[mythtv-users] Compiler optimizations

[Andrew Dodd]

Quoting Michael McConnell <mmcconnell at uniserve.com>:

> Hello Joseph,
> 
>     The reason I'm surprised, and suspect that it can't be true, is
> "generally" compiler optimizations of this nature only yield very minimal
> improvement; improvements in the range of 1% - 3% would generally be gained
> on some trivial operation calls via compiler optimization. It's very, very
> rare to see an improvement of 15% - 20% just through compiler optimization,
> however I clearly, and continually see the optimizations made a significant
> different. I'm trying to track down where, and what operation call could be
> used so heavily in MythTV that a PentiumPro to Pentium3 + FastMath compile
> could yield such an improvement.
> 
> Ultimately until I can track the point, which result in such different, I'm
> still a bit of a sceptic and I see it with my own eyes! (-; Clearly it will
> require a "some what trivial" rewrite of whatever is causing this. It must
> be bad code.
Especially considering that most video compression/decompression algorithms are
integer-heavy...  (Which is why MMX actually is such a benefit for many of them.)

For floating-point heavy apps on a Pentium 4, I would not be surprised if
-march=pentium4 gave a significant speedup - I remember reading a while ago a
document on the architecture of the P4, which mentioned that its x87 pipeline is
crippled and Intel reccommends using SSE for all floating-point ops, even
non-vector ones.