On Sat, Jul 16, 2016 at 3:13 PM, Patrick Palka <patr...@parcs.ath.cx> wrote:
> On Sat, 16 Jul 2016, Patrick Palka wrote:
>
>> This patch makes the 0.5GB libbackend.a get built via "ar rcs" instead
>> of via "ar rc" + "ranlib", if possible. The problem with the latter is
>> that it's about twice as slow as the former and it makes libbackend.a
>> get written twice which is wasteful. On my machine this optimization
>> reduces rebuild time in a --disable-bootstrap tree after touching tree.c
>> by around 25%, from ~28s to ~21s. This optimization is only performed
>> when the current ar is GNU ar where we can be reasonably sure that the
>> "rcs" directive is supported.
>>
>> Does this change look reasonable? I am currently bootstrapping it on
>> x86_64-pc-linux-gnu.
>>
>> gcc/ChangeLog:
>>
>> * configure.ac (ar_is_gnu): New variable. AC_SUBST it.
>> * configure: Regenerate.
>> * Makefile.in (AR_IS_GNU): New variable.
>> (USE_AR_RCS): New variable.
>> (libbackend.a): When building this archive, if USE_AR_RCS then
>> just invoke "ar rcs" instead of invoking "ar rc" followed by
>> "ranlib".
>> ---
>> gcc/Makefile.in | 18 ++++++++++++++++++
>> gcc/configure | 15 +++++++++++++--
>> gcc/configure.ac | 8 ++++++++
>> 3 files changed, 39 insertions(+), 2 deletions(-)
>>
>> diff --git a/gcc/Makefile.in b/gcc/Makefile.in
>> index 0786fa3..5ae6100 100644
>> --- a/gcc/Makefile.in
>> +++ b/gcc/Makefile.in
>> @@ -237,10 +237,22 @@ FLEX = @FLEX@
>> FLEXFLAGS =
>> AR = @AR@
>> AR_FLAGS = rc
>> +AR_IS_GNU = @ar_is_gnu@
>> NM = @NM@
>> RANLIB = @RANLIB@
>> RANLIB_FLAGS = @ranlib_flags@
>>
>> +# Whether invocations of ar + ranlib can be replaced by a single
>> +# invocation of "ar rcs".
>> +USE_AR_RCS = no
>> +ifeq ($(AR_IS_GNU),yes)
>> +ifeq ($(AR_FLAGS),rc)
>> +ifeq ($(RANLIB_FLAGS),)
>> +USE_AR_RCS = yes
>> +endif
>> +endif
>> +endif
>> +
>> # Libraries to use on the host.
>> HOST_LIBS = @HOST_LIBS@
>>
>> @@ -1882,8 +1894,14 @@ compilations: $(BACKEND)
>> # This archive is strictly for the host.
>> libbackend.a: $(OBJS)
>> -rm -rf libbackend.a
>> + @# Elide the invocation of ranlib if possible, as doing so
>> + @# significantly reduces the build time of this archive.
>> +ifeq ($(USE_AR_RCS),yes)
>> + $(AR) $(AR_FLAGS)s libbackend.a $(OBJS)
>> +else
>> $(AR) $(AR_FLAGS) libbackend.a $(OBJS)
>> -$(RANLIB) $(RANLIB_FLAGS) libbackend.a
>> +endif
>>
>> libcommon-target.a: $(OBJS-libcommon-target)
>> -rm -rf libcommon-target.a
>> diff --git a/gcc/configure b/gcc/configure
>> index ed44472..db761c8 100755
>> --- a/gcc/configure
>> +++ b/gcc/configure
>> @@ -749,6 +749,7 @@ CXXDEPMODE
>> DEPDIR
>> am__leading_dot
>> doc_build_sys
>> +ar_is_gnu
>> AR
>> NM
>> BISON
>> @@ -8459,6 +8460,16 @@ fi
>>
>> fi
>>
>> +{ $as_echo "$as_me:${as_lineno-$LINENO}: checking whether ar is GNU ar" >&5
>> +$as_echo_n "checking whether ar is GNU ar... " >&6; }
>> +ar_is_gnu=no
>> +if $AR --version 2>/dev/null | sed 1q | grep "GNU ar" > /dev/null; then
>> + ar_is_gnu=yes
>> +fi
>> +{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ar_is_gnu" >&5
>> +$as_echo "$ar_is_gnu" >&6; }
>> +
>> +
>> # The jit documentation looks better if built with sphinx, but can be
>> # built with texinfo if sphinx is not available.
