[RFC] How to add vector math functions to Glibc

[Joseph S. Myers]

On Tue, 21 Oct 2014, Andrew Senkevich wrote:

> > There shouldn't be such conditionals.  It should be arranged that if
> > there isn't a relevant vector version of a particular function, running
> > vector tests for that function does nothing - so there are no conditionals
> > on which *_test functions to run, and none inside those functions, just
> > conditionals affecting what the test macros do (by means of conditionals
> > inside them such as if (HAVE_VECTOR_cos_double_vlen4), for example,
> > resulting from appropriate concatenations).
> 
> With HAVE_VECTOR_cos_double_vlen4 we need to have such macros with
> zero for all set of not vector functions which is huge.

But I'd hope such macros could be generated by gen-libm-test.pl (or some 
such script, anyway) rather than needing lots of repetitive definitions to 
be maintained by hand and checked in.

Essentially:

* The architecture-specific headers (installed headers, or possibly 
non-installed ones used only by the testsuite in some cases) contain the 
information about what vector versions of what functions are available.  
Things are designed so that they only need to contain definitions where 
vector functions are available, not where they aren't (to avoid needing to 
repeat slightly different huge lists for each architecture).

* Where a default definition to 0 is needed in any cases, the relevant 
definitions are generated automatically.  (Indeed, this might make sense 
for a header included by bits/mathcalls.h, so that __MATHCALL can expand 
to include the right __DECL_SIMD_*, which might end up empty, rather than 
needing lots of #if conditionals before every function declaration there.)

(Incidentally, there have been so many different patch fragments posted in 
this discussion that it's hard to follow what you're proposing, if e.g. in 
the discussion of testing it's relevant to look at what you're proposing 
for installed headers.  I think it would help if you had a git branch with 
the current set of proposed changes, that you frequently rebase so it 
always shows what you currently propose.)

> May be more suitable way is to have determined name of vector function
> wrapper and selection based on function name?
> I mean to have something like this in test driver test-double-vlen4.c:
> 
> #define HAVE_VECTOR 1
> #define VEC_PREFIX_STR "VECTOR_LEN_"
> #define cos VECTOR_LEN_4_cos
> #include "libm-test.c"
> 
> in test-double-vlen4-wrapper.c:
> 
> VECTOR_WRAPPER(VECTOR_LEN_4_cos,_ZGVdN4v_cos)

If you did something like that, I think it would still be desirable to 
have some form of automatic generation of a list of defines, one per 
function and conditional as needed on whether the relevant vector version 
of that function exists.

-- 
Joseph S. Myers
joseph@codesourcery.com