1 /*
2 * Copyright (c) 2007, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
23
24 #include "precompiled.hpp"
25 #include "memory/allocation.inline.hpp"
26 #include "opto/connode.hpp"
27 #include "opto/vectornode.hpp"
28
29 //------------------------------VectorNode--------------------------------------
30
31 // Return the vector operator for the specified scalar operation
32 // and vector length.
33 int VectorNode::opcode(int sopc, BasicType bt) {
34 switch (sopc) {
35 case Op_AddI:
36 switch (bt) {
37 case T_BOOLEAN:
38 case T_BYTE: return Op_AddVB;
39 case T_CHAR:
40 case T_SHORT: return Op_AddVS;
41 case T_INT: return Op_AddVI;
42 }
43 ShouldNotReachHere();
44 case Op_AddL:
45 assert(bt == T_LONG, "must be");
46 return Op_AddVL;
47 case Op_AddF:
48 assert(bt == T_FLOAT, "must be");
49 return Op_AddVF;
50 case Op_AddD:
51 assert(bt == T_DOUBLE, "must be");
52 return Op_AddVD;
53 case Op_SubI:
54 switch (bt) {
55 case T_BOOLEAN:
56 case T_BYTE: return Op_SubVB;
57 case T_CHAR:
58 case T_SHORT: return Op_SubVS;
59 case T_INT: return Op_SubVI;
60 }
61 ShouldNotReachHere();
62 case Op_SubL:
63 assert(bt == T_LONG, "must be");
64 return Op_SubVL;
65 case Op_SubF:
66 assert(bt == T_FLOAT, "must be");
67 return Op_SubVF;
68 case Op_SubD:
69 assert(bt == T_DOUBLE, "must be");
70 return Op_SubVD;
71 case Op_MulI:
72 switch (bt) {
73 case T_BOOLEAN:
74 case T_BYTE: return 0; // Unimplemented
75 case T_CHAR:
76 case T_SHORT: return Op_MulVS;
77 case T_INT: return Op_MulVI;
78 }
79 ShouldNotReachHere();
80 case Op_MulL:
81 assert(bt == T_LONG, "must be");
82 return Op_MulVL;
83 case Op_MulF:
84 assert(bt == T_FLOAT, "must be");
85 return Op_MulVF;
86 case Op_MulD:
87 assert(bt == T_DOUBLE, "must be");
88 return Op_MulVD;
89 case Op_DivF:
90 assert(bt == T_FLOAT, "must be");
91 return Op_DivVF;
92 case Op_DivD:
93 assert(bt == T_DOUBLE, "must be");
94 return Op_DivVD;
95 case Op_AbsF:
96 assert(bt == T_FLOAT, "must be");
97 return Op_AbsVF;
98 case Op_AbsD:
99 assert(bt == T_DOUBLE, "must be");
100 return Op_AbsVD;
101 case Op_NegF:
102 assert(bt == T_FLOAT, "must be");
103 return Op_NegVF;
104 case Op_NegD:
105 assert(bt == T_DOUBLE, "must be");
106 return Op_NegVD;
107 case Op_SqrtD:
108 assert(bt == T_DOUBLE, "must be");
109 return Op_SqrtVD;
110 case Op_LShiftI:
111 switch (bt) {
112 case T_BOOLEAN:
113 case T_BYTE: return Op_LShiftVB;
114 case T_CHAR:
115 case T_SHORT: return Op_LShiftVS;
116 case T_INT: return Op_LShiftVI;
117 }
118 ShouldNotReachHere();
119 case Op_LShiftL:
120 assert(bt == T_LONG, "must be");
121 return Op_LShiftVL;
122 case Op_RShiftI:
123 switch (bt) {
124 case T_BOOLEAN:return Op_URShiftVB; // boolean is unsigned value
125 case T_CHAR: return Op_URShiftVS; // char is unsigned value
126 case T_BYTE: return Op_RShiftVB;
127 case T_SHORT: return Op_RShiftVS;
128 case T_INT: return Op_RShiftVI;
129 }
130 ShouldNotReachHere();
131 case Op_RShiftL:
132 assert(bt == T_LONG, "must be");
133 return Op_RShiftVL;
134 case Op_URShiftI:
135 switch (bt) {
136 case T_BOOLEAN:return Op_URShiftVB;
137 case T_CHAR: return Op_URShiftVS;
138 case T_BYTE:
139 case T_SHORT: return 0; // Vector logical right shift for signed short
140 // values produces incorrect Java result for
141 // negative data because java code should convert
142 // a short value into int value with sign
143 // extension before a shift.
