https://codereview.appspot.com/11648043/diff/6001/src/pkg/crypto/md5/md5block_arm.s
File src/pkg/crypto/md5/md5block_arm.s (right):
https://codereview.appspot.com/11648043/diff/6001/src/pkg/crypto/md5/md5block_arm.s#newcode78
src/pkg/crypto/md5/md5block_arm.s:78: ADD R(const), R(t0) ; \
On 2013/07/22 18:41:51, rsc wrote:
If you do
ADD $0x######, R(t0)
This generatesIf you do
ADD $0x######, R(t0)
LDR r11, 0xxxx(PC)
ADD R(t0), R(t0), r11
MOVW $0x######(R(t0)), R(t0)
I imagine that generates the same code but I haven't tried it. Thereisn't really any alternative in native instructions anyway since the
constants in question don't fit into an immediate load.
does the linker emit a load of a PC-relative constant into R11 and
then use R11
in the ADD?
If so, is that faster than doing the MOVM.IA.W to load 4 at a time
from farther
away?
Ignoring cache effects, the LDR/MOVW takes one cycle on an ARM A8 (forthen use R11
in the ADD?
If so, is that faster than doing the MOVM.IA.W to load 4 at a time
from farther
away?
instance) whereas the 4 register MOVM/LDM load only takes 2 cycles which
is a win of 2 cycles and 12 bytes of instruction cache.
I did in fact try the ADD version you suggested and it was quite a bit
slower in the benchmarks (I don't have the exact figures any more
though) - it was one of my intermediate versions.
In old and modern ARM MOVM is the same or faster than multiple MOVW s if
you are loading >= 2 registers as far as I know.
https://codereview.appspot.com/11648043/diff/6001/src/pkg/crypto/md5/md5block_arm.s#newcode233
src/pkg/crypto/md5/md5block_arm.s:233: TEXT _md5_block_table(SB),7,$-4
On 2013/07/22 18:41:51, rsc wrote:
Please make this (read-only) data. See the masks and shifts tables
used for AES
in ../../runtime/*.s.
OK I'll have a go with that.Please make this (read-only) data. See the masks and shifts tables
used for AES
in ../../runtime/*.s.
In an ideal world one would quad word align the table to minimise the
number of cache lines it uses but I don't think that is possible is it?
https://codereview.appspot.com/11648043/
--
---
You received this message because you are subscribed to the Google Groups "golang-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to golang-dev+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
