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. 11// 12//===----------------------------------------------------------------------===// 13 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" 22#include "llvm/Support/Valgrind.h" 23#include <cstdlib> 24#include <cstring> 25using namespace llvm; 26 27// Determine the platform we're running on 28#if defined (__x86_64__) || defined (_M_AMD64) || defined (_M_X64) 29# define X86_64_JIT 30#elif defined(__i386__) || defined(i386) || defined(_M_IX86) 31# define X86_32_JIT 32#endif 33 34void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) { 35 unsigned char *OldByte = (unsigned char *)Old; 36 *OldByte++ = 0xE9; // Emit JMP opcode. 37 unsigned *OldWord = (unsigned *)OldByte; 38 unsigned NewAddr = (intptr_t)New; 39 unsigned OldAddr = (intptr_t)OldWord; 40 *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code. 41 42 // X86 doesn't need to invalidate the processor cache, so just invalidate 43 // Valgrind's cache directly. 44 sys::ValgrindDiscardTranslations(Old, 5); 45} 46 47 48/// JITCompilerFunction - This contains the address of the JIT function used to 49/// compile a function lazily. 50static TargetJITInfo::JITCompilerFn JITCompilerFunction; 51 52// Get the ASMPREFIX for the current host. This is often '_'. 53#ifndef __USER_LABEL_PREFIX__ 54#define __USER_LABEL_PREFIX__ 55#endif 56#define GETASMPREFIX2(X) #X 57#define GETASMPREFIX(X) GETASMPREFIX2(X) 58#define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__) 59 60// For ELF targets, use a .size and .type directive, to let tools 61// know the extent of functions defined in assembler. 62#if defined(__ELF__) 63# define SIZE(sym) ".size " #sym ", . - " #sym "\n" 64# define TYPE_FUNCTION(sym) ".type " #sym ", @function\n" 65#else 66# define SIZE(sym) 67# define TYPE_FUNCTION(sym) 68#endif 69 70// Provide a convenient way for disabling usage of CFI directives. 71// This is needed for old/broken assemblers (for example, gas on 72// Darwin is pretty old and doesn't support these directives) 73#if defined(__APPLE__) 74# define CFI(x) 75#else 76// FIXME: Disable this until we really want to use it. Also, we will 77// need to add some workarounds for compilers, which support 78// only subset of these directives. 79# define CFI(x) 80#endif 81 82// Provide a wrapper for X86CompilationCallback2 that saves non-traditional 83// callee saved registers, for the fastcc calling convention. 84extern "C" { 85#if defined(X86_64_JIT) 86# ifndef _MSC_VER 87 // No need to save EAX/EDX for X86-64. 88 void X86CompilationCallback(void); 89 asm( 90 ".text\n" 91 ".align 8\n" 92 ".globl " ASMPREFIX "X86CompilationCallback\n" 93 TYPE_FUNCTION(X86CompilationCallback) 94 ASMPREFIX "X86CompilationCallback:\n" 95 CFI(".cfi_startproc\n") 96 // Save RBP 97 "pushq %rbp\n" 98 CFI(".cfi_def_cfa_offset 16\n") 99 CFI(".cfi_offset %rbp, -16\n") 100 // Save RSP 101 "movq %rsp, %rbp\n" 102 CFI(".cfi_def_cfa_register %rbp\n") 103 // Save all int arg registers 104 "pushq %rdi\n" 105 CFI(".cfi_rel_offset %rdi, 0\n") 106 "pushq %rsi\n" 107 