Adjustor.c 45.8 KB
Newer Older
1
2
3
4
5
6
/* -----------------------------------------------------------------------------
 * Foreign export adjustor thunks
 *
 * Copyright (c) 1998.
 *
 * ---------------------------------------------------------------------------*/
sof's avatar
sof committed
7

8
/* A little bit of background...
9
   
10
An adjustor thunk is a dynamically allocated code snippet that allows
sof's avatar
sof committed
11
12
13
Haskell closures to be viewed as C function pointers. 

Stable pointers provide a way for the outside world to get access to,
14
15
and evaluate, Haskell heap objects, with the RTS providing a small
range of ops for doing so. So, assuming we've got a stable pointer in
sof's avatar
sof committed
16
17
18
19
20
21
our hand in C, we can jump into the Haskell world and evaluate a callback
procedure, say. This works OK in some cases where callbacks are used, but
does require the external code to know about stable pointers and how to deal
with them. We'd like to hide the Haskell-nature of a callback and have it
be invoked just like any other C function pointer. 

22
23
24
25
26
27
Enter adjustor thunks. An adjustor thunk is a little piece of code
that's generated on-the-fly (one per Haskell closure being exported)
that, when entered using some 'universal' calling convention (e.g., the
C calling convention on platform X), pushes an implicit stable pointer
(to the Haskell callback) before calling another (static) C function stub
which takes care of entering the Haskell code via its stable pointer.
sof's avatar
sof committed
28
29
30
31
32

An adjustor thunk is allocated on the C heap, and is called from within
Haskell just before handing out the function pointer to the Haskell (IO)
action. User code should never have to invoke it explicitly.

33
An adjustor thunk differs from a C function pointer in one respect: when
sof's avatar
sof committed
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 will result in memory leaks on both the C and
sof's avatar
sof committed
36
Haskell side.
37
*/
38

39
#include "PosixSource.h"
40
#include "Rts.h"
Simon Marlow's avatar
Simon Marlow committed
41

42
#include "RtsUtils.h"
Simon Marlow's avatar
Simon Marlow committed
43
#include "Stable.h"
sof's avatar
sof committed
44

45
46
#if defined(USE_LIBFFI_FOR_ADJUSTORS)
#include "ffi.h"
47
#include <string.h>
Ian Lynagh's avatar
Ian Lynagh committed
48
49
#endif

50
#if defined(i386_HOST_ARCH)
Ian Lynagh's avatar
Ian Lynagh committed
51
52
53
54
55
56
57
extern void adjustorCode(void);
#elif defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH)
// from AdjustorAsm.s
// not declared as a function so that AIX-style
// fundescs can never get in the way.
extern void *adjustorCode;
#endif
58

Ian Lynagh's avatar
Ian Lynagh committed
59
#if defined(USE_LIBFFI_FOR_ADJUSTORS)
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
void
freeHaskellFunctionPtr(void* ptr)
{
    ffi_closure *cl;

    cl = (ffi_closure*)ptr;
    freeStablePtr(cl->user_data);
    stgFree(cl->cif->arg_types);
    stgFree(cl->cif);
    freeExec(cl);
}

static ffi_type * char_to_ffi_type(char c)
{
    switch (c) {
    case 'v':  return &ffi_type_void;
    case 'f':  return &ffi_type_float;
    case 'd':  return &ffi_type_double;
    case 'L':  return &ffi_type_sint64;
    case 'l':  return &ffi_type_uint64;
    case 'W':  return &ffi_type_sint32;
    case 'w':  return &ffi_type_uint32;
    case 'S':  return &ffi_type_sint16;
    case 's':  return &ffi_type_uint16;
    case 'B':  return &ffi_type_sint8;
    case 'b':  return &ffi_type_uint8;
86
    case 'p':  return &ffi_type_pointer;
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
    default:   barf("char_to_ffi_type: unknown type '%c'", c);
    }
}

void*
createAdjustor (int cconv, 
                StgStablePtr hptr,
                StgFunPtr wptr,
                char *typeString)
{
    ffi_cif *cif;
    ffi_type **arg_types;
    nat n_args, i;
    ffi_type *result_type;
    ffi_closure *cl;
    int r, abi;
103
    void *code;
104
105
106
107
108
109
110
111
112
113

    n_args = strlen(typeString) - 1;
    cif = stgMallocBytes(sizeof(ffi_cif), "createAdjustor");
    arg_types = stgMallocBytes(n_args * sizeof(ffi_type*), "createAdjustor");

    result_type = char_to_ffi_type(typeString[0]);
    for (i=0; i < n_args; i++) {
        arg_types[i] = char_to_ffi_type(typeString[i+1]);
    }
    switch (cconv) {
114
#if defined(mingw32_HOST_OS) && defined(i386_HOST_ARCH)
115
116
117
118
119
120
121
122
123
124
125
126
127
128
    case 0: /* stdcall */
        abi = FFI_STDCALL;
        break;
#endif
    case 1: /* ccall */
        abi = FFI_DEFAULT_ABI;
        break;
    default:
        barf("createAdjustor: convention %d not supported on this platform", cconv);
    }

    r = ffi_prep_cif(cif, abi, n_args, result_type, arg_types);
    if (r != FFI_OK) barf("ffi_prep_cif failed: %d", r);
    
