1 //===-- X86JITInfo.cpp - Implement the JIT interfaces for the X86 target --===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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 convenient way for disabling usage of CFI directives.
51 // This is needed for old/broken assemblers (for example, gas on
52 // Darwin is pretty old and doesn't support these directives)
53 #if defined(__APPLE__)
56 // FIXME: Disable this until we really want to use it. Also, we will
57 // need to add some workarounds for compilers, which support
58 // only subset of these directives.
62 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
63 // callee saved registers, for the fastcc calling convention.
65 #if defined(__x86_64__)
66 // No need to save EAX/EDX for X86-64.
67 void X86CompilationCallback(void);
71 ".globl " ASMPREFIX "X86CompilationCallback\n"
72 ASMPREFIX "X86CompilationCallback:\n"
73 CFI(".cfi_startproc\n")
76 CFI(".cfi_def_cfa_offset 16\n")
77 CFI(".cfi_offset %rbp, -16\n")
80 CFI(".cfi_def_cfa_register %rbp\n")
81 // Save all int arg registers
83 CFI(".cfi_rel_offset %rdi, 0\n")
85 CFI(".cfi_rel_offset %rsi, 8\n")
87 CFI(".cfi_rel_offset %rdx, 16\n")
89 CFI(".cfi_rel_offset %rcx, 24\n")
91 CFI(".cfi_rel_offset %r8, 32\n")
93 CFI(".cfi_rel_offset %r9, 40\n")
94 // Align stack on 16-byte boundary. ESP might not be properly aligned
95 // (8 byte) if this is called from an indirect stub.
97 // Save all XMM arg registers
99 "movaps %xmm0, (%rsp)\n"
100 "movaps %xmm1, 16(%rsp)\n"
101 "movaps %xmm2, 32(%rsp)\n"
102 "movaps %xmm3, 48(%rsp)\n"
103 "movaps %xmm4, 64(%rsp)\n"
104 "movaps %xmm5, 80(%rsp)\n"
105 "movaps %xmm6, 96(%rsp)\n"
106 "movaps %xmm7, 112(%rsp)\n"
108 "movq %rbp, %rdi\n" // Pass prev frame and return address
109 "movq 8(%rbp), %rsi\n"
110 "call " ASMPREFIX "X86CompilationCallback2\n"
111 // Restore all XMM arg registers
112 "movaps 112(%rsp), %xmm7\n"
113 "movaps 96(%rsp), %xmm6\n"
114 "movaps 80(%rsp), %xmm5\n"
115 "movaps 64(%rsp), %xmm4\n"
116 "movaps 48(%rsp), %xmm3\n"
117 "movaps 32(%rsp), %xmm2\n"
118 "movaps 16(%rsp), %xmm1\n"
119 "movaps (%rsp), %xmm0\n"
122 CFI(".cfi_def_cfa_register %rsp\n")
123 // Restore all int arg registers
125 CFI(".cfi_adjust_cfa_offset 48\n")
127 CFI(".cfi_adjust_cfa_offset -8\n")
128 CFI(".cfi_restore %r9\n")
130 CFI(".cfi_adjust_cfa_offset -8\n")
131 CFI(".cfi_restore %r8\n")
133 CFI(".cfi_adjust_cfa_offset -8\n")
134 CFI(".cfi_restore %rcx\n")
136 CFI(".cfi_adjust_cfa_offset -8\n")
137 CFI(".cfi_restore %rdx\n")
139 CFI(".cfi_adjust_cfa_offset -8\n")
140 CFI(".cfi_restore %rsi\n")
142 CFI(".cfi_adjust_cfa_offset -8\n")
143 CFI(".cfi_restore %rdi\n")
146 CFI(".cfi_adjust_cfa_offset -8\n")
147 CFI(".cfi_restore %rbp\n")
149 CFI(".cfi_endproc\n")
151 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
153 void X86CompilationCallback(void);
157 ".globl " ASMPREFIX "X86CompilationCallback\n"
158 ASMPREFIX "X86CompilationCallback:\n"
159 CFI(".cfi_startproc\n")
161 CFI(".cfi_def_cfa_offset 8\n")
162 CFI(".cfi_offset %ebp, -8\n")
163 "movl %esp, %ebp\n" // Standard prologue
164 CFI(".cfi_def_cfa_register %ebp\n")
166 CFI(".cfi_rel_offset %eax, 0\n")
167 "pushl %edx\n" // Save EAX/EDX/ECX
168 CFI(".cfi_rel_offset %edx, 4\n")
170 CFI(".cfi_rel_offset %ecx, 8\n")
171 #if defined(__APPLE__)
172 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
175 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
176 "movl %eax, 4(%esp)\n"
177 "movl %ebp, (%esp)\n"
178 "call " ASMPREFIX "X86CompilationCallback2\n"
179 "movl %ebp, %esp\n" // Restore ESP
180 CFI(".cfi_def_cfa_register %esp\n")
182 CFI(".cfi_adjust_cfa_offset 12\n")
184 CFI(".cfi_adjust_cfa_offset -4\n")
185 CFI(".cfi_restore %ecx\n")
187 CFI(".cfi_adjust_cfa_offset -4\n")
188 CFI(".cfi_restore %edx\n")
190 CFI(".cfi_adjust_cfa_offset -4\n")
191 CFI(".cfi_restore %eax\n")
193 CFI(".cfi_adjust_cfa_offset -4\n")
194 CFI(".cfi_restore %ebp\n")
196 CFI(".cfi_endproc\n")
199 // Same as X86CompilationCallback but also saves XMM argument registers.
