So we have a more recent version of SPIRV-Tools and also don't
have to recompile Mesa to test llvmpipe. This fixes a few failing
tests, but also breaks a couple.
In theory commit 7b07d77396 already
did that, but in practice it ended up picking a commit from the
1.8.2405 branch, where a few of our tests fail. Since I hope to
soon enable DXC for macOS again, it's useful to fix that oversight.
For the build-mingw-32 and build-mingw-64 jobs. These currently end up
picking up Linux pkg-config instead of MinGW pkg-config, which in turn
causes configure to pick up Linux OpenGL and XCB. We happen to get away
with that at the moment because none of the code using HAVE_OPENGL or
HAVE_XCB ends up getting built for Windows, but that's about to change.
Specifically, we'd like to build the vkd3d versions of the demos for
Windows.
With this we can finally disallow failure for the macOS CI script,
which is more valuable than checking DXC. Eventually DXC tests
will have to be fixed, though.
ERR is used to indicate internal inconsistencies in vkd3d. Here that's
not the case, we simply have to forward the error condition to the
caller.
This fixes failures on the CI with llvmpipe, because the build we use is
compiled without support for VK_KHR_surface and related extensions.
Unfortuantely different versions of DXC accept or fail differently
in some cases. We don't care too much about validating the DXC
behavior itself, but it's useful that all the CI jobs use the
same version so that we don't have to complicate the shader runner
language.
The macOS version is currently bound to be pretty recent because
otherwise libdxil.dylib is not provided. So I'm updating the
Linux and Windows version as well.
I don't expect it should be particularly hard for other maintainers
to keep up with the DXC updates, since it just amounts to
downloading a ZIP file and extracting two libraries.
This unfortunately introduces a lot of failures, because it turns
out there is still work to do there. But at least we can
estimate how bad we're doing.
