ec/asm/ecp_nistz256-x86_64.pl: addition to perform stricter reduction.

Addition was not preserving inputs' property of being fully reduced.

Thanks to Brian Smith for reporting this.

Reviewed-by: Rich Salz <rsalz@openssl.org>
(cherry picked from commit b62b2454fadfccaf5e055a1810d72174c2633b8f)

2 files changed, 84 insertions, 62 deletions

diff --git a/crypto/ec/asm/ecp_nistz256-x86_64.pl b/crypto/ec/asm/ecp_nistz256-x86_64.pl
index 7140860e24..2b82a81a36 100755
--- a/crypto/ec/asm/ecp_nistz256-x86_64.pl
+++ b/crypto/ec/asm/ecp_nistz256-x86_64.pl
@@ -128,6 +128,7 @@ ecp_nistz256_mul_by_2:
 	push	%r13
 
 	mov	8*0($a_ptr), $a0
+	xor	$t4,$t4
 	mov	8*1($a_ptr), $a1
 	add	$a0, $a0		# a0:a3+a0:a3
 	mov	8*2($a_ptr), $a2
@@ -138,7 +139,7 @@ ecp_nistz256_mul_by_2:
 	adc	$a2, $a2
 	adc	$a3, $a3
 	 mov	$a1, $t1
-	sbb	$t4, $t4
+	adc	\$0, $t4
 
 	sub	8*0($a_ptr), $a0
 	 mov	$a2, $t2
@@ -146,14 +147,14 @@ ecp_nistz256_mul_by_2:
 	sbb	8*2($a_ptr), $a2
 	 mov	$a3, $t3
 	sbb	8*3($a_ptr), $a3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $a0
-	cmovz	$t1, $a1
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
 	mov	$a0, 8*0($r_ptr)
-	cmovz	$t2, $a2
+	cmovb	$t2, $a2
 	mov	$a1, 8*1($r_ptr)
-	cmovz	$t3, $a3
+	cmovb	$t3, $a3
 	mov	$a2, 8*2($r_ptr)
 	mov	$a3, 8*3($r_ptr)
 
@@ -250,12 +251,12 @@ ecp_nistz256_mul_by_3:
 	sbb	\$0, $a2
 	 mov	$a3, $t3
 	sbb	.Lpoly+8*3(%rip), $a3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $a0
-	cmovz	$t1, $a1
-	cmovz	$t2, $a2
-	cmovz	$t3, $a3
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
+	cmovb	$t2, $a2
+	cmovb	$t3, $a3
 
 	xor	$t4, $t4
 	add	8*0($a_ptr), $a0	# a0:a3+=a_ptr[0:3]
@@ -272,14 +273,14 @@ ecp_nistz256_mul_by_3:
 	sbb	\$0, $a2
 	 mov	$a3, $t3
 	sbb	.Lpoly+8*3(%rip), $a3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $a0
-	cmovz	$t1, $a1
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
 	mov	$a0, 8*0($r_ptr)
-	cmovz	$t2, $a2
+	cmovb	$t2, $a2
 	mov	$a1, 8*1($r_ptr)
-	cmovz	$t3, $a3
+	cmovb	$t3, $a3
 	mov	$a2, 8*2($r_ptr)
 	mov	$a3, 8*3($r_ptr)
 
@@ -318,14 +319,14 @@ ecp_nistz256_add:
 	sbb	8*2($a_ptr), $a2
 	 mov	$a3, $t3
 	sbb	8*3($a_ptr), $a3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $a0
-	cmovz	$t1, $a1
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
 	mov	$a0, 8*0($r_ptr)
-	cmovz	$t2, $a2
+	cmovb	$t2, $a2
 	mov	$a1, 8*1($r_ptr)
-	cmovz	$t3, $a3
+	cmovb	$t3, $a3
 	mov	$a2, 8*2($r_ptr)
 	mov	$a3, 8*3($r_ptr)
 
@@ -1840,13 +1841,14 @@ $code.=<<___;
 .type	__ecp_nistz256_add_toq,\@abi-omnipotent
 .align	32
 __ecp_nistz256_add_toq:
+	xor	$t4,$t4
 	add	8*0($b_ptr), $a0
 	adc	8*1($b_ptr), $a1
 	 mov	$a0, $t0
 	adc	8*2($b_ptr), $a2
 	adc	8*3($b_ptr), $a3
 	 mov	$a1, $t1
-	sbb	$t4, $t4
+	adc	\$0, $t4
 
 	sub	\$-1, $a0
 	 mov	$a2, $t2
@@ -1854,14 +1856,14 @@ __ecp_nistz256_add_toq:
 	sbb	\$0, $a2
 	 mov	$a3, $t3
 	sbb	$poly3, $a3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $a0
-	cmovz	$t1, $a1
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
 	mov	$a0, 8*0($r_ptr)
-	cmovz	$t2, $a2
+	cmovb	$t2, $a2
 	mov	$a1, 8*1($r_ptr)
-	cmovz	$t3, $a3
+	cmovb	$t3, $a3
 	mov	$a2, 8*2($r_ptr)
 	mov	$a3, 8*3($r_ptr)
 
