1 /* -----------------------------------------------------------------------
2 ffi.c - (c) 2003-2004 Randolph Chung <tausq@debian.org>
4 HPPA Foreign Function Interface
5 HP-UX PA ABI support (c) 2006 Free Software Foundation, Inc.
7 Permission is hereby granted, free of charge, to any person obtaining
8 a copy of this software and associated documentation files (the
9 ``Software''), to deal in the Software without restriction, including
10 without limitation the rights to use, copy, modify, merge, publish,
11 distribute, sublicense, and/or sell copies of the Software, and to
12 permit persons to whom the Software is furnished to do so, subject to
13 the following conditions:
15 The above copyright notice and this permission notice shall be included
16 in all copies or substantial portions of the Software.
18 THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL CYGNUS SOLUTIONS BE LIABLE FOR ANY CLAIM, DAMAGES OR
22 OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
23 ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
24 OTHER DEALINGS IN THE SOFTWARE.
25 ----------------------------------------------------------------------- */
28 #include <ffi_common.h>
33 #define ROUND_UP(v, a) (((size_t)(v) + (a) - 1) & ~((a) - 1))
35 #define MIN_STACK_SIZE 64
36 #define FIRST_ARG_SLOT 9
39 #define fldw(addr, fpreg) \
40 __asm__ volatile ("fldw 0(%0), %%" #fpreg "L" : : "r"(addr) : #fpreg)
41 #define fstw(fpreg, addr) \
42 __asm__ volatile ("fstw %%" #fpreg "L, 0(%0)" : : "r"(addr))
43 #define fldd(addr, fpreg) \
44 __asm__ volatile ("fldd 0(%0), %%" #fpreg : : "r"(addr) : #fpreg)
45 #define fstd(fpreg, addr) \
46 __asm__ volatile ("fstd %%" #fpreg "L, 0(%0)" : : "r"(addr))
48 #define debug(lvl, x...) do { if (lvl <= DEBUG_LEVEL) { printf(x); } } while (0)
50 static inline int ffi_struct_type(ffi_type *t)
54 /* Small structure results are passed in registers,
55 larger ones are passed by pointer. Note that
56 small structures of size 2, 4 and 8 differ from
57 the corresponding integer types in that they have
58 different alignment requirements. */
61 return FFI_TYPE_UINT8;
63 return FFI_TYPE_SMALL_STRUCT2;
65 return FFI_TYPE_SMALL_STRUCT3;
67 return FFI_TYPE_SMALL_STRUCT4;
69 return FFI_TYPE_SMALL_STRUCT5;
71 return FFI_TYPE_SMALL_STRUCT6;
73 return FFI_TYPE_SMALL_STRUCT7;
75 return FFI_TYPE_SMALL_STRUCT8;
77 return FFI_TYPE_STRUCT; /* else, we pass it by pointer. */
80 /* PA has a downward growing stack, which looks like this:
84 SP = (4*(n+9)) arg word N
97 The first four argument words on the stack are reserved for use by
98 the callee. Instead, the general and floating registers replace
99 the first four argument slots. Non FP arguments are passed solely
100 in the general registers. FP arguments are passed in both general
101 and floating registers when using libffi.
103 Non-FP 32-bit args are passed in gr26, gr25, gr24 and gr23.
104 Non-FP 64-bit args are passed in register pairs, starting
105 on an odd numbered register (i.e. r25+r26 and r23+r24).
106 FP 32-bit arguments are passed in fr4L, fr5L, fr6L and fr7L.
107 FP 64-bit arguments are passed in fr5 and fr7.
109 The registers are allocated in the same manner as stack slots.
110 This allows the callee to save its arguments on the stack if
113 arg word 3 -> gr23 or fr7L
114 arg word 2 -> gr24 or fr6L or fr7R
115 arg word 1 -> gr25 or fr5L
116 arg word 0 -> gr26 or fr4L or fr5R
118 Note that fr4R and fr6R are never used for arguments (i.e.,
119 doubles are not passed in fr4 or fr6).
