[ctypes-commit] ctypes/source/gcc/libffi/src/ia64 ffi.c,NONE,1.1 ffitarget.h,NONE,1.1 ia64_flags.h,NONE,1.1 unix.S,NONE,1.1

[Thomas H. <th...@us...>]

Update of /cvsroot/ctypes/ctypes/source/gcc/libffi/src/ia64
In directory sc8-pr-cvs1.sourceforge.net:/tmp/cvs-serv12302/gcc/libffi/src/ia64

Added Files:
	ffi.c ffitarget.h ia64_flags.h unix.S 
Log Message:
Importing libffi sources and build tools.


--- NEW FILE: unix.S ---
/* -----------------------------------------------------------------------
   unix.S - Copyright (c) 1998 Red Hat, Inc.
            Copyright (c) 2000 Hewlett Packard Company
   
   IA64/unix Foreign Function Interface 

   Primary author: Hans Boehm, HP Labs

   Loosely modeled on Cygnus code for other platforms.

   Permission is hereby granted, free of charge, to any person obtaining
   a copy of this software and associated documentation files (the
   ``Software''), to deal in the Software without restriction, including
   without limitation the rights to use, copy, modify, merge, publish,
   distribute, sublicense, and/or sell copies of the Software, and to
   permit persons to whom the Software is furnished to do so, subject to
   the following conditions:

   The above copyright notice and this permission notice shall be included
   in all copies or substantial portions of the Software.

   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND, EXPRESS
   OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
   IN NO EVENT SHALL CYGNUS SOLUTIONS BE LIABLE FOR ANY CLAIM, DAMAGES OR
   OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
   ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
   OTHER DEALINGS IN THE SOFTWARE.
   ----------------------------------------------------------------------- */

#define LIBFFI_ASM	
#include <fficonfig.h>
#include <ffi.h>
#include "ia64_flags.h"

/* parameters:	*/
#define callback	in0
#define ecifp		in1
#define bytes		in2
#define flags		in3
#define raddr		in4
#define fn		in5

#define FLOAT_SZ	8 /* in-memory size of fp operands	*/

/* Allocate an ia64_args structure on the stack; call ffi_prep_args	*/
/* to fill it in with argument values; copy those to the real 		*/
/* registers, leaving overflow arguments on the stack.  Then call fn	*/
/* and move the result from registers into *raddr.			*/
	.pred.safe_across_calls p1-p5,p16-p63
.text
        .align 16
        .global ffi_call_unix
        .proc ffi_call_unix
ffi_call_unix:
	.prologue
	.save	ar.pfs,r38 /* loc0 */
	alloc   loc0=ar.pfs,6,6,8,0
	.save	rp,loc1
	mov 	loc1=b0;
	.vframe	loc5
	mov	loc5=sp;
	.body
	sub	sp=sp,bytes
	mov	loc4=r1		/* Save gp 	*/
	ld8	r8=[callback],8	/* code address of callback	*/
	;;
	mov 	out0=sp
	mov	out1=ecifp
	mov	out2=bytes
	ld8	r1=[callback]	/* Set up gp for callback.  Unnecessary? */
	mov	b6=r8
	;;
	br.call.sptk.many b0 = b6	/* call ffi_prep_args		*/
	cmp.eq	p6,p0=0,r8		/* r8 nonzero ==> need fp regs	*/
 	;;
(p6)	add	loc2=32+8*FLOAT_SZ,sp
(p6)	br.cond.dptk.many	fp_done
	;;	/* Quiets warning; needed?	*/
	add	loc2=32,sp
	add	loc3=32+FLOAT_SZ,sp
	;;
	ldfd	f8=[loc2],2*FLOAT_SZ
	ldfd	f9=[loc3],2*FLOAT_SZ
	;;
	ldfd	f10=[loc2],2*FLOAT_SZ
	ldfd	f11=[loc3],2*FLOAT_SZ
	;;
	ldfd	f12=[loc2],2*FLOAT_SZ
	ldfd	f13=[loc3],2*FLOAT_SZ
	;;
	ldfd	f14=[loc2],2*FLOAT_SZ
	ldfd	f15=[loc3]
	;;
fp_done:
	add	r9=16,sp	/* Pointer to r8_contents	*/
	/* loc2 points at first integer register value.  */
	add	loc3=8,loc2
	;;
	ld8	r8=[r9]		/* Just in case we return large struct */
	ld8	out0=[loc2],16
	ld8	out1=[loc3],16
	;;
	ld8	out2=[loc2],16
	ld8	out3=[loc3],16
	;;
	ld8	out4=[loc2],16
	ld8	out5=[loc3],16
	;;
	ld8	out6=[loc2]
	ld8	out7=[loc3]
        /* Set sp to 16 bytes below the first stack parameter.  This    */
        /* is the value currently in loc2.                              */
	mov	sp=loc2
	
	ld8 	r8=[fn],8
	;;
	ld8	r1=[fn]		/* Set up gp */
	mov	b6=r8;;
	br.call.sptk.many b0 = b6	/* call fn	*/
	