144 case T_INT: return Op_URShiftVI;
145 }
146 ShouldNotReachHere();
147 case Op_URShiftL:
148 assert(bt == T_LONG, "must be");
149 return Op_URShiftVL;
150 case Op_AndI:
151 case Op_AndL:
152 return Op_AndV;
153 case Op_OrI:
154 case Op_OrL:
155 return Op_OrV;
156 case Op_XorI:
157 case Op_XorL:
158 return Op_XorV;
159
160 case Op_LoadB:
161 case Op_LoadUB:
162 case Op_LoadUS:
163 case Op_LoadS:
164 case Op_LoadI:
165 case Op_LoadL:
166 case Op_LoadF:
167 case Op_LoadD:
168 return Op_LoadVector;
169
170 case Op_StoreB:
171 case Op_StoreC:
172 case Op_StoreI:
173 case Op_StoreL:
174 case Op_StoreF:
175 case Op_StoreD:
176 return Op_StoreVector;
177 }
178 return 0; // Unimplemented
179 }
180
181 // Also used to check if the code generator
182 // supports the vector operation.
183 bool VectorNode::implemented(int opc, uint vlen, BasicType bt) {
184 if (is_java_primitive(bt) &&
185 (vlen > 1) && is_power_of_2(vlen) &&
186 Matcher::vector_size_supported(bt, vlen)) {
187 int vopc = VectorNode::opcode(opc, bt);
188 return vopc > 0 && Matcher::match_rule_supported(vopc);
189 }
190 return false;
191 }
192
193 bool VectorNode::is_shift(Node* n) {
194 switch (n->Opcode()) {
195 case Op_LShiftI:
196 case Op_LShiftL:
197 case Op_RShiftI:
198 case Op_RShiftL:
199 case Op_URShiftI:
200 case Op_URShiftL:
201 return true;
202 }
203 return false;
204 }
205
206 // Check if input is loop invariant vector.
207 bool VectorNode::is_invariant_vector(Node* n) {
208 // Only Replicate vector nodes are loop invariant for now.
209 switch (n->Opcode()) {
210 case Op_ReplicateB:
211 case Op_ReplicateS:
212 case Op_ReplicateI:
213 case Op_ReplicateL:
214 case Op_ReplicateF:
215 case Op_ReplicateD:
216 return true;
217 }
218 return false;
219 }
220
221 // [Start, end) half-open range defining which operands are vectors
222 void VectorNode::vector_operands(Node* n, uint* start, uint* end) {
223 switch (n->Opcode()) {
224 case Op_LoadB: case Op_LoadUB:
225 case Op_LoadS: case Op_LoadUS:
226 case Op_LoadI: case Op_LoadL:
227 case Op_LoadF: case Op_LoadD:
228 case Op_LoadP: case Op_LoadN:
229 *start = 0;
230 *end = 0; // no vector operands
231 break;
232 case Op_StoreB: case Op_StoreC:
233 case Op_StoreI: case Op_StoreL:
234 case Op_StoreF: case Op_StoreD:
235 case Op_StoreP: case Op_StoreN:
236 *start = MemNode::ValueIn;
237 *end = MemNode::ValueIn + 1; // 1 vector operand
238 break;
239 case Op_LShiftI: case Op_LShiftL:
240 case Op_RShiftI: case Op_RShiftL:
241 case Op_URShiftI: case Op_URShiftL:
242 *start = 1;
243 *end = 2; // 1 vector operand
244 break;
245 case Op_AddI: case Op_AddL: case Op_AddF: case Op_AddD:
246 case Op_SubI: case Op_SubL: case Op_SubF: case Op_SubD:
247 case Op_MulI: case Op_MulL: case Op_MulF: case Op_MulD:
248 case Op_DivF: case Op_DivD:
249 case Op_AndI: case Op_AndL:
250 case Op_OrI: case Op_OrL:
251 case Op_XorI: case Op_XorL:
252 *start = 1;
253 *end = 3; // 2 vector operands
254 break;
255 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
256 *start = 2;
257 *end = n->req();
258 break;
259 default:
260 *start = 1;
261 *end = n->req(); // default is all operands
262 }
263 }
264
265 // Return the vector version of a scalar operation node.
