lib/ExecutionEngine/Orc/OrcTargetSupport.cpp

   1 #include "llvm/ADT/Triple.h"
   2 #include "llvm/ExecutionEngine/Orc/OrcTargetSupport.h"
   3 #include "llvm/Support/Process.h"
   4 #include <array>
   5
   6
   7 using namespace llvm::orc;
   8
   9 namespace {
  10
  11 uint64_t executeCompileCallback(JITCompileCallbackManagerBase *JCBM,
  12                                 TargetAddress CallbackID) {
  13   return JCBM->executeCompileCallback(CallbackID);
  14 }
  15
  16 }
  17
  18 namespace llvm {
  19 namespace orc {
  20
  21 const char* OrcX86_64::ResolverBlockName = "orc_resolver_block";
  22
  23 void OrcX86_64::insertResolverBlock(
  24     Module &M, JITCompileCallbackManagerBase &JCBM) {
  25
  26   // Trampoline code-sequence length, used to get trampoline address from return
  27   // address.
  28   const unsigned X86_64_TrampolineLength = 6;
  29
  30   // List of x86-64 GPRs to save. Note - RBP saved separately below.
  31   std::array<const char *, 14> GPRs = {{
  32       "rax", "rbx", "rcx", "rdx",
  33       "rsi", "rdi", "r8", "r9",
  34       "r10", "r11", "r12", "r13",
  35       "r14", "r15"
  36     }};
  37
  38   // Address of the executeCompileCallback function.
  39   uint64_t CallbackAddr =
  40       static_cast<uint64_t>(
  41         reinterpret_cast<uintptr_t>(executeCompileCallback));
  42
  43   std::ostringstream AsmStream;
  44   Triple TT(M.getTargetTriple());
  45
  46   // Switch to text section.
  47   if (TT.getOS() == Triple::Darwin)
  48     AsmStream << ".section __TEXT,__text,regular,pure_instructions\n"
  49               << ".align 4, 0x90\n";
  50   else
  51     AsmStream << ".text\n"
  52               << ".align 16, 0x90\n";
  53
  54   // Bake in a pointer to the callback manager immediately before the
  55   // start of the resolver function.
  56   AsmStream << "jit_callback_manager_addr:\n"
  57             << "  .quad " << &JCBM << "\n";
  58
  59   // Start the resolver function.
  60   AsmStream << ResolverBlockName << ":\n"
  61             << "  pushq     %rbp\n"
  62             << "  movq      %rsp, %rbp\n";
  63
  64   // Store the GPRs.
  65   for (const auto &GPR : GPRs)
  66     AsmStream << "  pushq     %" << GPR << "\n";
  67
  68   // Store floating-point state with FXSAVE.
  69   // Note: We need to keep the stack 16-byte aligned, so if we've emitted an odd
  70   //       number of 64-bit pushes so far (GPRs.size() plus 1 for RBP) then add
  71   //       an extra 64 bits of padding to the FXSave area.
  72   unsigned Padding = (GPRs.size() + 1) % 2 ? 8 : 0;
  73   unsigned FXSaveSize = 512 + Padding;
  74   AsmStream << "  subq      $" << FXSaveSize << ", %rsp\n"
  75             << "  fxsave64  (%rsp)\n"
  76
  77   // Load callback manager address, compute trampoline address, call JIT.
  78             << "  lea       jit_callback_manager_addr(%rip), %rdi\n"
  79             << "  movq      (%rdi), %rdi\n"
  80             << "  movq      0x8(%rbp), %rsi\n"
  81             << "  subq      $" << X86_64_TrampolineLength << ", %rsi\n"
  82             << "  movabsq   $" << CallbackAddr << ", %rax\n"
  83             << "  callq     *%rax\n"
  84
  85   // Replace the return to the trampoline with the return address of the
  86   // compiled function body.
  87             << "  movq      %rax, 0x8(%rbp)\n"
  88
  89   // Restore the floating point state.
  90             << "  fxrstor64 (%rsp)\n"
  91             << "  addq      $" << FXSaveSize << ", %rsp\n";
  92
  93   for (const auto &GPR : make_range(GPRs.rbegin(), GPRs.rend()))
  94     AsmStream << "  popq      %" << GPR << "\n";
  95
  96   // Restore original RBP and return to compiled function body.
