[f-cpu] statistics of direct indexing usage

[devik <devik@cdi.cz>]

Hi,
only quick result I just got. I tested with large source file
(combine.c from gcc) and tried to see what if I add load/store
with immediate offset to F-CPU ISA.
Without imm:
total insns: 29283
Labels: 2058, JMP: 2568, Conditional jumps: 3034, loadaddr: 3543
Simple L/S: 4881, post-update L/S: 618, imm-offsets:

And with 6bit offset (+- 32 words):
total insns: 27307
Labels: 2122, JMP: 2611, Conditional jumps: 3033, loadaddr: 3515
Simple L/S: 2236, post-update L/S: 558, imm-offsets: 2594

Also all postupdates can be converted to imms (I didn't it
yet - they are in prolog/epilogs).
You can see that about 70% of all load/stores are offsetted
in this small range. Using it saved about 2000 instructions
(9% in this file). I also expect higher speed because in
original file you can see too many address "adds" tightly
one after other.
Even if +-32 bytes (not words) displacement it is:
total insns: 27509
Labels: 2111, JMP: 2606, Conditional jumps: 3033, loadaddr: 3512
Simple L/S: 2511, post-update L/S: 561, imm-offsets: 2348

It is only about 10% worse than +-32 words imediate but still
very good and resulting address is in range of up to two
cachelines - so that it should not be so hard to implement it !

Your opinions ?

devik

*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/