xref: /openbmc/linux/arch/nios2/kernel/insnemu.S (revision b4e18b29)
1/* SPDX-License-Identifier: GPL-2.0-or-later */
2/*
3 *  Copyright (C) 2003-2013 Altera Corporation
4 *  All rights reserved.
5 */
6
7
8#include <linux/linkage.h>
9#include <asm/entry.h>
10
11.set noat
12.set nobreak
13
14/*
15* Explicitly allow the use of r1 (the assembler temporary register)
16* within this code. This register is normally reserved for the use of
17* the compiler.
18*/
19
20ENTRY(instruction_trap)
21	ldw	r1, PT_R1(sp)		// Restore registers
22	ldw	r2, PT_R2(sp)
23	ldw	r3, PT_R3(sp)
24	ldw	r4, PT_R4(sp)
25	ldw	r5, PT_R5(sp)
26	ldw	r6, PT_R6(sp)
27	ldw	r7, PT_R7(sp)
28	ldw	r8, PT_R8(sp)
29	ldw	r9, PT_R9(sp)
30	ldw	r10, PT_R10(sp)
31	ldw	r11, PT_R11(sp)
32	ldw	r12, PT_R12(sp)
33	ldw	r13, PT_R13(sp)
34	ldw	r14, PT_R14(sp)
35	ldw	r15, PT_R15(sp)
36	ldw	ra, PT_RA(sp)
37	ldw	fp, PT_FP(sp)
38	ldw	gp, PT_GP(sp)
39	ldw	et, PT_ESTATUS(sp)
40	wrctl	estatus, et
41	ldw	ea, PT_EA(sp)
42	ldw	et, PT_SP(sp)		/* backup sp in et */
43
44	addi	sp, sp, PT_REGS_SIZE
45
46	/* INSTRUCTION EMULATION
47	*  ---------------------
48	*
49	* Nios II processors generate exceptions for unimplemented instructions.
50	* The routines below emulate these instructions.  Depending on the
51	* processor core, the only instructions that might need to be emulated
52	* are div, divu, mul, muli, mulxss, mulxsu, and mulxuu.
53	*
54	* The emulations match the instructions, except for the following
55	* limitations:
56	*
57	* 1) The emulation routines do not emulate the use of the exception
58	*    temporary register (et) as a source operand because the exception
59	*    handler already has modified it.
60	*
61	* 2) The routines do not emulate the use of the stack pointer (sp) or
62	*    the exception return address register (ea) as a destination because
63	*    modifying these registers crashes the exception handler or the
64	*    interrupted routine.
65	*
66	* Detailed Design
67	* ---------------
68	*
69	* The emulation routines expect the contents of integer registers r0-r31
70	* to be on the stack at addresses sp, 4(sp), 8(sp), ... 124(sp).  The
71	* routines retrieve source operands from the stack and modify the
72	* destination register's value on the stack prior to the end of the
73	* exception handler.  Then all registers except the destination register
74	* are restored to their previous values.
75	*
76	* The instruction that causes the exception is found at address -4(ea).
77	* The instruction's OP and OPX fields identify the operation to be
78	* performed.
79	*
80	* One instruction, muli, is an I-type instruction that is identified by
81	* an OP field of 0x24.
82	*
83	* muli   AAAAA,BBBBB,IIIIIIIIIIIIIIII,-0x24-
84	*           27    22                6      0    <-- LSB of field
85	*
86	* The remaining emulated instructions are R-type and have an OP field
87	* of 0x3a.  Their OPX fields identify them.
88	*
89	* R-type AAAAA,BBBBB,CCCCC,XXXXXX,NNNNN,-0x3a-
90	*           27    22    17     11     6      0  <-- LSB of field
91	*
92	*
93	* Opcode Encoding.  muli is identified by its OP value.  Then OPX & 0x02
94	* is used to differentiate between the division opcodes and the
95	* remaining multiplication opcodes.
