1 |
282 |
jeremybenn |
;; ARM NEON coprocessor Machine Description
|
2 |
|
|
;; Copyright (C) 2006, 2007, 2008, 2009 Free Software Foundation, Inc.
|
3 |
|
|
;; Written by CodeSourcery.
|
4 |
|
|
;;
|
5 |
|
|
;; This file is part of GCC.
|
6 |
|
|
;;
|
7 |
|
|
;; GCC is free software; you can redistribute it and/or modify it
|
8 |
|
|
;; under the terms of the GNU General Public License as published by
|
9 |
|
|
;; the Free Software Foundation; either version 3, or (at your option)
|
10 |
|
|
;; any later version.
|
11 |
|
|
;;
|
12 |
|
|
;; GCC is distributed in the hope that it will be useful, but
|
13 |
|
|
;; WITHOUT ANY WARRANTY; without even the implied warranty of
|
14 |
|
|
;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
15 |
|
|
;; General Public License for more details.
|
16 |
|
|
;;
|
17 |
|
|
;; You should have received a copy of the GNU General Public License
|
18 |
|
|
;; along with GCC; see the file COPYING3. If not see
|
19 |
|
|
;; .
|
20 |
|
|
|
21 |
|
|
;; Constants for unspecs.
|
22 |
|
|
(define_constants
|
23 |
|
|
[(UNSPEC_ASHIFT_SIGNED 65)
|
24 |
|
|
(UNSPEC_ASHIFT_UNSIGNED 66)
|
25 |
|
|
(UNSPEC_VABA 67)
|
26 |
|
|
(UNSPEC_VABAL 68)
|
27 |
|
|
(UNSPEC_VABD 69)
|
28 |
|
|
(UNSPEC_VABDL 70)
|
29 |
|
|
(UNSPEC_VABS 71)
|
30 |
|
|
(UNSPEC_VADD 72)
|
31 |
|
|
(UNSPEC_VADDHN 73)
|
32 |
|
|
(UNSPEC_VADDL 74)
|
33 |
|
|
(UNSPEC_VADDW 75)
|
34 |
|
|
(UNSPEC_VAND 76)
|
35 |
|
|
(UNSPEC_VBIC 77)
|
36 |
|
|
(UNSPEC_VBSL 78)
|
37 |
|
|
(UNSPEC_VCAGE 79)
|
38 |
|
|
(UNSPEC_VCAGT 80)
|
39 |
|
|
(UNSPEC_VCEQ 81)
|
40 |
|
|
(UNSPEC_VCGE 82)
|
41 |
|
|
(UNSPEC_VCGT 83)
|
42 |
|
|
(UNSPEC_VCLS 84)
|
43 |
|
|
(UNSPEC_VCLZ 85)
|
44 |
|
|
(UNSPEC_VCNT 86)
|
45 |
|
|
(UNSPEC_VCOMBINE 87)
|
46 |
|
|
(UNSPEC_VCVT 88)
|
47 |
|
|
(UNSPEC_VCVT_N 89)
|
48 |
|
|
(UNSPEC_VDUP_LANE 90)
|
49 |
|
|
(UNSPEC_VDUP_N 91)
|
50 |
|
|
(UNSPEC_VEOR 92)
|
51 |
|
|
(UNSPEC_VEXT 93)
|
52 |
|
|
(UNSPEC_VGET_HIGH 94)
|
53 |
|
|
(UNSPEC_VGET_LANE 95)
|
54 |
|
|
(UNSPEC_VGET_LOW 96)
|
55 |
|
|
(UNSPEC_VHADD 97)
|
56 |
|
|
(UNSPEC_VHSUB 98)
|
57 |
|
|
(UNSPEC_VLD1 99)
|
58 |
|
|
(UNSPEC_VLD1_DUP 100)
|
59 |
|
|
(UNSPEC_VLD1_LANE 101)
|
60 |
|
|
(UNSPEC_VLD2 102)
|
61 |
|
|
(UNSPEC_VLD2_DUP 103)
|
62 |
|
|
(UNSPEC_VLD2_LANE 104)
|
63 |
|
|
(UNSPEC_VLD3 105)
|
64 |
|
|
(UNSPEC_VLD3A 106)
|
65 |
|
|
(UNSPEC_VLD3B 107)
|
66 |
|
|
(UNSPEC_VLD3_DUP 108)
|
67 |
|
|
(UNSPEC_VLD3_LANE 109)
|
68 |
|
|
(UNSPEC_VLD4 110)
|
69 |
|
|
(UNSPEC_VLD4A 111)
|
70 |
|
|
(UNSPEC_VLD4B 112)
|
71 |
|
|
(UNSPEC_VLD4_DUP 113)
|
72 |
|
|
(UNSPEC_VLD4_LANE 114)
|
73 |
|
|
(UNSPEC_VMAX 115)
|
74 |
|
|
(UNSPEC_VMIN 116)
|
75 |
|
|
(UNSPEC_VMLA 117)
|
76 |
|
|
(UNSPEC_VMLAL 118)
|
77 |
|
|
(UNSPEC_VMLA_LANE 119)
|
78 |
|
|
(UNSPEC_VMLAL_LANE 120)
|
79 |
|
|
(UNSPEC_VMLS 121)
|
80 |
|
|
(UNSPEC_VMLSL 122)
|
81 |
|
|
(UNSPEC_VMLS_LANE 123)
|
82 |
|
|
(UNSPEC_VMLSL_LANE 124)
|
83 |
|
|
(UNSPEC_VMOVL 125)
|
84 |
|
|
(UNSPEC_VMOVN 126)
|
85 |
|
|
(UNSPEC_VMUL 127)
|
86 |
|
|
(UNSPEC_VMULL 128)
|
87 |
|
|
(UNSPEC_VMUL_LANE 129)
|
88 |
|
|
(UNSPEC_VMULL_LANE 130)
|
89 |
|
|
(UNSPEC_VMUL_N 131)
|
90 |
|
|
(UNSPEC_VMVN 132)
|
91 |
|
|
(UNSPEC_VORN 133)
|
92 |
|
|
(UNSPEC_VORR 134)
|
93 |
|
|
(UNSPEC_VPADAL 135)
|
94 |
|
|
(UNSPEC_VPADD 136)
|
95 |
|
|
(UNSPEC_VPADDL 137)
|
96 |
|
|
(UNSPEC_VPMAX 138)
|
97 |
|
|
(UNSPEC_VPMIN 139)
|
98 |
|
|
(UNSPEC_VPSMAX 140)
|
99 |
|
|
(UNSPEC_VPSMIN 141)
|
100 |
|
|
(UNSPEC_VPUMAX 142)
|
101 |
|
|
(UNSPEC_VPUMIN 143)
|
102 |
|
|
(UNSPEC_VQABS 144)
|
103 |
|
|
(UNSPEC_VQADD 145)
|
104 |
|
|
(UNSPEC_VQDMLAL 146)
|
105 |
|
|
(UNSPEC_VQDMLAL_LANE 147)
|
106 |
|
|
(UNSPEC_VQDMLSL 148)
|
107 |
|
|
(UNSPEC_VQDMLSL_LANE 149)
|
108 |
|
|
(UNSPEC_VQDMULH 150)
|
109 |
|
|
(UNSPEC_VQDMULH_LANE 151)
|
110 |
|
|
(UNSPEC_VQDMULL 152)
|
111 |
|
|
(UNSPEC_VQDMULL_LANE 153)
|
112 |
|
|
(UNSPEC_VQMOVN 154)
|
113 |
|
|
(UNSPEC_VQMOVUN 155)
|
114 |
|
|
(UNSPEC_VQNEG 156)
|
115 |
|
|
(UNSPEC_VQSHL 157)
|
116 |
|
|
(UNSPEC_VQSHL_N 158)
|
117 |
|
|
(UNSPEC_VQSHLU_N 159)
|
118 |
|
|
(UNSPEC_VQSHRN_N 160)
|
119 |
|
|
(UNSPEC_VQSHRUN_N 161)
|
120 |
|
|
(UNSPEC_VQSUB 162)
|
121 |
|
|
(UNSPEC_VRECPE 163)
|
122 |
|
|
(UNSPEC_VRECPS 164)
|
123 |
|
|
(UNSPEC_VREV16 165)
|
124 |
|
|
(UNSPEC_VREV32 166)
|
125 |
|
|
(UNSPEC_VREV64 167)
|
126 |
|
|
(UNSPEC_VRSQRTE 168)
|
127 |
|
|
(UNSPEC_VRSQRTS 169)
|
128 |
|
|
(UNSPEC_VSET_LANE 170)
|
129 |
|
|
(UNSPEC_VSHL 171)
|
130 |
|
|
(UNSPEC_VSHLL_N 172)
|
131 |
|
|
(UNSPEC_VSHL_N 173)
|
132 |
|
|
(UNSPEC_VSHR_N 174)
|
133 |
|
|
(UNSPEC_VSHRN_N 175)
|
134 |
|
|
(UNSPEC_VSLI 176)
|
135 |
|
|
(UNSPEC_VSRA_N 177)
|
136 |
|
|
(UNSPEC_VSRI 178)
|
137 |
|
|
(UNSPEC_VST1 179)
|
138 |
|
|
(UNSPEC_VST1_LANE 180)
|
139 |
|
|
(UNSPEC_VST2 181)
|
140 |
|
|
(UNSPEC_VST2_LANE 182)
|
141 |
|
|
(UNSPEC_VST3 183)
|
142 |
|
|
(UNSPEC_VST3A 184)
|
143 |
|
|
(UNSPEC_VST3B 185)
|
144 |
|
|
(UNSPEC_VST3_LANE 186)
|
145 |
|
|
(UNSPEC_VST4 187)
|
146 |
|
|
(UNSPEC_VST4A 188)
|
147 |
|
|
(UNSPEC_VST4B 189)
|
148 |
|
|
(UNSPEC_VST4_LANE 190)
|
149 |
|
|
(UNSPEC_VSTRUCTDUMMY 191)
|
150 |
|
|
(UNSPEC_VSUB 192)
|
151 |
|
|
(UNSPEC_VSUBHN 193)
|
152 |
|
|
(UNSPEC_VSUBL 194)
|
153 |
|
|
(UNSPEC_VSUBW 195)
|
154 |
|
|
(UNSPEC_VTBL 196)
|
155 |
|
|
(UNSPEC_VTBX 197)
|
156 |
|
|
(UNSPEC_VTRN1 198)
|
157 |
|
|
(UNSPEC_VTRN2 199)
|
158 |
|
|
(UNSPEC_VTST 200)
|
159 |
|
|
(UNSPEC_VUZP1 201)
|
160 |
|
|
(UNSPEC_VUZP2 202)
|
161 |
|
|
(UNSPEC_VZIP1 203)
|
162 |
|
|
(UNSPEC_VZIP2 204)])
|
163 |
|
|
|
164 |
|
|
;; Double-width vector modes.
|
165 |
|
|
(define_mode_iterator VD [V8QI V4HI V2SI V2SF])
|
166 |
|
|
|
167 |
|
|
;; Double-width vector modes plus 64-bit elements.
|
168 |
|
|
(define_mode_iterator VDX [V8QI V4HI V2SI V2SF DI])
|
169 |
|
|
|
170 |
|
|
;; Same, without floating-point elements.
|
171 |
|
|
(define_mode_iterator VDI [V8QI V4HI V2SI])
|
172 |
|
|
|
173 |
|
|
;; Quad-width vector modes.
|
174 |
|
|
(define_mode_iterator VQ [V16QI V8HI V4SI V4SF])
|
175 |
|
|
|
176 |
|
|
;; Quad-width vector modes plus 64-bit elements.
|
177 |
|
|
(define_mode_iterator VQX [V16QI V8HI V4SI V4SF V2DI])
|
178 |
|
|
|
179 |
|
|
;; Same, without floating-point elements.
|
180 |
|
|
(define_mode_iterator VQI [V16QI V8HI V4SI])
|
181 |
|
|
|
182 |
|
|
;; Same, with TImode added, for moves.
|
183 |
|
|
(define_mode_iterator VQXMOV [V16QI V8HI V4SI V4SF V2DI TI])
|
184 |
|
|
|
185 |
|
|
;; Opaque structure types wider than TImode.
|
186 |
|
|
(define_mode_iterator VSTRUCT [EI OI CI XI])
|
187 |
|
|
|
188 |
|
|
;; Opaque structure types used in table lookups (except vtbl1/vtbx1).
|
189 |
|
|
(define_mode_iterator VTAB [TI EI OI])
|
190 |
|
|
|
191 |
|
|
;; vtbl suffix for above modes.
|
192 |
|
|
(define_mode_attr VTAB_n [(TI "2") (EI "3") (OI "4")])
|
193 |
|
|
|
194 |
|
|
;; Widenable modes.
|
195 |
|
|
(define_mode_iterator VW [V8QI V4HI V2SI])
|
196 |
|
|
|
197 |
|
|
;; Narrowable modes.
|
198 |
|
|
(define_mode_iterator VN [V8HI V4SI V2DI])
|
199 |
|
|
|
200 |
|
|
;; All supported vector modes (except singleton DImode).
|
201 |
|
|
(define_mode_iterator VDQ [V8QI V16QI V4HI V8HI V2SI V4SI V2SF V4SF V2DI])
|
202 |
|
|
|
203 |
|
|
;; All supported vector modes (except those with 64-bit integer elements).
|
204 |
|
|
(define_mode_iterator VDQW [V8QI V16QI V4HI V8HI V2SI V4SI V2SF V4SF])
|
205 |
|
|
|
206 |
|
|
;; Supported integer vector modes (not 64 bit elements).
|
207 |
|
|
(define_mode_iterator VDQIW [V8QI V16QI V4HI V8HI V2SI V4SI])
|
208 |
|
|
|
209 |
|
|
;; Supported integer vector modes (not singleton DI)
|
210 |
|
|
(define_mode_iterator VDQI [V8QI V16QI V4HI V8HI V2SI V4SI V2DI])
|
211 |
|
|
|
212 |
|
|
;; Vector modes, including 64-bit integer elements.
|
213 |
|
|
(define_mode_iterator VDQX [V8QI V16QI V4HI V8HI V2SI V4SI V2SF V4SF DI V2DI])
|
214 |
|
|
|
215 |
|
|
;; Vector modes including 64-bit integer elements, but no floats.
|
216 |
|
|
(define_mode_iterator VDQIX [V8QI V16QI V4HI V8HI V2SI V4SI DI V2DI])
|
217 |
|
|
|
218 |
|
|
;; Vector modes for float->int conversions.
|
219 |
|
|
(define_mode_iterator VCVTF [V2SF V4SF])
|
220 |
|
|
|
221 |
|
|
;; Vector modes form int->float conversions.
|
222 |
|
|
(define_mode_iterator VCVTI [V2SI V4SI])
|
223 |
|
|
|
224 |
|
|
;; Vector modes for doubleword multiply-accumulate, etc. insns.
|
225 |
|
|
(define_mode_iterator VMD [V4HI V2SI V2SF])
|
226 |
|
|
|
227 |
|
|
;; Vector modes for quadword multiply-accumulate, etc. insns.
|
228 |
|
|
(define_mode_iterator VMQ [V8HI V4SI V4SF])
|
229 |
|
|
|
230 |
|
|
;; Above modes combined.
|
231 |
|
|
(define_mode_iterator VMDQ [V4HI V2SI V2SF V8HI V4SI V4SF])
|
232 |
|
|
|
233 |
|
|
;; As VMD, but integer modes only.
|
234 |
|
|
(define_mode_iterator VMDI [V4HI V2SI])
|
235 |
|
|
|
236 |
|
|
;; As VMQ, but integer modes only.
|
237 |
|
|
(define_mode_iterator VMQI [V8HI V4SI])
|
238 |
|
|
|
239 |
|
|
;; Above modes combined.
|
240 |
|
|
(define_mode_iterator VMDQI [V4HI V2SI V8HI V4SI])
|
241 |
|
|
|
242 |
|
|
;; Modes with 8-bit and 16-bit elements.
|
243 |
|
|
(define_mode_iterator VX [V8QI V4HI V16QI V8HI])
|
244 |
|
|
|
245 |
|
|
;; Modes with 8-bit elements.
|
246 |
|
|
(define_mode_iterator VE [V8QI V16QI])
|
247 |
|
|
|
248 |
|
|
;; Modes with 64-bit elements only.
|
249 |
|
|
(define_mode_iterator V64 [DI V2DI])
|
250 |
|
|
|
251 |
|
|
;; Modes with 32-bit elements only.
|
252 |
|
|
(define_mode_iterator V32 [V2SI V2SF V4SI V4SF])
|
253 |
|
|
|
254 |
|
|
;; (Opposite) mode to convert to/from for above conversions.
|
255 |
|
|
(define_mode_attr V_CVTTO [(V2SI "V2SF") (V2SF "V2SI")
|
256 |
|
|
(V4SI "V4SF") (V4SF "V4SI")])
|
257 |
|
|
|
258 |
|
|
;; Define element mode for each vector mode.
|
259 |
|
|
(define_mode_attr V_elem [(V8QI "QI") (V16QI "QI")
|
260 |
|
|
(V4HI "HI") (V8HI "HI")
|
261 |
|
|
(V2SI "SI") (V4SI "SI")
|
262 |
|
|
(V2SF "SF") (V4SF "SF")
|
263 |
|
|
(DI "DI") (V2DI "DI")])
|
264 |
|
|
|
265 |
|
|
;; Element modes for vector extraction, padded up to register size.
|
266 |
|
|
|
267 |
|
|
(define_mode_attr V_ext [(V8QI "SI") (V16QI "SI")
|
268 |
|
|
(V4HI "SI") (V8HI "SI")
|
269 |
|
|
(V2SI "SI") (V4SI "SI")
|
270 |
|
|
(V2SF "SF") (V4SF "SF")
|
271 |
|
|
(DI "DI") (V2DI "DI")])
|
272 |
|
|
|
273 |
|
|
;; Mode of pair of elements for each vector mode, to define transfer
|
274 |
|
|
;; size for structure lane/dup loads and stores.
|
275 |
|
|
(define_mode_attr V_two_elem [(V8QI "HI") (V16QI "HI")
|
276 |
|
|
(V4HI "SI") (V8HI "SI")
|
277 |
|
|
(V2SI "V2SI") (V4SI "V2SI")
|
278 |
|
|
(V2SF "V2SF") (V4SF "V2SF")
|
279 |
|
|
(DI "V2DI") (V2DI "V2DI")])
|
280 |
|
|
|
281 |
|
|
;; Similar, for three elements.
|
282 |
|
|
;; ??? Should we define extra modes so that sizes of all three-element
|
283 |
|
|
;; accesses can be accurately represented?
|
284 |
|
|
(define_mode_attr V_three_elem [(V8QI "SI") (V16QI "SI")
|
285 |
|
|
(V4HI "V4HI") (V8HI "V4HI")
|
286 |
|
|
(V2SI "V4SI") (V4SI "V4SI")
|
287 |
|
|
(V2SF "V4SF") (V4SF "V4SF")
|
288 |
|
|
(DI "EI") (V2DI "EI")])
|
289 |
|
|
|
290 |
|
|
;; Similar, for four elements.
|
291 |
|
|
(define_mode_attr V_four_elem [(V8QI "SI") (V16QI "SI")
|
292 |
|
|
(V4HI "V4HI") (V8HI "V4HI")
|
293 |
|
|
(V2SI "V4SI") (V4SI "V4SI")
|
294 |
|
|
(V2SF "V4SF") (V4SF "V4SF")
|
295 |
|
|
(DI "OI") (V2DI "OI")])
|
296 |
|
|
|
297 |
|
|
;; Register width from element mode
|
298 |
|
|
(define_mode_attr V_reg [(V8QI "P") (V16QI "q")
|
299 |
|
|
(V4HI "P") (V8HI "q")
|
300 |
|
|
(V2SI "P") (V4SI "q")
|
301 |
|
|
(V2SF "P") (V4SF "q")
|
302 |
|
|
(DI "P") (V2DI "q")])
|
303 |
|
|
|
304 |
|
|
;; Wider modes with the same number of elements.
|
305 |
|
|
(define_mode_attr V_widen [(V8QI "V8HI") (V4HI "V4SI") (V2SI "V2DI")])
|
306 |
|
|
|
307 |
|
|
;; Narrower modes with the same number of elements.
|
308 |
|
|
(define_mode_attr V_narrow [(V8HI "V8QI") (V4SI "V4HI") (V2DI "V2SI")])
|
309 |
|
|
|
310 |
|
|
;; Modes with half the number of equal-sized elements.
|
311 |
|
|
(define_mode_attr V_HALF [(V16QI "V8QI") (V8HI "V4HI")
|
312 |
|
|
(V4SI "V2SI") (V4SF "V2SF")
|
313 |
|
|
(V2DI "DI")])
|
314 |
|
|
|
315 |
|
|
;; Same, but lower-case.
|
316 |
|
|
(define_mode_attr V_half [(V16QI "v8qi") (V8HI "v4hi")
|
317 |
|
|
(V4SI "v2si") (V4SF "v2sf")
|
318 |
|
|
(V2DI "di")])
|
319 |
|
|
|
320 |
|
|
;; Modes with twice the number of equal-sized elements.
|
321 |
|
|
(define_mode_attr V_DOUBLE [(V8QI "V16QI") (V4HI "V8HI")
|
322 |
|
|
(V2SI "V4SI") (V2SF "V4SF")
|
323 |
|
|
(DI "V2DI")])
|
324 |
|
|
|
325 |
|
|
;; Same, but lower-case.
|
326 |
|
|
(define_mode_attr V_double [(V8QI "v16qi") (V4HI "v8hi")
|
327 |
|
|
(V2SI "v4si") (V2SF "v4sf")
|
328 |
|
|
(DI "v2di")])
|
329 |
|
|
|
330 |
|
|
;; Modes with double-width elements.
|
331 |
|
|
(define_mode_attr V_double_width [(V8QI "V4HI") (V16QI "V8HI")
|
332 |
|
|
(V4HI "V2SI") (V8HI "V4SI")
|
333 |
|
|
(V2SI "DI") (V4SI "V2DI")])
|
334 |
|
|
|
335 |
|
|
;; Mode of result of comparison operations (and bit-select operand 1).
|
336 |
|
|
(define_mode_attr V_cmp_result [(V8QI "V8QI") (V16QI "V16QI")
|
337 |
|
|
(V4HI "V4HI") (V8HI "V8HI")
|
338 |
|
|
(V2SI "V2SI") (V4SI "V4SI")
|
339 |
|
|
(V2SF "V2SI") (V4SF "V4SI")
|
340 |
|
|
(DI "DI") (V2DI "V2DI")])
|
341 |
|
|
|
342 |
|
|
;; Get element type from double-width mode, for operations where we don't care
|
343 |
|
|
;; about signedness.
|
344 |
|
|
(define_mode_attr V_if_elem [(V8QI "i8") (V16QI "i8")
|
345 |
|
|
(V4HI "i16") (V8HI "i16")
|
346 |
|
|
(V2SI "i32") (V4SI "i32")
|
347 |
|
|
(DI "i64") (V2DI "i64")
|
348 |
|
|
(V2SF "f32") (V4SF "f32")])
|
349 |
|
|
|
350 |
|
|
;; Same, but for operations which work on signed values.
|
351 |
|
|
(define_mode_attr V_s_elem [(V8QI "s8") (V16QI "s8")
|
352 |
|
|
(V4HI "s16") (V8HI "s16")
|
353 |
|
|
(V2SI "s32") (V4SI "s32")
|
354 |
|
|
(DI "s64") (V2DI "s64")
|
355 |
|
|
(V2SF "f32") (V4SF "f32")])
|
356 |
|
|
|
357 |
|
|
;; Same, but for operations which work on unsigned values.
|
358 |
|
|
(define_mode_attr V_u_elem [(V8QI "u8") (V16QI "u8")
|
359 |
|
|
(V4HI "u16") (V8HI "u16")
|
360 |
|
|
(V2SI "u32") (V4SI "u32")
|
361 |
|
|
(DI "u64") (V2DI "u64")
|
362 |
|
|
(V2SF "f32") (V4SF "f32")])
|
363 |
|
|
|
364 |
|
|
;; Element types for extraction of unsigned scalars.
|
365 |
|
|
(define_mode_attr V_uf_sclr [(V8QI "u8") (V16QI "u8")
|
366 |
|
|
(V4HI "u16") (V8HI "u16")
|
367 |
|
|
(V2SI "32") (V4SI "32")
|
368 |
|
|
(V2SF "32") (V4SF "32")])
|
369 |
|
|
|
370 |
|
|
(define_mode_attr V_sz_elem [(V8QI "8") (V16QI "8")
|
371 |
|
|
(V4HI "16") (V8HI "16")
|
372 |
|
|
(V2SI "32") (V4SI "32")
|
373 |
|
|
(DI "64") (V2DI "64")
|
374 |
|
|
(V2SF "32") (V4SF "32")])
|
375 |
|
|
|
376 |
|
|
;; Element sizes for duplicating ARM registers to all elements of a vector.
|
377 |
|
|
(define_mode_attr VD_dup [(V8QI "8") (V4HI "16") (V2SI "32") (V2SF "32")])
|
378 |
|
|
|
379 |
|
|
;; Opaque integer types for results of pair-forming intrinsics (vtrn, etc.)
|
380 |
|
|
(define_mode_attr V_PAIR [(V8QI "TI") (V16QI "OI")
|
381 |
|
|
(V4HI "TI") (V8HI "OI")
|
382 |
|
|
(V2SI "TI") (V4SI "OI")
|
383 |
|
|
(V2SF "TI") (V4SF "OI")
|
384 |
|
|
(DI "TI") (V2DI "OI")])
|
385 |
|
|
|
386 |
|
|
;; Same, but lower-case.
|
387 |
|
|
(define_mode_attr V_pair [(V8QI "ti") (V16QI "oi")
|
388 |
|
|
(V4HI "ti") (V8HI "oi")
|
389 |
|
|
(V2SI "ti") (V4SI "oi")
|
390 |
|
|
(V2SF "ti") (V4SF "oi")
|
391 |
|
|
(DI "ti") (V2DI "oi")])
|
392 |
|
|
|
393 |
|
|
;; Operations on two halves of a quadword vector.
|
394 |
|
|
(define_code_iterator vqh_ops [plus smin smax umin umax])
|
395 |
|
|
|
396 |
|
|
;; Same, without unsigned variants (for use with *SFmode pattern).
|
397 |
|
|
(define_code_iterator vqhs_ops [plus smin smax])
|
398 |
|
|
|
399 |
|
|
;; Assembler mnemonics for above codes.
|
400 |
|
|
(define_code_attr VQH_mnem [(plus "vadd") (smin "vmin") (smax "vmax")
|
401 |
|
|
(umin "vmin") (umax "vmax")])
|
402 |
|
|
|
403 |
|
|
;; Signs of above, where relevant.
|
404 |
|
|
(define_code_attr VQH_sign [(plus "i") (smin "s") (smax "s") (umin "u")
|
405 |
|
|
(umax "u")])
|
406 |
|
|
|
407 |
|
|
;; Extra suffix on some 64-bit insn names (to avoid collision with standard
|
408 |
|
|
;; names which we don't want to define).
|
409 |
|
|
(define_mode_attr V_suf64 [(V8QI "") (V16QI "")
|
410 |
|
|
(V4HI "") (V8HI "")
|
411 |
|
|
(V2SI "") (V4SI "")
|
412 |
|
|
(V2SF "") (V4SF "")
|
413 |
|
|
(DI "_neon") (V2DI "")])
|
414 |
|
|
|
415 |
|
|
;; Scalars to be presented to scalar multiplication instructions
|
416 |
|
|
;; must satisfy the following constraints.
|
417 |
|
|
;; 1. If the mode specifies 16-bit elements, the scalar must be in D0-D7.
|
418 |
|
|
;; 2. If the mode specifies 32-bit elements, the scalar must be in D0-D15.
|
419 |
|
|
;; This mode attribute is used to obtain the correct register constraints.
|
420 |
|
|
(define_mode_attr scalar_mul_constraint [(V4HI "x") (V2SI "t") (V2SF "t")
|
421 |
|
|
(V8HI "x") (V4SI "t") (V4SF "t")])
|
422 |
|
|
|
423 |
|
|
;; Attribute used to permit string comparisons against in
|
424 |
|
|
;; neon_type attribute definitions.
|
425 |
|
|
(define_attr "vqh_mnem" "vadd,vmin,vmax" (const_string "vadd"))
|
426 |
|
|
|
427 |
|
|
;; Predicates used for setting neon_type
|
428 |
|
|
|
429 |
|
|
(define_mode_attr Is_float_mode [(V8QI "false") (V16QI "false")
|
430 |
|
|
(V4HI "false") (V8HI "false")
|
431 |
|
|
(V2SI "false") (V4SI "false")
|
432 |
|
|
(V2SF "true") (V4SF "true")
|
433 |
|
|
(DI "false") (V2DI "false")])
|
434 |
|
|
|
435 |
|
|
(define_mode_attr Scalar_mul_8_16 [(V8QI "true") (V16QI "true")
|
436 |
|
|
(V4HI "true") (V8HI "true")
|
437 |
|
|
(V2SI "false") (V4SI "false")
|
438 |
|
|
(V2SF "false") (V4SF "false")
|
439 |
|
|
(DI "false") (V2DI "false")])
|
440 |
|
|
|
441 |
|
|
|
442 |
|
|
(define_mode_attr Is_d_reg [(V8QI "true") (V16QI "false")
|
443 |
|
|
(V4HI "true") (V8HI "false")
|
444 |
|
|
(V2SI "true") (V4SI "false")
|
445 |
|
|
(V2SF "true") (V4SF "false")
|
446 |
|
|
(DI "true") (V2DI "false")])
|
447 |
|
|
|
448 |
|
|
(define_mode_attr V_mode_nunits [(V8QI "8") (V16QI "16")
|
449 |
|
|
(V4HI "4") (V8HI "8")
|
450 |
|
|
(V2SI "2") (V4SI "4")
|
451 |
|
|
(V2SF "2") (V4SF "4")
|
452 |
|
|
(DI "1") (V2DI "2")])
|
453 |
|
|
|
454 |
|
|
(define_insn "*neon_mov"
|
455 |
|
|
[(set (match_operand:VD 0 "nonimmediate_operand"
|
456 |
|
|
"=w,Uv,w, w, ?r,?w,?r,?r, ?Us")
|
457 |
|
|
(match_operand:VD 1 "general_operand"
|
458 |
|
|
" w,w, Dn,Uvi, w, r, r, Usi,r"))]
|
459 |
|
|
"TARGET_NEON
|
460 |
|
|
&& (register_operand (operands[0], mode)
|
461 |
|
|
|| register_operand (operands[1], mode))"
|
462 |
|
|
{
|
463 |
|
|
if (which_alternative == 2)
|
464 |
|
|
{
|
465 |
|
|
int width, is_valid;
|
466 |
|
|
static char templ[40];
|
467 |
|
|
|
468 |
|
|
is_valid = neon_immediate_valid_for_move (operands[1], mode,
|
469 |
|
|
&operands[1], &width);
|
470 |
|
|
|
471 |
|
|
gcc_assert (is_valid != 0);
|
472 |
|
|
|
473 |
|
|
if (width == 0)
|
474 |
|
|
return "vmov.f32\t%P0, %1 @ ";
|
475 |
|
|
else
|
476 |
|
|
sprintf (templ, "vmov.i%d\t%%P0, %%1 @ ", width);
|
477 |
|
|
|
478 |
|
|
return templ;
|
479 |
|
|
}
|
480 |
|
|
|
481 |
|
|
/* FIXME: If the memory layout is changed in big-endian mode, output_move_vfp
|
482 |
|
|
below must be changed to output_move_neon (which will use the
|
483 |
|
|
element/structure loads/stores), and the constraint changed to 'Um' instead
|
484 |
|
|
of 'Uv'. */
|
485 |
|
|
|
486 |
|
|
switch (which_alternative)
|
487 |
|
|
{
|
488 |
|
|
case 0: return "vmov\t%P0, %P1 @ ";
|
489 |
|
|
case 1: case 3: return output_move_vfp (operands);
|
490 |
|
|
case 2: gcc_unreachable ();
|
491 |
|
|
case 4: return "vmov\t%Q0, %R0, %P1 @ ";
|
492 |
|
|
case 5: return "vmov\t%P0, %Q1, %R1 @ ";
|
493 |
|
|
default: return output_move_double (operands);
|
494 |
|
|
}
|
495 |
|
|
}
|
496 |
|
|
[(set_attr "neon_type" "neon_int_1,*,neon_vmov,*,neon_mrrc,neon_mcr_2_mcrr,*,*,*")
|
497 |
|
|
(set_attr "type" "*,f_stored,*,f_loadd,*,*,alu,load2,store2")
|
498 |
|
|
(set_attr "insn" "*,*,*,*,*,*,mov,*,*")
|
499 |
|
|
(set_attr "length" "4,4,4,4,4,4,8,8,8")
|
500 |
|
|
(set_attr "pool_range" "*,*,*,1020,*,*,*,1020,*")
|
501 |
|
|
(set_attr "neg_pool_range" "*,*,*,1008,*,*,*,1008,*")])
|
502 |
|
|
|
503 |
|
|
(define_insn "*neon_mov"
|
504 |
|
|
[(set (match_operand:VQXMOV 0 "nonimmediate_operand"
|
505 |
|
|
"=w,Un,w, w, ?r,?w,?r,?r, ?Us")
|
506 |
|
|
(match_operand:VQXMOV 1 "general_operand"
|
507 |
|
|
" w,w, Dn,Uni, w, r, r, Usi, r"))]
|
508 |
|
|
"TARGET_NEON
|
509 |
|
|
&& (register_operand (operands[0], mode)
|
510 |
|
|
|| register_operand (operands[1], mode))"
|
511 |
|
|
{
|
512 |
|
|
if (which_alternative == 2)
|
513 |
|
|
{
|
514 |
|
|
int width, is_valid;
|
515 |
|
|
static char templ[40];
|
516 |
|
|
|
517 |
|
|
is_valid = neon_immediate_valid_for_move (operands[1], mode,
|
518 |
|
|
&operands[1], &width);
|
519 |
|
|
|
520 |
|
|
gcc_assert (is_valid != 0);
|
521 |
|
|
|
522 |
|
|
if (width == 0)
|
523 |
|
|
return "vmov.f32\t%q0, %1 @ ";
|
524 |
|
|
else
|
525 |
|
|
sprintf (templ, "vmov.i%d\t%%q0, %%1 @ ", width);
|
526 |
|
|
|
527 |
|
|
return templ;
|
528 |
|
|
}
|
529 |
|
|
|
530 |
|
|
switch (which_alternative)
|
531 |
|
|
{
|
532 |
|
|
case 0: return "vmov\t%q0, %q1 @ ";
|
533 |
|
|
case 1: case 3: return output_move_neon (operands);
|
534 |
|
|
case 2: gcc_unreachable ();
|
535 |
|
|
case 4: return "vmov\t%Q0, %R0, %e1 @ \;vmov\t%J0, %K0, %f1";
|
536 |
|
|
case 5: return "vmov\t%e0, %Q1, %R1 @ \;vmov\t%f0, %J1, %K1";
|
537 |
|
|
default: return output_move_quad (operands);
|
538 |
|
|
}
|
539 |
|
|
}
|
540 |
|
|
[(set_attr "neon_type" "neon_int_1,neon_stm_2,neon_vmov,neon_ldm_2,\
|
541 |
|
|
neon_mrrc,neon_mcr_2_mcrr,*,*,*")
|
542 |
|
|
(set_attr "type" "*,*,*,*,*,*,alu,load4,store4")
|
543 |
|
|
(set_attr "insn" "*,*,*,*,*,*,mov,*,*")
|
544 |
|
|
(set_attr "length" "4,8,4,8,8,8,16,8,16")
|
545 |
|
|
(set_attr "pool_range" "*,*,*,1020,*,*,*,1020,*")
|
546 |
|
|
(set_attr "neg_pool_range" "*,*,*,1008,*,*,*,1008,*")])
|
547 |
|
|
|
548 |
|
|
(define_expand "movti"
|
549 |
|
|
[(set (match_operand:TI 0 "nonimmediate_operand" "")
|
550 |
|
|
(match_operand:TI 1 "general_operand" ""))]
|
551 |
|
|
"TARGET_NEON"
|
552 |
|
|
{
|
553 |
|
|
if (can_create_pseudo_p ())
|
554 |
|
|
{
|
555 |
|
|
if (GET_CODE (operands[0]) != REG)
|
556 |
|
|
operands[1] = force_reg (TImode, operands[1]);
|
557 |
|
|
}
|
558 |
|
|
})
|
559 |
|
|
|
560 |
|
|
(define_expand "mov"
|
561 |
|
|
[(set (match_operand:VSTRUCT 0 "nonimmediate_operand" "")
|
562 |
|
|
(match_operand:VSTRUCT 1 "general_operand" ""))]
|
563 |
|
|
"TARGET_NEON"
|
564 |
|
|
{
|
565 |
|
|
if (can_create_pseudo_p ())
|
566 |
|
|
{
|
567 |
|
|
if (GET_CODE (operands[0]) != REG)
|
568 |
|
|
operands[1] = force_reg (mode, operands[1]);
|
569 |
|
|
}
|
570 |
|
|
})
|
571 |
|
|
|
572 |
|
|
(define_insn "*neon_mov"
|
573 |
|
|
[(set (match_operand:VSTRUCT 0 "nonimmediate_operand" "=w,Ut,w")
|
574 |
|
|
(match_operand:VSTRUCT 1 "general_operand" " w,w, Ut"))]
|
575 |
|
|
"TARGET_NEON
|
576 |
|
|
&& (register_operand (operands[0], mode)
|
577 |
|
|
|| register_operand (operands[1], mode))"
|
578 |
|
|
{
|
579 |
|
|
switch (which_alternative)
|
580 |
|
|
{
|
581 |
|
|
case 0: return "#";
|
582 |
|
|
case 1: case 2: return output_move_neon (operands);
|
583 |
|
|
default: gcc_unreachable ();
|
584 |
|
|
}
|
585 |
|
|
}
|
586 |
|
|
[(set_attr "neon_type" "neon_int_1,neon_stm_2,neon_ldm_2")
|
587 |
|
|
(set (attr "length") (symbol_ref "arm_attr_length_move_neon (insn)"))])
|
588 |
|
|
|
589 |
|
|
(define_split
|
590 |
|
|
[(set (match_operand:EI 0 "s_register_operand" "")
|
591 |
|
|
(match_operand:EI 1 "s_register_operand" ""))]
|
592 |
|
|
"TARGET_NEON && reload_completed"
|
593 |
|
|
[(set (match_dup 0) (match_dup 1))
|
594 |
|
|
(set (match_dup 2) (match_dup 3))]
|
595 |
|
|
{
|
596 |
|
|
int rdest = REGNO (operands[0]);
|
597 |
|
|
int rsrc = REGNO (operands[1]);
|
598 |
|
|
rtx dest[2], src[2];
|
599 |
|
|
|
600 |
|
|
dest[0] = gen_rtx_REG (TImode, rdest);
|
601 |
|
|
src[0] = gen_rtx_REG (TImode, rsrc);
|
602 |
|
|
dest[1] = gen_rtx_REG (DImode, rdest + 4);
|
603 |
|
|
src[1] = gen_rtx_REG (DImode, rsrc + 4);
|
604 |
|
|
|
605 |
|
|
neon_disambiguate_copy (operands, dest, src, 2);
|
606 |
|
|
})
|
607 |
|
|
|
608 |
|
|
(define_split
|
609 |
|
|
[(set (match_operand:OI 0 "s_register_operand" "")
|
610 |
|
|
(match_operand:OI 1 "s_register_operand" ""))]
|
611 |
|
|
"TARGET_NEON && reload_completed"
|
612 |
|
|
[(set (match_dup 0) (match_dup 1))
|
613 |
|
|
(set (match_dup 2) (match_dup 3))]
|
614 |
|
|
{
|
615 |
|
|
int rdest = REGNO (operands[0]);
|
616 |
|
|
int rsrc = REGNO (operands[1]);
|
617 |
|
|
rtx dest[2], src[2];
|
618 |
|
|
|
619 |
|
|
dest[0] = gen_rtx_REG (TImode, rdest);
|
620 |
|
|
src[0] = gen_rtx_REG (TImode, rsrc);
|
621 |
|
|
dest[1] = gen_rtx_REG (TImode, rdest + 4);
|
622 |
|
|
src[1] = gen_rtx_REG (TImode, rsrc + 4);
|
623 |
|
|
|
624 |
|
|
neon_disambiguate_copy (operands, dest, src, 2);
|
625 |
|
|
})
|
626 |
|
|
|
627 |
|
|
(define_split
|
628 |
|
|
[(set (match_operand:CI 0 "s_register_operand" "")
|
629 |
|
|
(match_operand:CI 1 "s_register_operand" ""))]
|
630 |
|
|
"TARGET_NEON && reload_completed"
|
631 |
|
|
[(set (match_dup 0) (match_dup 1))
|
632 |
|
|
(set (match_dup 2) (match_dup 3))
|
633 |
|
|
(set (match_dup 4) (match_dup 5))]
|
634 |
|
|
{
|
635 |
|
|
int rdest = REGNO (operands[0]);
|
636 |
|
|
int rsrc = REGNO (operands[1]);
|
637 |
|
|
rtx dest[3], src[3];
|
638 |
|
|
|
639 |
|
|
dest[0] = gen_rtx_REG (TImode, rdest);
|
640 |
|
|
src[0] = gen_rtx_REG (TImode, rsrc);
|
641 |
|
|
dest[1] = gen_rtx_REG (TImode, rdest + 4);
|
642 |
|
|
src[1] = gen_rtx_REG (TImode, rsrc + 4);
|
643 |
|
|
dest[2] = gen_rtx_REG (TImode, rdest + 8);
|
644 |
|
|
src[2] = gen_rtx_REG (TImode, rsrc + 8);
|
645 |
|
|
|
646 |
|
|
neon_disambiguate_copy (operands, dest, src, 3);
|
647 |
|
|
})
|
648 |
|
|
|
649 |
|
|
(define_split
|
650 |
|
|
[(set (match_operand:XI 0 "s_register_operand" "")
|
651 |
|
|
(match_operand:XI 1 "s_register_operand" ""))]
|
652 |
|
|
"TARGET_NEON && reload_completed"
|
653 |
|
|
[(set (match_dup 0) (match_dup 1))
|
654 |
|
|
(set (match_dup 2) (match_dup 3))
|
655 |
|
|
(set (match_dup 4) (match_dup 5))
|
656 |
|
|
(set (match_dup 6) (match_dup 7))]
|
657 |
|
|
{
|
658 |
|
|
int rdest = REGNO (operands[0]);
|
659 |
|
|
int rsrc = REGNO (operands[1]);
|
660 |
|
|
rtx dest[4], src[4];
|
661 |
|
|
|
662 |
|
|
dest[0] = gen_rtx_REG (TImode, rdest);
|
663 |
|
|
src[0] = gen_rtx_REG (TImode, rsrc);
|
664 |
|
|
dest[1] = gen_rtx_REG (TImode, rdest + 4);
|
665 |
|
|
src[1] = gen_rtx_REG (TImode, rsrc + 4);
|
666 |
|
|
dest[2] = gen_rtx_REG (TImode, rdest + 8);
|
667 |
|
|
src[2] = gen_rtx_REG (TImode, rsrc + 8);
|
668 |
|
|
dest[3] = gen_rtx_REG (TImode, rdest + 12);
|
669 |
|
|
src[3] = gen_rtx_REG (TImode, rsrc + 12);
|
670 |
|
|
|
671 |
|
|
neon_disambiguate_copy (operands, dest, src, 4);
|
672 |
|
|
})
|
673 |
|
|
|
674 |
|
|
(define_insn "vec_set_internal"
|
675 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
676 |
|
|
(vec_merge:VD
|
677 |
|
|
(vec_duplicate:VD
|
678 |
|
|
(match_operand: 1 "s_register_operand" "r"))
|
679 |
|
|
(match_operand:VD 3 "s_register_operand" "0")
|
680 |
|
|
(match_operand:SI 2 "immediate_operand" "i")))]
|
681 |
|
|
"TARGET_NEON"
|
682 |
|
|
{
|
683 |
|
|
int elt = ffs ((int) INTVAL (operands[2]) - 1);
|
684 |
|
|
if (BYTES_BIG_ENDIAN)
|
685 |
|
|
elt = GET_MODE_NUNITS (mode) - 1 - elt;
|
686 |
|
|
operands[2] = GEN_INT (elt);
|
687 |
|
|
|
688 |
|
|
return "vmov%?.\t%P0[%c2], %1";
|
689 |
|
|
}
|
690 |
|
|
[(set_attr "predicable" "yes")
|
691 |
|
|
(set_attr "neon_type" "neon_mcr")])
|
692 |
|
|
|
693 |
|
|
(define_insn "vec_set_internal"
|
694 |
|
|
[(set (match_operand:VQ 0 "s_register_operand" "=w")
|
695 |
|
|
(vec_merge:VQ
|
696 |
|
|
(vec_duplicate:VQ
|
697 |
|
|
(match_operand: 1 "s_register_operand" "r"))
|
698 |
|
|
(match_operand:VQ 3 "s_register_operand" "0")
|
699 |
|
|
(match_operand:SI 2 "immediate_operand" "i")))]
|
700 |
|
|
"TARGET_NEON"
|
701 |
|
|
{
|
702 |
|
|
HOST_WIDE_INT elem = ffs ((int) INTVAL (operands[2])) - 1;
|
703 |
|
|
int half_elts = GET_MODE_NUNITS (mode) / 2;
|
704 |
|
|
int elt = elem % half_elts;
|
705 |
|
|
int hi = (elem / half_elts) * 2;
|
706 |
|
|
int regno = REGNO (operands[0]);
|
707 |
|
|
|
708 |
|
|
if (BYTES_BIG_ENDIAN)
|
709 |
|
|
elt = half_elts - 1 - elt;
|
710 |
|
|
|
711 |
|
|
operands[0] = gen_rtx_REG (mode, regno + hi);
|
712 |
|
|
operands[2] = GEN_INT (elt);
|
713 |
|
|
|
714 |
|
|
return "vmov%?.\t%P0[%c2], %1";
|
715 |
|
|
}
|
716 |
|
|
[(set_attr "predicable" "yes")
|
717 |
|
|
(set_attr "neon_type" "neon_mcr")]
|
718 |
|
|
)
|
719 |
|
|
|
720 |
|
|
(define_insn "vec_setv2di_internal"
|
721 |
|
|
[(set (match_operand:V2DI 0 "s_register_operand" "=w")
|
722 |
|
|
(vec_merge:V2DI
|
723 |
|
|
(vec_duplicate:V2DI
|
724 |
|
|
(match_operand:DI 1 "s_register_operand" "r"))
|
725 |
|
|
(match_operand:V2DI 3 "s_register_operand" "0")
|
726 |
|
|
(match_operand:SI 2 "immediate_operand" "i")))]
|
727 |
|
|
"TARGET_NEON"
|
728 |
|
|
{
|
729 |
|
|
HOST_WIDE_INT elem = ffs ((int) INTVAL (operands[2])) - 1;
|
730 |
|
|
int regno = REGNO (operands[0]) + 2 * elem;
|
731 |
|
|
|
732 |
|
|
operands[0] = gen_rtx_REG (DImode, regno);
|
733 |
|
|
|
734 |
|
|
return "vmov%?.64\t%P0, %Q1, %R1";
|
735 |
|
|
}
|
736 |
|
|
[(set_attr "predicable" "yes")
|
737 |
|
|
(set_attr "neon_type" "neon_mcr_2_mcrr")]
|
738 |
|
|
)
|
739 |
|
|
|
740 |
|
|
(define_expand "vec_set"
|
741 |
|
|
[(match_operand:VDQ 0 "s_register_operand" "")
|
742 |
|
|
(match_operand: 1 "s_register_operand" "")
|
743 |
|
|
(match_operand:SI 2 "immediate_operand" "")]
|
744 |
|
|
"TARGET_NEON"
|
745 |
|
|
{
|
746 |
|
|
HOST_WIDE_INT elem = (HOST_WIDE_INT) 1 << INTVAL (operands[2]);
|
747 |
|
|
emit_insn (gen_vec_set_internal (operands[0], operands[1],
|
748 |
|
|
GEN_INT (elem), operands[0]));
|
749 |
|
|
DONE;
|
750 |
|
|
})
|
751 |
|
|
|
752 |
|
|
(define_insn "vec_extract"
|
753 |
|
|
[(set (match_operand: 0 "s_register_operand" "=r")
|
754 |
|
|
(vec_select:
|
755 |
|
|
(match_operand:VD 1 "s_register_operand" "w")
|
756 |
|
|
(parallel [(match_operand:SI 2 "immediate_operand" "i")])))]
|
757 |
|
|
"TARGET_NEON"
|
758 |
|
|
{
|
759 |
|
|
if (BYTES_BIG_ENDIAN)
|
760 |
|
|
{
|
761 |
|
|
int elt = INTVAL (operands[2]);
|
762 |
|
|
elt = GET_MODE_NUNITS (mode) - 1 - elt;
|
763 |
|
|
operands[2] = GEN_INT (elt);
|
764 |
|
|
}
|
765 |
|
|
return "vmov%?.\t%0, %P1[%c2]";
|
766 |
|
|
}
|
767 |
|
|
[(set_attr "predicable" "yes")
|
768 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
769 |
|
|
)
|
770 |
|
|
|
771 |
|
|
(define_insn "vec_extract"
|
772 |
|
|
[(set (match_operand: 0 "s_register_operand" "=r")
|
773 |
|
|
(vec_select:
|
774 |
|
|
(match_operand:VQ 1 "s_register_operand" "w")
|
775 |
|
|
(parallel [(match_operand:SI 2 "immediate_operand" "i")])))]
|
776 |
|
|
"TARGET_NEON"
|
777 |
|
|
{
|
778 |
|
|
int half_elts = GET_MODE_NUNITS (mode) / 2;
|
779 |
|
|
int elt = INTVAL (operands[2]) % half_elts;
|
780 |
|
|
int hi = (INTVAL (operands[2]) / half_elts) * 2;
|
781 |
|
|
int regno = REGNO (operands[1]);
|
782 |
|
|
|
783 |
|
|
if (BYTES_BIG_ENDIAN)
|
784 |
|
|
elt = half_elts - 1 - elt;
|
785 |
|
|
|
786 |
|
|
operands[1] = gen_rtx_REG (mode, regno + hi);
|
787 |
|
|
operands[2] = GEN_INT (elt);
|
788 |
|
|
|
789 |
|
|
return "vmov%?.\t%0, %P1[%c2]";
|
790 |
|
|
}
|
791 |
|
|
[(set_attr "predicable" "yes")
|
792 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
793 |
|
|
)
|
794 |
|
|
|
795 |
|
|
(define_insn "vec_extractv2di"
|
796 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=r")
|
797 |
|
|
(vec_select:DI
|
798 |
|
|
(match_operand:V2DI 1 "s_register_operand" "w")
|
799 |
|
|
(parallel [(match_operand:SI 2 "immediate_operand" "i")])))]
|
800 |
|
|
"TARGET_NEON"
|
801 |
|
|
{
|
802 |
|
|
int regno = REGNO (operands[1]) + INTVAL (operands[2]);
|
803 |
|
|
|
804 |
|
|
operands[1] = gen_rtx_REG (DImode, regno);
|
805 |
|
|
|
806 |
|
|
return "vmov%?.64\t%Q0, %R0, %P1";
|
807 |
|
|
}
|
808 |
|
|
[(set_attr "predicable" "yes")
|
809 |
|
|
(set_attr "neon_type" "neon_int_1")]
|
810 |
|
|
)
|
811 |
|
|
|
812 |
|
|
(define_expand "vec_init"
|
813 |
|
|
[(match_operand:VDQ 0 "s_register_operand" "")
|
814 |
|
|
(match_operand 1 "" "")]
|
815 |
|
|
"TARGET_NEON"
|
816 |
|
|
{
|
817 |
|
|
neon_expand_vector_init (operands[0], operands[1]);
|
818 |
|
|
DONE;
|
819 |
|
|
})
|
820 |
|
|
|
821 |
|
|
;; Doubleword and quadword arithmetic.
|
822 |
|
|
|
823 |
|
|
;; NOTE: vadd/vsub and some other instructions also support 64-bit integer
|
824 |
|
|
;; element size, which we could potentially use for "long long" operations. We
|
825 |
|
|
;; don't want to do this at present though, because moving values from the
|
826 |
|
|
;; vector unit to the ARM core is currently slow and 64-bit addition (etc.) is
|
827 |
|
|
;; easy to do with ARM instructions anyway.
|
828 |
|
|
|
829 |
|
|
(define_insn "*add3_neon"
|
830 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
831 |
|
|
(plus:VDQ (match_operand:VDQ 1 "s_register_operand" "w")
|
832 |
|
|
(match_operand:VDQ 2 "s_register_operand" "w")))]
|
833 |
|
|
"TARGET_NEON"
|
834 |
|
|
"vadd.\t%0, %1, %2"
|
835 |
|
|
[(set (attr "neon_type")
|
836 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
837 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
838 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
839 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
840 |
|
|
(const_string "neon_int_1")))]
|
841 |
|
|
)
|
842 |
|
|
|
843 |
|
|
(define_insn "*sub3_neon"
|
844 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
845 |
|
|
(minus:VDQ (match_operand:VDQ 1 "s_register_operand" "w")
|
846 |
|
|
(match_operand:VDQ 2 "s_register_operand" "w")))]
|
847 |
|
|
"TARGET_NEON"
|
848 |
|
|
"vsub.\t%0, %1, %2"
|
849 |
|
|
[(set (attr "neon_type")
|
850 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
851 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
852 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
853 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
854 |
|
|
(const_string "neon_int_2")))]
|
855 |
|
|
)
|
856 |
|
|
|
857 |
|
|
(define_insn "*mul3_neon"
|
858 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
859 |
|
|
(mult:VDQ (match_operand:VDQ 1 "s_register_operand" "w")
|
860 |
|
|
(match_operand:VDQ 2 "s_register_operand" "w")))]
|
861 |
|
|
"TARGET_NEON"
|
862 |
|
|
"vmul.\t%0, %1, %2"
|
863 |
|
|
[(set (attr "neon_type")
|
864 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
865 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
866 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
867 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
868 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
869 |
|
|
(if_then_else
|
870 |
|
|
(ne (symbol_ref "") (const_int 0))
|
871 |
|
|
(const_string "neon_mul_ddd_8_16_qdd_16_8_long_32_16_long")
|
872 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32"))
|
873 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
874 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32")
|
875 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32")))))]
|
876 |
|
|
)
|
877 |
|
|
|
878 |
|
|
(define_insn "*mul3add_neon"
|
879 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
880 |
|
|
(plus:VDQ (mult:VDQ (match_operand:VDQ 2 "s_register_operand" "w")
|
881 |
|
|
(match_operand:VDQ 3 "s_register_operand" "w"))
|
882 |
|
|
(match_operand:VDQ 1 "s_register_operand" "0")))]
|
883 |
|
|
"TARGET_NEON"
|
884 |
|
|
"vmla.\t%0, %2, %3"
|
885 |
|
|
[(set (attr "neon_type")
|
886 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
887 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
888 |
|
|
(const_string "neon_fp_vmla_ddd")
|
889 |
|
|
(const_string "neon_fp_vmla_qqq"))
|
890 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
891 |
|
|
(if_then_else
|
892 |
|
|
(ne (symbol_ref "") (const_int 0))
|
893 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
894 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long"))
|
895 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
896 |
|
|
(const_string "neon_mla_qqq_8_16")
|
897 |
|
|
(const_string "neon_mla_qqq_32_qqd_32_scalar")))))]
|
898 |
|
|
)
|
899 |
|
|
|
900 |
|
|
(define_insn "*mul3negadd_neon"
|
901 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
902 |
|
|
(minus:VDQ (match_operand:VDQ 1 "s_register_operand" "0")
|
903 |
|
|
(mult:VDQ (match_operand:VDQ 2 "s_register_operand" "w")
|
904 |
|
|
(match_operand:VDQ 3 "s_register_operand" "w"))))]
|
905 |
|
|
"TARGET_NEON"
|
906 |
|
|
"vmls.\t%0, %2, %3"
|
907 |
|
|
[(set (attr "neon_type")
|
908 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
909 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
910 |
|
|
(const_string "neon_fp_vmla_ddd")
|
911 |
|
|
(const_string "neon_fp_vmla_qqq"))
|
912 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
913 |
|
|
(if_then_else
|
914 |
|
|
(ne (symbol_ref "") (const_int 0))
|
915 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
916 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long"))
|
917 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
918 |
|
|
(const_string "neon_mla_qqq_8_16")
|
919 |
|
|
(const_string "neon_mla_qqq_32_qqd_32_scalar")))))]
|
920 |
|
|
)
|
921 |
|
|
|
922 |
|
|
(define_insn "ior3"
|
923 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w,w")
|
924 |
|
|
(ior:VDQ (match_operand:VDQ 1 "s_register_operand" "w,0")
|
925 |
|
|
(match_operand:VDQ 2 "neon_logic_op2" "w,Dl")))]
|
926 |
|
|
"TARGET_NEON"
|
927 |
|
|
{
|
928 |
|
|
switch (which_alternative)
|
929 |
|
|
{
|
930 |
|
|
case 0: return "vorr\t%0, %1, %2";
|
931 |
|
|
case 1: return neon_output_logic_immediate ("vorr", &operands[2],
|
932 |
|
|
mode, 0, VALID_NEON_QREG_MODE (mode));
|
933 |
|
|
default: gcc_unreachable ();
|
934 |
|
|
}
|
935 |
|
|
}
|
936 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
937 |
|
|
)
|
938 |
|
|
|
939 |
|
|
(define_insn "iordi3_neon"
|
940 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w,w")
|
941 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "w,0")
|
942 |
|
|
(match_operand:DI 2 "neon_logic_op2" "w,Dl")]
|
943 |
|
|
UNSPEC_VORR))]
|
944 |
|
|
"TARGET_NEON"
|
945 |
|
|
{
|
946 |
|
|
switch (which_alternative)
|
947 |
|
|
{
|
948 |
|
|
case 0: return "vorr\t%P0, %P1, %P2";
|
949 |
|
|
case 1: return neon_output_logic_immediate ("vorr", &operands[2],
|
950 |
|
|
DImode, 0, VALID_NEON_QREG_MODE (DImode));
|
951 |
|
|
default: gcc_unreachable ();
|
952 |
|
|
}
|
953 |
|
|
}
|
954 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
955 |
|
|
)
|
956 |
|
|
|
957 |
|
|
;; The concrete forms of the Neon immediate-logic instructions are vbic and
|
958 |
|
|
;; vorr. We support the pseudo-instruction vand instead, because that
|
959 |
|
|
;; corresponds to the canonical form the middle-end expects to use for
|
960 |
|
|
;; immediate bitwise-ANDs.
|
961 |
|
|
|
962 |
|
|
(define_insn "and3"
|
963 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w,w")
|
964 |
|
|
(and:VDQ (match_operand:VDQ 1 "s_register_operand" "w,0")
|
965 |
|
|
(match_operand:VDQ 2 "neon_inv_logic_op2" "w,DL")))]
|
966 |
|
|
"TARGET_NEON"
|
967 |
|
|
{
|
968 |
|
|
switch (which_alternative)
|
969 |
|
|
{
|
970 |
|
|
case 0: return "vand\t%0, %1, %2";
|
971 |
|
|
case 1: return neon_output_logic_immediate ("vand", &operands[2],
|
972 |
|
|
mode, 1, VALID_NEON_QREG_MODE (mode));
|
973 |
|
|
default: gcc_unreachable ();
|
974 |
|
|
}
|
975 |
|
|
}
|
976 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
977 |
|
|
)
|
978 |
|
|
|
979 |
|
|
(define_insn "anddi3_neon"
|
980 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w,w")
|
981 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "w,0")
|
982 |
|
|
(match_operand:DI 2 "neon_inv_logic_op2" "w,DL")]
|
983 |
|
|
UNSPEC_VAND))]
|
984 |
|
|
"TARGET_NEON"
|
985 |
|
|
{
|
986 |
|
|
switch (which_alternative)
|
987 |
|
|
{
|
988 |
|
|
case 0: return "vand\t%P0, %P1, %P2";
|
989 |
|
|
case 1: return neon_output_logic_immediate ("vand", &operands[2],
|
990 |
|
|
DImode, 1, VALID_NEON_QREG_MODE (DImode));
|
991 |
|
|
default: gcc_unreachable ();
|
992 |
|
|
}
|
993 |
|
|
}
|
994 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
995 |
|
|
)
|
996 |
|
|
|
997 |
|
|
(define_insn "orn3_neon"
|
998 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
999 |
|
|
(ior:VDQ (match_operand:VDQ 1 "s_register_operand" "w")
|
1000 |
|
|
(not:VDQ (match_operand:VDQ 2 "s_register_operand" "w"))))]
|
1001 |
|
|
"TARGET_NEON"
|
1002 |
|
|
"vorn\t%0, %1, %2"
|
1003 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1004 |
|
|
)
|
1005 |
|
|
|
1006 |
|
|
(define_insn "orndi3_neon"
|
1007 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w")
|
1008 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "w")
|
1009 |
|
|
(match_operand:DI 2 "s_register_operand" "w")]
|
1010 |
|
|
UNSPEC_VORN))]
|
1011 |
|
|
"TARGET_NEON"
|
1012 |
|
|
"vorn\t%P0, %P1, %P2"
|
1013 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1014 |
|
|
)
|
1015 |
|
|
|
1016 |
|
|
(define_insn "bic3_neon"
|
1017 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
1018 |
|
|
(and:VDQ (match_operand:VDQ 1 "s_register_operand" "w")
|
1019 |
|
|
(not:VDQ (match_operand:VDQ 2 "s_register_operand" "w"))))]
|
1020 |
|
|
"TARGET_NEON"
|
1021 |
|
|
"vbic\t%0, %1, %2"
|
1022 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1023 |
|
|
)
|
1024 |
|
|
|
1025 |
|
|
(define_insn "bicdi3_neon"
|
1026 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w")
|
1027 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "w")
|
1028 |
|
|
(match_operand:DI 2 "s_register_operand" "w")]
|
1029 |
|
|
UNSPEC_VBIC))]
|
1030 |
|
|
"TARGET_NEON"
|
1031 |
|
|
"vbic\t%P0, %P1, %P2"
|
1032 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1033 |
|
|
)
|
1034 |
|
|
|
1035 |
|
|
(define_insn "xor3"
|
1036 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
1037 |
|
|
(xor:VDQ (match_operand:VDQ 1 "s_register_operand" "w")
|
1038 |
|
|
(match_operand:VDQ 2 "s_register_operand" "w")))]
|
1039 |
|
|
"TARGET_NEON"
|
1040 |
|
|
"veor\t%0, %1, %2"
|
1041 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1042 |
|
|
)
|
1043 |
|
|
|
1044 |
|
|
(define_insn "xordi3_neon"
|
1045 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w")
|
1046 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "w")
|
1047 |
|
|
(match_operand:DI 2 "s_register_operand" "w")]
|
1048 |
|
|
UNSPEC_VEOR))]
|
1049 |
|
|
"TARGET_NEON"
|
1050 |
|
|
"veor\t%P0, %P1, %P2"
|
1051 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1052 |
|
|
)
|
1053 |
|
|
|
1054 |
|
|
(define_insn "one_cmpl2"
|
1055 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
1056 |
|
|
(not:VDQ (match_operand:VDQ 1 "s_register_operand" "w")))]
|
1057 |
|
|
"TARGET_NEON"
|
1058 |
|
|
"vmvn\t%0, %1"
|
1059 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1060 |
|
|
)
|
1061 |
|
|
|
1062 |
|
|
(define_insn "abs2"
|
1063 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
1064 |
|
|
(abs:VDQW (match_operand:VDQW 1 "s_register_operand" "w")))]
|
1065 |
|
|
"TARGET_NEON"
|
1066 |
|
|
"vabs.\t%0, %1"
|
1067 |
|
|
[(set (attr "neon_type")
|
1068 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1069 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1070 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1071 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
1072 |
|
|
(const_string "neon_int_3")))]
|
1073 |
|
|
)
|
1074 |
|
|
|
1075 |
|
|
(define_insn "neg2"
|
1076 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
1077 |
|
|
(neg:VDQW (match_operand:VDQW 1 "s_register_operand" "w")))]
|
1078 |
|
|
"TARGET_NEON"
|
1079 |
|
|
"vneg.\t%0, %1"
|
1080 |
|
|
[(set (attr "neon_type")
|
1081 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1082 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1083 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1084 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
1085 |
|
|
(const_string "neon_int_3")))]
|
1086 |
|
|
)
|
1087 |
|
|
|
1088 |
|
|
(define_insn "*umin3_neon"
|
1089 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
1090 |
|
|
(umin:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w")
|
1091 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")))]
|
1092 |
|
|
"TARGET_NEON"
|
1093 |
|
|
"vmin.\t%0, %1, %2"
|
1094 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
1095 |
|
|
)
|
1096 |
|
|
|
1097 |
|
|
(define_insn "*umax3_neon"
|
1098 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
1099 |
|
|
(umax:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w")
|
1100 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")))]
|
1101 |
|
|
"TARGET_NEON"
|
1102 |
|
|
"vmax.\t%0, %1, %2"
|
1103 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
1104 |
|
|
)
|
1105 |
|
|
|
1106 |
|
|
(define_insn "*smin3_neon"
|
1107 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
1108 |
|
|
(smin:VDQW (match_operand:VDQW 1 "s_register_operand" "w")
|
1109 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")))]
|
1110 |
|
|
"TARGET_NEON"
|
1111 |
|
|
"vmin.\t%0, %1, %2"
|
1112 |
|
|
[(set (attr "neon_type")
|
1113 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1114 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1115 |
|
|
(const_string "neon_int_5")))]
|
1116 |
|
|
)
|
1117 |
|
|
|
1118 |
|
|
(define_insn "*smax3_neon"
|
1119 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
1120 |
|
|
(smax:VDQW (match_operand:VDQW 1 "s_register_operand" "w")
|
1121 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")))]
|
1122 |
|
|
"TARGET_NEON"
|
1123 |
|
|
"vmax.\t%0, %1, %2"
|
1124 |
|
|
[(set (attr "neon_type")
|
1125 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1126 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1127 |
|
|
(const_string "neon_int_5")))]
|
1128 |
|
|
)
|
1129 |
|
|
|
1130 |
|
|
; TODO: V2DI shifts are current disabled because there are bugs in the
|
1131 |
|
|
; generic vectorizer code. It ends up creating a V2DI constructor with
|
1132 |
|
|
; SImode elements.
|
1133 |
|
|
|
1134 |
|
|
(define_insn "vashl3"
|
1135 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
1136 |
|
|
(ashift:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w")
|
1137 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")))]
|
1138 |
|
|
"TARGET_NEON"
|
1139 |
|
|
"vshl.\t%0, %1, %2"
|
1140 |
|
|
[(set (attr "neon_type")
|
1141 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1142 |
|
|
(const_string "neon_vshl_ddd")
|
1143 |
|
|
(const_string "neon_shift_3")))]
|
1144 |
|
|
)
|
1145 |
|
|
|
1146 |
|
|
; Used for implementing logical shift-right, which is a left-shift by a negative
|
1147 |
|
|
; amount, with signed operands. This is essentially the same as ashl3
|
1148 |
|
|
; above, but using an unspec in case GCC tries anything tricky with negative
|
1149 |
|
|
; shift amounts.
|
1150 |
|
|
|
1151 |
|
|
(define_insn "ashl3_signed"
|
1152 |
|
|
[(set (match_operand:VDQI 0 "s_register_operand" "=w")
|
1153 |
|
|
(unspec:VDQI [(match_operand:VDQI 1 "s_register_operand" "w")
|
1154 |
|
|
(match_operand:VDQI 2 "s_register_operand" "w")]
|
1155 |
|
|
UNSPEC_ASHIFT_SIGNED))]
|
1156 |
|
|
"TARGET_NEON"
|
1157 |
|
|
"vshl.\t%0, %1, %2"
|
1158 |
|
|
[(set (attr "neon_type")
|
1159 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1160 |
|
|
(const_string "neon_vshl_ddd")
|
1161 |
|
|
(const_string "neon_shift_3")))]
|
1162 |
|
|
)
|
1163 |
|
|
|
1164 |
|
|
; Used for implementing logical shift-right, which is a left-shift by a negative
|
1165 |
|
|
; amount, with unsigned operands.
