Message ID | 1341939617-26450-2-git-send-email-amartin@nvidia.com |
---|---|
State | Changes Requested |
Headers | show |
On Tue, Jul 10, 2012 at 10:00:17AM -0700, Allen Martin wrote: > Disable sibling call optimization based on binutils version. This is > to work around a bug in the assember in binutils versions < 2.22. > Branches to weak symbols can be incorrectly optimized in thumb mode to > a short branch (b.n instruction) that won't reach when the symbol gets > preempted. > > http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > Signed-off-by: Allen Martin <amartin@nvidia.com> > --- > changes for v2: > -changed GAS_BUG_12532 from yes/no to y/n to be consistent > -added additional warning about code size increase > --- > arch/arm/config.mk | 21 +++++++++++++++++++++ > 1 file changed, 21 insertions(+) > > diff --git a/arch/arm/config.mk b/arch/arm/config.mk > index 3f4453a..bcabf17 100644 > --- a/arch/arm/config.mk > +++ b/arch/arm/config.mk > @@ -87,3 +87,24 @@ endif > ifndef CONFIG_NAND_SPL > LDFLAGS_u-boot += -pie > endif > + > +# > +# binutils versions < 2.22 have a bug in the assembler where branches > +# to weak symbols can be incorrectly optimized in thumb mode to a > +# short branch (b.n instruction) that won't reach when the symbol > +# gets preempted > +# > +# http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > +# > +ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > +ifeq ($(GAS_BUG_12532),) > +export GAS_BUG_12532:=$(shell if [ $(call binutils-version) -lt 0222 ] ; then echo y; else echo n; fi) Please don't squash out whitespace, just break this out into a multi-line test ala the others in the toplevel config.mk. > +ifeq ($(GAS_BUG_12532),y) > +$(warning *** disabling sibling call optimzation because binutils version < 2.22) > +$(warning *** code size may be slightly larger) > +endif > +endif > +ifeq ($(GAS_BUG_12532),y) > +PLATFORM_RELFLAGS += -fno-optimize-sibling-calls > +endif > +endif This seems a bit longer than needed. Just check that it's y once and do both things. Also, I _think_ we should be able to squish this down to: ifeq ($(CONFIG_SYS_THUMB_BUILD),y) GAS_BUG_12532 := $(shell if [ $(call binutils-version) -lt 0222 ] ; \ then echo y; else echo n; fi) ifeq ($(GAS_BUG_125320),y) $(warning ...) $(warning ...) PLATFORM_RELFLAGS += ... endif endif But I haven't tested it.
On Tue, Jul 10, 2012 at 10:25:05AM -0700, Tom Rini wrote: > * PGP Signed by an unknown key > > On Tue, Jul 10, 2012 at 10:00:17AM -0700, Allen Martin wrote: > > Disable sibling call optimization based on binutils version. This is > > to work around a bug in the assember in binutils versions < 2.22. > > Branches to weak symbols can be incorrectly optimized in thumb mode to > > a short branch (b.n instruction) that won't reach when the symbol gets > > preempted. > > > > http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > > > Signed-off-by: Allen Martin <amartin@nvidia.com> > > --- > > changes for v2: > > -changed GAS_BUG_12532 from yes/no to y/n to be consistent > > -added additional warning about code size increase > > --- > > arch/arm/config.mk | 21 +++++++++++++++++++++ > > 1 file changed, 21 insertions(+) > > > > diff --git a/arch/arm/config.mk b/arch/arm/config.mk > > index 3f4453a..bcabf17 100644 > > --- a/arch/arm/config.mk > > +++ b/arch/arm/config.mk > > @@ -87,3 +87,24 @@ endif > > ifndef CONFIG_NAND_SPL > > LDFLAGS_u-boot += -pie > > endif > > + > > +# > > +# binutils versions < 2.22 have a bug in the assembler where branches > > +# to weak symbols can be incorrectly optimized in thumb mode to a > > +# short branch (b.n instruction) that won't reach when the symbol > > +# gets preempted > > +# > > +# http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > +# > > +ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > > +ifeq ($(GAS_BUG_12532),) > > +export GAS_BUG_12532:=$(shell if [ $(call binutils-version) -lt 0222 ] ; then echo y; else echo n; fi) > > Please don't squash out whitespace, just break this out into a > multi-line test ala the others in the toplevel config.mk. Ok > > +ifeq ($(GAS_BUG_12532),y) > > +$(warning *** disabling sibling call optimzation because binutils version < 2.22) > > +$(warning *** code size may be slightly larger) > > +endif > > +endif > > +ifeq ($(GAS_BUG_12532),y) > > +PLATFORM_RELFLAGS += -fno-optimize-sibling-calls > > +endif > > +endif > > This seems a bit longer than needed. Just check that it's y once and do > both things. Also, I _think_ we should be able to squish this down to: > ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > GAS_BUG_12532 := $(shell if [ $(call binutils-version) -lt 0222 ] ; \ > then echo y; else echo n; fi) > ifeq ($(GAS_BUG_125320),y) > $(warning ...) > $(warning ...) > PLATFORM_RELFLAGS += ... > endif > endif > > But I haven't tested it. I wrote it the way I did because I only want the expensive shell/binutils-version to get evaluated once per "make". I think what you proposed would cause it to get re-evaluated every time the build goes down into a subdir. -Allen