96	*
97	* Instruction   OP      OPX    OPX & 0x02
98	* -----------   ----    ----   ----------
99	* muli          0x24
100	* divu          0x3a    0x24         0
101	* div           0x3a    0x25         0
102	* mul           0x3a    0x27      != 0
103	* mulxuu        0x3a    0x07      != 0
104	* mulxsu        0x3a    0x17      != 0
105	* mulxss        0x3a    0x1f      != 0
106	*/
107
108
109	/*
110	* Save everything on the stack to make it easy for the emulation
111	* routines to retrieve the source register operands.
112	*/
113
114	addi sp, sp, -128
115	stw zero, 0(sp)	/* Save zero on stack to avoid special case for r0. */
116	stw r1, 4(sp)
117	stw r2,  8(sp)
118	stw r3, 12(sp)
119	stw r4, 16(sp)
120	stw r5, 20(sp)
121	stw r6, 24(sp)
122	stw r7, 28(sp)
123	stw r8, 32(sp)
124	stw r9, 36(sp)
125	stw r10, 40(sp)
126	stw r11, 44(sp)
127	stw r12, 48(sp)
128	stw r13, 52(sp)
129	stw r14, 56(sp)
130	stw r15, 60(sp)
131	stw r16, 64(sp)
132	stw r17, 68(sp)
133	stw r18, 72(sp)
134	stw r19, 76(sp)
135	stw r20, 80(sp)
136	stw r21, 84(sp)
137	stw r22, 88(sp)
138	stw r23, 92(sp)
139		/* Don't bother to save et.  It's already been changed. */
140	rdctl r5, estatus
141	stw r5,  100(sp)
142
143	stw gp, 104(sp)
144	stw et, 108(sp)	/* et contains previous sp value. */
145	stw fp, 112(sp)
146	stw ea, 116(sp)
147	stw ra, 120(sp)
148
149
150	/*
151	* Split the instruction into its fields.  We need 4*A, 4*B, and 4*C as
152	* offsets to the stack pointer for access to the stored register values.
153	*/
154	ldw r2,-4(ea)	/* r2 = AAAAA,BBBBB,IIIIIIIIIIIIIIII,PPPPPP */
155	roli r3, r2, 7	/* r3 = BBB,IIIIIIIIIIIIIIII,PPPPPP,AAAAA,BB */
156	roli r4, r3, 3	/* r4 = IIIIIIIIIIIIIIII,PPPPPP,AAAAA,BBBBB */
157	roli r5, r4, 2	/* r5 = IIIIIIIIIIIIII,PPPPPP,AAAAA,BBBBB,II */
158	srai r4, r4, 16	/* r4 = (sign-extended) IMM16 */
159	roli r6, r5, 5	/* r6 = XXXX,NNNNN,PPPPPP,AAAAA,BBBBB,CCCCC,XX */
160	andi r2, r2, 0x3f	/* r2 = 00000000000000000000000000,PPPPPP */
161	andi r3, r3, 0x7c	/* r3 = 0000000000000000000000000,AAAAA,00 */
162	andi r5, r5, 0x7c	/* r5 = 0000000000000000000000000,BBBBB,00 */
163	andi r6, r6, 0x7c	/* r6 = 0000000000000000000000000,CCCCC,00 */
164
165	/* Now
166	* r2 = OP
167	* r3 = 4*A
168	* r4 = IMM16 (sign extended)
169	* r5 = 4*B
170	* r6 = 4*C
171	*/
172
173	/*
174	* Get the operands.
175	*
176	* It is necessary to check for muli because it uses an I-type
177	* instruction format, while the other instructions are have an R-type
178	* format.
179	*
180	*  Prepare for either multiplication or division loop.
181	*  They both loop 32 times.
182	*/
183	movi r14, 32
184
185	add  r3, r3, sp		/* r3 = address of A-operand. */
186	ldw  r3, 0(r3)		/* r3 = A-operand. */
187	movi r7, 0x24		/* muli opcode (I-type instruction format) */
188	beq r2, r7, mul_immed /* muli doesn't use the B register as a source */
189
190	add  r5, r5, sp		/* r5 = address of B-operand. */
191	ldw  r5, 0(r5)		/* r5 = B-operand. */
192				/* r4 = SSSSSSSSSSSSSSSS,-----IMM16------ */
193				/* IMM16 not needed, align OPX portion */
194				/* r4 = SSSSSSSSSSSSSSSS,CCCCC,-OPX--,00000 */
195	srli r4, r4, 5		/* r4 = 00000,SSSSSSSSSSSSSSSS,CCCCC,-OPX-- */
196	andi r4, r4, 0x3f	/* r4 = 00000000000000000000000000,-OPX-- */
197
198	/* Now
199	* r2 = OP
200	* r3 = src1
201	* r5 = src2
202	* r4 = OPX (no longer can be muli)
203	* r6 = 4*C
204	*/
205
206
207	/*
208	*  Multiply or Divide?
