This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Disable CSE skip-blocks


On 10/29/06, Richard Kenner <kenner@vlsi1.ultra.nyu.edu> wrote:
>        flag_cse_follow_jumps = 1;
> -      flag_cse_skip_blocks = 1;

I would really like to be able to eliminate both of these.

With just cse-follow-jumps, CSE works on extended basic blocks, which is really as simple as we should go I think. Doing CSE only locally requires more work (like fwprop) to catch some simple things.

I originally
wrote both pieces of code and they were meant as temporary until we had
"proper" global optimization.  Unfortunately, gcse wasn't good enough to
eliminate them.  Your numbers suggest that we now MIGHT be able to,

I do *not* want to enter that discussion. Period. We have discussed disabling CSE path following many times before, we know what we'd miss, we've written passes to catch most of the things we'd miss, and still the conclusion always was that it was not a good idea to disable path following completely.

So again, that is *not* my intent.  If someone wants to investigate
this, I can only encourage that, but I'm not going to do it.

All I want at this point, is to disable -fcse-skip-blocks, so that I
can work from there on the things that, I believe, do *really* matter,
which are:
- cfglayout mode in more passes
- improving the passes in gcse to  subsume  rtl jump threading

So I'd like to discuss the merits of disabling *just* -fcse-skip-blocks, please.

Gr.
Steven


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]