Merge remote-tracking branch 'qatar/release/0.8' into release/0.10

* qatar/release/0.8:
  arm: hpeldsp: fix put_pixels8_y2_{,no_rnd_}armv6
  arm: hpeldsp: prevent overreads in armv6 asm
  lagarith: reallocate rgb_planes when needed
  lagarith: avoid infinite loop in lag_rac_refill()

Conflicts:
	libavcodec/lagarith.c

Merged-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/arm/asm.S

@@ -132,6 +132,13 @@ T       ldr             \rt, [\rn]
 T       add             \rn, \rn, \rm
 .endm
 
+.macro  ldrc_pre        cc,  rt,  rn,  rm:vararg
+A       ldr\cc          \rt, [\rn, \rm]!
+T       itt             \cc
+T       add\cc          \rn, \rn, \rm
+T       ldr\cc          \rt, [\rn]
+.endm
+
 .macro  ldrd_reg        rt,  rt2, rn,  rm
 A       ldrd            \rt, \rt2, [\rn, \rm]
 T       add             \rt, \rn, \rm

libavcodec/arm/dsputil_armv6.S

@@ -146,10 +146,11 @@ function ff_put_pixels8_y2_armv6, export=1
         eor             r7,  r5,  r7
         uadd8           r10, r10, r6
         and             r7,  r7,  r12
-        ldr_pre         r6,  r1,  r2
+        ldrc_pre        ne,  r6,  r1,  r2
         uadd8           r11, r11, r7
         strd_post       r8,  r9,  r0,  r2
-        ldr             r7,  [r1, #4]
+        it              ne
+        ldrne           r7,  [r1, #4]
         strd_post       r10, r11, r0,  r2
         bne             1b
 
@@ -198,9 +199,10 @@ function ff_put_pixels8_y2_no_rnd_armv6, export=1
         uhadd8          r9,  r5,  r7
         ldr             r5,  [r1, #4]
         uhadd8          r12, r4,  r6
-        ldr_pre         r6,  r1,  r2
+        ldrc_pre        ne,  r6,  r1,  r2
         uhadd8          r14, r5,  r7
-        ldr             r7,  [r1, #4]
+        it              ne
+        ldrne           r7,  [r1, #4]
         stm             r0,  {r8,r9}
         add             r0,  r0,  r2
         stm             r0,  {r12,r14}

libavcodec/lagarith.c

@@ -52,6 +52,7 @@ typedef struct LagarithContext {
     int zeros;                  /**< number of consecutive zero bytes encountered */
     int zeros_rem;              /**< number of zero bytes remaining to output */
     uint8_t *rgb_planes;
+    int      rgb_planes_allocated;
     int rgb_stride;
 } LagarithContext;
 
@@ -507,13 +508,12 @@ static int lag_decode_frame(AVCodecContext *avctx,
         offs[2] = 13;
         offs[3] = AV_RL32(buf + 9);
 
+        l->rgb_stride = FFALIGN(avctx->width, 16);
+        av_fast_malloc(&l->rgb_planes, &l->rgb_planes_allocated,
+                       l->rgb_stride * avctx->height * 4 + 1);
         if (!l->rgb_planes) {
-            l->rgb_stride = FFALIGN(avctx->width, 16);
-            l->rgb_planes = av_malloc(l->rgb_stride * avctx->height * 4 + 1);
-            if (!l->rgb_planes) {
-                av_log(avctx, AV_LOG_ERROR, "cannot allocate temporary buffer\n");
-                return AVERROR(ENOMEM);
-            }
+            av_log(avctx, AV_LOG_ERROR, "cannot allocate temporary buffer\n");
+            return AVERROR(ENOMEM);
         }
         for (i = 0; i < 4; i++)
             srcs[i] = l->rgb_planes + (i + 1) * l->rgb_stride * avctx->height - l->rgb_stride;

libavcodec/lagarithrac.h

@@ -107,6 +107,9 @@ static inline uint8_t lag_get_rac(lag_rac *l)
         l->range -= range_scaled * l->prob[255];
     }
 
+    if (!l->range)
+        l->range = 0x80;
+
     l->low -= range_scaled * l->prob[val];
 
     return val;