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"
25 extern "C" void *_AddressOfReturnAddress(void);
26 #pragma intrinsic(_AddressOfReturnAddress)
29 void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) {
30 unsigned char *OldByte = (unsigned char *)Old;
31 *OldByte++ = 0xE9; // Emit JMP opcode.
32 unsigned *OldWord = (unsigned *)OldByte;
33 unsigned NewAddr = (intptr_t)New;
34 unsigned OldAddr = (intptr_t)OldWord;
35 *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code.
39 /// JITCompilerFunction - This contains the address of the JIT function used to
40 /// compile a function lazily.
41 static TargetJITInfo::JITCompilerFn JITCompilerFunction;
43 // Get the ASMPREFIX for the current host. This is often '_'.
44 #ifndef __USER_LABEL_PREFIX__
45 #define __USER_LABEL_PREFIX__
47 #define GETASMPREFIX2(X) #X
48 #define GETASMPREFIX(X) GETASMPREFIX2(X)
49 #define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__)
51 // Provide a convenient way for disabling usage of CFI directives.
52 // This is needed for old/broken assemblers (for example, gas on
53 // Darwin is pretty old and doesn't support these directives)
54 #if defined(__APPLE__)
57 // FIXME: Disable this until we really want to use it. Also, we will
58 // need to add some workarounds for compilers, which support
59 // only subset of these directives.
63 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
64 // callee saved registers, for the fastcc calling convention.
66 #if defined(__x86_64__)
67 // No need to save EAX/EDX for X86-64.
68 void X86CompilationCallback(void);
72 ".globl " ASMPREFIX "X86CompilationCallback\n"
73 ASMPREFIX "X86CompilationCallback:\n"
74 CFI(".cfi_startproc\n")
77 CFI(".cfi_def_cfa_offset 16\n")
78 CFI(".cfi_offset %rbp, -16\n")
81 CFI(".cfi_def_cfa_register %rbp\n")
82 // Save all int arg registers
84 CFI(".cfi_rel_offset %rdi, 0\n")
86 CFI(".cfi_rel_offset %rsi, 8\n")
88 CFI(".cfi_rel_offset %rdx, 16\n")
90 CFI(".cfi_rel_offset %rcx, 24\n")
92 CFI(".cfi_rel_offset %r8, 32\n")
94 CFI(".cfi_rel_offset %r9, 40\n")
95 // Align stack on 16-byte boundary. ESP might not be properly aligned
96 // (8 byte) if this is called from an indirect stub.
98 // Save all XMM arg registers
100 "movaps %xmm0, (%rsp)\n"
101 "movaps %xmm1, 16(%rsp)\n"
102 "movaps %xmm2, 32(%rsp)\n"
103 "movaps %xmm3, 48(%rsp)\n"
104 "movaps %xmm4, 64(%rsp)\n"
105 "movaps %xmm5, 80(%rsp)\n"
106 "movaps %xmm6, 96(%rsp)\n"
107 "movaps %xmm7, 112(%rsp)\n"
109 "movq %rbp, %rdi\n" // Pass prev frame and return address
110 "movq 8(%rbp), %rsi\n"
111 "call " ASMPREFIX "X86CompilationCallback2\n"
112 // Restore all XMM arg registers
113 "movaps 112(%rsp), %xmm7\n"
114 "movaps 96(%rsp), %xmm6\n"
115 "movaps 80(%rsp), %xmm5\n"
116 "movaps 64(%rsp), %xmm4\n"
117 "movaps 48(%rsp), %xmm3\n"
118 "movaps 32(%rsp), %xmm2\n"
119 "movaps 16(%rsp), %xmm1\n"
120 "movaps (%rsp), %xmm0\n"
123 CFI(".cfi_def_cfa_register %rsp\n")
124 // Restore all int arg registers
126 CFI(".cfi_adjust_cfa_offset 48\n")
128 CFI(".cfi_adjust_cfa_offset -8\n")
129 CFI(".cfi_restore %r9\n")
131 CFI(".cfi_adjust_cfa_offset -8\n")
132 CFI(".cfi_restore %r8\n")
134 CFI(".cfi_adjust_cfa_offset -8\n")
135 CFI(".cfi_restore %rcx\n")
137 CFI(".cfi_adjust_cfa_offset -8\n")
138 CFI(".cfi_restore %rdx\n")
140 CFI(".cfi_adjust_cfa_offset -8\n")
141 CFI(".cfi_restore %rsi\n")
143 CFI(".cfi_adjust_cfa_offset -8\n")
144 CFI(".cfi_restore %rdi\n")
147 CFI(".cfi_adjust_cfa_offset -8\n")
148 CFI(".cfi_restore %rbp\n")
150 CFI(".cfi_endproc\n")
