Re: [boost] [ptr_container] abstraction penalty

From: Michael Marcin (mmarcin_at_[hidden])
Date: 2007-06-20 12:00:42

Thorsten Ottosen wrote:
> Michael Marcin skrev:
>> Hi,
>>
>> I have a problem with my compiler (ARM's RVCT 2.2.1).
>> It is not generating optimal code for ptr_containers, specifically
>> I've been looking at ptr_vector.
>> I love using ptr_containers to convery ownership but I rely on them
>> being zero overhead over std::vector with explicit delete loops for
>> at least simple operations like a for each loop.
>>
>> I would hope these 2 files (attached) would generate identical
>> assembly for the foo function but they are not even close (attached
>> as txt).
>>
>> I know there are a lot of compilers that are a lot smarter than this
>> one. Unfortunately I'm stuck with this one so I need to teach this
>> compiler to generate better code (by modifying the ptr_container
>> library) or stop using ptr_containers (cry).
>
> Hi Michael,
>
> Some compilers have a problem doing near-perfect inlining,
> which can be a real killer.
>
> I'll try to investigate this.
>

I appreciate it.
If you have things you'd like me to try out I can compile them and send you 
the mixed assembly output.

One thing I noticed is that it seems to fail to eliminate the return value 
copy of the void_ptr_iterator::base call.  Accessing iter_ directly 
(unfortunately) produces much better code but still not the best.

                  _Z3foov PROC ; foo()
;;;14     void foo()
;;;15     {
000000  490d              LDR      r1,|L1.56|
000002  b50c              PUSH     {r2,r3,lr}
000004  6848              LDR      r0,[r1,#4]
000006  9001              STR      r0,[sp,#4]            ;181
000008  6808              LDR      r0,[r1,#0]
00000a  e005              B        |L1.24|
                  |L1.12|
00000c  6800              LDR      r0,[r0,#0]
00000e  6801              LDR      r1,[r0,#0]
000010  1c49              ADDS     r1,r1,#1
000012  6001              STR      r1,[r0,#0]
000014  9800              LDR      r0,[sp,#0]
000016  1d00              ADDS     r0,r0,#4
                  |L1.24|
000018  9901              LDR      r1,[sp,#4]
00001a  9000              STR      r0,[sp,#0]
00001c  4288              CMP      r0,r1
00001e  d1f5              BNE      |L1.12|
000020  bd0c              POP      {r2,r3,pc}            ;186
;;;16      boost::ptr_vector<b>::iterator iend = x.end();
;;;17      for( boost::ptr_vector<b>::iterator it = x.begin(); it != iend; 
++it )
;;;18      {
;;;19       b& b_ = *it;
;;;20       b_.bar();
;;;21      }
;;;22     }
                          ENDP

Thanks,

Michael Marcin