|
1166 |
|
|
|
1167 |
|
|
(define_insn "ashl3_unsigned"
|
1168 |
|
|
[(set (match_operand:VDQI 0 "s_register_operand" "=w")
|
1169 |
|
|
(unspec:VDQI [(match_operand:VDQI 1 "s_register_operand" "w")
|
1170 |
|
|
(match_operand:VDQI 2 "s_register_operand" "w")]
|
1171 |
|
|
UNSPEC_ASHIFT_UNSIGNED))]
|
1172 |
|
|
"TARGET_NEON"
|
1173 |
|
|
"vshl.\t%0, %1, %2"
|
1174 |
|
|
[(set (attr "neon_type")
|
1175 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1176 |
|
|
(const_string "neon_vshl_ddd")
|
1177 |
|
|
(const_string "neon_shift_3")))]
|
1178 |
|
|
)
|
1179 |
|
|
|
1180 |
|
|
(define_expand "vashr3"
|
1181 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "")
|
1182 |
|
|
(ashiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "")
|
1183 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "")))]
|
1184 |
|
|
"TARGET_NEON"
|
1185 |
|
|
{
|
1186 |
|
|
rtx neg = gen_reg_rtx (mode);
|
1187 |
|
|
|
1188 |
|
|
emit_insn (gen_neg2 (neg, operands[2]));
|
1189 |
|
|
emit_insn (gen_ashl3_signed (operands[0], operands[1], neg));
|
1190 |
|
|
|
1191 |
|
|
DONE;
|
1192 |
|
|
})
|
1193 |
|
|
|
1194 |
|
|
(define_expand "vlshr3"
|
1195 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "")
|
1196 |
|
|
(lshiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "")
|
1197 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "")))]
|
1198 |
|
|
"TARGET_NEON"
|
1199 |
|
|
{
|
1200 |
|
|
rtx neg = gen_reg_rtx (mode);
|
1201 |
|
|
|
1202 |
|
|
emit_insn (gen_neg2 (neg, operands[2]));
|
1203 |
|
|
emit_insn (gen_ashl3_unsigned (operands[0], operands[1], neg));
|
1204 |
|
|
|
1205 |
|
|
DONE;
|
1206 |
|
|
})
|
1207 |
|
|
|
1208 |
|
|
;; Widening operations
|
1209 |
|
|
|
1210 |
|
|
(define_insn "widen_ssum3"
|
1211 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1212 |
|
|
(plus: (sign_extend:
|
1213 |
|
|
(match_operand:VW 1 "s_register_operand" "%w"))
|
1214 |
|
|
(match_operand: 2 "s_register_operand" "w")))]
|
1215 |
|
|
"TARGET_NEON"
|
1216 |
|
|
"vaddw.\t%q0, %q2, %P1"
|
1217 |
|
|
[(set_attr "neon_type" "neon_int_3")]
|
1218 |
|
|
)
|
1219 |
|
|
|
1220 |
|
|
(define_insn "widen_usum3"
|
1221 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1222 |
|
|
(plus: (zero_extend:
|
1223 |
|
|
(match_operand:VW 1 "s_register_operand" "%w"))
|
1224 |
|
|
(match_operand: 2 "s_register_operand" "w")))]
|
1225 |
|
|
"TARGET_NEON"
|
1226 |
|
|
"vaddw.\t%q0, %q2, %P1"
|
1227 |
|
|
[(set_attr "neon_type" "neon_int_3")]
|
1228 |
|
|
)
|
1229 |
|
|
|
1230 |
|
|
;; VEXT can be used to synthesize coarse whole-vector shifts with 8-bit
|
1231 |
|
|
;; shift-count granularity. That's good enough for the middle-end's current
|
1232 |
|
|
;; needs.
|
1233 |
|
|
|
1234 |
|
|
(define_expand "vec_shr_"
|
1235 |
|
|
[(match_operand:VDQ 0 "s_register_operand" "")
|
1236 |
|
|
(match_operand:VDQ 1 "s_register_operand" "")
|
1237 |
|
|
(match_operand:SI 2 "const_multiple_of_8_operand" "")]
|
1238 |
|
|
"TARGET_NEON"
|
1239 |
|
|
{
|
1240 |
|
|
rtx zero_reg;
|
1241 |
|
|
HOST_WIDE_INT num_bits = INTVAL (operands[2]);
|
1242 |
|
|
const int width = GET_MODE_BITSIZE (mode);
|
1243 |
|
|
const enum machine_mode bvecmode = (width == 128) ? V16QImode : V8QImode;
|
1244 |
|
|
rtx (*gen_ext) (rtx, rtx, rtx, rtx) =
|
1245 |
|
|
(width == 128) ? gen_neon_vextv16qi : gen_neon_vextv8qi;
|
1246 |
|
|
|
1247 |
|
|
if (num_bits == width)
|
1248 |
|
|
{
|
1249 |
|
|
emit_move_insn (operands[0], operands[1]);
|
1250 |
|
|
DONE;
|
1251 |
|
|
}
|
1252 |
|
|
|
1253 |
|
|
zero_reg = force_reg (bvecmode, CONST0_RTX (bvecmode));
|
1254 |
|
|
operands[0] = gen_lowpart (bvecmode, operands[0]);
|
1255 |
|
|
operands[1] = gen_lowpart (bvecmode, operands[1]);
|
1256 |
|
|
|
1257 |
|
|
emit_insn (gen_ext (operands[0], operands[1], zero_reg,
|
1258 |
|
|
GEN_INT (num_bits / BITS_PER_UNIT)));
|
1259 |
|
|
DONE;
|
1260 |
|
|
})
|
1261 |
|
|
|
1262 |
|
|
(define_expand "vec_shl_"
|
1263 |
|
|
[(match_operand:VDQ 0 "s_register_operand" "")
|
1264 |
|
|
(match_operand:VDQ 1 "s_register_operand" "")
|
1265 |
|
|
(match_operand:SI 2 "const_multiple_of_8_operand" "")]
|
1266 |
|
|
"TARGET_NEON"
|
1267 |
|
|
{
|
1268 |
|
|
rtx zero_reg;
|
1269 |
|
|
HOST_WIDE_INT num_bits = INTVAL (operands[2]);
|
1270 |
|
|
const int width = GET_MODE_BITSIZE (mode);
|
1271 |
|
|
const enum machine_mode bvecmode = (width == 128) ? V16QImode : V8QImode;
|
1272 |
|
|
rtx (*gen_ext) (rtx, rtx, rtx, rtx) =
|
1273 |
|
|
(width == 128) ? gen_neon_vextv16qi : gen_neon_vextv8qi;
|
1274 |
|
|
|
1275 |
|
|
if (num_bits == 0)
|
1276 |
|
|
{
|
1277 |
|
|
emit_move_insn (operands[0], CONST0_RTX (mode));
|
1278 |
|
|
DONE;
|
1279 |
|
|
}
|
1280 |
|
|
|
1281 |
|
|
num_bits = width - num_bits;
|
1282 |
|
|
|
1283 |
|
|
zero_reg = force_reg (bvecmode, CONST0_RTX (bvecmode));
|
1284 |
|
|
operands[0] = gen_lowpart (bvecmode, operands[0]);
|
1285 |
|
|
operands[1] = gen_lowpart (bvecmode, operands[1]);
|
1286 |
|
|
|
1287 |
|
|
emit_insn (gen_ext (operands[0], zero_reg, operands[1],
|
1288 |
|
|
GEN_INT (num_bits / BITS_PER_UNIT)));
|
1289 |
|
|
DONE;
|
1290 |
|
|
})
|
1291 |
|
|
|
1292 |
|
|
;; Helpers for quad-word reduction operations
|
1293 |
|
|
|
1294 |
|
|
; Add (or smin, smax...) the low N/2 elements of the N-element vector
|
1295 |
|
|
; operand[1] to the high N/2 elements of same. Put the result in operand[0], an
|
1296 |
|
|
; N/2-element vector.
|
1297 |
|
|
|
1298 |
|
|
(define_insn "quad_halves_v4si"
|
1299 |
|
|
[(set (match_operand:V2SI 0 "s_register_operand" "=w")
|
1300 |
|
|
(vqh_ops:V2SI
|
1301 |
|
|
(vec_select:V2SI (match_operand:V4SI 1 "s_register_operand" "w")
|
1302 |
|
|
(parallel [(const_int 0) (const_int 1)]))
|
1303 |
|
|
(vec_select:V2SI (match_dup 1)
|
1304 |
|
|
(parallel [(const_int 2) (const_int 3)]))))]
|
1305 |
|
|
"TARGET_NEON"
|
1306 |
|
|
".32\t%P0, %e1, %f1"
|
1307 |
|
|
[(set_attr "vqh_mnem" "")
|
1308 |
|
|
(set (attr "neon_type")
|
1309 |
|
|
(if_then_else (eq_attr "vqh_mnem" "vadd")
|
1310 |
|
|
(const_string "neon_int_1") (const_string "neon_int_5")))]
|
1311 |
|
|
)
|
1312 |
|
|
|
1313 |
|
|
(define_insn "quad_halves_v4sf"
|
1314 |
|
|
[(set (match_operand:V2SF 0 "s_register_operand" "=w")
|
1315 |
|
|
(vqhs_ops:V2SF
|
1316 |
|
|
(vec_select:V2SF (match_operand:V4SF 1 "s_register_operand" "w")
|
1317 |
|
|
(parallel [(const_int 0) (const_int 1)]))
|
1318 |
|
|
(vec_select:V2SF (match_dup 1)
|
1319 |
|
|
(parallel [(const_int 2) (const_int 3)]))))]
|
1320 |
|
|
"TARGET_NEON"
|
1321 |
|
|
".f32\t%P0, %e1, %f1"
|
1322 |
|
|
[(set_attr "vqh_mnem" "")
|
1323 |
|
|
(set (attr "neon_type")
|
1324 |
|
|
(if_then_else (eq_attr "vqh_mnem" "vadd")
|
1325 |
|
|
(const_string "neon_int_1") (const_string "neon_int_5")))]
|
1326 |
|
|
)
|
1327 |
|
|
|
1328 |
|
|
(define_insn "quad_halves_v8hi"
|
1329 |
|
|
[(set (match_operand:V4HI 0 "s_register_operand" "+w")
|
1330 |
|
|
(vqh_ops:V4HI
|
1331 |
|
|
(vec_select:V4HI (match_operand:V8HI 1 "s_register_operand" "w")
|
1332 |
|
|
(parallel [(const_int 0) (const_int 1)
|
1333 |
|
|
(const_int 2) (const_int 3)]))
|
1334 |
|
|
(vec_select:V4HI (match_dup 1)
|
1335 |
|
|
(parallel [(const_int 4) (const_int 5)
|
1336 |
|
|
(const_int 6) (const_int 7)]))))]
|
1337 |
|
|
"TARGET_NEON"
|
1338 |
|
|
".16\t%P0, %e1, %f1"
|
1339 |
|
|
[(set_attr "vqh_mnem" "")
|
1340 |
|
|
(set (attr "neon_type")
|
1341 |
|
|
(if_then_else (eq_attr "vqh_mnem" "vadd")
|
1342 |
|
|
(const_string "neon_int_1") (const_string "neon_int_5")))]
|
1343 |
|
|
)
|
1344 |
|
|
|
1345 |
|
|
(define_insn "quad_halves_v16qi"
|
1346 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "+w")
|
1347 |
|
|
(vqh_ops:V8QI
|
1348 |
|
|
(vec_select:V8QI (match_operand:V16QI 1 "s_register_operand" "w")
|
1349 |
|
|
(parallel [(const_int 0) (const_int 1)
|
1350 |
|
|
(const_int 2) (const_int 3)
|
1351 |
|
|
(const_int 4) (const_int 5)
|
1352 |
|
|
(const_int 6) (const_int 7)]))
|
1353 |
|
|
(vec_select:V8QI (match_dup 1)
|
1354 |
|
|
(parallel [(const_int 8) (const_int 9)
|
1355 |
|
|
(const_int 10) (const_int 11)
|
1356 |
|
|
(const_int 12) (const_int 13)
|
1357 |
|
|
(const_int 14) (const_int 15)]))))]
|
1358 |
|
|
"TARGET_NEON"
|
1359 |
|
|
".8\t%P0, %e1, %f1"
|
1360 |
|
|
[(set_attr "vqh_mnem" "")
|
1361 |
|
|
(set (attr "neon_type")
|
1362 |
|
|
(if_then_else (eq_attr "vqh_mnem" "vadd")
|
1363 |
|
|
(const_string "neon_int_1") (const_string "neon_int_5")))]
|
1364 |
|
|
)
|
1365 |
|
|
|
1366 |
|
|
; FIXME: We wouldn't need the following insns if we could write subregs of
|
1367 |
|
|
; vector registers. Make an attempt at removing unnecessary moves, though
|
1368 |
|
|
; we're really at the mercy of the register allocator.
|
1369 |
|
|
|
1370 |
|
|
(define_insn "move_lo_quad_v4si"
|
1371 |
|
|
[(set (match_operand:V4SI 0 "s_register_operand" "+w")
|
1372 |
|
|
(vec_concat:V4SI
|
1373 |
|
|
(match_operand:V2SI 1 "s_register_operand" "w")
|
1374 |
|
|
(vec_select:V2SI (match_dup 0)
|
1375 |
|
|
(parallel [(const_int 2) (const_int 3)]))))]
|
1376 |
|
|
"TARGET_NEON"
|
1377 |
|
|
{
|
1378 |
|
|
int dest = REGNO (operands[0]);
|
1379 |
|
|
int src = REGNO (operands[1]);
|
1380 |
|
|
|
1381 |
|
|
if (dest != src)
|
1382 |
|
|
return "vmov\t%e0, %P1";
|
1383 |
|
|
else
|
1384 |
|
|
return "";
|
1385 |
|
|
}
|
1386 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
1387 |
|
|
)
|
1388 |
|
|
|
1389 |
|
|
(define_insn "move_lo_quad_v4sf"
|
1390 |
|
|
[(set (match_operand:V4SF 0 "s_register_operand" "+w")
|
1391 |
|
|
(vec_concat:V4SF
|
1392 |
|
|
(match_operand:V2SF 1 "s_register_operand" "w")
|
1393 |
|
|
(vec_select:V2SF (match_dup 0)
|
1394 |
|
|
(parallel [(const_int 2) (const_int 3)]))))]
|
1395 |
|
|
"TARGET_NEON"
|
1396 |
|
|
{
|
1397 |
|
|
int dest = REGNO (operands[0]);
|
1398 |
|
|
int src = REGNO (operands[1]);
|
1399 |
|
|
|
1400 |
|
|
if (dest != src)
|
1401 |
|
|
return "vmov\t%e0, %P1";
|
1402 |
|
|
else
|
1403 |
|
|
return "";
|
1404 |
|
|
}
|
1405 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
1406 |
|
|
)
|
1407 |
|
|
|
1408 |
|
|
(define_insn "move_lo_quad_v8hi"
|
1409 |
|
|
[(set (match_operand:V8HI 0 "s_register_operand" "+w")
|
1410 |
|
|
(vec_concat:V8HI
|
1411 |
|
|
(match_operand:V4HI 1 "s_register_operand" "w")
|
1412 |
|
|
(vec_select:V4HI (match_dup 0)
|
1413 |
|
|
(parallel [(const_int 4) (const_int 5)
|
1414 |
|
|
(const_int 6) (const_int 7)]))))]
|
1415 |
|
|
"TARGET_NEON"
|
1416 |
|
|
{
|
1417 |
|
|
int dest = REGNO (operands[0]);
|
1418 |
|
|
int src = REGNO (operands[1]);
|
1419 |
|
|
|
1420 |
|
|
if (dest != src)
|
1421 |
|
|
return "vmov\t%e0, %P1";
|
1422 |
|
|
else
|
1423 |
|
|
return "";
|
1424 |
|
|
}
|
1425 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
1426 |
|
|
)
|
1427 |
|
|
|
1428 |
|
|
(define_insn "move_lo_quad_v16qi"
|
1429 |
|
|
[(set (match_operand:V16QI 0 "s_register_operand" "+w")
|
1430 |
|
|
(vec_concat:V16QI
|
1431 |
|
|
(match_operand:V8QI 1 "s_register_operand" "w")
|
1432 |
|
|
(vec_select:V8QI (match_dup 0)
|
1433 |
|
|
(parallel [(const_int 8) (const_int 9)
|
1434 |
|
|
(const_int 10) (const_int 11)
|
1435 |
|
|
(const_int 12) (const_int 13)
|
1436 |
|
|
(const_int 14) (const_int 15)]))))]
|
1437 |
|
|
"TARGET_NEON"
|
1438 |
|
|
{
|
1439 |
|
|
int dest = REGNO (operands[0]);
|
1440 |
|
|
int src = REGNO (operands[1]);
|
1441 |
|
|
|
1442 |
|
|
if (dest != src)
|
1443 |
|
|
return "vmov\t%e0, %P1";
|
1444 |
|
|
else
|
1445 |
|
|
return "";
|
1446 |
|
|
}
|
1447 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
1448 |
|
|
)
|
1449 |
|
|
|
1450 |
|
|
;; Reduction operations
|
1451 |
|
|
|
1452 |
|
|
(define_expand "reduc_splus_"
|
1453 |
|
|
[(match_operand:VD 0 "s_register_operand" "")
|
1454 |
|
|
(match_operand:VD 1 "s_register_operand" "")]
|
1455 |
|
|
"TARGET_NEON"
|
1456 |
|
|
{
|
1457 |
|
|
neon_pairwise_reduce (operands[0], operands[1], mode,
|
1458 |
|
|
&gen_neon_vpadd_internal);
|
1459 |
|
|
DONE;
|
1460 |
|
|
})
|
1461 |
|
|
|
1462 |
|
|
(define_expand "reduc_splus_"
|
1463 |
|
|
[(match_operand:VQ 0 "s_register_operand" "")
|
1464 |
|
|
(match_operand:VQ 1 "s_register_operand" "")]
|
1465 |
|
|
"TARGET_NEON"
|
1466 |
|
|
{
|
1467 |
|
|
rtx step1 = gen_reg_rtx (mode);
|
1468 |
|
|
rtx res_d = gen_reg_rtx (mode);
|
1469 |
|
|
|
1470 |
|
|
emit_insn (gen_quad_halves_plus (step1, operands[1]));
|
1471 |
|
|
emit_insn (gen_reduc_splus_ (res_d, step1));
|
1472 |
|
|
emit_insn (gen_move_lo_quad_ (operands[0], res_d));
|
1473 |
|
|
|
1474 |
|
|
DONE;
|
1475 |
|
|
})
|
1476 |
|
|
|
1477 |
|
|
(define_insn "reduc_splus_v2di"
|
1478 |
|
|
[(set (match_operand:V2DI 0 "s_register_operand" "=w")
|
1479 |
|
|
(unspec:V2DI [(match_operand:V2DI 1 "s_register_operand" "w")]
|
1480 |
|
|
UNSPEC_VPADD))]
|
1481 |
|
|
"TARGET_NEON"
|
1482 |
|
|
"vadd.i64\t%e0, %e1, %f1"
|
1483 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
1484 |
|
|
)
|
1485 |
|
|
|
1486 |
|
|
;; NEON does not distinguish between signed and unsigned addition except on
|
1487 |
|
|
;; widening operations.
|
1488 |
|
|
(define_expand "reduc_uplus_"
|
1489 |
|
|
[(match_operand:VDQI 0 "s_register_operand" "")
|
1490 |
|
|
(match_operand:VDQI 1 "s_register_operand" "")]
|
1491 |
|
|
"TARGET_NEON"
|
1492 |
|
|
{
|
1493 |
|
|
emit_insn (gen_reduc_splus_ (operands[0], operands[1]));
|
1494 |
|
|
DONE;
|
1495 |
|
|
})
|
1496 |
|
|
|
1497 |
|
|
(define_expand "reduc_smin_"
|
1498 |
|
|
[(match_operand:VD 0 "s_register_operand" "")
|
1499 |
|
|
(match_operand:VD 1 "s_register_operand" "")]
|
1500 |
|
|
"TARGET_NEON"
|
1501 |
|
|
{
|
1502 |
|
|
neon_pairwise_reduce (operands[0], operands[1], mode,
|
1503 |
|
|
&gen_neon_vpsmin);
|
1504 |
|
|
DONE;
|
1505 |
|
|
})
|
1506 |
|
|
|
1507 |
|
|
(define_expand "reduc_smin_"
|
1508 |
|
|
[(match_operand:VQ 0 "s_register_operand" "")
|
1509 |
|
|
(match_operand:VQ 1 "s_register_operand" "")]
|
1510 |
|
|
"TARGET_NEON"
|
1511 |
|
|
{
|
1512 |
|
|
rtx step1 = gen_reg_rtx (mode);
|
1513 |
|
|
rtx res_d = gen_reg_rtx (mode);
|
1514 |
|
|
|
1515 |
|
|
emit_insn (gen_quad_halves_smin (step1, operands[1]));
|
1516 |
|
|
emit_insn (gen_reduc_smin_ (res_d, step1));
|
1517 |
|
|
emit_insn (gen_move_lo_quad_ (operands[0], res_d));
|
1518 |
|
|
|
1519 |
|
|
DONE;
|
1520 |
|
|
})
|
1521 |
|
|
|
1522 |
|
|
(define_expand "reduc_smax_"
|
1523 |
|
|
[(match_operand:VD 0 "s_register_operand" "")
|
1524 |
|
|
(match_operand:VD 1 "s_register_operand" "")]
|
1525 |
|
|
"TARGET_NEON"
|
1526 |
|
|
{
|
1527 |
|
|
neon_pairwise_reduce (operands[0], operands[1], mode,
|
1528 |
|
|
&gen_neon_vpsmax);
|
1529 |
|
|
DONE;
|
1530 |
|
|
})
|
1531 |
|
|
|
1532 |
|
|
(define_expand "reduc_smax_"
|
1533 |
|
|
[(match_operand:VQ 0 "s_register_operand" "")
|
1534 |
|
|
(match_operand:VQ 1 "s_register_operand" "")]
|
1535 |
|
|
"TARGET_NEON"
|
1536 |
|
|
{
|
1537 |
|
|
rtx step1 = gen_reg_rtx (mode);
|
1538 |
|
|
rtx res_d = gen_reg_rtx (mode);
|
1539 |
|
|
|
1540 |
|
|
emit_insn (gen_quad_halves_smax (step1, operands[1]));
|
1541 |
|
|
emit_insn (gen_reduc_smax_ (res_d, step1));
|
1542 |
|
|
emit_insn (gen_move_lo_quad_ (operands[0], res_d));
|
1543 |
|
|
|
1544 |
|
|
DONE;
|
1545 |
|
|
})
|
1546 |
|
|
|
1547 |
|
|
(define_expand "reduc_umin_"
|
1548 |
|
|
[(match_operand:VDI 0 "s_register_operand" "")
|
1549 |
|
|
(match_operand:VDI 1 "s_register_operand" "")]
|
1550 |
|
|
"TARGET_NEON"
|
1551 |
|
|
{
|
1552 |
|
|
neon_pairwise_reduce (operands[0], operands[1], mode,
|
1553 |
|
|
&gen_neon_vpumin);
|
1554 |
|
|
DONE;
|
1555 |
|
|
})
|
1556 |
|
|
|
1557 |
|
|
(define_expand "reduc_umin_"
|
1558 |
|
|
[(match_operand:VQI 0 "s_register_operand" "")
|
1559 |
|
|
(match_operand:VQI 1 "s_register_operand" "")]
|
1560 |
|
|
"TARGET_NEON"
|
1561 |
|
|
{
|
1562 |
|
|
rtx step1 = gen_reg_rtx (mode);
|
1563 |
|
|
rtx res_d = gen_reg_rtx (mode);
|
1564 |
|
|
|
1565 |
|
|
emit_insn (gen_quad_halves_umin (step1, operands[1]));
|
1566 |
|
|
emit_insn (gen_reduc_umin_ (res_d, step1));
|
1567 |
|
|
emit_insn (gen_move_lo_quad_ (operands[0], res_d));
|
1568 |
|
|
|
1569 |
|
|
DONE;
|
1570 |
|
|
})
|
1571 |
|
|
|
1572 |
|
|
(define_expand "reduc_umax_"
|
1573 |
|
|
[(match_operand:VDI 0 "s_register_operand" "")
|
1574 |
|
|
(match_operand:VDI 1 "s_register_operand" "")]
|
1575 |
|
|
"TARGET_NEON"
|
1576 |
|
|
{
|
1577 |
|
|
neon_pairwise_reduce (operands[0], operands[1], mode,
|
1578 |
|
|
&gen_neon_vpumax);
|
1579 |
|
|
DONE;
|
1580 |
|
|
})
|
1581 |
|
|
|
1582 |
|
|
(define_expand "reduc_umax_"
|
1583 |
|
|
[(match_operand:VQI 0 "s_register_operand" "")
|
1584 |
|
|
(match_operand:VQI 1 "s_register_operand" "")]
|
1585 |
|
|
"TARGET_NEON"
|
1586 |
|
|
{
|
1587 |
|
|
rtx step1 = gen_reg_rtx (mode);
|
1588 |
|
|
rtx res_d = gen_reg_rtx (mode);
|
1589 |
|
|
|
1590 |
|
|
emit_insn (gen_quad_halves_umax (step1, operands[1]));
|
1591 |
|
|
emit_insn (gen_reduc_umax_ (res_d, step1));
|
1592 |
|
|
emit_insn (gen_move_lo_quad_ (operands[0], res_d));
|
1593 |
|
|
|
1594 |
|
|
DONE;
|
1595 |
|
|
})
|
1596 |
|
|
|
1597 |
|
|
(define_insn "neon_vpadd_internal"
|
1598 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
1599 |
|
|
(unspec:VD [(match_operand:VD 1 "s_register_operand" "w")
|
1600 |
|
|
(match_operand:VD 2 "s_register_operand" "w")]
|
1601 |
|
|
UNSPEC_VPADD))]
|
1602 |
|
|
"TARGET_NEON"
|
1603 |
|
|
"vpadd.\t%P0, %P1, %P2"
|
1604 |
|
|
;; Assume this schedules like vadd.
|
1605 |
|
|
[(set (attr "neon_type")
|
1606 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1607 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1608 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1609 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
1610 |
|
|
(const_string "neon_int_1")))]
|
1611 |
|
|
)
|
1612 |
|
|
|
1613 |
|
|
(define_insn "neon_vpsmin"
|
1614 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
1615 |
|
|
(unspec:VD [(match_operand:VD 1 "s_register_operand" "w")
|
1616 |
|
|
(match_operand:VD 2 "s_register_operand" "w")]
|
1617 |
|
|
UNSPEC_VPSMIN))]
|
1618 |
|
|
"TARGET_NEON"
|
1619 |
|
|
"vpmin.\t%P0, %P1, %P2"
|
1620 |
|
|
;; Assume this schedules like vmin.
|
1621 |
|
|
[(set (attr "neon_type")
|
1622 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1623 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1624 |
|
|
(const_string "neon_int_5")))]
|
1625 |
|
|
)
|
1626 |
|
|
|
1627 |
|
|
(define_insn "neon_vpsmax"
|
1628 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
1629 |
|
|
(unspec:VD [(match_operand:VD 1 "s_register_operand" "w")
|
1630 |
|
|
(match_operand:VD 2 "s_register_operand" "w")]
|
1631 |
|
|
UNSPEC_VPSMAX))]
|
1632 |
|
|
"TARGET_NEON"
|
1633 |
|
|
"vpmax.\t%P0, %P1, %P2"
|
1634 |
|
|
;; Assume this schedules like vmax.
|
1635 |
|
|
[(set (attr "neon_type")
|
1636 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1637 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1638 |
|
|
(const_string "neon_int_5")))]
|
1639 |
|
|
)
|
1640 |
|
|
|
1641 |
|
|
(define_insn "neon_vpumin"
|
1642 |
|
|
[(set (match_operand:VDI 0 "s_register_operand" "=w")
|
1643 |
|
|
(unspec:VDI [(match_operand:VDI 1 "s_register_operand" "w")
|
1644 |
|
|
(match_operand:VDI 2 "s_register_operand" "w")]
|
1645 |
|
|
UNSPEC_VPUMIN))]
|
1646 |
|
|
"TARGET_NEON"
|
1647 |
|
|
"vpmin.\t%P0, %P1, %P2"
|
1648 |
|
|
;; Assume this schedules like umin.
|
1649 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
1650 |
|
|
)
|
1651 |
|
|
|
1652 |
|
|
(define_insn "neon_vpumax"
|
1653 |
|
|
[(set (match_operand:VDI 0 "s_register_operand" "=w")
|
1654 |
|
|
(unspec:VDI [(match_operand:VDI 1 "s_register_operand" "w")
|
1655 |
|
|
(match_operand:VDI 2 "s_register_operand" "w")]
|
1656 |
|
|
UNSPEC_VPUMAX))]
|
1657 |
|
|
"TARGET_NEON"
|
1658 |
|
|
"vpmax.\t%P0, %P1, %P2"
|
1659 |
|
|
;; Assume this schedules like umax.
|
1660 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
1661 |
|
|
)
|
1662 |
|
|
|
1663 |
|
|
;; Saturating arithmetic
|
1664 |
|
|
|
1665 |
|
|
; NOTE: Neon supports many more saturating variants of instructions than the
|
1666 |
|
|
; following, but these are all GCC currently understands.
|
1667 |
|
|
; FIXME: Actually, GCC doesn't know how to create saturating add/sub by itself
|
1668 |
|
|
; yet either, although these patterns may be used by intrinsics when they're
|
1669 |
|
|
; added.
|
1670 |
|
|
|
1671 |
|
|
(define_insn "*ss_add_neon"
|
1672 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
1673 |
|
|
(ss_plus:VD (match_operand:VD 1 "s_register_operand" "w")
|
1674 |
|
|
(match_operand:VD 2 "s_register_operand" "w")))]
|
1675 |
|
|
"TARGET_NEON"
|
1676 |
|
|
"vqadd.\t%P0, %P1, %P2"
|
1677 |
|
|
[(set_attr "neon_type" "neon_int_4")]
|
1678 |
|
|
)
|
1679 |
|
|
|
1680 |
|
|
(define_insn "*us_add_neon"
|
1681 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
1682 |
|
|
(us_plus:VD (match_operand:VD 1 "s_register_operand" "w")
|
1683 |
|
|
(match_operand:VD 2 "s_register_operand" "w")))]
|
1684 |
|
|
"TARGET_NEON"
|
1685 |
|
|
"vqadd.\t%P0, %P1, %P2"
|
1686 |
|
|
[(set_attr "neon_type" "neon_int_4")]
|
1687 |
|
|
)
|
1688 |
|
|
|
1689 |
|
|
(define_insn "*ss_sub_neon"
|
1690 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
1691 |
|
|
(ss_minus:VD (match_operand:VD 1 "s_register_operand" "w")
|
1692 |
|
|
(match_operand:VD 2 "s_register_operand" "w")))]
|
1693 |
|
|
"TARGET_NEON"
|
1694 |
|
|
"vqsub.\t%P0, %P1, %P2"
|
1695 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
1696 |
|
|
)
|
1697 |
|
|
|
1698 |
|
|
(define_insn "*us_sub_neon"
|
1699 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
1700 |
|
|
(us_minus:VD (match_operand:VD 1 "s_register_operand" "w")
|
1701 |
|
|
(match_operand:VD 2 "s_register_operand" "w")))]
|
1702 |
|
|
"TARGET_NEON"
|
1703 |
|
|
"vqsub.\t%P0, %P1, %P2"
|
1704 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
1705 |
|
|
)
|
1706 |
|
|
|
1707 |
|
|
;; Patterns for builtins.
|
1708 |
|
|
|
1709 |
|
|
; good for plain vadd, vaddq.
|
1710 |
|
|
|
1711 |
|
|
(define_insn "neon_vadd"
|
1712 |
|
|
[(set (match_operand:VDQX 0 "s_register_operand" "=w")
|
1713 |
|
|
(unspec:VDQX [(match_operand:VDQX 1 "s_register_operand" "w")
|
1714 |
|
|
(match_operand:VDQX 2 "s_register_operand" "w")
|
1715 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1716 |
|
|
UNSPEC_VADD))]
|
1717 |
|
|
"TARGET_NEON"
|
1718 |
|
|
"vadd.\t%0, %1, %2"
|
1719 |
|
|
[(set (attr "neon_type")
|
1720 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1721 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1722 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1723 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
1724 |
|
|
(const_string "neon_int_1")))]
|
1725 |
|
|
)
|
1726 |
|
|
|
1727 |
|
|
; operand 3 represents in bits:
|
1728 |
|
|
; bit 0: signed (vs unsigned).
|
1729 |
|
|
; bit 1: rounding (vs none).
|
1730 |
|
|
|
1731 |
|
|
(define_insn "neon_vaddl"
|
1732 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1733 |
|
|
(unspec: [(match_operand:VDI 1 "s_register_operand" "w")
|
1734 |
|
|
(match_operand:VDI 2 "s_register_operand" "w")
|
1735 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1736 |
|
|
UNSPEC_VADDL))]
|
1737 |
|
|
"TARGET_NEON"
|
1738 |
|
|
"vaddl.%T3%#\t%q0, %P1, %P2"
|
1739 |
|
|
[(set_attr "neon_type" "neon_int_3")]
|
1740 |
|
|
)
|
1741 |
|
|
|
1742 |
|
|
(define_insn "neon_vaddw"
|
1743 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1744 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "w")
|
1745 |
|
|
(match_operand:VDI 2 "s_register_operand" "w")
|
1746 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1747 |
|
|
UNSPEC_VADDW))]
|
1748 |
|
|
"TARGET_NEON"
|
1749 |
|
|
"vaddw.%T3%#\t%q0, %q1, %P2"
|
1750 |
|
|
[(set_attr "neon_type" "neon_int_2")]
|
1751 |
|
|
)
|
1752 |
|
|
|
1753 |
|
|
; vhadd and vrhadd.
