Re: Use floor functions not __floor functions in glibc libm

[Joseph Myers <joseph at codesourcery dot com>]

On Fri, 14 Sep 2018, Wilco Dijkstra wrote:

> Looks great, thanks for doing this! The more general mechanism means
> it should be much easier to do this for the remaining functions. Yes it
> sounds like a good idea to do this for copysign too.

One thing to consider is whether we actually need this mechanism for these 
functions after all, or whether we can use libm_hidden_proto / 
libm_hidden_def instead.

Using this mechanism actually results in smaller and simpler patches (even 
if it looks like a lot of floor etc. implementations are getting changed), 
because only C implementations of the functions need to get 
NO_MATH_REDIRECT added, whereas .S implementations would also need to get 
libm_hidden_def added (and handling libm_hidden_* may also be more 
complicated in IFUNC cases than simply disabling the redirection with 
NO_MATH_REDIRECT is).  For sqrt, the special mechanism is needed because, 
for static linking (when libm_hidden_proto / libm_hidden_def does 
nothing), we still want non-inlined calls to go to __ieee754_sqrt, not 
through the wrapper that sets errno.  But for the other functions, there 
are no wrappers and no reason to add such wrappers (although in principle 
if we do something like this for lrint or fma, while those don't have 
wrappers at present there are also open bugs for missing errno setting for 
those functions).

If you use libm_hidden_*, you need to consider namespace issues for static 
linking.  Those certainly don't apply for floor or ceil because those are 
C90 functions.  For rint / round / trunc you'd need to be more careful 
about any uses from C90 functions (and the possibility that someone might 
want to add such a use in future and run into namespace difficulties, even 
if there aren't such uses at present).  (In theory that applies for 
copysign as well and there are even __copysign uses from C90 functions at 
present in glibc - but copysign functions are in fact always inlined 
except for some IBM long double cases, and long double libm functions were 
only added in C99, so there wouldn't actually be namespace issues from 
using libm_hidden_* for copysign functions (although also being in libc 
complicates things there).)

-- 
Joseph S. Myers
joseph@codesourcery.com