linsolve bug

[parisse]

Giac is using 48 bits mantissa by default, unless it is compiled with the -DDOUBLEVAL flag. At least on linux. On windows, I don't remember, perhaps it is different for the 32 bits cygwin version, 64 bits cygwin version and 64 bits mingw version ("native").
This is done in order to have 5 bits to store the type of a gen (generic C type to represent an object) and keep sizeof(gen)=8 bytes=64 bits. If -DDOUBLEVAL is used, then sizeof(gen)=12 bytes on 32 bits architectures and =16 bytes on 64 bits architectures. This means that objects are larger, e.g. matrices take twice more space, which slow down computations (especially when the L1 cache limit is reached).
48 bits is sufficient except for bad conditionned computations, and if you need more then you can calculate with longfloats using e.g. evalf(.,30) to have 30 digits. Multiprecision floats are currently not available on the Prime, they could be sometimes in the future, with quickjs, already available on other calculator ports of giac.