1 |
709 |
jeremybenn |
;; Expander definitions for vector support between altivec & vsx. No
|
2 |
|
|
;; instructions are in this file, this file provides the generic vector
|
3 |
|
|
;; expander, and the actual vector instructions will be in altivec.md and
|
4 |
|
|
;; vsx.md
|
5 |
|
|
|
6 |
|
|
;; Copyright (C) 2009, 2010, 2011
|
7 |
|
|
;; Free Software Foundation, Inc.
|
8 |
|
|
;; Contributed by Michael Meissner
|
9 |
|
|
|
10 |
|
|
;; This file is part of GCC.
|
11 |
|
|
|
12 |
|
|
;; GCC is free software; you can redistribute it and/or modify it
|
13 |
|
|
;; under the terms of the GNU General Public License as published
|
14 |
|
|
;; by the Free Software Foundation; either version 3, or (at your
|
15 |
|
|
;; option) any later version.
|
16 |
|
|
|
17 |
|
|
;; GCC is distributed in the hope that it will be useful, but WITHOUT
|
18 |
|
|
;; ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
19 |
|
|
;; or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
|
20 |
|
|
;; License for more details.
|
21 |
|
|
|
22 |
|
|
;; You should have received a copy of the GNU General Public License
|
23 |
|
|
;; along with GCC; see the file COPYING3. If not see
|
24 |
|
|
;; .
|
25 |
|
|
|
26 |
|
|
|
27 |
|
|
;; Vector int modes
|
28 |
|
|
(define_mode_iterator VEC_I [V16QI V8HI V4SI])
|
29 |
|
|
|
30 |
|
|
;; Vector float modes
|
31 |
|
|
(define_mode_iterator VEC_F [V4SF V2DF])
|
32 |
|
|
|
33 |
|
|
;; Vector arithmetic modes
|
34 |
|
|
(define_mode_iterator VEC_A [V16QI V8HI V4SI V4SF V2DF])
|
35 |
|
|
|
36 |
|
|
;; Vector modes that need alginment via permutes
|
37 |
|
|
(define_mode_iterator VEC_K [V16QI V8HI V4SI V4SF])
|
38 |
|
|
|
39 |
|
|
;; Vector logical modes
|
40 |
|
|
(define_mode_iterator VEC_L [V16QI V8HI V4SI V2DI V4SF V2DF TI])
|
41 |
|
|
|
42 |
|
|
;; Vector modes for moves. Don't do TImode here.
|
43 |
|
|
(define_mode_iterator VEC_M [V16QI V8HI V4SI V2DI V4SF V2DF])
|
44 |
|
|
|
45 |
|
|
;; Vector modes for types that don't need a realignment under VSX
|
46 |
|
|
(define_mode_iterator VEC_N [V4SI V4SF V2DI V2DF])
|
47 |
|
|
|
48 |
|
|
;; Vector comparison modes
|
49 |
|
|
(define_mode_iterator VEC_C [V16QI V8HI V4SI V4SF V2DF])
|
50 |
|
|
|
51 |
|
|
;; Vector init/extract modes
|
52 |
|
|
(define_mode_iterator VEC_E [V16QI V8HI V4SI V2DI V4SF V2DF])
|
53 |
|
|
|
54 |
|
|
;; Vector modes for 64-bit base types
|
55 |
|
|
(define_mode_iterator VEC_64 [V2DI V2DF])
|
56 |
|
|
|
57 |
|
|
;; Vector reload iterator
|
58 |
|
|
(define_mode_iterator VEC_R [V16QI V8HI V4SI V2DI V4SF V2DF DF TI])
|
59 |
|
|
|
60 |
|
|
;; Base type from vector mode
|
61 |
|
|
(define_mode_attr VEC_base [(V16QI "QI")
|
62 |
|
|
(V8HI "HI")
|
63 |
|
|
(V4SI "SI")
|
64 |
|
|
(V2DI "DI")
|
65 |
|
|
(V4SF "SF")
|
66 |
|
|
(V2DF "DF")
|
67 |
|
|
(TI "TI")])
|
68 |
|
|
|
69 |
|
|
;; Same size integer type for floating point data
|
70 |
|
|
(define_mode_attr VEC_int [(V4SF "v4si")
|
71 |
|
|
(V2DF "v2di")])
|
72 |
|
|
|
73 |
|
|
(define_mode_attr VEC_INT [(V4SF "V4SI")
|
74 |
|
|
(V2DF "V2DI")])
|
75 |
|
|
|
76 |
|
|
;; constants for unspec
|
77 |
|
|
(define_c_enum "unspec" [UNSPEC_PREDICATE
|
78 |
|
|
UNSPEC_REDUC])
|
79 |
|
|
|
80 |
|
|
;; Vector reduction code iterators
|
81 |
|
|
(define_code_iterator VEC_reduc [plus smin smax])
|
82 |
|
|
|
83 |
|
|
(define_code_attr VEC_reduc_name [(plus "splus")
|
84 |
|
|
(smin "smin")
|
85 |
|
|
(smax "smax")])
|
86 |
|
|
|
87 |
|
|
(define_code_attr VEC_reduc_rtx [(plus "add")
|
88 |
|
|
(smin "smin")
|
89 |
|
|
(smax "smax")])
|
90 |
|
|
|
91 |
|
|
|
92 |
|
|
;; Vector move instructions.
|
93 |
|
|
(define_expand "mov"
|
94 |
|
|
[(set (match_operand:VEC_M 0 "nonimmediate_operand" "")
|
95 |
|
|
(match_operand:VEC_M 1 "any_operand" ""))]
|
96 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
97 |
|
|
{
|
98 |
|
|
if (can_create_pseudo_p ())
|
99 |
|
|
{
|
100 |
|
|
if (CONSTANT_P (operands[1])
|
101 |
|
|
&& !easy_vector_constant (operands[1], mode))
|
102 |
|
|
operands[1] = force_const_mem (mode, operands[1]);
|
103 |
|
|
|
104 |
|
|
else if (!vlogical_operand (operands[0], mode)
|
105 |
|
|
&& !vlogical_operand (operands[1], mode))
|
106 |
|
|
operands[1] = force_reg (mode, operands[1]);
|
107 |
|
|
}
|
108 |
|
|
})
|
109 |
|
|
|
110 |
|
|
;; Generic vector floating point load/store instructions. These will match
|
111 |
|
|
;; insns defined in vsx.md or altivec.md depending on the switches.
|
112 |
|
|
(define_expand "vector_load_"
|
113 |
|
|
[(set (match_operand:VEC_M 0 "vfloat_operand" "")
|
114 |
|
|
(match_operand:VEC_M 1 "memory_operand" ""))]
|
115 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
116 |
|
|
"")
|
117 |
|
|
|
118 |
|
|
(define_expand "vector_store_"
|
119 |
|
|
[(set (match_operand:VEC_M 0 "memory_operand" "")
|
120 |
|
|
(match_operand:VEC_M 1 "vfloat_operand" ""))]
|
121 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
122 |
|
|
"")
|
123 |
|
|
|
124 |
|
|
;; Splits if a GPR register was chosen for the move
|
125 |
|
|
(define_split
|
126 |
|
|
[(set (match_operand:VEC_L 0 "nonimmediate_operand" "")
|
127 |
|
|
(match_operand:VEC_L 1 "input_operand" ""))]
|
128 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)
|
129 |
|
|
&& reload_completed
|
130 |
|
|
&& gpr_or_gpr_p (operands[0], operands[1])"
|
131 |
|
|
[(pc)]
|
132 |
|
|
{
|
133 |
|
|
rs6000_split_multireg_move (operands[0], operands[1]);
|
134 |
|
|
DONE;
|
135 |
|
|
})
|
136 |
|
|
|
137 |
|
|
;; Vector floating point load/store instructions that uses the Altivec
|
138 |
|
|
;; instructions even if we are compiling for VSX, since the Altivec
|
139 |
|
|
;; instructions silently ignore the bottom 3 bits of the address, and VSX does
|
140 |
|
|
;; not.