	/* Handle return value. */
	cmp.eq	p6,p0=0,raddr
	cmp.eq	p7,p0=FFI_TYPE_INT,flags
	cmp.eq	p10,p0=FFI_IS_SMALL_STRUCT2,flags
	cmp.eq	p11,p0=FFI_IS_SMALL_STRUCT3,flags
	cmp.eq	p12,p0=FFI_IS_SMALL_STRUCT4,flags
	;;
(p6) 	br.cond.dpnt.few done		/* Dont copy ret values if raddr = 0 */
(p7)	br.cond.dptk.few copy1
(p10)	br.cond.dpnt.few copy2
(p11)	br.cond.dpnt.few copy3
(p12)	br.cond.dpnt.few copy4
	cmp.eq	p8,p0=FFI_TYPE_FLOAT,flags
	cmp.eq	p9,p0=FFI_TYPE_DOUBLE,flags
	tbit.nz	p6,p0=flags,FLOAT_FP_AGGREGATE_BIT
	tbit.nz	p7,p0=flags,DOUBLE_FP_AGGREGATE_BIT
	;;
(p8)	stfs	[raddr]=f8
(p9)	stfd	[raddr]=f8
	;;
	.label_state 1
(p6)	br.cond.dpnt.few handle_float_hfa
(p7)	br.cond.dpnt.few handle_double_hfa
	br done

copy4:
	add	loc3=24,raddr
	;;
	st8	[loc3]=r11
copy3:
	add	loc3=16,raddr
	;;
	st8	[loc3]=r10
copy2:
	add	loc3=8,raddr
	;;
	st8	[loc3]=r9
copy1:
	st8	[raddr]=r8
	/* In the big struct case, raddr was passed as an argument.	*/
	/* In the void case there was nothing to do.			*/

done:
	mov	r1=loc4		/* Restore gp	*/
	mov	ar.pfs = loc0
	mov	b0 = loc1
	.restore sp
	mov	sp = loc5
	br.ret.sptk.many b0

handle_double_hfa:
	.body
	.copy_state 1
	/* Homogeneous floating point array of doubles is returned in	*/
	/* registers f8-f15.  Save one at a time to return area.	*/
	and	flags=0xf,flags	/* Retrieve size	*/
	;;
	cmp.eq	p6,p0=2,flags
	cmp.eq	p7,p0=3,flags
	cmp.eq	p8,p0=4,flags
	cmp.eq	p9,p0=5,flags
	cmp.eq	p10,p0=6,flags
	cmp.eq	p11,p0=7,flags
	cmp.eq	p12,p0=8,flags
	;;
(p6)	br.cond.dptk.few	dhfa2
(p7)	br.cond.dptk.few	dhfa3
(p8)	br.cond.dptk.few	dhfa4
(p9)	br.cond.dptk.few	dhfa5
(p10)	br.cond.dptk.few	dhfa6
(p11)	br.cond.dptk.few	dhfa7
dhfa8:	add 	loc3=7*8,raddr
	;;
	stfd	[loc3]=f15
dhfa7:	add 	loc3=6*8,raddr
	;;
	stfd	[loc3]=f14
dhfa6:	add 	loc3=5*8,raddr
	;;
	stfd	[loc3]=f13
dhfa5:	add 	loc3=4*8,raddr
	;;
	stfd	[loc3]=f12
dhfa4:	add 	loc3=3*8,raddr
	;;
	stfd	[loc3]=f11
dhfa3:	add 	loc3=2*8,raddr
	;;
	stfd	[loc3]=f10
dhfa2:	add 	loc3=1*8,raddr
	;;
	stfd	[loc3]=f9
	stfd	[raddr]=f8
	br	done

handle_float_hfa:
	/* Homogeneous floating point array of floats is returned in	*/
	/* registers f8-f15.  Save one at a time to return area.	*/
	and	flags=0xf,flags	/* Retrieve size	*/
	;;
	cmp.eq	p6,p0=2,flags
	cmp.eq	p7,p0=3,flags
	cmp.eq	p8,p0=4,flags
	cmp.eq	p9,p0=5,flags
	cmp.eq	p10,p0=6,flags
	cmp.eq	p11,p0=7,flags
	cmp.eq	p12,p0=8,flags
	;;
(p6)	br.cond.dptk.few	shfa2
(p7)	br.cond.dptk.few	shfa3
(p8)	br.cond.dptk.few	shfa4
(p9)	br.cond.dptk.few	shfa5
(p10)	br.cond.dptk.few	shfa6
(p11)	br.cond.dptk.few	shfa7
shfa8:	add 	loc3=7*4,raddr
	;;
	stfd	[loc3]=f15
shfa7:	add 	loc3=6*4,raddr
	;;
	stfd	[loc3]=f14
shfa6:	add 	loc3=5*4,raddr
	;;
	stfd	[loc3]=f13
shfa5:	add 	loc3=4*4,raddr
	;;
	stfd	[loc3]=f12
shfa4:	add 	loc3=3*4,raddr
	;;
	stfd	[loc3]=f11
shfa3:	add 	loc3=2*4,raddr
	;;
	stfd	[loc3]=f10
shfa2:	add 	loc3=1*4,raddr
	;;
	stfd	[loc3]=f9
	stfd	[raddr]=f8
	br	done

