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/Function.h"
19 #include "llvm/CodeGen/MachineCodeEmitter.h"
20 #include "llvm/Config/alloca.h"
21 #include "llvm/Support/Compiler.h"
26 // Determine the platform we're running on
27 #if defined (__x86_64__) || defined (_M_AMD64)
29 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
33 void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) {
34 unsigned char *OldByte = (unsigned char *)Old;
35 *OldByte++ = 0xE9; // Emit JMP opcode.
36 unsigned *OldWord = (unsigned *)OldByte;
37 unsigned NewAddr = (intptr_t)New;
38 unsigned OldAddr = (intptr_t)OldWord;
39 *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code.
43 /// JITCompilerFunction - This contains the address of the JIT function used to
44 /// compile a function lazily.
45 static TargetJITInfo::JITCompilerFn JITCompilerFunction;
47 // Get the ASMPREFIX for the current host. This is often '_'.
48 #ifndef __USER_LABEL_PREFIX__
49 #define __USER_LABEL_PREFIX__
51 #define GETASMPREFIX2(X) #X
52 #define GETASMPREFIX(X) GETASMPREFIX2(X)
53 #define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__)
55 // Check if building with -fPIC
56 #if defined(__PIC__) && __PIC__ && defined(__linux__)
57 #define ASMCALLSUFFIX "@PLT"
62 // Provide a convenient way for disabling usage of CFI directives.
63 // This is needed for old/broken assemblers (for example, gas on
64 // Darwin is pretty old and doesn't support these directives)
65 #if defined(__APPLE__)
68 // FIXME: Disable this until we really want to use it. Also, we will
69 // need to add some workarounds for compilers, which support
70 // only subset of these directives.
74 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
75 // callee saved registers, for the fastcc calling convention.
77 #if defined(X86_64_JIT)
79 // No need to save EAX/EDX for X86-64.
80 void X86CompilationCallback(void);
84 ".globl " ASMPREFIX "X86CompilationCallback\n"
85 ASMPREFIX "X86CompilationCallback:\n"
86 CFI(".cfi_startproc\n")
89 CFI(".cfi_def_cfa_offset 16\n")
90 CFI(".cfi_offset %rbp, -16\n")
93 CFI(".cfi_def_cfa_register %rbp\n")
94 // Save all int arg registers
96 CFI(".cfi_rel_offset %rdi, 0\n")
98 CFI(".cfi_rel_offset %rsi, 8\n")
100 CFI(".cfi_rel_offset %rdx, 16\n")
102 CFI(".cfi_rel_offset %rcx, 24\n")
104 CFI(".cfi_rel_offset %r8, 32\n")
106 CFI(".cfi_rel_offset %r9, 40\n")
107 // Align stack on 16-byte boundary. ESP might not be properly aligned
108 // (8 byte) if this is called from an indirect stub.
