1 /* -----------------------------------------------------------------------------
2 * Foreign export adjustor thunks
6 * ---------------------------------------------------------------------------*/
8 /* A little bit of background...
10 An adjustor thunk is a dynamically allocated code snippet that allows
11 Haskell closures to be viewed as C function pointers.
13 Stable pointers provide a way for the outside world to get access to,
14 and evaluate, Haskell heap objects, with the RTS providing a small
15 range of ops for doing so. So, assuming we've got a stable pointer in
16 our hand in C, we can jump into the Haskell world and evaluate a callback
17 procedure, say. This works OK in some cases where callbacks are used, but
18 does require the external code to know about stable pointers and how to deal
19 with them. We'd like to hide the Haskell-nature of a callback and have it
20 be invoked just like any other C function pointer.
22 Enter adjustor thunks. An adjustor thunk is a little piece of code
23 that's generated on-the-fly (one per Haskell closure being exported)
24 that, when entered using some 'universal' calling convention (e.g., the
25 C calling convention on platform X), pushes an implicit stable pointer
26 (to the Haskell callback) before calling another (static) C function stub
27 which takes care of entering the Haskell code via its stable pointer.
29 An adjustor thunk is allocated on the C heap, and is called from within
30 Haskell just before handing out the function pointer to the Haskell (IO)
31 action. User code should never have to invoke it explicitly.
33 An adjustor thunk differs from a C function pointer in one respect: when
34 the code is through with it, it has to be freed in order to release Haskell
35 and C resources. Failure to do so result in memory leaks on both the C and
39 #include "PosixSource.h"
41 #include "RtsExternal.h"
49 /* Heavily arch-specific, I'm afraid.. */
57 * Function: execPage()
59 * Set the executable bit on page containing addr.
61 * TODO: Can the code span more than one page? If yes, we need to make two
65 execPage (void* addr, pageMode mode)
67 #if defined(i386_TARGET_ARCH) && defined(_WIN32) && 0
69 DWORD dwOldProtect = 0;
71 /* doesn't return a result, so presumably it can't fail... */
72 GetSystemInfo(&sInfo);
74 if ( VirtualProtect ( (void*)((unsigned long)addr & (sInfo.dwPageSize - 1)),
76 ( mode == pageExecuteReadWrite ? PAGE_EXECUTE_READWRITE : PAGE_EXECUTE_READ),
77 &dwOldProtect) == 0 ) {
79 DWORD rc = GetLastError();
80 prog_belch("execPage: failed to protect 0x%p; error=%lu; old protection: %lu\n", addr, rc, dwOldProtect);
86 (void)addr; (void)mode; /* keep gcc -Wall happy */
91 #if defined(i386_TARGET_ARCH)
92 static unsigned char *obscure_ccall_ret_code;
95 #if defined(alpha_TARGET_ARCH)
96 /* To get the definition of PAL_imb: */
97 # if defined(linux_TARGET_OS)
100 # include <machine/pal.h>
104 #if defined(ia64_TARGET_ARCH)
107 /* Layout of a function descriptor */
108 typedef struct _IA64FunDesc {
114 stgAllocStable(size_t size_in_bytes, StgStablePtr *stable)
117 nat data_size_in_words, total_size_in_words;
119 /* round up to a whole number of words */
120 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
121 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
123 /* allocate and fill it in */
124 arr = (StgArrWords *)allocate(total_size_in_words);
125 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
127 /* obtain a stable ptr */
128 *stable = getStablePtr((StgPtr)arr);
130 /* and return a ptr to the goods inside the array */
131 return(BYTE_ARR_CTS(arr));
136 createAdjustor(int cconv, StgStablePtr hptr, StgFunPtr wptr)
138 void *adjustor = NULL;
142 case 0: /* _stdcall */
143 #if defined(i386_TARGET_ARCH)
144 /* Magic constant computed by inspecting the code length of
145 the following assembly language snippet
146 (offset and machine code prefixed):
148 <0>: 58 popl %eax # temp. remove ret addr..
149 <1>: 68 fd fc fe fa pushl 0xfafefcfd # constant is large enough to
150 # hold a StgStablePtr
151 <6>: 50 pushl %eax # put back ret. addr
152 <7>: b8 fa ef ff 00 movl $0x00ffeffa, %eax # load up wptr
153 <c>: ff e0 jmp %eax # and jump to it.
