Taints the non-acquire RMW's store address with the load part
[oota-llvm.git] / lib / ExecutionEngine / Orc / OrcArchitectureSupport.cpp
1 //===------ OrcArchSupport.cpp - Architecture specific support code -------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9
10 #include "llvm/ADT/Triple.h"
11 #include "llvm/ExecutionEngine/Orc/OrcArchitectureSupport.h"
12 #include "llvm/Support/Process.h"
13 #include <array>
14
15 namespace llvm {
16 namespace orc {
17
18 void OrcX86_64::writeResolverCode(uint8_t *ResolverMem, JITReentryFn ReentryFn,
19                                   void *CallbackMgr) {
20
21   const uint8_t ResolverCode[] = {
22                                                // resolver_entry:
23     0x55,                                      // 0x00: pushq     %rbp
24     0x48, 0x89, 0xe5,                          // 0x01: movq      %rsp, %rbp
25     0x50,                                      // 0x04: pushq     %rax
26     0x53,                                      // 0x05: pushq     %rbx
27     0x51,                                      // 0x06: pushq     %rcx
28     0x52,                                      // 0x07: pushq     %rdx
29     0x56,                                      // 0x08: pushq     %rsi
30     0x57,                                      // 0x09: pushq     %rdi
31     0x41, 0x50,                                // 0x0a: pushq     %r8
32     0x41, 0x51,                                // 0x0c: pushq     %r9
33     0x41, 0x52,                                // 0x0e: pushq     %r10
34     0x41, 0x53,                                // 0x10: pushq     %r11
35     0x41, 0x54,                                // 0x12: pushq     %r12
36     0x41, 0x55,                                // 0x14: pushq     %r13
37     0x41, 0x56,                                // 0x16: pushq     %r14
38     0x41, 0x57,                                // 0x18: pushq     %r15
39     0x48, 0x81, 0xec, 0x08, 0x02, 0x00, 0x00,  // 0x1a: subq      20, %rsp
40     0x48, 0x0f, 0xae, 0x04, 0x24,              // 0x21: fxsave64  (%rsp)
41     0x48, 0x8d, 0x3d, 0x43, 0x00, 0x00, 0x00,  // 0x26: leaq      67(%rip), %rdi
42     0x48, 0x8b, 0x3f,                          // 0x2d: movq      (%rdi), %rdi
43     0x48, 0x8b, 0x75, 0x08,                    // 0x30: movq      8(%rbp), %rsi
44     0x48, 0x83, 0xee, 0x06,                    // 0x34: subq      $6, %rsi
45     0x48, 0xb8,                                // 0x38: movabsq   $0, %rax
46
47     // 0x3a: JIT re-entry fn addr:
48     0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
49
50     0xff, 0xd0,                                // 0x42: callq     *%rax
51     0x48, 0x89, 0x45, 0x08,                    // 0x44: movq      %rax, 8(%rbp)
52     0x48, 0x0f, 0xae, 0x0c, 0x24,              // 0x48: fxrstor64 (%rsp)
53     0x48, 0x81, 0xc4, 0x08, 0x02, 0x00, 0x00,  // 0x4d: addq      20, %rsp
54     0x41, 0x5f,                                // 0x54: popq      %r15
55     0x41, 0x5e,                                // 0x56: popq      %r14
56     0x41, 0x5d,                                // 0x58: popq      %r13
57     0x41, 0x5c,                                // 0x5a: popq      %r12
58     0x41, 0x5b,                                // 0x5c: popq      %r11
59     0x41, 0x5a,                                // 0x5e: popq      %r10
60     0x41, 0x59,                                // 0x60: popq      %r9
61     0x41, 0x58,                                // 0x62: popq      %r8
62     0x5f,                                      // 0x64: popq      %rdi
63     0x5e,                                      // 0x65: popq      %rsi
64     0x5a,                                      // 0x66: popq      %rdx
65     0x59,                                      // 0x67: popq      %rcx
66     0x5b,                                      // 0x68: popq      %rbx
67     0x58,                                      // 0x69: popq      %rax
68     0x5d,                                      // 0x6a: popq      %rbp
69     0xc3,                                      // 0x6b: retq
70     0x00, 0x00, 0x00, 0x00,                    // 0x6c: <padding>
71
72     // 0x70: Callback mgr address.
