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 #if defined(__APPLE__)
56 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
57 // callee saved registers, for the fastcc calling convention.
59 #if defined(__x86_64__)
60 // No need to save EAX/EDX for X86-64.
61 void X86CompilationCallback(void);
65 ".globl " ASMPREFIX "X86CompilationCallback\n"
66 ASMPREFIX "X86CompilationCallback:\n"
67 CFI(".cfi_startproc\n")
70 CFI(".cfi_def_cfa_offset 16\n")
71 CFI(".cfi_offset %rbp, -16\n")
74 CFI(".cfi_def_cfa_register %rbp\n")
75 // Save all int arg registers
77 CFI(".cfi_rel_offset %rdi, 0\n")
79 CFI(".cfi_rel_offset %rsi, 8\n")
81 CFI(".cfi_rel_offset %rdx, 16\n")
83 CFI(".cfi_rel_offset %rcx, 24\n")
85 CFI(".cfi_rel_offset %r8, 32\n")
87 CFI(".cfi_rel_offset %r9, 40\n")
88 // Align stack on 16-byte boundary. ESP might not be properly aligned
89 // (8 byte) if this is called from an indirect stub.
91 // Save all XMM arg registers
93 "movaps %xmm0, (%rsp)\n"
94 "movaps %xmm1, 16(%rsp)\n"
95 "movaps %xmm2, 32(%rsp)\n"
96 "movaps %xmm3, 48(%rsp)\n"
97 "movaps %xmm4, 64(%rsp)\n"
98 "movaps %xmm5, 80(%rsp)\n"
99 "movaps %xmm6, 96(%rsp)\n"
100 "movaps %xmm7, 112(%rsp)\n"
102 "movq %rbp, %rdi\n" // Pass prev frame and return address
103 "movq 8(%rbp), %rsi\n"
104 "call " ASMPREFIX "X86CompilationCallback2\n"
105 // Restore all XMM arg registers
106 "movaps 112(%rsp), %xmm7\n"
107 "movaps 96(%rsp), %xmm6\n"
108 "movaps 80(%rsp), %xmm5\n"
109 "movaps 64(%rsp), %xmm4\n"
110 "movaps 48(%rsp), %xmm3\n"
111 "movaps 32(%rsp), %xmm2\n"
112 "movaps 16(%rsp), %xmm1\n"
113 "movaps (%rsp), %xmm0\n"
116 CFI(".cfi_def_cfa_register esp\n")
117 // Restore all int arg registers
119 CFI(".cfi_adjust_cfa_offset 48\n")
121 CFI(".cfi_adjust_cfa_offset -8\n")
122 CFI(".cfi_restore %r9\n")
124 CFI(".cfi_adjust_cfa_offset -8\n")
125 CFI(".cfi_restore %r8\n")
127 CFI(".cfi_adjust_cfa_offset -8\n")
128 CFI(".cfi_restore %rcx\n")
130 CFI(".cfi_adjust_cfa_offset -8\n")
131 CFI(".cfi_restore %rdx\n")
133 CFI(".cfi_adjust_cfa_offset -8\n")
134 CFI(".cfi_restore %rsi\n")
136 CFI(".cfi_adjust_cfa_offset -8\n")
137 CFI(".cfi_restore %rdi\n")
140 CFI(".cfi_adjust_cfa_offset -8\n")
141 CFI(".cfi_restore %rbp\n")
143 CFI(".cfi_endproc\n")
145 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
147 void X86CompilationCallback(void);
151 ".globl " ASMPREFIX "X86CompilationCallback\n"
152 ASMPREFIX "X86CompilationCallback:\n"
153 CFI(".cfi_startproc\n")
155 CFI(".cfi_def_cfa_offset 8\n")
156 CFI(".cfi_offset %ebp, -8\n")
157 "movl %esp, %ebp\n" // Standard prologue
158 CFI(".cfi_def_cfa_register %ebp\n")
160 CFI(".cfi_rel_offset %eax, 0\n")
161 "pushl %edx\n" // Save EAX/EDX/ECX
162 CFI(".cfi_rel_offset %edx, 4\n")
164 CFI(".cfi_rel_offset %ecx, 8\n")
165 #if defined(__APPLE__)
166 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
169 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
170 "movl %eax, 4(%esp)\n"
171 "movl %ebp, (%esp)\n"
172 "call " ASMPREFIX "X86CompilationCallback2\n"
173 "movl %ebp, %esp\n" // Restore ESP
174 CFI(".cfi_def_cfa_register %esp\n")
176 CFI(".cfi_adjust_cfa_offset 12\n")
178 CFI(".cfi_adjust_cfa_offset -4\n")
179 CFI(".cfi_restore %ecx\n")
181 CFI(".cfi_adjust_cfa_offset -4\n")
182 CFI(".cfi_restore %edx\n")
184 CFI(".cfi_adjust_cfa_offset -4\n")
185 CFI(".cfi_restore %eax\n")
187 CFI(".cfi_adjust_cfa_offset -4\n")
188 CFI(".cfi_restore %ebp\n")
190 CFI(".cfi_endproc\n")
193 // Same as X86CompilationCallback but also saves XMM argument registers.
