checksum_32.S

/* SPDX-License-Identifier: GPL-2.0 */
/* checksum.S: Sparc optimized checksum code.
 *
 *  Copyright(C) 1995 Linus Torvalds
 *  Copyright(C) 1995 Miguel de Icaza
 *  Copyright(C) 1996 David S. Miller
 *  Copyright(C) 1997 Jakub Jelinek
 *
 * derived from:
 *	Linux/Alpha checksum c-code
 *      Linux/ix86 inline checksum assembly
 *      RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code)
 *	David Mosberger-Tang for optimized reference c-code
 *	BSD4.4 portable checksum routine
 */

#include <linux/export.h>
#include <asm/errno.h>

#define CSUM_BIGCHUNK(buf, offset, sum, t0, t1, t2, t3, t4, t5)	\
	ldd	[buf + offset + 0x00], t0;			\
	ldd	[buf + offset + 0x08], t2;			\
	addxcc	t0, sum, sum;					\
	addxcc	t1, sum, sum;					\
	ldd	[buf + offset + 0x10], t4;			\
	addxcc	t2, sum, sum;					\
	addxcc	t3, sum, sum;					\
	ldd	[buf + offset + 0x18], t0;			\
	addxcc	t4, sum, sum;					\
	addxcc	t5, sum, sum;					\
	addxcc	t0, sum, sum;					\
	addxcc	t1, sum, sum;

#define CSUM_LASTCHUNK(buf, offset, sum, t0, t1, t2, t3)	\
	ldd	[buf - offset - 0x08], t0;			\
	ldd	[buf - offset - 0x00], t2;			\
	addxcc	t0, sum, sum;					\
	addxcc	t1, sum, sum;					\
	addxcc	t2, sum, sum;					\
	addxcc	t3, sum, sum;

	/* Do end cruft out of band to get better cache patterns. */
csum_partial_end_cruft:
	be	1f				! caller asks %o1 & 0x8
	 andcc	%o1, 4, %g0			! nope, check for word remaining
	ldd	[%o0], %g2			! load two
	addcc	%g2, %o2, %o2			! add first word to sum
	addxcc	%g3, %o2, %o2			! add second word as well
	add	%o0, 8, %o0			! advance buf ptr
	addx	%g0, %o2, %o2			! add in final carry
	andcc	%o1, 4, %g0			! check again for word remaining
1:	be	1f				! nope, skip this code
	 andcc	%o1, 3, %o1			! check for trailing bytes
	ld	[%o0], %g2			! load it
	addcc	%g2, %o2, %o2			! add to sum
	add	%o0, 4, %o0			! advance buf ptr
	addx	%g0, %o2, %o2			! add in final carry
	andcc	%o1, 3, %g0			! check again for trailing bytes
1:	be	1f				! no trailing bytes, return
	 addcc	%o1, -1, %g0			! only one byte remains?
	bne	2f				! at least two bytes more
	 subcc	%o1, 2, %o1			! only two bytes more?
	b	4f				! only one byte remains
	 or	%g0, %g0, %o4			! clear fake hword value
2:	lduh	[%o0], %o4			! get hword
	be	6f				! jmp if only hword remains
	 add	%o0, 2, %o0			! advance buf ptr either way
	sll	%o4, 16, %o4			! create upper hword
4:	ldub	[%o0], %o5			! get final byte
	sll	%o5, 8, %o5			! put into place
	or	%o5, %o4, %o4			! coalese with hword (if any)
6:	addcc	%o4, %o2, %o2			! add to sum
1:	retl					! get outta here
	 addx	%g0, %o2, %o0			! add final carry into retval

	/* Also do alignment out of band to get better cache patterns. */
csum_partial_fix_alignment:
	cmp	%o1, 6
	bl	cpte - 0x4
	 andcc	%o0, 0x2, %g0
	be	1f
	 andcc	%o0, 0x4, %g0
	lduh	[%o0 + 0x00], %g2
	sub	%o1, 2, %o1
	add	%o0, 2, %o0
	sll	%g2, 16, %g2
	addcc	%g2, %o2, %o2
	srl	%o2, 16, %g3
	addx	%g0, %g3, %g2
	sll	%o2, 16, %o2
	sll	%g2, 16, %g3
	srl	%o2, 16, %o2
	andcc	%o0, 0x4, %g0
	or	%g3, %o2, %o2
1:	be	cpa
	 andcc	%o1, 0xffffff80, %o3
	ld	[%o0 + 0x00], %g2
	sub	%o1, 4, %o1
	addcc	%g2, %o2, %o2
	add	%o0, 4, %o0
	addx	%g0, %o2, %o2
	b	cpa
	 andcc	%o1, 0xffffff80, %o3