73     0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
74   };
75
76   const unsigned ReentryFnAddrOffset = 0x3a;
77   const unsigned CallbackMgrAddrOffset = 0x70;
78   
79   memcpy(ResolverMem, ResolverCode, sizeof(ResolverCode));
80   memcpy(ResolverMem + ReentryFnAddrOffset, &ReentryFn, sizeof(ReentryFn));
81   memcpy(ResolverMem + CallbackMgrAddrOffset, &CallbackMgr,
82          sizeof(CallbackMgr));
83 }
84
85 void OrcX86_64::writeTrampolines(uint8_t *TrampolineMem, void *ResolverAddr,
86                                  unsigned NumTrampolines) {
87
88   unsigned OffsetToPtr = NumTrampolines * TrampolineSize;
89
90   memcpy(TrampolineMem + OffsetToPtr, &ResolverAddr, sizeof(void*));
91
92   uint64_t *Trampolines = reinterpret_cast<uint64_t*>(TrampolineMem);
93   uint64_t CallIndirPCRel = 0xf1c40000000015ff;
94
95   for (unsigned I = 0; I < NumTrampolines; ++I, OffsetToPtr -= TrampolineSize)
96     Trampolines[I] = CallIndirPCRel | ((OffsetToPtr - 6) << 16);
97 }
98
99 std::error_code OrcX86_64::emitIndirectStubsBlock(IndirectStubsInfo &StubsInfo,
100                                                   unsigned MinStubs,
101                                                   void *InitialPtrVal) {
102   // Stub format is:
103   //
104   // .section __orc_stubs
105   // stub1:
106   //                 jmpq    *ptr1(%rip)
107   //                 .byte   0xC4         ; <- Invalid opcode padding.
108   //                 .byte   0xF1
109   // stub2:
110   //                 jmpq    *ptr2(%rip)
111   //
112   // ...
113   //
114   // .section __orc_ptrs
115   // ptr1:
116   //                 .quad 0x0
117   // ptr2:
118   //                 .quad 0x0
119   //
120   // ...
121
122   const unsigned StubSize = IndirectStubsInfo::StubSize;
123
124   // Emit at least MinStubs, rounded up to fill the pages allocated.
125   unsigned PageSize = sys::Process::getPageSize();
126   unsigned NumPages = ((MinStubs * StubSize) + (PageSize - 1)) / PageSize;
127   unsigned NumStubs = (NumPages * PageSize) / StubSize;
128
129   // Allocate memory for stubs and pointers in one call.
130   std::error_code EC;
131   auto StubsMem =
132     sys::OwningMemoryBlock(
133       sys::Memory::allocateMappedMemory(2 * NumPages * PageSize, nullptr,
134                                         sys::Memory::MF_READ |
135                                         sys::Memory::MF_WRITE,
136                                         EC));
137
138   if (EC)
139     return EC;
140
141   // Create separate MemoryBlocks representing the stubs and pointers.
142   sys::MemoryBlock StubsBlock(StubsMem.base(), NumPages * PageSize);
143   sys::MemoryBlock PtrsBlock(static_cast<char*>(StubsMem.base()) +
144                                NumPages * PageSize,
145                              NumPages * PageSize);
146
147   // Populate the stubs page stubs and mark it executable.
148   uint64_t *Stub = reinterpret_cast<uint64_t*>(StubsBlock.base());
149   uint64_t PtrOffsetField =
150     static_cast<uint64_t>(NumPages * PageSize - 6) << 16;
151   for (unsigned I = 0; I < NumStubs; ++I)
152     Stub[I] = 0xF1C40000000025ff | PtrOffsetField;
153
154   if (auto EC = sys::Memory::protectMappedMemory(StubsBlock,
155                                                  sys::Memory::MF_READ |
156                                                  sys::Memory::MF_EXEC))
157     return EC;
158
159   // Initialize all pointers to point at FailureAddress.
160   void **Ptr = reinterpret_cast<void**>(PtrsBlock.base());
161   for (unsigned I = 0; I < NumStubs; ++I)
162     Ptr[I] = InitialPtrVal;
163
164   StubsInfo.NumStubs = NumStubs;
165   StubsInfo.StubsMem = std::move(StubsMem);
166
167   return std::error_code();
168 }
169
170 } // End namespace orc.
171 } // End namespace llvm.