129
130
131
132
    cl = allocateExec(sizeof(ffi_closure), &code);
    if (cl == NULL) {
        barf("createAdjustor: failed to allocate memory");
    }
133

134
    r = ffi_prep_closure_loc(cl, cif, (void*)wptr, hptr/*userdata*/, code);
135
136
    if (r != FFI_OK) barf("ffi_prep_closure failed: %d", r);

137
    return (void*)code;
138
139
140
141
}

#else // To end of file...

sof's avatar
sof committed
142
143
144
145
#if defined(_WIN32)
#include <windows.h>
#endif

146
#if defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS)
147
148
149
#include <string.h>
#endif

150
151
152
153
154
#ifdef LEADING_UNDERSCORE
#define UNDERSCORE "_"
#else 
#define UNDERSCORE ""
#endif
Simon Marlow's avatar
Simon Marlow committed
155
156

#if defined(x86_64_HOST_ARCH)
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
/* 
  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.
*/
dons's avatar
dons committed
174
175
176
static void GNUC3_ATTRIBUTE(used) obscure_ccall_wrapper(void)
{
  __asm__ (
177
178
   ".globl " UNDERSCORE "obscure_ccall_ret_code\n"
   UNDERSCORE "obscure_ccall_ret_code:\n\t"
179
   "addq $0x8, %rsp\n\t"
Ian Lynagh's avatar
Ian Lynagh committed
180
181
182
183
184
185
#if defined(mingw32_HOST_OS)
   /* On Win64, we had to put the original return address after the
      arg 1-4 spill slots, ro now we have to move it back */
   "movq 0x20(%rsp), %rcx\n"
   "movq %rcx, (%rsp)\n"
#endif
186
187
   "ret"
  );
dons's avatar
dons committed
188
}
189
190
191
extern void obscure_ccall_ret_code(void);
#endif

192
#if defined(alpha_HOST_ARCH)
ken's avatar
ken committed
193
/* To get the definition of PAL_imb: */
194
# if defined(linux_HOST_OS)
195
196
197
198
#  include <asm/pal.h>
# else
#  include <machine/pal.h>
# endif
ken's avatar
ken committed
199
200
#endif

201
#if defined(ia64_HOST_ARCH)
202
203
204
205
206
207
208
209
210
211
212
213
214
215

/* 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 */
216
  data_size_in_words  = ROUNDUP_BYTES_TO_WDS(size_in_bytes);
217
218
219
220
  total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
  
  /* allocate and fill it in */
  arr = (StgArrWords *)allocate(total_size_in_words);
221
  SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, size_in_bytes);
222
223
224
225
226
 
  /* obtain a stable ptr */
  *stable = getStablePtr((StgPtr)arr);

  /* and return a ptr to the goods inside the array */
227
  return(&(arr->payload));
228
229
230
}
#endif

231
#if defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS)
232
233
234
235
236
237
238
239
__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

240
241
#if defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH)
#if !(defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS))
242
243
244
245
246
247
248

/* !!! !!! WARNING: !!! !!!
 * This structure is accessed from AdjustorAsm.s
 * Any changes here have to be mirrored in the offsets there.
 */

typedef struct AdjustorStub {
249
#if defined(powerpc_HOST_ARCH) && defined(darwin_HOST_OS)
250
251
252
253
254
255
    unsigned        lis;
    unsigned        ori;
    unsigned        lwz;
    unsigned        mtctr;
    unsigned        bctr;
    StgFunPtr       code;
256
#elif defined(powerpc64_HOST_ARCH) && defined(darwin_HOST_OS)
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
        /* 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;
280

281
#endif
282
283
#endif

284
#if defined(i386_HOST_ARCH)
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299