        .endp ffi_call_unix


	.pred.safe_across_calls p1-p5,p16-p63
.text
        .align 16
        .global ffi_closure_UNIX
        .proc ffi_closure_UNIX
ffi_closure_UNIX:
	.prologue
	.save 	ar.pfs,r40 /* loc0 */
	alloc   loc0=ar.pfs,8,3,2,0
	.save	rp,loc1
	mov	loc1=b0
	.vframe	loc2
	mov	loc2=sp
	/* Retrieve closure pointer and real gp.	*/
	mov	out0=gp
	add	gp=16,gp
	;;
	ld8	gp=[gp]
	/* Reserve a structia64_args on the stack such that arguments	*/
	/* past the first 8 are automatically placed in the right	*/
	/* slot.  Note that when we start the sp points at 2 8-byte	*/
	/* scratch words, followed by the extra arguments.		*/
#	define BASIC_ARGS_SZ (8*FLOAT_SZ+8*8+2*8)
#	define FIRST_FP_OFFSET (4*8)
	add	r14=-(BASIC_ARGS_SZ-FIRST_FP_OFFSET),sp
	add	r15=-(BASIC_ARGS_SZ-FIRST_FP_OFFSET-FLOAT_SZ),sp
	add	sp=-BASIC_ARGS_SZ,sp
	/* r14 points to fp_regs[0], r15 points to fp_regs[1]	*/
	;;
	stfd	[r14]=f8,2*FLOAT_SZ
	stfd	[r15]=f9,2*FLOAT_SZ
	;;
	stfd	[r14]=f10,2*FLOAT_SZ
	stfd	[r15]=f11,2*FLOAT_SZ
	;;
	stfd	[r14]=f12,2*FLOAT_SZ
	stfd	[r15]=f13,2*FLOAT_SZ
	;;
	stfd	[r14]=f14,2*FLOAT_SZ
	stfd	[r15]=f15,FLOAT_SZ+8
	;;
	/* r14 points to first parameter register area, r15 to second. */
	st8	[r14]=in0,2*8
	st8	[r15]=in1,2*8
	;;
	st8	[r14]=in2,2*8
	st8	[r15]=in3,2*8
	;;
	st8	[r14]=in4,2*8
	st8	[r15]=in5,2*8
	;;
	st8	[r14]=in6,2*8
	st8	[r15]=in7,2*8
	/* Call ffi_closure_UNIX_inner */
	mov	out1=sp
	br.call.sptk.many b0=ffi_closure_UNIX_inner
	;;
	mov	b0=loc1
	mov 	ar.pfs=loc0
	.restore sp
	mov	sp=loc2
	br.ret.sptk.many b0
	.endp ffi_closure_UNIX
	


--- NEW FILE: ffi.c ---
/* -----------------------------------------------------------------------
   ffi.c - Copyright (c) 1998 Red Hat, Inc.
	   Copyright (c) 2000 Hewlett Packard Company
   
   IA64 Foreign Function Interface 

   Permission is hereby granted, free of charge, to any person obtaining
   a copy of this software and associated documentation files (the
   ``Software''), to deal in the Software without restriction, including
   without limitation the rights to use, copy, modify, merge, publish,
   distribute, sublicense, and/or sell copies of the Software, and to
   permit persons to whom the Software is furnished to do so, subject to
   the following conditions:

   The above copyright notice and this permission notice shall be included
   in all copies or substantial portions of the Software.

   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND, EXPRESS
   OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
   IN NO EVENT SHALL CYGNUS SOLUTIONS BE LIABLE FOR ANY CLAIM, DAMAGES OR
   OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
   ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
   OTHER DEALINGS IN THE SOFTWARE.
   ----------------------------------------------------------------------- */

#include <ffi.h>
#include <ffi_common.h>

#include <stdlib.h>
#include <stdbool.h>

#include "ia64_flags.h"

/* Memory image of fp register contents.  Should eventually be an fp 	*/
/* type long enough to hold an entire register.  For now we use double.	*/
typedef double float80;

/* The stack layout at call to ffi_prep_args.  Other_args will remain	*/
/* on the stack for the actual call.  Everything else we be transferred	*/
/* to registers and popped by the assembly code.			*/

struct ia64_args {
    long scratch[2];	/* Two scratch words at top of stack.		*/
			/* Allows sp to be passed as arg pointer.	*/
    void * r8_contents;	/* Value to be passed in r8			*/
    long spare;		/* Not used.					*/
    float80 fp_regs[8]; /* Contents of 8 floating point argument 	*/
			/* registers.					*/
    long out_regs[8];	/* Contents of the 8 out registers used 	*/
			/* for integer parameters.			*/
    long other_args[0]; /* Arguments passed on stack, variable size	*/
			/* Treated as continuation of out_regs.		*/
};

static size_t float_type_size(unsigned short tp)
{
  switch(tp) {
    case FFI_TYPE_FLOAT:
      return sizeof(float);
    case FFI_TYPE_DOUBLE:
      return sizeof(double);
#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
    case FFI_TYPE_LONGDOUBLE:
      return sizeof(long double);
#endif
    default:
      FFI_ASSERT(0);
  }
}