|
141 |
|
|
(define_expand "vector_altivec_load_"
|
142 |
|
|
[(set (match_operand:VEC_M 0 "vfloat_operand" "")
|
143 |
|
|
(match_operand:VEC_M 1 "memory_operand" ""))]
|
144 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
145 |
|
|
"
|
146 |
|
|
{
|
147 |
|
|
gcc_assert (VECTOR_MEM_ALTIVEC_OR_VSX_P (mode));
|
148 |
|
|
|
149 |
|
|
if (VECTOR_MEM_VSX_P (mode))
|
150 |
|
|
{
|
151 |
|
|
operands[1] = rs6000_address_for_altivec (operands[1]);
|
152 |
|
|
emit_insn (gen_altivec_lvx_ (operands[0], operands[1]));
|
153 |
|
|
DONE;
|
154 |
|
|
}
|
155 |
|
|
}")
|
156 |
|
|
|
157 |
|
|
(define_expand "vector_altivec_store_"
|
158 |
|
|
[(set (match_operand:VEC_M 0 "memory_operand" "")
|
159 |
|
|
(match_operand:VEC_M 1 "vfloat_operand" ""))]
|
160 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
161 |
|
|
"
|
162 |
|
|
{
|
163 |
|
|
gcc_assert (VECTOR_MEM_ALTIVEC_OR_VSX_P (mode));
|
164 |
|
|
|
165 |
|
|
if (VECTOR_MEM_VSX_P (mode))
|
166 |
|
|
{
|
167 |
|
|
operands[0] = rs6000_address_for_altivec (operands[0]);
|
168 |
|
|
emit_insn (gen_altivec_stvx_ (operands[0], operands[1]));
|
169 |
|
|
DONE;
|
170 |
|
|
}
|
171 |
|
|
}")
|
172 |
|
|
|
173 |
|
|
|
174 |
|
|
|
175 |
|
|
;; Reload patterns for vector operations. We may need an addtional base
|
176 |
|
|
;; register to convert the reg+offset addressing to reg+reg for vector
|
177 |
|
|
;; registers and reg+reg or (reg+reg)&(-16) addressing to just an index
|
178 |
|
|
;; register for gpr registers.
|
179 |
|
|
(define_expand "reload___store"
|
180 |
|
|
[(parallel [(match_operand:VEC_R 0 "memory_operand" "m")
|
181 |
|
|
(match_operand:VEC_R 1 "gpc_reg_operand" "r")
|
182 |
|
|
(match_operand:P 2 "register_operand" "=&b")])]
|
183 |
|
|
""
|
184 |
|
|
{
|
185 |
|
|
rs6000_secondary_reload_inner (operands[1], operands[0], operands[2], true);
|
186 |
|
|
DONE;
|
187 |
|
|
})
|
188 |
|
|
|
189 |
|
|
(define_expand "reload___load"
|
190 |
|
|
[(parallel [(match_operand:VEC_R 0 "gpc_reg_operand" "=&r")
|
191 |
|
|
(match_operand:VEC_R 1 "memory_operand" "m")
|
192 |
|
|
(match_operand:P 2 "register_operand" "=&b")])]
|
193 |
|
|
""
|
194 |
|
|
{
|
195 |
|
|
rs6000_secondary_reload_inner (operands[0], operands[1], operands[2], false);
|
196 |
|
|
DONE;
|
197 |
|
|
})
|
198 |
|
|
|
199 |
|
|
;; Reload sometimes tries to move the address to a GPR, and can generate
|
200 |
|
|
;; invalid RTL for addresses involving AND -16. Allow addresses involving
|
201 |
|
|
;; reg+reg, reg+small constant, or just reg, all wrapped in an AND -16.
|
202 |
|
|
|
203 |
|
|
(define_insn_and_split "*vec_reload_and_plus_"
|
204 |
|
|
[(set (match_operand:P 0 "gpc_reg_operand" "=b")
|
205 |
|
|
(and:P (plus:P (match_operand:P 1 "gpc_reg_operand" "r")
|
206 |
|
|
(match_operand:P 2 "reg_or_cint_operand" "rI"))
|
207 |
|
|
(const_int -16)))]
|
208 |
|
|
"(TARGET_ALTIVEC || TARGET_VSX) && (reload_in_progress || reload_completed)"
|
209 |
|
|
"#"
|
210 |
|
|
"&& reload_completed"
|
211 |
|
|
[(set (match_dup 0)
|
212 |
|
|
(plus:P (match_dup 1)
|
213 |
|
|
(match_dup 2)))
|
214 |
|
|
(parallel [(set (match_dup 0)
|
215 |
|
|
(and:P (match_dup 0)
|
216 |
|
|
(const_int -16)))
|
217 |
|
|
(clobber:CC (scratch:CC))])])
|
218 |
|
|
|
219 |
|
|
;; The normal ANDSI3/ANDDI3 won't match if reload decides to move an AND -16
|
220 |
|
|
;; address to a register because there is no clobber of a (scratch), so we add
|
221 |
|
|
;; it here.
|
222 |
|
|
(define_insn_and_split "*vec_reload_and_reg_"
|
223 |
|
|
[(set (match_operand:P 0 "gpc_reg_operand" "=b")
|
224 |
|
|
(and:P (match_operand:P 1 "gpc_reg_operand" "r")
|
225 |
|
|
(const_int -16)))]
|
226 |
|
|
"(TARGET_ALTIVEC || TARGET_VSX) && (reload_in_progress || reload_completed)"
|
227 |
|
|
"#"
|
228 |
|
|
"&& reload_completed"
|
229 |
|
|
[(parallel [(set (match_dup 0)
|
230 |
|
|
(and:P (match_dup 1)
|
231 |
|
|
(const_int -16)))
|
232 |
|
|
(clobber:CC (scratch:CC))])])
|
233 |
|
|
|
234 |
|
|
;; Generic floating point vector arithmetic support
|
235 |
|
|
(define_expand "add3"
|
236 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
237 |
|
|
(plus:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")
|
238 |
|
|
(match_operand:VEC_F 2 "vfloat_operand" "")))]
|
239 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
240 |
|
|
"")
|
241 |
|
|
|
242 |
|
|
(define_expand "sub3"
|
243 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
244 |
|
|
(minus:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")
|
245 |
|
|
(match_operand:VEC_F 2 "vfloat_operand" "")))]
|
246 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
247 |
|
|
"")
|
248 |
|
|
|
249 |
|
|
(define_expand "mul3"
|
250 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
251 |
|
|
(mult:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")
|
252 |
|
|
(match_operand:VEC_F 2 "vfloat_operand" "")))]
|
253 |
|
|
"VECTOR_UNIT_VSX_P (mode) || VECTOR_UNIT_ALTIVEC_P (mode)"
|
254 |
|
|
{
|
255 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
256 |
|
|
{
|
257 |
|
|
emit_insn (gen_altivec_mulv4sf3 (operands[0], operands[1], operands[2]));
|
258 |
|
|
DONE;
|
259 |
|
|
}
|
260 |
|
|
})
|
261 |
|
|
|
262 |
|
|
(define_expand "div3"
|
263 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
264 |
|
|
(div:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")
|
265 |
|
|
(match_operand:VEC_F 2 "vfloat_operand" "")))]
|
266 |
|
|
"VECTOR_UNIT_VSX_P (mode)"
|
267 |
|
|
"")
|
268 |
|
|
|
269 |
|
|
(define_expand "neg2"
|
270 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
271 |
|
|
(neg:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")))]
|
272 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
273 |
|
|
"
|
274 |
|
|
{
|
275 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
276 |
|
|
{
|
277 |
|
|
emit_insn (gen_altivec_negv4sf2 (operands[0], operands[1]));
|
278 |
|
|
DONE;
|
279 |
|
|
}
|
280 |
|
|
}")
|
281 |
|
|
|
282 |
|
|
(define_expand "abs2"
|
283 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
284 |
|
|
(abs:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")))]
|
285 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
286 |
|
|
"
|
287 |
|
|
{
|
288 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
