1 /*
2 * Copyright (c) 2020, 2025, 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
25 #include "ci/ciSymbols.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "opto/library_call.hpp"
28 #include "opto/runtime.hpp"
29 #include "opto/vectornode.hpp"
30 #include "prims/vectorSupport.hpp"
31 #include "runtime/stubRoutines.hpp"
32
33 #ifdef ASSERT
34 static bool is_vector(ciKlass* klass) {
35 return klass->is_subclass_of(ciEnv::current()->vector_VectorPayload_klass());
36 }
37
38 static bool check_vbox(const TypeInstPtr* vbox_type) {
39 assert(vbox_type->klass_is_exact(), "");
40
41 ciInstanceKlass* ik = vbox_type->instance_klass();
42 assert(is_vector(ik), "not a vector");
43
44 ciField* fd1 = ik->get_field_by_name(ciSymbols::ETYPE_name(), ciSymbols::class_signature(), /* is_static */ true);
45 assert(fd1 != nullptr, "element type info is missing");
46
47 ciConstant val1 = fd1->constant_value();
48 BasicType elem_bt = val1.as_object()->as_instance()->java_mirror_type()->basic_type();
49 assert(is_java_primitive(elem_bt), "element type info is missing");
50
51 ciField* fd2 = ik->get_field_by_name(ciSymbols::VLENGTH_name(), ciSymbols::int_signature(), /* is_static */ true);
52 assert(fd2 != nullptr, "vector length info is missing");
53
54 ciConstant val2 = fd2->constant_value();
55 assert(val2.as_int() > 0, "vector length info is missing");
56
57 return true;
58 }
59 #endif
60
61 #define log_if_needed(...) \
62 if (C->print_intrinsics()) { \
63 tty->print_cr(__VA_ARGS__); \
64 }
65
66 #ifndef PRODUCT
67 #define non_product_log_if_needed(...) log_if_needed(__VA_ARGS__)
68 #else
69 #define non_product_log_if_needed(...)
70 #endif
71
72 static bool is_vector_mask(ciKlass* klass) {
73 return klass->is_subclass_of(ciEnv::current()->vector_VectorMask_klass());
74 }
75
76 bool LibraryCallKit::arch_supports_vector_rotate(int opc, int num_elem, BasicType elem_bt,
77 VectorMaskUseType mask_use_type, bool has_scalar_args) {
78 bool is_supported = true;
79
80 // has_scalar_args flag is true only for non-constant scalar shift count,
81 // since in this case shift needs to be broadcasted.
82 if (!Matcher::match_rule_supported_vector(opc, num_elem, elem_bt) ||
83 (has_scalar_args && !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed))) {
84 is_supported = false;
85 }
86
87 if (is_supported) {
88 // Check if mask unboxing is supported, this is a two step process which first loads the contents
89 // of boolean array into vector followed by either lane expansion to match the lane size of masked
90 // vector operation or populate the predicate register.
91 if ((mask_use_type & VecMaskUseLoad) != 0) {
92 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, elem_bt) ||
93 !Matcher::match_rule_supported_vector(Op_LoadVector, num_elem, T_BOOLEAN)) {
94 non_product_log_if_needed(" ** Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
95 NodeClassNames[Op_VectorLoadMask], type2name(elem_bt), num_elem);
96 return false;
97 }
98 }
99
100 if ((mask_use_type & VecMaskUsePred) != 0) {
101 if (!Matcher::has_predicated_vectors() ||
102 !Matcher::match_rule_supported_vector_masked(opc, num_elem, elem_bt)) {
103 non_product_log_if_needed("Rejected vector mask predicate using (%s,%s,%d) because architecture does not support it",
104 NodeClassNames[opc], type2name(elem_bt), num_elem);
105 return false;
106 }
107 }
108 }
109
110 int lshiftopc, rshiftopc;
111 switch(elem_bt) {
112 case T_BYTE:
113 lshiftopc = Op_LShiftI;
114 rshiftopc = Op_URShiftB;
115 break;
116 case T_SHORT:
117 lshiftopc = Op_LShiftI;
118 rshiftopc = Op_URShiftS;
119 break;
120 case T_INT:
121 lshiftopc = Op_LShiftI;
122 rshiftopc = Op_URShiftI;
123 break;
124 case T_LONG:
125 lshiftopc = Op_LShiftL;
126 rshiftopc = Op_URShiftL;
127 break;
128 default: fatal("Unexpected type: %s", type2name(elem_bt));
129 }
130 int lshiftvopc = VectorNode::opcode(lshiftopc, elem_bt);
131 int rshiftvopc = VectorNode::opcode(rshiftopc, elem_bt);
132 if (!is_supported &&
133 arch_supports_vector(lshiftvopc, num_elem, elem_bt, VecMaskNotUsed, has_scalar_args) &&
134 arch_supports_vector(rshiftvopc, num_elem, elem_bt, VecMaskNotUsed, has_scalar_args) &&
135 arch_supports_vector(Op_OrV, num_elem, elem_bt, VecMaskNotUsed)) {
136 is_supported = true;
137 }
138 return is_supported;
139 }
140
141 Node* GraphKit::box_vector(Node* vector, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem, bool deoptimize_on_exception) {
142 assert(EnableVectorSupport, "");
143
144 PreserveReexecuteState preexecs(this);
145 jvms()->set_should_reexecute(true);
146
147 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
148 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
149 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true, deoptimize_on_exception);
150 set_i_o(gvn().transform( new ProjNode(alloc, TypeFunc::I_O) ));
151 set_all_memory(gvn().transform( new ProjNode(alloc, TypeFunc::Memory) ));
152 Node* ret = gvn().transform(new ProjNode(alloc, TypeFunc::Parms));
153
154 assert(check_vbox(vbox_type), "");
155 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_type->instance_klass()));
156 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vt);
157 return gvn().transform(vbox);
158 }
159
160 Node* GraphKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) {
161 assert(EnableVectorSupport, "");
162 const TypeInstPtr* vbox_type_v = gvn().type(v)->isa_instptr();
163 if (vbox_type_v == nullptr || vbox_type->instance_klass() != vbox_type_v->instance_klass()) {
164 return nullptr; // arguments don't agree on vector shapes
165 }
166 if (vbox_type_v->maybe_null()) {
167 return nullptr; // no nulls are allowed
168 }
169 assert(check_vbox(vbox_type), "");
170 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_type->instance_klass()));
171 Node* unbox = gvn().transform(new VectorUnboxNode(C, vt, v, merged_memory()));
172 return unbox;
173 }
174
175 Node* GraphKit::vector_shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
176 assert(bt == T_INT || bt == T_LONG || bt == T_SHORT || bt == T_BYTE, "byte, short, long and int are supported");
177 juint mask = (type2aelembytes(bt) * BitsPerByte - 1);
178 Node* nmask = gvn().transform(ConNode::make(TypeInt::make(mask)));
179 Node* mcnt = gvn().transform(new AndINode(cnt, nmask));
180 return gvn().transform(VectorNode::shift_count(shift_op, mcnt, num_elem, bt));
181 }
182
183 bool LibraryCallKit::arch_supports_vector(int sopc, int num_elem, BasicType type, VectorMaskUseType mask_use_type, bool has_scalar_args) {
184 // Check that the operation is valid.
185 if (sopc <= 0) {
186 non_product_log_if_needed(" ** Rejected intrinsification because no valid vector op could be extracted");
187 return false;
188 }
189
190 if (VectorNode::is_vector_rotate(sopc)) {
191 if(!arch_supports_vector_rotate(sopc, num_elem, type, mask_use_type, has_scalar_args)) {
192 non_product_log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support variable vector shifts",
193 NodeClassNames[sopc], type2name(type), num_elem);
194 return false;
195 }
196 } else if (VectorNode::is_vector_integral_negate(sopc)) {
197 if (!VectorNode::is_vector_integral_negate_supported(sopc, num_elem, type, false)) {
198 non_product_log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support integral vector negate",
199 NodeClassNames[sopc], type2name(type), num_elem);
200 return false;
201 }
202 } else {
203 // Check that architecture supports this op-size-type combination.
204 if (!Matcher::match_rule_supported_vector(sopc, num_elem, type)) {
205 non_product_log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support it",
206 NodeClassNames[sopc], type2name(type), num_elem);
207 return false;
208 } else {
209 assert(Matcher::match_rule_supported(sopc), "must be supported");
210 }
211 }
212
213 if (num_elem == 1) {
214 if (mask_use_type != VecMaskNotUsed) {
215 non_product_log_if_needed(" ** Rejected vector mask op (%s,%s,%d) because architecture does not support it",
216 NodeClassNames[sopc], type2name(type), num_elem);
217 return false;
218 }
219
220 if (sopc != 0) {
221 if (sopc != Op_LoadVector && sopc != Op_StoreVector) {
222 non_product_log_if_needed(" ** Not a svml call or load/store vector op (%s,%s,%d)",
223 NodeClassNames[sopc], type2name(type), num_elem);
224 return false;
225 }
226 }
227 }
228
229 if (!has_scalar_args && VectorNode::is_vector_shift(sopc) &&
230 Matcher::supports_vector_variable_shifts() == false) {
231 log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support variable vector shifts",
232 NodeClassNames[sopc], type2name(type), num_elem);
233 return false;
234 }
235
236 // Check if mask unboxing is supported, this is a two step process which first loads the contents
237 // of boolean array into vector followed by either lane expansion to match the lane size of masked
238 // vector operation or populate the predicate register.
239 if ((mask_use_type & VecMaskUseLoad) != 0) {
240 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type) ||
241 !Matcher::match_rule_supported_vector(Op_LoadVector, num_elem, T_BOOLEAN)) {
242 non_product_log_if_needed(" ** Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
243 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
244 return false;
245 }
246 }
247
248 // Check if mask boxing is supported, this is a two step process which first stores the contents
249 // of mask vector / predicate register into a boolean vector followed by vector store operation to
250 // transfer the contents to underlined storage of mask boxes which is a boolean array.
251 if ((mask_use_type & VecMaskUseStore) != 0) {
252 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type) ||
253 !Matcher::match_rule_supported_vector(Op_StoreVector, num_elem, T_BOOLEAN)) {
254 non_product_log_if_needed("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
255 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
256 return false;
257 }
258 }
259
260 if ((mask_use_type & VecMaskUsePred) != 0) {
261 bool is_supported = false;
262 if (Matcher::has_predicated_vectors()) {
263 if (VectorNode::is_vector_integral_negate(sopc)) {
264 is_supported = VectorNode::is_vector_integral_negate_supported(sopc, num_elem, type, true);
265 } else {
266 is_supported = Matcher::match_rule_supported_vector_masked(sopc, num_elem, type);
267 }
268 }
269 is_supported |= Matcher::supports_vector_predicate_op_emulation(sopc, num_elem, type);
270
271 if (!is_supported) {
272 non_product_log_if_needed("Rejected vector mask predicate using (%s,%s,%d) because architecture does not support it",
273 NodeClassNames[sopc], type2name(type), num_elem);
274 return false;
275 }
276 }
277
278 return true;
279 }
280
281 static bool is_klass_initialized(const TypeInstPtr* vec_klass) {
282 if (vec_klass->const_oop() == nullptr) {
283 return false; // uninitialized or some kind of unsafe access
284 }
285 assert(vec_klass->const_oop()->as_instance()->java_lang_Class_klass() != nullptr, "klass instance expected");
286 ciInstanceKlass* klass = vec_klass->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
287 return klass->is_initialized();
288 }
289
290 // public static
291 // <V extends Vector<E>,
292 // M extends VectorMask<E>,
293 // E>
294 // V unaryOp(int oprId, Class<? extends V> vmClass, Class<? extends M> maskClass, Class<E> elementType,
295 // int length, V v, M m,
296 // UnaryOperation<V, M> defaultImpl)
297 //
298 // public static
299 // <V,
300 // M extends VectorMask<E>,
301 // E>
302 // V binaryOp(int oprId, Class<? extends V> vmClass, Class<? extends M> maskClass, Class<E> elementType,
303 // int length, V v1, V v2, M m,
304 // BinaryOperation<V, M> defaultImpl)
305 //
306 // public static
307 // <V extends Vector<E>,
308 // M extends VectorMask<E>,
309 // E>
310 // V ternaryOp(int oprId, Class<? extends V> vmClass, Class<? extends M> maskClass, Class<E> elementType,
311 // int length, V v1, V v2, V v3, M m,
312 // TernaryOperation<V, M> defaultImpl)
313 //
314 bool LibraryCallKit::inline_vector_nary_operation(int n) {
315 const TypeInt* opr = gvn().type(argument(0))->isa_int();
316 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
317 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
318 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
319 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
320
321 if (opr == nullptr || !opr->is_con() ||
322 vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
323 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
324 vlen == nullptr || !vlen->is_con()) {
325 log_if_needed(" ** missing constant: opr=%s vclass=%s etype=%s vlen=%s",
326 NodeClassNames[argument(0)->Opcode()],
327 NodeClassNames[argument(1)->Opcode()],
328 NodeClassNames[argument(3)->Opcode()],
329 NodeClassNames[argument(4)->Opcode()]);
330 return false; // not enough info for intrinsification
331 }
332
333 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
334 if (!elem_type->is_primitive_type()) {
335 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
336 return false; // should be primitive type
337 }
338 if (!is_klass_initialized(vector_klass)) {
339 log_if_needed(" ** klass argument not initialized");
340 return false;
341 }
342
343 // "argument(n + 5)" should be the mask object. We assume it is "null" when no mask
344 // is used to control this operation.
345 const Type* vmask_type = gvn().type(argument(n + 5));
346 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
347 if (is_masked_op) {
348 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
349 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(2)->Opcode()]);
350 return false; // not enough info for intrinsification
351 }
352
353 if (!is_klass_initialized(mask_klass)) {
354 log_if_needed(" ** mask klass argument not initialized");
355 return false;
356 }
357
358 if (vmask_type->maybe_null()) {
359 log_if_needed(" ** null mask values are not allowed for masked op");
360 return false;
361 }
362 }
363
364 BasicType elem_bt = elem_type->basic_type();
365 bool has_scalar_op = VectorSupport::has_scalar_op(opr->get_con());
366 bool is_unsigned = VectorSupport::is_unsigned_op(opr->get_con());
367
368 int num_elem = vlen->get_con();
369 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
370 int sopc = has_scalar_op ? VectorNode::opcode(opc, elem_bt) : opc;
371 if (sopc == 0 || num_elem == 1) {
372 log_if_needed(" ** operation not supported: arity=%d opc=%s[%d] vlen=%d etype=%s",
373 n, NodeClassNames[opc], opc, num_elem, type2name(elem_bt));
374 return false; // operation not supported
375 }
376 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
377 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
378
379 if (is_vector_mask(vbox_klass)) {
380 assert(!is_masked_op, "mask operations do not need mask to control");
381 }
382
383 // When using mask, mask use type needs to be VecMaskUseLoad.
384 VectorMaskUseType mask_use_type = is_vector_mask(vbox_klass) ? VecMaskUseAll
385 : is_masked_op ? VecMaskUseLoad : VecMaskNotUsed;
386 if (!arch_supports_vector(sopc, num_elem, elem_bt, mask_use_type)) {
387 log_if_needed(" ** not supported: arity=%d opc=%d vlen=%d etype=%s ismask=%d is_masked_op=%d",
388 n, sopc, num_elem, type2name(elem_bt),
389 is_vector_mask(vbox_klass) ? 1 : 0, is_masked_op ? 1 : 0);
390 return false; // not supported
391 }
392
393 // Return true if current platform has implemented the masked operation with predicate feature.
394 bool use_predicate = is_masked_op && arch_supports_vector(sopc, num_elem, elem_bt, VecMaskUsePred);
395 if (is_masked_op && !use_predicate && !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
396 log_if_needed(" ** not supported: arity=%d opc=%d vlen=%d etype=%s ismask=0 is_masked_op=1",
397 n, sopc, num_elem, type2name(elem_bt));
398 return false;
399 }
400
401 Node* opd1 = nullptr; Node* opd2 = nullptr; Node* opd3 = nullptr;
402 switch (n) {
403 case 3: {
404 opd3 = unbox_vector(argument(7), vbox_type, elem_bt, num_elem);
405 if (opd3 == nullptr) {
406 log_if_needed(" ** unbox failed v3=%s",
407 NodeClassNames[argument(7)->Opcode()]);
408 return false;
409 }
410 // fall-through
411 }
412 case 2: {
413 opd2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
414 if (opd2 == nullptr) {
415 log_if_needed(" ** unbox failed v2=%s",
416 NodeClassNames[argument(6)->Opcode()]);
417 return false;
418 }
419 // fall-through
420 }
421 case 1: {
422 opd1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
423 if (opd1 == nullptr) {
424 log_if_needed(" ** unbox failed v1=%s",
425 NodeClassNames[argument(5)->Opcode()]);
426 return false;
427 }
428 break;
429 }
430 default: fatal("unsupported arity: %d", n);
431 }
432
433 Node* mask = nullptr;
434 if (is_masked_op) {
435 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
436 assert(is_vector_mask(mbox_klass), "argument(2) should be a mask class");
437 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
438 mask = unbox_vector(argument(n + 5), mbox_type, elem_bt, num_elem);
439 if (mask == nullptr) {
440 log_if_needed(" ** unbox failed mask=%s",
441 NodeClassNames[argument(n + 5)->Opcode()]);
442 return false;
443 }
444 }
445
446 Node* operation = nullptr;
447 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_klass));
448 switch (n) {
449 case 1:
450 case 2: {
451 operation = VectorNode::make(sopc, opd1, opd2, vt, is_vector_mask(vbox_klass), VectorNode::is_shift_opcode(opc), is_unsigned);
452 break;
453 }
454 case 3: {
455 operation = VectorNode::make(sopc, opd1, opd2, opd3, vt);
456 break;
457 }
458 default: fatal("unsupported arity: %d", n);
459 }
460
461 if (is_masked_op && mask != nullptr) {
462 if (use_predicate) {
463 operation->add_req(mask);
464 operation->add_flag(Node::Flag_is_predicated_vector);
465 } else {
466 operation->add_flag(Node::Flag_is_predicated_using_blend);
467 operation = gvn().transform(operation);
468 operation = new VectorBlendNode(opd1, operation, mask);
469 }
470 }
471 operation = gvn().transform(operation);
472
473 // Wrap it up in VectorBox to keep object type information.
474 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
475 set_result(vbox);
476 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
477 return true;
478 }
479
480 // public static
481 // <V extends Vector<E>, E>
482 // V libraryUnaryOp(long address, Class<? extends V> vClass, Class<E> elementType, int length, String debugName,
483 // V v,
484 // UnaryOperation<V, ?> defaultImpl)
485 //
486 // public static
487 // <V extends VectorPayload, E>
488 // V libraryBinaryOp(long address, Class<? extends V> vClass, Class<E> elementType, int length, String debugName,
489 // V v1, V v2,
490 // BinaryOperation<V, ?> defaultImpl)
491 bool LibraryCallKit::inline_vector_call(int arity) {
492 assert(Matcher::supports_vector_calling_convention(), "required");
493
494 const TypeLong* entry = gvn().type(argument(0))->isa_long();
495 const TypeInstPtr* vector_klass = gvn().type(argument(2))->isa_instptr();
496 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
497 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
498 const TypeInstPtr* debug_name_oop = gvn().type(argument(5))->isa_instptr();
499
500 if (entry == nullptr || !entry->is_con() ||
501 vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
502 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
503 vlen == nullptr || !vlen->is_con() ||
504 debug_name_oop == nullptr || debug_name_oop->const_oop() == nullptr) {
505 log_if_needed(" ** missing constant: opr=%s vclass=%s etype=%s vlen=%s debug_name=%s",
506 NodeClassNames[argument(0)->Opcode()],
507 NodeClassNames[argument(2)->Opcode()],
508 NodeClassNames[argument(3)->Opcode()],
509 NodeClassNames[argument(4)->Opcode()],
510 NodeClassNames[argument(5)->Opcode()]);
511 return false; // not enough info for intrinsification
512 }
513
514 if (entry->get_con() == 0) {
515 log_if_needed(" ** missing entry point");
516 return false;
517 }
518
519 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
520 if (!elem_type->is_primitive_type()) {
521 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
522 return false; // should be primitive type
523 }
524 if (!is_klass_initialized(vector_klass)) {
525 log_if_needed(" ** klass argument not initialized");
526 return false;
527 }
528
529 BasicType elem_bt = elem_type->basic_type();
530 int num_elem = vlen->get_con();
531 if (!Matcher::vector_size_supported(elem_bt, num_elem)) {
532 log_if_needed(" ** vector size (vlen=%d, etype=%s) is not supported",
533 num_elem, type2name(elem_bt));
534 return false;
535 }
536
537 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
538 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
539
540 Node* opd1 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
541 if (opd1 == nullptr) {
542 log_if_needed(" ** unbox failed v1=%s", NodeClassNames[argument(6)->Opcode()]);
543 return false;
544 }
545
546 Node* opd2 = nullptr;
547 if (arity > 1) {
548 opd2 = unbox_vector(argument(7), vbox_type, elem_bt, num_elem);
549 if (opd2 == nullptr) {
550 log_if_needed(" ** unbox failed v2=%s", NodeClassNames[argument(7)->Opcode()]);
551 return false;
552 }
553 }
554 assert(arity == 1 || arity == 2, "arity %d not supported", arity);
555 const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
556 const TypeFunc* call_type = OptoRuntime::Math_Vector_Vector_Type(arity, vt, vt);
557 address entry_addr = (address)entry->get_con();
558
559 const char* debug_name = "<unknown>";
560 if (!debug_name_oop->const_oop()->is_null_object()) {
561 size_t buflen = 100;
562 char* buf = NEW_ARENA_ARRAY(C->comp_arena(), char, buflen);
563 debug_name = debug_name_oop->const_oop()->as_instance()->java_lang_String_str(buf, buflen);
564 }
565 Node* vcall = make_runtime_call(RC_VECTOR,
566 call_type,
567 entry_addr,
568 debug_name,
569 TypePtr::BOTTOM,
570 opd1,
571 opd2);
572
573 vcall = gvn().transform(new ProjNode(gvn().transform(vcall), TypeFunc::Parms));
574
575 // Wrap it up in VectorBox to keep object type information.
576 Node* vbox = box_vector(vcall, vbox_type, elem_bt, num_elem);
577 set_result(vbox);
578 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
579 return true;
580 }
581
582 // <E, M>
583 // long maskReductionCoerced(int oper, Class<? extends M> maskClass, Class<?> elemClass,
584 // int length, M m, VectorMaskOp<M> defaultImpl)
585 bool LibraryCallKit::inline_vector_mask_operation() {
586 const TypeInt* oper = gvn().type(argument(0))->isa_int();
587 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
588 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
589 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
590 Node* mask = argument(4);
591
592 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr ||
593 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
594 vlen == nullptr || !vlen->is_con() ||
595 oper == nullptr || !oper->is_con() ||
596 mask->is_top()) {
597 return false; // dead code
598 }
599
600 if (!is_klass_initialized(mask_klass)) {
601 log_if_needed(" ** klass argument not initialized");
602 return false;
603 }
604
605 int num_elem = vlen->get_con();
606 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
607 BasicType elem_bt = elem_type->basic_type();
608
609 int mopc = VectorSupport::vop2ideal(oper->get_con(), elem_bt);
610 if (!arch_supports_vector(mopc, num_elem, elem_bt, VecMaskUseLoad)) {
611 log_if_needed(" ** not supported: arity=1 op=cast#%d/3 vlen2=%d etype2=%s",
612 mopc, num_elem, type2name(elem_bt));
613 return false; // not supported
614 }
615
616 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
617 const TypeInstPtr* mask_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
618 Node* mask_vec = unbox_vector(mask, mask_box_type, elem_bt, num_elem);
619 if (mask_vec == nullptr) {
620 log_if_needed(" ** unbox failed mask=%s",
621 NodeClassNames[argument(4)->Opcode()]);
622 return false;
623 }
624
625 if (mask_vec->bottom_type()->isa_vectmask() == nullptr) {
626 mask_vec = gvn().transform(VectorStoreMaskNode::make(gvn(), mask_vec, elem_bt, num_elem));
627 }
628 const Type* maskoper_ty = mopc == Op_VectorMaskToLong ? (const Type*)TypeLong::LONG : (const Type*)TypeInt::INT;
629 Node* maskoper = gvn().transform(VectorMaskOpNode::make(mask_vec, maskoper_ty, mopc));
630 if (mopc != Op_VectorMaskToLong) {
631 maskoper = ConvI2L(maskoper);
632 }
633 set_result(maskoper);
634
635 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
636 return true;
637 }
638
639 // public static
640 // <M,
641 // S extends VectorSpecies<E>,
642 // E>
643 // M fromBitsCoerced(Class<? extends M> vmClass, Class<E> elementType, int length,
644 // long bits, int mode, S s,
645 // BroadcastOperation<M, E, S> defaultImpl)
646 bool LibraryCallKit::inline_vector_frombits_coerced() {
647 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
648 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
649 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
650 const TypeLong* bits_type = gvn().type(argument(3))->isa_long();
651 // Mode argument determines the mode of operation it can take following values:-
652 // MODE_BROADCAST for vector Vector.broadcast and VectorMask.maskAll operations.
653 // MODE_BITS_COERCED_LONG_TO_MASK for VectorMask.fromLong operation.
654 const TypeInt* mode = gvn().type(argument(5))->isa_int();
655
656 if (vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
657 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
658 vlen == nullptr || !vlen->is_con() ||
659 bits_type == nullptr ||
660 mode == nullptr || !mode->is_con()) {
661 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s bitwise=%s",
662 NodeClassNames[argument(0)->Opcode()],
663 NodeClassNames[argument(1)->Opcode()],
664 NodeClassNames[argument(2)->Opcode()],
665 NodeClassNames[argument(5)->Opcode()]);
666 return false; // not enough info for intrinsification
667 }
668
669 if (!is_klass_initialized(vector_klass)) {
670 log_if_needed(" ** klass argument not initialized");
671 return false;
672 }
673 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
674 if (!elem_type->is_primitive_type()) {
675 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
676 return false; // should be primitive type
677 }
678 BasicType elem_bt = elem_type->basic_type();
679 int num_elem = vlen->get_con();
680 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
681 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
682
683 bool is_mask = is_vector_mask(vbox_klass);
684 int bcast_mode = mode->get_con();
685 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_mask ? VecMaskUseAll : VecMaskNotUsed);
686 int opc = bcast_mode == VectorSupport::MODE_BITS_COERCED_LONG_TO_MASK ? Op_VectorLongToMask : Op_Replicate;
687
688 if (!arch_supports_vector(opc, num_elem, elem_bt, checkFlags, true /*has_scalar_args*/)) {
689 // If the input long sets or unsets all lanes and Replicate is supported,
690 // generate a MaskAll or Replicate instead.
691
692 // The "maskAll" API uses the corresponding integer types for floating-point data.
693 BasicType maskall_bt = elem_bt == T_DOUBLE ? T_LONG : (elem_bt == T_FLOAT ? T_INT: elem_bt);
694 if (!(opc == Op_VectorLongToMask &&
695 VectorNode::is_maskall_type(bits_type, num_elem) &&
696 arch_supports_vector(Op_Replicate, num_elem, maskall_bt, checkFlags, true /*has_scalar_args*/))) {
697 log_if_needed(" ** not supported: arity=0 op=broadcast vlen=%d etype=%s ismask=%d bcast_mode=%d",
698 num_elem, type2name(elem_bt),
699 is_mask ? 1 : 0,
700 bcast_mode);
701 return false; // not supported
702 }
703 }
704
705 Node* broadcast = nullptr;
706 Node* bits = argument(3);
707 Node* elem = bits;
708
709 if (opc == Op_VectorLongToMask) {
710 const TypeVect* vt = TypeVect::makemask(elem_bt, num_elem);
711 if (vt->isa_vectmask()) {
712 broadcast = gvn().transform(new VectorLongToMaskNode(elem, vt));
713 } else {
714 const TypeVect* mvt = TypeVect::make(T_BOOLEAN, num_elem);
715 broadcast = gvn().transform(new VectorLongToMaskNode(elem, mvt));
716 broadcast = gvn().transform(new VectorLoadMaskNode(broadcast, vt));
717 }
718 } else {
719 switch (elem_bt) {
720 case T_BOOLEAN: // fall-through
721 case T_BYTE: // fall-through
722 case T_SHORT: // fall-through
723 case T_CHAR: // fall-through
724 case T_INT: {
725 elem = gvn().transform(new ConvL2INode(bits));
726 break;
727 }
728 case T_DOUBLE: {
729 elem = gvn().transform(new MoveL2DNode(bits));
730 break;
731 }
732 case T_FLOAT: {
733 bits = gvn().transform(new ConvL2INode(bits));
734 elem = gvn().transform(new MoveI2FNode(bits));
735 break;
736 }
737 case T_LONG: {
738 // no conversion needed
739 break;
740 }
741 default: fatal("%s", type2name(elem_bt));
742 }
743 broadcast = VectorNode::scalar2vector(elem, num_elem, elem_bt, is_mask);
744 broadcast = gvn().transform(broadcast);
745 }
746
747 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
748 set_result(box);
749 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
750 return true;
751 }
752
753 static bool elem_consistent_with_arr(BasicType elem_bt, const TypeAryPtr* arr_type, bool mismatched_ms) {
754 assert(arr_type != nullptr, "unexpected");
755 BasicType arr_elem_bt = arr_type->elem()->array_element_basic_type();
756 if (elem_bt == arr_elem_bt) {
757 return true;
758 } else if (elem_bt == T_SHORT && arr_elem_bt == T_CHAR) {
759 // Load/store of short vector from/to char[] is supported
760 return true;
761 } else if (elem_bt == T_BYTE && arr_elem_bt == T_BOOLEAN) {
762 // Load/store of byte vector from/to boolean[] is supported
763 return true;
764 } else {
765 return mismatched_ms;
766 }
767 }
768
769 // public static
770 // <C,
771 // VM extends VectorPayload,
772 // E,
773 // S extends VectorSpecies<E>>
774 // VM load(Class<? extends VM> vmClass, Class<E> eClass,
775 // int length,
776 // Object base, long offset, // Unsafe addressing
777 // boolean fromSegment,
778 // C container, long index, S s, // Arguments for default implementation
779 // LoadOperation<C, VM, S> defaultImpl) {
780 // public static
781 // <C,
782 // V extends VectorPayload>
783 // void store(Class<?> vClass, Class<?> eClass,
784 // int length,
785 // Object base, long offset, // Unsafe addressing
786 // boolean fromSegment,
787 // V v, C container, long index, // Arguments for default implementation
788 // StoreVectorOperation<C, V> defaultImpl) {
789 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
790 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
791 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
792 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
793 const TypeInt* from_ms = gvn().type(argument(6))->isa_int();
794
795 if (vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
796 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
797 vlen == nullptr || !vlen->is_con() ||
798 from_ms == nullptr || !from_ms->is_con()) {
799 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s from_ms=%s",
800 NodeClassNames[argument(0)->Opcode()],
801 NodeClassNames[argument(1)->Opcode()],
802 NodeClassNames[argument(2)->Opcode()],
803 NodeClassNames[argument(6)->Opcode()]);
804 return false; // not enough info for intrinsification
805 }
806 if (!is_klass_initialized(vector_klass)) {
807 log_if_needed(" ** klass argument not initialized");
808 return false;
809 }
810
811 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
812 if (!elem_type->is_primitive_type()) {
813 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
814 return false; // should be primitive type
815 }
816 BasicType elem_bt = elem_type->basic_type();
817 int num_elem = vlen->get_con();
818
819 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
820 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
821 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s ismask=no",
822 is_store, is_store ? "store" : "load",
823 num_elem, type2name(elem_bt));
824 return false; // not supported
825 }
826
827 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
828 bool is_mask = is_vector_mask(vbox_klass);
829
830 Node* base = argument(3);
831 Node* offset = ConvL2X(argument(4));
832
833 // Save state and restore on bailout
834 SavedState old_state(this);
835
836 Node* addr = make_unsafe_address(base, offset, (is_mask ? T_BOOLEAN : elem_bt), true);
837
838 // The memory barrier checks are based on ones for unsafe access.
839 // This is not 1-1 implementation.
