This patch adds a config option for COMBINE_COMPILE, adds the __global
tag to compiler.h and makes EXPORT_SYMBOL automatically use it. It also
contains a crappy Makefile hack which uses -fwhole-program --combine for
all multi-obj .o files -- anything which was
obj-m := foo.o
foo-objs := bar.o wibble.o
... should now use gcc -o foo.o --combine -fwhole-program bar.c wibble.c
The makefile hack is known not to work properly for generated C files in
out-of-source-tree builds, and for multi-obj files where one of the
sources is _assembly_ instead of C. It's good enough for the proof of
concept, until someone more clueful can do it properly though. It would
be useful to make built-in.o build with --combine from _everything_
which uses standard CFLAGS, rather than doing just what I've done here.
For example only; do not apply as-is.
diff --git a/include/asm-i386/linkage.h b/include/asm-i386/linkage.h
index f4a6eba..8b06e3e 100644
--- a/include/asm-i386/linkage.h
+++ b/include/asm-i386/linkage.h
@@ -1,7 +1,7 @@
#ifndef __ASM_LINKAGE_H
#define __ASM_LINKAGE_H
-#define asmlinkage CPP_ASMLINKAGE __attribute__((regparm(0)))
+#define asmlinkage CPP_ASMLINKAGE __attribute__((regparm(0))) __global
#define FASTCALL(x) x __attribute__((regparm(3)))
#define fastcall __attribute__((regparm(3)))
diff --git a/include/linux/compiler.h b/include/linux/compiler.h
index 9b4f110..ed4cf0c 100644
--- a/include/linux/compiler.h
+++ b/include/linux/compiler.h
@@ -156,4 +156,10 @@ #ifndef __attribute_const__
# define __attribute_const__ /* unimplemented */
#endif
+#ifdef CONFIG_COMBINED_COMPILE
+#define __global __attribute__((externally_visible,used))
+#else
+#define __global
+#endif
+
#endif /* __LINUX_COMPILER_H */
diff --git a/include/linux/linkage.h b/include/linux/linkage.h
index 932021f..d78eec9 100644
--- a/include/linux/linkage.h
+++ b/include/linux/linkage.h
@@ -10,7 +10,7 @@ #define CPP_ASMLINKAGE
#endif
#ifndef asmlinkage
-#define asmlinkage CPP_ASMLINKAGE
+#define asmlinkage __global CPP_ASMLINKAGE
#endif
#ifndef prevent_tail_call
diff --git a/include/linux/module.h b/include/linux/module.h
index 0dfb794..725e5df 100644
--- a/include/linux/module.h
+++ b/include/linux/module.h
@@ -175,7 +175,7 @@ #ifdef CONFIG_MODVERSIONS
#define __CRC_SYMBOL(sym, sec) \
extern void *__crc_##sym __attribute__((weak)); \
static const unsigned long __kcrctab_##sym \
- __attribute_used__ \
+ __attribute_used__ __global \
__attribute__((section("__kcrctab" sec), unused)) \
= (unsigned long) &__crc_##sym;
#else
@@ -184,13 +184,15 @@ #endif
/* For every exported symbol, place a struct in the __ksymtab section */
#define __EXPORT_SYMBOL(sym, sec) \
- extern typeof(sym) sym; \
+ extern typeof(sym) sym __global; \
__CRC_SYMBOL(sym, sec) \
static const char __kstrtab_##sym[] \
+ asm("__kstrtab_" #sym) \
__attribute__((section("__ksymtab_strings"))) \
= MODULE_SYMBOL_PREFIX #sym; \
static const struct kernel_symbol __ksymtab_##sym \
- __attribute_used__ \
+ asm("__ksymtab_" #sym) \
