Re: [patch, ARM] VFP hard-float calling convention support

[Dan Witte <dwitte at mozilla dot com>]

Ping on the commit? This would be useful for us to have for Firefox Mobile. Any plans to land or otherwise sync?

Cheers. :)

Dan

----- Original Message -----
> Anthony Green wrote:
> > Chung-Lin Tang<cltang@codesourcery.com> writes:
> >
> >> Hi,
> >> this patch implements VFP hard-float calling conventions, as a new
> >> ABI
> >> for ARM. Asides from VFP calling convention bits, some additional
> >> optimizations have been added to the current softfp assembly code
> >> paths too.
> >
> > Chung-Lin,
> >
> >    Sorry for the long delay. I think this patch is fine. Do you have
> >    commit rights to GCC? If so, please commit yourself. Otherwise,
> >    I'll
> >    push it into libffi git tree and sync with GCC in a week or so.
> >
> > Thank you!
> >
> > Anthony Green
> > 650 352-3402
> 
> Hi Anthony,
> thanks for the review. I have commit rights, but I think it would be
> better to sync with the git tree.
> 
> Thanks!
> Chung-Lin
> 
> 
> 
> 
> >
> >
> >>
> >> The FFI_DEFAULT_ABI value is now defined depending on a new GCC
> >> built-in preprocessor symbol __ARM_PCS_VFP, which was only recently
> >> added into GCC trunk; this seems to be the only reliable way of
> >> detecting the FP calling convention from within the source.
> >>
> >> Variadic function tests (cls_double_va.c and cls_longdouble_va.c in
> >> testsuite/libffi.call) has been disabled for VFP hard-float; the
> >> rules
> >> of AAPCS states that variadic functions switch back to the base
> >> (softfp) ABI, a condition on function type that seems not currently
> >> detectable within libffi.
> >>
> >> Testsuite patches have been added to support a dg-skip-if option
> >> (adapted from the GCC testsuite), to turn off the above two
> >> mentioned
> >> tests based on compiler options (skip on -mfloat-abi=hard). Patch
> >> was
> >> regression tested under both -mfloat-abi=hard and softfp with no
> >> unexpected failures.
> >>
> >> Comments on the patch are of course welcome.
> >> If approved, please help to commit to git head and GCC trunk.
> >>
> >> Thanks,
> >> Chung-Lin
> >>
> >>
> >> 2010-08-16 Chung-Lin Tang<cltang@codesourcery.com>
> >>
> >>     * src/arm/ffi.c (ffi_prep_args): Add VFP register argument
> >>     handling
> >>     code, new parameter, and return value. Update comments.
> >>     (ffi_prep_cif_machdep): Add case for VFP struct return values.
> >>     Add
> >>     call to layout_vfp_args().
> >>     (ffi_call_SYSV): Update declaration.
> >>     (ffi_call_VFP): New declaration.
> >>     (ffi_call): Add VFP struct return conditions. Call
> >>     ffi_call_VFP()
> >>     when ABI is FFI_VFP.
> >>     (ffi_closure_VFP): New declaration.
> >>     (ffi_closure_SYSV_inner): Add new vfp_args parameter, update
> >>     call to
> >>     ffi_prep_incoming_args_SYSV().
> >>     (ffi_prep_incoming_args_SYSV): Update parameters. Add VFP
> >>     argument
> >>     case handling.
> >>     (ffi_prep_closure_loc): Pass ffi_closure_VFP to trampoline
> >>     construction under VFP hard-float.
> >>     (rec_vfp_type_p): New function.
> >>     (vfp_type_p): Same.
> >>     (place_vfp_arg): Same.
> >>     (layout_vfp_args): Same.
> >>     * src/arm/ffitarget.h (ffi_abi): Add FFI_VFP. Define
> >>     FFI_DEFAULT_ABI
> >>     based on __ARM_PCS_VFP.
> >>     (FFI_EXTRA_CIF_FIELDS): Define for adding VFP hard-float
> >>     specific
> >>     fields.
> >>     (FFI_TYPE_STRUCT_VFP_FLOAT): Define internally used type code.
> >>     (FFI_TYPE_STRUCT_VFP_DOUBLE): Same.
> >>     * src/arm/sysv.S (ffi_call_SYSV): Change call of
> >>     ffi_prep_args() to
> >>     direct call. Move function pointer load upwards.
> >>     (ffi_call_VFP): New function.
> >>     (ffi_closure_VFP): Same.