840 const Type *const base_type = gvn().type(base);
841
842 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
843 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
844
845 const bool in_native = TypePtr::NULL_PTR == base_type; // base always null
846 const bool in_heap = !TypePtr::NULL_PTR->higher_equal(base_type); // base never null
847
848 const bool is_mixed_access = !in_heap && !in_native;
849
850 const bool is_mismatched_access = in_heap && (addr_type->isa_aryptr() == nullptr);
851
852 const bool needs_cpu_membar = is_mixed_access || is_mismatched_access;
853
854 // For non-masked mismatched memory segment vector read/write accesses, intrinsification can continue
855 // with unknown backing storage type and compiler can skip inserting explicit reinterpretation IR after
856 // loading from or before storing to backing storage which is mandatory for semantic correctness of
857 // big-endian memory layout.
858 bool mismatched_ms = LITTLE_ENDIAN_ONLY(false)
859 BIG_ENDIAN_ONLY(from_ms->get_con() && !is_mask && arr_type != nullptr &&
860 arr_type->elem()->array_element_basic_type() != elem_bt);
861 BasicType mem_elem_bt = mismatched_ms ? arr_type->elem()->array_element_basic_type() : elem_bt;
862 if (!is_java_primitive(mem_elem_bt)) {
863 log_if_needed(" ** non-primitive array element type");
864 return false;
865 }
866 int mem_num_elem = mismatched_ms ? (num_elem * type2aelembytes(elem_bt)) / type2aelembytes(mem_elem_bt) : num_elem;
867 if (arr_type != nullptr && !is_mask && !elem_consistent_with_arr(elem_bt, arr_type, mismatched_ms)) {
868 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s atype=%s ismask=no",
869 is_store, is_store ? "store" : "load",
870 num_elem, type2name(elem_bt), type2name(arr_type->elem()->array_element_basic_type()));
871 return false;
872 }
873
874 // In case of mismatched memory segment accesses, we need to double check that the source type memory operations are supported by backend.
875 if (mismatched_ms) {
876 if (is_store) {
877 if (!arch_supports_vector(Op_StoreVector, num_elem, elem_bt, VecMaskNotUsed)
878 || !arch_supports_vector(Op_VectorReinterpret, mem_num_elem, mem_elem_bt, VecMaskNotUsed)) {
879 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d*8 etype=%s/8 ismask=no",
880 is_store, "store",
881 num_elem, type2name(elem_bt));
882 return false; // not supported
883 }
884 } else {
885 if (!arch_supports_vector(Op_LoadVector, mem_num_elem, mem_elem_bt, VecMaskNotUsed)
886 || !arch_supports_vector(Op_VectorReinterpret, num_elem, elem_bt, VecMaskNotUsed)) {
887 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d*8 etype=%s/8 ismask=no",
888 is_store, "load",
889 mem_num_elem, type2name(mem_elem_bt));
890 return false; // not supported
891 }
892 }
893 }
894 if (is_mask) {
895 if (!is_store) {
896 if (!arch_supports_vector(Op_LoadVector, num_elem, elem_bt, VecMaskUseLoad)) {
897 return false; // not supported
898 }
899 } else {
900 if (!arch_supports_vector(Op_StoreVector, num_elem, elem_bt, VecMaskUseStore)) {
901 return false; // not supported
902 }
903 }
904 }
905
906 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
907
908 if (needs_cpu_membar) {
909 insert_mem_bar(Op_MemBarCPUOrder);
910 }
911
912 if (is_store) {
913 Node* val = unbox_vector(argument(7), vbox_type, elem_bt, num_elem);
914 if (val == nullptr) {
915 return false; // operand unboxing failed
916 }
917 set_all_memory(reset_memory());
918
919 // In case the store needs to happen to byte array, reinterpret the incoming vector to byte vector.
920 int store_num_elem = num_elem;
921 if (mismatched_ms) {
922 store_num_elem = mem_num_elem;
923 const TypeVect* to_vect_type = TypeVect::make(mem_elem_bt, store_num_elem);
924 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type));
925 }
926 if (is_mask) {
927 val = gvn().transform(VectorStoreMaskNode::make(gvn(), val, elem_bt, num_elem));
928 }
929 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(addr), addr, addr_type, val, store_num_elem));
930 set_memory(vstore, addr_type);
931 } else {
932 // When using byte array, we need to load as byte then reinterpret the value. Otherwise, do a simple vector load.
933 Node* vload = nullptr;
934 if (mismatched_ms) {
935 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, mem_num_elem, mem_elem_bt));
936 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
937 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type));
938 } else {
939 // Special handle for masks
940 if (is_mask) {
941 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, T_BOOLEAN));
942 vload = gvn().transform(new VectorLoadMaskNode(vload, TypeVect::makemask(elem_bt, num_elem)));
943 } else {
944 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, elem_bt));
945 }
946 }
947 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
948 set_result(box);
949 }
950
951 old_state.discard();
952
953 if (needs_cpu_membar) {
954 insert_mem_bar(Op_MemBarCPUOrder);
955 }
956
957 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
958 return true;
959 }
960
961 // public static
962 // <C,
963 // V extends Vector<?>,
964 // E,
965 // S extends VectorSpecies<E>,
966 // M extends VectorMask<E>>
967 // V loadMasked(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass,
968 // int length, Object base, long offset, // Unsafe addressing
969 // boolean fromSegment,
970 // M m, int offsetInRange,
971 // C container, long index, S s, // Arguments for default implementation
972 // LoadVectorMaskedOperation<C, V, S, M> defaultImpl) {
973 // public static
974 // <C,
975 // V extends Vector<E>,
976 // M extends VectorMask<E>,
977 // E>
978 // void storeMasked(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass,
979 // int length,
980 // Object base, long offset, // Unsafe addressing
981 // boolean fromSegment,
982 // V v, M m, C container, long index, // Arguments for default implementation
983 // StoreVectorMaskedOperation<C, V, M> defaultImpl) {
984
985 bool LibraryCallKit::inline_vector_mem_masked_operation(bool is_store) {
986 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
987 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
988 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
989 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
990 const TypeInt* from_ms = gvn().type(argument(7))->isa_int();
991
992 if (vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
993 mask_klass == nullptr || mask_klass->const_oop() == nullptr ||
994 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
995 vlen == nullptr || !vlen->is_con() ||
996 from_ms == nullptr || !from_ms->is_con()) {
997 log_if_needed(" ** missing constant: vclass=%s mclass=%s etype=%s vlen=%s from_ms=%s",
998 NodeClassNames[argument(0)->Opcode()],
999 NodeClassNames[argument(1)->Opcode()],
1000 NodeClassNames[argument(2)->Opcode()],
1001 NodeClassNames[argument(3)->Opcode()],
1002 NodeClassNames[argument(7)->Opcode()]);
1003 return false; // not enough info for intrinsification
1004 }
1005 if (!is_klass_initialized(vector_klass)) {
1006 log_if_needed(" ** klass argument not initialized");
1007 return false;
1008 }
1009
1010 if (!is_klass_initialized(mask_klass)) {
1011 log_if_needed(" ** mask klass argument not initialized");
1012 return false;
1013 }
1014
1015 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1016 if (!elem_type->is_primitive_type()) {
1017 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1018 return false; // should be primitive type
1019 }
1020
1021 BasicType elem_bt = elem_type->basic_type();
1022 int num_elem = vlen->get_con();
1023
1024 Node* base = argument(4);
1025 Node* offset = ConvL2X(argument(5));
1026
1027 // Save state and restore on bailout
1028 SavedState old_state(this);
1029
1030 Node* addr = make_unsafe_address(base, offset, elem_bt, true);
1031 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
1032 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
1033
1034 bool mismatched_ms = from_ms->get_con() && arr_type != nullptr && arr_type->elem()->array_element_basic_type() != elem_bt;
1035 BIG_ENDIAN_ONLY(if (mismatched_ms) return false;)
1036 // If there is no consistency between array and vector element types, it must be special byte array case
1037 if (arr_type != nullptr && !elem_consistent_with_arr(elem_bt, arr_type, mismatched_ms)) {
1038 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s atype=%s",
1039 is_store, is_store ? "storeMasked" : "loadMasked",
1040 num_elem, type2name(elem_bt), type2name(arr_type->elem()->array_element_basic_type()));
1041 return false;
1042 }
1043
1044 int mem_num_elem = mismatched_ms ? num_elem * type2aelembytes(elem_bt) : num_elem;
1045 BasicType mem_elem_bt = mismatched_ms ? T_BYTE : elem_bt;
1046 bool supports_predicate = arch_supports_vector(is_store ? Op_StoreVectorMasked : Op_LoadVectorMasked,
1047 mem_num_elem, mem_elem_bt, VecMaskUseLoad);
1048
1049 // If current arch does not support the predicated operations, we have to bail
1050 // out when current case uses the predicate feature.
1051 if (!supports_predicate) {
1052 bool needs_predicate = false;
1053 if (is_store) {
1054 // Masked vector store always uses the predicated store.
1055 needs_predicate = true;
1056 } else {
1057 // Masked vector load with IOOBE always uses the predicated load.
1058 const TypeInt* offset_in_range = gvn().type(argument(9))->isa_int();
1059 if (!offset_in_range->is_con()) {
1060 log_if_needed(" ** missing constant: offsetInRange=%s",
1061 NodeClassNames[argument(8)->Opcode()]);
1062 return false;
1063 }
1064 needs_predicate = (offset_in_range->get_con() == 0);
1065 }
1066
1067 if (needs_predicate) {
1068 log_if_needed(" ** not supported: op=%s vlen=%d etype=%s mismatched_ms=%d",
1069 is_store ? "storeMasked" : "loadMasked",
1070 num_elem, type2name(elem_bt), mismatched_ms ? 1 : 0);
1071 return false;
1072 }
1073 }
1074
1075 // This only happens for masked vector load. If predicate is not supported, then check whether
1076 // the normal vector load and blend operations are supported by backend.
1077 if (!supports_predicate && (!arch_supports_vector(Op_LoadVector, mem_num_elem, mem_elem_bt, VecMaskNotUsed) ||
1078 !arch_supports_vector(Op_VectorBlend, mem_num_elem, mem_elem_bt, VecMaskUseLoad))) {
1079 log_if_needed(" ** not supported: op=loadMasked vlen=%d etype=%s mismatched_ms=%d",
1080 num_elem, type2name(elem_bt), mismatched_ms ? 1 : 0);
1081 return false;
1082 }
1083
1084 // Since we are using byte array, we need to double check that the vector reinterpret operation
1085 // with byte type is supported by backend.
1086 if (mismatched_ms) {
1087 if (!arch_supports_vector(Op_VectorReinterpret, mem_num_elem, T_BYTE, VecMaskNotUsed)) {
1088 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s mismatched_ms=1",
1089 is_store, is_store ? "storeMasked" : "loadMasked",
1090 num_elem, type2name(elem_bt));
1091 return false;
1092 }
1093 }
1094
1095 // Since it needs to unbox the mask, we need to double check that the related load operations
1096 // for mask are supported by backend.
1097 if (!arch_supports_vector(Op_LoadVector, num_elem, elem_bt, VecMaskUseLoad)) {
1098 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s",
1099 is_store, is_store ? "storeMasked" : "loadMasked",
1100 num_elem, type2name(elem_bt));
1101 return false;
1102 }
1103
1104 // Can base be null? Otherwise, always on-heap access.
1105 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(gvn().type(base));
1106 if (can_access_non_heap) {
1107 insert_mem_bar(Op_MemBarCPUOrder);
1108 }
1109
1110 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1111 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1112 assert(!is_vector_mask(vbox_klass) && is_vector_mask(mbox_klass), "Invalid class type");
1113 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1114 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1115
1116 Node* mask = unbox_vector(is_store ? argument(9) : argument(8), mbox_type, elem_bt, num_elem);
1117 if (mask == nullptr) {
1118 log_if_needed(" ** unbox failed mask=%s",
1119 is_store ? NodeClassNames[argument(9)->Opcode()]
1120 : NodeClassNames[argument(8)->Opcode()]);
1121 return false;
1122 }
1123
1124 if (is_store) {
1125 Node* val = unbox_vector(argument(8), vbox_type, elem_bt, num_elem);
1126 if (val == nullptr) {
1127 log_if_needed(" ** unbox failed vector=%s",
1128 NodeClassNames[argument(8)->Opcode()]);
1129 return false; // operand unboxing failed
1130 }
1131 set_all_memory(reset_memory());
1132
1133 if (mismatched_ms) {
1134 // Reinterpret the incoming vector to byte vector.
1135 const TypeVect* to_vect_type = TypeVect::make(mem_elem_bt, mem_num_elem);
1136 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type));
1137 // Reinterpret the vector mask to byte type.
1138 const TypeVect* from_mask_type = TypeVect::makemask(elem_bt, num_elem);
1139 const TypeVect* to_mask_type = TypeVect::makemask(mem_elem_bt, mem_num_elem);
1140 mask = gvn().transform(new VectorReinterpretNode(mask, from_mask_type, to_mask_type));
1141 }
1142 Node* vstore = gvn().transform(new StoreVectorMaskedNode(control(), memory(addr), addr, val, addr_type, mask));
1143 set_memory(vstore, addr_type);
1144 } else {
1145 Node* vload = nullptr;
1146
1147 if (mismatched_ms) {
1148 // Reinterpret the vector mask to byte type.
1149 const TypeVect* from_mask_type = TypeVect::makemask(elem_bt, num_elem);
1150 const TypeVect* to_mask_type = TypeVect::makemask(mem_elem_bt, mem_num_elem);
1151 mask = gvn().transform(new VectorReinterpretNode(mask, from_mask_type, to_mask_type));
1152 }
1153
1154 if (supports_predicate) {
1155 // Generate masked load vector node if predicate feature is supported.
1156 const TypeVect* vt = TypeVect::make(mem_elem_bt, mem_num_elem);
1157 vload = gvn().transform(new LoadVectorMaskedNode(control(), memory(addr), addr, addr_type, vt, mask));
1158 } else {
1159 // Use the vector blend to implement the masked load vector. The biased elements are zeros.
1160 Node* zero = gvn().transform(gvn().zerocon(mem_elem_bt));
1161 zero = gvn().transform(VectorNode::scalar2vector(zero, mem_num_elem, mem_elem_bt));
1162 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, mem_num_elem, mem_elem_bt));
1163 vload = gvn().transform(new VectorBlendNode(zero, vload, mask));
1164 }
1165
1166 if (mismatched_ms) {
1167 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
1168 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type));
1169 }
1170
1171 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
1172 set_result(box);
1173 }
1174
1175 old_state.discard();
1176
1177 if (can_access_non_heap) {
1178 insert_mem_bar(Op_MemBarCPUOrder);
1179 }
1180
1181 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1182 return true;
1183 }
1184
1185 //
1186 // <C,
1187 // V extends Vector<?>,
1188 // W extends Vector<Integer>,
1189 // S extends VectorSpecies<E>,
1190 // M extends VectorMask<E>,
1191 // E>
1192 // V loadWithMap(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass, int length,
1193 // Class<? extends Vector<Integer>> vectorIndexClass, int indexLength,
1194 // Object base, long offset,
1195 // W indexVector1, W indexVector2, W indexVector3, W indexVector4,
1196 // M m, C container, int index, int[] indexMap, int indexM, S s,
1197 // LoadVectorOperationWithMap<C, V, S, M> defaultImpl)
1198 //
1199 // <C,
1200 // V extends Vector<E>,
1201 // W extends Vector<Integer>,
1202 // M extends VectorMask<E>,
1203 // E>
1204 // void storeWithMap(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass, int length,
1205 // Class<? extends Vector<Integer>> vectorIndexClass, int indexLength,
1206 // Object base, long offset, // Unsafe addressing
1207 // W indexVector, V v, M m,
1208 // C container, int index, int[] indexMap, int indexM, // Arguments for default implementation
1209 // StoreVectorOperationWithMap<C, V, M> defaultImpl)
1210 //
1211 bool LibraryCallKit::inline_vector_gather_scatter(bool is_scatter) {
1212 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1213 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
1214 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1215 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1216 const TypeInstPtr* vector_idx_klass = gvn().type(argument(4))->isa_instptr();
1217 const TypeInt* idx_vlen = gvn().type(argument(5))->isa_int();
1218
1219 if (vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
1220 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
1221 vlen == nullptr || !vlen->is_con() ||
1222 vector_idx_klass == nullptr || vector_idx_klass->const_oop() == nullptr ||
1223 idx_vlen == nullptr || !idx_vlen->is_con()) {
1224 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s viclass=%s idx_vlen=%s",
1225 NodeClassNames[argument(0)->Opcode()],
1226 NodeClassNames[argument(2)->Opcode()],
1227 NodeClassNames[argument(3)->Opcode()],
1228 NodeClassNames[argument(4)->Opcode()],
1229 NodeClassNames[argument(5)->Opcode()]);
1230 return false; // not enough info for intrinsification
1231 }
1232
1233 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(vector_idx_klass)) {
1234 log_if_needed(" ** klass argument not initialized");
1235 return false;
1236 }
1237
1238 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1239 if (!elem_type->is_primitive_type()) {
1240 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1241 return false; // should be primitive type
1242 }
1243
1244 BasicType elem_bt = elem_type->basic_type();
1245 int num_elem = vlen->get_con();
1246 int idx_num_elem = idx_vlen->get_con();
1247
1248 Node* m = is_scatter ? argument(11) : argument(13);
1249 const Type* vmask_type = gvn().type(m);
1250 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
1251 if (is_masked_op) {
1252 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
1253 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(1)->Opcode()]);
1254 return false; // not enough info for intrinsification
1255 }
1256
1257 if (!is_klass_initialized(mask_klass)) {
1258 log_if_needed(" ** mask klass argument not initialized");
1259 return false;
1260 }
1261
1262 if (vmask_type->maybe_null()) {
1263 log_if_needed(" ** null mask values are not allowed for masked op");
1264 return false;
1265 }
1266
1267 // Check whether the predicated gather/scatter node is supported by architecture.
