Ryan C. Gordon wrote:
For what it’s worth, I think this is bunk. The CPU bottleneck is
almost never the code cache, it’s the data cache.
That’s true generally. I tend to write a lot of numerical analysis code,
so I tend to see a lot of simple procedures being repeated ad infinitum,
and I find that squeezing as much as I possible can into the i-cache is
a boon for me. YMMV.
Matt
In terms of high-end game development, it’s not unreasonable to cut the
G3 out at this point; Apple doesn’t ship them anymore since the iBook
went to the G4, and it’s a real pain in some cases to conditionalize
code based on the presence of a vector unit. Unlike MMX/SSE on x86
chips, you can’t use Altivec in a function that a G3 will enter, since
the compiler adds setup code at the start of the function when it sees a
vector variable/function. This means function call overhead and/or ugly
code.
So don’t use little tiny AltiVec functions. Use normal-sized ones,
cutting out the excess overhead. To avoid duplicating your code,
simply compile the functions twice:
#define foo foo_G3
#define bar bar_G3
#undef USE_ALTIVEC
#include “foobar.c”
#undef foo
#undef bar
#define foo foo_G4
#define bar bar_G4
#define USE_ALTIVEC
#include “foobar.c”
#undef USE_ALTIVEC
#undef foo
#undef bar
// some function pointers, initialized to G3 versions
int(*foo)(SDL_Surface *, const char *) = foo_G3;
void(*bar)(double, const void *) = bar_G3;
Furthermore, -Os can be one of the best optimizations you can apply. The
more you can fit into your i-cache the better. Apple recommends all
shipping code be -Os. So unless you know for certain that these binaries
are built optimized for size, you shouldn’t ship them.For what it’s worth, I think this is bunk. The CPU bottleneck is almost
never the code cache, it’s the data cache.
Most apps are fast enough, except at start-up. It is then that
you really want -Os for reducing cache misses, disk seeks, page
faults, TLB misses, and so on.
You can compile the 10% hottest paths with -O3 if you like.
Leaving the bulk of the app at -Os would be good.On Sun, 2005-01-23 at 08:02 -0500, Ryan C. Gordon wrote: