X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86JITInfo.cpp;h=0168d12231f70060deb42eecf78e73cf55244e56;hp=7f72bd473e37e692c61703e886694c5212a93353;hb=b83a67e1e3fe210bd99a82eccd3dc5b1b44f1503;hpb=51cc3c13eac78da242f0518fc42580e48dd5304f diff --git a/lib/Target/X86/X86JITInfo.cpp b/lib/Target/X86/X86JITInfo.cpp index 7f72bd473e3..0168d12231f 100644 --- a/lib/Target/X86/X86JITInfo.cpp +++ b/lib/Target/X86/X86JITInfo.cpp @@ -15,15 +15,17 @@ #include "X86JITInfo.h" #include "X86Relocations.h" #include "X86Subtarget.h" +#include "X86TargetMachine.h" #include "llvm/Function.h" -#include "llvm/CodeGen/MachineCodeEmitter.h" -#include "llvm/Config/alloca.h" +#include "llvm/Support/Compiler.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/Support/Valgrind.h" #include #include using namespace llvm; // Determine the platform we're running on -#if defined (__x86_64__) || defined (_M_AMD64) +#if defined (__x86_64__) || defined (_M_AMD64) || defined (_M_X64) # define X86_64_JIT #elif defined(__i386__) || defined(i386) || defined(_M_IX86) # define X86_32_JIT @@ -36,6 +38,10 @@ void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) { unsigned NewAddr = (intptr_t)New; unsigned OldAddr = (intptr_t)OldWord; *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code. + + // X86 doesn't need to invalidate the processor cache, so just invalidate + // Valgrind's cache directly. + sys::ValgrindDiscardTranslations(Old, 5); } @@ -51,6 +57,16 @@ static TargetJITInfo::JITCompilerFn JITCompilerFunction; #define GETASMPREFIX(X) GETASMPREFIX2(X) #define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__) +// For ELF targets, use a .size and .type directive, to let tools +// know the extent of functions defined in assembler. +#if defined(__ELF__) +# define SIZE(sym) ".size " #sym ", . - " #sym "\n" +# define TYPE_FUNCTION(sym) ".type " #sym ", @function\n" +#else +# define SIZE(sym) +# define TYPE_FUNCTION(sym) +#endif + // Provide a convenient way for disabling usage of CFI directives. // This is needed for old/broken assemblers (for example, gas on // Darwin is pretty old and doesn't support these directives) @@ -74,6 +90,7 @@ extern "C" { ".text\n" ".align 8\n" ".globl " ASMPREFIX "X86CompilationCallback\n" + TYPE_FUNCTION(X86CompilationCallback) ASMPREFIX "X86CompilationCallback:\n" CFI(".cfi_startproc\n") // Save RBP @@ -110,9 +127,17 @@ extern "C" { "movaps %xmm6, 96(%rsp)\n" "movaps %xmm7, 112(%rsp)\n" // JIT callee +#ifdef _WIN64 + "subq $32, %rsp\n" + "movq %rbp, %rcx\n" // Pass prev frame and return address + "movq 8(%rbp), %rdx\n" + "call " ASMPREFIX "X86CompilationCallback2\n" + "addq $32, %rsp\n" +#else "movq %rbp, %rdi\n" // Pass prev frame and return address "movq 8(%rbp), %rsi\n" "call " ASMPREFIX "X86CompilationCallback2\n" +#endif // Restore all XMM arg registers "movaps 112(%rsp), %xmm7\n" "movaps 96(%rsp), %xmm6\n" @@ -152,6 +177,7 @@ extern "C" { CFI(".cfi_restore %rbp\n") "ret\n" CFI(".cfi_endproc\n") + SIZE(X86CompilationCallback) ); # else // No inline assembler support on this platform. The routine is in external @@ -165,7 +191,8 @@ extern "C" { asm( ".text\n" ".align 8\n" - ".globl " ASMPREFIX "X86CompilationCallback\n" + ".globl " ASMPREFIX "X86CompilationCallback\n" + TYPE_FUNCTION(X86CompilationCallback) ASMPREFIX "X86CompilationCallback:\n" CFI(".cfi_startproc\n") "pushl %ebp\n" @@ -205,6 +232,7 @@ extern "C" { CFI(".cfi_restore %ebp\n") "ret\n" CFI(".cfi_endproc\n") + SIZE(X86CompilationCallback) ); // Same as X86CompilationCallback but also saves XMM argument registers. @@ -212,7 +240,8 @@ extern "C" { asm( ".text\n" ".align 8\n" - ".globl " ASMPREFIX "X86CompilationCallback_SSE\n" + ".globl " ASMPREFIX "X86CompilationCallback_SSE\n" + TYPE_FUNCTION(X86CompilationCallback_SSE) ASMPREFIX "X86CompilationCallback_SSE:\n" CFI(".cfi_startproc\n") "pushl %ebp\n" @@ -268,9 +297,13 @@ extern "C" { CFI(".cfi_restore %ebp\n") "ret\n" CFI(".cfi_endproc\n") + SIZE(X86CompilationCallback_SSE) ); # else - void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr); + // the following function is called only from this translation unit, + // unless we are under 64bit Windows with MSC, where there is + // no support for inline assembly + static void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr); _declspec(naked) void X86CompilationCallback(void) { __asm { @@ -280,6 +313,7 @@ extern "C" { push edx push ecx and esp, -16 + sub esp, 16 mov eax, dword ptr [ebp+4] mov dword ptr [esp+4], eax mov dword ptr [esp], ebp @@ -298,23 +332,30 @@ extern "C" { #else // Not an i386 host void X86CompilationCallback() { - assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n"); - abort(); + llvm_unreachable("Cannot call X86CompilationCallback() on a non-x86 arch!"); } #endif } -/// X86CompilationCallback - This is the target-specific function invoked by the +/// X86CompilationCallback2 - This is the target-specific function invoked by the /// function stub when we did not know the real target of a call. This function /// must locate the start of the stub or call site and pass it into the JIT /// compiler function. -extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) { +extern "C" { +#if !(defined (X86_64_JIT) && defined(_MSC_VER)) + // the following function is called only from this translation unit, + // unless we are under 64bit Windows with MSC, where there is + // no support for inline assembly +static +#endif +void LLVM_ATTRIBUTE_USED +X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) { intptr_t *RetAddrLoc = &StackPtr[1]; assert(*RetAddrLoc == RetAddr && "Could not find return address on the stack!"); // It's a stub if there is an interrupt marker after the call. - bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD; + bool isStub = ((unsigned char*)RetAddr)[0] == 0xCE; // The call instruction should have pushed the return value onto the stack... #if defined (X86_64_JIT) @@ -324,10 +365,10 @@ extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) { #endif #if 0 - DOUT << "In callback! Addr=" << (void*)RetAddr - << " ESP=" << (void*)StackPtr - << ": Resolving call to function: " - << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n"; + DEBUG(dbgs() << "In callback! Addr=" << (void*)RetAddr + << " ESP=" << (void*)StackPtr + << ": Resolving call to function: " + << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n"); #endif // Sanity check to make sure this really is a call instruction. @@ -343,7 +384,9 @@ extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) { // Rewrite the call target... so that we don't end up here every time we // execute the call. #if defined (X86_64_JIT) - *(intptr_t *)(RetAddr - 0xa) = NewVal; + assert(isStub && + "X86-64 doesn't support rewriting non-stub lazy compilation calls:" + " the call instruction varies too much."); #else *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4); #endif @@ -352,11 +395,24 @@ extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) { // If this is a stub, rewrite the call into an unconditional branch // instruction so that two return addresses are not pushed onto the stack // when the requested function finally gets called. This also makes the - // 0xCD byte (interrupt) dead, so the marker doesn't effect anything. + // 0xCE byte (interrupt) dead, so