154 # the callee cleans up the stack
156 if ((adjustor = stgMallocBytes(14, "createAdjustor")) != NULL) {
157 unsigned char *const adj_code = (unsigned char *)adjustor;
158 adj_code[0x00] = (unsigned char)0x58; /* popl %eax */
160 adj_code[0x01] = (unsigned char)0x68; /* pushl hptr (which is a dword immediate ) */
161 *((StgStablePtr*)(adj_code + 0x02)) = (StgStablePtr)hptr;
163 adj_code[0x06] = (unsigned char)0x50; /* pushl %eax */
165 adj_code[0x07] = (unsigned char)0xb8; /* movl $wptr, %eax */
166 *((StgFunPtr*)(adj_code + 0x08)) = (StgFunPtr)wptr;
168 adj_code[0x0c] = (unsigned char)0xff; /* jmp %eax */
169 adj_code[0x0d] = (unsigned char)0xe0;
171 execPage(adjustor, pageExecuteReadWrite);
177 #if defined(i386_TARGET_ARCH)
178 /* Magic constant computed by inspecting the code length of
179 the following assembly language snippet
180 (offset and machine code prefixed):
182 <00>: 68 ef be ad de pushl $0xdeadbeef # constant is large enough to
183 # hold a StgStablePtr
184 <05>: b8 fa ef ff 00 movl $0x00ffeffa, %eax # load up wptr
185 <0a>: 68 ef be ad de pushl $obscure_ccall_ret_code # push the return address
186 <0f>: ff e0 jmp *%eax # jump to wptr
188 The ccall'ing version is a tad different, passing in the return
189 address of the caller to the auto-generated C stub (which enters
190 via the stable pointer.) (The auto-generated C stub is in on this
191 game, don't worry :-)
193 See the comment next to obscure_ccall_ret_code why we need to
194 perform a tail jump instead of a call, followed by some C stack
197 Note: The adjustor makes the assumption that any return value
198 coming back from the C stub is not stored on the stack.
199 That's (thankfully) the case here with the restricted set of
200 return types that we support.
202 if ((adjustor = stgMallocBytes(17, "createAdjustor")) != NULL) {
203 unsigned char *const adj_code = (unsigned char *)adjustor;
205 adj_code[0x00] = (unsigned char)0x68; /* pushl hptr (which is a dword immediate ) */
206 *((StgStablePtr*)(adj_code+0x01)) = (StgStablePtr)hptr;
208 adj_code[0x05] = (unsigned char)0xb8; /* movl $wptr, %eax */
209 *((StgFunPtr*)(adj_code + 0x06)) = (StgFunPtr)wptr;
211 adj_code[0x0a] = (unsigned char)0x68; /* pushl obscure_ccall_ret_code */
212 *((StgFunPtr*)(adj_code + 0x0b)) = (StgFunPtr)obscure_ccall_ret_code;
214 adj_code[0x0f] = (unsigned char)0xff; /* jmp *%eax */
215 adj_code[0x10] = (unsigned char)0xe0;
217 execPage(adjustor, pageExecuteReadWrite);
219 #elif defined(sparc_TARGET_ARCH)
220 /* Magic constant computed by inspecting the code length of the following
221 assembly language snippet (offset and machine code prefixed):
223 <00>: 9C23A008 sub %sp, 8, %sp ! make room for %o4/%o5 in caller's frame
224 <04>: DA23A060 st %o5, [%sp + 96] ! shift registers by 2 positions
225 <08>: D823A05C st %o4, [%sp + 92]
226 <0C>: 9A10000B mov %o3, %o5
227 <10>: 9810000A mov %o2, %o4
228 <14>: 96100009 mov %o1, %o3
229 <18>: 94100008 mov %o0, %o2
230 <1C>: 13000000 sethi %hi(wptr), %o1 ! load up wptr (1 of 2)
231 <20>: 11000000 sethi %hi(hptr), %o0 ! load up hptr (1 of 2)
232 <24>: 81C26000 jmp %o1 + %lo(wptr) ! jump to wptr (load 2 of 2)
233 <28>: 90122000 or %o0, %lo(hptr), %o0 ! load up hptr (2 of 2, delay slot)
234 <2C> 00000000 ! place for getting hptr back easily
236 ccall'ing on SPARC is easy, because we are quite lucky to push a
237 multiple of 8 bytes (1 word hptr + 1 word dummy arg) in front of the
238 existing arguments (note that %sp must stay double-word aligned at
239 all times, see ABI spec at http://www.sparc.org/standards/psABI3rd.pdf).
