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 // Determine the platform we're running on
25 #if defined (__x86_64__) || defined (_M_AMD64)
27 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
31 void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) {
32 unsigned char *OldByte = (unsigned char *)Old;
33 *OldByte++ = 0xE9; // Emit JMP opcode.
34 unsigned *OldWord = (unsigned *)OldByte;
35 unsigned NewAddr = (intptr_t)New;
36 unsigned OldAddr = (intptr_t)OldWord;
37 *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code.
41 /// JITCompilerFunction - This contains the address of the JIT function used to
42 /// compile a function lazily.
43 static TargetJITInfo::JITCompilerFn JITCompilerFunction;
45 // Get the ASMPREFIX for the current host. This is often '_'.
46 #ifndef __USER_LABEL_PREFIX__
47 #define __USER_LABEL_PREFIX__
49 #define GETASMPREFIX2(X) #X
50 #define GETASMPREFIX(X) GETASMPREFIX2(X)
51 #define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__)
53 // Provide a convenient way for disabling usage of CFI directives.
54 // This is needed for old/broken assemblers (for example, gas on
55 // Darwin is pretty old and doesn't support these directives)
56 #if defined(__APPLE__)
59 // FIXME: Disable this until we really want to use it. Also, we will
60 // need to add some workarounds for compilers, which support
61 // only subset of these directives.
65 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
66 // callee saved registers, for the fastcc calling convention.
68 #if defined(X86_64_JIT)
70 // No need to save EAX/EDX for X86-64.
71 void X86CompilationCallback(void);
75 ".globl " ASMPREFIX "X86CompilationCallback\n"
76 ASMPREFIX "X86CompilationCallback:\n"
77 CFI(".cfi_startproc\n")
80 CFI(".cfi_def_cfa_offset 16\n")
81 CFI(".cfi_offset %rbp, -16\n")
84 CFI(".cfi_def_cfa_register %rbp\n")
85 // Save all int arg registers
87 CFI(".cfi_rel_offset %rdi, 0\n")
89 CFI(".cfi_rel_offset %rsi, 8\n")
91 CFI(".cfi_rel_offset %rdx, 16\n")
93 CFI(".cfi_rel_offset %rcx, 24\n")
95 CFI(".cfi_rel_offset %r8, 32\n")
97 CFI(".cfi_rel_offset %r9, 40\n")
98 // Align stack on 16-byte boundary. ESP might not be properly aligned
99 // (8 byte) if this is called from an indirect stub.
101 // Save all XMM arg registers
103 "movaps %xmm0, (%rsp)\n"
104 "movaps %xmm1, 16(%rsp)\n"
105 "movaps %xmm2, 32(%rsp)\n"
106 "movaps %xmm3, 48(%rsp)\n"
107 "movaps %xmm4, 64(%rsp)\n"
108 "movaps %xmm5, 80(%rsp)\n"
109 "movaps %xmm6, 96(%rsp)\n"
110 "movaps %xmm7, 112(%rsp)\n"
112 "movq %rbp, %rdi\n" // Pass prev frame and return address
113 "movq 8(%rbp), %rsi\n"
114 "call " ASMPREFIX "X86CompilationCallback2\n"
115 // Restore all XMM arg registers
116 "movaps 112(%rsp), %xmm7\n"
117 "movaps 96(%rsp), %xmm6\n"
118 "movaps 80(%rsp), %xmm5\n"
119 "movaps 64(%rsp), %xmm4\n"
120 "movaps 48(%rsp), %xmm3\n"
121 "movaps 32(%rsp), %xmm2\n"
122 "movaps 16(%rsp), %xmm1\n"
123 "movaps (%rsp), %xmm0\n"
126 CFI(".cfi_def_cfa_register %rsp\n")
127 // Restore all int arg registers
129 CFI(".cfi_adjust_cfa_offset 48\n")
131 CFI(".cfi_adjust_cfa_offset -8\n")
132 CFI(".cfi_restore %r9\n")
134 CFI(".cfi_adjust_cfa_offset -8\n")
135 CFI(".cfi_restore %r8\n")
137 CFI(".cfi_adjust_cfa_offset -8\n")
138 CFI(".cfi_restore %rcx\n")
140 CFI(".cfi_adjust_cfa_offset -8\n")
141 CFI(".cfi_restore %rdx\n")
143 CFI(".cfi_adjust_cfa_offset -8\n")
144 CFI(".cfi_restore %rsi\n")
146 CFI(".cfi_adjust_cfa_offset -8\n")
147 CFI(".cfi_restore %rdi\n")
150 CFI(".cfi_adjust_cfa_offset -8\n")
151 CFI(".cfi_restore %rbp\n")
153 CFI(".cfi_endproc\n")
156 // No inline assembler support on this platform. The routine is in external
158 void X86CompilationCallback();
161 #elif defined (X86_32_JIT)
163 void X86CompilationCallback(void);