194 void X86CompilationCallback_SSE(void);
198 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
199 ASMPREFIX "X86CompilationCallback_SSE:\n"
200 CFI(".cfi_startproc\n")
202 CFI(".cfi_def_cfa_offset 8\n")
203 CFI(".cfi_offset %ebp, -8\n")
204 "movl %esp, %ebp\n" // Standard prologue
205 CFI(".cfi_def_cfa_register %ebp\n")
207 CFI(".cfi_rel_offset %eax, 0\n")
208 "pushl %edx\n" // Save EAX/EDX/ECX
209 CFI(".cfi_rel_offset %edx, 4\n")
211 CFI(".cfi_rel_offset %ecx, 8\n")
212 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
213 // Save all XMM arg registers
215 // FIXME: provide frame move information for xmm registers.
216 // This can be tricky, because CFA register is ebp (unaligned)
217 // and we need to produce offsets relative to it.
218 "movaps %xmm0, (%esp)\n"
219 "movaps %xmm1, 16(%esp)\n"
220 "movaps %xmm2, 32(%esp)\n"
221 "movaps %xmm3, 48(%esp)\n"
223 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
224 "movl %eax, 4(%esp)\n"
225 "movl %ebp, (%esp)\n"
226 "call " ASMPREFIX "X86CompilationCallback2\n"
228 "movaps 48(%esp), %xmm3\n"
229 CFI(".cfi_restore %xmm3\n")
230 "movaps 32(%esp), %xmm2\n"
231 CFI(".cfi_restore %xmm2\n")
232 "movaps 16(%esp), %xmm1\n"
233 CFI(".cfi_restore %xmm1\n")
234 "movaps (%esp), %xmm0\n"
235 CFI(".cfi_restore %xmm0\n")
236 "movl %ebp, %esp\n" // Restore ESP
237 CFI(".cfi_def_cfa_register esp\n")
239 CFI(".cfi_adjust_cfa_offset 12\n")
241 CFI(".cfi_adjust_cfa_offset -4\n")
242 CFI(".cfi_restore %ecx\n")
244 CFI(".cfi_adjust_cfa_offset -4\n")
245 CFI(".cfi_restore %edx\n")
247 CFI(".cfi_adjust_cfa_offset -4\n")
248 CFI(".cfi_restore %eax\n")
250 CFI(".cfi_adjust_cfa_offset -4\n")
251 CFI(".cfi_restore %ebp\n")
253 CFI(".cfi_endproc\n")
256 void X86CompilationCallback2(void);
258 _declspec(naked) void X86CompilationCallback(void) {
263 call X86CompilationCallback2
272 #else // Not an i386 host
273 void X86CompilationCallback() {
274 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
280 /// X86CompilationCallback - This is the target-specific function invoked by the
281 /// function stub when we did not know the real target of a call. This function
282 /// must locate the start of the stub or call site and pass it into the JIT
283 /// compiler function.
