core/crypto/aegis: Initial import

core/crypto/_aes/ct64/ct64.odin

@@ -210,11 +210,8 @@ orthogonalize :: proc "contextless" (q: ^[8]u64) {
 }
 
 @(require_results)
-interleave_in :: proc "contextless" (w: []u32) -> (q0, q1: u64) #no_bounds_check {
-	if len(w) < 4 {
-		panic_contextless("aes/ct64: invalid input size")
-	}
-	x0, x1, x2, x3 := u64(w[0]), u64(w[1]), u64(w[2]), u64(w[3])
+interleave_in :: proc "contextless" (w0, w1, w2, w3: u32) -> (q0, q1: u64) #no_bounds_check {
+	x0, x1, x2, x3 := u64(w0), u64(w1), u64(w2), u64(w3)
 	x0 |= (x0 << 16)
 	x1 |= (x1 << 16)
 	x2 |= (x2 << 16)

core/crypto/_aes/ct64/ct64_keysched.odin

@@ -77,7 +77,7 @@ keysched :: proc(comp_skey: []u64, key: []byte) -> int {
 
 	q: [8]u64 = ---
 	for i, j := 0, 0; i < nkf; i, j = i + 4, j + 2 {
-		q[0], q[4] = interleave_in(skey[i:])
+		q[0], q[4] = interleave_in(skey[i], skey[i+1], skey[i+2], skey[i+3])
 		q[1] = q[0]
 		q[2] = q[0]
 		q[3] = q[0]
@@ -122,57 +122,3 @@ skey_expand :: proc "contextless" (skey, comp_skey: []u64, num_rounds: int) {
 		skey[v + 3] = (x3 << 4) - x3
 	}
 }
-
-orthogonalize_roundkey :: proc "contextless" (qq: []u64, key: []byte) {
-	if len(qq) < 8 || len(key) != 16 {
-		panic_contextless("aes/ct64: invalid round key size")
-	}
-
-	skey: [4]u32 = ---
-	skey[0] = endian.unchecked_get_u32le(key[0:])
-	skey[1] = endian.unchecked_get_u32le(key[4:])
-	skey[2] = endian.unchecked_get_u32le(key[8:])
-	skey[3] = endian.unchecked_get_u32le(key[12:])
-
-	q: [8]u64 = ---
-	q[0], q[4] = interleave_in(skey[:])
-	q[1] = q[0]
-	q[2] = q[0]
-	q[3] = q[0]
-	q[5] = q[4]
-	q[6] = q[4]
-	q[7] = q[4]
-	orthogonalize(&q)
-
-	comp_skey: [2]u64 = ---
-	comp_skey[0] =
-		(q[0] & 0x1111111111111111) |
-		(q[1] & 0x2222222222222222) |
-		(q[2] & 0x4444444444444444) |
-		(q[3] & 0x8888888888888888)
-	comp_skey[1] =
-		(q[4] & 0x1111111111111111) |
-		(q[5] & 0x2222222222222222) |
-		(q[6] & 0x4444444444444444) |
-		(q[7] & 0x8888888888888888)
-
-	for x, u in comp_skey {
-		x0 := x
-		x1, x2, x3 := x0, x0, x0
-		x0 &= 0x1111111111111111
-		x1 &= 0x2222222222222222
-		x2 &= 0x4444444444444444
-		x3 &= 0x8888888888888888
-		x1 >>= 1
-		x2 >>= 2
-		x3 >>= 3
-		qq[u * 4 + 0] = (x0 << 4) - x0
-		qq[u * 4 + 1] = (x1 << 4) - x1
-		qq[u * 4 + 2] = (x2 << 4) - x2
-		qq[u * 4 + 3] = (x3 << 4) - x3
-	}
-
-	mem.zero_explicit(&skey, size_of(skey))
-	mem.zero_explicit(&q, size_of(q))
-	mem.zero_explicit(&comp_skey, size_of(comp_skey))
-}

core/crypto/_aes/ct64/helpers.odin

@@ -3,17 +3,39 @@ package aes_ct64
 import "core:crypto/_aes"
 import "core:encoding/endian"
 