On Tue, Jul 10, 2012 at 10:44:33AM -0700, Allen Martin wrote: > On Tue, Jul 10, 2012 at 10:25:05AM -0700, Tom Rini wrote: > > * PGP Signed by an unknown key > > > > On Tue, Jul 10, 2012 at 10:00:17AM -0700, Allen Martin wrote: > > > Disable sibling call optimization based on binutils version. This is > > > to work around a bug in the assember in binutils versions < 2.22. > > > Branches to weak symbols can be incorrectly optimized in thumb mode to > > > a short branch (b.n instruction) that won't reach when the symbol gets > > > preempted. > > > > > > http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > > > > > Signed-off-by: Allen Martin <amartin@nvidia.com> > > > --- > > > changes for v2: > > > -changed GAS_BUG_12532 from yes/no to y/n to be consistent > > > -added additional warning about code size increase > > > --- > > > arch/arm/config.mk | 21 +++++++++++++++++++++ > > > 1 file changed, 21 insertions(+) > > > > > > diff --git a/arch/arm/config.mk b/arch/arm/config.mk > > > index 3f4453a..bcabf17 100644 > > > --- a/arch/arm/config.mk > > > +++ b/arch/arm/config.mk > > > @@ -87,3 +87,24 @@ endif > > > ifndef CONFIG_NAND_SPL > > > LDFLAGS_u-boot += -pie > > > endif > > > + > > > +# > > > +# binutils versions < 2.22 have a bug in the assembler where branches > > > +# to weak symbols can be incorrectly optimized in thumb mode to a > > > +# short branch (b.n instruction) that won't reach when the symbol > > > +# gets preempted > > > +# > > > +# http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > > +# > > > +ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > > > +ifeq ($(GAS_BUG_12532),) > > > +export GAS_BUG_12532:=$(shell if [ $(call binutils-version) -lt 0222 ] ; then echo y; else echo n; fi) > > > > Please don't squash out whitespace, just break this out into a > > multi-line test ala the others in the toplevel config.mk. > > Ok > > > > +ifeq ($(GAS_BUG_12532),y) > > > +$(warning *** disabling sibling call optimzation because binutils version < 2.22) > > > +$(warning *** code size may be slightly larger) > > > +endif > > > +endif > > > +ifeq ($(GAS_BUG_12532),y) > > > +PLATFORM_RELFLAGS += -fno-optimize-sibling-calls > > > +endif > > > +endif > > > > This seems a bit longer than needed. Just check that it's y once and do > > both things. Also, I _think_ we should be able to squish this down to: > > ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > > GAS_BUG_12532 := $(shell if [ $(call binutils-version) -lt 0222 ] ; \ > > then echo y; else echo n; fi) > > ifeq ($(GAS_BUG_125320),y) > > $(warning ...) > > $(warning ...) > > PLATFORM_RELFLAGS += ... > > endif > > endif > > > > But I haven't tested it. > > I wrote it the way I did because I only want the expensive > shell/binutils-version to get evaluated once per "make". I think what > you proposed would cause it to get re-evaluated every time the build > goes down into a subdir. It should still get evaluated once, I think. Kluding the binutils-version script to echo 1 >> /tmp/test every time it's run would tell tho for sure :)
On Tue, Jul 10, 2012 at 10:50:34AM -0700, Tom Rini wrote: > On Tue, Jul 10, 2012 at 10:44:33AM -0700, Allen Martin wrote: > > On Tue, Jul 10, 2012 at 10:25:05AM -0700, Tom Rini wrote: > > > * PGP Signed by an unknown key > > > > > > On Tue, Jul 10, 2012 at 10:00:17AM -0700, Allen Martin wrote: > > > > Disable sibling call optimization based on binutils version. This is > > > > to work around a bug in the assember in binutils versions < 2.22. > > > > Branches to weak symbols can be incorrectly optimized in thumb mode to > > > > a short branch (b.n instruction) that won't reach when the symbol gets > > > > preempted. > > > > > > > > http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > > > > > > > Signed-off-by: Allen Martin <amartin@nvidia.com> > > > > --- > > > > changes for v2: > > > > -changed GAS_BUG_12532 from yes/no to y/n to be consistent > > > > -added additional warning about code size increase > > > > --- > > > > arch/arm/config.mk | 21 +++++++++++++++++++++ > > > > 1 file changed, 21 insertions(+) > > > > > > > > diff --git