[Bug c++/35878] New: Useless NULL pointer check when constructing object

[Bug c++/35878] New: Useless NULL pointer check when constructing object

[ian at airs dot com]

Consider this trivial C++ test case:

#include <vector>
void foo(std::vector<int>* v, int i) { v->push_back(i); }

When I compile this with current trunk with -O2 on i686-pc-linux-gnu I see,
among other things, this:

        movl    4(%edi), %edx
        cmpl    8(%edi), %edx
        je      .L2
        testl   %edx, %edx
        je      .L3
        movl    -4(%edx), %eax
        movl    %eax, (%edx)
        movl    4(%edi), %edx
.L3:
        leal    4(%edx), %eax
        movl    %eax, 4(%edi)

Note that test for whether %edx is NULL.  That comes from this in the
.004t.gimple file:

  D.7266 = operator new (4, D.7265);
  D.7264 = (int *) D.7266;
  if (D.7264 != 0B)
    {
      try
        {
          D.7268 = *__val;
          *D.7264 = D.7268;
        }
      catch
        {
          operator delete (D.7264, D.7265);
        }
      iftmp.48 = D.7264;
    }
  else
    {
      iftmp.48 = D.7264;
    }

We get rid of the try/catch pretty quickly, but we never manage to get rid of
the NULL pointer check.

I believe the NULL pointer check is being generated by build_new_1 in cp/init.c
because we are using a type for which TYPE_NOTHROW_P is true.

This NULL pointer check serves no useful purpose.  It is essentially an
abtraction penalty.  We should figure out some way to get rid of it, either in
the C++ frontend, or in some later optimization pass.


-- 
           Summary: Useless NULL pointer check when constructing object
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: ian at airs dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35878

[Bug c++/35878] Useless NULL pointer check when constructing object

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2008-04-08 21:40 -------


*** This bug has been marked as a duplicate of 19476 ***


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |DUPLICATE


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35878

[Bug c++/35878] Useless NULL pointer check when constructing object

[redi at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35878

--- Comment #3 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Marc Glisse from comment #2)
> The only place I can think of where we could tell the compiler to skip this
> check is in the library. I don't know if we should add
> if(p==0)__builtin_unreachable(); somewhere, or create a new
> attribute((neverzero)) for the pointer in the vector, or something else, but
> the knowledge that the pointer is not zero belongs in the library.

Yes, I'm inclined to agree.

[Bug c++/35878] Useless NULL pointer check when constructing object

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35878

Marc Glisse <glisse at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |UNCONFIRMED
                 CC|                            |glisse at gcc dot gnu.org,
                   |                            |redi at gcc dot gnu.org
         Resolution|DUPLICATE                   |---

--- Comment #2 from Marc Glisse <glisse at gcc dot gnu.org> ---
I am separating this PR from PR 19476 because it really isn't the same at all.

Copying Ian's later comment so it isn't lost: "Note that bug 35878, which was
closed as a duplicate of this one, was a case of placement new.  For placement
new the check for a NULL pointer is particularly useless, as the language
standard says that placement new is required to return the pointer which was
passed in."

The check for a null pointer after a call to placement new is sadly required by
the standard (if anyone cares to try and get that changed, please go ahead).
See this link for instance for an analysis:

http://stackoverflow.com/questions/17571103/passing-null-pointer-to-placement-new

p=new(nullptr)int(42); is an absurd but valid way of saying p=0;.


For the compiler to know that the pointer is non-zero, it would need to see
where the pointer is coming from, which is not the case in this testcase.

The only place I can think of where we could tell the compiler to skip this
check is in the library. I don't know if we should add
if(p==0)__builtin_unreachable(); somewhere, or create a new
attribute((neverzero)) for the pointer in the vector, or something else, but
the knowledge that the pointer is not zero belongs in the library.

(leaving as unconfirmed and not setting the component to libstdc++ because I'd
like someone else to confirm this)