	/* The common case is to get called with a nicely aligned
	 * buffer of size 0x20.  Follow the code path for that case.
	 */
	.globl	csum_partial
	EXPORT_SYMBOL(csum_partial)
csum_partial:			/* %o0=buf, %o1=len, %o2=sum */
	andcc	%o0, 0x7, %g0				! alignment problems?
	bne	csum_partial_fix_alignment		! yep, handle it
	 sethi	%hi(cpte - 8), %g7			! prepare table jmp ptr
	andcc	%o1, 0xffffff80, %o3			! num loop iterations
cpa:	be	3f					! none to do
	 andcc	%o1, 0x70, %g1				! clears carry flag too
5:	CSUM_BIGCHUNK(%o0, 0x00, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
	CSUM_BIGCHUNK(%o0, 0x20, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
	CSUM_BIGCHUNK(%o0, 0x40, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
	CSUM_BIGCHUNK(%o0, 0x60, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
	addx	%g0, %o2, %o2				! sink in final carry
	subcc	%o3, 128, %o3				! detract from loop iters
	bne	5b					! more to do
	 add	%o0, 128, %o0				! advance buf ptr
	andcc	%o1, 0x70, %g1				! clears carry flag too
3:	be	cpte					! nope
	 andcc	%o1, 0xf, %g0				! anything left at all?
	srl	%g1, 1, %o4				! compute offset
	sub	%g7, %g1, %g7				! adjust jmp ptr
	sub	%g7, %o4, %g7				! final jmp ptr adjust
	jmp	%g7 + %lo(cpte - 8)			! enter the table
	 add	%o0, %g1, %o0				! advance buf ptr
cptbl:	CSUM_LASTCHUNK(%o0, 0x68, %o2, %g2, %g3, %g4, %g5)
	CSUM_LASTCHUNK(%o0, 0x58, %o2, %g2, %g3, %g4, %g5)
	CSUM_LASTCHUNK(%o0, 0x48, %o2, %g2, %g3, %g4, %g5)
	CSUM_LASTCHUNK(%o0, 0x38, %o2, %g2, %g3, %g4, %g5)
	CSUM_LASTCHUNK(%o0, 0x28, %o2, %g2, %g3, %g4, %g5)
	CSUM_LASTCHUNK(%o0, 0x18, %o2, %g2, %g3, %g4, %g5)
	CSUM_LASTCHUNK(%o0, 0x08, %o2, %g2, %g3, %g4, %g5)
	addx	%g0, %o2, %o2				! fetch final carry
	andcc	%o1, 0xf, %g0				! anything left at all?
cpte:	bne	csum_partial_end_cruft			! yep, handle it
	 andcc	%o1, 8, %g0				! check how much
cpout:	retl						! get outta here
	 mov	%o2, %o0				! return computed csum

/* Work around cpp -rob */
#define ALLOC #alloc
#define EXECINSTR #execinstr
#define EX(x,y)					\
98:     x,y;                                    \
        .section __ex_table,ALLOC;		\
        .align  4;                              \
        .word   98b, cc_fault;                   \
        .text;                                  \
        .align  4

	/* This aligned version executes typically in 8.5 superscalar cycles, this
	 * is the best I can do.  I say 8.5 because the final add will pair with
	 * the next ldd in the main unrolled loop.  Thus the pipe is always full.
	 * If you change these macros (including order of instructions),
	 * please check the fixup code below as well.
	 */
#define CSUMCOPY_BIGCHUNK_ALIGNED(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7)	\
	EX(ldd	[src + off + 0x00], t0);						\
	EX(ldd	[src + off + 0x08], t2);						\
	addxcc	t0, sum, sum;								\
	EX(ldd	[src + off + 0x10], t4);						\
	addxcc	t1, sum, sum;								\
	EX(ldd	[src + off + 0x18], t6);						\
	addxcc	t2, sum, sum;								\
	EX(std	t0, [dst + off + 0x00]);						\
	addxcc	t3, sum, sum;								\
	EX(std	t2, [dst + off + 0x08]);						\
	addxcc	t4, sum, sum;								\
	EX(std	t4, [dst + off + 0x10]);						\
	addxcc	t5, sum, sum;								\
	EX(std	t6, [dst + off + 0x18]);						\
	addxcc	t6, sum, sum;								\
	addxcc	t7, sum, sum;

