, // 0x14: sd s4,32(sp)
929 0x03513423, // 0x18: sd s5,40(sp)
930 0x03613823, // 0x1c: sd s6,48(sp)
931 0x03713c23, // 0x20: sd s7,56(sp)
932 0x05813023, // 0x24: sd s8,64(sp)
933 0x05913423, // 0x28: sd s9,72(sp)
934 0x05a13823, // 0x2c: sd s10,80(sp)
935 0x05b13c23, // 0x30: sd s11,88(sp)
936 0x06113023, // 0x34: sd ra,96(sp)
937 0x06a13423, // 0x38: sd a0,104(sp)
938 0x06b13823, // 0x3c: sd a1,112(sp)
939 0x06c13c23, // 0x40: sd a2,120(sp)
940 0x08d13023, // 0x44: sd a3,128(sp)
941 0x08e13423, // 0x48: sd a4,136(sp)
942 0x08f13823, // 0x4c: sd a5,144(sp)
943 0x09013c23, // 0x50: sd a6,152(sp)
944 0x0b113023, // 0x54: sd a7,160(sp)
945 0x0a813427, // 0x58: fsd fs0,168(sp)
946 0x0a913827, // 0x5c: fsd fs1,176(sp)
947 0x0b213c27, // 0x60: fsd fs2,184(sp)
948 0x0d313027, // 0x64: fsd fs3,192(sp)
949 0x0d413427, // 0x68: fsd fs4,200(sp)
950 0x0d513827, // 0x6c: fsd fs5,208(sp)
951 0x0d613c27, // 0x70: fsd fs6,216(sp)
952 0x0f713027, // 0x74: fsd fs7,224(sp)
953 0x0f813427, // 0x78: fsd fs8,232(sp)
954 0x0f913827, // 0x7c: fsd fs9,240(sp)
955 0x0fa13c27, // 0x80: fsd fs10,248(sp)
956 0x11b13027, // 0x84: fsd fs11,256(sp)
957 0x00000517, // 0x88: auipc a0,0x0
958 0x0b053503, // 0x8c: ld a0,176(a0) # 0x138
959 0x00030593, // 0x90: mv a1,t1
960 0xff458593, // 0x94: addi a1,a1,-12
961 0x00000617, // 0x98: auipc a2,0x0
962 0x0a863603, // 0x9c: ld a2,168(a2) # 0x140
963 0x000600e7, // 0xa0: jalr a2
964 0x00050293, // 0xa4: mv t0,a0
965 0x00013403, // 0xa8: ld s0,0(sp)
966 0x00813483, // 0xac: ld s1,8(sp)
967 0x01013903, // 0xb0: ld s2,16(sp)
968 0x01813983, // 0xb4: ld s3,24(sp)
969 0x02013a03, // 0xb8: ld s4,32(sp)
970 0x02813a83, // 0xbc: ld s5,40(sp)
971 0x03013b03, // 0xc0: ld s6,48(sp)
972 0x03813b83, // 0xc4: ld s7,56(sp)
973 0x04013c03, // 0xc8: ld s8,64(sp)
974 0x04813c83, // 0xcc: ld s9,72(sp)
975 0x05013d03, // 0xd0: ld s10,80(sp)
976 0x05813d83, // 0xd4: ld s11,88(sp)
977 0x06013083, // 0xd8: ld ra,96(sp)
978 0x06813503, // 0xdc: ld a0,104(sp)
979 0x07013583, // 0xe0: ld a1,112(sp)
980 0x07813603, // 0xe4: ld a2,120(sp)
981 0x08013683, // 0xe8: ld a3,128(sp)
982 0x08813703, // 0xec: ld a4,136(sp)
983 0x09013783, // 0xf0: ld a5,144(sp)
984 0x09813803, // 0xf4: ld a6,152(sp)
985 0x0a013883, // 0xf8: ld a7,160(sp)
986 0x0a813407, // 0xfc: fld fs0,168(sp)
987 0x0b013487, // 0x100: fld fs1,176(sp)
988 0x0b813907, // 0x104: fld fs2,184(sp)
989 0x0c013987, // 0x108: fld fs3,192(sp)
990 0x0c813a07, // 0x10c: fld fs4,200(sp)
991 0x0d013a87, // 0x110: fld fs5,208(sp)
992 0x0d813b07, // 0x114: fld fs6,216(sp)
993 0x0e013b87, // 0x118: fld fs7,224(sp)
994 0x0e813c07, // 0x11c: fld fs8,232(sp)
995 0x0f013c87, // 0x120: fld fs9,240(sp)
996 0x0f813d07, // 0x124: fld fs10,248(sp)
997 0x10013d87, // 0x128: fld fs11,256(sp)
998 0x10810113, // 0x12c: addi sp,sp,264
999 0x00028067, // 0x130: jr t0
1000 0x12345678, // 0x134: padding to align at 8 byte
1001 0x12345678, // 0x138: Lreentry_ctx_ptr:
1002 0xdeadbeef, // 0x13c: .quad 0