the marker doesn't effect anything. #if defined (X86_64_JIT) - ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6)); + // If the target address is within 32-bit range of the stub, use a + // PC-relative branch instead of loading the actual address. (This is + // considerably shorter than the 64-bit immediate load already there.) + // We assume here intptr_t is 64 bits. + intptr_t diff = NewVal-RetAddr+7; + if (diff >= -2147483648LL && diff <= 2147483647LL) { + *(unsigned char*)(RetAddr-0xc) = 0xE9; + *(intptr_t *)(RetAddr-0xb) = diff & 0xffffffff; + } else { + *(intptr_t *)(RetAddr - 0xa) = NewVal; + ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6)); + } + sys::ValgrindDiscardTranslations((void*)(RetAddr-0xc), 0xd); #else ((unsigned char*)RetAddr)[-1] = 0xE9; + sys::ValgrindDiscardTranslations((void*)(RetAddr-1), 5); #endif } @@ -367,96 +423,113 @@ extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) { *RetAddrLoc -= 5; #endif } +} TargetJITInfo::LazyResolverFn X86JITInfo::getLazyResolverFunction(JITCompilerFn F) { + TsanIgnoreWritesBegin(); JITCompilerFunction = F; + TsanIgnoreWritesEnd(); #if defined (X86_32_JIT) && !defined (_MSC_VER) - unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0; - union { - unsigned u[3]; - char c[12]; - } text; - - if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) { - // FIXME: support for AMD family of processors. - if (memcmp(text.c, "GenuineIntel", 12) == 0) { - X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX); - if ((EDX >> 25) & 0x1) - return X86CompilationCallback_SSE; - } - } + if (Subtarget->hasSSE1()) + return X86CompilationCallback_SSE; #endif return X86CompilationCallback; } -void *X86JITInfo::emitGlobalValueLazyPtr(const GlobalValue* GV, void *ptr, - MachineCodeEmitter &MCE) { +X86JITInfo::X86JITInfo(X86TargetMachine &tm) : TM(tm) { + Subtarget = &TM.getSubtarget(); + useGOT = 0; + TLSOffset = 0; +} + +void *X86JITInfo::emitGlobalValueIndirectSym(const GlobalValue* GV, void *ptr, + JITCodeEmitter &JCE) { #if defined (X86_64_JIT) - MCE.startFunctionStub(8, 8); - MCE.emitWordLE(((unsigned *)&ptr)[0]); - MCE.emitWordLE(((unsigned *)&ptr)[1]); + const unsigned Alignment = 8; + uint8_t Buffer[8]; + uint8_t *Cur = Buffer; + MachineCodeEmitter::emitWordLEInto(Cur, (unsigned)(intptr_t)ptr); + MachineCodeEmitter::emitWordLEInto(Cur, (unsigned)(((intptr_t)ptr) >> 32)); #else - MCE.startFunctionStub(GV, 4, 4); - MCE.emitWordLE((intptr_t)ptr); + const unsigned Alignment = 4; + uint8_t Buffer[4]; + uint8_t *Cur = Buffer; + MachineCodeEmitter::emitWordLEInto(Cur, (intptr_t)ptr); #endif - return MCE.finishFunctionStub(GV); + return JCE.allocIndirectGV(GV, Buffer, sizeof(Buffer), Alignment); } -void *X86JITInfo::emitFunctionStub(const Function* F, void *Fn, - MachineCodeEmitter &MCE) { - // Note, we cast to intptr_t here to silence a -pedantic warning that +TargetJITInfo::StubLayout X86JITInfo::getStubLayout() { + // The 64-bit stub contains: + // movabs r10 <- 8-byte-target-address # 10 bytes + // call|jmp *r10 # 3 bytes + // The 32-bit stub contains a 5-byte call|jmp. + // If the stub is a call to the compilation callback, an extra byte is added + // to mark it as a stub. + StubLayout Result = {14, 4}; + return Result; +} + +void *X86JITInfo::emitFunctionStub(const Function* F, void *Target, + JITCodeEmitter &JCE) { + // Note, we cast to intptr_t here to silence a -pedantic warning that // complains about casting a function pointer to a normal pointer. #if defined (X86_32_JIT) && !defined (_MSC_VER) - bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback && - Fn != (void*)(intptr_t)X86CompilationCallback_SSE); + bool