From: Linus Torvalds <email@example.com>
Subject: Re: gcc -O2 vs gcc -Os performance
Date: Thu, 6 Feb 2003 23:21:56 GMT
On Thu, 6 Feb 2003, Martin J. Bligh wrote:
> The observation re low repeat rate is interesting ... might be amusing
> to do some really basic profile-guided optimisation on this grounds,
> take readprofile / oprofile output, and compile the files that don't
> get hammered at all with -Os rather than -O2. Given their low frequency
> (by definition), I'm not sure that improving their icache footprint will
> have a measureable effect though.
Icache footprint has nothing to do with repeat rates, which is exactly why
repeat rates are interesting for -Os.
Icache footprint is directly proportional to the _static_ size of the code
(ie exactly the thing that -Os is supposed to optimize for), while
instruction-level performance measurement is only valid on the _dynamic_
And with modern CPU's with big caches, a _lot_ of cache misses are the
forced kind - the startup costs, not the actual runtime cost. That's not
always true (if you touch big data sets, you'll have replacement misses
too, of course), but it's not really false either.
So think of the I$ (and TLB, and page load/map - all the same) cost as a
fixed cost that will always be there, but that -Os tries to minimize.
That's _one_ dimension in the total cost.
The "traditional" -O2 kind of "try to make the code run fast"
optimizations tend to try to minimize a totally different dimension,
namely the dynamic code speed.
And the time required for running the program is the sum of the static and
dynamic factors. In other words, a _good_ optimization should try to
minimize not one or the other, but the sum.
And low repeat rates means that the dynamic component is smaller, which
clearly makes the static component more important.
For example, if you are doing mp3 encoding, the repeat rates for the core
loop are huge, and the code is small, so clearly the static component is
largely insignificant. Use -O2.
But if you're running a GUI program then just the loading time is often
quite noticeable, and if you can improve that by, say, 10%, then that can
_more_ than make up for almost any amount of stupidity in your code.
Especially since a lot of the code isn't even all that loopy and tends to
have low repeat rates. You're almost guaranteed to be better off using -Os
If you've got performance counter data, check the I$ and ITLB miss ratios,
and if they are at all noticeable, think about the fact that a I$ miss
tends to cost a lot more than a few more dynamic instructions.
I suspect the kernel I$ behaviour is generally pretty good, and the ITLB
behaviour is improved even further thanks to large pages etc. That said, a
user app that blows the I$ will blow the kernel out of the I$ too, so
small is always beautiful, even in the kernel.