Experimental Integer Compression (Status Unknown)

dec64.s

.global _pack32dec64

.text

.align 2
// Decode64 reads X2 amount of bytes at address X1 and it writes the 32
// encoded-values to address X0. The first delta is applied against X3.
// X2 must be a multiple of 4, and no more than 32 × 8 (bytes).
_pack32dec64:
	// TODO(pascaldekloe): validate X2
	and	x2, x2, 252
	adr	x5, decode_table 
	add	x5, x5, x2, lsl 1
	br	x5
decode_table:
	adr	x4, dec0b64
	br	x4
	adr	x4, dec1b64
	br	x4
	adr	x4, dec2b64
	br	x4
	adr	x4, dec3b64
	br	x4

.align 2
dec0b64: // no change; just copy offset
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	stp	x3, x3, [x0], 16
	ret

dec1b64: // 1-bit reverse delta zig-zag encoding

	ldr	w4, [x1]	// fetch all 4 bytes of input

	// A 1-bit zig-zag encoding of reverse delta is either 0 for
	// no change against the previous, or 1 for an increment.
	sbfx	x2, x4, 0, 1	// fetch bit #0
	add	x2, x2, x3	// against previous
	sbfx	x3, x4, 1, 1	// fetch bit #1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 2, 1
	add	x2, x2, x3
	sbfx	x3, x4, 3, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 4, 1
	add	x2, x2, x3
	sbfx	x3, x4, 5, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 6, 1
	add	x2, x2, x3
	sbfx	x3, x4, 7, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 8, 1
	add	x2, x2, x3
	sbfx	x3, x4, 9, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 10, 1
	add	x2, x2, x3
	sbfx	x3, x4, 11, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 12, 1
	add	x2, x2, x3
	sbfx	x3, x4, 13, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 14, 1
	add	x2, x2, x3
	sbfx	x3, x4, 15, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 16, 1
	add	x2, x2, x3
	sbfx	x3, x4, 17, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 18, 1
	add	x2, x2, x3
	sbfx	x3, x4, 19, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 20, 1
	add	x2, x2, x3
	sbfx	x3, x4, 21, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 22, 1
	add	x2, x2, x3
	sbfx	x3, x4, 23, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 24, 1
	add	x2, x2, x3
	sbfx	x3, x4, 25, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 26, 1
	add	x2, x2, x3
	sbfx	x3, x4, 27, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 28, 1
	add	x2, x2, x3
	sbfx	x3, x4, 29, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	sbfx	x2, x4, 30, 1
	add	x2, x2, x3
	sbfx	x3, x4, 31, 1
	add	x3, x3, x2
	stp	x2, x3, [x0], 16

	ret

dec2b64: // 2-bit reverse delta zig-zag encoding
	ldr	x4, [x1]	// fetch all 8 bytes of input

	sbfx	x5, x4, 0, 1	// sign-extend first bit (equal to NEG)
	ubfx	x2, x4, 1, 1	// shift extract remaining bit
	eor	x2, x2, x5
	sub	x2, x3, x2	// against previous
	sbfx	x5, x4, 2, 1	// again …
	ubfx	x3, x4, 3, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 4, 1
	ubfx	x2, x4, 5, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 6, 1
	ubfx	x3, x4, 7, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 8, 1
	ubfx	x2, x4, 9, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 10, 1
	ubfx	x3, x4, 11, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 12, 1
	ubfx	x2, x4, 13, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 14, 1
	ubfx	x3, x4, 15, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 16, 1
	ubfx	x2, x4, 17, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 18, 1
	ubfx	x3, x4, 19, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 20, 1
	ubfx	x2, x4, 21, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 22, 1
	ubfx	x3, x4, 23, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 24, 1
	ubfx	x2, x4, 25, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 26, 1
	ubfx	x3, x4, 27, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 28, 1
	ubfx	x2, x4, 29, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 30, 1
	ubfx	x3, x4, 31, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 32, 1
	ubfx	x2, x4, 33, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 34, 1
	ubfx	x3, x4, 35, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 36, 1
	ubfx	x2, x4, 37, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 37, 1
	ubfx	x3, x4, 39, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 40, 1
	ubfx	x2, x4, 41, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 42, 1
	ubfx	x3, x4, 43, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 44, 1
	ubfx	x2, x4, 45, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 46, 1
	ubfx	x3, x4, 47, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 48, 1
	ubfx	x2, x4, 49, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 50, 1
	ubfx	x3, x4, 51, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 52, 1
	ubfx	x2, x4, 53, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 54, 1
	ubfx	x3, x4, 55, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 56, 1
	ubfx	x2, x4, 57, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 58, 1
	ubfx	x3, x4, 59, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	sbfx	x5, x4, 60, 1
	ubfx	x2, x4, 61, 1
	eor	x2, x2, x5
	sub	x2, x3, x2
	sbfx	x5, x4, 62, 1
	ubfx	x3, x4, 63, 1
	eor	x3, x3, x5
	sub	x3, x2, x3
	stp	x2, x3, [x0], 16

