| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
| |
The Linux kernel has started using statements like
-Wa,macros.s for including assembler with the code.
Handle these similar to the `-imacros` statement,
and send them to the compile server for processing.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
2010-08-18 Simon Baldwin <simonb@google.com>
* include_server/compiler_defaults.py (_SystemSearchdirsGCC):
Add a 'sysroot' parameter. Add --sysroot, if needed, to the gcc
invocation that identifies search directories and default paths.
Add debugging output.
(SetSystemDirsDefaults): Add 'sysroot' parameter. Add another
dictionary level to system_dirs_default for sysroot values.
Add debugging output.
* include_server/parse_command.py (ParseCommandArgs): Identify
and handle arguments of the form '--opt=value'. Pass any
--sysroot value to compiler_defaults.SetSystemDirsDefaults, use
it in compiler_defaults.system_dirs_default lookups.
(CPP_OPTIONS_APPEARING_AS_ASSIGNMENTS): New.
* include_server/parse_command_test.py (Mock_SetSystemDirsDefaults):
Add sysroot parameter, verify it on function calls.
(ParseCommandUnitTest): Add --sysroot to mock gcc invocations.
|
| |
|
|
|
| |
so it wasn't testing anything useful.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
gvn --project https://distcc.googlecode.com/svn/trunk review klarlund/d3b3
Fix Issue 7: Compiling with -I/usr/include/FOO or ... in pump mode
Problem:
Default system include directories are the directories on the default include
search path, such as /usr/include, that is determined by the compiler. Distcc
will not send default system include directories to the distcc compilation
servers. Nevertheless, distcc on the server blindly rewrites -I options such
as -I/usr/include/foo to -I/tmp/distccNNN/usr/include/foo.
This doesn't work, since the files in /usr/include/foo are not sent to the
distccd server.
Solution:
The present solution keeps the rewriting on the server, because we would like
to not manage starting the compiler, parsing its output, and storing the
default system paths on the server
Instead, we use the existing mechanism for defining relative symbolic links
under the server root. We escape from the root by using a sufficient number
of '../'s.
All this is tremendously complicated by:
-- The possibility that other links encountered may point into the system
default dirs in which case the usual mirroring of the reachable places
should not take place, because the routing of such links will go through
the link created for the system directory.
-- The fact that determination of default-system-dirness is lazy: a
deeply-seated link in a default system dir may become obsolete if it is
later determined that a higher directory than the parent directory of the
link is in fact also a default system dir. In that case, a new symlink,
sitting in a place higher than the previous one will need to be created.
Tests:
make check
benchmarks: samba (still 3X faster than with distcc), linux 2.6 kernel, and
a couple of others
TODO: better testing
TODO:
> In compiler_defaults.py, _MakeLinkFromMirrorToRealLocation:
> Maybe comment each of the 'if' cases with an example of how this case
> might be triggered. eg the real_prefix == rooted_system_dir is
> triggered when we see /usr/include/foo, and the later see
> /usr/include.
|
| |
|
|
|
|
|
|
|
| |
This change is mostly reverted -- some variable renamings are not
changed back. A subsequent CL proposes a less arbitrary way to solve
the problem of -I<some_dir_in_default_place_where_compiler_looks>,
which involves sending no system default header files at all.
Tests: make pump-check and make include-server-check
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add a new piece of command line parse information: send_systemdirs. This
boolean is normally false. It is true if there is a -isystem option with a
default system directory (one known to the compiler).
With send_systemdirs true, the compiler headers of system headers are sent to
the servers and mounted under the server root. The isystem option is as usual
rewritten to be relative to the root. Without this flag setting, distcc quickly
decides that pump mode is not viable because remote compilations fail.
Also, fixed the comment:
# TODO(klarlund): Make mechanism for handling -U, -undef options, along with
# default symbols.
(-D options are handled.)
---------------
TESTS: make pump-maintainer-check
make include-server-maintainer-check
make 'linux kernel' in a directory made by the benchmark script (as shown below)
DISTCC_EMAILLOG_WHOM_TO_BLAME=klarlund@google.com DISTCC_ENABLE_DISCREPANCY_EMAIL=1 INCLUDE_SERVER_ARGS='-d1 -t --email --unsafe_absolute_includes --send_systemdirs --stat_reset_triggers=include/linux/*' DISTCC_HOSTS="--randomize `/home/klarlund/svn-distcc/distcc/lsdistcc -P3 -k100 -d`" PATH=/home/klarlund/svn-distcc/distcc:$PATH ./masquerade pump make -j120
REVIVIEWER: csilvers@google.com
TRIVIA: the about 1000 files in the kernel build involve on average 400
headers. The build time on a quad-cpu machine drops from about 1m40 s to 1m20s.
With pump-mode, the machine is underutilized judging from the total sys + user
time, which is around only 2m20s.
NOTE: this is a successor to 'g6n8', which simply introduced a command line
option to force sending of system files. A bug in gvn did not allow me to
resuse that change list.
|
| |
|
|
|
|
|
|
| |
(typically the only change is in the FSF street address). Add Google
copyright line in some places it was missing. Add GPLv2 notice to
avahi patches.
OKed by fergus
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This required the following changes:
- Compute the default search paths for each language lazily,
when the language is first encountered, rather than
computing it for 'c' and 'c++'.
- Make the tests of Objective C and Objective C++
conditional on the relevant compiler being installed.
Tested by "make check" on two systems:
- One with Objective C installed, but not Objective C++.
- One with neither installed.
I didn't test on a system with Objective C++ installed,
because I didn't have one easily available.
Reviewers: Craig Silverstein, Nils Klarlund.
|
| |
|
