Reshuffel instructions for better preformance

src/crypto/chacha20-mips32r2.S

@@ -554,7 +554,6 @@ END(poly1305_generic_blocks)



#define PT0  $t0
#define PT1  $t1
#define PT2  $t2
#define PT3  $t3
@@ -565,8 +564,8 @@ END(poly1305_generic_blocks)
#define PR3  $t7
#define PR4  $t8

#define PS0 $t0
#define PS1 $t1
#define PS0 $t9
#define PS1 $t0
#define PS2 $t2
#define PS3 $t3

@@ -591,42 +590,45 @@ END(poly1305_generic_blocks)
.ent poly1305_init
//.type poly1305_init, @function
ENTRY(poly1305_init)
	LOAD_key(PT(0), 0)
	LOAD_key(PR(0), 0)
	LOAD_key(PT(1), 1)
	LOAD_key(PT(2), 2)
	LOAD_key(PT(3), 3)
#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
	wsbh 	PT(0)
	wsbh 	PR(0)
	wsbh 	PT(1)
	wsbh 	PT(2)
	wsbh 	PT(3)
	rotr 	PT(0), 16
	rotr 	PR(0), 16
	rotr 	PT(1), 16
	rotr 	PT(2), 16
	rotr 	PT(3), 16
#endif
	# // ctx->r[0] = t0 & 0x3ffffff;
	ext 					PR(0), PT(0), 0, 26

	# // ctx->r[4] = (t3 >> 8) & 0x00fffff;
	// ctx->r[x] = shift and ANDing
	SRL_U64_LSB( 			PR(1), PR(0),  26 )
	ext 					PR(0), PR(0), 0, 26
	SRL_U64_LSB( 			PR(2), PT(1),  20 )
	SRL_U64_LSB( 			PR(3), PT(2),  14 )
	ext						PR(4), PT(3), 8, 20

	# // t0 >>= 26; t0 |= t1 << 6;		                 
	# // ctx->r[1] = t0 & 0x3ffff03;
	SRL_U64_LSB( 			PR(1), PT(0),  26 ) 
	SRL_U64_MSB_UPTO26BIT( 	PR(1), PT(1),  26 ) 
	AND_6BIT( 				PR(1), 28-26 )

	# // t1 >>= 20; t1 |= t2 << 12;
	# // ctx->r[2] = t1 & 0x3ffc0ff;
	SRL_U64_LSB( 			PR(2), PT(1),  20 ) 
	sw		$zero, 36($a0)
	sw		$zero, 40($a0)
	sw		$zero, 44($a0)
	sw		$zero, 48($a0)	

	SRL_U64_MSB_UPTO26BIT( 	PR(1), PT(1),  26 ) 
	SRL_U64_MSB_UPTO26BIT( 	PR(2), PT(2),  20 ) 
	AND_6BIT( 				PR(2), 28-20 )
	SRL_U64_MSB_UPTO26BIT( 	PR(3), PT(3),  14 )

	# // t2 >>= 14; t2 |= t3 << 18;
	# // ctx->r[3] = t2 & 0x3f03fff;
	SRL_U64_LSB( 			PR(3), PT(2),  14 ) 
	SRL_U64_MSB_UPTO26BIT( 	PR(3), PT(3),  14 ) 
	sw		$zero, 52($a0)
	sw		$zero, 56($a0)
	sw		$zero, 60($a0)
	sw		$zero, 64($a0)
	sw		$zero, 68($a0)
	sw		$zero, 72($a0)

	AND_6BIT( 				PR(1), 28-26 )
	AND_6BIT( 				PR(2), 28-20 )
	AND_6BIT( 				PR(3), 28-14 )	

	STORE_poly_r(0)
@@ -654,20 +656,6 @@ ENTRY(poly1305_init)
	STORE_poly_s(2)
	STORE_poly_s(3)

	// Clear rest of the struct
	// H[5]
	sw		$zero, 36($a0)
	sw		$zero, 40($a0)
	sw		$zero, 44($a0)
	sw		$zero, 48($a0)
	sw		$zero, 52($a0)
	// buf[4] (8*16)/32 = 4
	sw		$zero, 56($a0)
	sw		$zero, 60($a0)
	sw		$zero, 64($a0)
	sw		$zero, 68($a0)
	// buflen
	sw		$zero, 72($a0)
	// Jump Back
	jr	$ra
	nop