/*
 * Is type a struct containing at most n floats, doubles, or extended
 * doubles, all of the same fp type?
 * If so, set *element_type to the fp type.
 */
static bool is_homogeneous_fp_aggregate(ffi_type * type, int n,
				        unsigned short * element_type)
{
  ffi_type **ptr; 
  unsigned short element, struct_element;

  int type_set = 0;

  FFI_ASSERT(type != NULL);

  FFI_ASSERT(type->elements != NULL);

  ptr = &(type->elements[0]);

  while ((*ptr) != NULL)
    {
      switch((*ptr) -> type) {
	case FFI_TYPE_FLOAT:
	  if (type_set && element != FFI_TYPE_FLOAT) return 0;
	  if (--n < 0) return false;
	  type_set = 1;
	  element = FFI_TYPE_FLOAT;
	  break;
	case FFI_TYPE_DOUBLE:
	  if (type_set && element != FFI_TYPE_DOUBLE) return 0;
	  if (--n < 0) return false;
	  type_set = 1;
	  element = FFI_TYPE_DOUBLE;
	  break;
	case FFI_TYPE_STRUCT:
	  if (!is_homogeneous_fp_aggregate(type, n, &struct_element))
	      return false;
	  if (type_set && struct_element != element) return false;
	  n -= (type -> size)/float_type_size(element);
	  element = struct_element;
	  if (n < 0) return false;
	  break;
	/* case FFI_TYPE_LONGDOUBLE:
	  Not yet implemented.	*/
	default:
	  return false;
      }
      ptr++;
    }
  *element_type = element;
  return true;
   
} 

/* ffi_prep_args is called by the assembly routine once stack space
   has been allocated for the function's arguments.  It fills in
   the arguments in the structure referenced by stack. Returns nonzero
   if fp registers are used for arguments. */

static bool
ffi_prep_args(struct ia64_args *stack, extended_cif *ecif, int bytes)
{
  register long i, avn;
  register void **p_argv;
  register long *argp = stack -> out_regs;
  register float80 *fp_argp = stack -> fp_regs;
  register ffi_type **p_arg;

  /* For big return structs, r8 needs to contain the target address.	*/
  /* Since r8 is otherwise dead, we set it unconditionally.		*/
  stack -> r8_contents = ecif -> rvalue;
  i = 0;
  avn = ecif->cif->nargs;
  p_arg = ecif->cif->arg_types;
  p_argv = ecif->avalue;
  while (i < avn)
    {
      size_t z; /* z is in units of arg slots or words, not bytes.	*/

      switch ((*p_arg)->type)
	{
	case FFI_TYPE_SINT8:
	  z = 1;
	  *(SINT64 *) argp = *(SINT8 *)(* p_argv);
	  break;
		  
	case FFI_TYPE_UINT8:
	  z = 1;
	  *(UINT64 *) argp = *(UINT8 *)(* p_argv);
	  break;
		  
	case FFI_TYPE_SINT16:
	  z = 1;
	  *(SINT64 *) argp = *(SINT16 *)(* p_argv);
	  break;
		  
	case FFI_TYPE_UINT16:
	  z = 1;
	  *(UINT64 *) argp = *(UINT16 *)(* p_argv);
	  break;
		  
	case FFI_TYPE_SINT32:
	  z = 1;
	  *(SINT64 *) argp = *(SINT32 *)(* p_argv);
	  break;
		  
	case FFI_TYPE_UINT32:
	  z = 1;
	  *(UINT64 *) argp = *(UINT32 *)(* p_argv);
	  break;

	case FFI_TYPE_SINT64:
	case FFI_TYPE_UINT64:
	case FFI_TYPE_POINTER:
	  z = 1;
	  *(UINT64 *) argp = *(UINT64 *)(* p_argv);
	  break;

	case FFI_TYPE_FLOAT:
	  z = 1;
	  if (fp_argp - stack->fp_regs < 8)
	    {
	      /* Note the conversion -- all the fp regs are loaded as
		 doubles.  */
	      *fp_argp++ = *(float *)(* p_argv);
	    }
	  /* Also put it into the integer registers or memory: */
	  *(UINT64 *) argp = *(UINT32 *)(* p_argv);
	  break;

	case FFI_TYPE_DOUBLE:
	  z = 1;
	  if (fp_argp - stack->fp_regs < 8)
	    *fp_argp++ = *(double *)(* p_argv);
	  /* Also put it into the integer registers or memory: */
	  *(double *) argp = *(double *)(* p_argv);
	  break;