|
1754 |
|
|
|
1755 |
|
|
(define_insn "neon_vhadd"
|
1756 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
1757 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
|
1758 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")
|
1759 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1760 |
|
|
UNSPEC_VHADD))]
|
1761 |
|
|
"TARGET_NEON"
|
1762 |
|
|
"v%O3hadd.%T3%#\t%0, %1, %2"
|
1763 |
|
|
[(set_attr "neon_type" "neon_int_4")]
|
1764 |
|
|
)
|
1765 |
|
|
|
1766 |
|
|
(define_insn "neon_vqadd"
|
1767 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
1768 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
1769 |
|
|
(match_operand:VDQIX 2 "s_register_operand" "w")
|
1770 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1771 |
|
|
UNSPEC_VQADD))]
|
1772 |
|
|
"TARGET_NEON"
|
1773 |
|
|
"vqadd.%T3%#\t%0, %1, %2"
|
1774 |
|
|
[(set_attr "neon_type" "neon_int_4")]
|
1775 |
|
|
)
|
1776 |
|
|
|
1777 |
|
|
(define_insn "neon_vaddhn"
|
1778 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1779 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
1780 |
|
|
(match_operand:VN 2 "s_register_operand" "w")
|
1781 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1782 |
|
|
UNSPEC_VADDHN))]
|
1783 |
|
|
"TARGET_NEON"
|
1784 |
|
|
"v%O3addhn.\t%P0, %q1, %q2"
|
1785 |
|
|
[(set_attr "neon_type" "neon_int_4")]
|
1786 |
|
|
)
|
1787 |
|
|
|
1788 |
|
|
(define_insn "neon_vmul"
|
1789 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
1790 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "w")
|
1791 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
1792 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1793 |
|
|
UNSPEC_VMUL))]
|
1794 |
|
|
"TARGET_NEON"
|
1795 |
|
|
"vmul.%F3%#\t%0, %1, %2"
|
1796 |
|
|
[(set (attr "neon_type")
|
1797 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1798 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1799 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1800 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
1801 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1802 |
|
|
(if_then_else
|
1803 |
|
|
(ne (symbol_ref "") (const_int 0))
|
1804 |
|
|
(const_string "neon_mul_ddd_8_16_qdd_16_8_long_32_16_long")
|
1805 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32"))
|
1806 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1807 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32")
|
1808 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32")))))]
|
1809 |
|
|
)
|
1810 |
|
|
|
1811 |
|
|
(define_insn "neon_vmla"
|
1812 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
1813 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "0")
|
1814 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
1815 |
|
|
(match_operand:VDQW 3 "s_register_operand" "w")
|
1816 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
1817 |
|
|
UNSPEC_VMLA))]
|
1818 |
|
|
"TARGET_NEON"
|
1819 |
|
|
"vmla.\t%0, %2, %3"
|
1820 |
|
|
[(set (attr "neon_type")
|
1821 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1822 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1823 |
|
|
(const_string "neon_fp_vmla_ddd")
|
1824 |
|
|
(const_string "neon_fp_vmla_qqq"))
|
1825 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1826 |
|
|
(if_then_else
|
1827 |
|
|
(ne (symbol_ref "") (const_int 0))
|
1828 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
1829 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long"))
|
1830 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1831 |
|
|
(const_string "neon_mla_qqq_8_16")
|
1832 |
|
|
(const_string "neon_mla_qqq_32_qqd_32_scalar")))))]
|
1833 |
|
|
)
|
1834 |
|
|
|
1835 |
|
|
(define_insn "neon_vmlal"
|
1836 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1837 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
1838 |
|
|
(match_operand:VW 2 "s_register_operand" "w")
|
1839 |
|
|
(match_operand:VW 3 "s_register_operand" "w")
|
1840 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
1841 |
|
|
UNSPEC_VMLAL))]
|
1842 |
|
|
"TARGET_NEON"
|
1843 |
|
|
"vmlal.%T4%#\t%q0, %P2, %P3"
|
1844 |
|
|
[(set (attr "neon_type")
|
1845 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1846 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
1847 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
1848 |
|
|
)
|
1849 |
|
|
|
1850 |
|
|
(define_insn "neon_vmls"
|
1851 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
1852 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "0")
|
1853 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
1854 |
|
|
(match_operand:VDQW 3 "s_register_operand" "w")
|
1855 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
1856 |
|
|
UNSPEC_VMLS))]
|
1857 |
|
|
"TARGET_NEON"
|
1858 |
|
|
"vmls.\t%0, %2, %3"
|
1859 |
|
|
[(set (attr "neon_type")
|
1860 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1861 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1862 |
|
|
(const_string "neon_fp_vmla_ddd")
|
1863 |
|
|
(const_string "neon_fp_vmla_qqq"))
|
1864 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1865 |
|
|
(if_then_else
|
1866 |
|
|
(ne (symbol_ref "") (const_int 0))
|
1867 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
1868 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long"))
|
1869 |
|
|
(if_then_else
|
1870 |
|
|
(ne (symbol_ref "") (const_int 0))
|
1871 |
|
|
(const_string "neon_mla_qqq_8_16")
|
1872 |
|
|
(const_string "neon_mla_qqq_32_qqd_32_scalar")))))]
|
1873 |
|
|
)
|
1874 |
|
|
|
1875 |
|
|
(define_insn "neon_vmlsl"
|
1876 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1877 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
1878 |
|
|
(match_operand:VW 2 "s_register_operand" "w")
|
1879 |
|
|
(match_operand:VW 3 "s_register_operand" "w")
|
1880 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
1881 |
|
|
UNSPEC_VMLSL))]
|
1882 |
|
|
"TARGET_NEON"
|
1883 |
|
|
"vmlsl.%T4%#\t%q0, %P2, %P3"
|
1884 |
|
|
[(set (attr "neon_type")
|
1885 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1886 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
1887 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
1888 |
|
|
)
|
1889 |
|
|
|
1890 |
|
|
(define_insn "neon_vqdmulh"
|
1891 |
|
|
[(set (match_operand:VMDQI 0 "s_register_operand" "=w")
|
1892 |
|
|
(unspec:VMDQI [(match_operand:VMDQI 1 "s_register_operand" "w")
|
1893 |
|
|
(match_operand:VMDQI 2 "s_register_operand" "w")
|
1894 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1895 |
|
|
UNSPEC_VQDMULH))]
|
1896 |
|
|
"TARGET_NEON"
|
1897 |
|
|
"vq%O3dmulh.\t%0, %1, %2"
|
1898 |
|
|
[(set (attr "neon_type")
|
1899 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1900 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1901 |
|
|
(const_string "neon_mul_ddd_8_16_qdd_16_8_long_32_16_long")
|
1902 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32"))
|
1903 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1904 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32")
|
1905 |
|
|
(const_string "neon_mul_qqq_8_16_32_ddd_32"))))]
|
1906 |
|
|
)
|
1907 |
|
|
|
1908 |
|
|
(define_insn "neon_vqdmlal"
|
1909 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1910 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
1911 |
|
|
(match_operand:VMDI 2 "s_register_operand" "w")
|
1912 |
|
|
(match_operand:VMDI 3 "s_register_operand" "w")
|
1913 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
1914 |
|
|
UNSPEC_VQDMLAL))]
|
1915 |
|
|
"TARGET_NEON"
|
1916 |
|
|
"vqdmlal.\t%q0, %P2, %P3"
|
1917 |
|
|
[(set (attr "neon_type")
|
1918 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1919 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
1920 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
1921 |
|
|
)
|
1922 |
|
|
|
1923 |
|
|
(define_insn "neon_vqdmlsl"
|
1924 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1925 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
1926 |
|
|
(match_operand:VMDI 2 "s_register_operand" "w")
|
1927 |
|
|
(match_operand:VMDI 3 "s_register_operand" "w")
|
1928 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
1929 |
|
|
UNSPEC_VQDMLSL))]
|
1930 |
|
|
"TARGET_NEON"
|
1931 |
|
|
"vqdmlsl.\t%q0, %P2, %P3"
|
1932 |
|
|
[(set (attr "neon_type")
|
1933 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1934 |
|
|
(const_string "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long")
|
1935 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
1936 |
|
|
)
|
1937 |
|
|
|
1938 |
|
|
(define_insn "neon_vmull"
|
1939 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1940 |
|
|
(unspec: [(match_operand:VW 1 "s_register_operand" "w")
|
1941 |
|
|
(match_operand:VW 2 "s_register_operand" "w")
|
1942 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1943 |
|
|
UNSPEC_VMULL))]
|
1944 |
|
|
"TARGET_NEON"
|
1945 |
|
|
"vmull.%T3%#\t%q0, %P1, %P2"
|
1946 |
|
|
[(set (attr "neon_type")
|
1947 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1948 |
|
|
(const_string "neon_mul_ddd_8_16_qdd_16_8_long_32_16_long")
|
1949 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar")))]
|
1950 |
|
|
)
|
1951 |
|
|
|
1952 |
|
|
(define_insn "neon_vqdmull"
|
1953 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1954 |
|
|
(unspec: [(match_operand:VMDI 1 "s_register_operand" "w")
|
1955 |
|
|
(match_operand:VMDI 2 "s_register_operand" "w")
|
1956 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1957 |
|
|
UNSPEC_VQDMULL))]
|
1958 |
|
|
"TARGET_NEON"
|
1959 |
|
|
"vqdmull.\t%q0, %P1, %P2"
|
1960 |
|
|
[(set (attr "neon_type")
|
1961 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1962 |
|
|
(const_string "neon_mul_ddd_8_16_qdd_16_8_long_32_16_long")
|
1963 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar")))]
|
1964 |
|
|
)
|
1965 |
|
|
|
1966 |
|
|
(define_insn "neon_vsub"
|
1967 |
|
|
[(set (match_operand:VDQX 0 "s_register_operand" "=w")
|
1968 |
|
|
(unspec:VDQX [(match_operand:VDQX 1 "s_register_operand" "w")
|
1969 |
|
|
(match_operand:VDQX 2 "s_register_operand" "w")
|
1970 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1971 |
|
|
UNSPEC_VSUB))]
|
1972 |
|
|
"TARGET_NEON"
|
1973 |
|
|
"vsub.\t%0, %1, %2"
|
1974 |
|
|
[(set (attr "neon_type")
|
1975 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1976 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
1977 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
1978 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
1979 |
|
|
(const_string "neon_int_2")))]
|
1980 |
|
|
)
|
1981 |
|
|
|
1982 |
|
|
(define_insn "neon_vsubl"
|
1983 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1984 |
|
|
(unspec: [(match_operand:VDI 1 "s_register_operand" "w")
|
1985 |
|
|
(match_operand:VDI 2 "s_register_operand" "w")
|
1986 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1987 |
|
|
UNSPEC_VSUBL))]
|
1988 |
|
|
"TARGET_NEON"
|
1989 |
|
|
"vsubl.%T3%#\t%q0, %P1, %P2"
|
1990 |
|
|
[(set_attr "neon_type" "neon_int_2")]
|
1991 |
|
|
)
|
1992 |
|
|
|
1993 |
|
|
(define_insn "neon_vsubw"
|
1994 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
1995 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "w")
|
1996 |
|
|
(match_operand:VDI 2 "s_register_operand" "w")
|
1997 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
1998 |
|
|
UNSPEC_VSUBW))]
|
1999 |
|
|
"TARGET_NEON"
|
2000 |
|
|
"vsubw.%T3%#\t%q0, %q1, %P2"
|
2001 |
|
|
[(set_attr "neon_type" "neon_int_2")]
|
2002 |
|
|
)
|
2003 |
|
|
|
2004 |
|
|
(define_insn "neon_vqsub"
|
2005 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
2006 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
2007 |
|
|
(match_operand:VDQIX 2 "s_register_operand" "w")
|
2008 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2009 |
|
|
UNSPEC_VQSUB))]
|
2010 |
|
|
"TARGET_NEON"
|
2011 |
|
|
"vqsub.%T3%#\t%0, %1, %2"
|
2012 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
2013 |
|
|
)
|
2014 |
|
|
|
2015 |
|
|
(define_insn "neon_vhsub"
|
2016 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
2017 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
|
2018 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")
|
2019 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2020 |
|
|
UNSPEC_VHSUB))]
|
2021 |
|
|
"TARGET_NEON"
|
2022 |
|
|
"vhsub.%T3%#\t%0, %1, %2"
|
2023 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
2024 |
|
|
)
|
2025 |
|
|
|
2026 |
|
|
(define_insn "neon_vsubhn"
|
2027 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2028 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
2029 |
|
|
(match_operand:VN 2 "s_register_operand" "w")
|
2030 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2031 |
|
|
UNSPEC_VSUBHN))]
|
2032 |
|
|
"TARGET_NEON"
|
2033 |
|
|
"v%O3subhn.\t%P0, %q1, %q2"
|
2034 |
|
|
[(set_attr "neon_type" "neon_int_4")]
|
2035 |
|
|
)
|
2036 |
|
|
|
2037 |
|
|
(define_insn "neon_vceq"
|
2038 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2039 |
|
|
(unspec: [(match_operand:VDQW 1 "s_register_operand" "w")
|
2040 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
2041 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2042 |
|
|
UNSPEC_VCEQ))]
|
2043 |
|
|
"TARGET_NEON"
|
2044 |
|
|
"vceq.\t%0, %1, %2"
|
2045 |
|
|
[(set (attr "neon_type")
|
2046 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2047 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2048 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2049 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
2050 |
|
|
(const_string "neon_int_5")))]
|
2051 |
|
|
)
|
2052 |
|
|
|
2053 |
|
|
(define_insn "neon_vcge"
|
2054 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2055 |
|
|
(unspec: [(match_operand:VDQW 1 "s_register_operand" "w")
|
2056 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
2057 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2058 |
|
|
UNSPEC_VCGE))]
|
2059 |
|
|
"TARGET_NEON"
|
2060 |
|
|
"vcge.%T3%#\t%0, %1, %2"
|
2061 |
|
|
[(set (attr "neon_type")
|
2062 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2063 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2064 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2065 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
2066 |
|
|
(const_string "neon_int_5")))]
|
2067 |
|
|
)
|
2068 |
|
|
|
2069 |
|
|
(define_insn "neon_vcgt"
|
2070 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2071 |
|
|
(unspec: [(match_operand:VDQW 1 "s_register_operand" "w")
|
2072 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
2073 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2074 |
|
|
UNSPEC_VCGT))]
|
2075 |
|
|
"TARGET_NEON"
|
2076 |
|
|
"vcgt.%T3%#\t%0, %1, %2"
|
2077 |
|
|
[(set (attr "neon_type")
|
2078 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2079 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2080 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2081 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
2082 |
|
|
(const_string "neon_int_5")))]
|
2083 |
|
|
)
|
2084 |
|
|
|
2085 |
|
|
(define_insn "neon_vcage"
|
2086 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2087 |
|
|
(unspec: [(match_operand:VCVTF 1 "s_register_operand" "w")
|
2088 |
|
|
(match_operand:VCVTF 2 "s_register_operand" "w")
|
2089 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2090 |
|
|
UNSPEC_VCAGE))]
|
2091 |
|
|
"TARGET_NEON"
|
2092 |
|
|
"vacge.\t%0, %1, %2"
|
2093 |
|
|
[(set (attr "neon_type")
|
2094 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2095 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2096 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2097 |
|
|
)
|
2098 |
|
|
|
2099 |
|
|
(define_insn "neon_vcagt"
|
2100 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2101 |
|
|
(unspec: [(match_operand:VCVTF 1 "s_register_operand" "w")
|
2102 |
|
|
(match_operand:VCVTF 2 "s_register_operand" "w")
|
2103 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2104 |
|
|
UNSPEC_VCAGT))]
|
2105 |
|
|
"TARGET_NEON"
|
2106 |
|
|
"vacgt.\t%0, %1, %2"
|
2107 |
|
|
[(set (attr "neon_type")
|
2108 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2109 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2110 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2111 |
|
|
)
|
2112 |
|
|
|
2113 |
|
|
(define_insn "neon_vtst"
|
2114 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
2115 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
|
2116 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")
|
2117 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2118 |
|
|
UNSPEC_VTST))]
|
2119 |
|
|
"TARGET_NEON"
|
2120 |
|
|
"vtst.\t%0, %1, %2"
|
2121 |
|
|
[(set_attr "neon_type" "neon_int_4")]
|
2122 |
|
|
)
|
2123 |
|
|
|
2124 |
|
|
(define_insn "neon_vabd"
|
2125 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
2126 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "w")
|
2127 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
2128 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2129 |
|
|
UNSPEC_VABD))]
|
2130 |
|
|
"TARGET_NEON"
|
2131 |
|
|
"vabd.%T3%#\t%0, %1, %2"
|
2132 |
|
|
[(set (attr "neon_type")
|
2133 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2134 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2135 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2136 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
2137 |
|
|
(const_string "neon_int_5")))]
|
2138 |
|
|
)
|
2139 |
|
|
|
2140 |
|
|
(define_insn "neon_vabdl"
|
2141 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2142 |
|
|
(unspec: [(match_operand:VW 1 "s_register_operand" "w")
|
2143 |
|
|
(match_operand:VW 2 "s_register_operand" "w")
|
2144 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2145 |
|
|
UNSPEC_VABDL))]
|
2146 |
|
|
"TARGET_NEON"
|
2147 |
|
|
"vabdl.%T3%#\t%q0, %P1, %P2"
|
2148 |
|
|
[(set_attr "neon_type" "neon_int_5")]
|
2149 |
|
|
)
|
2150 |
|
|
|
2151 |
|
|
(define_insn "neon_vaba"
|
2152 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
2153 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "0")
|
2154 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")
|
2155 |
|
|
(match_operand:VDQIW 3 "s_register_operand" "w")
|
2156 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
2157 |
|
|
UNSPEC_VABA))]
|
2158 |
|
|
"TARGET_NEON"
|
2159 |
|
|
"vaba.%T4%#\t%0, %2, %3"
|
2160 |
|
|
[(set (attr "neon_type")
|
2161 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2162 |
|
|
(const_string "neon_vaba") (const_string "neon_vaba_qqq")))]
|
2163 |
|
|
)
|
2164 |
|
|
|
2165 |
|
|
(define_insn "neon_vabal"
|
2166 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2167 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
2168 |
|
|
(match_operand:VW 2 "s_register_operand" "w")
|
2169 |
|
|
(match_operand:VW 3 "s_register_operand" "w")
|
2170 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
2171 |
|
|
UNSPEC_VABAL))]
|
2172 |
|
|
"TARGET_NEON"
|
2173 |
|
|
"vabal.%T4%#\t%q0, %P2, %P3"
|
2174 |
|
|
[(set_attr "neon_type" "neon_vaba")]
|
2175 |
|
|
)
|
2176 |
|
|
|
2177 |
|
|
(define_insn "neon_vmax"
|
2178 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
2179 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "w")
|
2180 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
2181 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2182 |
|
|
UNSPEC_VMAX))]
|
2183 |
|
|
"TARGET_NEON"
|
2184 |
|
|
"vmax.%T3%#\t%0, %1, %2"
|
2185 |
|
|
[(set (attr "neon_type")
|
2186 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2187 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2188 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2189 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
2190 |
|
|
(const_string "neon_int_5")))]
|
2191 |
|
|
)
|
2192 |
|
|
|
2193 |
|
|
(define_insn "neon_vmin"
|
2194 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
2195 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "w")
|
2196 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")
|
2197 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2198 |
|
|
UNSPEC_VMIN))]
|
2199 |
|
|
"TARGET_NEON"
|
2200 |
|
|
"vmin.%T3%#\t%0, %1, %2"
|
2201 |
|
|
[(set (attr "neon_type")
|
2202 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2203 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2204 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2205 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
2206 |
|
|
(const_string "neon_int_5")))]
|
2207 |
|
|
)
|
2208 |
|
|
|
2209 |
|
|
(define_expand "neon_vpadd"
|
2210 |
|
|
[(match_operand:VD 0 "s_register_operand" "=w")
|
2211 |
|
|
(match_operand:VD 1 "s_register_operand" "w")
|
2212 |
|
|
(match_operand:VD 2 "s_register_operand" "w")
|
2213 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2214 |
|
|
"TARGET_NEON"
|
2215 |
|
|
{
|
2216 |
|
|
emit_insn (gen_neon_vpadd_internal (operands[0], operands[1],
|
2217 |
|
|
operands[2]));
|
2218 |
|
|
DONE;
|
2219 |
|
|
})
|
2220 |
|
|
|
2221 |
|
|
(define_insn "neon_vpaddl"
|
2222 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2223 |
|
|
(unspec: [(match_operand:VDQIW 1 "s_register_operand" "w")
|
2224 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2225 |
|
|
UNSPEC_VPADDL))]
|
2226 |
|
|
"TARGET_NEON"
|
2227 |
|
|
"vpaddl.%T2%#\t%0, %1"
|
2228 |
|
|
;; Assume this schedules like vaddl.
|
2229 |
|
|
[(set_attr "neon_type" "neon_int_3")]
|
2230 |
|
|
)
|
2231 |
|
|
|
2232 |
|
|
(define_insn "neon_vpadal"
|
2233 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2234 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
2235 |
|
|
(match_operand:VDQIW 2 "s_register_operand" "w")
|
2236 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2237 |
|
|
UNSPEC_VPADAL))]
|
2238 |
|
|
"TARGET_NEON"
|
2239 |
|
|
"vpadal.%T3%#\t%0, %2"
|
2240 |
|
|
;; Assume this schedules like vpadd.
|
2241 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
2242 |
|
|
)
|
2243 |
|
|
|
2244 |
|
|
(define_insn "neon_vpmax"
|
2245 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
2246 |
|
|
(unspec:VD [(match_operand:VD 1 "s_register_operand" "w")
|
2247 |
|
|
(match_operand:VD 2 "s_register_operand" "w")
|
2248 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2249 |
|
|
UNSPEC_VPMAX))]
|
2250 |
|
|
"TARGET_NEON"
|
2251 |
|
|
"vpmax.%T3%#\t%0, %1, %2"
|
2252 |
|
|
;; Assume this schedules like vmax.
|
2253 |
|
|
[(set (attr "neon_type")
|
2254 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2255 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2256 |
|
|
(const_string "neon_int_5")))]
|
2257 |
|
|
)
|
2258 |
|
|
|
2259 |
|
|
(define_insn "neon_vpmin"
|
2260 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
2261 |
|
|
(unspec:VD [(match_operand:VD 1 "s_register_operand" "w")
|
2262 |
|
|
(match_operand:VD 2 "s_register_operand" "w")
|
2263 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2264 |
|
|
UNSPEC_VPMIN))]
|
2265 |
|
|
"TARGET_NEON"
|
2266 |
|
|
"vpmin.%T3%#\t%0, %1, %2"
|
2267 |
|
|
;; Assume this schedules like vmin.
|
2268 |
|
|
[(set (attr "neon_type")
|
2269 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2270 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2271 |
|
|
(const_string "neon_int_5")))]
|
2272 |
|
|
)
|
2273 |
|
|
|
2274 |
|
|
(define_insn "neon_vrecps"
|
2275 |
|
|
[(set (match_operand:VCVTF 0 "s_register_operand" "=w")
|
2276 |
|
|
(unspec:VCVTF [(match_operand:VCVTF 1 "s_register_operand" "w")
|
2277 |
|
|
(match_operand:VCVTF 2 "s_register_operand" "w")
|
2278 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2279 |
|
|
UNSPEC_VRECPS))]
|
2280 |
|
|
"TARGET_NEON"
|
2281 |
|
|
"vrecps.\t%0, %1, %2"
|
2282 |
|
|
[(set (attr "neon_type")
|
2283 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2284 |
|
|
(const_string "neon_fp_vrecps_vrsqrts_ddd")
|
2285 |
|
|
(const_string "neon_fp_vrecps_vrsqrts_qqq")))]
|
2286 |
|
|
)
|
2287 |
|
|
|
2288 |
|
|
(define_insn "neon_vrsqrts"
|
2289 |
|
|
[(set (match_operand:VCVTF 0 "s_register_operand" "=w")
|
2290 |
|
|
(unspec:VCVTF [(match_operand:VCVTF 1 "s_register_operand" "w")
|
2291 |
|
|
(match_operand:VCVTF 2 "s_register_operand" "w")
|
2292 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2293 |
|
|
UNSPEC_VRSQRTS))]
|
2294 |
|
|
"TARGET_NEON"
|
2295 |
|
|
"vrsqrts.\t%0, %1, %2"
|
2296 |
|
|
[(set (attr "neon_type")
|
2297 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2298 |
|
|
(const_string "neon_fp_vrecps_vrsqrts_ddd")
|
2299 |
|
|
(const_string "neon_fp_vrecps_vrsqrts_qqq")))]
|
2300 |
|
|
)
|
2301 |
|
|
|
2302 |
|
|
(define_insn "neon_vabs"
|
2303 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
2304 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "w")
|
2305 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2306 |
|
|
UNSPEC_VABS))]
|
2307 |
|
|
"TARGET_NEON"
|
2308 |
|
|
"vabs.\t%0, %1"
|
2309 |
|
|
[(set (attr "neon_type")
|
2310 |
|
|
(if_then_else (ior (ne (symbol_ref "") (const_int 0))
|
2311 |
|
|
(ne (symbol_ref "") (const_int 0)))
|
2312 |
|
|
(if_then_else
|
2313 |
|
|
(ne (symbol_ref "") (const_int 0))
|
2314 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2315 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq"))
|
2316 |
|
|
(const_string "neon_vqneg_vqabs")))]
|
2317 |
|
|
)
|
2318 |
|
|
|
2319 |
|
|
(define_insn "neon_vqabs"
|
2320 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
2321 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
|
2322 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2323 |
|
|
UNSPEC_VQABS))]
|
2324 |
|
|
"TARGET_NEON"
|
2325 |
|
|
"vqabs.\t%0, %1"
|
2326 |
|
|
[(set_attr "neon_type" "neon_vqneg_vqabs")]
|
2327 |
|
|
)
|
2328 |
|
|
|
2329 |
|
|
(define_expand "neon_vneg"
|
2330 |
|
|
[(match_operand:VDQW 0 "s_register_operand" "")
|
2331 |
|
|
(match_operand:VDQW 1 "s_register_operand" "")
|
2332 |
|
|
(match_operand:SI 2 "immediate_operand" "")]
|
2333 |
|
|
"TARGET_NEON"
|
2334 |
|
|
{
|
2335 |
|
|
emit_insn (gen_neg2 (operands[0], operands[1]));
|
2336 |
|
|
DONE;
|
2337 |
|
|
})
|
2338 |
|
|
|
2339 |
|
|
(define_insn "neon_vqneg"
|
2340 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
2341 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
|
2342 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2343 |
|
|
UNSPEC_VQNEG))]
|
2344 |
|
|
"TARGET_NEON"
|
2345 |
|
|
"vqneg.\t%0, %1"
|
2346 |
|
|
[(set_attr "neon_type" "neon_vqneg_vqabs")]
|
2347 |
|
|
)
|
2348 |
|
|
|
2349 |
|
|
(define_insn "neon_vcls"
|
2350 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
2351 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
|
2352 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2353 |
|
|
UNSPEC_VCLS))]
|
2354 |
|
|
"TARGET_NEON"
|
2355 |
|
|
"vcls.\t%0, %1"
|
2356 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
2357 |
|
|
)
|
2358 |
|
|
|
2359 |
|
|
(define_insn "neon_vclz"
|
2360 |
|
|
[(set (match_operand:VDQIW 0 "s_register_operand" "=w")
|
2361 |
|
|
(unspec:VDQIW [(match_operand:VDQIW 1 "s_register_operand" "w")
|
2362 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2363 |
|
|
UNSPEC_VCLZ))]
|
2364 |
|
|
"TARGET_NEON"
|
2365 |
|
|
"vclz.\t%0, %1"
|
2366 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
2367 |
|
|
)
|
2368 |
|
|
|
2369 |
|
|
(define_insn "neon_vcnt"
|
2370 |
|
|
[(set (match_operand:VE 0 "s_register_operand" "=w")
|
2371 |
|
|
(unspec:VE [(match_operand:VE 1 "s_register_operand" "w")
|
2372 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2373 |
|
|
UNSPEC_VCNT))]
|
2374 |
|
|
"TARGET_NEON"
|
2375 |
|
|
"vcnt.\t%0, %1"
|
2376 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
2377 |
|
|
)
|
2378 |
|
|
|
2379 |
|
|
(define_insn "neon_vrecpe"
|
2380 |
|
|
[(set (match_operand:V32 0 "s_register_operand" "=w")
|
2381 |
|
|
(unspec:V32 [(match_operand:V32 1 "s_register_operand" "w")
|
2382 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2383 |
|
|
UNSPEC_VRECPE))]
|
2384 |
|
|
"TARGET_NEON"
|
2385 |
|
|
"vrecpe.\t%0, %1"
|
2386 |
|
|
[(set (attr "neon_type")
|
2387 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2388 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2389 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2390 |
|
|
)
|
2391 |
|
|
|
2392 |
|
|
(define_insn "neon_vrsqrte"
|
2393 |
|
|
[(set (match_operand:V32 0 "s_register_operand" "=w")
|
2394 |
|
|
(unspec:V32 [(match_operand:V32 1 "s_register_operand" "w")
|
2395 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2396 |
|
|
UNSPEC_VRSQRTE))]
|
2397 |
|
|
"TARGET_NEON"
|
2398 |
|
|
"vrsqrte.\t%0, %1"
|
2399 |
|
|
[(set (attr "neon_type")
|
2400 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2401 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2402 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2403 |
|
|
)
|
2404 |
|
|
|
2405 |
|
|
(define_expand "neon_vmvn"
|
2406 |
|
|
[(match_operand:VDQIW 0 "s_register_operand" "")
|
2407 |
|
|
(match_operand:VDQIW 1 "s_register_operand" "")
|
2408 |
|
|
(match_operand:SI 2 "immediate_operand" "")]
|
2409 |
|
|
"TARGET_NEON"
|
2410 |
|
|
{
|
2411 |
|
|
emit_insn (gen_one_cmpl2 (operands[0], operands[1]));
|
2412 |
|
|
DONE;
|
2413 |
|
|
})
|
2414 |
|
|
|
2415 |
|
|
(define_insn "neon_vget_lane_sext_internal"
|
2416 |
|
|
[(set (match_operand:SI 0 "s_register_operand" "=r")
|
2417 |
|
|
(sign_extend:SI
|
2418 |
|
|
(vec_select:
|
2419 |
|
|
(match_operand:VD 1 "s_register_operand" "w")
|
2420 |
|
|
(parallel [(match_operand:SI 2 "immediate_operand" "i")]))))]
|
2421 |
|
|
"TARGET_NEON"
|
2422 |
|
|
{
|
2423 |
|
|
if (BYTES_BIG_ENDIAN)
|
2424 |
|
|
{
|
2425 |
|
|
int elt = INTVAL (operands[2]);
|
2426 |
|
|
elt = GET_MODE_NUNITS (mode) - 1 - elt;
|
2427 |
|
|
operands[2] = GEN_INT (elt);
|
2428 |
|
|
}
|
2429 |
|
|
return "vmov%?.s\t%0, %P1[%c2]";
|
2430 |
|
|
}
|
2431 |
|
|
[(set_attr "predicable" "yes")
|
2432 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2433 |
|
|
)
|
2434 |
|
|
|
2435 |
|
|
(define_insn "neon_vget_lane_zext_internal"
|
2436 |
|
|
[(set (match_operand:SI 0 "s_register_operand" "=r")
|
2437 |
|
|
(zero_extend:SI
|
2438 |
|
|
(vec_select:
|
2439 |
|
|
(match_operand:VD 1 "s_register_operand" "w")
|
2440 |
|
|
(parallel [(match_operand:SI 2 "immediate_operand" "i")]))))]
|
2441 |
|
|
"TARGET_NEON"
|
2442 |
|
|
{
|
2443 |
|
|
if (BYTES_BIG_ENDIAN)
|
2444 |
|
|
{
|
2445 |
|
|
int elt = INTVAL (operands[2]);
|
2446 |
|
|
elt = GET_MODE_NUNITS (mode) - 1 - elt;
|
2447 |
|
|
operands[2] = GEN_INT (elt);
|
2448 |
|
|
}
|
2449 |
|
|
return "vmov%?.u\t%0, %P1[%c2]";
|
2450 |
|
|
}
|
2451 |
|
|
[(set_attr "predicable" "yes")
|
2452 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2453 |
|
|
)
|
2454 |
|
|
|
2455 |
|
|
(define_insn "neon_vget_lane_sext_internal"
|
2456 |
|
|
[(set (match_operand:SI 0 "s_register_operand" "=r")
|
2457 |
|
|
(sign_extend:SI
|
2458 |
|
|
(vec_select:
|
2459 |
|
|
(match_operand:VQ 1 "s_register_operand" "w")
|
2460 |
|
|
(parallel [(match_operand:SI 2 "immediate_operand" "i")]))))]
|
2461 |
|
|
"TARGET_NEON"
|
2462 |
|
|
{
|
2463 |
|
|
rtx ops[3];
|
2464 |
|
|
int regno = REGNO (operands[1]);
|
2465 |
|
|
unsigned int halfelts = GET_MODE_NUNITS (mode) / 2;
|
2466 |
|
|
unsigned int elt = INTVAL (operands[2]);
|
2467 |
|
|
unsigned int elt_adj = elt % halfelts;
|
2468 |
|
|
|
2469 |
|
|
if (BYTES_BIG_ENDIAN)
|
2470 |
|
|
elt_adj = halfelts - 1 - elt_adj;
|
2471 |
|
|
|
2472 |
|
|
ops[0] = operands[0];
|
2473 |
|
|
ops[1] = gen_rtx_REG (mode, regno + 2 * (elt / halfelts));
|
2474 |
|
|
ops[2] = GEN_INT (elt_adj);
|
2475 |
|
|
output_asm_insn ("vmov%?.s\t%0, %P1[%c2]", ops);
|
2476 |
|
|
|
2477 |
|
|
return "";
|
2478 |
|
|
}
|
2479 |
|
|
[(set_attr "predicable" "yes")
|
2480 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2481 |
|
|
)
|
2482 |
|
|
|
2483 |
|
|
(define_insn "neon_vget_lane_zext_internal"
|
2484 |
|
|
[(set (match_operand:SI 0 "s_register_operand" "=r")
|
2485 |
|
|
(zero_extend:SI
|
2486 |
|
|
(vec_select:
|
2487 |
|
|
(match_operand:VQ 1 "s_register_operand" "w")
|
2488 |
|
|
(parallel [(match_operand:SI 2 "immediate_operand" "i")]))))]
|
2489 |
|
|
"TARGET_NEON"
|
2490 |
|
|
{
|
2491 |
|
|
rtx ops[3];
|
2492 |
|
|
int regno = REGNO (operands[1]);
|
2493 |
|
|
unsigned int halfelts = GET_MODE_NUNITS (mode) / 2;
|
2494 |
|
|
unsigned int elt = INTVAL (operands[2]);
|
2495 |
|
|
unsigned int elt_adj = elt % halfelts;
|
2496 |
|
|
|
2497 |
|
|
if (BYTES_BIG_ENDIAN)
|
2498 |
|
|
elt_adj = halfelts - 1 - elt_adj;
|
2499 |
|
|
|
2500 |
|
|
ops[0] = operands[0];
|
2501 |
|
|
ops[1] = gen_rtx_REG (mode, regno + 2 * (elt / halfelts));
|
2502 |
|
|
ops[2] = GEN_INT (elt_adj);
|
2503 |
|
|
output_asm_insn ("vmov%?.u\t%0, %P1[%c2]", ops);
|
2504 |
|
|
|
2505 |
|
|
return "";
|
2506 |
|
|
}
|
2507 |
|
|
[(set_attr "predicable" "yes")
|
2508 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2509 |
|
|
)
|
2510 |
|
|
|
2511 |
|
|
(define_expand "neon_vget_lane"
|
2512 |
|
|
[(match_operand: 0 "s_register_operand" "")
|
2513 |
|
|
(match_operand:VDQW 1 "s_register_operand" "")
|
2514 |
|
|
(match_operand:SI 2 "immediate_operand" "")
|
2515 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
2516 |
|
|
"TARGET_NEON"
|
2517 |
|
|
{
|
2518 |
|
|
HOST_WIDE_INT magic = INTVAL (operands[3]);
|
2519 |
|
|
rtx insn;
|
2520 |
|
|
|
2521 |
|
|
neon_lane_bounds (operands[2], 0, GET_MODE_NUNITS (mode));
|
2522 |
|
|
|
2523 |
|
|
if (BYTES_BIG_ENDIAN)
|
2524 |
|
|
{
|
2525 |
|
|
/* The intrinsics are defined in terms of a model where the
|
2526 |
|
|
element ordering in memory is vldm order, whereas the generic
|
2527 |
|
|
RTL is defined in terms of a model where the element ordering
|
2528 |
|
|
in memory is array order. Convert the lane number to conform
|
2529 |
|
|
to this model. */
|
2530 |
|
|
unsigned int elt = INTVAL (operands[2]);
|
2531 |
|
|
unsigned int reg_nelts
|
2532 |
|
|
= 64 / GET_MODE_BITSIZE (GET_MODE_INNER (mode));
|
2533 |
|
|
elt ^= reg_nelts - 1;
|
2534 |
|
|
operands[2] = GEN_INT (elt);
|
2535 |
|
|
}
|
2536 |
|
|
|
2537 |
|
|
if ((magic & 3) == 3 || GET_MODE_BITSIZE (GET_MODE_INNER (mode)) == 32)
|
2538 |
|
|
insn = gen_vec_extract (operands[0], operands[1], operands[2]);
|
2539 |
|
|
else
|
2540 |
|
|
{
|
2541 |
|
|
if ((magic & 1) != 0)
|
2542 |
|
|
insn = gen_neon_vget_lane_sext_internal (operands[0], operands[1],
|
2543 |
|
|
operands[2]);
|
2544 |
|
|
else
|
2545 |
|
|
insn = gen_neon_vget_lane_zext_internal (operands[0], operands[1],
|
2546 |
|
|
operands[2]);
|
2547 |
|
|
}
|
2548 |
|
|
emit_insn (insn);
|
2549 |
|
|
DONE;
|
2550 |
|
|
})
|
2551 |
|
|
|
2552 |
|
|
; Operand 3 (info word) is ignored because it does nothing useful with 64-bit
|
2553 |
|
|
; elements.