289 |
|
|
{
|
290 |
|
|
emit_insn (gen_altivec_absv4sf2 (operands[0], operands[1]));
|
291 |
|
|
DONE;
|
292 |
|
|
}
|
293 |
|
|
}")
|
294 |
|
|
|
295 |
|
|
(define_expand "smin3"
|
296 |
|
|
[(set (match_operand:VEC_F 0 "register_operand" "")
|
297 |
|
|
(smin:VEC_F (match_operand:VEC_F 1 "register_operand" "")
|
298 |
|
|
(match_operand:VEC_F 2 "register_operand" "")))]
|
299 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
300 |
|
|
"")
|
301 |
|
|
|
302 |
|
|
(define_expand "smax3"
|
303 |
|
|
[(set (match_operand:VEC_F 0 "register_operand" "")
|
304 |
|
|
(smax:VEC_F (match_operand:VEC_F 1 "register_operand" "")
|
305 |
|
|
(match_operand:VEC_F 2 "register_operand" "")))]
|
306 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
307 |
|
|
"")
|
308 |
|
|
|
309 |
|
|
|
310 |
|
|
(define_expand "sqrt2"
|
311 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
312 |
|
|
(sqrt:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")))]
|
313 |
|
|
"VECTOR_UNIT_VSX_P (mode)"
|
314 |
|
|
"")
|
315 |
|
|
|
316 |
|
|
(define_expand "rsqrte2"
|
317 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
318 |
|
|
(unspec:VEC_F [(match_operand:VEC_F 1 "vfloat_operand" "")]
|
319 |
|
|
UNSPEC_RSQRT))]
|
320 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
321 |
|
|
"")
|
322 |
|
|
|
323 |
|
|
(define_expand "re2"
|
324 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
325 |
|
|
(unspec:VEC_F [(match_operand:VEC_F 1 "vfloat_operand" "f")]
|
326 |
|
|
UNSPEC_FRES))]
|
327 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
328 |
|
|
"")
|
329 |
|
|
|
330 |
|
|
(define_expand "ftrunc2"
|
331 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
332 |
|
|
(fix:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")))]
|
333 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
334 |
|
|
"")
|
335 |
|
|
|
336 |
|
|
(define_expand "vector_ceil2"
|
337 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
338 |
|
|
(unspec:VEC_F [(match_operand:VEC_F 1 "vfloat_operand" "")]
|
339 |
|
|
UNSPEC_FRIP))]
|
340 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
341 |
|
|
"")
|
342 |
|
|
|
343 |
|
|
(define_expand "vector_floor2"
|
344 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
345 |
|
|
(unspec:VEC_F [(match_operand:VEC_F 1 "vfloat_operand" "")]
|
346 |
|
|
UNSPEC_FRIM))]
|
347 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
348 |
|
|
"")
|
349 |
|
|
|
350 |
|
|
(define_expand "vector_btrunc2"
|
351 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
352 |
|
|
(fix:VEC_F (match_operand:VEC_F 1 "vfloat_operand" "")))]
|
353 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
354 |
|
|
"")
|
355 |
|
|
|
356 |
|
|
(define_expand "vector_copysign3"
|
357 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
358 |
|
|
(unspec:VEC_F [(match_operand:VEC_F 1 "vfloat_operand" "")
|
359 |
|
|
(match_operand:VEC_F 2 "vfloat_operand" "")] UNSPEC_COPYSIGN))]
|
360 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
361 |
|
|
"
|
362 |
|
|
{
|
363 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
364 |
|
|
{
|
365 |
|
|
emit_insn (gen_altivec_copysign_v4sf3 (operands[0], operands[1],
|
366 |
|
|
operands[2]));
|
367 |
|
|
DONE;
|
368 |
|
|
}
|
369 |
|
|
}")
|
370 |
|
|
|
371 |
|
|
|
372 |
|
|
;; Vector comparisons
|
373 |
|
|
(define_expand "vcond"
|
374 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
375 |
|
|
(if_then_else:VEC_F
|
376 |
|
|
(match_operator 3 "comparison_operator"
|
377 |
|
|
[(match_operand:VEC_F 4 "vfloat_operand" "")
|
378 |
|
|
(match_operand:VEC_F 5 "vfloat_operand" "")])
|
379 |
|
|
(match_operand:VEC_F 1 "vfloat_operand" "")
|
380 |
|
|
(match_operand:VEC_F 2 "vfloat_operand" "")))]
|
381 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
382 |
|
|
"
|
383 |
|
|
{
|
384 |
|
|
if (rs6000_emit_vector_cond_expr (operands[0], operands[1], operands[2],
|
385 |
|
|
operands[3], operands[4], operands[5]))
|
386 |
|
|
DONE;
|
387 |
|
|
else
|
388 |
|
|
FAIL;
|
389 |
|
|
}")
|
390 |
|
|
|
391 |
|
|
(define_expand "vcond"
|
392 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
393 |
|
|
(if_then_else:VEC_I
|
394 |
|
|
(match_operator 3 "comparison_operator"
|
395 |
|
|
[(match_operand:VEC_I 4 "vint_operand" "")
|
396 |
|
|
(match_operand:VEC_I 5 "vint_operand" "")])
|
397 |
|
|
(match_operand:VEC_I 1 "vint_operand" "")
|
398 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
399 |
|
|
"VECTOR_UNIT_ALTIVEC_P (mode)"
|
400 |
|
|
"
|
401 |
|
|
{
|
402 |
|
|
if (rs6000_emit_vector_cond_expr (operands[0], operands[1], operands[2],
|
403 |
|
|
operands[3], operands[4], operands[5]))
|
404 |
|
|
DONE;
|
405 |
|
|
else
|
406 |
|
|
FAIL;
|
407 |
|
|
}")
|
408 |
|
|
|
409 |
|
|
(define_expand "vcondv4sfv4si"
|
410 |
|
|
[(set (match_operand:V4SF 0 "vfloat_operand" "")
|
411 |
|
|
(if_then_else:V4SF
|
412 |
|
|
(match_operator 3 "comparison_operator"
|
413 |
|
|
[(match_operand:V4SI 4 "vint_operand" "")
|
414 |
|
|
(match_operand:V4SI 5 "vint_operand" "")])
|
415 |
|
|
(match_operand:V4SF 1 "vfloat_operand" "")
|
416 |
|
|
(match_operand:V4SF 2 "vfloat_operand" "")))]
|
417 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SFmode)
|
418 |
|
|
&& VECTOR_UNIT_ALTIVEC_P (V4SImode)"
|
419 |
|
|
"
|
420 |
|
|
{
|
421 |
|
|
if (rs6000_emit_vector_cond_expr (operands[0], operands[1], operands[2],
|
422 |
|
|
operands[3], operands[4], operands[5]))
|
423 |
|
|
DONE;
|
424 |
|
|
else
|
425 |
|
|
FAIL;
|
426 |
|
|
}")
|
427 |
|
|
|
428 |
|
|
(define_expand "vcondv4siv4sf"
|
429 |
|
|
[(set (match_operand:V4SI 0 "vint_operand" "")
|
430 |
|
|
(if_then_else:V4SI
|
431 |
|
|
(match_operator 3 "comparison_operator"
|
432 |
|
|
[(match_operand:V4SF 4 "vfloat_operand" "")
|
433 |
|
|
(match_operand:V4SF 5 "vfloat_operand" "")])
|
434 |
|
|
(match_operand:V4SI 1 "vint_operand" "")
|
435 |
|
|
(match_operand:V4SI 2 "vint_operand" "")))]
|
436 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SFmode)
|
437 |
|
|
&& VECTOR_UNIT_ALTIVEC_P (V4SImode)"
|
438 |
|
|
"
|
439 |
|
|
{
|
440 |
|
|
if (rs6000_emit_vector_cond_expr (operands[0], operands[1], operands[2],
|
441 |
|
|
operands[3], operands[4], operands[5]))
|
442 |
|
|
DONE;
|
443 |
|
|
else
|
444 |
|
|