110 // Save all XMM arg registers
112 "movaps %xmm0, (%rsp)\n"
113 "movaps %xmm1, 16(%rsp)\n"
114 "movaps %xmm2, 32(%rsp)\n"
115 "movaps %xmm3, 48(%rsp)\n"
116 "movaps %xmm4, 64(%rsp)\n"
117 "movaps %xmm5, 80(%rsp)\n"
118 "movaps %xmm6, 96(%rsp)\n"
119 "movaps %xmm7, 112(%rsp)\n"
121 "movq %rbp, %rdi\n" // Pass prev frame and return address
122 "movq 8(%rbp), %rsi\n"
123 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
124 // Restore all XMM arg registers
125 "movaps 112(%rsp), %xmm7\n"
126 "movaps 96(%rsp), %xmm6\n"
127 "movaps 80(%rsp), %xmm5\n"
128 "movaps 64(%rsp), %xmm4\n"
129 "movaps 48(%rsp), %xmm3\n"
130 "movaps 32(%rsp), %xmm2\n"
131 "movaps 16(%rsp), %xmm1\n"
132 "movaps (%rsp), %xmm0\n"
135 CFI(".cfi_def_cfa_register %rsp\n")
136 // Restore all int arg registers
138 CFI(".cfi_adjust_cfa_offset 48\n")
140 CFI(".cfi_adjust_cfa_offset -8\n")
141 CFI(".cfi_restore %r9\n")
143 CFI(".cfi_adjust_cfa_offset -8\n")
144 CFI(".cfi_restore %r8\n")
146 CFI(".cfi_adjust_cfa_offset -8\n")
147 CFI(".cfi_restore %rcx\n")
149 CFI(".cfi_adjust_cfa_offset -8\n")
150 CFI(".cfi_restore %rdx\n")
152 CFI(".cfi_adjust_cfa_offset -8\n")
153 CFI(".cfi_restore %rsi\n")
155 CFI(".cfi_adjust_cfa_offset -8\n")
156 CFI(".cfi_restore %rdi\n")
159 CFI(".cfi_adjust_cfa_offset -8\n")
160 CFI(".cfi_restore %rbp\n")
162 CFI(".cfi_endproc\n")
165 // No inline assembler support on this platform. The routine is in external
167 void X86CompilationCallback();
170 #elif defined (X86_32_JIT)
172 void X86CompilationCallback(void);
176 ".globl " ASMPREFIX "X86CompilationCallback\n"
177 ASMPREFIX "X86CompilationCallback:\n"
178 CFI(".cfi_startproc\n")
180 CFI(".cfi_def_cfa_offset 8\n")
181 CFI(".cfi_offset %ebp, -8\n")
182 "movl %esp, %ebp\n" // Standard prologue
183 CFI(".cfi_def_cfa_register %ebp\n")
185 CFI(".cfi_rel_offset %eax, 0\n")
186 "pushl %edx\n" // Save EAX/EDX/ECX
187 CFI(".cfi_rel_offset %edx, 4\n")
189 CFI(".cfi_rel_offset %ecx, 8\n")
190 # if defined(__APPLE__)
191 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
194 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
195 "movl %eax, 4(%esp)\n"
196 "movl %ebp, (%esp)\n"
197 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
198 "movl %ebp, %esp\n" // Restore ESP
199 CFI(".cfi_def_cfa_register %esp\n")
201 CFI(".cfi_adjust_cfa_offset 12\n")
203 CFI(".cfi_adjust_cfa_offset -4\n")
204 CFI(".cfi_restore %ecx\n")
206 CFI(".cfi_adjust_cfa_offset -4\n")
207 CFI(".cfi_restore %edx\n")
209 CFI(".cfi_adjust_cfa_offset -4\n")
210 CFI(".cfi_restore %eax\n")
212 CFI(".cfi_adjust_cfa_offset -4\n")
213 CFI(".cfi_restore %ebp\n")
215 CFI(".cfi_endproc\n")
218 // Same as X86CompilationCallback but also saves XMM argument registers.
219 void X86CompilationCallback_SSE(void);
223 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
224 ASMPREFIX "X86CompilationCallback_SSE:\n"
225 CFI(".cfi_startproc\n")
227 CFI(".cfi_def_cfa_offset 8\n")
228 CFI(".cfi_offset %ebp, -8\n")
229 "movl %esp, %ebp\n" // Standard prologue
230 CFI(".cfi_def_cfa_register %ebp\n")
232 CFI(".cfi_rel_offset %eax, 0\n")
233 "pushl %edx\n" // Save EAX/EDX/ECX
234 CFI(".cfi_rel_offset %edx, 4\n")
236 CFI(".cfi_rel_offset %ecx, 8\n")
237 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
238 // Save all XMM arg registers
240 // FIXME: provide frame move information for xmm registers.
241 // This can be tricky, because CFA register is ebp (unaligned)
242 // and we need to produce offsets relative to it.
