On 2020-05-12 10:08, Magnus Ihse Bursie wrote:
On 2020-05-12 17:38, Erik Joelsson wrote:
On 2020-05-12 08:29, Magnus Ihse Bursie wrote:
We've broken our golden rule in SetupJavaCompilation, that all
command lines should be copy/paste:able to re-execute them from the
command line. The reason is that the file containing the source code
file names has a temporary name, which is renamed after the
compilation. This is just a messy way to use the file both as input
to javac and as a marker that the compilation has succeeded. By
splitting up these two roles, we get more readable code and a
re-executable command line.
Bug: https://bugs.openjdk.java.net/browse/JDK-8244844
WebRev:
http://cr.openjdk.java.net/~ihse/JDK-8244844-make-javac-re-executable/webrev.01
I really like seeing this fixed, but not sure about this patch. By
moving the filelist to a separate rule and only having the source
files as prereq, there will be cases where the source files
timestamps have not changed, but the list of files may have. I think
this would be safely covered if you moved VARDEPS_FILE as prereq to
FILELIST.
I see your point. Will do that.
I was also a bit divided about whether the actual javac invocation
rule should keep $$($1_SRCS) as a dependency. Technically, it's not
needed, since changes in the sources will result in the filelist being
updated, but maybe the intention will be better expressed. And if the
set of files has not changed, then we will write the same files back
to the filelist, which effectively just updates the timestamp in a
roundabout way. If we keep the list of source files as dependency, we
are free to implement a future optimization where the filelist is not
updated if the set of files has not changed.
Otoh, I don't know if a huge list of dependencies (the number of
sources files for e.g. java.base is non-trivial!) affects make
performance, so listing it twice just to make it "look good" maybe
isn't a good idea.
Do you have any opinion?
In general I like to list prereqs if they logically are direct prereqs.
The sources are used as input in the compilation recipe, so that would
be a direct prereq. If a prereq is really only transitive, then I don't
want to list it. I don't know what the performance impact would be, I'm
guessing it's not even measurable, but if it is, then maybe save some
processing by not adding the sources to both rules. No strong opinion in
this case.
/Erik
/Magnus
/Erik
