Re: [PATCH] define auto_vec copy ctor and assignment (PR 90904)

[Richard Biener <richard.guenther@gmail.com>]

On Fri, Jun 25, 2021 at 10:52 PM Martin Sebor <msebor@gmail.com> wrote:
>
> On 6/1/21 3:38 PM, Jason Merrill wrote:
> > On 6/1/21 3:56 PM, Martin Sebor wrote:
> >> On 5/27/21 2:53 PM, Jason Merrill wrote:
> >>> On 4/27/21 11:52 AM, Martin Sebor via Gcc-patches wrote:
> >>>> On 4/27/21 8:04 AM, Richard Biener wrote:
> >>>>> On Tue, Apr 27, 2021 at 3:59 PM Martin Sebor <msebor@gmail.com> wrote:
> >>>>>>
> >>>>>> On 4/27/21 1:58 AM, Richard Biener wrote:
> >>>>>>> On Tue, Apr 27, 2021 at 2:46 AM Martin Sebor via Gcc-patches
> >>>>>>> <gcc-patches@gcc.gnu.org> wrote:
> >>>>>>>>
> >>>>>>>> PR 90904 notes that auto_vec is unsafe to copy and assign because
> >>>>>>>> the class manages its own memory but doesn't define (or delete)
> >>>>>>>> either special function.  Since I first ran into the problem,
> >>>>>>>> auto_vec has grown a move ctor and move assignment from
> >>>>>>>> a dynamically-allocated vec but still no copy ctor or copy
> >>>>>>>> assignment operator.
> >>>>>>>>
> >>>>>>>> The attached patch adds the two special functions to auto_vec along
> >>>>>>>> with a few simple tests.  It makes auto_vec safe to use in
> >>>>>>>> containers
> >>>>>>>> that expect copyable and assignable element types and passes
> >>>>>>>> bootstrap
> >>>>>>>> and regression testing on x86_64-linux.
> >>>>>>>
> >>>>>>> The question is whether we want such uses to appear since those
> >>>>>>> can be quite inefficient?  Thus the option is to delete those
> >>>>>>> operators?
> >>>>>>
> >>>>>> I would strongly prefer the generic vector class to have the
> >>>>>> properties
> >>>>>> expected of any other generic container: copyable and assignable.  If
> >>>>>> we also want another vector type with this restriction I suggest
> >>>>>> to add
> >>>>>> another "noncopyable" type and make that property explicit in its
> >>>>>> name.
> >>>>>> I can submit one in a followup patch if you think we need one.
> >>>>>
> >>>>> I'm not sure (and not strictly against the copy and assign).
> >>>>> Looking around
> >>>>> I see that vec<> does not do deep copying.  Making auto_vec<> do it
> >>>>> might be surprising (I added the move capability to match how vec<>
> >>>>> is used - as "reference" to a vector)
> >>>>
> >>>> The vec base classes are special: they have no ctors at all (because
> >>>> of their use in unions).  That's something we might have to live with
> >>>> but it's not a model to follow in ordinary containers.
> >>>
> >>> I don't think we have to live with it anymore, now that we're writing
> >>> C++11.
> >>>
> >>>> The auto_vec class was introduced to fill the need for a conventional
> >>>> sequence container with a ctor and dtor.  The missing copy ctor and
> >>>> assignment operators were an oversight, not a deliberate feature.
> >>>> This change fixes that oversight.
> >>>>
> >>>> The revised patch also adds a copy ctor/assignment to the auto_vec
> >>>> primary template (that's also missing it).  In addition, it adds
> >>>> a new class called auto_vec_ncopy that disables copying and
> >>>> assignment as you prefer.
> >>>
> >>> Hmm, adding another class doesn't really help with the confusion
> >>> richi mentions.  And many uses of auto_vec will pass them as vec,
> >>> which will still do a shallow copy.  I think it's probably better to
> >>> disable the copy special members for auto_vec until we fix vec<>.
> >>
> >> There are at least a couple of problems that get in the way of fixing
> >> all of vec to act like a well-behaved C++ container:
> >>
> >> 1) The embedded vec has a trailing "flexible" array member with its
> >> instances having different size.  They're initialized by memset and
> >> copied by memcpy.  The class can't have copy ctors or assignments
> >> but it should disable/delete them instead.
> >>
> >> 2) The heap-based vec is used throughout GCC with the assumption of
> >> shallow copy semantics (not just as function arguments but also as
> >> members of other such POD classes).  This can be changed by providing
