On Sat, Aug 27, 2016 at 8:26 PM, Don Lewis <truck...@apache.org> wrote:
> On 17 Aug, To: dev@openoffice.apache.org wrote: > > On 16 Aug, Don Lewis wrote: > >> On 7 Aug, Damjan Jovanovic wrote: > >>> Hi > >>> > >>> branches/gbuild-reintegration has been merged to trunk. It has > 136 > >>> patches and makes many changes throughout the code, so you probably > need to > >>> "dmake clean" before rebuilding. > >>> > >>> I haven't fixed the Windows build performance regression caused by > issue > >>> #117845, so if you're building on Windows use --disable-pch which only > >>> extends your build time by around 2 hours instead of 6. If you aren't > happy > >>> with this, please help me figure out the problem. > >> > >> I did some sleuthing ... > >> > >> I compared build logs between --enable-pch and --disable-pch builds. The > >> compiler command line for the --enable-pch build has the additional > >> option > >> -DPRECOMPILED_HEADERS > >> which causes precompiled_foo.hxx to pull in a whole wad of headers even > >> if they aren't needed to compile an individual .cxx file. That's why > >> the build is taking longer. A precompiled version of > >> precompiled_foo.hxx is build beforehand by passing > >> -Ycprecompiled_foo.hxx to the compiler, but gbuild neglects to tell the > >> compiler to use it when building all the .cxx files because it never > >> passes -Yuprecompiled_foo.hxx to the compiler. > >> > >> That option is supposed to come from > >> gb_PrecompiledHeader_get_enableflags or > >> gb_NoexPrecompiledHeader_get_enableflags in LinkTarget.mk, but for some > >> reason that isn't getting activated. > >> > >> I don't see any obvious recent changes that might have broken this. > > > > More digging ... > > > > LinkTarget.mk contains some logic to that only enables the use of > > precompiled headers when the compiler flags that were used to build the > > precompiled header match the flags used to compile the target source > > file. If I compare the compiler flags used to build the precompiled > > header with the flags used to compile one of the source files, I see > > these differences: > > > > @@ -33,7 +33,6 @@ > > -D_DLL_ > > -DPRECOMPILED_HEADERS > > -DSW_DLLIMPLEMENTATION > > --DSW_DLLIMPLEMENTATION > > -Gd > > -GR > > -Gs > > @@ -78,13 +77,13 @@ > > -wd4826 > > -Zc:wchar_t- > > -Zm500 > > +-DEXCEPTIONS_ON > > +-EHa > > -Ob1 > > -Oxs > > -Oy- > > --DEXCEPTIONS_ON > > --EHa > > --Fd$W/LinkTarget/pdb/Library/isw.lib.pdb > > --I$S/sw/inc/pch/ > > +-Fd > > +-I$S/sw/source/core/access/ > > -IC:/aoo/aoo-trunk/main/solver/420/wntmsci12.pro/inc/stl > > -I$S/sw/source/core/inc > > -I$S/sw/source/filter/inc > > @@ -108,7 +107,5 @@ > > -I$O/inc/udkapi > > -I$O/inc/offapi > > -c > > -$S/sw/inc/pch/precompiled_sw.cxx > > --Ycprecompiled_sw.hxx > > --Fp$W/PrecompiledHeader/nodebug/precompiled_sw.hxx.pch > > --Fo$W/PrecompiledHeader/nodebug/precompiled_sw.hxx.pch.obj > > +$S/sw/source/core/access/acccell.cxx > > +-Fo$W/CxxObject/sw/source/core/access/acccell.o > > > > The potentially important differences are: > > > > * When building the precompiled header, -DSW_DLLIMPLEMENTATION is > > repeated > > > > * These flags are used when building the precompiled header: > > -Fd$W/LinkTarget/pdb/Library/isw.lib.pdb -I$S/sw/inc/pch/ > > whereas these flags are used when building the target: > > -Fd -I$S/sw/source/core/access/ > > > > Another observation is that if an override of the compiler flags for > > a particular target disables the use of precompiled headers, then > > -DPRECOMPILED_HEADERS should not be passed to the compiler if -Yu is > > not. That would avoid the pessimization of including all of those extra > > headers that are not used. > > The answer turns out to be none of the above. I didn't find the cause > for the duplication of -DSW_DLLIMPLEMENTATION, but it doesn't matter > because the duplicate is squashed by the comparison which uses > $(sort ...). I don't think the latter difference is part of the > comparison. > > The cause turns out to be that $(PCH_NAME) is not inherited by the > compile, so the -Yu compile flag is not used. I've fixed this in > r1758061. > > Doing "build --all -P2 -- -P2" on a two-core machine now takes: > 377 minutes with --disable-pch > 278 minutes with --enable-pch > > Fantastic work Don, thank you so much! Damjan
