diff mbox

arm64: Add support ARCH_SUPPORTS_INT128

Message ID 1466492123-3356-1-git-send-email-wangkefeng.wang@huawei.com
State Accepted
Commit 6b14c517a2a866c8407e9864c1cd3fcc6fed55ab
Headers show

Commit Message

Kefeng Wang June 21, 2016, 6:55 a.m. UTC 
The gcc support __SIZEOF_INT128__ and __int128 in arm64, thus,
enable ARCH_SUPPORTS_INT128 to make mul_u64_u32_shr() a bit
more efficient in scheduler.

Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>

---
 arch/arm64/Kconfig | 1 +
 1 file changed, 1 insertion(+)

-- 
1.7.12.4


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Comments

Catalin Marinas June 21, 2016, 3:14 p.m. UTC | #1 
On Tue, Jun 21, 2016 at 02:55:23PM +0800, Kefeng Wang wrote:
> The gcc support __SIZEOF_INT128__ and __int128 in arm64, thus,

> enable ARCH_SUPPORTS_INT128 to make mul_u64_u32_shr() a bit

> more efficient in scheduler.

> 

> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>


Queued for 4.8. Thanks.

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Catalin Marinas June 22, 2016, 10:57 a.m. UTC | #2 
On Tue, Jun 21, 2016 at 02:55:23PM +0800, Kefeng Wang wrote:
> The gcc support __SIZEOF_INT128__ and __int128 in arm64, thus,

> enable ARCH_SUPPORTS_INT128 to make mul_u64_u32_shr() a bit

> more efficient in scheduler.

> 

> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>

> ---

>  arch/arm64/Kconfig | 1 +

>  1 file changed, 1 insertion(+)

> 

> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig

> index 5a0a691..ab319eb 100644

> --- a/arch/arm64/Kconfig

> +++ b/arch/arm64/Kconfig

> @@ -11,6 +11,7 @@ config ARM64

>  	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST

>  	select ARCH_USE_CMPXCHG_LOCKREF

>  	select ARCH_SUPPORTS_ATOMIC_RMW

> +	select ARCH_SUPPORTS_INT128

>  	select ARCH_SUPPORTS_NUMA_BALANCING

>  	select ARCH_WANT_OPTIONAL_GPIOLIB

>  	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION


With this patch and UBSAN+KASAN enabled, Linux fails to link with:

lib/built-in.o: In function `get_signed_val':
lib/ubsan.c:93: undefined reference to `__ashlti3'
lib/ubsan.c:93: undefined reference to `__ashrti3'

It succeeds if I revert commit d67703a8a69e ("arm64: kill off the libgcc
dependency"). We may have to revisit that decision or implement the
required library functions in the kernel. In general, I'd like to reduce
the amount of code duplication if the functionality can be found
elsewhere like in libgcc. There is a risk of not knowing precisely what
ligbcc relies on (like libc functions), though there are other
architectures linking against it.

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Will Deacon June 22, 2016, 12:43 p.m. UTC | #3 
On Wed, Jun 22, 2016 at 11:57:50AM +0100, Catalin Marinas wrote:
> On Tue, Jun 21, 2016 at 02:55:23PM +0800, Kefeng Wang wrote:

> > The gcc support __SIZEOF_INT128__ and __int128 in arm64, thus,

> > enable ARCH_SUPPORTS_INT128 to make mul_u64_u32_shr() a bit

> > more efficient in scheduler.

> > 

> > Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>

> > ---

> >  arch/arm64/Kconfig | 1 +

> >  1 file changed, 1 insertion(+)

> > 

> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig

> > index 5a0a691..ab319eb 100644

> > --- a/arch/arm64/Kconfig

> > +++ b/arch/arm64/Kconfig

> > @@ -11,6 +11,7 @@ config ARM64

> >  	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST

> >  	select ARCH_USE_CMPXCHG_LOCKREF

> >  	select ARCH_SUPPORTS_ATOMIC_RMW

> > +	select ARCH_SUPPORTS_INT128

> >  	select ARCH_SUPPORTS_NUMA_BALANCING

> >  	select ARCH_WANT_OPTIONAL_GPIOLIB

> >  	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION

> 

> With this patch and UBSAN+KASAN enabled, Linux fails to link with:

> 

> lib/built-in.o: In function `get_signed_val':

> lib/ubsan.c:93: undefined reference to `__ashlti3'

> lib/ubsan.c:93: undefined reference to `__ashrti3'

> 

> It succeeds if I revert commit d67703a8a69e ("arm64: kill off the libgcc

> dependency"). We may have to revisit that decision or implement the

> required library functions in the kernel. In general, I'd like to reduce

> the amount of code duplication if the functionality can be found

> elsewhere like in libgcc. There is a risk of not knowing precisely what

> ligbcc relies on (like libc functions), though there are other

> architectures linking against it.


