Re: [PR24444] speed up locview resolution wiht relaxable frags

[Alan Modra <amodra at gmail dot com>]

On Tue, Apr 16, 2019 at 10:24:04AM -0300, Alexandre Oliva wrote:
> On Apr 14, 2019, Alan Modra <amodra@gmail.com> wrote:
> 
> > Yes, I believe so.  In fact, fr_fix is really unsigned.  Also, it is
> > an error for rs_org frags to go backwards.
> 
> Nice, that makes things simpler.
> 
> >> Tested on x86_64-linux-gnu, native and cross to xtensa-elf.  No new
> >> testcase; the issue was just performance, not output correctness.  Ok to
> >> install?
> 
> > Um, the testcase object file after this patch differs from the
> > original.  First readelf -wi difference shown below.
> 
> Which testcase was that?  I didn't catch that one.

The https://sourceware.org/bugzilla/show_bug.cgi?id=24444 reproducer
attachment.

I see differences in DW_AT_GNU_entry_view, in all cases substituting
3 for 10, or 2 for 9 (lower number after your patch).  I don't know
enough to say whether this is a reasonable result.

> 
> >> +  /* If any frag from frag2, inclusive, to frag1, exclusive, has a
> >> +     nonzero fixed length, the addresses won't be the same as long as
> >> +     we reach frag1.  */
> >> +  bfd_boolean maybe_same_addr = off1 == 0 && off2 == (valueT)frag2->fr_fix;
> 
> > Ins't it true that if the symbol offsets are not at exactly the above
> > values, then you already know the result and there's no need to
> > traverse the frags?  This is assuming you can't .org backwards, which
> > is the case.
> 
> Not entirely, no.  Although the comments state we assume a certain
> ordering of frags, the assumption is only valid for O_gt operations
> arising from the location view machinery.  It doesn't hurt if we resolve
> other O_gt operations if we can find the result, but those could have
> frags in the opposite order.

Thanks for reminding me.  I wrote that comment before writing the
simplification of your patch, and there you'll note I did chase down
the second frag..

> Now, O_gt is not exactly the operation the location view machinery
> needs; O_ne (or, after negation, O_eq) would be more like it.  IIRC the
> reason I went for O_gt was just that it was resolved early in a lot more
> cases than O_eq.  Now, with this kind of fallback, we could actually
> introduce another operation kind that (i) resolves to zero if frags are
> not in the same (sub)section, and (ii) can safely make the optimization
> you suggested for frags in the same (sub)section, with the extra benefit
> that we know we won't ever search all the way to the end of the frag
> linked list without finding the other frag: we could assert-check that,
> and stop the search at the first nonzero-fr_fix.
> 
> Would a new op with this semantics (O_incview?, vs reset view) be
> preferred?

No, if O_gt works let's go with that.

> > The following makes the changes I'm suggesting, and simplifies a few
> > more things.
> 
> Thanks, I'll wait for a response to the question above before taking
> further action on this, but I'll likely integrate your change (credited
> in the ChangeLog) in the patch regardless.

-- 
Alan Modra
Australia Development Lab, IBM