[PATCH] Early Target Dependent DAG Combine

[Dan Gohman]

On Wed, Jan 30, 2013 at 1:56 PM, Quentin Colombet <qcolombet at apple.com>wrote:

> Hi Dan,
>
> Thanks for the review.
>
> In fact, I'm already aware of the problem you are reporting, i.e.,
> instcombine disrupts the BFI pattern.
> I'm not trying to fix this issue here.
>
> Instead, I was trying to expose the kind of thing that can be achieved
> with the proposed hook using an in-tree target.
>
> Unfortunately, I don't know enough the existing problems for in-tree
> targets to come with a better example, but I'm sure somebody else run into
> this kind of problems and the proposed hook could have ease their
> development.
>
> What do you think of the proposed hook?
>

In the abstract, it's counter to dagcombine's design. Dagcombine is
designed to make use of canonicalization, in a similar fashion to
instcombine. Opening up a mechanism to let targets make exceptions to this
rule opens up a new kind of potential complexity. And, it adds yet another
context to the already-known-to-be-confusing set of contexts in which
target hooks can be called.

I don't think dagcombine's current design needs to remain fixed, however
changing it really ought to be done with a clear motivation.

Dan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130130/348f0639/attachment.html>