Re:Some small suggestions for the Intel instruction set
by haldean on Sunday May 18, 2014 @12:00PM
(#7763383)
I believe that's his point; I read that section not as "please make these instructions constant-time" and more as "please document whether these instructions are constant-time, so that we can rely on them being constant-time". If an instruction is constant time but then a new chipset comes out with early-exit semantics for multiplication, everything previously compiled for that architecture is suddenly vulnerable to timing attacks when it wasn't previously. He's just looking for a commitment for the leak and timing semantics of various instructions, even if the commitment is "you cannot rely upon this being constant-time".
I disagree with this and think he has it backwards; the majority of applications will benefit from early-out multiplication, and it's only crypto that wants constant-time. Also, MUL has never been constant-time on x86, ever since the 8086. My suggestion would be to add a control register bit that puts the CPU into a mode where the guarantees needed for crypto are true. All that software would need to do is set the bit before performing sensitive operations, and clear it afterwards.