/* !!! !!! 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

300
#if defined(i386_HOST_ARCH) || defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH)
301
302
303
304
305
306
307
308
309
310
311
312
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':
313
            case 'L':
314
315
316
317
318
319
320
321
322
323
324
325
326
327
                if(sizeof(void*) == 4)
                {
                    sz += 2;
                    break;
                }
                // everything else is one word.
            default:
                sz += 1;
        }
    }
    return sz;
}
#endif

sof's avatar
sof committed
328
void*
329
createAdjustor(int cconv, StgStablePtr hptr,
Ian Lynagh's avatar
Ian Lynagh committed
330
331
               StgFunPtr wptr,
               char *typeString
332
#if !defined(powerpc_HOST_ARCH) && !defined(powerpc64_HOST_ARCH) && !defined(x86_64_HOST_ARCH)
Ian Lynagh's avatar
Ian Lynagh committed
333
                  STG_UNUSED
334
335
#endif
              )
sof's avatar
sof committed
336
{
ken's avatar
ken committed
337
  void *adjustor = NULL;
338
  void *code;
339

ken's avatar
ken committed
340
341
342
  switch (cconv)
  {
  case 0: /* _stdcall */
343
#if defined(i386_HOST_ARCH) && !defined(darwin_HOST_OS)
sof's avatar
sof committed
344
345
346
347
    /* Magic constant computed by inspecting the code length of
       the following assembly language snippet
       (offset and machine code prefixed):

Ian Lynagh's avatar
Ian Lynagh committed
348
349
350
351
352
353
354
     <0>:       58                popl   %eax              # temp. remove ret addr..
     <1>:       68 fd fc fe fa    pushl  0xfafefcfd        # constant is large enough to
                                                           # hold a StgStablePtr
     <6>:       50                pushl  %eax              # put back ret. addr
     <7>:       b8 fa ef ff 00    movl   $0x00ffeffa, %eax # load up wptr
     <c>:       ff e0             jmp    %eax              # and jump to it.
                # the callee cleans up the stack
sof's avatar
sof committed
355
    */
356
    adjustor = allocateExec(14,&code);
357
    {
Ian Lynagh's avatar
Ian Lynagh committed
358
359
        unsigned char *const adj_code = (unsigned char *)adjustor;
        adj_code[0x00] = (unsigned char)0x58;  /* popl %eax  */
sof's avatar
sof committed
360

Ian Lynagh's avatar
Ian Lynagh committed
361
362
        adj_code[0x01] = (unsigned char)0x68;  /* pushl hptr (which is a dword immediate ) */
        *((StgStablePtr*)(adj_code + 0x02)) = (StgStablePtr)hptr;
sof's avatar
sof committed
363

Ian Lynagh's avatar
Ian Lynagh committed
364
        adj_code[0x06] = (unsigned char)0x50; /* pushl %eax */
sof's avatar
sof committed
365

Ian Lynagh's avatar
Ian Lynagh committed
366
367
        adj_code[0x07] = (unsigned char)0xb8; /* movl  $wptr, %eax */
        *((StgFunPtr*)(adj_code + 0x08)) = (StgFunPtr)wptr;
sof's avatar
sof committed
368

Ian Lynagh's avatar
Ian Lynagh committed
369
370
        adj_code[0x0c] = (unsigned char)0xff; /* jmp %eax */
        adj_code[0x0d] = (unsigned char)0xe0;
371
    }
ken's avatar
ken committed
372
373
#endif
    break;
sof's avatar
sof committed
374

ken's avatar
ken committed
375
  case 1: /* _ccall */
376
#if defined(i386_HOST_ARCH)
377
378
    {
        /*
379
380
381
382
383
          Most of the trickiness here is due to the need to keep the
          stack pointer 16-byte aligned (see #5250).  That means we
          can't just push another argument on the stack and call the
          wrapper, we may have to shuffle the whole argument block.

384
385
          We offload most of the work to AdjustorAsm.S.
        */
386
        AdjustorStub *adjustorStub = allocateExec(sizeof(AdjustorStub),&code);
387
388
389
390
391
        adjustor = adjustorStub;

        int sz = totalArgumentSize(typeString);
        
        adjustorStub->call[0] = 0xe8;
392
        *(long*)&adjustorStub->call[1] = ((char*)&adjustorCode) - ((char*)code + 5);
393
394
        adjustorStub->hptr = hptr;
        adjustorStub->wptr = wptr;
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
        
            // 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; 
410
411
412
        adjustorStub->argument_size = sz;
    }
    
413
#elif defined(x86_64_HOST_ARCH)
Ian Lynagh's avatar
Ian Lynagh committed
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457

# if defined(mingw32_HOST_OS)
    /*
      stack at call:
               argn
               ...
               arg5
               return address
               %rcx,%rdx,%r8,%r9 = arg1..arg4

      if there are <4 integer args, then we can just push the
      StablePtr into %rcx and shuffle the other args up.

      If there are >=4 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
             ...
             arg5
             return address  *** <-- dummy arg in stub fn.
             arg4
             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 4th integer argument.