@@ -1929,13 +1931,14 @@ __ecp_nistz256_subq:
 .type	__ecp_nistz256_mul_by_2q,\@abi-omnipotent
 .align	32
 __ecp_nistz256_mul_by_2q:
+	xor	$t4, $t4
 	add	$a0, $a0		# a0:a3+a0:a3
 	adc	$a1, $a1
 	 mov	$a0, $t0
 	adc	$a2, $a2
 	adc	$a3, $a3
 	 mov	$a1, $t1
-	sbb	$t4, $t4
+	adc	\$0, $t4
 
 	sub	\$-1, $a0
 	 mov	$a2, $t2
@@ -1943,14 +1946,14 @@ __ecp_nistz256_mul_by_2q:
 	sbb	\$0, $a2
 	 mov	$a3, $t3
 	sbb	$poly3, $a3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $a0
-	cmovz	$t1, $a1
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
 	mov	$a0, 8*0($r_ptr)
-	cmovz	$t2, $a2
+	cmovb	$t2, $a2
 	mov	$a1, 8*1($r_ptr)
-	cmovz	$t3, $a3
+	cmovb	$t3, $a3
 	mov	$a2, 8*2($r_ptr)
 	mov	$a3, 8*3($r_ptr)
 
@@ -2405,6 +2408,7 @@ $code.=<<___;
 	#lea	$Hsqr(%rsp), $r_ptr	# 2*U1*H^2
 	#call	__ecp_nistz256_mul_by_2	# ecp_nistz256_mul_by_2(Hsqr, U2);
 
+	xor	$t4, $t4
 	add	$acc0, $acc0		# a0:a3+a0:a3
 	lea	$Rsqr(%rsp), $a_ptr
 	adc	$acc1, $acc1
@@ -2412,7 +2416,7 @@ $code.=<<___;
 	adc	$acc2, $acc2
 	adc	$acc3, $acc3
 	 mov	$acc1, $t1
-	sbb	$t4, $t4
+	adc	\$0, $t4
 
 	sub	\$-1, $acc0
 	 mov	$acc2, $t2
@@ -2420,15 +2424,15 @@ $code.=<<___;
 	sbb	\$0, $acc2
 	 mov	$acc3, $t3
 	sbb	$poly3, $acc3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $acc0
+	cmovb	$t0, $acc0
 	mov	8*0($a_ptr), $t0
-	cmovz	$t1, $acc1
+	cmovb	$t1, $acc1
 	mov	8*1($a_ptr), $t1
-	cmovz	$t2, $acc2
+	cmovb	$t2, $acc2
 	mov	8*2($a_ptr), $t2
-	cmovz	$t3, $acc3
+	cmovb	$t3, $acc3
 	mov	8*3($a_ptr), $t3
 
 	call	__ecp_nistz256_sub$x		# p256_sub(res_x, Rsqr, Hsqr);
@@ -2710,6 +2714,7 @@ $code.=<<___;
 	#lea	$Hsqr(%rsp), $r_ptr	# 2*U1*H^2
 	#call	__ecp_nistz256_mul_by_2	# ecp_nistz256_mul_by_2(Hsqr, U2);
 
+	xor	$t4, $t4
 	add	$acc0, $acc0		# a0:a3+a0:a3
 	lea	$Rsqr(%rsp), $a_ptr
 	adc	$acc1, $acc1
@@ -2717,7 +2722,7 @@ $code.=<<___;
 	adc	$acc2, $acc2
 	adc	$acc3, $acc3
 	 mov	$acc1, $t1
-	sbb	$t4, $t4
+	adc	\$0, $t4
 
 	sub	\$-1, $acc0
 	 mov	$acc2, $t2
@@ -2725,15 +2730,15 @@ $code.=<<___;
 	sbb	\$0, $acc2
 	 mov	$acc3, $t3
 	sbb	$poly3, $acc3
-	test	$t4, $t4
+	sbb	\$0, $t4
 
