LLVM  15.0.0git
OrcABISupport.cpp
Go to the documentation of this file.
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/Process.h"
12 #include "llvm/Support/raw_ostream.h"
13 
14 #define DEBUG_TYPE "orc"
15 
16 using namespace llvm;
17 
18 template <typename ORCABI>
19 bool stubAndPointerRangesOk(JITTargetAddress StubBlockAddr,
20  JITTargetAddress PointerBlockAddr,
21  unsigned NumStubs) {
22  constexpr unsigned MaxDisp = ORCABI::StubToPointerMaxDisplacement;
23  JITTargetAddress FirstStub = StubBlockAddr;
24  JITTargetAddress LastStub = FirstStub + ((NumStubs - 1) * ORCABI::StubSize);
25  JITTargetAddress FirstPointer = PointerBlockAddr;
26  JITTargetAddress LastPointer =
27  FirstPointer + ((NumStubs - 1) * ORCABI::StubSize);
28 
29  if (FirstStub < FirstPointer) {
30  if (LastStub >= FirstPointer)
31  return false; // Ranges overlap.
32  return (FirstPointer - FirstStub <= MaxDisp) &&
33  (LastPointer - LastStub <= MaxDisp); // out-of-range.
34  }
35 
36  if (LastPointer >= FirstStub)
37  return false; // Ranges overlap.
38 
39  return (FirstStub - FirstPointer <= MaxDisp) &&
40  (LastStub - LastPointer <= MaxDisp);
41 }
42 
43 namespace llvm {
44 namespace orc {
45 
46 void OrcAArch64::writeResolverCode(char *ResolverWorkingMem,
47  JITTargetAddress ResolverTargetAddress,
48  JITTargetAddress ReentryFnAddr,
49  JITTargetAddress ReentryCtxAddr) {
50 
51  const uint32_t ResolverCode[] = {
52  // resolver_entry:
53  0xa9bf47fd, // 0x000: stp x29, x17, [sp, #-16]!
54  0x910003fd, // 0x004: mov x29, sp
55  0xa9bf73fb, // 0x008: stp x27, x28, [sp, #-16]!
56  0xa9bf6bf9, // 0x00c: stp x25, x26, [sp, #-16]!
57  0xa9bf63f7, // 0x010: stp x23, x24, [sp, #-16]!
58  0xa9bf5bf5, // 0x014: stp x21, x22, [sp, #-16]!
59  0xa9bf53f3, // 0x018: stp x19, x20, [sp, #-16]!
60  0xa9bf3fee, // 0x01c: stp x14, x15, [sp, #-16]!
61  0xa9bf37ec, // 0x020: stp x12, x13, [sp, #-16]!
62  0xa9bf2fea, // 0x024: stp x10, x11, [sp, #-16]!
63  0xa9bf27e8, // 0x028: stp x8, x9, [sp, #-16]!
64  0xa9bf1fe6, // 0x02c: stp x6, x7, [sp, #-16]!
65  0xa9bf17e4, // 0x030: stp x4, x5, [sp, #-16]!
66  0xa9bf0fe2, // 0x034: stp x2, x3, [sp, #-16]!
67  0xa9bf07e0, // 0x038: stp x0, x1, [sp, #-16]!
68  0xadbf7ffe, // 0x03c: stp q30, q31, [sp, #-32]!
69  0xadbf77fc, // 0x040: stp q28, q29, [sp, #-32]!
70  0xadbf6ffa, // 0x044: stp q26, q27, [sp, #-32]!
71  0xadbf67f8, // 0x048: stp q24, q25, [sp, #-32]!
72  0xadbf5ff6, // 0x04c: stp q22, q23, [sp, #-32]!
73  0xadbf57f4, // 0x050: stp q20, q21, [sp, #-32]!
74  0xadbf4ff2, // 0x054: stp q18, q19, [sp, #-32]!
75  0xadbf47f0, // 0x058: stp q16, q17, [sp, #-32]!
76  0xadbf3fee, // 0x05c: stp q14, q15, [sp, #-32]!
77  0xadbf37ec, // 0x060: stp q12, q13, [sp, #-32]!
78  0xadbf2fea, // 0x064: stp q10, q11, [sp, #-32]!
79  0xadbf27e8, // 0x068: stp q8, q9, [sp, #-32]!
80  0xadbf1fe6, // 0x06c: stp q6, q7, [sp, #-32]!
81  0xadbf17e4, // 0x070: stp q4, q5, [sp, #-32]!
82  0xadbf0fe2, // 0x074: stp q2, q3, [sp, #-32]!
83  0xadbf07e0, // 0x078: stp q0, q1, [sp, #-32]!
84  0x580004e0, // 0x07c: ldr x0, Lreentry_ctx_ptr
85  0xaa1e03e1, // 0x080: mov x1, x30
86  0xd1003021, // 0x084: sub x1, x1, #12
87  0x58000442, // 0x088: ldr x2, Lreentry_fn_ptr
88  0xd63f0040, // 0x08c: blr x2
89  0xaa0003f1, // 0x090: mov x17, x0
90  0xacc107e0, // 0x094: ldp q0, q1, [sp], #32
91  0xacc10fe2, // 0x098: ldp q2, q3, [sp], #32
92  0xacc117e4, // 0x09c: ldp q4, q5, [sp], #32
93  0xacc11fe6, // 0x0a0: ldp q6, q7, [sp], #32
94  0xacc127e8, // 0x0a4: ldp q8, q9, [sp], #32
95  0xacc12fea, // 0x0a8: ldp q10, q11, [sp], #32
96  0xacc137ec, // 0x0ac: ldp q12, q13, [sp], #32
97  0xacc13fee, // 0x0b0: ldp q14, q15, [sp], #32
98  0xacc147f0, // 0x0b4: ldp q16, q17, [sp], #32
99  0xacc14ff2, // 0x0b8: ldp q18, q19, [sp], #32
100  0xacc157f4, // 0x0bc: ldp q20, q21, [sp], #32
101  0xacc15ff6, // 0x0c0: ldp q22, q23, [sp], #32
102  0xacc167f8, // 0x0c4: ldp q24, q25, [sp], #32
103  0xacc16ffa, // 0x0c8: ldp q26, q27, [sp], #32
104  0xacc177fc, // 0x0cc: ldp q28, q29, [sp], #32
105  0xacc17ffe, // 0x0d0: ldp q30, q31, [sp], #32
106  0xa8c107e0, // 0x0d4: ldp x0, x1, [sp], #16
107  0xa8c10fe2, // 0x0d8: ldp x2, x3, [sp], #16
108  0xa8c117e4, // 0x0dc: ldp x4, x5, [sp], #16
109  0xa8c11fe6, // 0x0e0: ldp x6, x7, [sp], #16
110  0xa8c127e8, // 0x0e4: ldp x8, x9, [sp], #16
111  0xa8c12fea, // 0x0e8: ldp x10, x11, [sp], #16
112  0xa8c137ec, // 0x0ec: ldp x12, x13, [sp], #16
113  0xa8c13fee, // 0x0f0: ldp x14, x15, [sp], #16
114  0xa8c153f3, // 0x0f4: ldp x19, x20, [sp], #16
115  0xa8c15bf5, // 0x0f8: ldp x21, x22, [sp], #16
116  0xa8c163f7, // 0x0fc: ldp x23, x24, [sp], #16
117  0xa8c16bf9, // 0x100: ldp x25, x26, [sp], #16
118  0xa8c173fb, // 0x104: ldp x27, x28, [sp], #16
119  0xa8c17bfd, // 0x108: ldp x29, x30, [sp], #16
120  0xd65f0220, // 0x10c: ret x17
121  0x01234567, // 0x110: Lreentry_fn_ptr:
122  0xdeadbeef, // 0x114: .quad 0
123  0x98765432, // 0x118: Lreentry_ctx_ptr:
124  0xcafef00d // 0x11c: .quad 0
125  };
126 
127  const unsigned ReentryFnAddrOffset = 0x110;
128  const unsigned ReentryCtxAddrOffset = 0x118;
129 
130  memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
131  memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
132  sizeof(uint64_t));
133  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
134  sizeof(uint64_t));
135 }
136 
137 void OrcAArch64::writeTrampolines(char *TrampolineBlockWorkingMem,
138  JITTargetAddress TrampolineBlockTargetAddress,
139  JITTargetAddress ResolverAddr,
140  unsigned NumTrampolines) {
141 
142  unsigned OffsetToPtr = alignTo(NumTrampolines * TrampolineSize, 8);
143 
144  memcpy(TrampolineBlockWorkingMem + OffsetToPtr, &ResolverAddr,
145  sizeof(uint64_t));
146 
147  // OffsetToPtr is actually the offset from the PC for the 2nd instruction, so
148  // subtract 32-bits.