	/* 12 superscalar cycles seems to be the limit for this case,
	 * because of this we thus do all the ldd's together to get
	 * Viking MXCC into streaming mode.  Ho hum...
	 */
#define CSUMCOPY_BIGCHUNK(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7)	\
	EX(ldd	[src + off + 0x00], t0);					\
	EX(ldd	[src + off + 0x08], t2);					\
	EX(ldd	[src + off + 0x10], t4);					\
	EX(ldd	[src + off + 0x18], t6);					\
	EX(st	t0, [dst + off + 0x00]);					\
	addxcc	t0, sum, sum;							\
	EX(st	t1, [dst + off + 0x04]);					\
	addxcc	t1, sum, sum;							\
	EX(st	t2, [dst + off + 0x08]);					\
	addxcc	t2, sum, sum;							\
	EX(st	t3, [dst + off + 0x0c]);					\
	addxcc	t3, sum, sum;							\
	EX(st	t4, [dst + off + 0x10]);					\
	addxcc	t4, sum, sum;							\
	EX(st	t5, [dst + off + 0x14]);					\
	addxcc	t5, sum, sum;							\
	EX(st	t6, [dst + off + 0x18]);					\
	addxcc	t6, sum, sum;							\
	EX(st	t7, [dst + off + 0x1c]);					\
	addxcc	t7, sum, sum;

	/* Yuck, 6 superscalar cycles... */
#define CSUMCOPY_LASTCHUNK(src, dst, sum, off, t0, t1, t2, t3)	\
	EX(ldd	[src - off - 0x08], t0);			\
	EX(ldd	[src - off - 0x00], t2);			\
	addxcc	t0, sum, sum;					\
	EX(st	t0, [dst - off - 0x08]);			\
	addxcc	t1, sum, sum;					\
	EX(st	t1, [dst - off - 0x04]);			\
	addxcc	t2, sum, sum;					\
	EX(st	t2, [dst - off - 0x00]);			\
	addxcc	t3, sum, sum;					\
	EX(st	t3, [dst - off + 0x04]);

	/* Handle the end cruft code out of band for better cache patterns. */
cc_end_cruft:
	be	1f
	 andcc	%o3, 4, %g0
	EX(ldd	[%o0 + 0x00], %g2)
	add	%o1, 8, %o1
	addcc	%g2, %g7, %g7
	add	%o0, 8, %o0
	addxcc	%g3, %g7, %g7
	EX(st	%g2, [%o1 - 0x08])
	addx	%g0, %g7, %g7
	andcc	%o3, 4, %g0
	EX(st	%g3, [%o1 - 0x04])
1:	be	1f
	 andcc	%o3, 3, %o3
	EX(ld	[%o0 + 0x00], %g2)
	add	%o1, 4, %o1
	addcc	%g2, %g7, %g7
	EX(st	%g2, [%o1 - 0x04])
	addx	%g0, %g7, %g7
	andcc	%o3, 3, %g0
	add	%o0, 4, %o0
1:	be	1f
	 addcc	%o3, -1, %g0
	bne	2f
	 subcc	%o3, 2, %o3
	b	4f
	 or	%g0, %g0, %o4
2:	EX(lduh	[%o0 + 0x00], %o4)
	add	%o0, 2, %o0
	EX(sth	%o4, [%o1 + 0x00])
	be	6f
	 add	%o1, 2, %o1
	sll	%o4, 16, %o4
4:	EX(ldub	[%o0 + 0x00], %o5)
	EX(stb	%o5, [%o1 + 0x00])
	sll	%o5, 8, %o5
	or	%o5, %o4, %o4
6:	addcc	%o4, %g7, %g7
1:	retl
	 addx	%g0, %g7, %o0

