xref: /openbmc/linux/arch/xtensa/lib/umulsidi3.S (revision 9aecda97)
1/* SPDX-License-Identifier: GPL-2.0-or-later WITH GCC-exception-2.0 */
2#include <linux/linkage.h>
3#include <asm/asmmacro.h>
4#include <asm/core.h>
5
6#if XCHAL_HAVE_MUL16 || XCHAL_HAVE_MUL32 || XCHAL_HAVE_MAC16
7#define XCHAL_NO_MUL 0
8#else
9#define XCHAL_NO_MUL 1
10#endif
11
12ENTRY(__umulsidi3)
13
14#ifdef __XTENSA_CALL0_ABI__
15	abi_entry(32)
16	s32i	a12, sp, 16
17	s32i	a13, sp, 20
18	s32i	a14, sp, 24
19	s32i	a15, sp, 28
20#elif XCHAL_NO_MUL
21	/* This is not really a leaf function; allocate enough stack space
22	   to allow CALL12s to a helper function.  */
23	abi_entry(32)
24#else
25	abi_entry_default
26#endif
27
28#ifdef __XTENSA_EB__
29#define wh a2
30#define wl a3
31#else
32#define wh a3
33#define wl a2
34#endif /* __XTENSA_EB__ */
35
36	/* This code is taken from the mulsf3 routine in ieee754-sf.S.
37	   See more comments there.  */
38
39#if XCHAL_HAVE_MUL32_HIGH
40	mull	a6, a2, a3
41	muluh	wh, a2, a3
42	mov	wl, a6
43
44#else /* ! MUL32_HIGH */
45
46#if defined(__XTENSA_CALL0_ABI__) && XCHAL_NO_MUL
47	/* a0 and a8 will be clobbered by calling the multiply function
48	   but a8 is not used here and need not be saved.  */
49	s32i	a0, sp, 0
50#endif
51
52#if XCHAL_HAVE_MUL16 || XCHAL_HAVE_MUL32
53
54#define a2h a4
55#define a3h a5
56
57	/* Get the high halves of the inputs into registers.  */
58	srli	a2h, a2, 16
59	srli	a3h, a3, 16
60
61#define a2l a2
62#define a3l a3
63
64#if XCHAL_HAVE_MUL32 && !XCHAL_HAVE_MUL16
65	/* Clear the high halves of the inputs.  This does not matter
66	   for MUL16 because the high bits are ignored.  */
67	extui	a2, a2, 0, 16
68	extui	a3, a3, 0, 16
69#endif
70#endif /* MUL16 || MUL32 */
71
72
73#if XCHAL_HAVE_MUL16
74
75#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
76	mul16u	dst, xreg ## xhalf, yreg ## yhalf
77
78#elif XCHAL_HAVE_MUL32
79
80#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
81	mull	dst, xreg ## xhalf, yreg ## yhalf
82
83#elif XCHAL_HAVE_MAC16
84
85/* The preprocessor insists on inserting a space when concatenating after
86   a period in the definition of do_mul below.  These macros are a workaround
87   using underscores instead of periods when doing the concatenation.  */
88#define umul_aa_ll umul.aa.ll
89#define umul_aa_lh umul.aa.lh
90#define umul_aa_hl umul.aa.hl
91#define umul_aa_hh umul.aa.hh
92
93#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
94	umul_aa_ ## xhalf ## yhalf	xreg, yreg; \
95	rsr	dst, ACCLO
96
97#else /* no multiply hardware */
98
99#define set_arg_l(dst, src) \
100	extui	dst, src, 0, 16
101#define set_arg_h(dst, src) \
102	srli	dst, src, 16
103
104#ifdef __XTENSA_CALL0_ABI__
105#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
106	set_arg_ ## xhalf (a13, xreg); \
107	set_arg_ ## yhalf (a14, yreg); \
108	call0	.Lmul_mulsi3; \
109	mov	dst, a12
110#else
111#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
112	set_arg_ ## xhalf (a14, xreg); \
