[Bug libc/14059] New: HAS_FMA4 check needs to also check for AVX

[Bug libc/14059] New: HAS_FMA4 check needs to also check for AVX

[jwestfall at surrealistic dot net]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

             Bug #: 14059
           Summary: HAS_FMA4 check needs to also check for AVX
           Product: glibc
           Version: 2.15
            Status: NEW
          Severity: normal
          Priority: P2
         Component: libc
        AssignedTo: unassigned@sourceware.org
        ReportedBy: jwestfall@surrealistic.net
                CC: drepper.fsp@gmail.com
    Classification: Unclassified


Hi

In number of the sysdeps/x86_64/fpu/multiarch/*.c files there is code like this


libm_ifunc (__ieee754_exp,
            HAS_FMA4 ? __ieee754_exp_fma4
            : (HAS_AVX ? __ieee754_exp_avx : __ieee754_exp_sse2));

The HAS_FMA4 check is only looking at the FMA4 bit from cpuid() to determine if
it runs the fma4 version of the function, but fma4 instructions are dependent
on avx instructions.  This can result in the following invalid opcode if avx
isn't available.

(gdb) exec-file python-dbg
(gdb) run
Starting program: /usr/bin/python-dbg
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
Python 2.7.3 (default, Apr 20 2012, 22:01:19)
[GCC 4.6.3] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> print test

Program received signal SIGILL, Illegal instruction.
__ieee754_exp_fma4 (x=-0.5) at ../sysdeps/ieee754/dbl-64/e_exp.c:63
63 ../sysdeps/ieee754/dbl-64/e_exp.c: No such file or directory.
(gdb) bt
#0 __ieee754_exp_fma4 (x=-0.5) at ../sysdeps/ieee754/dbl-64/e_exp.c:63
#1 0x000000000058509f in ?? ()
#2 0x00000000009bde80 in ?? ()
#3 0x0000000100c52a10 in ?? ()
#4 0x0000000000417380 in ?? ()
#5 0x00000000009d2970 in ?? ()
#6 0x0000000000000000 in ?? ()
(gdb) info address __ieee754_exp_fma4
Symbol "__ieee754_exp_fma4" is a function at address 0x7ffff6cc35f0.
(gdb) disassemble 0x7ffff6cc35f0,+30
Dump of assembler code from 0x7ffff6cc35f0 to 0x7ffff6cc360e:
   0x00007ffff6cc35f0 <__ieee754_exp_fma4+0>: push %rbp
   0x00007ffff6cc35f1 <__ieee754_exp_fma4+1>: mov %rsp,%rbp
   0x00007ffff6cc35f4 <__ieee754_exp_fma4+4>: and $0xffffffffffffffe0,%rsp
   0x00007ffff6cc35f8 <__ieee754_exp_fma4+8>: add $0x10,%rsp
=> 0x00007ffff6cc35fc <__ieee754_exp_fma4+12>: vmovsd %xmm0,-0x20(%rsp)
   0x00007ffff6cc3602 <__ieee754_exp_fma4+18>: mov -0x20(%rsp),%rax
   0x00007ffff6cc3607 <__ieee754_exp_fma4+23>: mov %rax,%rcx
   0x00007ffff6cc360a <__ieee754_exp_fma4+26>: shr $0x20,%rcx


thanks
jim

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[jwestfall at surrealistic dot net]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

Jim Westfall <jwestfall at surrealistic dot net> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jwestfall at surrealistic
                   |                            |dot net

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[aj at suse dot de]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

Andreas Jaeger <aj at suse dot de> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |aj at suse dot de

--- Comment #1 from Andreas Jaeger <aj at suse dot de> 2012-05-07 19:06:59 UTC ---
This looks related to http://sourceware.org/bugzilla/show_bug.cgi?id=13583
which was fixed with commits afc5ed09cbce5d6fd48b3a8c5ec427b31f996880.

Can you confirm that that patch - most probably together with
6ee65ed6ddbf04402fad0bec6aa9c73b9d982ae4 - fixes your problem? Or is this
really a separate issue?