      Code for the simple case:

   0:   4d 89 c1                mov    %r8,%r9
   3:   49 89 d0                mov    %rdx,%r8
   6:   48 89 ca                mov    %rcx,%rdx
   9:   f2 0f 10 da             movsd  %xmm2,%xmm3
   d:   f2 0f 10 d1             movsd  %xmm1,%xmm2
  11:   f2 0f 10 c8             movsd  %xmm0,%xmm1
  15:   48 8b 0d 0c 00 00 00    mov    0xc(%rip),%rcx    # 28 <.text+0x28>
  1c:   ff 25 0e 00 00 00       jmpq   *0xe(%rip)        # 30 <.text+0x30>
  22:   90                      nop
  [...]


  And the version for >=4 integer arguments:

Ian Lynagh's avatar
Ian Lynagh committed
458
459
460
461
462
463
464
[we want to push the 4th argument (either %r9 or %xmm3, depending on
 whether it is a floating arg or not) and the return address onto the
 stack. However, slots 1-4 are reserved for code we call to spill its
 args 1-4 into, so we can't just push them onto the bottom of the stack.
 So first put the 4th argument onto the stack, above what will be the
 spill slots.]
   0:   48 83 ec 08             sub    $0x8,%rsp
465
466
467
468
469
470
[if non-floating arg, then do this:]
   4:   90                      nop
   5:   4c 89 4c 24 20          mov    %r9,0x20(%rsp)
[else if floating arg then do this:]
   4:   f2 0f 11 5c 24 20       movsd  %xmm3,0x20(%rsp)
[end if]
Ian Lynagh's avatar
Ian Lynagh committed
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
[Now push the new return address onto the stack]
   a:   ff 35 30 00 00 00       pushq  0x30(%rip)        # 40 <.text+0x40>
[But the old return address has been moved up into a spill slot, so
 we need to move it above them]
  10:   4c 8b 4c 24 10          mov    0x10(%rsp),%r9
  15:   4c 89 4c 24 30          mov    %r9,0x30(%rsp)
[Now we do the normal register shuffle-up etc]
  1a:   4d 89 c1                mov    %r8,%r9
  1d:   49 89 d0                mov    %rdx,%r8
  20:   48 89 ca                mov    %rcx,%rdx
  23:   f2 0f 10 da             movsd  %xmm2,%xmm3
  27:   f2 0f 10 d1             movsd  %xmm1,%xmm2
  2b:   f2 0f 10 c8             movsd  %xmm0,%xmm1
  2f:   48 8b 0d 12 00 00 00    mov    0x12(%rip),%rcx        # 48 <.text+0x48>
  36:   ff 25 14 00 00 00       jmpq   *0x14(%rip)        # 50 <.text+0x50>
  3c:   90                      nop
  3d:   90                      nop
  3e:   90                      nop
  3f:   90                      nop
Ian Lynagh's avatar
Ian Lynagh committed
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
  [...]

    */
    {  
        int i = 0;
        int fourthFloating;
        char *c;
        StgWord8 *adj_code;

        // determine whether we have 4 or more integer arguments,
        // and therefore need to flush one to the stack.
        for (c = typeString; *c != '\0'; c++) {
            i++;
            if (i == 4) {
                fourthFloating = (*c == 'f' || *c == 'd');
                break;
            }
        }

        if (i < 4) {
            adjustor = allocateExec(0x38,&code);
            adj_code = (StgWord8*)adjustor;

            *(StgInt32 *)adj_code        = 0x49c1894d;
            *(StgInt32 *)(adj_code+0x4)  = 0x8948d089;
            *(StgInt32 *)(adj_code+0x8)  = 0x100ff2ca;
            *(StgInt32 *)(adj_code+0xc)  = 0x100ff2da;
            *(StgInt32 *)(adj_code+0x10) = 0x100ff2d1;
            *(StgInt32 *)(adj_code+0x14) = 0x0d8b48c8;
            *(StgInt32 *)(adj_code+0x18) = 0x0000000c;