149  OffsetToPtr -= 4;
150 
151  uint32_t *Trampolines =
152  reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
153 
154  for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize) {
155  Trampolines[3 * I + 0] = 0xaa1e03f1; // mov x17, x30
156  Trampolines[3 * I + 1] = 0x58000010 | (OffsetToPtr << 3); // adr x16, Lptr
157  Trampolines[3 * I + 2] = 0xd63f0200; // blr x16
158  }
159 }
160 
161 void OrcAArch64::writeIndirectStubsBlock(
162  char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress,
163  JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs) {
164  // Stub format is:
165  //
166  // .section __orc_stubs
167  // stub1:
168  // ldr x0, ptr1 ; PC-rel load of ptr1
169  // br x0 ; Jump to resolver
170  // stub2:
171  // ldr x0, ptr2 ; PC-rel load of ptr2
172  // br x0 ; Jump to resolver
173  //
174  // ...
175  //
176  // .section __orc_ptrs
177  // ptr1:
178  // .quad 0x0
179  // ptr2:
180  // .quad 0x0
181  //
182  // ...
183 
184  static_assert(StubSize == PointerSize,
185  "Pointer and stub size must match for algorithm below");
186  assert(stubAndPointerRangesOk<OrcAArch64>(
187  StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
188  "PointersBlock is out of range");
189  uint64_t PtrDisplacement =
190  PointersBlockTargetAddress - StubsBlockTargetAddress;
191  uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlockWorkingMem);
192  uint64_t PtrOffsetField = PtrDisplacement << 3;
193 
194  for (unsigned I = 0; I < NumStubs; ++I)
195  Stub[I] = 0xd61f020058000010 | PtrOffsetField;
196 }
197 
198 void OrcX86_64_Base::writeTrampolines(
199  char *TrampolineBlockWorkingMem,
200  JITTargetAddress TrampolineBlockTargetAddress,
201  JITTargetAddress ResolverAddr, unsigned NumTrampolines) {
202 
203  unsigned OffsetToPtr = NumTrampolines * TrampolineSize;
204 
205  memcpy(TrampolineBlockWorkingMem + OffsetToPtr, &ResolverAddr,
206  sizeof(uint64_t));
207 
208  uint64_t *Trampolines =
209  reinterpret_cast<uint64_t *>(TrampolineBlockWorkingMem);
210  uint64_t CallIndirPCRel = 0xf1c40000000015ff;
211 
212  for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize)
213  Trampolines[I] = CallIndirPCRel | ((OffsetToPtr - 6) << 16);
214 }
215 
216 void OrcX86_64_Base::writeIndirectStubsBlock(
217  char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress,
218  JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs) {
219  // Stub format is:
220  //
221  // .section __orc_stubs
222  // stub1:
223  // jmpq *ptr1(%rip)
224  // .byte 0xC4 ; <- Invalid opcode padding.
225  // .byte 0xF1
226  // stub2:
227  // jmpq *ptr2(%rip)
228  //
229  // ...
230  //
231  // .section __orc_ptrs
232  // ptr1:
233  // .quad 0x0
234  // ptr2:
235  // .quad 0x0
236  //
237  // ...
238 
239  // Populate the stubs page stubs and mark it executable.
240  static_assert(StubSize == PointerSize,
241  "Pointer and stub size must match for algorithm below");
242  assert(stubAndPointerRangesOk<OrcX86_64_Base>(
243  StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
244  "PointersBlock is out of range");
245  uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlockWorkingMem);
246  uint64_t PtrOffsetField =
247  (PointersBlockTargetAddress - StubsBlockTargetAddress - 6) << 16;
248  for (unsigned I = 0; I < NumStubs; ++I)
249  Stub[I] = 0xF1C40000000025ff | PtrOffsetField;
250 }
251 
252 void OrcX86_64_SysV::writeResolverCode(char *ResolverWorkingMem,
253  JITTargetAddress ResolverTargetAddress,
254  JITTargetAddress ReentryFnAddr,
255  JITTargetAddress ReentryCtxAddr) {
256 
257  LLVM_DEBUG({
258  dbgs() << "Writing resolver code to "
259  << formatv("{0:x16}", ResolverTargetAddress) << "\n";
260  });
261 
262  const uint8_t ResolverCode[] = {
263  // resolver_entry:
264  0x55, // 0x00: pushq %rbp
265  0x48, 0x89, 0xe5, // 0x01: movq %rsp, %rbp
266  0x50, // 0x04: pushq %rax
267  0x53, // 0x05: pushq %rbx
268  0x51, // 0x06: pushq %rcx
269  0x52, // 0x07: pushq %rdx
270  0x56, // 0x08: pushq %rsi
271  0x57, // 0x09: pushq %rdi
272  0x41, 0x50, // 0x0a: pushq %r8
273  0x41, 0x51, // 0x0c: pushq %r9
274  0x41, 0x52, // 0x0e: pushq %r10
275  0x41, 0x53, // 0x10: pushq %r11
276  0x41, 0x54, // 0x12: pushq %r12
277  0x41, 0x55, // 0x14: pushq %r13
278  0x41, 0x56, // 0x16: pushq %r14
279  0x41, 0x57, // 0x18: pushq %r15
280  0x48, 0x81, 0xec, 0x08, 0x02, 0x00, 0x00, // 0x1a: subq 0x208, %rsp
281  0x48, 0x0f, 0xae, 0x04, 0x24, // 0x21: fxsave64 (%rsp)
282  0x48, 0xbf, // 0x26: movabsq <CBMgr>, %rdi
283 
284  // 0x28: JIT re-entry ctx addr.
285  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
286 
287  0x48, 0x8b, 0x75, 0x08, // 0x30: movq 8(%rbp), %rsi
288  0x48, 0x83, 0xee, 0x06, // 0x34: subq $6, %rsi
289  0x48, 0xb8, // 0x38: movabsq <REntry>, %rax
290 
291  // 0x3a: JIT re-entry fn addr:
292  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
293 
294  0xff, 0xd0, // 0x42: callq *%rax
295  0x48, 0x89, 0x45, 0x08, // 0x44: movq %rax, 8(%rbp)
296  0x48, 0x0f, 0xae, 0x0c, 0x24, // 0x48: fxrstor64 (%rsp)
297  0x48, 0x81, 0xc4, 0x08, 0x02, 0x00, 0x00, // 0x4d: addq 0x208, %rsp
298  0x41, 0x5f, // 0x54: popq %r15
299  0x41, 0x5e, // 0x56: popq %r14
300  0x41, 0x5d, // 0x58: popq %r13
301  0x41, 0x5c, // 0x5a: popq %r12
302  0x41, 0x5b, // 0x5c: popq %r11
303  0x41, 0x5a, // 0x5e: popq %r10
304  0x41, 0x59, // 0x60: popq %r9
305  0x41, 0x58, // 0x62: popq %r8
306  0x5f, // 0x64: popq %rdi
307  0x5e, // 0x65: popq %rsi
308  0x5a, // 0x66: popq %rdx
309  0x59, // 0x67: popq %rcx
310  0x5b, // 0x68: popq %rbx
311  0x58, // 0x69: popq %rax
312  0x5d, // 0x6a: popq %rbp
313  0xc3, // 0x6b: retq
314  };
315 
316  const unsigned ReentryFnAddrOffset = 0x3a;
317  const unsigned ReentryCtxAddrOffset = 0x28;
318 
319  memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
320  memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
321  sizeof(uint64_t));
322  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
323  sizeof(uint64_t));
324 }
325 
326 void OrcX86_64_Win32::writeResolverCode(char *ResolverWorkingMem,
327  JITTargetAddress ResolverTargetAddress,
328  JITTargetAddress ReentryFnAddr,
329  JITTargetAddress ReentryCtxAddr) {
330 
331  // resolverCode is similar to OrcX86_64 with differences specific to windows
332  // x64 calling convention: arguments go into rcx, rdx and come in reverse
333  // order, shadow space allocation on stack
334  const uint8_t ResolverCode[] = {
335  // resolver_entry:
336  0x55, // 0x00: pushq %rbp
337  0x48, 0x89, 0xe5, // 0x01: movq %rsp, %rbp
338  0x50, // 0x04: pushq %rax
339  0x53, // 0x05: pushq %rbx
340  0x51, // 0x06: pushq %rcx
341  0x52, // 0x07: pushq %rdx
342  0x56, // 0x08: pushq %rsi
343  0x57, // 0x09: pushq %rdi
344  0x41, 0x50, // 0x0a: pushq %r8
345  0x41, 0x51, // 0x0c: pushq %r9
346  0x41, 0x52, // 0x0e: pushq %r10
347  0x41, 0x53, // 0x10: pushq %r11
348  0x41, 0x54, // 0x12: pushq %r12
349  0x41, 0x55, // 0x14: pushq %r13
350  0x41, 0x56, // 0x16: pushq %r14
351  0x41, 0x57, // 0x18: pushq %r15
352  0x48, 0x81, 0xec, 0x08, 0x02, 0x00, 0x00, // 0x1a: subq 0x208, %rsp
353  0x48, 0x0f, 0xae, 0x04, 0x24, // 0x21: fxsave64 (%rsp)
354 
355  0x48, 0xb9, // 0x26: movabsq <CBMgr>, %rcx
356  // 0x28: JIT re-entry ctx addr.
