optimized speed and memory usage of salsa20.d and actions.d. Some more small fixes

source/actions.d

@@ -10,6 +10,7 @@ import cipher;
 public:
 
 enum chunkSize = 4096;
+enum base64ChunkSize = 57;
 
 enum string randomDeviceName = "/dev/random";
 
@@ -19,43 +20,70 @@ void encrypt(string keyFileName, Cipher cipher, bool armor)
     ubyte[8] nonce;
     if (armor)
     {
-        ubyte[] buf;
+        ubyte[base64ChunkSize] buf;
+        char[Base64.encodeLength(base64ChunkSize)] buf2;
+        int counter;
         foreach (b; stdin.byChunk(chunkSize).joiner.cipherFunction(key, nonce, cipher))
         {
-            buf ~= [b];
+            buf[counter] = b;
-            if (buf.length == 57)
+            ++counter;
+            if (counter == 57)
             {
-                stdout.writeln(Base64.encode(buf));
+                Base64.encode(buf, buf2[]);
-                buf = [];
+                stdout.writeln(buf2);
+                counter = 0;
             }
         }
-        if (buf !is null)
+        if (counter > 0)
-            stdout.writeln(Base64.encode(buf));
+        {
+            Base64.encode(buf[0..counter], buf2[]);
+            stdout.writeln(buf2[0..Base64.encodeLength(counter)]);
+        }
     }
     else
     {
+        ubyte[chunkSize] buf;
+        int counter;
         foreach (b; stdin.byChunk(chunkSize).joiner.cipherFunction(key, nonce, cipher))
-            stdout.rawWrite([b]);
+        {
+            buf[counter] = b;
+            ++counter;
+            if (counter == chunkSize)
+            {
+                stdout.rawWrite(buf);
+                counter = 0;
+            }
+        }
+        if (counter > 0)
+            stdout.rawWrite(buf[0..counter]);
     }
 }
 
 void decrypt(string keyFileName, Cipher cipher, bool armor)
 {
+    import std.range;
     auto key = loadKey(keyFileName, armor);
     ubyte[8] nonce;
+    ubyte[chunkSize] buf;
+    int counter;
+    InputRange!ubyte r;
     if (armor)
+        r = Base64.decoder(stdin.byLine).joiner.cipherFunction(key, nonce, cipher).inputRangeObject;
+    else
+        r = stdin.byChunk(chunkSize).joiner.cipherFunction(key, nonce, cipher).inputRangeObject;
+
+    foreach (b; r)
     {
-        ubyte[] buf;
+        buf[counter] = b;
-        foreach (b; Base64.decoder(stdin.byLine).joiner.cipherFunction(key, nonce, cipher))
+        ++counter;
+        if (counter == 4096)
         {
-            stdout.rawWrite([b]);
+            stdout.rawWrite(buf);
+            counter = 0;
         }
     }
-    else
+    if(counter > 0)
-    {
+        stdout.rawWrite(buf[0..counter]);
-        foreach (b; stdin.byChunk(chunkSize).joiner.cipherFunction(key, nonce, cipher))
-            stdout.rawWrite([b]);
-    }
 }
 
 void generateKey(bool armor)
@@ -90,12 +118,11 @@ ubyte[32] loadKey(string filename, bool armor)
     scope (exit)
         keyFile.close();
     ubyte[32] key;
+    ubyte[33] temp;
     if (armor)
     {
-        ubyte[] tempKey;
+        import std.string : chomp;
-        foreach (line; keyFile.byLine)
+        key = Base64.decode(keyFile.readln.chomp, temp[]);
-            tempKey ~= Base64.decode(line);
-        key = tempKey;
     }
     else
     {

source/app.d

@@ -7,7 +7,7 @@ import actions;
 
 int main(string[] args)
 {
-    bool[string] actions = ["genKey"  : false, "encrypt" : false, "decrypt" : false];
+    bool[string] actions = ["genKey" : false, "encrypt" : false, "decrypt" : false];
 
