1 /* -----------------------------------------------------------------------
2 ffi.c - (c) 2003-2004 Randolph Chung <tausq@debian.org>
5 HPPA Foreign Function Interface
6 HP-UX PA ABI support (c) 2006 Free Software Foundation, Inc.
8 Permission is hereby granted, free of charge, to any person obtaining
9 a copy of this software and associated documentation files (the
10 ``Software''), to deal in the Software without restriction, including
11 without limitation the rights to use, copy, modify, merge, publish,
12 distribute, sublicense, and/or sell copies of the Software, and to
13 permit persons to whom the Software is furnished to do so, subject to
14 the following conditions:
16 The above copyright notice and this permission notice shall be included
17 in all copies or substantial portions of the Software.
19 THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
20 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
22 NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
23 HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
24 WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
25 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
26 DEALINGS IN THE SOFTWARE.
27 ----------------------------------------------------------------------- */
30 #include <ffi_common.h>
35 #define ROUND_UP(v, a) (((size_t)(v) + (a) - 1) & ~((a) - 1))
37 #define MIN_STACK_SIZE 64
38 #define FIRST_ARG_SLOT 9
41 #define fldw(addr, fpreg) \
42 __asm__ volatile ("fldw 0(%0), %%" #fpreg "L" : : "r"(addr) : #fpreg)
43 #define fstw(fpreg, addr) \
44 __asm__ volatile ("fstw %%" #fpreg "L, 0(%0)" : : "r"(addr))
45 #define fldd(addr, fpreg) \
46 __asm__ volatile ("fldd 0(%0), %%" #fpreg : : "r"(addr) : #fpreg)
47 #define fstd(fpreg, addr) \
48 __asm__ volatile ("fstd %%" #fpreg "L, 0(%0)" : : "r"(addr))
50 #define debug(lvl, x...) do { if (lvl <= DEBUG_LEVEL) { printf(x); } } while (0)
52 static inline int ffi_struct_type(ffi_type *t)
56 /* Small structure results are passed in registers,
57 larger ones are passed by pointer. Note that
58 small structures of size 2, 4 and 8 differ from
59 the corresponding integer types in that they have
60 different alignment requirements. */
63 return FFI_TYPE_UINT8;
65 return FFI_TYPE_SMALL_STRUCT2;
67 return FFI_TYPE_SMALL_STRUCT3;
69 return FFI_TYPE_SMALL_STRUCT4;
71 return FFI_TYPE_SMALL_STRUCT5;
73 return FFI_TYPE_SMALL_STRUCT6;
75 return FFI_TYPE_SMALL_STRUCT7;
77 return FFI_TYPE_SMALL_STRUCT8;
79 return FFI_TYPE_STRUCT; /* else, we pass it by pointer. */
82 /* PA has a downward growing stack, which looks like this:
86 SP = (4*(n+9)) arg word N
99 The first four argument words on the stack are reserved for use by
100 the callee. Instead, the general and floating registers replace
101 the first four argument slots. Non FP arguments are passed solely
102 in the general registers. FP arguments are passed in both general
103 and floating registers when using libffi.
105 Non-FP 32-bit args are passed in gr26, gr25, gr24 and gr23.
106 Non-FP 64-bit args are passed in register pairs, starting
107 on an odd numbered register (i.e. r25+r26 and r23+r24).
108 FP 32-bit arguments are passed in fr4L, fr5L, fr6L and fr7L.
109 FP 64-bit arguments are passed in fr5 and fr7.
111 The registers are allocated in the same manner as stack slots.
112 This allows the callee to save its arguments on the stack if
115 arg word 3 -> gr23 or fr7L
116 arg word 2 -> gr24 or fr6L or fr7R
117 arg word 1 -> gr25 or fr5L
118 arg word 0 -> gr26 or fr4L or fr5R
120 Note that fr4R and fr6R are never used for arguments (i.e.,
121 doubles are not passed in fr4 or fr6).
