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