X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=ea91d9045d60bd813db17a1d5875c0619f13786f;hb=45bb77c6deeb9fc61e31e5b6975fe694d6642aba;hp=2ea1240975567c309b4e3a612a68d96b1e05be60;hpb=5926dc51030efef71abedb53d7d17371d48ea2c2;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index 2ea1240..ea91d90 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -38,42 +38,98 @@ Haskell side. */ #include "PosixSource.h" #include "Rts.h" +#include "RtsExternal.h" #include "RtsUtils.h" -#include "RtsFlags.h" - #include +#if defined(_WIN32) +#include +#endif + /* Heavily arch-specific, I'm afraid.. */ +typedef enum { + pageExecuteRead, + pageExecuteReadWrite +} pageMode; + +/* + * Function: execPage() + * + * Set the executable bit on page containing addr. + * + * TODO: Can the code span more than one page? If yes, we need to make two + * pages executable! + */ +static rtsBool +execPage (void* addr, pageMode mode) +{ +#if defined(i386_TARGET_ARCH) && defined(_WIN32) && 0 + SYSTEM_INFO sInfo; + DWORD dwOldProtect = 0; + + /* doesn't return a result, so presumably it can't fail... */ + GetSystemInfo(&sInfo); + + if ( VirtualProtect ( (void*)((unsigned long)addr & (sInfo.dwPageSize - 1)), + sInfo.dwPageSize, + ( mode == pageExecuteReadWrite ? PAGE_EXECUTE_READWRITE : PAGE_EXECUTE_READ), + &dwOldProtect) == 0 ) { +# if 1 + DWORD rc = GetLastError(); + prog_belch("execPage: failed to protect 0x%p; error=%lu; old protection: %lu\n", addr, rc, dwOldProtect); +# endif + return rtsFalse; + } + return rtsTrue; +#else + (void)addr; (void)mode; /* keep gcc -Wall happy */ + return rtsTrue; +#endif +} + #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. - - For this to work we make the assumption that bytes in .data - are considered executable. -*/ -static unsigned char __obscure_ccall_ret_code [] = - { 0x83, 0xc4, 0x04 /* addl $0x4, %esp */ - , 0xc3 /* ret */ - }; +static unsigned char *obscure_ccall_ret_code; #endif #if defined(alpha_TARGET_ARCH) /* To get the definition of PAL_imb: */ -#include +# if defined(linux_TARGET_OS) +# include +# else +# include +# endif +#endif + +#if defined(ia64_TARGET_ARCH) +#include "Storage.h" + +/* Layout of a function descriptor */ +typedef struct _IA64FunDesc { + StgWord64 ip; + StgWord64 gp; +} IA64FunDesc; + +static void * +stgAllocStable(size_t size_in_bytes, StgStablePtr *stable) +{ + StgArrWords* arr; + nat data_size_in_words, total_size_in_words; + + /* round up to a whole number of words */ + data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_); + total_size_in_words = sizeofW(StgArrWords) + data_size_in_words; + + /* allocate and fill it in */ + arr = (StgArrWords *)allocate(total_size_in_words); + SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words); + + /* obtain a stable ptr */ + *stable = getStablePtr((StgPtr)arr); + + /* and return a ptr to the goods inside the array */ + return(BYTE_ARR_CTS(arr)); +} #endif void* @@ -111,6 +167,8 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) adj_code[0x0c] = (unsigned char)0xff; /* jmp %eax */ adj_code[0x0d] = (unsigned char)0xe0; + + execPage(adjustor, pageExecuteReadWrite); } #endif break; @@ -124,7 +182,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) <00>: 68 ef be ad de pushl $0xdeadbeef # constant is large enough to # hold a StgStablePtr <05>: b8 fa ef ff 00 movl $0x00ffeffa, %eax # load up wptr - <0a>: 68 ef be ad de pushl $__obscure_ccall_ret_code # push the return address + <0a>: 68 ef be ad de pushl $obscure_ccall_ret_code # push the return address <0f>: ff e0 jmp *%eax # jump to wptr The ccall'ing version is a tad different, passing in the return @@ -132,7 +190,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) via the stable pointer.) (The auto-generated C stub is in on this game, don't worry :-) - See the comment next to __obscure_ccall_ret_code why we need to + See the comment next to obscure_ccall_ret_code why we need to perform a tail jump instead of a call, followed by some C stack fixup. @@ -150,11 +208,13 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) adj_code[0x05] = (unsigned char)0xb8; /* movl $wptr, %eax */ *((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; + adj_code[0x0a] = (unsigned char)0x68; /* pushl obscure_ccall_ret_code */ + *((StgFunPtr*)(adj_code + 0x0b)) = (StgFunPtr)obscure_ccall_ret_code; adj_code[0x0f] = (unsigned char)0xff; /* jmp *%eax */ adj_code[0x10] = (unsigned char)0xe0; + + execPage(adjustor, pageExecuteReadWrite); } #elif defined(sparc_TARGET_ARCH) /* Magic constant computed by inspecting the code length of