> >>
> >>     * testsuite/lib/libffi-dg.exp (check-flags): New function.
> >>     (dg-skip-if): New function.
> >>     * testsuite/libffi.call/cls_double_va.c: Skip if target is
> >>     arm*-*-*
> >>     and compiler options include -mfloat-abi=hard.
> >>     * testsuite/libffi.call/cls_longdouble_va.c: Same.
> >>
> >> diff --git a/src/arm/ffi.c b/src/arm/ffi.c
> >> index f6a6475..9a0a53c 100644
> >> --- a/src/arm/ffi.c
> >> +++ b/src/arm/ffi.c
> >> @@ -29,12 +29,20 @@
> >>
> >>   #include<stdlib.h>
> >>
> >> -/* ffi_prep_args is called by the assembly routine once stack
> >> space
> >> - has been allocated for the function's arguments */
> >> +/* Forward declares. */
> >> +static int vfp_type_p (ffi_type *);
> >> +static void layout_vfp_args (ffi_cif *);
> >>
> >> -void ffi_prep_args(char *stack, extended_cif *ecif)
> >> +/* ffi_prep_args is called by the assembly routine once stack
> >> space
> >> + has been allocated for the function's arguments
> >> +
> >> + The vfp_space parameter is the load area for VFP regs, the return
> >> + value is cif->vfp_used (word bitset of VFP regs used for passing
> >> + arguments). These are only used for the VFP hard-float ABI.
> >> +*/
> >> +int ffi_prep_args(char *stack, extended_cif *ecif, float
> >> *vfp_space)
> >>   {
> >> - register unsigned int i;
> >> + register unsigned int i, vi = 0;
> >>     register void **p_argv;
> >>     register char *argp;
> >>     register ffi_type **p_arg;
> >> @@ -54,6 +62,21 @@ void ffi_prep_args(char *stack, extended_cif
> >> *ecif)
> >>       {
> >>         size_t z;
> >>
> >> + /* Allocated in VFP registers. */
> >> + if (ecif->cif->abi == FFI_VFP
> >> + && vi< ecif->cif->vfp_nargs&& vfp_type_p (*p_arg))
> >> + {
> >> + float* vfp_slot = vfp_space + ecif->cif->vfp_args[vi++];
> >> + if ((*p_arg)->type == FFI_TYPE_FLOAT)
> >> + *((float*)vfp_slot) = *((float*)*p_argv);
> >> + else if ((*p_arg)->type == FFI_TYPE_DOUBLE)
> >> + *((double*)vfp_slot) = *((double*)*p_argv);
> >> + else
> >> + memcpy(vfp_slot, *p_argv, (*p_arg)->size);
> >> + p_argv++;
> >> + continue;
> >> + }
> >> +
> >>         /* Align if necessary */
> >>         if (((*p_arg)->alignment - 1)& (unsigned) argp) {
> >>   	argp = (char *) ALIGN(argp, (*p_arg)->alignment);
> >> @@ -103,13 +126,15 @@ void ffi_prep_args(char *stack, extended_cif
> >> *ecif)
> >>   	  p_argv++;
> >>   	  argp += z;
> >>       }
> >> -
> >> - return;
> >> +
> >> + /* Indicate the VFP registers used. */
> >> + return ecif->cif->vfp_used;
> >>   }
> >>
> >>   /* Perform machine dependent cif processing */
> >>   ffi_status ffi_prep_cif_machdep(ffi_cif *cif)
> >>   {
> >> + int type_code;
> >>     /* Round the stack up to a multiple of 8 bytes. This isn't
> >>     needed
> >>        everywhere, but it is on some platforms, and it doesn't harm
> >>        anything
> >>        when it isn't needed. */
> >> @@ -130,7 +155,14 @@ ffi_status ffi_prep_cif_machdep(ffi_cif *cif)
> >>         break;
> >>
> >>       case FFI_TYPE_STRUCT:
> >> - if (cif->rtype->size<= 4)
> >> + if (cif->abi == FFI_VFP
> >> + && (type_code = vfp_type_p (cif->rtype)) != 0)
> >> + {
> >> + /* A Composite Type passed in VFP registers, either
> >> + FFI_TYPE_STRUCT_VFP_FLOAT or FFI_TYPE_STRUCT_VFP_DOUBLE. */
> >> + cif->flags = (unsigned) type_code;
> >> + }
> >> + else if (cif->rtype->size<= 4)
> >>   	/* A Composite Type not larger than 4 bytes is returned in r0.