200 void X86CompilationCallback_SSE(void);
204 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
205 ASMPREFIX "X86CompilationCallback_SSE:\n"
206 CFI(".cfi_startproc\n")
208 CFI(".cfi_def_cfa_offset 8\n")
209 CFI(".cfi_offset %ebp, -8\n")
210 "movl %esp, %ebp\n" // Standard prologue
211 CFI(".cfi_def_cfa_register %ebp\n")
213 CFI(".cfi_rel_offset %eax, 0\n")
214 "pushl %edx\n" // Save EAX/EDX/ECX
215 CFI(".cfi_rel_offset %edx, 4\n")
217 CFI(".cfi_rel_offset %ecx, 8\n")
218 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
219 // Save all XMM arg registers
221 // FIXME: provide frame move information for xmm registers.
222 // This can be tricky, because CFA register is ebp (unaligned)
223 // and we need to produce offsets relative to it.
224 "movaps %xmm0, (%esp)\n"
225 "movaps %xmm1, 16(%esp)\n"
226 "movaps %xmm2, 32(%esp)\n"
227 "movaps %xmm3, 48(%esp)\n"
229 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
230 "movl %eax, 4(%esp)\n"
231 "movl %ebp, (%esp)\n"
232 "call " ASMPREFIX "X86CompilationCallback2\n"
234 "movaps 48(%esp), %xmm3\n"
235 CFI(".cfi_restore %xmm3\n")
236 "movaps 32(%esp), %xmm2\n"
237 CFI(".cfi_restore %xmm2\n")
238 "movaps 16(%esp), %xmm1\n"
239 CFI(".cfi_restore %xmm1\n")
240 "movaps (%esp), %xmm0\n"
241 CFI(".cfi_restore %xmm0\n")
242 "movl %ebp, %esp\n" // Restore ESP
243 CFI(".cfi_def_cfa_register esp\n")
245 CFI(".cfi_adjust_cfa_offset 12\n")
247 CFI(".cfi_adjust_cfa_offset -4\n")
248 CFI(".cfi_restore %ecx\n")
250 CFI(".cfi_adjust_cfa_offset -4\n")
251 CFI(".cfi_restore %edx\n")
253 CFI(".cfi_adjust_cfa_offset -4\n")
254 CFI(".cfi_restore %eax\n")
256 CFI(".cfi_adjust_cfa_offset -4\n")
257 CFI(".cfi_restore %ebp\n")
259 CFI(".cfi_endproc\n")
262 void X86CompilationCallback2(void);
264 _declspec(naked) void X86CompilationCallback(void) {
269 call X86CompilationCallback2
278 #else // Not an i386 host
279 void X86CompilationCallback() {
280 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
286 /// X86CompilationCallback - This is the target-specific function invoked by the
287 /// function stub when we did not know the real target of a call. This function
288 /// must locate the start of the stub or call site and pass it into the JIT
289 /// compiler function.
291 extern "C" void X86CompilationCallback2() {
292 assert(sizeof(size_t) == 4); // FIXME: handle Win64
293 intptr_t *RetAddrLoc = (intptr_t *)_AddressOfReturnAddress();
294 RetAddrLoc += 4; // skip over ret addr, edx, eax, ecx
295 intptr_t RetAddr = *RetAddrLoc;
297 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
298 intptr_t *RetAddrLoc = &StackPtr[1];
300 assert(*RetAddrLoc == RetAddr &&
301 "Could not find return address on the stack!");
303 // It's a stub if there is an interrupt marker after the call.
