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"
25 // Determine the platform we're running on
26 #if defined (__x86_64__) || defined (_M_AMD64)
28 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
32 void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) {
33 unsigned char *OldByte = (unsigned char *)Old;
34 *OldByte++ = 0xE9; // Emit JMP opcode.
35 unsigned *OldWord = (unsigned *)OldByte;
36 unsigned NewAddr = (intptr_t)New;
37 unsigned OldAddr = (intptr_t)OldWord;
38 *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code.
42 /// JITCompilerFunction - This contains the address of the JIT function used to
43 /// compile a function lazily.
44 static TargetJITInfo::JITCompilerFn JITCompilerFunction;
46 // Get the ASMPREFIX for the current host. This is often '_'.
47 #ifndef __USER_LABEL_PREFIX__
48 #define __USER_LABEL_PREFIX__
50 #define GETASMPREFIX2(X) #X
51 #define GETASMPREFIX(X) GETASMPREFIX2(X)
52 #define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__)
54 // Check if building with -fPIC
55 #if defined(__PIC__) && __PIC__ && defined(__linux__)
56 #define ASMCALLSUFFIX "@PLT"
61 // Provide a convenient way for disabling usage of CFI directives.
62 // This is needed for old/broken assemblers (for example, gas on
63 // Darwin is pretty old and doesn't support these directives)
64 #if defined(__APPLE__)
67 // FIXME: Disable this until we really want to use it. Also, we will
68 // need to add some workarounds for compilers, which support
69 // only subset of these directives.
73 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
74 // callee saved registers, for the fastcc calling convention.
76 #if defined(X86_64_JIT)
78 // No need to save EAX/EDX for X86-64.
79 void X86CompilationCallback(void);
83 ".globl " ASMPREFIX "X86CompilationCallback\n"
84 ASMPREFIX "X86CompilationCallback:\n"
85 CFI(".cfi_startproc\n")
88 CFI(".cfi_def_cfa_offset 16\n")
89 CFI(".cfi_offset %rbp, -16\n")
92 CFI(".cfi_def_cfa_register %rbp\n")
93 // Save all int arg registers
95 CFI(".cfi_rel_offset %rdi, 0\n")
97 CFI(".cfi_rel_offset %rsi, 8\n")
99 CFI(".cfi_rel_offset %rdx, 16\n")
101 CFI(".cfi_rel_offset %rcx, 24\n")
103 CFI(".cfi_rel_offset %r8, 32\n")
105 CFI(".cfi_rel_offset %r9, 40\n")
106 // Align stack on 16-byte boundary. ESP might not be properly aligned
107 // (8 byte) if this is called from an indirect stub.
109 // Save all XMM arg registers
111 "movaps %xmm0, (%rsp)\n"
112 "movaps %xmm1, 16(%rsp)\n"
113 "movaps %xmm2, 32(%rsp)\n"
114 "movaps %xmm3, 48(%rsp)\n"
115 "movaps %xmm4, 64(%rsp)\n"
116 "movaps %xmm5, 80(%rsp)\n"
117 "movaps %xmm6, 96(%rsp)\n"
118 "movaps %xmm7, 112(%rsp)\n"
120 "movq %rbp, %rdi\n" // Pass prev frame and return address
121 "movq 8(%rbp), %rsi\n"
122 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
123 // Restore all XMM arg registers
124 "movaps 112(%rsp), %xmm7\n"
125 "movaps 96(%rsp), %xmm6\n"
126 "movaps 80(%rsp), %xmm5\n"
127 "movaps 64(%rsp), %xmm4\n"
128 "movaps 48(%rsp), %xmm3\n"
129 "movaps 32(%rsp), %xmm2\n"
130 "movaps 16(%rsp), %xmm1\n"
131 "movaps (%rsp), %xmm0\n"
134 CFI(".cfi_def_cfa_register %rsp\n")
135 // Restore all int arg registers
137 CFI(".cfi_adjust_cfa_offset 48\n")
139 CFI(".cfi_adjust_cfa_offset -8\n")
140 CFI(".cfi_restore %r9\n")
142 CFI(".cfi_adjust_cfa_offset -8\n")
143 CFI(".cfi_restore %r8\n")
145 CFI(".cfi_adjust_cfa_offset -8\n")
146 CFI(".cfi_restore %rcx\n")
148 CFI(".cfi_adjust_cfa_offset -8\n")
149 CFI(".cfi_restore %rdx\n")
151 CFI(".cfi_adjust_cfa_offset -8\n")
152 CFI(".cfi_restore %rsi\n")
154 CFI(".cfi_adjust_cfa_offset -8\n")
155 CFI(".cfi_restore %rdi\n")
158 CFI(".cfi_adjust_cfa_offset -8\n")
159 CFI(".cfi_restore %rbp\n")
161 CFI(".cfi_endproc\n")
164 // No inline assembler support on this platform. The routine is in external
166 void X86CompilationCallback();
169 #elif defined (X86_32_JIT)