            *(StgInt32 *)(adj_code+0x1c) = 0x000e25ff;
            *(StgInt32 *)(adj_code+0x20) = 0x00000000;
            *(StgInt64 *)(adj_code+0x28) = (StgInt64)hptr;
            *(StgInt64 *)(adj_code+0x30) = (StgInt64)wptr;
        }
        else
        {
Ian Lynagh's avatar
Ian Lynagh committed
528
            adjustor = allocateExec(0x58,&code);
Ian Lynagh's avatar
Ian Lynagh committed
529
            adj_code = (StgWord8*)adjustor;
530
531
532
533
            *(StgInt32 *)adj_code        = 0x08ec8348;
            *(StgInt32 *)(adj_code+0x4)  = fourthFloating ? 0x5c110ff2
                                                          : 0x4c894c90;
            *(StgInt32 *)(adj_code+0x8)  = 0x35ff2024;
Ian Lynagh's avatar
Ian Lynagh committed
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
            *(StgInt32 *)(adj_code+0xc)  = 0x00000030;
            *(StgInt32 *)(adj_code+0x10) = 0x244c8b4c;
            *(StgInt32 *)(adj_code+0x14) = 0x4c894c10;
            *(StgInt32 *)(adj_code+0x18) = 0x894d3024;
            *(StgInt32 *)(adj_code+0x1c) = 0xd08949c1;
            *(StgInt32 *)(adj_code+0x20) = 0xf2ca8948;
            *(StgInt32 *)(adj_code+0x24) = 0xf2da100f;
            *(StgInt32 *)(adj_code+0x28) = 0xf2d1100f;
            *(StgInt32 *)(adj_code+0x2c) = 0x48c8100f;
            *(StgInt32 *)(adj_code+0x30) = 0x00120d8b;
            *(StgInt32 *)(adj_code+0x34) = 0x25ff0000;
            *(StgInt32 *)(adj_code+0x38) = 0x00000014;
            *(StgInt32 *)(adj_code+0x3c) = 0x90909090;
            *(StgInt64 *)(adj_code+0x40) = (StgInt64)obscure_ccall_ret_code;
            *(StgInt64 *)(adj_code+0x48) = (StgInt64)hptr;
            *(StgInt64 *)(adj_code+0x50) = (StgInt64)wptr;
Ian Lynagh's avatar
Ian Lynagh committed
550
551
        }
    }
Ian Lynagh's avatar
Ian Lynagh committed
552

Ian Lynagh's avatar
Ian Lynagh committed
553
# else
554
555
556
    /*
      stack at call:
               argn
Ian Lynagh's avatar
Ian Lynagh committed
557
558
               ...
               arg7
559
               return address
Ian Lynagh's avatar
Ian Lynagh committed
560
               %rdi,%rsi,%rdx,%rcx,%r8,%r9 = arg1..arg6
561
562
563
564
565
566
567
568
569

      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
Ian Lynagh's avatar
Ian Lynagh committed
570
571
572
573
574
             ...
             arg7
             return address  *** <-- dummy arg in stub fn.
             arg6
             obscure_ccall_ret_code
575
576
577
578
579
580
581
582
583
584
585
586
587

      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
588
  16:   ff 25 0c 00 00 00       jmpq   *12(%rip)
589
590
  ... 
  20: .quad 0  # aligned on 8-byte boundary
591
  28: .quad 0  # aligned on 8-byte boundary
592
593
594
595
596


  And the version for >=6 integer arguments:

   0:   41 51                   push   %r9
597
598
599
600
601
602
603
604
   2:   ff 35 20 00 00 00       pushq  32(%rip)        # 28 <ccall_adjustor+0x28>
   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 <ccall_adjustor+0x30>
  1e:   ff 25 14 00 00 00       jmpq   *20(%rip)        # 38 <ccall_adjustor+0x38>
605
606
  ...
  28: .quad 0  # aligned on 8-byte boundary
607
608
  30: .quad 0  # aligned on 8-byte boundary
  38: .quad 0  # aligned on 8-byte boundary
609
610
611
    */

    {  
Ian Lynagh's avatar
Ian Lynagh committed
612
613
614
615
616
617
618
619
620
621
622
623
624
        int i = 0;
        char *c;
        StgWord8 *adj_code;

        // 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 != 'f' && *c != 'd') i++;
            if (i == 6) break;
        }

        if (i < 6) {
            adjustor = allocateExec(0x30,&code);
Simon Marlow's avatar
Simon Marlow committed
625
626
            adj_code = (StgWord8*)adjustor;

Ian Lynagh's avatar
Ian Lynagh committed
627
628
629
630
631
632
633
634
635
636
637
638
639
            *(StgInt32 *)adj_code        = 0x49c1894d;
            *(StgInt32 *)(adj_code+0x4)  = 0x8948c889;
            *(StgInt32 *)(adj_code+0x8)  = 0xf28948d1;
            *(StgInt32 *)(adj_code+0xc)  = 0x48fe8948;
            *(StgInt32 *)(adj_code+0x10) = 0x000a3d8b;
            *(StgInt32 *)(adj_code+0x14) = 0x25ff0000;
            *(StgInt32 *)(adj_code+0x18) = 0x0000000c;
            *(StgInt64 *)(adj_code+0x20) = (StgInt64)hptr;
            *(StgInt64 *)(adj_code+0x28) = (StgInt64)wptr;
        }
        else
        {
            adjustor = allocateExec(0x40,&code);
Simon Marlow's avatar
Simon Marlow committed
640
641
            adj_code = (StgWord8*)adjustor;