+ __attribute_used__ __global \
__attribute__((section("__ksymtab" sec), unused)) \
= { (unsigned long)&sym, __kstrtab_##sym }
@@ -208,8 +210,8 @@ #ifdef CONFIG_UNUSED_SYMBOLS
#define EXPORT_UNUSED_SYMBOL(sym) __EXPORT_SYMBOL(sym, "_unused")
#define EXPORT_UNUSED_SYMBOL_GPL(sym) __EXPORT_SYMBOL(sym, "_unused_gpl")
#else
-#define EXPORT_UNUSED_SYMBOL(sym)
-#define EXPORT_UNUSED_SYMBOL_GPL(sym)
+#define EXPORT_UNUSED_SYMBOL(sym) extern typeof(sym) sym __global
+#define EXPORT_UNUSED_SYMBOL_GPL(sym) extern typeof(sym) sym __global
#endif
#endif
@@ -470,11 +472,12 @@ void module_add_driver(struct module *,
void module_remove_driver(struct device_driver *);
#else /* !CONFIG_MODULES... */
-#define EXPORT_SYMBOL(sym)
-#define EXPORT_SYMBOL_GPL(sym)
-#define EXPORT_SYMBOL_GPL_FUTURE(sym)
-#define EXPORT_UNUSED_SYMBOL(sym)
-#define EXPORT_UNUSED_SYMBOL_GPL(sym)
+
+#define EXPORT_SYMBOL(sym) extern typeof(sym) sym __global
+#define EXPORT_SYMBOL_GPL(sym) EXPORT_SYMBOL(sym)
+#define EXPORT_SYMBOL_GPL_FUTURE(sym) EXPORT_SYMBOL(sym)
+#define EXPORT_UNUSED_SYMBOL(sym) EXPORT_SYMBOL(sym)
+#define EXPORT_UNUSED_SYMBOL_GPL(sym) EXPORT_SYMBOL(sym)
/* Given an address, look for it in the exception tables. */
static inline const struct exception_table_entry *
diff --git a/init/Kconfig b/init/Kconfig
index a099fc6..b19cfd1 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -468,6 +468,16 @@ config MODULE_FORCE_UNLOAD
rmmod). This is mainly for kernel developers and desperate users.
If unsure, say N.
+config COMBINED_COMPILE
+ bool "Compile multiple source files at once"
+ help
+ GCC can perform optimisation between different C files by
+ using its '--combine' option to compile them all at once. By
+ also adding the '-fwhole-program' option to make global symbols
+ static, further optimisations are made possible. Say N unless you
+ have a version of GCC in which at least PR27898, PR28706, PR28712,
+ PR28744 and PR28779 are fixed.
+
config MODVERSIONS
bool "Module versioning support"
depends on MODULES
diff --git a/scripts/Makefile.build b/scripts/Makefile.build
index 3cb445c..2f2075d 100644
--- a/scripts/Makefile.build
+++ b/scripts/Makefile.build
@@ -128,6 +128,12 @@ modname = $(basetarget)
$(multi-objs-y:.o=.s) : modname = $(modname-multi)
$(multi-objs-y:.o=.lst) : modname = $(modname-multi)
+ifdef CONFIG_COMBINED_COMPILE
+$(multi-used-m) : modkern_cflags += $(CFLAGS_MODULE) -fwhole-program --combine $(sort $(addprefix $(srctree)/$(obj)/,$($(subst $(obj)/,,$(@:.o=-y)):.o=.c) $($(subst $(obj)/,,$(@:.o=-objs)):.o=.c)))
+
+$(multi-used-y) : CFLAGS += -fwhole-program --combine $(sort $(addprefix $(srctree)/$(obj)/,$($(subst $(obj)/,,$(@:.o=-y)):.o=.c) $($(subst $(obj)/,,$(@:.o=-objs)):.o=.c)))
+endif
+
quiet_cmd_cc_s_c = CC $(quiet_modtag) $@
cmd_cc_s_c = $(CC) $(c_flags) -fverbose-asm -S -o $@ $<
@@ -283,6 +289,23 @@ cmd_link_l_target = rm -f $@; $(AR) $(EX
targets += $(lib-target)
endif
+ifdef CONFIG_COMBINED_COMPILE