	ret

dec3b64: // 3-bit reverse delta zig-zag encoding
	ret



NLHTSP:
	.ascii	"\n\t "

.align 2
// Dump64 writes the (32) 64-bit values from the stack in hexadecimal.
_dump64:
	sub	sp, sp, #16	// grow stack
	str	lr, [sp]	// persist procedure call link register

	// print tab ident
	mov	x0, 1		// file descriptor 1 is standard output
	adr	x1, NLHTSP+1	// data pointer
	mov	x2, #1		// data size
	mov	x16, #4		// write defined in <syscall.h>
	svc	#0x80		// invoke syscall

	add	x13, sp, #16	// first 64-bit integer
	add	x14, sp, #34*8	//  last …
dump_next:
	ldr	x7, [x13], 8
	bl	print_x7_hex
	// print space suffix
	mov	x0, 1		// file descriptor 1 is standard output
	adr	x1, NLHTSP+2	// data pointer
	mov	x2, #1		// data size, read-only
	mov	x16, #4		// write defined in <syscall.h>
	svc	#0x80		// invoke syscall
	cmp	x13, x14
	ble	dump_next

	// print newline end
	mov	x0, 1		// file descriptor 1 is standard output
	adr	x1, NLHTSP	// data pointer
	mov	x2, #1		// data size, read-only
	mov	x16, #4		// write defined in <syscall.h>
	svc	#0x80		// invoke syscall

	ldr	lr, [sp]	// restore procedure call link register
	add	sp, sp, #16	// free stack
	ret

hex_dict:
	.ascii "0123456789abcdefgh"

.align 2
// Print_x7_hex writes the register value in hexadecimal to standard ouput
print_x7_hex:
	adr	x6, hex_dict	// digits

	// map 8 most-significant nibbles into x8
	lsr	x8, x7, #60
	ldrb	w8, [x6, x8]	// map to digit in 32-bit register
	lsr	x5, x7, #56	// 2nd nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x8, x8, x5, lsl 8
	lsr	x5, x7, #52	// 3rd nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x8, x8, x5, lsl 16
	lsr	x5, x7, #48	// 4th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x8, x8, x5, lsl 24
	lsr	x5, x7, #44	// 5th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x8, x8, x5, lsl 32
	lsr	x5, x7, #40	// 6th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x8, x8, x5, lsl 40
	lsr	x5, x7, #36	// 7th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x8, x8, x5, lsl 48
	lsr	x5, x7, #32	// 8th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x8, x8, x5, lsl 56

	// map 8 least-significant nibbles into x9
	lsr	x9, x7, #28	// 9th nibble
	and	x9, x9, 0xf
	ldrb	w9, [x6, x9]
	lsr	x5, x7, #24	// 10th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x9, x9, x5, lsl 8
	lsr	x5, x7, #20	// 11th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x9, x9, x5, lsl 16
	lsr	x5, x7, #16	// 12th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x9, x9, x5, lsl 24
	lsr	x5, x7, #12	// 13th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x9, x9, x5, lsl 32
	lsr	x5, x7, #8	// 14th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x9, x9, x5, lsl 40
	lsr	x5, x7, #4	// 15th nibble
	and	x5, x5, 0xf
	ldrb	w5, [x6, x5]
	orr	x9, x9, x5, lsl 48
	and	x5, x7, 0xf	// 15th nibble
	ldrb	w5, [x6, x5]
	orr	x9, x9, x5, lsl 56

	sub	sp, sp, 16	// grow stack
	stp	x8, x9, [sp]	// write hex string

	mov	x0, 1		// file descriptor 1 is standard output
	mov	x1, sp		// data pointer
	mov	x2, #16		// data size, read-only
	mov	x16, #4		// write defined in <syscall.h>
	svc	#0x80		// invoke syscall

	add	sp, sp, #16	// free stack
	ret