Snap for 5853686 from 5d2e0a05b7df5aca4525cc547cd8538f3339b526 to qt-qpr1-releaseandroid-10.0.0_r29android-10.0.0_r28android-10.0.0_r27android-10.0.0_r26android-10.0.0_r25android-10.0.0_r24android-10.0.0_r23android-10.0.0_r22android-10.0.0_r21android-10.0.0_r20android-10.0.0_r19android-10.0.0_r18android-10.0.0_r16android-10.0.0_r15android10-qpr1-releaseandroid10-qpr1-d-releaseandroid10-qpr1-c-s1-releaseandroid10-qpr1-c-releaseandroid10-qpr1-b-s1-releaseandroid10-qpr1-b-release

Change-Id: Ifa15d1f6ebf7bb99a48193b0763b532292f9b3a6

1 files changed, 8 insertions, 3 deletions

diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c
index ee06f986..6f3be025 100644
--- a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c
+++ b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c
@@ -84,7 +84,9 @@ void silk_warped_autocorrelation_FIX_neon(
         silk_assert( ( order & 1 ) == 0 );
         silk_assert( 2 * QS - QC >= 0 );
 
-        ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 );
+        /* The additional +4 is to ensure a later vld1q_s32 call does not overflow.               */
+        /* Strictly, only +3 is needed but +4 simplifies initialization using the 4x32 neon load. */
+        ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER + 4, opus_int32 );
 
         input_QS = input_QST;
         /* input_QS has zero paddings in the beginning and end. */
@@ -121,6 +123,8 @@ void silk_warped_autocorrelation_FIX_neon(
         vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
         input_QS += 4;
         vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
+        input_QS += 4;
+        vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
         input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT;
 
         /* The following loop runs ( length + order ) times, with ( order ) extra epilogues.                  */
@@ -153,7 +157,8 @@ void silk_warped_autocorrelation_FIX_neon(
             opus_int o = orderT;
             int32x4_t state_QS_s32x4[ 3 ][ 2 ];
 
-            ALLOC( state, length + orderT, opus_int32 );
+            /* The additional +4 is to ensure a later vld1q_s32 call does not overflow. */
+            ALLOC( state, length + order + 4, opus_int32 );
             state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 );
 
             /* Calculate 8 taps of all inputs in each loop. */
@@ -172,7 +177,7 @@ void silk_warped_autocorrelation_FIX_neon(
                     state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 );
                     state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
                     state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ];
-                } while( ++n < ( length + order - 3) );
+                } while( ++n < ( length + order ) );
                 in = state;
                 o -= 8;
             } while( o > 4 );