+@(require_results)
+load_interleaved :: proc "contextless" (src: []byte) -> (u64, u64) {
+	w0 := endian.unchecked_get_u32le(src[0:])
+	w1 := endian.unchecked_get_u32le(src[4:])
+	w2 := endian.unchecked_get_u32le(src[8:])
+	w3 := endian.unchecked_get_u32le(src[12:])
+	return interleave_in(w0, w1, w2, w3)
+}
+
+store_interleaved :: proc "contextless" (dst: []byte, a0, a1: u64) {
+	w0, w1, w2, w3 := interleave_out(a0, a1)
+	endian.unchecked_put_u32le(dst[0:], w0)
+	endian.unchecked_put_u32le(dst[4:], w1)
+	endian.unchecked_put_u32le(dst[8:], w2)
+	endian.unchecked_put_u32le(dst[12:], w3)
+}
+
+@(require_results)
+xor_interleaved :: #force_inline proc "contextless" (a0, a1, b0, b1: u64) -> (u64, u64) {
+	return a0 ~ b0, a1 ~ b1
+}
+
+@(require_results)
+and_interleaved :: #force_inline proc "contextless" (a0, a1, b0, b1: u64) -> (u64, u64) {
+	return a0 & b0, a1 & b1
+}
+
 load_blockx1 :: proc "contextless" (q: ^[8]u64, src: []byte) {
 	if len(src) != _aes.BLOCK_SIZE {
 		panic_contextless("aes/ct64: invalid block size")
 	}
 
-	w: [4]u32 = ---
-	w[0] = endian.unchecked_get_u32le(src[0:])
-	w[1] = endian.unchecked_get_u32le(src[4:])
-	w[2] = endian.unchecked_get_u32le(src[8:])
-	w[3] = endian.unchecked_get_u32le(src[12:])
-	q[0], q[4] = interleave_in(w[:])
+	q[0], q[4] = #force_inline load_interleaved(src)
 	orthogonalize(q)
 }
 
@@ -23,11 +45,7 @@ store_blockx1 :: proc "contextless" (dst: []byte, q: ^[8]u64) {
 	}
 
 	orthogonalize(q)
-	w0, w1, w2, w3 := interleave_out(q[0], q[4])
-	endian.unchecked_put_u32le(dst[0:], w0)
-	endian.unchecked_put_u32le(dst[4:], w1)
-	endian.unchecked_put_u32le(dst[8:], w2)
-	endian.unchecked_put_u32le(dst[12:], w3)
+	#force_inline store_interleaved(dst, q[0], q[4])
 }
 
 load_blocks :: proc "contextless" (q: ^[8]u64, src: [][]byte) {
@@ -35,17 +53,11 @@ load_blocks :: proc "contextless" (q: ^[8]u64, src: [][]byte) {
 		panic_contextless("aes/ct64: invalid block(s) size")
 	}
 
-	w: [4]u32 = ---
 	for s, i in src {
 		if len(s) != _aes.BLOCK_SIZE {
 			panic_contextless("aes/ct64: invalid block size")
 		}
-
-		w[0] = endian.unchecked_get_u32le(s[0:])
-		w[1] = endian.unchecked_get_u32le(s[4:])
-		w[2] = endian.unchecked_get_u32le(s[8:])
-		w[3] = endian.unchecked_get_u32le(s[12:])
-		q[i], q[i + 4] = interleave_in(w[:])
+		q[i], q[i + 4] = #force_inline load_interleaved(s)
 	}
 	orthogonalize(q)
 }
@@ -64,11 +76,6 @@ store_blocks :: proc "contextless" (dst: [][]byte, q: ^[8]u64) {
 		if len(d) != _aes.BLOCK_SIZE {
 			panic_contextless("aes/ct64: invalid block size")
 		}
-
-		w0, w1, w2, w3 := interleave_out(q[i], q[i + 4])
-		endian.unchecked_put_u32le(d[0:], w0)
-		endian.unchecked_put_u32le(d[4:], w1)
-		endian.unchecked_put_u32le(d[8:], w2)
-		endian.unchecked_put_u32le(d[12:], w3)
+		#force_inline store_interleaved(d, q[i], q[i + 4])
 	}
 }