Ian Lynagh's avatar
Ian Lynagh committed
642
643
644
645
646
647
648
649
650
651
652
653
654
655
            *(StgInt32 *)adj_code        = 0x35ff5141;
            *(StgInt32 *)(adj_code+0x4)  = 0x00000020;
            *(StgInt32 *)(adj_code+0x8)  = 0x49c1894d;
            *(StgInt32 *)(adj_code+0xc)  = 0x8948c889;
            *(StgInt32 *)(adj_code+0x10) = 0xf28948d1;
            *(StgInt32 *)(adj_code+0x14) = 0x48fe8948;
            *(StgInt32 *)(adj_code+0x18) = 0x00123d8b;
            *(StgInt32 *)(adj_code+0x1c) = 0x25ff0000;
            *(StgInt32 *)(adj_code+0x20) = 0x00000014;
            
            *(StgInt64 *)(adj_code+0x28) = (StgInt64)obscure_ccall_ret_code;
            *(StgInt64 *)(adj_code+0x30) = (StgInt64)hptr;
            *(StgInt64 *)(adj_code+0x38) = (StgInt64)wptr;
        }
656
    }
Ian Lynagh's avatar
Ian Lynagh committed
657
658
659
# endif


660
#elif defined(sparc_HOST_ARCH)
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
  /* 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! :-)
689
  */
690
    adjustor = allocateExec(4*(11+1),&code);
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
    {
        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");
    }
725
#elif defined(alpha_HOST_ARCH)
ken's avatar
ken committed
726
727
728
729
730
  /* Magic constant computed by inspecting the code length of
     the following assembly language snippet
     (offset and machine code prefixed; note that the machine code
     shown is longwords stored in little-endian order):

Ian Lynagh's avatar
Ian Lynagh committed
731
732
733
734
735
736
737
738
  <00>: 46520414        mov     a2, a4
  <04>: 46100412        mov     a0, a2
  <08>: a61b0020        ldq     a0, 0x20(pv)    # load up hptr
  <0c>: 46730415        mov     a3, a5
  <10>: a77b0028        ldq     pv, 0x28(pv)    # load up wptr
  <14>: 46310413        mov     a1, a3
  <18>: 6bfb----        jmp     (pv), <hint>    # jump to wptr (with hint)
  <1c>: 00000000                                # padding for alignment
ken's avatar
ken committed
739
740
  <20>: [8 bytes for hptr quadword]
  <28>: [8 bytes for wptr quadword]
ken's avatar
ken committed
741
742
743
744
745
746

     The "computed" jump at <08> above is really a jump to a fixed
     location.  Accordingly, we place an always-correct hint in the
     jump instruction, namely the address offset from <0c> to wptr,
     divided by 4, taking the lowest 14 bits.

ken's avatar
ken committed
747
     We only support passing 4 or fewer argument words, for the same
748
     reason described under sparc_HOST_ARCH above by JRS, 21 Aug 01.
ken's avatar
ken committed
749
750
751
752
753
754
755
756
757
758
759
760
     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.

     On the Alpha the calling convention is so complex and dependent
     on the callee's signature -- for example, the stack pointer has
     to be a multiple of 16 -- that it seems impossible to me [ccshan]
     to handle the general case correctly without changing how the
     adjustor is called from C.  For now, our solution of shuffling
     registers only and ignoring the stack only works if the original
     caller passed 4 or fewer argument words.

ken's avatar
ken committed
761
762
763
764
765
766
TODO: Depending on how much allocation overhead stgMallocBytes uses for
      header information (more precisely, if the overhead is no more than
      4 bytes), we should move the first three instructions above down by
      4 bytes (getting rid of the nop), hence saving memory. [ccshan]
  */
    ASSERT(((StgWord64)wptr & 3) == 0);
767
    adjustor = allocateExec(48,&code);
768
    {
Ian Lynagh's avatar
Ian Lynagh committed
769
        StgWord64 *const code = (StgWord64 *)adjustor;
770

Ian Lynagh's avatar
Ian Lynagh committed
771
772
773
774
775
        code[0] = 0x4610041246520414L;
        code[1] = 0x46730415a61b0020L;
        code[2] = 0x46310413a77b0028L;
        code[3] = 0x000000006bfb0000L
                | (((StgWord32*)(wptr) - (StgWord32*)(code) - 3) & 0x3fff);
ken's avatar
ken committed
776

Ian Lynagh's avatar
Ian Lynagh committed
777
778
        code[4] = (StgWord64)hptr;
        code[5] = (StgWord64)wptr;
ken's avatar
ken committed
779

Ian Lynagh's avatar
Ian Lynagh committed
780
781
        /* Ensure that instruction cache is consistent with our new code */
        __asm__ volatile("call_pal %0" : : "i" (PAL_imb));
782
    }
783
#elif defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS)
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817

