X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=59119bfd0796307fba0ea4f14069a4cf02639292;hb=153b9cb9b11e05c4edb1b6bc0a7b972660e41f70;hp=b0448c794261dc73ff6a1cbb9b121b354fd410a4;hpb=b45a6d2d410e10f971086e3d8269f444d1823268;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index b0448c7..59119bf 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -34,56 +34,178 @@ An adjustor thunk differs from a C function pointer in one respect: when the code is through with it, it has to be freed in order to release Haskell and C resources. Failure to do so result in memory leaks on both the C and Haskell side. - */ + #include "PosixSource.h" #include "Rts.h" +#include "RtsExternal.h" #include "RtsUtils.h" -#include "RtsFlags.h" +#include + +#if defined(_WIN32) +#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.. */ -#if defined(i386_TARGET_ARCH) || defined(sparc_TARGET_ARCH) || defined(alpha_TARGET_ARCH) || defined(powerpc_TARGET_ARCH) - -#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 */ - }; + +/* + * 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) +{ + 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; +} + +#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: */ -#include +# if defined(linux_HOST_OS) +# include +# else +# include +# endif +#endif + +#if defined(ia64_HOST_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 + +#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 + * Any changes here have to be mirrored in the offsets there. + */ + +typedef struct AdjustorStub { +#if defined(powerpc_HOST_ARCH) && defined(darwin_HOST_OS) + unsigned lis; + unsigned ori; + unsigned lwz; + unsigned mtctr; + unsigned bctr; + StgFunPtr code; +#elif defined(powerpc64_HOST_ARCH) && defined(darwin_HOST_OS) + /* powerpc64-darwin: just guessing that it won't use fundescs. */ + unsigned lis; + unsigned ori; + unsigned rldimi; + unsigned oris; + unsigned ori2; + unsigned lwz; + unsigned mtctr; + unsigned bctr; + StgFunPtr code; +#else + /* fundesc-based ABIs */ +#define FUNDESCS + StgFunPtr code; + struct AdjustorStub + *toc; + void *env; +#endif + StgStablePtr hptr; + StgFunPtr wptr; + StgInt negative_framesize; + StgInt extrawords_plus_one; +} AdjustorStub; + +#endif #endif void* -createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) +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): @@ -96,7 +218,8 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) : ff e0 jmp %eax # and jump to it. # the callee cleans up the stack */ - if ((adjustor = stgMallocBytes(14, "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(14); + { unsigned char *const adj_code = (unsigned char *)adjustor; adj_code[0x00] = (unsigned char)0x58; /* popl %eax */ @@ -115,7 +238,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) 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): @@ -123,7 +246,7 @@ 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 @@ -131,7 +254,7 @@ 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. @@ -140,7 +263,8 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) That's (thankfully) the case here with the restricted set of return types that we support. */ - if ((adjustor = stgMallocBytes(17, "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(17); + { unsigned char *const adj_code = (unsigned char *)adjustor; adj_code[0x00] = (unsigned char)0x68; /* pushl hptr (which is a dword immediate ) */ @@ -149,81 +273,78 @@ 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; } -#elif defined(sparc_TARGET_ARCH) - /* Magic constant computed by inspecting the code length of - the following assembly language snippet - (offset and machine code prefixed): - - <00>: BA 10 00 1B mov %i3, %i5 - <04>: B8 10 00 1A mov %i2, %i4 - <08>: B6 10 00 19 mov %i1, %i3 - <0c>: B4 10 00 18 mov %i0, %i2 - <10>: 13 00 3f fb sethi %hi(0x00ffeffa), %o1 # load up wptr (1 of 2) - <14>: 11 37 ab 6f sethi %hi(0xdeadbeef), %o0 # load up hptr (1 of 2) - <18>: 81 c2 63 fa jmp %o1+%lo(0x00ffeffa) # jump to wptr (load 2 of 2) - <1c>: 90 12 22 ef or %o0, %lo(0xdeadbeef), %o0 # load up hptr (2 of 2) - # [in delay slot] - <20>: de ad be ef # Place the value of the StgStablePtr somewhere readable - - ccall'ing on a SPARC leaves little to be performed by the