204 lines
4.2 KiB
ArmAsm
204 lines
4.2 KiB
ArmAsm
/*
|
|
* arch/alpha/lib/ev6-copy_page.S
|
|
*
|
|
* Copy an entire page.
|
|
*/
|
|
|
|
/* The following comparison of this routine vs the normal copy_page.S
|
|
was written by an unnamed ev6 hardware designer and forwarded to me
|
|
via Steven Hobbs <hobbs@steven.zko.dec.com>.
|
|
|
|
First Problem: STQ overflows.
|
|
-----------------------------
|
|
|
|
It would be nice if EV6 handled every resource overflow efficiently,
|
|
but for some it doesn't. Including store queue overflows. It causes
|
|
a trap and a restart of the pipe.
|
|
|
|
To get around this we sometimes use (to borrow a term from a VSSAD
|
|
researcher) "aeration". The idea is to slow the rate at which the
|
|
processor receives valid instructions by inserting nops in the fetch
|
|
path. In doing so, you can prevent the overflow and actually make
|
|
the code run faster. You can, of course, take advantage of the fact
|
|
that the processor can fetch at most 4 aligned instructions per cycle.
|
|
|
|
I inserted enough nops to force it to take 10 cycles to fetch the
|
|
loop code. In theory, EV6 should be able to execute this loop in
|
|
9 cycles but I was not able to get it to run that fast -- the initial
|
|
conditions were such that I could not reach this optimum rate on
|
|
(chaotic) EV6. I wrote the code such that everything would issue
|
|
in order.
|
|
|
|
Second Problem: Dcache index matches.
|
|
-------------------------------------
|
|
|
|
If you are going to use this routine on random aligned pages, there
|
|
is a 25% chance that the pages will be at the same dcache indices.
|
|
This results in many nasty memory traps without care.
|
|
|
|
The solution is to schedule the prefetches to avoid the memory
|
|
conflicts. I schedule the wh64 prefetches farther ahead of the
|
|
read prefetches to avoid this problem.
|
|
|
|
Third Problem: Needs more prefetching.
|
|
--------------------------------------
|
|
|
|
In order to improve the code I added deeper prefetching to take the
|
|
most advantage of EV6's bandwidth.
|
|
|
|
I also prefetched the read stream. Note that adding the read prefetch
|
|
forced me to add another cycle to the inner-most kernel - up to 11
|
|
from the original 8 cycles per iteration. We could improve performance
|
|
further by unrolling the loop and doing multiple prefetches per cycle.
|
|
|
|
I think that the code below will be very robust and fast code for the
|
|
purposes of copying aligned pages. It is slower when both source and
|
|
destination pages are in the dcache, but it is my guess that this is
|
|
less important than the dcache miss case. */
|
|
|
|
|
|
.text
|
|
.align 4
|
|
.global copy_page
|
|
.ent copy_page
|
|
copy_page:
|
|
.prologue 0
|
|
|
|
/* Prefetch 5 read cachelines; write-hint 10 cache lines. */
|
|
wh64 ($16)
|
|
ldl $31,0($17)
|
|
ldl $31,64($17)
|
|
lda $1,1*64($16)
|
|
|
|
wh64 ($1)
|
|
ldl $31,128($17)
|
|
ldl $31,192($17)
|
|
lda $1,2*64($16)
|
|
|
|
wh64 ($1)
|
|
ldl $31,256($17)
|
|
lda $18,118
|
|
lda $1,3*64($16)
|
|
|
|
wh64 ($1)
|
|
nop
|
|
lda $1,4*64($16)
|
|
lda $2,5*64($16)
|
|
|
|
wh64 ($1)
|
|
wh64 ($2)
|
|
lda $1,6*64($16)
|
|
lda $2,7*64($16)
|
|
|
|
wh64 ($1)
|
|
wh64 ($2)
|
|
lda $1,8*64($16)
|
|
lda $2,9*64($16)
|
|
|
|
wh64 ($1)
|
|
wh64 ($2)
|
|
lda $19,10*64($16)
|
|
nop
|
|
|
|
/* Main prefetching/write-hinting loop. */
|
|
1: ldq $0,0($17)
|
|
ldq $1,8($17)
|
|
unop
|
|
unop
|
|
|
|
unop
|
|
unop
|
|
ldq $2,16($17)
|
|
ldq $3,24($17)
|
|
|
|
ldq $4,32($17)
|
|
ldq $5,40($17)
|
|
unop
|
|
unop
|
|
|
|
unop
|
|
unop
|
|
ldq $6,48($17)
|
|
ldq $7,56($17)
|
|
|
|
ldl $31,320($17)
|
|
unop
|
|
unop
|
|
unop
|
|
|
|
/* This gives the extra cycle of aeration above the minimum. */
|
|
unop
|
|
unop
|
|
unop
|
|
unop
|
|
|
|
wh64 ($19)
|
|
unop
|
|
unop
|
|
unop
|
|
|
|
stq $0,0($16)
|
|
subq $18,1,$18
|
|
stq $1,8($16)
|
|
unop
|
|
|
|
unop
|
|
stq $2,16($16)
|
|
addq $17,64,$17
|
|
stq $3,24($16)
|
|
|
|
stq $4,32($16)
|
|
stq $5,40($16)
|
|
addq $19,64,$19
|
|
unop
|
|
|
|
stq $6,48($16)
|
|
stq $7,56($16)
|
|
addq $16,64,$16
|
|
bne $18, 1b
|
|
|
|
/* Prefetch the final 5 cache lines of the read stream. */
|
|
lda $18,10
|
|
ldl $31,320($17)
|
|
ldl $31,384($17)
|
|
ldl $31,448($17)
|
|
|
|
ldl $31,512($17)
|
|
ldl $31,576($17)
|
|
nop
|
|
nop
|
|
|
|
/* Non-prefetching, non-write-hinting cleanup loop for the
|
|
final 10 cache lines. */
|
|
2: ldq $0,0($17)
|
|
ldq $1,8($17)
|
|
ldq $2,16($17)
|
|
ldq $3,24($17)
|
|
|
|
ldq $4,32($17)
|
|
ldq $5,40($17)
|
|
ldq $6,48($17)
|
|
ldq $7,56($17)
|
|
|
|
stq $0,0($16)
|
|
subq $18,1,$18
|
|
stq $1,8($16)
|
|
addq $17,64,$17
|
|
|
|
stq $2,16($16)
|
|
stq $3,24($16)
|
|
stq $4,32($16)
|
|
stq $5,40($16)
|
|
|
|
stq $6,48($16)
|
|
stq $7,56($16)
|
|
addq $16,64,$16
|
|
bne $18, 2b
|
|
|
|
ret
|
|
nop
|
|
unop
|
|
nop
|
|
|
|
.end copy_page
|