357  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
358 
359  0x48, 0x8B, 0x55, 0x08, // 0x30: mov rdx, [rbp+0x8]
360  0x48, 0x83, 0xea, 0x06, // 0x34: sub rdx, 0x6
361 
362  0x48, 0xb8, // 0x38: movabsq <REntry>, %rax
363  // 0x3a: JIT re-entry fn addr:
364  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
365 
366  // 0x42: sub rsp, 0x20 (Allocate shadow space)
367  0x48, 0x83, 0xEC, 0x20,
368  0xff, 0xd0, // 0x46: callq *%rax
369 
370  // 0x48: add rsp, 0x20 (Free shadow space)
371  0x48, 0x83, 0xC4, 0x20,
372 
373  0x48, 0x89, 0x45, 0x08, // 0x4C: movq %rax, 8(%rbp)
374  0x48, 0x0f, 0xae, 0x0c, 0x24, // 0x50: fxrstor64 (%rsp)
375  0x48, 0x81, 0xc4, 0x08, 0x02, 0x00, 0x00, // 0x55: addq 0x208, %rsp
376  0x41, 0x5f, // 0x5C: popq %r15
377  0x41, 0x5e, // 0x5E: popq %r14
378  0x41, 0x5d, // 0x60: popq %r13
379  0x41, 0x5c, // 0x62: popq %r12
380  0x41, 0x5b, // 0x64: popq %r11
381  0x41, 0x5a, // 0x66: popq %r10
382  0x41, 0x59, // 0x68: popq %r9
383  0x41, 0x58, // 0x6a: popq %r8
384  0x5f, // 0x6c: popq %rdi
385  0x5e, // 0x6d: popq %rsi
386  0x5a, // 0x6e: popq %rdx
387  0x59, // 0x6f: popq %rcx
388  0x5b, // 0x70: popq %rbx
389  0x58, // 0x71: popq %rax
390  0x5d, // 0x72: popq %rbp
391  0xc3, // 0x73: retq
392  };
393 
394  const unsigned ReentryFnAddrOffset = 0x3a;
395  const unsigned ReentryCtxAddrOffset = 0x28;
396 
397  memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
398  memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
399  sizeof(uint64_t));
400  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
401  sizeof(uint64_t));
402 }
403 
404 void OrcI386::writeResolverCode(char *ResolverWorkingMem,
405  JITTargetAddress ResolverTargetAddress,
406  JITTargetAddress ReentryFnAddr,
407  JITTargetAddress ReentryCtxAddr) {
408 
409  assert((ReentryFnAddr >> 32) == 0 && "ReentryFnAddr out of range");
410  assert((ReentryCtxAddr >> 32) == 0 && "ReentryCtxAddr out of range");
411 
412  const uint8_t ResolverCode[] = {
413  // resolver_entry:
414  0x55, // 0x00: pushl %ebp
415  0x89, 0xe5, // 0x01: movl %esp, %ebp
416  0x54, // 0x03: pushl %esp
417  0x83, 0xe4, 0xf0, // 0x04: andl $-0x10, %esp
418  0x50, // 0x07: pushl %eax
419  0x53, // 0x08: pushl %ebx
420  0x51, // 0x09: pushl %ecx
421  0x52, // 0x0a: pushl %edx
422  0x56, // 0x0b: pushl %esi
423  0x57, // 0x0c: pushl %edi
424  0x81, 0xec, 0x18, 0x02, 0x00, 0x00, // 0x0d: subl $0x218, %esp
425  0x0f, 0xae, 0x44, 0x24, 0x10, // 0x13: fxsave 0x10(%esp)
426  0x8b, 0x75, 0x04, // 0x18: movl 0x4(%ebp), %esi
427  0x83, 0xee, 0x05, // 0x1b: subl $0x5, %esi
428  0x89, 0x74, 0x24, 0x04, // 0x1e: movl %esi, 0x4(%esp)
429  0xc7, 0x04, 0x24, 0x00, 0x00, 0x00,
430  0x00, // 0x22: movl <cbmgr>, (%esp)
431  0xb8, 0x00, 0x00, 0x00, 0x00, // 0x29: movl <reentry>, %eax
432  0xff, 0xd0, // 0x2e: calll *%eax
433  0x89, 0x45, 0x04, // 0x30: movl %eax, 0x4(%ebp)
434  0x0f, 0xae, 0x4c, 0x24, 0x10, // 0x33: fxrstor 0x10(%esp)
435  0x81, 0xc4, 0x18, 0x02, 0x00, 0x00, // 0x38: addl $0x218, %esp
436  0x5f, // 0x3e: popl %edi
437  0x5e, // 0x3f: popl %esi
438  0x5a, // 0x40: popl %edx
439  0x59, // 0x41: popl %ecx
440  0x5b, // 0x42: popl %ebx
441  0x58, // 0x43: popl %eax
442  0x8b, 0x65, 0xfc, // 0x44: movl -0x4(%ebp), %esp
443  0x5d, // 0x48: popl %ebp
444  0xc3 // 0x49: retl
445  };
446 
447  const unsigned ReentryFnAddrOffset = 0x2a;
448  const unsigned ReentryCtxAddrOffset = 0x25;
449 
450  memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
451  memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
452  sizeof(uint32_t));
453  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
454  sizeof(uint32_t));
455 }
456 
457 void OrcI386::writeTrampolines(char *TrampolineWorkingMem,
458  JITTargetAddress TrampolineBlockTargetAddress,
459  JITTargetAddress ResolverAddr,
460  unsigned NumTrampolines) {
461  assert((ResolverAddr >> 32) == 0 && "ResolverAddr out of range");
462 
463  uint64_t CallRelImm = 0xF1C4C400000000e8;
464  uint64_t ResolverRel = ResolverAddr - TrampolineBlockTargetAddress - 5;
465 
466  uint64_t *Trampolines = reinterpret_cast<uint64_t *>(TrampolineWorkingMem);
467  for (unsigned I = 0; I < NumTrampolines; ++I, ResolverRel -= TrampolineSize)
468  Trampolines[I] = CallRelImm | (ResolverRel << 8);
469 }
470 
471 void OrcI386::writeIndirectStubsBlock(
472  char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress,
473  JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs) {
474  assert((StubsBlockTargetAddress >> 32) == 0 &&
475  "StubsBlockTargetAddress is out of range");
476  assert((PointersBlockTargetAddress >> 32) == 0 &&
477  "PointersBlockTargetAddress is out of range");
478 
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  assert(stubAndPointerRangesOk<OrcI386>(
500  StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
501  "PointersBlock is out of range");
502 
503  uint64_t *Stub = reinterpret_cast<uint64_t *>(StubsBlockWorkingMem);
504  uint64_t PtrAddr = PointersBlockTargetAddress;
505  for (unsigned I = 0; I < NumStubs; ++I, PtrAddr += 4)
506  Stub[I] = 0xF1C40000000025ff | (PtrAddr << 16);
507 }
508 
509 void OrcMips32_Base::writeResolverCode(char *ResolverWorkingMem,
510  JITTargetAddress ResolverTargetAddress,
511  JITTargetAddress ReentryFnAddr,
512  JITTargetAddress ReentryCtxAddr,
513  bool isBigEndian) {
514 
515  const uint32_t ResolverCode[] = {
516  // resolver_entry:
517  0x27bdff98, // 0x00: addiu $sp,$sp,-104
518  0xafa20000, // 0x04: sw $v0,0($sp)
519  0xafa30004, // 0x08: sw $v1,4($sp)
520  0xafa40008, // 0x0c: sw $a0,8($sp)
521  0xafa5000c, // 0x10: sw $a1,12($sp)
522  0xafa60010, // 0x14: sw $a2,16($sp)
523  0xafa70014, // 0x18: sw $a3,20($sp)
524  0xafb00018, // 0x1c: sw $s0,24($sp)
525  0xafb1001c, // 0x20: sw $s1,28($sp)
526  0xafb20020, // 0x24: sw $s2,32($sp)
527  0xafb30024, // 0x28: sw $s3,36($sp)
528  0xafb40028, // 0x2c: sw $s4,40($sp)
529  0xafb5002c, // 0x30: sw $s5,44($sp)
530  0xafb60030, // 0x34: sw $s6,48($sp)
531  0xafb70034, // 0x38: sw $s7,52($sp)
532  0xafa80038, // 0x3c: sw $t0,56($sp)
533  0xafa9003c, // 0x40: sw $t1,60($sp)
534  0xafaa0040, // 0x44: sw $t2,64($sp)
535  0xafab0044, // 0x48: sw $t3,68($sp)
536  0xafac0048, // 0x4c: sw $t4,72($sp)
537  0xafad004c, // 0x50: sw $t5,76($sp)
538  0xafae0050, // 0x54: sw $t6,80($sp)
539  0xafaf0054, // 0x58: sw $t7,84($sp)
540  0xafb80058, // 0x5c: sw $t8,88($sp)
541  0xafb9005c, // 0x60: sw $t9,92($sp)
542  0xafbe0060, // 0x64: sw $fp,96($sp)
543  0xafbf0064, // 0x68: sw $ra,100($sp)
544 
545  // JIT re-entry ctx addr.
