Merge branch 'opt.TrkExSTEP_Propagator-20210325' into 'master'

TrkExSTEP_Propagator: Optimize some functions, even in dbg builds.

See merge request atlas/athena!42000

Tracking/TrkExtrapolation/TrkExSTEP_Propagator/CMakeLists.txt

-# Copyright (C) 2002-2020 CERN for the benefit of the ATLAS collaboration
+# Copyright (C) 2002-2021 CERN for the benefit of the ATLAS collaboration
 
 # Declare the package name:
 atlas_subdir( TrkExSTEP_Propagator )
@@ -14,3 +14,17 @@ atlas_add_component( TrkExSTEP_Propagator
                      LINK_LIBRARIES ${CLHEP_LIBRARIES} AthenaBaseComps AthenaKernel GeoPrimitives EventPrimitives TrkGeometry TrkEventPrimitives
 		             TrkMaterialOnTrack TrkParameters TrkExInterfaces TrkExUtils GaudiKernel TrkDetDescrUtils TrkSurfaces TrkTrack
 		             MagFieldElements MagFieldConditions CxxUtils)
+
+
+# Code in this file makes heavy use of eigen and runs orders of magnitude
+# more slowly without optimization.  So force this to be optimized even
+# in debug builds.  If you need to debug it you might want to change this.
+# Specifying optimization via an attribute on the particular
+# function didn't work, because that still didn't allow inlining.
+if ( "${CMAKE_BUILD_TYPE}" STREQUAL "Debug" )
+  set_source_files_properties(
+     ${CMAKE_CURRENT_SOURCE_DIR}/src/STEP_Propagator.cxx
+     PROPERTIES
+     COMPILE_FLAGS "${CMAKE_CXX_FLAGS_RELWITHDEBINFO}"
+     COMPILE_DEFINITIONS "FLATTEN" )
+endif()

Tracking/TrkExtrapolation/TrkExSTEP_Propagator/src/STEP_Propagator.cxx

@@ -1924,12 +1924,13 @@ Trk::STEP_Propagator::propagateWithJacobian (Cache& cache,
 // This is the default STEP method
 /////////////////////////////////////////////////////////////////////////////////
 
-// Force this function to be optimized, even in unoptimized builds:
-// this makes heavy use of Eigen, and Eigen is extremely slow if optimization
-// is disabled.  If you do need to debug this function, you can comment
-// this out.
-#if defined(__GNUC__) && !defined(__OPTIMIZE__)
-__attribute__ ((optimize(2)))
+#if defined(FLATTEN) && defined(__GNUC__)
+// We compile this package with optimization, even in debug builds; otherwise,
+// the heavy use of Eigen makes it too slow.  However, from here we may call
+// to out-of-line Eigen code that is linked from other DSOs; in that case,
+// it would not be optimized.  Avoid this by forcing all Eigen code
+// to be inlined here if possible.
+__attribute__ ((flatten))
 #endif
 bool
 Trk::STEP_Propagator::rungeKuttaStep( Cache& cache,