test_ec_compressed_points is very slow in SGX

[jethrogb]

This test takes over 20 seconds in SGX on my machine but less than 1 second on Linux. Changing the optimization level seems to have no effect.

[zugzwang]

Adding some info we gathered back then with @MihirLuthra.

MbedTLS implements comb methods for scalar multiplication (Algorithm 4 due to an attack by Louis Goubin, MbedTLS code). The first step is to precompute a list of generator multiples that is memory intensive (the list is thus fixed per curve).

To optimize one scalar mult, it could be useful to tune flags related to this algorithm (see MBEDTLS_ECP_WINDOW_SIZE (and COMB_MAX_PRE)).

To optimize several scalar mults: Precomputed points are cached within an EcGroup within each key. It would be great if MbedTLS reused static precomputation tables. This idea was discussed here and added to 3.X here.