GCC 14 inlining failed in call to always_inline

[tksuoran]

I can build Jolt by itself fine using GCC 14.
When I try to use Jolt in my own project using CMake fetchcontent, I get the following compilation error with GCC 14:

[build] [27/415] Building CXX object _deps/joltphysics-build/CMakeFiles/Jolt.dir/__/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp.o
[build] FAILED: _deps/joltphysics-build/CMakeFiles/Jolt.dir/__/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp.o 
[build] /usr/bin/g++-14 -DERHE_AUDIO_LIBRARY_NONE -DERHE_DLOAD_ALL_GL_SYMBOLS -DERHE_FONT_RASTERIZATION_LIBRARY_FREETYPE -DERHE_GLTF_LIBRARY_CGLTF -DERHE_GUI_LIBRARY_IMGUI -DERHE_OS_LINUX -DERHE_PNG_LIBRARY_MANGO -DERHE_PROFILE_LIBRARY_TRACY -DERHE_RAYTRACE_LIBRARY_BVH -DERHE_SVG_LIBRARY_LUNASVG -DERHE_TERMINAL_LIBRARY_NONE -DERHE_TEXT_LAYOUT_LIBRARY_HARFBUZZ -DERHE_WINDOW_LIBRARY_GLFW -DFMT_HEADER_ONLY -DGLM_ENABLE_EXPERIMENTAL -DGLM_FORCE_CXX2A=1 -DGLM_FORCE_DEPTH_ZERO_TO_ONE -DJPH_DEBUG_RENDERER -DJPH_FLOATING_POINT_EXCEPTIONS_ENABLED -DJPH_PROFILE_ENABLED -DJPH_USE_AVX -DJPH_USE_AVX2 -DJPH_USE_F16C -DJPH_USE_FMADD -DJPH_USE_LZCNT -DJPH_USE_SSE4_1 -DJPH_USE_SSE4_2 -DJPH_USE_TZCNT -DSPDLOG_FMT_EXTERNAL -D_DEBUG -I"/home/tksuoran/git/erhe/build/Configure GCC 14 Debug/_deps/joltphysics-src/Build/.." -Wno-comment -ffp-contract=off -Wno-inline -mavx2 -msse4.2 -mpopcnt -mlzcnt -mbmi -mf16c -mfma -Wall -Werror -g -Wno-stringop-overflow -ffp-contract=off -std=c++17 -Woverloaded-virtual -Wno-empty-body -O0 -g3 -mavx2 -mbmi -mpopcnt -mlzcnt -mf16c -mfma -mfpmath=sse -Winvalid-pch -include "/home/tksuoran/git/erhe/build/Configure GCC 14 Debug/_deps/joltphysics-build/CMakeFiles/Jolt.dir/cmake_pch.hxx" -MD -MT _deps/joltphysics-build/CMakeFiles/Jolt.dir/__/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp.o -MF _deps/joltphysics-build/CMakeFiles/Jolt.dir/__/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp.o.d -o _deps/joltphysics-build/CMakeFiles/Jolt.dir/__/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp.o -c '/home/tksuoran/git/erhe/build/Configure GCC 14 Debug/_deps/joltphysics-src/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp'
[build] /home/tksuoran/git/erhe/build/Configure GCC 14 Debug/_deps/joltphysics-src/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp: In member function ‘virtual bool JPH::StaticCompoundShape::CastRay(const JPH::RayCast&, const JPH::SubShapeIDCreator&, JPH::RayCastResult&) const’:
[build] /home/tksuoran/git/erhe/build/Configure GCC 14 Debug/_deps/joltphysics-src/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp:430:39: error: inlining failed in call to ‘always_inline’ ‘JPH::StaticCompoundShape::CastRay(const JPH::RayCast&, const JPH::SubShapeIDCreator&, JPH::RayCastResult&) const::Visitor::Visitor(const JPH::RayCast&, const JPH::CompoundShape*, const JPH::SubShapeIDCreator&, JPH::RayCastResult&) [inherited from JPH::CompoundShape::CastRayVisitor]’: function not considered for inlining
[build]   430 |                 using CastRayVisitor::CastRayVisitor;
[build]       |                                       ^~~~~~~~~~~~~~
[build] /home/tksuoran/git/erhe/build/Configure GCC 14 Debug/_deps/joltphysics-src/Jolt/Physics/Collision/Shape/StaticCompoundShape.cpp:449:64: note: called from here
[build]   449 |         Visitor visitor(inRay, this, inSubShapeIDCreator, ioHit);
[build]       |                                                                ^

The compiler arguments may be slightly different from what Jolt uses by itself, even though I've tried to make them compatible.

It seems that MutableCompoundShape::CastRay() uses CompoundShape::CastRayVisitor::CastRayVisitor which is marked with JPH_INLINE which is defined to __inline__ __attribute__((always_inline)) for clang/GCC.

Any suggestions how to deal with this? Thanks!

[tksuoran]

Using JPH_NO_FORCE_INLINE works around the issue, but I am still curious if there is something better.

[jrouwe]

I can reproduce this by just building Jolt in debug mode on gcc 14 (doesn't happen in release mode). I think that when specifying -O0 code shouldn't be inlined but it appears that the compiler gets confused and tries to inline these 2 functions anyway and then fails. I've created a workaround in #1099.

Note that it assumes that JPH_DEBUG will be on when -O0 is specified, but this doesn't need to be true, I did a quick search but can't find a macro that would allow me to detect this in a better way.