Re: G++ 4.0.3 - NRVO (named return value optimization) does not work if destructor/copy ctor is implicit

[Michael Kilburn]

On Jul 15, 5:30 pm, Vlad <vladimir.sakha...@gmail.com> wrote:
> On Jul 14, 4:41 pm, Vlad <vladimir.sakha...@gmail.com> wrote:
>
>
>
> > On Jul 12, 12:24 pm, Michael Kilburn <crusader.m...@gmail.com> wrote:
>
> > > Hi,
>
> > > I have got a problem with GCC 4.0.3 -- if class does not have both
> > > explicit dtor and cctor, NRVO does not work for it. I suspect this
> > > problem applies to other GCC versions. This (of course) adversely
> > > affects compiled code performance.
> > > I can not explain why it happens and looking for:
> > > - an explanation how implicit cctor/dtor affect NRVO-related decisions
> > > made by compiler
> > > - some kind of fix/workaround
>
> > > Thanks!
>
> > > (Btw, I checked this in MSVC 2005 -- everything is fine there)
>
> > > Here is test code (uncomment either 1. or 2. or both and first test
> > > will become as fast as second one):
>
> > > -------- code start -------
>
> > > #include <stdio.h>
> > > #include <stdlib.h>
> > > #include <time.h>
> > > #include <string.h>
>
> > > struct TestX
> > > {
> > >     /************* TEST **************/
> > >     // if you add any of these we get fast:
>
> > >     // 1.
> > >     //TestX() {}
> > >     //TestX(const TestX& rhs) { memcpy(data, rhs.data, sizeof
> > > (data)); }
>
> > >     // 2.
> > >     //~TestX() {}
> > >     /************ END TEST **************/
>
> > >     unsigned data[1000];
>
> > >     void run_test()
> > >     {
> > >         memset(data, 0, sizeof(data));
> > >     }
>
> > > };
>
> > > class RVO_Test
> > > {
> > > public:
> > >     TestX fill(unsigned i)
> > >     {
> > >         TestX data;
> > >         data.run_test();
> > >         return data;
> > >     }
>
> > >     void fill(unsigned i, TestX &data)
> > >     {
> > >         data.run_test();
> > >     }
>
> > > };
>
> > > int main()
> > > {
> > >     const unsigned MAX = 500000;
>
> > >     // test #1
> > >     {
> > >         printf("Running Option 1\n");
> > >         clock_t start = clock();
>
> > >         for(int k = 0; k < 10; k++)
> > >         {
> > >             // slow or fast depending on cc.h TEST changes
> > >             unsigned count = 0;
> > >             RVO_Test a;
> > >             for ( unsigned i = 0; i < MAX; ++i ) {
> > >                 TestX x = a.fill( MAX );
> > >                 count += x.data[5];
> > >             }
> > >             printf("Count: %i\n", count);
> > >         }
>
> > >         printf("Time: %f\n", double(clock() - start)/CLOCKS_PER_SEC);
> > >     }
>
> > >     // test #2
> > >     {
> > >         printf("Running Option 2\n");
> > >         clock_t start = clock();
>
> > >         for(int k = 0; k < 10; k++)
> > >         {
> > >             // fast
> > >             unsigned count = 0;
> > >             RVO_Test a;
> > >             for ( unsigned i = 0; i < MAX; ++i ) {
> > >                 TestX x;
> > >                 a.fill( MAX, x );
> > >                 count += x.data[5];
> > >             }
> > >             printf("Count: %i\n", count);
> > >         }
>
> > >         printf("Time: %f\n", double(clock() - start)/CLOCKS_PER_SEC);
> > >     }
>
> > >     return 0;
>
> > > }
>
> > > -------- code end -------
>
> > > Sincerely yours,
> > > Michael.
>
> > I'll check asm later but from what I see all your functions are
> > inline. So it is possible there is no NRVO at all.
> > It could be simple loop optimization (moving out of loop). ie
> > for(...) char a;
> > vs
> > char a;
> > for(...);
>
> After reviewing asm, good and bad news.
> Bad news, memcpy is indeed present. So it is problem.
> Good news, gcc 4.1.1 does not have this issue.
>
> So either upgrade or use ur workaround.


Thanks for help, Vlad! (sorry for late post, just returned from
vacations)