xref: /openbmc/linux/arch/sparc/lib/checksum_32.S (revision 11a163f2) 
1/* SPDX-License-Identifier: GPL-2.0 */
2/* checksum.S: Sparc optimized checksum code.
3 *
4 *  Copyright(C) 1995 Linus Torvalds
5 *  Copyright(C) 1995 Miguel de Icaza
6 *  Copyright(C) 1996 David S. Miller
7 *  Copyright(C) 1997 Jakub Jelinek
8 *
9 * derived from:
10 *	Linux/Alpha checksum c-code
11 *      Linux/ix86 inline checksum assembly
12 *      RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code)
13 *	David Mosberger-Tang for optimized reference c-code
14 *	BSD4.4 portable checksum routine
15 */
16
17#include <asm/errno.h>
18#include <asm/export.h>
19
20#define CSUM_BIGCHUNK(buf, offset, sum, t0, t1, t2, t3, t4, t5)	\
21	ldd	[buf + offset + 0x00], t0;			\
22	ldd	[buf + offset + 0x08], t2;			\
23	addxcc	t0, sum, sum;					\
24	addxcc	t1, sum, sum;					\
25	ldd	[buf + offset + 0x10], t4;			\
26	addxcc	t2, sum, sum;					\
27	addxcc	t3, sum, sum;					\
28	ldd	[buf + offset + 0x18], t0;			\
29	addxcc	t4, sum, sum;					\
30	addxcc	t5, sum, sum;					\
31	addxcc	t0, sum, sum;					\
32	addxcc	t1, sum, sum;
33
34#define CSUM_LASTCHUNK(buf, offset, sum, t0, t1, t2, t3)	\
35	ldd	[buf - offset - 0x08], t0;			\
36	ldd	[buf - offset - 0x00], t2;			\
37	addxcc	t0, sum, sum;					\
38	addxcc	t1, sum, sum;					\
39	addxcc	t2, sum, sum;					\
40	addxcc	t3, sum, sum;
41
42	/* Do end cruft out of band to get better cache patterns. */
43csum_partial_end_cruft:
44	be	1f				! caller asks %o1 & 0x8
45	 andcc	%o1, 4, %g0			! nope, check for word remaining
46	ldd	[%o0], %g2			! load two
47	addcc	%g2, %o2, %o2			! add first word to sum
48	addxcc	%g3, %o2, %o2			! add second word as well
49	add	%o0, 8, %o0			! advance buf ptr
50	addx	%g0, %o2, %o2			! add in final carry
51	andcc	%o1, 4, %g0			! check again for word remaining
521:	be	1f				! nope, skip this code
53	 andcc	%o1, 3, %o1			! check for trailing bytes
54	ld	[%o0], %g2			! load it
55	addcc	%g2, %o2, %o2			! add to sum
56	add	%o0, 4, %o0			! advance buf ptr
57	addx	%g0, %o2, %o2			! add in final carry
58	andcc	%o1, 3, %g0			! check again for trailing bytes
591:	be	1f				! no trailing bytes, return
60	 addcc	%o1, -1, %g0			! only one byte remains?
61	bne	2f				! at least two bytes more
62	 subcc	%o1, 2, %o1			! only two bytes more?
63	b	4f				! only one byte remains
64	 or	%g0, %g0, %o4			! clear fake hword value
652:	lduh	[%o0], %o4			! get hword
66	be	6f				! jmp if only hword remains
67	 add	%o0, 2, %o0			! advance buf ptr either way
68	sll	%o4, 16, %o4			! create upper hword
694:	ldub	[%o0], %o5			! get final byte
70	sll	%o5, 8, %o5			! put into place
71	or	%o5, %o4, %o4			! coalese with hword (if any)
726:	addcc	%o4, %o2, %o2			! add to sum
731:	retl					! get outta here
74	 addx	%g0, %o2, %o0			! add final carry into retval
75
76	/* Also do alignment out of band to get better cache patterns. */
77csum_partial_fix_alignment:
78	cmp	%o1, 6
79	bl	cpte - 0x4
80	 andcc	%o0, 0x2, %g0
81	be	1f
82	 andcc	%o0, 0x4, %g0
83	lduh	[%o0 + 0x00], %g2
84	sub	%o1, 2, %o1
85	add	%o0, 2, %o0
86	sll	%g2, 16, %g2
87	addcc	%g2, %o2, %o2
88	srl	%o2, 16, %g3
89	addx	%g0, %g3, %g2
90	sll	%o2, 16, %o2
91	sll	%g2, 16, %g3
92	srl	%o2, 16, %o2
93	andcc	%o0, 0x4, %g0
94	or	%g3, %o2, %o2
951:	be	cpa
96	 andcc	%o1, 0xffffff80, %o3
97	ld	[%o0 + 0x00], %g2
98	sub	%o1, 4, %o1
99	addcc	%g2, %o2, %o2
100	add	%o0, 4, %o0
101	addx	%g0, %o2, %o2
102	b	cpa
103	 andcc	%o1, 0xffffff80, %o3
104
105	/* The common case is to get called with a nicely aligned
106	 * buffer of size 0x20.  Follow the code path for that case.
