ref: e41a2a1fe0e10a8b2e22f238acb596c35b2f2f7f
parent: c48ea15febf8f6a3a34e998f04c36861f0fc4aed
author: Martin Storsjö <martin@martin.st>
date: Thu Nov 19 07:45:50 EST 2020
arm32: cdef: Simplify some cases in the padding function
--- a/src/arm/32/cdef.S
+++ b/src/arm/32/cdef.S
@@ -190,11 +190,9 @@
beq 1f
// CDEF_HAVE_LEFT+CDEF_HAVE_RIGHT
0:
- ldrh r12, [r3], #2
- vldr \n1, [r1]
- vdup.16 d2, r12
+ vld1.16 {d2[]}, [r3, :16]!ldrh r12, [r1, #\w]
- add r1, r1, r2
+ load_n_incr d0, r1, r2, \w
subs r5, r5, #1
vmov.16 d2[1], r12
vmovl.u8 q0, d0
@@ -207,9 +205,8 @@
b 3f
1:
// CDEF_HAVE_LEFT+!CDEF_HAVE_RIGHT
- ldrh r12, [r3], #2
+ vld1.16 {d2[]}, [r3, :16]!load_n_incr d0, r1, r2, \w
- vdup.16 d2, r12
subs r5, r5, #1
vmovl.u8 q0, d0
vmovl.u8 q1, d2