Boost logo

Boost :

Subject: [boost] [optional] generates unnessesary code for trivial types
From: Hite, Christopher (Christopher.Hite_at_[hidden])
Date: 2012-01-25 12:28:45


When decompiling my code I noticed a bunch of unnessesary code caused by boost::optional.

1) deconstruction
typedef boost::optional<int> optional_int;

void deconstruct_boost_optional(optional_int& o){
        o.~optional_int();
}

One would expect this to do nothing. Instead gcc 4.6.0 with O3 generates:

        if(m_initialized){
                // do nothing
                m_initialized = false;
        }

00000000 <deconstruct_boost_optional(boost::optional<int>&)>:
   0: 8b 44 24 04 mov 0x4(%esp),%eax
   4: 80 38 00 cmpb $0x0,(%eax)
   7: 74 03 je c <deconstruct_boost_optional(boost::optional<int>&)+0xc>
   9: c6 00 00 movb $0x0,(%eax)
   c: f3 c3 repz ret

This one could be easily fixed by removing the bit that sets m_initialized to false, since we're deconstructing anyway.

2) assignment also generates these problems:

void assign_boost_optional(optional_int& o){
        o=13;
}

Here there's a semantic issue: we have to decide to use the copy constructor or operator=. This is also wasteful for POD types or any type which has_trivial_copy<>.

3) Even more expensive is if we want to copy an optional<int>

void assign_boost_optional(optional_int& a,optional_int& b){
        a=b;
}

00000000 <assign_boost_optional(boost::optional<int>&, boost::optional<int>&)>:
   0: 8b 44 24 04 mov 0x4(%esp),%eax
   4: 8b 54 24 08 mov 0x8(%esp),%edx
   8: 80 38 00 cmpb $0x0,(%eax)
   b: 74 0b je 18 <assign_boost_optional(boost::optional<int>&, boost::optional<int>&)+0x18>
   d: 80 3a 00 cmpb $0x0,(%edx)
  10: 75 16 jne 28 <assign_boost_optional(boost::optional<int>&, boost::optional<int>&)+0x28>
  12: c6 00 00 movb $0x0,(%eax)
  15: c3 ret
  16: 66 90 xchg %ax,%ax
  18: 80 3a 00 cmpb $0x0,(%edx)
  1b: 74 09 je 26 <assign_boost_optional(boost::optional<int>&, boost::optional<int>&)+0x26>
  1d: 8b 52 04 mov 0x4(%edx),%edx
  20: c6 00 01 movb $0x1,(%eax)
  23: 89 50 04 mov %edx,0x4(%eax)
  26: f3 c3 repz ret
  28: 8b 52 04 mov 0x4(%edx),%edx
  2b: 89 50 04 mov %edx,0x4(%eax)
  2e: c3 ret

Three possible branches! Theoretically single 64 bit copy do the job. I'm tempted to say: it would be best if for any T has_trivial_copy< optional<T> > iff has_trivial_copy<T>. It might make a sense to make an exception for huge T, where the copying an unused T is more expensive than the branching.

4) has_trivial_destructor<T> should impl has_trivial_destructor< optional<T> > , but this is hard to implement without specialization of optional.

Checking has_trivial_destructor might take care of the complexity of optional<T&> since has_trivial_destructor< T& >.

I'd be willing to fix #1. The other issues need some discussion.

Chris


Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk