Weird performance degradation when using vector ops

[Andrej Mitrovic]

I was testing out array concatenation and reassignment performance when I ran into some weird performance degradation when using vector ops.

I have two dchar[][] buffers, one is initialized via vector ops, the other via a foreach loop. Both are initialized to the same values, and they compare equal via the assert.

After that, I compare concatenation and reassignment of both buffers. The buffer which was initialized via vector ops operates much slower for some reason.

On my machine, the following code gives these average results:
buffer1 done in usecs: 19820.
buffer2 done in usecs: 650.

Download: https://gist.github.com/965883
Pasted here:

import std.stdio;
import std.algorithm;
import std.array;
import std.datetime;

void testPerf(ref dchar[][] buffer)
{
    buffer = buffer[0..buffer.length/2] ~ "test"d.dup ~ buffer[buffer.length/2 .. $];
}

void main()
{
    dchar[][] buffer1;
    dchar[][] buffer2;
    buffer1.length = 100_000;
    buffer2.length = 100_000;
    
    buffer1[] = "foo"d.replicate(100).dup;
    
    foreach (ref buf; buffer2)
    {
        buf = "foo"d.replicate(100).dup;
    }
    
    assert(buffer1 == buffer2);    // will take a few seconds..
    
    StopWatch sw;
    sw.start();
    testPerf(buffer1);
    sw.stop();
    writefln("buffer1 done in usecs: %s.", sw.peek().usecs);
    
    sw.reset();
    sw.start();
    testPerf(buffer2);
    sw.stop();
    writefln("buffer2 done in usecs: %s.", sw.peek().usecs);
}

The actual initialization is much faster for vector ops than using the foreach loop, but that's not what is being tested here.