Re: F25 builds a library with AVX, causes SIGILL on AMD

[Pete Zaitcev]

On Tue, 23 May 2017 09:30:14 +0300
Benson Muite  wrote:

> Will this mean that repository builds will typically have poor 
> performance to support portability?

Yes. In extreme cases, where the difference in performance is measurable,
the library may need to be modified to select a proper implementation
at runtime.

Ironically, liberasurecode crashed on the part that's only invoked
once per data block, so obviously it's not performance critical to
such a degree. In addition, I'm not sure a conversion to floating
point is even necessary there.

Note that in case of liberasurecode, the library itself is by definition
is a switcher shim. Those that want performance specify the fastest Galois
Field implementation and liberasurecode then uses -ldl. The fastsest
known module for Intel CPUs is their in-house ISA-L library.

-- Pete
___
devel mailing list -- devel@lists.fedoraproject.org
To unsubscribe send an email to devel-le...@lists.fedoraproject.org

Re: F25 builds a library with AVX, causes SIGILL on AMD

[Dominik 'Rathann' Mierzejewski]

On Tuesday, 23 May 2017 at 08:30, Benson Muite wrote:
> Will this mean that repository builds will typically have poor performance
> to support portability?

It depends on how much gain is there from enabling AVX instructions, but
yes, we have to disable any non-runtime-detected usage of instructions
supported only by certain processors.

Regards,
Dominik
-- 
Fedora http://fedoraproject.org/wiki/User:Rathann
RPMFusion http://rpmfusion.org
"Faith manages."
-- Delenn to Lennier in Babylon 5:"Confessions and Lamentations"
___
devel mailing list -- devel@lists.fedoraproject.org
To unsubscribe send an email to devel-le...@lists.fedoraproject.org

Re: F25 builds a library with AVX, causes SIGILL on AMD

[Björn 'besser82' Esser]

Am 23.05.2017 um 07:00 schrieb Pete Zaitcev:

Hi, All:

I'm going to re-state the problem in short, but in case, URL
  https://bugzilla.redhat.com/show_bug.cgi?id=1454543

- I maintain a library (liberasurecode)
- It contains a call to ceill() from -lm
- When Koji builds the code, something makes gcc to inline AVX
- The result blows up:
Program received signal SIGILL, Illegal instruction.
0x7fafaf9e04bd in liberasurecode_get_aligned_data_size (
=> 0x778c44bd <+61>:vxorpd %xmm0,%xmm0,%xmm0

As much as I can tell, there's nothing custom in CFLAGS in Makefile.am,
everything is inherited from RPM somehow. Here's how the actual line
looks like in Koji logs:

libtool: compile:  gcc -DHAVE_CONFIG_H -I. -I../include 
-I../include/erasurecode -I../include/xor_codes -I../include/rs_vand 
-I../include/isa_l -I../include/shss -Werror -O2 -g -Werror -D_GNU_SOURCE=1 
-Wall -pedantic -std=c99 -O2 -g -pipe -Wall -Werror=format-security 
-Wp,-D_FORTIFY_SOURCE=2 -fexceptions -fstack-protector-strong 
--param=ssp-buffer-size=4 -grecord-gcc-switches 
-specs=/usr/lib/rpm/redhat/redhat-hardened-cc1 -m64 -mtune=generic -mmmx 
-DINTEL_MMX -msse -DINTEL_SSE -msse2 -DINTEL_SSE2 -msse3 -DINTEL_SSE3 -mssse3 
-DINTEL_SSSE3 -msse4.1 -DINTEL_SSE41 -msse4.2 -DINTEL_SSE42 -mavx -DINTEL_AVX 
-DARCH_64 -c erasurecode.c  -fPIC -DPIC -o .libs/liberasurecode_la-erasurecode.o

So, what now? I have no clue where to start fixing this.

Thanks in advance for any suggestions,
-- Pete



Hi Pete,

for some reason upstream appends `-mmmx -DINTEL_MMX -msse -DINTEL_SSE 
-msse2 -DINTEL_SSE2 -msse3 -DINTEL_SSE3 -mssse3 -DINTEL_SSSE3 -msse4.1 
-DINTEL_SSE41 -msse4.2 -DINTEL_SSE42 -mavx -DINTEL_AVX` to CFLAGS, 
activating avx and several different sse versions…


You must remove remove those (except for sse <= 2 on x86_64), to make 
the compiled binaries portable between different (older) CPUs; those 
might be set in configure.ac [1] as well…


Cheers,
  Björn


[1] 
https://github.com/openstack/liberasurecode/blob/master/configure.ac#L155

___
devel mailing list -- devel@lists.fedoraproject.org
To unsubscribe send an email to devel-le...@lists.fedoraproject.org

Re: F25 builds a library with AVX, causes SIGILL on AMD

[Benson Muite]

Hi,

Will this mean that repository builds will typically have poor 
performance to support portability?


