
18 Sep
2005
18 Sep
'05
10:48 a.m.
After almost two days of struggle, it is now certain: the version of GCC that, at the moment of writing this message, is the standard one in Fedora Core 4/i386 miscompiles the PPL, at least at -O2 or greater optimization level.
The problem can be seen by compiling tests/Polyhedron/relations17.cc: at -O2 and above the machine instruction
movl $2, 4
is generated, and this will of course cause a segmentation fault. At -O1 the right instruction, i.e.,
movl $2, 4(%eax)
is generated. Many thanks to Abramo who found that "smoking gun".
Luckily, the problem seems not to be present in the prerelease of GCC 4.0.2 of September 13th. Cheers,
Roberto
--
Prof. Roberto Bagnara
Computer Science Group
Department of Mathematics, University of Parma, Italy
http://www.cs.unipr.it/~bagnara/
mailto:bagnara@cs.unipr.it