1268 VectorMaskUseType mask = (VectorMaskUseType) (VecMaskUseLoad | VecMaskUsePred);
1269 if (!arch_supports_vector(is_scatter ? Op_StoreVectorScatterMasked : Op_LoadVectorGatherMasked, num_elem, elem_bt, mask)) {
1270 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s is_masked_op=1",
1271 is_scatter, is_scatter ? "scatterMasked" : "gatherMasked",
1272 num_elem, type2name(elem_bt));
1273 return false; // not supported
1274 }
1275 } else {
1276 // Check whether the normal gather/scatter node is supported for non-masked operation.
1277 if (!arch_supports_vector(is_scatter ? Op_StoreVectorScatter : Op_LoadVectorGather, num_elem, elem_bt, VecMaskNotUsed)) {
1278 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s is_masked_op=0",
1279 is_scatter, is_scatter ? "scatter" : "gather",
1280 num_elem, type2name(elem_bt));
1281 return false; // not supported
1282 }
1283 }
1284
1285 // Check that the vector holding indices is supported by architecture
1286 // For sub-word gathers expander receive index array.
1287 if (!is_subword_type(elem_bt) && !arch_supports_vector(Op_LoadVector, idx_num_elem, T_INT, VecMaskNotUsed)) {
1288 log_if_needed(" ** not supported: arity=%d op=%s/loadindex vlen=%d etype=int is_masked_op=%d",
1289 is_scatter, is_scatter ? "scatter" : "gather",
1290 idx_num_elem, is_masked_op ? 1 : 0);
1291 return false; // not supported
1292 }
1293
1294 Node* base = argument(6);
1295 Node* offset = ConvL2X(argument(7));
1296
1297 // Save state and restore on bailout
1298 SavedState old_state(this);
1299
1300 Node* addr = nullptr;
1301 if (!is_subword_type(elem_bt)) {
1302 addr = make_unsafe_address(base, offset, elem_bt, true);
1303 } else {
1304 assert(!is_scatter, "Only supports gather operation for subword types now");
1305 uint header = arrayOopDesc::base_offset_in_bytes(elem_bt);
1306 assert(offset->is_Con() && offset->bottom_type()->is_long()->get_con() == header,
1307 "offset must be the array base offset");
1308 Node* index = argument(15);
1309 addr = array_element_address(base, index, elem_bt);
1310 }
1311
1312 const TypePtr* addr_type = gvn().type(addr)->isa_ptr();
1313 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
1314
1315 // The array must be consistent with vector type
1316 if (arr_type == nullptr || (arr_type != nullptr && !elem_consistent_with_arr(elem_bt, arr_type, false))) {
1317 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s atype=%s ismask=no",
1318 is_scatter, is_scatter ? "scatter" : "gather",
1319 num_elem, type2name(elem_bt), type2name(arr_type->elem()->array_element_basic_type()));
1320 return false;
1321 }
1322
1323 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1324 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1325 ciKlass* vbox_idx_klass = vector_idx_klass->const_oop()->as_instance()->java_lang_Class_klass();
1326 if (vbox_idx_klass == nullptr) {
1327 return false;
1328 }
1329
1330 // Get the indexes for gather/scatter.
1331 Node* indexes = nullptr;
1332 const TypeInstPtr* vbox_idx_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_idx_klass);
1333 if (is_subword_type(elem_bt)) {
1334 Node* indexMap = argument(16);
1335 Node* indexM = argument(17);
1336 indexes = array_element_address(indexMap, indexM, T_INT);
1337 } else {
1338 // Get the first index vector.
1339 indexes = unbox_vector(argument(9), vbox_idx_type, T_INT, idx_num_elem);
1340 if (indexes == nullptr) {
1341 return false;
1342 }
1343 }
1344
1345 // Get the vector mask value.
1346 Node* mask = nullptr;
1347 if (is_masked_op) {
1348 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1349 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1350 mask = unbox_vector(m, mbox_type, elem_bt, num_elem);
1351 if (mask == nullptr) {
1352 log_if_needed(" ** unbox failed mask=%s", NodeClassNames[m->Opcode()]);
1353 return false;
1354 }
1355 }
1356
1357 const TypeVect* vector_type = TypeVect::make(elem_bt, num_elem);
1358 if (is_scatter) {
1359 Node* val = unbox_vector(argument(10), vbox_type, elem_bt, num_elem);
1360 if (val == nullptr) {
1361 return false; // operand unboxing failed
1362 }
1363 set_all_memory(reset_memory());
1364
1365 Node* vstore = nullptr;
1366 if (mask != nullptr) {
1367 vstore = gvn().transform(new StoreVectorScatterMaskedNode(control(), memory(addr), addr, addr_type, val, indexes, mask));
1368 } else {
1369 vstore = gvn().transform(new StoreVectorScatterNode(control(), memory(addr), addr, addr_type, val, indexes));
1370 }
1371 set_memory(vstore, addr_type);
1372 } else {
1373 Node* vload = nullptr;
1374 if (mask != nullptr) {
1375 vload = gvn().transform(new LoadVectorGatherMaskedNode(control(), memory(addr), addr, addr_type, vector_type, indexes, mask));
1376 } else {
1377 vload = gvn().transform(new LoadVectorGatherNode(control(), memory(addr), addr, addr_type, vector_type, indexes));
1378 }
1379 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
1380 set_result(box);
1381 }
1382
1383 old_state.discard();
1384 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1385 return true;
1386 }
1387
1388 // public static
1389 // <V extends Vector<E>,
1390 // M extends VectorMask<E>,
1391 // E>
1392 // long reductionCoerced(int oprId, Class<? extends V> vectorClass, Class<? extends M> maskClass,
1393 // Class<E> elementType, int length, V v, M m,
1394 // ReductionOperation<V, M> defaultImpl)
1395 bool LibraryCallKit::inline_vector_reduction() {
1396 const TypeInt* opr = gvn().type(argument(0))->isa_int();
1397 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
1398 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
1399 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
1400 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
1401
1402 if (opr == nullptr || !opr->is_con() ||
1403 vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
1404 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
1405 vlen == nullptr || !vlen->is_con()) {
1406 log_if_needed(" ** missing constant: opr=%s vclass=%s etype=%s vlen=%s",
1407 NodeClassNames[argument(0)->Opcode()],
1408 NodeClassNames[argument(1)->Opcode()],
1409 NodeClassNames[argument(3)->Opcode()],
1410 NodeClassNames[argument(4)->Opcode()]);
1411 return false; // not enough info for intrinsification
1412 }
1413 if (!is_klass_initialized(vector_klass)) {
1414 log_if_needed(" ** klass argument not initialized");
1415 return false;
1416 }
1417 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1418 if (!elem_type->is_primitive_type()) {
1419 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1420 return false; // should be primitive type
1421 }
1422
1423 const Type* vmask_type = gvn().type(argument(6));
1424 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
1425 if (is_masked_op) {
1426 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
1427 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(2)->Opcode()]);
1428 return false; // not enough info for intrinsification
1429 }
1430
1431 if (!is_klass_initialized(mask_klass)) {
1432 log_if_needed(" ** mask klass argument not initialized");
1433 return false;
1434 }
1435
1436 if (vmask_type->maybe_null()) {
1437 log_if_needed(" ** null mask values are not allowed for masked op");
1438 return false;
1439 }
1440 }
1441
1442 BasicType elem_bt = elem_type->basic_type();
1443 int num_elem = vlen->get_con();
1444 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
1445 int sopc = ReductionNode::opcode(opc, elem_bt);
1446
1447 // Ensure reduction operation for lanewise operation
1448 // When using mask, mask use type needs to be VecMaskUseLoad.
1449 if (sopc == opc || !arch_supports_vector(sopc, num_elem, elem_bt, is_masked_op ? VecMaskUseLoad : VecMaskNotUsed)) {
1450 log_if_needed(" ** not supported: arity=1 op=%d/reduce vlen=%d etype=%s is_masked_op=%d",
1451 sopc, num_elem, type2name(elem_bt), is_masked_op ? 1 : 0);
1452 return false;
1453 }
1454
1455 // Return true if current platform has implemented the masked operation with predicate feature.
1456 bool use_predicate = is_masked_op && arch_supports_vector(sopc, num_elem, elem_bt, VecMaskUsePred);
1457 if (is_masked_op && !use_predicate && !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
1458 log_if_needed(" ** not supported: arity=1 op=%d/reduce vlen=%d etype=%s is_masked_op=1",
1459 sopc, num_elem, type2name(elem_bt));
1460 return false;
1461 }
1462
1463 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1464 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1465
1466 Node* opd = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1467 if (opd == nullptr) {
1468 return false; // operand unboxing failed
1469 }
1470
1471 Node* mask = nullptr;
1472 if (is_masked_op) {
1473 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1474 assert(is_vector_mask(mbox_klass), "argument(2) should be a mask class");
1475 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1476 mask = unbox_vector(argument(6), mbox_type, elem_bt, num_elem);
1477 if (mask == nullptr) {
1478 log_if_needed(" ** unbox failed mask=%s",
1479 NodeClassNames[argument(6)->Opcode()]);
1480 return false;
1481 }
1482 }
1483
1484 Node* init = ReductionNode::make_identity_con_scalar(gvn(), opc, elem_bt);
1485 Node* value = opd;
1486
1487 assert(mask != nullptr || !is_masked_op, "Masked op needs the mask value never null");
1488 if (mask != nullptr && !use_predicate) {
1489 Node* reduce_identity = gvn().transform(VectorNode::scalar2vector(init, num_elem, elem_bt));
1490 value = gvn().transform(new VectorBlendNode(reduce_identity, value, mask));
1491 }
1492
1493 // Make an unordered Reduction node. This affects only AddReductionVF/VD and MulReductionVF/VD,
1494 // as these operations are allowed to be associative (not requiring strict order) in VectorAPI.
1495 value = ReductionNode::make(opc, nullptr, init, value, elem_bt, /* requires_strict_order */ false);
1496
1497 if (mask != nullptr && use_predicate) {
1498 value->add_req(mask);
1499 value->add_flag(Node::Flag_is_predicated_vector);
1500 }
1501
1502 value = gvn().transform(value);
1503
1504 Node* bits = nullptr;
1505 switch (elem_bt) {
1506 case T_BYTE:
1507 case T_SHORT:
1508 case T_INT: {
1509 bits = gvn().transform(new ConvI2LNode(value));
1510 break;
1511 }
1512 case T_FLOAT: {
1513 value = gvn().transform(new MoveF2INode(value));
1514 bits = gvn().transform(new ConvI2LNode(value));
1515 break;
1516 }
1517 case T_DOUBLE: {
1518 bits = gvn().transform(new MoveD2LNode(value));
1519 break;
1520 }
1521 case T_LONG: {
1522 bits = value; // no conversion needed
1523 break;
1524 }
1525 default: fatal("%s", type2name(elem_bt));
1526 }
1527 set_result(bits);
1528 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1529 return true;
1530 }
1531
1532 // public static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
1533 // V v1, V v2,
1534 // BiFunction<V, V, Boolean> defaultImpl)
1535 //
1536 bool LibraryCallKit::inline_vector_test() {
1537 const TypeInt* cond = gvn().type(argument(0))->isa_int();
1538 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
1539 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1540 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1541
1542 if (cond == nullptr || !cond->is_con() ||
1543 vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
1544 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
1545 vlen == nullptr || !vlen->is_con()) {
1546 log_if_needed(" ** missing constant: cond=%s vclass=%s etype=%s vlen=%s",
1547 NodeClassNames[argument(0)->Opcode()],
1548 NodeClassNames[argument(1)->Opcode()],
1549 NodeClassNames[argument(2)->Opcode()],
1550 NodeClassNames[argument(3)->Opcode()]);
1551 return false; // not enough info for intrinsification
1552 }
1553 if (!is_klass_initialized(vector_klass)) {
1554 log_if_needed(" ** klass argument not initialized");
1555 return false;
1556 }
1557 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1558 if (!elem_type->is_primitive_type()) {
1559 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1560 return false; // should be primitive type
1561 }
1562 BasicType elem_bt = elem_type->basic_type();
1563 int num_elem = vlen->get_con();
1564 BoolTest::mask booltest = (BoolTest::mask)cond->get_con();
1565 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1566 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1567
1568 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, is_vector_mask(vbox_klass) ? VecMaskUseLoad : VecMaskNotUsed)) {
1569 log_if_needed(" ** not supported: arity=2 op=test/%d vlen=%d etype=%s ismask=%d",
1570 cond->get_con(), num_elem, type2name(elem_bt),
1571 is_vector_mask(vbox_klass));
1572 return false;
1573 }
1574
1575 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
1576 Node* opd2;
1577 if (Matcher::vectortest_needs_second_argument(booltest == BoolTest::overflow,
1578 opd1->bottom_type()->isa_vectmask())) {
1579 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1580 } else {
1581 opd2 = opd1;
1582 }
1583 if (opd1 == nullptr || opd2 == nullptr) {
1584 return false; // operand unboxing failed
1585 }
1586
1587 Node* cmp = gvn().transform(new VectorTestNode(opd1, opd2, booltest));
1588 BoolTest::mask test = Matcher::vectortest_mask(booltest == BoolTest::overflow,
1589 opd1->bottom_type()->isa_vectmask(), num_elem);
1590 Node* bol = gvn().transform(new BoolNode(cmp, test));
1591 Node* res = gvn().transform(new CMoveINode(bol, gvn().intcon(0), gvn().intcon(1), TypeInt::BOOL));
1592
1593 set_result(res);
1594 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1595 return true;
1596 }
1597
1598 // public static
1599 // <V extends Vector<E>,
1600 // M extends VectorMask<E>,
1601 // E>
1602 // V blend(Class<? extends V> vectorClass, Class<M> maskClass, Class<E> elementType, int vlen,
1603 // V v1, V v2, M m,
1604 // VectorBlendOp<V, M, E> defaultImpl)
1605 bool LibraryCallKit::inline_vector_blend() {
1606 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1607 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
1608 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1609 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1610
1611 if (mask_klass == nullptr || vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
1612 return false; // dead code
1613 }
1614 if (mask_klass->const_oop() == nullptr || vector_klass->const_oop() == nullptr ||
1615 elem_klass->const_oop() == nullptr || !vlen->is_con()) {
1616 log_if_needed(" ** missing constant: vclass=%s mclass=%s etype=%s vlen=%s",
1617 NodeClassNames[argument(0)->Opcode()],
1618 NodeClassNames[argument(1)->Opcode()],
1619 NodeClassNames[argument(2)->Opcode()],
1620 NodeClassNames[argument(3)->Opcode()]);
1621 return false; // not enough info for intrinsification
1622 }
1623 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(mask_klass)) {
1624 log_if_needed(" ** klass argument not initialized");
1625 return false;
1626 }
1627 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1628 if (!elem_type->is_primitive_type()) {
1629 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1630 return false; // should be primitive type
1631 }
1632 BasicType elem_bt = elem_type->basic_type();
1633 BasicType mask_bt = elem_bt;
1634 int num_elem = vlen->get_con();
1635
1636 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
1637 log_if_needed(" ** not supported: arity=2 op=blend vlen=%d etype=%s ismask=useload",
1638 num_elem, type2name(elem_bt));
1639 return false; // not supported
1640 }
1641 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1642 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1643
1644 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1645 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1646
1647 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