285 extern "C" void X86CompilationCallback2() {
286 assert(sizeof(size_t) == 4); // FIXME: handle Win64
287 intptr_t *RetAddrLoc = (intptr_t *)_AddressOfReturnAddress();
288 RetAddrLoc += 4; // skip over ret addr, edx, eax, ecx
289 intptr_t RetAddr = *RetAddrLoc;
291 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
292 intptr_t *RetAddrLoc = &StackPtr[1];
294 assert(*RetAddrLoc == RetAddr &&
295 "Could not find return address on the stack!");
297 // It's a stub if there is an interrupt marker after the call.
298 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
300 // The call instruction should have pushed the return value onto the stack...
302 RetAddr--; // Backtrack to the reference itself...
304 RetAddr -= 4; // Backtrack to the reference itself...
308 DOUT << "In callback! Addr=" << (void*)RetAddr
309 << " ESP=" << (void*)StackPtr
310 << ": Resolving call to function: "
311 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
314 // Sanity check to make sure this really is a call instruction.
316 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
317 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
319 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
322 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
324 // Rewrite the call target... so that we don't end up here every time we
327 *(intptr_t *)(RetAddr - 0xa) = NewVal;
329 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
333 // If this is a stub, rewrite the call into an unconditional branch
334 // instruction so that two return addresses are not pushed onto the stack
335 // when the requested function finally gets called. This also makes the
336 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
338 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
340 ((unsigned char*)RetAddr)[-1] = 0xE9;
344 // Change the return address to reexecute the call instruction...
352 TargetJITInfo::LazyResolverFn
353 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
354 JITCompilerFunction = F;
356 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
357 !defined(_MSC_VER) && !defined(__x86_64__)
358 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
364 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
365 // FIXME: support for AMD family of processors.
366 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
367 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
368 if ((EDX >> 25) & 0x1)
369 return X86CompilationCallback_SSE;
374 return X86CompilationCallback;
377 void *X86JITInfo::emitFunctionStub(void *Fn, MachineCodeEmitter &MCE) {
378 // Note, we cast to intptr_t here to silence a -pedantic warning that
379 // complains about casting a function pointer to a normal pointer.
380 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
381 !defined(_MSC_VER) && !defined(__x86_64__)
382 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
383 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
385 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
389 MCE.startFunctionStub(13, 4);
390 MCE.emitByte(0x49); // REX prefix
391 MCE.emitByte(0xB8+2); // movabsq r10
392 MCE.emitWordLE(((unsigned *)&Fn)[0]);
393 MCE.emitWordLE(((unsigned *)&Fn)[1]);
394 MCE.emitByte(0x41); // REX prefix
395 MCE.emitByte(0xFF); // jmpq *r10
396 MCE.emitByte(2 | (4 << 3) | (3 << 6));
398 MCE.startFunctionStub(5, 4);
400 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
402 return MCE.finishFunctionStub(0);
406 MCE.startFunctionStub(14, 4);
407 MCE.emitByte(0x49); // REX prefix
408 MCE.emitByte(0xB8+2); // movabsq r10
409 MCE.emitWordLE(((unsigned *)&Fn)[0]);
410 MCE.emitWordLE(((unsigned *)&Fn)[1]);
411 MCE.emitByte(0x41); // REX prefix
412 MCE.emitByte(0xFF); // callq *r10
413 MCE.emitByte(2 | (2 << 3) | (3 << 6));
415 MCE.startFunctionStub(6, 4);
416 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
418 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
421 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
422 return MCE.finishFunctionStub(0);
425 /// relocate - Before the JIT can run a block of code that has been emitted,
426 /// it must rewrite the code to contain the actual addresses of any
427 /// referenced global symbols.
428 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
429 unsigned NumRelocs, unsigned char* GOTBase) {
430 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
431 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
432 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
433 switch ((X86::RelocationType)MR->getRelocationType()) {
434 case X86::reloc_pcrel_word: {
435 // PC relative relocation, add the relocated value to the value already in
436 // memory, after we adjust it for where the PC is.
437 ResultPtr = ResultPtr-(intptr_t)RelocPos-4-MR->getConstantVal();
438 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
441 case X86::reloc_absolute_word:
442 // Absolute relocation, just add the relocated value to the value already
444 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
446 case X86::reloc_absolute_dword:
447 *((intptr_t*)RelocPos) += ResultPtr;