From bcc0efc32770b48700adf2a4d8e93aa27796a9e2 Mon Sep 17 00:00:00 2001
From: Stephen Heumann <stephenheumann@gmail.com>
Date: Mon, 3 Jul 2017 22:58:58 -0500
Subject: [PATCH] Add zero bytes around SHA-256 state variables to optimize
 rotates.

---
 sha256.asm    | 46 +++++++++++++++++++++++++++++++---------------
 sha256.h      |  6 ++----
 sha256.macros | 18 +++---------------
 sha256test.c  |  8 +-------
 4 files changed, 37 insertions(+), 41 deletions(-)

diff --git a/sha256.asm b/sha256.asm
index 44d5972..2a7940f 100644
--- a/sha256.asm
+++ b/sha256.asm
@@ -24,19 +24,23 @@
 * Direct page locations	
 length	gequ	0
 extra	gequ	8
-idx	gequ	10
+zero	gequ	10
 a_	gequ	12	; elements of state
-b	gequ	16
-c	gequ	20
-d	gequ	24
-e	gequ	28
-f	gequ	32
-g	gequ	36
-h	gequ	40
-temp1	gequ	44
-temp2	gequ	48
-ch	gequ	52
-maj	gequ	56
+zero2	gequ	16
+b	gequ	18
+zero3	gequ	22
+c	gequ	24
+zero4	gequ	28
+d	gequ	30
+zero5	gequ	34
+e	gequ	36
+zero6	gequ	40
+f	gequ	42
+zero7	gequ	46
+g	gequ	48
+zero8	gequ	52
+h	gequ	54
+zero9	gequ	58
 h0	gequ	60
 h1	gequ	64
 h2	gequ	68
@@ -46,10 +50,10 @@ h5	gequ	80
 h6	gequ	84
 h7	gequ	88
 w	gequ	92
-temp3	gequ	156
-temp4	gequ	160
+ch	gequ	156
+maj	gequ	160
 k_ptr	gequ	164
-zero	gequ	168
+idx	gequ	168
 two	gequ	170
 four	gequ	172
 six	gequ	174
@@ -65,6 +69,10 @@ twentyfour gequ 192
 twentysix gequ	194
 twentyeight gequ 196
 thirty	gequ	198
+temp1	gequ	200
+temp2	gequ	204
+temp3	gequ	208
+temp4	gequ	212
 
 
 k private
@@ -133,6 +141,14 @@ SHA256_INIT start
 	stz	extra
 	
 	stz	zero
+	stz	zero2
+	stz	zero3
+	stz	zero4
+	stz	zero5
+	stz	zero6
+	stz	zero7
+	stz	zero8
+	stz	zero9
 	lda	#2
 	sta	two
 	lda	#4
diff --git a/sha256.h b/sha256.h
index 6062dbe..2b15c07 100644
--- a/sha256.h
+++ b/sha256.h
@@ -18,12 +18,10 @@ struct sha256_context {
 	unsigned long length;
 	unsigned long length2;
 	unsigned short extra;
-	unsigned short idx;
-	unsigned long vars[8];
-	unsigned char reserved1[16];
+	unsigned char reserved1[50];
 	unsigned char hash[32];
 	unsigned char block[64];
-	unsigned char reserved2[44];
+	unsigned char reserved2[60];
 };
 
 /*
diff --git a/sha256.macros b/sha256.macros
index 9bbcf9d..e6b3af4 100644
--- a/sha256.macros
+++ b/sha256.macros
@@ -359,21 +359,13 @@
 	sta	temp1
 	sta	temp2
 	lda	&e-1
-	and	#$FF00
-	sta	temp1+2
-	lda	&e+3
-	and	#$00FF
-	ora	temp1+2
+	ora	&e+3
 	sta	temp1+2
 	sta	temp2+2
 	ROTR4CONT temp2,3
 	ROTL4	temp1,2
 	lda	&e-1
-	and	#$FF00
-	sta	temp3
-	lda	&e+3
-	and	#$00FF
-	ora	temp3
+	ora	&e+3
 	sta	temp3
 	lda	&e+1
 	sta	temp3+2
@@ -442,11 +434,7 @@
 	lda	&a+1
 	sta	temp4+2
 	lda	&a-1
-	and	#$FF00
-	sta	temp4
-	lda	&a+3
-	and	#$00FF
-	ora	temp4
+	ora	&a+3
 	sta	temp4
 	ROTL4CONT temp4,2
 	lda	temp2
diff --git a/sha256test.c b/sha256test.c
index 3be187b..13d2311 100644
--- a/sha256test.c
+++ b/sha256test.c
@@ -27,7 +27,7 @@ int main(int argc, char **argv) {
     long double bytes_per_sec;
     
     struct sha256_context *context, **context_hndl;
-    struct sha256_context context_init = {0,0,0,0, {0}, {0}, {0},
+    struct sha256_context context_init = {0,0,0, {0}, {0},
         {0x61,0x62,0x63,0x80,
          0x00,0x00,0x00,0x00,
          0x00,0x00,0x00,0x00,
@@ -57,12 +57,6 @@ int main(int argc, char **argv) {
     sha256_init(context);
     sha256_processblock(context);
     
-    printf("a...h = ");
-    for (i = 0; i < 8; i++) {
-        printf("%08lx ", context->vars[i]);
-    }
-    printf("\n");
-    
     printf("h[..] = %02x%02x%02x%02x %02x%02x%02x%02x %02x%02x%02x%02x %02x%02x%02x%02x "
            "%02x%02x%02x%02x %02x%02x%02x%02x %02x%02x%02x%02x %02x%02x%02x%02x\n", 
            context->hash[3], context->hash[2], context->hash[1], context->hash[0],