symm/salsa20-*.S: Fix vector diagrams to be little-endian.

Missed these in the previous pass.

diff --git a/symm/salsa20-arm-neon.S b/symm/salsa20-arm-neon.S
index 9725e8a..334e182 100644 (file)
--- a/symm/salsa20-arm-neon.S
+++ b/symm/salsa20-arm-neon.S
@@ -204,7 +204,7 @@ FUNC(salsa20_core_arm_neon)
        // Almost there.  Firstly the feedfoward addition.  Also, establish a
        // constant which will be useful later.
 9:     vadd.u32 q0, q8, q12                    //  0,  5, 10, 15
-        vmov.i64 q12, #0xffffffff              // = (-1, 0, -1, 0)
+        vmov.i64 q12, #0xffffffff              // = (0, -1; 0, -1)
        vadd.u32 q1, q9, q13                    //  4,  9, 14,  3
        vadd.u32 q2, q10, q14                   //  8, 13,  2,  7
        vadd.u32 q3, q11, q15                   // 12,  1,  6, 11

diff --git a/symm/salsa20-arm64.S b/symm/salsa20-arm64.S
index a71e53b..a3bce4e 100644 (file)
--- a/symm/salsa20-arm64.S
+++ b/symm/salsa20-arm64.S
@@ -203,8 +203,8 @@ FUNC(salsa20_core_arm64)
        // Almost there.  Firstly the feedfoward addition.  Also, establish
        // constants which will be useful later.
 9:     add     v0.4s, v0.4s, v4.4s             //  0,  5, 10, 15
-        movi   v16.2d, #0xffffffff             // = (-1, 0, -1, 0)
-        movi   d17, #-1                        // = (-1, -1, 0, 0)
+        movi   v16.2d, #0xffffffff             // = (0, -1; 0, -1)
+        movi   d17, #-1                        // = (0, 0; -1, -1)
        add     v1.4s, v1.4s, v5.4s             //  4,  9, 14,  3
        add     v2.4s, v2.4s, v6.4s             //  8, 13,  2,  7
        add     v3.4s, v3.4s, v7.4s             // 12,  1,  6, 11