Tweak BB analysis for dr_analyze_innermost
Bin.Cheng
amker.cheng@gmail.com
Wed Jun 28 21:06:00 GMT 2017
On Wed, Jun 28, 2017 at 8:06 PM, Richard Sandiford
<richard.sandiford@linaro.org> wrote:
> "Bin.Cheng" <amker.cheng@gmail.com> writes:
>> On Wed, Jun 28, 2017 at 5:56 PM, Richard Sandiford
>> <richard.sandiford@linaro.org> wrote:
>>> "Bin.Cheng" <amker.cheng@gmail.com> writes:
>>>> Question is what would happen if simple_iv succeeds with non-ZERO step
>>>> when called with nest==NULL? The patch skips simple_iv and forces
>>>> ZERO step?
>>>
>>> Yeah, I mentioned that in the covering note:
>>>
>>> The handling seemed strange though. If the DR was part of a loop,
>>> we still tried to express the base and offset values as IVs, potentially
>>> giving a nonzero step. If that failed for any reason, we'd revert to
>>> using the original base and offset, just as we would if we hadn't asked
>>> for an IV in the first place.
>> Hmm, it says "If that failed for any reason, .... revert to using the
>> original base and offset, just as we would if we hadn't asked for an
>> IV in the first place", but question is what would happen if simple_iv
>> succeeds with nest==NULL. After change, the successful simple_iv is
>> bypassed. It's likely this case can't happen in reality.
>
> I suspect we're really in violent agreement here :-) but the reason
> I was surprised at the question was that you seemed to be treating
> the nest==NULL && simple_iv/nonzero-step combination as a corner case
> that hadn't really been addressed, whereas removing that combination is
> the main point of the patch. The reason for removing it isn't that it
Yes, this is where I misunderstood the patch. Thanks very much for
your patient explanation. Now I can see it IS a nice simplification.
Thanks,
Bin
> can't happen (bb-slp-pr65935.c proves that it did, to detrimental effect).
> The reason is that IMO it isn't useful. I don't think it makes sense to
> set the step to a nonzero value for "BB analysis".
>
> Thanks,
> Richard
More information about the Gcc-patches
mailing list