[Bug tree-optimization/80293] [6/7 Regression] unnecessary code at -O2 (-O1 is fine)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80293

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Martin Jambor from comment #3)
> (In reply to Richard Biener from comment #2)
> > We have another bugreport that complains about SRA and Martin said he had
> > patches but intended to wait for GCC 8.
> 
> My patch for PR 78687 unfortunately won't help here, that one is aimed
> at making the original aggregates disappear when in fact we want to
> leave the intermediate aggregate in a chain of assignments intact,
> quite the opposite of a normal mode of operation.
> 
> > Note that I believe that store-merging, SRA (and bswap) are all passes that
> > should be merged given they perform similar analyses and transforms.
> 
> Yeah, and I am (slowly) starting to look in that direction.
> 
> However, with the totally flow-insensitive SRA we have meanwhile, the
> only remedy for this that I can think of is to throttle down total
> scalarization here to never produce more than, say,
> DECL_SIZE (candidate) / BITS_PER_WORD scalar replacements.
> 
> What do you think?

Yeah, but that's not sth for GCC 7 either.  Note that we did get better
both in FRE and DSE so we may no longer need SRA to perform some of the
optimizations [in the early pipeline].

That said, the array cases we've seen are somewhat disturbing...  maybe
we can disable total scalarization for those?