xref: /openbmc/linux/arch/sparc/lib/checksum_32.S (revision ccc319dc) 
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	/* This aligned version executes typically in 8.5 superscalar cycles, this
159	 * is the best I can do.  I say 8.5 because the final add will pair with
160	 * the next ldd in the main unrolled loop.  Thus the pipe is always full.
161	 * If you change these macros (including order of instructions),
162	 * please check the fixup code below as well.
163	 */
164#define CSUMCOPY_BIGCHUNK_ALIGNED(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7)	\
165	EX(ldd	[src + off + 0x00], t0);						\
166	EX(ldd	[src + off + 0x08], t2);						\
167	addxcc	t0, sum, sum;								\
168	EX(ldd	[src + off + 0x10], t4);						\
169	addxcc	t1, sum, sum;								\
170	EX(ldd	[src + off + 0x18], t6);						\
171	addxcc	t2, sum, sum;								\
172	EX(std	t0, [dst + off + 0x00]);						\
173	addxcc	t3, sum, sum;								\
174	EX(std	t2, [dst + off + 0x08]);						\
175	addxcc	t4, sum, sum;								\
176	EX(std	t4, [dst + off + 0x10]);						\
177	addxcc	t5, sum, sum;								\
178	EX(std	t6, [dst + off + 0x18]);						\
179	addxcc	t6, sum, sum;								\
180	addxcc	t7, sum, sum;
181
182	/* 12 superscalar cycles seems to be the limit for this case,
183	 * because of this we thus do all the ldd's together to get
184	 * Viking MXCC into streaming mode.  Ho hum...
185	 */
186#define CSUMCOPY_BIGCHUNK(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7)	\
187	EX(ldd	[src + off + 0x00], t0);					\
188	EX(ldd	[src + off + 0x08], t2);					\
189	EX(ldd	[src + off + 0x10], t4);					\
190	EX(ldd	[src + off + 0x18], t6);					\
191	EX(st	t0, [dst + off + 0x00]);					\
192	addxcc	t0, sum, sum;							\
193	EX(st	t1, [dst + off + 0x04]);					\
194	addxcc	t1, sum, sum;							\
195	EX(st	t2, [dst + off + 0x08]);					\
196	addxcc	t2, sum, sum;							\
197	EX(st	t3, [dst + off + 0x0c]);					\
198	addxcc	t3, sum, sum;							\
199	EX(st	t4, [dst + off + 0x10]);					\
200	addxcc	t4, sum, sum;							\
201	EX(st	t5, [dst + off + 0x14]);					\
202	addxcc	t5, sum, sum;							\
203	EX(st	t6, [dst + off + 0x18]);					\
204	addxcc	t6, sum, sum;							\
205	EX(st	t7, [dst + off + 0x1c]);					\
206	addxcc	t7, sum, sum;
207
208	/* Yuck, 6 superscalar cycles... */
209#define CSUMCOPY_LASTCHUNK(src, dst, sum, off, t0, t1, t2, t3)	\
210	EX(ldd	[src - off - 0x08], t0);			\
211	EX(ldd	[src - off - 0x00], t2);			\
212	addxcc	t0, sum, sum;					\
213	EX(st	t0, [dst - off - 0x08]);			\
214	addxcc	t1, sum, sum;					\
215	EX(st	t1, [dst - off - 0x04]);			\
216	addxcc	t2, sum, sum;					\
217	EX(st	t2, [dst - off - 0x00]);			\
218	addxcc	t3, sum, sum;					\
219	EX(st	t3, [dst - off + 0x04]);
220
221	/* Handle the end cruft code out of band for better cache patterns. */
222cc_end_cruft:
223	be	1f
224	 andcc	%o3, 4, %g0
225	EX(ldd	[%o0 + 0x00], %g2)
226	add	%o1, 8, %o1
227	addcc	%g2, %g7, %g7
228	add	%o0, 8, %o0
229	addxcc	%g3, %g7, %g7
230	EX(st	%g2, [%o1 - 0x08])
231	addx	%g0, %g7, %g7
232	andcc	%o3, 4, %g0
233	EX(st	%g3, [%o1 - 0x04])
2341:	be	1f
235	 andcc	%o3, 3, %o3
236	EX(ld	[%o0 + 0x00], %g2)
237	add	%o1, 4, %o1
238	addcc	%g2, %g7, %g7
239	EX(st	%g2, [%o1 - 0x04])
240	addx	%g0, %g7, %g7
241	andcc	%o3, 3, %g0
242	add	%o0, 4, %o0
2431:	be	1f