     Cipher cipher = Cipher.chacha20;
     string keyFileName = "symkey.asc";
@@ -16,14 +16,13 @@ int main(string[] args)
     GetoptResult result;
     try
     {
-        result = getopt(args,
+        result = getopt(args, std.getopt.config.bundling, "gen-key|g",
-                        std.getopt.config.bundling,
+                "Generate a new 256 bit key.", &actions["genKey"],
-                        "gen-key|g", "Generate a new 256 bit key.", &actions["genKey"],
+                "encrypt|e", "Encrypt a message.", &actions["encrypt"],
-                        "encrypt|e", "Encrypt a message.", &actions["encrypt"],
+                "decrypt|d", "Decrypt a message.", &actions["decrypt"],
-                        "decrypt|d", "Decrypt a message.", &actions["decrypt"],
+                "cipher|c", "The cipher to use (default: %s).".format(cipher), &cipher, "key|k",
-                        "cipher|c", "The cipher to use (default: %s).".format(cipher), &cipher,
+                "The file which contains the key (default: %s).".format(keyFileName),
-                        "key|k", "The file which contains the key (default: %s).".format(keyFileName), &keyFileName,
+                &keyFileName, "armor|a", "use ascii-armored I/O.", &armor);
-                        "armor|a", "use ascii-armored I/O.", &armor);
     }
     catch (Exception e)
     {
@@ -71,12 +70,8 @@ int main(string[] args)
 
 void printHelp(Option[] options)
 {
-    defaultGetoptPrinter("Usage: ./learncrypt [options]\n\nCommon options:",
+    defaultGetoptPrinter("Usage: ./learncrypt [options]\n\nCommon options:", options[$ - 1 .. $]);
-                         options[$-1..$]);
+    defaultGetoptPrinter("\nGlobal options:", options[$ - 2 .. $ - 1]);
-    defaultGetoptPrinter("\nGlobal options:",
+    defaultGetoptPrinter("\nActions:", options[0 .. 3]);
-                         options[$-2..$-1]);
+    defaultGetoptPrinter("\nAction options:", options[3 .. 5]);
-    defaultGetoptPrinter("\nActions:",
-                         options[0..3]);
-    defaultGetoptPrinter("\nAction options:",
-                         options[3..5]);
 }

source/cipher.d

@@ -1,9 +1,12 @@
 module cipher;
 
-private import std.range : isInputRange, isForwardRange, ElementType, InputRange, ForwardRange, inputRangeObject;
+private import std.range : isInputRange, isForwardRange, ElementType,
+    InputRange, ForwardRange, inputRangeObject;
+
 private import std.string : format;
 
 private import salsa20;
+
 private import chacha20;
 
 public:
@@ -15,9 +18,9 @@ enum Cipher
 }
 
 template cipherFunction(R)
-if(isInputRange!R && is(ElementType!R : ubyte))
+    if (isInputRange!R && is(ElementType!R : ubyte))
 {
-    static if(isForwardRange!R)
+    static if (isForwardRange!R)
         alias ReturnType = ForwardRange;
     else
         alias ReturnType = InputRange;
@@ -26,10 +29,10 @@ if(isInputRange!R && is(ElementType!R : ubyte))
     {
         final switch (cipher)
         {
-            case Cipher.salsa20:
+        case Cipher.salsa20:
-                return range.salsa20Cipher(key, nonce).inputRangeObject;
+            return range.salsa20Cipher(key, nonce).inputRangeObject;
-            case Cipher.chacha20:
+        case Cipher.chacha20:
-                return range.chacha20Cipher(key, nonce).inputRangeObject;
+            return range.chacha20Cipher(key, nonce).inputRangeObject;
         }
     }
 }
@@ -37,9 +40,9 @@ if(isInputRange!R && is(ElementType!R : ubyte))
 unittest
 {
     import std.array;
-    
+
-    ubyte[32] key;
+    immutable ubyte[32] key;
-    ubyte[8] nonce;
+    immutable ubyte[8] nonce;
     ubyte[] a = [1, 2, 3];
     auto b = a.cipherFunction(key, nonce, Cipher.salsa20);
     InputRange!ubyte c = a.inputRangeObject;

source/salsa20.d

@@ -35,8 +35,10 @@ auto salsa20Cipher(R)(R range, ubyte[32] key, ubyte[8] nonce)
             salsaSection.popFront();
             if (salsaSection.empty)
             {
-                ++count;
+                ubyte[16] n;
-                salsaSection = salsa20Exp(key, nonce ~ littleEndianInv(count));
+                n[0..8] = nonce;
+                n[8..16] = littleEndianInv(0UL);
+                salsaSection = salsa20Exp(key, n).dup;
             }
             range.popFront();
         }
@@ -48,7 +50,10 @@ auto salsa20Cipher(R)(R range, ubyte[32] key, ubyte[8] nonce)
             }
         }
     }
-    return rangeResult(0UL, range, salsa20Exp(key, nonce ~ littleEndianInv(0UL)));
+    ubyte[16] n;
+    n[0..8] = nonce;
+    n[8..16] = littleEndianInv(0UL);
+    return rangeResult(0UL, range, salsa20Exp(key, n).dup);
 }
 
 private:
@@ -176,11 +181,11 @@ unittest
     mixin(colRound!(x00, x01, x02, x03, x04, x05, x06, x07,
                     x08, x09, x10, x11, x12, x13, x14, x15));
 
