1. If when generating something, we find more that one suitable generators,
run them and more then one return something, immediately report ambiguity.
Don't care if the produced targets are the same. This is better that
running several generators all the time, performance wise.
2. Remove the notion of 'intermediate' virtual-targets. IIRC, they were used
to prevent staging of RSP files, and we don't stage them anyway now.
[SVN r29491]
When a library is present in sources of a static library, return it
to dependents via usage requirements, not by adding it to the list
of created targets.
[SVN r29362]
* If a generator was given a source it could not handle, it used to return
that source together with generated targets. This was nice for some use
cases, but no very nice for others, and this behaviour could not be turned
off. One use case where it worked bad was:
lib plugin : plugin.cpp helper ;
lib helper : helper.cpp ;
On windows, 'plugin' would link to the 'import library' and pass the DLL
target though. So, when installing 'plugin', we'd also install 'helper.dll',
and it was not possible to do anything about it.
* If we asked generators.construct to produce sources of type CPP,
and the selected generator produced both targets of type CPP, and of
some other type, we'd try again to convert those other targets to CPP.
This simply complicated the logic for no good reason.
* Most generator function had 'multiple' parameter, which function
was long forgotten by anybody.
As a bit of history, I believe some of the above decisions were due to a
certain use case:
CPP <------- WHL
\
WD
/
CPP <------- DLP
Here, a source file is converted to two targets with one command, and each
produced file is converted to CPP. Our generators search would notice that
there are two generators for CPP: the WHL->CPP and DPL->CPP
generators. Neither is better that the other so both are tried, and produce
(CPP, DPL) and (CPP, WHL) pairs of targets. To avoid reporting an ambiguity,
we'd try to convert, DLP to CPP and WHL to CPP, do it successfully, notice
that produced targets are the same and decide that there's no ambiguity.
However, this is rather complex logic for a relatively rare case. It can
be handled by writing another WD->CPP generator that would handle
disambiguation itself.
This commit has one user-visible change. The code:
exe a : a.cpp b ;
obj b : b.cpp helper;
lib helper ;
No longer works -- the 'a' target won't link to 'helper'. However, this is
pretty stange code and worked before almost by accident.
[SVN r29361]
Otherwise, the added path elements won't be seen when loading user-config.jam
and site-config.jam and we'll get error when those files are present only
in V2 directories.
[SVN r29360]
Not sure why it worked this way in the first place. The problem is that
after inheriting parent rules no longer have access to variables in original
module, like "RM", and no longer work.
[SVN r29325]
I think that Dave committed then unintentionally in revision 1.36 and they
cause C++ files to be always recompiled. A better version is in the works
but cannot be committed yet, so I'll just restore the previous one for now.
[SVN r29198]