121 The rest of the arguments are passed on the stack starting at SP-52,
122 but 64-bit arguments need to be aligned to an 8-byte boundary
124 This means we can have holes either in the register allocation,
127 /* ffi_prep_args is called by the assembly routine once stack space
128 has been allocated for the function's arguments
130 The following code will put everything into the stack frame
131 (which was allocated by the asm routine), and on return
132 the asm routine will load the arguments that should be
133 passed by register into the appropriate registers
135 NOTE: We load floating point args in this function... that means we
136 assume gcc will not mess with fp regs in here. */
139 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(*p_argv, fr4); break;
200 case 1: fldw(*p_argv, fr5); break;
201 case 2: fldw(*p_argv, fr6); break;
202 case 3: fldw(*p_argv, 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(*p_argv, fr5); break;
215 case 3: fldd(*p_argv, 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);
372 extern void ffi_call_pa32(void (*)(UINT32 *, extended_cif *, unsigned),
373 /*@out@*/ extended_cif *,
375 /*@out@*/ unsigned *,
380 void ffi_call(/*@dependent@*/ ffi_cif *cif,
382 /*@out@*/ void *rvalue,
383 /*@dependent@*/ void **avalue)
388 ecif.avalue = avalue;
390 /* If the return value is a struct and we don't have a return
391 value address then we need to make one. */
395 && (cif->rtype->type == FFI_TYPE_STRUCT
396 || cif->rtype->type == FFI_TYPE_LONGDOUBLE))
398 && cif->rtype->type == FFI_TYPE_STRUCT)
402 ecif.rvalue = alloca(cif->rtype->size);
406 ecif.rvalue = rvalue;
413 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);
414 ffi_call_pa32(ffi_prep_args_pa32, &ecif, cif->bytes,
415 cif->flags, ecif.rvalue, fn);
426 /* This is more-or-less an inverse of ffi_call -- we have arguments on
427 the stack, and we need to fill them into a cif structure and invoke
428 the user function. This really ought to be in asm to make sure
429 the compiler doesn't do things we don't expect. */
430 ffi_status ffi_closure_inner_pa32(ffi_closure *closure, UINT32 *stack)
435 UINT32 ret[2]; /* function can return up to 64-bits in registers */
439 unsigned int slot = FIRST_ARG_SLOT;
440 register UINT32 r28 asm("r28");
444 /* If returning via structure, callee will write to our pointer. */
445 if (cif->flags == FFI_TYPE_STRUCT)
446 rvalue = (void *)r28;
450 avalue = (void **)alloca(cif->nargs * FFI_SIZEOF_ARG);
452 p_arg = cif->arg_types;
454 for (i = 0; i < avn; i++)
456 int type = (*p_arg)->type;
462 case FFI_TYPE_SINT16:
463 case FFI_TYPE_UINT16:
464 case FFI_TYPE_SINT32:
465 case FFI_TYPE_UINT32:
466 case FFI_TYPE_POINTER:
467 avalue[i] = (char *)(stack - slot) + sizeof(UINT32) - (*p_arg)->size;
470 case FFI_TYPE_SINT64:
471 case FFI_TYPE_UINT64:
472 slot += (slot & 1) ? 1 : 2;
473 avalue[i] = (void *)(stack - slot);
478 /* The closure call is indirect. In Linux, floating point
479 arguments in indirect calls with a prototype are passed
480 in the floating point registers instead of the general
481 registers. So, we need to replace what was previously
482 stored in the current slot with the value in the
483 corresponding floating point register. */
484 switch (slot - FIRST_ARG_SLOT)
486 case 0: fstw(fr4, (void *)(stack - slot)); break;
487 case 1: fstw(fr5, (void *)(stack - slot)); break;
488 case 2: fstw(fr6, (void *)(stack - slot)); break;
489 case 3: fstw(fr7, (void *)(stack - slot)); break;
492 avalue[i] = (void *)(stack - slot);
495 case FFI_TYPE_DOUBLE:
496 slot += (slot & 1) ? 1 : 2;
498 /* See previous comment for FFI_TYPE_FLOAT. */
499 switch (slot - FIRST_ARG_SLOT)
501 case 1: fstd(fr5, (void *)(stack - slot)); break;
502 case 3: fstd(fr7, (void *)(stack - slot)); break;
505 avalue[i] = (void *)(stack - slot);
508 case FFI_TYPE_STRUCT:
509 /* Structs smaller or equal than 4 bytes are passed in one
510 register. Structs smaller or equal 8 bytes are passed in two
511 registers. Larger structures are passed by pointer. */
