[PATCH] D42323: [DAGCombiner] filter out denorm inputs when calculating sqrt estimate (PR34994)

[Sanjay Patel via llvm-commits]

Thanks for checking that, Evandro. I see what's happening - on AArch64, we
don't generate the generic N-R refinement because we have 'frsqrts'. So the
'Iterations' value that is guarding this codegen in generic DAGCombiner is
'0'. AArch64 will need to duplicate more code, or (better) we should move
the !Reciprocal portion of this sequence out where all targets can use it,
and we'll kill the existing duplicated code from
AArch64TargetLowering::getSqrtEstimate().

Let me see if the latter can be done cleanly.

On Wed, Jan 31, 2018 at 4:09 PM, Evandro Menezes <e.menezes at samsung.com>
wrote:

> Hi, Sanjay.
>
> The patch LGTM, but, when trying on AArch64, I didn't see any difference
> in the AArch64 test case in attach.  I'm confused...
>
> --
> Evandro Menezes
>
> On 01/26/2018 12:01 PM, Sanjay Patel via Phabricator via llvm-commits
> wrote:
>
> spatel added a reviewer: bkramer.
> spatel added a comment.
>
> Ping.
>
> Note that on x86 Linux, specifying -fdenormal-fp-math=ieee doesn't inhibit linking crtfastmath.o ( https://reviews.llvm.org/rL165240 / https://bugs.llvm.org/show_bug.cgi?id=14024 ), but I think that's a separate issue. (No matter what we do here in codegen, denormals will be off on that platform.)
>
> https://reviews.llvm.org/D42323
>
>
>
>
> _______________________________________________
> llvm-commits mailing listllvm-commits at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180131/e87d8aed/attachment.html>