#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
            {
818
                if((t == 'l' || t == 'L') && src_gpr <= 9)
819
820
821
822
823
824
                {
                    if((src_gpr & 1) == 0)
                        src_gpr++;
                    src_locs[i] = -src_gpr;
                    src_gpr += 2;
                }
825
                else if((t == 'w' || t == 'W') && src_gpr <= 10)
826
827
828
829
830
                {
                    src_locs[i] = -(src_gpr++);
                }
                else
                {
831
                    if(t == 'l' || t == 'L' || t == 'd')
832
833
834
835
836
                    {
                        if(src_offset % 8)
                            src_offset += 4;
                    }
                    src_locs[i] = src_offset;
837
                    src_offset += (t == 'l' || t == 'L' || t == 'd') ? 8 : 4;
838
839
                }

840
                    if((t == 'l' || t == 'L') && dst_gpr <= 9)
841
842
843
844
845
846
                {
                    if((dst_gpr & 1) == 0)
                        dst_gpr++;
                    dst_locs[i] = -dst_gpr;
                    dst_gpr += 2;
                }
847
                else if((t == 'w' || t == 'W') && dst_gpr <= 10)
848
849
850
851
852
                {
                    dst_locs[i] = -(dst_gpr++);
                }
                else
                {
853
                    if(t == 'l' || t == 'L' || t == 'd')
854
855
856
857
858
                    {
                        if(dst_offset % 8)
                            dst_offset += 4;
                    }
                    dst_locs[i] = dst_offset;
859
                    dst_offset += (t == 'l' || t == 'L' || t == 'd') ? 8 : 4;
860
861
862
863
864
865
866
867
868
869
870
871
                }
            }
        }

        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.
872
        adjustor = allocateExec(4 * (4*n + 14),&code);
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
        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.

901
                    if(typeString[i] == 'l' || typeString[i] == 'L')
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
                    {
                            // 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
                {
917
                    if(typeString[i] == 'l' || typeString[i] == 'L')
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
                    {
                            // 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');

936
                if(typeString[i] == 'l' || typeString[i] == 'L')
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
                {
                    // 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");
        }
    }
991

992
#elif defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH)
993
994
995
        
#define OP_LO(op,lo)  ((((unsigned)(op)) << 16) | (((unsigned)(lo)) & 0xFFFF))
#define OP_HI(op,hi)  ((((unsigned)(op)) << 16) | (((unsigned)(hi)) >> 16))
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
    {
        /* 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;

1023
#ifdef FUNDESCS
1024
        adjustorStub = stgMallocBytes(sizeof(AdjustorStub), "createAdjustor");
1025
#else
1026
        adjustorStub = allocateExec(sizeof(AdjustorStub),&code);
1027
#endif
1028
1029
1030
        adjustor = adjustorStub;
            
        adjustorStub->code = (void*) &adjustorCode;
1031
1032

#ifdef FUNDESCS
1033
1034
1035
            // function descriptors are a cool idea.
            // We don't need to generate any code at runtime.
        adjustorStub->toc = adjustorStub;
1036
1037
#else

1038
1039
            // no function descriptors :-(
            // We need to do things "by hand".
1040
#if defined(powerpc_HOST_ARCH)
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
            // 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;
1052
#else
1053
        barf("adjustor creation not supported on this platform");
1054
1055
#endif

1056
1057
1058
1059
1060
        // Flush the Instruction cache:
        {
            int n = sizeof(AdjustorStub)/sizeof(unsigned);
            unsigned *p = (unsigned*)adjustor;
            while(n--)
1061
            {
1062
1063
1064
                __asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0"
                                    : : "r" (p));
                p++;
1065
            }
1066
1067
            __asm__ volatile ("sync\n\tisync");
        }
1068
1069
#endif

1070
            // Calculate the size of the stack frame, in words.
1071
1072
        sz = totalArgumentSize(typeString);
        
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
            // 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;
    }
1098

1099
#elif defined(ia64_HOST_ARCH)
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
/*
    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.

Ian Lynagh's avatar
Ian Lynagh committed
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
        [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;;
1134
1135
1136
*/

/* These macros distribute a long constant into the two words of an MLX bundle */
Ian Lynagh's avatar
Ian Lynagh committed
1137
1138
1139
1140
1141
1142
1143
1144
#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,0,7)    << 36)      \
                                | (BITS(val,7,9)    << 50)      \
                                | (BITS(val,16,5)   << 45)      \
                                | (BITS(val,21,1)   << 44)      \
                                | (BITS(val,40,23))             \
                                | (BITS(val,63,1)    << 59))
1145
1146