209	*/
210	andi r7, r4, 0x02	/* For R-type multiply instructions,
211				   OPX & 0x02 != 0 */
212	bne r7, zero, multiply
213
214
215	/* DIVISION
216	*
217	* Divide an unsigned dividend by an unsigned divisor using
218	* a shift-and-subtract algorithm.  The example below shows
219	* 43 div 7 = 6 for 8-bit integers.  This classic algorithm uses a
220	* single register to store both the dividend and the quotient,
221	* allowing both values to be shifted with a single instruction.
222	*
223	*                               remainder dividend:quotient
224	*                               --------- -----------------
225	*   initialize                   00000000     00101011:
226	*   shift                        00000000     0101011:_
227	*   remainder >= divisor? no     00000000     0101011:0
228	*   shift                        00000000     101011:0_
229	*   remainder >= divisor? no     00000000     101011:00
230	*   shift                        00000001     01011:00_
231	*   remainder >= divisor? no     00000001     01011:000
232	*   shift                        00000010     1011:000_
233	*   remainder >= divisor? no     00000010     1011:0000
234	*   shift                        00000101     011:0000_
235	*   remainder >= divisor? no     00000101     011:00000
236	*   shift                        00001010     11:00000_
237	*   remainder >= divisor? yes    00001010     11:000001
238	*       remainder -= divisor   - 00000111
239	*                              ----------
240	*                                00000011     11:000001
241	*   shift                        00000111     1:000001_
242	*   remainder >= divisor? yes    00000111     1:0000011
243	*       remainder -= divisor   - 00000111
244	*                              ----------
245	*                                00000000     1:0000011
246	*   shift                        00000001     :0000011_
247	*   remainder >= divisor? no     00000001     :00000110
248	*
249	* The quotient is 00000110.
250	*/
251
252divide:
253	/*
254	*  Prepare for division by assuming the result
255	*  is unsigned, and storing its "sign" as 0.
256	*/
257	movi r17, 0
258
259
260	/* Which division opcode? */
261	xori r7, r4, 0x25		/* OPX of div */
262	bne r7, zero, unsigned_division
263
264
265	/*
266	*  OPX is div.  Determine and store the sign of the quotient.
267	*  Then take the absolute value of both operands.
268	*/
269	xor r17, r3, r5		/* MSB contains sign of quotient */
270	bge r3,zero,dividend_is_nonnegative
271	sub r3, zero, r3	/* -r3 */
272dividend_is_nonnegative:
273	bge r5, zero, divisor_is_nonnegative
274	sub r5, zero, r5	/* -r5 */
275divisor_is_nonnegative:
276
277
278unsigned_division:
279	/* Initialize the unsigned-division loop. */
280	movi r13, 0	/* remainder = 0 */
281
282	/* Now
283	* r3 = dividend : quotient
284	* r4 = 0x25 for div, 0x24 for divu
285	* r5 = divisor
286	* r13 = remainder
287	* r14 = loop counter (already initialized to 32)
288	* r17 = MSB contains sign of quotient
289	*/
290
291
292	/*
293	*   for (count = 32; count > 0; --count)
294	*   {
295	*/
296divide_loop:
297
298	/*
299	*       Division:
300	*
301	*       (remainder:dividend:quotient) <<= 1;
302	*/
303	slli r13, r13, 1
304	cmplt r7, r3, zero	/* r7 = MSB of r3 */
305	or r13, r13, r7
306	slli r3, r3, 1
307
308
309	/*
310	*       if (remainder >= divisor)
311	*       {
312	*           set LSB of quotient
313	*           remainder -= divisor;
314	*       }
315	*/
316	bltu r13, r5, div_skip
317	ori r3, r3, 1
318	sub r13, r13, r5
319div_skip:
320
321	/*
322	*   }
323	*/
324	subi r14, r14, 1
325	bne r14, zero, divide_loop
326
327
328	/* Now
329	* r3 = quotient
330	* r4 = 0x25 for div, 0x24 for divu
331	* r6 = 4*C
332	* r17 = MSB contains sign of quotient
333	*/
334
335
336	/*
337	*  Conditionally negate signed quotient.  If quotient is unsigned,
338	*  the sign already is initialized to 0.