304 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
306 // The call instruction should have pushed the return value onto the stack...
308 RetAddr--; // Backtrack to the reference itself...
310 RetAddr -= 4; // Backtrack to the reference itself...
314 DOUT << "In callback! Addr=" << (void*)RetAddr
315 << " ESP=" << (void*)StackPtr
316 << ": Resolving call to function: "
317 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
320 // Sanity check to make sure this really is a call instruction.
322 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
323 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
325 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
328 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
330 // Rewrite the call target... so that we don't end up here every time we
333 *(intptr_t *)(RetAddr - 0xa) = NewVal;
335 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
339 // If this is a stub, rewrite the call into an unconditional branch
340 // instruction so that two return addresses are not pushed onto the stack
341 // when the requested function finally gets called. This also makes the
342 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
344 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
346 ((unsigned char*)RetAddr)[-1] = 0xE9;
350 // Change the return address to reexecute the call instruction...
358 TargetJITInfo::LazyResolverFn
359 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
360 JITCompilerFunction = F;
362 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
363 !defined(_MSC_VER) && !defined(__x86_64__)
364 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
370 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
371 // FIXME: support for AMD family of processors.
372 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
373 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
374 if ((EDX >> 25) & 0x1)
375 return X86CompilationCallback_SSE;
380 return X86CompilationCallback;
383 void *X86JITInfo::emitGlobalValueLazyPtr(void *GV, MachineCodeEmitter &MCE) {
385 MCE.startFunctionStub(8, 8);
386 MCE.emitWordLE(((unsigned *)&GV)[0]);
387 MCE.emitWordLE(((unsigned *)&GV)[1]);
389 MCE.startFunctionStub(4, 4);
390 MCE.emitWordLE((intptr_t)GV);
392 return MCE.finishFunctionStub(0);
395 void *X86JITInfo::emitFunctionStub(void *Fn, MachineCodeEmitter &MCE) {
396 // Note, we cast to intptr_t here to silence a -pedantic warning that
397 // complains about casting a function pointer to a normal pointer.
398 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
399 !defined(_MSC_VER) && !defined(__x86_64__)
400 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
401 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
403 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
407 MCE.startFunctionStub(13, 4);
408 MCE.emitByte(0x49); // REX prefix
409 MCE.emitByte(0xB8+2); // movabsq r10
410 MCE.emitWordLE(((unsigned *)&Fn)[0]);
411 MCE.emitWordLE(((unsigned *)&Fn)[1]);
412 MCE.emitByte(0x41); // REX prefix
413 MCE.emitByte(0xFF); // jmpq *r10
414 MCE.emitByte(2 | (4 << 3) | (3 << 6));
416 MCE.startFunctionStub(5, 4);
418 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
420 return MCE.finishFunctionStub(0);
424 MCE.startFunctionStub(14, 4);
425 MCE.emitByte(0x49); // REX prefix
426 MCE.emitByte(0xB8+2); // movabsq r10
427 MCE.emitWordLE(((unsigned *)&Fn)[0]);
428 MCE.emitWordLE(((unsigned *)&Fn)[1]);
429 MCE.emitByte(0x41); // REX prefix
430 MCE.emitByte(0xFF); // callq *r10
431 MCE.emitByte(2 | (2 << 3) | (3 << 6));
433 MCE.startFunctionStub(6, 4);
434 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
436 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
439 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
440 return MCE.finishFunctionStub(0);
443 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
444 /// specific basic block.
445 intptr_t X86JITInfo::getPICJumpTableEntry(intptr_t BB, intptr_t Entry) {
449 /// relocate - Before the JIT can run a block of code that has been emitted,
450 /// it must rewrite the code to contain the actual addresses of any
451 /// referenced global symbols.
452 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
453 unsigned NumRelocs, unsigned char* GOTBase) {
454 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
455 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
456 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
457 switch ((X86::RelocationType)MR->getRelocationType()) {
458 case X86::reloc_pcrel_word: {
459 // PC relative relocation, add the relocated value to the value already in
460 // memory, after we adjust it for where the PC is.
461 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
462 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
465 case X86::reloc_picrel_word: {
466 // PIC base relative relocation, add the relocated value to the value
467 // already in memory, after we adjust it for where the PIC base is.
468 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
469 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
472 case X86::reloc_absolute_word:
473 // Absolute relocation, just add the relocated value to the value already
475 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
477 case X86::reloc_absolute_dword:
478 *((intptr_t*)RelocPos) += ResultPtr;