X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=f3e5bfe6aac1f592e3aec0d5829de67c757d52f4;hb=6a710247bc1a7b9bced7a4e6f143de91eb4f9552;hp=b798d86e2dcb2a19fc33701ed9bf24ac4aa3f838;hpb=0aca2f00f9bddce3624c1c99e9d2373a3a10a6c3;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index b798d86..f3e5bfe 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -46,64 +46,73 @@ Haskell side. #include #endif -#if defined(openbsd_TARGET_OS) -#include -#include -#include - -/* no C99 header stdint.h on OpenBSD? */ -typedef unsigned long my_uintptr_t; +#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_TARGET_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_TARGET_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 + +#endif -#if defined(i386_TARGET_ARCH) -static unsigned char *obscure_ccall_ret_code; +#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 */ @@ -130,12 +139,21 @@ 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(powerpc64_TARGET_ARCH) -#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" + "mtlr 0\n\t" + "blr"); +extern void obscure_ccall_ret_code(void); +#endif + +#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 @@ -143,14 +161,14 @@ stgAllocStable(size_t size_in_bytes, StgStablePtr *stable) */ 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; @@ -178,15 +196,64 @@ 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) +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) && !defined(darwin_HOST_OS) /* Magic constant computed by inspecting the code length of the following assembly language snippet (offset and machine code prefixed): @@ -199,7 +266,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) : 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 */ @@ -219,7 +286,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) break; case 1: /* _ccall */ -#if defined(i386_TARGET_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): @@ -244,7 +311,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) 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; @@ -255,12 +322,159 @@ 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(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): @@ -290,7 +504,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr, char *typeString) 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; @@ -325,7 +539,7 @@ 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 @@ -348,7 +562,7 @@ 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. @@ -367,7 +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; @@ -383,169 +597,328 @@ 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) -/* - For PowerPC Linux, the following code is used: - - mr r10,r8 - mr r9,r7 - mr r8,r6 - mr r7,r5 - mr r6,r4 - mr r5,r3 - lis r0,0xDEAD ;hi(wptr) - lis r3,0xDEAF ;hi(hptr) - ori r0,r0,0xBEEF ; lo(wptr) - ori r3,r3,0xFACE ; lo(hptr) - mtctr r0 - bctr - - The arguments (passed in registers r3 - r10) are shuffled along by two to - make room for hptr and a dummy argument. As r9 and r10 are overwritten by - this code, it only works for up to 6 arguments (when floating point arguments - are involved, this may be more or less, depending on the exact situation). -*/ - adjustor = mallocBytesRWX(4*13); - { - unsigned long *const adj_code = (unsigned long *)adjustor; - - // make room for extra arguments - adj_code[0] = 0x7d0a4378; //mr r10,r8 - adj_code[1] = 0x7ce93b78; //mr r9,r7 - adj_code[2] = 0x7cc83378; //mr r8,r6 - adj_code[3] = 0x7ca72b78; //mr r7,r5 - adj_code[4] = 0x7c862378; //mr r6,r4 - adj_code[5] = 0x7c651b78; //mr r5,r3 - - adj_code[6] = 0x3c000000; //lis r0,hi(wptr) - adj_code[6] |= ((unsigned long)wptr) >> 16; - - adj_code[7] = 0x3c600000; //lis r3,hi(hptr) - adj_code[7] |= ((unsigned long)hptr) >> 16; - - adj_code[8] = 0x60000000; //ori r0,r0,lo(wptr) - adj_code[8] |= ((unsigned long)wptr) & 