Fix segmentation fault caused by hls-prepare for some files.

As discussed in github issue #11, some input files cause segmentation
faults during hls-prepare. In these cases the reason is that we access
data after the end of the audio master file vector.

This commit fixes the problem by appending zeros if the audio master is
too short to provide the context for a segment. This also allows us to
unconditionally using 3 seconds of context after every segment (even
after the last segment).
Signed-off-by: Stefan Westerfeld <stefan@space.twc.de>

src/hls.cc

@@ -542,7 +542,7 @@ hls_prepare (const string& in_dir, const string& out_dir, const string& filename
       /* store 3 seconds of the context before this segment and after this segment (if available) */
       const size_t ctx_3sec = 3 * out.sample_rate();
       const size_t prev_size = min<size_t> (start_pos, ctx_3sec);
-      const size_t next_size = min<size_t> (audio_master_data.n_frames() - (segment.size + start_pos), ctx_3sec);
+      const size_t segment_size_with_ctx = prev_size + segment.size + ctx_3sec;
 
       segment.vars["start_pos"] = string_printf ("%zd", start_pos);
       segment.vars["size"] = string_printf ("%zd", segment.size);
@@ -550,12 +550,15 @@ hls_prepare (const string& in_dir, const string& out_dir, const string& filename
       segment.vars["bit_rate"] = string_printf ("%d", bit_rate);
 
       /* write audio segment with context */
-      const size_t start_point = start_pos - prev_size;
-      const size_t end_point = start_point + prev_size + segment.size + next_size;
+      const size_t start_point = min (start_pos - prev_size, audio_master_data.n_frames());
+      const size_t end_point = min (start_point + segment_size_with_ctx, audio_master_data.n_frames());
 
       vector<float> out_signal (audio_master_data.samples().begin() + start_point * audio_master_data.n_channels(),
                                 audio_master_data.samples().begin() + end_point * audio_master_data.n_channels());
 
+      // append zeros if audio master is too short to provide segment with context
+      out_signal.resize (segment_size_with_ctx * audio_master_data.n_channels());
+
       vector<unsigned char> full_flac_mem;
       SFOutputStream out_stream;
       err = out_stream.open (&full_flac_mem,