[Bug c++/57175] New: NRVO and alignment

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57175

             Bug #: 57175
           Summary: NRVO and alignment
    Classification: Unclassified
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: c++
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: glisse@gcc.gnu.org


Hello,

in typeck.c (check_return_expr), before applying the NRV optimization, there is
a check: DECL_ALIGN (retval) >= DECL_ALIGN (result)

It seems to me that this check is backwards and should be <= instead (compare
with tree_nrv in tree-nrv.c which seems correct).

#include <iostream>
struct A {
  A(){std::cerr<<"A()\n";}
  ~A(){std::cerr<<"~A()\n";}
  A(A&&){std::cerr<<"A(A&&)\n";}
  A(A const&){std::cerr<<"A(A const&)\n";}
};
A f(){
  alignas(32) A x;
  return x;
}
int main(){
  f();
}

If I understand NRVO properly, main reserves memory for the return value of f
and passes f the address of that memory. f then decides to use that memory
directly for x instead of allocating new memory that would have to be copied
later. So the return memory needs to be at least as aligned as x, not the
reverse (I think retval corresponds to x here).

Current g++ does the elision on this code. If I change >= to <=, it inhibits
the elision.