[PATCH] Workaround premature bitfield folding (PR c++/79681)

[Jakub Jelinek]

On Tue, Feb 28, 2017 at 04:43:19PM +0100, Richard Biener wrote:
> > This patch just attempts to reuse as much as possible from orig_inner and
> > only use adjusted BIT_FIELD_REF on that when in GENERIC (first I've tried to
> > use instead COMPONENT_REF with DECL_BIT_FIELD_REPRESENTATIVE, but that isn't
> > handled by constexpr.c either).
> > Bootstrapped/regtested on x86_64/i686-linux, ok for trunk?
> 
> Hmm, I'd rather not do anything is_gimple_form specific.

I think doing what the patch for TREE_CODE (orig_inner) == COMPONENT_REF
is probably fine as well, at least assuming say SCCVN uses
get_inner_reference or something similar to query to ultimate bases and
offsets for CSE purposes.  Though, for non-bitfield refs we still use
in various places x.foo.bar.baz and in other places MEM_REF[&x + xyz],
so I think SCCVN needs to deal with those already anyway.

>  Can't we
> restrict this transform to work on the outermost handled-component only?

That is essentially what the patch does; get_inner_reference returns
the ultimate base and the patch re-adds the handled-components other than
the outermost one back if it yields the same positions/sizes.
Perhaps an alternative would be not to call get_inner_reference at all,
just something that does handle only the outermost handled component
(or add get_inner_reference argument to do that).

> Ideally we'd want to find a common base of course.
> 
> The folding is also guarded with optimize != 0 so maybe simply disable
> it from the FEs?  (but yes, you say it does a bad job on GIMPLE

No, I'm afraid we wouldn't fold it at all then.

> which is quite understandable as there's no TRUTH_*, but we can
> eventually call it from gimplification instead?)

The optimize_bit_field_compare perhaps can work in some cases on GIMPLE,
the truth_fold_andor* I think would need something like tree-ssa-reassoc.c
infrastructure or something similar to gather many comparisons from multiple
basic blocks.

	Jakub