1 //===-- X86JITInfo.cpp - Implement the JIT interfaces for the X86 target --===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the JIT interfaces for the X86 target.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "jit"
15 #include "X86JITInfo.h"
16 #include "X86Relocations.h"
17 #include "llvm/CodeGen/MachineCodeEmitter.h"
18 #include "llvm/Config/alloca.h"
24 extern "C" void *_AddressOfReturnAddress(void);
25 #pragma intrinsic(_AddressOfReturnAddress)
28 void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) {
29 unsigned char *OldByte = (unsigned char *)Old;
30 *OldByte++ = 0xE9; // Emit JMP opcode.
31 unsigned *OldWord = (unsigned *)OldByte;
32 unsigned NewAddr = (intptr_t)New;
33 unsigned OldAddr = (intptr_t)OldWord;
34 *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code.
38 /// JITCompilerFunction - This contains the address of the JIT function used to
39 /// compile a function lazily.
40 static TargetJITInfo::JITCompilerFn JITCompilerFunction;
42 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
43 // callee saved registers, for the fastcc calling convention.
45 #if defined(__i386__) || defined(i386) || defined(_M_IX86)
47 void X86CompilationCallback(void);
51 #if defined(__CYGWIN__) || defined(__APPLE__) || defined(__MINGW32__)
52 ".globl _X86CompilationCallback\n"
53 "_X86CompilationCallback:\n"
55 ".globl X86CompilationCallback\n"
56 "X86CompilationCallback:\n"
59 "movl %esp, %ebp\n" // Standard prologue
60 #if FASTCC_NUM_INT_ARGS_INREGS > 0
62 "pushl %edx\n" // Save EAX/EDX
64 #if defined(__APPLE__)
65 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
67 #if defined(__CYGWIN__) || defined(__MINGW32__) || defined(__APPLE__)
68 "call _X86CompilationCallback2\n"
70 "call X86CompilationCallback2\n"
72 #if defined(__APPLE__)
73 "movl %ebp, %esp\n" // Restore ESP
75 #if FASTCC_NUM_INT_ARGS_INREGS > 0
76 #if defined(__APPLE__)
85 void X86CompilationCallback2(void);
87 _declspec(naked) void X86CompilationCallback(void) {
91 call X86CompilationCallback2
99 #else // Not an i386 host
100 void X86CompilationCallback() {
101 std::cerr << "Cannot call X86CompilationCallback() on a non-x86 arch!\n";
107 /// X86CompilationCallback - This is the target-specific function invoked by the
108 /// function stub when we did not know the real target of a call. This function
109 /// must locate the start of the stub or call site and pass it into the JIT
110 /// compiler function.
111 extern "C" void X86CompilationCallback2() {
113 assert(sizeof(size_t) == 4); // FIXME: handle Win64
114 unsigned *RetAddrLoc = (unsigned *)_AddressOfReturnAddress();
115 RetAddrLoc += 3; // skip over ret addr, edx, eax
116 unsigned RetAddr = *RetAddrLoc;
118 unsigned *StackPtr = (unsigned*)__builtin_frame_address(1);
119 unsigned RetAddr = (unsigned)(intptr_t)__builtin_return_address(1);
120 unsigned *RetAddrLoc = &StackPtr[1];
122 // NOTE: __builtin_frame_address doesn't work if frame pointer elimination has
123 // been performed. Having a variable sized alloca disables frame pointer
124 // elimination currently, even if it's dead. This is a gross hack.
125 alloca(10+(RetAddr >> 31));
128 assert(*RetAddrLoc == RetAddr &&
129 "Could not find return address on the stack!");
131 // It's a stub if there is an interrupt marker after the call.
132 bool isStub = ((unsigned char*)(intptr_t)RetAddr)[0] == 0xCD;
134 // The call instruction should have pushed the return value onto the stack...
135 RetAddr -= 4; // Backtrack to the reference itself...
138 DEBUG(std::cerr << "In callback! Addr=" << (void*)RetAddr
139 << " ESP=" << (void*)StackPtr
140 << ": Resolving call to function: "
141 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n");
144 // Sanity check to make sure this really is a call instruction.
145 assert(((unsigned char*)(intptr_t)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
147 unsigned NewVal = (intptr_t)JITCompilerFunction((void*)(intptr_t)RetAddr);
149 // Rewrite the call target... so that we don't end up here every time we
151 *(unsigned*)(intptr_t)RetAddr = NewVal-RetAddr-4;
154 // If this is a stub, rewrite the call into an unconditional branch
155 // instruction so that two return addresses are not pushed onto the stack
156 // when the requested function finally gets called. This also makes the
157 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
158 ((unsigned char*)(intptr_t)RetAddr)[-1] = 0xE9;
161 // Change the return address to reexecute the call instruction...
165 TargetJITInfo::LazyResolverFn
166 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
167 JITCompilerFunction = F;
168 return X86CompilationCallback;
171 void *X86JITInfo::emitFunctionStub(void *Fn, MachineCodeEmitter &MCE) {
172 // Note, we cast to intptr_t here to silence a -pedantic warning that
173 // complains about casting a function pointer to a normal pointer.
174 if (Fn != (void*)(intptr_t)X86CompilationCallback) {
175 MCE.startFunctionStub(5);
177 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
178 return MCE.finishFunctionStub(0);
181 MCE.startFunctionStub(6);
182 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
184 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
186 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
187 return MCE.finishFunctionStub(0);
190 /// relocate - Before the JIT can run a block of code that has been emitted,
191 /// it must rewrite the code to contain the actual addresses of any
192 /// referenced global symbols.
193 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
194 unsigned NumRelocs, unsigned char* GOTBase) {
195 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
196 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
197 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
198 switch ((X86::RelocationType)MR->getRelocationType()) {
199 case X86::reloc_pcrel_word:
200 // PC relative relocation, add the relocated value to the value already in
201 // memory, after we adjust it for where the PC is.
202 ResultPtr = ResultPtr-(intptr_t)RelocPos-4;
203 *((intptr_t*)RelocPos) += ResultPtr;
205 case X86::reloc_absolute_word:
206 // Absolute relocation, just add the relocated value to the value already
208 *((intptr_t*)RelocPos) += ResultPtr;