FAIL;
|
445 |
|
|
}")
|
446 |
|
|
|
447 |
|
|
(define_expand "vcondu"
|
448 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
449 |
|
|
(if_then_else:VEC_I
|
450 |
|
|
(match_operator 3 "comparison_operator"
|
451 |
|
|
[(match_operand:VEC_I 4 "vint_operand" "")
|
452 |
|
|
(match_operand:VEC_I 5 "vint_operand" "")])
|
453 |
|
|
(match_operand:VEC_I 1 "vint_operand" "")
|
454 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
455 |
|
|
"VECTOR_UNIT_ALTIVEC_P (mode)"
|
456 |
|
|
"
|
457 |
|
|
{
|
458 |
|
|
if (rs6000_emit_vector_cond_expr (operands[0], operands[1], operands[2],
|
459 |
|
|
operands[3], operands[4], operands[5]))
|
460 |
|
|
DONE;
|
461 |
|
|
else
|
462 |
|
|
FAIL;
|
463 |
|
|
}")
|
464 |
|
|
|
465 |
|
|
(define_expand "vconduv4sfv4si"
|
466 |
|
|
[(set (match_operand:V4SF 0 "vfloat_operand" "")
|
467 |
|
|
(if_then_else:V4SF
|
468 |
|
|
(match_operator 3 "comparison_operator"
|
469 |
|
|
[(match_operand:V4SI 4 "vint_operand" "")
|
470 |
|
|
(match_operand:V4SI 5 "vint_operand" "")])
|
471 |
|
|
(match_operand:V4SF 1 "vfloat_operand" "")
|
472 |
|
|
(match_operand:V4SF 2 "vfloat_operand" "")))]
|
473 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SFmode)
|
474 |
|
|
&& VECTOR_UNIT_ALTIVEC_P (V4SImode)"
|
475 |
|
|
"
|
476 |
|
|
{
|
477 |
|
|
if (rs6000_emit_vector_cond_expr (operands[0], operands[1], operands[2],
|
478 |
|
|
operands[3], operands[4], operands[5]))
|
479 |
|
|
DONE;
|
480 |
|
|
else
|
481 |
|
|
FAIL;
|
482 |
|
|
}")
|
483 |
|
|
|
484 |
|
|
(define_expand "vector_eq"
|
485 |
|
|
[(set (match_operand:VEC_C 0 "vlogical_operand" "")
|
486 |
|
|
(eq:VEC_C (match_operand:VEC_C 1 "vlogical_operand" "")
|
487 |
|
|
(match_operand:VEC_C 2 "vlogical_operand" "")))]
|
488 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
489 |
|
|
"")
|
490 |
|
|
|
491 |
|
|
(define_expand "vector_gt"
|
492 |
|
|
[(set (match_operand:VEC_C 0 "vlogical_operand" "")
|
493 |
|
|
(gt:VEC_C (match_operand:VEC_C 1 "vlogical_operand" "")
|
494 |
|
|
(match_operand:VEC_C 2 "vlogical_operand" "")))]
|
495 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
496 |
|
|
"")
|
497 |
|
|
|
498 |
|
|
(define_expand "vector_ge"
|
499 |
|
|
[(set (match_operand:VEC_C 0 "vlogical_operand" "")
|
500 |
|
|
(ge:VEC_C (match_operand:VEC_C 1 "vlogical_operand" "")
|
501 |
|
|
(match_operand:VEC_C 2 "vlogical_operand" "")))]
|
502 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
503 |
|
|
"")
|
504 |
|
|
|
505 |
|
|
(define_expand "vector_gtu"
|
506 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
507 |
|
|
(gtu:VEC_I (match_operand:VEC_I 1 "vint_operand" "")
|
508 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
509 |
|
|
"VECTOR_UNIT_ALTIVEC_P (mode)"
|
510 |
|
|
"")
|
511 |
|
|
|
512 |
|
|
(define_expand "vector_geu"
|
513 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
514 |
|
|
(geu:VEC_I (match_operand:VEC_I 1 "vint_operand" "")
|
515 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
516 |
|
|
"VECTOR_UNIT_ALTIVEC_P (mode)"
|
517 |
|
|
"")
|
518 |
|
|
|
519 |
|
|
;; Note the arguments for __builtin_altivec_vsel are op2, op1, mask
|
520 |
|
|
;; which is in the reverse order that we want
|
521 |
|
|
(define_expand "vector_select_"
|
522 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
523 |
|
|
(if_then_else:VEC_L
|
524 |
|
|
(ne:CC (match_operand:VEC_L 3 "vlogical_operand" "")
|
525 |
|
|
(match_dup 4))
|
526 |
|
|
(match_operand:VEC_L 2 "vlogical_operand" "")
|
527 |
|
|
(match_operand:VEC_L 1 "vlogical_operand" "")))]
|
528 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
529 |
|
|
"operands[4] = CONST0_RTX (mode);")
|
530 |
|
|
|
531 |
|
|
(define_expand "vector_select__uns"
|
532 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
533 |
|
|
(if_then_else:VEC_L
|
534 |
|
|
(ne:CCUNS (match_operand:VEC_L 3 "vlogical_operand" "")
|
535 |
|
|
(match_dup 4))
|
536 |
|
|
(match_operand:VEC_L 2 "vlogical_operand" "")
|
537 |
|
|
(match_operand:VEC_L 1 "vlogical_operand" "")))]
|
538 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
539 |
|
|
"operands[4] = CONST0_RTX (mode);")
|
540 |
|
|
|
541 |
|
|
;; Expansions that compare vectors producing a vector result and a predicate,
|
542 |
|
|
;; setting CR6 to indicate a combined status
|
543 |
|
|
(define_expand "vector_eq__p"
|
544 |
|
|
[(parallel
|
545 |
|
|
[(set (reg:CC 74)
|
546 |
|
|
(unspec:CC [(eq:CC (match_operand:VEC_A 1 "vlogical_operand" "")
|
547 |
|
|
(match_operand:VEC_A 2 "vlogical_operand" ""))]
|
548 |
|
|
UNSPEC_PREDICATE))
|
549 |
|
|
(set (match_operand:VEC_A 0 "vlogical_operand" "")
|
550 |
|
|
(eq:VEC_A (match_dup 1)
|
551 |
|
|
(match_dup 2)))])]
|
552 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
553 |
|
|
"")
|
554 |
|
|
|
555 |
|
|
(define_expand "vector_gt__p"
|
556 |
|
|
[(parallel
|
557 |
|
|
[(set (reg:CC 74)
|
558 |
|
|
(unspec:CC [(gt:CC (match_operand:VEC_A 1 "vlogical_operand" "")
|
559 |
|
|
(match_operand:VEC_A 2 "vlogical_operand" ""))]
|
560 |
|
|
UNSPEC_PREDICATE))
|
561 |
|
|
(set (match_operand:VEC_A 0 "vlogical_operand" "")
|
562 |
|
|
(gt:VEC_A (match_dup 1)
|
563 |
|
|
(match_dup 2)))])]
|
564 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
565 |
|
|
"")
|
566 |
|
|
|
567 |
|
|
(define_expand "vector_ge__p"
|
568 |
|
|
[(parallel
|
569 |
|
|
[(set (reg:CC 74)
|
570 |
|
|
(unspec:CC [(ge:CC (match_operand:VEC_F 1 "vfloat_operand" "")
|
571 |
|
|
(match_operand:VEC_F 2 "vfloat_operand" ""))]
|
572 |
|
|
UNSPEC_PREDICATE))
|
573 |
|
|
(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
574 |
|
|
(ge:VEC_F (match_dup 1)
|
575 |
|
|
(match_dup 2)))])]
|
576 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
577 |
|
|
"")
|
578 |
|
|
|
579 |
|
|
(define_expand "vector_gtu__p"
|
580 |
|
|
[(parallel
|
581 |
|
|
[(set (reg:CC 74)
|
582 |
|
|
(unspec:CC [(gtu:CC (match_operand:VEC_I 1 "vint_operand" "")
|
583 |
|
|
(match_operand:VEC_I 2 "vint_operand" ""))]
|
584 |
|
|
UNSPEC_PREDICATE))
|
585 |
|
|
(set (match_operand:VEC_I 0 "vlogical_operand" "")
|
586 |
|
|
(gtu:VEC_I (match_dup 1)
|
587 |
|
|
(match_dup 2)))])]
|
588 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
589 |
|
|
"")
|
590 |
|
|
|
591 |
|
|
;; AltiVec/VSX predicates.