167 ".globl " ASMPREFIX "X86CompilationCallback\n"
168 ASMPREFIX "X86CompilationCallback:\n"
169 CFI(".cfi_startproc\n")
171 CFI(".cfi_def_cfa_offset 8\n")
172 CFI(".cfi_offset %ebp, -8\n")
173 "movl %esp, %ebp\n" // Standard prologue
174 CFI(".cfi_def_cfa_register %ebp\n")
176 CFI(".cfi_rel_offset %eax, 0\n")
177 "pushl %edx\n" // Save EAX/EDX/ECX
178 CFI(".cfi_rel_offset %edx, 4\n")
180 CFI(".cfi_rel_offset %ecx, 8\n")
181 # if defined(__APPLE__)
182 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
185 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
186 "movl %eax, 4(%esp)\n"
187 "movl %ebp, (%esp)\n"
188 "call " ASMPREFIX "X86CompilationCallback2\n"
189 "movl %ebp, %esp\n" // Restore ESP
190 CFI(".cfi_def_cfa_register %esp\n")
192 CFI(".cfi_adjust_cfa_offset 12\n")
194 CFI(".cfi_adjust_cfa_offset -4\n")
195 CFI(".cfi_restore %ecx\n")
197 CFI(".cfi_adjust_cfa_offset -4\n")
198 CFI(".cfi_restore %edx\n")
200 CFI(".cfi_adjust_cfa_offset -4\n")
201 CFI(".cfi_restore %eax\n")
203 CFI(".cfi_adjust_cfa_offset -4\n")
204 CFI(".cfi_restore %ebp\n")
206 CFI(".cfi_endproc\n")
209 // Same as X86CompilationCallback but also saves XMM argument registers.
210 void X86CompilationCallback_SSE(void);
214 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
215 ASMPREFIX "X86CompilationCallback_SSE:\n"
216 CFI(".cfi_startproc\n")
218 CFI(".cfi_def_cfa_offset 8\n")
219 CFI(".cfi_offset %ebp, -8\n")
220 "movl %esp, %ebp\n" // Standard prologue
221 CFI(".cfi_def_cfa_register %ebp\n")
223 CFI(".cfi_rel_offset %eax, 0\n")
224 "pushl %edx\n" // Save EAX/EDX/ECX
225 CFI(".cfi_rel_offset %edx, 4\n")
227 CFI(".cfi_rel_offset %ecx, 8\n")
228 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
229 // Save all XMM arg registers
231 // FIXME: provide frame move information for xmm registers.
232 // This can be tricky, because CFA register is ebp (unaligned)
233 // and we need to produce offsets relative to it.
234 "movaps %xmm0, (%esp)\n"
235 "movaps %xmm1, 16(%esp)\n"
236 "movaps %xmm2, 32(%esp)\n"
237 "movaps %xmm3, 48(%esp)\n"
239 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
240 "movl %eax, 4(%esp)\n"
241 "movl %ebp, (%esp)\n"
242 "call " ASMPREFIX "X86CompilationCallback2\n"
244 "movaps 48(%esp), %xmm3\n"
245 CFI(".cfi_restore %xmm3\n")
246 "movaps 32(%esp), %xmm2\n"
247 CFI(".cfi_restore %xmm2\n")
248 "movaps 16(%esp), %xmm1\n"
249 CFI(".cfi_restore %xmm1\n")
250 "movaps (%esp), %xmm0\n"
251 CFI(".cfi_restore %xmm0\n")
252 "movl %ebp, %esp\n" // Restore ESP
253 CFI(".cfi_def_cfa_register esp\n")
255 CFI(".cfi_adjust_cfa_offset 12\n")
257 CFI(".cfi_adjust_cfa_offset -4\n")
258 CFI(".cfi_restore %ecx\n")
260 CFI(".cfi_adjust_cfa_offset -4\n")
261 CFI(".cfi_restore %edx\n")
263 CFI(".cfi_adjust_cfa_offset -4\n")
264 CFI(".cfi_restore %eax\n")
266 CFI(".cfi_adjust_cfa_offset -4\n")
267 CFI(".cfi_restore %ebp\n")
269 CFI(".cfi_endproc\n")
272 void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr);
274 _declspec(naked) void X86CompilationCallback(void) {
282 mov eax, dword ptr [ebp+4]
283 mov dword ptr [esp+4], eax
284 mov dword ptr [esp], ebp
285 call X86CompilationCallback2
298 #else // Not an i386 host
299 void X86CompilationCallback() {
300 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
306 /// X86CompilationCallback - This is the target-specific function invoked by the
307 /// function stub when we did not know the real target of a call. This function
308 /// must locate the start of the stub or call site and pass it into the JIT
309 /// compiler function.
310 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
311 intptr_t *RetAddrLoc = &StackPtr[1];
312 assert(*RetAddrLoc == RetAddr &&
313 "Could not find return address on the stack!");
315 // It's a stub if there is an interrupt marker after the call.