1648 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1649 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
1650
1651 if (v1 == nullptr || v2 == nullptr || mask == nullptr) {
1652 return false; // operand unboxing failed
1653 }
1654
1655 Node* blend = gvn().transform(new VectorBlendNode(v1, v2, mask));
1656
1657 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
1658 set_result(box);
1659 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1660 return true;
1661 }
1662
1663 // public static
1664 // <V extends Vector<E>,
1665 // M extends VectorMask<E>,
1666 // E>
1667 // M compare(int cond, Class<? extends V> vectorClass, Class<M> maskClass, Class<E> elementType, int vlen,
1668 // V v1, V v2, M m,
1669 // VectorCompareOp<V,M> defaultImpl)
1670 bool LibraryCallKit::inline_vector_compare() {
1671 const TypeInt* cond = gvn().type(argument(0))->isa_int();
1672 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
1673 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
1674 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
1675 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
1676
1677 if (cond == nullptr || vector_klass == nullptr || mask_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
1678 return false; // dead code
1679 }
1680 if (!cond->is_con() || vector_klass->const_oop() == nullptr || mask_klass->const_oop() == nullptr ||
1681 elem_klass->const_oop() == nullptr || !vlen->is_con()) {
1682 log_if_needed(" ** missing constant: cond=%s vclass=%s mclass=%s etype=%s vlen=%s",
1683 NodeClassNames[argument(0)->Opcode()],
1684 NodeClassNames[argument(1)->Opcode()],
1685 NodeClassNames[argument(2)->Opcode()],
1686 NodeClassNames[argument(3)->Opcode()],
1687 NodeClassNames[argument(4)->Opcode()]);
1688 return false; // not enough info for intrinsification
1689 }
1690 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(mask_klass)) {
1691 log_if_needed(" ** klass argument not initialized");
1692 return false;
1693 }
1694 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1695 if (!elem_type->is_primitive_type()) {
1696 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1697 return false; // should be primitive type
1698 }
1699
1700 int num_elem = vlen->get_con();
1701 BasicType elem_bt = elem_type->basic_type();
1702 BasicType mask_bt = elem_bt;
1703
1704 if ((cond->get_con() & BoolTest::unsigned_compare) != 0) {
1705 if (!Matcher::supports_vector_comparison_unsigned(num_elem, elem_bt)) {
1706 log_if_needed(" ** not supported: unsigned comparison op=comp/%d vlen=%d etype=%s ismask=usestore",
1707 cond->get_con() & (BoolTest::unsigned_compare - 1), num_elem, type2name(elem_bt));
1708 return false;
1709 }
1710 }
1711
1712 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
1713 log_if_needed(" ** not supported: arity=2 op=comp/%d vlen=%d etype=%s ismask=usestore",
1714 cond->get_con(), num_elem, type2name(elem_bt));
1715 return false;
1716 }
1717
1718 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1719 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1720
1721 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1722 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1723
1724 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1725 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
1726
1727 bool is_masked_op = argument(7)->bottom_type() != TypePtr::NULL_PTR;
1728 Node* mask = is_masked_op ? unbox_vector(argument(7), mbox_type, elem_bt, num_elem) : nullptr;
1729 if (is_masked_op && mask == nullptr) {
1730 log_if_needed(" ** not supported: mask = null arity=2 op=comp/%d vlen=%d etype=%s ismask=usestore is_masked_op=1",
1731 cond->get_con(), num_elem, type2name(elem_bt));
1732 return false;
1733 }
1734
1735 bool use_predicate = is_masked_op && arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUsePred);
1736 if (is_masked_op && !use_predicate && !arch_supports_vector(Op_AndV, num_elem, elem_bt, VecMaskUseLoad)) {
1737 log_if_needed(" ** not supported: arity=2 op=comp/%d vlen=%d etype=%s ismask=usestore is_masked_op=1",
1738 cond->get_con(), num_elem, type2name(elem_bt));
1739 return false;
1740 }
1741
1742 if (v1 == nullptr || v2 == nullptr) {
1743 return false; // operand unboxing failed
1744 }
1745 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
1746 ConINode* pred_node = (ConINode*)gvn().makecon(cond);
1747
1748 const TypeVect* vmask_type = TypeVect::makemask(mask_bt, num_elem);
1749 Node* operation = new VectorMaskCmpNode(pred, v1, v2, pred_node, vmask_type);
1750
1751 if (is_masked_op) {
1752 if (use_predicate) {
1753 operation->add_req(mask);
1754 operation->add_flag(Node::Flag_is_predicated_vector);
1755 } else {
1756 operation = gvn().transform(operation);
1757 operation = VectorNode::make(Op_AndV, operation, mask, vmask_type);
1758 }
1759 }
1760
1761 operation = gvn().transform(operation);
1762
1763 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
1764 set_result(box);
1765 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1766 return true;
1767 }
1768
1769 // public static
1770 // <V extends Vector<E>,
1771 // Sh extends VectorShuffle<E>,
1772 // M extends VectorMask<E>,
1773 // E>
1774 // V rearrangeOp(Class<? extends V> vectorClass, Class<Sh> shuffleClass, Class<M> maskClass, Class<E> elementType, int vlen,
1775 // V v1, Sh sh, M m,
1776 // VectorRearrangeOp<V, Sh, M, E> defaultImpl)
1777 bool LibraryCallKit::inline_vector_rearrange() {
1778 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1779 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->isa_instptr();
1780 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
1781 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
1782 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
1783
1784 if (vector_klass == nullptr || shuffle_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
1785 return false; // dead code
1786 }
1787 if (shuffle_klass->const_oop() == nullptr ||
1788 vector_klass->const_oop() == nullptr ||
1789 elem_klass->const_oop() == nullptr ||
1790 !vlen->is_con()) {
1791 log_if_needed(" ** missing constant: vclass=%s sclass=%s etype=%s vlen=%s",
1792 NodeClassNames[argument(0)->Opcode()],
1793 NodeClassNames[argument(1)->Opcode()],
1794 NodeClassNames[argument(3)->Opcode()],
1795 NodeClassNames[argument(4)->Opcode()]);
1796 return false; // not enough info for intrinsification
1797 }
1798 if (!is_klass_initialized(vector_klass) ||
1799 !is_klass_initialized(shuffle_klass)) {
1800 log_if_needed(" ** klass argument not initialized");
1801 return false;
1802 }
1803 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1804 if (!elem_type->is_primitive_type()) {
1805 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1806 return false; // should be primitive type
1807 }
1808
1809 BasicType elem_bt = elem_type->basic_type();
1810 BasicType shuffle_bt = elem_bt;
1811 if (shuffle_bt == T_FLOAT) {
1812 shuffle_bt = T_INT;
1813 } else if (shuffle_bt == T_DOUBLE) {
1814 shuffle_bt = T_LONG;
1815 }
1816
1817 int num_elem = vlen->get_con();
1818 bool need_load_shuffle = Matcher::vector_rearrange_requires_load_shuffle(shuffle_bt, num_elem);
1819
1820 if (need_load_shuffle && !arch_supports_vector(Op_VectorLoadShuffle, num_elem, shuffle_bt, VecMaskNotUsed)) {
1821 if (C->print_intrinsics()) {
1822 tty->print_cr(" ** not supported: arity=0 op=load/shuffle vlen=%d etype=%s ismask=no",
1823 num_elem, type2name(shuffle_bt));
1824 }
1825 return false; // not supported
1826 }
1827
1828 bool is_masked_op = argument(7)->bottom_type() != TypePtr::NULL_PTR;
1829 bool use_predicate = is_masked_op;
1830 if (is_masked_op &&
1831 (mask_klass == nullptr ||
1832 mask_klass->const_oop() == nullptr ||
1833 !is_klass_initialized(mask_klass))) {
1834 log_if_needed(" ** mask_klass argument not initialized");
1835 }
1836 if (!arch_supports_vector(Op_AndV, num_elem, elem_bt, VecMaskNotUsed)) {
1837 log_if_needed(" ** not supported: arity=2 op=and vlen=%d etype=%s ismask=no",
1838 num_elem, type2name(elem_bt));
1839 return false;
1840 }
1841 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_masked_op ? (VecMaskUseLoad | VecMaskUsePred) : VecMaskNotUsed);
1842 if (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, checkFlags)) {
1843 use_predicate = false;
1844 if(!is_masked_op ||
1845 (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed) ||
1846 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad) ||
1847 !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed))) {
1848 log_if_needed(" ** not supported: arity=2 op=shuffle/rearrange vlen=%d etype=%s ismask=no",
1849 num_elem, type2name(elem_bt));
1850 return false; // not supported
1851 }
1852 }
1853 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1854 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1855
1856 ciKlass* shbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
1857 const TypeInstPtr* shbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, shbox_klass);
1858
1859 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1860 Node* shuffle = unbox_vector(argument(6), shbox_type, shuffle_bt, num_elem);
1861 const TypeVect* st = TypeVect::make(shuffle_bt, num_elem);
1862
1863 if (v1 == nullptr || shuffle == nullptr) {
1864 return false; // operand unboxing failed
1865 }
1866
1867 assert(is_power_of_2(num_elem), "wrapping invalid");
1868 Node* wrapping_mask_elem = gvn().makecon(TypeInteger::make(num_elem - 1, num_elem - 1, Type::WidenMin, shuffle_bt == T_LONG ? T_LONG : T_INT));
1869 Node* wrapping_mask = gvn().transform(VectorNode::scalar2vector(wrapping_mask_elem, num_elem, shuffle_bt));
1870 shuffle = gvn().transform(new AndVNode(shuffle, wrapping_mask, st));
1871
1872 Node* mask = nullptr;
1873 if (is_masked_op) {
1874 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1875 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1876 mask = unbox_vector(argument(7), mbox_type, elem_bt, num_elem);
1877 if (mask == nullptr) {
1878 log_if_needed(" ** not supported: arity=3 op=shuffle/rearrange vlen=%d etype=%s ismask=useload is_masked_op=1",
1879 num_elem, type2name(elem_bt));
1880 return false;
1881 }
1882 }
1883
1884 if (need_load_shuffle) {
1885 shuffle = gvn().transform(new VectorLoadShuffleNode(shuffle, st));
1886 }
1887
1888 Node* rearrange = new VectorRearrangeNode(v1, shuffle);
1889 if (is_masked_op) {
1890 if (use_predicate) {
1891 rearrange->add_req(mask);
1892 rearrange->add_flag(Node::Flag_is_predicated_vector);
1893 } else {
1894 rearrange = gvn().transform(rearrange);
1895 Node* zero = gvn().makecon(Type::get_zero_type(elem_bt));
1896 Node* zerovec = gvn().transform(VectorNode::scalar2vector(zero, num_elem, elem_bt));
1897 rearrange = new VectorBlendNode(zerovec, rearrange, mask);
1898 }
1899 }
1900 rearrange = gvn().transform(rearrange);
1901
1902 Node* box = box_vector(rearrange, vbox_type, elem_bt, num_elem);
1903 set_result(box);
1904 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1905 return true;
1906 }
1907
1908 // public static
1909 // <V extends Vector<E>,
1910 // M extends VectorMask<E>,
1911 // E>
1912 // V selectFromOp(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass,
1913 // int length, V v1, V v2, M m,
1914 // VectorSelectFromOp<V, M> defaultImpl)
1915 bool LibraryCallKit::inline_vector_select_from() {
1916 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1917 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
1918 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1919 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1920
1921 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
1922 vector_klass->const_oop() == nullptr ||
1923 elem_klass->const_oop() == nullptr ||
1924 !vlen->is_con()) {
1925 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s",
1926 NodeClassNames[argument(0)->Opcode()],
1927 NodeClassNames[argument(2)->Opcode()],
1928 NodeClassNames[argument(3)->Opcode()]);
1929 return false; // not enough info for intrinsification
1930 }
1931 if (!is_klass_initialized(vector_klass)) {
1932 log_if_needed(" ** klass argument not initialized");
1933 return false;
1934 }
1935 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1936 if (!elem_type->is_primitive_type()) {
1937 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1938 return false; // should be primitive type
1939 }
1940 BasicType elem_bt = elem_type->basic_type();
1941 int num_elem = vlen->get_con();
1942 if (!is_power_of_2(num_elem)) {
1943 log_if_needed(" ** vlen not power of two=%d", num_elem);
1944 return false;
1945 }
1946
1947 BasicType shuffle_bt = elem_bt;
1948 if (shuffle_bt == T_FLOAT) {
1949 shuffle_bt = T_INT;
1950 } else if (shuffle_bt == T_DOUBLE) {
1951 shuffle_bt = T_LONG;
1952 }
1953 bool need_load_shuffle = Matcher::vector_rearrange_requires_load_shuffle(shuffle_bt, num_elem);
1954
1955 int cast_vopc = VectorCastNode::opcode(-1, elem_bt); // from vector of type elem_bt
1956 if ((need_load_shuffle && !arch_supports_vector(Op_VectorLoadShuffle, num_elem, elem_bt, VecMaskNotUsed)) ||
1957 (elem_bt != shuffle_bt && !arch_supports_vector(cast_vopc, num_elem, shuffle_bt, VecMaskNotUsed)) ||
1958 !arch_supports_vector(Op_AndV, num_elem, shuffle_bt, VecMaskNotUsed) ||
1959 !arch_supports_vector(Op_Replicate, num_elem, shuffle_bt, VecMaskNotUsed)) {
1960 log_if_needed(" ** not supported: arity=0 op=selectFrom vlen=%d etype=%s ismask=no",
1961 num_elem, type2name(elem_bt));
1962 return false; // not supported
1963 }
1964
1965 bool is_masked_op = argument(6)->bottom_type() != TypePtr::NULL_PTR;
1966 bool use_predicate = is_masked_op;
1967 if (is_masked_op &&
1968 (mask_klass == nullptr ||
1969 mask_klass->const_oop() == nullptr ||
1970 !is_klass_initialized(mask_klass))) {
1971 log_if_needed(" ** mask_klass argument not initialized");
1972 return false; // not supported
1973 }
1974 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_masked_op ? (VecMaskUseLoad | VecMaskUsePred) : VecMaskNotUsed);
1975 if (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, checkFlags)) {
1976 use_predicate = false;
1977 if(!is_masked_op ||
1978 (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed) ||
1979 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad) ||
1980 !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed))) {
1981 log_if_needed(" ** not supported: op=selectFrom vlen=%d etype=%s is_masked_op=%d",
1982 num_elem, type2name(elem_bt), is_masked_op);
1983 return false; // not supported
1984 }
1985 }
1986 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1987 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1988
1989 // v1 is the index vector
1990 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
1991 // v2 is the vector being rearranged
1992 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1993
1994 if (v1 == nullptr) {
1995 log_if_needed(" ** unbox failed v1=%s", NodeClassNames[argument(4)->Opcode()]);
1996 return false; // operand unboxing failed
1997 }
1998
1999 if (v2 == nullptr) {
2000 log_if_needed(" ** unbox failed v2=%s", NodeClassNames[argument(5)->Opcode()]);
2001 return false; // operand unboxing failed
2002 }
2003
2004 Node* mask = nullptr;
2005 if (is_masked_op) {
2006 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
2007 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
2008 mask = unbox_vector(argument(6), mbox_type, elem_bt, num_elem);
2009 if (mask == nullptr) {
2010 log_if_needed(" ** unbox failed mask=%s", NodeClassNames[argument(6)->Opcode()]);
2011 return false;
2012 }
2013 }
2014
2015 // cast index vector from elem_bt vector to byte vector
2016 const TypeVect* shuffle_vt = TypeVect::make(shuffle_bt, num_elem);
2017 Node* shuffle = v1;
2018
2019 if (shuffle_bt != elem_bt) {
2020 shuffle = gvn().transform(VectorCastNode::make(cast_vopc, v1, shuffle_bt, num_elem));