|
592 |
|
|
|
593 |
|
|
(define_expand "cr6_test_for_zero"
|
594 |
|
|
[(set (match_operand:SI 0 "register_operand" "=r")
|
595 |
|
|
(eq:SI (reg:CC 74)
|
596 |
|
|
(const_int 0)))]
|
597 |
|
|
"TARGET_ALTIVEC || TARGET_VSX"
|
598 |
|
|
"")
|
599 |
|
|
|
600 |
|
|
(define_expand "cr6_test_for_zero_reverse"
|
601 |
|
|
[(set (match_operand:SI 0 "register_operand" "=r")
|
602 |
|
|
(eq:SI (reg:CC 74)
|
603 |
|
|
(const_int 0)))
|
604 |
|
|
(set (match_dup 0) (minus:SI (const_int 1) (match_dup 0)))]
|
605 |
|
|
"TARGET_ALTIVEC || TARGET_VSX"
|
606 |
|
|
"")
|
607 |
|
|
|
608 |
|
|
(define_expand "cr6_test_for_lt"
|
609 |
|
|
[(set (match_operand:SI 0 "register_operand" "=r")
|
610 |
|
|
(lt:SI (reg:CC 74)
|
611 |
|
|
(const_int 0)))]
|
612 |
|
|
"TARGET_ALTIVEC || TARGET_VSX"
|
613 |
|
|
"")
|
614 |
|
|
|
615 |
|
|
(define_expand "cr6_test_for_lt_reverse"
|
616 |
|
|
[(set (match_operand:SI 0 "register_operand" "=r")
|
617 |
|
|
(lt:SI (reg:CC 74)
|
618 |
|
|
(const_int 0)))
|
619 |
|
|
(set (match_dup 0) (minus:SI (const_int 1) (match_dup 0)))]
|
620 |
|
|
"TARGET_ALTIVEC || TARGET_VSX"
|
621 |
|
|
"")
|
622 |
|
|
|
623 |
|
|
|
624 |
|
|
;; Vector logical instructions
|
625 |
|
|
(define_expand "xor3"
|
626 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
627 |
|
|
(xor:VEC_L (match_operand:VEC_L 1 "vlogical_operand" "")
|
628 |
|
|
(match_operand:VEC_L 2 "vlogical_operand" "")))]
|
629 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
630 |
|
|
"")
|
631 |
|
|
|
632 |
|
|
(define_expand "ior3"
|
633 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
634 |
|
|
(ior:VEC_L (match_operand:VEC_L 1 "vlogical_operand" "")
|
635 |
|
|
(match_operand:VEC_L 2 "vlogical_operand" "")))]
|
636 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
637 |
|
|
"")
|
638 |
|
|
|
639 |
|
|
(define_expand "and3"
|
640 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
641 |
|
|
(and:VEC_L (match_operand:VEC_L 1 "vlogical_operand" "")
|
642 |
|
|
(match_operand:VEC_L 2 "vlogical_operand" "")))]
|
643 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
644 |
|
|
"")
|
645 |
|
|
|
646 |
|
|
(define_expand "one_cmpl2"
|
647 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
648 |
|
|
(not:VEC_L (match_operand:VEC_L 1 "vlogical_operand" "")))]
|
649 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
650 |
|
|
"")
|
651 |
|
|
|
652 |
|
|
(define_expand "nor3"
|
653 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
654 |
|
|
(not:VEC_L (ior:VEC_L (match_operand:VEC_L 1 "vlogical_operand" "")
|
655 |
|
|
(match_operand:VEC_L 2 "vlogical_operand" ""))))]
|
656 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
657 |
|
|
"")
|
658 |
|
|
|
659 |
|
|
(define_expand "andc3"
|
660 |
|
|
[(set (match_operand:VEC_L 0 "vlogical_operand" "")
|
661 |
|
|
(and:VEC_L (not:VEC_L (match_operand:VEC_L 2 "vlogical_operand" ""))
|
662 |
|
|
(match_operand:VEC_L 1 "vlogical_operand" "")))]
|
663 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
664 |
|
|
"")
|
665 |
|
|
|
666 |
|
|
;; Same size conversions
|
667 |
|
|
(define_expand "float2"
|
668 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
669 |
|
|
(float:VEC_F (match_operand: 1 "vint_operand" "")))]
|
670 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
671 |
|
|
"
|
672 |
|
|
{
|
673 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
674 |
|
|
{
|
675 |
|
|
emit_insn (gen_altivec_vcfsx (operands[0], operands[1], const0_rtx));
|
676 |
|
|
DONE;
|
677 |
|
|
}
|
678 |
|
|
}")
|
679 |
|
|
|
680 |
|
|
(define_expand "floatuns2"
|
681 |
|
|
[(set (match_operand:VEC_F 0 "vfloat_operand" "")
|
682 |
|
|
(unsigned_float:VEC_F (match_operand: 1 "vint_operand" "")))]
|
683 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
684 |
|
|
"
|
685 |
|
|
{
|
686 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
687 |
|
|
{
|
688 |
|
|
emit_insn (gen_altivec_vcfux (operands[0], operands[1], const0_rtx));
|
689 |
|
|
DONE;
|
690 |
|
|
}
|
691 |
|
|
}")
|
692 |
|
|
|
693 |
|
|
(define_expand "fix_trunc2"
|
694 |
|
|
[(set (match_operand: 0 "vint_operand" "")
|
695 |
|
|
(fix: (match_operand:VEC_F 1 "vfloat_operand" "")))]
|
696 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
697 |
|
|
"
|
698 |
|
|
{
|
699 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
700 |
|
|
{
|
701 |
|
|
emit_insn (gen_altivec_vctsxs (operands[0], operands[1], const0_rtx));
|
702 |
|
|
DONE;
|
703 |
|
|
}
|
704 |
|
|
}")
|
705 |
|
|
|
706 |
|
|
(define_expand "fixuns_trunc2"
|
707 |
|
|
[(set (match_operand: 0 "vint_operand" "")
|
708 |
|
|
(unsigned_fix: (match_operand:VEC_F 1 "vfloat_operand" "")))]
|
709 |
|
|
"VECTOR_UNIT_ALTIVEC_OR_VSX_P (mode)"
|
710 |
|
|
"
|
711 |
|
|
{
|
712 |
|
|
if (mode == V4SFmode && VECTOR_UNIT_ALTIVEC_P (mode))
|
713 |
|
|
{
|
714 |
|
|
emit_insn (gen_altivec_vctuxs (operands[0], operands[1], const0_rtx));
|
715 |
|
|
DONE;
|
716 |
|
|
}
|
717 |
|
|
}")
|
718 |
|
|
|
719 |
|
|
|
720 |
|
|
;; Vector initialization, set, extract
|
721 |
|
|
(define_expand "vec_init"
|
722 |
|
|
[(match_operand:VEC_E 0 "vlogical_operand" "")
|
723 |
|
|
(match_operand:VEC_E 1 "" "")]
|
724 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
725 |
|
|
{
|
726 |
|
|
rs6000_expand_vector_init (operands[0], operands[1]);
|
727 |
|
|
DONE;
|
728 |
|
|
})
|
729 |
|
|
|
730 |
|
|
(define_expand "vec_set"
|
731 |
|
|
[(match_operand:VEC_E 0 "vlogical_operand" "")
|
732 |
|
|
(match_operand: 1 "register_operand" "")
|
733 |
|
|
(match_operand 2 "const_int_operand" "")]
|
734 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
735 |
|
|
{
|
736 |
|
|
