daanx
commented
2 years ago Thank you for your detailed report @thehans -- and really nice to see that openSCAD is otherwise working well with mimalloc.
The issue is that the recommended way to use mimalloc on Windows is through the visual C++ or clang that both generate standard windows executables using the universal C runtime that is part of the OS. The redirection dll (mimalloc-redirect.dll) is made to work in this setting.
Now, I have not tested with mingw64 and I really should; if I have some time I will look into it as it would be great to support it well. With the mingw64 toolchain, I believe it links by default to the old msvcrt dll's instead of ucrt.dll (I might be wrong here). I read somewhere about using -D_UCRT -lucrt as command line options when building? or gcc -mcrtdll=ucrt? Not sure.
With regard to console output -- this is just output from mimalloc right? This is because on Windows (since mingw sets -DWIN32) mimalloc uses direct-to-the-console output (_cputs) to avoid deadlock in the universal C runtime. A way around this is to call mi_register_output(&myoutput,NULL) at the start of main and define:
static void myoutput( const char* msg, void* arg ) {
printf(stderr, msg);
}for example.
to be continued; I'll try to set up mingw and see if I can see what is going on.
thehans
I have been working on replacing malloc with mimalloc for OpenSCAD on all our various supported platforms. So far there have been great improvements to performance on Linux and Mac, but I have not been able to get something fully working on Windows.
Most recently I attempted the recommended method of dynamic override / redirect with DLLs. Our builds which target Windows use MXE to cross compile from Linux This creates a mingw64/gcc based toolchain. Our application is built with cmake, with mimalloc included as a git submodule (using cmake's
add_subdirectory()) from our project's CMakeLists.txtIt successfully builds a DLL and links our application to it, but when I run the program, there are a number of issues:
1) No console output. I have tried setting
MIMALLOC_VERBOSE,MIMALLOC_SHOW_STATS, etc. and am unable to get any output from the mimalloc library on Windows (from any of cmd.exe, powershell or even git Bash from Git for Windows).2) Standard malloc is not redirected! (on wine at least) Here is a log from wine, which is the only way I was able to see any console messages from mimalloc:
3) I keep seeing references to
mimalloc-override.dll; In documentation, the warnings above, and various github issue comments, but cmake never generates such a file. In fact it was generating alibmimalloc.dllfor me, and when I ran the build I got a message which said something like "redirect: unable to find module". It may just be a unique consequence of us cross-compiling from linux, but I basically ended up adding this to our cmake, just beforeadd_subdirectory(mimalloc)Now it generates
mimalloc.dlland links to that, and instead of "unable to find module" I see the messages which I pasted above for point 2, but... does it actually need to be named:mimalloc-override.dllfor override to work?4) Despite the dynamic override/redirect not working, I still do a manual override of allocation functions specifically for GMP. When I run an interactive session, with Qt based GUI, it functions ok, but upon closing I often get a crash, pasted below. (This may just be a bug in Wine? I don't recall if I got similar crashes on native windows, I didn't do much testing there due to the lack of console output)