X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=59119bfd0796307fba0ea4f14069a4cf02639292;hb=153b9cb9b11e05c4edb1b6bc0a7b972660e41f70;hp=387d792f949ef7d9e90b7debce27243d8937dc26;hpb=4b086a750debcb4a5e1738c00a2f0e6c97c44a23;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index 387d792..59119bf 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -46,7 +46,7 @@ Haskell side. #include #endif -#if defined(openbsd_TARGET_OS) +#if defined(openbsd_HOST_OS) #include #include #include @@ -55,7 +55,7 @@ Haskell side. typedef unsigned long my_uintptr_t; #endif -#if defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS) +#if defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS) #include #endif @@ -72,7 +72,7 @@ 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 +80,7 @@ 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) /* malloced memory isn't executable by default on OpenBSD */ my_uintptr_t pageSize = sysconf(_SC_PAGESIZE); my_uintptr_t mask = ~(pageSize - 1); @@ -94,20 +94,20 @@ mallocBytesRWX(int len) return addr; } -#if defined(i386_TARGET_ARCH) +#if defined(i386_HOST_ARCH) static unsigned char *obscure_ccall_ret_code; #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 */ @@ -138,7 +138,7 @@ stgAllocStable(size_t size_in_bytes, StgStablePtr *stable) } #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 +147,8 @@ __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 +156,14 @@ 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 +192,20 @@ 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) + 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 +238,7 @@ 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): @@ -273,7 +279,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) adj_code[0x0f] = (unsigned char)0xff; /* jmp *%eax */ adj_code[0x10] = (unsigned char)0xe0; } -#elif defined(sparc_TARGET_ARCH) +#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 +344,7 @@ 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 +367,7 @@ 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 +402,7 @@ 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)) @@ -605,103 +611,135 @@ 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)) - { - AdjustorStub *adjustorStub; - int sz = 0, extra_sz, total_sz; - - // from AdjustorAsm.s - // not declared as a function so that AIX-style - // fundescs can never get in the way. - extern void *adjustorCode; - + { + /* The following code applies to all PowerPC and PowerPC64 platforms + whose stack layout is based on the AIX ABI. + + Besides (obviously) AIX, this includes + Mac OS 9 and BeOS/PPC (may they rest in peace), + which use the 32-bit AIX ABI + powerpc64-linux, + which uses the 64-bit AIX ABI + and Darwin (Mac OS X), + which uses the same stack layout as AIX, + but no function descriptors. + + The actual stack-frame shuffling is implemented out-of-line + in the function adjustorCode, in AdjustorAsm.S. + Here, we set up an AdjustorStub structure, which + is a function descriptor (on platforms that have function + descriptors) or a short piece of stub code (on Darwin) to call + adjustorCode with a pointer to the AdjustorStub struct loaded + into register r2. + + One nice thing about this is that there is _no_ code generated at + runtime on the platforms that have function descriptors. + */ + AdjustorStub *adjustorStub; + int sz = 0, extra_sz, total_sz; + + // from AdjustorAsm.s + // not declared as a function so that AIX-style + // fundescs can never get in the way. + extern void *adjustorCode; + #ifdef FUNDESCS - adjustorStub = stgMallocBytes(sizeof(AdjustorStub), "createAdjustor"); + adjustorStub = stgMallocBytes(sizeof(AdjustorStub), "createAdjustor"); #else - adjustorStub = mallocBytesRWX(sizeof(AdjustorStub)); + adjustorStub = mallocBytesRWX(sizeof(AdjustorStub)); #endif - adjustor = adjustorStub; - - adjustorStub->code = (void*) &adjustorCode; + adjustor = adjustorStub; + + adjustorStub->code = (void*) &adjustorCode; #ifdef FUNDESCS - // function descriptors are a cool idea. - // We don't need to generate any code at runtime. - adjustorStub->toc = adjustorStub; + // function descriptors are a cool idea. + // We don't need to