commit 2d7f323d10f256ab4d9630d6f663c6db4af98856
author Xin Li <delphij@google.com> Thu Sep 05 12:31:37 2019 -0700
committer android-build-merger <android-build-merger@google.com> Thu Sep 05 12:31:37 2019 -0700
tree d63135df0e611e34319dc4f50f04fa715a1a17bf
parent 53c13b251e1cc840458482c1efc9a73a67e8bc5c
parent e4530e29dfd6179e6a3c4a48549575a6ab198f97

[automerger skipped] DO NOT MERGE - Merge Android 10 into master am: 244dc17f49 -s ours am: 5ac4c8b94f
am: e4530e29df

Change-Id: Ib3a550bb3f1e2f8970c496eb83448320f07bcb72

silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c

diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c
index ee06f98..6f3be02 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 @@
         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 @@
         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 @@
             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 @@
                     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 );