> >>   	*/
> >>   	cif->flags = (unsigned)FFI_TYPE_INT;
> >>         else
> >> @@ -145,11 +177,18 @@ ffi_status ffi_prep_cif_machdep(ffi_cif *cif)
> >>         break;
> >>       }
> >>
> >> + /* Map out the register placements of VFP register args.
> >> + The VFP hard-float calling conventions are slightly more
> >> sophisticated than
> >> + the base calling conventions, so we do it here instead of in
> >> ffi_prep_args(). */
> >> + if (cif->abi == FFI_VFP)
> >> + layout_vfp_args (cif);
> >> +
> >>     return FFI_OK;
> >>   }
> >>
> >> -extern void ffi_call_SYSV(void (*)(char *, extended_cif *),
> >> extended_cif *,
> >> - unsigned, unsigned, unsigned *, void (*fn)(void));
> >> +/* Prototypes for assembly functions, in sysv.S */
> >> +extern void ffi_call_SYSV (void (*fn)(void), extended_cif *,
> >> unsigned, unsigned, unsigned *);
> >> +extern void ffi_call_VFP (void (*fn)(void), extended_cif *,
> >> unsigned, unsigned, unsigned *);
> >>
> >>   void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void
> >>   **avalue)
> >>   {
> >> @@ -157,6 +196,8 @@ void ffi_call(ffi_cif *cif, void (*fn)(void),
> >> void *rvalue, void **avalue)
> >>
> >>     int small_struct = (cif->flags == FFI_TYPE_INT
> >>   		&& cif->rtype->type == FFI_TYPE_STRUCT);
> >> + int vfp_struct = (cif->flags == FFI_TYPE_STRUCT_VFP_FLOAT
> >> + || cif->flags == FFI_TYPE_STRUCT_VFP_DOUBLE);
> >>
> >>     ecif.cif = cif;
> >>     ecif.avalue = avalue;
> >> @@ -173,38 +214,51 @@ void ffi_call(ffi_cif *cif, void (*fn)(void),
> >> void *rvalue, void **avalue)
> >>       }
> >>     else if (small_struct)
> >>       ecif.rvalue =&temp;
> >> + else if (vfp_struct)
> >> + {
> >> + /* Largest case is double x 4. */
> >> + ecif.rvalue = alloca(32);
> >> + }
> >>     else
> >>       ecif.rvalue = rvalue;
> >>
> >>     switch (cif->abi)
> >>       {
> >>       case FFI_SYSV:
> >> - ffi_call_SYSV(ffi_prep_args,&ecif, cif->bytes, cif->flags,
> >> ecif.rvalue,
> >> - fn);
> >> + ffi_call_SYSV (fn,&ecif, cif->bytes, cif->flags, ecif.rvalue);
> >> + break;
> >>
> >> + case FFI_VFP:
> >> + ffi_call_VFP (fn,&ecif, cif->bytes, cif->flags, ecif.rvalue);
> >>         break;
> >> +
> >>       default:
> >>         FFI_ASSERT(0);
> >>         break;
> >>       }
> >>     if (small_struct)
> >>       memcpy (rvalue,&temp, cif->rtype->size);
> >> + else if (vfp_struct)
> >> + memcpy (rvalue, ecif.rvalue, cif->rtype->size);
> >>   }
> >>
> >>   /** private members **/
> >>
> >>   static void ffi_prep_incoming_args_SYSV (char *stack, void **ret,
> >> - void** args, ffi_cif* cif);
> >> + void** args, ffi_cif* cif, float *vfp_stack);
> >>
> >>   void ffi_closure_SYSV (ffi_closure *);
> >>
> >> +void ffi_closure_VFP (ffi_closure *);
> >> +
> >>   /* This function is jumped to by the trampoline */
> >>
> >>   unsigned int
> >> -ffi_closure_SYSV_inner (closure, respp, args)
> >> +ffi_closure_SYSV_inner (closure, respp, args, vfp_args)
> >>        ffi_closure *closure;
> >>        void **respp;
> >>        void *args;
> >> + void *vfp_args;
> >>   {
> >>     // our various things...