rs6000_expand_vector_set (operands[0], operands[1], INTVAL (operands[2]));
|
737 |
|
|
DONE;
|
738 |
|
|
})
|
739 |
|
|
|
740 |
|
|
(define_expand "vec_extract"
|
741 |
|
|
[(match_operand: 0 "register_operand" "")
|
742 |
|
|
(match_operand:VEC_E 1 "vlogical_operand" "")
|
743 |
|
|
(match_operand 2 "const_int_operand" "")]
|
744 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
745 |
|
|
{
|
746 |
|
|
rs6000_expand_vector_extract (operands[0], operands[1],
|
747 |
|
|
INTVAL (operands[2]));
|
748 |
|
|
DONE;
|
749 |
|
|
})
|
750 |
|
|
|
751 |
|
|
;; Convert double word types to single word types
|
752 |
|
|
(define_expand "vec_pack_trunc_v2df"
|
753 |
|
|
[(match_operand:V4SF 0 "vfloat_operand" "")
|
754 |
|
|
(match_operand:V2DF 1 "vfloat_operand" "")
|
755 |
|
|
(match_operand:V2DF 2 "vfloat_operand" "")]
|
756 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && TARGET_ALTIVEC"
|
757 |
|
|
{
|
758 |
|
|
rtx r1 = gen_reg_rtx (V4SFmode);
|
759 |
|
|
rtx r2 = gen_reg_rtx (V4SFmode);
|
760 |
|
|
|
761 |
|
|
emit_insn (gen_vsx_xvcvdpsp (r1, operands[1]));
|
762 |
|
|
emit_insn (gen_vsx_xvcvdpsp (r2, operands[2]));
|
763 |
|
|
rs6000_expand_extract_even (operands[0], r1, r2);
|
764 |
|
|
DONE;
|
765 |
|
|
})
|
766 |
|
|
|
767 |
|
|
(define_expand "vec_pack_sfix_trunc_v2df"
|
768 |
|
|
[(match_operand:V4SI 0 "vint_operand" "")
|
769 |
|
|
(match_operand:V2DF 1 "vfloat_operand" "")
|
770 |
|
|
(match_operand:V2DF 2 "vfloat_operand" "")]
|
771 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && TARGET_ALTIVEC"
|
772 |
|
|
{
|
773 |
|
|
rtx r1 = gen_reg_rtx (V4SImode);
|
774 |
|
|
rtx r2 = gen_reg_rtx (V4SImode);
|
775 |
|
|
|
776 |
|
|
emit_insn (gen_vsx_xvcvdpsxws (r1, operands[1]));
|
777 |
|
|
emit_insn (gen_vsx_xvcvdpsxws (r2, operands[2]));
|
778 |
|
|
rs6000_expand_extract_even (operands[0], r1, r2);
|
779 |
|
|
DONE;
|
780 |
|
|
})
|
781 |
|
|
|
782 |
|
|
(define_expand "vec_pack_ufix_trunc_v2df"
|
783 |
|
|
[(match_operand:V4SI 0 "vint_operand" "")
|
784 |
|
|
(match_operand:V2DF 1 "vfloat_operand" "")
|
785 |
|
|
(match_operand:V2DF 2 "vfloat_operand" "")]
|
786 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && TARGET_ALTIVEC"
|
787 |
|
|
{
|
788 |
|
|
rtx r1 = gen_reg_rtx (V4SImode);
|
789 |
|
|
rtx r2 = gen_reg_rtx (V4SImode);
|
790 |
|
|
|
791 |
|
|
emit_insn (gen_vsx_xvcvdpuxws (r1, operands[1]));
|
792 |
|
|
emit_insn (gen_vsx_xvcvdpuxws (r2, operands[2]));
|
793 |
|
|
rs6000_expand_extract_even (operands[0], r1, r2);
|
794 |
|
|
DONE;
|
795 |
|
|
})
|
796 |
|
|
|
797 |
|
|
;; Convert single word types to double word
|
798 |
|
|
(define_expand "vec_unpacks_hi_v4sf"
|
799 |
|
|
[(match_operand:V2DF 0 "vfloat_operand" "")
|
800 |
|
|
(match_operand:V4SF 1 "vfloat_operand" "")]
|
801 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SFmode)"
|
802 |
|
|
{
|
803 |
|
|
rtx reg = gen_reg_rtx (V4SFmode);
|
804 |
|
|
|
805 |
|
|
rs6000_expand_interleave (reg, operands[1], operands[1], true);
|
806 |
|
|
emit_insn (gen_vsx_xvcvspdp (operands[0], reg));
|
807 |
|
|
DONE;
|
808 |
|
|
})
|
809 |
|
|
|
810 |
|
|
(define_expand "vec_unpacks_lo_v4sf"
|
811 |
|
|
[(match_operand:V2DF 0 "vfloat_operand" "")
|
812 |
|
|
(match_operand:V4SF 1 "vfloat_operand" "")]
|
813 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SFmode)"
|
814 |
|
|
{
|
815 |
|
|
rtx reg = gen_reg_rtx (V4SFmode);
|
816 |
|
|
|
817 |
|
|
rs6000_expand_interleave (reg, operands[1], operands[1], false);
|
818 |
|
|
emit_insn (gen_vsx_xvcvspdp (operands[0], reg));
|
819 |
|
|
DONE;
|
820 |
|
|
})
|
821 |
|
|
|
822 |
|
|
(define_expand "vec_unpacks_float_hi_v4si"
|
823 |
|
|
[(match_operand:V2DF 0 "vfloat_operand" "")
|
824 |
|
|
(match_operand:V4SI 1 "vint_operand" "")]
|
825 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SImode)"
|
826 |
|
|
{
|
827 |
|
|
rtx reg = gen_reg_rtx (V4SImode);
|
828 |
|
|
|
829 |
|
|
rs6000_expand_interleave (reg, operands[1], operands[1], true);
|
830 |
|
|
emit_insn (gen_vsx_xvcvsxwdp (operands[0], reg));
|
831 |
|
|
DONE;
|
832 |
|
|
})
|
833 |
|
|
|
834 |
|
|
(define_expand "vec_unpacks_float_lo_v4si"
|
835 |
|
|
[(match_operand:V2DF 0 "vfloat_operand" "")
|
836 |
|
|
(match_operand:V4SI 1 "vint_operand" "")]
|
837 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SImode)"
|
838 |
|
|
{
|
839 |
|
|
rtx reg = gen_reg_rtx (V4SImode);
|
840 |
|
|
|
841 |
|
|
rs6000_expand_interleave (reg, operands[1], operands[1], false);
|
842 |
|
|
emit_insn (gen_vsx_xvcvsxwdp (operands[0], reg));
|
843 |
|
|
DONE;
|
844 |
|
|
})
|
845 |
|
|
|
846 |
|
|
(define_expand "vec_unpacku_float_hi_v4si"
|
847 |
|
|
[(match_operand:V2DF 0 "vfloat_operand" "")
|
848 |
|
|
(match_operand:V4SI 1 "vint_operand" "")]
|
849 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SImode)"
|
850 |
|
|
{
|
851 |
|
|
rtx reg = gen_reg_rtx (V4SImode);
|
852 |
|
|
|
853 |
|
|
rs6000_expand_interleave (reg, operands[1], operands[1], true);
|
854 |
|
|
emit_insn (gen_vsx_xvcvuxwdp (operands[0], reg));
|
855 |
|
|
DONE;
|
856 |
|
|
})
|
857 |
|
|
|
858 |
|
|
(define_expand "vec_unpacku_float_lo_v4si"
|
859 |
|
|
[(match_operand:V2DF 0 "vfloat_operand" "")
|
860 |
|
|
(match_operand:V4SI 1 "vint_operand" "")]
|
861 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode) && VECTOR_UNIT_ALTIVEC_OR_VSX_P (V4SImode)"
|
862 |
|
|
{
|
863 |
|
|
rtx reg = gen_reg_rtx (V4SImode);
|
864 |
|
|
|
865 |
|
|
rs6000_expand_interleave (reg, operands[1], operands[1], false);
|
866 |
|
|
emit_insn (gen_vsx_xvcvuxwdp (operands[0], reg));
|
867 |
|
|
DONE;
|
868 |
|
|
})
|
869 |
|
|
|
870 |
|
|
|
871 |
|
|
;; Align vector loads with a permute.
