On Mon, 7 Nov 2016, Paul Burton wrote: > Instead of using the at/$1 register (which does not form part of the > typical calling convention) to provide the end of the source region to > __copy_user* functions, use the a3/$7 register. This prepares us for > being able to call __copy_user* with a standard function call. > > Signed-off-by: Paul Burton <paul.burton@xxxxxxxxxx> > --- > > arch/mips/cavium-octeon/octeon-memcpy.S | 8 ++++---- > arch/mips/include/asm/uaccess.h | 21 ++++++++++++--------- > arch/mips/lib/memcpy.S | 8 ++++---- > 3 files changed, 20 insertions(+), 17 deletions(-) > [...] > diff --git a/arch/mips/lib/memcpy.S b/arch/mips/lib/memcpy.S > index 48684c4..5af9f03 100644 > --- a/arch/mips/lib/memcpy.S > +++ b/arch/mips/lib/memcpy.S > @@ -70,13 +70,13 @@ > > /* > * The exception handler for loads requires that: > - * 1- AT contain the address of the byte just past the end of the source > + * 1- a3 contain the address of the byte just past the end of the source > * of the copy, > - * 2- src_entry <= src < AT, and > + * 2- src_entry <= src < a3, and > * 3- (dst - src) == (dst_entry - src_entry), > * The _entry suffix denotes values when __copy_user was called. > * > - * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user > + * (1) is set up up by uaccess.h and maintained by not writing a3 in copy_user > * (2) is met by incrementing src by the number of bytes copied > * (3) is met by not doing loads between a pair of increments of dst and src > * > @@ -549,7 +549,7 @@ > nop > LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address > nop > - SUB len, AT, t0 # len number of uncopied bytes > + SUB len, a3, t0 # len number of uncopied bytes > bnez ta2, .Ldone\@ /* Skip the zeroing part if inatomic */ > /* > * Here's where we rely on src and dst being incremented in tandem, With the lone explicit use of $at gone from this code I think you can remove `.set noat/at=v1' pseudo-ops across this source file as well. I think it would be good actually to do both changes with a single patch as it will ensure that whoever comes across them in the future in a look through our repo history will know immediately that one is a direct consequence of the other (i.e. that we only have those `.set noat/at=v1' pseudo-ops because of the special use of $at in this code). Thanks for doing these clean-ups; I actually have found this use of $at here particularly irritating. Maciej