2021 }
2022
2023 // wrap the byte vector lanes to (num_elem - 1) to form the shuffle vector where num_elem is vector length
2024 // this is a simple AND operation as we come here only for power of two vector length
2025 Node* mod_val = gvn().makecon(TypeInteger::make(num_elem - 1, num_elem - 1, Type::WidenMin, shuffle_bt == T_LONG ? T_LONG : T_INT));
2026 Node* bcast_mod = gvn().transform(VectorNode::scalar2vector(mod_val, num_elem, shuffle_bt));
2027 shuffle = gvn().transform(VectorNode::make(Op_AndV, shuffle, bcast_mod, shuffle_vt));
2028
2029 // load the shuffle to use in rearrange
2030 if (need_load_shuffle) {
2031 shuffle = gvn().transform(new VectorLoadShuffleNode(shuffle, shuffle_vt));
2032 }
2033
2034 // and finally rearrange
2035 Node* rearrange = new VectorRearrangeNode(v2, shuffle);
2036 if (is_masked_op) {
2037 if (use_predicate) {
2038 // masked rearrange is supported so use that directly
2039 rearrange->add_req(mask);
2040 rearrange->add_flag(Node::Flag_is_predicated_vector);
2041 } else {
2042 // masked rearrange is not supported so emulate usig blend
2043 const TypeVect* vt = v1->bottom_type()->is_vect();
2044 rearrange = gvn().transform(rearrange);
2045
2046 // create a zero vector with each lane element set as zero
2047 Node* zero = gvn().makecon(Type::get_zero_type(elem_bt));
2048 Node* zerovec = gvn().transform(VectorNode::scalar2vector(zero, num_elem, elem_bt));
2049
2050 // For each lane for which mask is set, blend in the rearranged lane into zero vector
2051 rearrange = new VectorBlendNode(zerovec, rearrange, mask);
2052 }
2053 }
2054 rearrange = gvn().transform(rearrange);
2055
2056 // box the result
2057 Node* box = box_vector(rearrange, vbox_type, elem_bt, num_elem);
2058 set_result(box);
2059
2060 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2061 return true;
2062 }
2063
2064 // public static
2065 // <V extends Vector<E>,
2066 // M extends VectorMask<E>,
2067 // E>
2068 // V broadcastInt(int opr, Class<? extends V> vectorClass, Class<? extends M> maskClass,
2069 // Class<E> elementType, int length,
2070 // V v, int n, M m,
2071 // VectorBroadcastIntOp<V, M> defaultImpl)
2072 bool LibraryCallKit::inline_vector_broadcast_int() {
2073 const TypeInt* opr = gvn().type(argument(0))->isa_int();
2074 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
2075 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
2076 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
2077 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
2078
2079 if (opr == nullptr || vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
2080 return false; // dead code
2081 }
2082 if (!opr->is_con() || vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
2083 log_if_needed(" ** missing constant: opr=%s vclass=%s etype=%s vlen=%s",
2084 NodeClassNames[argument(0)->Opcode()],
2085 NodeClassNames[argument(1)->Opcode()],
2086 NodeClassNames[argument(3)->Opcode()],
2087 NodeClassNames[argument(4)->Opcode()]);
2088 return false; // not enough info for intrinsification
2089 }
2090 if (!is_klass_initialized(vector_klass)) {
2091 log_if_needed(" ** klass argument not initialized");
2092 return false;
2093 }
2094
2095 const Type* vmask_type = gvn().type(argument(7));
2096 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
2097 if (is_masked_op) {
2098 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
2099 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(2)->Opcode()]);
2100 return false; // not enough info for intrinsification
2101 }
2102
2103 if (!is_klass_initialized(mask_klass)) {
2104 log_if_needed(" ** mask klass argument not initialized");
2105 return false;
2106 }
2107
2108 if (vmask_type->maybe_null()) {
2109 log_if_needed(" ** null mask values are not allowed for masked op");
2110 return false;
2111 }
2112 }
2113
2114 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2115 if (!elem_type->is_primitive_type()) {
2116 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2117 return false; // should be primitive type
2118 }
2119
2120 int num_elem = vlen->get_con();
2121 BasicType elem_bt = elem_type->basic_type();
2122 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
2123
2124 bool is_shift = VectorNode::is_shift_opcode(opc);
2125 bool is_rotate = VectorNode::is_rotate_opcode(opc);
2126
2127 if (opc == 0 || (!is_shift && !is_rotate)) {
2128 log_if_needed(" ** operation not supported: op=%d bt=%s", opr->get_con(), type2name(elem_bt));
2129 return false; // operation not supported
2130 }
2131
2132 int sopc = VectorNode::opcode(opc, elem_bt);
2133 if (sopc == 0) {
2134 log_if_needed(" ** operation not supported: opc=%s bt=%s", NodeClassNames[opc], type2name(elem_bt));
2135 return false; // operation not supported
2136 }
2137
2138 Node* cnt = argument(6);
2139 const TypeInt* cnt_type = cnt->bottom_type()->isa_int();
2140
2141 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2142 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2143
2144 // If CPU supports vector constant rotate instructions pass it directly
2145 bool is_const_rotate = is_rotate && cnt_type && cnt_type->is_con() &&
2146 Matcher::supports_vector_constant_rotates(cnt_type->get_con());
2147 bool has_scalar_args = is_rotate ? !is_const_rotate : true;
2148
2149 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_masked_op ? (VecMaskUseLoad | VecMaskUsePred) : VecMaskNotUsed);
2150 bool use_predicate = is_masked_op;
2151
2152 if (!arch_supports_vector(sopc, num_elem, elem_bt, checkFlags, has_scalar_args)) {
2153 use_predicate = false;
2154 if (!is_masked_op ||
2155 (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed, has_scalar_args) ||
2156 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad))) {
2157
2158 log_if_needed(" ** not supported: arity=0 op=int/%d vlen=%d etype=%s is_masked_op=%d",
2159 sopc, num_elem, type2name(elem_bt), is_masked_op ? 1 : 0);
2160 return false; // not supported
2161 }
2162 }
2163
2164 Node* opd1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
2165 Node* opd2 = nullptr;
2166 if (is_shift) {
2167 opd2 = vector_shift_count(cnt, opc, elem_bt, num_elem);
2168 } else {
2169 assert(is_rotate, "unexpected operation");
2170 if (!is_const_rotate) {
2171 cnt = elem_bt == T_LONG ? gvn().transform(new ConvI2LNode(cnt)) : cnt;
2172 opd2 = gvn().transform(VectorNode::scalar2vector(cnt, num_elem, elem_bt));
2173 } else {
2174 // Constant shift value.
2175 opd2 = cnt;
2176 }
2177 }
2178
2179 if (opd1 == nullptr || opd2 == nullptr) {
2180 return false;
2181 }
2182
2183 Node* mask = nullptr;
2184 if (is_masked_op) {
2185 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
2186 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
2187 mask = unbox_vector(argument(7), mbox_type, elem_bt, num_elem);
2188 if (mask == nullptr) {
2189 log_if_needed(" ** unbox failed mask=%s", NodeClassNames[argument(7)->Opcode()]);
2190 return false;
2191 }
2192 }
2193
2194 Node* operation = VectorNode::make(opc, opd1, opd2, num_elem, elem_bt);
2195 if (is_masked_op && mask != nullptr) {
2196 if (use_predicate) {
2197 operation->add_req(mask);
2198 operation->add_flag(Node::Flag_is_predicated_vector);
2199 } else {
2200 operation = gvn().transform(operation);
2201 operation = new VectorBlendNode(opd1, operation, mask);
2202 }
2203 }
2204 operation = gvn().transform(operation);
2205 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
2206 set_result(vbox);
2207 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2208 return true;
2209 }
2210
2211 // public static <VOUT extends VectorPayload,
2212 // VIN extends VectorPayload,
2213 // S extends VectorSpecies>
2214 // VOUT convert(int oprId,
2215 // Class<?> fromVectorClass, Class<?> fromElementType, int fromVLen,
2216 // Class<?> toVectorClass, Class<?> toElementType, int toVLen,
2217 // VIN v, S s,
2218 // VectorConvertOp<VOUT, VIN, S> defaultImpl)
2219 //
2220 bool LibraryCallKit::inline_vector_convert() {
2221 const TypeInt* opr = gvn().type(argument(0))->isa_int();
2222
2223 const TypeInstPtr* vector_klass_from = gvn().type(argument(1))->isa_instptr();
2224 const TypeInstPtr* elem_klass_from = gvn().type(argument(2))->isa_instptr();
2225 const TypeInt* vlen_from = gvn().type(argument(3))->isa_int();
2226
2227 const TypeInstPtr* vector_klass_to = gvn().type(argument(4))->isa_instptr();
2228 const TypeInstPtr* elem_klass_to = gvn().type(argument(5))->isa_instptr();
2229 const TypeInt* vlen_to = gvn().type(argument(6))->isa_int();
2230
2231 if (opr == nullptr ||
2232 vector_klass_from == nullptr || elem_klass_from == nullptr || vlen_from == nullptr ||
2233 vector_klass_to == nullptr || elem_klass_to == nullptr || vlen_to == nullptr) {
2234 return false; // dead code
2235 }
2236 if (!opr->is_con() ||
2237 vector_klass_from->const_oop() == nullptr || elem_klass_from->const_oop() == nullptr || !vlen_from->is_con() ||
2238 vector_klass_to->const_oop() == nullptr || elem_klass_to->const_oop() == nullptr || !vlen_to->is_con()) {
2239 log_if_needed(" ** missing constant: opr=%s vclass_from=%s etype_from=%s vlen_from=%s vclass_to=%s etype_to=%s vlen_to=%s",
2240 NodeClassNames[argument(0)->Opcode()],
2241 NodeClassNames[argument(1)->Opcode()],
2242 NodeClassNames[argument(2)->Opcode()],
2243 NodeClassNames[argument(3)->Opcode()],
2244 NodeClassNames[argument(4)->Opcode()],
2245 NodeClassNames[argument(5)->Opcode()],
2246 NodeClassNames[argument(6)->Opcode()]);
2247 return false; // not enough info for intrinsification
2248 }
2249 if (!is_klass_initialized(vector_klass_from) || !is_klass_initialized(vector_klass_to)) {
2250 log_if_needed(" ** klass argument not initialized");
2251 return false;
2252 }
2253
2254 assert(opr->get_con() == VectorSupport::VECTOR_OP_CAST ||
2255 opr->get_con() == VectorSupport::VECTOR_OP_UCAST ||
2256 opr->get_con() == VectorSupport::VECTOR_OP_REINTERPRET, "wrong opcode");
2257 bool is_cast = (opr->get_con() == VectorSupport::VECTOR_OP_CAST || opr->get_con() == VectorSupport::VECTOR_OP_UCAST);
2258 bool is_ucast = (opr->get_con() == VectorSupport::VECTOR_OP_UCAST);
2259
2260 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
2261 ciKlass* vbox_klass_to = vector_klass_to->const_oop()->as_instance()->java_lang_Class_klass();
2262
2263 bool is_mask = is_vector_mask(vbox_klass_from);
2264
2265 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
2266 if (!elem_type_from->is_primitive_type()) {
2267 return false; // should be primitive type
2268 }
2269 BasicType elem_bt_from = elem_type_from->basic_type();
2270 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
2271 if (!elem_type_to->is_primitive_type()) {
2272 return false; // should be primitive type
2273 }
2274 BasicType elem_bt_to = elem_type_to->basic_type();
2275
2276 int num_elem_from = vlen_from->get_con();
2277 int num_elem_to = vlen_to->get_con();
2278
2279 // Check whether we can unbox to appropriate size. Even with casting, checking for reinterpret is needed
2280 // since we may need to change size.
2281 if (!arch_supports_vector(Op_VectorReinterpret,
2282 num_elem_from,
2283 elem_bt_from,
2284 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
2285 log_if_needed(" ** not supported: arity=1 op=%s/1 vlen1=%d etype1=%s ismask=%d",
2286 is_cast ? "cast" : "reinterpret",
2287 num_elem_from, type2name(elem_bt_from), is_mask);
2288 return false;
2289 }
2290
2291 // Check whether we can support resizing/reinterpreting to the new size.
2292 if (!arch_supports_vector(Op_VectorReinterpret,
2293 num_elem_to,
2294 elem_bt_to,
2295 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
2296 log_if_needed(" ** not supported: arity=1 op=%s/2 vlen2=%d etype2=%s ismask=%d",
2297 is_cast ? "cast" : "reinterpret",
2298 num_elem_to, type2name(elem_bt_to), is_mask);
2299 return false;
2300 }
2301
2302 // At this point, we know that both input and output vector registers are supported
2303 // by the architecture. Next check if the casted type is simply to same type - which means
2304 // that it is actually a resize and not a cast.
2305 if (is_cast && elem_bt_from == elem_bt_to) {
2306 is_cast = false;
2307 }
2308
2309 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
2310
2311 Node* opd1 = unbox_vector(argument(7), vbox_type_from, elem_bt_from, num_elem_from);
2312 if (opd1 == nullptr) {
2313 return false;
2314 }
2315
2316 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from, is_mask);
2317 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to, is_mask);
2318
2319 // Safety check to prevent casting if source mask is of type vector
2320 // and destination mask of type predicate vector and vice-versa.
2321 // From X86 standpoint, this case will only arise over KNL target,
2322 // where certain masks (depending on the species) are either propagated
2323 // through a vector or predicate register.
2324 if (is_mask &&
2325 ((src_type->isa_vectmask() == nullptr && dst_type->isa_vectmask()) ||
2326 (dst_type->isa_vectmask() == nullptr && src_type->isa_vectmask()))) {
2327 return false;
2328 }
2329
2330 Node* op = opd1;
2331 if (is_cast) {
2332 assert(!is_mask || num_elem_from == num_elem_to, "vector mask cast needs the same elem num");
2333 int cast_vopc = VectorCastNode::opcode(-1, elem_bt_from, !is_ucast);
2334
2335 // Make sure that vector cast is implemented to particular type/size combination if it is
2336 // not a mask casting.
2337 if (!is_mask && !arch_supports_vector(cast_vopc, num_elem_to, elem_bt_to, VecMaskNotUsed)) {
2338 log_if_needed(" ** not supported: arity=1 op=cast#%d/3 vlen2=%d etype2=%s ismask=%d",
2339 cast_vopc, num_elem_to, type2name(elem_bt_to), is_mask);
2340 return false;
2341 }
2342
2343 if (num_elem_from < num_elem_to) {
2344 // Since input and output number of elements are not consistent, we need to make sure we
2345 // properly size. Thus, first make a cast that retains the number of elements from source.
2346 int num_elem_for_cast = num_elem_from;
2347
2348 // It is possible that arch does not support this intermediate vector size
2349 // TODO More complex logic required here to handle this corner case for the sizes.
2350 if (!arch_supports_vector(cast_vopc, num_elem_for_cast, elem_bt_to, VecMaskNotUsed)) {
2351 log_if_needed(" ** not supported: arity=1 op=cast#%d/4 vlen1=%d etype2=%s ismask=%d",
2352 cast_vopc,
2353 num_elem_for_cast, type2name(elem_bt_to), is_mask);
2354 return false;
2355 }
2356
2357 op = gvn().transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_for_cast));
2358 // Now ensure that the destination gets properly resized to needed size.
2359 op = gvn().transform(new VectorReinterpretNode(op, op->bottom_type()->is_vect(), dst_type));
2360 } else if (num_elem_from > num_elem_to) {
2361 // Since number of elements from input is larger than output, simply reduce size of input
2362 // (we are supposed to drop top elements anyway).
2363 int num_elem_for_resize = num_elem_to;
2364
2365 // It is possible that arch does not support this intermediate vector size
2366 // TODO More complex logic required here to handle this corner case for the sizes.
2367 if (!arch_supports_vector(Op_VectorReinterpret,
2368 num_elem_for_resize,
2369 elem_bt_from,
2370 VecMaskNotUsed)) {
2371 log_if_needed(" ** not supported: arity=1 op=cast/5 vlen2=%d etype1=%s ismask=%d",
2372 num_elem_for_resize, type2name(elem_bt_from), is_mask);
2373 return false;
2374 }
2375
2376 const TypeVect* resize_type = TypeVect::make(elem_bt_from, num_elem_for_resize);
2377 op = gvn().transform(new VectorReinterpretNode(op, src_type, resize_type));
2378 op = gvn().transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
2379 } else { // num_elem_from == num_elem_to
2380 if (is_mask) {
2381 // Make sure that cast for vector mask is implemented to particular type/size combination.