123 The rest of the arguments are passed on the stack starting at SP-52,
124 but 64-bit arguments need to be aligned to an 8-byte boundary
126 This means we can have holes either in the register allocation,
129 /* ffi_prep_args is called by the assembly routine once stack space
130 has been allocated for the function's arguments
132 The following code will put everything into the stack frame
133 (which was allocated by the asm routine), and on return
134 the asm routine will load the arguments that should be
135 passed by register into the appropriate registers
137 NOTE: We load floating point args in this function... that means we
138 assume gcc will not mess with fp regs in here. */
140 void ffi_prep_args_pa32(UINT32 *stack, extended_cif *ecif, unsigned bytes)
142 register unsigned int i;
143 register ffi_type **p_arg;
144 register void **p_argv;
145 unsigned int slot = FIRST_ARG_SLOT;
149 debug(1, "%s: stack = %p, ecif = %p, bytes = %u\n", __FUNCTION__, stack,
152 p_arg = ecif->cif->arg_types;
153 p_argv = ecif->avalue;
155 for (i = 0; i < ecif->cif->nargs; i++)
157 int type = (*p_arg)->type;
162 *(SINT32 *)(stack - slot) = *(SINT8 *)(*p_argv);
166 *(UINT32 *)(stack - slot) = *(UINT8 *)(*p_argv);
169 case FFI_TYPE_SINT16:
170 *(SINT32 *)(stack - slot) = *(SINT16 *)(*p_argv);
173 case FFI_TYPE_UINT16:
174 *(UINT32 *)(stack - slot) = *(UINT16 *)(*p_argv);
177 case FFI_TYPE_UINT32:
178 case FFI_TYPE_SINT32:
179 case FFI_TYPE_POINTER:
180 debug(3, "Storing UINT32 %u in slot %u\n", *(UINT32 *)(*p_argv),
182 *(UINT32 *)(stack - slot) = *(UINT32 *)(*p_argv);
185 case FFI_TYPE_UINT64:
186 case FFI_TYPE_SINT64:
187 /* Align slot for 64-bit type. */
188 slot += (slot & 1) ? 1 : 2;
189 *(UINT64 *)(stack - slot) = *(UINT64 *)(*p_argv);
193 /* First 4 args go in fr4L - fr7L. */
194 debug(3, "Storing UINT32(float) in slot %u\n", slot);
195 *(UINT32 *)(stack - slot) = *(UINT32 *)(*p_argv);
196 switch (slot - FIRST_ARG_SLOT)
198 /* First 4 args go in fr4L - fr7L. */
199 case 0: fldw(stack - slot, fr4); break;
200 case 1: fldw(stack - slot, fr5); break;
201 case 2: fldw(stack - slot, fr6); break;
202 case 3: fldw(stack - slot, fr7); break;
206 case FFI_TYPE_DOUBLE:
207 /* Align slot for 64-bit type. */
208 slot += (slot & 1) ? 1 : 2;
209 debug(3, "Storing UINT64(double) at slot %u\n", slot);
210 *(UINT64 *)(stack - slot) = *(UINT64 *)(*p_argv);
211 switch (slot - FIRST_ARG_SLOT)
213 /* First 2 args go in fr5, fr7. */
214 case 1: fldd(stack - slot, fr5); break;
215 case 3: fldd(stack - slot, fr7); break;
220 case FFI_TYPE_LONGDOUBLE:
221 /* Long doubles are passed in the same manner as structures
222 larger than 8 bytes. */
223 *(UINT32 *)(stack - slot) = (UINT32)(*p_argv);
227 case FFI_TYPE_STRUCT:
229 /* Structs smaller or equal than 4 bytes are passed in one
230 register. Structs smaller or equal 8 bytes are passed in two
231 registers. Larger structures are passed by pointer. */
233 len = (*p_arg)->size;
236 dest_cpy = (char *)(stack - slot) + 4 - len;
237 memcpy(dest_cpy, (char *)*p_argv, len);
241 slot += (slot & 1) ? 1 : 2;
242 dest_cpy = (char *)(stack - slot) + 8 - len;
243 memcpy(dest_cpy, (char *)*p_argv, len);
246 *(UINT32 *)(stack - slot) = (UINT32)(*p_argv);
258 /* Make sure we didn't mess up and scribble on the stack. */
262 debug(5, "Stack setup:\n");
263 for (n = 0; n < (bytes + 3) / 4; n++)
265 if ((n%4) == 0) { debug(5, "\n%08x: ", (unsigned int)(stack - n)); }
266 debug(5, "%08x ", *(stack - n));