	case FFI_TYPE_STRUCT:
	  {
	      size_t sz = (*p_arg)->size;
	      unsigned short element_type;
              z = ((*p_arg)->size + FFI_SIZEOF_ARG - 1)/FFI_SIZEOF_ARG;
	      if (is_homogeneous_fp_aggregate(*p_arg, 8, &element_type)) {
		int i;
		int nelements = sz/float_type_size(element_type);
		for (i = 0; i < nelements; ++i) {
		  switch (element_type) {
		    case FFI_TYPE_FLOAT:
		      if (fp_argp - stack->fp_regs < 8)
			*fp_argp++ = ((float *)(* p_argv))[i];
		      break;
		    case FFI_TYPE_DOUBLE:
		      if (fp_argp - stack->fp_regs < 8)
			*fp_argp++ = ((double *)(* p_argv))[i];
		      break;
		    default:
			/* Extended precision not yet implemented. */
			abort();
		  }
		}
	      }
	      /* And pass it in integer registers as a struct, with	*/
	      /* its actual field sizes packed into registers.		*/
	      memcpy(argp, *p_argv, (*p_arg)->size);
	  }
	  break;

	default:
	  FFI_ASSERT(0);
	}

      argp += z;
      i++, p_arg++, p_argv++;
    }
  return (fp_argp != stack -> fp_regs);
}

/* Perform machine dependent cif processing */
ffi_status
ffi_prep_cif_machdep(ffi_cif *cif)
{
  long i, avn;
  bool is_simple = true;
  long simple_flag = FFI_SIMPLE_V;
  /* Adjust cif->bytes to include space for the 2 scratch words,
     r8 register contents, spare word,
     the 8 fp register contents, and all 8 integer register contents.
     This will be removed before the call, though 2 scratch words must
     remain.  */

  cif->bytes += 4*sizeof(long) + 8 *sizeof(float80);
  if (cif->bytes < sizeof(struct ia64_args))
    cif->bytes = sizeof(struct ia64_args);

  /* The stack must be double word aligned, so round bytes up
     appropriately. */

  cif->bytes = ALIGN(cif->bytes, 2*sizeof(void*));

  avn = cif->nargs;
  if (avn <= 2) {
    for (i = 0; i < avn; ++i) {
      switch(cif -> arg_types[i] -> type) {
	case FFI_TYPE_SINT32:
	  simple_flag = FFI_ADD_INT_ARG(simple_flag);
	  break;
	case FFI_TYPE_SINT64:
	case FFI_TYPE_UINT64:
	case FFI_TYPE_POINTER:
	  simple_flag = FFI_ADD_LONG_ARG(simple_flag);
	  break;
	default:
	  is_simple = false;
      }
    }
  } else {
    is_simple = false;
  }

  /* Set the return type flag */
  switch (cif->rtype->type)
    {
    case FFI_TYPE_VOID:
      cif->flags = FFI_TYPE_VOID;
      break;

    case FFI_TYPE_STRUCT:
      {
        size_t sz = cif -> rtype -> size;
  	unsigned short element_type;

	is_simple = false;
  	if (is_homogeneous_fp_aggregate(cif -> rtype, 8, &element_type)) {
	  int nelements = sz/float_type_size(element_type);
	  if (nelements <= 1) {
	    if (0 == nelements) {
	      cif -> flags = FFI_TYPE_VOID;
	    } else {
	      cif -> flags = element_type;
	    }
	  } else {
	    switch(element_type) {
	      case FFI_TYPE_FLOAT:
	        cif -> flags = FFI_IS_FLOAT_FP_AGGREGATE | nelements;
		break;
	      case FFI_TYPE_DOUBLE:
	        cif -> flags = FFI_IS_DOUBLE_FP_AGGREGATE | nelements;
		break;
	      default:
		/* long double NYI */
		abort();
	    }
	  }
	  break;
        }
        if (sz <= 32) {
	  if (sz <= 8) {
              cif->flags = FFI_TYPE_INT;
  	  } else if (sz <= 16) {
              cif->flags = FFI_IS_SMALL_STRUCT2;
  	  } else if (sz <= 24) {
              cif->flags = FFI_IS_SMALL_STRUCT3;
	  } else {
              cif->flags = FFI_IS_SMALL_STRUCT4;
	  }
        } else {
          cif->flags = FFI_TYPE_STRUCT;
	}
      }
      break;

    case FFI_TYPE_FLOAT:
      is_simple = false;
      cif->flags = FFI_TYPE_FLOAT;
      break;

    case FFI_TYPE_DOUBLE:
      is_simple = false;
      cif->flags = FFI_TYPE_DOUBLE;
      break;

    default:
      cif->flags = FFI_TYPE_INT;
      /* This seems to depend on little endian mode, and the fact that	*/
      /* the return pointer always points to at least 8 bytes.  But 	*/
      /* that also seems to be true for other platforms.		*/
      break;
    }
  
  if (is_simple) cif -> flags |= simple_flag;
  return FFI_OK;
}

extern int ffi_call_unix(bool (*)(struct ia64_args *, extended_cif *, int), 
			 extended_cif *, unsigned, 
			 unsigned, unsigned *, void (*)());

void
ffi_call(ffi_cif *cif, void (*fn)(), void *rvalue, void **avalue)
{
  extended_cif ecif;
  long simple = cif -> flags & FFI_SIMPLE;

