1 /*
2 * Copyright (c) 2020, 2024, 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 "precompiled.hpp"
26 #include "ci/ciSymbols.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "opto/library_call.hpp"
29 #include "opto/runtime.hpp"
30 #include "opto/vectornode.hpp"
31 #include "prims/vectorSupport.hpp"
32 #include "runtime/stubRoutines.hpp"
33
34 #ifdef ASSERT
35 static bool is_vector(ciKlass* klass) {
36 return klass->is_subclass_of(ciEnv::current()->vector_VectorPayload_klass());
37 }
38
39 static bool check_vbox(const TypeInstPtr* vbox_type) {
40 assert(vbox_type->klass_is_exact(), "");
41
42 ciInstanceKlass* ik = vbox_type->instance_klass();
43 assert(is_vector(ik), "not a vector");
44
45 ciField* fd1 = ik->get_field_by_name(ciSymbols::ETYPE_name(), ciSymbols::class_signature(), /* is_static */ true);
46 assert(fd1 != nullptr, "element type info is missing");
47
48 ciConstant val1 = fd1->constant_value();
49 BasicType elem_bt = val1.as_object()->as_instance()->java_mirror_type()->basic_type();
50 assert(is_java_primitive(elem_bt), "element type info is missing");
51
52 ciField* fd2 = ik->get_field_by_name(ciSymbols::VLENGTH_name(), ciSymbols::int_signature(), /* is_static */ true);
53 assert(fd2 != nullptr, "vector length info is missing");
54
55 ciConstant val2 = fd2->constant_value();
56 assert(val2.as_int() > 0, "vector length info is missing");
57
58 return true;
59 }
60 #endif
61
62 #define log_if_needed(...) \
63 if (C->print_intrinsics()) { \
64 tty->print_cr(__VA_ARGS__); \
65 }
66
67 #ifndef PRODUCT
68 #define non_product_log_if_needed(...) log_if_needed(__VA_ARGS__)
69 #else
70 #define non_product_log_if_needed(...)
71 #endif
72
73 static bool is_vector_mask(ciKlass* klass) {
74 return klass->is_subclass_of(ciEnv::current()->vector_VectorMask_klass());
75 }
76
77 static bool is_vector_shuffle(ciKlass* klass) {
78 return klass->is_subclass_of(ciEnv::current()->vector_VectorShuffle_klass());
79 }
80
81 bool LibraryCallKit::arch_supports_vector_rotate(int opc, int num_elem, BasicType elem_bt,
82 VectorMaskUseType mask_use_type, bool has_scalar_args) {
83 bool is_supported = true;
84
85 // has_scalar_args flag is true only for non-constant scalar shift count,
86 // since in this case shift needs to be broadcasted.
87 if (!Matcher::match_rule_supported_vector(opc, num_elem, elem_bt) ||
88 (has_scalar_args && !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed))) {
89 is_supported = false;
90 }
91
92 if (is_supported) {
93 // Check if mask unboxing is supported, this is a two step process which first loads the contents
94 // of boolean array into vector followed by either lane expansion to match the lane size of masked
95 // vector operation or populate the predicate register.
96 if ((mask_use_type & VecMaskUseLoad) != 0) {
97 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, elem_bt) ||
98 !Matcher::match_rule_supported_vector(Op_LoadVector, num_elem, T_BOOLEAN)) {
99 non_product_log_if_needed(" ** Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
100 NodeClassNames[Op_VectorLoadMask], type2name(elem_bt), num_elem);
101 return false;
102 }
103 }
104
105 if ((mask_use_type & VecMaskUsePred) != 0) {
106 if (!Matcher::has_predicated_vectors() ||
107 !Matcher::match_rule_supported_vector_masked(opc, num_elem, elem_bt)) {
108 non_product_log_if_needed("Rejected vector mask predicate using (%s,%s,%d) because architecture does not support it",
109 NodeClassNames[opc], type2name(elem_bt), num_elem);
110 return false;
111 }
112 }
113 }
114
115 int lshiftopc, rshiftopc;
116 switch(elem_bt) {
117 case T_BYTE:
118 lshiftopc = Op_LShiftI;
119 rshiftopc = Op_URShiftB;
120 break;
121 case T_SHORT:
122 lshiftopc = Op_LShiftI;
123 rshiftopc = Op_URShiftS;
124 break;
125 case T_INT:
126 lshiftopc = Op_LShiftI;
127 rshiftopc = Op_URShiftI;
128 break;
129 case T_LONG:
130 lshiftopc = Op_LShiftL;
131 rshiftopc = Op_URShiftL;
132 break;
133 default: fatal("Unexpected type: %s", type2name(elem_bt));
134 }
135 int lshiftvopc = VectorNode::opcode(lshiftopc, elem_bt);
136 int rshiftvopc = VectorNode::opcode(rshiftopc, elem_bt);
137 if (!is_supported &&
138 arch_supports_vector(lshiftvopc, num_elem, elem_bt, VecMaskNotUsed, has_scalar_args) &&
139 arch_supports_vector(rshiftvopc, num_elem, elem_bt, VecMaskNotUsed, has_scalar_args) &&
140 arch_supports_vector(Op_OrV, num_elem, elem_bt, VecMaskNotUsed)) {
141 is_supported = true;
142 }
143 return is_supported;
144 }
145
146 Node* GraphKit::box_vector(Node* vector, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem, bool deoptimize_on_exception) {
147 assert(EnableVectorSupport, "");
148
149 PreserveReexecuteState preexecs(this);
150 jvms()->set_should_reexecute(true);
151
152 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
153 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
154 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true, deoptimize_on_exception);
155 set_i_o(gvn().transform( new ProjNode(alloc, TypeFunc::I_O) ));
156 set_all_memory(gvn().transform( new ProjNode(alloc, TypeFunc::Memory) ));
157 Node* ret = gvn().transform(new ProjNode(alloc, TypeFunc::Parms));
158
159 assert(check_vbox(vbox_type), "");
160 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_type->instance_klass()));
161 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vt);
162 return gvn().transform(vbox);
163 }
164
165 Node* GraphKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem, bool shuffle_to_vector) {
166 assert(EnableVectorSupport, "");
167 const TypeInstPtr* vbox_type_v = gvn().type(v)->isa_instptr();
168 if (vbox_type_v == nullptr || vbox_type->instance_klass() != vbox_type_v->instance_klass()) {
169 return nullptr; // arguments don't agree on vector shapes
170 }
171 if (vbox_type_v->maybe_null()) {
172 return nullptr; // no nulls are allowed
173 }
174 assert(check_vbox(vbox_type), "");
175 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_type->instance_klass()));
176 Node* unbox = gvn().transform(new VectorUnboxNode(C, vt, v, merged_memory(), shuffle_to_vector));
177 return unbox;
178 }
179
180 Node* GraphKit::vector_shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
181 assert(bt == T_INT || bt == T_LONG || bt == T_SHORT || bt == T_BYTE, "byte, short, long and int are supported");
182 juint mask = (type2aelembytes(bt) * BitsPerByte - 1);
183 Node* nmask = gvn().transform(ConNode::make(TypeInt::make(mask)));
184 Node* mcnt = gvn().transform(new AndINode(cnt, nmask));
185 return gvn().transform(VectorNode::shift_count(shift_op, mcnt, num_elem, bt));
186 }
187
188 bool LibraryCallKit::arch_supports_vector(int sopc, int num_elem, BasicType type, VectorMaskUseType mask_use_type, bool has_scalar_args) {
189 // Check that the operation is valid.
190 if (sopc <= 0) {
191 non_product_log_if_needed(" ** Rejected intrinsification because no valid vector op could be extracted");
192 return false;
193 }
194
195 if (VectorNode::is_vector_rotate(sopc)) {
196 if(!arch_supports_vector_rotate(sopc, num_elem, type, mask_use_type, has_scalar_args)) {
197 non_product_log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support variable vector shifts",
198 NodeClassNames[sopc], type2name(type), num_elem);
199 return false;
200 }
201 } else if (VectorNode::is_vector_integral_negate(sopc)) {
202 if (!VectorNode::is_vector_integral_negate_supported(sopc, num_elem, type, false)) {
203 non_product_log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support integral vector negate",
204 NodeClassNames[sopc], type2name(type), num_elem);
205 return false;
206 }
207 } else {
208 // Check that architecture supports this op-size-type combination.
209 if (!Matcher::match_rule_supported_vector(sopc, num_elem, type)) {
210 non_product_log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support it",
211 NodeClassNames[sopc], type2name(type), num_elem);
212 return false;
213 } else {
214 assert(Matcher::match_rule_supported(sopc), "must be supported");
215 }
216 }
217
218 if (num_elem == 1) {
219 if (mask_use_type != VecMaskNotUsed) {
220 non_product_log_if_needed(" ** Rejected vector mask op (%s,%s,%d) because architecture does not support it",
221 NodeClassNames[sopc], type2name(type), num_elem);
222 return false;
223 }
224
225 if (sopc != 0) {
226 if (sopc != Op_LoadVector && sopc != Op_StoreVector) {
227 non_product_log_if_needed(" ** Not a svml call or load/store vector op (%s,%s,%d)",
228 NodeClassNames[sopc], type2name(type), num_elem);
229 return false;
230 }
231 }
232 }
233
234 if (!has_scalar_args && VectorNode::is_vector_shift(sopc) &&
235 Matcher::supports_vector_variable_shifts() == false) {
236 log_if_needed(" ** Rejected vector op (%s,%s,%d) because architecture does not support variable vector shifts",
237 NodeClassNames[sopc], type2name(type), num_elem);
238 return false;
239 }
240
241 // Check if mask unboxing is supported, this is a two step process which first loads the contents
242 // of boolean array into vector followed by either lane expansion to match the lane size of masked
243 // vector operation or populate the predicate register.
244 if ((mask_use_type & VecMaskUseLoad) != 0) {
245 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type) ||
246 !Matcher::match_rule_supported_vector(Op_LoadVector, num_elem, T_BOOLEAN)) {
247 non_product_log_if_needed(" ** Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
248 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
249 return false;
250 }
251 }
252
253 // Check if mask boxing is supported, this is a two step process which first stores the contents
254 // of mask vector / predicate register into a boolean vector followed by vector store operation to
255 // transfer the contents to underlined storage of mask boxes which is a boolean array.
256 if ((mask_use_type & VecMaskUseStore) != 0) {
257 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type) ||
258 !Matcher::match_rule_supported_vector(Op_StoreVector, num_elem, T_BOOLEAN)) {
259 non_product_log_if_needed("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
260 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
261 return false;
262 }
263 }
264
265 if ((mask_use_type & VecMaskUsePred) != 0) {
266 bool is_supported = false;
267 if (Matcher::has_predicated_vectors()) {
268 if (VectorNode::is_vector_integral_negate(sopc)) {
269 is_supported = VectorNode::is_vector_integral_negate_supported(sopc, num_elem, type, true);
270 } else {
271 is_supported = Matcher::match_rule_supported_vector_masked(sopc, num_elem, type);
272 }
273 }
274 is_supported |= Matcher::supports_vector_predicate_op_emulation(sopc, num_elem, type);
275
276 if (!is_supported) {
277 non_product_log_if_needed("Rejected vector mask predicate using (%s,%s,%d) because architecture does not support it",
278 NodeClassNames[sopc], type2name(type), num_elem);
279 return false;
280 }
281 }
282
283 return true;
284 }
285
286 static bool is_klass_initialized(const TypeInstPtr* vec_klass) {
287 if (vec_klass->const_oop() == nullptr) {
288 return false; // uninitialized or some kind of unsafe access
289 }
290 assert(vec_klass->const_oop()->as_instance()->java_lang_Class_klass() != nullptr, "klass instance expected");
291 ciInstanceKlass* klass = vec_klass->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
292 return klass->is_initialized();
293 }
294
295 // public static
296 // <V extends Vector<E>,
297 // M extends VectorMask<E>,
298 // E>
299 // V unaryOp(int oprId, Class<? extends V> vmClass, Class<? extends M> maskClass, Class<E> elementType,
300 // int length, V v, M m,
301 // UnaryOperation<V, M> defaultImpl)
302 //
303 // public static
304 // <V,
305 // M extends VectorMask<E>,
306 // E>
307 // V binaryOp(int oprId, Class<? extends V> vmClass, Class<? extends M> maskClass, Class<E> elementType,
308 // int length, V v1, V v2, M m,
309 // BinaryOperation<V, M> defaultImpl)
310 //
311 // public static
312 // <V extends Vector<E>,
313 // M extends VectorMask<E>,
314 // E>
315 // V ternaryOp(int oprId, Class<? extends V> vmClass, Class<? extends M> maskClass, Class<E> elementType,
316 // int length, V v1, V v2, V v3, M m,
317 // TernaryOperation<V, M> defaultImpl)
318 //
319 bool LibraryCallKit::inline_vector_nary_operation(int n) {
320 const TypeInt* opr = gvn().type(argument(0))->isa_int();
321 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
322 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
323 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
324 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
325
326 if (opr == nullptr || vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
327 !opr->is_con() || vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
328 log_if_needed(" ** missing constant: opr=%s vclass=%s etype=%s vlen=%s",
329 NodeClassNames[argument(0)->Opcode()],
330 NodeClassNames[argument(1)->Opcode()],
331 NodeClassNames[argument(3)->Opcode()],
332 NodeClassNames[argument(4)->Opcode()]);
333 return false; // not enough info for intrinsification
334 }
335
336 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
337 if (!elem_type->is_primitive_type()) {
338 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
339 return false; // should be primitive type
340 }
341 if (!is_klass_initialized(vector_klass)) {
342 log_if_needed(" ** klass argument not initialized");
343 return false;
344 }
345
346 // "argument(n + 5)" should be the mask object. We assume it is "null" when no mask
347 // is used to control this operation.
