byteswap.h

byteswap.h

[fab10]

I had a look at byteswap.h and it seems to me that the code is not very 
efficient. Every function in this header could be translated in a single 
assembly instruction with the gcc intrinsics:

__builtin_bswap16
__builtin_bswap32
__builtin_bswap64

Is there a reason to not use these gcc intrinsics?

Bye

Re: byteswap.h

[Jon Chesterfield]

[-- Attachment #1: Type: text/plain, Size: 529 bytes --]

Have you tested whether the intrinsics improve codegen on a recent
compiler? There's a good chance they make the code non-portable while
generating the same asm.

Jon


On 7 Apr 2017 18:53, "fab10" <0xfab10@gmail.com> wrote:

I had a look at byteswap.h and it seems to me that the code is not very
efficient. Every function in this header could be translated in a single
assembly instruction with the gcc intrinsics:

__builtin_bswap16
__builtin_bswap32
__builtin_bswap64

Is there a reason to not use these gcc intrinsics?

Bye

[-- Attachment #2: Type: text/html, Size: 923 bytes --]

Re: byteswap.h

[Rich Felker]

On Fri, Apr 07, 2017 at 07:53:09PM +0200, fab10 wrote:
> I had a look at byteswap.h and it seems to me that the code is not
> very efficient. Every function in this header could be translated in
> a single assembly instruction with the gcc intrinsics:
> 
> __builtin_bswap16
> __builtin_bswap32
> __builtin_bswap64
> 
> Is there a reason to not use these gcc intrinsics?

Because the compile should be (and as far as I know, is) generating
the optimal single instructions from the portable C, without the need
for hackish non-portable intrinsics.

Rich

Re: byteswap.h

[Joakim Sindholt]

On Fri, Apr 07, 2017 at 07:53:09PM +0200, fab10 wrote:
> I had a look at byteswap.h and it seems to me that the code is not very 
> efficient. Every function in this header could be translated in a single 
> assembly instruction with the gcc intrinsics:
> 
> __builtin_bswap16
> __builtin_bswap32
> __builtin_bswap64
> 
> Is there a reason to not use these gcc intrinsics?
> 
> Bye
> 

While musl does employ GNU C features in quite a few places it's just
plain unnecessary here.

https://godbolt.org/g/eLZWwI

When optimized those functions yield a single bswap instruction anyway,
because GCC is smart enough to deduce this.

Re: byteswap.h

[Szabolcs Nagy]

* Joakim Sindholt <opensource@zhasha.com> [2017-04-07 20:07:42 +0200]:
> On Fri, Apr 07, 2017 at 07:53:09PM +0200, fab10 wrote:
> > I had a look at byteswap.h and it seems to me that the code is not very 
> > efficient. Every function in this header could be translated in a single 
> > assembly instruction with the gcc intrinsics:
> > 
> > __builtin_bswap16
> > __builtin_bswap32
> > __builtin_bswap64
> > 
> > Is there a reason to not use these gcc intrinsics?
> > 
> > Bye
> > 
> 
> While musl does employ GNU C features in quite a few places it's just
> plain unnecessary here.
> 
rarely in public headers though which have to
be portable to all c source parser tools.
builtins  would need to be conditional on __GNUC__
and we still need the portable code as a fallback
which is harder to maintain and test

> https://godbolt.org/g/eLZWwI
> 
> When optimized those functions yield a single bswap instruction anyway,
> because GCC is smart enough to deduce this.