Implement optimized square function for nacl.

diff --git a/src/lib/nano-nacl.ts b/src/lib/nano-nacl.ts
index 789f52adffeae0e29a43ab4819a8d03330033242..1840a11ff170810343c1a91e2122f63a6f1fb128 100644 (file)
--- a/src/lib/nano-nacl.ts
+++ b/src/lib/nano-nacl.ts
@@ -139,7 +139,8 @@ export class NanoNaCl {
        }\r
 \r
        static Multiply (o: Float64Array, a: Float64Array, b: Float64Array): void {\r
-               let v, c, s = 1 << 16\r
+               let v, c\r
+               const s = 1 << 16\r
                const t = new Array(31)\r
                t.fill(0)\r
 \r
@@ -178,7 +179,46 @@ export class NanoNaCl {
        }\r
 \r
        static Square (o: Float64Array, a: Float64Array): void {\r
-               this.Multiply(o, a, a)\r
+               let v, c\r
+               const s = 1 << 16\r
+               const t = new Array(31)\r
+               t.fill(0)\r
+\r
+               // init t values, same as Multiply except we can skip some iterations of\r
+               // the inner loop since a[x]*a[y] + a[y]*a[x] = 2*a[x]*a[y]\r
+               for (let i = 0; i < 16; i++) {\r
+                       for (let j = i; j < 16; j++) {\r
+                               //@ts-expect-error\r
+                               t[i + j] += a[i] * a[j] * ((i < j) + 1)\r
+                       }\r
+               }\r
+\r
+               for (let i = 0; i < 15; i++) {\r
+                       t[i] += 38 * t[i + 16]\r
+               }\r
+               // t15 left as is\r
+\r
+               // first carry\r
+               c = 1\r
+               for (let i = 0; i < 16; i++) {\r
+                       v = t[i] + c + s - 1\r
+                       c = Math.floor(v / s)\r
+                       t[i] = v - c * s\r
+               }\r
+               t[0] += 38 * (c - 1)\r
+\r
+               // second carry\r
+               c = 1\r
+               for (let i = 0; i < 16; i++) {\r
+                       v = t[i] + c + s - 1\r
+                       c = Math.floor(v / s)\r
+                       t[i] = v - c * s\r
+               }\r
+               t[0] += 38 * (c - 1)\r
+\r
+               // assign result to output\r
+               o.set(t.slice(0, 16), 0)\r
+               // this.Multiply(o, a, a)\r
        }\r
 \r
        static inv25519 (o: Float64Array, i: Float64Array): void {\r