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[jwestfall at surrealistic dot net]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

--- Comment #2 from Jim Westfall <jwestfall at surrealistic dot net> 2012-05-07 20:01:11 UTC ---
Created attachment 6398
  --> http://sourceware.org/bugzilla/attachment.cgi?id=6398
patch sent to ubuntu folks to fix their glibc build

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[jwestfall at surrealistic dot net]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

--- Comment #3 from Jim Westfall <jwestfall at surrealistic dot net> 2012-05-07 20:03:21 UTC ---
Hi

My backtrace is actually from 2.15ubuntu10 which has
afc5ed09cbce5d6fd48b3a8c5ec427b31f996880 applied.  

Something similar to that would be needed for disabling FMA4 if AVX is
unavailable.  Attached is the patch I submitted to the ubuntu folks to fix the
issue.

I will do a fresh build against master and will report back.

I am a bit dubious that master is broken with AVX runtime detection because of
the following 2 commits

08cf777f9e7f6d826658a99c7d77a359f73a45bf
56f6f6a2403cfa7267cad722597113be35ecf70d

Specifically 56f6f6a2403cfa7267cad722597113be35ecf70d reverts HAS_YMM_USABLE
back to HAS_AVX, but it didn't also revert the code in
sysdeps/x86_64/multiarch/init-arch.c to unset the AVX bit.  So It now appears
that HAS_AVX will be true when OSXSAVE is disabled, which would be
re-introducing the issue from bug 13583.

thanks
jim

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[jwestfall at surrealistic dot net]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

--- Comment #4 from Jim Westfall <jwestfall at surrealistic dot net> 2012-05-07 22:46:18 UTC ---
Hi

Here is what I see with master/head.  My test case is simply this

#include <math.h>

int main(void) {
  double x = 0.5;
  exp(x);
  return 0;
}
gcc -o exp-test exp-test.c -lm

Intel i5-2405S dom0 under xen.  AVX runtime detection is re-broken.

Program terminated with signal 4, Illegal instruction.
#0  0x00007f00d708ff71 in __ieee754_exp_avx (x=0.5) at
../sysdeps/ieee754/dbl-64/e_exp.c:55
55    __ieee754_exp(double x) {
(gdb) info address __ieee754_exp_avx
Symbol "__ieee754_exp_avx" is a function at address 0x7f00d708ff70.
(gdb) disassemble __ieee754_exp_avx,+30
Dump of assembler code from 0x7f00d708ff70 to 0x7f00d708ff8e:
   0x00007f00d708ff70 <__ieee754_exp_avx+0>:    push   %rbx
=> 0x00007f00d708ff71 <__ieee754_exp_avx+1>:    vmovapd %xmm0,%xmm2
   0x00007f00d708ff75 <__ieee754_exp_avx+5>:    sub    $0x20,%rsp
   0x00007f00d708ff79 <__ieee754_exp_avx+9>:    vstmxcsr 0x1c(%rsp)
   0x00007f00d708ff7f <__ieee754_exp_avx+15>:    mov    0x1c(%rsp),%ebx
   0x00007f00d708ff83 <__ieee754_exp_avx+19>:    mov    %ebx,%eax
   0x00007f00d708ff85 <__ieee754_exp_avx+21>:    and    $0x9f,%ah
   0x00007f00d708ff88 <__ieee754_exp_avx+24>:    mov    %eax,0x1c(%rsp)
   0x00007f00d708ff8c <__ieee754_exp_avx+28>:    vldmxcsr 0x1c(%rsp)
End of assembler dump.