	/* Also, handle the alignment code out of band. */
cc_dword_align:
	cmp	%g1, 16
	bge	1f
	 srl	%g1, 1, %o3
2:	cmp	%o3, 0
	be,a	ccte
	 andcc	%g1, 0xf, %o3
	andcc	%o3, %o0, %g0	! Check %o0 only (%o1 has the same last 2 bits)
	be,a	2b
	 srl	%o3, 1, %o3
1:	andcc	%o0, 0x1, %g0
	bne	ccslow
	 andcc	%o0, 0x2, %g0
	be	1f
	 andcc	%o0, 0x4, %g0
	EX(lduh	[%o0 + 0x00], %g4)
	sub	%g1, 2, %g1
	EX(sth	%g4, [%o1 + 0x00])
	add	%o0, 2, %o0
	sll	%g4, 16, %g4
	addcc	%g4, %g7, %g7
	add	%o1, 2, %o1
	srl	%g7, 16, %g3
	addx	%g0, %g3, %g4
	sll	%g7, 16, %g7
	sll	%g4, 16, %g3
	srl	%g7, 16, %g7
	andcc	%o0, 0x4, %g0
	or	%g3, %g7, %g7
1:	be	3f
	 andcc	%g1, 0xffffff80, %g0
	EX(ld	[%o0 + 0x00], %g4)
	sub	%g1, 4, %g1
	EX(st	%g4, [%o1 + 0x00])
	add	%o0, 4, %o0
	addcc	%g4, %g7, %g7
	add	%o1, 4, %o1
	addx	%g0, %g7, %g7
	b	3f
	 andcc	%g1, 0xffffff80, %g0

	/* Sun, you just can't beat me, you just can't.  Stop trying,
	 * give up.  I'm serious, I am going to kick the living shit
	 * out of you, game over, lights out.
	 */
	.align	8
	.globl	__csum_partial_copy_sparc_generic
	EXPORT_SYMBOL(__csum_partial_copy_sparc_generic)
__csum_partial_copy_sparc_generic:
					/* %o0=src, %o1=dest, %g1=len, %g7=sum */
	xor	%o0, %o1, %o4		! get changing bits
	andcc	%o4, 3, %g0		! check for mismatched alignment
	bne	ccslow			! better this than unaligned/fixups
	 andcc	%o0, 7, %g0		! need to align things?
	bne	cc_dword_align		! yes, we check for short lengths there
	 andcc	%g1, 0xffffff80, %g0	! can we use unrolled loop?
3:	be	3f			! nope, less than one loop remains
	 andcc	%o1, 4, %g0		! dest aligned on 4 or 8 byte boundary?
	be	ccdbl + 4		! 8 byte aligned, kick ass
5:	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	sub	%g1, 128, %g1		! detract from length
	addx	%g0, %g7, %g7		! add in last carry bit
	andcc	%g1, 0xffffff80, %g0	! more to csum?
	add	%o0, 128, %o0		! advance src ptr
	bne	5b			! we did not go negative, continue looping
	 add	%o1, 128, %o1		! advance dest ptr
3:	andcc	%g1, 0x70, %o2		! can use table?
ccmerge:be	ccte			! nope, go and check for end cruft
	 andcc	%g1, 0xf, %o3		! get low bits of length (clears carry btw)
	srl	%o2, 1, %o4		! begin negative offset computation
	sethi	%hi(12f), %o5		! set up table ptr end
	add	%o0, %o2, %o0		! advance src ptr
	sub	%o5, %o4, %o5		! continue table calculation
	sll	%o2, 1, %g2		! constant multiplies are fun...
	sub	%o5, %g2, %o5		! some more adjustments
	jmp	%o5 + %lo(12f)		! jump into it, duff style, wheee...
	 add	%o1, %o2, %o1		! advance dest ptr (carry is clear btw)
cctbl:	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x68,%g2,%g3,%g4,%g5)
	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x58,%g2,%g3,%g4,%g5)
	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x48,%g2,%g3,%g4,%g5)
	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x38,%g2,%g3,%g4,%g5)
	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x28,%g2,%g3,%g4,%g5)
	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x18,%g2,%g3,%g4,%g5)
	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x08,%g2,%g3,%g4,%g5)
12:	addx	%g0, %g7, %g7
	andcc	%o3, 0xf, %g0		! check for low bits set
ccte:	bne	cc_end_cruft		! something left, handle it out of band
	 andcc	%o3, 8, %g0		! begin checks for that code
	retl				! return
	 mov	%g7, %o0		! give em the computed checksum
ccdbl:	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
	sub	%g1, 128, %g1		! detract from length
	addx	%g0, %g7, %g7		! add in last carry bit
	andcc	%g1, 0xffffff80, %g0	! more to csum?
	add	%o0, 128, %o0		! advance src ptr
	bne	ccdbl			! we did not go negative, continue looping
	 add	%o1, 128, %o1		! advance dest ptr
	b	ccmerge			! finish it off, above
	 andcc	%g1, 0x70, %o2		! can use table? (clears carry btw)