  /* Should this also check for Unix ABI? */
  /* This is almost, but not quite, machine independent.  Note that	*/
  /* we can get away with not caring about length of the result because	*/
  /* we assume we are little endian, and the result buffer is large 	*/
  /* enough.								*/
  /* This needs work for HP/UX.						*/
  if (simple) {
    long (*lfn)() = (long (*)())fn;
    long result;
    switch(simple) {
      case FFI_SIMPLE_V:
	result = lfn();
	break;
      case FFI_SIMPLE_I:
	result = lfn(*(int *)avalue[0]);
	break;
      case FFI_SIMPLE_L:
	result = lfn(*(long *)avalue[0]);
	break;
      case FFI_SIMPLE_II:
	result = lfn(*(int *)avalue[0], *(int *)avalue[1]);
	break;
      case FFI_SIMPLE_IL:
	result = lfn(*(int *)avalue[0], *(long *)avalue[1]);
	break;
      case FFI_SIMPLE_LI:
	result = lfn(*(long *)avalue[0], *(int *)avalue[1]);
	break;
      case FFI_SIMPLE_LL:
	result = lfn(*(long *)avalue[0], *(long *)avalue[1]);
	break;
    }
    if ((cif->flags & ~FFI_SIMPLE) != FFI_TYPE_VOID && 0 != rvalue) {
      * (long *)rvalue = result;
    }
    return;
  }
  ecif.cif = cif;
  ecif.avalue = avalue;
  
  /* If the return value is a struct and we don't have a return
     value address then we need to make one.  */
  
  if (rvalue == NULL && cif->rtype->type == FFI_TYPE_STRUCT)
    ecif.rvalue = alloca(cif->rtype->size);
  else
    ecif.rvalue = rvalue;
    
  switch (cif->abi) 
    {
    case FFI_UNIX:
      ffi_call_unix(ffi_prep_args, &ecif, cif->bytes,
		    cif->flags, rvalue, fn);
      break;

    default:
      FFI_ASSERT(0);
      break;
    }
}

/*
 * Closures represent a pair consisting of a function pointer, and
 * some user data.  A closure is invoked by reinterpreting the closure
 * as a function pointer, and branching to it.  Thus we can make an
 * interpreted function callable as a C function:  We turn the interpreter
 * itself, together with a pointer specifying the interpreted procedure,
 * into a closure.
 * On X86, the first few words of the closure structure actually contain code,
 * which will do the right thing.  On most other architectures, this
 * would raise some Icache/Dcache coherence issues (which can be solved, but
 * often not cheaply).
 * For IA64, function pointer are already pairs consisting of a code
 * pointer, and a gp pointer.  The latter is needed to access global variables.
 * Here we set up such a pair as the first two words of the closure (in
 * the "trampoline" area), but we replace the gp pointer with a pointer
 * to the closure itself.  We also add the real gp pointer to the
 * closure.  This allows the function entry code to both retrieve the
 * user data, and to restire the correct gp pointer.
 */

static void 
ffi_prep_incoming_args_UNIX(struct ia64_args *args, void **rvalue,
			    void **avalue, ffi_cif *cif);

/* This function is entered with the doctored gp (r1) value.
 * This code is extremely gcc specific.  There is some argument that
 * it should really be written in assembly code, since it depends on
 * gcc properties that might change over time.
 */

/* ffi_closure_UNIX is an assembly routine, which copies the register 	*/
/* state into a struct ia64_args, and then invokes			*/
/* ffi_closure_UNIX_inner.  It also recovers the closure pointer	*/
/* from its fake gp pointer.						*/
void ffi_closure_UNIX();

#ifndef __GNUC__
#   error This requires gcc
#endif
void
ffi_closure_UNIX_inner (ffi_closure *closure, struct ia64_args * args)
/* Hopefully declaring this as a varargs function will force all args	*/
/* to memory.								*/
{
  // this is our return value storage
  long double    res;

  // our various things...
  ffi_cif       *cif;
  unsigned short rtype;
  void          *resp;
  void		**arg_area;

  resp = (void*)&res;
  cif         = closure->cif;
  arg_area    = (void**) alloca (cif->nargs * sizeof (void*));  

  /* this call will initialize ARG_AREA, such that each
   * element in that array points to the corresponding 
   * value on the stack; and if the function returns
   * a structure, it will re-set RESP to point to the
   * structure return address.  */

  ffi_prep_incoming_args_UNIX(args, (void**)&resp, arg_area, cif);
  
  (closure->fun) (cif, resp, arg_area, closure->user_data);

  rtype = cif->flags;

  /* now, do a generic return based on the value of rtype */
  if (rtype == FFI_TYPE_INT)
    {
      asm volatile ("ld8 r8=[%0]" : : "r" (resp) : "r8");
    }
  else if (rtype == FFI_TYPE_FLOAT)
    {
      asm volatile ("ldfs f8=[%0]" : : "r" (resp) : "f8");
    }
  else if (rtype == FFI_TYPE_DOUBLE)
    {
      asm volatile ("ldfd f8=[%0]" : : "r" (resp) : "f8");
    }
  else if (rtype == FFI_IS_SMALL_STRUCT2)
    {
      asm volatile ("ld8 r8=[%0]; ld8 r9=[%1]"
		    : : "r" (resp), "r" (resp+8) : "r8","r9");
    }
  else if (rtype == FFI_IS_SMALL_STRUCT3)
    {
      asm volatile ("ld8 r8=[%0]; ld8 r9=[%1]; ld8 r10=[%2]"
		    : : "r" (resp), "r" (resp+8), "r" (resp+16)
		    : "r8","r9","r10");
    }
  else if (rtype == FFI_IS_SMALL_STRUCT4)
    {
      asm volatile ("ld8 r8=[%0]; ld8 r9=[%1]; ld8 r10=[%2]; ld8 r11=[%3]"
		    : : "r" (resp), "r" (resp+8), "r" (resp+16), "r" (resp+24)
		    : "r8","r9","r10","r11");
    }
  else if (rtype != FFI_TYPE_VOID && rtype != FFI_TYPE_STRUCT)
    {
      /* Can only happen for homogeneous FP aggregates?	*/
      abort();
    }
}