546  0x00000000, // 0x6c: lui $a0,ctx
547  0x00000000, // 0x70: addiu $a0,$a0,ctx
548 
549  0x03e02825, // 0x74: move $a1, $ra
550  0x24a5ffec, // 0x78: addiu $a1,$a1,-20
551 
552  // JIT re-entry fn addr:
553  0x00000000, // 0x7c: lui $t9,reentry
554  0x00000000, // 0x80: addiu $t9,$t9,reentry
555 
556  0x0320f809, // 0x84: jalr $t9
557  0x00000000, // 0x88: nop
558  0x8fbf0064, // 0x8c: lw $ra,100($sp)
559  0x8fbe0060, // 0x90: lw $fp,96($sp)
560  0x8fb9005c, // 0x94: lw $t9,92($sp)
561  0x8fb80058, // 0x98: lw $t8,88($sp)
562  0x8faf0054, // 0x9c: lw $t7,84($sp)
563  0x8fae0050, // 0xa0: lw $t6,80($sp)
564  0x8fad004c, // 0xa4: lw $t5,76($sp)
565  0x8fac0048, // 0xa8: lw $t4,72($sp)
566  0x8fab0044, // 0xac: lw $t3,68($sp)
567  0x8faa0040, // 0xb0: lw $t2,64($sp)
568  0x8fa9003c, // 0xb4: lw $t1,60($sp)
569  0x8fa80038, // 0xb8: lw $t0,56($sp)
570  0x8fb70034, // 0xbc: lw $s7,52($sp)
571  0x8fb60030, // 0xc0: lw $s6,48($sp)
572  0x8fb5002c, // 0xc4: lw $s5,44($sp)
573  0x8fb40028, // 0xc8: lw $s4,40($sp)
574  0x8fb30024, // 0xcc: lw $s3,36($sp)
575  0x8fb20020, // 0xd0: lw $s2,32($sp)
576  0x8fb1001c, // 0xd4: lw $s1,28($sp)
577  0x8fb00018, // 0xd8: lw $s0,24($sp)
578  0x8fa70014, // 0xdc: lw $a3,20($sp)
579  0x8fa60010, // 0xe0: lw $a2,16($sp)
580  0x8fa5000c, // 0xe4: lw $a1,12($sp)
581  0x8fa40008, // 0xe8: lw $a0,8($sp)
582  0x27bd0068, // 0xec: addiu $sp,$sp,104
583  0x0300f825, // 0xf0: move $ra, $t8
584  0x03200008, // 0xf4: jr $t9
585  0x00000000, // 0xf8: move $t9, $v0/v1
586  };
587 
588  const unsigned ReentryFnAddrOffset = 0x7c; // JIT re-entry fn addr lui
589  const unsigned ReentryCtxAddrOffset = 0x6c; // JIT re-entry context addr lui
590  const unsigned Offsett = 0xf8;
591 
592  memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
593 
594  // Depending on endian return value will be in v0 or v1.
595  uint32_t MoveVxT9 = isBigEndian ? 0x0060c825 : 0x0040c825;
596  memcpy(ResolverWorkingMem + Offsett, &MoveVxT9, sizeof(MoveVxT9));
597 
598  uint32_t ReentryCtxLUi =
599  0x3c040000 | (((ReentryCtxAddr + 0x8000) >> 16) & 0xFFFF);
600  uint32_t ReentryCtxADDiu = 0x24840000 | ((ReentryCtxAddr)&0xFFFF);
601  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxLUi,
602  sizeof(ReentryCtxLUi));
603  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset + 4, &ReentryCtxADDiu,
604  sizeof(ReentryCtxADDiu));
605 
606  uint32_t ReentryFnLUi =
607  0x3c190000 | (((ReentryFnAddr + 0x8000) >> 16) & 0xFFFF);
608  uint32_t ReentryFnADDiu = 0x27390000 | ((ReentryFnAddr)&0xFFFF);
609  memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnLUi,
610  sizeof(ReentryFnLUi));
611  memcpy(ResolverWorkingMem + ReentryFnAddrOffset + 4, &ReentryFnADDiu,
612  sizeof(ReentryFnADDiu));
613 }
614 
615 void OrcMips32_Base::writeTrampolines(
616  char *TrampolineBlockWorkingMem,
617  JITTargetAddress TrampolineBlockTargetAddress,
618  JITTargetAddress ResolverAddr, unsigned NumTrampolines) {
619 
620  assert((ResolverAddr >> 32) == 0 && "ResolverAddr out of range");
621 
622  uint32_t *Trampolines =
623  reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
624  uint32_t RHiAddr = ((ResolverAddr + 0x8000) >> 16);
625 
626  for (unsigned I = 0; I < NumTrampolines; ++I) {
627  // move $t8,$ra
628  // lui $t9,ResolverAddr
629  // addiu $t9,$t9,ResolverAddr
630  // jalr $t9
631  // nop
632  Trampolines[5 * I + 0] = 0x03e0c025;
633  Trampolines[5 * I + 1] = 0x3c190000 | (RHiAddr & 0xFFFF);
634  Trampolines[5 * I + 2] = 0x27390000 | (ResolverAddr & 0xFFFF);
635  Trampolines[5 * I + 3] = 0x0320f809;
636  Trampolines[5 * I + 4] = 0x00000000;
637  }
638 }
639 
640 void OrcMips32_Base::writeIndirectStubsBlock(
641  char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress,
642  JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs) {
643  assert((StubsBlockTargetAddress >> 32) == 0 &&
644  "InitialPtrVal is out of range");
645 
646  // Stub format is:
647  //
648  // .section __orc_stubs
649  // stub1:
650  // lui $t9, ptr1
651  // lw $t9, %lo(ptr1)($t9)
652  // jr $t9
653  // stub2:
654  // lui $t9, ptr2
655  // lw $t9,%lo(ptr1)($t9)
656  // jr $t9
657  //
658  // ...