> >>     ffi_cif *cif;
> >> @@ -219,7 +273,7 @@ ffi_closure_SYSV_inner (closure, respp, args)
> >>      * a structure, it will re-set RESP to point to the
> >>      * structure return address. */
> >>
> >> - ffi_prep_incoming_args_SYSV(args, respp, arg_area, cif);
> >> + ffi_prep_incoming_args_SYSV(args, respp, arg_area, cif,
> >> vfp_args);
> >>
> >>     (closure->fun) (cif, *respp, arg_area, closure->user_data);
> >>
> >> @@ -229,10 +283,12 @@ ffi_closure_SYSV_inner (closure, respp, args)
> >>   /*@-exportheader@*/
> >>   static void
> >>   ffi_prep_incoming_args_SYSV(char *stack, void **rvalue,
> >> - void **avalue, ffi_cif *cif)
> >> + void **avalue, ffi_cif *cif,
> >> + /* Used only under VFP hard-float ABI. */
> >> + float *vfp_stack)
> >>   /*@=exportheader@*/
> >>   {
> >> - register unsigned int i;
> >> + register unsigned int i, vi = 0;
> >>     register void **p_argv;
> >>     register char *argp;
> >>     register ffi_type **p_arg;
> >> @@ -249,8 +305,16 @@ ffi_prep_incoming_args_SYSV(char *stack, void
> >> **rvalue,
> >>     for (i = cif->nargs, p_arg = cif->arg_types; (i != 0); i--,
> >>     p_arg++)
> >>       {
> >>         size_t z;
> >> -
> >> - size_t alignment = (*p_arg)->alignment;
> >> + size_t alignment;
> >> +
> >> + if (cif->abi == FFI_VFP
> >> + && vi< cif->vfp_nargs&& vfp_type_p (*p_arg))
> >> + {
> >> + *p_argv++ = (void*)(vfp_stack + cif->vfp_args[vi++]);
> >> + continue;
> >> + }
> >> +
> >> + alignment = (*p_arg)->alignment;
> >>         if (alignment< 4)
> >>   	alignment = 4;
> >>         /* Align if necessary */
> >> @@ -295,10 +359,17 @@ ffi_prep_closure_loc (ffi_closure* closure,
> >>   		      void *user_data,
> >>   		      void *codeloc)
> >>   {
> >> - FFI_ASSERT (cif->abi == FFI_SYSV);
> >> + void (*closure_func)(ffi_closure*) = NULL;
> >>
> >> + if (cif->abi == FFI_SYSV)
> >> + closure_func =&ffi_closure_SYSV;
> >> + else if (cif->abi == FFI_VFP)
> >> + closure_func =&ffi_closure_VFP;
> >> + else
> >> + FFI_ASSERT (0);
> >> +
> >>     FFI_INIT_TRAMPOLINE (&closure->tramp[0], \
> >> - &ffi_closure_SYSV, \
> >> + closure_func, \
> >>   		       codeloc);
> >>
> >>     closure->cif = cif;
> >> @@ -307,3 +378,123 @@ ffi_prep_closure_loc (ffi_closure* closure,
> >>
> >>     return FFI_OK;
> >>   }
> >> +
> >> +/* Below are routines for VFP hard-float support. */
> >> +
> >> +static int rec_vfp_type_p (ffi_type *t, int *elt, int *elnum)
> >> +{
> >> + switch (t->type)
> >> + {
> >> + case FFI_TYPE_FLOAT:
> >> + case FFI_TYPE_DOUBLE:
> >> + *elt = (int) t->type;
> >> + *elnum = 1;
> >> + return 1;
> >> +
> >> + case FFI_TYPE_STRUCT_VFP_FLOAT:
> >> + *elt = FFI_TYPE_FLOAT;
> >> + *elnum = t->size / sizeof (float);
> >> + return 1;
> >> +
> >> + case FFI_TYPE_STRUCT_VFP_DOUBLE:
> >> + *elt = FFI_TYPE_DOUBLE;
> >> + *elnum = t->size / sizeof (double);
> >> + return 1;
> >> +
> >> + case FFI_TYPE_STRUCT:;
> >> + {
> >> + int base_elt = 0, total_elnum = 0;
> >> + ffi_type **el = t->elements;
> >> + while (*el)
> >> + {
> >> + int el_elt = 0, el_elnum = 0;
> >> + if (! rec_vfp_type_p (*el,&el_elt,&el_elnum)
> >> + || (base_elt&& base_elt != el_elt)
> >> + || total_elnum + el_elnum> 4)
> >> + return 0;
> >> + base_elt = el_elt;
> >> + total_elnum += el_elnum;
> >> + el++;
> >> + }
> >> + *elnum = total_elnum;