339	*/
340	bge r17, zero, quotient_is_nonnegative
341	sub r3, zero, r3		/* -r3 */
342	quotient_is_nonnegative:
343
344
345	/*
346	*  Final quotient is in r3.
347	*/
348	add r6, r6, sp
349	stw r3, 0(r6)	/* write quotient to stack */
350	br restore_registers
351
352
353
354
355	/* MULTIPLICATION
356	*
357	* A "product" is the number that one gets by summing a "multiplicand"
358	* several times.  The "multiplier" specifies the number of copies of the
359	* multiplicand that are summed.
360	*
361	* Actual multiplication algorithms don't use repeated addition, however.
362	* Shift-and-add algorithms get the same answer as repeated addition, and
363	* they are faster.  To compute the lower half of a product (pppp below)
364	* one shifts the product left before adding in each of the partial
365	* products (a * mmmm) through (d * mmmm).
366	*
367	* To compute the upper half of a product (PPPP below), one adds in the
368	* partial products (d * mmmm) through (a * mmmm), each time following
369	* the add by a right shift of the product.
370	*
371	*     mmmm
372	*   * abcd
373	*   ------
374	*     ####  = d * mmmm
375	*    ####   = c * mmmm
376	*   ####    = b * mmmm
377	*  ####     = a * mmmm
378	* --------
379	* PPPPpppp
380	*
381	* The example above shows 4 partial products.  Computing actual Nios II
382	* products requires 32 partials.
383	*
384	* It is possible to compute the result of mulxsu from the result of
385	* mulxuu because the only difference between the results of these two
386	* opcodes is the value of the partial product associated with the sign
387	* bit of rA.
388	*
389	*   mulxsu = mulxuu - (rA < 0) ? rB : 0;
390	*
391	* It is possible to compute the result of mulxss from the result of
392	* mulxsu because the only difference between the results of these two
393	* opcodes is the value of the partial product associated with the sign
394	* bit of rB.
395	*
396	*   mulxss = mulxsu - (rB < 0) ? rA : 0;
397	*
398	*/
399
400mul_immed:
401	/* Opcode is muli.  Change it into mul for remainder of algorithm. */
402	mov r6, r5		/* Field B is dest register, not field C. */
403	mov r5, r4		/* Field IMM16 is src2, not field B. */
404	movi r4, 0x27		/* OPX of mul is 0x27 */
405
406multiply:
407	/* Initialize the multiplication loop. */
408	movi r9, 0	/* mul_product    = 0 */
409	movi r10, 0	/* mulxuu_product = 0 */
410	mov r11, r5	/* save original multiplier for mulxsu and mulxss */
411	mov r12, r5	/* mulxuu_multiplier (will be shifted) */
412	movi r16, 1	/* used to create "rori B,A,1" from "ror B,A,r16" */
413
414	/* Now
415	* r3 = multiplicand
416	* r5 = mul_multiplier
417	* r6 = 4 * dest_register (used later as offset to sp)
418	* r7 = temp
419	* r9 = mul_product
420	* r10 = mulxuu_product
421	* r11 = original multiplier
422	* r12 = mulxuu_multiplier
423	* r14 = loop counter (already initialized)
424	* r16 = 1
425	*/
426
427
428	/*
429	*   for (count = 32; count > 0; --count)
430	*   {
431	*/
432multiply_loop:
433
434	/*
435	*       mul_product <<= 1;