659  //
660  // .section __orc_ptrs
661  // ptr1:
662  // .word 0x0
663  // ptr2:
664  // .word 0x0
665  //
666  // i..
667 
668  assert(stubAndPointerRangesOk<OrcAArch64>(
669  StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
670  "PointersBlock is out of range");
671 
672  // Populate the stubs page stubs and mark it executable.
673  uint32_t *Stub = reinterpret_cast<uint32_t *>(StubsBlockWorkingMem);
674  uint64_t PtrAddr = PointersBlockTargetAddress;
675 
676  for (unsigned I = 0; I < NumStubs; ++I) {
677  uint32_t HiAddr = ((PtrAddr + 0x8000) >> 16);
678  Stub[4 * I + 0] = 0x3c190000 | (HiAddr & 0xFFFF); // lui $t9,ptr1
679  Stub[4 * I + 1] = 0x8f390000 | (PtrAddr & 0xFFFF); // lw $t9,%lo(ptr1)($t9)
680  Stub[4 * I + 2] = 0x03200008; // jr $t9
681  Stub[4 * I + 3] = 0x00000000; // nop
682  PtrAddr += 4;
683  }
684 }
685 
686 void OrcMips64::writeResolverCode(char *ResolverWorkingMem,
687  JITTargetAddress ResolverTargetAddress,
688  JITTargetAddress ReentryFnAddr,
689  JITTargetAddress ReentryCtxAddr) {
690 
691  const uint32_t ResolverCode[] = {
692  //resolver_entry:
693  0x67bdff30, // 0x00: daddiu $sp,$sp,-208
694  0xffa20000, // 0x04: sd v0,0(sp)
695  0xffa30008, // 0x08: sd v1,8(sp)
696  0xffa40010, // 0x0c: sd a0,16(sp)
697  0xffa50018, // 0x10: sd a1,24(sp)
698  0xffa60020, // 0x14: sd a2,32(sp)
699  0xffa70028, // 0x18: sd a3,40(sp)
700  0xffa80030, // 0x1c: sd a4,48(sp)
701  0xffa90038, // 0x20: sd a5,56(sp)
702  0xffaa0040, // 0x24: sd a6,64(sp)
703  0xffab0048, // 0x28: sd a7,72(sp)
704  0xffac0050, // 0x2c: sd t0,80(sp)
705  0xffad0058, // 0x30: sd t1,88(sp)
706  0xffae0060, // 0x34: sd t2,96(sp)
707  0xffaf0068, // 0x38: sd t3,104(sp)
708  0xffb00070, // 0x3c: sd s0,112(sp)
709  0xffb10078, // 0x40: sd s1,120(sp)
710  0xffb20080, // 0x44: sd s2,128(sp)
711  0xffb30088, // 0x48: sd s3,136(sp)
712  0xffb40090, // 0x4c: sd s4,144(sp)
713  0xffb50098, // 0x50: sd s5,152(sp)
714  0xffb600a0, // 0x54: sd s6,160(sp)
715  0xffb700a8, // 0x58: sd s7,168(sp)
716  0xffb800b0, // 0x5c: sd t8,176(sp)
717  0xffb900b8, // 0x60: sd t9,184(sp)
718  0xffbe00c0, // 0x64: sd fp,192(sp)
719  0xffbf00c8, // 0x68: sd ra,200(sp)
720 
721  // JIT re-entry ctx addr.
722  0x00000000, // 0x6c: lui $a0,heighest(ctx)
723  0x00000000, // 0x70: daddiu $a0,$a0,heigher(ctx)
724  0x00000000, // 0x74: dsll $a0,$a0,16
725  0x00000000, // 0x78: daddiu $a0,$a0,hi(ctx)
726  0x00000000, // 0x7c: dsll $a0,$a0,16
727  0x00000000, // 0x80: daddiu $a0,$a0,lo(ctx)
728 
729  0x03e02825, // 0x84: move $a1, $ra
730  0x64a5ffdc, // 0x88: daddiu $a1,$a1,-36
731 
732  // JIT re-entry fn addr:
733  0x00000000, // 0x8c: lui $t9,reentry
734  0x00000000, // 0x90: daddiu $t9,$t9,reentry
735  0x00000000, // 0x94: dsll $t9,$t9,
736  0x00000000, // 0x98: daddiu $t9,$t9,
737  0x00000000, // 0x9c: dsll $t9,$t9,
738  0x00000000, // 0xa0: daddiu $t9,$t9,
739  0x0320f809, // 0xa4: jalr $t9
740  0x00000000, // 0xa8: nop
741  0xdfbf00c8, // 0xac: ld ra, 200(sp)
742  0xdfbe00c0, // 0xb0: ld fp, 192(sp)
743  0xdfb900b8, // 0xb4: ld t9, 184(sp)
744  0xdfb800b0, // 0xb8: ld t8, 176(sp)
745  0xdfb700a8, // 0xbc: ld s7, 168(sp)
746  0xdfb600a0, // 0xc0: ld s6, 160(sp)
747  0xdfb50098, // 0xc4: ld s5, 152(sp)
748  0xdfb40090, // 0xc8: ld s4, 144(sp)
749  0xdfb30088, // 0xcc: ld s3, 136(sp)
750  0xdfb20080, // 0xd0: ld s2, 128(sp)
751  0xdfb10078, // 0xd4: ld s1, 120(sp)
752  0xdfb00070, // 0xd8: ld s0, 112(sp)
753  0xdfaf0068, // 0xdc: ld t3, 104(sp)
754  0xdfae0060, // 0xe0: ld t2, 96(sp)
755  0xdfad0058, // 0xe4: ld t1, 88(sp)
756  0xdfac0050, // 0xe8: ld t0, 80(sp)
757  0xdfab0048, // 0xec: ld a7, 72(sp)
758  0xdfaa0040, // 0xf0: ld a6, 64(sp)
759  0xdfa90038, // 0xf4: ld a5, 56(sp)
760  0xdfa80030, // 0xf8: ld a4, 48(sp)
761  0xdfa70028, // 0xfc: ld a3, 40(sp)
762  0xdfa60020, // 0x100: ld a2, 32(sp)
763  0xdfa50018, // 0x104: ld a1, 24(sp)
764  0xdfa40010, // 0x108: ld a0, 16(sp)
765  0xdfa30008, // 0x10c: ld v1, 8(sp)
766  0x67bd00d0, // 0x110: daddiu $sp,$sp,208
767  0x0300f825, // 0x114: move $ra, $t8
768  0x03200008, // 0x118: jr $t9
769  0x0040c825, // 0x11c: move $t9, $v0
770  };
771 
772  const unsigned ReentryFnAddrOffset = 0x8c; // JIT re-entry fn addr lui
773  const unsigned ReentryCtxAddrOffset = 0x6c; // JIT re-entry ctx addr lui
774 
775  memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
776 
777  uint32_t ReentryCtxLUi =
778  0x3c040000 | (((ReentryCtxAddr + 0x800080008000) >> 48) & 0xFFFF);
779  uint32_t ReentryCtxDADDiu =
780  0x64840000 | (((ReentryCtxAddr + 0x80008000) >> 32) & 0xFFFF);
781  uint32_t ReentryCtxDSLL = 0x00042438;
782  uint32_t ReentryCtxDADDiu2 =
783  0x64840000 | ((((ReentryCtxAddr + 0x8000) >> 16) & 0xFFFF));
784  uint32_t ReentryCtxDSLL2 = 0x00042438;
785  uint32_t ReentryCtxDADDiu3 = 0x64840000 | ((ReentryCtxAddr)&0xFFFF);
786 
787  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxLUi,
788  sizeof(ReentryCtxLUi));
789  memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 4), &ReentryCtxDADDiu,
790  sizeof(ReentryCtxDADDiu));
791  memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 8), &ReentryCtxDSLL,
792  sizeof(ReentryCtxDSLL));
793  memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 12), &ReentryCtxDADDiu2,
794  sizeof(ReentryCtxDADDiu2));
795  memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 16), &ReentryCtxDSLL2,
796  sizeof(ReentryCtxDSLL2));
797  memcpy(ResolverWorkingMem + (ReentryCtxAddrOffset + 20), &ReentryCtxDADDiu3,
798  sizeof(ReentryCtxDADDiu3));
799 
800  uint32_t ReentryFnLUi =
801  0x3c190000 | (((ReentryFnAddr + 0x800080008000) >> 48) & 0xFFFF);
802 
803  uint32_t ReentryFnDADDiu =
804  0x67390000 | (((ReentryFnAddr + 0x80008000) >> 32) & 0xFFFF);
805 
806  uint32_t ReentryFnDSLL = 0x0019cc38;
807 
808  uint32_t ReentryFnDADDiu2 =
809  0x67390000 | (((ReentryFnAddr + 0x8000) >> 16) & 0xFFFF);
