X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=0c45f8c7bf7a4382a50d775938b196d609b72be6;hb=7bb0f34f7e8e75d19a730891de6fe76ce96860d5;hp=bf4540af19727e57d097e2bd728f4886436d9b6b;hpb=25cd0a322939cd661cddb9c8a2264ae9d7eedaaf;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index bf4540a..0c45f8c 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -46,16 +46,21 @@ Haskell side. #include #endif -#if defined(openbsd_TARGET_OS) +#if defined(openbsd_HOST_OS) || defined(linux_HOST_OS) #include #include #include /* no C99 header stdint.h on OpenBSD? */ +#if defined(openbsd_HOST_OS) typedef unsigned long my_uintptr_t; +#else +#include +typedef uintptr_t my_uintptr_t; +#endif #endif -#if defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS) +#if defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS) #include #endif @@ -72,7 +77,7 @@ static void* mallocBytesRWX(int len) { void *addr = stgMallocBytes(len, "mallocBytesRWX"); -#if defined(i386_TARGET_ARCH) && defined(_WIN32) +#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; @@ -80,7 +85,7 @@ mallocBytesRWX(int len) barf("mallocBytesRWX: failed to protect 0x%p; error=%lu; old protection: %lu\n", addr, (unsigned long)GetLastError(), (unsigned long)dwOldProtect); } -#elif defined(openbsd_TARGET_OS) +#elif defined(openbsd_HOST_OS) || defined(linux_HOST_OS) /* malloced memory isn't executable by default on OpenBSD */ my_uintptr_t pageSize = sysconf(_SC_PAGESIZE); my_uintptr_t mask = ~(pageSize - 1); @@ -94,20 +99,69 @@ mallocBytesRWX(int len) return addr; } -#if defined(i386_TARGET_ARCH) -static unsigned char *obscure_ccall_ret_code; +#ifdef LEADING_UNDERSCORE +#define UNDERSCORE "_" +#else +#define UNDERSCORE "" +#endif +#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. +*/ +static void GNUC3_ATTRIBUTE(used) obscure_ccall_wrapper(void) +{ + __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 + +#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_TARGET_ARCH) +#if defined(alpha_HOST_ARCH) /* To get the definition of PAL_imb: */ -# if defined(linux_TARGET_OS) +# if defined(linux_HOST_OS) # include # else # include # endif #endif -#if defined(ia64_TARGET_ARCH) +#if defined(ia64_HOST_ARCH) #include "Storage.h" /* Layout of a function descriptor */ @@ -134,11 +188,11 @@ 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 -#if defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS) +#if defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS) __asm__("obscure_ccall_ret_code:\n\t" "lwz 1,0(1)\n\t" "lwz 0,4(1)\n\t" @@ -147,8 +201,8 @@ __asm__("obscure_ccall_ret_code:\n\t" extern void obscure_ccall_ret_code(void); #endif -#if defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH) -#if !(defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS)) +#if defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH) +#if !(defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS)) /* !!! !!! WARNING: !!! !!! * This structure is accessed from AdjustorAsm.s @@ -156,14 +210,14 @@ extern void obscure_ccall_ret_code(void); */ typedef struct AdjustorStub { -#if defined(powerpc_TARGET_ARCH) && defined(darwin_TARGET_OS) +#if defined(powerpc_HOST_ARCH) && defined(darwin_HOST_OS) unsigned lis; unsigned ori; unsigned lwz; unsigned mtctr; unsigned bctr; StgFunPtr code; -#elif defined(powerpc64_TARGET_ARCH) && defined(darwin_TARGET_OS) +#elif defined(powerpc64_HOST_ARCH) && defined(darwin_HOST_OS) /* powerpc64-darwin: just guessing that it won't use fundescs. */ unsigned lis; unsigned ori; @@ -192,14 +246,20 @@ typedef struct AdjustorStub { #endif void* -createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) +createAdjustor(int cconv, StgStablePtr hptr, + StgFunPtr wptr, + char *typeString +#if !defined(powerpc_HOST_ARCH) && !defined(powerpc64_HOST_ARCH) && !defined(x86_64_HOST_ARCH) + STG_UNUSED +#endif + ) { void *adjustor = NULL; switch (cconv) { case 0: /* _stdcall */ -#if defined(i386_TARGET_ARCH) +#if defined(i386_HOST_ARCH) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed): @@ -232,7 +292,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) break; case 1: /* _ccall */ -#if defined(i386_TARGET_ARCH) +#if defined(i386_HOST_ARCH) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed): @@ -268,12 +328,122 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) *((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(sparc_TARGET_ARCH) +#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: e9 00 00 00 00 jmpq stub_function + ... + 20: .quad 0 # aligned on 8-byte boundary + + + And the version for >=6 integer arguments: + + 0: 41 51 push %r9 + 2: 68 00 00 00 00 pushq $obscure_ccall_ret_code + 7: 4d 89 c1 mov %r8,%r9 + a: 49 89 c8 mov %rcx,%r8 + d: 48 89 d1 mov %rdx,%rcx + 10: 48 89 f2 mov %rsi,%rdx + 13: 48 89 fe mov %rdi,%rsi + 16: 48 8b 3d 0b 00 00 00 mov 11(%rip),%rdi + 1d: e9 00 00 00 00 jmpq stub_function + ... + 28: .