Fix commit e40d2cb3af (Xcode: Add embed resources support, 2023-07-31,
v3.28.0-rc1~281^2). The implementation should not name the `_PATH`
suffix explicitly. That variant is automatically handled by
`cmGlobalXCodeGenerator::AddEmbeddedObjects`.
Makefiles do not have a per-object sense of where they come from, so
forwarding any module information here would end up with incorrect
module file path construction by consuming targets. Leave a TODO item in
its place.
Fortran modules provided by objects added as linked items via
`$<TARGET_OBJECTS>` should also be considered as "linked targets" for
collation purposes. As C++ modules have their own visibility rules
through their `FILE_SET` feature, do not expose these for C++ module
collation.
When a target uses objects from another target which provides modules as
sources, the modules provided by the referenced target must also be
treated as if they were provided by the referencing target. Add the
concept of "forwarding" modules so that consumers can use modules
created by these sources as well.
Note that this is only sensible for Fortran where module usages are
implicit as far as CMake's visibility model is concerned. C++ modules
have their own concept of visibility which does not require or support
such `$<TARGET_OBJECTS>` reuse in this way.
When the link was updated to `https` by commit 52eac4573d (Help: Fix
link to cdash.org from CTest manual, 2021-04-21, v3.21.0-rc1~262^2~2)
the markup was incorrectly adjusted to show the link as part of the
"See Also" section. It is meant to be the link destination for links
in prose elsewhere in the manual. Fix the markup and move it to a
clearer location. Also update the link to resolve a redirect.
Revert commit b6a5382217 (Ninja: depend on language module information
files directly, 2023-02-10, v3.27.0-rc1~502^2) from !8197.
This reverts the "exact dependencies" for collation inputs and returns
to "get all targets" and target-ordering. Use of exact dependencies
caused a parade of use cases that had not been tested previously to be
found and need fixing over the 3.27 release series. To stop the flow on
3.27, revert to the 3.26 strategy. We will continue in 3.28.
Note that this is a restoration of 3.26 semantics where incremental
rebuilds may be subtly incorrect in the presence of stale
`<LANG>Modules.json` files. However, since C++ support is experimental
and Fortran has always had this problem as of 3.27, it is not considered
a regression.
See: #25112
See: #25123
See: #25252
See: #25365
See: #25417
See: #25425
Targets only using Fortran modules via `$<TARGET_OBJECTS>` also need a
collation step to be performed. Check for this case and trigger the
depends rule to be used.
Targets only using Fortran modules via `$<TARGET_OBJECTS>` also need a
collation step to be performed. Check for this case and trigger the
collation rule to be added and used.
Fortran modules provided by objects added as sources via
`$<TARGET_OBJECTS>` should also be considered as "linked targets" for
collation purposes. As C++ modules have their own visibility rules
through their `FILE_SET` feature, do not expose these for C++ module
collation.
This will be eventually be used to inform the collator of this
information so that Fortran modules provided by the resulting objects
can also be used as intended.
This avoids having to do manual "is already present" checks. The order
the targets are processed does not need to be preserved because the
resulting `languages` result is already a `set`.
* 'revert-exact-collation-depends-3.27' (early part):
Tests/FortranModules: add test for TARGET_OBJECTS-as-linked-items module usage
Tests/FortranModules: add test for TARGET_OBJECTS-as-sources module usage
When `clang-scan-deps` fails to scan (e.g., bad source syntax, junk
flags, etc.), the redirection unconditionally updates the file. If this
fails, the `.ddi` file timestamp is updated. If the state is then
reverted (e.g., the command line returns to the state of the last
successful build), the updated file is not useful, but `ninja` does not
rerun because:
- the command hash matches the last successful run
- the output file is newer than its inputs
However, since the `.ddi` file has been updated with bogus contents from
a failed scan, collation fails as the `rules` array is empty (or
incomplete from a batch scan).
If `clang-scan-deps` were properly aware of its output file, it could
use this to not write the file if any inner scan fails. Requested in
https://github.com/llvm/llvm-project/issues/72875.
See: https://github.com/llvm/llvm-project/issues/72875Fixes: #25429
halx99