2382 if (!arch_supports_vector(Op_VectorMaskCast, num_elem_to, elem_bt_to, VecMaskNotUsed)) {
2383 log_if_needed(" ** not supported: arity=1 op=maskcast vlen2=%d etype2=%s ismask=%d",
2384 num_elem_to, type2name(elem_bt_to), is_mask);
2385 return false;
2386 }
2387 op = gvn().transform(new VectorMaskCastNode(op, dst_type));
2388 } else {
2389 // Since input and output number of elements match, and since we know this vector size is
2390 // supported, simply do a cast with no resize needed.
2391 op = gvn().transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
2392 }
2393 }
2394 } else if (!Type::equals(src_type, dst_type)) {
2395 assert(!is_cast, "must be reinterpret");
2396 op = gvn().transform(new VectorReinterpretNode(op, src_type, dst_type));
2397 }
2398
2399 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
2400 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
2401 set_result(vbox);
2402 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem_to * type2aelembytes(elem_bt_to))));
2403 return true;
2404 }
2405
2406 // public static
2407 // <V extends Vector<E>,
2408 // E>
2409 // V insert(Class<? extends V> vectorClass, Class<E> elementType, int vlen,
2410 // V vec, int ix, long val,
2411 // VecInsertOp<V> defaultImpl)
2412 bool LibraryCallKit::inline_vector_insert() {
2413 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2414 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
2415 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
2416 const TypeInt* idx = gvn().type(argument(4))->isa_int();
2417
2418 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr || idx == nullptr) {
2419 return false; // dead code
2420 }
2421 if (vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con() || !idx->is_con()) {
2422 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s idx=%s",
2423 NodeClassNames[argument(0)->Opcode()],
2424 NodeClassNames[argument(1)->Opcode()],
2425 NodeClassNames[argument(2)->Opcode()],
2426 NodeClassNames[argument(4)->Opcode()]);
2427 return false; // not enough info for intrinsification
2428 }
2429 if (!is_klass_initialized(vector_klass)) {
2430 log_if_needed(" ** klass argument not initialized");
2431 return false;
2432 }
2433 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2434 if (!elem_type->is_primitive_type()) {
2435 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2436 return false; // should be primitive type
2437 }
2438 BasicType elem_bt = elem_type->basic_type();
2439 int num_elem = vlen->get_con();
2440 if (!arch_supports_vector(Op_VectorInsert, num_elem, elem_bt, VecMaskNotUsed)) {
2441 log_if_needed(" ** not supported: arity=1 op=insert vlen=%d etype=%s ismask=no",
2442 num_elem, type2name(elem_bt));
2443 return false; // not supported
2444 }
2445
2446 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2447 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2448
2449 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2450 if (opd == nullptr) {
2451 return false;
2452 }
2453
2454 Node* insert_val = argument(5);
2455 assert(gvn().type(insert_val)->isa_long() != nullptr, "expected to be long");
2456
2457 // Convert insert value back to its appropriate type.
2458 switch (elem_bt) {
2459 case T_BYTE:
2460 insert_val = gvn().transform(new ConvL2INode(insert_val, TypeInt::BYTE));
2461 break;
2462 case T_SHORT:
2463 insert_val = gvn().transform(new ConvL2INode(insert_val, TypeInt::SHORT));
2464 break;
2465 case T_INT:
2466 insert_val = gvn().transform(new ConvL2INode(insert_val));
2467 break;
2468 case T_FLOAT:
2469 insert_val = gvn().transform(new ConvL2INode(insert_val));
2470 insert_val = gvn().transform(new MoveI2FNode(insert_val));
2471 break;
2472 case T_DOUBLE:
2473 insert_val = gvn().transform(new MoveL2DNode(insert_val));
2474 break;
2475 case T_LONG:
2476 // no conversion needed
2477 break;
2478 default: fatal("%s", type2name(elem_bt)); break;
2479 }
2480
2481 Node* operation = gvn().transform(VectorInsertNode::make(opd, insert_val, idx->get_con(), gvn()));
2482
2483 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
2484 set_result(vbox);
2485 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2486 return true;
2487 }
2488
2489 // public static
2490 // <VM extends VectorPayload,
2491 // E>
2492 // long extract(Class<? extends VM> vClass, Class<E> eClass,
2493 // int length,
2494 // VM vm, int i,
2495 // VecExtractOp<VM> defaultImpl)
2496 bool LibraryCallKit::inline_vector_extract() {
2497 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2498 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
2499 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
2500 const TypeInt* idx = gvn().type(argument(4))->isa_int();
2501
2502 if (vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
2503 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
2504 vlen == nullptr || !vlen->is_con() ||
2505 idx == nullptr || !idx->is_con()) {
2506 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s",
2507 NodeClassNames[argument(0)->Opcode()],
2508 NodeClassNames[argument(1)->Opcode()],
2509 NodeClassNames[argument(2)->Opcode()]);
2510 return false; // not enough info for intrinsification
2511 }
2512 if (!is_klass_initialized(vector_klass)) {
2513 log_if_needed(" ** klass argument not initialized");
2514 return false;
2515 }
2516 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2517 if (!elem_type->is_primitive_type()) {
2518 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2519 return false; // should be primitive type
2520 }
2521 BasicType elem_bt = elem_type->basic_type();
2522 int num_elem = vlen->get_con();
2523
2524 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2525 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2526
2527 Node* opd = nullptr;
2528
2529 if (is_vector_mask(vbox_klass)) {
2530 // vbox_klass is mask. This is used for VectorMask.laneIsSet(int).
2531
2532 Node* pos = argument(4); // can be variable
2533 if (arch_supports_vector(Op_ExtractUB, num_elem, elem_bt, VecMaskUseAll)) {
2534 // Transform mask to vector with type of boolean and utilize ExtractUB node.
2535 opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2536 if (opd == nullptr) {
2537 return false;
2538 }
2539 opd = gvn().transform(VectorStoreMaskNode::make(gvn(), opd, elem_bt, num_elem));
2540 opd = gvn().transform(new ExtractUBNode(opd, pos));
2541 opd = gvn().transform(new ConvI2LNode(opd));
2542 } else if (arch_supports_vector(Op_VectorMaskToLong, num_elem, elem_bt, VecMaskUseLoad)) {
2543 opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2544 if (opd == nullptr) {
2545 return false;
2546 }
2547 // VectorMaskToLongNode requires the input is either a mask or a vector with BOOLEAN type.
2548 if (opd->bottom_type()->isa_vectmask() == nullptr) {
2549 opd = gvn().transform(VectorStoreMaskNode::make(gvn(), opd, elem_bt, num_elem));
2550 }
2551 // ((toLong() >>> pos) & 1L
2552 opd = gvn().transform(new VectorMaskToLongNode(opd, TypeLong::LONG));
2553 opd = gvn().transform(new URShiftLNode(opd, pos));
2554 opd = gvn().transform(new AndLNode(opd, gvn().makecon(TypeLong::ONE)));
2555 } else {
2556 log_if_needed(" ** Rejected mask extraction because architecture does not support it");
2557 return false; // not supported
2558 }
2559 } else {
2560 // vbox_klass is vector. This is used for Vector.lane(int).
2561 if (!idx->is_con()) {
2562 log_if_needed(" ** missing constant: idx=%s", NodeClassNames[argument(4)->Opcode()]);
2563 return false; // not enough info for intrinsification
2564 }
2565
2566 int vopc = ExtractNode::opcode(elem_bt);
2567 if (!arch_supports_vector(vopc, num_elem, elem_bt, VecMaskNotUsed)) {
2568 log_if_needed(" ** not supported: arity=1 op=extract vlen=%d etype=%s ismask=no",
2569 num_elem, type2name(elem_bt));
2570 return false; // not supported
2571 }
2572
2573 opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2574 if (opd == nullptr) {
2575 return false;
2576 }
2577 ConINode* idx_con = gvn().intcon(idx->get_con())->as_ConI();
2578
2579 opd = gvn().transform(ExtractNode::make(opd, idx_con, elem_bt));
2580 switch (elem_bt) {
2581 case T_BYTE:
2582 case T_SHORT:
2583 case T_INT: {
2584 opd = gvn().transform(new ConvI2LNode(opd));
2585 break;
2586 }
2587 case T_FLOAT: {
2588 opd = gvn().transform(new MoveF2INode(opd));
2589 opd = gvn().transform(new ConvI2LNode(opd));
2590 break;
2591 }
2592 case T_DOUBLE: {
2593 opd = gvn().transform(new MoveD2LNode(opd));
2594 break;
2595 }
2596 case T_LONG: {
2597 // no conversion needed
2598 break;
2599 }
2600 default: fatal("%s", type2name(elem_bt));
2601 }
2602 }
2603 set_result(opd);
2604 return true;
2605 }
2606
2607 static Node* LowerSelectFromTwoVectorOperation(PhaseGVN& phase, Node* index_vec, Node* src1, Node* src2, const TypeVect* vt) {
2608 int num_elem = vt->length();
2609 BasicType elem_bt = vt->element_basic_type();
2610
2611 // Lower selectFrom operation into its constituent operations.
2612 // SelectFromTwoVectorNode =
2613 // (VectorBlend
2614 // (VectorRearrange SRC1 (WRAPED_INDEX AND (VLEN-1))
2615 // (VectorRearrange SRC2 (WRAPED_INDEX AND (VLEN-1))
2616 // MASK)
2617 // Where
2618 // WRAPED_INDEX are computed by wrapping incoming indexes
2619 // to two vector index range [0, VLEN*2) and
2620 // MASK = WRAPED_INDEX < VLEN
2621 //
2622 // IR lowering prevents intrinsification failure and associated argument
2623 // boxing penalties.
2624 //
2625
2626 BasicType shuffle_bt = elem_bt;
2627 if (shuffle_bt == T_FLOAT) {
2628 shuffle_bt = T_INT;
2629 } else if (shuffle_bt == T_DOUBLE) {
2630 shuffle_bt = T_LONG;
2631 }
2632 const TypeVect* st = TypeVect::make(shuffle_bt, num_elem);
2633
2634 // Cast index vector to the corresponding bit type
2635 if (elem_bt != shuffle_bt) {
2636 int cast_vopc = VectorCastNode::opcode(0, elem_bt, true);
2637 index_vec = phase.transform(VectorCastNode::make(cast_vopc, index_vec, shuffle_bt, num_elem));
2638 }
2639
2640 // Wrap indexes into two vector index range [0, VLEN * 2)
2641 Node* two_vect_lane_cnt_m1 = phase.makecon(TypeInteger::make(2 * num_elem - 1, 2 * num_elem - 1, Type::WidenMin, shuffle_bt == T_LONG ? T_LONG : T_INT));
2642 Node* bcast_two_vect_lane_cnt_m1_vec = phase.transform(VectorNode::scalar2vector(two_vect_lane_cnt_m1, num_elem,
2643 shuffle_bt, false));
2644 index_vec = phase.transform(VectorNode::make(Op_AndV, index_vec, bcast_two_vect_lane_cnt_m1_vec, st));
2645
2646 // Compute the blend mask for merging two independently permitted vectors
2647 // using shuffle index in two vector index range [0, VLEN * 2).
2648 BoolTest::mask pred = BoolTest::le;
2649 ConINode* pred_node = phase.makecon(TypeInt::make(pred))->as_ConI();
2650 const TypeVect* vmask_type = TypeVect::makemask(shuffle_bt, num_elem);
2651 Node* lane_cnt_m1 = phase.makecon(TypeInteger::make(num_elem - 1, num_elem - 1, Type::WidenMin, shuffle_bt == T_LONG ? T_LONG : T_INT));
2652 Node* bcast_lane_cnt_m1_vec = phase.transform(VectorNode::scalar2vector(lane_cnt_m1, num_elem, shuffle_bt, false));
2653 Node* mask = phase.transform(new VectorMaskCmpNode(pred, index_vec, bcast_lane_cnt_m1_vec, pred_node, vmask_type));
2654
2655 // Rearrange expects the indexes to lie within single vector index range [0, VLEN).
2656 Node* wrapped_index_vec = phase.transform(VectorNode::make(Op_AndV, index_vec, bcast_lane_cnt_m1_vec, st));
2657
2658 // Load indexes from byte vector and appropriately transform them to target
2659 // specific permutation index format.
2660 if (Matcher::vector_rearrange_requires_load_shuffle(shuffle_bt, num_elem)) {
2661 wrapped_index_vec = phase.transform(new VectorLoadShuffleNode(wrapped_index_vec, st));
2662 }
2663
2664 vmask_type = TypeVect::makemask(elem_bt, num_elem);
2665 mask = phase.transform(new VectorMaskCastNode(mask, vmask_type));
2666
2667 Node* p1 = phase.transform(new VectorRearrangeNode(src1, wrapped_index_vec));
2668 Node* p2 = phase.transform(new VectorRearrangeNode(src2, wrapped_index_vec));
2669
2670 return new VectorBlendNode(p2, p1, mask);
2671 }
2672
2673 // public static
2674 // <V extends Vector<E>,
2675 // E>
2676 // V selectFromTwoVectorOp(Class<? extends V> vClass, Class<E> eClass, int length,
2677 // V v1, V v2, V v3,
2678 // SelectFromTwoVector<V> defaultImpl)
2679 bool LibraryCallKit::inline_vector_select_from_two_vectors() {
2680 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2681 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
2682 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
2683
2684 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr || vector_klass->const_oop() == nullptr ||
2685 elem_klass->const_oop() == nullptr ||!vlen->is_con()) {
2686 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s",
2687 NodeClassNames[argument(0)->Opcode()],
2688 NodeClassNames[argument(1)->Opcode()],
2689 NodeClassNames[argument(2)->Opcode()]);
2690 return false; // not enough info for intrinsification
2691 }
2692
2693 if (!is_klass_initialized(vector_klass)) {
2694 log_if_needed(" ** klass argument not initialized");
2695 return false;
2696 }
2697
2698 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2699 if (!elem_type->is_primitive_type()) {
2700 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2701 return false; // should be primitive type
2702 }
2703
2704 int num_elem = vlen->get_con();
2705 if (!is_power_of_2(num_elem)) {
2706 log_if_needed(" ** vlen is not power of two=%d", num_elem);
2707 return false;
2708 }
2709
2710 BasicType elem_bt = elem_type->basic_type();
2711 BasicType index_elem_bt = elem_bt;
2712 if (elem_bt == T_FLOAT) {
2713 index_elem_bt = T_INT;
2714 } else if (elem_bt == T_DOUBLE) {
2715 index_elem_bt = T_LONG;
2716 }
2717
2718 // Check if the platform requires a VectorLoadShuffle node to be generated
2719 bool need_load_shuffle = Matcher::vector_rearrange_requires_load_shuffle(index_elem_bt, num_elem);
2720
2721 bool lowerSelectFromOp = false;
2722 if (!arch_supports_vector(Op_SelectFromTwoVector, num_elem, elem_bt, VecMaskNotUsed)) {
2723 int cast_vopc = VectorCastNode::opcode(-1, elem_bt, true);
2724 if ((elem_bt != index_elem_bt && !arch_supports_vector(cast_vopc, num_elem, index_elem_bt, VecMaskNotUsed)) ||
2725 !arch_supports_vector(Op_VectorMaskCmp, num_elem, index_elem_bt, VecMaskNotUsed) ||
2726 !arch_supports_vector(Op_AndV, num_elem, index_elem_bt, VecMaskNotUsed) ||
2727 !arch_supports_vector(Op_VectorMaskCast, num_elem, elem_bt, VecMaskNotUsed) ||
2728 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad) ||
2729 !arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed) ||
2730 (need_load_shuffle && !arch_supports_vector(Op_VectorLoadShuffle, num_elem, index_elem_bt, VecMaskNotUsed)) ||
2731 !arch_supports_vector(Op_Replicate, num_elem, index_elem_bt, VecMaskNotUsed)) {
2732 log_if_needed(" ** not supported: opc=%d vlen=%d etype=%s ismask=useload",
2733 Op_SelectFromTwoVector, num_elem, type2name(elem_bt));
2734 return false; // not supported
2735 }
2736 lowerSelectFromOp = true;
2737 }
2738
2739 int cast_vopc = VectorCastNode::opcode(-1, elem_bt, true);
2740 if (!lowerSelectFromOp) {
2741 if (!arch_supports_vector(Op_AndV, num_elem, index_elem_bt, VecMaskNotUsed) ||
2742 !arch_supports_vector(Op_Replicate, num_elem, index_elem_bt, VecMaskNotUsed) ||
2743 (is_floating_point_type(elem_bt) &&
2744 !arch_supports_vector(cast_vopc, num_elem, index_elem_bt, VecMaskNotUsed))) {
2745 log_if_needed(" ** index wrapping not supported: vlen=%d etype=%s" ,
2746 num_elem, type2name(elem_bt));
2747 return false; // not supported
2748 }
2749 }
2750
2751 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2752 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2753
2754 Node* opd1 = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2755 if (opd1 == nullptr) {
2756 log_if_needed(" ** unbox failed v1=%s",
2757 NodeClassNames[argument(3)->Opcode()]);
2758 return false;
2759 }
2760 Node* opd2 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
2761 if (opd2 == nullptr) {
2762 log_if_needed(" ** unbox failed v2=%s",
2763 NodeClassNames[argument(4)->Opcode()]);
2764 return false;
2765 }
2766 Node* opd3 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
2767 if (opd3 == nullptr) {
2768 log_if_needed(" ** unbox failed v3=%s",
2769 NodeClassNames[argument(5)->Opcode()]);
2770 return false;
2771 }
2772
2773 const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
2774
2775 Node* operation = nullptr;
2776 if (lowerSelectFromOp) {
2777 operation = gvn().transform(LowerSelectFromTwoVectorOperation(gvn(), opd1, opd2, opd3, vt));
2778 } else {
2779 if (index_elem_bt != elem_bt) {
2780 opd1 = gvn().transform(VectorCastNode::make(cast_vopc, opd1, index_elem_bt, num_elem));
2781 }
2782 int indexRangeMask = 2 * num_elem - 1;
2783 Node* wrap_mask = gvn().makecon(TypeInteger::make(indexRangeMask, indexRangeMask, Type::WidenMin, index_elem_bt != T_LONG ? T_INT : index_elem_bt));
2784 Node* wrap_mask_vec = gvn().transform(VectorNode::scalar2vector(wrap_mask, num_elem, index_elem_bt, false));
2785 opd1 = gvn().transform(VectorNode::make(Op_AndV, opd1, wrap_mask_vec, opd1->bottom_type()->is_vect()));
2786 operation = gvn().transform(VectorNode::make(Op_SelectFromTwoVector, opd1, opd2, opd3, vt));
2787 }
2788
2789 // Wrap it up in VectorBox to keep object type information.
