ev6-copy_page.S
上传用户:lgb322
上传日期:2013-02-24
资源大小:30529k
文件大小:4k
- /*
- * 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