244	 addcc	%o3, -1, %g0
245	bne	2f
246	 subcc	%o3, 2, %o3
247	b	4f
248	 or	%g0, %g0, %o4
2492:	EX(lduh	[%o0 + 0x00], %o4)
250	add	%o0, 2, %o0
251	EX(sth	%o4, [%o1 + 0x00])
252	be	6f
253	 add	%o1, 2, %o1
254	sll	%o4, 16, %o4
2554:	EX(ldub	[%o0 + 0x00], %o5)
256	EX(stb	%o5, [%o1 + 0x00])
257	sll	%o5, 8, %o5
258	or	%o5, %o4, %o4
2596:	addcc	%o4, %g7, %g7
2601:	retl
261	 addx	%g0, %g7, %o0
262
263	/* Also, handle the alignment code out of band. */
264cc_dword_align:
265	cmp	%g1, 16
266	bge	1f
267	 srl	%g1, 1, %o3
2682:	cmp	%o3, 0
269	be,a	ccte
270	 andcc	%g1, 0xf, %o3
271	andcc	%o3, %o0, %g0	! Check %o0 only (%o1 has the same last 2 bits)
272	be,a	2b
273	 srl	%o3, 1, %o3
2741:	andcc	%o0, 0x1, %g0
275	bne	ccslow
276	 andcc	%o0, 0x2, %g0
277	be	1f
278	 andcc	%o0, 0x4, %g0
279	EX(lduh	[%o0 + 0x00], %g4)
280	sub	%g1, 2, %g1
281	EX(sth	%g4, [%o1 + 0x00])
282	add	%o0, 2, %o0
283	sll	%g4, 16, %g4
284	addcc	%g4, %g7, %g7
285	add	%o1, 2, %o1
286	srl	%g7, 16, %g3
287	addx	%g0, %g3, %g4
288	sll	%g7, 16, %g7
289	sll	%g4, 16, %g3
290	srl	%g7, 16, %g7
291	andcc	%o0, 0x4, %g0
292	or	%g3, %g7, %g7
2931:	be	3f
294	 andcc	%g1, 0xffffff80, %g0
295	EX(ld	[%o0 + 0x00], %g4)
296	sub	%g1, 4, %g1
297	EX(st	%g4, [%o1 + 0x00])
298	add	%o0, 4, %o0
299	addcc	%g4, %g7, %g7
300	add	%o1, 4, %o1
301	addx	%g0, %g7, %g7
302	b	3f
303	 andcc	%g1, 0xffffff80, %g0
304
305	/* Sun, you just can't beat me, you just can't.  Stop trying,
306	 * give up.  I'm serious, I am going to kick the living shit
307	 * out of you, game over, lights out.
308	 */
309	.align	8
310	.globl	__csum_partial_copy_sparc_generic
311	EXPORT_SYMBOL(__csum_partial_copy_sparc_generic)
312__csum_partial_copy_sparc_generic:
313					/* %o0=src, %o1=dest, %g1=len, %g7=sum */
314	xor	%o0, %o1, %o4		! get changing bits
315	andcc	%o4, 3, %g0		! check for mismatched alignment
316	bne	ccslow			! better this than unaligned/fixups
317	 andcc	%o0, 7, %g0		! need to align things?
318	bne	cc_dword_align		! yes, we check for short lengths there
319	 andcc	%g1, 0xffffff80, %g0	! can we use unrolled loop?
3203:	be	3f			! nope, less than one loop remains
321	 andcc	%o1, 4, %g0		! dest aligned on 4 or 8 byte boundary?
322	be	ccdbl + 4		! 8 byte aligned, kick ass
3235:	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
324	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
325	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
326	CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
327	sub	%g1, 128, %g1		! detract from length
328	addx	%g0, %g7, %g7		! add in last carry bit
329	andcc	%g1, 0xffffff80, %g0	! more to csum?
330	add	%o0, 128, %o0		! advance src ptr
331	bne	5b			! we did not go negative, continue looping
332	 add	%o1, 128, %o1		! advance dest ptr
3333:	andcc	%g1, 0x70, %o2		! can use table?
334ccmerge:be	ccte			! nope, go and check for end cruft
335	 andcc	%g1, 0xf, %o3		! get low bits of length (clears carry btw)
336	srl	%o2, 1, %o4		! begin negative offset computation
337	sethi	%hi(12f), %o5		! set up table ptr end
338	add	%o0, %o2, %o0		! advance src ptr
339	sub	%o5, %o4, %o5		! continue table calculation
340	sll	%o2, 1, %g2		! constant multiplies are fun...
341	sub	%o5, %g2, %o5		! some more adjustments
342	jmp	%o5 + %lo(12f)		! jump into it, duff style, wheee...
