On Sat, 03 Oct 2015 09:56:23 -0400 Wayne Sallee <wa...@waynesallee.com> wrote:
> 0x00007fe021043dde <__gmpn_mul_1+94>: mulx (%rsi),%rbx,%rax Wayne, That's it! That's the offender -> mulx It's a new instruction for Intel's Haswell processors: http://www.intel.com/content/www/us/en/intelligent-systems/intel-technology/ia-large-integer-arithmetic-paper.html https://en.wikipedia.org/wiki/Haswell_%28microarchitecture%29 But, not every Intel CPU that may be thought of as being of "Haswell" class actually supports it. Which CPU are you running here? Do a cat /proc/cpuinfo To support mulx, the CPU must list bmi2 in the cpuinfo flags section: http://unix.stackexchange.com/questions/43539/what-do-the-flags-in-proc-cpuinfo-mean (see under "Intel-defined CPU features, CPUID level 0x00000007:0 (ebx)") There was a bug in the detection code of gmp before 6.0.0a which affected Intel G3420 CPUs: https://gmplib.org/list-archives/gmp-devel/2014-February/003696.html In that thread, take note that the correct patch (which is already in the latest gmp, 6.0.0a) is here: https://gmplib.org/list-archives/gmp-devel/attachments/20140227/db077071/attachment.obj (which is a text/diff file) You should be able to recompile and reinstall gmp 6.0.0a without having to redo all of gcc. http://www.linuxfromscratch.org/lfs/view/development/chapter06/gmp.html http://ftp.gnu.org/gnu/gmp/ If the CPU BMI detection problem persists on your machine even with gmp 6.0.0a (not withstanding the problem of compiling on one machine and moving the result to another, of course), it should be reported to the gmp maintainers. Although there isn't a configure option to lockout the use of bmi2 instructions in gmp, you can force their lockout with with a code change to config.guess where around line 815 where: if (strcmp (modelstr, "coreihwl") == 0) { /* Some Haswells lack BMI2. Let them appear as Sandybridges for now. */ CPUID (dummy_string, 7); if ((dummy_string[0 + 8 / 8] & (1 << (8 % 8))) == 0) modelstr = "coreisbr"; } can be changed simply to modelstr = "coreisbr"; and also in mpn/x86_64/fat/fat.c around line 293 delete if ((dummy_string[0 + 8 / 8] & (1 << (8 % 8))) != 0) CPUVEC_SETUP_coreihwl; The idea here being is to force the use of coreisbr (which does not use the bmi2 instructions) rather than coreihwl. Cheers, Mike -- http://lists.linuxfromscratch.org/listinfo/blfs-support FAQ: http://www.linuxfromscratch.org/blfs/faq.html Unsubscribe: See the above information page