0xFFFF; - - adj_code[9] = 0x60630000; //ori r3,r3,lo(hptr) - adj_code[9] |= ((unsigned long)hptr) & 0xFFFF; - - adj_code[10] = 0x7c0903a6; //mtctr r0 - adj_code[11] = 0x4e800420; //bctr - adj_code[12] = (unsigned long)hptr; - - // Flush the Instruction cache: - // MakeDataExecutable(adjustor,4*13); - /* This would require us to link with CoreServices.framework */ - { /* this should do the same: */ - int n = 13; - unsigned long *p = adj_code; - while(n--) - { - __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0" - : : "r" (p)); - p++; - } - __asm__ volatile ("sync\n\tisync"); - } - } +#elif defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS) -#elif defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH) - #define OP_LO(op,lo) ((((unsigned)(op)) << 16) | (((unsigned)(lo)) & 0xFFFF)) #define OP_HI(op,hi) ((((unsigned)(op)) << 16) | (((unsigned)(hi)) >> 16)) + { + /* The PowerPC Linux (32-bit) calling convention is annoyingly complex. + We need to calculate all the details of the stack frame layout, + taking into account the types of all the arguments, and then + generate code on the fly. */ + + int src_gpr = 3, dst_gpr = 5; + int fpr = 3; + int src_offset = 0, dst_offset = 0; + int n = strlen(typeString),i; + int src_locs[n], dst_locs[n]; + int frameSize; + unsigned *code; + + /* Step 1: + Calculate where the arguments should go. + src_locs[] will contain the locations of the arguments in the + original stack frame passed to the adjustor. + dst_locs[] will contain the locations of the arguments after the + adjustor runs, on entry to the wrapper proc pointed to by wptr. + + This algorithm is based on the one described on page 3-19 of the + System V ABI PowerPC Processor Supplement. + */ + for(i=0;typeString[i];i++) + { + char t = typeString[i]; + if((t == 'f' || t == 'd') && fpr <= 8) + src_locs[i] = dst_locs[i] = -32-(fpr++); + else + { + if(t == 'l' && src_gpr <= 9) + { + if((src_gpr & 1) == 0) + src_gpr++; + src_locs[i] = -src_gpr; + src_gpr += 2; + } + else if(t == 'i' && src_gpr <= 10) + { + src_locs[i] = -(src_gpr++); + } + else + { + if(t == 'l' || t == 'd') + { + if(src_offset % 8) + src_offset += 4; + } + src_locs[i] = src_offset; + src_offset += (t == 'l' || t == 'd') ? 8 : 4; + } + + if(t == 'l' && dst_gpr <= 9) + { + if((dst_gpr & 1) == 0) + dst_gpr++; + dst_locs[i] = -dst_gpr; + dst_gpr += 2; + } + else if(t == 'i' && dst_gpr <= 10) + { + dst_locs[i] = -(dst_gpr++); + } + else + { + if(t == 'l' || t == 'd') + { + if(dst_offset % 8) + dst_offset += 4; + } + dst_locs[i] = dst_offset; + dst_offset += (t == 'l' || t == 'd') ? 8 : 4; + } + } + } + + frameSize = dst_offset + 8; + frameSize = (frameSize+15) & ~0xF; + + /* Step 2: + Build the adjustor. + */ + // allocate space for at most 4 insns per parameter + // plus 14 more instructions. + adjustor = stgMallocBytesRWX(4 * (4*n + 14)); + code = (unsigned*)adjustor; + + *code++ = 0x48000008; // b *+8 + // * Put the hptr in a place where freeHaskellFunctionPtr + // can get at it. + *code++ = (unsigned) hptr; + + // * save the link register + *code++ = 0x7c0802a6; // mflr r0; + *code++ = 0x90010004; // stw r0, 4(r1); + // * and build a new stack frame + *code++ = OP_LO(0x9421, -frameSize); // stwu r1, -frameSize(r1) + + // * now generate instructions to copy arguments + // from the old stack frame into the new stack frame. + for(i=n-1;i>=0;i--) { - AdjustorStub *adjustorStub; - int sz = 0, extra_sz, total_sz; + if(src_locs[i] < -32) + ASSERT(dst_locs[i] == src_locs[i]); + else if(src_locs[i] < 0) + { + // source in GPR. + ASSERT(typeString[i] != 'f' && typeString[i] != 'd'); + if(dst_locs[i] < 0) + { + ASSERT(dst_locs[i] > -32); + // dst is in GPR, too. + + if(typeString[i] == 'l') + { + // mr dst+1, src+1 + *code++ = 0x7c000378 + | ((-dst_locs[i]+1) << 16) + | ((-src_locs[i]+1) << 