107	 */
108	.globl	csum_partial
109	EXPORT_SYMBOL(csum_partial)
110csum_partial:			/* %o0=buf, %o1=len, %o2=sum */
111	andcc	%o0, 0x7, %g0				! alignment problems?
112	bne	csum_partial_fix_alignment		! yep, handle it
113	 sethi	%hi(cpte - 8), %g7			! prepare table jmp ptr
114	andcc	%o1, 0xffffff80, %o3			! num loop iterations
115cpa:	be	3f					! none to do
116	 andcc	%o1, 0x70, %g1				! clears carry flag too
1175:	CSUM_BIGCHUNK(%o0, 0x00, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
118	CSUM_BIGCHUNK(%o0, 0x20, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
119	CSUM_BIGCHUNK(%o0, 0x40, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
120	CSUM_BIGCHUNK(%o0, 0x60, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
121	addx	%g0, %o2, %o2				! sink in final carry
122	subcc	%o3, 128, %o3				! detract from loop iters
123	bne	5b					! more to do
124	 add	%o0, 128, %o0				! advance buf ptr
125	andcc	%o1, 0x70, %g1				! clears carry flag too
1263:	be	cpte					! nope
127	 andcc	%o1, 0xf, %g0				! anything left at all?
128	srl	%g1, 1, %o4				! compute offset
129	sub	%g7, %g1, %g7				! adjust jmp ptr
130	sub	%g7, %o4, %g7				! final jmp ptr adjust
131	jmp	%g7 + %lo(cpte - 8)			! enter the table
132	 add	%o0, %g1, %o0				! advance buf ptr
133cptbl:	CSUM_LASTCHUNK(%o0, 0x68, %o2, %g2, %g3, %g4, %g5)
134	CSUM_LASTCHUNK(%o0, 0x58, %o2, %g2, %g3, %g4, %g5)
135	CSUM_LASTCHUNK(%o0, 0x48, %o2, %g2, %g3, %g4, %g5)
136	CSUM_LASTCHUNK(%o0, 0x38, %o2, %g2, %g3, %g4, %g5)
137	CSUM_LASTCHUNK(%o0, 0x28, %o2, %g2, %g3, %g4, %g5)
138	CSUM_LASTCHUNK(%o0, 0x18, %o2, %g2, %g3, %g4, %g5)
139	CSUM_LASTCHUNK(%o0, 0x08, %o2, %g2, %g3, %g4, %g5)
140	addx	%g0, %o2, %o2				! fetch final carry
141	andcc	%o1, 0xf, %g0				! anything left at all?
142cpte:	bne	csum_partial_end_cruft			! yep, handle it
143	 andcc	%o1, 8, %g0				! check how much
144cpout:	retl						! get outta here
145	 mov	%o2, %o0				! return computed csum
146
147/* Work around cpp -rob */
148#define ALLOC #alloc
149#define EXECINSTR #execinstr
150#define EX(x,y)					\
15198:     x,y;                                    \
152        .section __ex_table,ALLOC;		\
153        .align  4;                              \
154        .word   98b, cc_fault;                   \
155        .text;                                  \
156        .align  4
157
158#define EXT(start,end)				\
159        .section __ex_table,ALLOC;		\
160        .align  4;                              \
161        .word   start, 0, end, cc_fault;         \
162        .text;                                  \
163        .align  4
164
165	/* This aligned version executes typically in 8.5 superscalar cycles, this
166	 * is the best I can do.  I say 8.5 because the final add will pair with
167	 * the next ldd in the main unrolled loop.  Thus the pipe is always full.
168	 * If you change these macros (including order of instructions),
169	 * please check the fixup code below as well.
