Implement Spectre Defenses

[defuse]

Given the existence of transient execution attacks Spectre and Meltdown, the inadequacies of current defenses, and unknowns about what other kinds of attacks researchers will discover in the future, it's probably best to tell users "assume your keys can be stolen by other users on the system" as we currently do in security-warnings.md.

However, GCC has a feature we might be able to take advantage of. Quoting from that paper:

Speculative Load Hardening (SLH) is an approach used
by  LLVM  and  was  proposed  by  Carruth  [12].   Using  this
idea, loads are checked using branchless code to ensure that
they  are  executing  along  a  valid  control  flow  path.   To  do
this,  they transform the code at the compiler level and in-
troduce a data dependency on the condition.  In the case of
misspeculation, the pointer is zeroed out, preventing it from
leaking data through speculative execution. One prerequisite
for this approach is hardware that allows implementation of
a branchless and unpredicted conditional update of a regis-
ter’s value. As of now, the feature is only available in LLVM
for x86 as the patch for ARM is still under review.   GCC
adopted the idea of SLH for their implementation, support-
ing both x86 and ARM. They provide a builtin function to
either emit a speculation barrier or return a safe value if it
determines that the instruction is transient [17]

[daira]

What's the typical performance impact of this approach?

(It's not that I value performance above security, but:

• this particular approach may or may not be effective;
• there might be other approaches with less performance impact;
• we need to know the performance impact anyway.)

Also, does this interact or is it mutually exclusive with other approaches to mitigating transient execution side channels or other attacks? (See the linked paper for examples of such interactions.)

[daira]

Note that according to Table 7 of the "Systematic Evaluation" paper, SLH only attempts to address Spectre-PHT attacks (the classic attack that tries to bypass bounds checks by exploiting the Pattern History Table, also called Spectre Variant 1).