  97   AsmStream << "  popq      %rbp\n"
  98             << "  retq\n";
  99
 100   M.appendModuleInlineAsm(AsmStream.str());
 101 }
 102
 103 OrcX86_64::LabelNameFtor
 104 OrcX86_64::insertCompileCallbackTrampolines(Module &M,
 105                                             TargetAddress ResolverBlockAddr,
 106                                             unsigned NumCalls,
 107                                             unsigned StartIndex) {
 108   const char *ResolverBlockPtrName = "Lorc_resolve_block_addr";
 109
 110   std::ostringstream AsmStream;
 111   Triple TT(M.getTargetTriple());
 112
 113   if (TT.getOS() == Triple::Darwin)
 114     AsmStream << ".section __TEXT,__text,regular,pure_instructions\n"
 115               << ".align 4, 0x90\n";
 116   else
 117     AsmStream << ".text\n"
 118               << ".align 16, 0x90\n";
 119
 120   AsmStream << ResolverBlockPtrName << ":\n"
 121             << "  .quad " << ResolverBlockAddr << "\n";
 122
 123   auto GetLabelName =
 124     [=](unsigned I) {
 125       std::ostringstream LabelStream;
 126       LabelStream << "orc_jcc_" << (StartIndex + I);
 127       return LabelStream.str();
 128   };
 129
 130   for (unsigned I = 0; I < NumCalls; ++I)
 131     AsmStream << GetLabelName(I) << ":\n"
 132               << "  callq *" << ResolverBlockPtrName << "(%rip)\n";
 133
 134   M.appendModuleInlineAsm(AsmStream.str());
 135
 136   return GetLabelName;
 137 }
 138
 139 OrcX86_64::IndirectStubsInfo::~IndirectStubsInfo() {
 140   sys::Memory::releaseMappedMemory(StubsBlock);
 141   sys::Memory::releaseMappedMemory(PtrsBlock);
 142 }
 143
 144 std::error_code OrcX86_64::emitIndirectStubsBlock(IndirectStubsInfo &StubsInfo,
 145                                                   unsigned MinStubs,
 146                                                   void *InitialPtrVal) {
 147   // Stub format is:
 148   //
 149   // .section __orc_stubs
 150   // stub1:
 151   //                 jmpq    *ptr1(%rip)
 152   //                 .byte   0xC4         ; <- Invalid opcode padding.
 153   //                 .byte   0xF1
 154   // stub2:
 155   //                 jmpq    *ptr2(%rip)
 156   //
 157   // ...
 158   //
 159   // .section __orc_ptrs
 160   // ptr1:
 161   //                 .quad 0x0
 162   // ptr2:
 163   //                 .quad 0x0
 164   //
 165   // ...
 166
 167   const unsigned StubSize = IndirectStubsInfo::StubSize;
 168
 169   // Emit at least MinStubs, rounded up to fill the pages allocated.
 170   unsigned PageSize = sys::Process::getPageSize();
 171   unsigned NumPages = ((MinStubs * StubSize) + (PageSize - 1)) / PageSize;
 172   unsigned NumStubs = (NumPages * PageSize) / StubSize;
 173
 174   // Allocate memory for stubs and pointers in one call.
 175   std::error_code EC;
 176   auto InitialBlock = sys::Memory::allocateMappedMemory(2 * NumPages * PageSize,
 177                                                         nullptr,
 178                                                         sys::Memory::MF_READ |
 179                                                         sys::Memory::MF_WRITE,
 180                                                         EC);
 181
 182   if (EC)
 183     return EC;
 184
 185   // Create separate MemoryBlocks representing the stubs and pointers.
 186   sys::MemoryBlock StubsBlock(InitialBlock.base(), NumPages * PageSize);
 187   sys::MemoryBlock PtrsBlock(static_cast<char*>(InitialBlock.base()) +
 188                              NumPages * PageSize,
 189                              NumPages * PageSize);
 190
 191   // Populate the stubs page stubs and mark it executable.
 192   uint64_t *Stub = reinterpret_cast<uint64_t*>(StubsBlock.base());
 193   uint64_t PtrOffsetField =
 194     static_cast<uint64_t>(NumPages * PageSize - 6) << 16;
 195   for (unsigned I = 0; I < NumStubs; ++I)
 196     Stub[I] = 0xF1C40000000025ff | PtrOffsetField;
 197
 198   if (auto EC = sys::Memory::protectMappedMemory(StubsBlock,
 199                                                  sys::Memory::MF_READ |
 200                                                  sys::Memory::MF_EXEC))
 201     return EC;
 202
 203   // Initialize all pointers to point at FailureAddress.
 204   void **Ptr = reinterpret_cast<void**>(PtrsBlock.base());
 205   for (unsigned I = 0; I < NumStubs; ++I)
 206     Ptr[I] = InitialPtrVal;
 207
 208   StubsInfo.NumStubs = NumStubs;
 209   StubsInfo.StubsBlock = std::move(StubsBlock);
 210   StubsInfo.PtrsBlock = std::move(PtrsBlock);
 211
 212   return std::error_code();
 213 }
 214
 215 } // End namespace orc.
 216 } // End namespace llvm.