348 const Type* vmask_type = gvn().type(argument(n + 5));
349 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
350 if (is_masked_op) {
351 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
352 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(2)->Opcode()]);
353 return false; // not enough info for intrinsification
354 }
355
356 if (!is_klass_initialized(mask_klass)) {
357 log_if_needed(" ** mask klass argument not initialized");
358 return false;
359 }
360
361 if (vmask_type->maybe_null()) {
362 log_if_needed(" ** null mask values are not allowed for masked op");
363 return false;
364 }
365 }
366
367 BasicType elem_bt = elem_type->basic_type();
368 int num_elem = vlen->get_con();
369 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
370 int sopc = VectorNode::opcode(opc, elem_bt);
371 if ((opc != Op_CallLeafVector) && (sopc == 0)) {
372 log_if_needed(" ** operation not supported: opc=%s bt=%s", NodeClassNames[opc], type2name(elem_bt));
373 return false; // operation not supported
374 }
375 if (num_elem == 1) {
376 if (opc != Op_CallLeafVector || elem_bt != T_DOUBLE) {
377 log_if_needed(" ** not a svml call: arity=%d opc=%d vlen=%d etype=%s",
378 n, opc, num_elem, type2name(elem_bt));
379 return false;
380 }
381 }
382 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
383 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
384
385 if (is_vector_mask(vbox_klass)) {
386 assert(!is_masked_op, "mask operations do not need mask to control");
387 }
388
389 if (opc == Op_CallLeafVector) {
390 if (!UseVectorStubs) {
391 log_if_needed(" ** vector stubs support is disabled");
392 return false;
393 }
394 if (!Matcher::supports_vector_calling_convention()) {
395 log_if_needed(" ** no vector calling conventions supported");
396 return false;
397 }
398 if (!Matcher::vector_size_supported(elem_bt, num_elem)) {
399 log_if_needed(" ** vector size (vlen=%d, etype=%s) is not supported",
400 num_elem, type2name(elem_bt));
401 return false;
402 }
403 }
404
405 // When using mask, mask use type needs to be VecMaskUseLoad.
406 VectorMaskUseType mask_use_type = is_vector_mask(vbox_klass) ? VecMaskUseAll
407 : is_masked_op ? VecMaskUseLoad : VecMaskNotUsed;
408 if ((sopc != 0) && !arch_supports_vector(sopc, num_elem, elem_bt, mask_use_type)) {
409 log_if_needed(" ** not supported: arity=%d opc=%d vlen=%d etype=%s ismask=%d is_masked_op=%d",
410 n, sopc, num_elem, type2name(elem_bt),
411 is_vector_mask(vbox_klass) ? 1 : 0, is_masked_op ? 1 : 0);
412 return false; // not supported
413 }
414
415 // Return true if current platform has implemented the masked operation with predicate feature.
416 bool use_predicate = is_masked_op && sopc != 0 && arch_supports_vector(sopc, num_elem, elem_bt, VecMaskUsePred);
417 if (is_masked_op && !use_predicate && !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
418 log_if_needed(" ** not supported: arity=%d opc=%d vlen=%d etype=%s ismask=0 is_masked_op=1",
419 n, sopc, num_elem, type2name(elem_bt));
420 return false;
421 }
422
423 Node* opd1 = nullptr; Node* opd2 = nullptr; Node* opd3 = nullptr;
424 switch (n) {
425 case 3: {
426 opd3 = unbox_vector(argument(7), vbox_type, elem_bt, num_elem);
427 if (opd3 == nullptr) {
428 log_if_needed(" ** unbox failed v3=%s",
429 NodeClassNames[argument(7)->Opcode()]);
430 return false;
431 }
432 // fall-through
433 }
434 case 2: {
435 opd2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
436 if (opd2 == nullptr) {
437 log_if_needed(" ** unbox failed v2=%s",
438 NodeClassNames[argument(6)->Opcode()]);
439 return false;
440 }
441 // fall-through
442 }
443 case 1: {
444 opd1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
445 if (opd1 == nullptr) {
446 log_if_needed(" ** unbox failed v1=%s",
447 NodeClassNames[argument(5)->Opcode()]);
448 return false;
449 }
450 break;
451 }
452 default: fatal("unsupported arity: %d", n);
453 }
454
455 Node* mask = nullptr;
456 if (is_masked_op) {
457 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
458 assert(is_vector_mask(mbox_klass), "argument(2) should be a mask class");
459 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
460 mask = unbox_vector(argument(n + 5), mbox_type, elem_bt, num_elem);
461 if (mask == nullptr) {
462 log_if_needed(" ** unbox failed mask=%s",
463 NodeClassNames[argument(n + 5)->Opcode()]);
464 return false;
465 }
466 }
467
468 Node* operation = nullptr;
469 if (opc == Op_CallLeafVector) {
470 assert(UseVectorStubs, "sanity");
471 operation = gen_call_to_vector_math(opr->get_con(), elem_bt, num_elem, opd1, opd2);
472 if (operation == nullptr) {
473 log_if_needed(" ** Vector math call failed for %s_%s_%d",
474 (elem_bt == T_FLOAT) ? "float" : "double",
475 VectorSupport::mathname[opr->get_con() - VectorSupport::VECTOR_OP_MATH_START],
476 num_elem * type2aelembytes(elem_bt));
477 return false;
478 }
479 } else {
480 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, is_vector_mask(vbox_klass));
481 switch (n) {
482 case 1:
483 case 2: {
484 operation = VectorNode::make(sopc, opd1, opd2, vt, is_vector_mask(vbox_klass), VectorNode::is_shift_opcode(opc));
485 break;
486 }
487 case 3: {
488 operation = VectorNode::make(sopc, opd1, opd2, opd3, vt);
489 break;
490 }
491 default: fatal("unsupported arity: %d", n);
492 }
493 }
494
495 if (is_masked_op && mask != nullptr) {
496 if (use_predicate) {
497 operation->add_req(mask);
498 operation->add_flag(Node::Flag_is_predicated_vector);
499 } else {
500 operation->add_flag(Node::Flag_is_predicated_using_blend);
501 operation = gvn().transform(operation);
502 operation = new VectorBlendNode(opd1, operation, mask);
503 }
504 }
505 operation = gvn().transform(operation);
506
507 // Wrap it up in VectorBox to keep object type information.
508 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
509 set_result(vbox);
510 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
511 return true;
512 }
513
514 // Following routine generates IR corresponding to AbstractShuffle::partiallyWrapIndex method,
515 // which partially wraps index by modulo VEC_LENGTH and generates a negative index value if original
516 // index is out of valid index range [0, VEC_LENGTH)
517 //
518 // wrapped_index = (VEC_LENGTH - 1) & index
519 // if (index u> VEC_LENGTH) {
520 // wrapped_index -= VEC_LENGTH;
521 //
522 // Note: Unsigned greater than comparison treat both <0 and >VEC_LENGTH indices as out-of-bound
523 // indexes.
524 Node* LibraryCallKit::partially_wrap_indexes(Node* index_vec, int num_elem, BasicType elem_bt) {
525 assert(elem_bt == T_BYTE, "Shuffles use byte array based backing storage.");
526 const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
527 const Type* type_bt = Type::get_const_basic_type(elem_bt);
528
529 Node* mod_mask = gvn().makecon(TypeInt::make(num_elem-1));
530 Node* bcast_mod_mask = gvn().transform(VectorNode::scalar2vector(mod_mask, num_elem, type_bt));
531
532 BoolTest::mask pred = BoolTest::ugt;
533 ConINode* pred_node = (ConINode*)gvn().makecon(TypeInt::make(pred));
534 Node* lane_cnt = gvn().makecon(TypeInt::make(num_elem));
535 Node* bcast_lane_cnt = gvn().transform(VectorNode::scalar2vector(lane_cnt, num_elem, type_bt));
536 const TypeVect* vmask_type = TypeVect::makemask(type_bt, num_elem);
537 Node* mask = gvn().transform(new VectorMaskCmpNode(pred, bcast_lane_cnt, index_vec, pred_node, vmask_type));
538
539 // Make the indices greater than lane count as -ve values to match the java side implementation.
540 index_vec = gvn().transform(VectorNode::make(Op_AndV, index_vec, bcast_mod_mask, vt));
541 Node* biased_val = gvn().transform(VectorNode::make(Op_SubVB, index_vec, bcast_lane_cnt, vt));
542 return gvn().transform(new VectorBlendNode(biased_val, index_vec, mask));
543 }
544
545 // <Sh extends VectorShuffle<E>, E>
546 // Sh ShuffleIota(Class<?> E, Class<?> shuffleClass, Vector.Species<E> s, int length,
547 // int start, int step, int wrap, ShuffleIotaOperation<Sh, E> defaultImpl)
548 bool LibraryCallKit::inline_vector_shuffle_iota() {
549 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->isa_instptr();
550 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
551 const TypeInt* start_val = gvn().type(argument(4))->isa_int();
552 const TypeInt* step_val = gvn().type(argument(5))->isa_int();
553 const TypeInt* wrap = gvn().type(argument(6))->isa_int();
554
555 if (shuffle_klass == nullptr || shuffle_klass->const_oop() == nullptr ||
556 vlen == nullptr || !vlen->is_con() || start_val == nullptr || step_val == nullptr ||
557 wrap == nullptr || !wrap->is_con()) {
558 return false; // not enough info for intrinsification
559 }
560
561 if (!is_klass_initialized(shuffle_klass)) {
562 log_if_needed(" ** klass argument not initialized");
563 return false;
564 }
565
566 int do_wrap = wrap->get_con();
567 int num_elem = vlen->get_con();
568 BasicType elem_bt = T_BYTE;
569
570 bool effective_indices_in_range = false;
571 if (start_val->is_con() && step_val->is_con()) {
572 int effective_min_index = start_val->get_con();
573 int effective_max_index = start_val->get_con() + step_val->get_con() * (num_elem - 1);
574 effective_indices_in_range = effective_max_index >= effective_min_index && effective_min_index >= -128 && effective_max_index <= 127;
575 }
576
577 if (!do_wrap && !effective_indices_in_range) {
578 // Disable instrinsification for unwrapped shuffle iota if start/step
579 // values are non-constant OR if intermediate result overflows byte value range.
580 return false;
581 }
582
583 if (!arch_supports_vector(Op_AddVB, num_elem, elem_bt, VecMaskNotUsed) ||
584 !arch_supports_vector(Op_AndV, num_elem, elem_bt, VecMaskNotUsed) ||
585 !arch_supports_vector(Op_VectorLoadConst, num_elem, elem_bt, VecMaskNotUsed) ||
586 !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed)) {
587 return false;
588 }
589
590 if (!do_wrap &&
591 (!arch_supports_vector(Op_SubVB, num_elem, elem_bt, VecMaskNotUsed) ||
592 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskNotUsed) ||
593 !arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskNotUsed))) {
594 return false;
595 }
596
597 bool step_multiply = !step_val->is_con() || !is_power_of_2(step_val->get_con());
598 if ((step_multiply && !arch_supports_vector(Op_MulVB, num_elem, elem_bt, VecMaskNotUsed)) ||
599 (!step_multiply && !arch_supports_vector(Op_LShiftVB, num_elem, elem_bt, VecMaskNotUsed))) {
600 return false;
601 }
602
603 const Type * type_bt = Type::get_const_basic_type(elem_bt);
604 const TypeVect * vt = TypeVect::make(type_bt, num_elem);
605
606 Node* res = gvn().transform(new VectorLoadConstNode(gvn().makecon(TypeInt::ZERO), vt));
607
608 Node* start = argument(4);
609 Node* step = argument(5);
610
611 if (step_multiply) {
612 Node* bcast_step = gvn().transform(VectorNode::scalar2vector(step, num_elem, type_bt));
613 res = gvn().transform(VectorNode::make(Op_MulVB, res, bcast_step, vt));
614 } else if (step_val->get_con() > 1) {
615 Node* cnt = gvn().makecon(TypeInt::make(log2i_exact(step_val->get_con())));
616 Node* shift_cnt = vector_shift_count(cnt, Op_LShiftI, elem_bt, num_elem);
617 res = gvn().transform(VectorNode::make(Op_LShiftVB, res, shift_cnt, vt));
618 }
619
620 if (!start_val->is_con() || start_val->get_con() != 0) {
621 Node* bcast_start = gvn().transform(VectorNode::scalar2vector(start, num_elem, type_bt));
622 res = gvn().transform(VectorNode::make(Op_AddVB, res, bcast_start, vt));
623 }
624
625 Node * mod_val = gvn().makecon(TypeInt::make(num_elem-1));
626 Node * bcast_mod = gvn().transform(VectorNode::scalar2vector(mod_val, num_elem, type_bt));
627
628 if (do_wrap) {
629 // Wrap the indices greater than lane count.