Linking agianst libgcc is pretty scary if we're starting to use plugins
for the compiler[1], but IANAL.

Will

[1] https://lwn.net/Articles/691102/

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Ard Biesheuvel June 22, 2016, 1:23 p.m. UTC | #4 
On 22 June 2016 at 12:57, Catalin Marinas <catalin.marinas@arm.com> wrote:
> On Tue, Jun 21, 2016 at 02:55:23PM +0800, Kefeng Wang wrote:

>> The gcc support __SIZEOF_INT128__ and __int128 in arm64, thus,

>> enable ARCH_SUPPORTS_INT128 to make mul_u64_u32_shr() a bit

>> more efficient in scheduler.

>>

>> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>

>> ---

>>  arch/arm64/Kconfig | 1 +

>>  1 file changed, 1 insertion(+)

>>

>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig

>> index 5a0a691..ab319eb 100644

>> --- a/arch/arm64/Kconfig

>> +++ b/arch/arm64/Kconfig

>> @@ -11,6 +11,7 @@ config ARM64

>>       select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST

>>       select ARCH_USE_CMPXCHG_LOCKREF

>>       select ARCH_SUPPORTS_ATOMIC_RMW

>> +     select ARCH_SUPPORTS_INT128

>>       select ARCH_SUPPORTS_NUMA_BALANCING

>>       select ARCH_WANT_OPTIONAL_GPIOLIB

>>       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION

>

> With this patch and UBSAN+KASAN enabled, Linux fails to link with:

>

> lib/built-in.o: In function `get_signed_val':

> lib/ubsan.c:93: undefined reference to `__ashlti3'

> lib/ubsan.c:93: undefined reference to `__ashrti3'

>

> It succeeds if I revert commit d67703a8a69e ("arm64: kill off the libgcc

> dependency"). We may have to revisit that decision or implement the

> required library functions in the kernel. In general, I'd like to reduce

> the amount of code duplication if the functionality can be found

> elsewhere like in libgcc. There is a risk of not knowing precisely what

> ligbcc relies on (like libc functions), though there are other

> architectures linking against it.

>


I wonder if this change is such a huge win if we simply end up
invoking support routines in libgcc to implement the 128-bit
arithmetic. Do we have any benchmark results?

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Catalin Marinas June 22, 2016, 1:43 p.m. UTC | #5 
On Wed, Jun 22, 2016 at 01:43:23PM +0100, Will Deacon wrote:
> On Wed, Jun 22, 2016 at 11:57:50AM +0100, Catalin Marinas wrote:

> > On Tue, Jun 21, 2016 at 02:55:23PM +0800, Kefeng Wang wrote:

> > > The gcc support __SIZEOF_INT128__ and __int128 in arm64, thus,

> > > enable ARCH_SUPPORTS_INT128 to make mul_u64_u32_shr() a bit

> > > more efficient in scheduler.

> > > 

> > > Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>

> > > ---

> > >  arch/arm64/Kconfig | 1 +

> > >  1 file changed, 1 insertion(+)

> > > 

> > > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig

> > > index 5a0a691..ab319eb 100644

> > > --- a/arch/arm64/Kconfig

> > > +++ b/arch/arm64/Kconfig

> > > @@ -11,6 +11,7 @@ config ARM64

> > >  	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST

> > >  	select ARCH_USE_CMPXCHG_LOCKREF

> > >  	select ARCH_SUPPORTS_ATOMIC_RMW

> > > +	select ARCH_SUPPORTS_INT128

> > >  	select ARCH_SUPPORTS_NUMA_BALANCING

> > >  	select ARCH_WANT_OPTIONAL_GPIOLIB

> > >  	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION

> > 

> > With this patch and UBSAN+KASAN enabled, Linux fails to link with:

> > 

> > lib/built-in.o: In function `get_signed_val':

> > lib/ubsan.c:93: undefined reference to `__ashlti3'

> > lib/ubsan.c:93: undefined reference to `__ashrti3'

> > 

> > It succeeds if I revert commit d67703a8a69e ("arm64: kill off the libgcc

> > dependency"). We may have to revisit that decision or implement the

> > required library functions in the kernel. In general, I'd like to reduce

> > the amount of code duplication if the functionality can be found

> > elsewhere like in libgcc. There is a risk of not knowing precisely what

> > ligbcc relies on (like libc functions), though there are other

> > architectures linking against it.

> 

> Linking agianst libgcc is pretty scary if we're starting to use plugins

> for the compiler[1], but IANAL.