caller. - The callee shifts the window on entry and restores it on exit. - Input paramters and results are passed via registers. (%o0 in the - code above contains the input paramter to wptr.) The return address - is stored in %o7/%i7. Since we don't shift the window in this code, - the return address is preserved and wptr will return to our caller. - - JRS, 21 Aug 01: the above para is a fiction. The caller passes - args in %i0 .. %i5 and then the rest at [%sp+92]. We want to - tailjump to wptr, passing hptr as the new first arg, and a dummy - second arg, which would be where the return address is on x86. - That means we have to shuffle the original caller's args along by - two. - - We do a half-correct solution which works only if the original - caller passed 4 or fewer arg words. Move %i0 .. %i3 into %i3 - .. %i6, so we can park hptr in %i0 and a bogus arg in %i1. The - fully correct solution would be to subtract 8 from %sp and then - place %i4 and %i5 at [%sp+92] and [%sp+96] respectively. This - machinery should then work in all cases. (Or would it? Perhaps - it would trash parts of the caller's frame. Dunno). +#elif defined(sparc_HOST_ARCH) + /* Magic constant computed by inspecting the code length of the following + assembly language snippet (offset and machine code prefixed): + + <00>: 9C23A008 sub %sp, 8, %sp ! make room for %o4/%o5 in caller's frame + <04>: DA23A060 st %o5, [%sp + 96] ! shift registers by 2 positions + <08>: D823A05C st %o4, [%sp + 92] + <0C>: 9A10000B mov %o3, %o5 + <10>: 9810000A mov %o2, %o4 + <14>: 96100009 mov %o1, %o3 + <18>: 94100008 mov %o0, %o2 + <1C>: 13000000 sethi %hi(wptr), %o1 ! load up wptr (1 of 2) + <20>: 11000000 sethi %hi(hptr), %o0 ! load up hptr (1 of 2) + <24>: 81C26000 jmp %o1 + %lo(wptr) ! jump to wptr (load 2 of 2) + <28>: 90122000 or %o0, %lo(hptr), %o0 ! load up hptr (2 of 2, delay slot) + <2C> 00000000 ! place for getting hptr back easily + + ccall'ing on SPARC is easy, because we are quite lucky to push a + multiple of 8 bytes (1 word hptr + 1 word dummy arg) in front of the + existing arguments (note that %sp must stay double-word aligned at + all times, see ABI spec at http://www.sparc.org/standards/psABI3rd.pdf). + To do this, we extend the *caller's* stack frame by 2 words and shift + the output registers used for argument passing (%o0 - %o5, we are a *leaf* + procedure because of the tail-jump) by 2 positions. This makes room in + %o0 and %o1 for the additinal arguments, namely hptr and a dummy (used + for destination addr of jump on SPARC, return address on x86, ...). This + shouldn't cause any problems for a C-like caller: alloca is implemented + similarly, and local variables should be accessed via %fp, not %sp. In a + nutshell: This should work! (Famous last words! :-) */ - if ((adjustor = stgMallocBytes(4*(8+1), "createAdjustor")) != NULL) { - unsigned long *const adj_code = (unsigned long *)adjustor; - - /* mov %o3, %o5 */ - adj_code[0] = (unsigned long)0x9A10000B; - /* mov %o2, %o4 */ - adj_code[1] = (unsigned long)0x9810000A; - /* mov %o1, %o3 */ - adj_code[2] = (unsigned long)0x96100009; - /* mov %o0, %o2 */ - adj_code[3] = (unsigned long)0x94100008; - - /* sethi %hi(wptr), %o1 */ - adj_code[4] = (unsigned long)0x13000000; - adj_code[4] |= ((unsigned long)wptr) >> 10; - - /* sethi %hi(hptr), %o0 */ - adj_code[5] = (unsigned long)0x11000000; - adj_code[5] |= ((unsigned long)hptr) >> 10; - - /* jmp %o1+%lo(wptr) */ - adj_code[6] = (unsigned long)0x81c26000; - adj_code[6] |= ((unsigned long)wptr) & 0x000003ff; - - /* or %o0, %lo(hptr), %o0 */ - adj_code[7] = (unsigned long)0x90122000; - adj_code[7] |= ((unsigned long)hptr) & 0x000003ff; - - adj_code[8] = (StgStablePtr)hptr; + adjustor = mallocBytesRWX(4*(11+1)); + { + unsigned long *const adj_code = (unsigned long *)adjustor; + + adj_code[ 0] = 0x9C23A008UL; /* sub %sp, 8, %sp */ + adj_code[ 1] = 0xDA23A060UL; /* st %o5, [%sp + 96] */ + adj_code[ 2] = 0xD823A05CUL; /* st %o4, [%sp + 92] */ + adj_code[ 3] = 0x9A10000BUL; /* mov %o3, %o5 */ + adj_code[ 4] = 0x9810000AUL; /* mov %o2, %o4 */ + adj_code[ 5] = 0x96100009UL; /* mov %o1, %o3 */ + adj_code[ 6] = 0x94100008UL; /* mov %o0, %o2 */ + adj_code[ 7] = 0x13000000UL; /* sethi %hi(wptr), %o1 */ + adj_code[ 7] |= ((unsigned long)wptr) >> 10; + adj_code[ 8] = 0x11000000UL; /* sethi %hi(hptr), %o0 */ + adj_code[ 8] |= ((unsigned long)hptr) >> 10; + adj_code[ 9] = 0x81C26000UL; /* jmp %o1 + %lo(wptr) */ + adj_code[ 9] |= ((unsigned long)wptr) & 0x000003FFUL; + adj_code[10] = 0x90122000UL; /* or %o0, %lo(hptr), %o0 */ + adj_code[10] |= ((unsigned long)hptr) & 0x000003FFUL; + + adj_code[11] = (unsigned long)hptr; + + /* flush cache */ + asm("flush %0" : : "r" (adj_code )); + asm("flush %0" : : "r" (adj_code + 2)); + asm("flush %0" : : "r" (adj_code + 4)); + asm("flush %0" : : "r" (adj_code + 6)); + asm("flush %0" : : "r" (adj_code + 10)); + + /* max. 5 instructions latency, and we need at >= 1 for returning */ + asm("nop"); + asm("nop"); + 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 @@ -246,7 +367,7 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) 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. @@ -265,7 +386,8 @@ 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); - if ((adjustor = stgMallocBytes(48, "createAdjustor")) != NULL) { + adjustor = mallocBytesRWX(48); + { StgWord64 *const code = (StgWord64 *)adjustor; code[0] = 0x4610041246520414L; @@ -280,72 +402,424 @@ 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) +#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)) + { + /* 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 = mallocBytesRWX(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--) + { + 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'); + + 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"); +#else + adjustorStub = mallocBytesRWX(sizeof(AdjustorStub)); +#endif + 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; +#else + + // 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"); +#endif + + // Flush the Instruction cache: + { + 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"); + } +#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; + } + } + // 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_HOST_ARCH) /* - For PowerPC, 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). + 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;; */ - if ((adjustor = stgMallocBytes(4*13, "createAdjustor")) != NULL) { - 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[6] |= ((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" - : : "g" (p)); - p++; - } - __asm__ volatile ("sync\n\tisync"); - } - } + +/* 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 -#error Adjustor creation is not supported on this platform. + barf("adjustor creation not supported on this platform"); #endif break; @@ -358,15 +832,14 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for return adjustor; } -#endif void freeHaskellFunctionPtr(void* ptr) { -#if defined(i386_TARGET_ARCH) +#if defined(i386_HOST_ARCH) if ( *(unsigned char*)ptr != 0x68 && *(unsigned char*)ptr != 0x58 ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } @@ -376,33 +849,87 @@ freeHaskellFunctionPtr(void* ptr) } else { freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02))); } -#elif defined(sparc_TARGET_ARCH) - if ( *(unsigned char*)ptr != 0x13 ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); +#elif defined(sparc_HOST_ARCH) + if ( *(unsigned long*)ptr != 0x9C23A008UL ) { + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } /* Free the stable pointer first..*/ - freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10))); -#elif defined(alpha_TARGET_ARCH) + freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 11))); +#elif defined(alpha_HOST_ARCH) if ( *(StgWord64*)ptr != 0xa77b0018a61b0010L ) { - fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); + errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } /* Free the stable pointer first..*/ 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); +#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))); + 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_HOST_ARCH) + IA64FunDesc *fdesc = (IA64FunDesc *)ptr; + StgWord64 *code = (StgWord64 *)(fdesc+1); + + if (fdesc->ip != (StgWord64)code) { + errorBelch("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.) + */ +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 */ +#endif +}