Open nbuwe opened 1 year ago
As a side note, precompiled header that is forced to be included with -include
on the command line makes #pragma interface
in the source files invalid (hence the ifdef around them). this makes all inline functions from the headers to be emitted in every object file. For the RelWithDebInfo (i.e. optimized) this gives about 4MB Self binary, but the object files total 83MB of code, i.e. about 95% is duplicates.
If I don't fix up the precompiled header rule manually and so the precompiled header is not used, the slowdown is not too bad - about 4% user time (1% wall time in my -j4 build).
Would be interesting to test with proper include hygiene, but that requires a lot of menial changes, so not a quick test to try.
I have tried to compile with #pragma implementation
in effect and it reduces object file sizes and compilation time quite a bit. On Linux with -j4 a RelWithDebInfo
build with manually fixed pch rule goes:
$ tail -3 build-reldbg/.log
real 20m35.995s
user 72m48.945s
sys 2m45.777s
$ size build-reldbg/vm/Self
text data bss dec hex filename
4287395 38344 151148 4476887 444fd7 build-reldbg/vm/Self
$ ll build-reldbg/vm/Self
-rwxrwxr-x 1 uwe uwe 373488596 Aug 12 01:54 build-reldbg/vm/Self*
$ find build-reldbg/vm/CMakeFiles/Self.dir -name '*.o' | xargs size -t | tail -1
87429670 36053 150638 87616361 538eb69 (TOTALS)
vs.
$ tail -3 build-reldbg-pragma/.log
real 6m18.492s
user 21m9.525s
sys 1m5.130s
$ size build-reldbg-pragma/vm/Self
text data bss dec hex filename
4397634 38344 151148 4587126 45fe76 build-reldbg-pragma/vm/Self
$ ll build-reldbg-pragma/vm/Self
-rwxrwxr-x 1 uwe uwe 91423604 Aug 12 23:55 build-reldbg-pragma/vm/Self*
$ find build-reldbg-pragma/vm/CMakeFiles/Self.dir -name '*.o' | xargs size -t | tail -1
21322576 36053 150638 21509267 1483493 (TOTALS)
To do this test I have added a bunch of #include "_precompiled.hh"
to the cpp files and dropped -include
for it from the command line. This might be easier to do by including _precompiled.hh
in the generated _<file>.cpp.incl
files, but it would have taken me longer to figure out how to do that.
I haven't yet checked where the 110K text size increase comes from. But otherwise this looks like something worth pursuing.
Ah, the build sets -fkeep-inline-functions
for gcc. And I used to wonder, why clang was so lightning fast... Dropping that flag brings the build time down to just a few minutes
real 1m49.889s
user 5m48.046s
sys 0m34.977s
and object files are now tiny.
add_pch_rule
invm/cmake/functions.cmake
doesn't correctly add build type specific flags to the rule that produces the precompiled header. this makes the compiler complain later on, when the precompiled header is used, that the flags used to create the header and the flags it is used with are different and that the precompiled header will be ignored.The flags that are missing are stuff from
COMPILE_DEFINITIONS_RELEASE
and friends set invm/cmake/configurations.cmake
(likeGENERATE_DEBUGGING_AIDS
, etc)I don't have any cmake-fu to fix this. I figure it needs
gather_flags
but my naive attemtps fail with "get_property could not find TARGET .../self/vm. Perhaps it has not yet been created."