Unexpected srfi-4 C Interface Change in 2.0.10

Unexpected srfi-4 C Interface Change in 2.0.10

[Daniel Llorens]

> Message: 2
> Date: Wed, 19 Mar 2014 10:22:40 -0400
> From: Barry Fishman <barry_fishman@acm.org>
> To: guile-devel@gnu.org
> Subject: Unexpected srfi-4 C Interface Change in 2.0.10
> Message-ID: <m37g7qxo7j.fsf@barry_fishman.acm.org>
> Content-Type: text/plain
> 
> When building C interface code using scm_f32vector_writable_elements()
> in 2.0.10 and on the trunk I found:
> 
> For a vector generated in guile with:
> 
>  (f32vector 0.25 0.25 0.25 1.0)
> 
> I found in the C interface:
> 
>  scm_t_array_handle handle;
>  size_t vlen, vinc;
>  float *arrayp;
> 
>  arrayp = scm_f32vector_writable_elements(s_uvec, &handle,
>                                           &vlen, &vinc)
> 
> When I test vlen it now seems to contain the number of bytes (16)
> rather than the number of elements (4) in the f32vector.
> 
> It was 4 in 2.0.9 and the guile trunk I built on Jan 8.
> It was 16 in 2.0.10 and the guile trunk I built on March 17.
> I assume it relates to the Feb 8 changes, where:
> (in commit dc65b88d839c326889618112c4870ad3a64e9446)
> 
>    *lenp = scm_c_bytevector_length (uvec) / width;
> 
> and width is set to 1 in the macro:
> 
> DEFINE_SRFI_4_C_FUNCS (F32, f32, float, 1);

This is a bug. Is there a test setup for the C interface?

The last argument of DEFINE_SRFI_4_C_FUNCS (1) is correct; that can be 1 or 2, because complex types use a pointer to real. We could use C99 complex types and get rid of it. Just cast the pointer for C<99 users (?)

The bug appeared in replacing the deprecated scm_uniform_vector_elements() with the bytevector functions. Generally lengths are computed with scm_i_array_element_type_sizes[] in uniform.c (e.g. see make_bytevector() in bytevector.c).

So the patch would be

   *lenp = scm_c_bytevector_length (uvec) / (width * scm_i_array_element_type_sizes[ETYPE (TAG)]) ;

Actually I'd do

   *lenp = scm_c_bytevector_length (uvec) / (width * sizeof(ctype))

(Both untested).

Regards,

	Daniel

Re: Unexpected srfi-4 C Interface Change in 2.0.10

[Mark H Weaver]

Daniel Llorens <daniel.llorens@bluewin.ch> writes:

>> Message: 2
>> Date: Wed, 19 Mar 2014 10:22:40 -0400
>> From: Barry Fishman <barry_fishman@acm.org>
>> To: guile-devel@gnu.org
>> Subject: Unexpected srfi-4 C Interface Change in 2.0.10
>> Message-ID: <m37g7qxo7j.fsf@barry_fishman.acm.org>
>> Content-Type: text/plain
>> 
>> When building C interface code using scm_f32vector_writable_elements()
>> in 2.0.10 and on the trunk I found:
>> 
>> For a vector generated in guile with:
>> 
>>  (f32vector 0.25 0.25 0.25 1.0)
>> 
>> I found in the C interface:
>> 
>>  scm_t_array_handle handle;
>>  size_t vlen, vinc;
>>  float *arrayp;
>> 
>>  arrayp = scm_f32vector_writable_elements(s_uvec, &handle,
>>                                           &vlen, &vinc)
>> 
>> When I test vlen it now seems to contain the number of bytes (16)
>> rather than the number of elements (4) in the f32vector.
>> 
>> It was 4 in 2.0.9 and the guile trunk I built on Jan 8.
>> It was 16 in 2.0.10 and the guile trunk I built on March 17.
>> I assume it relates to the Feb 8 changes, where:
>> (in commit dc65b88d839c326889618112c4870ad3a64e9446)
>> 
>>    *lenp = scm_c_bytevector_length (uvec) / width;
>> 
>> and width is set to 1 in the macro:
>> 
>> DEFINE_SRFI_4_C_FUNCS (F32, f32, float, 1);
>
> This is a bug. Is there a test setup for the C interface?
>
> The last argument of DEFINE_SRFI_4_C_FUNCS (1) is correct; that can be 1 or 2, because complex types use a pointer to real. We could use C99 complex types and get rid of it. Just cast the pointer for C<99 users (?)
>
> The bug appeared in replacing the deprecated scm_uniform_vector_elements() with the bytevector functions. Generally lengths are computed with scm_i_array_element_type_sizes[] in uniform.c (e.g. see make_bytevector() in bytevector.c).
>
> So the patch would be
>
>    *lenp = scm_c_bytevector_length (uvec) / (width * scm_i_array_element_type_sizes[ETYPE (TAG)]) ;
>
> Actually I'd do
>
>    *lenp = scm_c_bytevector_length (uvec) / (width * sizeof(ctype))

This last one looks good to me.  The length check 4 lines above this
also needs to be fixed:

    if (!scm_is_bytevector (uvec)                                       \
        || (scm_c_bytevector_length (uvec) % (width * sizeof(ctype))))  \
      scm_wrong_type_arg_msg (NULL, 0, uvec, #tag "vector");            \

I guess it's time to release 2.0.11 :-/

     Mark

Re: Unexpected srfi-4 C Interface Change in 2.0.10

[Andy Wingo]

Hi,

I'm very sorry about this breakage to the SRFI-4 C interface.  I've
pushed a fix (2be7131ee0c38336483226657872a8faa62a2562) and a test case
and we'll roll another release shortly; perhaps tomorrow.  Apologies
again.

Regards,

Andy
-- 
http://wingolog.org/

Unexpected srfi-4 C Interface Change in 2.0.10

[Barry Fishman]

When building C interface code using scm_f32vector_writable_elements()
in 2.0.10 and on the trunk I found:

For a vector generated in guile with:

  (f32vector 0.25 0.25 0.25 1.0)

I found in the C interface:

  scm_t_array_handle handle;
  size_t vlen, vinc;
  float *arrayp;
  
  arrayp = scm_f32vector_writable_elements(s_uvec, &handle,
                                           &vlen, &vinc)

When I test vlen it now seems to contain the number of bytes (16)
rather than the number of elements (4) in the f32vector.

It was 4 in 2.0.9 and the guile trunk I built on Jan 8.
It was 16 in 2.0.10 and the guile trunk I built on March 17.
I assume it relates to the Feb 8 changes, where:
 (in commit dc65b88d839c326889618112c4870ad3a64e9446)

    *lenp = scm_c_bytevector_length (uvec) / width;

and width is set to 1 in the macro:

DEFINE_SRFI_4_C_FUNCS (F32, f32, float, 1);

--
Barry Fishman