630 res = gvn().transform(VectorNode::make(Op_AndV, res, bcast_mod, vt));
631 } else {
632 res = partially_wrap_indexes(res, num_elem, elem_bt);
633 }
634
635 ciKlass* sbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
636 const TypeInstPtr* shuffle_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, sbox_klass);
637
638 // Wrap it up in VectorBox to keep object type information.
639 res = box_vector(res, shuffle_box_type, elem_bt, num_elem);
640 set_result(res);
641 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
642 return true;
643 }
644
645 // <E, M>
646 // long maskReductionCoerced(int oper, Class<? extends M> maskClass, Class<?> elemClass,
647 // int length, M m, VectorMaskOp<M> defaultImpl)
648 bool LibraryCallKit::inline_vector_mask_operation() {
649 const TypeInt* oper = gvn().type(argument(0))->isa_int();
650 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
651 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
652 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
653 Node* mask = argument(4);
654
655 if (mask_klass == nullptr || elem_klass == nullptr || mask->is_top() || vlen == nullptr) {
656 return false; // dead code
657 }
658
659 if (!is_klass_initialized(mask_klass)) {
660 log_if_needed(" ** klass argument not initialized");
661 return false;
662 }
663
664 int num_elem = vlen->get_con();
665 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
666 BasicType elem_bt = elem_type->basic_type();
667
668 int mopc = VectorSupport::vop2ideal(oper->get_con(), elem_bt);
669 if (!arch_supports_vector(mopc, num_elem, elem_bt, VecMaskUseLoad)) {
670 log_if_needed(" ** not supported: arity=1 op=cast#%d/3 vlen2=%d etype2=%s",
671 mopc, num_elem, type2name(elem_bt));
672 return false; // not supported
673 }
674
675 const Type* elem_ty = Type::get_const_basic_type(elem_bt);
676 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
677 const TypeInstPtr* mask_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
678 Node* mask_vec = unbox_vector(mask, mask_box_type, elem_bt, num_elem, true);
679 if (mask_vec == nullptr) {
680 log_if_needed(" ** unbox failed mask=%s",
681 NodeClassNames[argument(4)->Opcode()]);
682 return false;
683 }
684
685 if (mask_vec->bottom_type()->isa_vectmask() == nullptr) {
686 mask_vec = gvn().transform(VectorStoreMaskNode::make(gvn(), mask_vec, elem_bt, num_elem));
687 }
688 const Type* maskoper_ty = mopc == Op_VectorMaskToLong ? (const Type*)TypeLong::LONG : (const Type*)TypeInt::INT;
689 Node* maskoper = gvn().transform(VectorMaskOpNode::make(mask_vec, maskoper_ty, mopc));
690 if (mopc != Op_VectorMaskToLong) {
691 maskoper = ConvI2L(maskoper);
692 }
693 set_result(maskoper);
694
695 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
696 return true;
697 }
698
699 // public static
700 // <V,
701 // Sh extends VectorShuffle<E>,
702 // E>
703 // V shuffleToVector(Class<? extends Vector<E>> vclass, Class<E> elementType,
704 // Class<? extends Sh> shuffleClass, Sh s, int length,
705 // ShuffleToVectorOperation<V, Sh, E> defaultImpl)
706 bool LibraryCallKit::inline_vector_shuffle_to_vector() {
707 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
708 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
709 const TypeInstPtr* shuffle_klass = gvn().type(argument(2))->isa_instptr();
710 Node* shuffle = argument(3);
711 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
712
713 if (vector_klass == nullptr || elem_klass == nullptr || shuffle_klass == nullptr || shuffle->is_top() || vlen == nullptr) {
714 return false; // dead code
715 }
716 if (!vlen->is_con() || vector_klass->const_oop() == nullptr || shuffle_klass->const_oop() == nullptr) {
717 return false; // not enough info for intrinsification
718 }
719 if (!is_klass_initialized(shuffle_klass) || !is_klass_initialized(vector_klass) ) {
720 log_if_needed(" ** klass argument not initialized");
721 return false;
722 }
723
724 int num_elem = vlen->get_con();
725 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
726 BasicType elem_bt = elem_type->basic_type();
727
728 if (num_elem < 4) {
729 return false;
730 }
731
732 int cast_vopc = VectorCastNode::opcode(-1, T_BYTE); // from shuffle of type T_BYTE
733 // Make sure that cast is implemented to particular type/size combination.
734 if (!arch_supports_vector(cast_vopc, num_elem, elem_bt, VecMaskNotUsed)) {
735 log_if_needed(" ** not supported: arity=1 op=cast#%d/3 vlen2=%d etype2=%s",
736 cast_vopc, num_elem, type2name(elem_bt));
737 return false;
738 }
739
740 ciKlass* sbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
741 const TypeInstPtr* shuffle_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, sbox_klass);
742
743 // Unbox shuffle with true flag to indicate its load shuffle to vector
744 // shuffle is a byte array
745 Node* shuffle_vec = unbox_vector(shuffle, shuffle_box_type, T_BYTE, num_elem, true);
746
747 // cast byte to target element type
748 shuffle_vec = gvn().transform(VectorCastNode::make(cast_vopc, shuffle_vec, elem_bt, num_elem));
749
750 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
751 const TypeInstPtr* vec_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
752
753 // Box vector
754 Node* res = box_vector(shuffle_vec, vec_box_type, elem_bt, num_elem);
755 set_result(res);
756 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
757 return true;
758 }
759
760 // public static
761 // <E,
762 // SH extends VectorShuffle<E>>
763 // SH wrapShuffleIndexes(Class<E> eClass, Class<? extends SH> shClass, SH sh, int length,
764 // ShuffleWrapIndexesOperation<SH> defaultImpl)
765 bool LibraryCallKit::inline_vector_wrap_shuffle_indexes() {
766 const TypeInstPtr* elem_klass = gvn().type(argument(0))->isa_instptr();
767 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->isa_instptr();
768 Node* shuffle = argument(2);
769 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
770
771 if (elem_klass == nullptr || shuffle_klass == nullptr || shuffle->is_top() || vlen == nullptr ||
772 !vlen->is_con() || shuffle_klass->const_oop() == nullptr) {
773 // not enough info for intrinsification
774 return false;
775 }
776
777 if (!is_klass_initialized(shuffle_klass)) {
778 log_if_needed(" ** klass argument not initialized");
779 return false;
780 }
781
782 int num_elem = vlen->get_con();
783 if ((num_elem < 4) || !is_power_of_2(num_elem)) {
784 log_if_needed(" ** vlen < 4 or not power of two=%d", num_elem);
785 return false;
786 }
787
788 // Shuffles use byte array based backing storage
789 BasicType shuffle_bt = T_BYTE;
790 if (!arch_supports_vector(Op_AndV, num_elem, shuffle_bt, VecMaskNotUsed) ||
791 !arch_supports_vector(Op_Replicate, num_elem, shuffle_bt, VecMaskNotUsed)) {
792 log_if_needed(" ** not supported: op=wrapShuffleIndexes vlen=%d etype=%s",
793 num_elem, type2name(shuffle_bt));
794 return false;
795 }
796
797 ciKlass* sbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
798 const TypeInstPtr* shuffle_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, sbox_klass);
799
800 // Unbox shuffle with true flag to indicate its load shuffle to vector
801 // shuffle is a byte array
802 Node* shuffle_vec = unbox_vector(shuffle, shuffle_box_type, shuffle_bt, num_elem, true);
803
804 const TypeVect* vt = TypeVect::make(shuffle_bt, num_elem);
805 const Type* shuffle_type_bt = Type::get_const_basic_type(shuffle_bt);
806 Node* mod_mask = gvn().makecon(TypeInt::make(num_elem-1));
807 Node* bcast_mod_mask = gvn().transform(VectorNode::scalar2vector(mod_mask, num_elem, shuffle_type_bt));
808 // Wrap the indices greater than lane count.
809 Node* res = gvn().transform(VectorNode::make(Op_AndV, shuffle_vec, bcast_mod_mask, vt));
810
811 // Wrap it up in VectorBox to keep object type information.
812 res = box_vector(res, shuffle_box_type, shuffle_bt, num_elem);
813 set_result(res);
814 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(shuffle_bt))));
815 return true;
816 }
817
818 // public static
819 // <M,
820 // S extends VectorSpecies<E>,
821 // E>
822 // M fromBitsCoerced(Class<? extends M> vmClass, Class<E> elementType, int length,
823 // long bits, int mode, S s,
824 // BroadcastOperation<M, E, S> defaultImpl)
825 bool LibraryCallKit::inline_vector_frombits_coerced() {
826 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
827 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
828 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
829 const TypeLong* bits_type = gvn().type(argument(3))->isa_long();
830 // Mode argument determines the mode of operation it can take following values:-
831 // MODE_BROADCAST for vector Vector.broadcast and VectorMask.maskAll operations.
832 // MODE_BITS_COERCED_LONG_TO_MASK for VectorMask.fromLong operation.
833 const TypeInt* mode = gvn().type(argument(5))->isa_int();
834
835 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr || mode == nullptr ||
836 bits_type == nullptr || vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr ||
837 !vlen->is_con() || !mode->is_con()) {
838 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s bitwise=%s",
839 NodeClassNames[argument(0)->Opcode()],
840 NodeClassNames[argument(1)->Opcode()],
841 NodeClassNames[argument(2)->Opcode()],
842 NodeClassNames[argument(5)->Opcode()]);
843 return false; // not enough info for intrinsification
844 }
845
846 if (!is_klass_initialized(vector_klass)) {
847 log_if_needed(" ** klass argument not initialized");
848 return false;
849 }
850 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
851 if (!elem_type->is_primitive_type()) {
852 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
853 return false; // should be primitive type
854 }
855 BasicType elem_bt = elem_type->basic_type();
856 int num_elem = vlen->get_con();
857 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
858 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
859
860 bool is_mask = is_vector_mask(vbox_klass);
861 int bcast_mode = mode->get_con();
862 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_mask ? VecMaskUseAll : VecMaskNotUsed);
863 int opc = bcast_mode == VectorSupport::MODE_BITS_COERCED_LONG_TO_MASK ? Op_VectorLongToMask : Op_Replicate;
864
865 if (!arch_supports_vector(opc, num_elem, elem_bt, checkFlags, true /*has_scalar_args*/)) {
866 log_if_needed(" ** not supported: arity=0 op=broadcast vlen=%d etype=%s ismask=%d bcast_mode=%d",
867 num_elem, type2name(elem_bt),
868 is_mask ? 1 : 0,
869 bcast_mode);
870 return false; // not supported
871 }
872
873 Node* broadcast = nullptr;
874 Node* bits = argument(3);
875 Node* elem = bits;
876
877 if (opc == Op_VectorLongToMask) {
878 const TypeVect* vt = TypeVect::makemask(elem_bt, num_elem);
879 if (vt->isa_vectmask()) {
880 broadcast = gvn().transform(new VectorLongToMaskNode(elem, vt));
881 } else {
882 const TypeVect* mvt = TypeVect::make(T_BOOLEAN, num_elem);
883 broadcast = gvn().transform(new VectorLongToMaskNode(elem, mvt));
884 broadcast = gvn().transform(new VectorLoadMaskNode(broadcast, vt));
885 }
886 } else {
887 switch (elem_bt) {
888 case T_BOOLEAN: // fall-through
889 case T_BYTE: // fall-through
890 case T_SHORT: // fall-through
891 case T_CHAR: // fall-through
892 case T_INT: {
893 elem = gvn().transform(new ConvL2INode(bits));
894 break;
895 }
896 case T_DOUBLE: {
897 elem = gvn().transform(new MoveL2DNode(bits));
898 break;
899 }
900 case T_FLOAT: {
901 bits = gvn().transform(new ConvL2INode(bits));
902 elem = gvn().transform(new MoveI2FNode(bits));
903 break;
904 }
905 case T_LONG: {
906 // no conversion needed
907 break;
908 }
909 default: fatal("%s", type2name(elem_bt));
910 }
911 broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt), is_mask);
912 broadcast = gvn().transform(broadcast);
913 }
914
915 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
916 set_result(box);
917 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
918 return true;
919 }
920
921 static bool elem_consistent_with_arr(BasicType elem_bt, const TypeAryPtr* arr_type, bool mismatched_ms) {
922 assert(arr_type != nullptr, "unexpected");
923 BasicType arr_elem_bt = arr_type->elem()->array_element_basic_type();
924 if (elem_bt == arr_elem_bt) {
925 return true;
926 } else if (elem_bt == T_SHORT && arr_elem_bt == T_CHAR) {
927 // Load/store of short vector from/to char[] is supported
928 return true;
929 } else if (elem_bt == T_BYTE && arr_elem_bt == T_BOOLEAN) {
930 // Load/store of byte vector from/to boolean[] is supported
931 return true;
932 } else {
933 return mismatched_ms;
934 }
935 }
936
937 // public static
938 // <C,
939 // VM extends VectorPayload,
940 // E,
941 // S extends VectorSpecies<E>>
942 // VM load(Class<? extends VM> vmClass, Class<E> eClass,
943 // int length,
944 // Object base, long offset, // Unsafe addressing
945 // boolean fromSegment,
946 // C container, long index, S s, // Arguments for default implementation
947 // LoadOperation<C, VM, S> defaultImpl) {
948 // public static
949 // <C,
950 // V extends VectorPayload>
951 // void store(Class<?> vClass, Class<?> eClass,
952 // int length,
953 // Object base, long offset, // Unsafe addressing
954 // boolean fromSegment,
955 // V v, C container, long index, // Arguments for default implementation
956 // StoreVectorOperation<C, V> defaultImpl) {
957 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
958 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
959 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
960 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
961 const TypeInt* from_ms = gvn().type(argument(6))->isa_int();
962
963 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr || !from_ms->is_con() ||
964 vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
965 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s from_ms=%s",
966 NodeClassNames[argument(0)->Opcode()],
967 NodeClassNames[argument(1)->Opcode()],
968 NodeClassNames[argument(2)->Opcode()],
969 NodeClassNames[argument(6)->Opcode()]);
970 return false; // not enough info for intrinsification
971 }
972 if (!is_klass_initialized(vector_klass)) {
973 log_if_needed(" ** klass argument not initialized");
974 return false;
975 }
976
977 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
978 if (!elem_type->is_primitive_type()) {
979 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
980 return false; // should be primitive type
981 }
982 BasicType elem_bt = elem_type->basic_type();
983 int num_elem = vlen->get_con();
984
985 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
986 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
987 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s ismask=no",
988 is_store, is_store ? "store" : "load",
989 num_elem, type2name(elem_bt));
990 return false; // not supported
991 }
992
993 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
994 bool is_mask = is_vector_mask(vbox_klass);
995
996 Node* base = argument(3);
997 Node* offset = ConvL2X(argument(4));
998
999 // Save state and restore on bailout
1000 uint old_sp = sp();
1001 SafePointNode* old_map = clone_map();
1002
1003 Node* addr = make_unsafe_address(base, offset, (is_mask ? T_BOOLEAN : elem_bt), true);
1004
1005 // The memory barrier checks are based on ones for unsafe access.