170	 */
171#define CSUMCOPY_BIGCHUNK_ALIGNED(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7)	\
172	ldd	[src + off + 0x00], t0;							\
173	ldd	[src + off + 0x08], t2;							\
174	addxcc	t0, sum, sum;								\
175	ldd	[src + off + 0x10], t4;							\
176	addxcc	t1, sum, sum;								\
177	ldd	[src + off + 0x18], t6;							\
178	addxcc	t2, sum, sum;								\
179	std	t0, [dst + off + 0x00];							\
180	addxcc	t3, sum, sum;								\
181	std	t2, [dst + off + 0x08];							\
182	addxcc	t4, sum, sum;								\
183	std	t4, [dst + off + 0x10];							\
184	addxcc	t5, sum, sum;								\
185	std	t6, [dst + off + 0x18];							\
186	addxcc	t6, sum, sum;								\
187	addxcc	t7, sum, sum;
188
189	/* 12 superscalar cycles seems to be the limit for this case,
190	 * because of this we thus do all the ldd's together to get
191	 * Viking MXCC into streaming mode.  Ho hum...
192	 */
193#define CSUMCOPY_BIGCHUNK(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7)	\
194	ldd	[src + off + 0x00], t0;						\
195	ldd	[src + off + 0x08], t2;						\
196	ldd	[src + off + 0x10], t4;						\
197	ldd	[src + off + 0x18], t6;						\
198	st	t0, [dst + off + 0x00];						\
199	addxcc	t0, sum, sum;							\
200	st	t1, [dst + off + 0x04];						\
201	addxcc	t1, sum, sum;							\
202	st	t2, [dst + off + 0x08];						\
203	addxcc	t2, sum, sum;							\
204	st	t3, [dst + off + 0x0c];						\
205	addxcc	t3, sum, sum;							\
206	st	t4, [dst + off + 0x10];						\
207	addxcc	t4, sum, sum;							\
208	st	t5, [dst + off + 0x14];						\
209	addxcc	t5, sum, sum;							\
210	st	t6, [dst + off + 0x18];						\
211	addxcc	t6, sum, sum;							\
212	st	t7, [dst + off + 0x1c];						\
213	addxcc	t7, sum, sum;
214
215	/* Yuck, 6 superscalar cycles... */
216#define CSUMCOPY_LASTCHUNK(src, dst, sum, off, t0, t1, t2, t3)	\
217	ldd	[src - off - 0x08], t0;				\
218	ldd	[src - off - 0x00], t2;				\
219	addxcc	t0, sum, sum;					\
220	st	t0, [dst - off - 0x08];				\
221	addxcc	t1, sum, sum;					\
222	st	t1, [dst - off - 0x04];				\
223	addxcc	t2, sum, sum;					\
224	st	t2, [dst - off - 0x00];				\
225	addxcc	t3, sum, sum;					\
226	st	t3, [dst - off + 0x04];
227
228	/* Handle the end cruft code out of band for better cache patterns. */
229cc_end_cruft:
230	be	1f
231	 andcc	%o3, 4, %g0
232	EX(ldd	[%o0 + 0x00], %g2)
233	add	%o1, 8, %o1
234	addcc	%g2, %g7, %g7
235	add	%o0, 8, %o0
236	addxcc	%g3, %g7, %g7
237	EX(st	%g2, [%o1 - 0x08])
238	addx	%g0, %g7, %g7
239	andcc	%o3, 4, %g0
240	EX(st	%g3, [%o1 - 0x04])
2411:	be	1f
242	 andcc	%o3, 3, %o3
243	EX(ld	[%o0 + 0x00], %g2)
244	add	%o1, 4, %o1
245	addcc	%g2, %g7, %g7
246	EX(st	%g2, [%o1 - 0x04])
247	addx	%g0, %g7, %g7
248	andcc	%o3, 3, %g0
249	add	%o0, 4, %o0
2501:	be	1f
251	 addcc	%o3, -1, %g0
252	bne	2f
253	 subcc	%o3, 2, %o3
254	b	4f
255	 or	%g0, %g0, %o4
2562:	EX(lduh	[%o0 + 0x00], %o4)
257	add	%o0, 2, %o0
258	EX(sth	%o4, [%o1 + 0x00])
259	be	6f
260	 add	%o1, 2, %o1
261	sll	%o4, 16, %o4
2624:	EX(ldub	[%o0 + 0x00], %o5)
263	EX(stb	%o5, [%o1 + 0x00])
264	sll	%o5, 8, %o5