-    uint[] test0 = [y00, y01, y02, y03, y04, y05, y06, y07,
+    immutable uint[] test0 = [y00, y01, y02, y03, y04, y05, y06, y07,
-                    y08, y09, y10, y11, y12, y13, y14, y15];
+                              y08, y09, y10, y11, y12, y13, y14, y15];
 
-    uint[] test1 = [x00, x01, x02, x03, x04, x05, x06, x07,
+    immutable uint[] test1 = [x00, x01, x02, x03, x04, x05, x06, x07,
-                    x08, x09, x10, x11, x12, x13, x14, x15];
+                              x08, x09, x10, x11, x12, x13, x14, x15];
 
     assert(test0 == [0x10090288, 0x00000000, 0x00000000, 0x00000000,
                      0x00000101, 0x00000000, 0x00000000, 0x00000000,
@@ -193,8 +198,8 @@ unittest
                      0x481c2027, 0x53a8e4b5, 0x4c1f89c5, 0x3f78c9c8]);
 }
 
-ubyte[] salsa20(in ubyte[] input) pure nothrow @safe
+ubyte[64] salsa20(in ubyte[64] input) pure nothrow @safe
-in
+/*in
 {
     assert(input.length == 64);
 }
@@ -202,7 +207,7 @@ out(result)
 {
     assert(result.length == 64);
 }
-body
+body*/
 {
     auto x00 = littleEndian(input[00..04]), x01 = littleEndian(input[04..08]), 
          x02 = littleEndian(input[08..12]), x03 = littleEndian(input[12..16]), 
@@ -227,37 +232,55 @@ body
                         x08, x09, x10, x11, x12, x13, x14, x15));
     }
 
-    return littleEndianInv(x00 + y00) ~ littleEndianInv(x01 + y01) ~
+    ubyte[64] buf;
+    buf[00..04] = littleEndianInv(x00 + y00);
+    buf[04..08] = littleEndianInv(x01 + y01);
+    buf[08..12] = littleEndianInv(x02 + y02);
+    buf[12..16] = littleEndianInv(x03 + y03);
+    buf[16..20] = littleEndianInv(x04 + y04);
+    buf[20..24] = littleEndianInv(x05 + y05);
+    buf[24..28] = littleEndianInv(x06 + y06);
+    buf[28..32] = littleEndianInv(x07 + y07);
+    buf[32..36] = littleEndianInv(x08 + y08);
+    buf[36..40] = littleEndianInv(x09 + y09);
+    buf[40..44] = littleEndianInv(x10 + y10);
+    buf[44..48] = littleEndianInv(x11 + y11);
+    buf[48..52] = littleEndianInv(x12 + y12);
+    buf[52..56] = littleEndianInv(x13 + y13);
+    buf[56..60] = littleEndianInv(x14 + y14);
+    buf[60..64] = littleEndianInv(x15 + y15);
+    return buf;
+/*    return littleEndianInv(x00 + y00) ~ littleEndianInv(x01 + y01) ~
            littleEndianInv(x02 + y02) ~ littleEndianInv(x03 + y03) ~ 
            littleEndianInv(x04 + y04) ~ littleEndianInv(x05 + y05) ~ 
            littleEndianInv(x06 + y06) ~ littleEndianInv(x07 + y07) ~ 
            littleEndianInv(x08 + y08) ~ littleEndianInv(x09 + y09) ~ 
            littleEndianInv(x10 + y10) ~ littleEndianInv(x11 + y11) ~ 
            littleEndianInv(x12 + y12) ~ littleEndianInv(x13 + y13) ~ 
-           littleEndianInv(x14 + y14) ~ littleEndianInv(x15 + y15);
+           littleEndianInv(x14 + y14) ~ littleEndianInv(x15 + y15);*/
 }
 
 unittest
 {
-    ubyte[] test0 = [  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+    ubyte[64] test0 = [  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
                        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
                        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
                        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0];
     test0 = salsa20(test0);
 
-    ubyte[] test1 = [211,159, 13,115, 76, 55, 82,183,  3,117,222, 37,191,187,234,136,
+    ubyte[64] test1 = [211,159, 13,115, 76, 55, 82,183,  3,117,222, 37,191,187,234,136,
                       49,237,179, 48,  1,106,178,219,175,199,166, 48, 86, 16,179,207,
                       31,240, 32, 63, 15, 83, 93,161,116,147, 48,113,238, 55,204, 36,
                       79,201,235, 79,  3, 81,156, 47,203, 26,244,243, 88,118,104, 54];
     test1 = salsa20(test1);
 
