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 "X86TargetMachine.h"
19 #include "llvm/Function.h"
20 #include "llvm/Support/Compiler.h"
21 #include "llvm/Support/ErrorHandling.h"
26 // Determine the platform we're running on
27 #if defined (__x86_64__) || defined (_M_AMD64) || defined (_M_X64)
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 // For ELF targets, use a .size and .type directive, to let tools
56 // know the extent of functions defined in assembler.
58 # define SIZE(sym) ".size " #sym ", . - " #sym "\n"
59 # define TYPE_FUNCTION(sym) ".type " #sym ", @function\n"
62 # define TYPE_FUNCTION(sym)
65 // Provide a convenient way for disabling usage of CFI directives.
66 // This is needed for old/broken assemblers (for example, gas on
67 // Darwin is pretty old and doesn't support these directives)
68 #if defined(__APPLE__)
71 // FIXME: Disable this until we really want to use it. Also, we will
72 // need to add some workarounds for compilers, which support
73 // only subset of these directives.
77 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
78 // callee saved registers, for the fastcc calling convention.
80 #if defined(X86_64_JIT)
82 // No need to save EAX/EDX for X86-64.
83 void X86CompilationCallback(void);
87 ".globl " ASMPREFIX "X86CompilationCallback\n"
88 TYPE_FUNCTION(X86CompilationCallback)
89 ASMPREFIX "X86CompilationCallback:\n"
90 CFI(".cfi_startproc\n")
93 CFI(".cfi_def_cfa_offset 16\n")
94 CFI(".cfi_offset %rbp, -16\n")
97 CFI(".cfi_def_cfa_register %rbp\n")
98 // Save all int arg registers
100 CFI(".cfi_rel_offset %rdi, 0\n")
102 CFI(".cfi_rel_offset %rsi, 8\n")
104 CFI(".cfi_rel_offset %rdx, 16\n")
106 CFI(".cfi_rel_offset %rcx, 24\n")
108 CFI(".cfi_rel_offset %r8, 32\n")
110 CFI(".cfi_rel_offset %r9, 40\n")
111 // Align stack on 16-byte boundary. ESP might not be properly aligned
112 // (8 byte) if this is called from an indirect stub.
114 // Save all XMM arg registers
116 "movaps %xmm0, (%rsp)\n"
117 "movaps %xmm1, 16(%rsp)\n"
118 "movaps %xmm2, 32(%rsp)\n"
119 "movaps %xmm3, 48(%rsp)\n"
120 "movaps %xmm4, 64(%rsp)\n"
121 "movaps %xmm5, 80(%rsp)\n"
122 "movaps %xmm6, 96(%rsp)\n"
123 "movaps %xmm7, 112(%rsp)\n"
125 "movq %rbp, %rdi\n" // Pass prev frame and return address
126 "movq 8(%rbp), %rsi\n"
127 "call " ASMPREFIX "X86CompilationCallback2\n"
128 // Restore all XMM arg registers
129 "movaps 112(%rsp), %xmm7\n"
130 "movaps 96(%rsp), %xmm6\n"
131 "movaps 80(%rsp), %xmm5\n"
132 "movaps 64(%rsp), %xmm4\n"
133 "movaps 48(%rsp), %xmm3\n"
134 "movaps 32(%rsp), %xmm2\n"
135 "movaps 16(%rsp), %xmm1\n"
136 "movaps (%rsp), %xmm0\n"
139 CFI(".cfi_def_cfa_register %rsp\n")
140 // Restore all int arg registers
142 CFI(".cfi_adjust_cfa_offset 48\n")
144 CFI(".cfi_adjust_cfa_offset -8\n")
145 CFI(".cfi_restore %r9\n")
147 CFI(".cfi_adjust_cfa_offset -8\n")
148 CFI(".cfi_restore %r8\n")
150 CFI(".cfi_adjust_cfa_offset -8\n")
151 CFI(".cfi_restore %rcx\n")
153 CFI(".cfi_adjust_cfa_offset -8\n")
154 CFI(".cfi_restore %rdx\n")
156 CFI(".cfi_adjust_cfa_offset -8\n")
157 CFI(".cfi_restore %rsi\n")
159 CFI(".cfi_adjust_cfa_offset -8\n")
160 CFI(".cfi_restore %rdi\n")
163 CFI(".cfi_adjust_cfa_offset -8\n")
164 CFI(".cfi_restore %rbp\n")
166 CFI(".cfi_endproc\n")
167 SIZE(X86CompilationCallback)
170 // No inline assembler support on this platform. The routine is in external
172 void X86CompilationCallback();
175 #elif defined (X86_32_JIT)
177 void X86CompilationCallback(void);
181 ".globl " ASMPREFIX "X86CompilationCallback\n"
182 TYPE_FUNCTION(X86CompilationCallback)
183 ASMPREFIX "X86CompilationCallback:\n"
