Avoid leaking intermediate states in point doubling special case.

Cherry picked from
https://github.com/google/boringssl/commit/12d9ed670da3edd64ce8175cfe0e091982989c18

Reviewed-by: Nicola Tuveri <nic.tuv@gmail.com>
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/9239)

(cherry picked from commit 2baea7c7e0896658b74956cac6084dd7e82e8c1b)

diff --git a/crypto/ec/asm/ecp_nistz256-x86_64.pl b/crypto/ec/asm/ecp_nistz256-x86_64.pl
index 533da506adba12d83063e96dd7b70b15b51975b3..7b8329079e1e82eee872121c1d2aff1ec5a7ed36 100755 (executable)
--- a/crypto/ec/asm/ecp_nistz256-x86_64.pl
+++ b/crypto/ec/asm/ecp_nistz256-x86_64.pl
@@ -3626,17 +3626,26 @@ $code.=<<___;
 
        or      $acc5, $acc4                    # see if result is zero
        or      $acc0, $acc4
-       or      $acc1, $acc4
+       or      $acc1, $acc4                    # !is_equal(U1, U2)
 
-       .byte   0x3e                            # predict taken
-       jnz     .Ladd_proceed$x                 # is_equal(U1,U2)?
        movq    %xmm2, $acc0
        movq    %xmm3, $acc1
-       test    $acc0, $acc0
-       jnz     .Ladd_proceed$x                 # (in1infty || in2infty)?
+
+       or      $acc0, $acc4
+       .byte   0x3e                            # predict taken
+       jnz     .Ladd_proceed$x                 # !is_equal(U1, U2) || in1infty || in2infty
+
+       # We now know A = B or A = -B and neither is infinity. Compare the
+       # y-coordinates via S1 and S2.
        test    $acc1, $acc1
-       jz      .Ladd_double$x                  # is_equal(S1,S2)?
+       jz      .Ladd_double$x                  # is_equal(S1, S2)
 
+       # A = -B, so the result is infinity.
+       #
+       # TODO: see https://github.com/google/boringssl/blob/12d9ed670da3edd64ce8175cfe0e091982989c18/crypto/fipsmodule/ec/asm/p256-x86_64-asm.pl#L3128-L3132
+       # Does .Ladd_proceed handle this case? It seems to, in
+       # which case we should eliminate this special-case and simplify the
+       # timing analysis.
        movq    %xmm0, $r_ptr                   # restore $r_ptr
        pxor    %xmm0, %xmm0
        movdqu  %xmm0, 0x00($r_ptr)