2790 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
2791 set_result(vbox);
2792 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2793 return true;
2794 }
2795
2796 // public static
2797 // <V extends Vector<E>,
2798 // M extends VectorMask<E>,
2799 // E>
2800 // V compressExpandOp(int opr,
2801 // Class<? extends V> vClass, Class<? extends M> mClass, Class<E> eClass,
2802 // int length, V v, M m,
2803 // CompressExpandOperation<V, M> defaultImpl)
2804 bool LibraryCallKit::inline_vector_compress_expand() {
2805 const TypeInt* opr = gvn().type(argument(0))->isa_int();
2806 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
2807 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
2808 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
2809 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
2810
2811 if (opr == nullptr || !opr->is_con() ||
2812 vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
2813 mask_klass == nullptr || mask_klass->const_oop() == nullptr ||
2814 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
2815 vlen == nullptr || !vlen->is_con()) {
2816 log_if_needed(" ** missing constant: opr=%s vclass=%s mclass=%s etype=%s vlen=%s",
2817 NodeClassNames[argument(0)->Opcode()],
2818 NodeClassNames[argument(1)->Opcode()],
2819 NodeClassNames[argument(2)->Opcode()],
2820 NodeClassNames[argument(3)->Opcode()],
2821 NodeClassNames[argument(4)->Opcode()]);
2822 return false; // not enough info for intrinsification
2823 }
2824
2825 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(mask_klass)) {
2826 log_if_needed(" ** klass argument not initialized");
2827 return false;
2828 }
2829
2830 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2831 if (!elem_type->is_primitive_type()) {
2832 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2833 return false; // should be primitive type
2834 }
2835
2836 int num_elem = vlen->get_con();
2837 BasicType elem_bt = elem_type->basic_type();
2838 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
2839
2840 if (!arch_supports_vector(opc, num_elem, elem_bt, VecMaskUseLoad)) {
2841 log_if_needed(" ** not supported: opc=%d vlen=%d etype=%s ismask=useload",
2842 opc, num_elem, type2name(elem_bt));
2843 return false; // not supported
2844 }
2845
2846 Node* opd1 = nullptr;
2847 const TypeInstPtr* vbox_type = nullptr;
2848 if (opc != Op_CompressM) {
2849 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2850 vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2851 opd1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
2852 if (opd1 == nullptr) {
2853 log_if_needed(" ** unbox failed vector=%s",
2854 NodeClassNames[argument(5)->Opcode()]);
2855 return false;
2856 }
2857 }
2858
2859 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
2860 assert(is_vector_mask(mbox_klass), "argument(6) should be a mask class");
2861 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
2862
2863 Node* mask = unbox_vector(argument(6), mbox_type, elem_bt, num_elem);
2864 if (mask == nullptr) {
2865 log_if_needed(" ** unbox failed mask=%s",
2866 NodeClassNames[argument(6)->Opcode()]);
2867 return false;
2868 }
2869
2870 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, opc == Op_CompressM);
2871 Node* operation = gvn().transform(VectorNode::make(opc, opd1, mask, vt));
2872
2873 // Wrap it up in VectorBox to keep object type information.
2874 const TypeInstPtr* box_type = opc == Op_CompressM ? mbox_type : vbox_type;
2875 Node* vbox = box_vector(operation, box_type, elem_bt, num_elem);
2876 set_result(vbox);
2877 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2878 return true;
2879 }
2880
2881 // public static
2882 // <V extends Vector<E>,
2883 // E,
2884 // S extends VectorSpecies<E>>
2885 // V indexVector(Class<? extends V> vClass, Class<E> eClass,
2886 // int length,
2887 // V v, int step, S s,
2888 // IndexOperation<V, S> defaultImpl)
2889 bool LibraryCallKit::inline_index_vector() {
2890 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2891 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
2892 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
2893
2894 if (vector_klass == nullptr || vector_klass->const_oop() == nullptr ||
2895 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
2896 vlen == nullptr || !vlen->is_con() ) {
2897 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s",
2898 NodeClassNames[argument(0)->Opcode()],
2899 NodeClassNames[argument(1)->Opcode()],
2900 NodeClassNames[argument(2)->Opcode()]);
2901 return false; // not enough info for intrinsification
2902 }
2903
2904 if (!is_klass_initialized(vector_klass)) {
2905 log_if_needed(" ** klass argument not initialized");
2906 return false;
2907 }
2908
2909 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2910 if (!elem_type->is_primitive_type()) {
2911 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2912 return false; // should be primitive type
2913 }
2914
2915 int num_elem = vlen->get_con();
2916 BasicType elem_bt = elem_type->basic_type();
2917
2918 // Check whether the iota index generation op is supported by the current hardware
2919 if (!arch_supports_vector(Op_VectorLoadConst, num_elem, elem_bt, VecMaskNotUsed)) {
2920 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
2921 return false; // not supported
2922 }
2923
2924 int mul_op = VectorSupport::vop2ideal(VectorSupport::VECTOR_OP_MUL, elem_bt);
2925 int vmul_op = VectorNode::opcode(mul_op, elem_bt);
2926 bool needs_mul = true;
2927 Node* scale = argument(4);
2928 const TypeInt* scale_type = gvn().type(scale)->isa_int();
2929 // Multiply is not needed if the scale is a constant "1".
2930 if (scale_type && scale_type->is_con() && scale_type->get_con() == 1) {
2931 needs_mul = false;
2932 } else {
2933 // Check whether the vector multiply op is supported by the current hardware
2934 if (!arch_supports_vector(vmul_op, num_elem, elem_bt, VecMaskNotUsed)) {
2935 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
2936 return false; // not supported
2937 }
2938
2939 // Check whether the scalar cast op is supported by the current hardware
2940 if (is_floating_point_type(elem_bt) || elem_bt == T_LONG) {
2941 int cast_op = elem_bt == T_LONG ? Op_ConvI2L :
2942 elem_bt == T_FLOAT? Op_ConvI2F : Op_ConvI2D;
2943 if (!Matcher::match_rule_supported(cast_op)) {
2944 log_if_needed(" ** Rejected op (%s) because architecture does not support it",
2945 NodeClassNames[cast_op]);
2946 return false; // not supported
2947 }
2948 }
2949 }
2950
2951 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2952 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2953 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2954 if (opd == nullptr) {
2955 log_if_needed(" ** unbox failed vector=%s",
2956 NodeClassNames[argument(3)->Opcode()]);
2957 return false;
2958 }
2959
2960 int add_op = VectorSupport::vop2ideal(VectorSupport::VECTOR_OP_ADD, elem_bt);
2961 int vadd_op = VectorNode::opcode(add_op, elem_bt);
2962 bool needs_add = true;
2963 // The addition is not needed if all the element values of "opd" are zero
2964 if (VectorNode::is_all_zeros_vector(opd)) {
2965 needs_add = false;
2966 } else {
2967 // Check whether the vector addition op is supported by the current hardware
2968 if (!arch_supports_vector(vadd_op, num_elem, elem_bt, VecMaskNotUsed)) {
2969 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
2970 return false; // not supported
2971 }
2972 }
2973
2974 // Compute the iota indice vector
2975 const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
2976 Node* index = gvn().transform(new VectorLoadConstNode(gvn().makecon(TypeInt::ZERO), vt));
2977
2978 // Broadcast the "scale" to a vector, and multiply the "scale" with iota indice vector.
2979 if (needs_mul) {
2980 switch (elem_bt) {
2981 case T_BOOLEAN: // fall-through
2982 case T_BYTE: // fall-through
2983 case T_SHORT: // fall-through
2984 case T_CHAR: // fall-through
2985 case T_INT: {
2986 // no conversion needed
2987 break;
2988 }
2989 case T_LONG: {
2990 scale = gvn().transform(new ConvI2LNode(scale));
2991 break;
2992 }
2993 case T_FLOAT: {
2994 scale = gvn().transform(new ConvI2FNode(scale));
2995 break;
2996 }
2997 case T_DOUBLE: {
2998 scale = gvn().transform(new ConvI2DNode(scale));
2999 break;
3000 }
3001 default: fatal("%s", type2name(elem_bt));
3002 }
3003 scale = gvn().transform(VectorNode::scalar2vector(scale, num_elem, elem_bt));
3004 index = gvn().transform(VectorNode::make(vmul_op, index, scale, vt));
3005 }
3006
3007 // Add "opd" if addition is needed.
3008 if (needs_add) {
3009 index = gvn().transform(VectorNode::make(vadd_op, opd, index, vt));
3010 }
3011 Node* vbox = box_vector(index, vbox_type, elem_bt, num_elem);
3012 set_result(vbox);
3013 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
3014 return true;
3015 }
3016
3017 // public static
3018 // <E,
3019 // M extends VectorMask<E>>
3020 // M indexPartiallyInUpperRange(Class<? extends M> mClass, Class<E> eClass, int length,
3021 // long offset, long limit,
3022 // IndexPartiallyInUpperRangeOperation<E, M> defaultImpl)
3023 bool LibraryCallKit::inline_index_partially_in_upper_range() {
3024 const TypeInstPtr* mask_klass = gvn().type(argument(0))->isa_instptr();
3025 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
3026 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
3027
3028 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr ||
3029 elem_klass == nullptr || elem_klass->const_oop() == nullptr ||
3030 vlen == nullptr || !vlen->is_con()) {
3031 log_if_needed(" ** missing constant: mclass=%s etype=%s vlen=%s",
3032 NodeClassNames[argument(0)->Opcode()],
3033 NodeClassNames[argument(1)->Opcode()],
3034 NodeClassNames[argument(2)->Opcode()]);
3035 return false; // not enough info for intrinsification
3036 }
3037
3038 if (!is_klass_initialized(mask_klass)) {
3039 log_if_needed(" ** klass argument not initialized");
3040 return false;
3041 }
3042
3043 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
3044 if (!elem_type->is_primitive_type()) {
3045 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
3046 return false; // should be primitive type
3047 }
3048
3049 int num_elem = vlen->get_con();
3050 BasicType elem_bt = elem_type->basic_type();
3051
3052 // Check whether the necessary ops are supported by current hardware.
3053 bool supports_mask_gen = arch_supports_vector(Op_VectorMaskGen, num_elem, elem_bt, VecMaskUseStore);
3054 if (!supports_mask_gen) {
3055 if (!arch_supports_vector(Op_VectorLoadConst, num_elem, elem_bt, VecMaskNotUsed) ||
3056 !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed) ||
3057 !arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
3058 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
3059 return false; // not supported
3060 }
3061
3062 // Check whether the scalar cast operation is supported by current hardware.
3063 if (elem_bt != T_LONG) {
3064 int cast_op = is_integral_type(elem_bt) ? Op_ConvL2I
3065 : (elem_bt == T_FLOAT ? Op_ConvL2F : Op_ConvL2D);
3066 if (!Matcher::match_rule_supported(cast_op)) {
3067 log_if_needed(" ** Rejected op (%s) because architecture does not support it",
3068 NodeClassNames[cast_op]);
3069 return false; // not supported
3070 }
3071 }
3072 }
3073
3074 Node* offset = argument(3);
3075 Node* limit = argument(5);
3076 if (offset == nullptr || limit == nullptr) {
3077 log_if_needed(" ** offset or limit argument is null");
3078 return false; // not supported
3079 }
3080
3081 ciKlass* box_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
3082 assert(is_vector_mask(box_klass), "argument(0) should be a mask class");
3083 const TypeInstPtr* box_type = TypeInstPtr::make_exact(TypePtr::NotNull, box_klass);
3084
3085 // We assume "offset > 0 && limit >= offset && limit - offset < num_elem".
3086 // So directly get indexLimit with "indexLimit = limit - offset".
3087 Node* indexLimit = gvn().transform(new SubLNode(limit, offset));
3088 Node* mask = nullptr;
3089 if (supports_mask_gen) {
3090 mask = gvn().transform(VectorMaskGenNode::make(indexLimit, elem_bt, num_elem));
3091 } else {
3092 // Generate the vector mask based on "mask = iota < indexLimit".
3093 // Broadcast "indexLimit" to a vector.
3094 switch (elem_bt) {
3095 case T_BOOLEAN: // fall-through
3096 case T_BYTE: // fall-through
3097 case T_SHORT: // fall-through
3098 case T_CHAR: // fall-through
3099 case T_INT: {
3100 indexLimit = gvn().transform(new ConvL2INode(indexLimit));
3101 break;
3102 }
3103 case T_DOUBLE: {
3104 indexLimit = gvn().transform(new ConvL2DNode(indexLimit));
3105 break;
3106 }
3107 case T_FLOAT: {
3108 indexLimit = gvn().transform(new ConvL2FNode(indexLimit));
3109 break;
3110 }
3111 case T_LONG: {
3112 // no conversion needed
3113 break;
3114 }
3115 default: fatal("%s", type2name(elem_bt));
3116 }
3117 indexLimit = gvn().transform(VectorNode::scalar2vector(indexLimit, num_elem, elem_bt));
3118
3119 // Load the "iota" vector.
3120 const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
3121 Node* iota = gvn().transform(new VectorLoadConstNode(gvn().makecon(TypeInt::ZERO), vt));
3122
3123 // Compute the vector mask with "mask = iota < indexLimit".
3124 ConINode* pred_node = (ConINode*)gvn().makecon(TypeInt::make(BoolTest::lt));
3125 const TypeVect* vmask_type = TypeVect::makemask(elem_bt, num_elem);
3126 mask = gvn().transform(new VectorMaskCmpNode(BoolTest::lt, iota, indexLimit, pred_node, vmask_type));
3127 }
3128 Node* vbox = box_vector(mask, box_type, elem_bt, num_elem);
3129 set_result(vbox);
3130 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
3131 return true;
3132 }
3133
3134 #undef non_product_log_if_needed
3135 #undef log_if_needed