1006 // This is not 1-1 implementation.
1007 const Type *const base_type = gvn().type(base);
1008
1009 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
1010 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
1011
1012 const bool in_native = TypePtr::NULL_PTR == base_type; // base always null
1013 const bool in_heap = !TypePtr::NULL_PTR->higher_equal(base_type); // base never null
1014
1015 const bool is_mixed_access = !in_heap && !in_native;
1016
1017 const bool is_mismatched_access = in_heap && (addr_type->isa_aryptr() == nullptr);
1018
1019 const bool needs_cpu_membar = is_mixed_access || is_mismatched_access;
1020
1021 // For non-masked mismatched memory segment vector read/write accesses, intrinsification can continue
1022 // with unknown backing storage type and compiler can skip inserting explicit reinterpretation IR after
1023 // loading from or before storing to backing storage which is mandatory for semantic correctness of
1024 // big-endian memory layout.
1025 bool mismatched_ms = LITTLE_ENDIAN_ONLY(false)
1026 BIG_ENDIAN_ONLY(from_ms->get_con() && !is_mask && arr_type != nullptr &&
1027 arr_type->elem()->array_element_basic_type() != elem_bt);
1028 BasicType mem_elem_bt = mismatched_ms ? arr_type->elem()->array_element_basic_type() : elem_bt;
1029 if (!is_java_primitive(mem_elem_bt)) {
1030 log_if_needed(" ** non-primitive array element type");
1031 return false;
1032 }
1033 int mem_num_elem = mismatched_ms ? (num_elem * type2aelembytes(elem_bt)) / type2aelembytes(mem_elem_bt) : num_elem;
1034 if (arr_type != nullptr && !is_mask && !elem_consistent_with_arr(elem_bt, arr_type, mismatched_ms)) {
1035 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s atype=%s ismask=no",
1036 is_store, is_store ? "store" : "load",
1037 num_elem, type2name(elem_bt), type2name(arr_type->elem()->array_element_basic_type()));
1038 set_map(old_map);
1039 set_sp(old_sp);
1040 return false;
1041 }
1042
1043 // In case of mismatched memory segment accesses, we need to double check that the source type memory operations are supported by backend.
1044 if (mismatched_ms) {
1045 if (is_store) {
1046 if (!arch_supports_vector(Op_StoreVector, num_elem, elem_bt, VecMaskNotUsed)
1047 || !arch_supports_vector(Op_VectorReinterpret, mem_num_elem, mem_elem_bt, VecMaskNotUsed)) {
1048 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d*8 etype=%s/8 ismask=no",
1049 is_store, "store",
1050 num_elem, type2name(elem_bt));
1051 set_map(old_map);
1052 set_sp(old_sp);
1053 return false; // not supported
1054 }
1055 } else {
1056 if (!arch_supports_vector(Op_LoadVector, mem_num_elem, mem_elem_bt, VecMaskNotUsed)
1057 || !arch_supports_vector(Op_VectorReinterpret, num_elem, elem_bt, VecMaskNotUsed)) {
1058 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d*8 etype=%s/8 ismask=no",
1059 is_store, "load",
1060 mem_num_elem, type2name(mem_elem_bt));
1061 set_map(old_map);
1062 set_sp(old_sp);
1063 return false; // not supported
1064 }
1065 }
1066 }
1067 if (is_mask) {
1068 if (!is_store) {
1069 if (!arch_supports_vector(Op_LoadVector, num_elem, elem_bt, VecMaskUseLoad)) {
1070 set_map(old_map);
1071 set_sp(old_sp);
1072 return false; // not supported
1073 }
1074 } else {
1075 if (!arch_supports_vector(Op_StoreVector, num_elem, elem_bt, VecMaskUseStore)) {
1076 set_map(old_map);
1077 set_sp(old_sp);
1078 return false; // not supported
1079 }
1080 }
1081 }
1082
1083 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1084
1085 if (needs_cpu_membar) {
1086 insert_mem_bar(Op_MemBarCPUOrder);
1087 }
1088
1089 if (is_store) {
1090 Node* val = unbox_vector(argument(7), vbox_type, elem_bt, num_elem);
1091 if (val == nullptr) {
1092 set_map(old_map);
1093 set_sp(old_sp);
1094 return false; // operand unboxing failed
1095 }
1096 set_all_memory(reset_memory());
1097
1098 // In case the store needs to happen to byte array, reinterpret the incoming vector to byte vector.
1099 int store_num_elem = num_elem;
1100 if (mismatched_ms) {
1101 store_num_elem = mem_num_elem;
1102 const TypeVect* to_vect_type = TypeVect::make(mem_elem_bt, store_num_elem);
1103 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type));
1104 }
1105 if (is_mask) {
1106 val = gvn().transform(VectorStoreMaskNode::make(gvn(), val, elem_bt, num_elem));
1107 }
1108 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(addr), addr, addr_type, val, store_num_elem));
1109 set_memory(vstore, addr_type);
1110 } else {
1111 // When using byte array, we need to load as byte then reinterpret the value. Otherwise, do a simple vector load.
1112 Node* vload = nullptr;
1113 if (mismatched_ms) {
1114 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, mem_num_elem, mem_elem_bt));
1115 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
1116 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type));
1117 } else {
1118 // Special handle for masks
1119 if (is_mask) {
1120 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, T_BOOLEAN));
1121 vload = gvn().transform(new VectorLoadMaskNode(vload, TypeVect::makemask(elem_bt, num_elem)));
1122 } else {
1123 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, elem_bt));
1124 }
1125 }
1126 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
1127 set_result(box);
1128 }
1129
1130 destruct_map_clone(old_map);
1131
1132 if (needs_cpu_membar) {
1133 insert_mem_bar(Op_MemBarCPUOrder);
1134 }
1135
1136 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1137 return true;
1138 }
1139
1140 // public static
1141 // <C,
1142 // V extends Vector<?>,
1143 // E,
1144 // S extends VectorSpecies<E>,
1145 // M extends VectorMask<E>>
1146 // V loadMasked(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass,
1147 // int length, Object base, long offset, // Unsafe addressing
1148 // boolean fromSegment,
1149 // M m, int offsetInRange,
1150 // C container, long index, S s, // Arguments for default implementation
1151 // LoadVectorMaskedOperation<C, V, S, M> defaultImpl) {
1152 // public static
1153 // <C,
1154 // V extends Vector<E>,
1155 // M extends VectorMask<E>,
1156 // E>
1157 // void storeMasked(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass,
1158 // int length,
1159 // Object base, long offset, // Unsafe addressing
1160 // boolean fromSegment,
1161 // V v, M m, C container, long index, // Arguments for default implementation
1162 // StoreVectorMaskedOperation<C, V, M> defaultImpl) {
1163
1164 bool LibraryCallKit::inline_vector_mem_masked_operation(bool is_store) {
1165 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1166 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
1167 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1168 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1169 const TypeInt* from_ms = gvn().type(argument(7))->isa_int();
1170
1171 if (vector_klass == nullptr || mask_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
1172 vector_klass->const_oop() == nullptr || mask_klass->const_oop() == nullptr || from_ms == nullptr ||
1173 elem_klass->const_oop() == nullptr || !vlen->is_con() || !from_ms->is_con()) {
1174 log_if_needed(" ** missing constant: vclass=%s mclass=%s etype=%s vlen=%s from_ms=%s",
1175 NodeClassNames[argument(0)->Opcode()],
1176 NodeClassNames[argument(1)->Opcode()],
1177 NodeClassNames[argument(2)->Opcode()],
1178 NodeClassNames[argument(3)->Opcode()],
1179 NodeClassNames[argument(7)->Opcode()]);
1180 return false; // not enough info for intrinsification
1181 }
1182 if (!is_klass_initialized(vector_klass)) {
1183 log_if_needed(" ** klass argument not initialized");
1184 return false;
1185 }
1186
1187 if (!is_klass_initialized(mask_klass)) {
1188 log_if_needed(" ** mask klass argument not initialized");
1189 return false;
1190 }
1191
1192 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1193 if (!elem_type->is_primitive_type()) {
1194 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1195 return false; // should be primitive type
1196 }
1197
1198 BasicType elem_bt = elem_type->basic_type();
1199 int num_elem = vlen->get_con();
1200
1201 Node* base = argument(4);
1202 Node* offset = ConvL2X(argument(5));
1203
1204 // Save state and restore on bailout
1205 uint old_sp = sp();
1206 SafePointNode* old_map = clone_map();
1207
1208 Node* addr = make_unsafe_address(base, offset, elem_bt, true);
1209 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
1210 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
1211
1212 bool mismatched_ms = from_ms->get_con() && arr_type != nullptr && arr_type->elem()->array_element_basic_type() != elem_bt;
1213 BIG_ENDIAN_ONLY(if (mismatched_ms) return false;)
1214 // If there is no consistency between array and vector element types, it must be special byte array case
1215 if (arr_type != nullptr && !elem_consistent_with_arr(elem_bt, arr_type, mismatched_ms)) {
1216 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s atype=%s",
1217 is_store, is_store ? "storeMasked" : "loadMasked",
1218 num_elem, type2name(elem_bt), type2name(arr_type->elem()->array_element_basic_type()));
1219 set_map(old_map);
1220 set_sp(old_sp);
1221 return false;
1222 }
1223
1224 int mem_num_elem = mismatched_ms ? num_elem * type2aelembytes(elem_bt) : num_elem;
1225 BasicType mem_elem_bt = mismatched_ms ? T_BYTE : elem_bt;
1226 bool supports_predicate = arch_supports_vector(is_store ? Op_StoreVectorMasked : Op_LoadVectorMasked,
1227 mem_num_elem, mem_elem_bt, VecMaskUseLoad);
1228
1229 // If current arch does not support the predicated operations, we have to bail
1230 // out when current case uses the predicate feature.