152 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
154 void X86CompilationCallback(void);
158 ".globl " ASMPREFIX "X86CompilationCallback\n"
159 ASMPREFIX "X86CompilationCallback:\n"
160 CFI(".cfi_startproc\n")
162 CFI(".cfi_def_cfa_offset 8\n")
163 CFI(".cfi_offset %ebp, -8\n")
164 "movl %esp, %ebp\n" // Standard prologue
165 CFI(".cfi_def_cfa_register %ebp\n")
167 CFI(".cfi_rel_offset %eax, 0\n")
168 "pushl %edx\n" // Save EAX/EDX/ECX
169 CFI(".cfi_rel_offset %edx, 4\n")
171 CFI(".cfi_rel_offset %ecx, 8\n")
172 #if defined(__APPLE__)
173 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
176 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
177 "movl %eax, 4(%esp)\n"
178 "movl %ebp, (%esp)\n"
179 "call " ASMPREFIX "X86CompilationCallback2\n"
180 "movl %ebp, %esp\n" // Restore ESP
181 CFI(".cfi_def_cfa_register %esp\n")
183 CFI(".cfi_adjust_cfa_offset 12\n")
185 CFI(".cfi_adjust_cfa_offset -4\n")
186 CFI(".cfi_restore %ecx\n")
188 CFI(".cfi_adjust_cfa_offset -4\n")
189 CFI(".cfi_restore %edx\n")
191 CFI(".cfi_adjust_cfa_offset -4\n")
192 CFI(".cfi_restore %eax\n")
194 CFI(".cfi_adjust_cfa_offset -4\n")
195 CFI(".cfi_restore %ebp\n")
197 CFI(".cfi_endproc\n")
200 // Same as X86CompilationCallback but also saves XMM argument registers.
201 void X86CompilationCallback_SSE(void);
205 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
206 ASMPREFIX "X86CompilationCallback_SSE:\n"
207 CFI(".cfi_startproc\n")
209 CFI(".cfi_def_cfa_offset 8\n")
210 CFI(".cfi_offset %ebp, -8\n")
211 "movl %esp, %ebp\n" // Standard prologue
212 CFI(".cfi_def_cfa_register %ebp\n")
214 CFI(".cfi_rel_offset %eax, 0\n")
215 "pushl %edx\n" // Save EAX/EDX/ECX
216 CFI(".cfi_rel_offset %edx, 4\n")
218 CFI(".cfi_rel_offset %ecx, 8\n")
219 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
220 // Save all XMM arg registers
222 // FIXME: provide frame move information for xmm registers.
223 // This can be tricky, because CFA register is ebp (unaligned)
224 // and we need to produce offsets relative to it.
225 "movaps %xmm0, (%esp)\n"
226 "movaps %xmm1, 16(%esp)\n"
227 "movaps %xmm2, 32(%esp)\n"
228 "movaps %xmm3, 48(%esp)\n"
230 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
231 "movl %eax, 4(%esp)\n"
232 "movl %ebp, (%esp)\n"
233 "call " ASMPREFIX "X86CompilationCallback2\n"
235 "movaps 48(%esp), %xmm3\n"
236 CFI(".cfi_restore %xmm3\n")
237 "movaps 32(%esp), %xmm2\n"
238 CFI(".cfi_restore %xmm2\n")
239 "movaps 16(%esp), %xmm1\n"
240 CFI(".cfi_restore %xmm1\n")
241 "movaps (%esp), %xmm0\n"
242 CFI(".cfi_restore %xmm0\n")
243 "movl %ebp, %esp\n" // Restore ESP
244 CFI(".cfi_def_cfa_register esp\n")
246 CFI(".cfi_adjust_cfa_offset 12\n")
248 CFI(".cfi_adjust_cfa_offset -4\n")
249 CFI(".cfi_restore %ecx\n")
251 CFI(".cfi_adjust_cfa_offset -4\n")
252 CFI(".cfi_restore %edx\n")
254 CFI(".cfi_adjust_cfa_offset -4\n")
255 CFI(".cfi_restore %eax\n")
257 CFI(".cfi_adjust_cfa_offset -4\n")
258 CFI(".cfi_restore %ebp\n")
260 CFI(".cfi_endproc\n")
263 void X86CompilationCallback2(void);
265 _declspec(naked) void X86CompilationCallback(void) {
270 call X86CompilationCallback2
279 #else // Not an i386 host
280 void X86CompilationCallback() {
281 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
287 /// X86CompilationCallback - This is the target-specific function invoked by the
288 /// function stub when we did not know the real target of a call. This function
289 /// must locate the start of the stub or call site and pass it into the JIT
290 /// compiler function.