266 VectorNode* VectorNode::make(int opc, Node* n1, Node* n2, uint vlen, BasicType bt) {
267 const TypeVect* vt = TypeVect::make(bt, vlen);
268 int vopc = VectorNode::opcode(opc, bt);
269 // This method should not be called for unimplemented vectors.
270 guarantee(vopc > 0, err_msg_res("Vector for '%s' is not implemented", NodeClassNames[opc]));
271 switch (vopc) {
272 case Op_AddVB: return new AddVBNode(n1, n2, vt);
273 case Op_AddVS: return new AddVSNode(n1, n2, vt);
274 case Op_AddVI: return new AddVINode(n1, n2, vt);
275 case Op_AddVL: return new AddVLNode(n1, n2, vt);
276 case Op_AddVF: return new AddVFNode(n1, n2, vt);
277 case Op_AddVD: return new AddVDNode(n1, n2, vt);
278
279 case Op_SubVB: return new SubVBNode(n1, n2, vt);
280 case Op_SubVS: return new SubVSNode(n1, n2, vt);
281 case Op_SubVI: return new SubVINode(n1, n2, vt);
282 case Op_SubVL: return new SubVLNode(n1, n2, vt);
283 case Op_SubVF: return new SubVFNode(n1, n2, vt);
284 case Op_SubVD: return new SubVDNode(n1, n2, vt);
285
286 case Op_MulVS: return new MulVSNode(n1, n2, vt);
287 case Op_MulVI: return new MulVINode(n1, n2, vt);
288 case Op_MulVL: return new MulVLNode(n1, n2, vt);
289 case Op_MulVF: return new MulVFNode(n1, n2, vt);
290 case Op_MulVD: return new MulVDNode(n1, n2, vt);
291
292 case Op_DivVF: return new DivVFNode(n1, n2, vt);
293 case Op_DivVD: return new DivVDNode(n1, n2, vt);
294
295 case Op_AbsVF: return new AbsVFNode(n1, vt);
296 case Op_AbsVD: return new AbsVDNode(n1, vt);
297
298 case Op_NegVF: return new NegVFNode(n1, vt);
299 case Op_NegVD: return new NegVDNode(n1, vt);
300
301 // Currently only supports double precision sqrt
302 case Op_SqrtVD: return new SqrtVDNode(n1, vt);
303
304 case Op_LShiftVB: return new LShiftVBNode(n1, n2, vt);
305 case Op_LShiftVS: return new LShiftVSNode(n1, n2, vt);
306 case Op_LShiftVI: return new LShiftVINode(n1, n2, vt);
307 case Op_LShiftVL: return new LShiftVLNode(n1, n2, vt);
308
309 case Op_RShiftVB: return new RShiftVBNode(n1, n2, vt);
310 case Op_RShiftVS: return new RShiftVSNode(n1, n2, vt);
311 case Op_RShiftVI: return new RShiftVINode(n1, n2, vt);
312 case Op_RShiftVL: return new RShiftVLNode(n1, n2, vt);
313
314 case Op_URShiftVB: return new URShiftVBNode(n1, n2, vt);
315 case Op_URShiftVS: return new URShiftVSNode(n1, n2, vt);
316 case Op_URShiftVI: return new URShiftVINode(n1, n2, vt);
317 case Op_URShiftVL: return new URShiftVLNode(n1, n2, vt);
318
319 case Op_AndV: return new AndVNode(n1, n2, vt);
320 case Op_OrV: return new OrVNode (n1, n2, vt);
321 case Op_XorV: return new XorVNode(n1, n2, vt);
322 }
323 fatal(err_msg_res("Missed vector creation for '%s'", NodeClassNames[vopc]));
324 return NULL;
325
326 }
327
328 // Scalar promotion
329 VectorNode* VectorNode::scalar2vector(Node* s, uint vlen, const Type* opd_t) {
330 BasicType bt = opd_t->array_element_basic_type();
331 const TypeVect* vt = opd_t->singleton() ? TypeVect::make(opd_t, vlen)
332 : TypeVect::make(bt, vlen);
333 switch (bt) {
334 case T_BOOLEAN:
335 case T_BYTE:
336 return new ReplicateBNode(s, vt);
337 case T_CHAR:
338 case T_SHORT:
339 return new ReplicateSNode(s, vt);
340 case T_INT:
341 return new ReplicateINode(s, vt);
342 case T_LONG:
343 return new ReplicateLNode(s, vt);
344 case T_FLOAT:
345 return new ReplicateFNode(s, vt);
346 case T_DOUBLE:
347 return new ReplicateDNode(s, vt);
348 }
349 fatal(err_msg_res("Type '%s' is not supported for vectors", type2name(bt)));
350 return NULL;
351 }
352
353 VectorNode* VectorNode::shift_count(Node* shift, Node* cnt, uint vlen, BasicType bt) {
354 assert(VectorNode::is_shift(shift) && !cnt->is_Con(), "only variable shift count");
355 // Match shift count type with shift vector type.