CFI(".cfi_rel_offset %rsi, 8\n") 108 "pushq %rdx\n" 109 CFI(".cfi_rel_offset %rdx, 16\n") 110 "pushq %rcx\n" 111 CFI(".cfi_rel_offset %rcx, 24\n") 112 "pushq %r8\n" 113 CFI(".cfi_rel_offset %r8, 32\n") 114 "pushq %r9\n" 115 CFI(".cfi_rel_offset %r9, 40\n") 116 // Align stack on 16-byte boundary. ESP might not be properly aligned 117 // (8 byte) if this is called from an indirect stub. 118 "andq $-16, %rsp\n" 119 // Save all XMM arg registers 120 "subq $128, %rsp\n" 121 "movaps %xmm0, (%rsp)\n" 122 "movaps %xmm1, 16(%rsp)\n" 123 "movaps %xmm2, 32(%rsp)\n" 124 "movaps %xmm3, 48(%rsp)\n" 125 "movaps %xmm4, 64(%rsp)\n" 126 "movaps %xmm5, 80(%rsp)\n" 127 "movaps %xmm6, 96(%rsp)\n" 128 "movaps %xmm7, 112(%rsp)\n" 129 // JIT callee 130#ifdef _WIN64 131 "subq $32, %rsp\n" 132 "movq %rbp, %rcx\n" // Pass prev frame and return address 133 "movq 8(%rbp), %rdx\n" 134 "call " ASMPREFIX "X86CompilationCallback2\n" 135 "addq $32, %rsp\n" 136#else 137 "movq %rbp, %rdi\n" // Pass prev frame and return address 138 "movq 8(%rbp), %rsi\n" 139 "call " ASMPREFIX "X86CompilationCallback2\n" 140#endif 141 // Restore all XMM arg registers 142 "movaps 112(%rsp), %xmm7\n" 143 "movaps 96(%rsp), %xmm6\n" 144 "movaps 80(%rsp), %xmm5\n" 145 "movaps 64(%rsp), %xmm4\n" 146 "movaps 48(%rsp), %xmm3\n" 147 "movaps 32(%rsp), %xmm2\n" 148 "movaps 16(%rsp), %xmm1\n" 149 "movaps (%rsp), %xmm0\n" 150 // Restore RSP 151 "movq %rbp, %rsp\n" 152 CFI(".cfi_def_cfa_register %rsp\n") 153 // Restore all int arg registers 154 "subq $48, %rsp\n" 155 CFI(".cfi_adjust_cfa_offset 48\n") 156 "popq %r9\n" 157 CFI(".cfi_adjust_cfa_offset -8\n") 158 CFI(".cfi_restore %r9\n") 159 "popq %r8\n" 160 CFI(".cfi_adjust_cfa_offset -8\n") 161 CFI(".cfi_restore %r8\n") 162 "popq %rcx\n" 163 CFI(".cfi_adjust_cfa_offset -8\n") 164 CFI(".cfi_restore %rcx\n") 165 "popq %rdx\n" 166 CFI(".cfi_adjust_cfa_offset -8\n") 167 CFI(".cfi_restore %rdx\n") 168 "popq %rsi\n" 169 CFI(".cfi_adjust_cfa_offset -8\n") 170 CFI(".cfi_restore %rsi\n") 171 "popq %rdi\n" 172 CFI(".cfi_adjust_cfa_offset -8\n") 173 CFI(".cfi_restore %rdi\n") 174 // Restore RBP 175 "popq %rbp\n" 176 CFI(".cfi_adjust_cfa_offset -8\n") 177 CFI(".cfi_restore %rbp\n") 178 "ret\n" 179 CFI(".cfi_endproc\n") 180 SIZE(X86CompilationCallback) 181 ); 182# else 183 // No inline assembler support on this platform. The routine is in external 184 // file. 185 void X86CompilationCallback(); 186 187# endif 188#elif defined (X86_32_JIT) 189# ifndef _MSC_VER 190 void X86CompilationCallback(void); 191 asm( 192 ".text\n" 193 ".align 8\n" 194 ".globl " ASMPREFIX "X86CompilationCallback\n" 195 TYPE_FUNCTION(X86CompilationCallback) 196 ASMPREFIX "X86CompilationCallback:\n" 197 CFI(".cfi_startproc\n") 198 "pushl %ebp\n" 199 CFI(".cfi_def_cfa_offset 8\n") 200 CFI(".cfi_offset %ebp, -8\n") 201 "movl %esp, %ebp\n" // Standard prologue 202 CFI(".cfi_def_cfa_register %ebp\n") 203 "pushl %eax\n" 204 CFI(".cfi_rel_offset %eax, 0\n") 205 "pushl %edx\n" // Save EAX/EDX/ECX 206 CFI(".cfi_rel_offset %edx, 4\n") 207 "pushl %ecx\n" 208 CFI(".cfi_rel_offset %ecx, 8\n") 209# if defined(__APPLE__) 210 "andl $-16, %esp\n" // Align ESP on 16-byte boundary 211# endif 212 "subl $16, %esp\n" 213 "movl 4(%ebp), %eax\n" // Pass prev frame and return address 214 "movl %eax, 4(%esp)\n" 215 "movl %ebp, (%esp)\n" 216 "call " ASMPREFIX "X86CompilationCallback2\n" 217 "movl %ebp, %esp\n" // Restore ESP 218 CFI(".cfi_def_cfa_register %esp\n") 219 "subl $12, %esp\n" 220 CFI(".cfi_adjust_cfa_offset 12\n") 221 "popl %ecx\n" 222 CFI(".cfi_adjust_cfa_offset -4\n") 223 CFI(".cfi_restore %ecx\n") 224 "popl %edx\n" 225 CFI(".cfi_adjust_cfa_offset -4\n") 226 CFI(".cfi_restore %edx\n") 227 "popl %eax\n" 228 CFI(".cfi_adjust_cfa_offset -4\n") 229 CFI(".cfi_restore %eax\n") 230 "popl %ebp\n" 231 CFI(".cfi_adjust_cfa_offset -4\n") 232 CFI(".cfi_restore %ebp\n") 233 "ret\n" 234 CFI(".cfi_endproc\n") 235 SIZE(X86CompilationCallback) 236 ); 237 238 // Same as X86CompilationCallback but also saves XMM argument registers. 239 void X86CompilationCallback_SSE(void); 240 asm( 241 ".text\n" 242 ".align 8\n" 243 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n" 244 TYPE_FUNCTION(X86CompilationCallback_SSE) 245 ASMPREFIX "X86CompilationCallback_SSE:\n" 246 CFI(".cfi_startproc\n") 247 "pushl %ebp\n" 248 CFI(".cfi_def_cfa_offset 8\n") 249 CFI(".cfi_offset %ebp, -8\n") 250 "movl %esp, %ebp\n" // Standard prologue 251 CFI(".cfi_def_cfa_register %ebp\n") 252 "pushl %eax\n" 253 CFI(".cfi_rel_offset %eax, 0\n") 254 "pushl %edx\n" // Save EAX/EDX/ECX 255 CFI(".cfi_rel_offset %edx, 4\n") 256 "pushl %ecx\n" 257 CFI(".cfi_rel_offset %ecx, 8\n") 258 "andl $-16, %esp\n" // Align ESP on 16-byte boundary 259 // Save all XMM arg registers 260 "subl $64, %esp\n" 261 // FIXME: provide frame move information for xmm registers. 262 // This can be tricky, because CFA register is ebp (unaligned) 263 // and we need to produce offsets relative to it. 264 "movaps %xmm0, (%esp)\n" 265 "movaps %xmm1, 16(%esp)\n" 266 "movaps %xmm2, 32(%esp)\n" 267 "movaps %xmm3, 48(%esp)\n" 268 "subl $16, %esp\n" 269 "movl 4(%ebp), %eax\n" // Pass prev frame and return address 270 "movl %eax, 4(%esp)\n" 271 "movl %ebp, (%esp)\n" 272 "call " ASMPREFIX "X86CompilationCallback2\n" 273 "addl $16, %esp\n" 274 "movaps 48(%esp), %xmm3\n" 275 CFI(".cfi_restore %xmm3\n") 276 "movaps 32(%esp), %xmm2\n" 277 CFI(".cfi_restore %xmm2\n") 278 "movaps 16(%esp), %xmm1\n" 279 CFI(".cfi_restore %xmm1\n") 280 "movaps (%esp), %xmm0\n" 281 CFI(".cfi_restore %xmm0\n") 282 "movl %ebp, %esp\n" // Restore ESP 283 CFI(".cfi_def_cfa_register esp\n") 284 "subl $12, %esp\n" 285 CFI(".cfi_adjust_cfa_offset 12\n") 286 "popl %ecx\n" 287 CFI(".cfi_adjust_cfa_offset -4\n") 288 CFI(".cfi_restore %ecx\n") 289 "popl %edx\n" 290 CFI(".cfi_adjust_cfa_offset -4\n") 291 CFI(".cfi_restore %edx\n") 292 "popl %eax\n" 293 CFI(".cfi_adjust_cfa_offset -4\n") 294 CFI(".cfi_restore %eax\n") 295 "popl %ebp\n" 296 CFI(".cfi_adjust_cfa_offset -4\n") 297 CFI(".cfi_restore %ebp\n") 298 "ret\n" 299 CFI(".cfi_endproc\n") 300 SIZE(X86CompilationCallback_SSE) 301 ); 302# else 303 