Messages in this thread | | | Date | Mon, 16 Nov 1998 22:05:53 +0100 (CET) | From | Rik van Riel <> | Subject | Re: Linux-asm (was A patch for linux 2.1.127) |
| |
On Mon, 16 Nov 1998, Jamie Lokier wrote: > On Sun, Nov 15, 1998 at 10:25:37AM -0500, Richard B. Johnson wrote: > > The demonstration code was an example of how to pipeline > > something that is, buy it's nature, capable of being > > pipelined. It is not a fix for something that is not > > broken. > > If you want pipelining, why not help the GCC folks implement better > instruction scheduling? One guy is claiming speedups of 30% or more > just from better experimental scheduling for an AMD K6.
Things like writing 'unrolled loop' code and making sure there are multiple things to be done at once (making good multi-pipelining possible) are also nice things to do.
But I agree that we really should be working on gcc/egcs to: - add machine descriptions for AMD/Rise/Cyrix/Centaur/etc. so the compiler can choose the proper instructions on the proper chip (different cost for operands on different chips) - have the compiler do more efficient multipipelining, this will be especially useful for the Rise chip - replace one expensive opcode by multiple (parallelizible) cheap ones when useful (-frisc for pgcc???) - make egcs do better instruction scheduling by avoiding unneeded branches and/or doing predicative branching (a'la Merced) in software (if it's worth it at all:-)
Rik -- slowly getting used to dvorak kbd layout... +-------------------------------------------------------------------+ | Linux memory management tour guide. H.H.vanRiel@phys.uu.nl | | Scouting Vries cubscout leader. http://www.phys.uu.nl/~riel/ | +-------------------------------------------------------------------+
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |