[PATCH] Use the 'arcp' fast-math-flag when combining repeated FP divisors

[Akira Hatanaka]

On Tue, May 12, 2015 at 1:04 PM, Sanjay Patel <spatel at rotateright.com>
wrote:

> In http://reviews.llvm.org/D9708#171355, @ahatanak wrote:
>
> > To preserve backward compatibility, I think we should make sure
> functions using fast math attributes (e.g., "unsafe-fp-math" in
> test/CodeGen/PowerPC/fdiv-combine.ll) still get optimized after we migrate
> to modeling those flags at the node-level (using autoupgrade?).
>
>
> Thanks, Akira. If I understand correctly, I should create this autoupgrade
> patch first then. That would allow us to remove the Options.UnsafeFPMath
> checks as we update all of the places in the backend to use the node-level
> flags.
>
>
The autoupgrade patch will be needed when UnsafeFPMath is removed, but I'm
not sure removing it completely should be a prerequisite to committing this
patch. If it turns out to be not trivial to remove it completely, it might
be acceptable to do it later.


> Is a separate mechanism needed to handle the llc command-line flag
> "-enable-unsafe-fp-math"?
>
>
If we decide to keep the command line option, yes, we have to fix the
handling of the flag.


>
> http://reviews.llvm.org/D9708
>
> EMAIL PREFERENCES
>   http://reviews.llvm.org/settings/panel/emailpreferences/
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150512/a6886e37/attachment.html>