diff --git a/src/libslic3r/GCode.cpp b/src/libslic3r/GCode.cpp
index 14b1f87448..cb0ffc01d3 100644
--- a/src/libslic3r/GCode.cpp
+++ b/src/libslic3r/GCode.cpp
@@ -1099,6 +1099,11 @@ void GCode::_do_export(Print& print, GCodeOutputStream &file, ThumbnailsGenerato
     // modifies m_silent_time_estimator_enabled
     DoExport::init_gcode_processor(print.config(), m_processor, m_silent_time_estimator_enabled);
 
+    if (! print.config().gcode_substitutions.values.empty()) {
+        m_find_replace = make_unique<GCodeFindReplace>(print.config());
+        file.set_find_replace(m_find_replace.get());
+    }
+
     // resets analyzer's tracking data
     m_last_height  = 0.f;
     m_last_layer_z = 0.f;
@@ -1155,9 +1160,6 @@ void GCode::_do_export(Print& print, GCodeOutputStream &file, ThumbnailsGenerato
     m_enable_extrusion_role_markers = false;
 #endif /* HAS_PRESSURE_EQUALIZER */
 
-    if (! print.config().gcode_substitutions.values.empty())
-        m_find_replace = make_unique<GCodeFindReplace>(print.config());
-
     // Write information on the generator.
     file.write_format("; %s\n\n", Slic3r::header_slic3r_generated().c_str());
 
@@ -1572,6 +1574,7 @@ void GCode::process_layers(
     );
 
     // The pipeline elements are joined using const references, thus no copying is performed.
+    output_stream.set_find_replace(nullptr);
     if (m_spiral_vase && m_find_replace)
         tbb::parallel_pipeline(12, generator & spiral_vase & cooling & find_replace & output);
     else if (m_spiral_vase)
@@ -1580,6 +1583,7 @@ void GCode::process_layers(
         tbb::parallel_pipeline(12, generator & cooling & find_replace & output);
     else
         tbb::parallel_pipeline(12, generator & cooling & output);
+    output_stream.set_find_replace(m_find_replace.get());
 }
 
 // Process all layers of a single object instance (sequential mode) with a parallel pipeline:
@@ -1623,6 +1627,7 @@ void GCode::process_layers(
     );
 
     // The pipeline elements are joined using const references, thus no copying is performed.
+    output_stream.set_find_replace(nullptr);
     if (m_spiral_vase && m_find_replace)
         tbb::parallel_pipeline(12, generator & spiral_vase & cooling & find_replace & output);
     else if (m_spiral_vase)
@@ -1631,6 +1636,7 @@ void GCode::process_layers(
         tbb::parallel_pipeline(12, generator & cooling & find_replace & output);
     else
         tbb::parallel_pipeline(12, generator & cooling & output);
+    output_stream.set_find_replace(m_find_replace.get());
 }
 
 std::string GCode::placeholder_parser_process(const std::string &name, const std::string &templ, unsigned int current_extruder_id, const DynamicConfig *config_override)
@@ -2847,7 +2853,10 @@ void GCode::GCodeOutputStream::write(const char *what)
         // writes string to file
         fwrite(gcode, 1, ::strlen(gcode), this->f);
         //FIXME don't allocate a string, maybe process a batch of lines?
-        m_processor.process_buffer(std::string(gcode));
+        if (m_find_replace)
+            m_processor.process_buffer(m_find_replace->process_layer(std::string(gcode)));
+        else
+            m_processor.process_buffer(std::string(gcode));
     }
 }
 
diff --git a/src/libslic3r/GCode.hpp b/src/libslic3r/GCode.hpp
index 94bb47f85c..f46558c353 100644
--- a/src/libslic3r/GCode.hpp
+++ b/src/libslic3r/GCode.hpp
@@ -190,6 +190,11 @@ private:
         GCodeOutputStream(FILE *f, GCodeProcessor &processor) : f(f), m_processor(processor) {}
         ~GCodeOutputStream() { this->close(); }
 
+        // Set a find-replace post-processor to modify the G-code before GCodePostProcessor.
+        // It is being set to null inside process_layers(), because the find-replace process
+        // is being called on a secondary thread to improve performance.
+        void set_find_replace(GCodeFindReplace *find_replace) { m_find_replace = find_replace; }
+
         bool is_open() const { return f; }
         bool is_error() const;
         
@@ -209,8 +214,10 @@ private:
         void write_format(const char* format, ...);
 
     private:
-        FILE *f = nullptr;
-        GCodeProcessor &m_processor;
+        FILE             *f { nullptr };
+        // Find-replace post-processor to be called before GCodePostProcessor.
+        GCodeFindReplace *m_find_replace { nullptr };
+        GCodeProcessor   &m_processor;
     };
     void            _do_export(Print &print, GCodeOutputStream &file, ThumbnailsGeneratorCallback thumbnail_cb);
 
diff --git a/src/libslic3r/GCode/FindReplace.cpp b/src/libslic3r/GCode/FindReplace.cpp
index 6d489b9489..e78e34c114 100644
--- a/src/libslic3r/GCode/FindReplace.cpp
+++ b/src/libslic3r/GCode/FindReplace.cpp
@@ -14,7 +14,7 @@ GCodeFindReplace::GCodeFindReplace(const PrintConfig &print_config)
     for (size_t i = 0; i < subst.size(); i += 3) {
         boost::regex pattern;
         try {
-            pattern.assign(subst[i]);
+            pattern.assign(subst[i], boost::regex::optimize); // boost::regex::icase
         } catch (const std::exception &ex) {
             throw RuntimeError(std::string("Invalid gcode_substitutions parameter, failed to compile regular expression: ") + ex.what());
         }
@@ -60,7 +60,7 @@ std::string GCodeFindReplace::process_layer(const std::string &ain)
         temp.clear();
         temp.reserve(in->size());
         boost::regex_replace(ToStringIterator(temp), in->begin(), in->end(),
-            substitution.pattern, substitution.format, boost::match_default | boost::format_all);
+            substitution.pattern, substitution.format, boost::match_default | boost::match_not_dot_newline | boost::match_not_dot_null | boost::format_all);
         std::swap(out, temp);
         in = &out;
     }