|
2554 |
|
|
|
2555 |
|
|
(define_insn "neon_vget_lanedi"
|
2556 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=r")
|
2557 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "w")
|
2558 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
2559 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2560 |
|
|
UNSPEC_VGET_LANE))]
|
2561 |
|
|
"TARGET_NEON"
|
2562 |
|
|
{
|
2563 |
|
|
neon_lane_bounds (operands[2], 0, 1);
|
2564 |
|
|
return "vmov%?\t%Q0, %R0, %P1 @ di";
|
2565 |
|
|
}
|
2566 |
|
|
[(set_attr "predicable" "yes")
|
2567 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2568 |
|
|
)
|
2569 |
|
|
|
2570 |
|
|
(define_insn "neon_vget_lanev2di"
|
2571 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=r")
|
2572 |
|
|
(unspec:DI [(match_operand:V2DI 1 "s_register_operand" "w")
|
2573 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
2574 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2575 |
|
|
UNSPEC_VGET_LANE))]
|
2576 |
|
|
"TARGET_NEON"
|
2577 |
|
|
{
|
2578 |
|
|
rtx ops[2];
|
2579 |
|
|
unsigned int regno = REGNO (operands[1]);
|
2580 |
|
|
unsigned int elt = INTVAL (operands[2]);
|
2581 |
|
|
|
2582 |
|
|
neon_lane_bounds (operands[2], 0, 2);
|
2583 |
|
|
|
2584 |
|
|
ops[0] = operands[0];
|
2585 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2 * elt);
|
2586 |
|
|
output_asm_insn ("vmov%?\t%Q0, %R0, %P1 @ v2di", ops);
|
2587 |
|
|
|
2588 |
|
|
return "";
|
2589 |
|
|
}
|
2590 |
|
|
[(set_attr "predicable" "yes")
|
2591 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2592 |
|
|
)
|
2593 |
|
|
|
2594 |
|
|
(define_insn "neon_vset_lane"
|
2595 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
2596 |
|
|
(unspec:VD [(match_operand: 1 "s_register_operand" "r")
|
2597 |
|
|
(match_operand:VD 2 "s_register_operand" "0")
|
2598 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2599 |
|
|
UNSPEC_VSET_LANE))]
|
2600 |
|
|
"TARGET_NEON"
|
2601 |
|
|
{
|
2602 |
|
|
neon_lane_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
2603 |
|
|
return "vmov%?.\t%P0[%c3], %1";
|
2604 |
|
|
}
|
2605 |
|
|
[(set_attr "predicable" "yes")
|
2606 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2607 |
|
|
)
|
2608 |
|
|
|
2609 |
|
|
; See neon_vget_lanedi comment for reasons operands 2 & 3 are ignored.
|
2610 |
|
|
|
2611 |
|
|
(define_insn "neon_vset_lanedi"
|
2612 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w")
|
2613 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "r")
|
2614 |
|
|
(match_operand:DI 2 "s_register_operand" "0")
|
2615 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2616 |
|
|
UNSPEC_VSET_LANE))]
|
2617 |
|
|
"TARGET_NEON"
|
2618 |
|
|
{
|
2619 |
|
|
neon_lane_bounds (operands[3], 0, 1);
|
2620 |
|
|
return "vmov%?\t%P0, %Q1, %R1 @ di";
|
2621 |
|
|
}
|
2622 |
|
|
[(set_attr "predicable" "yes")
|
2623 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2624 |
|
|
)
|
2625 |
|
|
|
2626 |
|
|
(define_insn "neon_vset_lane"
|
2627 |
|
|
[(set (match_operand:VQ 0 "s_register_operand" "=w")
|
2628 |
|
|
(unspec:VQ [(match_operand: 1 "s_register_operand" "r")
|
2629 |
|
|
(match_operand:VQ 2 "s_register_operand" "0")
|
2630 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2631 |
|
|
UNSPEC_VSET_LANE))]
|
2632 |
|
|
"TARGET_NEON"
|
2633 |
|
|
{
|
2634 |
|
|
rtx ops[4];
|
2635 |
|
|
unsigned int regno = REGNO (operands[0]);
|
2636 |
|
|
unsigned int halfelts = GET_MODE_NUNITS (mode) / 2;
|
2637 |
|
|
unsigned int elt = INTVAL (operands[3]);
|
2638 |
|
|
|
2639 |
|
|
neon_lane_bounds (operands[3], 0, halfelts * 2);
|
2640 |
|
|
|
2641 |
|
|
ops[0] = gen_rtx_REG (mode, regno + 2 * (elt / halfelts));
|
2642 |
|
|
ops[1] = operands[1];
|
2643 |
|
|
ops[2] = GEN_INT (elt % halfelts);
|
2644 |
|
|
output_asm_insn ("vmov%?.\t%P0[%c2], %1", ops);
|
2645 |
|
|
|
2646 |
|
|
return "";
|
2647 |
|
|
}
|
2648 |
|
|
[(set_attr "predicable" "yes")
|
2649 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2650 |
|
|
)
|
2651 |
|
|
|
2652 |
|
|
(define_insn "neon_vset_lanev2di"
|
2653 |
|
|
[(set (match_operand:V2DI 0 "s_register_operand" "=w")
|
2654 |
|
|
(unspec:V2DI [(match_operand:DI 1 "s_register_operand" "r")
|
2655 |
|
|
(match_operand:V2DI 2 "s_register_operand" "0")
|
2656 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2657 |
|
|
UNSPEC_VSET_LANE))]
|
2658 |
|
|
"TARGET_NEON"
|
2659 |
|
|
{
|
2660 |
|
|
rtx ops[2];
|
2661 |
|
|
unsigned int regno = REGNO (operands[0]);
|
2662 |
|
|
unsigned int elt = INTVAL (operands[3]);
|
2663 |
|
|
|
2664 |
|
|
neon_lane_bounds (operands[3], 0, 2);
|
2665 |
|
|
|
2666 |
|
|
ops[0] = gen_rtx_REG (DImode, regno + 2 * elt);
|
2667 |
|
|
ops[1] = operands[1];
|
2668 |
|
|
output_asm_insn ("vmov%?\t%P0, %Q1, %R1 @ v2di", ops);
|
2669 |
|
|
|
2670 |
|
|
return "";
|
2671 |
|
|
}
|
2672 |
|
|
[(set_attr "predicable" "yes")
|
2673 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2674 |
|
|
)
|
2675 |
|
|
|
2676 |
|
|
(define_expand "neon_vcreate"
|
2677 |
|
|
[(match_operand:VDX 0 "s_register_operand" "")
|
2678 |
|
|
(match_operand:DI 1 "general_operand" "")]
|
2679 |
|
|
"TARGET_NEON"
|
2680 |
|
|
{
|
2681 |
|
|
rtx src = gen_lowpart (mode, operands[1]);
|
2682 |
|
|
emit_move_insn (operands[0], src);
|
2683 |
|
|
DONE;
|
2684 |
|
|
})
|
2685 |
|
|
|
2686 |
|
|
(define_insn "neon_vdup_n"
|
2687 |
|
|
[(set (match_operand:VX 0 "s_register_operand" "=w")
|
2688 |
|
|
(unspec:VX [(match_operand: 1 "s_register_operand" "r")]
|
2689 |
|
|
UNSPEC_VDUP_N))]
|
2690 |
|
|
"TARGET_NEON"
|
2691 |
|
|
"vdup%?.\t%0, %1"
|
2692 |
|
|
;; Assume this schedules like vmov.
|
2693 |
|
|
[(set_attr "predicable" "yes")
|
2694 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2695 |
|
|
)
|
2696 |
|
|
|
2697 |
|
|
(define_insn "neon_vdup_n"
|
2698 |
|
|
[(set (match_operand:V32 0 "s_register_operand" "=w,w")
|
2699 |
|
|
(unspec:V32 [(match_operand: 1 "s_register_operand" "r,t")]
|
2700 |
|
|
UNSPEC_VDUP_N))]
|
2701 |
|
|
"TARGET_NEON"
|
2702 |
|
|
"@
|
2703 |
|
|
vdup%?.\t%0, %1
|
2704 |
|
|
vdup%?.\t%0, %y1"
|
2705 |
|
|
;; Assume this schedules like vmov.
|
2706 |
|
|
[(set_attr "predicable" "yes")
|
2707 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2708 |
|
|
)
|
2709 |
|
|
|
2710 |
|
|
(define_insn "neon_vdup_ndi"
|
2711 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w")
|
2712 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "r")]
|
2713 |
|
|
UNSPEC_VDUP_N))]
|
2714 |
|
|
"TARGET_NEON"
|
2715 |
|
|
"vmov%?\t%P0, %Q1, %R1"
|
2716 |
|
|
[(set_attr "predicable" "yes")
|
2717 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2718 |
|
|
)
|
2719 |
|
|
|
2720 |
|
|
(define_insn "neon_vdup_nv2di"
|
2721 |
|
|
[(set (match_operand:V2DI 0 "s_register_operand" "=w")
|
2722 |
|
|
(unspec:V2DI [(match_operand:DI 1 "s_register_operand" "r")]
|
2723 |
|
|
UNSPEC_VDUP_N))]
|
2724 |
|
|
"TARGET_NEON"
|
2725 |
|
|
"vmov%?\t%e0, %Q1, %R1\;vmov%?\t%f0, %Q1, %R1"
|
2726 |
|
|
[(set_attr "predicable" "yes")
|
2727 |
|
|
(set_attr "length" "8")
|
2728 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2729 |
|
|
)
|
2730 |
|
|
|
2731 |
|
|
(define_insn "neon_vdup_lane"
|
2732 |
|
|
[(set (match_operand:VD 0 "s_register_operand" "=w")
|
2733 |
|
|
(unspec:VD [(match_operand:VD 1 "s_register_operand" "w")
|
2734 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2735 |
|
|
UNSPEC_VDUP_LANE))]
|
2736 |
|
|
"TARGET_NEON"
|
2737 |
|
|
{
|
2738 |
|
|
neon_lane_bounds (operands[2], 0, GET_MODE_NUNITS (mode));
|
2739 |
|
|
return "vdup.\t%P0, %P1[%c2]";
|
2740 |
|
|
}
|
2741 |
|
|
;; Assume this schedules like vmov.
|
2742 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
2743 |
|
|
)
|
2744 |
|
|
|
2745 |
|
|
(define_insn "neon_vdup_lane"
|
2746 |
|
|
[(set (match_operand:VQ 0 "s_register_operand" "=w")
|
2747 |
|
|
(unspec:VQ [(match_operand: 1 "s_register_operand" "w")
|
2748 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2749 |
|
|
UNSPEC_VDUP_LANE))]
|
2750 |
|
|
"TARGET_NEON"
|
2751 |
|
|
{
|
2752 |
|
|
neon_lane_bounds (operands[2], 0, GET_MODE_NUNITS (mode));
|
2753 |
|
|
return "vdup.\t%q0, %P1[%c2]";
|
2754 |
|
|
}
|
2755 |
|
|
;; Assume this schedules like vmov.
|
2756 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
2757 |
|
|
)
|
2758 |
|
|
|
2759 |
|
|
; Scalar index is ignored, since only zero is valid here.
|
2760 |
|
|
(define_expand "neon_vdup_lanedi"
|
2761 |
|
|
[(set (match_operand:DI 0 "s_register_operand" "=w")
|
2762 |
|
|
(unspec:DI [(match_operand:DI 1 "s_register_operand" "w")
|
2763 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2764 |
|
|
UNSPEC_VDUP_LANE))]
|
2765 |
|
|
"TARGET_NEON"
|
2766 |
|
|
{
|
2767 |
|
|
neon_lane_bounds (operands[2], 0, 1);
|
2768 |
|
|
emit_move_insn (operands[0], operands[1]);
|
2769 |
|
|
DONE;
|
2770 |
|
|
})
|
2771 |
|
|
|
2772 |
|
|
; Likewise.
|
2773 |
|
|
(define_insn "neon_vdup_lanev2di"
|
2774 |
|
|
[(set (match_operand:V2DI 0 "s_register_operand" "=w")
|
2775 |
|
|
(unspec:V2DI [(match_operand:DI 1 "s_register_operand" "w")
|
2776 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2777 |
|
|
UNSPEC_VDUP_LANE))]
|
2778 |
|
|
"TARGET_NEON"
|
2779 |
|
|
{
|
2780 |
|
|
neon_lane_bounds (operands[2], 0, 1);
|
2781 |
|
|
return "vmov\t%e0, %P1\;vmov\t%f0, %P1";
|
2782 |
|
|
}
|
2783 |
|
|
[(set_attr "length" "8")
|
2784 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2785 |
|
|
)
|
2786 |
|
|
|
2787 |
|
|
;; In this insn, operand 1 should be low, and operand 2 the high part of the
|
2788 |
|
|
;; dest vector.
|
2789 |
|
|
;; FIXME: A different implementation of this builtin could make it much
|
2790 |
|
|
;; more likely that we wouldn't actually need to output anything (we could make
|
2791 |
|
|
;; it so that the reg allocator puts things in the right places magically
|
2792 |
|
|
;; instead). Lack of subregs for vectors makes that tricky though, I think.
|
2793 |
|
|
|
2794 |
|
|
(define_insn "neon_vcombine"
|
2795 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2796 |
|
|
(unspec: [(match_operand:VDX 1 "s_register_operand" "w")
|
2797 |
|
|
(match_operand:VDX 2 "s_register_operand" "w")]
|
2798 |
|
|
UNSPEC_VCOMBINE))]
|
2799 |
|
|
"TARGET_NEON"
|
2800 |
|
|
{
|
2801 |
|
|
int dest = REGNO (operands[0]);
|
2802 |
|
|
int src1 = REGNO (operands[1]);
|
2803 |
|
|
int src2 = REGNO (operands[2]);
|
2804 |
|
|
rtx destlo;
|
2805 |
|
|
|
2806 |
|
|
if (src1 == dest && src2 == dest + 2)
|
2807 |
|
|
return "";
|
2808 |
|
|
else if (src2 == dest && src1 == dest + 2)
|
2809 |
|
|
/* Special case of reversed high/low parts. */
|
2810 |
|
|
return "vswp\t%P1, %P2";
|
2811 |
|
|
|
2812 |
|
|
destlo = gen_rtx_REG (mode, dest);
|
2813 |
|
|
|
2814 |
|
|
if (!reg_overlap_mentioned_p (operands[2], destlo))
|
2815 |
|
|
{
|
2816 |
|
|
/* Try to avoid unnecessary moves if part of the result is in the right
|
2817 |
|
|
place already. */
|
2818 |
|
|
if (src1 != dest)
|
2819 |
|
|
output_asm_insn ("vmov\t%e0, %P1", operands);
|
2820 |
|
|
if (src2 != dest + 2)
|
2821 |
|
|
output_asm_insn ("vmov\t%f0, %P2", operands);
|
2822 |
|
|
}
|
2823 |
|
|
else
|
2824 |
|
|
{
|
2825 |
|
|
if (src2 != dest + 2)
|
2826 |
|
|
output_asm_insn ("vmov\t%f0, %P2", operands);
|
2827 |
|
|
if (src1 != dest)
|
2828 |
|
|
output_asm_insn ("vmov\t%e0, %P1", operands);
|
2829 |
|
|
}
|
2830 |
|
|
|
2831 |
|
|
return "";
|
2832 |
|
|
}
|
2833 |
|
|
;; We set the neon_type attribute based on the vmov instructions above.
|
2834 |
|
|
[(set_attr "length" "8")
|
2835 |
|
|
(set_attr "neon_type" "neon_bp_simple")]
|
2836 |
|
|
)
|
2837 |
|
|
|
2838 |
|
|
(define_insn "neon_vget_high"
|
2839 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2840 |
|
|
(unspec: [(match_operand:VQX 1 "s_register_operand" "w")]
|
2841 |
|
|
UNSPEC_VGET_HIGH))]
|
2842 |
|
|
"TARGET_NEON"
|
2843 |
|
|
{
|
2844 |
|
|
int dest = REGNO (operands[0]);
|
2845 |
|
|
int src = REGNO (operands[1]);
|
2846 |
|
|
|
2847 |
|
|
if (dest != src + 2)
|
2848 |
|
|
return "vmov\t%P0, %f1";
|
2849 |
|
|
else
|
2850 |
|
|
return "";
|
2851 |
|
|
}
|
2852 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
2853 |
|
|
)
|
2854 |
|
|
|
2855 |
|
|
(define_insn "neon_vget_low"
|
2856 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2857 |
|
|
(unspec: [(match_operand:VQX 1 "s_register_operand" "w")]
|
2858 |
|
|
UNSPEC_VGET_LOW))]
|
2859 |
|
|
"TARGET_NEON"
|
2860 |
|
|
{
|
2861 |
|
|
int dest = REGNO (operands[0]);
|
2862 |
|
|
int src = REGNO (operands[1]);
|
2863 |
|
|
|
2864 |
|
|
if (dest != src)
|
2865 |
|
|
return "vmov\t%P0, %e1";
|
2866 |
|
|
else
|
2867 |
|
|
return "";
|
2868 |
|
|
}
|
2869 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
2870 |
|
|
)
|
2871 |
|
|
|
2872 |
|
|
(define_insn "neon_vcvt"
|
2873 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2874 |
|
|
(unspec: [(match_operand:VCVTF 1 "s_register_operand" "w")
|
2875 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2876 |
|
|
UNSPEC_VCVT))]
|
2877 |
|
|
"TARGET_NEON"
|
2878 |
|
|
"vcvt.%T2%#32.f32\t%0, %1"
|
2879 |
|
|
[(set (attr "neon_type")
|
2880 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2881 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2882 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2883 |
|
|
)
|
2884 |
|
|
|
2885 |
|
|
(define_insn "neon_vcvt"
|
2886 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2887 |
|
|
(unspec: [(match_operand:VCVTI 1 "s_register_operand" "w")
|
2888 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2889 |
|
|
UNSPEC_VCVT))]
|
2890 |
|
|
"TARGET_NEON"
|
2891 |
|
|
"vcvt.f32.%T2%#32\t%0, %1"
|
2892 |
|
|
[(set (attr "neon_type")
|
2893 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2894 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2895 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2896 |
|
|
)
|
2897 |
|
|
|
2898 |
|
|
(define_insn "neon_vcvt_n"
|
2899 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2900 |
|
|
(unspec: [(match_operand:VCVTF 1 "s_register_operand" "w")
|
2901 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
2902 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2903 |
|
|
UNSPEC_VCVT_N))]
|
2904 |
|
|
"TARGET_NEON"
|
2905 |
|
|
{
|
2906 |
|
|
neon_const_bounds (operands[2], 1, 33);
|
2907 |
|
|
return "vcvt.%T3%#32.f32\t%0, %1, %2";
|
2908 |
|
|
}
|
2909 |
|
|
[(set (attr "neon_type")
|
2910 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2911 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2912 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2913 |
|
|
)
|
2914 |
|
|
|
2915 |
|
|
(define_insn "neon_vcvt_n"
|
2916 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2917 |
|
|
(unspec: [(match_operand:VCVTI 1 "s_register_operand" "w")
|
2918 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
2919 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
2920 |
|
|
UNSPEC_VCVT_N))]
|
2921 |
|
|
"TARGET_NEON"
|
2922 |
|
|
{
|
2923 |
|
|
neon_const_bounds (operands[2], 1, 33);
|
2924 |
|
|
return "vcvt.f32.%T3%#32\t%0, %1, %2";
|
2925 |
|
|
}
|
2926 |
|
|
[(set (attr "neon_type")
|
2927 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2928 |
|
|
(const_string "neon_fp_vadd_ddd_vabs_dd")
|
2929 |
|
|
(const_string "neon_fp_vadd_qqq_vabs_qq")))]
|
2930 |
|
|
)
|
2931 |
|
|
|
2932 |
|
|
(define_insn "neon_vmovn"
|
2933 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2934 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
2935 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2936 |
|
|
UNSPEC_VMOVN))]
|
2937 |
|
|
"TARGET_NEON"
|
2938 |
|
|
"vmovn.\t%P0, %q1"
|
2939 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
2940 |
|
|
)
|
2941 |
|
|
|
2942 |
|
|
(define_insn "neon_vqmovn"
|
2943 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2944 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
2945 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2946 |
|
|
UNSPEC_VQMOVN))]
|
2947 |
|
|
"TARGET_NEON"
|
2948 |
|
|
"vqmovn.%T2%#\t%P0, %q1"
|
2949 |
|
|
[(set_attr "neon_type" "neon_shift_2")]
|
2950 |
|
|
)
|
2951 |
|
|
|
2952 |
|
|
(define_insn "neon_vqmovun"
|
2953 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2954 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
2955 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2956 |
|
|
UNSPEC_VQMOVUN))]
|
2957 |
|
|
"TARGET_NEON"
|
2958 |
|
|
"vqmovun.\t%P0, %q1"
|
2959 |
|
|
[(set_attr "neon_type" "neon_shift_2")]
|
2960 |
|
|
)
|
2961 |
|
|
|
2962 |
|
|
(define_insn "neon_vmovl"
|
2963 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
2964 |
|
|
(unspec: [(match_operand:VW 1 "s_register_operand" "w")
|
2965 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
2966 |
|
|
UNSPEC_VMOVL))]
|
2967 |
|
|
"TARGET_NEON"
|
2968 |
|
|
"vmovl.%T2%#\t%q0, %P1"
|
2969 |
|
|
[(set_attr "neon_type" "neon_shift_1")]
|
2970 |
|
|
)
|
2971 |
|
|
|
2972 |
|
|
(define_insn "neon_vmul_lane"
|
2973 |
|
|
[(set (match_operand:VMD 0 "s_register_operand" "=w")
|
2974 |
|
|
(unspec:VMD [(match_operand:VMD 1 "s_register_operand" "w")
|
2975 |
|
|
(match_operand:VMD 2 "s_register_operand"
|
2976 |
|
|
"")
|
2977 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
2978 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
2979 |
|
|
UNSPEC_VMUL_LANE))]
|
2980 |
|
|
"TARGET_NEON"
|
2981 |
|
|
{
|
2982 |
|
|
neon_lane_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
2983 |
|
|
return "vmul.\t%P0, %P1, %P2[%c3]";
|
2984 |
|
|
}
|
2985 |
|
|
[(set (attr "neon_type")
|
2986 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2987 |
|
|
(const_string "neon_fp_vmul_ddd")
|
2988 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
2989 |
|
|
(const_string "neon_mul_ddd_16_scalar_32_16_long_scalar")
|
2990 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar"))))]
|
2991 |
|
|
)
|
2992 |
|
|
|
2993 |
|
|
(define_insn "neon_vmul_lane"
|
2994 |
|
|
[(set (match_operand:VMQ 0 "s_register_operand" "=w")
|
2995 |
|
|
(unspec:VMQ [(match_operand:VMQ 1 "s_register_operand" "w")
|
2996 |
|
|
(match_operand: 2 "s_register_operand"
|
2997 |
|
|
"")
|
2998 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
2999 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
3000 |
|
|
UNSPEC_VMUL_LANE))]
|
3001 |
|
|
"TARGET_NEON"
|
3002 |
|
|
{
|
3003 |
|
|
neon_lane_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
3004 |
|
|
return "vmul.\t%q0, %q1, %P2[%c3]";
|
3005 |
|
|
}
|
3006 |
|
|
[(set (attr "neon_type")
|
3007 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3008 |
|
|
(const_string "neon_fp_vmul_qqd")
|
3009 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3010 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar")
|
3011 |
|
|
(const_string "neon_mul_qqd_32_scalar"))))]
|
3012 |
|
|
)
|
3013 |
|
|
|
3014 |
|
|
(define_insn "neon_vmull_lane"
|
3015 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3016 |
|
|
(unspec: [(match_operand:VMDI 1 "s_register_operand" "w")
|
3017 |
|
|
(match_operand:VMDI 2 "s_register_operand"
|
3018 |
|
|
"")
|
3019 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
3020 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
3021 |
|
|
UNSPEC_VMULL_LANE))]
|
3022 |
|
|
"TARGET_NEON"
|
3023 |
|
|
{
|
3024 |
|
|
neon_lane_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
3025 |
|
|
return "vmull.%T4%#\t%q0, %P1, %P2[%c3]";
|
3026 |
|
|
}
|
3027 |
|
|
[(set (attr "neon_type")
|
3028 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3029 |
|
|
(const_string "neon_mul_ddd_16_scalar_32_16_long_scalar")
|
3030 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar")))]
|
3031 |
|
|
)
|
3032 |
|
|
|
3033 |
|
|
(define_insn "neon_vqdmull_lane"
|
3034 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3035 |
|
|
(unspec: [(match_operand:VMDI 1 "s_register_operand" "w")
|
3036 |
|
|
(match_operand:VMDI 2 "s_register_operand"
|
3037 |
|
|
"")
|
3038 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
3039 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
3040 |
|
|
UNSPEC_VQDMULL_LANE))]
|
3041 |
|
|
"TARGET_NEON"
|
3042 |
|
|
{
|
3043 |
|
|
neon_lane_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
3044 |
|
|
return "vqdmull.\t%q0, %P1, %P2[%c3]";
|
3045 |
|
|
}
|
3046 |
|
|
[(set (attr "neon_type")
|
3047 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3048 |
|
|
(const_string "neon_mul_ddd_16_scalar_32_16_long_scalar")
|
3049 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar")))]
|
3050 |
|
|
)
|
3051 |
|
|
|
3052 |
|
|
(define_insn "neon_vqdmulh_lane"
|
3053 |
|
|
[(set (match_operand:VMQI 0 "s_register_operand" "=w")
|
3054 |
|
|
(unspec:VMQI [(match_operand:VMQI 1 "s_register_operand" "w")
|
3055 |
|
|
(match_operand: 2 "s_register_operand"
|
3056 |
|
|
"")
|
3057 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
3058 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
3059 |
|
|
UNSPEC_VQDMULH_LANE))]
|
3060 |
|
|
"TARGET_NEON"
|
3061 |
|
|
{
|
3062 |
|
|
neon_lane_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
3063 |
|
|
return "vq%O4dmulh.%T4%#\t%q0, %q1, %P2[%c3]";
|
3064 |
|
|
}
|
3065 |
|
|
[(set (attr "neon_type")
|
3066 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3067 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar")
|
3068 |
|
|
(const_string "neon_mul_qqd_32_scalar")))]
|
3069 |
|
|
)
|
3070 |
|
|
|
3071 |
|
|
(define_insn "neon_vqdmulh_lane"
|
3072 |
|
|
[(set (match_operand:VMDI 0 "s_register_operand" "=w")
|
3073 |
|
|
(unspec:VMDI [(match_operand:VMDI 1 "s_register_operand" "w")
|
3074 |
|
|
(match_operand:VMDI 2 "s_register_operand"
|
3075 |
|
|
"")
|
3076 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
3077 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
3078 |
|
|
UNSPEC_VQDMULH_LANE))]
|
3079 |
|
|
"TARGET_NEON"
|
3080 |
|
|
{
|
3081 |
|
|
neon_lane_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
3082 |
|
|
return "vq%O4dmulh.%T4%#\t%P0, %P1, %P2[%c3]";
|
3083 |
|
|
}
|
3084 |
|
|
[(set (attr "neon_type")
|
3085 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3086 |
|
|
(const_string "neon_mul_ddd_16_scalar_32_16_long_scalar")
|
3087 |
|
|
(const_string "neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar")))]
|
3088 |
|
|
)
|
3089 |
|
|
|
3090 |
|
|
(define_insn "neon_vmla_lane"
|
3091 |
|
|
[(set (match_operand:VMD 0 "s_register_operand" "=w")
|
3092 |
|
|
(unspec:VMD [(match_operand:VMD 1 "s_register_operand" "0")
|
3093 |
|
|
(match_operand:VMD 2 "s_register_operand" "w")
|
3094 |
|
|
(match_operand:VMD 3 "s_register_operand"
|
3095 |
|
|
"")
|
3096 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3097 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3098 |
|
|
UNSPEC_VMLA_LANE))]
|
3099 |
|
|
"TARGET_NEON"
|
3100 |
|
|
{
|
3101 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3102 |
|
|
return "vmla.\t%P0, %P2, %P3[%c4]";
|
3103 |
|
|
}
|
3104 |
|
|
[(set (attr "neon_type")
|
3105 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3106 |
|
|
(const_string "neon_fp_vmla_ddd_scalar")
|
3107 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3108 |
|
|
(const_string "neon_mla_ddd_16_scalar_qdd_32_16_long_scalar")
|
3109 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long"))))]
|
3110 |
|
|
)
|
3111 |
|
|
|
3112 |
|
|
(define_insn "neon_vmla_lane"
|
3113 |
|
|
[(set (match_operand:VMQ 0 "s_register_operand" "=w")
|
3114 |
|
|
(unspec:VMQ [(match_operand:VMQ 1 "s_register_operand" "0")
|
3115 |
|
|
(match_operand:VMQ 2 "s_register_operand" "w")
|
3116 |
|
|
(match_operand: 3 "s_register_operand"
|
3117 |
|
|
"")
|
3118 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3119 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3120 |
|
|
UNSPEC_VMLA_LANE))]
|
3121 |
|
|
"TARGET_NEON"
|
3122 |
|
|
{
|
3123 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3124 |
|
|
return "vmla.\t%q0, %q2, %P3[%c4]";
|
3125 |
|
|
}
|
3126 |
|
|
[(set (attr "neon_type")
|
3127 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3128 |
|
|
(const_string "neon_fp_vmla_qqq_scalar")
|
3129 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3130 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")
|
3131 |
|
|
(const_string "neon_mla_qqq_32_qqd_32_scalar"))))]
|
3132 |
|
|
)
|
3133 |
|
|
|
3134 |
|
|
(define_insn "neon_vmlal_lane"
|
3135 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3136 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
3137 |
|
|
(match_operand:VMDI 2 "s_register_operand" "w")
|
3138 |
|
|
(match_operand:VMDI 3 "s_register_operand"
|
3139 |
|
|
"")
|
3140 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3141 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3142 |
|
|
UNSPEC_VMLAL_LANE))]
|
3143 |
|
|
"TARGET_NEON"
|
3144 |
|
|
{
|
3145 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3146 |
|
|
return "vmlal.%T5%#\t%q0, %P2, %P3[%c4]";
|
3147 |
|
|
}
|
3148 |
|
|
[(set (attr "neon_type")
|
3149 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3150 |
|
|
(const_string "neon_mla_ddd_16_scalar_qdd_32_16_long_scalar")
|
3151 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
3152 |
|
|
)
|
3153 |
|
|
|
3154 |
|
|
(define_insn "neon_vqdmlal_lane"
|
3155 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3156 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
3157 |
|
|
(match_operand:VMDI 2 "s_register_operand" "w")
|
3158 |
|
|
(match_operand:VMDI 3 "s_register_operand"
|
3159 |
|
|
"")
|
3160 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3161 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3162 |
|
|
UNSPEC_VQDMLAL_LANE))]
|
3163 |
|
|
"TARGET_NEON"
|
3164 |
|
|
{
|
3165 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3166 |
|
|
return "vqdmlal.\t%q0, %P2, %P3[%c4]";
|
3167 |
|
|
}
|
3168 |
|
|
[(set (attr "neon_type")
|
3169 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3170 |
|
|
(const_string "neon_mla_ddd_16_scalar_qdd_32_16_long_scalar")
|
3171 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
3172 |
|
|
)
|
3173 |
|
|
|
3174 |
|
|
(define_insn "neon_vmls_lane"
|
3175 |
|
|
[(set (match_operand:VMD 0 "s_register_operand" "=w")
|
3176 |
|
|
(unspec:VMD [(match_operand:VMD 1 "s_register_operand" "0")
|
3177 |
|
|
(match_operand:VMD 2 "s_register_operand" "w")
|
3178 |
|
|
(match_operand:VMD 3 "s_register_operand"
|
3179 |
|
|
"")
|
3180 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3181 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3182 |
|
|
UNSPEC_VMLS_LANE))]
|
3183 |
|
|
"TARGET_NEON"
|
3184 |
|
|
{
|
3185 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3186 |
|
|
return "vmls.\t%P0, %P2, %P3[%c4]";
|
3187 |
|
|
}
|
3188 |
|
|
[(set (attr "neon_type")
|
3189 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3190 |
|
|
(const_string "neon_fp_vmla_ddd_scalar")
|
3191 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3192 |
|
|
(const_string "neon_mla_ddd_16_scalar_qdd_32_16_long_scalar")
|
3193 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long"))))]
|
3194 |
|
|
)
|
3195 |
|
|
|
3196 |
|
|
(define_insn "neon_vmls_lane"
|
3197 |
|
|
[(set (match_operand:VMQ 0 "s_register_operand" "=w")
|
3198 |
|
|
(unspec:VMQ [(match_operand:VMQ 1 "s_register_operand" "0")
|
3199 |
|
|
(match_operand:VMQ 2 "s_register_operand" "w")
|
3200 |
|
|
(match_operand: 3 "s_register_operand"
|
3201 |
|
|
"")
|
3202 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3203 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3204 |
|
|
UNSPEC_VMLS_LANE))]
|
3205 |
|
|
"TARGET_NEON"
|
3206 |
|
|
{
|
3207 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3208 |
|
|
return "vmls.\t%q0, %q2, %P3[%c4]";
|
3209 |
|
|
}
|
3210 |
|
|
[(set (attr "neon_type")
|
3211 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3212 |
|
|
(const_string "neon_fp_vmla_qqq_scalar")
|
3213 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3214 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")
|
3215 |
|
|
(const_string "neon_mla_qqq_32_qqd_32_scalar"))))]
|
3216 |
|
|
)
|
3217 |
|
|
|
3218 |
|
|
(define_insn "neon_vmlsl_lane"
|
3219 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3220 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
3221 |
|
|
(match_operand:VMDI 2 "s_register_operand" "w")
|
3222 |
|
|
(match_operand:VMDI 3 "s_register_operand"
|
3223 |
|
|
"")
|
3224 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3225 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3226 |
|
|
UNSPEC_VMLSL_LANE))]
|
3227 |
|
|
"TARGET_NEON"
|
3228 |
|
|
{
|
3229 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3230 |
|
|
return "vmlsl.%T5%#\t%q0, %P2, %P3[%c4]";
|
3231 |
|
|
}
|
3232 |
|
|
[(set (attr "neon_type")
|
3233 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3234 |
|
|
(const_string "neon_mla_ddd_16_scalar_qdd_32_16_long_scalar")
|
3235 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
3236 |
|
|
)
|
3237 |
|
|
|
3238 |
|
|
(define_insn "neon_vqdmlsl_lane"
|
3239 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3240 |
|
|
(unspec: [(match_operand: 1 "s_register_operand" "0")
|
3241 |
|
|
(match_operand:VMDI 2 "s_register_operand" "w")
|
3242 |
|
|
(match_operand:VMDI 3 "s_register_operand"
|
3243 |
|
|
"")
|
3244 |
|
|
(match_operand:SI 4 "immediate_operand" "i")
|
3245 |
|
|
(match_operand:SI 5 "immediate_operand" "i")]
|
3246 |
|
|
UNSPEC_VQDMLSL_LANE))]
|
3247 |
|
|
"TARGET_NEON"
|
3248 |
|
|
{
|
3249 |
|
|
neon_lane_bounds (operands[4], 0, GET_MODE_NUNITS (mode));
|
3250 |
|
|
return "vqdmlsl.\t%q0, %P2, %P3[%c4]";
|
3251 |
|
|
}
|
3252 |
|
|
[(set (attr "neon_type")
|
3253 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3254 |
|
|
(const_string "neon_mla_ddd_16_scalar_qdd_32_16_long_scalar")
|
3255 |
|
|
(const_string "neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long")))]
|
3256 |
|
|
)
|
3257 |
|
|
|
3258 |
|
|
; FIXME: For the "_n" multiply/multiply-accumulate insns, we copy a value in a
|
3259 |
|
|
; core register into a temp register, then use a scalar taken from that. This
|
3260 |
|
|
; isn't an optimal solution if e.g. the scalar has just been read from memory
|
3261 |
|
|
; or extracted from another vector. The latter case it's currently better to
|
3262 |
|
|
; use the "_lane" variant, and the former case can probably be implemented
|
3263 |
|
|
; using vld1_lane, but that hasn't been done yet.