240 To do this, we extend the *caller's* stack frame by 2 words and shift
241 the output registers used for argument passing (%o0 - %o5, we are a *leaf*
242 procedure because of the tail-jump) by 2 positions. This makes room in
243 %o0 and %o1 for the additinal arguments, namely hptr and a dummy (used
244 for destination addr of jump on SPARC, return address on x86, ...). This
245 shouldn't cause any problems for a C-like caller: alloca is implemented
246 similarly, and local variables should be accessed via %fp, not %sp. In a
247 nutshell: This should work! (Famous last words! :-)
249 if ((adjustor = stgMallocBytes(4*(11+1), "createAdjustor")) != NULL) {
250 unsigned long *const adj_code = (unsigned long *)adjustor;
252 adj_code[ 0] = 0x9C23A008UL; /* sub %sp, 8, %sp */
253 adj_code[ 1] = 0xDA23A060UL; /* st %o5, [%sp + 96] */
254 adj_code[ 2] = 0xD823A05CUL; /* st %o4, [%sp + 92] */
255 adj_code[ 3] = 0x9A10000BUL; /* mov %o3, %o5 */
256 adj_code[ 4] = 0x9810000AUL; /* mov %o2, %o4 */
257 adj_code[ 5] = 0x96100009UL; /* mov %o1, %o3 */
258 adj_code[ 6] = 0x94100008UL; /* mov %o0, %o2 */
259 adj_code[ 7] = 0x13000000UL; /* sethi %hi(wptr), %o1 */
260 adj_code[ 7] |= ((unsigned long)wptr) >> 10;
261 adj_code[ 8] = 0x11000000UL; /* sethi %hi(hptr), %o0 */
262 adj_code[ 8] |= ((unsigned long)hptr) >> 10;
263 adj_code[ 9] = 0x81C26000UL; /* jmp %o1 + %lo(wptr) */
264 adj_code[ 9] |= ((unsigned long)wptr) & 0x000003FFUL;
265 adj_code[10] = 0x90122000UL; /* or %o0, %lo(hptr), %o0 */
266 adj_code[10] |= ((unsigned long)hptr) & 0x000003FFUL;
268 adj_code[11] = (unsigned long)hptr;
271 asm("flush %0" : : "r" (adj_code ));
272 asm("flush %0" : : "r" (adj_code + 2));
273 asm("flush %0" : : "r" (adj_code + 4));
274 asm("flush %0" : : "r" (adj_code + 6));
275 asm("flush %0" : : "r" (adj_code + 10));
277 /* max. 5 instructions latency, and we need at >= 1 for returning */
283 #elif defined(alpha_TARGET_ARCH)
284 /* Magic constant computed by inspecting the code length of
285 the following assembly language snippet
286 (offset and machine code prefixed; note that the machine code
287 shown is longwords stored in little-endian order):
289 <00>: 46520414 mov a2, a4
290 <04>: 46100412 mov a0, a2
291 <08>: a61b0020 ldq a0, 0x20(pv) # load up hptr
292 <0c>: 46730415 mov a3, a5
293 <10>: a77b0028 ldq pv, 0x28(pv) # load up wptr
294 <14>: 46310413 mov a1, a3
295 <18>: 6bfb---- jmp (pv), <hint> # jump to wptr (with hint)
296 <1c>: 00000000 # padding for alignment
297 <20>: [8 bytes for hptr quadword]
298 <28>: [8 bytes for wptr quadword]
300 The "computed" jump at <08> above is really a jump to a fixed
301 location. Accordingly, we place an always-correct hint in the
302 jump instruction, namely the address offset from <0c> to wptr,
303 divided by 4, taking the lowest 14 bits.
305 We only support passing 4 or fewer argument words, for the same
306 reason described under sparc_TARGET_ARCH above by JRS, 21 Aug 01.
307 On the Alpha the first 6 integer arguments are in a0 through a5,
308 and the rest on the stack. Hence we want to shuffle the original
309 caller's arguments by two.
311 On the Alpha the calling convention is so complex and dependent
312 on the callee's signature -- for example, the stack pointer has
313 to be a multiple of 16 -- that it seems impossible to me [ccshan]
314 to handle the general case correctly without changing how the
315 adjustor is called from C. For now, our solution of shuffling
316 registers only and ignoring the stack only works if the original
317 caller passed 4 or fewer argument words.
