| Message ID | 1457046400-29276-2-git-send-email-cyrilbur@gmail.com (mailing list archive) |
|---|---|
| State | Accepted |
| Headers | show |
On Thu, 2016-03-03 at 23:06:40 UTC, Cyril Bur wrote: > Currently it doesn't appear the resulting binary actually uses any Altivec > or VSX instructions the solution is to explicitly tell GCC to use vector > instructions and use vector types in the code. > > Part of this this issue can be GCC version specific: > > GCC 4.9.x is happy to use Altivec and VSX instructions if altivec.h is > includedi (and possibly if vector types are used), this also means that > 4.9.x will use VSX instructions even if only -maltivec is passed. It is > also possible that Altivec instructions will be used even without -maltivec > or -mabi=altivec. > > GCC 5.2.x complains about the lack of -maltivec parameter if altivec.h is > included and will not use VSX unless -mvsx is present on commandline. > > GCC 5.3.0 has a regression that means __attribute__((__target__("no-vsx")) > fails to build. A fix is targeted for 5.4. > > Furthermore LTO (Link Time Optimisation) doesn't play well with > __attribute__((__target__("no-vsx")), LTO can cause GCC to forget about the > attribute and compile with VSX instructions regardless. Be weary when > enabling -flfo for this test. > > Signed-off-by: Cyril Bur <cyrilbur@gmail.com> Applied to powerpc next, thanks. https://git.kernel.org/powerpc/c/f2418ae8a81760b4dec8d5e3e7 cheers
Hi Cyril, On 03/03/2016 11:06 PM, Cyril Bur wrote: > Currently it doesn't appear the resulting binary actually uses any Altivec > or VSX instructions the solution is to explicitly tell GCC to use vector > instructions and use vector types in the code. > > Part of this this issue can be GCC version specific: > > GCC 4.9.x is happy to use Altivec and VSX instructions if altivec.h is > includedi (and possibly if vector types are used), this also means that > 4.9.x will use VSX instructions even if only -maltivec is passed. It is > also possible that Altivec instructions will be used even without -maltivec > or -mabi=altivec. > > GCC 5.2.x complains about the lack of -maltivec parameter if altivec.h is > included and will not use VSX unless -mvsx is present on commandline. > > GCC 5.3.0 has a regression that means __attribute__((__target__("no-vsx")) > fails to build. A fix is targeted for 5.4. > > Furthermore LTO (Link Time Optimisation) doesn't play well with > __attribute__((__target__("no-vsx")), LTO can cause GCC to forget about the > attribute and compile with VSX instructions regardless. Be weary when > enabling -flfo for this test. > > Signed-off-by: Cyril Bur <cyrilbur@gmail.com> This patch breaks the build on my setup: ppc-linux-gcc -std=gnu99 -O2 -Wall -Werror -DGIT_VERSION='"v5.0-rc3-560-ge0ce62731d77"' -I/root/linux-powerpc/tools/testing/selftests/powerpc/include -O2 -maltivec -mvsx -mabi=altivec context_switch.c ../harness.c ../utils.c -lpthread -o /root/linux-powerpc/tools/testing/selftests/powerpc/benchmarks/context_switch context_switch.c:1:0: error: -mvsx requires hardware floating point [-Werror] /* ^ cc1: all warnings being treated as errors ../harness.c:1:0: error: -mvsx requires hardware floating point [-Werror] /* ^ cc1: all warnings being treated as errors ../utils.c:1:0: error: -mvsx requires hardware floating point [-Werror] /* ^ cc1: all warnings being treated as errors make[1]: *** [/root/linux-powerpc/tools/testing/selftests/powerpc/benchmarks/context_switch] Error 1 By removing the -mvsx option, it compiles just fine. Is that option really required ? According to gcc doc, it is automatically selected when compiling for cpus that support it. I'm using: [root@po16846vm linux-powerpc]# ppc-linux-gcc -v Using built-in specs. COLLECT_GCC=ppc-linux-gcc COLLECT_LTO_WRAPPER=/opt/cldk-1.4.0/libexec/gcc/ppc-linux/5.4.0/lto-wrapper Target: ppc-linux Configured with: /root/cldk/gcc-5.4.0/configure --target=ppc-linux --with-headers=yes --with-cpu=860 --prefix=/opt/cldk-1.4.0 --bindir=/opt/cldk-1.4.0/bin --sbindir=/opt/cldk-1.4.0/sbin --libexecdir=/opt/cldk-1.4.0/libexec --datadir=/opt/cldk-1.4.0/share --sysconfdir=/opt/cldk-1.4.0/etc --libdir=/opt/cldk-1.4.0/lib --includedir=/opt/cldk-1.4.0/usr/include --oldincludedir=/opt/cldk-1.4.0/usr/include --infodir=/opt/cldk-1.4.0/share/info --mandir=/opt/cldk-1.4.0/share/man --with-glibc-version=2.18 --enable-languages=c,c++ Thread model: posix gcc version 5.4.0 (GCC) Christophe > --- > tools/testing/selftests/powerpc/benchmarks/Makefile | 1 + > tools/testing/selftests/powerpc/benchmarks/context_switch.c | 11 ++++++++--- > 2 files changed, 9 insertions(+), 3 deletions(-) > > diff --git a/tools/testing/selftests/powerpc/benchmarks/Makefile b/tools/testing/selftests/powerpc/benchmarks/Makefile > index 912445f..6816fc2 100644 > --- a/tools/testing/selftests/powerpc/benchmarks/Makefile > +++ b/tools/testing/selftests/powerpc/benchmarks/Makefile > @@ -7,6 +7,7 @@ all: $(TEST_PROGS) > $(TEST_PROGS): ../harness.c > > context_switch: ../utils.c > +context_switch: CFLAGS += -maltivec -mvsx -mabi=altivec > context_switch: LDLIBS += -lpthread > > include ../../lib.mk > diff --git a/tools/testing/selftests/powerpc/benchmarks/context_switch.c b/tools/testing/selftests/powerpc/benchmarks/context_switch.c > index e6af382..a36883a 100644 > --- a/tools/testing/selftests/powerpc/benchmarks/context_switch.c > +++ b/tools/testing/selftests/powerpc/benchmarks/context_switch.c > @@ -25,7 +25,9 @@ > #include <sys/types.h> > #include <sys/shm.h> > #include <linux/futex.h> > - > +#ifdef __powerpc__ > +#include <altivec.h> > +#endif > #include "../utils.h" > > static unsigned int timeout = 30; > @@ -37,12 +39,15 @@ static int touch_fp = 1; > double fp; > > static int touch_vector = 1; > -typedef int v4si __attribute__ ((vector_size (16))); > -v4si a, b, c; > +vector int a, b, c; > > #ifdef __powerpc__ > static int touch_altivec = 1; > > +/* > + * Note: LTO (Link Time Optimisation) doesn't play well with this function > + * attribute. Be very careful enabling LTO for this test. > + */ > static void __attribute__((__target__("no-vsx"))) altivec_touch_fn(void) > { > c = a + b; >
diff --git a/tools/testing/selftests/powerpc/benchmarks/Makefile b/tools/testing/selftests/powerpc/benchmarks/Makefile index 912445f..6816fc2 100644 --- a/tools/testing/selftests/powerpc/benchmarks/Makefile +++ b/tools/testing/selftests/powerpc/benchmarks/Makefile @@ -7,6 +7,7 @@ all: $(TEST_PROGS) $(TEST_PROGS): ../harness.c context_switch: ../utils.c +context_switch: CFLAGS += -maltivec -mvsx -mabi=altivec context_switch: LDLIBS += -lpthread include ../../lib.mk diff --git a/tools/testing/selftests/powerpc/benchmarks/context_switch.c b/tools/testing/selftests/powerpc/benchmarks/context_switch.c index e6af382..a36883a 100644 --- a/tools/testing/selftests/powerpc/benchmarks/context_switch.c +++ b/tools/testing/selftests/powerpc/benchmarks/context_switch.c @@ -25,7 +25,9 @@ #include <sys/types.h> #include <sys/shm.h> #include <linux/futex.h> - +#ifdef __powerpc__ +#include <altivec.h> +#endif #include "../utils.h" static unsigned int timeout = 30; @@ -37,12 +39,15 @@ static int touch_fp = 1; double fp; static int touch_vector = 1; -typedef int v4si __attribute__ ((vector_size (16))); -v4si a, b, c; +vector int a, b, c; #ifdef __powerpc__ static int touch_altivec = 1; +/* + * Note: LTO (Link Time Optimisation) doesn't play well with this function + * attribute. Be very careful enabling LTO for this test. + */ static void __attribute__((__target__("no-vsx"))) altivec_touch_fn(void) { c = a + b;
Currently it doesn't appear the resulting binary actually uses any Altivec or VSX instructions the solution is to explicitly tell GCC to use vector instructions and use vector types in the code. Part of this this issue can be GCC version specific: GCC 4.9.x is happy to use Altivec and VSX instructions if altivec.h is includedi (and possibly if vector types are used), this also means that 4.9.x will use VSX instructions even if only -maltivec is passed. It is also possible that Altivec instructions will be used even without -maltivec or -mabi=altivec. GCC 5.2.x complains about the lack of -maltivec parameter if altivec.h is included and will not use VSX unless -mvsx is present on commandline. GCC 5.3.0 has a regression that means __attribute__((__target__("no-vsx")) fails to build. A fix is targeted for 5.4. Furthermore LTO (Link Time Optimisation) doesn't play well with __attribute__((__target__("no-vsx")), LTO can cause GCC to forget about the attribute and compile with VSX instructions regardless. Be weary when enabling -flfo for this test. Signed-off-by: Cyril Bur <cyrilbur@gmail.com> --- tools/testing/selftests/powerpc/benchmarks/Makefile | 1 + tools/testing/selftests/powerpc/benchmarks/context_switch.c | 11 ++++++++--- 2 files changed, 9 insertions(+), 3 deletions(-)