356 const TypeVect* vt = TypeVect::make(bt, vlen);
357 switch (shift->Opcode()) {
358 case Op_LShiftI:
359 case Op_LShiftL:
360 return new LShiftCntVNode(cnt, vt);
361 case Op_RShiftI:
362 case Op_RShiftL:
363 case Op_URShiftI:
364 case Op_URShiftL:
365 return new RShiftCntVNode(cnt, vt);
366 }
367 fatal(err_msg_res("Missed vector creation for '%s'", NodeClassNames[shift->Opcode()]));
368 return NULL;
369 }
370
371 // Return initial Pack node. Additional operands added with add_opd() calls.
372 PackNode* PackNode::make(Node* s, uint vlen, BasicType bt) {
373 const TypeVect* vt = TypeVect::make(bt, vlen);
374 switch (bt) {
375 case T_BOOLEAN:
376 case T_BYTE:
377 return new PackBNode(s, vt);
378 case T_CHAR:
379 case T_SHORT:
380 return new PackSNode(s, vt);
381 case T_INT:
382 return new PackINode(s, vt);
383 case T_LONG:
384 return new PackLNode(s, vt);
385 case T_FLOAT:
386 return new PackFNode(s, vt);
387 case T_DOUBLE:
388 return new PackDNode(s, vt);
389 }
390 fatal(err_msg_res("Type '%s' is not supported for vectors", type2name(bt)));
391 return NULL;
392 }
393
394 // Create a binary tree form for Packs. [lo, hi) (half-open) range
395 PackNode* PackNode::binary_tree_pack(int lo, int hi) {
396 int ct = hi - lo;
397 assert(is_power_of_2(ct), "power of 2");
398 if (ct == 2) {
399 PackNode* pk = PackNode::make(in(lo), 2, vect_type()->element_basic_type());
400 pk->add_opd(in(lo+1));
401 return pk;
402
403 } else {
404 int mid = lo + ct/2;
405 PackNode* n1 = binary_tree_pack(lo, mid);
406 PackNode* n2 = binary_tree_pack(mid, hi );
407
408 BasicType bt = n1->vect_type()->element_basic_type();
409 assert(bt == n2->vect_type()->element_basic_type(), "should be the same");
410 switch (bt) {
411 case T_BOOLEAN:
412 case T_BYTE:
413 return new PackSNode(n1, n2, TypeVect::make(T_SHORT, 2));
414 case T_CHAR:
415 case T_SHORT:
416 return new PackINode(n1, n2, TypeVect::make(T_INT, 2));
417 case T_INT:
418 return new PackLNode(n1, n2, TypeVect::make(T_LONG, 2));
419 case T_LONG:
420 return new Pack2LNode(n1, n2, TypeVect::make(T_LONG, 2));
421 case T_FLOAT:
422 return new PackDNode(n1, n2, TypeVect::make(T_DOUBLE, 2));
423 case T_DOUBLE:
424 return new Pack2DNode(n1, n2, TypeVect::make(T_DOUBLE, 2));
425 }
426 fatal(err_msg_res("Type '%s' is not supported for vectors", type2name(bt)));
427 }
428 return NULL;
429 }
430
431 // Return the vector version of a scalar load node.
432 LoadVectorNode* LoadVectorNode::make(int opc, Node* ctl, Node* mem,
433 Node* adr, const TypePtr* atyp,
434 uint vlen, BasicType bt,
435 ControlDependency control_dependency) {
436 const TypeVect* vt = TypeVect::make(bt, vlen);
437 return new LoadVectorNode(ctl, mem, adr, atyp, vt, control_dependency);
438 }
439
440 // Return the vector version of a scalar store node.
