Fix verify_type ICE during Ada bootstrap

> On Tue, 24 Nov 2015, Jan Hubicka wrote:
> 
> > > > 
> > > > We do already wrap all bases into MEM_REFs at streaming time, it would
> > > > be easy to adjust it to make it effectively alias-set zero.  But of
> > > > course the overhead and the downstream effects of having more MEM_REFs
> > > > (we strip the unneeded ones at stream-in) are unknown (compared to
> > > > the effect of disabling inlining).
> > > 
> > > Hmm, I can test in on Firefox (once I get it back to working condition).
> > 
> > One way would be to keep current MEM_REFS stripping and conditoinal in
> > get_alias_set on strict aliasing, but extend inliner to introduce them at a
> > point -fno-strict-aliasing is inlined to -fstrict-aliasing.  That way we could
> > drop the code in lto-streamer-out that forcingly set alias set to 0 when
> > get_alias_set == 0 and hopefully get all code transitions right.
> 
> Yeah, that could also work.  We can also rewrite overflow stuff
> this way to do overflow related inlining (in one direction only?).
> That is, when inlining !strict-overflow into strict-overflow code
> re-write arithmetic to unsigned during inlining.
> 
> Sth for next stage1.
> 
> Maybe you can open an enhancement PR for these cases.

I will certainly do.
sadly more I understand the implementation the easier I can consturct wrong
code examples (see testcases bellow).

I think the whole idea of storing TYPE_ALIAS_SET at streaming out time is not
working well. First of all it does not solve optimization attribute and second
we can randomly lose the info (on prestreamed type or types where canonical
type merging prevails with non-0 alias set type) or push random type to alias
set 0 (where canonical type merging prevail the oposite direction).  I do not
see how to easily fix it: canonical type merging can not make difference between
alias set 0 types and others unless we make it clear that the derived types
can not alias (which I think they can).  I suppose only way here would be to
force all alias set 0 types to be variant and revisit all the code to check
it before going to main vairant&canonical type.  Compared to that I like
the solution with flag in MEM_REF better, but that of course is an invasive
change where we will need to revisit all MEM_REF construction to set the
flag correctly.

I wonder what would you think of the following patch.  It basically makes
type representation to be completely agnostic of -fstrict-aliasing (it should
be because -fstrict-aliasing is function local property, while types are
not) and makes -fstrict-aliasing to be purely evaulated at a time we ask
TBAA oracle. I disabled the TYPE_ALIAS_SET streaming and instead I assert
that LTO's implementation will be compatible (which caught some surprises
where we tamper with alias set in rtti.c and free_lang_data)

I modifed inliner to make the -fstrict-aliasing infectious. That is instead of
forbidding the inlinng it simply drops the flags in caller. Moreover I play
the game with COMDATs and the fact that whenever you inline comdat w/o
explicit optimization attribute to caller w/o explicit optimization
attribute you may assume that the function body is valid under caller flags.
We already play this game in can_inline_p.  This means that no inlines are
blocked in firefox.

Of course it is always dodgy to change optimization flags after ealry
optimizations that may have made code previously valid wrt -fstrict-aliasing
invalid. I reviewed the 26 uses of -fstrict-aliasing in the compiler and it
seems that only ipa-icf and fold-const may result in such transform. So I
disabled them pre-inlining (which I think is good idea for time being until we
make -fstrict-aliasing part of MEM_REF: both transforms are far less important
than inlining). Once we update to MEM_REF we can easilu drop this.

The patch bootstraps/regtests x86_64-linux and seems to do decent job on
Firefox (actually increasing effectivity of TBAA, only 26 functions are demoted
to -fno-strict-aliasing because of the new code in ipa-inline-transform).  I
plan to do more testing tomorrow (I still can't build the firefox binary to do
some benchamrks).

Honza

	* ipa-inline-transform.c (inline_call): Merge -fno-strict-aliasing
	if needed.
	* ipa-icf-gimple.c (func_checker::compatible_types_p): Pass true
	to get_alias_set.
	* alias.c (get_alias_set): Add new strict flag.
	(new_alias_set): Always produce new set.
	(record_component_aliases): Pass true to get_alias_set.
	* alias.h (get_alias_set): New parameter STRICT which is false by
	default.
	* fold-const.c (operand_equal_p): Before inlining don not permit
	any transformations that would be invalid if code became strict-aliasing
	* tree-streamer-out.c
	 (pack_ts_type_common_value_fields): Do not stream TYPE_ALIAS_SET;
	sanity check that no alias set 0 info is lost.
	* tree-streamer-in.c (unpack_ts_type_common_value_fields): Do not
	stream in TYPE_ALIAS_SET.
	* tree.c (free_lang_data): Pass true to get_alias_set.

	* lto-symtab.c (warn_type_compatibility_p): Pass true to
	get_alias_set

	* c-common.c (parse_optimize_options): Remove ugly hack that makes
	strict-aliasing changes to be silently ignored.

	* gcc.dg/lto/alias-1_0.c: New testcase.
	* gcc.dg/lto/alias-1_1.c: New testcase.
	* gcc.c-torture/execute/alias-1.c: New testcase.

Fix verify_type ICE during Ada bootstrap

Commit Message

Patch