271 FFI_ASSERT(slot * 4 <= bytes);
276 static void ffi_size_stack_pa32(ffi_cif *cif)
280 int z = 0; /* # stack slots */
282 for (ptr = cif->arg_types, i = 0; i < cif->nargs; ptr++, i++)
284 int type = (*ptr)->type;
288 case FFI_TYPE_DOUBLE:
289 case FFI_TYPE_UINT64:
290 case FFI_TYPE_SINT64:
291 z += 2 + (z & 1); /* must start on even regs, so we may waste one */
295 case FFI_TYPE_LONGDOUBLE:
297 case FFI_TYPE_STRUCT:
298 z += 1; /* pass by ptr, callee will copy */
301 default: /* <= 32-bit values */
306 /* We can fit up to 6 args in the default 64-byte stack frame,
307 if we need more, we need more stack. */
309 cif->bytes = MIN_STACK_SIZE; /* min stack size */
311 cif->bytes = 64 + ROUND_UP((z - 6) * sizeof(UINT32), MIN_STACK_SIZE);
313 debug(3, "Calculated stack size is %u bytes\n", cif->bytes);
316 /* Perform machine dependent cif processing. */
317 ffi_status ffi_prep_cif_machdep(ffi_cif *cif)
319 /* Set the return type flag */
320 switch (cif->rtype->type)
324 case FFI_TYPE_DOUBLE:
325 cif->flags = (unsigned) cif->rtype->type;
329 case FFI_TYPE_LONGDOUBLE:
330 /* Long doubles are treated like a structure. */
331 cif->flags = FFI_TYPE_STRUCT;
335 case FFI_TYPE_STRUCT:
336 /* For the return type we have to check the size of the structures.
337 If the size is smaller or equal 4 bytes, the result is given back
338 in one register. If the size is smaller or equal 8 bytes than we
339 return the result in two registers. But if the size is bigger than
340 8 bytes, we work with pointers. */
341 cif->flags = ffi_struct_type(cif->rtype);
344 case FFI_TYPE_UINT64:
345 case FFI_TYPE_SINT64:
346 cif->flags = FFI_TYPE_UINT64;
350 cif->flags = FFI_TYPE_INT;
354 /* Lucky us, because of the unique PA ABI we get to do our
359 ffi_size_stack_pa32(cif);
370 extern void ffi_call_pa32(void (*)(UINT32 *, extended_cif *, unsigned),
371 extended_cif *, unsigned, unsigned, unsigned *,
374 void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
379 ecif.avalue = avalue;
381 /* If the return value is a struct and we don't have a return
382 value address then we need to make one. */
386 && (cif->rtype->type == FFI_TYPE_STRUCT
387 || cif->rtype->type == FFI_TYPE_LONGDOUBLE))
389 && cif->rtype->type == FFI_TYPE_STRUCT)
392 ecif.rvalue = alloca(cif->rtype->size);
395 ecif.rvalue = rvalue;
401 debug(3, "Calling ffi_call_pa32: ecif=%p, bytes=%u, flags=%u, rvalue=%p, fn=%p\n", &ecif, cif->bytes, cif->flags, ecif.rvalue, (void *)fn);
402 ffi_call_pa32(ffi_prep_args_pa32, &ecif, cif->bytes,
403 cif->flags, ecif.rvalue, fn);
413 /* This is more-or-less an inverse of ffi_call -- we have arguments on
414 the stack, and we need to fill them into a cif structure and invoke
415 the user function. This really ought to be in asm to make sure
416 the compiler doesn't do things we don't expect. */
417 ffi_status ffi_closure_inner_pa32(ffi_closure *closure, UINT32 *stack)
422 UINT32 ret[2]; /* function can return up to 64-bits in registers */
426 unsigned int slot = FIRST_ARG_SLOT;
427 register UINT32 r28 asm("r28");
431 /* If returning via structure, callee will write to our pointer. */
432 if (cif->flags == FFI_TYPE_STRUCT)
433 rvalue = (void *)r28;
437 avalue = (void **)alloca(cif->nargs * FFI_SIZEOF_ARG);
439 p_arg = cif->arg_types;
441 for (i = 0; i < avn; i++)
443 int type = (*p_arg)->type;
449 case FFI_TYPE_SINT16:
450 case FFI_TYPE_UINT16:
451 case FFI_TYPE_SINT32:
452 case FFI_TYPE_UINT32:
453 case FFI_TYPE_POINTER:
454 avalue[i] = (char *)(stack - slot) + sizeof(UINT32) - (*p_arg)->size;
457 case FFI_TYPE_SINT64:
458 case FFI_TYPE_UINT64:
459 slot += (slot & 1) ? 1 : 2;
460 avalue[i] = (void *)(stack - slot);
465 /* The closure call is indirect. In Linux, floating point
466 arguments in indirect calls with a prototype are passed
467 in the floating point registers instead of the general
468 registers. So, we need to replace what was previously
469 stored in the current slot with the value in the
470 corresponding floating point register. */
471 switch (slot - FIRST_ARG_SLOT)
473 case 0: fstw(fr4, (void *)(stack - slot)); break;
474 case 1: fstw(fr5, (void *)(stack - slot)); break;
475 case 2: fstw(fr6, (void *)(stack - slot)); break;
476 case 3: fstw(fr7, (void *)(stack - slot)); break;
479 avalue[i] = (void *)(stack - slot);
482 case FFI_TYPE_DOUBLE:
483 slot += (slot & 1) ? 1 : 2;
485 /* See previous comment for FFI_TYPE_FLOAT. */
486 switch (slot - FIRST_ARG_SLOT)
488 case 1: fstd(fr5, (void *)(stack - slot)); break;
489 case 3: fstd(fr7, (void *)(stack - slot)); break;
492 avalue[i] = (void *)(stack - slot);
496 case FFI_TYPE_LONGDOUBLE:
497 /* Long doubles are treated like a big structure. */
498 avalue[i] = (void *) *(stack - slot);
502 case FFI_TYPE_STRUCT:
503 /* Structs smaller or equal than 4 bytes are passed in one
504 register. Structs smaller or equal 8 bytes are passed in two
505 registers. Larger structures are passed by pointer. */
506 if((*p_arg)->size <= 4)
508 avalue[i] = (void *)(stack - slot) + sizeof(UINT32) -
511 else if ((*p_arg)->size <= 8)
513 slot += (slot & 1) ? 1 : 2;
514 avalue[i] = (void *)(stack - slot) + sizeof(UINT64) -
518 avalue[i] = (void *) *(stack - slot);
529 /* Invoke the closure. */
530 (closure->fun) (cif, rvalue, avalue, closure->user_data);
532 debug(3, "after calling function, ret[0] = %08x, ret[1] = %08x\n", ret[0],
535 /* Store the result using the lower 2 bytes of the flags. */
539 *(stack - FIRST_ARG_SLOT) = (UINT8)(ret[0] >> 24);
542 *(stack - FIRST_ARG_SLOT) = (SINT8)(ret[0] >> 24);
544 case FFI_TYPE_UINT16:
545 *(stack - FIRST_ARG_SLOT) = (UINT16)(ret[0] >> 16);
547 case FFI_TYPE_SINT16:
548 *(stack - FIRST_ARG_SLOT) = (SINT16)(ret[0] >> 16);
551 case FFI_TYPE_SINT32:
552 case FFI_TYPE_UINT32:
553 *(stack - FIRST_ARG_SLOT) = ret[0];
555 case FFI_TYPE_SINT64:
556 case FFI_TYPE_UINT64:
557 *(stack - FIRST_ARG_SLOT) = ret[0];
558 *(stack - FIRST_ARG_SLOT - 1) = ret[1];
561 case FFI_TYPE_DOUBLE:
569 case FFI_TYPE_STRUCT:
570 /* Don't need a return value, done by caller. */
573 case FFI_TYPE_SMALL_STRUCT2:
574 case FFI_TYPE_SMALL_STRUCT3:
575 case FFI_TYPE_SMALL_STRUCT4:
576 tmp = (void*)(stack - FIRST_ARG_SLOT);
577 tmp += 4 - cif->rtype->size;
578 memcpy((void*)tmp, &ret[0], cif->rtype->size);
581 case FFI_TYPE_SMALL_STRUCT5:
582 case FFI_TYPE_SMALL_STRUCT6:
583 case FFI_TYPE_SMALL_STRUCT7:
584 case FFI_TYPE_SMALL_STRUCT8:
586 unsigned int ret2[2];
589 /* Right justify ret[0] and ret[1] */
592 case FFI_TYPE_SMALL_STRUCT5: off = 3; break;
593 case FFI_TYPE_SMALL_STRUCT6: off = 2; break;
594 case FFI_TYPE_SMALL_STRUCT7: off = 1; break;
595 default: off = 0; break;
598 memset (ret2, 0, sizeof (ret2));
599 memcpy ((char *)ret2 + off, ret, 8 - off);
601 *(stack - FIRST_ARG_SLOT) = ret2[0];
602 *(stack - FIRST_ARG_SLOT - 1) = ret2[1];
606 case FFI_TYPE_POINTER:
611 debug(0, "assert with cif->flags: %d\n",cif->flags);
618 /* Fill in a closure to refer to the specified fun and user_data.