184 CFI(".cfi_startproc\n")
186 CFI(".cfi_def_cfa_offset 8\n")
187 CFI(".cfi_offset %ebp, -8\n")
188 "movl %esp, %ebp\n" // Standard prologue
189 CFI(".cfi_def_cfa_register %ebp\n")
191 CFI(".cfi_rel_offset %eax, 0\n")
192 "pushl %edx\n" // Save EAX/EDX/ECX
193 CFI(".cfi_rel_offset %edx, 4\n")
195 CFI(".cfi_rel_offset %ecx, 8\n")
196 # if defined(__APPLE__)
197 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
200 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
201 "movl %eax, 4(%esp)\n"
202 "movl %ebp, (%esp)\n"
203 "call " ASMPREFIX "X86CompilationCallback2\n"
204 "movl %ebp, %esp\n" // Restore ESP
205 CFI(".cfi_def_cfa_register %esp\n")
207 CFI(".cfi_adjust_cfa_offset 12\n")
209 CFI(".cfi_adjust_cfa_offset -4\n")
210 CFI(".cfi_restore %ecx\n")
212 CFI(".cfi_adjust_cfa_offset -4\n")
213 CFI(".cfi_restore %edx\n")
215 CFI(".cfi_adjust_cfa_offset -4\n")
216 CFI(".cfi_restore %eax\n")
218 CFI(".cfi_adjust_cfa_offset -4\n")
219 CFI(".cfi_restore %ebp\n")
221 CFI(".cfi_endproc\n")
222 SIZE(X86CompilationCallback)
225 // Same as X86CompilationCallback but also saves XMM argument registers.
226 void X86CompilationCallback_SSE(void);
230 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
231 TYPE_FUNCTION(X86CompilationCallback_SSE)
232 ASMPREFIX "X86CompilationCallback_SSE:\n"
233 CFI(".cfi_startproc\n")
235 CFI(".cfi_def_cfa_offset 8\n")
236 CFI(".cfi_offset %ebp, -8\n")
237 "movl %esp, %ebp\n" // Standard prologue
238 CFI(".cfi_def_cfa_register %ebp\n")
240 CFI(".cfi_rel_offset %eax, 0\n")
241 "pushl %edx\n" // Save EAX/EDX/ECX
242 CFI(".cfi_rel_offset %edx, 4\n")
244 CFI(".cfi_rel_offset %ecx, 8\n")
245 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
246 // Save all XMM arg registers
248 // FIXME: provide frame move information for xmm registers.
249 // This can be tricky, because CFA register is ebp (unaligned)
250 // and we need to produce offsets relative to it.
251 "movaps %xmm0, (%esp)\n"
252 "movaps %xmm1, 16(%esp)\n"
253 "movaps %xmm2, 32(%esp)\n"
254 "movaps %xmm3, 48(%esp)\n"
256 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
257 "movl %eax, 4(%esp)\n"
258 "movl %ebp, (%esp)\n"
259 "call " ASMPREFIX "X86CompilationCallback2\n"
261 "movaps 48(%esp), %xmm3\n"
262 CFI(".cfi_restore %xmm3\n")
263 "movaps 32(%esp), %xmm2\n"
264 CFI(".cfi_restore %xmm2\n")
265 "movaps 16(%esp), %xmm1\n"
266 CFI(".cfi_restore %xmm1\n")
267 "movaps (%esp), %xmm0\n"
268 CFI(".cfi_restore %xmm0\n")
269 "movl %ebp, %esp\n" // Restore ESP
270 CFI(".cfi_def_cfa_register esp\n")
272 CFI(".cfi_adjust_cfa_offset 12\n")
274 CFI(".cfi_adjust_cfa_offset -4\n")
275 CFI(".cfi_restore %ecx\n")
277 CFI(".cfi_adjust_cfa_offset -4\n")
278 CFI(".cfi_restore %edx\n")
280 CFI(".cfi_adjust_cfa_offset -4\n")
281 CFI(".cfi_restore %eax\n")
283 CFI(".cfi_adjust_cfa_offset -4\n")
284 CFI(".cfi_restore %ebp\n")
286 CFI(".cfi_endproc\n")
287 SIZE(X86CompilationCallback_SSE)
290 void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr);
292 _declspec(naked) void X86CompilationCallback(void) {
300 mov eax, dword ptr [ebp+4]
301 mov dword ptr [esp+4], eax
302 mov dword ptr [esp], ebp
303 call X86CompilationCallback2
316 #else // Not an i386 host
317 void X86CompilationCallback() {
318 llvm_unreachable("Cannot call X86CompilationCallback() on a non-x86 arch!");
323 /// X86CompilationCallback2 - This is the target-specific function invoked by the
324 /// function stub when we did not know the real target of a call. This function
325 /// must locate the start of the stub or call site and pass it into the JIT
326 /// compiler function.