> >> + *elt = base_elt;
> >> + return 1;
> >> + }
> >> + default: ;
> >> + }
> >> + return 0;
> >> +}
> >> +
> >> +static int vfp_type_p (ffi_type *t)
> >> +{
> >> + int elt, elnum;
> >> + if (rec_vfp_type_p (t,&elt,&elnum))
> >> + {
> >> + if (t->type == FFI_TYPE_STRUCT)
> >> + {
> >> + if (elnum == 1)
> >> + t->type = elt;
> >> + else
> >> + t->type = (elt == FFI_TYPE_FLOAT
> >> + ? FFI_TYPE_STRUCT_VFP_FLOAT
> >> + : FFI_TYPE_STRUCT_VFP_DOUBLE);
> >> + }
> >> + return (int) t->type;
> >> + }
> >> + return 0;
> >> +}
> >> +
> >> +static void place_vfp_arg (ffi_cif *cif, ffi_type *t)
> >> +{
> >> + int reg = cif->vfp_reg_free;
> >> + int nregs = t->size / sizeof (float);
> >> + int align = ((t->type == FFI_TYPE_STRUCT_VFP_FLOAT
> >> + || t->type == FFI_TYPE_FLOAT) ? 1 : 2);
> >> + /* Align register number. */
> >> + if ((reg& 1)&& align == 2)
> >> + reg++;
> >> + while (reg + nregs<= 16)
> >> + {
> >> + int s, new_used = 0;
> >> + for (s = reg; s< reg + nregs; s++)
> >> + {
> >> + new_used |= (1<< s);
> >> + if (cif->vfp_used& (1<< s))
> >> + {
> >> + reg += align;
> >> + goto next_reg;
> >> + }
> >> + }
> >> + /* Found regs to allocate. */
> >> + cif->vfp_used |= new_used;
> >> + cif->vfp_args[cif->vfp_nargs++] = reg;
> >> +
> >> + /* Update vfp_reg_free. */
> >> + if (cif->vfp_used& (1<< cif->vfp_reg_free))
> >> + {
> >> + reg += nregs;
> >> + while (cif->vfp_used& (1<< reg))
> >> + reg += 1;
> >> + cif->vfp_reg_free = reg;
> >> + }
> >> + return;
> >> + next_reg: ;
> >> + }
> >> +}
> >> +
> >> +static void layout_vfp_args (ffi_cif *cif)
> >> +{
> >> + int i;
> >> + /* Init VFP fields */
> >> + cif->vfp_used = 0;
> >> + cif->vfp_nargs = 0;
> >> + cif->vfp_reg_free = 0;
> >> + memset (cif->vfp_args, -1, 16); /* Init to -1. */
> >> +
> >> + for (i = 0; i< cif->nargs; i++)
> >> + {
> >> + ffi_type *t = cif->arg_types[i];
> >> + if (vfp_type_p (t))
> >> + place_vfp_arg (cif, t);
> >> + }
> >> +}
> >> diff --git a/src/arm/ffitarget.h b/src/arm/ffitarget.h
> >> index eede79f..aceb10b 100644
> >> --- a/src/arm/ffitarget.h
> >> +++ b/src/arm/ffitarget.h
> >> @@ -34,11 +34,25 @@ typedef signed long ffi_sarg;
> >>   typedef enum ffi_abi {
> >>     FFI_FIRST_ABI = 0,
> >>     FFI_SYSV,
> >> + FFI_VFP,
> >>     FFI_LAST_ABI,
> >> - FFI_DEFAULT_ABI = FFI_SYSV
> >> +#ifdef __ARM_PCS_VFP
> >> + FFI_DEFAULT_ABI = FFI_VFP,
> >> +#else
> >> + FFI_DEFAULT_ABI = FFI_SYSV,
> >> +#endif
> >>   } ffi_abi;
> >>   #endif
> >>
> >> +#define FFI_EXTRA_CIF_FIELDS \
> >> + int vfp_used; \
> >> + short vfp_reg_free, vfp_nargs; \
> >> + signed char vfp_args[16] \
> >> +
> >> +/* Internally used. */
> >> +#define FFI_TYPE_STRUCT_VFP_FLOAT (FFI_TYPE_LAST + 1)
> >> +#define FFI_TYPE_STRUCT_VFP_DOUBLE (FFI_TYPE_LAST + 2)
> >> +
> >>   /* ---- Definitions for closures
> >>   ----------------------------------------- */
> >>
> >>   #define FFI_CLOSURES 1
> >> diff --git a/src/arm/sysv.S b/src/arm/sysv.S
> >> index 9064318..7bce727 100644
> >> --- a/src/arm/sysv.S
> >> +++ b/src/arm/sysv.S
> >> @@ -142,12 +142,11 @@ _L__\name:
> >>   .endm
> >>
> >>
> >> - @ r0: ffi_prep_args
> >> + @ r0: fn
> >>   	@ r1:&ecif
> >>   	@ r2: cif->bytes
> >>   	@ r3: fig->flags
> >>   	@ sp+0: ecif.rvalue
> >> - @ sp+4: fn
> >>
> >>   	@ This assumes we are using gas.