810 
811  uint32_t ReentryFnDSLL2 = 0x0019cc38;
812 
813  uint32_t ReentryFnDADDiu3 = 0x67390000 | ((ReentryFnAddr)&0xFFFF);
814 
815  memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnLUi,
816  sizeof(ReentryFnLUi));
817  memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 4), &ReentryFnDADDiu,
818  sizeof(ReentryFnDADDiu));
819  memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 8), &ReentryFnDSLL,
820  sizeof(ReentryFnDSLL));
821  memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 12), &ReentryFnDADDiu2,
822  sizeof(ReentryFnDADDiu2));
823  memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 16), &ReentryFnDSLL2,
824  sizeof(ReentryFnDSLL2));
825  memcpy(ResolverWorkingMem + (ReentryFnAddrOffset + 20), &ReentryFnDADDiu3,
826  sizeof(ReentryFnDADDiu3));
827 }
828 
829 void OrcMips64::writeTrampolines(char *TrampolineBlockWorkingMem,
830  JITTargetAddress TrampolineBlockTargetAddress,
831  JITTargetAddress ResolverAddr,
832  unsigned NumTrampolines) {
833 
834  uint32_t *Trampolines =
835  reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
836 
837  uint64_t HeighestAddr = ((ResolverAddr + 0x800080008000) >> 48);
838  uint64_t HeigherAddr = ((ResolverAddr + 0x80008000) >> 32);
839  uint64_t HiAddr = ((ResolverAddr + 0x8000) >> 16);
840 
841  for (unsigned I = 0; I < NumTrampolines; ++I) {
842  Trampolines[10 * I + 0] = 0x03e0c025; // move $t8,$ra
843  Trampolines[10 * I + 1] = 0x3c190000 | (HeighestAddr & 0xFFFF); // lui $t9,resolveAddr
844  Trampolines[10 * I + 2] = 0x67390000 | (HeigherAddr & 0xFFFF); // daddiu $t9,$t9,%higher(resolveAddr)
845  Trampolines[10 * I + 3] = 0x0019cc38; // dsll $t9,$t9,16
846  Trampolines[10 * I + 4] = 0x67390000 | (HiAddr & 0xFFFF); // daddiu $t9,$t9,%hi(ptr)
847  Trampolines[10 * I + 5] = 0x0019cc38; // dsll $t9,$t9,16
848  Trampolines[10 * I + 6] =
849  0x67390000 | (ResolverAddr & 0xFFFF); // daddiu $t9,$t9,%lo(ptr)
850  Trampolines[10 * I + 7] = 0x0320f809; // jalr $t9
851  Trampolines[10 * I + 8] = 0x00000000; // nop
852  Trampolines[10 * I + 9] = 0x00000000; // nop
853  }
854 }
855 
856 void OrcMips64::writeIndirectStubsBlock(
857  char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress,
858  JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs) {
859  // Stub format is:
860  //
861  // .section __orc_stubs
862  // stub1:
863  // lui $t9,ptr1
864  // dsll $t9,$t9,16
865  // daddiu $t9,$t9,%hi(ptr)
866  // dsll $t9,$t9,16
867  // ld $t9,%lo(ptr)
868  // jr $t9
869  // stub2:
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  //
877  // ...
878  //
879  // .section __orc_ptrs
880  // ptr1:
881  // .dword 0x0
882  // ptr2:
883  // .dword 0x0
884  //
885  // ...
886 
887  assert(stubAndPointerRangesOk<OrcAArch64>(
888  StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
889  "PointersBlock is out of range");
890 
891  // Populate the stubs page stubs and mark it executable.
892  uint32_t *Stub = reinterpret_cast<uint32_t *>(StubsBlockWorkingMem);
893  uint64_t PtrAddr = PointersBlockTargetAddress;
894 
895  for (unsigned I = 0; I < NumStubs; ++I, PtrAddr += 8) {
896  uint64_t HeighestAddr = ((PtrAddr + 0x800080008000) >> 48);
897  uint64_t HeigherAddr = ((PtrAddr + 0x80008000) >> 32);
898  uint64_t HiAddr = ((PtrAddr + 0x8000) >> 16);
899  Stub[8 * I + 0] = 0x3c190000 | (HeighestAddr & 0xFFFF); // lui $t9,ptr1
900  Stub[8 * I + 1] = 0x67390000 | (HeigherAddr & 0xFFFF); // daddiu $t9,$t9,%higher(ptr)
901  Stub[8 * I + 2] = 0x0019cc38; // dsll $t9,$t9,16
902  Stub[8 * I + 3] = 0x67390000 | (HiAddr & 0xFFFF); // daddiu $t9,$t9,%hi(ptr)
903  Stub[8 * I + 4] = 0x0019cc38; // dsll $t9,$t9,16
904  Stub[8 * I + 5] = 0xdf390000 | (PtrAddr & 0xFFFF); // ld $t9,%lo(ptr)
905  Stub[8 * I + 6] = 0x03200008; // jr $t9
906  Stub[8 * I + 7] = 0x00000000; // nop
907  }
908 }
909 
910 void OrcRiscv64::writeResolverCode(char *ResolverWorkingMem,
911  JITTargetAddress ResolverTargetAddress,
912  JITTargetAddress ReentryFnAddr,
913  JITTargetAddress ReentryCtxAddr) {
914 
915  const uint32_t ResolverCode[] = {
916  0xef810113, // 0x00: addi sp,sp,-264
917  0x00813023, // 0x04: sd s0,0(sp)
918  0x00913423, // 0x08: sd s1,8(sp)
919  0x01213823, // 0x0c: sd s2,16(sp)
920  0x01313c23, // 0x10: sd s3,24(sp)
921  0x03413023, // 0x14: sd s4,32(sp)
922  0x03513423, // 0x18: sd s5,40(sp)
923  0x03613823, // 0x1c: sd s6,48(sp)
924  0x03713c23, // 0x20: sd s7,56(sp)
925  0x05813023, // 0x24: sd s8,64(sp)
926  0x05913423, // 0x28: sd s9,72(sp)
927  0x05a13823, // 0x2c: sd s10,80(sp)
928  0x05b13c23, // 0x30: sd s11,88(sp)
929  0x06113023, // 0x34: sd ra,96(sp)
930  0x06a13423, // 0x38: sd a0,104(sp)
931  0x06b13823, // 0x3c: sd a1,112(sp)
932  0x06c13c23, // 0x40: sd a2,120(sp)
933  0x08d13023, // 0x44: sd a3,128(sp)
934  0x08e13423, // 0x48: sd a4,136(sp)
935  0x08f13823, // 0x4c: sd a5,144(sp)
936  0x09013c23, // 0x50: sd a6,152(sp)
937  0x0b113023, // 0x54: sd a7,160(sp)
938  0x0a813427, // 0x58: fsd fs0,168(sp)
939  0x0a913827, // 0x5c: fsd fs1,176(sp)
940  0x0b213c27, // 0x60: fsd fs2,184(sp)
941  0x0d313027, // 0x64: fsd fs3,192(sp)
942  0x0d413427, // 0x68: fsd fs4,200(sp)
943  0x0d513827, // 0x6c: fsd fs5,208(sp)
944  0x0d613c27, // 0x70: fsd fs6,216(sp)
945  0x0f713027, // 0x74: fsd fs7,224(sp)
946  0x0f813427, // 0x78: fsd fs8,232(sp)
947  0x0f913827, // 0x7c: fsd fs9,240(sp)
948  0x0fa13c27, // 0x80: fsd fs10,248(sp)
949  0x11b13027, // 0x84: fsd fs11,256(sp)
950  0x00000517, // 0x88: auipc a0,0x0
951  0x0b053503, // 0x8c: ld a0,176(a0) # 0x138
952  0x00030593, // 0x90: mv a1,t1
953  0xff458593, // 0x94: addi a1,a1,-12
954  0x00000617, // 0x98: auipc a2,0x0
955  0x0a863603, // 0x9c: ld a2,168(a2) # 0x140
956  0x000600e7, // 0xa0: jalr a2
957  0x00050293, // 0xa4: mv t0,a0
958  0x00013403, // 0xa8: ld s0,0(sp)