-	cmovz	$t0, $acc0
+	cmovb	$t0, $acc0
 	mov	8*0($a_ptr), $t0
-	cmovz	$t1, $acc1
+	cmovb	$t1, $acc1
 	mov	8*1($a_ptr), $t1
-	cmovz	$t2, $acc2
+	cmovb	$t2, $acc2
 	mov	8*2($a_ptr), $t2
-	cmovz	$t3, $acc3
+	cmovb	$t3, $acc3
 	mov	8*3($a_ptr), $t3
 
 	call	__ecp_nistz256_sub$x		# p256_sub(res_x, Rsqr, Hsqr);
@@ -2885,14 +2890,14 @@ __ecp_nistz256_add_tox:
 	sbb	\$0, $a2
 	 mov	$a3, $t3
 	sbb	$poly3, $a3
+	sbb	\$0, $t4
 
-	bt	\$0, $t4
-	cmovnc	$t0, $a0
-	cmovnc	$t1, $a1
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
 	mov	$a0, 8*0($r_ptr)
-	cmovnc	$t2, $a2
+	cmovb	$t2, $a2
 	mov	$a1, 8*1($r_ptr)
-	cmovnc	$t3, $a3
+	cmovb	$t3, $a3
 	mov	$a2, 8*2($r_ptr)
 	mov	$a3, 8*3($r_ptr)
 
@@ -2980,14 +2985,14 @@ __ecp_nistz256_mul_by_2x:
 	sbb	\$0, $a2
 	 mov	$a3, $t3
 	sbb	$poly3, $a3
+	sbb	\$0, $t4
 
-	bt	\$0, $t4
-	cmovnc	$t0, $a0
-	cmovnc	$t1, $a1
+	cmovb	$t0, $a0
+	cmovb	$t1, $a1
 	mov	$a0, 8*0($r_ptr)
-	cmovnc	$t2, $a2
+	cmovb	$t2, $a2
 	mov	$a1, 8*1($r_ptr)
-	cmovnc	$t3, $a3
+	cmovb	$t3, $a3
 	mov	$a2, 8*2($r_ptr)
 	mov	$a3, 8*3($r_ptr)
 
diff --git a/crypto/ec/ecp_nistz256.c b/crypto/ec/ecp_nistz256.c
index 1459980e87..7687490bfd 100644
--- a/crypto/ec/ecp_nistz256.c
+++ b/crypto/ec/ecp_nistz256.c
@@ -82,19 +82,36 @@ typedef struct ec_pre_comp_st {
 } EC_PRE_COMP;
 
 /* Functions implemented in assembly */
+/*
+ * Most of below mentioned functions *preserve* the property of inputs
+ * being fully reduced, i.e. being in [0, modulus) range. Simply put if
+ * inputs are fully reduced, then output is too. Note that reverse is
+ * not true, in sense that given partially reduced inputs output can be
+ * either, not unlikely reduced. And "most" in first sentence refers to
+ * the fact that given the calculations flow one can tolerate that
+ * addition, 1st function below, produces partially reduced result *if*
+ * multiplications by 2 and 3, which customarily use addition, fully
+ * reduce it. This effectively gives two options: a) addition produces
+ * fully reduced result [as long as inputs are, just like remaining
+ * functions]; b) addition is allowed to produce partially reduced
+ * result, but multiplications by 2 and 3 perform additional reduction
+ * step. Choice between the two can be platform-specific, but it was a)
+ * in all cases so far...
+ */
+/* Modular add: res = a+b mod P   */
+void ecp_nistz256_add(BN_ULONG res[P256_LIMBS],
+                      const BN_ULONG a[P256_LIMBS],
+                      const BN_ULONG b[P256_LIMBS]);
 /* Modular mul by 2: res = 2*a mod P */
 void ecp_nistz256_mul_by_2(BN_ULONG res[P256_LIMBS],
                            const BN_ULONG a[P256_LIMBS]);
-/* Modular div by 2: res = a/2 mod P */
-void ecp_nistz256_div_by_2(BN_ULONG res[P256_LIMBS],
-                           const BN_ULONG a[P256_LIMBS]);
 /* Modular mul by 3: res = 3*a mod P */
 void ecp_nistz256_mul_by_3(BN_ULONG res[P256_LIMBS],
                            const BN_ULONG a[P256_LIMBS]);
-/* Modular add: res = a+b mod P   */
-void ecp_nistz256_add(BN_ULONG res[P256_LIMBS],
-                      const BN_ULONG a[P256_LIMBS],
-                      const BN_ULONG b[P256_LIMBS]);
+
+/* Modular div by 2: res = a/2 mod P */
+void ecp_nistz256_div_by_2(BN_ULONG res[P256_LIMBS],
+                           const BN_ULONG a[P256_LIMBS]);
 /* Modular sub: res = a-b mod P   */
 void ecp_nistz256_sub(BN_ULONG res[P256_LIMBS],
                       const BN_ULONG a[P256_LIMBS],