1003 0x98765432, // 0x140: Lreentry_fn_ptr:
1004 0xcafef00d // 0x144: .quad 0
1005 };
1006
1007 const unsigned ReentryCtxAddrOffset = 0x138;
1008 const unsigned ReentryFnAddrOffset = 0x140;
1009
1010 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
1011 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
1012 sizeof(uint64_t));
1013 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
1014 sizeof(uint64_t));
1015 }
1016
1017 void OrcRiscv64::writeTrampolines(char *TrampolineBlockWorkingMem,
1018 ExecutorAddr TrampolineBlockTargetAddress,
1019 ExecutorAddr ResolverAddr,
1020 unsigned NumTrampolines) {
1021
1022 unsigned OffsetToPtr = alignTo(NumTrampolines * TrampolineSize, 8);
1023
1024 memcpy(TrampolineBlockWorkingMem + OffsetToPtr, &ResolverAddr,
1025 sizeof(uint64_t));
1026
1027 uint32_t *Trampolines =
1028 reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
1029 for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize) {
1030 uint32_t Hi20 = (OffsetToPtr + 0x800) & 0xFFFFF000;
1031 uint32_t Lo12 = OffsetToPtr - Hi20;
1032 Trampolines[4 * I + 0] = 0x00000297 | Hi20; // auipc t0, %hi(Lptr)
1033 Trampolines[4 * I + 1] =
1034 0x0002b283 | ((Lo12 & 0xFFF) << 20); // ld t0, %lo(Lptr)
1035 Trampolines[4 * I + 2] = 0x00028367; // jalr t1, t0
1036 Trampolines[4 * I + 3] = 0xdeadface; // padding
1037 }
1038 }
1039
1040 void OrcRiscv64::writeIndirectStubsBlock(
1041 char *StubsBlockWorkingMem, ExecutorAddr StubsBlockTargetAddress,
1042 ExecutorAddr PointersBlockTargetAddress, unsigned NumStubs) {
1043 // Stub format is:
1044 //
1045 // .section __orc_stubs
1046 // stub1:
1047 // auipc t0, %hi(ptr1) ; PC-rel load of ptr1
1048 // ld t0, %lo(t0)
1049 // jr t0 ; Jump to resolver
1050 // .quad 0 ; Pad to 16 bytes
1051 // stub2:
1052 // auipc t0, %hi(ptr1) ; PC-rel load of ptr1
1053 // ld t0, %lo(t0)
1054 // jr t0 ; Jump to resolver
1055 // .quad 0
1056 //
1057 // ...
1058 //
1059 // .section __orc_ptrs
1060 // ptr1:
1061 // .quad 0x0
1062 // ptr2:
1063 // .quad 0x0
1064 //
1065 // ...
1066
1067 assert(stubAndPointerRangesOk<OrcRiscv64>(
1068 StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
1069 "PointersBlock is out of range");
1070
1071 uint32_t *Stub = reinterpret_cast<uint32_t *>(StubsBlockWorkingMem);
1072
1073 for (unsigned I = 0; I < NumStubs; ++I) {
1074 uint64_t PtrDisplacement =
1075 PointersBlockTargetAddress - StubsBlockTargetAddress;
1076 uint32_t Hi20 = (PtrDisplacement + 0x800) & 0xFFFFF000;
1077 uint32_t Lo12 = PtrDisplacement - Hi20;
1078 Stub[4 * I + 0] = 0x00000297 | Hi20; // auipc t0, %hi(Lptr)
1079 Stub[4 * I + 1] = 0x0002b283 | ((Lo12 & 0xFFF) << 20); // ld t0, %lo(Lptr)
1080 Stub[4 * I + 2] = 0x00028067; // jr t0
1081 Stub[4 * I + 3] = 0xfeedbeef; // padding
1082 PointersBlockTargetAddress += PointerSize;
1083 StubsBlockTargetAddress += StubSize;
1084 }
1085 }
1086
1087 void OrcLoongArch64::writeResolverCode(char *ResolverWorkingMem,
1088 ExecutorAddr ResolverTargetAddress,
1089 ExecutorAddr ReentryFnAddr,
