This reverts commit a1fe671db8.
Lue suggested reverting #661 because ZLS worked around the issue of @src
being relative in that environment: https://github.com/zigtools/zls/pull/898
This is not a perfect solution (what zls did seems to be a workaround), but
is good enough for us until Zig gets an official package manager.
* all: build: fix sdkPath for relative @src.file
Prior to this commit, the build system heavily assumed that the result
`@src.file` would always be absolute, but this is no longer
guaranteed, likely due to there being no such thing as an "absolute
path" in WASI.
It appears that for normal invocations of `zig build`, it is safe to
assume that `@src.file` is absolute. However, when ZLS uses a custom
`build_runner.zig` to collect build configuration, `@src.file` is
actually relative to the current working directory, at least on my
system. For a while, this led to ZLS completions breaking entirely,
but presently it actually causes ZLS to crash!
The solution is not as simple as using relative `sdkPath` results
as-is, because the build system may attempt to resolve these paths
relative to build root, when the paths are actually relative to the
current working directory.
This leads to a sticky situation: the current working directory is a
runtime concept, but `@src.file` is resolved at compile time. However,
it appears that the build runner does not change current working
directory in between compilation and execution, so it is probably safe
to calculate `sdkPath` using runtime current working directory.
Still, this requires major changes with how `sdkPath` works, since
runtime computation and allocations are required. So pretty much
anything that relied on `sdkPath` being comptime-known has been
refactored in this commit.
The most severe result of this is that, for example, `gpu.pkg` can no
longer be a comptime-known constant: it has to be a runtime function
that takes a `*Builder` and returns a `Pkg`.
This commit deals with usages of `*.pkg` and `sdkPath` within Mach
itself, but projects that depend on Mach such as `mach-examples` will
almost certainly require changes as well.
* all: update README to reflect change in pkg usage
For details on updating your code to use this version, see: 88b1106953
Signed-off-by: Stephen Gutekanst <stephen@hexops.com>
Co-authored-by: Stephen Gutekanst <stephen@hexops.com>
This declaration itself was nonsensical, as the file being imported was
from the harfbuzz package (instead users should import harfbuzz
directly). But as well as being nonsensical, this line was occasionally
triggering a stage2 bug which we haven't quite tracked down yet which
seems to be something to do with conflicting modules importing a file.
This issue meant that projects with a specific dependency pattern on
freetype and harfbuzz would sometimes randomly get a nonsensical
compilation error. This change works around that issue for
mach-freetype.
With almost all tests/examples working on all platforms now with the new compiler,
https://github.com/hexops/mach/issues/180, it's time to remove stage1 support.
Signed-off-by: Stephen Gutekanst <stephen@hexops.com>
