Use more efficient approach for addressing into k array in SHA-256 computation.

sha256.asm

@@ -50,6 +50,22 @@ temp3	gequ	156
 temp4	gequ	160
 k_ptr	gequ	164
 zero	gequ	168
+two	gequ	170
+four	gequ	172
+six	gequ	174
+eight	gequ	176
+ten	gequ	178
+twelve	gequ	180
+fourteen gequ	182
+sixteen	gequ	184
+eighteen gequ	186
+twenty	gequ	188
+twentytwo gequ	190
+twentyfour gequ 192
+twentysix gequ	194
+twentyeight gequ 196
+thirty	gequ	198
+
 
 k private
 	dc i4'$428a2f98, $71374491, $b5c0fbcf, $e9b5dba5'
@@ -117,6 +133,37 @@ SHA256_INIT start
 	stz	extra
 	
 	stz	zero
+	lda	#2
+	sta	two
+	lda	#4
+	sta	four
+	lda	#6
+	sta	six
+	lda	#8
+	sta	eight
+	lda	#10
+	sta	ten
+	lda	#12
+	sta	twelve
+	lda	#14
+	sta	fourteen
+	lda	#16
+	sta	sixteen
+	lda	#18
+	sta	eighteen
+	lda	#20
+	sta	twenty
+	lda	#22
+	sta	twentytwo
+	lda	#24
+	sta	twentyfour
+	lda	#26
+	sta	twentysix
+	lda	#28
+	sta	twentyeight
+	lda	#30
+	sta	thirty
+	
 	rtl
 	end
 

sha256.h

@@ -23,7 +23,7 @@ struct sha256_context {
 	unsigned char reserved1[16];
 	unsigned char hash[32];
 	unsigned char block[64];
-	unsigned char reserved2[16];
+	unsigned char reserved2[44];
 };
 
 /*

sha256.macros

@@ -353,7 +353,7 @@
 * The a,b,c,d,e,f,g,h variables are given as parameters so we can avoid
 * cycling them.
 	macro
-	BlockLoopIter &a,&b,&c,&d,&e,&f,&g,&h,&iter
+	BlockLoopIter &a,&b,&c,&d,&e,&f,&g,&h,&iter,&k_idx1,&k_idx2
 	
 ; Sigma_1+w[i] computation
 	lda	&e+1
@@ -418,17 +418,12 @@
 	clc
 	tya
 	ldy	k_ptr
-	adc	(zero),y
+	adc	(&k_idx1),y
 	sta	temp1
 	txa
-	iny
+	adc	(&k_idx2),y
-	iny
-	adc	(zero),y
 	sta	temp1+2
 	tax
-	iny
-	iny
-	sty	k_ptr
 	
 	clc
 	lda	temp1
@@ -498,14 +493,19 @@
 	stz	idx
 loop&part anop
 	
-	BlockLoopIter a_,b,c,d,e,f,g,h,0
+	BlockLoopIter a_,b,c,d,e,f,g,h,0,zero,two
-	BlockLoopIter h,a_,b,c,d,e,f,g,1
+	BlockLoopIter h,a_,b,c,d,e,f,g,1,four,six
-	BlockLoopIter g,h,a_,b,c,d,e,f,2
+	BlockLoopIter g,h,a_,b,c,d,e,f,2,eight,ten
-	BlockLoopIter f,g,h,a_,b,c,d,e,3
+	BlockLoopIter f,g,h,a_,b,c,d,e,3,twelve,fourteen
-	BlockLoopIter e,f,g,h,a_,b,c,d,4
+	BlockLoopIter e,f,g,h,a_,b,c,d,4,sixteen,eighteen
-	BlockLoopIter d,e,f,g,h,a_,b,c,5
+	BlockLoopIter d,e,f,g,h,a_,b,c,5,twenty,twentytwo
-	BlockLoopIter c,d,e,f,g,h,a_,b,6
+	BlockLoopIter c,d,e,f,g,h,a_,b,6,twentyfour,twentysix
-	BlockLoopIter b,c,d,e,f,g,h,a_,7
+	BlockLoopIter b,c,d,e,f,g,h,a_,7,twentyeight,thirty
+
+	clc
+	lda	k_ptr
+	adc	#4*8
+	sta	k_ptr
 
 	clc
 	lda	idx