|
3264 |
|
|
|
3265 |
|
|
(define_expand "neon_vmul_n"
|
3266 |
|
|
[(match_operand:VMD 0 "s_register_operand" "")
|
3267 |
|
|
(match_operand:VMD 1 "s_register_operand" "")
|
3268 |
|
|
(match_operand: 2 "s_register_operand" "")
|
3269 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
3270 |
|
|
"TARGET_NEON"
|
3271 |
|
|
{
|
3272 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3273 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[2], tmp, const0_rtx));
|
3274 |
|
|
emit_insn (gen_neon_vmul_lane (operands[0], operands[1], tmp,
|
3275 |
|
|
const0_rtx, const0_rtx));
|
3276 |
|
|
DONE;
|
3277 |
|
|
})
|
3278 |
|
|
|
3279 |
|
|
(define_expand "neon_vmul_n"
|
3280 |
|
|
[(match_operand:VMQ 0 "s_register_operand" "")
|
3281 |
|
|
(match_operand:VMQ 1 "s_register_operand" "")
|
3282 |
|
|
(match_operand: 2 "s_register_operand" "")
|
3283 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
3284 |
|
|
"TARGET_NEON"
|
3285 |
|
|
{
|
3286 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3287 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[2], tmp, const0_rtx));
|
3288 |
|
|
emit_insn (gen_neon_vmul_lane (operands[0], operands[1], tmp,
|
3289 |
|
|
const0_rtx, const0_rtx));
|
3290 |
|
|
DONE;
|
3291 |
|
|
})
|
3292 |
|
|
|
3293 |
|
|
(define_expand "neon_vmull_n"
|
3294 |
|
|
[(match_operand: 0 "s_register_operand" "")
|
3295 |
|
|
(match_operand:VMDI 1 "s_register_operand" "")
|
3296 |
|
|
(match_operand: 2 "s_register_operand" "")
|
3297 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
3298 |
|
|
"TARGET_NEON"
|
3299 |
|
|
{
|
3300 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3301 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[2], tmp, const0_rtx));
|
3302 |
|
|
emit_insn (gen_neon_vmull_lane (operands[0], operands[1], tmp,
|
3303 |
|
|
const0_rtx, operands[3]));
|
3304 |
|
|
DONE;
|
3305 |
|
|
})
|
3306 |
|
|
|
3307 |
|
|
(define_expand "neon_vqdmull_n"
|
3308 |
|
|
[(match_operand: 0 "s_register_operand" "")
|
3309 |
|
|
(match_operand:VMDI 1 "s_register_operand" "")
|
3310 |
|
|
(match_operand: 2 "s_register_operand" "")
|
3311 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
3312 |
|
|
"TARGET_NEON"
|
3313 |
|
|
{
|
3314 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3315 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[2], tmp, const0_rtx));
|
3316 |
|
|
emit_insn (gen_neon_vqdmull_lane (operands[0], operands[1], tmp,
|
3317 |
|
|
const0_rtx, const0_rtx));
|
3318 |
|
|
DONE;
|
3319 |
|
|
})
|
3320 |
|
|
|
3321 |
|
|
(define_expand "neon_vqdmulh_n"
|
3322 |
|
|
[(match_operand:VMDI 0 "s_register_operand" "")
|
3323 |
|
|
(match_operand:VMDI 1 "s_register_operand" "")
|
3324 |
|
|
(match_operand: 2 "s_register_operand" "")
|
3325 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
3326 |
|
|
"TARGET_NEON"
|
3327 |
|
|
{
|
3328 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3329 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[2], tmp, const0_rtx));
|
3330 |
|
|
emit_insn (gen_neon_vqdmulh_lane (operands[0], operands[1], tmp,
|
3331 |
|
|
const0_rtx, operands[3]));
|
3332 |
|
|
DONE;
|
3333 |
|
|
})
|
3334 |
|
|
|
3335 |
|
|
(define_expand "neon_vqdmulh_n"
|
3336 |
|
|
[(match_operand:VMQI 0 "s_register_operand" "")
|
3337 |
|
|
(match_operand:VMQI 1 "s_register_operand" "")
|
3338 |
|
|
(match_operand: 2 "s_register_operand" "")
|
3339 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
3340 |
|
|
"TARGET_NEON"
|
3341 |
|
|
{
|
3342 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3343 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[2], tmp, const0_rtx));
|
3344 |
|
|
emit_insn (gen_neon_vqdmulh_lane (operands[0], operands[1], tmp,
|
3345 |
|
|
const0_rtx, operands[3]));
|
3346 |
|
|
DONE;
|
3347 |
|
|
})
|
3348 |
|
|
|
3349 |
|
|
(define_expand "neon_vmla_n"
|
3350 |
|
|
[(match_operand:VMD 0 "s_register_operand" "")
|
3351 |
|
|
(match_operand:VMD 1 "s_register_operand" "")
|
3352 |
|
|
(match_operand:VMD 2 "s_register_operand" "")
|
3353 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3354 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3355 |
|
|
"TARGET_NEON"
|
3356 |
|
|
{
|
3357 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3358 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3359 |
|
|
emit_insn (gen_neon_vmla_lane (operands[0], operands[1], operands[2],
|
3360 |
|
|
tmp, const0_rtx, operands[4]));
|
3361 |
|
|
DONE;
|
3362 |
|
|
})
|
3363 |
|
|
|
3364 |
|
|
(define_expand "neon_vmla_n"
|
3365 |
|
|
[(match_operand:VMQ 0 "s_register_operand" "")
|
3366 |
|
|
(match_operand:VMQ 1 "s_register_operand" "")
|
3367 |
|
|
(match_operand:VMQ 2 "s_register_operand" "")
|
3368 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3369 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3370 |
|
|
"TARGET_NEON"
|
3371 |
|
|
{
|
3372 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3373 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3374 |
|
|
emit_insn (gen_neon_vmla_lane (operands[0], operands[1], operands[2],
|
3375 |
|
|
tmp, const0_rtx, operands[4]));
|
3376 |
|
|
DONE;
|
3377 |
|
|
})
|
3378 |
|
|
|
3379 |
|
|
(define_expand "neon_vmlal_n"
|
3380 |
|
|
[(match_operand: 0 "s_register_operand" "")
|
3381 |
|
|
(match_operand: 1 "s_register_operand" "")
|
3382 |
|
|
(match_operand:VMDI 2 "s_register_operand" "")
|
3383 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3384 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3385 |
|
|
"TARGET_NEON"
|
3386 |
|
|
{
|
3387 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3388 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3389 |
|
|
emit_insn (gen_neon_vmlal_lane (operands[0], operands[1], operands[2],
|
3390 |
|
|
tmp, const0_rtx, operands[4]));
|
3391 |
|
|
DONE;
|
3392 |
|
|
})
|
3393 |
|
|
|
3394 |
|
|
(define_expand "neon_vqdmlal_n"
|
3395 |
|
|
[(match_operand: 0 "s_register_operand" "")
|
3396 |
|
|
(match_operand: 1 "s_register_operand" "")
|
3397 |
|
|
(match_operand:VMDI 2 "s_register_operand" "")
|
3398 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3399 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3400 |
|
|
"TARGET_NEON"
|
3401 |
|
|
{
|
3402 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3403 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3404 |
|
|
emit_insn (gen_neon_vqdmlal_lane (operands[0], operands[1], operands[2],
|
3405 |
|
|
tmp, const0_rtx, operands[4]));
|
3406 |
|
|
DONE;
|
3407 |
|
|
})
|
3408 |
|
|
|
3409 |
|
|
(define_expand "neon_vmls_n"
|
3410 |
|
|
[(match_operand:VMD 0 "s_register_operand" "")
|
3411 |
|
|
(match_operand:VMD 1 "s_register_operand" "")
|
3412 |
|
|
(match_operand:VMD 2 "s_register_operand" "")
|
3413 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3414 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3415 |
|
|
"TARGET_NEON"
|
3416 |
|
|
{
|
3417 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3418 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3419 |
|
|
emit_insn (gen_neon_vmls_lane (operands[0], operands[1], operands[2],
|
3420 |
|
|
tmp, const0_rtx, operands[4]));
|
3421 |
|
|
DONE;
|
3422 |
|
|
})
|
3423 |
|
|
|
3424 |
|
|
(define_expand "neon_vmls_n"
|
3425 |
|
|
[(match_operand:VMQ 0 "s_register_operand" "")
|
3426 |
|
|
(match_operand:VMQ 1 "s_register_operand" "")
|
3427 |
|
|
(match_operand:VMQ 2 "s_register_operand" "")
|
3428 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3429 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3430 |
|
|
"TARGET_NEON"
|
3431 |
|
|
{
|
3432 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3433 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3434 |
|
|
emit_insn (gen_neon_vmls_lane (operands[0], operands[1], operands[2],
|
3435 |
|
|
tmp, const0_rtx, operands[4]));
|
3436 |
|
|
DONE;
|
3437 |
|
|
})
|
3438 |
|
|
|
3439 |
|
|
(define_expand "neon_vmlsl_n"
|
3440 |
|
|
[(match_operand: 0 "s_register_operand" "")
|
3441 |
|
|
(match_operand: 1 "s_register_operand" "")
|
3442 |
|
|
(match_operand:VMDI 2 "s_register_operand" "")
|
3443 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3444 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3445 |
|
|
"TARGET_NEON"
|
3446 |
|
|
{
|
3447 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3448 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3449 |
|
|
emit_insn (gen_neon_vmlsl_lane (operands[0], operands[1], operands[2],
|
3450 |
|
|
tmp, const0_rtx, operands[4]));
|
3451 |
|
|
DONE;
|
3452 |
|
|
})
|
3453 |
|
|
|
3454 |
|
|
(define_expand "neon_vqdmlsl_n"
|
3455 |
|
|
[(match_operand: 0 "s_register_operand" "")
|
3456 |
|
|
(match_operand: 1 "s_register_operand" "")
|
3457 |
|
|
(match_operand:VMDI 2 "s_register_operand" "")
|
3458 |
|
|
(match_operand: 3 "s_register_operand" "")
|
3459 |
|
|
(match_operand:SI 4 "immediate_operand" "")]
|
3460 |
|
|
"TARGET_NEON"
|
3461 |
|
|
{
|
3462 |
|
|
rtx tmp = gen_reg_rtx (mode);
|
3463 |
|
|
emit_insn (gen_neon_vset_lane (tmp, operands[3], tmp, const0_rtx));
|
3464 |
|
|
emit_insn (gen_neon_vqdmlsl_lane (operands[0], operands[1], operands[2],
|
3465 |
|
|
tmp, const0_rtx, operands[4]));
|
3466 |
|
|
DONE;
|
3467 |
|
|
})
|
3468 |
|
|
|
3469 |
|
|
(define_insn "neon_vext"
|
3470 |
|
|
[(set (match_operand:VDQX 0 "s_register_operand" "=w")
|
3471 |
|
|
(unspec:VDQX [(match_operand:VDQX 1 "s_register_operand" "w")
|
3472 |
|
|
(match_operand:VDQX 2 "s_register_operand" "w")
|
3473 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3474 |
|
|
UNSPEC_VEXT))]
|
3475 |
|
|
"TARGET_NEON"
|
3476 |
|
|
{
|
3477 |
|
|
neon_const_bounds (operands[3], 0, GET_MODE_NUNITS (mode));
|
3478 |
|
|
return "vext.\t%0, %1, %2, %3";
|
3479 |
|
|
}
|
3480 |
|
|
[(set (attr "neon_type")
|
3481 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3482 |
|
|
(const_string "neon_bp_simple")
|
3483 |
|
|
(const_string "neon_bp_2cycle")))]
|
3484 |
|
|
)
|
3485 |
|
|
|
3486 |
|
|
(define_insn "neon_vrev64"
|
3487 |
|
|
[(set (match_operand:VDQ 0 "s_register_operand" "=w")
|
3488 |
|
|
(unspec:VDQ [(match_operand:VDQ 1 "s_register_operand" "w")
|
3489 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
3490 |
|
|
UNSPEC_VREV64))]
|
3491 |
|
|
"TARGET_NEON"
|
3492 |
|
|
"vrev64.\t%0, %1"
|
3493 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
3494 |
|
|
)
|
3495 |
|
|
|
3496 |
|
|
(define_insn "neon_vrev32"
|
3497 |
|
|
[(set (match_operand:VX 0 "s_register_operand" "=w")
|
3498 |
|
|
(unspec:VX [(match_operand:VX 1 "s_register_operand" "w")
|
3499 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
3500 |
|
|
UNSPEC_VREV32))]
|
3501 |
|
|
"TARGET_NEON"
|
3502 |
|
|
"vrev32.\t%0, %1"
|
3503 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
3504 |
|
|
)
|
3505 |
|
|
|
3506 |
|
|
(define_insn "neon_vrev16"
|
3507 |
|
|
[(set (match_operand:VE 0 "s_register_operand" "=w")
|
3508 |
|
|
(unspec:VE [(match_operand:VE 1 "s_register_operand" "w")
|
3509 |
|
|
(match_operand:SI 2 "immediate_operand" "i")]
|
3510 |
|
|
UNSPEC_VREV16))]
|
3511 |
|
|
"TARGET_NEON"
|
3512 |
|
|
"vrev16.\t%0, %1"
|
3513 |
|
|
[(set_attr "neon_type" "neon_bp_simple")]
|
3514 |
|
|
)
|
3515 |
|
|
|
3516 |
|
|
; vbsl_* intrinsics may compile to any of vbsl/vbif/vbit depending on register
|
3517 |
|
|
; allocation. For an intrinsic of form:
|
3518 |
|
|
; rD = vbsl_* (rS, rN, rM)
|
3519 |
|
|
; We can use any of:
|
3520 |
|
|
; vbsl rS, rN, rM (if D = S)
|
3521 |
|
|
; vbit rD, rN, rS (if D = M, so 1-bits in rS choose bits from rN, else rM)
|
3522 |
|
|
; vbif rD, rM, rS (if D = N, so 0-bits in rS choose bits from rM, else rN)
|
3523 |
|
|
|
3524 |
|
|
(define_insn "neon_vbsl_internal"
|
3525 |
|
|
[(set (match_operand:VDQX 0 "s_register_operand" "=w,w,w")
|
3526 |
|
|
(unspec:VDQX [(match_operand:VDQX 1 "s_register_operand" " 0,w,w")
|
3527 |
|
|
(match_operand:VDQX 2 "s_register_operand" " w,w,0")
|
3528 |
|
|
(match_operand:VDQX 3 "s_register_operand" " w,0,w")]
|
3529 |
|
|
UNSPEC_VBSL))]
|
3530 |
|
|
"TARGET_NEON"
|
3531 |
|
|
"@
|
3532 |
|
|
vbsl\t%0, %2, %3
|
3533 |
|
|
vbit\t%0, %2, %1
|
3534 |
|
|
vbif\t%0, %3, %1"
|
3535 |
|
|
[(set_attr "neon_type" "neon_int_1")]
|
3536 |
|
|
)
|
3537 |
|
|
|
3538 |
|
|
(define_expand "neon_vbsl"
|
3539 |
|
|
[(set (match_operand:VDQX 0 "s_register_operand" "")
|
3540 |
|
|
(unspec:VDQX [(match_operand: 1 "s_register_operand" "")
|
3541 |
|
|
(match_operand:VDQX 2 "s_register_operand" "")
|
3542 |
|
|
(match_operand:VDQX 3 "s_register_operand" "")]
|
3543 |
|
|
UNSPEC_VBSL))]
|
3544 |
|
|
"TARGET_NEON"
|
3545 |
|
|
{
|
3546 |
|
|
/* We can't alias operands together if they have different modes. */
|
3547 |
|
|
operands[1] = gen_lowpart (mode, operands[1]);
|
3548 |
|
|
})
|
3549 |
|
|
|
3550 |
|
|
(define_insn "neon_vshl"
|
3551 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3552 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
3553 |
|
|
(match_operand:VDQIX 2 "s_register_operand" "w")
|
3554 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3555 |
|
|
UNSPEC_VSHL))]
|
3556 |
|
|
"TARGET_NEON"
|
3557 |
|
|
"v%O3shl.%T3%#\t%0, %1, %2"
|
3558 |
|
|
[(set (attr "neon_type")
|
3559 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3560 |
|
|
(const_string "neon_vshl_ddd")
|
3561 |
|
|
(const_string "neon_shift_3")))]
|
3562 |
|
|
)
|
3563 |
|
|
|
3564 |
|
|
(define_insn "neon_vqshl"
|
3565 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3566 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
3567 |
|
|
(match_operand:VDQIX 2 "s_register_operand" "w")
|
3568 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3569 |
|
|
UNSPEC_VQSHL))]
|
3570 |
|
|
"TARGET_NEON"
|
3571 |
|
|
"vq%O3shl.%T3%#\t%0, %1, %2"
|
3572 |
|
|
[(set (attr "neon_type")
|
3573 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3574 |
|
|
(const_string "neon_shift_2")
|
3575 |
|
|
(const_string "neon_vqshl_vrshl_vqrshl_qqq")))]
|
3576 |
|
|
)
|
3577 |
|
|
|
3578 |
|
|
(define_insn "neon_vshr_n"
|
3579 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3580 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
3581 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3582 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3583 |
|
|
UNSPEC_VSHR_N))]
|
3584 |
|
|
"TARGET_NEON"
|
3585 |
|
|
{
|
3586 |
|
|
neon_const_bounds (operands[2], 1, neon_element_bits (mode) + 1);
|
3587 |
|
|
return "v%O3shr.%T3%#\t%0, %1, %2";
|
3588 |
|
|
}
|
3589 |
|
|
[(set_attr "neon_type" "neon_shift_1")]
|
3590 |
|
|
)
|
3591 |
|
|
|
3592 |
|
|
(define_insn "neon_vshrn_n"
|
3593 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3594 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
3595 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3596 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3597 |
|
|
UNSPEC_VSHRN_N))]
|
3598 |
|
|
"TARGET_NEON"
|
3599 |
|
|
{
|
3600 |
|
|
neon_const_bounds (operands[2], 1, neon_element_bits (mode) / 2 + 1);
|
3601 |
|
|
return "v%O3shrn.\t%P0, %q1, %2";
|
3602 |
|
|
}
|
3603 |
|
|
[(set_attr "neon_type" "neon_shift_1")]
|
3604 |
|
|
)
|
3605 |
|
|
|
3606 |
|
|
(define_insn "neon_vqshrn_n"
|
3607 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3608 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
3609 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3610 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3611 |
|
|
UNSPEC_VQSHRN_N))]
|
3612 |
|
|
"TARGET_NEON"
|
3613 |
|
|
{
|
3614 |
|
|
neon_const_bounds (operands[2], 1, neon_element_bits (mode) / 2 + 1);
|
3615 |
|
|
return "vq%O3shrn.%T3%#\t%P0, %q1, %2";
|
3616 |
|
|
}
|
3617 |
|
|
[(set_attr "neon_type" "neon_shift_2")]
|
3618 |
|
|
)
|
3619 |
|
|
|
3620 |
|
|
(define_insn "neon_vqshrun_n"
|
3621 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3622 |
|
|
(unspec: [(match_operand:VN 1 "s_register_operand" "w")
|
3623 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3624 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3625 |
|
|
UNSPEC_VQSHRUN_N))]
|
3626 |
|
|
"TARGET_NEON"
|
3627 |
|
|
{
|
3628 |
|
|
neon_const_bounds (operands[2], 1, neon_element_bits (mode) / 2 + 1);
|
3629 |
|
|
return "vq%O3shrun.%T3%#\t%P0, %q1, %2";
|
3630 |
|
|
}
|
3631 |
|
|
[(set_attr "neon_type" "neon_shift_2")]
|
3632 |
|
|
)
|
3633 |
|
|
|
3634 |
|
|
(define_insn "neon_vshl_n"
|
3635 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3636 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
3637 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3638 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3639 |
|
|
UNSPEC_VSHL_N))]
|
3640 |
|
|
"TARGET_NEON"
|
3641 |
|
|
{
|
3642 |
|
|
neon_const_bounds (operands[2], 0, neon_element_bits (mode));
|
3643 |
|
|
return "vshl.\t%0, %1, %2";
|
3644 |
|
|
}
|
3645 |
|
|
[(set_attr "neon_type" "neon_shift_1")]
|
3646 |
|
|
)
|
3647 |
|
|
|
3648 |
|
|
(define_insn "neon_vqshl_n"
|
3649 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3650 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
3651 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3652 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3653 |
|
|
UNSPEC_VQSHL_N))]
|
3654 |
|
|
"TARGET_NEON"
|
3655 |
|
|
{
|
3656 |
|
|
neon_const_bounds (operands[2], 0, neon_element_bits (mode));
|
3657 |
|
|
return "vqshl.%T3%#\t%0, %1, %2";
|
3658 |
|
|
}
|
3659 |
|
|
[(set_attr "neon_type" "neon_shift_2")]
|
3660 |
|
|
)
|
3661 |
|
|
|
3662 |
|
|
(define_insn "neon_vqshlu_n"
|
3663 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3664 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "w")
|
3665 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3666 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3667 |
|
|
UNSPEC_VQSHLU_N))]
|
3668 |
|
|
"TARGET_NEON"
|
3669 |
|
|
{
|
3670 |
|
|
neon_const_bounds (operands[2], 0, neon_element_bits (mode));
|
3671 |
|
|
return "vqshlu.%T3%#\t%0, %1, %2";
|
3672 |
|
|
}
|
3673 |
|
|
[(set_attr "neon_type" "neon_shift_2")]
|
3674 |
|
|
)
|
3675 |
|
|
|
3676 |
|
|
(define_insn "neon_vshll_n"
|
3677 |
|
|
[(set (match_operand: 0 "s_register_operand" "=w")
|
3678 |
|
|
(unspec: [(match_operand:VW 1 "s_register_operand" "w")
|
3679 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
3680 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3681 |
|
|
UNSPEC_VSHLL_N))]
|
3682 |
|
|
"TARGET_NEON"
|
3683 |
|
|
{
|
3684 |
|
|
/* The boundaries are: 0 < imm <= size. */
|
3685 |
|
|
neon_const_bounds (operands[2], 0, neon_element_bits (mode) + 1);
|
3686 |
|
|
return "vshll.%T3%#\t%q0, %P1, %2";
|
3687 |
|
|
}
|
3688 |
|
|
[(set_attr "neon_type" "neon_shift_1")]
|
3689 |
|
|
)
|
3690 |
|
|
|
3691 |
|
|
(define_insn "neon_vsra_n"
|
3692 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3693 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "0")
|
3694 |
|
|
(match_operand:VDQIX 2 "s_register_operand" "w")
|
3695 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
3696 |
|
|
(match_operand:SI 4 "immediate_operand" "i")]
|
3697 |
|
|
UNSPEC_VSRA_N))]
|
3698 |
|
|
"TARGET_NEON"
|
3699 |
|
|
{
|
3700 |
|
|
neon_const_bounds (operands[3], 1, neon_element_bits (mode) + 1);
|
3701 |
|
|
return "v%O4sra.%T4%#\t%0, %2, %3";
|
3702 |
|
|
}
|
3703 |
|
|
[(set_attr "neon_type" "neon_vsra_vrsra")]
|
3704 |
|
|
)
|
3705 |
|
|
|
3706 |
|
|
(define_insn "neon_vsri_n"
|
3707 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3708 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "0")
|
3709 |
|
|
(match_operand:VDQIX 2 "s_register_operand" "w")
|
3710 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3711 |
|
|
UNSPEC_VSRI))]
|
3712 |
|
|
"TARGET_NEON"
|
3713 |
|
|
{
|
3714 |
|
|
neon_const_bounds (operands[3], 1, neon_element_bits (mode) + 1);
|
3715 |
|
|
return "vsri.\t%0, %2, %3";
|
3716 |
|
|
}
|
3717 |
|
|
[(set (attr "neon_type")
|
3718 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3719 |
|
|
(const_string "neon_shift_1")
|
3720 |
|
|
(const_string "neon_shift_3")))]
|
3721 |
|
|
)
|
3722 |
|
|
|
3723 |
|
|
(define_insn "neon_vsli_n"
|
3724 |
|
|
[(set (match_operand:VDQIX 0 "s_register_operand" "=w")
|
3725 |
|
|
(unspec:VDQIX [(match_operand:VDQIX 1 "s_register_operand" "0")
|
3726 |
|
|
(match_operand:VDQIX 2 "s_register_operand" "w")
|
3727 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
3728 |
|
|
UNSPEC_VSLI))]
|
3729 |
|
|
"TARGET_NEON"
|
3730 |
|
|
{
|
3731 |
|
|
neon_const_bounds (operands[3], 0, neon_element_bits (mode));
|
3732 |
|
|
return "vsli.\t%0, %2, %3";
|
3733 |
|
|
}
|
3734 |
|
|
[(set (attr "neon_type")
|
3735 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3736 |
|
|
(const_string "neon_shift_1")
|
3737 |
|
|
(const_string "neon_shift_3")))]
|
3738 |
|
|
)
|
3739 |
|
|
|
3740 |
|
|
(define_insn "neon_vtbl1v8qi"
|
3741 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3742 |
|
|
(unspec:V8QI [(match_operand:V8QI 1 "s_register_operand" "w")
|
3743 |
|
|
(match_operand:V8QI 2 "s_register_operand" "w")]
|
3744 |
|
|
UNSPEC_VTBL))]
|
3745 |
|
|
"TARGET_NEON"
|
3746 |
|
|
"vtbl.8\t%P0, {%P1}, %P2"
|
3747 |
|
|
[(set_attr "neon_type" "neon_bp_2cycle")]
|
3748 |
|
|
)
|
3749 |
|
|
|
3750 |
|
|
(define_insn "neon_vtbl2v8qi"
|
3751 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3752 |
|
|
(unspec:V8QI [(match_operand:TI 1 "s_register_operand" "w")
|
3753 |
|
|
(match_operand:V8QI 2 "s_register_operand" "w")]
|
3754 |
|
|
UNSPEC_VTBL))]
|
3755 |
|
|
"TARGET_NEON"
|
3756 |
|
|
{
|
3757 |
|
|
rtx ops[4];
|
3758 |
|
|
int tabbase = REGNO (operands[1]);
|
3759 |
|
|
|
3760 |
|
|
ops[0] = operands[0];
|
3761 |
|
|
ops[1] = gen_rtx_REG (V8QImode, tabbase);
|
3762 |
|
|
ops[2] = gen_rtx_REG (V8QImode, tabbase + 2);
|
3763 |
|
|
ops[3] = operands[2];
|
3764 |
|
|
output_asm_insn ("vtbl.8\t%P0, {%P1, %P2}, %P3", ops);
|
3765 |
|
|
|
3766 |
|
|
return "";
|
3767 |
|
|
}
|
3768 |
|
|
[(set_attr "neon_type" "neon_bp_2cycle")]
|
3769 |
|
|
)
|
3770 |
|
|
|
3771 |
|
|
(define_insn "neon_vtbl3v8qi"
|
3772 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3773 |
|
|
(unspec:V8QI [(match_operand:EI 1 "s_register_operand" "w")
|
3774 |
|
|
(match_operand:V8QI 2 "s_register_operand" "w")]
|
3775 |
|
|
UNSPEC_VTBL))]
|
3776 |
|
|
"TARGET_NEON"
|
3777 |
|
|
{
|
3778 |
|
|
rtx ops[5];
|
3779 |
|
|
int tabbase = REGNO (operands[1]);
|
3780 |
|
|
|
3781 |
|
|
ops[0] = operands[0];
|
3782 |
|
|
ops[1] = gen_rtx_REG (V8QImode, tabbase);
|
3783 |
|
|
ops[2] = gen_rtx_REG (V8QImode, tabbase + 2);
|
3784 |
|
|
ops[3] = gen_rtx_REG (V8QImode, tabbase + 4);
|
3785 |
|
|
ops[4] = operands[2];
|
3786 |
|
|
output_asm_insn ("vtbl.8\t%P0, {%P1, %P2, %P3}, %P4", ops);
|
3787 |
|
|
|
3788 |
|
|
return "";
|
3789 |
|
|
}
|
3790 |
|
|
[(set_attr "neon_type" "neon_bp_3cycle")]
|
3791 |
|
|
)
|
3792 |
|
|
|
3793 |
|
|
(define_insn "neon_vtbl4v8qi"
|
3794 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3795 |
|
|
(unspec:V8QI [(match_operand:OI 1 "s_register_operand" "w")
|
3796 |
|
|
(match_operand:V8QI 2 "s_register_operand" "w")]
|
3797 |
|
|
UNSPEC_VTBL))]
|
3798 |
|
|
"TARGET_NEON"
|
3799 |
|
|
{
|
3800 |
|
|
rtx ops[6];
|
3801 |
|
|
int tabbase = REGNO (operands[1]);
|
3802 |
|
|
|
3803 |
|
|
ops[0] = operands[0];
|
3804 |
|
|
ops[1] = gen_rtx_REG (V8QImode, tabbase);
|
3805 |
|
|
ops[2] = gen_rtx_REG (V8QImode, tabbase + 2);
|
3806 |
|
|
ops[3] = gen_rtx_REG (V8QImode, tabbase + 4);
|
3807 |
|
|
ops[4] = gen_rtx_REG (V8QImode, tabbase + 6);
|
3808 |
|
|
ops[5] = operands[2];
|
3809 |
|
|
output_asm_insn ("vtbl.8\t%P0, {%P1, %P2, %P3, %P4}, %P5", ops);
|
3810 |
|
|
|
3811 |
|
|
return "";
|
3812 |
|
|
}
|
3813 |
|
|
[(set_attr "neon_type" "neon_bp_3cycle")]
|
3814 |
|
|
)
|
3815 |
|
|
|
3816 |
|
|
(define_insn "neon_vtbx1v8qi"
|
3817 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3818 |
|
|
(unspec:V8QI [(match_operand:V8QI 1 "s_register_operand" "0")
|
3819 |
|
|
(match_operand:V8QI 2 "s_register_operand" "w")
|
3820 |
|
|
(match_operand:V8QI 3 "s_register_operand" "w")]
|
3821 |
|
|
UNSPEC_VTBX))]
|
3822 |
|
|
"TARGET_NEON"
|
3823 |
|
|
"vtbx.8\t%P0, {%P2}, %P3"
|
3824 |
|
|
[(set_attr "neon_type" "neon_bp_2cycle")]
|
3825 |
|
|
)
|
3826 |
|
|
|
3827 |
|
|
(define_insn "neon_vtbx2v8qi"
|
3828 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3829 |
|
|
(unspec:V8QI [(match_operand:V8QI 1 "s_register_operand" "0")
|
3830 |
|
|
(match_operand:TI 2 "s_register_operand" "w")
|
3831 |
|
|
(match_operand:V8QI 3 "s_register_operand" "w")]
|
3832 |
|
|
UNSPEC_VTBX))]
|
3833 |
|
|
"TARGET_NEON"
|
3834 |
|
|
{
|
3835 |
|
|
rtx ops[4];
|
3836 |
|
|
int tabbase = REGNO (operands[2]);
|
3837 |
|
|
|
3838 |
|
|
ops[0] = operands[0];
|
3839 |
|
|
ops[1] = gen_rtx_REG (V8QImode, tabbase);
|
3840 |
|
|
ops[2] = gen_rtx_REG (V8QImode, tabbase + 2);
|
3841 |
|
|
ops[3] = operands[3];
|
3842 |
|
|
output_asm_insn ("vtbx.8\t%P0, {%P1, %P2}, %P3", ops);
|
3843 |
|
|
|
3844 |
|
|
return "";
|
3845 |
|
|
}
|
3846 |
|
|
[(set_attr "neon_type" "neon_bp_2cycle")]
|
3847 |
