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 // Provide a convenient way for disabling usage of CFI directives.
55 // This is needed for old/broken assemblers (for example, gas on
56 // Darwin is pretty old and doesn't support these directives)
57 #if defined(__APPLE__)
60 // FIXME: Disable this until we really want to use it. Also, we will
61 // need to add some workarounds for compilers, which support
62 // only subset of these directives.
66 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
67 // callee saved registers, for the fastcc calling convention.
69 #if defined(X86_64_JIT)
71 // No need to save EAX/EDX for X86-64.
72 void X86CompilationCallback(void);
76 ".globl " ASMPREFIX "X86CompilationCallback\n"
77 ASMPREFIX "X86CompilationCallback:\n"
78 CFI(".cfi_startproc\n")
81 CFI(".cfi_def_cfa_offset 16\n")
82 CFI(".cfi_offset %rbp, -16\n")
85 CFI(".cfi_def_cfa_register %rbp\n")
86 // Save all int arg registers
88 CFI(".cfi_rel_offset %rdi, 0\n")
90 CFI(".cfi_rel_offset %rsi, 8\n")
92 CFI(".cfi_rel_offset %rdx, 16\n")
94 CFI(".cfi_rel_offset %rcx, 24\n")
96 CFI(".cfi_rel_offset %r8, 32\n")
98 CFI(".cfi_rel_offset %r9, 40\n")
99 // Align stack on 16-byte boundary. ESP might not be properly aligned
100 // (8 byte) if this is called from an indirect stub.
102 // Save all XMM arg registers
104 "movaps %xmm0, (%rsp)\n"
105 "movaps %xmm1, 16(%rsp)\n"
106 "movaps %xmm2, 32(%rsp)\n"
107 "movaps %xmm3, 48(%rsp)\n"
108 "movaps %xmm4, 64(%rsp)\n"
109 "movaps %xmm5, 80(%rsp)\n"
110 "movaps %xmm6, 96(%rsp)\n"
111 "movaps %xmm7, 112(%rsp)\n"
113 "movq %rbp, %rdi\n" // Pass prev frame and return address
114 "movq 8(%rbp), %rsi\n"
115 "call " ASMPREFIX "X86CompilationCallback2\n"
116 // Restore all XMM arg registers
117 "movaps 112(%rsp), %xmm7\n"
118 "movaps 96(%rsp), %xmm6\n"
119 "movaps 80(%rsp), %xmm5\n"
120 "movaps 64(%rsp), %xmm4\n"
121 "movaps 48(%rsp), %xmm3\n"
122 "movaps 32(%rsp), %xmm2\n"
123 "movaps 16(%rsp), %xmm1\n"
124 "movaps (%rsp), %xmm0\n"
127 CFI(".cfi_def_cfa_register %rsp\n")
128 // Restore all int arg registers
130 CFI(".cfi_adjust_cfa_offset 48\n")
132 CFI(".cfi_adjust_cfa_offset -8\n")
133 CFI(".cfi_restore %r9\n")
135 CFI(".cfi_adjust_cfa_offset -8\n")
136 CFI(".cfi_restore %r8\n")
138 CFI(".cfi_adjust_cfa_offset -8\n")
139 CFI(".cfi_restore %rcx\n")
141 CFI(".cfi_adjust_cfa_offset -8\n")
142 CFI(".cfi_restore %rdx\n")
144 CFI(".cfi_adjust_cfa_offset -8\n")
145 CFI(".cfi_restore %rsi\n")
147 CFI(".cfi_adjust_cfa_offset -8\n")
148 CFI(".cfi_restore %rdi\n")
151 CFI(".cfi_adjust_cfa_offset -8\n")
152 CFI(".cfi_restore %rbp\n")
154 CFI(".cfi_endproc\n")
157 // No inline assembler support on this platform. The routine is in external
159 void X86CompilationCallback();
162 #elif defined (X86_32_JIT)
164 void X86CompilationCallback(void);
168 ".globl " ASMPREFIX "X86CompilationCallback\n"
169 ASMPREFIX "X86CompilationCallback:\n"
170 CFI(".cfi_startproc\n")
172 CFI(".cfi_def_cfa_offset 8\n")
173 CFI(".cfi_offset %ebp, -8\n")
174 "movl %esp, %ebp\n" // Standard prologue
175 CFI(".cfi_def_cfa_register %ebp\n")
177 CFI(".cfi_rel_offset %eax, 0\n")
178 "pushl %edx\n" // Save EAX/EDX/ECX
179 CFI(".cfi_rel_offset %edx, 4\n")
181 CFI(".cfi_rel_offset %ecx, 8\n")
182 # if defined(__APPLE__)
183 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
186 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
187 "movl %eax, 4(%esp)\n"
188 "movl %ebp, (%esp)\n"
189 "call " ASMPREFIX "X86CompilationCallback2\n"
190 "movl %ebp, %esp\n" // Restore ESP
191 CFI(".cfi_def_cfa_register %esp\n")
193 CFI(".cfi_adjust_cfa_offset 12\n")
195 CFI(".cfi_adjust_cfa_offset -4\n")
196 CFI(".cfi_restore %ecx\n")
198 CFI(".cfi_adjust_cfa_offset -4\n")
199 CFI(".cfi_restore %edx\n")
201 CFI(".cfi_adjust_cfa_offset -4\n")
202 CFI(".cfi_restore %eax\n")
204 CFI(".cfi_adjust_cfa_offset -4\n")
205 CFI(".cfi_restore %ebp\n")
207 CFI(".cfi_endproc\n")
210 // Same as X86CompilationCallback but also saves XMM argument registers.