> >>   ARM_FUNC_START ffi_call_SYSV
> >> @@ -162,24 +161,23 @@ ARM_FUNC_START ffi_call_SYSV
> >>   	sub sp, fp, r2
> >>
> >>   	@ Place all of the ffi_prep_args in position
> >> - mov ip, r0
> >>   	mov r0, sp
> >>   	@ r1 already set
> >>
> >>   	@ Call ffi_prep_args(stack,&ecif)
> >> - call_reg(ip)
> >> + bl ffi_prep_args
> >>
> >>   	@ move first 4 parameters in registers
> >>   	ldmia sp, {r0-r3}
> >>
> >>   	@ and adjust stack
> >> - ldr ip, [fp, #8]
> >> - cmp ip, #16
> >> - movhs ip, #16
> >> - add sp, sp, ip
> >> + sub lr, fp, sp @ cif->bytes == fp - sp
> >> + ldr ip, [fp] @ load fn() in advance
> >> + cmp lr, #16
> >> + movhs lr, #16
> >> + add sp, sp, lr
> >>
> >>   	@ call (fn) (...)
> >> - ldr ip, [fp, #28]
> >>   	call_reg(ip)
> >>
> >>   	@ Remove the space we pushed for the args
> >> @@ -230,6 +228,101 @@ LSYM(Lepilogue):
> >>   	UNWIND .fnend
> >>           .size
> >>           CNAME(ffi_call_SYSV),.ffi_call_SYSV_end-CNAME(ffi_call_SYSV)
> >>
> >> +
> >> + @ r0: fn
> >> + @ r1:&ecif
> >> + @ r2: cif->bytes
> >> + @ r3: fig->flags
> >> + @ sp+0: ecif.rvalue
> >> +
> >> +ARM_FUNC_START ffi_call_VFP
> >> + @ Save registers
> >> + stmfd sp!, {r0-r3, fp, lr}
> >> + UNWIND .save {r0-r3, fp, lr}
> >> + mov fp, sp
> >> + UNWIND .setfp fp, sp
> >> +
> >> + @ Make room for all of the new args.
> >> + sub sp, sp, r2
> >> +
> >> + @ Make room for loading VFP args
> >> + sub sp, sp, #64
> >> +
> >> + @ Place all of the ffi_prep_args in position
> >> + mov r0, sp
> >> + @ r1 already set
> >> + sub r2, fp, #64 @ VFP scratch space
> >> +
> >> + @ Call ffi_prep_args(stack,&ecif, vfp_space)
> >> + bl ffi_prep_args
> >> +
> >> + @ Load VFP register args if needed
> >> + cmp r0, #0
> >> + beq LSYM(Lbase_args)
> >> +
> >> + @ Load only d0 if possible
> >> + cmp r0, #3
> >> + sub ip, fp, #64
> >> + flddle d0, [ip]
> >> + fldmiadgt ip, {d0-d7}
> >> +
> >> +LSYM(Lbase_args):
> >> + @ move first 4 parameters in registers
> >> + ldmia sp, {r0-r3}
> >> +
> >> + @ and adjust stack
> >> + sub lr, ip, sp @ cif->bytes == (fp - 64) - sp
> >> + ldr ip, [fp] @ load fn() in advance
> >> + cmp lr, #16
> >> + movhs lr, #16
> >> + add sp, sp, lr
> >> +
> >> + @ call (fn) (...)
> >> + call_reg(ip)
> >> +
> >> + @ Remove the space we pushed for the args
> >> + mov sp, fp
> >> +
> >> + @ Load r2 with the pointer to storage for
> >> + @ the return value
> >> + ldr r2, [sp, #24]
> >> +
> >> + @ Load r3 with the return type code
> >> + ldr r3, [sp, #12]
> >> +
> >> + @ If the return value pointer is NULL,
> >> + @ assume no return value.