> >> copy and move ctors and assignment operators for it, and also for
> >> some of the classes in which it's a member and that are used with
> >> the same assumption.
> >>
> >> 3) The heap-based vec::block_remove() assumes its elements are PODs.
> >> That breaks in VEC_ORDERED_REMOVE_IF (used in gcc/dwarf2cfi.c:2862
> >> and tree-vect-patterns.c).
> >>
> >> I took a stab at both and while (1) is easy, (2) is shaping up to
> >> be a big and tricky project.  Tricky because it involves using
> >> std::move in places where what's moved is subsequently still used.
> >> I can keep plugging away at it but it won't change the fact that
> >> the embedded and heap-based vecs have different requirements.
> >>
> >> It doesn't seem to me that having a safely copyable auto_vec needs
> >> to be put on hold until the rats nest above is untangled.  It won't
> >> make anything worse than it is.  (I have a project that depends on
> >> a sane auto_vec working).
> >>
> >> A couple of alternatives to solving this are to use std::vector or
> >> write an equivalent vector class just for GCC.
> >
> > It occurs to me that another way to work around the issue of passing an
> > auto_vec by value as a vec, and thus doing a shallow copy, would be to
> > add a vec ctor taking an auto_vec, and delete that.  This would mean if
> > you want to pass an auto_vec to a vec interface, it needs to be by
> > reference.  We might as well do the same for operator=, though that
> > isn't as important.
>
> Thanks, that sounds like a good idea.  Attached is an implementation
> of this change.  Since the auto_vec copy ctor and assignment have
> been deleted by someone else in the interim, this patch doesn't
> reverse that.  I will propose it separately after these changes
> are finalized.
>
> My approach was to 1) disable the auto_vec to vec conversion,
> 2) introduce an auto_vec::to_vec() to make the conversion possible
> explicitly, and 3) resolve compilation errors by either changing
> APIs to take a vec by reference or callers to convert auto_vec to
> vec explicitly by to_vec().  In (3) I tried to minimize churn while
> improving the const-correctness of the APIs.
>
> When changing a vec API from by-value to by-reference I used vec &
> and didn't try to use vec * even though the latter seems to be more
> common (I avoided this to minimize the scope of changes).  If we
> want to do this I suggest handling that in a separate change.
>
> It's possible to get rid of vNULL in initialization.  It can be
> replaced with value-initialization in declarations but it's still
> handy when passing "empty" AKA nil vecs as arguments.  I didn't
> do it except in one instance (vec::copy) as a POC.
>
> I only touched APIs that needed to be adjusted so there may still
> be others that take vec by value that didn't change because they're
> not passed an auto_vec.  My hope is that we'll all pitch in and
> convert those over time.
>
> I kept an eye on the indirection due to changing vec to vec & that
> Richi's concerned about but I didn't look into the cost in any
> depth.  I don't have the sense that it's a significant issue.  In
> small static functions it shouldn't matter if they get inlined and
> in bigger ones I'd expect the cost of the indirection to be
> outweighed by other costs.  If there are cases where it should
> be felt to be significant I'll adjust those if you point thm out
> to me.

The cases I'm concerned about are when we for example
load from the vec<> in a loop but we can't hoist the extra
dereference because of aliasing issues (that is, the load
of the m_vec member).  It's enough for the loop to store
to a char * to make TBAA inefficient here.  At least the
situation with TBAA for pointers isn't as bad as it was a
few years ago (all pointers aliased - see get_alias_set
and the large comment before POINTER_TYPE_P
handling).  Still void * aliases all pointers.

From a performance perspective a direct load-to-load
dependence is always quite bad.

Richard.

> Tested on x86_64-linux.
>
> Martin