[Bug c++/68148] New: Devirtualization only applies to last of multiple successive calls

[Bug c++/68148] New: Devirtualization only applies to last of multiple successive calls

[matt at godbolt dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68148

            Bug ID: 68148
           Summary: Devirtualization only applies to last of multiple
                    successive calls
           Product: gcc
           Version: 5.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: matt at godbolt dot org
  Target Milestone: ---

Given the code:

----
struct Interface {
  virtual ~Interface() {}
  virtual void virtualFunc() = 0;
  virtual void virtualFunc2() = 0;
};

struct Concrete : Interface {
  int counter_;
  Concrete() : counter_(0) {}
  void virtualFunc() { counter_++; }
  void virtualFunc2() { counter_++; }
};

void test(Interface &c) {
  c.virtualFunc();
  c.virtualFunc2();
}
----

(Compiled at -O3 -fdevirtualize-speculatively)

Speculative devirtualization is applied to the call to virtualFunc2, but not to
virtualFunc. (See https://goo.gl/Vtx5Fe). If one comments out the call to
virtualFunc2, then the virtualFunc() call *is* speculatively devirtualized
(https://goo.gl/G8f505).

It seems to me that either both should be spec devirtualized, or none. Or
perhaps even more generally, if the vtable pointer is that of "Concrete" then
both calls can be inlined in one and converted to counter+=2 (provided
inspection proved that Concrete's virtualFunc() does not modify the vtable,
which I believe is otherwise a barrier to this kind of optimization).

Am I missing something here, or is this a missed opportunity?

Thanks, Matt

[Bug ipa/68148] Devirtualization only applies to last of multiple successive calls

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68148

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
                 CC|                            |hubicka at gcc dot gnu.org
          Component|c++                         |ipa

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
I suppose it's simply that the devirt machinery operates at a call level and
not
a vtable level.

[Bug ipa/68148] Devirtualization only applies to last of multiple successive calls

[matt at godbolt dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68148

--- Comment #2 from Matt Godbolt <matt at godbolt dot org> ---
That would make sense for my second observation; but if it's a per-call thing
why is only one of the calls speculatively devirtualized?