1090 ExecutorAddr ReentryCtxAddr) {
1091
1092 LLVM_DEBUG({
1093 dbgs() << "Writing resolver code to "
1094 << formatv("{0:x16}", ResolverTargetAddress) << "\n";
1095 });
1096
1097 const uint32_t ResolverCode[] = {
1098 0x02fde063, // 0x0: addi.d $sp, $sp, -136(0xf78)
1099 0x29c00061, // 0x4: st.d $ra, $sp, 0
1100 0x29c02064, // 0x8: st.d $a0, $sp, 8(0x8)
1101 0x29c04065, // 0xc: st.d $a1, $sp, 16(0x10)
1102 0x29c06066, // 0x10: st.d $a2, $sp, 24(0x18)
1103 0x29c08067, // 0x14: st.d $a3, $sp, 32(0x20)
1104 0x29c0a068, // 0x18: st.d $a4, $sp, 40(0x28)
1105 0x29c0c069, // 0x1c: st.d $a5, $sp, 48(0x30)
1106 0x29c0e06a, // 0x20: st.d $a6, $sp, 56(0x38)
1107 0x29c1006b, // 0x24: st.d $a7, $sp, 64(0x40)
1108 0x2bc12060, // 0x28: fst.d $fa0, $sp, 72(0x48)
1109 0x2bc14061, // 0x2c: fst.d $fa1, $sp, 80(0x50)
1110 0x2bc16062, // 0x30: fst.d $fa2, $sp, 88(0x58)
1111 0x2bc18063, // 0x34: fst.d $fa3, $sp, 96(0x60)
1112 0x2bc1a064, // 0x38: fst.d $fa4, $sp, 104(0x68)
1113 0x2bc1c065, // 0x3c: fst.d $fa5, $sp, 112(0x70)
1114 0x2bc1e066, // 0x40: fst.d $fa6, $sp, 120(0x78)
1115 0x2bc20067, // 0x44: fst.d $fa7, $sp, 128(0x80)
1116 0x1c000004, // 0x48: pcaddu12i $a0, 0
1117 0x28c1c084, // 0x4c: ld.d $a0, $a0, 112(0x70)
1118 0x001501a5, // 0x50: move $a1, $t1
1119 0x02ffd0a5, // 0x54: addi.d $a1, $a1, -12(0xff4)
1120 0x1c000006, // 0x58: pcaddu12i $a2, 0
1121 0x28c1a0c6, // 0x5c: ld.d $a2, $a2, 104(0x68)
1122 0x4c0000c1, // 0x60: jirl $ra, $a2, 0
1123 0x0015008c, // 0x64: move $t0, $a0
1124 0x2b820067, // 0x68: fld.d $fa7, $sp, 128(0x80)
1125 0x2b81e066, // 0x6c: fld.d $fa6, $sp, 120(0x78)
1126 0x2b81c065, // 0x70: fld.d $fa5, $sp, 112(0x70)
1127 0x2b81a064, // 0x74: fld.d $fa4, $sp, 104(0x68)
1128 0x2b818063, // 0x78: fld.d $fa3, $sp, 96(0x60)
1129 0x2b816062, // 0x7c: fld.d $fa2, $sp, 88(0x58)
1130 0x2b814061, // 0x80: fld.d $fa1, $sp, 80(0x50)
1131 0x2b812060, // 0x84: fld.d $fa0, $sp, 72(0x48)
1132 0x28c1006b, // 0x88: ld.d $a7, $sp, 64(0x40)
1133 0x28c0e06a, // 0x8c: ld.d $a6, $sp, 56(0x38)
1134 0x28c0c069, // 0x90: ld.d $a5, $sp, 48(0x30)
1135 0x28c0a068, // 0x94: ld.d $a4, $sp, 40(0x28)
1136 0x28c08067, // 0x98: ld.d $a3, $sp, 32(0x20)
1137 0x28c06066, // 0x9c: ld.d $a2, $sp, 24(0x18)
1138 0x28c04065, // 0xa0: ld.d $a1, $sp, 16(0x10)
1139 0x28c02064, // 0xa4: ld.d $a0, $sp, 8(0x8)
1140 0x28c00061, // 0xa8: ld.d $ra, $sp, 0
1141 0x02c22063, // 0xac: addi.d $sp, $sp, 136(0x88)
1142 0x4c000180, // 0xb0: jr $t0
1143 0x00000000, // 0xb4: padding to align at 8 bytes
1144 0x01234567, // 0xb8: Lreentry_ctx_ptr:
1145 0xdeedbeef, // 0xbc: .dword 0
1146 0x98765432, // 0xc0: Lreentry_fn_ptr:
1147 0xcafef00d, // 0xc4: .dword 0
1148 };
1149
1150 const unsigned ReentryCtxAddrOffset = 0xb8;
1151 const unsigned ReentryFnAddrOffset = 0xc0;
1152
1153 memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
1154 memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
1155 sizeof(uint64_t));
1156 memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