At a quick grep, there are 9 other architectures linking the kernel
image against libgcc. I wonder what they do about it.

For the time being, I'm reverting this patch. It's probably not that
hard to add the int128 variants of these functions to the kernel (in a
GPLv2 variant).

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Kefeng Wang June 23, 2016, 3:23 a.m. UTC | #6 
On 2016/6/22 18:57, Catalin Marinas wrote:
> On Tue, Jun 21, 2016 at 02:55:23PM +0800, Kefeng Wang wrote:

>> The gcc support __SIZEOF_INT128__ and __int128 in arm64, thus,

>> enable ARCH_SUPPORTS_INT128 to make mul_u64_u32_shr() a bit

>> more efficient in scheduler.

>>

>> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>

>> ---

>>  arch/arm64/Kconfig | 1 +

>>  1 file changed, 1 insertion(+)

>>

>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig

>> index 5a0a691..ab319eb 100644

>> --- a/arch/arm64/Kconfig

>> +++ b/arch/arm64/Kconfig

>> @@ -11,6 +11,7 @@ config ARM64

>>  	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST

>>  	select ARCH_USE_CMPXCHG_LOCKREF

>>  	select ARCH_SUPPORTS_ATOMIC_RMW

>> +	select ARCH_SUPPORTS_INT128

>>  	select ARCH_SUPPORTS_NUMA_BALANCING

>>  	select ARCH_WANT_OPTIONAL_GPIOLIB

>>  	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION

> 

> With this patch and UBSAN+KASAN enabled, Linux fails to link with:


Hi Catalin,

I use an old gcc version[1], after enable CONFIG_UBSAN, I got some build error,

wkf@linux:~/work/linux/build-64> ../scripts/diffconfig .config*
 UBSAN n -> y
+UBSAN_ALIGNMENT n
+UBSAN_SANITIZE_ALL y

arch/arm64/kernel/built-in.o: In function `c_show':
/home/wkf/work/linux/build-64/../arch/arm64/kernel/cpuinfo.c:125:(.text+0xfb28): relocation truncated to fit: R_AARCH64_LDST64_ABS_LO12_NC against symbol `loops_per_jiffy' defined in .data section in init/built-in.o
/home/wkf/work/linux/build-64/../arch/arm64/kernel/cpuinfo.c:125:(.text+0xfccc): relocation truncated to fit: R_AARCH64_LDST64_ABS_LO12_NC against symbol `loops_per_jiffy' defined in .data section in init/built-in.o
drivers/built-in.o: In function `_until_dmac_idle':
/home/wkf/work/linux/build-64/../drivers/dma/pl330.c:916:(.text+0x1920e0): relocation truncated to fit: R_AARCH64_LDST64_ABS_LO12_NC against symbol `loops_per_jiffy' defined in .data section in init/built-in.o
drivers/built-in.o: In function `xen_alloc_coherent_pages':
/home/wkf/work/linux/build-64/../arch/arm64/include/../../arm/include/asm/xen/page-coherent.h:24:(.text+0x1f3b2c): relocation truncated to fit: R_AARCH64_LDST64_ABS_LO12_NC against symbol `dummy_dma_ops' defined in .data section in arch/arm64/mm/built-in.o
drivers/built-in.o: In function `flush':
/home/wkf/work/linux/build-64/../drivers/spi/spi-pl022.c:549:(.text+0x4080f8): relocation truncated to fit: R_AARCH64_LDST64_ABS_LO12_NC against symbol `loops_per_jiffy' defined in .data section in init/built-in.o
drivers/built-in.o: In function `e1000_clean_rx_irq_ps':
/home/wkf/work/linux/build-64/../drivers/net/ethernet/intel/e1000e/netdev.c:1394:(.text+0x49a864): relocation truncated to fit: R_AARCH64_LDST32_ABS_LO12_NC against symbol `copybreak' defined in .data section in drivers/built-in.o
drivers/built-in.o: In function `e1000_clean_rx_irq':
/home/wkf/work/linux/build-64/../drivers/net/ethernet/intel/e1000e/netdev.c:1018:(.text+0x49c408): relocation truncated to fit: R_AARCH64_LDST32_ABS_LO12_NC against symbol `copybreak' defined in .data section in drivers/built-in.o
drivers/built-in.o: In function `e1000_clean_jumbo_rx_irq':
/home/wkf/work/linux/build-64/../drivers/net/ethernet/intel/e1000e/netdev.c:1609:(.text+0x49d28c): relocation truncated to fit: R_AARCH64_LDST32_ABS_LO12_NC against symbol `copybreak' defined in .data section in drivers/built-in.o
drivers/built-in.o: In function `led_trigger_unregister':
/home/wkf/work/linux/build-64/../drivers/leds/led-triggers.c:233:(.text+0x7865e0): relocation truncated to fit: R_AARCH64_LDST64_ABS_LO12_NC against symbol `leds_list' defined in .data section in drivers/built-in.o
drivers/built-in.o: In function `led_trigger_register':
/home/wkf/work/linux/build-64/../drivers/leds/led-triggers.c:206:(.text+0x78693c): relocation truncated to fit: R_AARCH64_LDST64_ABS_LO12_NC against symbol `leds_list' defined in .data section in drivers/built-in.o
drivers/built-in.o: In function `tegra124_132_clock_init_post':
/home/wkf/work/linux/build-64/../drivers/clk/tegra/clk-tegra124.c:1641:(.init.text+0x3bd34): additional relocation overflows omitted from the output
make[2]: *** [vmlinux] Error 1
make[1]: *** [sub-make] Error 2
make: *** [__sub-make] Error 2