11) + | ((-src_locs[i]+1) << 21); + } + // mr dst, src + *code++ = 0x7c000378 + | ((-dst_locs[i]) << 16) + | ((-src_locs[i]) << 11) + | ((-src_locs[i]) << 21); + } + else + { + if(typeString[i] == 'l') + { + // stw src+1, dst_offset+4(r1) + *code++ = 0x90010000 + | ((-src_locs[i]+1) << 21) + | (dst_locs[i] + 4); + } + + // stw src, dst_offset(r1) + *code++ = 0x90010000 + | ((-src_locs[i]) << 21) + | (dst_locs[i] + 8); + } + } + else + { + ASSERT(dst_locs[i] >= 0); + ASSERT(typeString[i] != 'f' && typeString[i] != 'd'); - // from AdjustorAsm.s - // not declared as a function so that AIX-style - // fundescs can never get in the way. - extern void *adjustorCode; - + if(typeString[i] == 'l') + { + // lwz r0, src_offset(r1) + *code++ = 0x80010000 + | (src_locs[i] + frameSize + 8 + 4); + // stw r0, dst_offset(r1) + *code++ = 0x90010000 + | (dst_locs[i] + 8 + 4); + } + // lwz r0, src_offset(r1) + *code++ = 0x80010000 + | (src_locs[i] + frameSize + 8); + // stw r0, dst_offset(r1) + *code++ = 0x90010000 + | (dst_locs[i] + 8); + } + } + + // * hptr will be the new first argument. + // lis r3, hi(hptr) + *code++ = OP_HI(0x3c60, hptr); + // ori r3,r3,lo(hptr) + *code++ = OP_LO(0x6063, hptr); + + // * we need to return to a piece of code + // which will tear down the stack frame. + // lis r11,hi(obscure_ccall_ret_code) + *code++ = OP_HI(0x3d60, obscure_ccall_ret_code); + // ori r11,r11,lo(obscure_ccall_ret_code) + *code++ = OP_LO(0x616b, obscure_ccall_ret_code); + // mtlr r11 + *code++ = 0x7d6803a6; + + // * jump to wptr + // lis r11,hi(wptr) + *code++ = OP_HI(0x3d60, wptr); + // ori r11,r11,lo(wptr) + *code++ = OP_LO(0x616b, wptr); + // mtctr r11 + *code++ = 0x7d6903a6; + // bctr + *code++ = 0x4e800420; + + // Flush the Instruction cache: + { + unsigned *p = adjustor; + while(p < code) + { + __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0" + : : "r" (p)); + p++; + } + __asm__ volatile ("sync\n\tisync"); + } + } + +#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)) + { + /* 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 = stgMallocBytesRWX(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++; - - 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; -#endif - case 'f': sz += 1; break; - case 'i': sz += 1; break; - } - } - 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; - } + // Calculate the size of the stack frame, in words. + 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 + // 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. @@ -642,7 +1015,7 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for void freeHaskellFunctionPtr(void* ptr) { -#if defined(i386_TARGET_ARCH) +#if defined(i386_HOST_ARCH) && !defined(darwin_HOST_OS) if ( *(unsigned char*)ptr != 0x68 && *(unsigned char*)ptr != 0x58 ) { errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); @@ -654,8 +1027,23 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x01))); } else { freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02))); - } -#elif defined(sparc_TARGET_ARCH) + } +#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); return; @@ -663,7 +1051,7 @@ 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; @@ -671,20 +1059,20 @@ freeHaskellFunctionPtr(void* ptr) /* Free the stable pointer first..*/ freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10))); -#elif defined(powerpc_TARGET_ARCH) && defined(linux_TARGET_OS) - if ( *(StgWord*)ptr != 0x7d0a4378 ) { +#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*)((unsigned char*)ptr + 4*12))); -#elif defined(powerpc_TARGET_ARCH) || defined(powerpc64_TARGET_ARCH) + freeStablePtr(((StgStablePtr*)ptr)[1]); +#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); @@ -712,30 +1100,11 @@ 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 = 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 }