> >> + cmp r2, #0
> >> + beq LSYM(Lepilogue_vfp)
> >> +
> >> + cmp r3, #FFI_TYPE_INT
> >> + streq r0, [r2]
> >> + beq LSYM(Lepilogue_vfp)
> >> +
> >> + cmp r3, #FFI_TYPE_SINT64
> >> + stmeqia r2, {r0, r1}
> >> + beq LSYM(Lepilogue_vfp)
> >> +
> >> + cmp r3, #FFI_TYPE_FLOAT
> >> + fstseq s0, [r2]
> >> + beq LSYM(Lepilogue_vfp)
> >> +
> >> + cmp r3, #FFI_TYPE_DOUBLE
> >> + fstdeq d0, [r2]
> >> + beq LSYM(Lepilogue_vfp)
> >> +
> >> + cmp r3, #FFI_TYPE_STRUCT_VFP_FLOAT
> >> + cmpne r3, #FFI_TYPE_STRUCT_VFP_DOUBLE
> >> + fstmiadeq r2, {d0-d3}
> >> +
> >> +LSYM(Lepilogue_vfp):
> >> + RETLDM "r0-r3,fp"
> >> +
> >> +.ffi_call_VFP_end:
> >> + UNWIND .fnend
> >> + .size CNAME(ffi_call_VFP),.ffi_call_VFP_end-CNAME(ffi_call_VFP)
> >> +
> >> +
> >>   /*
> >>   	unsigned int FFI_HIDDEN
> >>   	ffi_closure_SYSV_inner (closure, respp, args)
> >> @@ -302,6 +395,68 @@ ARM_FUNC_START ffi_closure_SYSV
> >>   	UNWIND .fnend
> >>           .size
> >>           CNAME(ffi_closure_SYSV),.ffi_closure_SYSV_end-CNAME(ffi_closure_SYSV)
> >>
> >> +
> >> +ARM_FUNC_START ffi_closure_VFP
> >> + fstmfdd sp!, {d0-d7}
> >> + @ r0-r3, then d0-d7
> >> + UNWIND .pad #80
> >> + add ip, sp, #80
> >> + stmfd sp!, {ip, lr}
> >> + UNWIND .save {r0, lr}
> >> + add r2, sp, #72
> >> + add r3, sp, #8
> >> + .pad #72
> >> + sub sp, sp, #72
> >> + str sp, [sp, #64]
> >> + add r1, sp, #64
> >> + bl ffi_closure_SYSV_inner
> >> +
> >> + cmp r0, #FFI_TYPE_INT
> >> + beq .Lretint_vfp
> >> +
> >> + cmp r0, #FFI_TYPE_FLOAT
> >> + beq .Lretfloat_vfp
> >> +
> >> + cmp r0, #FFI_TYPE_DOUBLE
> >> + cmpne r0, #FFI_TYPE_LONGDOUBLE
> >> + beq .Lretdouble_vfp
> >> +
> >> + cmp r0, #FFI_TYPE_SINT64
> >> + beq .Lretlonglong_vfp
> >> +
> >> + cmp r0, #FFI_TYPE_STRUCT_VFP_FLOAT
> >> + beq .Lretfloat_struct_vfp
> >> +
> >> + cmp r0, #FFI_TYPE_STRUCT_VFP_DOUBLE
> >> + beq .Lretdouble_struct_vfp
> >> +
> >> +.Lclosure_epilogue_vfp:
> >> + add sp, sp, #72
> >> + ldmfd sp, {sp, pc}
> >> +
> >> +.Lretfloat_vfp:
> >> + flds s0, [sp]
> >> + b .Lclosure_epilogue_vfp
> >> +.Lretdouble_vfp:
> >> + fldd d0, [sp]
> >> + b .Lclosure_epilogue_vfp
> >> +.Lretint_vfp:
> >> + ldr r0, [sp]
> >> + b .Lclosure_epilogue_vfp
> >> +.Lretlonglong_vfp:
> >> + ldmia sp, {r0, r1}
> >> + b .Lclosure_epilogue_vfp
> >> +.Lretfloat_struct_vfp:
> >> + fldmiad sp, {d0-d1}
> >> + b .Lclosure_epilogue_vfp
> >> +.Lretdouble_struct_vfp:
> >> + fldmiad sp, {d0-d3}
> >> + b .Lclosure_epilogue_vfp
> >> +
> >> +.ffi_closure_VFP_end:
> >> + UNWIND .fnend
> >> + .size
> >> CNAME(ffi_closure_VFP),.ffi_closure_VFP_end-CNAME(ffi_closure_VFP)
> >> +
> >>   #if defined __ELF__&& defined __linux__
> >>   	.section .note.GNU-stack,"",%progbits
> >>   #endif
> >> diff --git a/testsuite/lib/libffi-dg.exp
> >> b/testsuite/lib/libffi-dg.exp
> >> index 838a306..ca5f751 100644
> >> --- a/testsuite/lib/libffi-dg.exp
> >> +++ b/testsuite/lib/libffi-dg.exp
> >> @@ -266,6 +266,56 @@ proc dg-xfail-if { args } {
> >>       }
> >>   }
> >>
> >> +proc check-flags { args } {
> >> +
> >> + # The args are within another list; pull them out.
