[llvm] r257400 - [LibCallSimplifier] use instruction-level fast-math-flags to transform sqrt calls

[Sanjay Patel via llvm-commits]

Thanks, Joerg.

Looks like the sqrt path is missing a check that the other mathlib
optimizations have: make sure the parameter and return type are FP. The
change to use instruction-level FMF exposed the existing bug.

I'm working on a fix now.


On Thu, Jan 14, 2016 at 8:48 AM, Joerg Sonnenberger <joerg at britannica.bec.de
> wrote:

> On Mon, Jan 11, 2016 at 10:34:20PM -0000, Sanjay Patel via llvm-commits
> wrote:
> > Author: spatel
> > Date: Mon Jan 11 16:34:19 2016
> > New Revision: 257400
> >
> > URL: http://llvm.org/viewvc/llvm-project?rev=257400&view=rev
> > Log:
> > [LibCallSimplifier] use instruction-level fast-math-flags to transform
> sqrt calls
>
> This crashes witht -O2 on typical configure tests like:
>
>   char sqrt ();
>   int main () { return sqrt (); ; return 0; }
>
> A definite show stopper for 3.8...
>
> Joerg
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160114/92691d0d/attachment.html>