Fix for left shift of negative value in NetEq.

webrtc/common_audio/signal_processing/levinson_durbin.c

Index: webrtc/common_audio/signal_processing/levinson_durbin.c
diff --git a/webrtc/common_audio/signal_processing/levinson_durbin.c b/webrtc/common_audio/signal_processing/levinson_durbin.c
index e890c806cbac26fb330920c95815f28dbc8f83b7..27cc2eca3fa33e22b0e9c8cbac6eaaf6ebade0bf 100644
--- a/webrtc/common_audio/signal_processing/levinson_durbin.c
+++ b/webrtc/common_audio/signal_processing/levinson_durbin.c
@@ -51,8 +51,7 @@ int16_t WebRtcSpl_LevinsonDurbin(const int32_t* R, int16_t* A, int16_t* K,
 
     // K = A[1] = -R[1] / R[0]
 
-    temp2W32 = WEBRTC_SPL_LSHIFT_W32((int32_t)R_hi[1],16)
-            + WEBRTC_SPL_LSHIFT_W32((int32_t)R_low[1],1); // R[1] in Q31

[ivoc, 2016/12/14 09:10:44]

This code is putting the upper and lower bits together again into one number,
but only for R[1]. It seems a bit pointless to me. Maybe the author thought
shifting and adding is cheaper than multiplying?

+    temp2W32 = (R[1] * (1 << norm)); // R[1] in Q31

[hlundin-webrtc, 2016/12/14 08:24:05]

You can drop the outer parentheses.

[hlundin-webrtc, 2016/12/14 08:24:05]

Since the code at line 49 is quite obscure, did you verify that you get the same
value for temp2W32 before and after your change?

[ivoc, 2016/12/14 09:10:44]

I verified that it gives exactly the same values. It also still passes the
bitexactness tests.

     temp3W32 = WEBRTC_SPL_ABS_W32(temp2W32); // abs R[1]
     temp1W32 = WebRtcSpl_DivW32HiLow(temp3W32, R_hi[0], R_low[0]); // abs(R[1])/R[0] in Q31
     // Put back the sign on R[1]