211 void X86CompilationCallback_SSE(void);
215 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
216 ASMPREFIX "X86CompilationCallback_SSE:\n"
217 CFI(".cfi_startproc\n")
219 CFI(".cfi_def_cfa_offset 8\n")
220 CFI(".cfi_offset %ebp, -8\n")
221 "movl %esp, %ebp\n" // Standard prologue
222 CFI(".cfi_def_cfa_register %ebp\n")
224 CFI(".cfi_rel_offset %eax, 0\n")
225 "pushl %edx\n" // Save EAX/EDX/ECX
226 CFI(".cfi_rel_offset %edx, 4\n")
228 CFI(".cfi_rel_offset %ecx, 8\n")
229 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
230 // Save all XMM arg registers
232 // FIXME: provide frame move information for xmm registers.
233 // This can be tricky, because CFA register is ebp (unaligned)
234 // and we need to produce offsets relative to it.
235 "movaps %xmm0, (%esp)\n"
236 "movaps %xmm1, 16(%esp)\n"
237 "movaps %xmm2, 32(%esp)\n"
238 "movaps %xmm3, 48(%esp)\n"
240 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
241 "movl %eax, 4(%esp)\n"
242 "movl %ebp, (%esp)\n"
243 "call " ASMPREFIX "X86CompilationCallback2\n"
245 "movaps 48(%esp), %xmm3\n"
246 CFI(".cfi_restore %xmm3\n")
247 "movaps 32(%esp), %xmm2\n"
248 CFI(".cfi_restore %xmm2\n")
249 "movaps 16(%esp), %xmm1\n"
250 CFI(".cfi_restore %xmm1\n")
251 "movaps (%esp), %xmm0\n"
252 CFI(".cfi_restore %xmm0\n")
253 "movl %ebp, %esp\n" // Restore ESP
254 CFI(".cfi_def_cfa_register esp\n")
256 CFI(".cfi_adjust_cfa_offset 12\n")
258 CFI(".cfi_adjust_cfa_offset -4\n")
259 CFI(".cfi_restore %ecx\n")
261 CFI(".cfi_adjust_cfa_offset -4\n")
262 CFI(".cfi_restore %edx\n")
264 CFI(".cfi_adjust_cfa_offset -4\n")
265 CFI(".cfi_restore %eax\n")
267 CFI(".cfi_adjust_cfa_offset -4\n")
268 CFI(".cfi_restore %ebp\n")
270 CFI(".cfi_endproc\n")
273 void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr);
275 _declspec(naked) void X86CompilationCallback(void) {
283 mov eax, dword ptr [ebp+4]
284 mov dword ptr [esp+4], eax
285 mov dword ptr [esp], ebp
286 call X86CompilationCallback2
299 #else // Not an i386 host
300 void X86CompilationCallback() {
301 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
307 /// X86CompilationCallback - This is the target-specific function invoked by the
308 /// function stub when we did not know the real target of a call. This function
309 /// must locate the start of the stub or call site and pass it into the JIT
310 /// compiler function.
311 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
312 intptr_t *RetAddrLoc = &StackPtr[1];
313 assert(*RetAddrLoc == RetAddr &&
314 "Could not find return address on the stack!");
316 // It's a stub if there is an interrupt marker after the call.
317 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
319 // The call instruction should have pushed the return value onto the stack...