|
872 |
|
|
(define_expand "vec_realign_load_"
|
873 |
|
|
[(match_operand:VEC_K 0 "vlogical_operand" "")
|
874 |
|
|
(match_operand:VEC_K 1 "vlogical_operand" "")
|
875 |
|
|
(match_operand:VEC_K 2 "vlogical_operand" "")
|
876 |
|
|
(match_operand:V16QI 3 "vlogical_operand" "")]
|
877 |
|
|
"VECTOR_MEM_ALTIVEC_OR_VSX_P (mode)"
|
878 |
|
|
{
|
879 |
|
|
emit_insn (gen_altivec_vperm_ (operands[0], operands[1], operands[2],
|
880 |
|
|
operands[3]));
|
881 |
|
|
DONE;
|
882 |
|
|
})
|
883 |
|
|
|
884 |
|
|
;; Under VSX, vectors of 4/8 byte alignments do not need to be aligned
|
885 |
|
|
;; since the load already handles it.
|
886 |
|
|
(define_expand "movmisalign"
|
887 |
|
|
[(set (match_operand:VEC_N 0 "nonimmediate_operand" "")
|
888 |
|
|
(match_operand:VEC_N 1 "any_operand" ""))]
|
889 |
|
|
"VECTOR_MEM_VSX_P (mode) && TARGET_ALLOW_MOVMISALIGN"
|
890 |
|
|
"")
|
891 |
|
|
|
892 |
|
|
|
893 |
|
|
;; Vector shift left in bits. Currently supported ony for shift
|
894 |
|
|
;; amounts that can be expressed as byte shifts (divisible by 8).
|
895 |
|
|
;; General shift amounts can be supported using vslo + vsl. We're
|
896 |
|
|
;; not expecting to see these yet (the vectorizer currently
|
897 |
|
|
;; generates only shifts divisible by byte_size).
|
898 |
|
|
(define_expand "vec_shl_"
|
899 |
|
|
[(match_operand:VEC_L 0 "vlogical_operand" "")
|
900 |
|
|
(match_operand:VEC_L 1 "vlogical_operand" "")
|
901 |
|
|
(match_operand:QI 2 "reg_or_short_operand" "")]
|
902 |
|
|
"TARGET_ALTIVEC"
|
903 |
|
|
"
|
904 |
|
|
{
|
905 |
|
|
rtx bitshift = operands[2];
|
906 |
|
|
rtx shift;
|
907 |
|
|
rtx insn;
|
908 |
|
|
HOST_WIDE_INT bitshift_val;
|
909 |
|
|
HOST_WIDE_INT byteshift_val;
|
910 |
|
|
|
911 |
|
|
if (! CONSTANT_P (bitshift))
|
912 |
|
|
FAIL;
|
913 |
|
|
bitshift_val = INTVAL (bitshift);
|
914 |
|
|
if (bitshift_val & 0x7)
|
915 |
|
|
FAIL;
|
916 |
|
|
byteshift_val = bitshift_val >> 3;
|
917 |
|
|
if (TARGET_VSX && (byteshift_val & 0x3) == 0)
|
918 |
|
|
{
|
919 |
|
|
shift = gen_rtx_CONST_INT (QImode, byteshift_val >> 2);
|
920 |
|
|
insn = gen_vsx_xxsldwi_ (operands[0], operands[1], operands[1],
|
921 |
|
|
shift);
|
922 |
|
|
}
|
923 |
|
|
else
|
924 |
|
|
{
|
925 |
|
|
shift = gen_rtx_CONST_INT (QImode, byteshift_val);
|
926 |
|
|
insn = gen_altivec_vsldoi_ (operands[0], operands[1], operands[1],
|
927 |
|
|
shift);
|
928 |
|
|
}
|
929 |
|
|
|
930 |
|
|
emit_insn (insn);
|
931 |
|
|
DONE;
|
932 |
|
|
}")
|
933 |
|
|
|
934 |
|
|
;; Vector shift right in bits. Currently supported ony for shift
|
935 |
|
|
;; amounts that can be expressed as byte shifts (divisible by 8).
|
936 |
|
|
;; General shift amounts can be supported using vsro + vsr. We're
|
937 |
|
|
;; not expecting to see these yet (the vectorizer currently
|
938 |
|
|
;; generates only shifts divisible by byte_size).
|
939 |
|
|
(define_expand "vec_shr_"
|
940 |
|
|
[(match_operand:VEC_L 0 "vlogical_operand" "")
|
941 |
|
|
(match_operand:VEC_L 1 "vlogical_operand" "")
|
942 |
|
|
(match_operand:QI 2 "reg_or_short_operand" "")]
|
943 |
|
|
"TARGET_ALTIVEC"
|
944 |
|
|
"
|
945 |
|
|
{
|
946 |
|
|
rtx bitshift = operands[2];
|
947 |
|
|
rtx shift;
|
948 |
|
|
rtx insn;
|
949 |
|
|
HOST_WIDE_INT bitshift_val;
|
950 |
|
|
HOST_WIDE_INT byteshift_val;
|
951 |
|
|
|
952 |
|
|
if (! CONSTANT_P (bitshift))
|
953 |
|
|
FAIL;
|
954 |
|
|
bitshift_val = INTVAL (bitshift);
|
955 |
|
|
if (bitshift_val & 0x7)
|
956 |
|
|
FAIL;
|
957 |
|
|
byteshift_val = 16 - (bitshift_val >> 3);
|
958 |
|
|
if (TARGET_VSX && (byteshift_val & 0x3) == 0)
|
959 |
|
|
{
|
960 |
|
|
shift = gen_rtx_CONST_INT (QImode, byteshift_val >> 2);
|
961 |
|
|
insn = gen_vsx_xxsldwi_ (operands[0], operands[1], operands[1],
|
962 |
|
|
shift);
|
963 |
|
|
}
|
964 |
|
|
else
|
965 |
|
|
{
|
966 |
|
|
shift = gen_rtx_CONST_INT (QImode, byteshift_val);
|
967 |
|
|
insn = gen_altivec_vsldoi_ (operands[0], operands[1], operands[1],
|
968 |
|
|
shift);
|
969 |
|
|
}
|
970 |
|
|
|
971 |
|
|
emit_insn (insn);
|
972 |
|
|
DONE;
|
973 |
|
|
}")
|
974 |
|
|
|
975 |
|
|
;; Expanders for rotate each element in a vector
|
976 |
|
|
(define_expand "vrotl3"
|
977 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
978 |
|
|
(rotate:VEC_I (match_operand:VEC_I 1 "vint_operand" "")
|
979 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
980 |
|
|
"TARGET_ALTIVEC"
|
981 |
|
|
"")
|
982 |
|
|
|
983 |
|
|
;; Expanders for arithmetic shift left on each vector element
|
984 |
|
|
(define_expand "vashl3"
|
985 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
986 |
|
|
(ashift:VEC_I (match_operand:VEC_I 1 "vint_operand" "")
|
987 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
988 |
|
|
"TARGET_ALTIVEC"
|
989 |
|
|
"")
|
990 |
|
|
|
991 |
|
|
;; Expanders for logical shift right on each vector element
|
992 |
|
|
(define_expand "vlshr3"
|
993 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
994 |
|
|
(lshiftrt:VEC_I (match_operand:VEC_I 1 "vint_operand" "")
|
995 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
996 |
|
|
"TARGET_ALTIVEC"
|
997 |
|
|
"")
|
998 |
|
|
|
999 |
|
|
;; Expanders for arithmetic shift right on each vector element
|
1000 |
|
|
(define_expand "vashr3"
|
1001 |
|
|
[(set (match_operand:VEC_I 0 "vint_operand" "")
|
1002 |
|
|
(ashiftrt:VEC_I (match_operand:VEC_I 1 "vint_operand" "")
|
1003 |
|
|
(match_operand:VEC_I 2 "vint_operand" "")))]
|
1004 |
|
|
"TARGET_ALTIVEC"
|
1005 |
|
|
"")
|
1006 |
|
|
|
1007 |
|
|
;; Vector reduction expanders for VSX
|
1008 |
|
|
|
1009 |
|
|
(define_expand "reduc__v2df"
|
1010 |
|
|
[(parallel [(set (match_operand:V2DF 0 "vfloat_operand" "")
|
1011 |
|
|
(VEC_reduc:V2DF
|
1012 |
|
|
(vec_concat:V2DF
|
1013 |
|
|
(vec_select:DF
|
1014 |
|
|
(match_operand:V2DF 1 "vfloat_operand" "")
|
1015 |
|
|
(parallel [(const_int 1)]))
|
1016 |
|
|
(vec_select:DF
|
1017 |
|
|
(match_dup 1)
|
1018 |
|
|
(parallel [(const_int 0)])))
|
1019 |
|
|
(match_dup 1)))
|
1020 |
|
|
(clobber (match_scratch:V2DF 2 ""))])]
|
1021 |
|
|
"VECTOR_UNIT_VSX_P (V2DFmode)"
|
1022 |
|
|
"")
|
1023 |
|
|
|
1024 |
|
|
; The (VEC_reduc:V4SF
|
1025 |
|
|
; (op1)
|
1026 |
|
|
; (unspec:V4SF [(const_int 0)] UNSPEC_REDUC))
|
1027 |
|
|
;
|
1028 |
|
|
; is to allow us to use a code iterator, but not completely list all of the
|
1029 |
|
|
; vector rotates, etc. to prevent canonicalization
|
1030 |
|
|
|
1031 |
|
|
(define_expand "reduc__v4sf"
|
1032 |
|
|
[(parallel [(set (match_operand:V4SF 0 "vfloat_operand" "")
|
1033 |
|
|
(VEC_reduc:V4SF
|
1034 |
|
|
(unspec:V4SF [(const_int 0)] UNSPEC_REDUC)
|
1035 |
|
|
(match_operand:V4SF 1 "vfloat_operand" "")))
|
1036 |
|
|
(clobber (match_scratch:V4SF 2 ""))
|
1037 |
|
|
(clobber (match_scratch:V4SF 3 ""))])]
|
1038 |
|
|
"VECTOR_UNIT_VSX_P (V4SFmode)"
|
1039 |
|
|
"")
|
1040 |
|
|
|
1041 |
|
|
|
1042 |
|
|
;;; Expanders for vector insn patterns shared between the SPE and TARGET_PAIRED systems.
