Hanle i8 returns
[llvm/msp430.git] / lib / Target / X86 / X86JITInfo.cpp
blob4bbccf380c3fae347612a75d9d23e779b42a7866
1 //===-- X86JITInfo.cpp - Implement the JIT interfaces for the X86 target --===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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"
22 #include <cstdlib>
23 #include <cstring>
24 using namespace llvm;
26 // Determine the platform we're running on
27 #if defined (__x86_64__) || defined (_M_AMD64)
28 # define X86_64_JIT
29 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
30 # define X86_32_JIT
31 #endif
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__
50 #endif
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"
58 #else
59 #define ASMCALLSUFFIX
60 #endif
62 // For ELF targets, use a .size and .type directive, to let tools
63 // know the extent of functions defined in assembler.
64 #if defined(__ELF__)
65 # define SIZE(sym) ".size " #sym ", . - " #sym "\n"
66 # define TYPE_FUNCTION(sym) ".type " #sym ", @function\n"
67 #else
68 # define SIZE(sym)
69 # define TYPE_FUNCTION(sym)
70 #endif
72 // Provide a convenient way for disabling usage of CFI directives.
73 // This is needed for old/broken assemblers (for example, gas on
74 // Darwin is pretty old and doesn't support these directives)
75 #if defined(__APPLE__)
76 # define CFI(x)
77 #else
78 // FIXME: Disable this until we really want to use it. Also, we will
79 // need to add some workarounds for compilers, which support
80 // only subset of these directives.
81 # define CFI(x)
82 #endif
84 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
85 // callee saved registers, for the fastcc calling convention.
86 extern "C" {
87 #if defined(X86_64_JIT)
88 # ifndef _MSC_VER
89 // No need to save EAX/EDX for X86-64.
90 void X86CompilationCallback(void);
91 asm(
92 ".text\n"
93 ".align 8\n"
94 ".globl " ASMPREFIX "X86CompilationCallback\n"
95 TYPE_FUNCTION(X86CompilationCallback)
96 ASMPREFIX "X86CompilationCallback:\n"
97 CFI(".cfi_startproc\n")
98 // Save RBP
99 "pushq %rbp\n"
100 CFI(".cfi_def_cfa_offset 16\n")
101 CFI(".cfi_offset %rbp, -16\n")
102 // Save RSP
103 "movq %rsp, %rbp\n"
104 CFI(".cfi_def_cfa_register %rbp\n")
105 // Save all int arg registers
106 "pushq %rdi\n"
107 CFI(".cfi_rel_offset %rdi, 0\n")
108 "pushq %rsi\n"
109 CFI(".cfi_rel_offset %rsi, 8\n")
110 "pushq %rdx\n"
111 CFI(".cfi_rel_offset %rdx, 16\n")
112 "pushq %rcx\n"
113 CFI(".cfi_rel_offset %rcx, 24\n")
114 "pushq %r8\n"
115 CFI(".cfi_rel_offset %r8, 32\n")
116 "pushq %r9\n"
117 CFI(".cfi_rel_offset %r9, 40\n")
118 // Align stack on 16-byte boundary. ESP might not be properly aligned
119 // (8 byte) if this is called from an indirect stub.