the following @@ -335,12 +395,91 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for while(n--) { __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0" - : : "g" (p)); + : : "r" (p)); p++; } __asm__ volatile ("sync\n\tisync"); } } +#elif defined(ia64_TARGET_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. + We then shuffle the others along by 4 (taking 2 registers for ourselves + to save return address and previous function state - we need to come back + here on the way out to restore the stack, so this is a real function + rather than just a trampoline). + + The function descriptor we create contains the gp of the target function + so gp is already loaded correctly. + + [MLX] alloc r16=ar.pfs,10,2,0 + movl r17=wptr + [MII] st8.spill [r12]=r38,8 // spill in6 (out4) + mov r41=r37 // out7 = in5 (out3) + mov r40=r36;; // out6 = in4 (out2) + [MII] st8.spill [r12]=r39 // spill in7 (out5) + mov.sptk b6=r17,50 + mov r38=r34;; // out4 = in2 (out0) + [MII] mov r39=r35 // out5 = in3 (out1) + mov r37=r33 // out3 = in1 (loc1) + mov r36=r32 // out2 = in0 (loc0) + [MLX] adds r12=-24,r12 // update sp + movl r34=hptr;; // out0 = hptr + [MIB] mov r33=r16 // loc1 = ar.pfs + mov r32=b0 // loc0 = retaddr + br.call.sptk.many b0=b6;; + + [MII] adds r12=-16,r12 + mov b0=r32 + mov.i ar.pfs=r33 + [MFB] nop.m 0x0 + nop.f 0x0 + br.ret.sptk.many b0;; +*/ + +/* These macros distribute a long constant into the two words of an MLX bundle */ +#define BITS(val,start,count) (((val) >> (start)) & ((1 << (count))-1)) +#define MOVL_LOWORD(val) (BITS(val,22,18) << 46) +#define MOVL_HIWORD(val) (BITS(val,40,23) | (BITS(val,0,7) << 36) | (BITS(val,7,9) << 50) \ + | (BITS(val,16,5) << 55) | (BITS(val,21,1) << 44) | BITS(val,63,1) << 59) + + { + StgStablePtr stable; + IA64FunDesc *wdesc = (IA64FunDesc *)wptr; + StgWord64 wcode = wdesc->ip; + IA64FunDesc *fdesc; + StgWord64 *code; + + /* we allocate on the Haskell heap since malloc'd memory isn't executable - argh */ + adjustor = stgAllocStable(sizeof(IA64FunDesc)+18*8, &stable); + + fdesc = (IA64FunDesc *)adjustor; + code = (StgWord64 *)(fdesc + 1); + fdesc->ip = (StgWord64)code; + fdesc->gp = wdesc->gp; + + code[0] = 0x0000058004288004 | MOVL_LOWORD(wcode); + code[1] = 0x6000000220000000 | MOVL_HIWORD(wcode); + code[2] = 0x029015d818984001; + code[3] = 0x8401200500420094; + code[4] = 0x886011d8189c0001; + code[5] = 0x84011004c00380c0; + code[6] = 0x0250210046013800; + code[7] = 0x8401000480420084; + code[8] = 0x0000233f19a06005 | MOVL_LOWORD((StgWord64)hptr); + code[9] = 0x6000000440000000 | MOVL_HIWORD((StgWord64)hptr); + code[10] = 0x0200210020010811; + code[11] = 0x1080006800006200; + code[12] = 0x0000210018406000; + code[13] = 0x00aa021000038005; + code[14] = 0x000000010000001d; + code[15] = 0x0084000880000200; + + /* save stable pointers in convenient form */ + code[16] = (StgWord64)hptr; + code[17] = (StgWord64)stable; + } #else barf("adjustor creation not supported on this platform"); #endif @@ -362,7 +501,7 @@ freeHaskellFunctionPtr(void* ptr) #if defined(i386_TARGET_ARCH) if ( *(unsigned char*)ptr != 0x68 && *(unsigned char*)ptr != 0x58 ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } @@ -374,7 +513,7 @@ freeHaskellFunctionPtr(void* ptr) } #elif defined(sparc_TARGET_ARCH) if ( *(unsigned long*)ptr != 0x9C23A008UL ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } @@ -382,7 +521,7 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 11))); #elif defined(alpha_TARGET_ARCH) if ( *(StgWord64*)ptr != 0xa77b0018a61b0010L ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } @@ -390,15 +529,65 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10))); #elif defined(powerpc_TARGET_ARCH) if ( *(StgWord*)ptr != 0x7d0a4378 ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 4*12))); +#elif defined(ia64_TARGET_ARCH) + IA64FunDesc *fdesc = (IA64FunDesc *)ptr; + StgWord64 *code = (StgWord64 *)(fdesc+1); + + if (fdesc->ip != (StgWord64)code) { + prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + return; + } + freeStablePtr((StgStablePtr)code[16]); + freeStablePtr((StgStablePtr)code[17]); + return; #else ASSERT(0); #endif *((unsigned char*)ptr) = '\0'; - free(ptr); + stgFree(ptr); } + +/* + * Function: initAdjustor() + * + * Perform initialisation of adjustor thunk layer (if needed.) + */ +rtsBool +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 = stgMallocBytes(4, "initAdjustor"); + + 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 */ + + execPage(obscure_ccall_ret_code, pageExecuteRead); +#endif + return rtsTrue; +}