328 #if !(defined (X86_64_JIT) && defined(_MSC_VER))
329 // the following function is called only from this translation unit,
330 // unless we are under 64bit Windows with MSC, where there is
331 // no support for inline assembly
335 X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
336 intptr_t *RetAddrLoc = &StackPtr[1];
337 assert(*RetAddrLoc == RetAddr &&
338 "Could not find return address on the stack!");
340 // It's a stub if there is an interrupt marker after the call.
341 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCE;
343 // The call instruction should have pushed the return value onto the stack...
344 #if defined (X86_64_JIT)
345 RetAddr--; // Backtrack to the reference itself...
347 RetAddr -= 4; // Backtrack to the reference itself...
351 DEBUG(errs() << "In callback! Addr=" << (void*)RetAddr
352 << " ESP=" << (void*)StackPtr
353 << ": Resolving call to function: "
354 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n");
357 // Sanity check to make sure this really is a call instruction.
358 #if defined (X86_64_JIT)
359 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
360 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
362 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
365 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
367 // Rewrite the call target... so that we don't end up here every time we
369 #if defined (X86_64_JIT)
371 "X86-64 doesn't support rewriting non-stub lazy compilation calls:"
372 " the call instruction varies too much.");
374 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
378 // If this is a stub, rewrite the call into an unconditional branch
379 // instruction so that two return addresses are not pushed onto the stack
380 // when the requested function finally gets called. This also makes the
381 // 0xCE byte (interrupt) dead, so the marker doesn't effect anything.
382 #if defined (X86_64_JIT)
383 // If the target address is within 32-bit range of the stub, use a
384 // PC-relative branch instead of loading the actual address. (This is
385 // considerably shorter than the 64-bit immediate load already there.)
386 // We assume here intptr_t is 64 bits.
387 intptr_t diff = NewVal-RetAddr+7;
388 if (diff >= -2147483648LL && diff <= 2147483647LL) {
389 *(unsigned char*)(RetAddr-0xc) = 0xE9;
390 *(intptr_t *)(RetAddr-0xb) = diff & 0xffffffff;
392 *(intptr_t *)(RetAddr - 0xa) = NewVal;
393 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
396 ((unsigned char*)RetAddr)[-1] = 0xE9;
400 // Change the return address to reexecute the call instruction...
401 #if defined (X86_64_JIT)
409 TargetJITInfo::LazyResolverFn
410 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
411 JITCompilerFunction = F;
413 #if defined (X86_32_JIT) && !defined (_MSC_VER)
414 if (Subtarget->hasSSE1())
415 return X86CompilationCallback_SSE;
418 return X86CompilationCallback;
421 X86JITInfo::X86JITInfo(X86TargetMachine &tm) : TM(tm) {
422 Subtarget = &TM.getSubtarget<X86Subtarget>();
427 void *X86JITInfo::emitGlobalValueIndirectSym(const GlobalValue* GV, void *ptr,
428 JITCodeEmitter &JCE) {
429 #if defined (X86_64_JIT)
430 const unsigned Alignment = 8;
432 uint8_t *Cur = Buffer;
433 MachineCodeEmitter::emitWordLEInto(Cur, (unsigned)(intptr_t)ptr);
434 MachineCodeEmitter::emitWordLEInto(Cur, (unsigned)(((intptr_t)ptr) >> 32));
436 const unsigned Alignment = 4;
438 uint8_t *Cur = Buffer;
439 MachineCodeEmitter::emitWordLEInto(Cur, (intptr_t)ptr);
441 return JCE.allocIndirectGV(GV, Buffer, sizeof(Buffer), Alignment);
444 TargetJITInfo::StubLayout X86JITInfo::getStubLayout() {
445 // The 64-bit stub contains:
446 // movabs r10 <- 8-byte-target-address # 10 bytes
447 // call|jmp *r10 # 3 bytes
448 // The 32-bit stub contains a 5-byte call|jmp.