120 "andq $-16, %rsp\n"
121 // Save all XMM arg registers
122 "subq $128, %rsp\n"
123 "movaps %xmm0, (%rsp)\n"
124 "movaps %xmm1, 16(%rsp)\n"
125 "movaps %xmm2, 32(%rsp)\n"
126 "movaps %xmm3, 48(%rsp)\n"
127 "movaps %xmm4, 64(%rsp)\n"
128 "movaps %xmm5, 80(%rsp)\n"
129 "movaps %xmm6, 96(%rsp)\n"
130 "movaps %xmm7, 112(%rsp)\n"
131 // JIT callee
132 "movq %rbp, %rdi\n" // Pass prev frame and return address
133 "movq 8(%rbp), %rsi\n"
134 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
135 // Restore all XMM arg registers
136 "movaps 112(%rsp), %xmm7\n"
137 "movaps 96(%rsp), %xmm6\n"
138 "movaps 80(%rsp), %xmm5\n"
139 "movaps 64(%rsp), %xmm4\n"
140 "movaps 48(%rsp), %xmm3\n"
141 "movaps 32(%rsp), %xmm2\n"
142 "movaps 16(%rsp), %xmm1\n"
143 "movaps (%rsp), %xmm0\n"
144 // Restore RSP
145 "movq %rbp, %rsp\n"
146 CFI(".cfi_def_cfa_register %rsp\n")
147 // Restore all int arg registers
148 "subq $48, %rsp\n"
149 CFI(".cfi_adjust_cfa_offset 48\n")
150 "popq %r9\n"
151 CFI(".cfi_adjust_cfa_offset -8\n")
152 CFI(".cfi_restore %r9\n")
153 "popq %r8\n"
154 CFI(".cfi_adjust_cfa_offset -8\n")
155 CFI(".cfi_restore %r8\n")
156 "popq %rcx\n"
157 CFI(".cfi_adjust_cfa_offset -8\n")
158 CFI(".cfi_restore %rcx\n")
159 "popq %rdx\n"
160 CFI(".cfi_adjust_cfa_offset -8\n")
161 CFI(".cfi_restore %rdx\n")
162 "popq %rsi\n"
163 CFI(".cfi_adjust_cfa_offset -8\n")
164 CFI(".cfi_restore %rsi\n")
165 "popq %rdi\n"
166 CFI(".cfi_adjust_cfa_offset -8\n")
167 CFI(".cfi_restore %rdi\n")
168 // Restore RBP
169 "popq %rbp\n"
170 CFI(".cfi_adjust_cfa_offset -8\n")
171 CFI(".cfi_restore %rbp\n")
172 "ret\n"
173 CFI(".cfi_endproc\n")
174 SIZE(X86CompilationCallback)
176 # else
177 // No inline assembler support on this platform. The routine is in external
178 // file.
179 void X86CompilationCallback();
181 # endif
182 #elif defined (X86_32_JIT)
183 # ifndef _MSC_VER
184 void X86CompilationCallback(void);
185 asm(
186 ".text\n"
187 ".align 8\n"
188 ".globl " ASMPREFIX "X86CompilationCallback\n"
189 TYPE_FUNCTION(X86CompilationCallback)
190 ASMPREFIX "X86CompilationCallback:\n"
191 CFI(".cfi_startproc\n")
192 "pushl %ebp\n"
193 CFI(".cfi_def_cfa_offset 8\n")
194 CFI(".cfi_offset %ebp, -8\n")
195 "movl %esp, %ebp\n" // Standard prologue
196 CFI(".cfi_def_cfa_register %ebp\n")
197 "pushl %eax\n"
198 CFI(".cfi_rel_offset %eax, 0\n")
199 "pushl %edx\n" // Save EAX/EDX/ECX
200 CFI(".cfi_rel_offset %edx, 4\n")
201 "pushl %ecx\n"
202 CFI(".cfi_rel_offset %ecx, 8\n")
203 # if defined(__APPLE__)
204 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
205 # endif
206 "subl $16, %esp\n"
207 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
208 "movl %eax, 4(%esp)\n"
209 "movl %ebp, (%esp)\n"
210 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
211 "movl %ebp, %esp\n" // Restore ESP
212 CFI(".cfi_def_cfa_register %esp\n")
213 "subl $12, %esp\n"
214 CFI(".cfi_adjust_cfa_offset 12\n")
215 "popl %ecx\n"
216 CFI(".cfi_adjust_cfa_offset -4\n")
217 CFI(".cfi_restore %ecx\n")
218 "popl %edx\n"
219 CFI(".cfi_adjust_cfa_offset -4\n")
220 CFI(".cfi_restore %edx\n")
221 "popl %eax\n"
222 CFI(".cfi_adjust_cfa_offset -4\n")
223 CFI(".cfi_restore %eax\n")
224 "popl %ebp\n"
225 CFI(".cfi_adjust_cfa_offset -4\n")
226 CFI(".cfi_restore %ebp\n")
227 "ret\n"
228 CFI(".cfi_endproc\n")
229 SIZE(X86CompilationCallback)
232 // Same as X86CompilationCallback but also saves XMM argument registers.