113	set_arg_ ## yhalf (a15, yreg); \
114	call12	.Lmul_mulsi3; \
115	mov	dst, a14
116#endif /* __XTENSA_CALL0_ABI__ */
117
118#endif /* no multiply hardware */
119
120	/* Add pp1 and pp2 into a6 with carry-out in a9.  */
121	do_mul(a6, a2, l, a3, h)	/* pp 1 */
122	do_mul(a11, a2, h, a3, l)	/* pp 2 */
123	movi	a9, 0
124	add	a6, a6, a11
125	bgeu	a6, a11, 1f
126	addi	a9, a9, 1
1271:
128	/* Shift the high half of a9/a6 into position in a9.  Note that
129	   this value can be safely incremented without any carry-outs.  */
130	ssai	16
131	src	a9, a9, a6
132
133	/* Compute the low word into a6.  */
134	do_mul(a11, a2, l, a3, l)	/* pp 0 */
135	sll	a6, a6
136	add	a6, a6, a11
137	bgeu	a6, a11, 1f
138	addi	a9, a9, 1
1391:
140	/* Compute the high word into wh.  */
141	do_mul(wh, a2, h, a3, h)	/* pp 3 */
142	add	wh, wh, a9
143	mov	wl, a6
144
145#endif /* !MUL32_HIGH */
146
147#if defined(__XTENSA_CALL0_ABI__) && XCHAL_NO_MUL
148	/* Restore the original return address.  */
149	l32i	a0, sp, 0
150#endif
151#ifdef __XTENSA_CALL0_ABI__
152	l32i	a12, sp, 16
153	l32i	a13, sp, 20
154	l32i	a14, sp, 24
155	l32i	a15, sp, 28
156	abi_ret(32)
157#else
158	abi_ret_default
159#endif
160
161#if XCHAL_NO_MUL
162
163	.macro	do_addx2 dst, as, at, tmp
164#if XCHAL_HAVE_ADDX
165	addx2	\dst, \as, \at
166#else
167	slli	\tmp, \as, 1
168	add	\dst, \tmp, \at
169#endif
170	.endm
171
172	.macro	do_addx4 dst, as, at, tmp
173#if XCHAL_HAVE_ADDX
174	addx4	\dst, \as, \at
175#else
176	slli	\tmp, \as, 2
177	add	\dst, \tmp, \at
178#endif
179	.endm
180
181	.macro	do_addx8 dst, as, at, tmp
182#if XCHAL_HAVE_ADDX
183	addx8	\dst, \as, \at
184#else
185	slli	\tmp, \as, 3
186	add	\dst, \tmp, \at
187#endif
188	.endm
189
190	/* For Xtensa processors with no multiply hardware, this simplified
191	   version of _mulsi3 is used for multiplying 16-bit chunks of
192	   the floating-point mantissas.  When using CALL0, this function
193	   uses a custom ABI: the inputs are passed in a13 and a14, the
194	   result is returned in a12, and a8 and a15 are clobbered.  */
195	.align	4
196.Lmul_mulsi3:
197	abi_entry_default
198
199	.macro mul_mulsi3_body dst, src1, src2, tmp1, tmp2
200	movi	\dst, 0
2011:	add	\tmp1, \src2, \dst
202	extui	\tmp2, \src1, 0, 1
203	movnez	\dst, \tmp1, \tmp2
204
205	do_addx2 \tmp1, \src2, \dst, \tmp1
206	extui	\tmp2, \src1, 1, 1
207	movnez	\dst, \tmp1, \tmp2
208
209	do_addx4 \tmp1, \src2, \dst, \tmp1
210	extui	\tmp2, \src1, 2, 1
211	movnez	\dst, \tmp1, \tmp2
212
213	do_addx8 \tmp1, \src2, \dst, \tmp1
214	extui	\tmp2, \src1, 3, 1
215	movnez	\dst, \tmp1, \tmp2
216
217	srli	\src1, \src1, 4
218	slli	\src2, \src2, 4
219	bnez	\src1, 1b
220	.endm
221
222#ifdef __XTENSA_CALL0_ABI__
223	mul_mulsi3_body a12, a13, a14, a15, a8
224#else
225	/* The result will be written into a2, so save that argument in a4.  */
226	mov	a4, a2
227	mul_mulsi3_body a2, a4, a3, a5, a6
228#endif
229	abi_ret_default
230#endif /* XCHAL_NO_MUL */
231
232ENDPROC(__umulsidi3)
233EXPORT_SYMBOL(__umulsidi3)
234