arm64/sve: Write ZCR_EL1 on context switch only if changed

Writes to ZCR_EL1 are self-synchronising, and so may be expensive
in typical implementations.

This patch adopts the approach used for costly system register
writes elsewhere in the kernel: the system register write is
suppressed if it would not change the stored value.

Since the common case will be that of switching between tasks that
use the same vector length as one another, prediction hit rates on
the conditional branch should be reasonably good, with lower
expected amortised cost than the unconditional execution of a
heavyweight self-synchronising instruction.

Signed-off-by: Dave Martin <Dave.Martin@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This commit is contained in:
Dave Martin 2018-05-14 18:51:09 +01:00 committed by Catalin Marinas
parent 37c3ec2d81
commit 159fd7b8d3
2 changed files with 8 additions and 6 deletions

View file

@ -207,12 +207,14 @@
str w\nxtmp, [\xpfpsr, #4]
.endm
.macro sve_load nxbase, xpfpsr, xvqminus1, nxtmp
.macro sve_load nxbase, xpfpsr, xvqminus1, nxtmp, xtmp2
mrs_s x\nxtmp, SYS_ZCR_EL1
bic x\nxtmp, x\nxtmp, ZCR_ELx_LEN_MASK
orr x\nxtmp, x\nxtmp, \xvqminus1
msr_s SYS_ZCR_EL1, x\nxtmp // self-synchronising
bic \xtmp2, x\nxtmp, ZCR_ELx_LEN_MASK
orr \xtmp2, \xtmp2, \xvqminus1
cmp \xtmp2, x\nxtmp
b.eq 921f
msr_s SYS_ZCR_EL1, \xtmp2 // self-synchronising
921:
_for n, 0, 31, _sve_ldr_v \n, \nxbase, \n - 34
_sve_ldr_p 0, \nxbase
_sve_wrffr 0

View file

@ -49,7 +49,7 @@ ENTRY(sve_save_state)
ENDPROC(sve_save_state)
ENTRY(sve_load_state)
sve_load 0, x1, x2, 3
sve_load 0, x1, x2, 3, x4
ret
ENDPROC(sve_load_state)