i#6635 core filter, part 7: Improve init-idle subclass support

clients/drcachesim/tools/filter/record_filter.cpp

@@ -182,6 +182,17 @@ record_filter_t::initialize_shard_type(shard_type_t shard_type)
     return "";
 }
 
+std::string
+record_filter_t::get_output_basename(memtrace_stream_t *shard_stream)
+{
+    if (shard_type_ == SHARD_BY_CORE) {
+        return output_dir_ + DIRSEP + "drmemtrace.core." +
+            std::to_string(shard_stream->get_shard_index()) + ".trace";
+    } else {
+        return output_dir_ + DIRSEP + shard_stream->get_stream_name();
+    }
+}
+
 std::string
 record_filter_t::initialize_shard_output(per_shard_t *per_shard,
                                          memtrace_stream_t *shard_stream)
@@ -192,8 +203,7 @@ record_filter_t::initialize_shard_output(per_shard_t *per_shard,
         // Since some shards may not have inputs, we need to synchronize determining
         // the file extension.
         // First, get our path without the extension, so we can add it later.
-        per_shard->output_path = output_dir_ + DIRSEP + "drmemtrace.core." +
-            std::to_string(shard_stream->get_shard_index()) + ".trace";
+        per_shard->output_path = get_output_basename(shard_stream);
         std::string input_name = shard_stream->get_stream_name();
         // Now synchronize determining the extension.
         auto lock = std::unique_lock<std::mutex>(input_info_mutex_);
@@ -218,7 +228,7 @@ record_filter_t::initialize_shard_output(per_shard_t *per_shard,
             lock.unlock();
         }
     } else {
-        per_shard->output_path = output_dir_ + DIRSEP + shard_stream->get_stream_name();
+        per_shard->output_path = get_output_basename(shard_stream);
     }
     return "";
 }
@@ -228,40 +238,6 @@ record_filter_t::get_writer(per_shard_t *per_shard, memtrace_stream_t *shard_str
 {
     if (per_shard->output_path.empty())
         return "Error: output_path is empty";
-    if (shard_type_ == SHARD_BY_CORE) {
-        // Each output is a mix of inputs so we do not want to reuse the input
-        // names with tids.
-        // Since some shards may not have inputs, we need to synchronize determining
-        // the file extension.
-        // First, get our path without the extension, so we can add it later.
-        per_shard->output_path = output_dir_ + DIRSEP + "drmemtrace.core." +
-            std::to_string(shard_stream->get_shard_index()) + ".trace";
-        std::string input_name = shard_stream->get_stream_name();
-        // Now synchronize determining the extension.
-        auto lock = std::unique_lock<std::mutex>(input_info_mutex_);
-        if (!output_ext_.empty()) {
-            per_shard->output_path += output_ext_;
-            lock.unlock();
-        } else if (!input_name.empty()) {
-            size_t last_dot = input_name.rfind('.');
-            if (last_dot == std::string::npos)
-                return "Failed to determine filename type from extension";
-            output_ext_ = input_name.substr(last_dot);
-            // Set the other key input data.
-            version_ = shard_stream->get_version();
-            filetype_ = add_to_filetype(shard_stream->get_filetype());
-            per_shard->output_path += output_ext_;
-            lock.unlock();
-            input_info_cond_var_.notify_all();
-        } else {
-            // We have to wait for another shard with an input to set output_ext_.
-            input_info_cond_var_.wait(lock, [this] { return !output_ext_.empty(); });
-            per_shard->output_path += output_ext_;
-            lock.unlock();
-        }
-    } else {
-        per_shard->output_path = output_dir_ + DIRSEP + shard_stream->get_stream_name();
-    }
 #ifdef HAS_ZLIB
     if (ends_with(per_shard->output_path, ".gz")) {
         VPRINT(this, 3, "Using the gzip writer for %s\n", per_shard->output_path.c_str());

clients/drcachesim/tools/filter/record_filter.h

@@ -208,12 +208,27 @@ class record_filter_t : public record_analysis_tool_t {
     std::string
     process_delayed_encodings(per_shard_t *per_shard, trace_entry_t &entry, bool output);
 
+    // Computes the output path without the extension output_ext_ which is added
+    // separately after determining the input path extension.
+    virtual std::string
+    get_output_basename(memtrace_stream_t *shard_stream);
+
     std::unordered_map<int, per_shard_t *> shard_map_;
     // This mutex is only needed in parallel_shard_init. In all other accesses
     // to shard_map (print_results) we are single-threaded.
     std::mutex shard_map_mutex_;
     shard_type_t shard_type_ = SHARD_BY_THREAD;
 
+    // For core-sharded we don't have a 1:1 input:output file mapping.
+    // Thus, some shards may not have an input stream at init time, and
+    // need to figure out their file extension and header info from other shards.
+    std::mutex input_info_mutex_;
+    std::condition_variable input_info_cond_var_;
+    // The above locks guard these fields:
+    std::string output_ext_;
+    uint64_t version_ = 0;
+    uint64_t filetype_ = 0;
+
 private:
     virtual bool
     write_trace_entry(per_shard_t *shard, const trace_entry_t &entry);
@@ -254,16 +269,6 @@ class record_filter_t : public record_analysis_tool_t {
     // XXX: We could use a read-write lock but C++11 doesn't have a ready-made one.
     // If we had the input count we could use an array and atomic reads.
     std::unordered_map<int64_t, std::unique_ptr<per_input_t>> input2info_;
-
-    // For core-sharded we don't have a 1:1 input:output file mapping.
-    // Thus, some shards may not have an input stream at init time, and
-    // need to figure out their file extension and header info from other shards.
-    std::mutex input_info_mutex_;
-    std::condition_variable input_info_cond_var_;
-    // The above locks guard these fields:
-    std::string output_ext_;
-    uint64_t version_ = 0;
-    uint64_t filetype_ = 0;
 };
 
 } // namespace drmemtrace