320 #if defined (X86_64_JIT)
321 RetAddr--; // Backtrack to the reference itself...
323 RetAddr -= 4; // Backtrack to the reference itself...
327 DOUT << "In callback! Addr=" << (void*)RetAddr
328 << " ESP=" << (void*)StackPtr
329 << ": Resolving call to function: "
330 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
333 // Sanity check to make sure this really is a call instruction.
334 #if defined (X86_64_JIT)
335 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
336 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
338 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
341 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
343 // Rewrite the call target... so that we don't end up here every time we
345 #if defined (X86_64_JIT)
346 *(intptr_t *)(RetAddr - 0xa) = NewVal;
348 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
352 // If this is a stub, rewrite the call into an unconditional branch
353 // instruction so that two return addresses are not pushed onto the stack
354 // when the requested function finally gets called. This also makes the
355 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
356 #if defined (X86_64_JIT)
357 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
359 ((unsigned char*)RetAddr)[-1] = 0xE9;
363 // Change the return address to reexecute the call instruction...
364 #if defined (X86_64_JIT)
371 TargetJITInfo::LazyResolverFn
372 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
373 JITCompilerFunction = F;
375 #if defined (X86_32_JIT) && !defined (_MSC_VER)
376 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
382 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
383 // FIXME: support for AMD family of processors.
384 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
385 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
386 if ((EDX >> 25) & 0x1)
387 return X86CompilationCallback_SSE;
392 return X86CompilationCallback;
395 void *X86JITInfo::emitGlobalValueLazyPtr(const GlobalValue* GV, void *ptr,
396 MachineCodeEmitter &MCE) {
397 #if defined (X86_64_JIT)
398 MCE.startFunctionStub(GV, 8, 8);
399 MCE.emitWordLE((unsigned)(intptr_t)ptr);
400 MCE.emitWordLE((unsigned)(((intptr_t)ptr) >> 32));
402 MCE.startFunctionStub(GV, 4, 4);
403 MCE.emitWordLE((intptr_t)ptr);
405 return MCE.finishFunctionStub(GV);
408 void *X86JITInfo::emitFunctionStub(const Function* F, void *Fn,
409 MachineCodeEmitter &MCE) {
410 // Note, we cast to intptr_t here to silence a -pedantic warning that
411 // complains about casting a function pointer to a normal pointer.
412 #if defined (X86_32_JIT) && !defined (_MSC_VER)
413 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
414 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
416 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
419 #if defined (X86_64_JIT)
420 MCE.startFunctionStub(F, 13, 4);
421 MCE.emitByte(0x49); // REX prefix
422 MCE.emitByte(0xB8+2); // movabsq r10
423 MCE.emitWordLE((unsigned)(intptr_t)Fn);
424 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
425 MCE.emitByte(0x41); // REX prefix
426 MCE.emitByte(0xFF); // jmpq *r10
427 MCE.emitByte(2 | (4 << 3) | (3 << 6));
429 MCE.startFunctionStub(F, 5, 4);
431 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
433 return MCE.finishFunctionStub(F);
436 #if defined (X86_64_JIT)
437 MCE.startFunctionStub(F, 14, 4);
438 MCE.emitByte(0x49); // REX prefix
439 MCE.emitByte(0xB8+2); // movabsq r10
440 MCE.emitWordLE((unsigned)(intptr_t)Fn);
441 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
442 MCE.emitByte(0x41); // REX prefix
443 MCE.emitByte(0xFF); // callq *r10
444 MCE.emitByte(2 | (2 << 3) | (3 << 6));
446 MCE.startFunctionStub(F, 6, 4);
447 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
449 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
452 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
453 return MCE.finishFunctionStub(F);
456 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
457 /// specific basic block.
458 intptr_t X86JITInfo::getPICJumpTableEntry(intptr_t BB, intptr_t Entry) {
462 /// relocate - Before the JIT can run a block of code that has been emitted,
463 /// it must rewrite the code to contain the actual addresses of any
464 /// referenced global symbols.
465 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
466 unsigned NumRelocs, unsigned char* GOTBase) {
467 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
468 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
469 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
470 switch ((X86::RelocationType)MR->getRelocationType()) {
471 case X86::reloc_pcrel_word: {
472 // PC relative relocation, add the relocated value to the value already in
473 // memory, after we adjust it for where the PC is.
474 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
475 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
478 case X86::reloc_picrel_word: {
479 // PIC base relative relocation, add the relocated value to the value
480 // already in memory, after we adjust it for where the PIC base is.
481 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
482 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
485 case X86::reloc_absolute_word:
486 // Absolute relocation, just add the relocated value to the value already
488 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
490 case X86::reloc_absolute_dword:
491 *((intptr_t*)RelocPos) += ResultPtr;