|
|
)
|
3848 |
|
|
|
3849 |
|
|
(define_insn "neon_vtbx3v8qi"
|
3850 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3851 |
|
|
(unspec:V8QI [(match_operand:V8QI 1 "s_register_operand" "0")
|
3852 |
|
|
(match_operand:EI 2 "s_register_operand" "w")
|
3853 |
|
|
(match_operand:V8QI 3 "s_register_operand" "w")]
|
3854 |
|
|
UNSPEC_VTBX))]
|
3855 |
|
|
"TARGET_NEON"
|
3856 |
|
|
{
|
3857 |
|
|
rtx ops[5];
|
3858 |
|
|
int tabbase = REGNO (operands[2]);
|
3859 |
|
|
|
3860 |
|
|
ops[0] = operands[0];
|
3861 |
|
|
ops[1] = gen_rtx_REG (V8QImode, tabbase);
|
3862 |
|
|
ops[2] = gen_rtx_REG (V8QImode, tabbase + 2);
|
3863 |
|
|
ops[3] = gen_rtx_REG (V8QImode, tabbase + 4);
|
3864 |
|
|
ops[4] = operands[3];
|
3865 |
|
|
output_asm_insn ("vtbx.8\t%P0, {%P1, %P2, %P3}, %P4", ops);
|
3866 |
|
|
|
3867 |
|
|
return "";
|
3868 |
|
|
}
|
3869 |
|
|
[(set_attr "neon_type" "neon_bp_3cycle")]
|
3870 |
|
|
)
|
3871 |
|
|
|
3872 |
|
|
(define_insn "neon_vtbx4v8qi"
|
3873 |
|
|
[(set (match_operand:V8QI 0 "s_register_operand" "=w")
|
3874 |
|
|
(unspec:V8QI [(match_operand:V8QI 1 "s_register_operand" "0")
|
3875 |
|
|
(match_operand:OI 2 "s_register_operand" "w")
|
3876 |
|
|
(match_operand:V8QI 3 "s_register_operand" "w")]
|
3877 |
|
|
UNSPEC_VTBX))]
|
3878 |
|
|
"TARGET_NEON"
|
3879 |
|
|
{
|
3880 |
|
|
rtx ops[6];
|
3881 |
|
|
int tabbase = REGNO (operands[2]);
|
3882 |
|
|
|
3883 |
|
|
ops[0] = operands[0];
|
3884 |
|
|
ops[1] = gen_rtx_REG (V8QImode, tabbase);
|
3885 |
|
|
ops[2] = gen_rtx_REG (V8QImode, tabbase + 2);
|
3886 |
|
|
ops[3] = gen_rtx_REG (V8QImode, tabbase + 4);
|
3887 |
|
|
ops[4] = gen_rtx_REG (V8QImode, tabbase + 6);
|
3888 |
|
|
ops[5] = operands[3];
|
3889 |
|
|
output_asm_insn ("vtbx.8\t%P0, {%P1, %P2, %P3, %P4}, %P5", ops);
|
3890 |
|
|
|
3891 |
|
|
return "";
|
3892 |
|
|
}
|
3893 |
|
|
[(set_attr "neon_type" "neon_bp_3cycle")]
|
3894 |
|
|
)
|
3895 |
|
|
|
3896 |
|
|
(define_insn "neon_vtrn_internal"
|
3897 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
3898 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "0")]
|
3899 |
|
|
UNSPEC_VTRN1))
|
3900 |
|
|
(set (match_operand:VDQW 2 "s_register_operand" "=w")
|
3901 |
|
|
(unspec:VDQW [(match_operand:VDQW 3 "s_register_operand" "2")]
|
3902 |
|
|
UNSPEC_VTRN2))]
|
3903 |
|
|
"TARGET_NEON"
|
3904 |
|
|
"vtrn.\t%0, %2"
|
3905 |
|
|
[(set (attr "neon_type")
|
3906 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3907 |
|
|
(const_string "neon_bp_simple")
|
3908 |
|
|
(const_string "neon_bp_3cycle")))]
|
3909 |
|
|
)
|
3910 |
|
|
|
3911 |
|
|
(define_expand "neon_vtrn"
|
3912 |
|
|
[(match_operand:SI 0 "s_register_operand" "r")
|
3913 |
|
|
(match_operand:VDQW 1 "s_register_operand" "w")
|
3914 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")]
|
3915 |
|
|
"TARGET_NEON"
|
3916 |
|
|
{
|
3917 |
|
|
neon_emit_pair_result_insn (mode, gen_neon_vtrn_internal,
|
3918 |
|
|
operands[0], operands[1], operands[2]);
|
3919 |
|
|
DONE;
|
3920 |
|
|
})
|
3921 |
|
|
|
3922 |
|
|
(define_insn "neon_vzip_internal"
|
3923 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
3924 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "0")]
|
3925 |
|
|
UNSPEC_VZIP1))
|
3926 |
|
|
(set (match_operand:VDQW 2 "s_register_operand" "=w")
|
3927 |
|
|
(unspec:VDQW [(match_operand:VDQW 3 "s_register_operand" "2")]
|
3928 |
|
|
UNSPEC_VZIP2))]
|
3929 |
|
|
"TARGET_NEON"
|
3930 |
|
|
"vzip.\t%0, %2"
|
3931 |
|
|
[(set (attr "neon_type")
|
3932 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3933 |
|
|
(const_string "neon_bp_simple")
|
3934 |
|
|
(const_string "neon_bp_3cycle")))]
|
3935 |
|
|
)
|
3936 |
|
|
|
3937 |
|
|
(define_expand "neon_vzip"
|
3938 |
|
|
[(match_operand:SI 0 "s_register_operand" "r")
|
3939 |
|
|
(match_operand:VDQW 1 "s_register_operand" "w")
|
3940 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")]
|
3941 |
|
|
"TARGET_NEON"
|
3942 |
|
|
{
|
3943 |
|
|
neon_emit_pair_result_insn (mode, gen_neon_vzip_internal,
|
3944 |
|
|
operands[0], operands[1], operands[2]);
|
3945 |
|
|
DONE;
|
3946 |
|
|
})
|
3947 |
|
|
|
3948 |
|
|
(define_insn "neon_vuzp_internal"
|
3949 |
|
|
[(set (match_operand:VDQW 0 "s_register_operand" "=w")
|
3950 |
|
|
(unspec:VDQW [(match_operand:VDQW 1 "s_register_operand" "0")]
|
3951 |
|
|
UNSPEC_VUZP1))
|
3952 |
|
|
(set (match_operand:VDQW 2 "s_register_operand" "=w")
|
3953 |
|
|
(unspec:VDQW [(match_operand:VDQW 3 "s_register_operand" "2")]
|
3954 |
|
|
UNSPEC_VUZP2))]
|
3955 |
|
|
"TARGET_NEON"
|
3956 |
|
|
"vuzp.\t%0, %2"
|
3957 |
|
|
[(set (attr "neon_type")
|
3958 |
|
|
(if_then_else (ne (symbol_ref "") (const_int 0))
|
3959 |
|
|
(const_string "neon_bp_simple")
|
3960 |
|
|
(const_string "neon_bp_3cycle")))]
|
3961 |
|
|
)
|
3962 |
|
|
|
3963 |
|
|
(define_expand "neon_vuzp"
|
3964 |
|
|
[(match_operand:SI 0 "s_register_operand" "r")
|
3965 |
|
|
(match_operand:VDQW 1 "s_register_operand" "w")
|
3966 |
|
|
(match_operand:VDQW 2 "s_register_operand" "w")]
|
3967 |
|
|
"TARGET_NEON"
|
3968 |
|
|
{
|
3969 |
|
|
neon_emit_pair_result_insn (mode, gen_neon_vuzp_internal,
|
3970 |
|
|
operands[0], operands[1], operands[2]);
|
3971 |
|
|
DONE;
|
3972 |
|
|
})
|
3973 |
|
|
|
3974 |
|
|
(define_expand "neon_vreinterpretv8qi"
|
3975 |
|
|
[(match_operand:V8QI 0 "s_register_operand" "")
|
3976 |
|
|
(match_operand:VDX 1 "s_register_operand" "")]
|
3977 |
|
|
"TARGET_NEON"
|
3978 |
|
|
{
|
3979 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
3980 |
|
|
DONE;
|
3981 |
|
|
})
|
3982 |
|
|
|
3983 |
|
|
(define_expand "neon_vreinterpretv4hi"
|
3984 |
|
|
[(match_operand:V4HI 0 "s_register_operand" "")
|
3985 |
|
|
(match_operand:VDX 1 "s_register_operand" "")]
|
3986 |
|
|
"TARGET_NEON"
|
3987 |
|
|
{
|
3988 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
3989 |
|
|
DONE;
|
3990 |
|
|
})
|
3991 |
|
|
|
3992 |
|
|
(define_expand "neon_vreinterpretv2si"
|
3993 |
|
|
[(match_operand:V2SI 0 "s_register_operand" "")
|
3994 |
|
|
(match_operand:VDX 1 "s_register_operand" "")]
|
3995 |
|
|
"TARGET_NEON"
|
3996 |
|
|
{
|
3997 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
3998 |
|
|
DONE;
|
3999 |
|
|
})
|
4000 |
|
|
|
4001 |
|
|
(define_expand "neon_vreinterpretv2sf"
|
4002 |
|
|
[(match_operand:V2SF 0 "s_register_operand" "")
|
4003 |
|
|
(match_operand:VDX 1 "s_register_operand" "")]
|
4004 |
|
|
"TARGET_NEON"
|
4005 |
|
|
{
|
4006 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
4007 |
|
|
DONE;
|
4008 |
|
|
})
|
4009 |
|
|
|
4010 |
|
|
(define_expand "neon_vreinterpretdi"
|
4011 |
|
|
[(match_operand:DI 0 "s_register_operand" "")
|
4012 |
|
|
(match_operand:VDX 1 "s_register_operand" "")]
|
4013 |
|
|
"TARGET_NEON"
|
4014 |
|
|
{
|
4015 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
4016 |
|
|
DONE;
|
4017 |
|
|
})
|
4018 |
|
|
|
4019 |
|
|
(define_expand "neon_vreinterpretv16qi"
|
4020 |
|
|
[(match_operand:V16QI 0 "s_register_operand" "")
|
4021 |
|
|
(match_operand:VQX 1 "s_register_operand" "")]
|
4022 |
|
|
"TARGET_NEON"
|
4023 |
|
|
{
|
4024 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
4025 |
|
|
DONE;
|
4026 |
|
|
})
|
4027 |
|
|
|
4028 |
|
|
(define_expand "neon_vreinterpretv8hi"
|
4029 |
|
|
[(match_operand:V8HI 0 "s_register_operand" "")
|
4030 |
|
|
(match_operand:VQX 1 "s_register_operand" "")]
|
4031 |
|
|
"TARGET_NEON"
|
4032 |
|
|
{
|
4033 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
4034 |
|
|
DONE;
|
4035 |
|
|
})
|
4036 |
|
|
|
4037 |
|
|
(define_expand "neon_vreinterpretv4si"
|
4038 |
|
|
[(match_operand:V4SI 0 "s_register_operand" "")
|
4039 |
|
|
(match_operand:VQX 1 "s_register_operand" "")]
|
4040 |
|
|
"TARGET_NEON"
|
4041 |
|
|
{
|
4042 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
4043 |
|
|
DONE;
|
4044 |
|
|
})
|
4045 |
|
|
|
4046 |
|
|
(define_expand "neon_vreinterpretv4sf"
|
4047 |
|
|
[(match_operand:V4SF 0 "s_register_operand" "")
|
4048 |
|
|
(match_operand:VQX 1 "s_register_operand" "")]
|
4049 |
|
|
"TARGET_NEON"
|
4050 |
|
|
{
|
4051 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
4052 |
|
|
DONE;
|
4053 |
|
|
})
|
4054 |
|
|
|
4055 |
|
|
(define_expand "neon_vreinterpretv2di"
|
4056 |
|
|
[(match_operand:V2DI 0 "s_register_operand" "")
|
4057 |
|
|
(match_operand:VQX 1 "s_register_operand" "")]
|
4058 |
|
|
"TARGET_NEON"
|
4059 |
|
|
{
|
4060 |
|
|
neon_reinterpret (operands[0], operands[1]);
|
4061 |
|
|
DONE;
|
4062 |
|
|
})
|
4063 |
|
|
|
4064 |
|
|
(define_insn "neon_vld1"
|
4065 |
|
|
[(set (match_operand:VDQX 0 "s_register_operand" "=w")
|
4066 |
|
|
(unspec:VDQX [(mem:VDQX (match_operand:SI 1 "s_register_operand" "r"))]
|
4067 |
|
|
UNSPEC_VLD1))]
|
4068 |
|
|
"TARGET_NEON"
|
4069 |
|
|
"vld1.\t%h0, [%1]"
|
4070 |
|
|
[(set_attr "neon_type" "neon_vld1_1_2_regs")]
|
4071 |
|
|
)
|
4072 |
|
|
|
4073 |
|
|
(define_insn "neon_vld1_lane"
|
4074 |
|
|
[(set (match_operand:VDX 0 "s_register_operand" "=w")
|
4075 |
|
|
(unspec:VDX [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4076 |
|
|
(match_operand:VDX 2 "s_register_operand" "0")
|
4077 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
4078 |
|
|
UNSPEC_VLD1_LANE))]
|
4079 |
|
|
"TARGET_NEON"
|
4080 |
|
|
{
|
4081 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4082 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4083 |
|
|
if (lane < 0 || lane >= max)
|
4084 |
|
|
error ("lane out of range");
|
4085 |
|
|
if (max == 1)
|
4086 |
|
|
return "vld1.\t%P0, [%1]";
|
4087 |
|
|
else
|
4088 |
|
|
return "vld1.\t{%P0[%c3]}, [%1]";
|
4089 |
|
|
}
|
4090 |
|
|
[(set (attr "neon_type")
|
4091 |
|
|
(if_then_else (eq (const_string "") (const_int 2))
|
4092 |
|
|
(const_string "neon_vld1_1_2_regs")
|
4093 |
|
|
(const_string "neon_vld1_vld2_lane")))]
|
4094 |
|
|
)
|
4095 |
|
|
|
4096 |
|
|
(define_insn "neon_vld1_lane"
|
4097 |
|
|
[(set (match_operand:VQX 0 "s_register_operand" "=w")
|
4098 |
|
|
(unspec:VQX [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4099 |
|
|
(match_operand:VQX 2 "s_register_operand" "0")
|
4100 |
|
|
(match_operand:SI 3 "immediate_operand" "i")]
|
4101 |
|
|
UNSPEC_VLD1_LANE))]
|
4102 |
|
|
"TARGET_NEON"
|
4103 |
|
|
{
|
4104 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4105 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4106 |
|
|
int regno = REGNO (operands[0]);
|
4107 |
|
|
if (lane < 0 || lane >= max)
|
4108 |
|
|
error ("lane out of range");
|
4109 |
|
|
else if (lane >= max / 2)
|
4110 |
|
|
{
|
4111 |
|
|
lane -= max / 2;
|
4112 |
|
|
regno += 2;
|
4113 |
|
|
operands[3] = GEN_INT (lane);
|
4114 |
|
|
}
|
4115 |
|
|
operands[0] = gen_rtx_REG (mode, regno);
|
4116 |
|
|
if (max == 2)
|
4117 |
|
|
return "vld1.\t%P0, [%1]";
|
4118 |
|
|
else
|
4119 |
|
|
return "vld1.\t{%P0[%c3]}, [%1]";
|
4120 |
|
|
}
|
4121 |
|
|
[(set (attr "neon_type")
|
4122 |
|
|
(if_then_else (eq (const_string "") (const_int 2))
|
4123 |
|
|
(const_string "neon_vld1_1_2_regs")
|
4124 |
|
|
(const_string "neon_vld1_vld2_lane")))]
|
4125 |
|
|
)
|
4126 |
|
|
|
4127 |
|
|
(define_insn "neon_vld1_dup"
|
4128 |
|
|
[(set (match_operand:VDX 0 "s_register_operand" "=w")
|
4129 |
|
|
(unspec:VDX [(mem: (match_operand:SI 1 "s_register_operand" "r"))]
|
4130 |
|
|
UNSPEC_VLD1_DUP))]
|
4131 |
|
|
"TARGET_NEON"
|
4132 |
|
|
{
|
4133 |
|
|
if (GET_MODE_NUNITS (mode) > 1)
|
4134 |
|
|
return "vld1.\t{%P0[]}, [%1]";
|
4135 |
|
|
else
|
4136 |
|
|
return "vld1.\t%h0, [%1]";
|
4137 |
|
|
}
|
4138 |
|
|
[(set (attr "neon_type")
|
4139 |
|
|
(if_then_else (gt (const_string "") (const_string "1"))
|
4140 |
|
|
(const_string "neon_vld2_2_regs_vld1_vld2_all_lanes")
|
4141 |
|
|
(const_string "neon_vld1_1_2_regs")))]
|
4142 |
|
|
)
|
4143 |
|
|
|
4144 |
|
|
(define_insn "neon_vld1_dup"
|
4145 |
|
|
[(set (match_operand:VQX 0 "s_register_operand" "=w")
|
4146 |
|
|
(unspec:VQX [(mem: (match_operand:SI 1 "s_register_operand" "r"))]
|
4147 |
|
|
UNSPEC_VLD1_DUP))]
|
4148 |
|
|
"TARGET_NEON"
|
4149 |
|
|
{
|
4150 |
|
|
if (GET_MODE_NUNITS (mode) > 2)
|
4151 |
|
|
return "vld1.\t{%e0[], %f0[]}, [%1]";
|
4152 |
|
|
else
|
4153 |
|
|
return "vld1.\t%h0, [%1]";
|
4154 |
|
|
}
|
4155 |
|
|
[(set (attr "neon_type")
|
4156 |
|
|
(if_then_else (gt (const_string "") (const_string "1"))
|
4157 |
|
|
(const_string "neon_vld2_2_regs_vld1_vld2_all_lanes")
|
4158 |
|
|
(const_string "neon_vld1_1_2_regs")))]
|
4159 |
|
|
)
|
4160 |
|
|
|
4161 |
|
|
(define_insn "neon_vst1"
|
4162 |
|
|
[(set (mem:VDQX (match_operand:SI 0 "s_register_operand" "r"))
|
4163 |
|
|
(unspec:VDQX [(match_operand:VDQX 1 "s_register_operand" "w")]
|
4164 |
|
|
UNSPEC_VST1))]
|
4165 |
|
|
"TARGET_NEON"
|
4166 |
|
|
"vst1.\t%h1, [%0]"
|
4167 |
|
|
[(set_attr "neon_type" "neon_vst1_1_2_regs_vst2_2_regs")])
|
4168 |
|
|
|
4169 |
|
|
(define_insn "neon_vst1_lane"
|
4170 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4171 |
|
|
(vec_select:
|
4172 |
|
|
(match_operand:VDX 1 "s_register_operand" "w")
|
4173 |
|
|
(parallel [(match_operand:SI 2 "neon_lane_number" "i")])))]
|
4174 |
|
|
"TARGET_NEON"
|
4175 |
|
|
{
|
4176 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4177 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4178 |
|
|
if (lane < 0 || lane >= max)
|
4179 |
|
|
error ("lane out of range");
|
4180 |
|
|
if (max == 1)
|
4181 |
|
|
return "vst1.\t{%P1}, [%0]";
|
4182 |
|
|
else
|
4183 |
|
|
return "vst1.\t{%P1[%c2]}, [%0]";
|
4184 |
|
|
}
|
4185 |
|
|
[(set (attr "neon_type")
|
4186 |
|
|
(if_then_else (eq (const_string "") (const_int 1))
|
4187 |
|
|
(const_string "neon_vst1_1_2_regs_vst2_2_regs")
|
4188 |
|
|
(const_string "neon_vst1_vst2_lane")))])
|
4189 |
|
|
|
4190 |
|
|
(define_insn "neon_vst1_lane"
|
4191 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4192 |
|
|
(vec_select:
|
4193 |
|
|
(match_operand:VQX 1 "s_register_operand" "w")
|
4194 |
|
|
(parallel [(match_operand:SI 2 "neon_lane_number" "i")])))]
|
4195 |
|
|
"TARGET_NEON"
|
4196 |
|
|
{
|
4197 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4198 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4199 |
|
|
int regno = REGNO (operands[1]);
|
4200 |
|
|
if (lane < 0 || lane >= max)
|
4201 |
|
|
error ("lane out of range");
|
4202 |
|
|
else if (lane >= max / 2)
|
4203 |
|
|
{
|
4204 |
|
|
lane -= max / 2;
|
4205 |
|
|
regno += 2;
|
4206 |
|
|
operands[2] = GEN_INT (lane);
|
4207 |
|
|
}
|
4208 |
|
|
operands[1] = gen_rtx_REG (mode, regno);
|
4209 |
|
|
if (max == 2)
|
4210 |
|
|
return "vst1.\t{%P1}, [%0]";
|
4211 |
|
|
else
|
4212 |
|
|
return "vst1.\t{%P1[%c2]}, [%0]";
|
4213 |
|
|
}
|
4214 |
|
|
[(set_attr "neon_type" "neon_vst1_vst2_lane")]
|
4215 |
|
|
)
|
4216 |
|
|
|
4217 |
|
|
(define_insn "neon_vld2"
|
4218 |
|
|
[(set (match_operand:TI 0 "s_register_operand" "=w")
|
4219 |
|
|
(unspec:TI [(mem:TI (match_operand:SI 1 "s_register_operand" "r"))
|
4220 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4221 |
|
|
UNSPEC_VLD2))]
|
4222 |
|
|
"TARGET_NEON"
|
4223 |
|
|
{
|
4224 |
|
|
if ( == 64)
|
4225 |
|
|
return "vld1.64\t%h0, [%1]";
|
4226 |
|
|
else
|
4227 |
|
|
return "vld2.\t%h0, [%1]";
|
4228 |
|
|
}
|
4229 |
|
|
[(set (attr "neon_type")
|
4230 |
|
|
(if_then_else (eq (const_string "") (const_string "64"))
|
4231 |
|
|
(const_string "neon_vld1_1_2_regs")
|
4232 |
|
|
(const_string "neon_vld2_2_regs_vld1_vld2_all_lanes")))]
|
4233 |
|
|
)
|
4234 |
|
|
|
4235 |
|
|
(define_insn "neon_vld2"
|
4236 |
|
|
[(set (match_operand:OI 0 "s_register_operand" "=w")
|
4237 |
|
|
(unspec:OI [(mem:OI (match_operand:SI 1 "s_register_operand" "r"))
|
4238 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4239 |
|
|
UNSPEC_VLD2))]
|
4240 |
|
|
"TARGET_NEON"
|
4241 |
|
|
"vld2.\t%h0, [%1]"
|
4242 |
|
|
[(set_attr "neon_type" "neon_vld2_2_regs_vld1_vld2_all_lanes")])
|
4243 |
|
|
|
4244 |
|
|
(define_insn "neon_vld2_lane"
|
4245 |
|
|
[(set (match_operand:TI 0 "s_register_operand" "=w")
|
4246 |
|
|
(unspec:TI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4247 |
|
|
(match_operand:TI 2 "s_register_operand" "0")
|
4248 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
4249 |
|
|
(unspec:VD [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4250 |
|
|
UNSPEC_VLD2_LANE))]
|
4251 |
|
|
"TARGET_NEON"
|
4252 |
|
|
{
|
4253 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4254 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4255 |
|
|
int regno = REGNO (operands[0]);
|
4256 |
|
|
rtx ops[4];
|
4257 |
|
|
if (lane < 0 || lane >= max)
|
4258 |
|
|
error ("lane out of range");
|
4259 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4260 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2);
|
4261 |
|
|
ops[2] = operands[1];
|
4262 |
|
|
ops[3] = operands[3];
|
4263 |
|
|
output_asm_insn ("vld2.\t{%P0[%c3], %P1[%c3]}, [%2]", ops);
|
4264 |
|
|
return "";
|
4265 |
|
|
}
|
4266 |
|
|
[(set_attr "neon_type" "neon_vld1_vld2_lane")]
|
4267 |
|
|
)
|
4268 |
|
|
|
4269 |
|
|
(define_insn "neon_vld2_lane"
|
4270 |
|
|
[(set (match_operand:OI 0 "s_register_operand" "=w")
|
4271 |
|
|
(unspec:OI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4272 |
|
|
(match_operand:OI 2 "s_register_operand" "0")
|
4273 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
4274 |
|
|
(unspec:VMQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4275 |
|
|
UNSPEC_VLD2_LANE))]
|
4276 |
|
|
"TARGET_NEON"
|
4277 |
|
|
{
|
4278 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4279 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4280 |
|
|
int regno = REGNO (operands[0]);
|
4281 |
|
|
rtx ops[4];
|
4282 |
|
|
if (lane < 0 || lane >= max)
|
4283 |
|
|
error ("lane out of range");
|
4284 |
|
|
else if (lane >= max / 2)
|
4285 |
|
|
{
|
4286 |
|
|
lane -= max / 2;
|
4287 |
|
|
regno += 2;
|
4288 |
|
|
}
|
4289 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4290 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 4);
|
4291 |
|
|
ops[2] = operands[1];
|
4292 |
|
|
ops[3] = GEN_INT (lane);
|
4293 |
|
|
output_asm_insn ("vld2.\t{%P0[%c3], %P1[%c3]}, [%2]", ops);
|
4294 |
|
|
return "";
|
4295 |
|
|
}
|
4296 |
|
|
[(set_attr "neon_type" "neon_vld1_vld2_lane")]
|
4297 |
|
|
)
|
4298 |
|
|
|
4299 |
|
|
(define_insn "neon_vld2_dup"
|
4300 |
|
|
[(set (match_operand:TI 0 "s_register_operand" "=w")
|
4301 |
|
|
(unspec:TI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4302 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4303 |
|
|
UNSPEC_VLD2_DUP))]
|
4304 |
|
|
"TARGET_NEON"
|
4305 |
|
|
{
|
4306 |
|
|
if (GET_MODE_NUNITS (mode) > 1)
|
4307 |
|
|
return "vld2.\t{%e0[], %f0[]}, [%1]";
|
4308 |
|
|
else
|
4309 |
|
|
return "vld1.\t%h0, [%1]";
|
4310 |
|
|
}
|
4311 |
|
|
[(set (attr "neon_type")
|
4312 |
|
|
(if_then_else (gt (const_string "") (const_string "1"))
|
4313 |
|
|
(const_string "neon_vld2_2_regs_vld1_vld2_all_lanes")
|
4314 |
|
|
(const_string "neon_vld1_1_2_regs")))]
|
4315 |
|
|
)
|
4316 |
|
|
|
4317 |
|
|
(define_insn "neon_vst2"
|
4318 |
|
|
[(set (mem:TI (match_operand:SI 0 "s_register_operand" "r"))
|
4319 |
|
|
(unspec:TI [(match_operand:TI 1 "s_register_operand" "w")
|
4320 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4321 |
|
|
UNSPEC_VST2))]
|
4322 |
|
|
"TARGET_NEON"
|
4323 |
|
|
{
|
4324 |
|
|
if ( == 64)
|
4325 |
|
|
return "vst1.64\t%h1, [%0]";
|
4326 |
|
|
else
|
4327 |
|
|
return "vst2.\t%h1, [%0]";
|
4328 |
|
|
}
|
4329 |
|
|
[(set (attr "neon_type")
|
4330 |
|
|
(if_then_else (eq (const_string "") (const_string "64"))
|
4331 |
|
|
(const_string "neon_vst1_1_2_regs_vst2_2_regs")
|
4332 |
|
|
(const_string "neon_vst1_1_2_regs_vst2_2_regs")))]
|
4333 |
|
|
)
|
4334 |
|
|
|
4335 |
|
|
(define_insn "neon_vst2"
|
4336 |
|
|
[(set (mem:OI (match_operand:SI 0 "s_register_operand" "r"))
|
4337 |
|
|
(unspec:OI [(match_operand:OI 1 "s_register_operand" "w")
|
4338 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4339 |
|
|
UNSPEC_VST2))]
|
4340 |
|
|
"TARGET_NEON"
|
4341 |
|
|
"vst2.\t%h1, [%0]"
|
4342 |
|
|
[(set_attr "neon_type" "neon_vst1_1_2_regs_vst2_2_regs")]
|
4343 |
|
|
)
|
4344 |
|
|
|
4345 |
|
|
(define_insn "neon_vst2_lane"
|
4346 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4347 |
|
|
(unspec:
|
4348 |
|
|
[(match_operand:TI 1 "s_register_operand" "w")
|
4349 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
4350 |
|
|
(unspec:VD [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4351 |
|
|
UNSPEC_VST2_LANE))]
|
4352 |
|
|
"TARGET_NEON"
|
4353 |
|
|
{
|
4354 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4355 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4356 |
|
|
int regno = REGNO (operands[1]);
|
4357 |
|
|
rtx ops[4];
|
4358 |
|
|
if (lane < 0 || lane >= max)
|
4359 |
|
|
error ("lane out of range");
|
4360 |
|
|
ops[0] = operands[0];
|
4361 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4362 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 2);
|
4363 |
|
|
ops[3] = operands[2];
|
4364 |
|
|
output_asm_insn ("vst2.\t{%P1[%c3], %P2[%c3]}, [%0]", ops);
|
4365 |
|
|
return "";
|
4366 |
|
|
}
|
4367 |
|
|
[(set_attr "neon_type" "neon_vst1_vst2_lane")]
|
4368 |
|
|
)
|
4369 |
|
|
|
4370 |
|
|
(define_insn "neon_vst2_lane"
|
4371 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4372 |
|
|
(unspec:
|
4373 |
|
|
[(match_operand:OI 1 "s_register_operand" "w")
|
4374 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
4375 |
|
|
(unspec:VMQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4376 |
|
|
UNSPEC_VST2_LANE))]
|
4377 |
|
|
"TARGET_NEON"
|
4378 |
|
|
{
|
4379 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4380 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4381 |
|
|
int regno = REGNO (operands[1]);
|
4382 |
|
|
rtx ops[4];
|
4383 |
|
|
if (lane < 0 || lane >= max)
|
4384 |
|
|
error ("lane out of range");
|
4385 |
|
|
else if (lane >= max / 2)
|
4386 |
|
|
{
|
4387 |
|
|
lane -= max / 2;
|
4388 |
|
|
regno += 2;
|
4389 |
|
|
}
|
4390 |
|
|
ops[0] = operands[0];
|
4391 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4392 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4393 |
|
|
ops[3] = GEN_INT (lane);
|
4394 |
|
|
output_asm_insn ("vst2.\t{%P1[%c3], %P2[%c3]}, [%0]", ops);
|
4395 |
|
|
return "";
|
4396 |
|
|
}
|
4397 |
|
|
[(set_attr "neon_type" "neon_vst1_vst2_lane")]
|
4398 |
|
|
)
|
4399 |
|
|
|
4400 |
|
|
(define_insn "neon_vld3"
|
4401 |
|
|
[(set (match_operand:EI 0 "s_register_operand" "=w")
|
4402 |
|
|
(unspec:EI [(mem:EI (match_operand:SI 1 "s_register_operand" "r"))
|
4403 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4404 |
|
|
UNSPEC_VLD3))]
|
4405 |
|
|
"TARGET_NEON"
|
4406 |
|
|
{
|
4407 |
|
|
if ( == 64)
|
4408 |
|
|
return "vld1.64\t%h0, [%1]";
|
4409 |
|
|
else
|
4410 |
|
|
return "vld3.\t%h0, [%1]";
|
4411 |
|
|
}
|
4412 |
|
|
[(set (attr "neon_type")
|
4413 |
|
|
(if_then_else (eq (const_string "") (const_string "64"))
|
4414 |
|
|
(const_string "neon_vld1_1_2_regs")
|
4415 |
|
|
(const_string "neon_vld3_vld4")))]
|
4416 |
|
|
)
|
4417 |
|
|
|
4418 |
|
|
(define_expand "neon_vld3"
|
4419 |
|
|
[(match_operand:CI 0 "s_register_operand" "=w")
|
4420 |
|
|
(match_operand:SI 1 "s_register_operand" "+r")
|
4421 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4422 |
|
|
"TARGET_NEON"
|
4423 |
|
|
{
|
4424 |
|
|
emit_insn (gen_neon_vld3qa (operands[0], operands[0],
|
4425 |
|
|
operands[1], operands[1]));
|
4426 |
|
|
emit_insn (gen_neon_vld3qb (operands[0], operands[0],
|
4427 |
|
|
operands[1], operands[1]));
|
4428 |
|
|
DONE;
|
4429 |
|
|
})
|
4430 |
|
|
|
4431 |
|
|
(define_insn "neon_vld3qa"
|
4432 |
|
|
[(set (match_operand:CI 0 "s_register_operand" "=w")
|
4433 |
|
|
(unspec:CI [(mem:CI (match_operand:SI 3 "s_register_operand" "2"))
|
4434 |
|
|
(match_operand:CI 1 "s_register_operand" "0")
|
4435 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4436 |
|
|
UNSPEC_VLD3A))
|
4437 |
|
|
(set (match_operand:SI 2 "s_register_operand" "=r")
|
4438 |
|
|
(plus:SI (match_dup 3)
|
4439 |
|
|
(const_int 24)))]
|
4440 |
|
|
"TARGET_NEON"
|
4441 |
|
|
{
|
4442 |
|
|
int regno = REGNO (operands[0]);
|
4443 |
|
|
rtx ops[4];
|
4444 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4445 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 4);
|
4446 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 8);
|
4447 |
|
|
ops[3] = operands[2];
|
4448 |
|
|
output_asm_insn ("vld3.\t{%P0, %P1, %P2}, [%3]!", ops);
|
4449 |
|
|
return "";
|
4450 |
|
|
}
|
4451 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4")]
|
4452 |
|
|
)
|
4453 |
|
|
|
4454 |
|
|
(define_insn "neon_vld3qb"
|
4455 |
|
|
[(set (match_operand:CI 0 "s_register_operand" "=w")
|
4456 |
|
|
(unspec:CI [(mem:CI (match_operand:SI 3 "s_register_operand" "2"))
|
4457 |
|
|
(match_operand:CI 1 "s_register_operand" "0")
|
4458 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4459 |
|
|
UNSPEC_VLD3B))
|
4460 |
|
|
(set (match_operand:SI 2 "s_register_operand" "=r")
|
4461 |
|
|
(plus:SI (match_dup 3)
|