233 void X86CompilationCallback_SSE(void);
234 asm(
235 ".text\n"
236 ".align 8\n"
237 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
238 TYPE_FUNCTION(X86CompilationCallback_SSE)
239 ASMPREFIX "X86CompilationCallback_SSE:\n"
240 CFI(".cfi_startproc\n")
241 "pushl %ebp\n"
242 CFI(".cfi_def_cfa_offset 8\n")
243 CFI(".cfi_offset %ebp, -8\n")
244 "movl %esp, %ebp\n" // Standard prologue
245 CFI(".cfi_def_cfa_register %ebp\n")
246 "pushl %eax\n"
247 CFI(".cfi_rel_offset %eax, 0\n")
248 "pushl %edx\n" // Save EAX/EDX/ECX
249 CFI(".cfi_rel_offset %edx, 4\n")
250 "pushl %ecx\n"
251 CFI(".cfi_rel_offset %ecx, 8\n")
252 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
253 // Save all XMM arg registers
254 "subl $64, %esp\n"
255 // FIXME: provide frame move information for xmm registers.
256 // This can be tricky, because CFA register is ebp (unaligned)
257 // and we need to produce offsets relative to it.
258 "movaps %xmm0, (%esp)\n"
259 "movaps %xmm1, 16(%esp)\n"
260 "movaps %xmm2, 32(%esp)\n"
261 "movaps %xmm3, 48(%esp)\n"
262 "subl $16, %esp\n"
263 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
264 "movl %eax, 4(%esp)\n"
265 "movl %ebp, (%esp)\n"
266 "call " ASMPREFIX "X86CompilationCallback2" ASMCALLSUFFIX "\n"
267 "addl $16, %esp\n"
268 "movaps 48(%esp), %xmm3\n"
269 CFI(".cfi_restore %xmm3\n")
270 "movaps 32(%esp), %xmm2\n"
271 CFI(".cfi_restore %xmm2\n")
272 "movaps 16(%esp), %xmm1\n"
273 CFI(".cfi_restore %xmm1\n")
274 "movaps (%esp), %xmm0\n"
275 CFI(".cfi_restore %xmm0\n")
276 "movl %ebp, %esp\n" // Restore ESP
277 CFI(".cfi_def_cfa_register esp\n")
278 "subl $12, %esp\n"
279 CFI(".cfi_adjust_cfa_offset 12\n")
280 "popl %ecx\n"
281 CFI(".cfi_adjust_cfa_offset -4\n")
282 CFI(".cfi_restore %ecx\n")
283 "popl %edx\n"
284 CFI(".cfi_adjust_cfa_offset -4\n")
285 CFI(".cfi_restore %edx\n")
286 "popl %eax\n"
287 CFI(".cfi_adjust_cfa_offset -4\n")
288 CFI(".cfi_restore %eax\n")
289 "popl %ebp\n"
290 CFI(".cfi_adjust_cfa_offset -4\n")
291 CFI(".cfi_restore %ebp\n")
292 "ret\n"
293 CFI(".cfi_endproc\n")
294 SIZE(X86CompilationCallback_SSE)
296 # else
297 void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr);
299 _declspec(naked) void X86CompilationCallback(void) {
300 __asm {
301 push ebp
302 mov ebp, esp
303 push eax
304 push edx
305 push ecx
306 and esp, -16
307 mov eax, dword ptr [ebp+4]
308 mov dword ptr [esp+4], eax
309 mov dword ptr [esp], ebp
310 call X86CompilationCallback2
311 mov esp, ebp
312 sub esp, 12
313 pop ecx
314 pop edx
315 pop eax
316 pop ebp
321 # endif // _MSC_VER
323 #else // Not an i386 host
324 void X86CompilationCallback() {
325 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
326 abort();
328 #endif
331 /// X86CompilationCallback2 - This is the target-specific function invoked by the
332 /// function stub when we did not know the real target of a call. This function
333 /// must locate the start of the stub or call site and pass it into the JIT
334 /// compiler function.