171 void X86CompilationCallback(void);
175 ".globl " ASMPREFIX "X86CompilationCallback\n"
176 ASMPREFIX "X86CompilationCallback:\n"
177 CFI(".cfi_startproc\n")
179 CFI(".cfi_def_cfa_offset 8\n")
180 CFI(".cfi_offset %ebp, -8\n")
181 "movl %esp, %ebp\n" // Standard prologue
182 CFI(".cfi_def_cfa_register %ebp\n")
184 CFI(".cfi_rel_offset %eax, 0\n")
185 "pushl %edx\n" // Save EAX/EDX/ECX
186 CFI(".cfi_rel_offset %edx, 4\n")
188 CFI(".cfi_rel_offset %ecx, 8\n")
189 # if defined(__APPLE__)
190 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
193 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
194 "movl %eax, 4(%esp)\n"
195 "movl %ebp, (%esp)\n"
196 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
197 "movl %ebp, %esp\n" // Restore ESP
198 CFI(".cfi_def_cfa_register %esp\n")
200 CFI(".cfi_adjust_cfa_offset 12\n")
202 CFI(".cfi_adjust_cfa_offset -4\n")
203 CFI(".cfi_restore %ecx\n")
205 CFI(".cfi_adjust_cfa_offset -4\n")
206 CFI(".cfi_restore %edx\n")
208 CFI(".cfi_adjust_cfa_offset -4\n")
209 CFI(".cfi_restore %eax\n")
211 CFI(".cfi_adjust_cfa_offset -4\n")
212 CFI(".cfi_restore %ebp\n")
214 CFI(".cfi_endproc\n")
217 // Same as X86CompilationCallback but also saves XMM argument registers.
218 void X86CompilationCallback_SSE(void);
222 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
223 ASMPREFIX "X86CompilationCallback_SSE:\n"
224 CFI(".cfi_startproc\n")
226 CFI(".cfi_def_cfa_offset 8\n")
227 CFI(".cfi_offset %ebp, -8\n")
228 "movl %esp, %ebp\n" // Standard prologue
229 CFI(".cfi_def_cfa_register %ebp\n")
231 CFI(".cfi_rel_offset %eax, 0\n")
232 "pushl %edx\n" // Save EAX/EDX/ECX
233 CFI(".cfi_rel_offset %edx, 4\n")
235 CFI(".cfi_rel_offset %ecx, 8\n")
236 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
237 // Save all XMM arg registers
239 // FIXME: provide frame move information for xmm registers.
240 // This can be tricky, because CFA register is ebp (unaligned)
241 // and we need to produce offsets relative to it.
242 "movaps %xmm0, (%esp)\n"
243 "movaps %xmm1, 16(%esp)\n"
244 "movaps %xmm2, 32(%esp)\n"
245 "movaps %xmm3, 48(%esp)\n"
247 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
248 "movl %eax, 4(%esp)\n"
249 "movl %ebp, (%esp)\n"
250 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
252 "movaps 48(%esp), %xmm3\n"
253 CFI(".cfi_restore %xmm3\n")
254 "movaps 32(%esp), %xmm2\n"
255 CFI(".cfi_restore %xmm2\n")
256 "movaps 16(%esp), %xmm1\n"
257 CFI(".cfi_restore %xmm1\n")
258 "movaps (%esp), %xmm0\n"
259 CFI(".cfi_restore %xmm0\n")
260 "movl %ebp, %esp\n" // Restore ESP
261 CFI(".cfi_def_cfa_register esp\n")
263 CFI(".cfi_adjust_cfa_offset 12\n")
265 CFI(".cfi_adjust_cfa_offset -4\n")
266 CFI(".cfi_restore %ecx\n")
268 CFI(".cfi_adjust_cfa_offset -4\n")
269 CFI(".cfi_restore %edx\n")
271 CFI(".cfi_adjust_cfa_offset -4\n")
272 CFI(".cfi_restore %eax\n")
274 CFI(".cfi_adjust_cfa_offset -4\n")
275 CFI(".cfi_restore %ebp\n")
277 CFI(".cfi_endproc\n")
280 void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr);
282 _declspec(naked) void X86CompilationCallback(void) {
290 mov eax, dword ptr [ebp+4]
291 mov dword ptr [esp+4], eax
292 mov dword ptr [esp], ebp
293 call X86CompilationCallback2
306 #else // Not an i386 host
307 void X86CompilationCallback() {
308 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
314 /// X86CompilationCallback - This is the target-specific function invoked by the
315 /// function stub when we did not know the real target of a call. This function
316 /// must locate the start of the stub or call site and pass it into the JIT
317 /// compiler function.
318 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
319 intptr_t *RetAddrLoc = &StackPtr[1];
320 assert(*RetAddrLoc == RetAddr &&
321 "Could not find return address on the stack!");
323 // It's a stub if there is an interrupt marker after the call.