|
1043 |
|
|
|
1044 |
|
|
(define_expand "absv2sf2"
|
1045 |
|
|
[(set (match_operand:V2SF 0 "gpc_reg_operand" "")
|
1046 |
|
|
(abs:V2SF (match_operand:V2SF 1 "gpc_reg_operand" "")))]
|
1047 |
|
|
"TARGET_PAIRED_FLOAT || TARGET_SPE"
|
1048 |
|
|
"")
|
1049 |
|
|
|
1050 |
|
|
(define_expand "negv2sf2"
|
1051 |
|
|
[(set (match_operand:V2SF 0 "gpc_reg_operand" "")
|
1052 |
|
|
(neg:V2SF (match_operand:V2SF 1 "gpc_reg_operand" "")))]
|
1053 |
|
|
"TARGET_PAIRED_FLOAT || TARGET_SPE"
|
1054 |
|
|
"")
|
1055 |
|
|
|
1056 |
|
|
(define_expand "addv2sf3"
|
1057 |
|
|
[(set (match_operand:V2SF 0 "gpc_reg_operand" "")
|
1058 |
|
|
(plus:V2SF (match_operand:V2SF 1 "gpc_reg_operand" "")
|
1059 |
|
|
(match_operand:V2SF 2 "gpc_reg_operand" "")))]
|
1060 |
|
|
"TARGET_PAIRED_FLOAT || TARGET_SPE"
|
1061 |
|
|
"
|
1062 |
|
|
{
|
1063 |
|
|
if (TARGET_SPE)
|
1064 |
|
|
{
|
1065 |
|
|
/* We need to make a note that we clobber SPEFSCR. */
|
1066 |
|
|
rtx par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (2));
|
1067 |
|
|
|
1068 |
|
|
XVECEXP (par, 0, 0) = gen_rtx_SET (VOIDmode, operands[0],
|
1069 |
|
|
gen_rtx_PLUS (V2SFmode, operands[1], operands[2]));
|
1070 |
|
|
XVECEXP (par, 0, 1) = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, SPEFSCR_REGNO));
|
1071 |
|
|
emit_insn (par);
|
1072 |
|
|
DONE;
|
1073 |
|
|
}
|
1074 |
|
|
}")
|
1075 |
|
|
|
1076 |
|
|
(define_expand "subv2sf3"
|
1077 |
|
|
[(set (match_operand:V2SF 0 "gpc_reg_operand" "")
|
1078 |
|
|
(minus:V2SF (match_operand:V2SF 1 "gpc_reg_operand" "")
|
1079 |
|
|
(match_operand:V2SF 2 "gpc_reg_operand" "")))]
|
1080 |
|
|
"TARGET_PAIRED_FLOAT || TARGET_SPE"
|
1081 |
|
|
"
|
1082 |
|
|
{
|
1083 |
|
|
if (TARGET_SPE)
|
1084 |
|
|
{
|
1085 |
|
|
/* We need to make a note that we clobber SPEFSCR. */
|
1086 |
|
|
rtx par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (2));
|
1087 |
|
|
|
1088 |
|
|
XVECEXP (par, 0, 0) = gen_rtx_SET (VOIDmode, operands[0],
|
1089 |
|
|
gen_rtx_MINUS (V2SFmode, operands[1], operands[2]));
|
1090 |
|
|
XVECEXP (par, 0, 1) = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, SPEFSCR_REGNO));
|
1091 |
|
|
emit_insn (par);
|
1092 |
|
|
DONE;
|
1093 |
|
|
}
|
1094 |
|
|
}")
|
1095 |
|
|
|
1096 |
|
|
(define_expand "mulv2sf3"
|
1097 |
|
|
[(set (match_operand:V2SF 0 "gpc_reg_operand" "")
|
1098 |
|
|
(mult:V2SF (match_operand:V2SF 1 "gpc_reg_operand" "")
|
1099 |
|
|
(match_operand:V2SF 2 "gpc_reg_operand" "")))]
|
1100 |
|
|
"TARGET_PAIRED_FLOAT || TARGET_SPE"
|
1101 |
|
|
"
|
1102 |
|
|
{
|
1103 |
|
|
if (TARGET_SPE)
|
1104 |
|
|
{
|
1105 |
|
|
/* We need to make a note that we clobber SPEFSCR. */
|
1106 |
|
|
rtx par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (2));
|
1107 |
|
|
|
1108 |
|
|
XVECEXP (par, 0, 0) = gen_rtx_SET (VOIDmode, operands[0],
|
1109 |
|
|
gen_rtx_MULT (V2SFmode, operands[1], operands[2]));
|
1110 |
|
|
XVECEXP (par, 0, 1) = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, SPEFSCR_REGNO));
|
1111 |
|
|
emit_insn (par);
|
1112 |
|
|
DONE;
|
1113 |
|
|
}
|
1114 |
|
|
}")
|
1115 |
|
|
|
1116 |
|
|
(define_expand "divv2sf3"
|
1117 |
|
|
[(set (match_operand:V2SF 0 "gpc_reg_operand" "")
|
1118 |
|
|
(div:V2SF (match_operand:V2SF 1 "gpc_reg_operand" "")
|
1119 |
|
|
(match_operand:V2SF 2 "gpc_reg_operand" "")))]
|
1120 |
|
|
"TARGET_PAIRED_FLOAT || TARGET_SPE"
|
1121 |
|
|
"
|
1122 |
|
|
{
|
1123 |
|
|
if (TARGET_SPE)
|
1124 |
|
|
{
|
1125 |
|
|
/* We need to make a note that we clobber SPEFSCR. */
|
1126 |
|
|
rtx par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (2));
|
1127 |
|
|
|
1128 |
|
|
XVECEXP (par, 0, 0) = gen_rtx_SET (VOIDmode, operands[0],
|
1129 |
|
|
gen_rtx_DIV (V2SFmode, operands[1], operands[2]));
|
1130 |
|
|
XVECEXP (par, 0, 1) = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, SPEFSCR_REGNO));
|
1131 |
|
|
emit_insn (par);
|
1132 |
|
|
DONE;
|
1133 |
|
|
}
|
1134 |
|
|
}")
|