-    ubyte[] test2 = [ 88,118,104, 54, 79,201,235, 79,  3, 81,156, 47,203, 26,244,243,
+    ubyte[64] test2 = [ 88,118,104, 54, 79,201,235, 79,  3, 81,156, 47,203, 26,244,243,
                      191,187,234,136,211,159, 13,115, 76, 55, 82,183,  3,117,222, 37,
                       86, 16,179,207, 49,237,179, 48,  1,106,178,219,175,199,166, 48,
                      238, 55,204, 36, 31,240, 32, 63, 15, 83, 93,161,116,147, 48,113];
     test2 = salsa20(test2);
 
-    ubyte[] test3 = [  6,124, 83,146, 38,191,  9, 50,  4,161, 47,222,122,182,223,185,
+    ubyte[64] test3 = [  6,124, 83,146, 38,191,  9, 50,  4,161, 47,222,122,182,223,185,
                       75, 27,  0,216, 16,122,  7, 89,162,104,101,147,213, 21, 54, 95,
                      225,253,139,176,105,132, 23,116, 76, 41,176,207,221, 34,157,108,
                       94, 94, 99, 52, 90,117, 91,220,146,190,239,143,196,176,130,186];
@@ -285,13 +308,24 @@ unittest
                      122,127,195,185,185,204,188, 90,245,  9,183,248,226, 85,245,104]);
 }
 
-enum ubyte[4] σ0 = [101, 120, 112,  97];
+static immutable ubyte[4] sigma0 = [101, 120, 112,  97];
-enum ubyte[4] σ1 = [110, 100,  32,  51];
+static immutable ubyte[4] sigma1 = [110, 100,  32,  51];
-enum ubyte[4] σ2 = [ 50,  45,  98, 121];
+static immutable ubyte[4] sigma2 = [ 50,  45,  98, 121];
-enum ubyte[4] σ3 = [116, 101,  32, 107];
+static immutable ubyte[4] sigma3 = [116, 101,  32, 107];
 
-ubyte[] salsa20Exp(in ubyte[] key, in ubyte[] n) @safe pure nothrow
+import std.range;
-in
+import std.algorithm;
+import std.typetuple;
+
+template expand(alias A, alias C="A[I]") {
+  auto ref M(alias I)() @property { return mixin(C); }
+  mixin(q{alias expand = TypeTuple!(}
+         ~ iota(A.length).map!q{"M!"~text(a)}().join(",")
+         ~ q{);});
+}
+
+ubyte[64] salsa20Exp(in ubyte[32] key, in ubyte[16] n) @safe pure nothrow
+/*in
 {
     assert(key.length == 32);
     assert(n.length == 16);
@@ -300,24 +334,41 @@ out(result)
 {
     assert(result.length == 64);
 }
-body
+body*/
 {
-    return salsa20(σ0 ~ key[0..16] ~ σ1 ~ n ~ σ2 ~ key[16..$] ~ σ3);
+    ubyte[64] buf = [
+        expand!sigma0,
+        expand!key[0..16],
+        expand!sigma1,
+        expand!n,
+        expand!sigma2,
+        expand!key[16..$],
+        expand!sigma3
+    ];
+/*    buf[0..4] = sigma0;
+    buf[4..20] = key[0..16];
+    buf[20..24] = sigma1;
+    buf[24..40] = n;
+    buf[40..44] = sigma2;
+    buf[44..60] = key[16..$];
+    buf[60..64] = sigma3;*/
+    return salsa20(buf);
+    //return salsa20(σ0 ~ key[0..16] ~ σ1 ~ n ~ σ2 ~ key[16..$] ~ σ3);
 }
 
 unittest
 {
-    ubyte[] key;
+    ubyte[32] key;
-    ubyte[] n;
+    ubyte[16] n;
-    key.length = 32;
+/*    key.length = 32;
-    n.length = 16;
+    n.length = 16;*/
     foreach (i; 0..16)
         key[i] = cast(ubyte)(i + 1);
     foreach (i; 16..32)
         key[i] = cast(ubyte)(i + 200 - 15);
 
     foreach (i; 0..16)
-        n[i] = cast(ubyte)(i + 1+ 100);
+        n[i] = cast(ubyte)(i + 1 + 100);
 
     assert(salsa20Exp(key, n) == [ 69, 37, 68, 39, 41, 15,107,193,255,139,122,  6,170,233,217, 98,
                                    89,144,182,106, 21, 51,200, 65,239, 49,222, 34,215,114, 40,126,