265	or	%o5, %o4, %o4
2666:	addcc	%o4, %g7, %g7
2671:	retl
268	 addx	%g0, %g7, %o0
269
270	/* Also, handle the alignment code out of band. */
271cc_dword_align:
272	cmp	%g1, 16
273	bge	1f
274	 srl	%g1, 1, %o3
2752:	cmp	%o3, 0
276	be,a	ccte
277	 andcc	%g1, 0xf, %o3
278	andcc	%o3, %o0, %g0	! Check %o0 only (%o1 has the same last 2 bits)
279	be,a	2b
280	 srl	%o3, 1, %o3
2811:	andcc	%o0, 0x1, %g0
282	bne	ccslow
283	 andcc	%o0, 0x2, %g0
284	be	1f
285	 andcc	%o0, 0x4, %g0
286	EX(lduh	[%o0 + 0x00], %g4)
287	sub	%g1, 2, %g1
288	EX(sth	%g4, [%o1 + 0x00])
289	add	%o0, 2, %o0
290	sll	%g4, 16, %g4
291	addcc	%g4, %g7, %g7
292	add	%o1, 2, %o1
293	srl	%g7, 16, %g3
294	addx	%g0, %g3, %g4
295	sll	%g7, 16, %g7
296	sll	%g4, 16, %g3
297	srl	%g7, 16, %g7
298	andcc	%o0, 0x4, %g0
299	or	%g3, %g7, %g7
3001:	be	3f
301	 andcc	%g1, 0xffffff80, %g0
302	EX(ld	[%o0 + 0x00], %g4)
303	sub	%g1, 4, %g1
304	EX(st	%g4, [%o1 + 0x00])
305	add	%o0, 4, %o0
306	addcc	%g4, %g7, %g7
307	add	%o1, 4, %o1
308	addx	%g0, %g7, %g7
309	b	3f
310	 andcc	%g1, 0xffffff80, %g0
311
312	/* Sun, you just can't beat me, you just can't.  Stop trying,
313	 * give up.  I'm serious, I am going to kick the living shit
314	 * out of you, game over, lights out.
315	 */
316	.align	8
317	.globl	__csum_partial_copy_sparc_generic
318	EXPORT_SYMBOL(__csum_partial_copy_sparc_generic)
319__csum_partial_copy_sparc_generic:
320					/* %o0=src, %o1=dest, %g1=len, %g7=sum */
321	xor	%o0, %o1, %o4		! get changing bits
322	andcc	%o4, 3, %g0		! check for mismatched alignment
323	bne	ccslow			! better this than unaligned/fixups
324	 andcc	%o0, 7, %g0		! need to align things?
325	bne	cc_dword_align		! yes, we check for short lengths there
326	 andcc	%g1, 0xffffff80, %g0	! can we use unrolled loop?
3273:	be	3f			! nope, less than one loop remains
328	 andcc	%o1, 4, %g0		! dest aligned on 4 or 8 byte boundary?
329	be	ccdbl + 4		! 8 byte aligned, kick ass
3305:	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
331	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
332	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
333	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
33410:	EXT(5b, 10b)			! note for exception handling
335	sub	%g1, 128, %g1		! detract from length
336	addx	%g0, %g7, %g7		! add in last carry bit
337	andcc	%g1, 0xffffff80, %g0	! more to csum?
338	add	%o0, 128, %o0		! advance src ptr
339	bne	5b			! we did not go negative, continue looping
340	 add	%o1, 128, %o1		! advance dest ptr
3413:	andcc	%g1, 0x70, %o2		! can use table?
342ccmerge:be	ccte			! nope, go and check for end cruft
343	 andcc	%g1, 0xf, %o3		! get low bits of length (clears carry btw)
344	srl	%o2, 1, %o4		! begin negative offset computation
345	sethi	%hi(12f), %o5		! set up table ptr end
346	add	%o0, %o2, %o0		! advance src ptr
347	sub	%o5, %o4, %o5		! continue table calculation
348	sll	%o2, 1, %g2		! constant multiplies are fun...
349	sub	%o5, %g2, %o5		! some more adjustments
350	jmp	%o5 + %lo(12f)		! jump into it, duff style, wheee...