335 extern "C" void ATTRIBUTE_USED
336 X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
337 intptr_t *RetAddrLoc = &StackPtr[1];
338 assert(*RetAddrLoc == RetAddr &&
339 "Could not find return address on the stack!");
341 // It's a stub if there is an interrupt marker after the call.
342 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
344 // The call instruction should have pushed the return value onto the stack...
345 #if defined (X86_64_JIT)
346 RetAddr--; // Backtrack to the reference itself...
347 #else
348 RetAddr -= 4; // Backtrack to the reference itself...
349 #endif
351 #if 0
352 DOUT << "In callback! Addr=" << (void*)RetAddr
353 << " ESP=" << (void*)StackPtr
354 << ": Resolving call to function: "
355 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
356 #endif
358 // Sanity check to make sure this really is a call instruction.
359 #if defined (X86_64_JIT)
360 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
361 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
362 #else
363 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
364 #endif
366 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
368 // Rewrite the call target... so that we don't end up here every time we
369 // execute the call.
370 #if defined (X86_64_JIT)
371 if (!isStub)
372 *(intptr_t *)(RetAddr - 0xa) = NewVal;
373 #else
374 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
375 #endif
377 if (isStub) {
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 // 0xCD 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;
391 } else {
392 *(intptr_t *)(RetAddr - 0xa) = NewVal;
393 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
395 #else
396 ((unsigned char*)RetAddr)[-1] = 0xE9;
397 #endif
400 // Change the return address to reexecute the call instruction...
401 #if defined (X86_64_JIT)
402 *RetAddrLoc -= 0xd;
403 #else
404 *RetAddrLoc -= 5;
405 #endif
408 TargetJITInfo::LazyResolverFn
409 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
410 JITCompilerFunction = F;
412 #if defined (X86_32_JIT) && !defined (_MSC_VER)
413 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
414 union {
415 unsigned u[3];
416 char c[12];
417 } text;
419 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
420 // FIXME: support for AMD family of processors.
421 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
422 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
423 if ((EDX >> 25) & 0x1)
424 return X86CompilationCallback_SSE;
427 #endif
429 return X86CompilationCallback;
432 void *X86JITInfo::emitGlobalValueIndirectSym(const GlobalValue* GV, void *ptr,
433 MachineCodeEmitter &MCE) {
434 #if defined (X86_64_JIT)
435 MCE.startGVStub(GV, 8, 8);
436 MCE.emitWordLE((unsigned)(intptr_t)ptr);
437 MCE.emitWordLE((unsigned)(((intptr_t)ptr) >> 32));
438 #else
439 MCE.startGVStub(GV, 4, 4);
440 MCE.emitWordLE((intptr_t)ptr);
441 #endif
442 return MCE.finishGVStub(GV);
445 void *X86JITInfo::emitFunctionStub(const Function* F, void *Fn,
446 MachineCodeEmitter &MCE) {
447 // Note, we cast to intptr_t here to silence a -pedantic warning that
448 // complains about casting a function pointer to a normal pointer.