1231 if (!supports_predicate) {
1232 bool needs_predicate = false;
1233 if (is_store) {
1234 // Masked vector store always uses the predicated store.
1235 needs_predicate = true;
1236 } else {
1237 // Masked vector load with IOOBE always uses the predicated load.
1238 const TypeInt* offset_in_range = gvn().type(argument(9))->isa_int();
1239 if (!offset_in_range->is_con()) {
1240 log_if_needed(" ** missing constant: offsetInRange=%s",
1241 NodeClassNames[argument(8)->Opcode()]);
1242 set_map(old_map);
1243 set_sp(old_sp);
1244 return false;
1245 }
1246 needs_predicate = (offset_in_range->get_con() == 0);
1247 }
1248
1249 if (needs_predicate) {
1250 log_if_needed(" ** not supported: op=%s vlen=%d etype=%s mismatched_ms=%d",
1251 is_store ? "storeMasked" : "loadMasked",
1252 num_elem, type2name(elem_bt), mismatched_ms ? 1 : 0);
1253 set_map(old_map);
1254 set_sp(old_sp);
1255 return false;
1256 }
1257 }
1258
1259 // This only happens for masked vector load. If predicate is not supported, then check whether
1260 // the normal vector load and blend operations are supported by backend.
1261 if (!supports_predicate && (!arch_supports_vector(Op_LoadVector, mem_num_elem, mem_elem_bt, VecMaskNotUsed) ||
1262 !arch_supports_vector(Op_VectorBlend, mem_num_elem, mem_elem_bt, VecMaskUseLoad))) {
1263 log_if_needed(" ** not supported: op=loadMasked vlen=%d etype=%s mismatched_ms=%d",
1264 num_elem, type2name(elem_bt), mismatched_ms ? 1 : 0);
1265 set_map(old_map);
1266 set_sp(old_sp);
1267 return false;
1268 }
1269
1270 // Since we are using byte array, we need to double check that the vector reinterpret operation
1271 // with byte type is supported by backend.
1272 if (mismatched_ms) {
1273 if (!arch_supports_vector(Op_VectorReinterpret, mem_num_elem, T_BYTE, VecMaskNotUsed)) {
1274 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s mismatched_ms=1",
1275 is_store, is_store ? "storeMasked" : "loadMasked",
1276 num_elem, type2name(elem_bt));
1277 set_map(old_map);
1278 set_sp(old_sp);
1279 return false;
1280 }
1281 }
1282
1283 // Since it needs to unbox the mask, we need to double check that the related load operations
1284 // for mask are supported by backend.
1285 if (!arch_supports_vector(Op_LoadVector, num_elem, elem_bt, VecMaskUseLoad)) {
1286 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s",
1287 is_store, is_store ? "storeMasked" : "loadMasked",
1288 num_elem, type2name(elem_bt));
1289 set_map(old_map);
1290 set_sp(old_sp);
1291 return false;
1292 }
1293
1294 // Can base be null? Otherwise, always on-heap access.
1295 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(gvn().type(base));
1296 if (can_access_non_heap) {
1297 insert_mem_bar(Op_MemBarCPUOrder);
1298 }
1299
1300 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1301 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1302 assert(!is_vector_mask(vbox_klass) && is_vector_mask(mbox_klass), "Invalid class type");
1303 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1304 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1305
1306 Node* mask = unbox_vector(is_store ? argument(9) : argument(8), mbox_type, elem_bt, num_elem);
1307 if (mask == nullptr) {
1308 log_if_needed(" ** unbox failed mask=%s",
1309 is_store ? NodeClassNames[argument(9)->Opcode()]
1310 : NodeClassNames[argument(8)->Opcode()]);
1311 set_map(old_map);
1312 set_sp(old_sp);
1313 return false;
1314 }
1315
1316 if (is_store) {
1317 Node* val = unbox_vector(argument(8), vbox_type, elem_bt, num_elem);
1318 if (val == nullptr) {
1319 log_if_needed(" ** unbox failed vector=%s",
1320 NodeClassNames[argument(8)->Opcode()]);
1321 set_map(old_map);
1322 set_sp(old_sp);
1323 return false; // operand unboxing failed
1324 }
1325 set_all_memory(reset_memory());
1326
1327 if (mismatched_ms) {
1328 // Reinterpret the incoming vector to byte vector.
1329 const TypeVect* to_vect_type = TypeVect::make(mem_elem_bt, mem_num_elem);
1330 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type));
1331 // Reinterpret the vector mask to byte type.
1332 const TypeVect* from_mask_type = TypeVect::makemask(elem_bt, num_elem);
1333 const TypeVect* to_mask_type = TypeVect::makemask(mem_elem_bt, mem_num_elem);
1334 mask = gvn().transform(new VectorReinterpretNode(mask, from_mask_type, to_mask_type));
1335 }
1336 Node* vstore = gvn().transform(new StoreVectorMaskedNode(control(), memory(addr), addr, val, addr_type, mask));
1337 set_memory(vstore, addr_type);
1338 } else {
1339 Node* vload = nullptr;
1340
1341 if (mismatched_ms) {
1342 // Reinterpret the vector mask to byte type.
1343 const TypeVect* from_mask_type = TypeVect::makemask(elem_bt, num_elem);
1344 const TypeVect* to_mask_type = TypeVect::makemask(mem_elem_bt, mem_num_elem);
1345 mask = gvn().transform(new VectorReinterpretNode(mask, from_mask_type, to_mask_type));
1346 }
1347
1348 if (supports_predicate) {
1349 // Generate masked load vector node if predicate feature is supported.
1350 const TypeVect* vt = TypeVect::make(mem_elem_bt, mem_num_elem);
1351 vload = gvn().transform(new LoadVectorMaskedNode(control(), memory(addr), addr, addr_type, vt, mask));
1352 } else {
1353 // Use the vector blend to implement the masked load vector. The biased elements are zeros.
1354 Node* zero = gvn().transform(gvn().zerocon(mem_elem_bt));
1355 zero = gvn().transform(VectorNode::scalar2vector(zero, mem_num_elem, Type::get_const_basic_type(mem_elem_bt)));
1356 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, mem_num_elem, mem_elem_bt));
1357 vload = gvn().transform(new VectorBlendNode(zero, vload, mask));
1358 }
1359
1360 if (mismatched_ms) {
1361 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
1362 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type));
1363 }
1364
1365 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
1366 set_result(box);
1367 }
1368
1369 destruct_map_clone(old_map);
1370
1371 if (can_access_non_heap) {
1372 insert_mem_bar(Op_MemBarCPUOrder);
1373 }
1374
1375 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1376 return true;
1377 }
1378
1379 // <C,
1380 // V extends Vector<?>,
1381 // W extends Vector<Integer>,
1382 // S extends VectorSpecies<E>,
1383 // M extends VectorMask<E>,
1384 // E>
1385 // V loadWithMap(Class<? extends V> vectorClass, Class<M> maskClass, Class<E> elementType, int length,
1386 // Class<? extends Vector<Integer>> vectorIndexClass,
1387 // Object base, long offset, // Unsafe addressing
1388 // W index_vector, M m,
1389 // C container, int index, int[] indexMap, int indexM, S s, // Arguments for default implementation
1390 // LoadVectorOperationWithMap<C, V, E, S, M> defaultImpl)
1391 //
1392 // <C,
1393 // V extends Vector<E>,
1394 // W extends Vector<Integer>,
1395 // M extends VectorMask<E>,
1396 // E>
1397 // void storeWithMap(Class<? extends V> vectorClass, Class<M> maskClass, Class<E> elementType,
1398 // int length, Class<? extends Vector<Integer>> vectorIndexClass, Object base, long offset, // Unsafe addressing
1399 // W index_vector, V v, M m,
1400 // C container, int index, int[] indexMap, int indexM, // Arguments for default implementation
1401 // StoreVectorOperationWithMap<C, V, M, E> defaultImpl)
1402 //
1403 bool LibraryCallKit::inline_vector_gather_scatter(bool is_scatter) {
1404 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1405 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
1406 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1407 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1408 const TypeInstPtr* vector_idx_klass = gvn().type(argument(4))->isa_instptr();
1409
1410 if (vector_klass == nullptr || elem_klass == nullptr || vector_idx_klass == nullptr || vlen == nullptr ||
1411 vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || vector_idx_klass->const_oop() == nullptr || !vlen->is_con()) {
1412 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s viclass=%s",
1413 NodeClassNames[argument(0)->Opcode()],
1414 NodeClassNames[argument(2)->Opcode()],
1415 NodeClassNames[argument(3)->Opcode()],
1416 NodeClassNames[argument(4)->Opcode()]);
1417 return false; // not enough info for intrinsification
1418 }
1419
1420 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(vector_idx_klass)) {
1421 log_if_needed(" ** klass argument not initialized");
1422 return false;
1423 }
1424
1425 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1426 if (!elem_type->is_primitive_type()) {
1427 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1428 return false; // should be primitive type
1429 }
1430
1431 BasicType elem_bt = elem_type->basic_type();
1432 int num_elem = vlen->get_con();
1433
1434 const Type* vmask_type = gvn().type(is_scatter ? argument(10) : argument(9));
1435 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
1436 if (is_masked_op) {
1437 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
1438 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(1)->Opcode()]);
1439 return false; // not enough info for intrinsification
1440 }
1441
1442 if (!is_klass_initialized(mask_klass)) {
1443 log_if_needed(" ** mask klass argument not initialized");
1444 return false;
1445 }
1446
1447 if (vmask_type->maybe_null()) {
1448 log_if_needed(" ** null mask values are not allowed for masked op");
1449 return false;
1450 }
1451
1452 // Check whether the predicated gather/scatter node is supported by architecture.
1453 VectorMaskUseType mask = (VectorMaskUseType) (VecMaskUseLoad | VecMaskUsePred);
1454 if (!arch_supports_vector(is_scatter ? Op_StoreVectorScatterMasked : Op_LoadVectorGatherMasked, num_elem, elem_bt, mask)) {
1455 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s is_masked_op=1",
1456 is_scatter, is_scatter ? "scatterMasked" : "gatherMasked",
1457 num_elem, type2name(elem_bt));
1458 return false; // not supported
1459 }
1460 } else {
1461 // Check whether the normal gather/scatter node is supported for non-masked operation.
1462 if (!arch_supports_vector(is_scatter ? Op_StoreVectorScatter : Op_LoadVectorGather, num_elem, elem_bt, VecMaskNotUsed)) {
1463 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s is_masked_op=0",
1464 is_scatter, is_scatter ? "scatter" : "gather",
1465 num_elem, type2name(elem_bt));
1466 return false; // not supported
1467 }
1468 }
1469
1470 // Check that the vector holding indices is supported by architecture
1471 // For sub-word gathers expander receive index array.
