Re: [PATCH 4/5] tree-walk: unroll get_mode since loop boundaries are well-known

> On Sat, Apr 2, 2011 at 4:28 AM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
>> On Thu, Mar 31, 2011 at 8:38 AM, Dan McGee <dpmcgee@gmail.com> wrote:
>>> We know our mode entry in our tree objects should be 5 or 6 characters
>>> long. This change both enforces this fact and also unrolls the parsing
>>> of the information giving the compiler more room for optimization of the
>>> operations.
>>
>> I'm skeptical. Did you measure signficant gain after this patch? I
>> looked at asm output with -O3 and failed to see the compiler doing
>> anything fancy. Perhaps it's because I'm on x86 with quite small
>> register set.
>
> I'm on x86_64 and was just using -O2; -O3 produces the same output
> actually. You can see it below. I had taken a look at this before I
> submitted, and noticed a few things:
> 1. We do use multiple registers now since we aren't constrained to a loop.
> 2. movzbl (for the string parts) and cmb instructions tend to get
> clustered first.
> 3. mozbl (for the mode shifting) and leal instructions tend to get
> clustered later.
> 4. The normal case now involves no conditional jumps until the ' '
> (space) comparison.
>
> Call these "trivial", but on my worst case operation times went from
> (shown below) 27.41 secs to 26.49 secs. Considering this operation is
> called 530,588,868 times (that is not a typo) during this operation,
> every saved instruction or non-missed branch prediction does seem to
> make a difference.