Re: Remove powerpc32 sqrt wrappers

[Joseph Myers <joseph at codesourcery dot com>]

On Thu, 24 Aug 2017, Gabriel F. T. Gomes wrote:

> On Thu, 24 Aug 2017 14:23:36 +0000
> Joseph Myers <joseph@codesourcery.com> wrote:
> 
> >Would any powerpc people like to comment on this patch 
> ><https://sourceware.org/ml/libc-alpha/2017-08/msg01028.html>?
> 
> I had never ran benchtests of functions from math (only from string)
> before you sent this patch (so I'm spending some time learning about its
> outputs). I would like to comment on it in the following days.

Note that benchtests of sqrt deliberately use -fno-builtin, specifically 
because the point is to benchmark glibc's implementations rather than the 
compiler's inline code.

Thus those benchtests may indicate how much performance (if any) is gained 
(given -fno-builtin) from the custom sqrt wrappers, compared with code 
generated by a modern GCC version that can do shrink wrapping (the 
comments in the powerpc32 wrappers say "Until gcc supports prologue 
shrink-wrapping this is the best we can do.", i.e. they were intended to 
improve on the performance of code built with old GCC).  But they are also 
fairly peripheral to my main point (that in the normal case, without 
-fno-builtin, these wrappers will only ever be called for 
non-performance-relevant error cases, and not even those with 
-fno-math-errno).

-- 
Joseph S. Myers
joseph@codesourcery.com