linux-kernel - Re: [PATCH next] tools/nolibc: Optimise and common up number to ascii functions

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <20260207222329.38f7c6d6@pumpkin>
Date: Sat, 7 Feb 2026 22:23:29 +0000
From: David Laight <david.laight.linux@...il.com>
To: Willy Tarreau <w@....eu>
Cc: Thomas Weißschuh <linux@...ssschuh.net>,
 linux-kernel@...r.kernel.org, Cheng Li <lechain@...il.com>
Subject: Re: [PATCH next] tools/nolibc: Optimise and common up number to
 ascii functions

On Sat, 7 Feb 2026 16:06:51 +0100
Willy Tarreau <w@....eu> wrote:

> Hi David,
> 
> On Tue, Feb 03, 2026 at 03:13:15PM +0000, david.laight.linux@...il.com wrote:
> > From: David Laight <david.laight.linux@...il.com>
> > 
> > Implement u[64]to[ah]_r() using a common function that uses multiply
> > by reciprocal to generate the least significant digit first and then
> > reverses the string.
> > 
> > On 32bit this is five multiplies (with 64bit product) for each output
> > digit. I think the old utoa_r() always did 36 multiplies and a lot
> > of subtracts - so this is likely faster even for 32bit values.
> > Definitely better for 64bit values (especially small ones).
> > 
> > Clearly shifts are faster for base 16, but reversing the output buffer
> > makes a big difference.
> > 
> > Sharing the code reduces the footprint (unless gcc decides to constant
> > fold the functions).
> > Definitely helps vfprintf() where the constants get loaded and a single
> > call is down.
> > Also makes it cheap to add octal support to vfprintf for completeness.
> > 
> > Signed-off-by: David Laight <david.laight.linux@...il.com>  
> 
> OK, I had a long series of tests on it, including with older compilers
> going back to gcc-4.7 and on various archs. Except for code that would
> previously only use utoh(), the new code is slightly smaller in the vast
> majority of cases. And this combined with the added flexibility looks
> like a good addition. The code is not trivial (as every time we're
> dealing with number representation) but it's well documented, so I'm
> personally fine with the change.
> 
> I'm just having a few comments below:
> 
> > -static __attribute__((unused))
> > -int utoh_r(unsigned long in, char *buffer)
> > +#define __U64TOA_RECIP(base) ((base) & 1 ? ~0ull / (base) : (1ull << 63) / ((base) / 2))  
> 
> Please rename this macro to have _NOBLIC_ as a prefix.

Not hard :-)

> > +#if defined(__SIZEOF_INT128__) && !defined(__mips__)  
> 
> Out of curiosity, why !mips ? I tried with -mabi=64 and the function size
> dropped from 0x120 to 0xc0 (lost 1/3 of its size).

I think it is mips, some of the older versions of gcc emit a library call
even though the cpu has the required instruction.
Cropped up in the mul_u64_u64_div_u64 code.
I could look up the versions and add a comment.

> 
> > +		q = ((unsigned __int128)in * recip) >> 64;
> > +#else  
> (...)
> 
> Once the macro is renamed, feel free to add:
> 
> Acked-by: Willy Tarreau <w@....eu>

Thanks.

	David

> 
> Thanks!
> Willy