316 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
318 // The call instruction should have pushed the return value onto the stack...
319 #if defined (X86_64_JIT)
320 RetAddr--; // Backtrack to the reference itself...
322 RetAddr -= 4; // Backtrack to the reference itself...
326 DOUT << "In callback! Addr=" << (void*)RetAddr
327 << " ESP=" << (void*)StackPtr
328 << ": Resolving call to function: "
329 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
332 // Sanity check to make sure this really is a call instruction.
333 #if defined (X86_64_JIT)
334 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
335 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
337 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
340 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
342 // Rewrite the call target... so that we don't end up here every time we
344 #if defined (X86_64_JIT)
345 *(intptr_t *)(RetAddr - 0xa) = NewVal;
347 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
351 // If this is a stub, rewrite the call into an unconditional branch
352 // instruction so that two return addresses are not pushed onto the stack
353 // when the requested function finally gets called. This also makes the
354 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
355 #if defined (X86_64_JIT)
356 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
358 ((unsigned char*)RetAddr)[-1] = 0xE9;
362 // Change the return address to reexecute the call instruction...
363 #if defined (X86_64_JIT)
370 TargetJITInfo::LazyResolverFn
371 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
372 JITCompilerFunction = F;
374 #if defined (X86_32_JIT) && !defined (_MSC_VER)
375 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
381 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
382 // FIXME: support for AMD family of processors.
383 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
384 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
385 if ((EDX >> 25) & 0x1)
386 return X86CompilationCallback_SSE;
391 return X86CompilationCallback;
394 void *X86JITInfo::emitGlobalValueLazyPtr(void *GV, MachineCodeEmitter &MCE) {
395 #if defined (X86_64_JIT)
396 MCE.startFunctionStub(8, 8);
397 MCE.emitWordLE(((unsigned *)&GV)[0]);
398 MCE.emitWordLE(((unsigned *)&GV)[1]);
400 MCE.startFunctionStub(4, 4);
401 MCE.emitWordLE((intptr_t)GV);
403 return MCE.finishFunctionStub(0);
406 void *X86JITInfo::emitFunctionStub(void *Fn, MachineCodeEmitter &MCE) {
407 // Note, we cast to intptr_t here to silence a -pedantic warning that
408 // complains about casting a function pointer to a normal pointer.
409 #if defined (X86_32_JIT) && !defined (_MSC_VER)
410 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
411 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
413 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
416 #if defined (X86_64_JIT)
417 MCE.startFunctionStub(13, 4);
418 MCE.emitByte(0x49); // REX prefix
419 MCE.emitByte(0xB8+2); // movabsq r10
420 MCE.emitWordLE(((unsigned *)&Fn)[0]);
421 MCE.emitWordLE(((unsigned *)&Fn)[1]);
422 MCE.emitByte(0x41); // REX prefix
423 MCE.emitByte(0xFF); // jmpq *r10
424 MCE.emitByte(2 | (4 << 3) | (3 << 6));
426 MCE.startFunctionStub(5, 4);
428 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
430 return MCE.finishFunctionStub(0);
433 #if defined (X86_64_JIT)
434 MCE.startFunctionStub(14, 4);
435 MCE.emitByte(0x49); // REX prefix
436 MCE.emitByte(0xB8+2); // movabsq r10
437 MCE.emitWordLE(((unsigned *)&Fn)[0]);
438 MCE.emitWordLE(((unsigned *)&Fn)[1]);
439 MCE.emitByte(0x41); // REX prefix
440 MCE.emitByte(0xFF); // callq *r10
441 MCE.emitByte(2 | (2 << 3) | (3 << 6));
443 MCE.startFunctionStub(6, 4);
444 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
446 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
449 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
450 return MCE.finishFunctionStub(0);
453 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
454 /// specific basic block.
455 intptr_t X86JITInfo::getPICJumpTableEntry(intptr_t BB, intptr_t Entry) {
459 /// relocate - Before the JIT can run a block of code that has been emitted,
460 /// it must rewrite the code to contain the actual addresses of any
461 /// referenced global symbols.
462 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
463 unsigned NumRelocs, unsigned char* GOTBase) {
464 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
465 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
466 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
467 switch ((X86::RelocationType)MR->getRelocationType()) {
468 case X86::reloc_pcrel_word: {
469 // PC relative relocation, add the relocated value to the value already in
470 // memory, after we adjust it for where the PC is.
471 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
472 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
475 case X86::reloc_picrel_word: {
476 // PIC base relative relocation, add the relocated value to the value
477 // already in memory, after we adjust it for where the PIC base is.
478 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
479 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
482 case X86::reloc_absolute_word:
483 // Absolute relocation, just add the relocated value to the value already
485 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
487 case X86::reloc_absolute_dword:
488 *((intptr_t*)RelocPos) += ResultPtr;