void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr); 304 305 _declspec(naked) void X86CompilationCallback(void) { 306 __asm { 307 push ebp 308 mov ebp, esp 309 push eax 310 push edx 311 push ecx 312 and esp, -16 313 sub esp, 16 314 mov eax, dword ptr [ebp+4] 315 mov dword ptr [esp+4], eax 316 mov dword ptr [esp], ebp 317 call X86CompilationCallback2 318 mov esp, ebp 319 sub esp, 12 320 pop ecx 321 pop edx 322 pop eax 323 pop ebp 324 ret 325 } 326 } 327 328# endif // _MSC_VER 329 330#else // Not an i386 host 331 void X86CompilationCallback() { 332 llvm_unreachable("Cannot call X86CompilationCallback() on a non-x86 arch!"); 333 } 334#endif 335} 336 337/// X86CompilationCallback2 - This is the target-specific function invoked by the 338/// function stub when we did not know the real target of a call. This function 339/// must locate the start of the stub or call site and pass it into the JIT 340/// compiler function. 341extern "C" { 342void 343X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) { 344 intptr_t *RetAddrLoc = &StackPtr[1]; 345 assert(*RetAddrLoc == RetAddr && 346 "Could not find return address on the stack!"); 347 348 // It's a stub if there is an interrupt marker after the call. 349 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCE; 350 351 // The call instruction should have pushed the return value onto the stack... 352#if defined (X86_64_JIT) 353 RetAddr--; // Backtrack to the reference itself... 354#else 355 RetAddr -= 4; // Backtrack to the reference itself... 356#endif 357 358#if 0 359 DEBUG(dbgs() << "In callback! Addr=" << (void*)RetAddr 360 << " ESP=" << (void*)StackPtr 361 << ": Resolving call to function: " 362 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n"); 363#endif 364 365 // Sanity check to make sure this really is a call instruction. 366#if defined (X86_64_JIT) 367 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!"); 368 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!"); 369#else 370 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!"); 371#endif 372 373 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr); 374 375 // Rewrite the call target... so that we don't end up here every time we 376 // execute the call. 377#if defined (X86_64_JIT) 378 assert(isStub && 379 "X86-64 doesn't support rewriting non-stub lazy compilation calls:" 380 " the call instruction varies too much."); 381#else 382 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4); 383#endif 384 385 if (isStub) { 386 // If this is a stub, rewrite the call into an unconditional branch 387 // instruction so that two return addresses are not pushed onto the stack 388 // when the requested function finally gets called. This also makes the 389 // 0xCE byte (interrupt) dead, so the marker doesn't effect anything. 390#if defined (X86_64_JIT) 391 // If the target address is within 32-bit range of the stub, use a 392 // PC-relative branch instead of loading the actual address. (This is 393 // considerably shorter than the 64-bit immediate load already there.) 