X-Git-Url: https://git.cryptolib.org/?p=avr-crypto-lib.git;a=blobdiff_plain;f=keccak%2Fkeccak-asm.S;h=9d613f95042bdec0006cb4f1fb2d6ae8183fe195;hp=77dc9ce10bc99f92f67f5b7e43ddcf0155c5532c;hb=ddc66bc0543c50a0f36a3c26abf90dfef36dc3d9;hpb=4128060fcc4585ab8f40d01486baa4248b499160

diff --git a/keccak/keccak-asm.S b/keccak/keccak-asm.S
index 77dc9ce..9d613f9 100644
--- a/keccak/keccak-asm.S
+++ b/keccak/keccak-asm.S
@@ -40,6 +40,156 @@ rho_pi_idx_table:
 	.endr
 
 .align 2
+
+.global rotate64_1bit_left
+rotate64_1bit_left:
+	bst r25, 7
+	rol r18
+	rol r19
+	rol r20
+	rol r21
+	rol r22
+	rol r23
+	rol r24
+	rol r25
+	bld r18, 0
+	ret
+
+.global rotate64_1bit_right
+rotate64_1bit_right:
+	bst r18, 0
+	ror r25
+	ror r24
+	ror r23
+	ror r22
+	ror r21
+	ror r20
+	ror r19
+	ror r18
+	bld r25, 7
+	ret
+
+.global rotate64_nbit_autodir
+rotate64_nbit_autodir:
+	lsr r16
+	brcc rotate64_nbit_left
+.global rotate64_nbit_right
+rotate64_nbit_right:
+	ldi r30, pm_lo8(rotate64_1bit_right)
+	ldi r31, pm_hi8(rotate64_1bit_right)
+	rjmp icall_r16_times
+.global rotate64_nbit_left
+rotate64_nbit_left:
+	ldi r30, pm_lo8(rotate64_1bit_left)
+	ldi r31, pm_hi8(rotate64_1bit_left)
+icall_r16_times:
+1:	dec r16
+	brmi 2f
+	icall
+	rjmp 1b
+2:
+    ret
+
+rotate64_1byte_left:
+	mov r0, r25
+	mov r25, r24
+	mov r24, r23
+	mov r23, r22
+	mov r22, r21
+	mov r21, r20
+	mov r20, r19
+	mov r19, r18
+	mov r18, r0
+	ret
+
+rotate64_2byte_left:
+	movw r0, r24
+	movw r24, r22
+	movw r22, r20
+	movw r20, r18
+	movw r18, r0
+	ret
+
+rotate64_3byte_left:
+	mov r0, r25
+	mov r25, r22
+	mov r22, r19
+	mov r19, r24
+	mov r24, r21
+	mov r21, r18
+	mov r18, r23
+	mov r23, r20
+	mov r20, r0
+	ret
+
+rotate64_4byte_left:
+	movw r0, r24
+	movw r24, r20
+	movw r20, r0
+	movw r0, r22
+	movw r22, r18
+	movw r18, r0
+	ret
+
+rotate64_5byte_left:
+	mov r0, r25
+	mov r25, r20
+	mov r20, r23
+	mov r23, r18
+	mov r18, r21
+	mov r21, r24
+	mov r24, r19
+	mov r19, r22
+	mov r22, r0
+	ret
+
+rotate64_6byte_left:
+	movw r0, r18
+	movw r18, r20
+	movw r20, r22
+	movw r22, r24
+	movw r24, r0
+	ret
+
+rotate64_7byte_left:
+	mov r0, r18
+	mov r18, r19
+	mov r19, r20
+	mov r20, r21
+	mov r21, r22
+	mov r22, r23
+	mov r23, r24
+	mov r24, r25
+	mov r25, r0
+	ret
+
+
+byte_rot_jmp_table:
+	ret
+	rjmp rotate64_1byte_left
+	rjmp rotate64_2byte_left
+	rjmp rotate64_3byte_left
+	rjmp rotate64_4byte_left
+	rjmp rotate64_5byte_left
+	rjmp rotate64_6byte_left
+	rjmp rotate64_7byte_left
+
+.global rotate64left_code
+rotate64left_code:
+	ldi r30, pm_lo8(byte_rot_jmp_table)
+	ldi r31, pm_hi8(byte_rot_jmp_table)
+	mov r0, r16
+	andi r16, 0x70
+	swap r16
+	add r30, r16
+	adc r31, r1
+	mov r16, r0
+	andi r16, 0x0f
+	icall
+	clr r1
+	rjmp rotate64_nbit_autodir
+
+
 /*
 	void keccak_theta (uint64_t *a, uint64_t *b){
 	// uint64_t b[5][5];
@@ -166,14 +316,18 @@ chi_step:
 	brne 10b
 	ret
 
-.global keccak_theta
-keccak_theta:
-	push_range 2, 8
+.global keccak_f1600
+keccak_f1600:
+	push_range 2, 9
 	push r16
 	push_range 28, 29
 
+	stack_alloc_large 200, r26, r27
+	adiw XL, 1
+
+	clr r9
+5:
 	movw r30, r24 ; Z = a
-	movw r26, r22 ; X = b
 
 	ldi r19, 5
 10:
@@ -305,7 +459,7 @@ keccak_theta:
 	movw ZL, r2
 	lpm r16, Z+
 	movw r2, ZL
-	call rotate64left_code
+	rcall rotate64left_code
 	movw ZL, r4
 	lpm r16, Z+
 	movw r4, ZL
@@ -350,6 +504,7 @@ keccak_theta:
 
 	; Z points at b
 	movw XL, ZL
+	movw r4, ZL
 	adiw XL, 8
 	adiw ZL, 16
 	movw YL, r2
@@ -366,8 +521,51 @@ keccak_theta:
 	adiw ZL, 5 * 8
 	dec r18
 	brne 10b
+
+	/* -- iota -- */
+	ldi r30, lo8(keccak_rc_comp)
+	ldi r31, hi8(keccak_rc_comp)
+	add r30, r9
+	adc r31, __zero_reg__
+	lpm r20, Z+
+	movw YL, r2
+	ldi r21, 0x80
+	bst r20, 6
+	brtc 10f
+	ldd r22, Y+7
+	eor r22, r21
+	std Y+7, r22
+10:
+	bst r20, 5
+	brtc 10f
+	ldd r22, Y+3
+	eor r22, r21
+	std Y+3, r22
+10:
+	bst r20, 4
+	brtc 10f
+	ldd r22, Y+1
+	eor r22, r21
+	std Y+1, r22
+10:
+	andi r20, 0x8f
+	ld r22, Y
+	eor r22, r20
+	st Y, r22
+
+	inc r9
+	mov r16, r9
+	cpi r16, 24
+	breq 20f
+	movw r24, YL
+	movw r26, r4
+	rjmp 5b
+20:
+
+	stack_free_large3 200
+
 	pop_range 28, 29
 	pop r16
-	pop_range 2, 8
+	pop_range 2, 9
 
 	ret