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 = mallocBytesRWX(40); + + *(StgInt32 *)adjustor = 0x49c1894d; + *(StgInt32 *)(adjustor+4) = 0x8948c889; + *(StgInt32 *)(adjustor+8) = 0xf28948d1; + *(StgInt32 *)(adjustor+12) = 0x48fe8948; + *(StgInt32 *)(adjustor+16) = 0x000a3d8b; + *(StgInt32 *)(adjustor+20) = 0x00e90000; + + *(StgInt32 *)(adjustor+23) = + (StgInt32)((StgInt64)wptr - (StgInt64)adjustor - 27); + *(StgInt64 *)(adjustor+32) = (StgInt64)hptr; + } + else + { + adjustor = mallocBytesRWX(48); + + *(StgInt32 *)adjustor = 0x00685141; + *(StgInt32 *)(adjustor+4) = 0x4d000000; + *(StgInt32 *)(adjustor+8) = 0x8949c189; + *(StgInt32 *)(adjustor+12) = 0xd18948c8; + *(StgInt32 *)(adjustor+16) = 0x48f28948; + *(StgInt32 *)(adjustor+20) = 0x8b48fe89; + *(StgInt32 *)(adjustor+24) = 0x00000b3d; + *(StgInt32 *)(adjustor+28) = 0x0000e900; + + *(StgInt32 *)(adjustor+3) = + (StgInt32)(StgInt64)obscure_ccall_ret_code; + *(StgInt32 *)(adjustor+30) = + (StgInt32)((StgInt64)wptr - (StgInt64)adjustor - 34); + *(StgInt64 *)(adjustor+40) = (StgInt64)hptr; + } + } +#elif defined(sparc_HOST_ARCH) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed): @@ -338,7 +508,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) asm("nop"); asm("nop"); } -#elif defined(alpha_TARGET_ARCH) +#elif defined(alpha_HOST_ARCH) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed; note that the machine code @@ -361,7 +531,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) divided by 4, taking the lowest 14 bits. We only support passing 4 or fewer argument words, for the same - reason described under sparc_TARGET_ARCH above by JRS, 21 Aug 01. + reason described under sparc_HOST_ARCH above by JRS, 21 Aug 01. On the Alpha the first 6 integer arguments are in a0 through a5, and the rest on the stack. Hence we want to shuffle the original caller's arguments by two. @@ -396,7 +566,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for /* Ensure that instruction cache is consistent with our new code */ __asm__ volatile("call_pal %0" : : "i" (PAL_imb)); } -#elif defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS) +#elif defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS) #define OP_LO(op,lo) ((((unsigned)(op)) << 16) | (((unsigned)(lo)) & 0xFFFF)) #define OP_HI(op,hi) ((((unsigned)(op)) << 16) | (((unsigned)(hi)) >> 16)) @@ -605,7 +775,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for } } -#elif defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH) +#elif defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH) #define OP_LO(op,lo) ((((unsigned)(op)) << 16) | (((unsigned)(lo)) & 0xFFFF)) #define OP_HI(op,hi) ((((unsigned)(op)) << 16) | (((unsigned)(hi)) >> 16)) @@ -658,7 +828,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for // no function descriptors :-( // We need to do things "by hand". -#if defined(powerpc_TARGET_ARCH) +#if defined(powerpc_HOST_ARCH) // lis r2, hi(adjustorStub) adjustorStub->lis = OP_HI(0x3c40, adjustorStub); // ori r2, r2, lo(adjustorStub) @@ -695,7 +865,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for switch(t) { -#if defined(powerpc_TARGET_ARCH) +#if defined(powerpc_HOST_ARCH) // on 32-bit platforms, Double and Int64 occupy two words. case 'd': case 'l': @@ -733,7 +903,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for adjustorStub->extrawords_plus_one = extra_sz + 1; } -#elif defined(ia64_TARGET_ARCH) +#elif defined(ia64_HOST_ARCH) /* Up to 8 inputs are passed in registers. We flush the last two inputs to the stack, initially into the 16-byte scratch region left by the caller. @@ -830,7 +1000,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for void freeHaskellFunctionPtr(void* ptr) { -#if defined(i386_TARGET_ARCH) +#if defined(i386_HOST_ARCH) if ( *(unsigned char*)ptr != 0x68 && *(unsigned char*)ptr != 0x58 ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); @@ -842,8 +1012,17 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x01))); } else { freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02))); - } -#elif defined(sparc_TARGET_ARCH) + } +#elif defined(x86_64_HOST_ARCH) + if ( *(StgWord16 *)ptr == 0x894d ) { + freeStablePtr(*(StgStablePtr*)(ptr+32)); + } else if ( *(StgWord16 *)ptr == 0x5141 ) { + freeStablePtr(*(StgStablePtr*)(ptr+40)); + } 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); return; @@ -851,7 +1030,7 @@ freeHaskellFunctionPtr(void* ptr) /* Free the stable pointer first..*/ freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 11))); -#elif defined(alpha_TARGET_ARCH) +#elif defined(alpha_HOST_ARCH) if ( *(StgWord64*)ptr != 0xa77b0018a61b0010L ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; @@ -859,20 +1038,20 @@ freeHaskellFunctionPtr(void* ptr) /* Free the stable pointer first..*/ freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10))); -#elif defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS) +#elif defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS) if ( *(StgWord*)ptr != 0x48000008 ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } freeStablePtr(((StgStablePtr*)ptr)[1]); -#elif defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH) +#elif defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH) extern void* adjustorCode; if ( ((AdjustorStub*)ptr)->code != (StgFunPtr) &adjustorCode ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } freeStablePtr(((AdjustorStub*)ptr)->hptr); -#elif defined(ia64_TARGET_ARCH) +#elif defined(ia64_HOST_ARCH) IA64FunDesc *fdesc = (IA64FunDesc *)ptr; StgWord64 *code = (StgWord64 *)(fdesc+1); @@ -900,30 +1079,11 @@ freeHaskellFunctionPtr(void* ptr) void initAdjustor(void) { -#if defined(i386_TARGET_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 = mallocBytesRWX(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 }