243 "movaps %xmm0, (%esp)\n"
244 "movaps %xmm1, 16(%esp)\n"
245 "movaps %xmm2, 32(%esp)\n"
246 "movaps %xmm3, 48(%esp)\n"
248 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
249 "movl %eax, 4(%esp)\n"
250 "movl %ebp, (%esp)\n"
251 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
253 "movaps 48(%esp), %xmm3\n"
254 CFI(".cfi_restore %xmm3\n")
255 "movaps 32(%esp), %xmm2\n"
256 CFI(".cfi_restore %xmm2\n")
257 "movaps 16(%esp), %xmm1\n"
258 CFI(".cfi_restore %xmm1\n")
259 "movaps (%esp), %xmm0\n"
260 CFI(".cfi_restore %xmm0\n")
261 "movl %ebp, %esp\n" // Restore ESP
262 CFI(".cfi_def_cfa_register esp\n")
264 CFI(".cfi_adjust_cfa_offset 12\n")
266 CFI(".cfi_adjust_cfa_offset -4\n")
267 CFI(".cfi_restore %ecx\n")
269 CFI(".cfi_adjust_cfa_offset -4\n")
270 CFI(".cfi_restore %edx\n")
272 CFI(".cfi_adjust_cfa_offset -4\n")
273 CFI(".cfi_restore %eax\n")
275 CFI(".cfi_adjust_cfa_offset -4\n")
276 CFI(".cfi_restore %ebp\n")
278 CFI(".cfi_endproc\n")
281 void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr);
283 _declspec(naked) void X86CompilationCallback(void) {
291 mov eax, dword ptr [ebp+4]
292 mov dword ptr [esp+4], eax
293 mov dword ptr [esp], ebp
294 call X86CompilationCallback2
307 #else // Not an i386 host
308 void X86CompilationCallback() {
309 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
315 /// X86CompilationCallback - This is the target-specific function invoked by the
316 /// function stub when we did not know the real target of a call. This function
317 /// must locate the start of the stub or call site and pass it into the JIT
318 /// compiler function.
319 extern "C" void ATTRIBUTE_USED
320 X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
321 intptr_t *RetAddrLoc = &StackPtr[1];
322 assert(*RetAddrLoc == RetAddr &&
323 "Could not find return address on the stack!");
325 // It's a stub if there is an interrupt marker after the call.
326 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
328 // The call instruction should have pushed the return value onto the stack...
329 #if defined (X86_64_JIT)
330 RetAddr--; // Backtrack to the reference itself...
332 RetAddr -= 4; // Backtrack to the reference itself...
336 DOUT << "In callback! Addr=" << (void*)RetAddr
337 << " ESP=" << (void*)StackPtr
338 << ": Resolving call to function: "
339 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
342 // Sanity check to make sure this really is a call instruction.
343 #if defined (X86_64_JIT)
344 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
345 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
347 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
350 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
352 // Rewrite the call target... so that we don't end up here every time we
354 #if defined (X86_64_JIT)
356 *(intptr_t *)(RetAddr - 0xa) = NewVal;
358 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
362 // If this is a stub, rewrite the call into an unconditional branch
363 // instruction so that two return addresses are not pushed onto the stack
364 // when the requested function finally gets called. This also makes the
365 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
366 #if defined (X86_64_JIT)
367 // If the target address is within 32-bit range of the stub, use a
368 // PC-relative branch instead of loading the actual address. (This is
369 // considerably shorter than the 64-bit immediate load already there.)
370 // We assume here intptr_t is 64 bits.
371 intptr_t diff = NewVal-RetAddr+7;
372 if (diff >= -2147483648LL && diff <= 2147483647LL) {
373 *(unsigned char*)(RetAddr-0xc) = 0xE9;
374 *(intptr_t *)(RetAddr-0xb) = diff & 0xffffffff;
376 *(intptr_t *)(RetAddr - 0xa) = NewVal;
377 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
380 ((unsigned char*)RetAddr)[-1] = 0xE9;
384 // Change the return address to reexecute the call instruction...