+# We would rather have a list of rules like
+# foo.o: $(foo-objs)
+# but that's not so easy, so we rather make all composite objects depend
+# on the set of all their parts
+
+$(multi-used-y) : %.o: $(srctree)/dummy.c $(multi-objs-y:.o=.c) FORCE
+ $(call cmd,force_checksrc)
+ $(call if_changed_rule,cc_o_c)
+
+$(multi-used-m) : %.o: $(srctree)/dummy.c $(multi-objs-m:.o=.c) FORCE
+ $(call cmd,force_checksrc)
+ $(call if_changed_rule,cc_o_c)
+ @{ echo $(@:.o=.ko); echo $(link_multi_deps); } > $(MODVERDIR)/$(@F:.o=.mod)
+
+targets += $(multi-used-y) $(multi-used-m)
+else
#
# Rule to link composite objects
#
@@ -313,7 +336,7 @@ # on the set of all their parts
@{ echo $(@:.o=.ko); echo $(link_multi_deps); } > $(MODVERDIR)/$(@F:.o=.mod)
targets += $(multi-used-y) $(multi-used-m)
-
+endif
# Descending
# ---------------------------------------------------------------------------
--
dwmw2
On Thu, 2006-08-24 at 16:26 +0100, David Woodhouse wrote:
> This patch adds a config option for COMBINE_COMPILE, adds the __global
> tag to compiler.h and makes EXPORT_SYMBOL automatically use it. It also
> contains a crappy Makefile hack which uses -fwhole-program --combine for
> all multi-obj .o files -- anything which was
> obj-m := foo.o
> foo-objs := bar.o wibble.o
> ... should now use gcc -o foo.o --combine -fwhole-program bar.c wibble.c
>
> The makefile hack is known not to work properly for generated C files in
> out-of-source-tree builds, and for multi-obj files where one of the
> sources is _assembly_ instead of C. It's good enough for the proof of
> concept, until someone more clueful can do it properly though. It would
> be useful to make built-in.o build with --combine from _everything_
> which uses standard CFLAGS, rather than doing just what I've done here.
Would the generation and use of preprocessed source files ($x.c -> $x.i)
help solve the "standard CFLAGS" problem here, and allow a subsequent
single compilation of the entire kernel (sans modules) with --combine
-fwhole-program?
- Josh Triplett
On Thu, 2006-08-24 at 10:27 -0700, Josh Triplett wrote:
> Would the generation and use of preprocessed source files ($x.c -> $x.i)
> help solve the "standard CFLAGS" problem here, and allow a subsequent
> single compilation of the entire kernel (sans modules) with --combine
> -fwhole-program?
Not for stuff like CFLAGS_raid6altivec1.o := -maltivec -mabi-altivec
Stuff which is _purely_ #defines, such as...
ncr53c8xx-flags-$(CONFIG_SCSI_ZALON) \
:= -DCONFIG_NCR53C8XX_PREFETCH -DSCSI_NCR_BIG_ENDIAN \
-DCONFIG_SCSI_NCR53C8XX_NO_WORD_TRANSFERS
CFLAGS_ncr53c8xx.o := $(ncr53c8xx-flags-y) $(ncr53c8xx-flags-m)
... can probably be done differently anyway.
It's not entirely unreasonable to treat files with their own CFLAGS just
as we have to treat .S files -- compile them separately, then link.
--
dwmw2
On Thu, Aug 24, 2006 at 04:26:07PM +0100, David Woodhouse wrote:
>...
> index 0dfb794..725e5df 100644
> --- a/include/linux/module.h
> +++ b/include/linux/module.h
>...