959  0x00813483, // 0xac: ld s1,8(sp)
960  0x01013903, // 0xb0: ld s2,16(sp)
961  0x01813983, // 0xb4: ld s3,24(sp)
962  0x02013a03, // 0xb8: ld s4,32(sp)
963  0x02813a83, // 0xbc: ld s5,40(sp)
964  0x03013b03, // 0xc0: ld s6,48(sp)
965  0x03813b83, // 0xc4: ld s7,56(sp)
966  0x04013c03, // 0xc8: ld s8,64(sp)
967  0x04813c83, // 0xcc: ld s9,72(sp)
968  0x05013d03, // 0xd0: ld s10,80(sp)
969  0x05813d83, // 0xd4: ld s11,88(sp)
970  0x06013083, // 0xd8: ld ra,96(sp)
971  0x06813503, // 0xdc: ld a0,104(sp)
972  0x07013583, // 0xe0: ld a1,112(sp)
973  0x07813603, // 0xe4: ld a2,120(sp)
974  0x08013683, // 0xe8: ld a3,128(sp)
975  0x08813703, // 0xec: ld a4,136(sp)
976  0x09013783, // 0xf0: ld a5,144(sp)
977  0x09813803, // 0xf4: ld a6,152(sp)
978  0x0a013883, // 0xf8: ld a7,160(sp)
979  0x0a813407, // 0xfc: fld fs0,168(sp)
980  0x0b013487, // 0x100: fld fs1,176(sp)
981  0x0b813907, // 0x104: fld fs2,184(sp)
982  0x0c013987, // 0x108: fld fs3,192(sp)
983  0x0c813a07, // 0x10c: fld fs4,200(sp)
984  0x0d013a87, // 0x110: fld fs5,208(sp)
985  0x0d813b07, // 0x114: fld fs6,216(sp)
986  0x0e013b87, // 0x118: fld fs7,224(sp)
987  0x0e813c07, // 0x11c: fld fs8,232(sp)
988  0x0f013c87, // 0x120: fld fs9,240(sp)
989  0x0f813d07, // 0x124: fld fs10,248(sp)
990  0x10013d87, // 0x128: fld fs11,256(sp)
991  0x10810113, // 0x12c: addi sp,sp,264
992  0x00028067, // 0x130: jr t0
993  0x12345678, // 0x134: padding to align at 8 byte
994  0x12345678, // 0x138: Lreentry_ctx_ptr:
995  0xdeadbeef, // 0x13c: .quad 0
996  0x98765432, // 0x140: Lreentry_fn_ptr:
997  0xcafef00d // 0x144: .quad 0
998  };
999 
1000  const unsigned ReentryCtxAddrOffset = 0x138;
1001  const unsigned ReentryFnAddrOffset = 0x140;
1002 
1003  memcpy(ResolverWorkingMem, ResolverCode, sizeof(ResolverCode));
1004  memcpy(ResolverWorkingMem + ReentryFnAddrOffset, &ReentryFnAddr,
1005  sizeof(uint64_t));
1006  memcpy(ResolverWorkingMem + ReentryCtxAddrOffset, &ReentryCtxAddr,
1007  sizeof(uint64_t));
1008 }
1009 
1010 void OrcRiscv64::writeTrampolines(char *TrampolineBlockWorkingMem,
1011  JITTargetAddress TrampolineBlockTargetAddress,
1012  JITTargetAddress ResolverAddr,
1013  unsigned NumTrampolines) {
1014 
1015  unsigned OffsetToPtr = alignTo(NumTrampolines * TrampolineSize, 8);
1016 
1017  memcpy(TrampolineBlockWorkingMem + OffsetToPtr, &ResolverAddr,
1018  sizeof(uint64_t));
1019 
1020  uint32_t *Trampolines =
1021  reinterpret_cast<uint32_t *>(TrampolineBlockWorkingMem);
1022  for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize) {
1023  uint32_t Hi20 = (OffsetToPtr + 0x800) & 0xFFFFF000;
1024  uint32_t Lo12 = OffsetToPtr - Hi20;
1025  Trampolines[4 * I + 0] = 0x00000297 | Hi20; // auipc t0, %hi(Lptr)
1026  Trampolines[4 * I + 1] =
1027  0x0002b283 | ((Lo12 & 0xFFF) << 20); // ld t0, %lo(Lptr)
1028  Trampolines[4 * I + 2] = 0x00028367; // jalr t1, t0
1029  Trampolines[4 * I + 3] = 0xdeadface; // padding
1030  }
1031 }
1032 
1033 void OrcRiscv64::writeIndirectStubsBlock(
1034  char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress,
1035  JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs) {
1036  // Stub format is:
1037  //
1038  // .section __orc_stubs
1039  // stub1:
1040  // auipc t0, %hi(ptr1) ; PC-rel load of ptr1
1041  // ld t0, %lo(t0)
1042  // jr t0 ; Jump to resolver
1043  // .quad 0 ; Pad to 16 bytes
1044  // stub2:
1045  // auipc t0, %hi(ptr1) ; PC-rel load of ptr1
1046  // ld t0, %lo(t0)
1047  // jr t0 ; Jump to resolver
1048  // .quad 0
1049  //
1050  // ...
1051  //
1052  // .section __orc_ptrs
1053  // ptr1:
1054  // .quad 0x0
1055  // ptr2:
1056  // .quad 0x0
1057  //
1058  // ...
1059 
1060  assert(stubAndPointerRangesOk<OrcRiscv64>(
1061  StubsBlockTargetAddress, PointersBlockTargetAddress, NumStubs) &&
1062  "PointersBlock is out of range");
1063 
1064  uint32_t *Stub = reinterpret_cast<uint32_t *>(StubsBlockWorkingMem);
1065 
1066  for (unsigned I = 0; I < NumStubs; ++I) {
1067  uint64_t PtrDisplacement =
1068  PointersBlockTargetAddress - StubsBlockTargetAddress;
1069  uint32_t Hi20 = (PtrDisplacement + 0x800) & 0xFFFFF000;
1070  uint32_t Lo12 = PtrDisplacement - Hi20;
1071  Stub[4 * I + 0] = 0x00000297 | Hi20; // auipc t0, %hi(Lptr)
1072  Stub[4 * I + 1] = 0x0002b283 | ((Lo12 & 0xFFF) << 20); // ld t0, %lo(Lptr)
1073  Stub[4 * I + 2] = 0x00028067; // jr t0
1074  Stub[4 * I + 3] = 0xfeedbeef; // padding
1075  PointersBlockTargetAddress += PointerSize;
1076  StubsBlockTargetAddress += StubSize;
1077  }
1078 }
1079 
1080 } // End namespace orc.
1081 } // End namespace llvm.
llvm::alignTo
uint64_t alignTo(uint64_t Size, Align A)
Returns a multiple of A needed to store Size bytes.
Definition: Alignment.h:156
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:17
llvm::orc::OrcMips64::writeResolverCode
static void writeResolverCode(char *ResolverWorkingMem, JITTargetAddress ResolverTargetAddress, JITTargetAddress ReentryFnAddr, JITTargetAddress ReentryCtxAddr)
Write the resolver code into the given memory.
Definition: OrcABISupport.cpp:686
llvm::orc::OrcI386::TrampolineSize
static constexpr unsigned TrampolineSize
Definition: OrcABISupport.h:198
llvm::orc::OrcX86_64_Base::writeTrampolines
static void writeTrampolines(char *TrampolineBlockWorkingMem, JITTargetAddress TrampolineBlockTargetAddress, JITTargetAddress ResolverAddr, unsigned NumTrampolines)
Write the requested number of trampolines into the given memory, which must be big enough to hold 1 p...
Definition: OrcABISupport.cpp:198
llvm::orc::OrcX86_64_Base::PointerSize
static constexpr unsigned PointerSize
Definition: OrcABISupport.h:130
llvm::orc::OrcX86_64_Win32::writeResolverCode
static void writeResolverCode(char *ResolverWorkingMem, JITTargetAddress ResolverTargetAddress, JITTargetAddress ReentryFnAddr, JITTargetAddress ReentryCtxAddr)
Write the resolver code into the given memory.
