Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 22 Mar 2020 15:10:19 -0400
From: Rich Felker <dalias@...c.org>
To: musl@...ts.openwall.com
Subject: Re: Q: dealing with missing removal of excess precision

On Sun, Mar 22, 2020 at 09:51:09PM +0300, Alexander Monakov wrote:
> On Sun, 22 Mar 2020, Rich Felker wrote:
> 
> > > Nack for sqrt 'unsigned short' fix, I recommend to consider
> > > 'unsigned fpsr = 0' and "+a" constraint, the resulting assembly is
> > > much better (GCC doesn't seem to do a good job optimizing the 'unsigned short'
> > > variant at all).
> > 
> > Sorry, I forgot to disassemble and check after making that change, and
> > indeed it is very bad.
> > 
> > How about just leaving it as-is? The value is masked such that upper
> > bits don't matter, and "whatever the register happened to contain
> > before" is a valid albeit ugly output from inline asm -- it doesn't
> > admit any compiler transformations that would cause inconsistent value
> > to be observed or other problems.
> 
> Yes, I guess it's acceptable.
> 
> > > Actually, may I ask why the initial commit did not mention that it relies
> > > on this nontrivial property?
> > 
> > The need for fixup of double was only realized later, in commit
> > 809556e60a3359f646946879dd94c4760e5b8e84. It was discussed at the time
> > that no action was needed for single, but it seems since there was no
> > change there wasn't any mention of it in log.
> 
> Are you sure? single-precision sqrtf received a change just two days
> prior to that in commit e0a54e6725 ("correct rounding for i387 sqrtf function")
> which is the same day as when Stephen Canon supplied his answer in
> https://stackoverflow.com/questions/9678224/is-there-any-way-to-get-correct-rounding-with-the-i387-fsqrt-instruction
> (and also the same day you asked the question)

I just dug through the old gcc and glibc bz's it was mentioned on
(52593 and 14032 resp.) and didn't find anything, but I'm pretty sure
it was known in 2012 that it was a non-issue for single-precision.
What I didn't understand then was that the callee was responsible for
dropping the excess precision; I wrongly believed that "something" in
the caller would make it not-matter/get collapsed down to nominal
precision and rounded correctly. Commit e0a54e6725 and related commits
were fixing that mistake, which I'd since recognized was wrong but
hadn't yet recognized the urgency of fixing until I started seeing how
bad it could break the compiler.

Rich

Powered by blists - more mailing lists

Your e-mail address:

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.