> >> + set args [lindex $args 0]
> >> +
> >> + # The next two arguments are optional. If they were not
> >> specified,
> >> + # use the defaults.
> >> + if { [llength $args] == 2 } {
> >> + lappend $args [list "*"]
> >> + }
> >> + if { [llength $args] == 3 } {
> >> + lappend $args [list ""]
> >> + }
> >> +
> >> + # If the option strings are the defaults, or the same as the
> >> + # defaults, there is no need to call check_conditional_xfail to
> >> + # compare them to the actual options.
> >> + if { [string compare [lindex $args 2] "*"] == 0
> >> + && [string compare [lindex $args 3] "" ] == 0 } {
> >> + set result 1
> >> + } else {
> >> + # The target list might be an effective-target keyword, so
> >> replace
> >> + # the original list with "*-*-*", since we already know it
> >> matches.
> >> + set result [check_conditional_xfail [lreplace $args 1 1 "*-*-*"]]
> >> + }
> >> +
> >> + return $result
> >> +}
> >> +
> >> +proc dg-skip-if { args } {
> >> + # Verify the number of arguments. The last two are optional.
> >> + set args [lreplace $args 0 0]
> >> + if { [llength $args]< 2 || [llength $args]> 4 } {
> >> + error "dg-skip-if 2: need 2, 3, or 4 arguments"
> >> + }
> >> +
> >> + # Don't bother if we're already skipping the test.
> >> + upvar dg-do-what dg-do-what
> >> + if { [lindex ${dg-do-what} 1] == "N" } {
> >> + return
> >> + }
> >> +
> >> + set selector [list target [lindex $args 1]]
> >> + if { [dg-process-target $selector] == "S" } {
> >> + if [check-flags $args] {
> >> + upvar dg-do-what dg-do-what
> >> + set dg-do-what [list [lindex ${dg-do-what} 0] "N" "P"]
> >> + }
> >> + }
> >> +}
> >>
> >>   # We need to make sure that additional_files and
> >>   additional_sources
> >>   # are both cleared out after every test. It is not enough to
> >>   clear
> >> diff --git a/testsuite/libffi.call/cls_double_va.c
> >> b/testsuite/libffi.call/cls_double_va.c
> >> index 0695874..62bebbd 100644
> >> --- a/testsuite/libffi.call/cls_double_va.c
> >> +++ b/testsuite/libffi.call/cls_double_va.c
> >> @@ -6,6 +6,8 @@
> >>
> >>   /* { dg-do run { xfail strongarm*-*-* xscale*-*-* } } */
> >>   /* { dg-output "" { xfail avr32*-*-* } } */
> >> +/* { dg-skip-if "" arm*-*-* { "-mfloat-abi=hard" } { "" } } */
> >> +
> >>   #include "ffitest.h"
> >>
> >>   static void
> >> diff --git a/testsuite/libffi.call/cls_longdouble_va.c
> >> b/testsuite/libffi.call/cls_longdouble_va.c
> >> index 38564cb..b33b2b7 100644
> >> --- a/testsuite/libffi.call/cls_longdouble_va.c
> >> +++ b/testsuite/libffi.call/cls_longdouble_va.c
> >> @@ -6,6 +6,8 @@
> >>
> >>   /* { dg-do run { xfail strongarm*-*-* xscale*-*-* } } */
> >>   /* { dg-output "" { xfail avr32*-*-* x86_64-*-mingw* } } */
> >> +/* { dg-skip-if "" arm*-*-* { "-mfloat-abi=hard" } { "" } } */
> >> +
> >>   #include "ffitest.h"
> >>
> >>   static void