ccslow:	cmp	%g1, 0
	mov	0, %g5
	bleu	4f
	 andcc	%o0, 1, %o5		
	be,a	1f
	 srl	%g1, 1, %g4		
	sub	%g1, 1, %g1	
	EX(ldub	[%o0], %g5)
	add	%o0, 1, %o0	
	EX(stb	%g5, [%o1])
	srl	%g1, 1, %g4
	add	%o1, 1, %o1
1:	cmp	%g4, 0		
	be,a	3f
	 andcc	%g1, 1, %g0
	andcc	%o0, 2, %g0	
	be,a	1f
	 srl	%g4, 1, %g4
	EX(lduh	[%o0], %o4)
	sub	%g1, 2, %g1	
	srl	%o4, 8, %g2
	sub	%g4, 1, %g4	
	EX(stb	%g2, [%o1])
	add	%o4, %g5, %g5
	EX(stb	%o4, [%o1 + 1])
	add	%o0, 2, %o0	
	srl	%g4, 1, %g4
	add	%o1, 2, %o1
1:	cmp	%g4, 0		
	be,a	2f
	 andcc	%g1, 2, %g0
	EX(ld	[%o0], %o4)
5:	srl	%o4, 24, %g2
	srl	%o4, 16, %g3
	EX(stb	%g2, [%o1])
	srl	%o4, 8, %g2
	EX(stb	%g3, [%o1 + 1])
	add	%o0, 4, %o0
	EX(stb	%g2, [%o1 + 2])
	addcc	%o4, %g5, %g5
	EX(stb	%o4, [%o1 + 3])
	addx	%g5, %g0, %g5	! I am now to lazy to optimize this (question it
	add	%o1, 4, %o1	! is worthy). Maybe some day - with the sll/srl
	subcc	%g4, 1, %g4	! tricks
	bne,a	5b
	 EX(ld	[%o0], %o4)
	sll	%g5, 16, %g2
	srl	%g5, 16, %g5
	srl	%g2, 16, %g2
	andcc	%g1, 2, %g0
	add	%g2, %g5, %g5 
2:	be,a	3f		
	 andcc	%g1, 1, %g0
	EX(lduh	[%o0], %o4)
	andcc	%g1, 1, %g0
	srl	%o4, 8, %g2
	add	%o0, 2, %o0	
	EX(stb	%g2, [%o1])
	add	%g5, %o4, %g5
	EX(stb	%o4, [%o1 + 1])
	add	%o1, 2, %o1
3:	be,a	1f		
	 sll	%g5, 16, %o4
	EX(ldub	[%o0], %g2)
	sll	%g2, 8, %o4	
	EX(stb	%g2, [%o1])
	add	%g5, %o4, %g5
	sll	%g5, 16, %o4
1:	addcc	%o4, %g5, %g5
	srl	%g5, 16, %o4
	addx	%g0, %o4, %g5
	orcc	%o5, %g0, %g0
	be	4f
	 srl	%g5, 8, %o4
	and	%g5, 0xff, %g2
	and	%o4, 0xff, %o4
	sll	%g2, 8, %g2
	or	%g2, %o4, %g5
4:	addcc	%g7, %g5, %g7
	retl	
	 addx	%g0, %g7, %o0

/* We do these strange calculations for the csum_*_from_user case only, ie.
 * we only bother with faults on loads... */

cc_fault:
	retl
	 clr	%o0