llvm dev - Sep 2015 - fast-math-flags propagated to DAG nodes from IR instructions

Hi backend maintainers/hackers,

With r247815:
http://llvm.org/viewvc/llvm-project?view=revision&revision=247815

...we're propagating fast-math-flags (FMF) from IR to DAG nodes by default.
This has been reverted before...
But it's been one week since the commit and I haven't heard any
complaints
yet, so I'm hoping it sticks this time.

If you have target-specific DAG combines on FP nodes, you probably want to
propagate the optimization flags from incoming nodes to nodes that you are
creating. Otherwise, you may fail to get CSE optimizations that happened
before this change. See the test case in the patch for an example of how
that might manifest.

The code change when creating a node that should have FMF looks something
like this:

-      Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
+      Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0),
Tmp1, Node->getFlags());


FMF only applies to binary FP ops at the moment, but as noted in the commit
message, the plan is to extend the flags to more DAG node types.

Going forward, we can use the node-level flags to trigger optimizations
that the global target settings (eg, DAG.getTarget().Options.UnsafeFPMath)
might not allow. And some day, we may be able to remove global FP
optimization flags entirely.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150923/62eca1c4/attachment.html>

I confirmed that with a patch in our backend (an extremely FP-heavy GPU backend
with tons of combines) to propagate fast-math flags on our own target DAG
combines, the fast-math change Sanjay committed is NFC. Hopefully anyone else
having to do the same sort of thing with their backend has the same result!

—escha
> On Sep 23, 2015, at 9:48 AM, Sanjay Patel via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> Hi backend maintainers/hackers,
> 
> With r247815: 
> http://llvm.org/viewvc/llvm-project?view=revision&revision=247815
<http://llvm.org/viewvc/llvm-project?view=revision&revision=247815>
> 
> ...we're propagating fast-math-flags (FMF) from IR to DAG nodes by
default. This has been reverted before...
> But it's been one week since the commit and I haven't heard any
complaints yet, so I'm hoping it sticks this time.
> 
> If you have target-specific DAG combines on FP nodes, you probably want to
propagate the optimization flags from incoming nodes to nodes that you are
creating. Otherwise, you may fail to get CSE optimizations that happened before
this change. See the test case in the patch for an example of how that might
manifest.
> 
> The code change when creating a node that should have FMF looks something
like this:
> -      Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
> +      Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1,
Node->getFlags());
> 
> FMF only applies to binary FP ops at the moment, but as noted in the commit
message, the plan is to extend the flags to more DAG node types.
> 
> Going forward, we can use the node-level flags to trigger optimizations
that the global target settings (eg, DAG.getTarget().Options.UnsafeFPMath) might
not allow. And some day, we may be able to remove global FP optimization flags
entirely.
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150923/b184f852/attachment.html>