1157 sizeof(uint64_t));
1158 }
1159
1160 void OrcLoongArch64::writeTrampolines(char *TrampolineBlockWorkingMem,
1161 ExecutorAddr TrampolineBlockTargetAddress,
1162 ExecutorAddr ResolverAddr,
1163 unsigned NumTrampolines) {
1164
1165 LLVM_DEBUG({
1166 dbgs() << "Writing trampoline code to "
1167 << formatv("{0:x16}", TrampolineBlockTargetAddress) << "\n";
1168 });
1169
1170 unsigned OffsetToPtr = alignTo(NumTrampolines * TrampolineSize, 8);
1171
1172 memcpy(TrampolineBlockWorkingMem + OffsetToPtr, &ResolverAddr,
1173 sizeof(uint64_t));
1174
1175 uint32_t *Trampolines =
1176 reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
1177 for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize) {
1178 uint32_t Hi20 = (OffsetToPtr + 0x800) & 0xfffff000;
1179 uint32_t Lo12 = OffsetToPtr - Hi20;
1180 Trampolines[4 * I + 0] =
1181 0x1c00000c |
1182 (((Hi20 >> 12) & 0xfffff) << 5); // pcaddu12i $t0, %pc_hi20(Lptr)
1183 Trampolines[4 * I + 1] =
1184 0x28c0018c | ((Lo12 & 0xfff) << 10); // ld.d $t0, $t0, %pc_lo12(Lptr)
1185 Trampolines[4 * I + 2] = 0x4c00018d; // jirl $t1, $t0, 0
1186 Trampolines[4 * I + 3] = 0x0; // padding
1187 }
1188 }
1189
1190 void OrcLoongArch64::writeIndirectStubsBlock(
1191 char *StubsBlockWorkingMem, ExecutorAddr StubsBlockTargetAddress,
1192 ExecutorAddr PointersBlockTargetAddress, unsigned NumStubs) {
1193 // Stub format is:
1194 //
1195 // .section __orc_stubs
1196 // stub1:
1197 // pcaddu12i $t0, %pc_hi20(ptr1) ; PC-rel load of ptr1
1198 // ld.d $t0, $t0, %pc_lo12(ptr1)
1199 // jr $t0 ; Jump to resolver
1200 // .dword 0 ; Pad to 16 bytes
1201 // stub2:
1202 // pcaddu12i $t0, %pc_hi20(ptr2) ; PC-rel load of ptr2
1203 // ld.d $t0, $t0, %pc_lo12(ptr2)
1204 // jr $t0 ; Jump to resolver
1205 // .dword 0 ; Pad to 16 bytes
1206 // ...
1207 //
1208 // .section __orc_ptrs
1209 // ptr1:
1210 // .dword 0x0
1211 // ptr2:
1212 // .dword 0x0
1213 // ...
1214 LLVM_DEBUG({
1215 dbgs() << "Writing stubs code to "
1216 << formatv("{0:x16}", StubsBlockTargetAddress) << "\n";
1217 });
1218 assert(stubAndPointerRangesOk<OrcLoongArch64>(
1219 StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
1220 "PointersBlock is out of range");
1221
1222 uint32_t *Stub = reinterpret_cast<uint32_t *>(StubsBlockWorkingMem);
1223
1224 for (unsigned I = 0; I < NumStubs; ++I) {
1225 uint64_t PtrDisplacement =
1226 PointersBlockTargetAddress - StubsBlockTargetAddress;
1227 uint32_t Hi20 = (PtrDisplacement + 0x800) & 0xfffff000;
1228 uint32_t Lo12 = PtrDisplacement - Hi20;
1229 Stub[4 * I + 0] = 0x1c00000c | (((Hi20 >> 12) & 0xfffff)
1230 << 5); // pcaddu12i $t0, %pc_hi20(Lptr)
1231 Stub[4 * I + 1] =
1232 0x28c0018c | ((Lo12 & 0xfff) << 10); // ld.d $t0, $t0, %pc_lo12(Lptr)
1233 Stub[4 * I + 2] = 0x4c000180; // jr $t0
1234 Stub[4 * I + 3] = 0x0; // padding
1235 PointersBlockTargetAddress += PointerSize;
1236 StubsBlockTargetAddress += StubSize;
1237 }
1238 }
1239
1240 } // End namespace orc.
1241 } // End namespace llvm.
LLVM monorepo