generate any code at runtime. + adjustorStub->toc = adjustorStub; #else - // no function descriptors :-( - // We need to do things "by hand". -#if defined(powerpc_TARGET_ARCH) - // lis r2, hi(adjustorStub) - adjustorStub->lis = OP_HI(0x3c40, adjustorStub); - // ori r2, r2, lo(adjustorStub) - adjustorStub->ori = OP_LO(0x6042, adjustorStub); - // lwz r0, code(r2) - adjustorStub->lwz = OP_LO(0x8002, (char*)(&adjustorStub->code) - - (char*)adjustorStub); - // mtctr r0 - adjustorStub->mtctr = 0x7c0903a6; - // bctr - adjustorStub->bctr = 0x4e800420; + // no function descriptors :-( + // We need to do things "by hand". +#if defined(powerpc_HOST_ARCH) + // lis r2, hi(adjustorStub) + adjustorStub->lis = OP_HI(0x3c40, adjustorStub); + // ori r2, r2, lo(adjustorStub) + adjustorStub->ori = OP_LO(0x6042, adjustorStub); + // lwz r0, code(r2) + adjustorStub->lwz = OP_LO(0x8002, (char*)(&adjustorStub->code) + - (char*)adjustorStub); + // mtctr r0 + adjustorStub->mtctr = 0x7c0903a6; + // bctr + adjustorStub->bctr = 0x4e800420; #else - barf("adjustor creation not supported on this platform"); + barf("adjustor creation not supported on this platform"); #endif - // Flush the Instruction cache: + // Flush the Instruction cache: + { + int n = sizeof(AdjustorStub)/sizeof(unsigned); + unsigned *p = (unsigned*)adjustor; + while(n--) { - int n = sizeof(AdjustorStub)/sizeof(unsigned); - unsigned *p = (unsigned*)adjustor; - while(n--) - { - __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0" - : : "r" (p)); - p++; - } - __asm__ volatile ("sync\n\tisync"); + __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0" + : : "r" (p)); + p++; } + __asm__ volatile ("sync\n\tisync"); + } #endif - printf("createAdjustor: %s\n", typeString); - while(*typeString) - { - char t = *typeString++; + // Calculate the size of the stack frame, in words. + while(*typeString) + { + char t = *typeString++; - switch(t) - { -#if defined(powerpc64_TARGET_ARCH) - case 'd': sz += 1; break; - case 'l': sz += 1; break; -#else - case 'd': sz += 2; break; - case 'l': sz += 2; break; + 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 - case 'f': sz += 1; break; - case 'i': sz += 1; break; - } + // everything else is one word. + default: + sz += 1; } - extra_sz = sz - 8; - if(extra_sz < 0) - extra_sz = 0; - total_sz = (6 /* linkage area */ - + 8 /* minimum parameter area */ - + 2 /* two extra arguments */ - + extra_sz)*sizeof(StgWord); - - // align to 16 bytes. - // AIX only requires 8 bytes, but who cares? - total_sz = (total_sz+15) & ~0xF; - - adjustorStub->hptr = hptr; - adjustorStub->wptr = wptr; - adjustorStub->negative_framesize = -total_sz; - adjustorStub->extrawords_plus_one = extra_sz + 1; } + // 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 + // 8 words. + extra_sz = sz - 8; + if(extra_sz < 0) + extra_sz = 0; + + // Calculate the total size of the stack frame. + total_sz = (6 /* linkage area */ + + 8 /* minimum parameter area */ + + 2 /* two extra arguments */ + + extra_sz)*sizeof(StgWord); + + // align to 16 bytes. + // AIX only requires 8 bytes, but who cares? + total_sz = (total_sz+15) & ~0xF; + + // Fill in the information that adjustorCode in AdjustorAsm.S + // will use to create a new stack frame with the additional args. + adjustorStub->hptr = hptr; + adjustorStub->wptr = wptr; + adjustorStub->negative_framesize = -total_sz; + 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. @@ -798,7 +836,7 @@ 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); @@ -811,7 +849,7 @@ freeHaskellFunctionPtr(void* ptr) } else { freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02))); } -#elif defined(sparc_TARGET_ARCH) +#elif defined(sparc_HOST_ARCH) if ( *(unsigned long*)ptr != 0x9C23A008UL ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; @@ -819,7 +857,7 @@ 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; @@ -827,20 +865,20 @@ 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); @@ -868,7 +906,7 @@ freeHaskellFunctionPtr(void* ptr) void initAdjustor(void) { -#if defined(i386_TARGET_ARCH) +#if defined(i386_HOST_ARCH) /* Now here's something obscure for you: When generating an adjustor thunk that uses the C calling