343	 add	%o1, %o2, %o1		! advance dest ptr (carry is clear btw)
344cctbl:	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x68,%g2,%g3,%g4,%g5)
345	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x58,%g2,%g3,%g4,%g5)
346	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x48,%g2,%g3,%g4,%g5)
347	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x38,%g2,%g3,%g4,%g5)
348	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x28,%g2,%g3,%g4,%g5)
349	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x18,%g2,%g3,%g4,%g5)
350	CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x08,%g2,%g3,%g4,%g5)
35112:	addx	%g0, %g7, %g7
352	andcc	%o3, 0xf, %g0		! check for low bits set
353ccte:	bne	cc_end_cruft		! something left, handle it out of band
354	 andcc	%o3, 8, %g0		! begin checks for that code
355	retl				! return
356	 mov	%g7, %o0		! give em the computed checksum
357ccdbl:	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
358	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
359	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
360	CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
361	sub	%g1, 128, %g1		! detract from length
362	addx	%g0, %g7, %g7		! add in last carry bit
363	andcc	%g1, 0xffffff80, %g0	! more to csum?
364	add	%o0, 128, %o0		! advance src ptr
365	bne	ccdbl			! we did not go negative, continue looping
366	 add	%o1, 128, %o1		! advance dest ptr
367	b	ccmerge			! finish it off, above
368	 andcc	%g1, 0x70, %o2		! can use table? (clears carry btw)
369
370ccslow:	cmp	%g1, 0
371	mov	0, %g5
372	bleu	4f
373	 andcc	%o0, 1, %o5
374	be,a	1f
375	 srl	%g1, 1, %g4
376	sub	%g1, 1, %g1
377	EX(ldub	[%o0], %g5)
378	add	%o0, 1, %o0
379	EX(stb	%g5, [%o1])
380	srl	%g1, 1, %g4
381	add	%o1, 1, %o1
3821:	cmp	%g4, 0
383	be,a	3f
384	 andcc	%g1, 1, %g0
385	andcc	%o0, 2, %g0
386	be,a	1f
387	 srl	%g4, 1, %g4
388	EX(lduh	[%o0], %o4)
389	sub	%g1, 2, %g1
390	srl	%o4, 8, %g2
391	sub	%g4, 1, %g4
392	EX(stb	%g2, [%o1])
393	add	%o4, %g5, %g5
394	EX(stb	%o4, [%o1 + 1])
395	add	%o0, 2, %o0
396	srl	%g4, 1, %g4
397	add	%o1, 2, %o1
3981:	cmp	%g4, 0
399	be,a	2f
400	 andcc	%g1, 2, %g0
401	EX(ld	[%o0], %o4)
4025:	srl	%o4, 24, %g2
403	srl	%o4, 16, %g3
404	EX(stb	%g2, [%o1])
405	srl	%o4, 8, %g2
406	EX(stb	%g3, [%o1 + 1])
407	add	%o0, 4, %o0
408	EX(stb	%g2, [%o1 + 2])
409	addcc	%o4, %g5, %g5
410	EX(stb	%o4, [%o1 + 3])
411	addx	%g5, %g0, %g5	! I am now to lazy to optimize this (question it
412	add	%o1, 4, %o1	! is worthy). Maybe some day - with the sll/srl
413	subcc	%g4, 1, %g4	! tricks
414	bne,a	5b
415	 EX(ld	[%o0], %o4)
416	sll	%g5, 16, %g2
417	srl	%g5, 16, %g5
418	srl	%g2, 16, %g2
419	andcc	%g1, 2, %g0
420	add	%g2, %g5, %g5
4212:	be,a	3f
422	 andcc	%g1, 1, %g0
423	EX(lduh	[%o0], %o4)
424	andcc	%g1, 1, %g0
425	srl	%o4, 8, %g2
426	add	%o0, 2, %o0
427	EX(stb	%g2, [%o1])
428	add	%g5, %o4, %g5
429	EX(stb	%o4, [%o1 + 1])
430	add	%o1, 2, %o1
4313:	be,a	1f
432	 sll	%g5, 16, %o4
433	EX(ldub	[%o0], %g2)
434	sll	%g2, 8, %o4
435	EX(stb	%g2, [%o1])
436	add	%g5, %o4, %g5
437	sll	%g5, 16, %o4
4381:	addcc	%o4, %g5, %g5
439	srl	%g5, 16, %o4
440	addx	%g0, %o4, %g5
441	orcc	%o5, %g0, %g0
442	be	4f
443	 srl	%g5, 8, %o4
444	and	%g5, 0xff, %g2
445	and	%o4, 0xff, %o4
446	sll	%g2, 8, %g2
447	or	%g2, %o4, %g5
4484:	addcc	%g7, %g5, %g7
449	retl
450	 addx	%g0, %g7, %o0
451
452/* We do these strange calculations for the csum_*_from_user case only, ie.
453 * we only bother with faults on loads... */
454
455cc_fault:
456	ret
457	 clr	%o0
458