static void 
ffi_prep_incoming_args_UNIX(struct ia64_args *args, void **rvalue,
			    void **avalue, ffi_cif *cif)
{
  register unsigned int i;
  register unsigned int avn;
  register void **p_argv;
  register long *argp = args -> out_regs;
  unsigned fp_reg_num = 0;
  register ffi_type **p_arg;

  avn = cif->nargs;
  p_argv = avalue;

  for (i = cif->nargs, p_arg = cif->arg_types; i != 0; i--, p_arg++)
    {
      size_t z; /* In units of words or argument slots.	*/

      switch ((*p_arg)->type)
	{
	case FFI_TYPE_SINT8:
	case FFI_TYPE_UINT8:
	case FFI_TYPE_SINT16:
	case FFI_TYPE_UINT16:
	case FFI_TYPE_SINT32:
	case FFI_TYPE_UINT32:
	case FFI_TYPE_SINT64:
	case FFI_TYPE_UINT64:
	case FFI_TYPE_POINTER:
	  z = 1;
	  *p_argv = (void *)argp;
	  break;
		  
	case FFI_TYPE_FLOAT:
	  z = 1;
	  /* Convert argument back to float in place from the saved value */
	  if (argp - args->out_regs < 8 && fp_reg_num < 8) {
	      *(float *)argp = args -> fp_regs[fp_reg_num++];
	  }
	  *p_argv = (void *)argp;
	  break;

	case FFI_TYPE_DOUBLE:
	  z = 1;
	  if (argp - args->out_regs < 8 && fp_reg_num < 8) {
	      *p_argv = args -> fp_regs + fp_reg_num++;
	  } else {
	      *p_argv = (void *)argp;
	  }
	  break;

	case FFI_TYPE_STRUCT:
	  {
	      size_t sz = (*p_arg)->size;
	      unsigned short element_type;
              z = ((*p_arg)->size + FFI_SIZEOF_ARG - 1)/FFI_SIZEOF_ARG;
	      if (argp - args->out_regs < 8
		  && is_homogeneous_fp_aggregate(*p_arg, 8, &element_type)) {
		int nelements = sz/float_type_size(element_type);
		if (nelements + fp_reg_num >= 8) {
		  /* hard case NYI.	*/
		  abort();
		}
		if (element_type == FFI_TYPE_DOUBLE) {
	          *p_argv = args -> fp_regs + fp_reg_num;
		  fp_reg_num += nelements;
		  break;
		}
		if (element_type == FFI_TYPE_FLOAT) {
		  int j;
		  for (j = 0; j < nelements; ++ j) {
		     ((float *)argp)[j] = args -> fp_regs[fp_reg_num + j];
		  }
	          *p_argv = (void *)argp;
		  fp_reg_num += nelements;
		  break;
		}
		abort();  /* Other fp types NYI */
	      }
	  }
	  break;

	default:
	  FFI_ASSERT(0);
	}

      argp += z;
      p_argv++;

    }
  
  return;
}


/* Fill in a closure to refer to the specified fun and user_data.	*/
/* cif specifies the argument and result types for fun.			*/
/* the cif must already be prep'ed */

/* The layout of a function descriptor.  A C function pointer really 	*/
/* points to one of these.						*/
typedef struct ia64_fd_struct {
    void *code_pointer;
    void *gp;
} ia64_fd;

ffi_status
ffi_prep_closure (ffi_closure* closure,
		  ffi_cif* cif,
		  void (*fun)(ffi_cif*,void*,void**,void*),
		  void *user_data)
{
  struct ffi_ia64_trampoline_struct *tramp =
    (struct ffi_ia64_trampoline_struct *) (closure -> tramp);
  ia64_fd *fd = (ia64_fd *)(void *)ffi_closure_UNIX;

  FFI_ASSERT (cif->abi == FFI_UNIX);

  tramp -> code_pointer = fd -> code_pointer;
  tramp -> real_gp = fd -> gp;
  tramp -> fake_gp = closure;
  closure->cif  = cif;
  closure->user_data = user_data;
  closure->fun  = fun;

  return FFI_OK;
}



--- NEW FILE: ia64_flags.h ---
/* -----------------------------------------------------------------------
   ia64_flags.h - Copyright (c) 2000 Hewlett Packard Company
   
   IA64/unix Foreign Function Interface 

   Original author: Hans Boehm, HP Labs

   Permission is hereby granted, free of charge, to any person obtaining
   a copy of this software and associated documentation files (the
   ``Software''), to deal in the Software without restriction, including
   without limitation the rights to use, copy, modify, merge, publish,
   distribute, sublicense, and/or sell copies of the Software, and to
   permit persons to whom the Software is furnished to do so, subject to
   the following conditions:

   The above copyright notice and this permission notice shall be included
   in all copies or substantial portions of the Software.