619 cif specifies the argument and result types for fun.
620 The cif must already be prep'ed. */
622 extern void ffi_closure_pa32(void);
625 ffi_prep_closure_loc (ffi_closure* closure,
627 void (*fun)(ffi_cif*,void*,void**,void*),
631 UINT32 *tramp = (UINT32 *)(closure->tramp);
636 FFI_ASSERT (cif->abi == FFI_PA32);
638 /* Make a small trampoline that will branch to our
639 handler function. Use PC-relative addressing. */
642 tramp[0] = 0xeaa00000; /* b,l .+8,%r21 ; %r21 <- pc+8 */
643 tramp[1] = 0xd6a01c1e; /* depi 0,31,2,%r21 ; mask priv bits */
644 tramp[2] = 0x4aa10028; /* ldw 20(%r21),%r1 ; load plabel */
645 tramp[3] = 0x36b53ff1; /* ldo -8(%r21),%r21 ; get closure addr */
646 tramp[4] = 0x0c201096; /* ldw 0(%r1),%r22 ; address of handler */
647 tramp[5] = 0xeac0c000; /* bv%r0(%r22) ; branch to handler */
648 tramp[6] = 0x0c281093; /* ldw 4(%r1),%r19 ; GP of handler */
649 tramp[7] = ((UINT32)(ffi_closure_pa32) & ~2);
651 /* Flush d/icache -- have to flush up 2 two lines because of
656 "fic 0(%%sr4, %0)\n\t"
657 "fic %1(%%sr4, %0)\n\t"
667 : "r"((unsigned long)tramp & ~31),
673 tramp[0] = 0xeaa00000; /* b,l .+8,%r21 ; %r21 <- pc+8 */
674 tramp[1] = 0xd6a01c1e; /* depi 0,31,2,%r21 ; mask priv bits */
675 tramp[2] = 0x4aa10038; /* ldw 28(%r21),%r1 ; load plabel */
676 tramp[3] = 0x36b53ff1; /* ldo -8(%r21),%r21 ; get closure addr */
677 tramp[4] = 0x0c201096; /* ldw 0(%r1),%r22 ; address of handler */
678 tramp[5] = 0x02c010b4; /* ldsid (%r22),%r20 ; load space id */
679 tramp[6] = 0x00141820; /* mtsp %r20,%sr0 ; into %sr0 */
680 tramp[7] = 0xe2c00000; /* be 0(%sr0,%r22) ; branch to handler */
681 tramp[8] = 0x0c281093; /* ldw 4(%r1),%r19 ; GP of handler */
682 tramp[9] = ((UINT32)(ffi_closure_pa32) & ~2);
684 /* Flush d/icache -- have to flush three lines because of alignment. */
693 "fic,m %2(%%sr0,%0)\n\t"
694 "fic,m %2(%%sr0,%0)\n\t"
695 "fic,m %2(%%sr0,%0)\n\t"
704 : "=&r" ((unsigned long)tmp)
705 : "r" ((unsigned long)tramp & ~31),
711 closure->user_data = user_data;