cmake: refined conditions for math library linking on windows

[Xarbirus]

I encountered some problems when building llama.cpp on windows (for x64 and arm64) on a system where MSVC (with clang) and mingw are installed at the same time.

It seems to me that the condition for adding MATH_LIBRARY to the build is not quite correct. MATH_LIBRARY should not be added if

• Intel oneAPI is already used (since these libraries include optimized math libraries and tools)
• or the MSVC compiler is used

An important addition - in case clang is used in the MSVC pipeline, we should not add the library (condition CMAKE_CXX_SIMULATE_ID STREQUAL "MSVC" was added for this). At the same time, when building with mingw, the library can be added.

I think this would also be a solution to this issue (if it hadn't closed).

[slaren]

I encountered some problems when building llama.cpp on windows (for x64 and arm64) on a system where MSVC (with clang) and mingw are installed at the same time.

Can you suggest a way to reproduce the issue that you are seeing? Does it happen with a specific version of mingw? I have mingw installed alongside MSVC, and the github runners do as well, so at least it doesn't happen in every case.

[Xarbirus]

Hi @slaren in my case, the problem occurs when 3 compilers are installed on the system at once: msvc, mingw and clang.
I use vcvarsall.bat to set environment variables, and compile the project with the clang compiler.
The problem looks like this:

[51/134] Linking CXX executable bin\Release\llama-gguf.exe
FAILED: bin/Release/llama-gguf.exe
cmd.exe /C "cd . && C:\PROGRA~1\LLVM\bin\CLANG_~1.EXE --target=x86_64-pc-windows-msvc -fuse-ld=lld-link -nostartfiles -nostdlib -fms-compatibility  -fvectorize -ffp-model=fast -fno-finite-math-only -Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments -O3 -DNDEBUG -fms-runtime-lib=static -Xlinker /subsystem:console examples/gguf/CMakeFiles/llama-gguf.dir/Release/gguf.cpp.obj -o bin\Release\llama-gguf.exe -Xlinker /implib:examples\gguf\Release\llama-gguf.lib -Xlinker /pdb:bin\Release\llama-gguf.pdb -Xlinker /version:0.0   ggml/src/Release/ggml.lib  ggml/src/Release/ggml-cpu.lib  ggml/src/Release/ggml-base.lib  -lm.lib  -lkernel32 -luser32 -lgdi32 -lwinspool -lshell32 -lole32 -loleaut32 -luuid -lcomdlg32 -ladvapi32 -loldnames && cd ."
lld-link: error: could not open 'm.lib': no such file or directory

In my understanding the problem occurs because cmake finds libraries related to mingw (m.lib) because the path to mingw is in the path environment variable.

[Xarbirus]

Hi @slaren! I would like to clarify if there are any plans to merge this pr?

[slaren]

I was not able to reproduce the issue that you mentioned. I have all 3 compilers installed, and so does the github CI. I am not comfortable merging something that I don't understand.

[Xarbirus]

Okay, thanks, I'll try to dig deeper into our CI.