   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND, EXPRESS
   OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
   IN NO EVENT SHALL CYGNUS SOLUTIONS BE LIABLE FOR ANY CLAIM, DAMAGES OR
   OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
   ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
   OTHER DEALINGS IN THE SOFTWARE.
   ----------------------------------------------------------------------- */


/* Homogeneous Floating Point Aggregates (HFAs) which are returned	*/
/* in FP registers.  The least significant bits specify the size in 	*/
/* words.								*/
#define FFI_IS_FLOAT_FP_AGGREGATE 0x1000
#define FFI_IS_DOUBLE_FP_AGGREGATE 0x0800
#define FLOAT_FP_AGGREGATE_BIT 12
#define DOUBLE_FP_AGGREGATE_BIT 11

/* Small structures containing N words.  If N=1, they are returned	*/
/* as though they were integers.					*/
#define FFI_IS_SMALL_STRUCT2	0x40 /* Struct > 8, <=16 bytes	*/
#define FFI_IS_SMALL_STRUCT3	0x41 /* Struct > 16 <= 24 bytes	*/
#define FFI_IS_SMALL_STRUCT4	0x42 /* Struct > 24, <=32 bytes	*/

/* Flag values identifying particularly simple cases, which are 	*/
/* handled specially.  We treat functions as simple if they take all	*/
/* arguments can be passed as 32 or 64 bit integer quantities, there is	*/
/* either no return value or it can be treated as a 64bit integer, and	*/
/* if there are at most 2 arguments.					*/
/* This is OR'ed with the normal flag values.				*/
#define FFI_SIMPLE_V 0x10000	/* () -> X	*/
#define FFI_SIMPLE_I 0x20000	/* (int) -> X	*/
#define FFI_SIMPLE_L 0x30000	/* (long) -> X	*/
#define FFI_SIMPLE_II 0x40000	/* (int,int) -> X	*/
#define FFI_SIMPLE_IL 0x50000	/* (int,long) -> X	*/
#define FFI_SIMPLE_LI 0x60000	/* (long,int) -> X	*/
#define FFI_SIMPLE_LL 0x70000	/* (long,long) -> X	*/

/* Mask for all of the FFI_SIMPLE bits:	*/
#define FFI_SIMPLE 0xf0000

/* An easy way to build FFI_SIMPLE flags from FFI_SIMPLE_V:	*/
#define FFI_ADD_LONG_ARG(flag) (((flag) << 1) | 0x10000)
#define FFI_ADD_INT_ARG(flag) ((flag) << 1)

--- NEW FILE: ffitarget.h ---
/* -----------------------------------------------------------------*-C-*-
   ffitarget.h - Copyright (c) 1996-2003  Red Hat, Inc.
   Target configuration macros for IA-64.

   Permission is hereby granted, free of charge, to any person obtaining
   a copy of this software and associated documentation files (the
   ``Software''), to deal in the Software without restriction, including
   without limitation the rights to use, copy, modify, merge, publish,
   distribute, sublicense, and/or sell copies of the Software, and to
   permit persons to whom the Software is furnished to do so, subject to
   the following conditions:

   The above copyright notice and this permission notice shall be included
   in all copies or substantial portions of the Software.

   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND, EXPRESS
   OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
   IN NO EVENT SHALL CYGNUS SOLUTIONS BE LIABLE FOR ANY CLAIM, DAMAGES OR
   OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
   ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
   OTHER DEALINGS IN THE SOFTWARE.

   ----------------------------------------------------------------------- */

#ifndef LIBFFI_TARGET_H
#define LIBFFI_TARGET_H

#ifndef LIBFFI_ASM
typedef unsigned long          ffi_arg;
typedef signed long            ffi_sarg;

typedef enum ffi_abi {
  FFI_FIRST_ABI = 0,
  FFI_UNIX,   	/* Linux and all Unix variants use the same conventions	*/
  FFI_DEFAULT_ABI = FFI_UNIX,
  FFI_LAST_ABI = FFI_DEFAULT_ABI + 1
} ffi_abi;
#endif

/* ---- Definitions for closures ----------------------------------------- */

#define FFI_CLOSURES 1
#define FFI_TRAMPOLINE_SIZE 24  /* Really the following struct, which 	*/
				/* can be interpreted as a C function	*/
				/* descriptor:				*/

#ifndef LIBFFI_ASM
struct ffi_ia64_trampoline_struct {
    void * code_pointer;	/* Pointer to ffi_closure_UNIX	*/
    void * fake_gp;		/* Pointer to closure, installed as gp	*/
    void * real_gp;		/* Real gp value, reinstalled by 	*/
				/* ffi_closure_UNIX.			*/
};
#endif

#endif