Reland mysterious cast that improves performance.

webrtc/common_audio/signal_processing/cross_correlation.c

Index: webrtc/common_audio/signal_processing/cross_correlation.c
diff --git a/webrtc/common_audio/signal_processing/cross_correlation.c b/webrtc/common_audio/signal_processing/cross_correlation.c
index c14ce4c3220a53bb19f8b8530a7fd2d2a8e11ecd..898d934f2739be65ae64207d02db7b57e5af8746 100644
--- a/webrtc/common_audio/signal_processing/cross_correlation.c
+++ b/webrtc/common_audio/signal_processing/cross_correlation.c
@@ -22,9 +22,11 @@ void WebRtcSpl_CrossCorrelationC(int32_t* cross_correlation,
 
   for (i = 0; i < dim_cross_correlation; i++) {
     int32_t corr = 0;
-    /* Unrolling doesn't seem to improve performance. */
+    // Linux 64-bit performance is improved by the int16_t cast below.
+    // Presumably this is some sort of compiler bug, as there's no obvious
+    // reason why that should result in better code.
     for (j = 0; j < dim_seq; j++)
-      corr += (seq1[j] * seq2[j]) >> right_shifts;
+      corr += (seq1[j] * seq2[j]) >> (int16_t)right_shifts;
     seq2 += step_seq2;
     *cross_correlation++ = corr;
   }