a/arch/arm/config.mk b/arch/arm/config.mk > > > > index 3f4453a..bcabf17 100644 > > > > --- a/arch/arm/config.mk > > > > +++ b/arch/arm/config.mk > > > > @@ -87,3 +87,24 @@ endif > > > > ifndef CONFIG_NAND_SPL > > > > LDFLAGS_u-boot += -pie > > > > endif > > > > + > > > > +# > > > > +# binutils versions < 2.22 have a bug in the assembler where branches > > > > +# to weak symbols can be incorrectly optimized in thumb mode to a > > > > +# short branch (b.n instruction) that won't reach when the symbol > > > > +# gets preempted > > > > +# > > > > +# http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > > > +# > > > > +ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > > > > +ifeq ($(GAS_BUG_12532),) > > > > +export GAS_BUG_12532:=$(shell if [ $(call binutils-version) -lt 0222 ] ; then echo y; else echo n; fi) > > > > > > Please don't squash out whitespace, just break this out into a > > > multi-line test ala the others in the toplevel config.mk. > > > > Ok > > > > > > +ifeq ($(GAS_BUG_12532),y) > > > > +$(warning *** disabling sibling call optimzation because binutils version < 2.22) > > > > +$(warning *** code size may be slightly larger) > > > > +endif > > > > +endif > > > > +ifeq ($(GAS_BUG_12532),y) > > > > +PLATFORM_RELFLAGS += -fno-optimize-sibling-calls > > > > +endif > > > > +endif > > > > > > This seems a bit longer than needed. Just check that it's y once and do > > > both things. Also, I _think_ we should be able to squish this down to: > > > ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > > > GAS_BUG_12532 := $(shell if [ $(call binutils-version) -lt 0222 ] ; \ > > > then echo y; else echo n; fi) > > > ifeq ($(GAS_BUG_125320),y) > > > $(warning ...) > > > $(warning ...) > > > PLATFORM_RELFLAGS += ... > > > endif > > > endif > > > > > > But I haven't tested it. > > > > I wrote it the way I did because I only want the expensive > > shell/binutils-version to get evaluated once per "make". I think what > > you proposed would cause it to get re-evaluated every time the build > > goes down into a subdir. > > It should still get evaluated once, I think. Kluding the > binutils-version script to echo 1 >> /tmp/test every time it's run would > tell tho for sure :) I tried it out, it's obvious from the $(warning) it's getting reevaluated: $ ./MAKEALL -s tegra20 /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger Configuring for harmony board... text data bss dec hex filename 170334 4284 278352 452970 6e96a ./u-boot 11676 161 36 11873 2e61 ./spl/u-boot-spl /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger ... and on and on It makes sense since unless GAS_BUG_125320 is "export"ed to the new make, it's going to be unset again each the build starts a new "make -C" -Allen
On Tue, Jul 10, 2012 at 11:04:50AM -0700, Allen Martin wrote: > On Tue, Jul 10, 2012 at 10:50:34AM -0700, Tom Rini wrote: > > On Tue, Jul 10, 2012 at 10:44:33AM -0700, Allen Martin wrote: > > > On Tue, Jul 10, 2012 at 10:25:05AM -0700, Tom Rini wrote: > > > > * PGP Signed by an unknown key > > > > > > > > On Tue, Jul 10, 2012 at 10:00:17AM -0700, Allen Martin wrote: > > > > > Disable sibling call optimization based on binutils version. This is > > > > > to work around a bug in the assember in binutils versions < 2.22. > > > > > Branches to weak symbols can be incorrectly optimized in thumb mode to > > > > > a short branch (b.n instruction) that won't reach when the symbol gets > > > > > preempted. > > > > > > > > > > http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > > > > > > > > > Signed-off-by: Allen Martin <amartin@nvidia.com> > > > > > --- > > > > > changes for v2: > > > > > -changed GAS_BUG_12532 from yes/no to y/n to be consistent > > > > > -added additional warning about code size increase > > > > > --- > > > > > arch/arm/config.mk | 21 +++++++++++++++++++++ > > > > > 1 file changed, 21 insertions(+) > > > > > > > > > > diff --git a/arch/arm/config.mk b/arch/arm/config.mk > > > > > index 3f4453a..bcabf17 100644 > > > > > --- a/arch/arm/config.mk > > > > > +++ b/arch/arm/config.mk > > > > > @@ -87,3 +87,24 @@ endif > > > > > ifndef CONFIG_NAND_SPL > > > > > LDFLAGS_u-boot += -pie > > > > > endif > > > > > + > > > > > +# > > > > > +# binutils versions < 2.22 have a bug in the assembler where branches > > > > > +# to weak symbols can be incorrectly optimized in thumb mode to a > > > > > +# short branch (b.n instruction) that won't reach when the symbol > > > > > +# gets preempted > > > > > +# > > > > > +# http://sourceware.org/bugzilla/show_bug.cgi?id=12532 > > > > > +# > > > > > +ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > > > > > +ifeq ($(GAS_BUG_12532),) > > > > > +export GAS_BUG_12532:=$(shell if [ $(call binutils-version) -lt 0222 ] ; then echo y; else echo n; fi) > > > > > > > > Please don't squash out whitespace, just break this out into a > > > > multi-line test ala the others in the toplevel config.mk. > > > > > > Ok > > > > > > > > +ifeq ($(GAS_BUG_12532),y) > > > > > +$(warning *** disabling sibling call optimzation because binutils version < 2.22) > > > > > +$(warning *** code size may be slightly larger) > > > > > +endif > > > > > +endif > > > > > +ifeq ($(GAS_BUG_12532),y) > > > > > +PLATFORM_RELFLAGS += -fno-optimize-sibling-calls > > > > > +endif > > > > > +endif > > > > > > > > This seems a bit longer than needed. Just check that it's y once and do > > > > both things. Also, I _think_ we should be able to squish this down to: > > > > ifeq ($(CONFIG_SYS_THUMB_BUILD),y) > > > > GAS_BUG_12532 := $(shell if [ $(call binutils-version) -lt 0222 ] ; \ > > > > then echo y; else echo n; fi) > > > > ifeq ($(GAS_BUG_125320),y) > > > > $(warning ...) > > > > $(warning ...) > > > > PLATFORM_RELFLAGS += ... > > > > endif > > > > endif > > > > > > > > But I haven't tested it. > > > > > > I wrote it the way I did because I only want the expensive > > > shell/binutils-version to get evaluated once per "make". I think what > > > you proposed would cause it to get re-evaluated every time the build > > > goes down into a subdir. > > > > It should still get evaluated once, I think. Kluding the > > binutils-version script to echo 1 >> /tmp/test every time it's run would > > tell tho for sure :) > > I tried it out, it's obvious from the $(warning) it's getting > reevaluated: > > $ ./MAKEALL -s tegra20 > /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 > /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger > Configuring for harmony board... > text data bss dec hex filename > 170334 4284 278352 452970 6e96a ./u-boot > 11676 161 36 11873 2e61 ./spl/u-boot-spl > /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 > /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger > /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 > /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger > /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 > /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger > /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 > /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger > /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 > /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger > /home/arm/u-boot/arch/arm/config.mk:102: *** disabling sibling call optimzation because binutils version < 2.22 > /home/arm/u-boot/arch/arm/config.mk:103: *** code size may be slightly larger > ... and on and on > > > It makes sense since unless GAS_BUG_125320 is "export"ed to the new > make, it's going to be unset again each the build starts a new > "make -C" Thanks for proving me wrong. Just the white-space / multi line thing please, thanks :)
diff --git a/arch/arm/config.mk b/arch/arm/config.mk index 3f4453a..bcabf17 100644 --- a/arch/arm/config.mk +++ b/arch/arm/config.mk @@ -87,3 +87,24 @@ endif ifndef CONFIG_NAND_SPL LDFLAGS_u-boot += -pie endif + +# +# binutils versions < 2.22 have a bug in the assembler where branches +# to weak symbols can be incorrectly optimized in thumb mode to a +# short branch (b.n instruction) that won't reach when the symbol +# gets preempted +# +# http://sourceware.org/bugzilla/show_bug.cgi?id=12532 +# +ifeq ($(CONFIG_SYS_THUMB_BUILD),y) +ifeq ($(GAS_BUG_12532),) +export GAS_BUG_12532:=$(shell if [ $(call binutils-version) -lt 0222 ] ; then echo y; else echo n; fi) +ifeq ($(GAS_BUG_12532),y) +$(warning *** disabling sibling call optimzation because binutils version < 2.22) +$(warning *** code size may be slightly larger) +endif +endif +ifeq ($(GAS_BUG_12532),y) +PLATFORM_RELFLAGS += -fno-optimize-sibling-calls +endif +endif
Disable sibling call optimization based on binutils version. This is to work around a bug in the assember in binutils versions < 2.22. Branches to weak symbols can be incorrectly optimized in thumb mode to a short branch (b.n instruction) that won't reach when the symbol gets preempted. http://sourceware.org/bugzilla/show_bug.cgi?id=12532 Signed-off-by: Allen Martin <amartin@nvidia.com> --- changes for v2: -changed GAS_BUG_12532 from yes/no to y/n to be consistent -added additional warning about code size increase --- arch/arm/config.mk | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+)