ec/ecp_nistz256: harmonize is_infinity with ec_GFp_simple_is_at_infinity.

[openssl.git] / crypto / ec / asm / ecp_nistz256-armv8.pl

diff --git a/crypto/ec/asm/ecp_nistz256-armv8.pl b/crypto/ec/asm/ecp_nistz256-armv8.pl
index 13625861e86a4fdae1585e254077d23858a914d8..cdc91617ff13eea730a572238ff484d4597f23ec 100644 (file)
--- a/crypto/ec/asm/ecp_nistz256-armv8.pl
+++ b/crypto/ec/asm/ecp_nistz256-armv8.pl
@@ -862,46 +862,28 @@ ecp_nistz256_point_add:
        stp     x25,x26,[sp,#64]
        sub     sp,sp,#32*12
 
-       ldp     $a0,$a1,[$bp]
-       ldp     $a2,$a3,[$bp,#16]
-       ldp     $t0,$t1,[$bp,#32]
-       ldp     $t2,$t3,[$bp,#48]
+       ldp     $a0,$a1,[$bp,#64]       // in2_z
+       ldp     $a2,$a3,[$bp,#64+16]
         mov    $rp_real,$rp
         mov    $ap_real,$ap
         mov    $bp_real,$bp
-       orr     $a0,$a0,$a1
-       orr     $a2,$a2,$a3
-        ldp    $acc0,$acc1,[$ap]
-       orr     $t0,$t0,$t1
-       orr     $t2,$t2,$t3
-        ldp    $acc2,$acc3,[$ap,#16]
-       orr     $a0,$a0,$a2
-       orr     $t2,$t0,$t2
-        ldp    $t0,$t1,[$ap,#32]
-       orr     $in2infty,$a0,$t2
-       cmp     $in2infty,#0
-        ldp    $t2,$t3,[$ap,#48]
-       csetm   $in2infty,ne            // !in2infty
-
-        ldp    $a0,$a1,[$bp_real,#64]  // forward load for p256_sqr_mont
-       orr     $acc0,$acc0,$acc1
-       orr     $acc2,$acc2,$acc3
-        ldp    $a2,$a3,[$bp_real,#64+16]
-       orr     $t0,$t0,$t1
-       orr     $t2,$t2,$t3
-       orr     $acc0,$acc0,$acc2
-       orr     $t0,$t0,$t2
-       orr     $in1infty,$acc0,$t0
-       cmp     $in1infty,#0
         ldr    $poly1,.Lpoly+8
         ldr    $poly3,.Lpoly+24
-       csetm   $in1infty,ne            // !in1infty
-
+       orr     $t0,$a0,$a1
+       orr     $t2,$a2,$a3
+       orr     $in2infty,$t0,$t2
+       cmp     $in2infty,#0
+       csetm   $in2infty,ne            // !in2infty
        add     $rp,sp,#$Z2sqr
        bl      __ecp_nistz256_sqr_mont // p256_sqr_mont(Z2sqr, in2_z);
 
-       ldp     $a0,$a1,[$ap_real,#64]
+       ldp     $a0,$a1,[$ap_real,#64]  // in1_z
        ldp     $a2,$a3,[$ap_real,#64+16]
+       orr     $t0,$a0,$a1
+       orr     $t2,$a2,$a3
+       orr     $in1infty,$t0,$t2
+       cmp     $in1infty,#0
+       csetm   $in1infty,ne            // !in1infty
        add     $rp,sp,#$Z1sqr
        bl      __ecp_nistz256_sqr_mont // p256_sqr_mont(Z1sqr, in1_z);
 
@@ -1150,36 +1132,28 @@ ecp_nistz256_point_add_affine:
        ldr     $poly1,.Lpoly+8
        ldr     $poly3,.Lpoly+24
 
-       ldp     $a0,$a1,[$ap]
-       ldp     $a2,$a3,[$ap,#16]
-       ldp     $t0,$t1,[$ap,#32]
-       ldp     $t2,$t3,[$ap,#48]
-       orr     $a0,$a0,$a1
-       orr     $a2,$a2,$a3
-       orr     $t0,$t0,$t1
-       orr     $t2,$t2,$t3
-       orr     $a0,$a0,$a2
-       orr     $t0,$t0,$t2
-       orr     $in1infty,$a0,$t0
+       ldp     $a0,$a1,[$ap,#64]       // in1_z
+       ldp     $a2,$a3,[$ap,#64+16]
+       orr     $t0,$a0,$a1
+       orr     $t2,$a2,$a3
+       orr     $in1infty,$t0,$t2
        cmp     $in1infty,#0
        csetm   $in1infty,ne            // !in1infty
 
-       ldp     $a0,$a1,[$bp]
-       ldp     $a2,$a3,[$bp,#16]
-       ldp     $t0,$t1,[$bp,#32]
+       ldp     $acc0,$acc1,[$bp]       // in2_x
+       ldp     $acc2,$acc3,[$bp,#16]
+       ldp     $t0,$t1,[$bp,#32]       // in2_y
        ldp     $t2,$t3,[$bp,#48]
-       orr     $a0,$a0,$a1
-       orr     $a2,$a2,$a3
+       orr     $acc0,$acc0,$acc1
+       orr     $acc2,$acc2,$acc3
        orr     $t0,$t0,$t1
        orr     $t2,$t2,$t3
-       orr     $a0,$a0,$a2
+       orr     $acc0,$acc0,$acc2
        orr     $t0,$t0,$t2
-       orr     $in2infty,$a0,$t0
+       orr     $in2infty,$acc0,$t0
        cmp     $in2infty,#0
        csetm   $in2infty,ne            // !in2infty
 
-       ldp     $a0,$a1,[$ap_real,#64]
-       ldp     $a2,$a3,[$ap_real,#64+16]
        add     $rp,sp,#$Z1sqr
        bl      __ecp_nistz256_sqr_mont // p256_sqr_mont(Z1sqr, in1_z);