This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: PowerPC shrink-wrap support 3 of 3
On Wed, Oct 26, 2011 at 03:01:01PM +0200, Bernd Schmidt wrote:
> On 10/26/11 14:27, Alan Modra wrote:
> > Committed revision 180522. It turns out that shrink-wrapping isn't as
> > effective as it used to be with the 20110915 based sources I was using
> > originally. povray Ray_In_Bound no longer gets the benefit of shrink
> > wrap, likely due to some cfg optimization. We end up with a simple
> > block that just does r3=1 then jumps to last_bb being reached from
> > blocks that need a prologue as well as blocks that don't. That's
> > enough to kill our current shrink wrap implementation. What we need
> > is something to duplicate these tail blocks..
>
> Would it work to insert the epilogue on some edges to this R3=1 block,
> and not on the others?
Wouldn't you need to modify all the target epilogue code? Our
epilogues return.
> (How many edges of each kind are there?)
In the povray case there was one edge of each kind, but I have seen
other cases where there were 4 edges from blocks needing no prologue
and 2 edges from blocks needing a prologue. I can't tell you what the
testcase was now; It was something I looked at when ironing out bugs
in my code. You wouldn't believe how many ways it is possible to
write buggy cfg manipulation code..
I guess the tradeoff between the classic shrink-wrap epilogue scheme
and my duplicate tail idea is whether duplicating tail blocks adds
more code than duplicating epilogues. From what I've seen, the
duplicate tails are generally very small. I guess I should dump out
some info so we can get a better idea.
--
Alan Modra
Australia Development Lab, IBM