385 #if defined (X86_64_JIT)
392 TargetJITInfo::LazyResolverFn
393 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
394 JITCompilerFunction = F;
396 #if defined (X86_32_JIT) && !defined (_MSC_VER)
397 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
403 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
404 // FIXME: support for AMD family of processors.
405 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
406 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
407 if ((EDX >> 25) & 0x1)
408 return X86CompilationCallback_SSE;
413 return X86CompilationCallback;
416 void *X86JITInfo::emitGlobalValueIndirectSym(const GlobalValue* GV, void *ptr,
417 MachineCodeEmitter &MCE) {
418 #if defined (X86_64_JIT)
419 MCE.startGVStub(GV, 8, 8);
420 MCE.emitWordLE((unsigned)(intptr_t)ptr);
421 MCE.emitWordLE((unsigned)(((intptr_t)ptr) >> 32));
423 MCE.startGVStub(GV, 4, 4);
424 MCE.emitWordLE((intptr_t)ptr);
426 return MCE.finishGVStub(GV);
429 void *X86JITInfo::emitFunctionStub(const Function* F, void *Fn,
430 MachineCodeEmitter &MCE) {
431 // Note, we cast to intptr_t here to silence a -pedantic warning that
432 // complains about casting a function pointer to a normal pointer.
433 #if defined (X86_32_JIT) && !defined (_MSC_VER)
434 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
435 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
437 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
440 #if defined (X86_64_JIT)
441 MCE.startGVStub(F, 13, 4);
442 MCE.emitByte(0x49); // REX prefix
443 MCE.emitByte(0xB8+2); // movabsq r10
444 MCE.emitWordLE((unsigned)(intptr_t)Fn);
445 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
446 MCE.emitByte(0x41); // REX prefix
447 MCE.emitByte(0xFF); // jmpq *r10
448 MCE.emitByte(2 | (4 << 3) | (3 << 6));
450 MCE.startGVStub(F, 5, 4);
452 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
454 return MCE.finishGVStub(F);
457 #if defined (X86_64_JIT)
458 MCE.startGVStub(F, 14, 4);
459 MCE.emitByte(0x49); // REX prefix
460 MCE.emitByte(0xB8+2); // movabsq r10
461 MCE.emitWordLE((unsigned)(intptr_t)Fn);
462 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
463 MCE.emitByte(0x41); // REX prefix
464 MCE.emitByte(0xFF); // callq *r10
465 MCE.emitByte(2 | (2 << 3) | (3 << 6));
467 MCE.startGVStub(F, 6, 4);
468 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
470 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
473 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
474 return MCE.finishGVStub(F);
477 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
478 /// specific basic block.
479 uintptr_t X86JITInfo::getPICJumpTableEntry(uintptr_t BB, uintptr_t Entry) {
480 #if defined(X86_64_JIT)
487 /// relocate - Before the JIT can run a block of code that has been emitted,
488 /// it must rewrite the code to contain the actual addresses of any
489 /// referenced global symbols.
490 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
491 unsigned NumRelocs, unsigned char* GOTBase) {
492 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
493 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
494 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
495 switch ((X86::RelocationType)MR->getRelocationType()) {
496 case X86::reloc_pcrel_word: {
497 // PC relative relocation, add the relocated value to the value already in
498 // memory, after we adjust it for where the PC is.
499 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
500 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
503 case X86::reloc_picrel_word: {
504 // PIC base relative relocation, add the relocated value to the value
505 // already in memory, after we adjust it for where the PIC base is.
506 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
507 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
510 case X86::reloc_absolute_word:
511 // Absolute relocation, just add the relocated value to the value already
513 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
515 case X86::reloc_absolute_dword:
516 *((intptr_t*)RelocPos) += ResultPtr;
522 char* X86JITInfo::allocateThreadLocalMemory(size_t size) {
523 #if defined(X86_32_JIT) && !defined(__APPLE__) && !defined(_MSC_VER)
527 assert(0 && "Cannot allocate thread local storage on this arch!\n");