2004-04-09 10:33:18

by Herbert Xu

[permalink] [raw]
Subject: [ALPHA] Fix unaligned stxncpy again

Hi:

The current stxncpy on alpha is still broken when it comes to single
word, unaligned, src misalignment > dest misalignment copies.

I've attached a program which demonstrates this problem.

It looks like the code in glibc is correct and does not suffer from
the last problem that Richard fixed where the destination got
partially cleared.

So here is the patch to revert the unaligned case to use the same code
as glibc.

Cheers,
--
Debian GNU/Linux 3.0 is out! ( http://www.debian.org/ )
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt


Attachments:
(No filename) (659.00 B)
b.c (501.00 B)
p (1.47 kB)
Download all attachments

2004-04-09 19:35:20

by Ivan Kokshaysky

[permalink] [raw]
Subject: Re: [ALPHA] Fix unaligned stxncpy again

On Fri, Apr 09, 2004 at 08:32:44PM +1000, Herbert Xu wrote:
> The current stxncpy on alpha is still broken when it comes to single
> word, unaligned, src misalignment > dest misalignment copies.
>
> I've attached a program which demonstrates this problem.

Ugh, indeed. It fails when there is a zero byte before the data.
Thanks.

> So here is the patch to revert the unaligned case to use the same code
> as glibc.

Here is simpler equivalent of that and ev6 fix.

Ivan.

--- linux.orig/arch/alpha/lib/ev6-stxncpy.S Sun Apr 4 07:37:42 2004
+++ linux/arch/alpha/lib/ev6-stxncpy.S Fri Apr 9 22:13:40 2004
@@ -365,7 +365,7 @@ $unaligned:
andnot t2, t6, t12 # E : dest mask for a single word copy
or t8, t10, t5 # E : test for end-of-count too

- cmpbge zero, t2, t3 # E :
+ cmpbge zero, t12, t3 # E :
cmoveq a2, t5, t8 # E : Latency=2, extra map slot
nop # E : keep with cmoveq
andnot t8, t3, t8 # E : (stall)
--- linux.orig/arch/alpha/lib/stxncpy.S Sun Apr 4 07:38:22 2004
+++ linux/arch/alpha/lib/stxncpy.S Fri Apr 9 22:14:24 2004
@@ -317,7 +317,7 @@ $unaligned:
cmpbge zero, t1, t8 # .. e1 : is there a zero?
andnot t2, t6, t12 # e0 : dest mask for a single word copy
or t8, t10, t5 # .. e1 : test for end-of-count too
- cmpbge zero, t2, t3 # e0 :
+ cmpbge zero, t12, t3 # e0 :
cmoveq a2, t5, t8 # .. e1 :
andnot t8, t3, t8 # e0 :
beq t8, $u_head # .. e1 (zdb)

2004-04-09 23:51:10

by Herbert Xu

[permalink] [raw]
Subject: Re: [ALPHA] Fix unaligned stxncpy again

On Fri, Apr 09, 2004 at 11:35:11PM +0400, Ivan Kokshaysky wrote:
>
> Here is simpler equivalent of that and ev6 fix.

Thanks. But is there any reason why we need to have code that
is different from glibc?
--
Debian GNU/Linux 3.0 is out! ( http://www.debian.org/ )
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

2004-04-10 17:49:32

by Ivan Kokshaysky

[permalink] [raw]
Subject: Re: [ALPHA] Fix unaligned stxncpy again

On Sat, Apr 10, 2004 at 09:50:40AM +1000, Herbert Xu wrote:
> Thanks. But is there any reason why we need to have code that
> is different from glibc?

The code is different anyway because the glibc routine clobbers
t7 register which must not be touched in the kernel.
Apparently the whole mess started with a typo in the register
renaming patch...
Yes, ideally glibc and kernel stxncpy code can be exactly the same,
but this assumes that the glibc variant is updated to not use the t7.

Ivan.