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 "X86Subtarget.h"
18 #include "llvm/CodeGen/MachineCodeEmitter.h"
19 #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 // Get the ASMPREFIX for the current host. This is often '_'.
43 #ifndef __USER_LABEL_PREFIX__
44 #define __USER_LABEL_PREFIX__
46 #define GETASMPREFIX2(X) #X
47 #define GETASMPREFIX(X) GETASMPREFIX2(X)
48 #define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__)
50 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
51 // callee saved registers, for the fastcc calling convention.
53 #if defined(__x86_64__)
54 // No need to save EAX/EDX for X86-64.
55 void X86CompilationCallback(void);
59 ".globl " ASMPREFIX "X86CompilationCallback\n"
60 ASMPREFIX "X86CompilationCallback:\n"
65 // Save all int arg registers
72 // Align stack on 16-byte boundary. ESP might not be properly aligned
73 // (8 byte) if this is called from an indirect stub.
75 // Save all XMM arg registers
77 "movaps %xmm0, (%rsp)\n"
78 "movaps %xmm1, 16(%rsp)\n"
79 "movaps %xmm2, 32(%rsp)\n"
80 "movaps %xmm3, 48(%rsp)\n"
81 "movaps %xmm4, 64(%rsp)\n"
82 "movaps %xmm5, 80(%rsp)\n"
83 "movaps %xmm6, 96(%rsp)\n"
84 "movaps %xmm7, 112(%rsp)\n"
86 "movq %rbp, %rdi\n" // Pass prev frame and return address
87 "movq 8(%rbp), %rsi\n"
88 "call " ASMPREFIX "X86CompilationCallback2\n"
89 // Restore all XMM arg registers
90 "movaps 112(%rsp), %xmm7\n"
91 "movaps 96(%rsp), %xmm6\n"
92 "movaps 80(%rsp), %xmm5\n"
93 "movaps 64(%rsp), %xmm4\n"
94 "movaps 48(%rsp), %xmm3\n"
95 "movaps 32(%rsp), %xmm2\n"
96 "movaps 16(%rsp), %xmm1\n"
97 "movaps (%rsp), %xmm0\n"
100 // Restore all int arg registers
111 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
113 void X86CompilationCallback(void);
117 ".globl " ASMPREFIX "X86CompilationCallback\n"
118 ASMPREFIX "X86CompilationCallback:\n"
120 "movl %esp, %ebp\n" // Standard prologue
121 #if FASTCC_NUM_INT_ARGS_INREGS > 0
123 "pushl %edx\n" // Save EAX/EDX
125 #if defined(__APPLE__)
126 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
129 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
130 "movl %eax, 4(%esp)\n"
131 "movl %ebp, (%esp)\n"
132 "call " ASMPREFIX "X86CompilationCallback2\n"
133 "movl %ebp, %esp\n" // Restore ESP
134 #if FASTCC_NUM_INT_ARGS_INREGS > 0
142 // Same as X86CompilationCallback but also saves XMM argument registers.
143 void X86CompilationCallback_SSE(void);
147 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
148 ASMPREFIX "X86CompilationCallback_SSE:\n"
150 "movl %esp, %ebp\n" // Standard prologue
151 #if FASTCC_NUM_INT_ARGS_INREGS > 0
153 "pushl %edx\n" // Save EAX/EDX
155 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
156 // Save all XMM arg registers
158 "movaps %xmm0, (%esp)\n"
159 "movaps %xmm1, 16(%esp)\n"
160 "movaps %xmm2, 32(%esp)\n"
161 "movaps %xmm3, 48(%esp)\n"
163 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
164 "movl %eax, 4(%esp)\n"
165 "movl %ebp, (%esp)\n"
166 "call " ASMPREFIX "X86CompilationCallback2\n"
168 "movaps 48(%esp), %xmm3\n"
169 "movaps 32(%esp), %xmm2\n"
170 "movaps 16(%esp), %xmm1\n"
171 "movaps (%esp), %xmm0\n"
172 "movl %ebp, %esp\n" // Restore ESP
173 #if FASTCC_NUM_INT_ARGS_INREGS > 0
181 void X86CompilationCallback2(void);
183 _declspec(naked) void X86CompilationCallback(void) {
187 call X86CompilationCallback2
195 #else // Not an i386 host
196 void X86CompilationCallback() {
197 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
203 /// X86CompilationCallback - This is the target-specific function invoked by the
204 /// function stub when we did not know the real target of a call. This function
205 /// must locate the start of the stub or call site and pass it into the JIT
206 /// compiler function.