441 StoreVectorNode* StoreVectorNode::make(int opc, Node* ctl, Node* mem,
442 Node* adr, const TypePtr* atyp, Node* val,
443 uint vlen) {
444 return new StoreVectorNode(ctl, mem, adr, atyp, val);
445 }
446
447 // Extract a scalar element of vector.
448 Node* ExtractNode::make(Node* v, uint position, BasicType bt) {
449 assert((int)position < Matcher::max_vector_size(bt), "pos in range");
450 ConINode* pos = ConINode::make((int)position);
451 switch (bt) {
452 case T_BOOLEAN:
453 return new ExtractUBNode(v, pos);
454 case T_BYTE:
455 return new ExtractBNode(v, pos);
456 case T_CHAR:
457 return new ExtractCNode(v, pos);
458 case T_SHORT:
459 return new ExtractSNode(v, pos);
460 case T_INT:
461 return new ExtractINode(v, pos);
462 case T_LONG:
463 return new ExtractLNode(v, pos);
464 case T_FLOAT:
465 return new ExtractFNode(v, pos);
466 case T_DOUBLE:
467 return new ExtractDNode(v, pos);
468 }
469 fatal(err_msg_res("Type '%s' is not supported for vectors", type2name(bt)));
470 return NULL;
471 }
472
473 int ReductionNode::opcode(int opc, BasicType bt) {
474 int vopc = opc;
475 switch (opc) {
476 case Op_AddI:
477 assert(bt == T_INT, "must be");
478 vopc = Op_AddReductionVI;
479 break;
480 case Op_AddL:
481 assert(bt == T_LONG, "must be");
482 vopc = Op_AddReductionVL;
483 break;
484 case Op_AddF:
485 assert(bt == T_FLOAT, "must be");
486 vopc = Op_AddReductionVF;
487 break;
488 case Op_AddD:
489 assert(bt == T_DOUBLE, "must be");
490 vopc = Op_AddReductionVD;
491 break;
492 case Op_MulI:
493 assert(bt == T_INT, "must be");
494 vopc = Op_MulReductionVI;
495 break;
496 case Op_MulL:
497 assert(bt == T_LONG, "must be");
498 vopc = Op_MulReductionVL;
499 break;
500 case Op_MulF:
501 assert(bt == T_FLOAT, "must be");
502 vopc = Op_MulReductionVF;
503 break;
504 case Op_MulD:
505 assert(bt == T_DOUBLE, "must be");
506 vopc = Op_MulReductionVD;
507 break;
508 // TODO: add MulL for targets that support it
509 default:
510 break;
511 }
512 return vopc;
513 }
514
515 // Return the appropriate reduction node.
516 ReductionNode* ReductionNode::make(int opc, Node *ctrl, Node* n1, Node* n2, BasicType bt) {
517
518 int vopc = opcode(opc, bt);
519
520 // This method should not be called for unimplemented vectors.
521 guarantee(vopc != opc, err_msg_res("Vector for '%s' is not implemented", NodeClassNames[opc]));
522
523 switch (vopc) {
524 case Op_AddReductionVI: return new AddReductionVINode(ctrl, n1, n2);
525 case Op_AddReductionVL: return new AddReductionVLNode(ctrl, n1, n2);
526 case Op_AddReductionVF: return new AddReductionVFNode(ctrl, n1, n2);
527 case Op_AddReductionVD: return new AddReductionVDNode(ctrl, n1, n2);
528 case Op_MulReductionVI: return new MulReductionVINode(ctrl, n1, n2);
529 case Op_MulReductionVL: return new MulReductionVLNode(ctrl, n1, n2);
530 case Op_MulReductionVF: return new MulReductionVFNode(ctrl, n1, n2);
531 case Op_MulReductionVD: return new MulReductionVDNode(ctrl, n1, n2);
532 }
533 fatal(err_msg_res("Missed vector creation for '%s'", NodeClassNames[vopc]));
534 return NULL;
535 }
536
537 bool ReductionNode::implemented(int opc, uint vlen, BasicType bt) {
538 if (is_java_primitive(bt) &&
539 (vlen > 1) && is_power_of_2(vlen) &&
540 Matcher::vector_size_supported(bt, vlen)) {
541 int vopc = ReductionNode::opcode(opc, bt);
542 return vopc != opc && Matcher::match_rule_supported(vopc);
543 }
544 return false;
545 }
546