4462 |
|
|
(const_int 24)))]
|
4463 |
|
|
"TARGET_NEON"
|
4464 |
|
|
{
|
4465 |
|
|
int regno = REGNO (operands[0]);
|
4466 |
|
|
rtx ops[4];
|
4467 |
|
|
ops[0] = gen_rtx_REG (DImode, regno + 2);
|
4468 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 6);
|
4469 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 10);
|
4470 |
|
|
ops[3] = operands[2];
|
4471 |
|
|
output_asm_insn ("vld3.\t{%P0, %P1, %P2}, [%3]!", ops);
|
4472 |
|
|
return "";
|
4473 |
|
|
}
|
4474 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4")]
|
4475 |
|
|
)
|
4476 |
|
|
|
4477 |
|
|
(define_insn "neon_vld3_lane"
|
4478 |
|
|
[(set (match_operand:EI 0 "s_register_operand" "=w")
|
4479 |
|
|
(unspec:EI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4480 |
|
|
(match_operand:EI 2 "s_register_operand" "0")
|
4481 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
4482 |
|
|
(unspec:VD [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4483 |
|
|
UNSPEC_VLD3_LANE))]
|
4484 |
|
|
"TARGET_NEON"
|
4485 |
|
|
{
|
4486 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4487 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4488 |
|
|
int regno = REGNO (operands[0]);
|
4489 |
|
|
rtx ops[5];
|
4490 |
|
|
if (lane < 0 || lane >= max)
|
4491 |
|
|
error ("lane out of range");
|
4492 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4493 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2);
|
4494 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4495 |
|
|
ops[3] = operands[1];
|
4496 |
|
|
ops[4] = operands[3];
|
4497 |
|
|
output_asm_insn ("vld3.\t{%P0[%c4], %P1[%c4], %P2[%c4]}, [%3]",
|
4498 |
|
|
ops);
|
4499 |
|
|
return "";
|
4500 |
|
|
}
|
4501 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4_lane")]
|
4502 |
|
|
)
|
4503 |
|
|
|
4504 |
|
|
(define_insn "neon_vld3_lane"
|
4505 |
|
|
[(set (match_operand:CI 0 "s_register_operand" "=w")
|
4506 |
|
|
(unspec:CI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4507 |
|
|
(match_operand:CI 2 "s_register_operand" "0")
|
4508 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
4509 |
|
|
(unspec:VMQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4510 |
|
|
UNSPEC_VLD3_LANE))]
|
4511 |
|
|
"TARGET_NEON"
|
4512 |
|
|
{
|
4513 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4514 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4515 |
|
|
int regno = REGNO (operands[0]);
|
4516 |
|
|
rtx ops[5];
|
4517 |
|
|
if (lane < 0 || lane >= max)
|
4518 |
|
|
error ("lane out of range");
|
4519 |
|
|
else if (lane >= max / 2)
|
4520 |
|
|
{
|
4521 |
|
|
lane -= max / 2;
|
4522 |
|
|
regno += 2;
|
4523 |
|
|
}
|
4524 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4525 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 4);
|
4526 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 8);
|
4527 |
|
|
ops[3] = operands[1];
|
4528 |
|
|
ops[4] = GEN_INT (lane);
|
4529 |
|
|
output_asm_insn ("vld3.\t{%P0[%c4], %P1[%c4], %P2[%c4]}, [%3]",
|
4530 |
|
|
ops);
|
4531 |
|
|
return "";
|
4532 |
|
|
}
|
4533 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4_lane")]
|
4534 |
|
|
)
|
4535 |
|
|
|
4536 |
|
|
(define_insn "neon_vld3_dup"
|
4537 |
|
|
[(set (match_operand:EI 0 "s_register_operand" "=w")
|
4538 |
|
|
(unspec:EI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4539 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4540 |
|
|
UNSPEC_VLD3_DUP))]
|
4541 |
|
|
"TARGET_NEON"
|
4542 |
|
|
{
|
4543 |
|
|
if (GET_MODE_NUNITS (mode) > 1)
|
4544 |
|
|
{
|
4545 |
|
|
int regno = REGNO (operands[0]);
|
4546 |
|
|
rtx ops[4];
|
4547 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4548 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2);
|
4549 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4550 |
|
|
ops[3] = operands[1];
|
4551 |
|
|
output_asm_insn ("vld3.\t{%P0[], %P1[], %P2[]}, [%3]", ops);
|
4552 |
|
|
return "";
|
4553 |
|
|
}
|
4554 |
|
|
else
|
4555 |
|
|
return "vld1.\t%h0, [%1]";
|
4556 |
|
|
}
|
4557 |
|
|
[(set (attr "neon_type")
|
4558 |
|
|
(if_then_else (gt (const_string "") (const_string "1"))
|
4559 |
|
|
(const_string "neon_vld3_vld4_all_lanes")
|
4560 |
|
|
(const_string "neon_vld1_1_2_regs")))])
|
4561 |
|
|
|
4562 |
|
|
(define_insn "neon_vst3"
|
4563 |
|
|
[(set (mem:EI (match_operand:SI 0 "s_register_operand" "r"))
|
4564 |
|
|
(unspec:EI [(match_operand:EI 1 "s_register_operand" "w")
|
4565 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4566 |
|
|
UNSPEC_VST3))]
|
4567 |
|
|
"TARGET_NEON"
|
4568 |
|
|
{
|
4569 |
|
|
if ( == 64)
|
4570 |
|
|
return "vst1.64\t%h1, [%0]";
|
4571 |
|
|
else
|
4572 |
|
|
return "vst3.\t%h1, [%0]";
|
4573 |
|
|
}
|
4574 |
|
|
[(set (attr "neon_type")
|
4575 |
|
|
(if_then_else (eq (const_string "") (const_string "64"))
|
4576 |
|
|
(const_string "neon_vst1_1_2_regs_vst2_2_regs")
|
4577 |
|
|
(const_string "neon_vst2_4_regs_vst3_vst4")))])
|
4578 |
|
|
|
4579 |
|
|
(define_expand "neon_vst3"
|
4580 |
|
|
[(match_operand:SI 0 "s_register_operand" "+r")
|
4581 |
|
|
(match_operand:CI 1 "s_register_operand" "w")
|
4582 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4583 |
|
|
"TARGET_NEON"
|
4584 |
|
|
{
|
4585 |
|
|
emit_insn (gen_neon_vst3qa (operands[0], operands[0], operands[1]));
|
4586 |
|
|
emit_insn (gen_neon_vst3qb (operands[0], operands[0], operands[1]));
|
4587 |
|
|
DONE;
|
4588 |
|
|
})
|
4589 |
|
|
|
4590 |
|
|
(define_insn "neon_vst3qa"
|
4591 |
|
|
[(set (mem:EI (match_operand:SI 1 "s_register_operand" "0"))
|
4592 |
|
|
(unspec:EI [(match_operand:CI 2 "s_register_operand" "w")
|
4593 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4594 |
|
|
UNSPEC_VST3A))
|
4595 |
|
|
(set (match_operand:SI 0 "s_register_operand" "=r")
|
4596 |
|
|
(plus:SI (match_dup 1)
|
4597 |
|
|
(const_int 24)))]
|
4598 |
|
|
"TARGET_NEON"
|
4599 |
|
|
{
|
4600 |
|
|
int regno = REGNO (operands[2]);
|
4601 |
|
|
rtx ops[4];
|
4602 |
|
|
ops[0] = operands[0];
|
4603 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4604 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4605 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 8);
|
4606 |
|
|
output_asm_insn ("vst3.\t{%P1, %P2, %P3}, [%0]!", ops);
|
4607 |
|
|
return "";
|
4608 |
|
|
}
|
4609 |
|
|
[(set_attr "neon_type" "neon_vst2_4_regs_vst3_vst4")]
|
4610 |
|
|
)
|
4611 |
|
|
|
4612 |
|
|
(define_insn "neon_vst3qb"
|
4613 |
|
|
[(set (mem:EI (match_operand:SI 1 "s_register_operand" "0"))
|
4614 |
|
|
(unspec:EI [(match_operand:CI 2 "s_register_operand" "w")
|
4615 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4616 |
|
|
UNSPEC_VST3B))
|
4617 |
|
|
(set (match_operand:SI 0 "s_register_operand" "=r")
|
4618 |
|
|
(plus:SI (match_dup 1)
|
4619 |
|
|
(const_int 24)))]
|
4620 |
|
|
"TARGET_NEON"
|
4621 |
|
|
{
|
4622 |
|
|
int regno = REGNO (operands[2]);
|
4623 |
|
|
rtx ops[4];
|
4624 |
|
|
ops[0] = operands[0];
|
4625 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2);
|
4626 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 6);
|
4627 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 10);
|
4628 |
|
|
output_asm_insn ("vst3.\t{%P1, %P2, %P3}, [%0]!", ops);
|
4629 |
|
|
return "";
|
4630 |
|
|
}
|
4631 |
|
|
[(set_attr "neon_type" "neon_vst2_4_regs_vst3_vst4")]
|
4632 |
|
|
)
|
4633 |
|
|
|
4634 |
|
|
(define_insn "neon_vst3_lane"
|
4635 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4636 |
|
|
(unspec:
|
4637 |
|
|
[(match_operand:EI 1 "s_register_operand" "w")
|
4638 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
4639 |
|
|
(unspec:VD [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4640 |
|
|
UNSPEC_VST3_LANE))]
|
4641 |
|
|
"TARGET_NEON"
|
4642 |
|
|
{
|
4643 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4644 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4645 |
|
|
int regno = REGNO (operands[1]);
|
4646 |
|
|
rtx ops[5];
|
4647 |
|
|
if (lane < 0 || lane >= max)
|
4648 |
|
|
error ("lane out of range");
|
4649 |
|
|
ops[0] = operands[0];
|
4650 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4651 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 2);
|
4652 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 4);
|
4653 |
|
|
ops[4] = operands[2];
|
4654 |
|
|
output_asm_insn ("vst3.\t{%P1[%c4], %P2[%c4], %P3[%c4]}, [%0]",
|
4655 |
|
|
ops);
|
4656 |
|
|
return "";
|
4657 |
|
|
}
|
4658 |
|
|
[(set_attr "neon_type" "neon_vst3_vst4_lane")]
|
4659 |
|
|
)
|
4660 |
|
|
|
4661 |
|
|
(define_insn "neon_vst3_lane"
|
4662 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4663 |
|
|
(unspec:
|
4664 |
|
|
[(match_operand:CI 1 "s_register_operand" "w")
|
4665 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
4666 |
|
|
(unspec:VMQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4667 |
|
|
UNSPEC_VST3_LANE))]
|
4668 |
|
|
"TARGET_NEON"
|
4669 |
|
|
{
|
4670 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4671 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4672 |
|
|
int regno = REGNO (operands[1]);
|
4673 |
|
|
rtx ops[5];
|
4674 |
|
|
if (lane < 0 || lane >= max)
|
4675 |
|
|
error ("lane out of range");
|
4676 |
|
|
else if (lane >= max / 2)
|
4677 |
|
|
{
|
4678 |
|
|
lane -= max / 2;
|
4679 |
|
|
regno += 2;
|
4680 |
|
|
}
|
4681 |
|
|
ops[0] = operands[0];
|
4682 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4683 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4684 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 8);
|
4685 |
|
|
ops[4] = GEN_INT (lane);
|
4686 |
|
|
output_asm_insn ("vst3.\t{%P1[%c4], %P2[%c4], %P3[%c4]}, [%0]",
|
4687 |
|
|
ops);
|
4688 |
|
|
return "";
|
4689 |
|
|
}
|
4690 |
|
|
[(set_attr "neon_type" "neon_vst3_vst4_lane")])
|
4691 |
|
|
|
4692 |
|
|
(define_insn "neon_vld4"
|
4693 |
|
|
[(set (match_operand:OI 0 "s_register_operand" "=w")
|
4694 |
|
|
(unspec:OI [(mem:OI (match_operand:SI 1 "s_register_operand" "r"))
|
4695 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4696 |
|
|
UNSPEC_VLD4))]
|
4697 |
|
|
"TARGET_NEON"
|
4698 |
|
|
{
|
4699 |
|
|
if ( == 64)
|
4700 |
|
|
return "vld1.64\t%h0, [%1]";
|
4701 |
|
|
else
|
4702 |
|
|
return "vld4.\t%h0, [%1]";
|
4703 |
|
|
}
|
4704 |
|
|
[(set (attr "neon_type")
|
4705 |
|
|
(if_then_else (eq (const_string "") (const_string "64"))
|
4706 |
|
|
(const_string "neon_vld1_1_2_regs")
|
4707 |
|
|
(const_string "neon_vld3_vld4")))]
|
4708 |
|
|
)
|
4709 |
|
|
|
4710 |
|
|
(define_expand "neon_vld4"
|
4711 |
|
|
[(match_operand:XI 0 "s_register_operand" "=w")
|
4712 |
|
|
(match_operand:SI 1 "s_register_operand" "+r")
|
4713 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4714 |
|
|
"TARGET_NEON"
|
4715 |
|
|
{
|
4716 |
|
|
emit_insn (gen_neon_vld4qa (operands[0], operands[0],
|
4717 |
|
|
operands[1], operands[1]));
|
4718 |
|
|
emit_insn (gen_neon_vld4qb (operands[0], operands[0],
|
4719 |
|
|
operands[1], operands[1]));
|
4720 |
|
|
DONE;
|
4721 |
|
|
})
|
4722 |
|
|
|
4723 |
|
|
(define_insn "neon_vld4qa"
|
4724 |
|
|
[(set (match_operand:XI 0 "s_register_operand" "=w")
|
4725 |
|
|
(unspec:XI [(mem:XI (match_operand:SI 3 "s_register_operand" "2"))
|
4726 |
|
|
(match_operand:XI 1 "s_register_operand" "0")
|
4727 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4728 |
|
|
UNSPEC_VLD4A))
|
4729 |
|
|
(set (match_operand:SI 2 "s_register_operand" "=r")
|
4730 |
|
|
(plus:SI (match_dup 3)
|
4731 |
|
|
(const_int 32)))]
|
4732 |
|
|
"TARGET_NEON"
|
4733 |
|
|
{
|
4734 |
|
|
int regno = REGNO (operands[0]);
|
4735 |
|
|
rtx ops[5];
|
4736 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4737 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 4);
|
4738 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 8);
|
4739 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 12);
|
4740 |
|
|
ops[4] = operands[2];
|
4741 |
|
|
output_asm_insn ("vld4.\t{%P0, %P1, %P2, %P3}, [%4]!", ops);
|
4742 |
|
|
return "";
|
4743 |
|
|
}
|
4744 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4")]
|
4745 |
|
|
)
|
4746 |
|
|
|
4747 |
|
|
(define_insn "neon_vld4qb"
|
4748 |
|
|
[(set (match_operand:XI 0 "s_register_operand" "=w")
|
4749 |
|
|
(unspec:XI [(mem:XI (match_operand:SI 3 "s_register_operand" "2"))
|
4750 |
|
|
(match_operand:XI 1 "s_register_operand" "0")
|
4751 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4752 |
|
|
UNSPEC_VLD4B))
|
4753 |
|
|
(set (match_operand:SI 2 "s_register_operand" "=r")
|
4754 |
|
|
(plus:SI (match_dup 3)
|
4755 |
|
|
(const_int 32)))]
|
4756 |
|
|
"TARGET_NEON"
|
4757 |
|
|
{
|
4758 |
|
|
int regno = REGNO (operands[0]);
|
4759 |
|
|
rtx ops[5];
|
4760 |
|
|
ops[0] = gen_rtx_REG (DImode, regno + 2);
|
4761 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 6);
|
4762 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 10);
|
4763 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 14);
|
4764 |
|
|
ops[4] = operands[2];
|
4765 |
|
|
output_asm_insn ("vld4.\t{%P0, %P1, %P2, %P3}, [%4]!", ops);
|
4766 |
|
|
return "";
|
4767 |
|
|
}
|
4768 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4")]
|
4769 |
|
|
)
|
4770 |
|
|
|
4771 |
|
|
(define_insn "neon_vld4_lane"
|
4772 |
|
|
[(set (match_operand:OI 0 "s_register_operand" "=w")
|
4773 |
|
|
(unspec:OI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4774 |
|
|
(match_operand:OI 2 "s_register_operand" "0")
|
4775 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
4776 |
|
|
(unspec:VD [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4777 |
|
|
UNSPEC_VLD4_LANE))]
|
4778 |
|
|
"TARGET_NEON"
|
4779 |
|
|
{
|
4780 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4781 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4782 |
|
|
int regno = REGNO (operands[0]);
|
4783 |
|
|
rtx ops[6];
|
4784 |
|
|
if (lane < 0 || lane >= max)
|
4785 |
|
|
error ("lane out of range");
|
4786 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4787 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2);
|
4788 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4789 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 6);
|
4790 |
|
|
ops[4] = operands[1];
|
4791 |
|
|
ops[5] = operands[3];
|
4792 |
|
|
output_asm_insn ("vld4.\t{%P0[%c5], %P1[%c5], %P2[%c5], %P3[%c5]}, [%4]",
|
4793 |
|
|
ops);
|
4794 |
|
|
return "";
|
4795 |
|
|
}
|
4796 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4_lane")]
|
4797 |
|
|
)
|
4798 |
|
|
|
4799 |
|
|
(define_insn "neon_vld4_lane"
|
4800 |
|
|
[(set (match_operand:XI 0 "s_register_operand" "=w")
|
4801 |
|
|
(unspec:XI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4802 |
|
|
(match_operand:XI 2 "s_register_operand" "0")
|
4803 |
|
|
(match_operand:SI 3 "immediate_operand" "i")
|
4804 |
|
|
(unspec:VMQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4805 |
|
|
UNSPEC_VLD4_LANE))]
|
4806 |
|
|
"TARGET_NEON"
|
4807 |
|
|
{
|
4808 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[3]);
|
4809 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4810 |
|
|
int regno = REGNO (operands[0]);
|
4811 |
|
|
rtx ops[6];
|
4812 |
|
|
if (lane < 0 || lane >= max)
|
4813 |
|
|
error ("lane out of range");
|
4814 |
|
|
else if (lane >= max / 2)
|
4815 |
|
|
{
|
4816 |
|
|
lane -= max / 2;
|
4817 |
|
|
regno += 2;
|
4818 |
|
|
}
|
4819 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4820 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 4);
|
4821 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 8);
|
4822 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 12);
|
4823 |
|
|
ops[4] = operands[1];
|
4824 |
|
|
ops[5] = GEN_INT (lane);
|
4825 |
|
|
output_asm_insn ("vld4.\t{%P0[%c5], %P1[%c5], %P2[%c5], %P3[%c5]}, [%4]",
|
4826 |
|
|
ops);
|
4827 |
|
|
return "";
|
4828 |
|
|
}
|
4829 |
|
|
[(set_attr "neon_type" "neon_vld3_vld4_lane")]
|
4830 |
|
|
)
|
4831 |
|
|
|
4832 |
|
|
(define_insn "neon_vld4_dup"
|
4833 |
|
|
[(set (match_operand:OI 0 "s_register_operand" "=w")
|
4834 |
|
|
(unspec:OI [(mem: (match_operand:SI 1 "s_register_operand" "r"))
|
4835 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4836 |
|
|
UNSPEC_VLD4_DUP))]
|
4837 |
|
|
"TARGET_NEON"
|
4838 |
|
|
{
|
4839 |
|
|
if (GET_MODE_NUNITS (mode) > 1)
|
4840 |
|
|
{
|
4841 |
|
|
int regno = REGNO (operands[0]);
|
4842 |
|
|
rtx ops[5];
|
4843 |
|
|
ops[0] = gen_rtx_REG (DImode, regno);
|
4844 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2);
|
4845 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4846 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 6);
|
4847 |
|
|
ops[4] = operands[1];
|
4848 |
|
|
output_asm_insn ("vld4.\t{%P0[], %P1[], %P2[], %P3[]}, [%4]",
|
4849 |
|
|
ops);
|
4850 |
|
|
return "";
|
4851 |
|
|
}
|
4852 |
|
|
else
|
4853 |
|
|
return "vld1.\t%h0, [%1]";
|
4854 |
|
|
}
|
4855 |
|
|
[(set (attr "neon_type")
|
4856 |
|
|
(if_then_else (gt (const_string "") (const_string "1"))
|
4857 |
|
|
(const_string "neon_vld3_vld4_all_lanes")
|
4858 |
|
|
(const_string "neon_vld1_1_2_regs")))]
|
4859 |
|
|
)
|
4860 |
|
|
|
4861 |
|
|
(define_insn "neon_vst4"
|
4862 |
|
|
[(set (mem:OI (match_operand:SI 0 "s_register_operand" "r"))
|
4863 |
|
|
(unspec:OI [(match_operand:OI 1 "s_register_operand" "w")
|
4864 |
|
|
(unspec:VDX [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4865 |
|
|
UNSPEC_VST4))]
|
4866 |
|
|
"TARGET_NEON"
|
4867 |
|
|
{
|
4868 |
|
|
if ( == 64)
|
4869 |
|
|
return "vst1.64\t%h1, [%0]";
|
4870 |
|
|
else
|
4871 |
|
|
return "vst4.\t%h1, [%0]";
|
4872 |
|
|
}
|
4873 |
|
|
[(set (attr "neon_type")
|
4874 |
|
|
(if_then_else (eq (const_string "") (const_string "64"))
|
4875 |
|
|
(const_string "neon_vst1_1_2_regs_vst2_2_regs")
|
4876 |
|
|
(const_string "neon_vst2_4_regs_vst3_vst4")))]
|
4877 |
|
|
)
|
4878 |
|
|
|
4879 |
|
|
(define_expand "neon_vst4"
|
4880 |
|
|
[(match_operand:SI 0 "s_register_operand" "+r")
|
4881 |
|
|
(match_operand:XI 1 "s_register_operand" "w")
|
4882 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4883 |
|
|
"TARGET_NEON"
|
4884 |
|
|
{
|
4885 |
|
|
emit_insn (gen_neon_vst4qa (operands[0], operands[0], operands[1]));
|
4886 |
|
|
emit_insn (gen_neon_vst4qb (operands[0], operands[0], operands[1]));
|
4887 |
|
|
DONE;
|
4888 |
|
|
})
|
4889 |
|
|
|
4890 |
|
|
(define_insn "neon_vst4qa"
|
4891 |
|
|
[(set (mem:OI (match_operand:SI 1 "s_register_operand" "0"))
|
4892 |
|
|
(unspec:OI [(match_operand:XI 2 "s_register_operand" "w")
|
4893 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4894 |
|
|
UNSPEC_VST4A))
|
4895 |
|
|
(set (match_operand:SI 0 "s_register_operand" "=r")
|
4896 |
|
|
(plus:SI (match_dup 1)
|
4897 |
|
|
(const_int 32)))]
|
4898 |
|
|
"TARGET_NEON"
|
4899 |
|
|
{
|
4900 |
|
|
int regno = REGNO (operands[2]);
|
4901 |
|
|
rtx ops[5];
|
4902 |
|
|
ops[0] = operands[0];
|
4903 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4904 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4905 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 8);
|
4906 |
|
|
ops[4] = gen_rtx_REG (DImode, regno + 12);
|
4907 |
|
|
output_asm_insn ("vst4.\t{%P1, %P2, %P3, %P4}, [%0]!", ops);
|
4908 |
|
|
return "";
|
4909 |
|
|
}
|
4910 |
|
|
[(set_attr "neon_type" "neon_vst2_4_regs_vst3_vst4")]
|
4911 |
|
|
)
|
4912 |
|
|
|
4913 |
|
|
(define_insn "neon_vst4qb"
|
4914 |
|
|
[(set (mem:OI (match_operand:SI 1 "s_register_operand" "0"))
|
4915 |
|
|
(unspec:OI [(match_operand:XI 2 "s_register_operand" "w")
|
4916 |
|
|
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4917 |
|
|
UNSPEC_VST4B))
|
4918 |
|
|
(set (match_operand:SI 0 "s_register_operand" "=r")
|
4919 |
|
|
(plus:SI (match_dup 1)
|
4920 |
|
|
(const_int 32)))]
|
4921 |
|
|
"TARGET_NEON"
|
4922 |
|
|
{
|
4923 |
|
|
int regno = REGNO (operands[2]);
|
4924 |
|
|
rtx ops[5];
|
4925 |
|
|
ops[0] = operands[0];
|
4926 |
|
|
ops[1] = gen_rtx_REG (DImode, regno + 2);
|
4927 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 6);
|
4928 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 10);
|
4929 |
|
|
ops[4] = gen_rtx_REG (DImode, regno + 14);
|
4930 |
|
|
output_asm_insn ("vst4.\t{%P1, %P2, %P3, %P4}, [%0]!", ops);
|
4931 |
|
|
return "";
|
4932 |
|
|
}
|
4933 |
|
|
[(set_attr "neon_type" "neon_vst2_4_regs_vst3_vst4")]
|
4934 |
|
|
)
|
4935 |
|
|
|
4936 |
|
|
(define_insn "neon_vst4_lane"
|
4937 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4938 |
|
|
(unspec:
|
4939 |
|
|
[(match_operand:OI 1 "s_register_operand" "w")
|
4940 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
4941 |
|
|
(unspec:VD [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4942 |
|
|
UNSPEC_VST4_LANE))]
|
4943 |
|
|
"TARGET_NEON"
|
4944 |
|
|
{
|
4945 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4946 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4947 |
|
|
int regno = REGNO (operands[1]);
|
4948 |
|
|
rtx ops[6];
|
4949 |
|
|
if (lane < 0 || lane >= max)
|
4950 |
|
|
error ("lane out of range");
|
4951 |
|
|
ops[0] = operands[0];
|
4952 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4953 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 2);
|
4954 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 4);
|
4955 |
|
|
ops[4] = gen_rtx_REG (DImode, regno + 6);
|
4956 |
|
|
ops[5] = operands[2];
|
4957 |
|
|
output_asm_insn ("vst4.\t{%P1[%c5], %P2[%c5], %P3[%c5], %P4[%c5]}, [%0]",
|
4958 |
|
|
ops);
|
4959 |
|
|
return "";
|
4960 |
|
|
}
|
4961 |
|
|
[(set_attr "neon_type" "neon_vst3_vst4_lane")]
|
4962 |
|
|
)
|
4963 |
|
|
|
4964 |
|
|
(define_insn "neon_vst4_lane"
|
4965 |
|
|
[(set (mem: (match_operand:SI 0 "s_register_operand" "r"))
|
4966 |
|
|
(unspec:
|
4967 |
|
|
[(match_operand:XI 1 "s_register_operand" "w")
|
4968 |
|
|
(match_operand:SI 2 "immediate_operand" "i")
|
4969 |
|
|
(unspec:VMQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
|
4970 |
|
|
UNSPEC_VST4_LANE))]
|
4971 |
|
|
"TARGET_NEON"
|
4972 |
|
|
{
|
4973 |
|
|
HOST_WIDE_INT lane = INTVAL (operands[2]);
|
4974 |
|
|
HOST_WIDE_INT max = GET_MODE_NUNITS (mode);
|
4975 |
|
|
int regno = REGNO (operands[1]);
|
4976 |
|
|
rtx ops[6];
|
4977 |
|
|
if (lane < 0 || lane >= max)
|
4978 |
|
|
error ("lane out of range");
|
4979 |
|
|
else if (lane >= max / 2)
|
4980 |
|
|
{
|
4981 |
|
|
lane -= max / 2;
|
4982 |
|
|
regno += 2;
|
4983 |
|
|
}
|
4984 |
|
|
ops[0] = operands[0];
|
4985 |
|
|
ops[1] = gen_rtx_REG (DImode, regno);
|
4986 |
|
|
ops[2] = gen_rtx_REG (DImode, regno + 4);
|
4987 |
|
|
ops[3] = gen_rtx_REG (DImode, regno + 8);
|
4988 |
|
|
ops[4] = gen_rtx_REG (DImode, regno + 12);
|
4989 |
|
|
ops[5] = GEN_INT (lane);
|
4990 |
|
|
output_asm_insn ("vst4.\t{%P1[%c5], %P2[%c5], %P3[%c5], %P4[%c5]}, [%0]",
|
4991 |
|
|
ops);
|
4992 |
|
|
return "";
|
4993 |
|
|
}
|
4994 |
|
|
[(set_attr "neon_type" "neon_vst3_vst4_lane")]
|
4995 |
|
|
)
|
4996 |
|
|
|
4997 |
|
|
(define_expand "neon_vand"
|
4998 |
|
|
[(match_operand:VDQX 0 "s_register_operand" "")
|
4999 |
|
|
(match_operand:VDQX 1 "s_register_operand" "")
|
5000 |
|
|
(match_operand:VDQX 2 "neon_inv_logic_op2" "")
|
5001 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
5002 |
|
|
"TARGET_NEON"
|
5003 |
|
|
{
|
5004 |
|
|
emit_insn (gen_and3 (operands[0], operands[1], operands[2]));
|
5005 |
|
|
DONE;
|
5006 |
|
|
})
|
5007 |
|
|
|
5008 |
|
|
(define_expand "neon_vorr"
|
5009 |
|
|
[(match_operand:VDQX 0 "s_register_operand" "")
|
5010 |
|
|
(match_operand:VDQX 1 "s_register_operand" "")
|
5011 |
|
|
(match_operand:VDQX 2 "neon_logic_op2" "")
|
5012 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
5013 |
|
|
"TARGET_NEON"
|
5014 |
|
|
{
|
5015 |
|
|
emit_insn (gen_ior3 (operands[0], operands[1], operands[2]));
|
5016 |
|
|
DONE;
|
5017 |
|
|
})
|
5018 |
|
|
|
5019 |
|
|
(define_expand "neon_veor"
|
5020 |
|
|
[(match_operand:VDQX 0 "s_register_operand" "")
|
5021 |
|
|
(match_operand:VDQX 1 "s_register_operand" "")
|
5022 |
|
|
(match_operand:VDQX 2 "s_register_operand" "")
|
5023 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
5024 |
|
|
"TARGET_NEON"
|
5025 |
|
|
{
|
5026 |
|
|
emit_insn (gen_xor3 (operands[0], operands[1], operands[2]));
|
5027 |
|
|
DONE;
|
5028 |
|
|
})
|
5029 |
|
|
|
5030 |
|
|
(define_expand "neon_vbic"
|
5031 |
|
|
[(match_operand:VDQX 0 "s_register_operand" "")
|
5032 |
|
|
(match_operand:VDQX 1 "s_register_operand" "")
|
5033 |
|
|
(match_operand:VDQX 2 "neon_logic_op2" "")
|
5034 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
5035 |
|
|
"TARGET_NEON"
|
5036 |
|
|
{
|
5037 |
|
|
emit_insn (gen_bic3_neon (operands[0], operands[1], operands[2]));
|
5038 |
|
|
DONE;
|
5039 |
|
|
})
|
5040 |
|
|
|
5041 |
|
|
(define_expand "neon_vorn"
|
5042 |
|
|
[(match_operand:VDQX 0 "s_register_operand" "")
|
5043 |
|
|
(match_operand:VDQX 1 "s_register_operand" "")
|
5044 |
|
|
(match_operand:VDQX 2 "neon_inv_logic_op2" "")
|
5045 |
|
|
(match_operand:SI 3 "immediate_operand" "")]
|
5046 |
|
|
"TARGET_NEON"
|
5047 |
|
|
{
|
5048 |
|
|
emit_insn (gen_orn3_neon (operands[0], operands[1], operands[2]));
|
5049 |
|
|
DONE;
|
5050 |
|
|
})
|