Since process_attr/3 searches source code for attributes, it can
happen that it finds an attribute which is eventually not needed
by the compilation, e.g. hidden by ifdef macro - see enclosed test.
Such attribute can reference a file which is not in the path or
which even doesn't exist at all, and current code doesn't expect
such a situation. We fix things by simply ignoring such files.
These times already contain max modified info of their
dependencies. Therefore we no longer have to check in
internal_erl_compile whether the file really needs
recompiling, which simplifies the flow somewhat, because
the work with dependency graph is now localized to much
smaller space then before.
The precise algorithm is now following:
- Decide which files need to be compiled first (ErlFirstFiles).
- Split AllErlFiles which are not in ErlFirstFiles into two groups,
one consisting of files some other file depends on (DepErls)
and the rest (OtherErls).
- Final FirstErls are obtained as ErlFirstFiles (in original order)
plus toposorted DepErls. OtherErls are left as is since they can
be compiled in any (random) order.
Also create the erlcinfo graph as acyclic, since otherwise the
toposort could not work and we don't want to have cycles among
dependencies anyway.
Optimizing for best compression could be very slow for large
projects, so optimize rather for the speed. erlcinfo file
isn't that huge anyway and moreover difference between levels
2 and 9 isn't that big in practice (although there's quite big
difference between level 2 and no compression at all).
When the source file changes it could happen that some of its
dependecies get removed. In that case we should remove these
former dependencies from the graph, so that they don't influence
recompilation of the source file anymore.
Current separation of part of the logic into include_path
obscures the purpose of Dirs (see expand_file_names).
Moreover for each particular Erl its source file was included
twice in Dirs, which is now corrected.
Also InclDirs (specified in erl_opts) as part of erlcinfo
since they can affect resulting graph, which needs to be
therefore regenerated when InclDirs change. See added
test as an example.
As erlcinfo is only an internal optimization mechanism, user
should be only minimally bothered by issues with it. Especially
it shouldn't ever cause a fatal error and if it cannot be restored
properly, only a warning (not error) message should be emitted (if
any). Also it probably doesn't make sense for this warning message
to be too detailed - just state that something went wrong and
silently delete the erlcinfo file.
Expand test cases to handle .appup.src with
empty upgrade from and downgrade to lists, lift
larger .appup.src file from OTP's
lib/dialyzer/test/r9c_SUITE_data/src/asn1/asn1.appup.src
If I pass an expression like: "$PWD/deps/local/lib" in the port env
string, it expands to "/deps/local/lib", (variable is expanded to the
empty string) but if I pass "${PWD}/deps/local/lib" it expands
properly. I found that confusing.
This was because we require environment vars to end with whitespace,
while I think requiring a non-word character would be sufficient, since
the variables can only contain word characters.
This changes the expansion system to recognize variables that are
terminated by non word characters.
This includes some fixes as per some comments by tuncer.
Fixesrebar/rebar#457
Similar to how .app.src are being handled, all files in directory
.src ending with .appup.src are processed, checked for
correctness and copied to ebin without the .src extension.
By way of using rebar_config:consult_file/1, .appup.src.script
files are automatically supported in exactly the same manner
as in .app.src.script.
Remove possible last dot from deps dir path by calling
dirname on base_dir, deps_dir and dummy path part.
With this fix delete-deps command deletes dependencies
properly in case of deps dir path ended with dot.
Examples of deps_dir config value that triggers error:
".", "some/path/.".
Code that triggers error: rebar_deps:'delete-deps'/2,
code line lists:prefix(DepsDir, D#dep.dir).
If deps_dir = ".", cwd = "/root", dependency name = "app"
then DepsDir = "/root/." and D#dep.dir = "/root/app" so
prefix fun returns false but should be true.
Dependency app will not be removed because of this.
proto_opts config option contains the compiler directive
that defines which proto compiler to use, also contains a
src_dirs entry that defines a list of locations for the
.proto files to be processed.
Add integration test that compiles .proto files from src
and from proto directory specified in separate rebar
config files
For protocol buffer files, when there were gpb options to alter the
module name with prefix or suffix, recompilation was not properly
detected. This is now fixed. (Issue #384).
Use the rebar_base_compiler's ability to specify both source and
target file names, to be able to also support prefixes.
This also introduces a call to gpb_compile:format_error, so the xref
recipe needs to be updated to ignore it, to avoid false errors.
The rebar_base_compiler:run/5 is an alternative to run/8, where
instead of source/target dirs and file name extensions, it takes
a list of tuples: [{SourceFileName, TargetFileName}], after the
idea by Tuncer Ayaz in discussion with Luis Rascão and me.
This is useful for situations where the target file name cannot be
expressed as only a new suffix or extension. For example the gpb
protocol buffer compiler can take an option to add a file name prefix
as well as a suffix.
As a result, the values passed around inside the internal functions in
rebar_base_compiler are either source file names (eg. when the run/4
function has been called), or items, which are two-tuples
{SourceFileName,TargetFileName} (when the run/7,8 or /5 functions have
been called).
{recursive_cmds, []} was already documented as part of the core
rebar.config options that gets printed via 'rebar -h', but this wasn't
sufficiently informative. To fix this, explain the use of
recursive_cmds as part of the -r help string.
Reported-by: Stefan Strigler
* Refactor reltool version check to use vsn app key if ebin/reltool.app
exists.
OTP releases prior to R14 didn't install the app file, so we still
have to support getting the version string from the lib_dir path. If
you're using an R13B03 or R13B04 install that doesn't have version
strings in lib_dir paths (lib/reltool vs lib/reltool-0.5.3), we won't
be able to retrieve the version string. Given R13's age and that it's
unusual to have vsn-less lib_dir paths, this shouldn't be a problem.
* While at it, delete trailing white space.
This reverts commit 81063d33f0, reversing
changes made to 6584defdda.
This was reverted because it changed the gpb compiler to
not use rebar_base_compiler. That change was to add support for
target name prefixes, but it is better to add that support to
rebar_base_compiler.