Benson


On 05/23/2017 08:15 AM, Pete Zaitcev wrote:

On Mon, 22 May 2017 23:00:14 -0600
Pete Zaitcev  wrote:


As much as I can tell, there's nothing custom in CFLAGS in Makefile.am,
everything is inherited from RPM somehow. Here's how the actual line
looks like in Koji logs:

Hate to reply to myself, but I'm an idiot as usual. The culprit is in
configure.ac, not in Makefile.am

$CC - -E -mavx /dev/null 2>&1
if [[ $? == "0" ]]; then
   SUPPORTED_FLAGS="$SUPPORTED_FLAGS -mavx"
   AC_MSG_RESULT([$CC supports -mavx])
fi

I need to nail down this stuff to some generally portable subset.

Sorry for the noise,
-- Pete
___
devel mailing list -- devel@lists.fedoraproject.org
To unsubscribe send an email to devel-le...@lists.fedoraproject.org

___
devel mailing list -- devel@lists.fedoraproject.org
To unsubscribe send an email to devel-le...@lists.fedoraproject.org

Re: F25 builds a library with AVX, causes SIGILL on AMD

[Pete Zaitcev]

On Mon, 22 May 2017 23:00:14 -0600
Pete Zaitcev  wrote:

> As much as I can tell, there's nothing custom in CFLAGS in Makefile.am,
> everything is inherited from RPM somehow. Here's how the actual line
> looks like in Koji logs:

Hate to reply to myself, but I'm an idiot as usual. The culprit is in
configure.ac, not in Makefile.am

$CC - -E -mavx /dev/null 2>&1
if [[ $? == "0" ]]; then
  SUPPORTED_FLAGS="$SUPPORTED_FLAGS -mavx"
  AC_MSG_RESULT([$CC supports -mavx])
fi

I need to nail down this stuff to some generally portable subset.

Sorry for the noise,
-- Pete
___
devel mailing list -- devel@lists.fedoraproject.org
To unsubscribe send an email to devel-le...@lists.fedoraproject.org

F25 builds a library with AVX, causes SIGILL on AMD

[Pete Zaitcev]

Hi, All:

I'm going to re-state the problem in short, but in case, URL
 https://bugzilla.redhat.com/show_bug.cgi?id=1454543

- I maintain a library (liberasurecode)
- It contains a call to ceill() from -lm
- When Koji builds the code, something makes gcc to inline AVX
- The result blows up:
Program received signal SIGILL, Illegal instruction.  
0x7fafaf9e04bd in liberasurecode_get_aligned_data_size (
=> 0x778c44bd <+61>:vxorpd %xmm0,%xmm0,%xmm0

As much as I can tell, there's nothing custom in CFLAGS in Makefile.am,
everything is inherited from RPM somehow. Here's how the actual line
looks like in Koji logs:

libtool: compile:  gcc -DHAVE_CONFIG_H -I. -I../include 
-I../include/erasurecode -I../include/xor_codes -I../include/rs_vand 
-I../include/isa_l -I../include/shss -Werror -O2 -g -Werror -D_GNU_SOURCE=1 
-Wall -pedantic -std=c99 -O2 -g -pipe -Wall -Werror=format-security 
-Wp,-D_FORTIFY_SOURCE=2 -fexceptions -fstack-protector-strong 
--param=ssp-buffer-size=4 -grecord-gcc-switches 
-specs=/usr/lib/rpm/redhat/redhat-hardened-cc1 -m64 -mtune=generic -mmmx 
-DINTEL_MMX -msse -DINTEL_SSE -msse2 -DINTEL_SSE2 -msse3 -DINTEL_SSE3 -mssse3 
-DINTEL_SSSE3 -msse4.1 -DINTEL_SSE41 -msse4.2 -DINTEL_SSE42 -mavx -DINTEL_AVX 
-DARCH_64 -c erasurecode.c  -fPIC -DPIC -o .libs/liberasurecode_la-erasurecode.o

So, what now? I have no clue where to start fixing this.

Thanks in advance for any suggestions,
-- Pete
___
devel mailing list -- devel@lists.fedoraproject.org
To unsubscribe send an email to devel-le...@lists.fedoraproject.org