351	 add	%o1, %o2, %o1		! advance dest ptr (carry is clear btw)
352cctbl:	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x68,%g2,%g3,%g4,%g5)
353	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x58,%g2,%g3,%g4,%g5)
354	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x48,%g2,%g3,%g4,%g5)
355	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x38,%g2,%g3,%g4,%g5)
356	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x28,%g2,%g3,%g4,%g5)
357	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x18,%g2,%g3,%g4,%g5)
358	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x08,%g2,%g3,%g4,%g5)
35912:	EXT(cctbl, 12b)			! note for exception table handling
360	addx	%g0, %g7, %g7
361	andcc	%o3, 0xf, %g0		! check for low bits set
362ccte:	bne	cc_end_cruft		! something left, handle it out of band
363	 andcc	%o3, 8, %g0		! begin checks for that code
364	retl				! return
365	 mov	%g7, %o0		! give em the computed checksum
366ccdbl:	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
367	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
368	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
369	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
37011:	EXT(ccdbl, 11b)			! note for exception table handling
371	sub	%g1, 128, %g1		! detract from length
372	addx	%g0, %g7, %g7		! add in last carry bit
373	andcc	%g1, 0xffffff80, %g0	! more to csum?
374	add	%o0, 128, %o0		! advance src ptr
375	bne	ccdbl			! we did not go negative, continue looping
376	 add	%o1, 128, %o1		! advance dest ptr
377	b	ccmerge			! finish it off, above
378	 andcc	%g1, 0x70, %o2		! can use table? (clears carry btw)
379
380ccslow:	cmp	%g1, 0
381	mov	0, %g5
382	bleu	4f
383	 andcc	%o0, 1, %o5
384	be,a	1f
385	 srl	%g1, 1, %g4
386	sub	%g1, 1, %g1
387	EX(ldub	[%o0], %g5)
388	add	%o0, 1, %o0
389	EX(stb	%g5, [%o1])
390	srl	%g1, 1, %g4
391	add	%o1, 1, %o1
3921:	cmp	%g4, 0
393	be,a	3f
394	 andcc	%g1, 1, %g0
395	andcc	%o0, 2, %g0
396	be,a	1f
397	 srl	%g4, 1, %g4
398	EX(lduh	[%o0], %o4)
399	sub	%g1, 2, %g1
400	srl	%o4, 8, %g2
401	sub	%g4, 1, %g4
402	EX(stb	%g2, [%o1])
403	add	%o4, %g5, %g5
404	EX(stb	%o4, [%o1 + 1])
405	add	%o0, 2, %o0
406	srl	%g4, 1, %g4
407	add	%o1, 2, %o1
4081:	cmp	%g4, 0
409	be,a	2f
410	 andcc	%g1, 2, %g0
411	EX(ld	[%o0], %o4)
4125:	srl	%o4, 24, %g2
413	srl	%o4, 16, %g3
414	EX(stb	%g2, [%o1])
415	srl	%o4, 8, %g2
416	EX(stb	%g3, [%o1 + 1])
417	add	%o0, 4, %o0
418	EX(stb	%g2, [%o1 + 2])
419	addcc	%o4, %g5, %g5
420	EX(stb	%o4, [%o1 + 3])
421	addx	%g5, %g0, %g5	! I am now to lazy to optimize this (question it
422	add	%o1, 4, %o1	! is worthy). Maybe some day - with the sll/srl
423	subcc	%g4, 1, %g4	! tricks
424	bne,a	5b
425	 EX(ld	[%o0], %o4)
426	sll	%g5, 16, %g2
427	srl	%g5, 16, %g5
428	srl	%g2, 16, %g2
429	andcc	%g1, 2, %g0
430	add	%g2, %g5, %g5
4312:	be,a	3f
432	 andcc	%g1, 1, %g0
433	EX(lduh	[%o0], %o4)
434	andcc	%g1, 1, %g0
435	srl	%o4, 8, %g2
436	add	%o0, 2, %o0
437	EX(stb	%g2, [%o1])
438	add	%g5, %o4, %g5
439	EX(stb	%o4, [%o1 + 1])
440	add	%o1, 2, %o1
4413:	be,a	1f
442	 sll	%g5, 16, %o4
443	EX(ldub	[%o0], %g2)
444	sll	%g2, 8, %o4
445	EX(stb	%g2, [%o1])
446	add	%g5, %o4, %g5
447	sll	%g5, 16, %o4
4481:	addcc	%o4, %g5, %g5
449	srl	%g5, 16, %o4
450	addx	%g0, %o4, %g5
451	orcc	%o5, %g0, %g0
452	be	4f
453	 srl	%g5, 8, %o4
454	and	%g5, 0xff, %g2
455	and	%o4, 0xff, %o4
456	sll	%g2, 8, %g2
457	or	%g2, %o4, %g5
4584:	addcc	%g7, %g5, %g7
459	retl
460	 addx	%g0, %g7, %o0
461
462/* We do these strange calculations for the csum_*_from_user case only, ie.
463 * we only bother with faults on loads... */
464
465cc_fault:
466	ret
467	 clr	%o0
468