512 if((*p_arg)->size <= 4)
514 avalue[i] = (void *)(stack - slot) + sizeof(UINT32) -
517 else if ((*p_arg)->size <= 8)
519 slot += (slot & 1) ? 1 : 2;
520 avalue[i] = (void *)(stack - slot) + sizeof(UINT64) -
524 avalue[i] = (void *) *(stack - slot);
535 /* Invoke the closure. */
536 (closure->fun) (cif, rvalue, avalue, closure->user_data);
538 debug(3, "after calling function, ret[0] = %08x, ret[1] = %08x\n", ret[0],
541 /* Store the result using the lower 2 bytes of the flags. */
545 *(stack - FIRST_ARG_SLOT) = (UINT8)(ret[0] >> 24);
548 *(stack - FIRST_ARG_SLOT) = (SINT8)(ret[0] >> 24);
550 case FFI_TYPE_UINT16:
551 *(stack - FIRST_ARG_SLOT) = (UINT16)(ret[0] >> 16);
553 case FFI_TYPE_SINT16:
554 *(stack - FIRST_ARG_SLOT) = (SINT16)(ret[0] >> 16);
557 case FFI_TYPE_SINT32:
558 case FFI_TYPE_UINT32:
559 *(stack - FIRST_ARG_SLOT) = ret[0];
561 case FFI_TYPE_SINT64:
562 case FFI_TYPE_UINT64:
563 *(stack - FIRST_ARG_SLOT) = ret[0];
564 *(stack - FIRST_ARG_SLOT - 1) = ret[1];
567 case FFI_TYPE_DOUBLE:
575 case FFI_TYPE_STRUCT:
576 /* Don't need a return value, done by caller. */
579 case FFI_TYPE_SMALL_STRUCT2:
580 case FFI_TYPE_SMALL_STRUCT3:
581 case FFI_TYPE_SMALL_STRUCT4:
582 tmp = (void*)(stack - FIRST_ARG_SLOT);
583 tmp += 4 - cif->rtype->size;
584 memcpy((void*)tmp, &ret[0], cif->rtype->size);
587 case FFI_TYPE_SMALL_STRUCT5:
588 case FFI_TYPE_SMALL_STRUCT6:
589 case FFI_TYPE_SMALL_STRUCT7:
590 case FFI_TYPE_SMALL_STRUCT8:
592 unsigned int ret2[2];
595 /* Right justify ret[0] and ret[1] */
598 case FFI_TYPE_SMALL_STRUCT5: off = 3; break;
599 case FFI_TYPE_SMALL_STRUCT6: off = 2; break;
600 case FFI_TYPE_SMALL_STRUCT7: off = 1; break;
601 default: off = 0; break;
604 memset (ret2, 0, sizeof (ret2));
605 memcpy ((char *)ret2 + off, ret, 8 - off);
607 *(stack - FIRST_ARG_SLOT) = ret2[0];
608 *(stack - FIRST_ARG_SLOT - 1) = ret2[1];
612 case FFI_TYPE_POINTER:
617 debug(0, "assert with cif->flags: %d\n",cif->flags);
624 /* Fill in a closure to refer to the specified fun and user_data.
625 cif specifies the argument and result types for fun.
626 The cif must already be prep'ed. */
628 extern void ffi_closure_pa32(void);
631 ffi_prep_closure (ffi_closure* closure,
633 void (*fun)(ffi_cif*,void*,void**,void*),
636 UINT32 *tramp = (UINT32 *)(closure->tramp);
641 FFI_ASSERT (cif->abi == FFI_PA32);
643 /* Make a small trampoline that will branch to our
644 handler function. Use PC-relative addressing. */
647 tramp[0] = 0xeaa00000; /* b,l .+8,%r21 ; %r21 <- pc+8 */
648 tramp[1] = 0xd6a01c1e; /* depi 0,31,2,%r21 ; mask priv bits */
649 tramp[2] = 0x4aa10028; /* ldw 20(%r21),%r1 ; load plabel */
650 tramp[3] = 0x36b53ff1; /* ldo -8(%r21),%r21 ; get closure addr */
651 tramp[4] = 0x0c201096; /* ldw 0(%r1),%r22 ; address of handler */
652 tramp[5] = 0xeac0c000; /* bv%r0(%r22) ; branch to handler */
653 tramp[6] = 0x0c281093; /* ldw 4(%r1),%r19 ; GP of handler */
654 tramp[7] = ((UINT32)(ffi_closure_pa32) & ~2);
656 /* Flush d/icache -- have to flush up 2 two lines because of
661 "fic 0(%%sr4, %0)\n\t"
662 "fic %1(%%sr4, %0)\n\t"
672 : "r"((unsigned long)tramp & ~31),
678 tramp[0] = 0xeaa00000; /* b,l .+8,%r21 ; %r21 <- pc+8 */
679 tramp[1] = 0xd6a01c1e; /* depi 0,31,2,%r21 ; mask priv bits */
680 tramp[2] = 0x4aa10038; /* ldw 28(%r21),%r1 ; load plabel */
681 tramp[3] = 0x36b53ff1; /* ldo -8(%r21),%r21 ; get closure addr */
682 tramp[4] = 0x0c201096; /* ldw 0(%r1),%r22 ; address of handler */
683 tramp[5] = 0x02c010b4; /* ldsid (%r22),%r20 ; load space id */
684 tramp[6] = 0x00141820; /* mtsp %r20,%sr0 ; into %sr0 */
685 tramp[7] = 0xe2c00000; /* be 0(%sr0,%r22) ; branch to handler */
686 tramp[8] = 0x0c281093; /* ldw 4(%r1),%r19 ; GP of handler */
687 tramp[9] = ((UINT32)(ffi_closure_pa32) & ~2);
689 /* Flush d/icache -- have to flush three lines because of alignment. */
698 "fic,m %2(%%sr0,%0)\n\t"
699 "fic,m %2(%%sr0,%0)\n\t"
700 "fic,m %2(%%sr0,%0)\n\t"
709 : "=&r" ((unsigned long)tmp)
710 : "r" ((unsigned long)tramp & ~31),
716 closure->user_data = user_data;