436	*       lsb = multiplier & 1;
437	*/
438	slli r9, r9, 1
439	andi r7, r12, 1
440
441	/*
442	*       if (lsb == 1)
443	*       {
444	*           mulxuu_product += multiplicand;
445	*       }
446	*/
447	beq r7, zero, mulx_skip
448	add r10, r10, r3
449	cmpltu r7, r10, r3 /* Save the carry from the MSB of mulxuu_product. */
450	ror r7, r7, r16	/* r7 = 0x80000000 on carry, or else 0x00000000 */
451mulx_skip:
452
453	/*
454	*       if (MSB of mul_multiplier == 1)
455	*       {
456	*           mul_product += multiplicand;
457	*       }
458	*/
459	bge r5, zero, mul_skip
460	add r9, r9, r3
461mul_skip:
462
463	/*
464	*       mulxuu_product >>= 1;           logical shift
465	*       mul_multiplier <<= 1;           done with MSB
466	*       mulx_multiplier >>= 1;          done with LSB
467	*/
468	srli r10, r10, 1
469	or r10, r10, r7		/* OR in the saved carry bit. */
470	slli r5, r5, 1
471	srli r12, r12, 1
472
473
474	/*
475	*   }
476	*/
477	subi r14, r14, 1
478	bne r14, zero, multiply_loop
479
480
481	/*
482	*  Multiply emulation loop done.
483	*/
484
485	/* Now
486	* r3 = multiplicand
487	* r4 = OPX
488	* r6 = 4 * dest_register (used later as offset to sp)
489	* r7 = temp
490	* r9 = mul_product
491	* r10 = mulxuu_product
492	* r11 = original multiplier
493	*/
494
495
496	/* Calculate address for result from 4 * dest_register */
497	add r6, r6, sp
498
499
500	/*
501	* Select/compute the result based on OPX.
502	*/
503
504
505	/* OPX == mul?  Then store. */
506	xori r7, r4, 0x27
507	beq r7, zero, store_product
508
509	/* It's one of the mulx.. opcodes.  Move over the result. */
510	mov r9, r10
511
512	/* OPX == mulxuu?  Then store. */
513	xori r7, r4, 0x07
514	beq r7, zero, store_product
515
516	/* Compute mulxsu
517	 *
518	 * mulxsu = mulxuu - (rA < 0) ? rB : 0;
519	 */
520	bge r3, zero, mulxsu_skip
521	sub r9, r9, r11
522mulxsu_skip:
523
524	/* OPX == mulxsu?  Then store. */
525	xori r7, r4, 0x17
526	beq r7, zero, store_product
527
528	/* Compute mulxss
529	 *
530	 * mulxss = mulxsu - (rB < 0) ? rA : 0;
531	 */
532	bge r11,zero,mulxss_skip
533	sub r9, r9, r3
534mulxss_skip:
535	/* At this point, assume that OPX is mulxss, so store*/
536
537
538store_product:
539	stw r9, 0(r6)
540
541
542restore_registers:
543			/* No need to restore r0. */
544	ldw r5, 100(sp)
545	wrctl estatus, r5
546
547	ldw r1, 4(sp)
548	ldw r2, 8(sp)
549	ldw r3, 12(sp)
550	ldw r4, 16(sp)
551	ldw r5, 20(sp)
552	ldw r6, 24(sp)
553	ldw r7, 28(sp)
554	ldw r8, 32(sp)
555	ldw r9, 36(sp)
556	ldw r10, 40(sp)
557	ldw r11, 44(sp)
558	ldw r12, 48(sp)
559	ldw r13, 52(sp)
560	ldw r14, 56(sp)
561	ldw r15, 60(sp)
562	ldw r16, 64(sp)
563	ldw r17, 68(sp)
564	ldw r18, 72(sp)
565	ldw r19, 76(sp)
566	ldw r20, 80(sp)
567	ldw r21, 84(sp)
568	ldw r22, 88(sp)
569	ldw r23, 92(sp)
570			/* Does not need to restore et */
571	ldw gp, 104(sp)
572
573	ldw fp, 112(sp)
574	ldw ea, 116(sp)
575	ldw ra, 120(sp)
576	ldw sp, 108(sp)	/* last restore sp */
577	eret
578
579.set at
580.set break
581