[SciPy-user] How to get past the dfftpack build error "suffix or operands invalid for `movd'"

Steve Leibman sleibman@gmail....
Mon Oct 1 20:26:34 CDT 2007


While attempting to build scipy on an Intel core 2 duo box with fedora
core 5, I ran into a problem during the compilation of dfftpack.
I saw a few other people post this complaint to various lists, but no
published solutions, so this email is intended to show the workaround
that worked for me. I have not tried to reproduce it in a standalone
copy of fftpack from netlib, though presumably the bug is somewhere in
there.

The symptoms:
============

/tmp/ccCnF9WU.s: Assembler messages:
/tmp/ccCnF9WU.s:599: Error: suffix or operands invalid for `movd'
/tmp/ccCnF9WU.s:2982: Error: suffix or operands invalid for `movd'
/tmp/ccCnF9WU.s: Assembler messages:
/tmp/ccCnF9WU.s:599: Error: suffix or operands invalid for `movd'
/tmp/ccCnF9WU.s:2982: Error: suffix or operands invalid for `movd'
error: Command "/usr/bin/g77 -g -Wall -fno-second-underscore -fPIC -O2
-funroll-loops -march=pentium3 -mmmx -msse2 -msse -f
omit-frame-pointer -malign-double -c -c Lib/fftpack/dfftpack/zfftf1.f
-o build/temp.linux-i686-2.4/Lib/fftpack/dfftpack/zfftf1.o" failed
with exit status 1

The solution:
==========
The real solution is to figure out why the generated assembly has a
screwy instruction in it.

In the meantime... despite the fact that my machine supports the sse2
instructions, I was able to work around the issue by getting rid of
the -msse2 flag on the compile line in order to make it work (if you
have this problem, run the g77 compile line both with and without
"-msse2", to see whether this solution will work for you). The problem
is that there's no convenient way to forcefully remove the "-msse2"
option... it isn't written into anything civilized like a Makefile.
Instead, I temporarily modified my python numpy installation, changing
the _has_sse2() function in site-packages/numpy/distutils/cpuinfo.py,
such that it always returns False.
Pretty nasty, but it worked for me.

--
Steve Leibman
sleibman  +  gmail  +  com


More information about the SciPy-user mailing list