RegAllocGreedy behaves host-dependent (-mfpmath on i686)

Chandler Carruth chandlerc at google.com
Mon Feb 3 04:49:03 PST 2014


We should simply remove the use of floating point math from LLVM for
anything other than statistics. It has been a *very* long standing problem.


On Mon, Feb 3, 2014 at 4:44 AM, NAKAMURA Takumi <geek4civic at gmail.com>wrote:

> Jakob,
>
> On i686 host, I saw different output between llc compiled by
>  gcc-4.7 (-mfpmath=387 by default) and
>  llc compiled by clang (-mfpmath=sse, I guess).
> It seems "clang -mno-sse" can suppress this issue.
>
> Although I am still investigating, I could suggest (float)Hysteresis
> as more simple number
> to suppress this issue.
>
> I suggest (64225 / 65536.0f) or (2007 / 2048.0f). The latter is more
> simple :)
>
> This issue affects i686 selfhosting, for example;
> http://bb.pgr.jp/builders/clang-3stage-i686-cygwin
>
> Excuse me to bother you, thanks.
>
> ...Takumi
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140203/399392bf/attachment.html>


More information about the llvm-commits mailing list