324 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
326 // The call instruction should have pushed the return value onto the stack...
327 #if defined (X86_64_JIT)
328 RetAddr--; // Backtrack to the reference itself...
330 RetAddr -= 4; // Backtrack to the reference itself...
334 DOUT << "In callback! Addr=" << (void*)RetAddr
335 << " ESP=" << (void*)StackPtr
336 << ": Resolving call to function: "
337 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
340 // Sanity check to make sure this really is a call instruction.
341 #if defined (X86_64_JIT)
342 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
343 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
345 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
348 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
350 // Rewrite the call target... so that we don't end up here every time we
352 #if defined (X86_64_JIT)
353 *(intptr_t *)(RetAddr - 0xa) = NewVal;
355 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
359 // If this is a stub, rewrite the call into an unconditional branch
360 // instruction so that two return addresses are not pushed onto the stack
361 // when the requested function finally gets called. This also makes the
362 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
363 #if defined (X86_64_JIT)
364 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
366 ((unsigned char*)RetAddr)[-1] = 0xE9;
370 // Change the return address to reexecute the call instruction...
371 #if defined (X86_64_JIT)
378 TargetJITInfo::LazyResolverFn
379 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
380 JITCompilerFunction = F;
382 #if defined (X86_32_JIT) && !defined (_MSC_VER)
383 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
389 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
390 // FIXME: support for AMD family of processors.
391 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
392 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
393 if ((EDX >> 25) & 0x1)
394 return X86CompilationCallback_SSE;
399 return X86CompilationCallback;
402 void *X86JITInfo::emitGlobalValueLazyPtr(const GlobalValue* GV, void *ptr,
403 MachineCodeEmitter &MCE) {
404 #if defined (X86_64_JIT)
405 MCE.startFunctionStub(GV, 8, 8);
406 MCE.emitWordLE((unsigned)(intptr_t)ptr);
407 MCE.emitWordLE((unsigned)(((intptr_t)ptr) >> 32));
409 MCE.startFunctionStub(GV, 4, 4);
410 MCE.emitWordLE((intptr_t)ptr);
412 return MCE.finishFunctionStub(GV);
415 void *X86JITInfo::emitFunctionStub(const Function* F, void *Fn,
416 MachineCodeEmitter &MCE) {
417 // Note, we cast to intptr_t here to silence a -pedantic warning that
418 // complains about casting a function pointer to a normal pointer.
419 #if defined (X86_32_JIT) && !defined (_MSC_VER)
420 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
421 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
423 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
426 #if defined (X86_64_JIT)
427 MCE.startFunctionStub(F, 13, 4);
428 MCE.emitByte(0x49); // REX prefix
429 MCE.emitByte(0xB8+2); // movabsq r10
430 MCE.emitWordLE((unsigned)(intptr_t)Fn);
431 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
432 MCE.emitByte(0x41); // REX prefix
433 MCE.emitByte(0xFF); // jmpq *r10
434 MCE.emitByte(2 | (4 << 3) | (3 << 6));
436 MCE.startFunctionStub(F, 5, 4);
438 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
440 return MCE.finishFunctionStub(F);
443 #if defined (X86_64_JIT)
444 MCE.startFunctionStub(F, 14, 4);
445 MCE.emitByte(0x49); // REX prefix
446 MCE.emitByte(0xB8+2); // movabsq r10
447 MCE.emitWordLE((unsigned)(intptr_t)Fn);
448 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
449 MCE.emitByte(0x41); // REX prefix
450 MCE.emitByte(0xFF); // callq *r10
451 MCE.emitByte(2 | (2 << 3) | (3 << 6));
453 MCE.startFunctionStub(F, 6, 4);
454 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
456 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
459 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
460 return MCE.finishFunctionStub(F);
463 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
464 /// specific basic block.
465 intptr_t X86JITInfo::getPICJumpTableEntry(intptr_t BB, intptr_t Entry) {
466 #if defined(X86_64_JIT)
473 /// relocate - Before the JIT can run a block of code that has been emitted,
474 /// it must rewrite the code to contain the actual addresses of any
475 /// referenced global symbols.
476 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
477 unsigned NumRelocs, unsigned char* GOTBase) {
478 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
479 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
480 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
481 switch ((X86::RelocationType)MR->getRelocationType()) {
482 case X86::reloc_pcrel_word: {
483 // PC relative relocation, add the relocated value to the value already in
484 // memory, after we adjust it for where the PC is.
485 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
486 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
489 case X86::reloc_picrel_word: {
490 // PIC base relative relocation, add the relocated value to the value
491 // already in memory, after we adjust it for where the PIC base is.
492 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
493 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
496 case X86::reloc_absolute_word:
497 // Absolute relocation, just add the relocated value to the value already
499 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
501 case X86::reloc_absolute_dword:
502 *((intptr_t*)RelocPos) += ResultPtr;