X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FAdjustor.c;h=2ea1240975567c309b4e3a612a68d96b1e05be60;hb=99cccd623b63570df4d33cbe335faaee323826df;hp=85273a8eb8ac309d074dd8895e424ddbb12457b2;hpb=0e25b971fb86cd91767644301dcfa2c1c413e3a8;p=ghc-hetmet.git diff --git a/ghc/rts/Adjustor.c b/ghc/rts/Adjustor.c index 85273a8..2ea1240 100644 --- a/ghc/rts/Adjustor.c +++ b/ghc/rts/Adjustor.c @@ -41,8 +41,9 @@ Haskell side. #include "RtsUtils.h" #include "RtsFlags.h" +#include + /* Heavily arch-specific, I'm afraid.. */ -#if defined(i386_TARGET_ARCH) || defined(sparc_TARGET_ARCH) || defined(alpha_TARGET_ARCH) #if defined(i386_TARGET_ARCH) /* Now here's something obscure for you: @@ -156,72 +157,68 @@ createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr) 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). + /* 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; + if ((adjustor = stgMallocBytes(4*(11+1), "createAdjustor")) != NULL) { + 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) /* Magic constant computed by inspecting the code length of @@ -280,8 +277,72 @@ 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) +/* + 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). +*/ + 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[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" + : : "g" (p)); + p++; + } + __asm__ volatile ("sync\n\tisync"); + } + } #else -#error Adjustor creation is not supported on this platform. + barf("adjustor creation not supported on this platform"); #endif break; @@ -294,7 +355,6 @@ TODO: Depending on how much allocation overhead stgMallocBytes uses for return adjustor; } -#endif void freeHaskellFunctionPtr(void* ptr) @@ -313,14 +373,14 @@ freeHaskellFunctionPtr(void* ptr) freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02))); } #elif defined(sparc_TARGET_ARCH) - if ( *(unsigned char*)ptr != 0x13 ) { + if ( *(unsigned long*)ptr != 0x9C23A008UL ) { fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr); return; } /* Free the stable pointer first..*/ - freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10))); -#elif defined(sparc_TARGET_ARCH) + 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); return; @@ -328,6 +388,12 @@ freeHaskellFunctionPtr(void* ptr) /* 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); + return; + } + freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 4*12))); #else ASSERT(0); #endif