NotCC = (Target != (void*)(intptr_t)X86CompilationCallback && + Target != (void*)(intptr_t)X86CompilationCallback_SSE); #else - bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback; + bool NotCC = Target != (void*)(intptr_t)X86CompilationCallback; #endif + JCE.emitAlignment(4); + void *Result = (void*)JCE.getCurrentPCValue(); if (NotCC) { #if defined (X86_64_JIT) - MCE.startFunctionStub(F, 13, 4); - MCE.emitByte(0x49); // REX prefix - MCE.emitByte(0xB8+2); // movabsq r10 - MCE.emitWordLE(((unsigned *)&Fn)[0]); - MCE.emitWordLE(((unsigned *)&Fn)[1]); - MCE.emitByte(0x41); // REX prefix - MCE.emitByte(0xFF); // jmpq *r10 - MCE.emitByte(2 | (4 << 3) | (3 << 6)); + JCE.emitByte(0x49); // REX prefix + JCE.emitByte(0xB8+2); // movabsq r10 + JCE.emitWordLE((unsigned)(intptr_t)Target); + JCE.emitWordLE((unsigned)(((intptr_t)Target) >> 32)); + JCE.emitByte(0x41); // REX prefix + JCE.emitByte(0xFF); // jmpq *r10 + JCE.emitByte(2 | (4 << 3) | (3 << 6)); #else - MCE.startFunctionStub(F, 5, 4); - MCE.emitByte(0xE9); - MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4); + JCE.emitByte(0xE9); + JCE.emitWordLE((intptr_t)Target-JCE.getCurrentPCValue()-4); #endif - return MCE.finishFunctionStub(F); + return Result; } #if defined (X86_64_JIT) - MCE.startFunctionStub(F, 14, 4); - MCE.emitByte(0x49); // REX prefix - MCE.emitByte(0xB8+2); // movabsq r10 - MCE.emitWordLE(((unsigned *)&Fn)[0]); - MCE.emitWordLE(((unsigned *)&Fn)[1]); - MCE.emitByte(0x41); // REX prefix - MCE.emitByte(0xFF); // callq *r10 - MCE.emitByte(2 | (2 << 3) | (3 << 6)); + JCE.emitByte(0x49); // REX prefix + JCE.emitByte(0xB8+2); // movabsq r10 + JCE.emitWordLE((unsigned)(intptr_t)Target); + JCE.emitWordLE((unsigned)(((intptr_t)Target) >> 32)); + JCE.emitByte(0x41); // REX prefix + JCE.emitByte(0xFF); // callq *r10 + JCE.emitByte(2 | (2 << 3) | (3 << 6)); #else - MCE.startFunctionStub(F, 6, 4); - MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination... + JCE.emitByte(0xE8); // Call with 32 bit pc-rel destination... - MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4); + JCE.emitWordLE((intptr_t)Target-JCE.getCurrentPCValue()-4); #endif - MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub! - return MCE.finishFunctionStub(F); + // This used to use 0xCD, but that value is used by JITMemoryManager to + // initialize the buffer with garbage, which means it may follow a + // noreturn function call, confusing X86CompilationCallback2. PR 4929. + JCE.emitByte(0xCE); // Interrupt - Just a marker identifying the stub! + return Result; } /// getPICJumpTableEntry - Returns the value of the jumptable entry for the /// specific basic block. -intptr_t X86JITInfo::getPICJumpTableEntry(intptr_t BB, intptr_t Entry) { +uintptr_t X86JITInfo::getPICJumpTableEntry(uintptr_t BB, uintptr_t Entry) { +#if defined(X86_64_JIT) + return BB - Entry; +#else return BB - PICBase; +#endif } /// relocate - Before the JIT can run a block of code that has been emitted, @@ -483,6 +556,7 @@ void X86JITInfo::relocate(void *Function, MachineRelocation *MR, break; } case X86::reloc_absolute_word: + case X86::reloc_absolute_word_sext: // Absolute relocation, just add the relocated value to the value already // in memory. *((unsigned*)RelocPos) += (unsigned)ResultPtr; @@ -493,3 +567,12 @@ void X86JITInfo::relocate(void *Function, MachineRelocation *MR, } } } + +char* X86JITInfo::allocateThreadLocalMemory(size_t size) { +#if defined(X86_32_JIT) && !defined(__APPLE__) && !defined(_MSC_VER) + TLSOffset -= size; + return TLSOffset; +#else + llvm_unreachable("Cannot allocate thread local storage on this arch!"); +#endif +}