store pqcrystals polynomial in secure memory

[falko-strenzke]

No description provided.

[reneme]

Note that the PolynomialVector class is tightly coupled with the Polynomial class in a sense that it can borrow its byte storage (that's what the redundant std::span is for). I.e. also the std::vector in the PolynomialVector should be replaced by secure_vector then.

[randombit]

How much does this affect performance?

[randombit]

Actually @reneme isn't this always of known length namely Trait::N? Why not a std::array?

[reneme]

Actually @reneme isn't this always of known length namely Trait::N? Why not a std::array?

This was done to allow the Polynomial class to borrow its storage from the PolynomialVector, in which case the array would just be wasted stack space. Essentially, instances of polynomials can be free-standing (owning their own storage in the std::vector) or part of a vector of polynomials where the storage is coalesced into a single (appropriately larger) std::vector in the outer PolynomialVector instance. I'm not 100% sure anymore why we added this complexity, frankly. I believe it was faster.

Anyway, the computation is always implemented over the std::span<T, Trait::N> to let the compiler leverage the statically known bounds.

[coveralls]

coverage: 90.335% (-1.6%) from 91.974%
when pulling b312daa on polynomial-secmem
into b4032cd on master.

[falko-strenzke]

How much does this affect performance?

I tried to measure performance with the speed test. The natural variation of the timings is so large that observing any performance impact is impossible. If the performance aspect is really so important here, I would have to ask for the preferred exact method to measure it.