1472 if (!is_subword_type(elem_bt) && !arch_supports_vector(Op_LoadVector, num_elem, T_INT, VecMaskNotUsed)) {
1473 log_if_needed(" ** not supported: arity=%d op=%s/loadindex vlen=%d etype=int is_masked_op=%d",
1474 is_scatter, is_scatter ? "scatter" : "gather",
1475 num_elem, is_masked_op ? 1 : 0);
1476 return false; // not supported
1477 }
1478
1479 Node* base = argument(5);
1480 Node* offset = ConvL2X(argument(6));
1481
1482 // Save state and restore on bailout
1483 uint old_sp = sp();
1484 SafePointNode* old_map = clone_map();
1485
1486 Node* addr = make_unsafe_address(base, offset, elem_bt, true);
1487
1488 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
1489 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
1490
1491 // The array must be consistent with vector type
1492 if (arr_type == nullptr || (arr_type != nullptr && !elem_consistent_with_arr(elem_bt, arr_type, false))) {
1493 log_if_needed(" ** not supported: arity=%d op=%s vlen=%d etype=%s atype=%s ismask=no",
1494 is_scatter, is_scatter ? "scatter" : "gather",
1495 num_elem, type2name(elem_bt), type2name(arr_type->elem()->array_element_basic_type()));
1496 set_map(old_map);
1497 set_sp(old_sp);
1498 return false;
1499 }
1500
1501 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1502 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1503 ciKlass* vbox_idx_klass = vector_idx_klass->const_oop()->as_instance()->java_lang_Class_klass();
1504 if (vbox_idx_klass == nullptr) {
1505 set_map(old_map);
1506 set_sp(old_sp);
1507 return false;
1508 }
1509
1510 Node* index_vect = nullptr;
1511 const TypeInstPtr* vbox_idx_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_idx_klass);
1512 if (!is_subword_type(elem_bt)) {
1513 index_vect = unbox_vector(argument(8), vbox_idx_type, T_INT, num_elem);
1514 if (index_vect == nullptr) {
1515 set_map(old_map);
1516 set_sp(old_sp);
1517 return false;
1518 }
1519 }
1520
1521 Node* mask = nullptr;
1522 if (is_masked_op) {
1523 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1524 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1525 mask = unbox_vector(is_scatter ? argument(10) : argument(9), mbox_type, elem_bt, num_elem);
1526 if (mask == nullptr) {
1527 log_if_needed(" ** unbox failed mask=%s",
1528 is_scatter ? NodeClassNames[argument(10)->Opcode()]
1529 : NodeClassNames[argument(9)->Opcode()]);
1530 set_map(old_map);
1531 set_sp(old_sp);
1532 return false;
1533 }
1534 }
1535
1536 const TypeVect* vector_type = TypeVect::make(elem_bt, num_elem);
1537 if (is_scatter) {
1538 Node* val = unbox_vector(argument(9), vbox_type, elem_bt, num_elem);
1539 if (val == nullptr) {
1540 set_map(old_map);
1541 set_sp(old_sp);
1542 return false; // operand unboxing failed
1543 }
1544 set_all_memory(reset_memory());
1545
1546 Node* vstore = nullptr;
1547 if (mask != nullptr) {
1548 vstore = gvn().transform(new StoreVectorScatterMaskedNode(control(), memory(addr), addr, addr_type, val, index_vect, mask));
1549 } else {
1550 vstore = gvn().transform(new StoreVectorScatterNode(control(), memory(addr), addr, addr_type, val, index_vect));
1551 }
1552 set_memory(vstore, addr_type);
1553 } else {
1554 Node* vload = nullptr;
1555 Node* index = argument(11);
1556 Node* indexMap = argument(12);
1557 Node* indexM = argument(13);
1558 if (mask != nullptr) {
1559 if (is_subword_type(elem_bt)) {
1560 Node* index_arr_base = array_element_address(indexMap, indexM, T_INT);
1561 vload = gvn().transform(new LoadVectorGatherMaskedNode(control(), memory(addr), addr, addr_type, vector_type, index_arr_base, mask, index));
1562 } else {
1563 vload = gvn().transform(new LoadVectorGatherMaskedNode(control(), memory(addr), addr, addr_type, vector_type, index_vect, mask));
1564 }
1565 } else {
1566 if (is_subword_type(elem_bt)) {
1567 Node* index_arr_base = array_element_address(indexMap, indexM, T_INT);
1568 vload = gvn().transform(new LoadVectorGatherNode(control(), memory(addr), addr, addr_type, vector_type, index_arr_base, index));
1569 } else {
1570 vload = gvn().transform(new LoadVectorGatherNode(control(), memory(addr), addr, addr_type, vector_type, index_vect));
1571 }
1572 }
1573 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
1574 set_result(box);
1575 }
1576
1577 destruct_map_clone(old_map);
1578
1579 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1580 return true;
1581 }
1582
1583 // public static
1584 // <V extends Vector<E>,
1585 // M extends VectorMask<E>,
1586 // E>
1587 // long reductionCoerced(int oprId, Class<? extends V> vectorClass, Class<? extends M> maskClass,
1588 // Class<E> elementType, int length, V v, M m,
1589 // ReductionOperation<V, M> defaultImpl)
1590 bool LibraryCallKit::inline_vector_reduction() {
1591 const TypeInt* opr = gvn().type(argument(0))->isa_int();
1592 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
1593 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
1594 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
1595 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
1596
1597 if (opr == nullptr || vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
1598 !opr->is_con() || vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
1599 log_if_needed(" ** missing constant: opr=%s vclass=%s etype=%s vlen=%s",
1600 NodeClassNames[argument(0)->Opcode()],
1601 NodeClassNames[argument(1)->Opcode()],
1602 NodeClassNames[argument(3)->Opcode()],
1603 NodeClassNames[argument(4)->Opcode()]);
1604 return false; // not enough info for intrinsification
1605 }
1606 if (!is_klass_initialized(vector_klass)) {
1607 log_if_needed(" ** klass argument not initialized");
1608 return false;
1609 }
1610 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1611 if (!elem_type->is_primitive_type()) {
1612 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1613 return false; // should be primitive type
1614 }
1615
1616 const Type* vmask_type = gvn().type(argument(6));
1617 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
1618 if (is_masked_op) {
1619 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
1620 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(2)->Opcode()]);
1621 return false; // not enough info for intrinsification
1622 }
1623
1624 if (!is_klass_initialized(mask_klass)) {
1625 log_if_needed(" ** mask klass argument not initialized");
1626 return false;
1627 }
1628
1629 if (vmask_type->maybe_null()) {
1630 log_if_needed(" ** null mask values are not allowed for masked op");
1631 return false;
1632 }
1633 }
1634
1635 BasicType elem_bt = elem_type->basic_type();
1636 int num_elem = vlen->get_con();
1637 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
1638 int sopc = ReductionNode::opcode(opc, elem_bt);
1639
1640 // When using mask, mask use type needs to be VecMaskUseLoad.
1641 if (!arch_supports_vector(sopc, num_elem, elem_bt, is_masked_op ? VecMaskUseLoad : VecMaskNotUsed)) {
1642 log_if_needed(" ** not supported: arity=1 op=%d/reduce vlen=%d etype=%s is_masked_op=%d",
1643 sopc, num_elem, type2name(elem_bt), is_masked_op ? 1 : 0);
1644 return false;
1645 }
1646
1647 // Return true if current platform has implemented the masked operation with predicate feature.
1648 bool use_predicate = is_masked_op && arch_supports_vector(sopc, num_elem, elem_bt, VecMaskUsePred);
1649 if (is_masked_op && !use_predicate && !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
1650 log_if_needed(" ** not supported: arity=1 op=%d/reduce vlen=%d etype=%s is_masked_op=1",
1651 sopc, num_elem, type2name(elem_bt));
1652 return false;
1653 }
1654
1655 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1656 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1657
1658 Node* opd = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1659 if (opd == nullptr) {
1660 return false; // operand unboxing failed
1661 }
1662
1663 Node* mask = nullptr;
1664 if (is_masked_op) {
1665 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1666 assert(is_vector_mask(mbox_klass), "argument(2) should be a mask class");
1667 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1668 mask = unbox_vector(argument(6), mbox_type, elem_bt, num_elem);
1669 if (mask == nullptr) {
1670 log_if_needed(" ** unbox failed mask=%s",
1671 NodeClassNames[argument(6)->Opcode()]);
1672 return false;
1673 }
1674 }
1675
1676 Node* init = ReductionNode::make_identity_con_scalar(gvn(), opc, elem_bt);
1677 Node* value = opd;
1678
1679 assert(mask != nullptr || !is_masked_op, "Masked op needs the mask value never null");
1680 if (mask != nullptr && !use_predicate) {
1681 Node* reduce_identity = gvn().transform(VectorNode::scalar2vector(init, num_elem, Type::get_const_basic_type(elem_bt)));
1682 value = gvn().transform(new VectorBlendNode(reduce_identity, value, mask));
1683 }
1684
1685 // Make an unordered Reduction node. This affects only AddReductionVF/VD and MulReductionVF/VD,
1686 // as these operations are allowed to be associative (not requiring strict order) in VectorAPI.
1687 value = ReductionNode::make(opc, nullptr, init, value, elem_bt, /* requires_strict_order */ false);
1688
1689 if (mask != nullptr && use_predicate) {
1690 value->add_req(mask);
1691 value->add_flag(Node::Flag_is_predicated_vector);
1692 }
1693
1694 value = gvn().transform(value);
1695
1696 Node* bits = nullptr;
1697 switch (elem_bt) {
1698 case T_BYTE:
1699 case T_SHORT:
1700 case T_INT: {
1701 bits = gvn().transform(new ConvI2LNode(value));
1702 break;
1703 }
1704 case T_FLOAT: {
1705 value = gvn().transform(new MoveF2INode(value));
1706 bits = gvn().transform(new ConvI2LNode(value));
1707 break;
1708 }
1709 case T_DOUBLE: {
1710 bits = gvn().transform(new MoveD2LNode(value));
1711 break;
1712 }
1713 case T_LONG: {
1714 bits = value; // no conversion needed
1715 break;
1716 }
1717 default: fatal("%s", type2name(elem_bt));
1718 }
1719 set_result(bits);
1720 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1721 return true;
1722 }
1723
1724 // public static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
1725 // V v1, V v2,
1726 // BiFunction<V, V, Boolean> defaultImpl)
1727 //
1728 bool LibraryCallKit::inline_vector_test() {
1729 const TypeInt* cond = gvn().type(argument(0))->isa_int();
1730 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
1731 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1732 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1733
1734 if (cond == nullptr || vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
1735 !cond->is_con() || vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
1736 log_if_needed(" ** missing constant: cond=%s vclass=%s etype=%s vlen=%s",
1737 NodeClassNames[argument(0)->Opcode()],
1738 NodeClassNames[argument(1)->Opcode()],
1739 NodeClassNames[argument(2)->Opcode()],
1740 NodeClassNames[argument(3)->Opcode()]);
1741 return false; // not enough info for intrinsification
1742 }
1743 if (!is_klass_initialized(vector_klass)) {
1744 log_if_needed(" ** klass argument not initialized");
1745 return false;
1746 }
1747 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1748 if (!elem_type->is_primitive_type()) {
1749 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1750 return false; // should be primitive type
1751 }
1752 BasicType elem_bt = elem_type->basic_type();
1753 int num_elem = vlen->get_con();
1754 BoolTest::mask booltest = (BoolTest::mask)cond->get_con();
1755 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1756 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1757
1758 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, is_vector_mask(vbox_klass) ? VecMaskUseLoad : VecMaskNotUsed)) {
1759 log_if_needed(" ** not supported: arity=2 op=test/%d vlen=%d etype=%s ismask=%d",
1760 cond->get_con(), num_elem, type2name(elem_bt),
1761 is_vector_mask(vbox_klass));
1762 return false;
1763 }
1764
1765 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
1766 Node* opd2;
1767 if (Matcher::vectortest_needs_second_argument(booltest == BoolTest::overflow,
1768 opd1->bottom_type()->isa_vectmask())) {
1769 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1770 } else {
1771 opd2 = opd1;
1772 }
1773 if (opd1 == nullptr || opd2 == nullptr) {
1774 return false; // operand unboxing failed
1775 }
1776
1777 Node* cmp = gvn().transform(new VectorTestNode(opd1, opd2, booltest));
1778 BoolTest::mask test = Matcher::vectortest_mask(booltest == BoolTest::overflow,
1779 opd1->bottom_type()->isa_vectmask(), num_elem);
1780 Node* bol = gvn().transform(new BoolNode(cmp, test));
1781 Node* res = gvn().transform(new CMoveINode(bol, gvn().intcon(0), gvn().intcon(1), TypeInt::BOOL));
1782
1783 set_result(res);
1784 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1785 return true;
1786 }
1787
1788 // public static
1789 // <V extends Vector<E>,
1790 // M extends VectorMask<E>,
1791 // E>
1792 // V blend(Class<? extends V> vectorClass, Class<M> maskClass, Class<E> elementType, int vlen,
1793 // V v1, V v2, M m,
1794 // VectorBlendOp<V, M, E> defaultImpl)
1795 bool LibraryCallKit::inline_vector_blend() {
1796 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1797 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
1798 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
1799 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
1800
1801 if (mask_klass == nullptr || vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
1802 return false; // dead code
1803 }