394 // We assume here intptr_t is 64 bits. 395 intptr_t diff = NewVal-RetAddr+7; 396 if (diff >= -2147483648LL && diff <= 2147483647LL) { 397 *(unsigned char*)(RetAddr-0xc) = 0xE9; 398 *(intptr_t *)(RetAddr-0xb) = diff & 0xffffffff; 399 } else { 400 *(intptr_t *)(RetAddr - 0xa) = NewVal; 401 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6)); 402 } 403 sys::ValgrindDiscardTranslations((void*)(RetAddr-0xc), 0xd); 404#else 405 ((unsigned char*)RetAddr)[-1] = 0xE9; 406 sys::ValgrindDiscardTranslations((void*)(RetAddr-1), 5); 407#endif 408 } 409 410 // Change the return address to reexecute the call instruction... 411#if defined (X86_64_JIT) 412 *RetAddrLoc -= 0xd; 413#else 414 *RetAddrLoc -= 5; 415#endif 416} 417} 418 419TargetJITInfo::LazyResolverFn 420X86JITInfo::getLazyResolverFunction(JITCompilerFn F) { 421 JITCompilerFunction = F; 422 423#if defined (X86_32_JIT) && !defined (_MSC_VER) 424 if (Subtarget->hasSSE1()) 425 return X86CompilationCallback_SSE; 426#endif 427 428 return X86CompilationCallback; 429} 430 431X86JITInfo::X86JITInfo(X86TargetMachine &tm) : TM(tm) { 432 Subtarget = &TM.getSubtarget<X86Subtarget>(); 433 useGOT = 0; 434 TLSOffset = 0; 435} 436 437void *X86JITInfo::emitGlobalValueIndirectSym(const GlobalValue* GV, void *ptr, 438 JITCodeEmitter &JCE) { 439#if defined (X86_64_JIT) 440 const unsigned Alignment = 8; 441 uint8_t Buffer[8]; 442 uint8_t *Cur = Buffer; 443 MachineCodeEmitter::emitWordLEInto(Cur, (unsigned)(intptr_t)ptr); 444 MachineCodeEmitter::emitWordLEInto(Cur, (unsigned)(((intptr_t)ptr) >> 32)); 445#else 446 const unsigned Alignment = 4; 447 uint8_t Buffer[4]; 448 uint8_t *Cur = Buffer; 449 MachineCodeEmitter::emitWordLEInto(Cur, (intptr_t)ptr); 450#endif 451 return JCE.allocIndirectGV(GV, Buffer, sizeof(Buffer), Alignment); 452} 453 454TargetJITInfo::StubLayout X86JITInfo::getStubLayout() { 455 // The 64-bit stub contains: 456 // movabs r10 <- 8-byte-target-address # 10 bytes 457 // call|jmp *r10 # 3 bytes 458 // The 32-bit stub contains a 5-byte call|jmp. 459 // If the stub is a call to the compilation callback, an extra byte is added 460 // to mark it as a stub. 461 StubLayout Result = {14, 4}; 462 return Result; 463} 464 465void *X86JITInfo::emitFunctionStub(const Function* F, void *Target, 466 JITCodeEmitter &JCE) { 467 // Note, we cast to intptr_t here to silence a -pedantic warning that 468 // complains about casting a function pointer to a normal pointer. 469#if defined (X86_32_JIT) && !defined (_MSC_VER) 470 bool NotCC = (Target != (void*)(intptr_t)X86CompilationCallback && 471 Target != (void*)(intptr_t)X86CompilationCallback_SSE); 472#else 473 bool NotCC = Target != (void*)(intptr_t)X86CompilationCallback; 474#endif 475 JCE.emitAlignment(4); 476 void *Result = (void*)JCE.getCurrentPCValue(); 477 if (NotCC) { 478#if defined (X86_64_JIT) 479 JCE.emitByte(0x49); // REX prefix 480 JCE.emitByte(0xB8+2); // movabsq r10 481 JCE.emitWordLE((unsigned)(intptr_t)Target); 482 JCE.emitWordLE((unsigned)(((intptr_t)Target) >> 32)); 483 JCE.emitByte(0x41); // REX prefix 484 JCE.emitByte(0xFF); // jmpq *r10 485 JCE.emitByte(2 | (4 << 3) | (3 << 6)); 486#else 487 JCE.emitByte(0xE9); 488 JCE.emitWordLE((intptr_t)Target-JCE.getCurrentPCValue()-4); 489#endif 490 return Result; 491 } 492 493#if defined (X86_64_JIT) 494 JCE.emitByte(0x49); // REX prefix 495 JCE.emitByte(0xB8+2); // movabsq r10 496 JCE.emitWordLE((unsigned)(intptr_t)Target); 497 JCE.emitWordLE((unsigned)(((intptr_t)Target) >> 32)); 498 JCE.emitByte(0x41); // REX prefix 499 JCE.emitByte(0xFF); // callq *r10 500 JCE.emitByte(2 | (2 << 3) | (3 << 6)); 501#else 502 JCE.emitByte(0xE8); // Call with 32 bit pc-rel destination... 503 504 JCE.emitWordLE((intptr_t)Target-JCE.getCurrentPCValue()-4); 505#endif 506 507 // This used to use 0xCD, but that value is used by JITMemoryManager to 508 // initialize the buffer with garbage, which means it may follow a 509 // noreturn function call, confusing X86CompilationCallback2. PR 4929. 510 JCE.emitByte(0xCE); // Interrupt - Just a marker identifying the stub! 511 return Result; 512} 513 514/// getPICJumpTableEntry - Returns the value of the jumptable entry for the 515/// specific basic block. 516uintptr_t X86JITInfo::getPICJumpTableEntry(uintptr_t BB, uintptr_t Entry) { 517#if defined(X86_64_JIT) 518 return BB - Entry; 519#else 520 return BB - PICBase; 521#endif 522} 523 524/// relocate - Before the JIT can run a block of code that has been emitted, 525/// it must rewrite the code to contain the actual addresses of any 526/// referenced global symbols. 527void X86JITInfo::relocate(void *Function, MachineRelocation *MR, 528 unsigned NumRelocs, unsigned char* GOTBase) { 529 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) { 530 void *RelocPos = (char*)Function + MR->getMachineCodeOffset(); 531 intptr_t ResultPtr = (intptr_t)MR->getResultPointer(); 532 switch ((X86::RelocationType)MR->getRelocationType()) { 533 case X86::reloc_pcrel_word: { 534 // PC relative relocation, add the relocated value to the value already in 535 // memory, after we adjust it for where the PC is. 536 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal(); 537 *((unsigned*)RelocPos) += (unsigned)ResultPtr; 538 break; 539 } 540 case X86::reloc_picrel_word: { 541 // PIC base relative relocation, add the relocated value to the value 542 // already in memory, after we adjust it for where the PIC base is. 543 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal()); 544 *((unsigned*)RelocPos) += (unsigned)ResultPtr; 545 break; 546 } 547 case X86::reloc_absolute_word: 548 case X86::reloc_absolute_word_sext: 549 // Absolute relocation, just add the relocated value to the value already 550 // in memory. 551 *((unsigned*)RelocPos) += (unsigned)ResultPtr; 552 break; 553 case X86::reloc_absolute_dword: 554 *((intptr_t*)RelocPos) += ResultPtr; 555 break; 556 } 557 } 558} 559 560char* X86JITInfo::allocateThreadLocalMemory(size_t size) { 561#if defined(X86_32_JIT) && !defined(__APPLE__) && !defined(_MSC_VER) 562 TLSOffset -= size; 563 return TLSOffset; 564#else 565 llvm_unreachable("Cannot allocate thread local storage on this arch!"); 566 return 0; 567#endif 568} 569