Fix for fuzzer bug.

webrtc/modules/audio_coding/neteq/decision_logic_normal.cc

Index: webrtc/modules/audio_coding/neteq/decision_logic_normal.cc
diff --git a/webrtc/modules/audio_coding/neteq/decision_logic_normal.cc b/webrtc/modules/audio_coding/neteq/decision_logic_normal.cc
index c271bd969df2c4a5065334f2794edde776524ad7..78287e2da55d8b12d070ba1ca3d659c0e863dcdd 100644
--- a/webrtc/modules/audio_coding/neteq/decision_logic_normal.cc
+++ b/webrtc/modules/audio_coding/neteq/decision_logic_normal.cc
@@ -97,14 +97,17 @@ Operations DecisionLogicNormal::CngOperation(Modes prev_mode,
       available_timestamp);
   int32_t optimal_level_samp = static_cast<int32_t>(
       (delay_manager_->TargetLevel() * packet_length_samples_) >> 8);
-  int32_t excess_waiting_time_samp = -timestamp_diff - optimal_level_samp;
+  int64_t excess_waiting_time_samp =

[kwiberg-webrtc, 2017/05/15 09:02:45]

Add const while you're at it?

[ivoc, 2017/05/16 13:40:28]

Done.

+      -static_cast<int64_t>(timestamp_diff) - optimal_level_samp;
 
   if (excess_waiting_time_samp > optimal_level_samp / 2) {
     // The waiting time for this packet will be longer than 1.5
     // times the wanted buffer delay. Apply fast-forward to cut the
     // waiting time down to the optimal.
-    noise_fast_forward_ += excess_waiting_time_samp;
-    timestamp_diff += excess_waiting_time_samp;
+    noise_fast_forward_ = rtc::saturated_cast<size_t>(noise_fast_forward_ +

[ivoc, 2017/05/12 15:00:26]

Not sure if saturated_cast is the right thing to do here, wdyt?

[hlundin-webrtc, 2017/05/15 08:13:45]

I think you should use a dchecked_cast instead. This variable should represent
times on the order of seconds. Even on a 32-bit platform, maxing out a size_t
(32-bit) at 48000 samples per second is more than 24 hours.

[kwiberg-webrtc, 2017/05/15 09:02:45]

+1 for dchecked_cast

[ivoc, 2017/05/16 13:40:28]

Done.

+                                                      excess_waiting_time_samp);
+    timestamp_diff =
+        rtc::saturated_cast<int32_t>(timestamp_diff + excess_waiting_time_samp);

[ivoc, 2017/05/12 15:00:26]

Here I do think saturated_cast makes sense, since the only thing timestamp_diff
is used for below is to check if it's < 0, which should be preserved by
saturated_cast.

[hlundin-webrtc, 2017/05/15 08:13:45]

Acknowledged.

[kwiberg-webrtc, 2017/05/15 09:02:45]

Umm... excess_waiting_time_samp == -static_cast<int64_t>(timestamp_diff) -
optimal_level_samp. So why not avoid updating timestamp_diff, and replace

  timestamp_diff < 0

below with

  optimal_level_samp > 0

?

[hlundin-webrtc, 2017/05/15 09:48:45]

But the modification of timestamp_diff is only done conditioned on
excess_waiting_time_samp > optimal_level_samp / 2.

[kwiberg-webrtc, 2017/05/15 11:39:17]

Right. And you can't assign -timestamp_diff or
dchecked_cast<int32_t>(timestamp_diff + excess_waiting_time_samp), because those
two (equivalent) expressions may evaluate to -INT32_MIN...

Hmpf. You could make a bool that's set to optimal_level_samp > 0 in this branch
and timestamp_diff < 0 in the other, but I guess just using saturating_cast here
is just as good.

   }
 
   if (timestamp_diff < 0 && prev_mode == kModeRfc3389Cng) {