This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] x86-64: Add sinf with FMA

On Mon, Dec 4, 2017 at 12:59 PM, Joseph Myers <> wrote:
> On Mon, 4 Dec 2017, H.J. Lu wrote:
>> Is
>> (*__errno_location ()) = xxx
> If anything I'd expect a direct TLS initial-exec access to errno to be
> faster.

I will update x86-64 s_sinf.S to use errno.

>> faster?  On x86-64, there should be no call to __floor.
> The x86_64 __floor inline in math_private.h is only when compiling glibc
> for SSE4.1 or later.
> The case of inlining floor / __floor and related functions for x86_64
> without SSE4.1 is tricky.  Supposing we had appropriate asm redirects to
> allow libm to call floor / ceil / trunc etc. directly so the compiler
> could inline them but __* are still called if not inlined, the default
> SSE2 inlines would come into play.  But those inlines are slower on SSE4.1
> hardware than an out-of-line call to the floor / ceil / trunc IFUNC, so if
> you're building a generic SSE2 glibc that may well be used on SSE4.1
> hardware, you may wish either to avoid those inlines or, if there is a
> significant performance difference in benchmarks, have an SSE4.1 IFUNC of
> the calling function using floor (or __floor, with the present inline).
> The expf etc. set of optimized float functions have several different
> choices of how conversions to integer are handled, which may be configured
> by an architecture.  That may make sense in other cases as well.

x86-64 s_sinf.S avoids floor () with cvttsd2si.  I don't think it can be used
with generic sinf.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]