1804 if (mask_klass->const_oop() == nullptr || vector_klass->const_oop() == nullptr ||
1805 elem_klass->const_oop() == nullptr || !vlen->is_con()) {
1806 log_if_needed(" ** missing constant: vclass=%s mclass=%s etype=%s vlen=%s",
1807 NodeClassNames[argument(0)->Opcode()],
1808 NodeClassNames[argument(1)->Opcode()],
1809 NodeClassNames[argument(2)->Opcode()],
1810 NodeClassNames[argument(3)->Opcode()]);
1811 return false; // not enough info for intrinsification
1812 }
1813 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(mask_klass)) {
1814 log_if_needed(" ** klass argument not initialized");
1815 return false;
1816 }
1817 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1818 if (!elem_type->is_primitive_type()) {
1819 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1820 return false; // should be primitive type
1821 }
1822 BasicType elem_bt = elem_type->basic_type();
1823 BasicType mask_bt = elem_bt;
1824 int num_elem = vlen->get_con();
1825
1826 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
1827 log_if_needed(" ** not supported: arity=2 op=blend vlen=%d etype=%s ismask=useload",
1828 num_elem, type2name(elem_bt));
1829 return false; // not supported
1830 }
1831 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1832 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1833
1834 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1835 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1836
1837 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
1838 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1839 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
1840
1841 if (v1 == nullptr || v2 == nullptr || mask == nullptr) {
1842 return false; // operand unboxing failed
1843 }
1844
1845 Node* blend = gvn().transform(new VectorBlendNode(v1, v2, mask));
1846
1847 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
1848 set_result(box);
1849 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1850 return true;
1851 }
1852
1853 // public static
1854 // <V extends Vector<E>,
1855 // M extends VectorMask<E>,
1856 // E>
1857 // M compare(int cond, Class<? extends V> vectorClass, Class<M> maskClass, Class<E> elementType, int vlen,
1858 // V v1, V v2, M m,
1859 // VectorCompareOp<V,M> defaultImpl)
1860 bool LibraryCallKit::inline_vector_compare() {
1861 const TypeInt* cond = gvn().type(argument(0))->isa_int();
1862 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
1863 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
1864 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
1865 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
1866
1867 if (cond == nullptr || vector_klass == nullptr || mask_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
1868 return false; // dead code
1869 }
1870 if (!cond->is_con() || vector_klass->const_oop() == nullptr || mask_klass->const_oop() == nullptr ||
1871 elem_klass->const_oop() == nullptr || !vlen->is_con()) {
1872 log_if_needed(" ** missing constant: cond=%s vclass=%s mclass=%s etype=%s vlen=%s",
1873 NodeClassNames[argument(0)->Opcode()],
1874 NodeClassNames[argument(1)->Opcode()],
1875 NodeClassNames[argument(2)->Opcode()],
1876 NodeClassNames[argument(3)->Opcode()],
1877 NodeClassNames[argument(4)->Opcode()]);
1878 return false; // not enough info for intrinsification
1879 }
1880 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(mask_klass)) {
1881 log_if_needed(" ** klass argument not initialized");
1882 return false;
1883 }
1884 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1885 if (!elem_type->is_primitive_type()) {
1886 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1887 return false; // should be primitive type
1888 }
1889
1890 int num_elem = vlen->get_con();
1891 BasicType elem_bt = elem_type->basic_type();
1892 BasicType mask_bt = elem_bt;
1893
1894 if ((cond->get_con() & BoolTest::unsigned_compare) != 0) {
1895 if (!Matcher::supports_vector_comparison_unsigned(num_elem, elem_bt)) {
1896 log_if_needed(" ** not supported: unsigned comparison op=comp/%d vlen=%d etype=%s ismask=usestore",
1897 cond->get_con() & (BoolTest::unsigned_compare - 1), num_elem, type2name(elem_bt));
1898 return false;
1899 }
1900 }
1901
1902 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
1903 log_if_needed(" ** not supported: arity=2 op=comp/%d vlen=%d etype=%s ismask=usestore",
1904 cond->get_con(), num_elem, type2name(elem_bt));
1905 return false;
1906 }
1907
1908 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
1909 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
1910
1911 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
1912 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
1913
1914 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
1915 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
1916
1917 bool is_masked_op = argument(7)->bottom_type() != TypePtr::NULL_PTR;
1918 Node* mask = is_masked_op ? unbox_vector(argument(7), mbox_type, elem_bt, num_elem) : nullptr;
1919 if (is_masked_op && mask == nullptr) {
1920 log_if_needed(" ** not supported: mask = null arity=2 op=comp/%d vlen=%d etype=%s ismask=usestore is_masked_op=1",
1921 cond->get_con(), num_elem, type2name(elem_bt));
1922 return false;
1923 }
1924
1925 bool use_predicate = is_masked_op && arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUsePred);
1926 if (is_masked_op && !use_predicate && !arch_supports_vector(Op_AndV, num_elem, elem_bt, VecMaskUseLoad)) {
1927 log_if_needed(" ** not supported: arity=2 op=comp/%d vlen=%d etype=%s ismask=usestore is_masked_op=1",
1928 cond->get_con(), num_elem, type2name(elem_bt));
1929 return false;
1930 }
1931
1932 if (v1 == nullptr || v2 == nullptr) {
1933 return false; // operand unboxing failed
1934 }
1935 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
1936 ConINode* pred_node = (ConINode*)gvn().makecon(cond);
1937
1938 const TypeVect* vmask_type = TypeVect::makemask(mask_bt, num_elem);
1939 Node* operation = new VectorMaskCmpNode(pred, v1, v2, pred_node, vmask_type);
1940
1941 if (is_masked_op) {
1942 if (use_predicate) {
1943 operation->add_req(mask);
1944 operation->add_flag(Node::Flag_is_predicated_vector);
1945 } else {
1946 operation = gvn().transform(operation);
1947 operation = VectorNode::make(Op_AndV, operation, mask, vmask_type);
1948 }
1949 }
1950
1951 operation = gvn().transform(operation);
1952
1953 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
1954 set_result(box);
1955 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
1956 return true;
1957 }
1958
1959 // public static
1960 // <V extends Vector<E>,
1961 // Sh extends VectorShuffle<E>,
1962 // M extends VectorMask<E>,
1963 // E>
1964 // V rearrangeOp(Class<? extends V> vectorClass, Class<Sh> shuffleClass, Class<M> maskClass, Class<E> elementType, int vlen,
1965 // V v1, Sh sh, M m,
1966 // VectorRearrangeOp<V, Sh, M, E> defaultImpl)
1967 bool LibraryCallKit::inline_vector_rearrange() {
1968 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
1969 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->isa_instptr();
1970 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
1971 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
1972 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
1973
1974 if (vector_klass == nullptr || shuffle_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
1975 return false; // dead code
1976 }
1977 if (shuffle_klass->const_oop() == nullptr ||
1978 vector_klass->const_oop() == nullptr ||
1979 elem_klass->const_oop() == nullptr ||
1980 !vlen->is_con()) {
1981 log_if_needed(" ** missing constant: vclass=%s sclass=%s etype=%s vlen=%s",
1982 NodeClassNames[argument(0)->Opcode()],
1983 NodeClassNames[argument(1)->Opcode()],
1984 NodeClassNames[argument(3)->Opcode()],
1985 NodeClassNames[argument(4)->Opcode()]);
1986 return false; // not enough info for intrinsification
1987 }
1988 if (!is_klass_initialized(vector_klass) ||
1989 !is_klass_initialized(shuffle_klass)) {
1990 log_if_needed(" ** klass argument not initialized");
1991 return false;
1992 }
1993 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
1994 if (!elem_type->is_primitive_type()) {
1995 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
1996 return false; // should be primitive type
1997 }
1998 BasicType elem_bt = elem_type->basic_type();
1999 BasicType shuffle_bt = elem_bt;
2000 int num_elem = vlen->get_con();
2001
2002 if (!arch_supports_vector(Op_VectorLoadShuffle, num_elem, elem_bt, VecMaskNotUsed)) {
2003 log_if_needed(" ** not supported: arity=0 op=load/shuffle vlen=%d etype=%s ismask=no",
2004 num_elem, type2name(elem_bt));
2005 return false; // not supported
2006 }
2007
2008 bool is_masked_op = argument(7)->bottom_type() != TypePtr::NULL_PTR;
2009 bool use_predicate = is_masked_op;
2010 if (is_masked_op &&
2011 (mask_klass == nullptr ||
2012 mask_klass->const_oop() == nullptr ||
2013 !is_klass_initialized(mask_klass))) {
2014 log_if_needed(" ** mask_klass argument not initialized");
2015 }
2016 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_masked_op ? (VecMaskUseLoad | VecMaskUsePred) : VecMaskNotUsed);
2017 if (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, checkFlags)) {
2018 use_predicate = false;
2019 if(!is_masked_op ||
2020 (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed) ||
2021 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad) ||
2022 !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed))) {
2023 log_if_needed(" ** not supported: arity=2 op=shuffle/rearrange vlen=%d etype=%s ismask=no",
2024 num_elem, type2name(elem_bt));
2025 return false; // not supported
2026 }
2027 }
2028 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2029 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2030
2031 ciKlass* shbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
2032 const TypeInstPtr* shbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, shbox_klass);
2033
2034 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
2035 Node* shuffle = unbox_vector(argument(6), shbox_type, shuffle_bt, num_elem);
2036
2037 if (v1 == nullptr || shuffle == nullptr) {
2038 return false; // operand unboxing failed
2039 }
2040
2041 Node* mask = nullptr;
2042 if (is_masked_op) {
2043 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
2044 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
2045 mask = unbox_vector(argument(7), mbox_type, elem_bt, num_elem);
2046 if (mask == nullptr) {
2047 log_if_needed(" ** not supported: arity=3 op=shuffle/rearrange vlen=%d etype=%s ismask=useload is_masked_op=1",
2048 num_elem, type2name(elem_bt));
2049 return false;
2050 }
2051 }
2052
2053 Node* rearrange = new VectorRearrangeNode(v1, shuffle);
2054 if (is_masked_op) {
2055 if (use_predicate) {
2056 rearrange->add_req(mask);
2057 rearrange->add_flag(Node::Flag_is_predicated_vector);
2058 } else {
2059 const TypeVect* vt = v1->bottom_type()->is_vect();
2060 rearrange = gvn().transform(rearrange);
2061 Node* zero = gvn().makecon(Type::get_zero_type(elem_bt));
2062 Node* zerovec = gvn().transform(VectorNode::scalar2vector(zero, num_elem, Type::get_const_basic_type(elem_bt)));
2063 rearrange = new VectorBlendNode(zerovec, rearrange, mask);
2064 }
2065 }
2066 rearrange = gvn().transform(rearrange);
2067
2068 Node* box = box_vector(rearrange, vbox_type, elem_bt, num_elem);
2069 set_result(box);
2070 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2071 return true;
2072 }
2073
2074 static address get_vector_math_address(int vop, int bits, BasicType bt, char* name_ptr, int name_len) {
2075 address addr = nullptr;
2076 assert(UseVectorStubs, "sanity");
2077 assert(name_ptr != nullptr, "unexpected");
2078 assert((vop >= VectorSupport::VECTOR_OP_MATH_START) && (vop <= VectorSupport::VECTOR_OP_MATH_END), "unexpected");
2079 int op = vop - VectorSupport::VECTOR_OP_MATH_START;
2080
2081 switch(bits) {
2082 case 64: //fallthough
2083 case 128: //fallthough
2084 case 256: //fallthough
2085 case 512:
2086 if (bt == T_FLOAT) {
2087 snprintf(name_ptr, name_len, "vector_%s_float_%dbits_fixed", VectorSupport::mathname[op], bits);
2088 addr = StubRoutines::_vector_f_math[exact_log2(bits/64)][op];
2089 } else {
2090 assert(bt == T_DOUBLE, "must be FP type only");
2091 snprintf(name_ptr, name_len, "vector_%s_double_%dbits_fixed", VectorSupport::mathname[op], bits);
2092 addr = StubRoutines::_vector_d_math[exact_log2(bits/64)][op];
2093 }
2094 break;
2095 default:
2096 if (!Matcher::supports_scalable_vector() || !Matcher::vector_size_supported(bt, bits/type2aelembytes(bt)) ) {
2097 snprintf(name_ptr, name_len, "invalid");
2098 addr = nullptr;
2099 Unimplemented();
2100 }
2101 break;
2102 }
2103
2104 if (addr == nullptr && Matcher::supports_scalable_vector()) {
2105 if (bt == T_FLOAT) {
2106 snprintf(name_ptr, name_len, "vector_%s_float_%dbits_scalable", VectorSupport::mathname[op], bits);
2107 addr = StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_SCALABLE][op];
2108 } else {
2109 assert(bt == T_DOUBLE, "must be FP type only");
2110 snprintf(name_ptr, name_len, "vector_%s_double_%dbits_scalable", VectorSupport::mathname[op], bits);
2111 addr = StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_SCALABLE][op];
2112 }
2113 }
2114
2115 return addr;
2116 }
2117
2118 // public static
2119 // <V extends Vector<E>,
2120 // M extends VectorMask<E>,
2121 // E>
2122 // V selectFromOp(Class<? extends V> vClass, Class<M> mClass, Class<E> eClass,
