Clear FPU in i386/x86_64 longjmp()

[jdemeyer]

If an interrupt occurs during a 387 FPU computation or during an MMX computation (which also uses the FPU registers), then some of the FPU registers will be occupied. This may lead to failures with subsequent FPU compuations, see for example #12777. Some operating systems (e.g. Linux and OS X 10.6) restore the FPU state when an interrupt occurs, but on Solaris this is not the case.

The proposed solution is to clear the FPU "tag word" (using the EMMS or FFREE instructions) in the signal handler.

To see the problem in action, compile attachment: emms.c on i686 hardware with

gcc -march=pentium2 -m32 -mfpmath=387 -lm emms.c -o emms

Apply attachment: 12873_emms.patch

Obsoleted by #13076 and fixed in a better way.

CC: @nexttime

Component: c_lib

Author: Jeroen Demeyer

Reviewer: Volker Braun

Merged: sage-5.1.beta0

Issue created by migration from https://trac.sagemath.org/ticket/12873

[jdemeyer]

comment:1

Well, actually it's not so clear whether longjmp() should restore the FPU state. I'll try to see whether I can reproduce the problem on Linux i686.

[jdemeyer]

comment:3

I can't be sure that the problem is limited to Solaris. Using a test program (attached), I managed to reproduce the issue on Linux also.

[jdemeyer]

Test program

[jdemeyer]

comment:6

Attachment: emms.c.gz

[jdemeyer]

comment:9

Attachment: 12873_emms.patch.gz

[vbraun]

Reviewer: Volker Braun

[vbraun]

comment:11

The joys of the x86 architechture! ;-)

Patch looks good to me.

[jdemeyer]

Merged: sage-5.1.beta0

[jdemeyer]

Similar test program using setcontext()

[jdemeyer]

comment:14

Attachment: context.c.gz

#13076 fixes this problem in a much better way, so that ticket effectively undoes this.