The overread avoidance fix in cbddee1cca0ebd01e8c5aa694d31228eb4de4b41
broke the computation for the last row since it prevented the safe
reading from the height+1-th row.
---
I'm not fully sure how I came to the conclusion that those values are
not used during the last iteration. I either checked the wrong function
or I just looked above, saw that they are used and incorrectly assumed
that they are already loaded for the next iteration.
My testing was unfortunately flawed since CPUFLAGS=-neon doesn't masks
neon out and I tested on a CPU with neon.
libavcodec/arm/hpeldsp_armv6.S | 10 ++++------
1 file changed, 4 insertions(+), 6 deletions(-)
diff --git a/libavcodec/arm/hpeldsp_armv6.S b/libavcodec/arm/hpeldsp_armv6.S
index 984e0f0..f1abc32 100644
--- a/libavcodec/arm/hpeldsp_armv6.S
+++ b/libavcodec/arm/hpeldsp_armv6.S
@@ -132,12 +132,11 @@ function ff_put_pixels8_y2_armv6, export=1
uhadd8 r9, r5, r7
eor r11, r5, r7
and r10, r10, r12
- ldrc_pre ne, r4, r1, r2
+ ldr_pre r4, r1, r2
uadd8 r8, r8, r10
and r11, r11, r12
uadd8 r9, r9, r11
- it ne
- ldrne r5, [r1, #4]
+ ldr r5, [r1, #4]
uhadd8 r10, r4, r6
eor r6, r4, r6
uhadd8 r11, r5, r7
@@ -194,10 +193,9 @@ function ff_put_pixels8_y2_no_rnd_armv6, export=1
1:
subs r3, r3, #2
uhadd8 r8, r4, r6
- ldrc_pre ne, r4, r1, r2
+ ldr_pre r4, r1, r2
uhadd8 r9, r5, r7
- it ne
- ldrne r5, [r1, #4]
+ ldr r5, [r1, #4]
uhadd8 r12, r4, r6
ldrc_pre ne, r6, r1, r2
uhadd8 r14, r5, r7
--
1.9.0