[1] gcc version
wkf@linux:~/work/linux/build-64> aarch64-linux-gnu-gcc -v
Using built-in specs.
COLLECT_GCC=aarch64-linux-gnu-gcc
COLLECT_LTO_WRAPPER=/opt/gcc-linaro-4.9-2014.11-x86_64_aarch64-linux-gnu/bin/../libexec/gcc/aarch64-linux-gnu/4.9.3/lto-wrapper
Target: aarch64-linux-gnu
Configured with: /home/buildslave/workspace/BinaryRelease/label/x86_64/target/aarch64-linux-gnu/snapshots/gcc-linaro-4.9-2014.11/configure SHELL=/bin/bash --with-bugurl=https://bugs.linaro.org --with-mpc=/home/buildslave/workspace/BinaryRelease/label/x86_64/target/aarch64-linux-gnu/_build/builds/destdir/x86_64-unknown-linux-gnu --with-mpfr=/home/buildslave/workspace/BinaryRelease/label/x86_64/target/aarch64-linux-gnu/_build/builds/destdir/x86_64-unknown-linux-gnu --with-gmp=/home/buildslave/workspace/BinaryRelease/label/x86_64/target/aarch64-linux-gnu/_build/builds/destdir/x86_64-unknown-linux-gnu --with-gnu-as --with-gnu-ld --disable-libstdcxx-pch --disable-libmudflap --with-cloog=no --with-ppl=no --with-isl=no --disable-nls --enable-multiarch --disable-multilib --enable-c99 --with-arch=armv8-a --disable-shared --enable-static --with-build-sysroot=/home/buildslave/workspace/BinaryRelease/label/x86_64/target/aarch64-linux-gnu/_build/sysroots/aarch64-linux-gnu --enable-lto --!
enable-l
inker-build-id --enable-long-long --enable-shared --with-sysroot=/home/buildslave/workspace/BinaryRelease/label/x86_64/target/aarch64-linux-gnu/_build/builds/destdir/x86_64-unknown-linux-gnu/libc --enable-languages=c,c++,fortran,lto -enable-fix-cortex-a53-835769 --enable-checking=release --with-bugurl=https://bugs.linaro.org --with-pkgversion='Linaro GCC 2014.11' --build=x86_64-unknown-linux-gnu --host=x86_64-unknown-linux-gnu --target=aarch64-linux-gnu --prefix=/home/buildslave/workspace/BinaryRelease/label/x86_64/target/aarch64-linux-gnu/_build/builds/destdir/x86_64-unknown-linux-gnu
Thread model: posix
gcc version 4.9.3 20141031 (prerelease) (Linaro GCC 2014.11)


> 

> lib/built-in.o: In function `get_signed_val':

> lib/ubsan.c:93: undefined reference to `__ashlti3'

> lib/ubsan.c:93: undefined reference to `__ashrti3'


Will update the new gcc and try again.

Thanks,
Kefeng

> 

> It succeeds if I revert commit d67703a8a69e ("arm64: kill off the libgcc

> dependency"). We may have to revisit that decision or implement the

> required library functions in the kernel. In general, I'd like to reduce

> the amount of code duplication if the functionality can be found

> elsewhere like in libgcc. There is a risk of not knowing precisely what

> ligbcc relies on (like libc functions), though there are other

> architectures linking against it.

> 






_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
diff mbox

Patch

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 5a0a691..ab319eb 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -11,6 +11,7 @@  config ARM64
 	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST
 	select ARCH_USE_CMPXCHG_LOCKREF
 	select ARCH_SUPPORTS_ATOMIC_RMW
+	select ARCH_SUPPORTS_INT128
 	select ARCH_SUPPORTS_NUMA_BALANCING
 	select ARCH_WANT_OPTIONAL_GPIOLIB
 	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION