X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=f3e5bfe6aac1f592e3aec0d5829de67c757d52f4;hb=28a464a75e14cece5db40f2765a29348273ff2d2;hp=59119bfd0796307fba0ea4f14069a4cf02639292;hpb=153b9cb9b11e05c4edb1b6bc0a7b972660e41f70;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index 59119bf..f3e5bfe 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -46,56 +46,61 @@ Haskell side. #include #endif -#if defined(openbsd_HOST_OS) -#include -#include -#include - -/* no C99 header stdint.h on OpenBSD? */ -typedef unsigned long my_uintptr_t; -#endif - #if defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS) #include #endif -/* Heavily arch-specific, I'm afraid.. */ +#ifdef LEADING_UNDERSCORE +#define UNDERSCORE "_" +#else +#define UNDERSCORE "" +#endif +#if defined(i386_HOST_ARCH) && !defined(darwin_HOST_OS) +/* + Now here's something obscure for you: -/* - * Allocate len bytes which are readable, writable, and executable. - * - * ToDo: If this turns out to be a performance bottleneck, one could - * e.g. cache the last VirtualProtect/mprotect-ed region and do - * nothing in case of a cache hit. - */ -static void* -mallocBytesRWX(int len) + When generating an adjustor thunk that uses the C calling + convention, we have to make sure that the thunk kicks off + the process of jumping into Haskell with a tail jump. Why? + Because as a result of jumping in into Haskell we may end + up freeing the very adjustor thunk we came from using + freeHaskellFunctionPtr(). Hence, we better not return to + the adjustor code on our way out, since it could by then + point to junk. + + The fix is readily at hand, just include the opcodes + for the C stack fixup code that we need to perform when + returning in some static piece of memory and arrange + to return to it before tail jumping from the adjustor thunk. +*/ +static void GNUC3_ATTRIBUTE(used) obscure_ccall_wrapper(void) { - void *addr = stgMallocBytes(len, "mallocBytesRWX"); -#if defined(i386_HOST_ARCH) && defined(_WIN32) - /* This could be necessary for processors which distinguish between READ and - EXECUTE memory accesses, e.g. Itaniums. */ - DWORD dwOldProtect = 0; - if (VirtualProtect (addr, len, PAGE_EXECUTE_READWRITE, &dwOldProtect) == 0) { - barf("mallocBytesRWX: failed to protect 0x%p; error=%lu; old protection: %lu\n", - addr, (unsigned long)GetLastError(), (unsigned long)dwOldProtect); - } -#elif defined(openbsd_HOST_OS) - /* malloced memory isn't executable by default on OpenBSD */ - my_uintptr_t pageSize = sysconf(_SC_PAGESIZE); - my_uintptr_t mask = ~(pageSize - 1); - my_uintptr_t startOfFirstPage = ((my_uintptr_t)addr ) & mask; - my_uintptr_t startOfLastPage = ((my_uintptr_t)addr + len - 1) & mask; - my_uintptr_t size = startOfLastPage - startOfFirstPage + pageSize; - if (mprotect((void*)startOfFirstPage, (size_t)size, PROT_EXEC | PROT_READ | PROT_WRITE) != 0) { - barf("mallocBytesRWX: failed to protect 0x%p\n", addr); - } -#endif - return addr; + __asm__ ( + ".globl " UNDERSCORE "obscure_ccall_ret_code\n" + UNDERSCORE "obscure_ccall_ret_code:\n\t" + "addl $0x4, %esp\n\t" + "ret" + ); } +extern void obscure_ccall_ret_code(void); + +#if defined(openbsd_HOST_OS) +static unsigned char *obscure_ccall_ret_code_dyn; +#endif -#if defined(i386_HOST_ARCH) -static unsigned char *obscure_ccall_ret_code; +#endif + +#if defined(x86_64_HOST_ARCH) +static void GNUC3_ATTRIBUTE(used) obscure_ccall_wrapper(void) +{ + __asm__ ( + ".globl " UNDERSCORE "obscure_ccall_ret_code\n" + UNDERSCORE "obscure_ccall_ret_code:\n\t" + "addq $0x8, %rsp\n\t" + "ret" + ); +} +extern void obscure_ccall_ret_code(void); #endif #if defined(alpha_HOST_ARCH) @@ -134,7 +139,7 @@ stgAllocStable(size_t size_in_bytes, StgStablePtr *stable) *stable = getStablePtr((StgPtr)arr); /* and return a ptr to the goods inside the array */ - return(BYTE_ARR_CTS(arr)); + return(&(arr->payload)); } #endif @@ -191,11 +196,54 @@ typedef struct AdjustorStub { #endif #endif +#if defined(i386_HOST_ARCH) && defined(darwin_HOST_OS) + +/* !!! !!! WARNING: !!! !!! + * This structure is accessed from AdjustorAsm.s + * Any changes here have to be mirrored in the offsets there. + */ + +typedef struct AdjustorStub { + unsigned char call[8]; + StgStablePtr hptr; + StgFunPtr wptr; + StgInt frame_size; + StgInt argument_size; +} AdjustorStub; +#endif + +#if defined(darwin_HOST_OS) || defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH) +static int totalArgumentSize(char *typeString) +{ + int sz = 0; + while(*typeString) + { + char t = *typeString++; + + switch(t) + { + // on 32-bit platforms, Double and Int64 occupy two words. + case 'd': + case 'l': + if(sizeof(void*) == 4) + { + sz += 2; + break; + } + // everything else is one word. + default: + sz += 1; + } + } + return sz; +} +#endif + void* createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString -#if !defined(powerpc_HOST_ARCH) && !defined(powerpc64_HOST_ARCH) +#if !defined(powerpc_HOST_ARCH) && !defined(powerpc64_HOST_ARCH) && !defined(x86_64_HOST_ARCH) STG_UNUSED #endif ) @@ -205,7 +253,7 @@ createAdjustor(int cconv, StgStablePtr hptr, switch (cconv) { case 0: /* _stdcall */ -#if defined(i386_HOST_ARCH) +#if defined(i386_HOST_ARCH) && !defined(darwin_HOST_OS) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed): @@ -218,7 +266,7 @@ createAdjustor(int cconv, StgStablePtr hptr, : ff e0 jmp %eax # and jump to it. # the callee cleans up the stack */ - adjustor = mallocBytesRWX(14); + adjustor = stgMallocBytesRWX(14); { unsigned char *const adj_code = (unsigned char *)adjustor; adj_code[0x00] = (unsigned char)0x58; /* popl %eax */ @@ -238,7 +286,7 @@ createAdjustor(int cconv, StgStablePtr hptr, break; case 1: /* _ccall */ -#if defined(i386_HOST_ARCH) +#if defined(i386_HOST_ARCH) && !defined(darwin_HOST_OS) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed): @@ -263,7 +311,7 @@ createAdjustor(int cconv, StgStablePtr hptr, That's (thankfully) the case here with the restricted set of return types that we support. */ - adjustor = mallocBytesRWX(17); + adjustor = stgMallocBytesRWX(17); { unsigned char *const adj_code = (unsigned char *)adjustor; @@ -274,11 +322,158 @@ createAdjustor(int cconv, StgStablePtr hptr, *((StgFunPtr*)(adj_code + 0x06)) = (StgFunPtr)wptr; adj_code[0x0a] = (unsigned char)0x68; /* pushl obscure_ccall_ret_code */ - *((StgFunPtr*)(adj_code + 0x0b)) = (StgFunPtr)obscure_ccall_ret_code; + *((StgFunPtr*)(adj_code + 0x0b)) = +#if !defined(openbsd_HOST_OS) + (StgFunPtr)obscure_ccall_ret_code; +#else + (StgFunPtr)obscure_ccall_ret_code_dyn; +#endif adj_code[0x0f] = (unsigned char)0xff; /* jmp *%eax */ adj_code[0x10] = (unsigned char)0xe0; } +#elif defined(i386_HOST_ARCH) && defined(darwin_HOST_OS) + { + /* + What's special about Darwin/Mac OS X on i386? + It wants the stack to stay 16-byte aligned. + + We offload most of the work to AdjustorAsm.S. + */ + AdjustorStub *adjustorStub = stgMallocBytesRWX(sizeof(AdjustorStub)); + adjustor = adjustorStub; + + extern void adjustorCode(void); + int sz = totalArgumentSize(typeString); + + adjustorStub->call[0] = 0xe8; + *(long*)&adjustorStub->call[1] = ((char*)&adjustorCode) - ((char*)adjustorStub + 5); + adjustorStub->hptr = hptr; + adjustorStub->wptr = wptr; + + // The adjustor puts the following things on the stack: + // 1.) %ebp link + // 2.) padding and (a copy of) the arguments + // 3.) a dummy argument + // 4.) hptr + // 5.) return address (for returning to the adjustor) + // All these have to add up to a multiple of 16. + + // first, include everything in frame_size + adjustorStub->frame_size = sz * 4 + 16; + // align to 16 bytes + adjustorStub->frame_size = (adjustorStub->frame_size + 15) & ~15; + // only count 2.) and 3.) as part of frame_size + adjustorStub->frame_size -= 12; + adjustorStub->argument_size = sz; + } + +#elif defined(x86_64_HOST_ARCH) + /* + stack at call: + argn + ... + arg7 + return address + %rdi,%rsi,%rdx,%rcx,%r8,%r9 = arg0..arg6 + + if there are <6 integer args, then we can just push the + StablePtr into %edi and shuffle the other args up. + + If there are >=6 integer args, then we have to flush one arg + to the stack, and arrange to adjust the stack ptr on return. + The stack will be rearranged to this: + + argn + ... + arg7 + return address *** <-- dummy arg in stub fn. + arg6 + obscure_ccall_ret_code + + This unfortunately means that the type of the stub function + must have a dummy argument for the original return address + pointer inserted just after the 6th integer argument. + + Code for the simple case: + + 0: 4d 89 c1 mov %r8,%r9 + 3: 49 89 c8 mov %rcx,%r8 + 6: 48 89 d1 mov %rdx,%rcx + 9: 48 89 f2 mov %rsi,%rdx + c: 48 89 fe mov %rdi,%rsi + f: 48 8b 3d 0a 00 00 00 mov 10(%rip),%rdi + 16: ff 25 0c 00 00 00 jmpq *12(%rip) + ... + 20: .quad 0 # aligned on 8-byte boundary + 28: .quad 0 # aligned on 8-byte boundary + + + And the version for >=6 integer arguments: + + 0: 41 51 push %r9 + 2: ff 35 20 00 00 00 pushq 32(%rip) # 28 + 8: 4d 89 c1 mov %r8,%r9 + b: 49 89 c8 mov %rcx,%r8 + e: 48 89 d1 mov %rdx,%rcx + 11: 48 89 f2 mov %rsi,%rdx + 14: 48 89 fe mov %rdi,%rsi + 17: 48 8b 3d 12 00 00 00 mov 18(%rip),%rdi # 30 + 1e: ff 25 14 00 00 00 jmpq *20(%rip) # 38 + ... + 28: .quad 0 # aligned on 8-byte boundary + 30: .quad 0 # aligned on 8-byte boundary + 38: .quad 0 # aligned on 8-byte boundary + */ + + /* we assume the small code model (gcc -mcmmodel=small) where + * all symbols are <2^32, so hence wptr should fit into 32 bits. + */ + ASSERT(((long)wptr >> 32) == 0); + + { + int i = 0; + char *c; + + // determine whether we have 6 or more integer arguments, + // and therefore need to flush one to the stack. + for (c = typeString; *c != '\0'; c++) { + if (*c == 'i' || *c == 'l') i++; + if (i == 6) break; + } + + if (i < 6) { + adjustor = stgMallocBytesRWX(0x30); + + *(StgInt32 *)adjustor = 0x49c1894d; + *(StgInt32 *)(adjustor+0x4) = 0x8948c889; + *(StgInt32 *)(adjustor+0x8) = 0xf28948d1; + *(StgInt32 *)(adjustor+0xc) = 0x48fe8948; + *(StgInt32 *)(adjustor+0x10) = 0x000a3d8b; + *(StgInt32 *)(adjustor+0x14) = 0x25ff0000; + *(StgInt32 *)(adjustor+0x18) = 0x0000000c; + *(StgInt64 *)(adjustor+0x20) = (StgInt64)hptr; + *(StgInt64 *)(adjustor+0x28) = (StgInt64)wptr; + } + else + { + adjustor = stgMallocBytesRWX(0x40); + + *(StgInt32 *)adjustor = 0x35ff5141; + *(StgInt32 *)(adjustor+0x4) = 0x00000020; + *(StgInt32 *)(adjustor+0x8) = 0x49c1894d; + *(StgInt32 *)(adjustor+0xc) = 0x8948c889; + *(StgInt32 *)(adjustor+0x10) = 0xf28948d1; + *(StgInt32 *)(adjustor+0x14) = 0x48fe8948; + *(StgInt32 *)(adjustor+0x18) = 0x00123d8b; + *(StgInt32 *)(adjustor+0x1c) = 0x25ff0000; + *(StgInt32 *)(adjustor+0x20) = 0x00000014; + + *(StgInt64 *)(adjustor+0x28) = (StgInt64)obscure_ccall_ret_code; + *(StgInt64 *)(adjustor+0x30) = (StgInt64)hptr; + *(StgInt64 *)(adjustor+0x38) = (StgInt64)wptr; + } + } #elif defined(sparc_HOST_ARCH) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed): @@ -309,7 +504,7 @@ createAdjustor(int cconv, StgStablePtr hptr, similarly, and local variables should be accessed via %fp, not %sp. In a nutshell: This should work! (Famous last words! :-) */ - adjustor = mallocBytesRWX(4*(11+1)); + adjustor = stgMallocBytesRWX(4*(11+1)); { unsigned long *const adj_code = (unsigned long *)adjustor; @@ -386,7 +581,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for 4 bytes (getting rid of the nop), hence saving memory. [ccshan] */ ASSERT(((StgWord64)wptr & 3) == 0); - adjustor = mallocBytesRWX(48); + adjustor = stgMallocBytesRWX(48); { StgWord64 *const code = (StgWord64 *)adjustor; @@ -491,7 +686,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for */ // allocate space for at most 4 insns per parameter // plus 14 more instructions. - adjustor = mallocBytesRWX(4 * (4*n + 14)); + adjustor = stgMallocBytesRWX(4 * (4*n + 14)); code = (unsigned*)adjustor; *code++ = 0x48000008; // b *+8 @@ -650,7 +845,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for #ifdef FUNDESCS adjustorStub = stgMallocBytes(sizeof(AdjustorStub), "createAdjustor"); #else - adjustorStub = mallocBytesRWX(sizeof(AdjustorStub)); + adjustorStub = stgMallocBytesRWX(sizeof(AdjustorStub)); #endif adjustor = adjustorStub; @@ -695,24 +890,8 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for #endif // Calculate the size of the stack frame, in words. - while(*typeString) - { - char t = *typeString++; - - switch(t) - { -#if defined(powerpc_HOST_ARCH) - // on 32-bit platforms, Double and Int64 occupy two words. - case 'd': - case 'l': - sz += 2; - break; -#endif - // everything else is one word. - default: - sz += 1; - } - } + sz = totalArgumentSize(typeString); + // The first eight words of the parameter area // are just "backing store" for the parameters passed in // the GPRs. extra_sz is the number of words beyond those first @@ -836,7 +1015,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for void freeHaskellFunctionPtr(void* ptr) { -#if defined(i386_HOST_ARCH) +#if defined(i386_HOST_ARCH) && !defined(darwin_HOST_OS) if ( *(unsigned char*)ptr != 0x68 && *(unsigned char*)ptr != 0x58 ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); @@ -848,7 +1027,22 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x01))); } else { freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02))); - } + } +#elif defined(x86_TARGET_ARCH) && defined(darwin_HOST_OS) +if ( *(unsigned char*)ptr != 0xe8 ) { + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + return; + } + freeStablePtr(((AdjustorStub*)ptr)->hptr); +#elif defined(x86_64_HOST_ARCH) + if ( *(StgWord16 *)ptr == 0x894d ) { + freeStablePtr(*(StgStablePtr*)(ptr+0x20)); + } else if ( *(StgWord16 *)ptr == 0x5141 ) { + freeStablePtr(*(StgStablePtr*)(ptr+0x30)); + } else { + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + return; + } #elif defined(sparc_HOST_ARCH) if ( *(unsigned long*)ptr != 0x9C23A008UL ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); @@ -906,30 +1100,11 @@ freeHaskellFunctionPtr(void* ptr) void initAdjustor(void) { -#if defined(i386_HOST_ARCH) - /* Now here's something obscure for you: - - When generating an adjustor thunk that uses the C calling - convention, we have to make sure that the thunk kicks off - the process of jumping into Haskell with a tail jump. Why? - Because as a result of jumping in into Haskell we may end - up freeing the very adjustor thunk we came from using - freeHaskellFunctionPtr(). Hence, we better not return to - the adjustor code on our way out, since it could by then - point to junk. - - The fix is readily at hand, just include the opcodes - for the C stack fixup code that we need to perform when - returning in some static piece of memory and arrange - to return to it before tail jumping from the adjustor thunk. - */ - - obscure_ccall_ret_code = mallocBytesRWX(4); - - obscure_ccall_ret_code[0x00] = (unsigned char)0x83; /* addl $0x4, %esp */ - obscure_ccall_ret_code[0x01] = (unsigned char)0xc4; - obscure_ccall_ret_code[0x02] = (unsigned char)0x04; - - obscure_ccall_ret_code[0x03] = (unsigned char)0xc3; /* ret */ +#if defined(i386_HOST_ARCH) && defined(openbsd_HOST_OS) + obscure_ccall_ret_code_dyn = stgMallocBytesRWX(4); + obscure_ccall_ret_code_dyn[0] = ((unsigned char *)obscure_ccall_ret_code)[0]; + obscure_ccall_ret_code_dyn[1] = ((unsigned char *)obscure_ccall_ret_code)[1]; + obscure_ccall_ret_code_dyn[2] = ((unsigned char *)obscure_ccall_ret_code)[2]; + obscure_ccall_ret_code_dyn[3] = ((unsigned char *)obscure_ccall_ret_code)[3]; #endif }