AMD 6272 dom0 under xen
Program terminated with signal 4, Illegal instruction.
#0  0x00007fe07cf29d31 in __ieee754_exp_fma4 (x=0.5)
    at ../sysdeps/ieee754/dbl-64/e_exp.c:55
55    ../sysdeps/ieee754/dbl-64/e_exp.c: No such file or directory.
(gdb) disassemble __ieee754_exp_fma4,+30
Dump of assembler code from 0x7fe07cf29d30 to 0x7fe07cf29d4e:
   0x00007fe07cf29d30 <__ieee754_exp_fma4+0>:    push   %rbx
=> 0x00007fe07cf29d31 <__ieee754_exp_fma4+1>:    vmovapd %xmm0,%xmm1
   0x00007fe07cf29d35 <__ieee754_exp_fma4+5>:    sub    $0x20,%rsp
   0x00007fe07cf29d39 <__ieee754_exp_fma4+9>:    vstmxcsr 0x1c(%rsp)
   0x00007fe07cf29d3f <__ieee754_exp_fma4+15>:    mov    0x1c(%rsp),%ebx
   0x00007fe07cf29d43 <__ieee754_exp_fma4+19>:    mov    %ebx,%eax
   0x00007fe07cf29d45 <__ieee754_exp_fma4+21>:    and    $0x9f,%ah
   0x00007fe07cf29d48 <__ieee754_exp_fma4+24>:    mov    %eax,0x1c(%rsp)
   0x00007fe07cf29d4c <__ieee754_exp_fma4+28>:    vldmxcsr 0x1c(%rsp)
End of assembler dump.

jim

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[jwestfall at surrealistic dot net]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

--- Comment #5 from Jim Westfall <jwestfall at surrealistic dot net> 2012-05-07 22:50:27 UTC ---
Created attachment 6399
  --> http://sourceware.org/bugzilla/attachment.cgi?id=6399
patch that fixes avx/fma4 runtime detection in master

This patch fixes a (missed?) revert as part of
08cf777f9e7f6d826658a99c7d77a359f73a45bf and
56f6f6a2403cfa7267cad722597113be35ecf70d, and also disables fma4 if avx is
unavailable.

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[aj at suse dot de]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

Andreas Jaeger <aj at suse dot de> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Attachment #6399|0                           |1
           is patch|                            |
   Attachment #6399|application/octet-stream    |text/plain
          mime type|                            |

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[aj at suse dot de]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

--- Comment #6 from Andreas Jaeger <aj at suse dot de> 2012-05-08 08:14:38 UTC ---
Thanks for the patch.

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[carlos_odonell at mentor dot com]

http://sourceware.org/bugzilla/show_bug.cgi?id=14059

Carlos O'Donell <carlos_odonell at mentor dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
                 CC|                            |carlos_odonell at mentor
                   |                            |dot com
         Resolution|                            |FIXED
         AssignedTo|unassigned at sourceware    |carlos_odonell at mentor
                   |dot org                     |dot com

--- Comment #7 from Carlos O'Donell <carlos_odonell at mentor dot com> 2012-05-17 14:10:55 UTC ---
Fixed with this commit:

http://sources.redhat.com/git/gitweb.cgi?p=glibc.git;a=commitdiff;h=1a0994f5356214e8af8a1c1cc33fbf74a7ac8993

We've cleaned this up considerably, since technically FMA4 doesn't depend on
AVX, it just depends on YMM/XMM state being usable (saved and restored by the
OS) and the feature being present.

We also fix the AVX detection with this patch.

Tested on AVX-enabled and non-AVX-enabled hosts, and for good measure we now
have a regression test for this.

I'm going to backport this to 2.15.1 to fix the mess there, see BZ313753 for
the backport request.

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[jackie.rosen at hushmail dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=14059

Jackie Rosen <jackie.rosen at hushmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jackie.rosen at hushmail dot com

--- Comment #8 from Jackie Rosen <jackie.rosen at hushmail dot com> ---
*** Bug 260998 has been marked as a duplicate of this bug. ***
Seen from the domain http://volichat.com
Page where seen: http://volichat.com/adult-chat-rooms
Marked for reference. Resolved as fixed @bugzilla.

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[schwab at sourceware dot org]

https://sourceware.org/bugzilla/show_bug.cgi?id=14059

Andreas Schwab <schwab at sourceware dot org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|jackie.rosen at hushmail dot com   |

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug libc/14059] HAS_FMA4 check needs to also check for AVX

[fweimer at redhat dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=14059

Florian Weimer <fweimer at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
              Flags|                            |security-

-- 
You are receiving this mail because:
You are on the CC list for the bug.