292 extern "C" void X86CompilationCallback2() {
293 assert(sizeof(size_t) == 4); // FIXME: handle Win64
294 intptr_t *RetAddrLoc = (intptr_t *)_AddressOfReturnAddress();
295 RetAddrLoc += 4; // skip over ret addr, edx, eax, ecx
296 intptr_t RetAddr = *RetAddrLoc;
298 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
299 intptr_t *RetAddrLoc = &StackPtr[1];
301 assert(*RetAddrLoc == RetAddr &&
302 "Could not find return address on the stack!");
304 // It's a stub if there is an interrupt marker after the call.
305 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
307 // The call instruction should have pushed the return value onto the stack...
309 RetAddr--; // Backtrack to the reference itself...
311 RetAddr -= 4; // Backtrack to the reference itself...
315 DOUT << "In callback! Addr=" << (void*)RetAddr
316 << " ESP=" << (void*)StackPtr
317 << ": Resolving call to function: "
318 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
321 // Sanity check to make sure this really is a call instruction.
323 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
324 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
326 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
329 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
331 // Rewrite the call target... so that we don't end up here every time we
334 *(intptr_t *)(RetAddr - 0xa) = NewVal;
336 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
340 // If this is a stub, rewrite the call into an unconditional branch
341 // instruction so that two return addresses are not pushed onto the stack
342 // when the requested function finally gets called. This also makes the
343 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
345 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
347 ((unsigned char*)RetAddr)[-1] = 0xE9;
351 // Change the return address to reexecute the call instruction...
359 TargetJITInfo::LazyResolverFn
360 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
361 JITCompilerFunction = F;
363 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
364 !defined(_MSC_VER) && !defined(__x86_64__)
365 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
371 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
372 // FIXME: support for AMD family of processors.
373 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
374 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
375 if ((EDX >> 25) & 0x1)
376 return X86CompilationCallback_SSE;
381 return X86CompilationCallback;
384 void *X86JITInfo::emitGlobalValueLazyPtr(void *GV, MachineCodeEmitter &MCE) {
386 MCE.startFunctionStub(8, 8);
387 MCE.emitWordLE(((unsigned *)&GV)[0]);
388 MCE.emitWordLE(((unsigned *)&GV)[1]);
390 MCE.startFunctionStub(4, 4);
391 MCE.emitWordLE((intptr_t)GV);
393 return MCE.finishFunctionStub(0);
396 void *X86JITInfo::emitFunctionStub(void *Fn, MachineCodeEmitter &MCE) {
397 // Note, we cast to intptr_t here to silence a -pedantic warning that
398 // complains about casting a function pointer to a normal pointer.
399 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
400 !defined(_MSC_VER) && !defined(__x86_64__)
401 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
402 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
404 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
408 MCE.startFunctionStub(13, 4);
409 MCE.emitByte(0x49); // REX prefix
410 MCE.emitByte(0xB8+2); // movabsq r10
411 MCE.emitWordLE(((unsigned *)&Fn)[0]);
412 MCE.emitWordLE(((unsigned *)&Fn)[1]);
413 MCE.emitByte(0x41); // REX prefix
414 MCE.emitByte(0xFF); // jmpq *r10
415 MCE.emitByte(2 | (4 << 3) | (3 << 6));
417 MCE.startFunctionStub(5, 4);
419 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
421 return MCE.finishFunctionStub(0);
425 MCE.startFunctionStub(14, 4);
426 MCE.emitByte(0x49); // REX prefix
427 MCE.emitByte(0xB8+2); // movabsq r10
428 MCE.emitWordLE(((unsigned *)&Fn)[0]);
429 MCE.emitWordLE(((unsigned *)&Fn)[1]);
430 MCE.emitByte(0x41); // REX prefix
431 MCE.emitByte(0xFF); // callq *r10
432 MCE.emitByte(2 | (2 << 3) | (3 << 6));
434 MCE.startFunctionStub(6, 4);
435 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
437 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
440 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
441 return MCE.finishFunctionStub(0);
444 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
445 /// specific basic block.
446 intptr_t X86JITInfo::getPICJumpTableEntry(intptr_t BB, intptr_t Entry) {
450 /// relocate - Before the JIT can run a block of code that has been emitted,
451 /// it must rewrite the code to contain the actual addresses of any
452 /// referenced global symbols.
453 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
454 unsigned NumRelocs, unsigned char* GOTBase) {
455 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
456 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
457 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
458 switch ((X86::RelocationType)MR->getRelocationType()) {
459 case X86::reloc_pcrel_word: {
460 // PC relative relocation, add the relocated value to the value already in
461 // memory, after we adjust it for where the PC is.
462 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
463 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
466 case X86::reloc_picrel_word: {
467 // PIC base relative relocation, add the relocated value to the value
468 // already in memory, after we adjust it for where the PIC base is.
469 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
470 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
473 case X86::reloc_absolute_word:
474 // Absolute relocation, just add the relocated value to the value already
476 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
478 case X86::reloc_absolute_dword:
479 *((intptr_t*)RelocPos) += ResultPtr;