> @@ -470,11 +472,12 @@ void module_add_driver(struct module *,
> void module_remove_driver(struct device_driver *);
>
> #else /* !CONFIG_MODULES... */
> -#define EXPORT_SYMBOL(sym)
> -#define EXPORT_SYMBOL_GPL(sym)
> -#define EXPORT_SYMBOL_GPL_FUTURE(sym)
> -#define EXPORT_UNUSED_SYMBOL(sym)
> -#define EXPORT_UNUSED_SYMBOL_GPL(sym)
> +
> +#define EXPORT_SYMBOL(sym) extern typeof(sym) sym __global
> +#define EXPORT_SYMBOL_GPL(sym) EXPORT_SYMBOL(sym)
> +#define EXPORT_SYMBOL_GPL_FUTURE(sym) EXPORT_SYMBOL(sym)
> +#define EXPORT_UNUSED_SYMBOL(sym) EXPORT_SYMBOL(sym)
> +#define EXPORT_UNUSED_SYMBOL_GPL(sym) EXPORT_SYMBOL(sym)
>...
If a "build everything except for assembler files at once" approach is
possible, it should be possible to revert this and get even further
savings.
> dwmw2
cu
Adrian
--
Gentoo kernels are 42 times more popular than SUSE kernels among
KLive users (a service by SUSE contractor Andrea Arcangeli that
gathers data about kernels from many users worldwide).
There are three kinds of lies: Lies, Damn Lies, and Statistics.
Benjamin Disraeli
On Thu, 2006-08-24 at 23:33 +0200, Adrian Bunk wrote:
> If a "build everything except for assembler files at once" approach is
> possible, it should be possible to revert this and get even further
> savings.
Only if we build _everything_ at once, which may take an insane amount
of RAM. Doing it a directory at a time makes a certain amount of sense,
and tends to combine the most incestuous code -- although maybe
combinations like building arch/$ARCH/kernel/ with kernel/ (and likewise
mm) could be an interesting experiment.
I suspected that most of the 'further savings' to which you refer above
could be achieved more easily with -ffunction-sections -fdata-sections
--gc-sections
--
dwmw2
On Fri, Aug 25, 2006 at 10:37:23AM +0100, David Woodhouse wrote:
> On Thu, 2006-08-24 at 23:33 +0200, Adrian Bunk wrote:
> > If a "build everything except for assembler files at once" approach is
> > possible, it should be possible to revert this and get even further
> > savings.
>
> Only if we build _everything_ at once, which may take an insane amount
> of RAM. Doing it a directory at a time makes a certain amount of sense,
> and tends to combine the most incestuous code -- although maybe
> combinations like building arch/$ARCH/kernel/ with kernel/ (and likewise
> mm) could be an interesting experiment.
My hope is "insane" would be something like "1 GB of RAM" that is no
longer insane on current computers. [1]
> I suspected that most of the 'further savings' to which you refer above
> could be achieved more easily with -ffunction-sections -fdata-sections
> --gc-sections
AFAIR -ffunction-sections/-fdata-sections cause some overhead in the
resulting binary?
> dwmw2
cu
Adrian
[1] The interesting cases are embedded systems needing a small kernel
that gets built on a much bigger system.
Whether this should be the default compile mode for everyone is a
different issue.
--
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon Seed
On Fri, 2006-08-25 at 12:30 +0200, Adrian Bunk wrote:
> My hope is "insane" would be something like "1 GB of RAM" that is no
> longer insane on current computers. [1]
Maybe -- but still, that's quite a steep requirement, and would also
require a fairly major redesign of the current kbuild architecture.
I'm inclined to stick with per-directory builds for now, which is both
relatively simple to implement and where the _majority_ of the benefit
it likely to be seen.
Whole-kernel optimisation is something I'm inclined to leave until LTO
happens -- but don't let me stop you from investigating it.
> > I suspected that most of the 'further savings' to which you refer above
> > could be achieved more easily with -ffunction-sections -fdata-sections
> > --gc-sections
>
> AFAIR -ffunction-sections/-fdata-sections cause some overhead in the
> resulting binary?
Not that I'm aware of. There's overhead in the resulting ELF files, but
we combine sections in the vmlinux.lds so there isn't even that overhead
in the vmlinux.
--
dwmw2