    {
Ian Lynagh's avatar
Ian Lynagh committed
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
        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 */
        /* Allocated memory is word-aligned (8 bytes) but functions on ia64
         * must be aligned to 16 bytes.  We allocate an extra 8 bytes of
         * wiggle room so that we can put the code on a 16 byte boundary. */
        adjustor = stgAllocStable(sizeof(IA64FunDesc)+18*8+8, &stable);

        fdesc = (IA64FunDesc *)adjustor;
        code = (StgWord64 *)(fdesc + 1);
        /* add 8 bytes to code if needed to align to a 16-byte boundary */
        if ((StgWord64)code & 15) code++;
        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;
1187
    }
ken's avatar
ken committed
1188
#else
1189
    barf("adjustor creation not supported on this platform");
1190
#endif
ken's avatar
ken committed
1191
    break;
sof's avatar
sof committed
1192
  
ken's avatar
ken committed
1193
1194
1195
  default:
    ASSERT(0);
    break;
sof's avatar
sof committed
1196
1197
1198
  }

  /* Have fun! */
1199
  return code;
sof's avatar
sof committed
1200
1201
}

ken's avatar
ken committed
1202

sof's avatar
sof committed
1203
void
1204
freeHaskellFunctionPtr(void* ptr)
sof's avatar
sof committed
1205
{
1206
1207
#if defined(i386_HOST_ARCH)
 if ( *(unsigned char*)ptr != 0xe8 &&
sof's avatar
sof committed
1208
      *(unsigned char*)ptr != 0x58 ) {
1209
   errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
sof's avatar
sof committed
1210
1211
   return;
 }
1212
1213
 if (*(unsigned char*)ptr == 0xe8) { /* Aha, a ccall adjustor! */
     freeStablePtr(((AdjustorStub*)ptr)->hptr);
sof's avatar
sof committed
1214
 } else {
1215
    freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02)));
1216
1217
1218
 }
#elif defined(x86_64_HOST_ARCH)
 if ( *(StgWord16 *)ptr == 0x894d ) {
1219
1220
1221
1222
1223
1224
1225
     freeStablePtr(*(StgStablePtr*)((StgWord8*)ptr+
#if defined(mingw32_HOST_OS)
                                                   0x28
#else
                                                   0x20
#endif
                                                       ));
1226
#if !defined(mingw32_HOST_OS)
1227
 } else if ( *(StgWord16 *)ptr == 0x5141 ) {
1228
1229
     freeStablePtr(*(StgStablePtr*)((StgWord8*)ptr+0x30));
#endif
1230
#if defined(mingw32_HOST_OS)
1231
 } else if ( *(StgWord16 *)ptr == 0x8348 ) {
Ian Lynagh's avatar
Ian Lynagh committed
1232
     freeStablePtr(*(StgStablePtr*)((StgWord8*)ptr+0x48));
1233
#endif
1234
1235
1236
1237
 } else {
   errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
   return;
 }
1238
#elif defined(sparc_HOST_ARCH)
1239
 if ( *(unsigned long*)ptr != 0x9C23A008UL ) {
1240
   errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
1241
1242
1243
1244
   return;
 }

 /* Free the stable pointer first..*/
1245
 freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 11)));
1246
#elif defined(alpha_HOST_ARCH)
ken's avatar
ken committed
1247
 if ( *(StgWord64*)ptr != 0xa77b0018a61b0010L ) {
1248
   errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
ken's avatar
ken committed
1249
1250
1251
1252
1253
   return;
 }

 /* Free the stable pointer first..*/
 freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10)));
1254
#elif defined(powerpc_HOST_ARCH) && defined(linux_HOST_OS)
1255
 if ( *(StgWord*)ptr != 0x48000008 ) {
1256
   errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
1257
1258
   return;
 }
1259
 freeStablePtr(((StgStablePtr*)ptr)[1]);
1260
#elif defined(powerpc_HOST_ARCH) || defined(powerpc64_HOST_ARCH)
1261
1262
1263
1264
1265
 if ( ((AdjustorStub*)ptr)->code != (StgFunPtr) &adjustorCode ) {
   errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
   return;
 }
 freeStablePtr(((AdjustorStub*)ptr)->hptr);
1266
#elif defined(ia64_HOST_ARCH)
1267
1268
1269
1270
 IA64FunDesc *fdesc = (IA64FunDesc *)ptr;
 StgWord64 *code = (StgWord64 *)(fdesc+1);

 if (fdesc->ip != (StgWord64)code) {
1271
   errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
1272
1273
1274
1275
1276
   return;
 }
 freeStablePtr((StgStablePtr)code[16]);
 freeStablePtr((StgStablePtr)code[17]);
 return;
ken's avatar
ken committed
1277
1278
#else
 ASSERT(0);
1279
#endif
1280
1281
 // Can't write to this memory, it is only executable:
 // *((unsigned char*)ptr) = '\0';
sof's avatar
sof committed
1282

1283
 freeExec(ptr);
sof's avatar
sof committed
1284
}
1285

1286
#endif // !USE_LIBFFI_FOR_ADJUSTORS