Re: [PING][Patch] Add support for IEEE-conformant versions of scalar fmin* and fmax*

[Joseph Myers <joseph@codesourcery.com>]

On Wed, 19 Aug 2015, Richard Biener wrote:

> As an additional point for many math functions we have to support errno
> which means, like, BUILT_IN_SQRT can be rewritten to SQRT_EXPR
> only if -fno-math-errno is in effect.  But then code has to handle

I'd say that for functions like that (which can be expanded inline only 
for -fno-math-errno) there should be no-errno built-in function variants 
that users can call even if -fmath-errno (if not expanded inline, they'd 
still result in a call to a libm function that might set errno).

An example of a use for that is AArch64 sqrt intrinsics that need an 
architecture-specific built-in __builtin_aarch64_sqrtdf when 
__builtin_sqrt_noerrno would do just as well if it existed.  As another 
example: various libm functions are marked in builtins.def as not setting 
errno, even though their proper semantics mean they might set errno; see 
bug 64101 for the example of erf.  One such function is fma.  But if you 
limit fma inline expansion (for calls to fma / __builtin_fma in the user's 
program; obviously this doesn't affect expansion via contraction of a * b 
+ c) to allow for the possibility of errno setting, you definitely want a 
way for user programs to get back the efficient inline expansion if they 
don't need errno set; for example, glibc uses __builtin_fma in various 
cases if _FP_FAST_FMA, and does not need errno setting in those cases, so 
would want to use __builtin_fma_noerrno in the event of any such change.

-- 
Joseph S. Myers
joseph@codesourcery.com