Optimize SAIF writes (#5916)

include/verilated_saif_c.cpp

@@ -319,6 +319,7 @@ void VerilatedSaif::close() VL_MT_SAFE_EXCLUDES(m_mutex) {
     finalizeSaifFileContents();
     clearCurrentlyCollectedData();
 
+    writeBuffered(true);
     ::close(m_filep);
     m_isOpen = false;
 
@@ -437,9 +438,25 @@ void VerilatedSaif::clearCurrentlyCollectedData() {
     m_activityAccumulators.clear();
 }
 
-void VerilatedSaif::printStr(const char* str) { ::write(m_filep, str, strlen(str)); }
+void VerilatedSaif::printStr(const char* str) {
+    m_buffer.append(str);
+    writeBuffered(false);
+}
 
-void VerilatedSaif::printStr(const std::string& str) { ::write(m_filep, str.c_str(), str.size()); }
+void VerilatedSaif::printStr(const std::string& str) {
+    m_buffer.append(str);
+    writeBuffered(false);
+}
+
+void VerilatedSaif::writeBuffered(bool force) {
+    if (VL_UNLIKELY(m_buffer.size() >= WRITE_BUFFER_SIZE || force)) {
+        if (VL_UNLIKELY(!m_buffer.empty())) {
+            ::write(m_filep, m_buffer.data(), m_buffer.size());
+            m_buffer = "";
+            m_buffer.reserve(WRITE_BUFFER_SIZE * 2);
+        }
+    }
+}
 
 //=============================================================================
 // Definitions
@@ -454,7 +471,7 @@ void VerilatedSaif::incrementIndent() { m_indent += 1; }
 void VerilatedSaif::decrementIndent() { m_indent -= 1; }
 
 void VerilatedSaif::printIndent() {
-    for (int i = 0; i < m_indent; ++i) printStr(" ");
+    printStr(std::string(m_indent, ' '));  // Must use () constructor
 }
 
 void VerilatedSaif::pushPrefix(const std::string& name, VerilatedTracePrefixType type) {

include/verilated_saif_c.h

@@ -52,8 +52,10 @@ private:
     int m_filep = 0;  // File we're writing to
     bool m_isOpen = false;  // True indicates open file
     std::string m_filename;  // Filename we're writing to (if open)
+    std::string m_buffer;  // Write data buffer
 
     int m_indent = 0;  // Indentation size in spaces
+    static constexpr size_t WRITE_BUFFER_SIZE = 256 * 1024;  // Bytes between write calls
 
     // Currently active scope
     VerilatedSaifActivityScope* m_currentScope = nullptr;
@@ -92,6 +94,7 @@ private:
 
     void printStr(const char* str);
     void printStr(const std::string& str);
+    void writeBuffered(bool force);
 
     void clearCurrentlyCollectedData();
 

test_regress/driver.py

@@ -2382,6 +2382,7 @@ class VlTest:
         out = test.run_capture(cmd, check=True)
         if out != '':
             print(out)
+            self.copy_if_golden(fn1, fn2)
             self.error("SAIF files don't match!")
 
     def _vcd_read(self, filename: str) -> str:

test_regress/t/t_trace_array_saif.py

@@ -17,6 +17,7 @@ test.compile(verilator_flags2=['--cc --trace-saif --trace-structs'])
 
 test.execute()
 
-test.saif_identical(test.trace_filename, test.golden_filename)
+# saif_identical is very slow, so require exact match
+test.files_identical(test.trace_filename, test.golden_filename)
 
 test.passes()

test_regress/t/t_trace_array_saif_portable.py

@@ -17,6 +17,7 @@ test.compile(verilator_flags2=['--cc --trace-saif --trace-structs', '-CFLAGS -DV
 
 test.execute()
 
-test.saif_identical(test.trace_filename, test.golden_filename)
+# saif_identical is very slow, so require exact match
+test.files_identical(test.trace_filename, test.golden_filename)
 
 test.passes()

test_regress/t/t_trace_array_saif_threads_1.py

@@ -17,6 +17,7 @@ test.compile(verilator_flags2=['--cc --trace-saif --trace-threads 1 --trace-stru
 
 test.execute()
 
-test.saif_identical(test.trace_filename, test.golden_filename)
+# saif_identical is very slow, so require exact match
+test.files_identical(test.trace_filename, test.golden_filename)
 
 test.passes()

test_regress/t/t_trace_array_saif_threads_2.py

@@ -17,6 +17,7 @@ test.compile(verilator_flags2=['--cc --trace-saif --trace-threads 2 --trace-stru
 
 test.execute()
 
-test.saif_identical(test.trace_filename, test.golden_filename)
+# saif_identical is very slow, so require exact match
+test.files_identical(test.trace_filename, test.golden_filename)
 
 test.passes()