319 TODO: Depending on how much allocation overhead stgMallocBytes uses for
320 header information (more precisely, if the overhead is no more than
321 4 bytes), we should move the first three instructions above down by
322 4 bytes (getting rid of the nop), hence saving memory. [ccshan]
324 ASSERT(((StgWord64)wptr & 3) == 0);
325 if ((adjustor = stgMallocBytes(48, "createAdjustor")) != NULL) {
326 StgWord64 *const code = (StgWord64 *)adjustor;
328 code[0] = 0x4610041246520414L;
329 code[1] = 0x46730415a61b0020L;
330 code[2] = 0x46310413a77b0028L;
331 code[3] = 0x000000006bfb0000L
332 | (((StgWord32*)(wptr) - (StgWord32*)(code) - 3) & 0x3fff);
334 code[4] = (StgWord64)hptr;
335 code[5] = (StgWord64)wptr;
337 /* Ensure that instruction cache is consistent with our new code */
338 __asm__ volatile("call_pal %0" : : "i" (PAL_imb));
340 #elif defined(powerpc_TARGET_ARCH)
342 For PowerPC, the following code is used:
350 lis r0,0xDEAD ;hi(wptr)
351 lis r3,0xDEAF ;hi(hptr)
352 ori r0,r0,0xBEEF ; lo(wptr)
353 ori r3,r3,0xFACE ; lo(hptr)
357 The arguments (passed in registers r3 - r10) are shuffled along by two to
358 make room for hptr and a dummy argument. As r9 and r10 are overwritten by
359 this code, it only works for up to 6 arguments (when floating point arguments
360 are involved, this may be more or less, depending on the exact situation).
362 if ((adjustor = stgMallocBytes(4*13, "createAdjustor")) != NULL) {
363 unsigned long *const adj_code = (unsigned long *)adjustor;
365 // make room for extra arguments
366 adj_code[0] = 0x7d0a4378; //mr r10,r8
367 adj_code[1] = 0x7ce93b78; //mr r9,r7
368 adj_code[2] = 0x7cc83378; //mr r8,r6
369 adj_code[3] = 0x7ca72b78; //mr r7,r5
370 adj_code[4] = 0x7c862378; //mr r6,r4
371 adj_code[5] = 0x7c651b78; //mr r5,r3
373 adj_code[6] = 0x3c000000; //lis r0,hi(wptr)
374 adj_code[6] |= ((unsigned long)wptr) >> 16;
376 adj_code[7] = 0x3c600000; //lis r3,hi(hptr)
377 adj_code[7] |= ((unsigned long)hptr) >> 16;
379 adj_code[8] = 0x60000000; //ori r0,r0,lo(wptr)
380 adj_code[8] |= ((unsigned long)wptr) & 0xFFFF;
382 adj_code[9] = 0x60630000; //ori r3,r3,lo(hptr)
383 adj_code[9] |= ((unsigned long)hptr) & 0xFFFF;
385 adj_code[10] = 0x7c0903a6; //mtctr r0
386 adj_code[11] = 0x4e800420; //bctr
387 adj_code[12] = (unsigned long)hptr;
389 // Flush the Instruction cache:
390 // MakeDataExecutable(adjustor,4*13);
391 /* This would require us to link with CoreServices.framework */
392 { /* this should do the same: */
394 unsigned long *p = adj_code;
397 __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0"
401 __asm__ volatile ("sync\n\tisync");
404 #elif defined(ia64_TARGET_ARCH)
406 Up to 8 inputs are passed in registers. We flush the last two inputs to
407 the stack, initially into the 16-byte scratch region left by the caller.
408 We then shuffle the others along by 4 (taking 2 registers for ourselves
409 to save return address and previous function state - we need to come back
410 here on the way out to restore the stack, so this is a real function
411 rather than just a trampoline).
413 The function descriptor we create contains the gp of the target function
414 so gp is already loaded correctly.
