[nawb]

So this is all cool, and it's incredibly nifty to have the hardware insight while writing software. But these optimizations (loop tiling, loop fusion, etc) could (and, to my knowledge, ARE) part of basic gcc and java compilers. Why are they not more commonly used? Why do we have to specifically provide a flag to say "Hey btw I almost forgot, make my program 50 times faster."

I'm slightly in the dark as to why loop optimizations are not part of the default compile process.

[RogerL]

Optimization takes serious amounts of time - it's generally a combinatorial problem. If you are not building for production release, do you want to sit for 5 minutes (or 5 hours for a big system) while the compiler cranks away optimizing code that you are just going to rebuild in 20 minutes anyway? Plus, it just doesn't matter for a lot of production code. Finally, if the code is not in a hot path, it really doesn't matter if it is optimized or not. If you make a block of code 10x faster, but some other loop costs you 50% of the performance time, well, you really gained almost nothing. So, heavy optimization tends to be something that you turn on selectively and judiciously.

[nawb]

That's a good point. Thanks!

[username223]

You'd be surprised. Gcc usually won't do loop interchange, which makes a huge difference for many programs, and you have to treat it just right for vectorization to work. It tries, but it's easy to write code it won't optimize.

[fulafel]

C and Java aren't very optimizer friendly. Says something about stagnation of compiler tech that Fortran is stil the king in loop/vector opts.