Definition: OrcABISupport.cpp:326
llvm::orc::OrcMips64::writeIndirectStubsBlock
static void writeIndirectStubsBlock(char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress, JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs)
Write NumStubs indirect stubs to working memory at StubsBlockWorkingMem.
Definition: OrcABISupport.cpp:856
llvm::orc::OrcAArch64::writeTrampolines
static void writeTrampolines(char *TrampolineBlockWorkingMem, JITTargetAddress TrampolineBlockTargetAddress, JITTargetAddress ResolverAddr, unsigned NumTrampolines)
Write the requested number of trampolines into the given memory, which must be big enough to hold 1 p...
Definition: OrcABISupport.cpp:137
isBigEndian
static Optional< bool > isBigEndian(const SmallDenseMap< int64_t, int64_t, 8 > &MemOffset2Idx, int64_t LowestIdx)
Given a map from byte offsets in memory to indices in a load/store, determine if that map corresponds...
Definition: CombinerHelper.cpp:108
stubAndPointerRangesOk
bool stubAndPointerRangesOk(JITTargetAddress StubBlockAddr, JITTargetAddress PointerBlockAddr, unsigned NumStubs)
Definition: OrcABISupport.cpp:19
llvm::orc::OrcMips32_Base::writeTrampolines
static void writeTrampolines(char *TrampolineBlockWorkingMem, JITTargetAddress TrampolineBlockTargetAddress, JITTargetAddress ResolverAddr, unsigned NumTrampolines)
Write the requested number of trampolines into the given memory, which must be big enough to hold 1 p...
Definition: OrcABISupport.cpp:615
OrcABISupport.h
LLVM_DEBUG
#define LLVM_DEBUG(X)
Definition: Debug.h:101
llvm::orc::OrcAArch64::writeIndirectStubsBlock
static void writeIndirectStubsBlock(char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress, JITTargetAddress PointersBlockTargetAddress, unsigned MinStubs)
Write NumStubs indirect stubs to working memory at StubsBlockWorkingMem.
Definition: OrcABISupport.cpp:161
llvm::dbgs
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
llvm::formatv
auto formatv(const char *Fmt, Ts &&... Vals) -> formatv_object< decltype(std::make_tuple(detail::build_format_adapter(std::forward< Ts >(Vals))...))>
Definition: FormatVariadic.h:251
Process.h
llvm::orc::OrcI386::writeTrampolines
static void writeTrampolines(char *TrampolineBlockWorkingMem, JITTargetAddress TrampolineBlockTargetAddress, JITTargetAddress ResolverAddr, unsigned NumTrampolines)
Write the requested number of trampolines into the given memory, which must be big enough to hold 1 p...
Definition: OrcABISupport.cpp:457
llvm::orc::OrcRiscv64::writeResolverCode
static void writeResolverCode(char *ResolverWorkingMem, JITTargetAddress ResolverTargetAddress, JITTargetAddress ReentryFnAddr, JITTargetAddress ReentryCtxAddr)
Write the resolver code into the given memory.
Definition: OrcABISupport.cpp:910
llvm::orc::OrcRiscv64::StubSize
static constexpr unsigned StubSize
Definition: OrcABISupport.h:340
FormatVariadic.h
llvm::orc::OrcAArch64::PointerSize
static constexpr unsigned PointerSize
Definition: OrcABISupport.h:90
llvm::orc::OrcAArch64::writeResolverCode
static void writeResolverCode(char *ResolverWorkingMem, JITTargetAddress ResolverTargetAddress, JITTargetAddress ReentryFnAddr, JITTargetAddress RentryCtxAddr)
Write the resolver code into the given memory.
Definition: OrcABISupport.cpp:46
llvm::orc::OrcMips32_Base::writeIndirectStubsBlock
static void writeIndirectStubsBlock(char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress, JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs)
Write NumStubs indirect stubs to working memory at StubsBlockWorkingMem.
Definition: OrcABISupport.cpp:640
llvm::orc::OrcRiscv64::PointerSize
static constexpr unsigned PointerSize
Definition: OrcABISupport.h:338
uint64_t
I
#define I(x, y, z)
Definition: MD5.cpp:58
llvm::orc::OrcRiscv64::TrampolineSize
static constexpr unsigned TrampolineSize
Definition: OrcABISupport.h:339
assert
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
memcpy
<%struct.s * > cast struct s *S to sbyte *< sbyte * > sbyte uint cast struct s *agg result to sbyte *< sbyte * > sbyte uint cast struct s *memtmp to sbyte *< sbyte * > sbyte uint ret void llc ends up issuing two memcpy or custom lower memcpy(of small size) to be ldmia/stmia. I think option 2 is better but the current register allocator cannot allocate a chunk of registers at a time. A feasible temporary solution is to use specific physical registers at the lowering time for small(<
llvm::orc::OrcAArch64::TrampolineSize
static constexpr unsigned TrampolineSize
Definition: OrcABISupport.h:91
uint32_t
llvm::orc::OrcI386::writeIndirectStubsBlock
static void writeIndirectStubsBlock(char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress, JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs)
Write NumStubs indirect stubs to working memory at StubsBlockWorkingMem.
Definition: OrcABISupport.cpp:471
llvm::orc::OrcI386::writeResolverCode
static void writeResolverCode(char *ResolverWorkingMem, JITTargetAddress ResolverTargetAddress, JITTargetAddress ReentryFnAddr, JITTargetAddress ReentryCtxAddr)
Write the resolver code into the given memory.
Definition: OrcABISupport.cpp:404
llvm::orc::OrcMips64::writeTrampolines
static void writeTrampolines(char *TrampolineBlockWorkingMem, JITTargetAddress TrampolineBlockTargetAddress, JITTargetAddress ResolverFnAddr, unsigned NumTrampolines)
Write the requested number of trampolines into the given memory, which must be big enough to hold 1 p...
Definition: OrcABISupport.cpp:829
llvm::orc::OrcX86_64_Base::StubSize
static constexpr unsigned StubSize
Definition: OrcABISupport.h:132
llvm::orc::OrcX86_64_Base::TrampolineSize
static constexpr unsigned TrampolineSize
Definition: OrcABISupport.h:131
llvm::orc::OrcX86_64_Base::writeIndirectStubsBlock
static void writeIndirectStubsBlock(char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress, JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs)
Write NumStubs indirect stubs to working memory at StubsBlockWorkingMem.
Definition: OrcABISupport.cpp:216
llvm::orc::OrcRiscv64::writeIndirectStubsBlock
static void writeIndirectStubsBlock(char *StubsBlockWorkingMem, JITTargetAddress StubsBlockTargetAddress, JITTargetAddress PointersBlockTargetAddress, unsigned NumStubs)
Write NumStubs indirect stubs to working memory at StubsBlockWorkingMem.
Definition: OrcABISupport.cpp:1033
llvm::orc::OrcAArch64::StubSize
static constexpr unsigned StubSize
Definition: OrcABISupport.h:92
raw_ostream.h
llvm::orc::OrcMips32_Base::writeResolverCode
static void writeResolverCode(char *ResolverBlockWorkingMem, JITTargetAddress ResolverBlockTargetAddress, JITTargetAddress ReentryFnAddr, JITTargetAddress ReentryCtxAddr, bool isBigEndian)
Write the resolver code into the given memory.
Definition: OrcABISupport.cpp:509
llvm::orc::OrcRiscv64::writeTrampolines
static void writeTrampolines(char *TrampolineBlockWorkingMem, JITTargetAddress TrampolineBlockTargetAddress, JITTargetAddress ResolverFnAddr, unsigned NumTrampolines)
Write the requested number of trampolines into the given memory, which must be big enough to hold 1 p...
Definition: OrcABISupport.cpp:1010
llvm::orc::OrcX86_64_SysV::writeResolverCode
static void writeResolverCode(char *ResolverWorkingMem, JITTargetAddress ResolverTargetAddress, JITTargetAddress ReentryFnAddr, JITTargetAddress ReentryCtxAddr)
Write the resolver code into the given memory.
Definition: OrcABISupport.cpp:252
Generated on Thu Jun 30 2022 05:57:46 for LLVM by   doxygen 1.8.17