The previous code used comparison operators >= and == that are quite likely to
be compiled to branches by some compilers on some architectures (with some
optimisation levels).
For example, take the following function:
void old_update( size_t data_len, size_t *padlen )
{
*padlen *= ( data_len >= *padlen + 1 );
}
With Clang 3.8, let's compile it for the Arm v6-M architecture:
% clang --target=arm-none-eabi -march=armv6-m -Os foo.c -S -o - |
sed -n '/^old_update:$/,/\.size/p'
old_update:
.fnstart
@ BB#0:
.save {r4, lr}
push {r4, lr}
ldr r2, [r1]
adds r4, r2, #1
movs r3, #0
cmp r4, r0
bls .LBB0_2
@ BB#1:
mov r2, r3
.LBB0_2:
str r2, [r1]
pop {r4, pc}
.Lfunc_end0:
.size old_update, .Lfunc_end0-old_update
We can see an unbalanced secret-dependant branch, resulting in a total
execution time depends on the value of the secret (here padlen) in a
straightforward way.
The new version, based on bit operations, doesn't have this issue:
new_update:
.fnstart
@ BB#0:
ldr r2, [r1]
subs r0, r0, #1
subs r0, r0, r2
asrs r0, r0, #31
bics r2, r0
str r2, [r1]
bx lr
.Lfunc_end1:
.size new_update, .Lfunc_end1-new_update
(As a bonus, it's smaller and uses less stack.)
While there's no formal guarantee that the version based on bit operations in
C won't be translated using branches by the compiler, experiments tend to show
that's the case [1], and it is commonly accepted knowledge in the practical
crypto community that if we want to sick to C, bit operations are the safest
bet [2].
[1] https://github.com/mpg/ct/blob/master/results
[2] https://github.com/veorq/cryptocoding
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
2ddec4306f