416 [MLX] alloc r16=ar.pfs,10,2,0
418 [MII] st8.spill [r12]=r38,8 // spill in6 (out4)
419 mov r41=r37 // out7 = in5 (out3)
420 mov r40=r36;; // out6 = in4 (out2)
421 [MII] st8.spill [r12]=r39 // spill in7 (out5)
423 mov r38=r34;; // out4 = in2 (out0)
424 [MII] mov r39=r35 // out5 = in3 (out1)
425 mov r37=r33 // out3 = in1 (loc1)
426 mov r36=r32 // out2 = in0 (loc0)
427 [MLX] adds r12=-24,r12 // update sp
428 movl r34=hptr;; // out0 = hptr
429 [MIB] mov r33=r16 // loc1 = ar.pfs
430 mov r32=b0 // loc0 = retaddr
431 br.call.sptk.many b0=b6;;
433 [MII] adds r12=-16,r12
438 br.ret.sptk.many b0;;
441 /* These macros distribute a long constant into the two words of an MLX bundle */
442 #define BITS(val,start,count) (((val) >> (start)) & ((1 << (count))-1))
443 #define MOVL_LOWORD(val) (BITS(val,22,18) << 46)
444 #define MOVL_HIWORD(val) (BITS(val,40,23) | (BITS(val,0,7) << 36) | (BITS(val,7,9) << 50) \
445 | (BITS(val,16,5) << 55) | (BITS(val,21,1) << 44) | BITS(val,63,1) << 59)
449 IA64FunDesc *wdesc = (IA64FunDesc *)wptr;
450 StgWord64 wcode = wdesc->ip;
454 /* we allocate on the Haskell heap since malloc'd memory isn't executable - argh */
455 adjustor = stgAllocStable(sizeof(IA64FunDesc)+18*8, &stable);
457 fdesc = (IA64FunDesc *)adjustor;
458 code = (StgWord64 *)(fdesc + 1);
459 fdesc->ip = (StgWord64)code;
460 fdesc->gp = wdesc->gp;
462 code[0] = 0x0000058004288004 | MOVL_LOWORD(wcode);
463 code[1] = 0x6000000220000000 | MOVL_HIWORD(wcode);
464 code[2] = 0x029015d818984001;
465 code[3] = 0x8401200500420094;
466 code[4] = 0x886011d8189c0001;
467 code[5] = 0x84011004c00380c0;
468 code[6] = 0x0250210046013800;
469 code[7] = 0x8401000480420084;
470 code[8] = 0x0000233f19a06005 | MOVL_LOWORD((StgWord64)hptr);
471 code[9] = 0x6000000440000000 | MOVL_HIWORD((StgWord64)hptr);
472 code[10] = 0x0200210020010811;
473 code[11] = 0x1080006800006200;
474 code[12] = 0x0000210018406000;
475 code[13] = 0x00aa021000038005;
476 code[14] = 0x000000010000001d;
477 code[15] = 0x0084000880000200;
479 /* save stable pointers in convenient form */
480 code[16] = (StgWord64)hptr;
481 code[17] = (StgWord64)stable;
484 barf("adjustor creation not supported on this platform");
499 freeHaskellFunctionPtr(void* ptr)
501 #if defined(i386_TARGET_ARCH)
502 if ( *(unsigned char*)ptr != 0x68 &&
503 *(unsigned char*)ptr != 0x58 ) {
504 prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
508 /* Free the stable pointer first..*/
509 if (*(unsigned char*)ptr == 0x68) { /* Aha, a ccall adjustor! */
510 freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x01)));
512 freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02)));
514 #elif defined(sparc_TARGET_ARCH)
515 if ( *(unsigned long*)ptr != 0x9C23A008UL ) {
516 prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
520 /* Free the stable pointer first..*/
521 freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 11)));
522 #elif defined(alpha_TARGET_ARCH)
523 if ( *(StgWord64*)ptr != 0xa77b0018a61b0010L ) {
524 prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
528 /* Free the stable pointer first..*/
529 freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10)));
530 #elif defined(powerpc_TARGET_ARCH)
531 if ( *(StgWord*)ptr != 0x7d0a4378 ) {
532 prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
535 freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 4*12)));
536 #elif defined(ia64_TARGET_ARCH)
537 IA64FunDesc *fdesc = (IA64FunDesc *)ptr;
538 StgWord64 *code = (StgWord64 *)(fdesc+1);
540 if (fdesc->ip != (StgWord64)code) {
541 prog_belch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
544 freeStablePtr((StgStablePtr)code[16]);
545 freeStablePtr((StgStablePtr)code[17]);
550 *((unsigned char*)ptr) = '\0';
557 * Function: initAdjustor()
559 * Perform initialisation of adjustor thunk layer (if needed.)
564 #if defined(i386_TARGET_ARCH)
565 /* Now here's something obscure for you:
567 When generating an adjustor thunk that uses the C calling
568 convention, we have to make sure that the thunk kicks off
569 the process of jumping into Haskell with a tail jump. Why?
570 Because as a result of jumping in into Haskell we may end
571 up freeing the very adjustor thunk we came from using
572 freeHaskellFunctionPtr(). Hence, we better not return to
573 the adjustor code on our way out, since it could by then
576 The fix is readily at hand, just include the opcodes
577 for the C stack fixup code that we need to perform when
578 returning in some static piece of memory and arrange
579 to return to it before tail jumping from the adjustor thunk.
582 obscure_ccall_ret_code = stgMallocBytes(4, "initAdjustor");
584 obscure_ccall_ret_code[0x00] = (unsigned char)0x83; /* addl $0x4, %esp */
585 obscure_ccall_ret_code[0x01] = (unsigned char)0xc4;
586 obscure_ccall_ret_code[0x02] = (unsigned char)0x04;
588 obscure_ccall_ret_code[0x03] = (unsigned char)0xc3; /* ret */
590 execPage(obscure_ccall_ret_code, pageExecuteRead);