208 extern "C" void X86CompilationCallback2() {
209 assert(sizeof(size_t) == 4); // FIXME: handle Win64
210 unsigned *RetAddrLoc = (unsigned *)_AddressOfReturnAddress();
211 RetAddrLoc += 3; // skip over ret addr, edx, eax
212 unsigned RetAddr = *RetAddrLoc;
214 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
215 intptr_t *RetAddrLoc = &StackPtr[1];
217 assert(*RetAddrLoc == RetAddr &&
218 "Could not find return address on the stack!");
220 // It's a stub if there is an interrupt marker after the call.
221 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
223 // The call instruction should have pushed the return value onto the stack...
224 RetAddr -= 4; // Backtrack to the reference itself...
227 DOUT << "In callback! Addr=" << (void*)RetAddr
228 << " ESP=" << (void*)StackPtr
229 << ": Resolving call to function: "
230 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
233 // Sanity check to make sure this really is a call instruction.
234 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
236 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
238 // Rewrite the call target... so that we don't end up here every time we
240 *(unsigned *)RetAddr = (unsigned)(NewVal-RetAddr-4);
243 // If this is a stub, rewrite the call into an unconditional branch
244 // instruction so that two return addresses are not pushed onto the stack
245 // when the requested function finally gets called. This also makes the
246 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
247 ((unsigned char*)RetAddr)[-1] = 0xE9;
250 // Change the return address to reexecute the call instruction...
254 TargetJITInfo::LazyResolverFn
255 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
256 JITCompilerFunction = F;
258 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
259 !defined(_MSC_VER) && !defined(__x86_64__)
260 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
266 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
267 // FIXME: support for AMD family of processors.
268 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
269 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
270 if ((EDX >> 25) & 0x1)
271 return X86CompilationCallback_SSE;
276 return X86CompilationCallback;
279 void *X86JITInfo::emitFunctionStub(void *Fn, MachineCodeEmitter &MCE) {
280 // Note, we cast to intptr_t here to silence a -pedantic warning that
281 // complains about casting a function pointer to a normal pointer.
282 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
283 !defined(_MSC_VER) && !defined(__x86_64__)
284 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
285 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
287 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
290 MCE.startFunctionStub(5, 4);
292 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
293 return MCE.finishFunctionStub(0);
296 MCE.startFunctionStub(6, 4);
297 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
299 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
301 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
302 return MCE.finishFunctionStub(0);
305 /// relocate - Before the JIT can run a block of code that has been emitted,
306 /// it must rewrite the code to contain the actual addresses of any
307 /// referenced global symbols.
308 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
309 unsigned NumRelocs, unsigned char* GOTBase) {
310 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
311 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
312 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
313 switch ((X86::RelocationType)MR->getRelocationType()) {
314 case X86::reloc_pcrel_word: {
315 // PC relative relocation, add the relocated value to the value already in
316 // memory, after we adjust it for where the PC is.
317 ResultPtr = ResultPtr-(intptr_t)RelocPos-4-MR->getConstantVal();
318 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
321 case X86::reloc_absolute_word:
322 // Absolute relocation, just add the relocated value to the value already
324 *((unsigned*)RelocPos) += (unsigned)ResultPtr;