Peter Dimov

It's relevant here because I want to figure out what CMakeLists.txt file we want on the master branch. On develop, if I understand correctly, you want a combination of this...

Yes it's the right thing to do.

Even more interesting is `u1.segments() = u1.segments();`, which should do the same (that is, remove any unnecessary percent encoding.) Not sure about `u1.segments().back() = u1.segments().back();`. Does this work?

`configs/user-config-boost-1_70_0-common.jam` uses `%ARCH%` as the Clang version, which is not numeric and is therefore ignored by the mangling. That's why the toolset is encoded as just `clang` and not `clang8`...

No, just remove `%ARCH`, it should autodetect the version. `clang8` is not a valid numeric version either.

I don't know what versions of MSVC work correctly and what do not; which one did you test with when you added the workaround?

Turns out, it's pretty involved. The "normal" (not Clang) implementation works under VS2017 in both [C++14](https://godbolt.org/z/YzWjnfKzz) and [C++17](https://godbolt.org/z/ocMacEvMx) modes. It also works with VS2019 in [C++14](https://godbolt.org/z/o4fYP353o) and [C++20](https://godbolt.org/z/6hqYWen4Y) mode. It...

This is a problem in clang-win.jam that has been fixed in https://github.com/boostorg/build/commit/2c00d909f4f6c6faf7dce3d7359376e5fe753d67 on develop, but the fix has not yet been merged to master. There's an additional issue with clang-win...

`target-os` is `cygwin` though. ``` C:\boost-git\develop\libs\thread>b2 test toolset=gcc Performing configuration checks - default address-model : 32-bit (cached) - default architecture : x86 (cached) - default address-model : 32-bit (cached) -...

Would moving the threadapi detection to default-build fix it?