449 #if defined (X86_32_JIT) && !defined (_MSC_VER)
450 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
451 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
452 #else
453 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
454 #endif
455 if (NotCC) {
456 #if defined (X86_64_JIT)
457 MCE.startGVStub(F, 13, 4);
458 MCE.emitByte(0x49); // REX prefix
459 MCE.emitByte(0xB8+2); // movabsq r10
460 MCE.emitWordLE((unsigned)(intptr_t)Fn);
461 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
462 MCE.emitByte(0x41); // REX prefix
463 MCE.emitByte(0xFF); // jmpq *r10
464 MCE.emitByte(2 | (4 << 3) | (3 << 6));
465 #else
466 MCE.startGVStub(F, 5, 4);
467 MCE.emitByte(0xE9);
468 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
469 #endif
470 return MCE.finishGVStub(F);
473 #if defined (X86_64_JIT)
474 MCE.startGVStub(F, 14, 4);
475 MCE.emitByte(0x49); // REX prefix
476 MCE.emitByte(0xB8+2); // movabsq r10
477 MCE.emitWordLE((unsigned)(intptr_t)Fn);
478 MCE.emitWordLE((unsigned)(((intptr_t)Fn) >> 32));
479 MCE.emitByte(0x41); // REX prefix
480 MCE.emitByte(0xFF); // callq *r10
481 MCE.emitByte(2 | (2 << 3) | (3 << 6));
482 #else
483 MCE.startGVStub(F, 6, 4);
484 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
486 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
487 #endif
489 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
490 return MCE.finishGVStub(F);
493 void X86JITInfo::emitFunctionStubAtAddr(const Function* F, void *Fn, void *Stub,
494 MachineCodeEmitter &MCE) {
495 // Note, we cast to intptr_t here to silence a -pedantic warning that
496 // complains about casting a function pointer to a normal pointer.
497 MCE.startGVStub(F, Stub, 5);
498 MCE.emitByte(0xE9);
499 #if defined (X86_64_JIT)
500 assert(((((intptr_t)Fn-MCE.getCurrentPCValue()-5) << 32) >> 32) ==
501 ((intptr_t)Fn-MCE.getCurrentPCValue()-5)
502 && "PIC displacement does not fit in displacement field!");
503 #endif
504 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
505 MCE.finishGVStub(F);
508 /// getPICJumpTableEntry - Returns the value of the jumptable entry for the
509 /// specific basic block.
510 uintptr_t X86JITInfo::getPICJumpTableEntry(uintptr_t BB, uintptr_t Entry) {
511 #if defined(X86_64_JIT)
512 return BB - Entry;
513 #else
514 return BB - PICBase;
515 #endif
518 /// relocate - Before the JIT can run a block of code that has been emitted,
519 /// it must rewrite the code to contain the actual addresses of any
520 /// referenced global symbols.
521 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
522 unsigned NumRelocs, unsigned char* GOTBase) {
523 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
524 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
525 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
526 switch ((X86::RelocationType)MR->getRelocationType()) {
527 case X86::reloc_pcrel_word: {
528 // PC relative relocation, add the relocated value to the value already in
529 // memory, after we adjust it for where the PC is.
530 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
531 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
532 break;
534 case X86::reloc_picrel_word: {
535 // PIC base relative relocation, add the relocated value to the value
536 // already in memory, after we adjust it for where the PIC base is.
537 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
538 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
539 break;
541 case X86::reloc_absolute_word:
542 // Absolute relocation, just add the relocated value to the value already
543 // in memory.
544 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
545 break;
546 case X86::reloc_absolute_dword:
547 *((intptr_t*)RelocPos) += ResultPtr;
548 break;
553 char* X86JITInfo::allocateThreadLocalMemory(size_t size) {
554 #if defined(X86_32_JIT) && !defined(__APPLE__) && !defined(_MSC_VER)
555 TLSOffset -= size;
556 return TLSOffset;
557 #else
558 assert(0 && "Cannot allocate thread local storage on this arch!\n");
559 return 0;
560 #endif