>> # Set "doc_build_sys" to "sphinx" or "texinfo" accordingly.
>> @@ -18475,7 +18486,7 @@ else
>> lt_dlunknown=0; lt_dlno_uscore=1; lt_dlneed_uscore=2
>> lt_status=$lt_dlunknown
>> cat > conftest.$ac_ext <<_LT_EOF
>> -#line 18478 "configure"
>> +#line 18489 "configure"
>> #include "confdefs.h"
>>
>> #if HAVE_DLFCN_H
>> @@ -18581,7 +18592,7 @@ else
>> lt_dlunknown=0; lt_dlno_uscore=1; lt_dlneed_uscore=2
>> lt_status=$lt_dlunknown
>> cat > conftest.$ac_ext <<_LT_EOF
>> -#line 18584 "configure"
>> +#line 18595 "configure"
>> #include "confdefs.h"
>>
>> #if HAVE_DLFCN_H
>> diff --git a/gcc/configure.ac b/gcc/configure.ac
>> index 086d0fc..3749b21 100644
>> --- a/gcc/configure.ac
>> +++ b/gcc/configure.ac
>> @@ -1081,6 +1081,14 @@ else
>> AC_CHECK_PROG(AR, ar, ar, ${CONFIG_SHELL-/bin/sh} ${srcdir}/../missing ar)
>> fi
>>
>> +AC_MSG_CHECKING(whether ar is GNU ar)
>> +ar_is_gnu=no
>> +if $AR --version 2>/dev/null | sed 1q | grep "GNU ar" > /dev/null; then
>> + ar_is_gnu=yes
>> +fi
>> +AC_MSG_RESULT($ar_is_gnu)
>> +AC_SUBST(ar_is_gnu)
>> +
>> # The jit documentation looks better if built with sphinx, but can be
>> # built with texinfo if sphinx is not available.
>> # Set "doc_build_sys" to "sphinx" or "texinfo" accordingly.
>> --
>> 2.9.2.321.g3e7e728
>>
>>
>
> Even better, why build libbackend.a at all? Why not just pass the list
> of object files directly to the linker instead of going through an
> intermediate archive? The following patch cuts rebuild time by 50%,
> from 28s to 14s by making libbackend.a no longer get built. (The
> Make-lang.in hunk is apparently necessary to avoid duplicate-reference
> link errors since incpath.o is already listed in $(OBJS)).
There is an issue with this in that the number of files in OBJS can
increase to a large number.
Really we should be testing if response files (@file) support exist
and use that instead.
Thanks,
Andrew
>
> diff --git a/gcc/Makefile.in b/gcc/Makefile.in
> index 5ae6100..7b8ec16 100644
> --- a/gcc/Makefile.in
> +++ b/gcc/Makefile.in
> @@ -1591,7 +1591,7 @@ endif
> # compilation or not.
> ALL_HOST_OBJS = $(ALL_HOST_FRONTEND_OBJS) $(ALL_HOST_BACKEND_OBJS)
>
> -BACKEND = libbackend.a main.o libcommon-target.a libcommon.a \
> +BACKEND = $(OBJS) main.o libcommon-target.a libcommon.a \
> $(CPPLIB) $(LIBDECNUMBER)
>
> # This is defined to "yes" if Tree checking is enabled, which roughly means
> diff --git a/gcc/cp/Make-lang.in b/gcc/cp/Make-lang.in
> index cc3337d..d09b449 100644
> --- a/gcc/cp/Make-lang.in
> +++ b/gcc/cp/Make-lang.in
> @@ -67,7 +67,7 @@ g++-cross$(exeext): xg++$(exeext)
>
> # The compiler itself.
> # Shared with C front end:
> -CXX_C_OBJS = attribs.o incpath.o \
> +CXX_C_OBJS = attribs.o \
> $(C_COMMON_OBJS) $(CXX_TARGET_OBJS)
>
> # Language-specific object files for C++ and Objective C++.
