Date: Wed, 12 Sep 2012 19:08:33 -0700 From: Steve Kargl <sgk@troutmask.apl.washington.edu> To: Nathan Whitehorn <nwhitehorn@freebsd.org> Cc: toolchain@freebsd.org, Tijl Coosemans <tijl@coosemans.org>, Dimitry Andric <dim@freebsd.org>, current@freebsd.org Subject: Re: Clang as default compiler November 4th Message-ID: <20120913020833.GA8255@troutmask.apl.washington.edu> In-Reply-To: <505101C3.70203@freebsd.org> References: <20120910211207.GC64920@lor.one-eyed-alien.net> <20120911104518.GF37286@deviant.kiev.zoral.com.ua> <20120911120649.GA52235@freebsd.org> <20120911132410.GA87126@troutmask.apl.washington.edu> <504F4645.4070900@FreeBSD.org> <504F4A6B.4010001@coosemans.org> <504F5101.8090906@FreeBSD.org> <505101C3.70203@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, Sep 12, 2012 at 04:42:27PM -0500, Nathan Whitehorn wrote: > On 09/11/12 09:56, Dimitry Andric wrote: > >On 2012-09-11 16:27, Tijl Coosemans wrote:> On 11-09-2012 16:10, > >Dimitry Andric wrote: > >... > >>>Yes, maths support, specifically precision, is admittedly still one of > >>>clang's (really llvm's) weaker points. It is currently not really a > >>>high priority item for upstream. > >>> > >>>This is obviously something that a certain part of our userbase will > >>>care a lot about, while most of the time they won't care so much about > >>>licensing or politics. So those people are probably better off using > >>>gcc for the time being. > >> > >>Does it affect the accuracy of libm functions? > > > >It seems to, at least in specific cases; Steve posted about this in an > >earlier thread on -current: > > > > http://docs.freebsd.org/cgi/mid.cgi?20120905221310.GA97847 > >_______________________________________________ > > If true, this is a serious problem, especially for those of us who use > FreeBSD in a scientific computing environment. Just to clarify. I do not oppose switching the default compiler to clang as long as the proponents for the switch have shown adequate testing. Neither clang successfully building world nor clang building a working kernel are adequate testing (IMHO). Neither of those "benchmarks" use floating point, and AFAIK the libm built by clang during a buildworld is not (extensively?) exercised. As far as the URL above, I've been fixing accuracy issues in the j0f() function, and so, I have a program that allows me to exhaustively test all possible input values in the range reported. For my locally patched j0f(), I saw the issue as reported in the URL, but in doing additional development on j0f() I accidentally deletely/lost that specific version of the code. I hope to regenerate the code from my notes this weekend, and redo the tests. In regards to my initial post in this thread, I was just trying to assess whether any benchmarks have been performed on FreeBSD for floating point generated by clang. Other than the limited testing that I've done, it appears that the answer is 'no'. -- Steve
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20120913020833.GA8255>