449 // If the stub is a call to the compilation callback, an extra byte is added
450 // to mark it as a stub.
451 StubLayout Result = {14, 4};
455 void *X86JITInfo::emitFunctionStub(const Function* F, void *Target,
456 JITCodeEmitter &JCE) {
457 // Note, we cast to intptr_t here to silence a -pedantic warning that
458 // complains about casting a function pointer to a normal pointer.
459 #if defined (X86_32_JIT) && !defined (_MSC_VER)
460 bool NotCC = (Target != (void*)(intptr_t)X86CompilationCallback &&
461 Target != (void*)(intptr_t)X86CompilationCallback_SSE);
463 bool NotCC = Target != (void*)(intptr_t)X86CompilationCallback;
465 JCE.emitAlignment(4);
466 void *Result = (void*)JCE.getCurrentPCValue();
468 #if defined (X86_64_JIT)
469 JCE.emitByte(0x49); // REX prefix
470 JCE.emitByte(0xB8+2); // movabsq r10
471 JCE.emitWordLE((unsigned)(intptr_t)Target);
472 JCE.emitWordLE((unsigned)(((intptr_t)Target) >> 32));
473 JCE.emitByte(0x41); // REX prefix
474 JCE.emitByte(0xFF); // jmpq *r10
475 JCE.emitByte(2 | (4 << 3) | (3 << 6));
478 JCE.emitWordLE((intptr_t)Target-JCE.getCurrentPCValue()-4);
483 #if defined (X86_64_JIT)
484 JCE.emitByte(0x49); // REX prefix
485 JCE.emitByte(0xB8+2); // movabsq r10
486 JCE.emitWordLE((unsigned)(intptr_t)Target);
487 JCE.emitWordLE((unsigned)(((intptr_t)Target) >> 32));
488 JCE.emitByte(0x41); // REX prefix
489 JCE.emitByte(0xFF); // callq *r10
490 JCE.emitByte(2 | (2 << 3) | (3 << 6));
492 JCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
494 JCE.emitWordLE((intptr_t)Target-JCE.getCurrentPCValue()-4);
497 // This used to use 0xCD, but that value is used by JITMemoryManager to
498 // initialize the buffer with garbage, which means it may follow a
499 // noreturn function call, confusing X86CompilationCallback2. PR 4929.
500 JCE.emitByte(0xCE); // Interrupt - Just a marker identifying the stub!
504 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
505 /// specific basic block.
506 uintptr_t X86JITInfo::getPICJumpTableEntry(uintptr_t BB, uintptr_t Entry) {
507 #if defined(X86_64_JIT)
514 /// relocate - Before the JIT can run a block of code that has been emitted,
515 /// it must rewrite the code to contain the actual addresses of any
516 /// referenced global symbols.
517 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
518 unsigned NumRelocs, unsigned char* GOTBase) {
519 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
520 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
521 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
522 switch ((X86::RelocationType)MR->getRelocationType()) {
523 case X86::reloc_pcrel_word: {
524 // PC relative relocation, add the relocated value to the value already in
525 // memory, after we adjust it for where the PC is.
526 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
527 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
530 case X86::reloc_picrel_word: {
531 // PIC base relative relocation, add the relocated value to the value
532 // already in memory, after we adjust it for where the PIC base is.
533 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
534 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
537 case X86::reloc_absolute_word:
538 case X86::reloc_absolute_word_sext:
539 // Absolute relocation, just add the relocated value to the value already
541 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
543 case X86::reloc_absolute_dword:
544 *((intptr_t*)RelocPos) += ResultPtr;
550 char* X86JITInfo::allocateThreadLocalMemory(size_t size) {
551 #if defined(X86_32_JIT) && !defined(__APPLE__) && !defined(_MSC_VER)
555 llvm_unreachable("Cannot allocate thread local storage on this arch!");