sandy2x: align branch targets

src/libsodium/crypto_scalarmult/curve25519/sandy2x/fe51_nsquare.S

@@ -41,6 +41,8 @@ movq   %r9,16(%rdi)
 movq   %rax,24(%rdi)
 movq   %rsi,32(%rdi)
 mov  %rdx,%rsi
+
+.p2align 4
 ._loop:
 sub  $1,%rsi
 mov  %rcx,%rax
@@ -154,6 +156,7 @@ and  %rdx,%r10
 movq   %r10,32(%rdi)
 cmp  $0,%rsi
 jne ._loop
+
 movq   %rcx,0(%rdi)
 movq   %r8,8(%rdi)
 movq 0(%rsp),%r11

src/libsodium/crypto_scalarmult/curve25519/sandy2x/fe51_pack.S

@@ -35,6 +35,8 @@ movq   32(%rsi),%rsi
 movq REDMASK51(%rip),%rax
 lea  -18(%rax),%r10
 mov  $3,%r11
+
+.p2align 4
 ._reduceloop:
 mov  %rdx,%r12
 shr  $51,%r12
@@ -59,6 +61,7 @@ imulq  $19, %r12,%r12
 add  %r12,%rdx
 sub  $1,%r11
 ja ._reduceloop
+
 mov  $1,%r12
 cmp  %r10,%rdx
 cmovl %r11,%r12

src/libsodium/crypto_scalarmult/curve25519/sandy2x/ladder.S

@@ -110,6 +110,8 @@ xorq 16(%rsi),%r8
 xorq 24(%rsi),%r9
 leaq 800(%rsp),%rsi
 mov  $64,%rax
+
+.p2align 4
 ._ladder_small_loop:
 mov  %rdx,%r10
 mov  %rcx,%r11
@@ -136,6 +138,8 @@ sub  $1,%rax
 jne ._ladder_small_loop
 mov  $255,%rdx
 add  $760,%rsi
+
+.p2align 4
 ._ladder_loop:
 sub  $1,%rdx
 vbroadcastss 0(%rsi),%xmm10

src/libsodium/crypto_scalarmult/curve25519/sandy2x/ladder_base.S

@@ -61,6 +61,8 @@ xorq 16(%rsi),%r8
 xorq 24(%rsi),%r9
 leaq 512(%rsp),%rsi
 mov  $64,%rax
+
+.p2align 4
 ._ladder_base_small_loop:
 mov  %rdx,%r10
 mov  %rcx,%r11
@@ -87,6 +89,8 @@ sub  $1,%rax
 jne ._ladder_base_small_loop
 mov  $255,%rdx
 add  $760,%rsi
+
+.p2align 4
 ._ladder_base_loop:
 sub  $1,%rdx
 vbroadcastss 0(%rsi),%xmm10