2123 // int length, V v1, V v2, M m,
2124 // VectorSelectFromOp<V, M> defaultImpl)
2125 bool LibraryCallKit::inline_vector_select_from() {
2126 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2127 const TypeInstPtr* mask_klass = gvn().type(argument(1))->isa_instptr();
2128 const TypeInstPtr* elem_klass = gvn().type(argument(2))->isa_instptr();
2129 const TypeInt* vlen = gvn().type(argument(3))->isa_int();
2130
2131 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
2132 vector_klass->const_oop() == nullptr ||
2133 elem_klass->const_oop() == nullptr ||
2134 !vlen->is_con()) {
2135 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s",
2136 NodeClassNames[argument(0)->Opcode()],
2137 NodeClassNames[argument(2)->Opcode()],
2138 NodeClassNames[argument(3)->Opcode()]);
2139 return false; // not enough info for intrinsification
2140 }
2141 if (!is_klass_initialized(vector_klass)) {
2142 log_if_needed(" ** klass argument not initialized");
2143 return false;
2144 }
2145 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2146 if (!elem_type->is_primitive_type()) {
2147 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2148 return false; // should be primitive type
2149 }
2150 BasicType elem_bt = elem_type->basic_type();
2151 int num_elem = vlen->get_con();
2152 if (!is_power_of_2(num_elem)) {
2153 log_if_needed(" ** vlen not power of two=%d", num_elem);
2154 return false;
2155 }
2156
2157 int cast_vopc = VectorCastNode::opcode(-1, elem_bt); // from vector of type elem_bt
2158 if (!arch_supports_vector(Op_VectorLoadShuffle, num_elem, elem_bt, VecMaskNotUsed)||
2159 !arch_supports_vector(Op_AndV, num_elem, T_BYTE, VecMaskNotUsed) ||
2160 !arch_supports_vector(Op_Replicate, num_elem, T_BYTE, VecMaskNotUsed) ||
2161 !arch_supports_vector(cast_vopc, num_elem, T_BYTE, VecMaskNotUsed)) {
2162 log_if_needed(" ** not supported: arity=0 op=selectFrom vlen=%d etype=%s ismask=no",
2163 num_elem, type2name(elem_bt));
2164 return false; // not supported
2165 }
2166
2167 bool is_masked_op = argument(6)->bottom_type() != TypePtr::NULL_PTR;
2168 bool use_predicate = is_masked_op;
2169 if (is_masked_op &&
2170 (mask_klass == nullptr ||
2171 mask_klass->const_oop() == nullptr ||
2172 !is_klass_initialized(mask_klass))) {
2173 log_if_needed(" ** mask_klass argument not initialized");
2174 return false; // not supported
2175 }
2176 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_masked_op ? (VecMaskUseLoad | VecMaskUsePred) : VecMaskNotUsed);
2177 if (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, checkFlags)) {
2178 use_predicate = false;
2179 if(!is_masked_op ||
2180 (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed) ||
2181 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad) ||
2182 !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed))) {
2183 log_if_needed(" ** not supported: op=selectFrom vlen=%d etype=%s is_masked_op=%d",
2184 num_elem, type2name(elem_bt), is_masked_op);
2185 return false; // not supported
2186 }
2187 }
2188 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2189 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2190
2191 // v1 is the index vector
2192 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
2193 // v2 is the vector being rearranged
2194 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
2195
2196 if (v1 == nullptr) {
2197 log_if_needed(" ** unbox failed v1=%s", NodeClassNames[argument(4)->Opcode()]);
2198 return false; // operand unboxing failed
2199 }
2200
2201 if (v2 == nullptr) {
2202 log_if_needed(" ** unbox failed v2=%s", NodeClassNames[argument(5)->Opcode()]);
2203 return false; // operand unboxing failed
2204 }
2205
2206 Node* mask = nullptr;
2207 if (is_masked_op) {
2208 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
2209 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
2210 mask = unbox_vector(argument(6), mbox_type, elem_bt, num_elem);
2211 if (mask == nullptr) {
2212 log_if_needed(" ** unbox failed mask=%s", NodeClassNames[argument(6)->Opcode()]);
2213 return false;
2214 }
2215 }
2216
2217 // cast index vector from elem_bt vector to byte vector
2218 const Type * byte_bt = Type::get_const_basic_type(T_BYTE);
2219 const TypeVect * byte_vt = TypeVect::make(byte_bt, num_elem);
2220 Node* byte_shuffle = gvn().transform(VectorCastNode::make(cast_vopc, v1, T_BYTE, num_elem));
2221
2222 // wrap the byte vector lanes to (num_elem - 1) to form the shuffle vector where num_elem is vector length
2223 // this is a simple AND operation as we come here only for power of two vector length
2224 Node* mod_val = gvn().makecon(TypeInt::make(num_elem-1));
2225 Node* bcast_mod = gvn().transform(VectorNode::scalar2vector(mod_val, num_elem, byte_bt));
2226 byte_shuffle = gvn().transform(VectorNode::make(Op_AndV, byte_shuffle, bcast_mod, byte_vt));
2227
2228 // load the shuffle to use in rearrange
2229 const TypeVect * shuffle_vt = TypeVect::make(elem_bt, num_elem);
2230 Node* load_shuffle = gvn().transform(new VectorLoadShuffleNode(byte_shuffle, shuffle_vt));
2231
2232 // and finally rearrange
2233 Node* rearrange = new VectorRearrangeNode(v2, load_shuffle);
2234 if (is_masked_op) {
2235 if (use_predicate) {
2236 // masked rearrange is supported so use that directly
2237 rearrange->add_req(mask);
2238 rearrange->add_flag(Node::Flag_is_predicated_vector);
2239 } else {
2240 // masked rearrange is not supported so emulate usig blend
2241 const TypeVect* vt = v1->bottom_type()->is_vect();
2242 rearrange = gvn().transform(rearrange);
2243
2244 // create a zero vector with each lane element set as zero
2245 Node* zero = gvn().makecon(Type::get_zero_type(elem_bt));
2246 Node* zerovec = gvn().transform(VectorNode::scalar2vector(zero, num_elem, Type::get_const_basic_type(elem_bt)));
2247
2248 // For each lane for which mask is set, blend in the rearranged lane into zero vector
2249 rearrange = new VectorBlendNode(zerovec, rearrange, mask);
2250 }
2251 }
2252 rearrange = gvn().transform(rearrange);
2253
2254 // box the result
2255 Node* box = box_vector(rearrange, vbox_type, elem_bt, num_elem);
2256 set_result(box);
2257
2258 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2259 return true;
2260 }
2261
2262 Node* LibraryCallKit::gen_call_to_vector_math(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2) {
2263 assert(UseVectorStubs, "sanity");
2264 assert(vector_api_op_id >= VectorSupport::VECTOR_OP_MATH_START && vector_api_op_id <= VectorSupport::VECTOR_OP_MATH_END, "need valid op id");
2265 assert(opd1 != nullptr, "must not be null");
2266 const TypeVect* vt = TypeVect::make(bt, num_elem);
2267 const TypeFunc* call_type = OptoRuntime::Math_Vector_Vector_Type(opd2 != nullptr ? 2 : 1, vt, vt);
2268 char name[100] = "";
2269
2270 // Get address for vector math method.
2271 address addr = get_vector_math_address(vector_api_op_id, vt->length_in_bytes() * BitsPerByte, bt, name, 100);
2272
2273 if (addr == nullptr) {
2274 return nullptr;
2275 }
2276
2277 assert(name[0] != '\0', "name must not be null");
2278 Node* operation = make_runtime_call(RC_VECTOR,
2279 call_type,
2280 addr,
2281 name,
2282 TypePtr::BOTTOM,
2283 opd1,
2284 opd2);
2285 return gvn().transform(new ProjNode(gvn().transform(operation), TypeFunc::Parms));
2286 }
2287
2288 // public static
2289 // <V extends Vector<E>,
2290 // M extends VectorMask<E>,
2291 // E>
2292 // V broadcastInt(int opr, Class<? extends V> vectorClass, Class<? extends M> maskClass,
2293 // Class<E> elementType, int length,
2294 // V v, int n, M m,
2295 // VectorBroadcastIntOp<V, M> defaultImpl)
2296 bool LibraryCallKit::inline_vector_broadcast_int() {
2297 const TypeInt* opr = gvn().type(argument(0))->isa_int();
2298 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
2299 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
2300 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
2301 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
2302
2303 if (opr == nullptr || vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr) {
2304 return false; // dead code
2305 }
2306 if (!opr->is_con() || vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
2307 log_if_needed(" ** missing constant: opr=%s vclass=%s etype=%s vlen=%s",
2308 NodeClassNames[argument(0)->Opcode()],
2309 NodeClassNames[argument(1)->Opcode()],
2310 NodeClassNames[argument(3)->Opcode()],
2311 NodeClassNames[argument(4)->Opcode()]);
2312 return false; // not enough info for intrinsification
2313 }
2314 if (!is_klass_initialized(vector_klass)) {
2315 log_if_needed(" ** klass argument not initialized");
2316 return false;
2317 }
2318
2319 const Type* vmask_type = gvn().type(argument(7));
2320 bool is_masked_op = vmask_type != TypePtr::NULL_PTR;
2321 if (is_masked_op) {
2322 if (mask_klass == nullptr || mask_klass->const_oop() == nullptr) {
2323 log_if_needed(" ** missing constant: maskclass=%s", NodeClassNames[argument(2)->Opcode()]);
2324 return false; // not enough info for intrinsification
2325 }
2326
2327 if (!is_klass_initialized(mask_klass)) {
2328 log_if_needed(" ** mask klass argument not initialized");
2329 return false;
2330 }
2331
2332 if (vmask_type->maybe_null()) {
2333 log_if_needed(" ** null mask values are not allowed for masked op");
2334 return false;
2335 }
2336 }
2337
2338 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2339 if (!elem_type->is_primitive_type()) {
2340 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2341 return false; // should be primitive type
2342 }
2343
2344 int num_elem = vlen->get_con();
2345 BasicType elem_bt = elem_type->basic_type();
2346 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
2347
2348 bool is_shift = VectorNode::is_shift_opcode(opc);
2349 bool is_rotate = VectorNode::is_rotate_opcode(opc);
2350
2351 if (opc == 0 || (!is_shift && !is_rotate)) {
2352 log_if_needed(" ** operation not supported: op=%d bt=%s", opr->get_con(), type2name(elem_bt));
2353 return false; // operation not supported
2354 }
2355
2356 int sopc = VectorNode::opcode(opc, elem_bt);
2357 if (sopc == 0) {
2358 log_if_needed(" ** operation not supported: opc=%s bt=%s", NodeClassNames[opc], type2name(elem_bt));
2359 return false; // operation not supported
2360 }
2361
2362 Node* cnt = argument(6);
2363 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2364 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2365 const TypeInt* cnt_type = cnt->bottom_type()->isa_int();
2366
2367 // If CPU supports vector constant rotate instructions pass it directly
2368 bool is_const_rotate = is_rotate && cnt_type && cnt_type->is_con() &&
2369 Matcher::supports_vector_constant_rotates(cnt_type->get_con());
2370 bool has_scalar_args = is_rotate ? !is_const_rotate : true;
2371
2372 VectorMaskUseType checkFlags = (VectorMaskUseType)(is_masked_op ? (VecMaskUseLoad | VecMaskUsePred) : VecMaskNotUsed);
2373 bool use_predicate = is_masked_op;
2374
2375 if (!arch_supports_vector(sopc, num_elem, elem_bt, checkFlags, has_scalar_args)) {
2376 use_predicate = false;
2377 if (!is_masked_op ||
2378 (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed, has_scalar_args) ||
2379 !arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad))) {
2380
2381 log_if_needed(" ** not supported: arity=0 op=int/%d vlen=%d etype=%s is_masked_op=%d",
2382 sopc, num_elem, type2name(elem_bt), is_masked_op ? 1 : 0);
2383 return false; // not supported
2384 }
2385 }
2386
2387 Node* opd1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
2388 Node* opd2 = nullptr;
2389 if (is_shift) {
2390 opd2 = vector_shift_count(cnt, opc, elem_bt, num_elem);
2391 } else {
2392 assert(is_rotate, "unexpected operation");
2393 if (!is_const_rotate) {
2394 const Type * type_bt = Type::get_const_basic_type(elem_bt);
2395 cnt = elem_bt == T_LONG ? gvn().transform(new ConvI2LNode(cnt)) : cnt;
2396 opd2 = gvn().transform(VectorNode::scalar2vector(cnt, num_elem, type_bt));
2397 } else {
2398 // Constant shift value.
2399 opd2 = cnt;
2400 }
2401 }
2402
2403 if (opd1 == nullptr || opd2 == nullptr) {
2404 return false;
2405 }
2406
2407 Node* mask = nullptr;
2408 if (is_masked_op) {
2409 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
2410 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
2411 mask = unbox_vector(argument(7), mbox_type, elem_bt, num_elem);
2412 if (mask == nullptr) {
2413 log_if_needed(" ** unbox failed mask=%s", NodeClassNames[argument(7)->Opcode()]);
2414 return false;
2415 }
2416 }
2417
2418 Node* operation = VectorNode::make(opc, opd1, opd2, num_elem, elem_bt);
2419 if (is_masked_op && mask != nullptr) {
2420 if (use_predicate) {
2421 operation->add_req(mask);
2422 operation->add_flag(Node::Flag_is_predicated_vector);
2423 } else {
2424 operation = gvn().transform(operation);
2425 operation = new VectorBlendNode(opd1, operation, mask);
2426 }
2427 }
2428 operation = gvn().transform(operation);
2429 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
2430 set_result(vbox);
2431 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2432 return true;
2433 }
2434
2435 // public static <VOUT extends VectorPayload,
2436 // VIN extends VectorPayload,
2437 // S extends VectorSpecies>
2438 // VOUT convert(int oprId,
2439 // Class<?> fromVectorClass, Class<?> fromElementType, int fromVLen,
2440 // Class<?> toVectorClass, Class<?> toElementType, int toVLen,
2441 // VIN v, S s,
2442 // VectorConvertOp<VOUT, VIN, S> defaultImpl)
2443 //
2444 bool LibraryCallKit::inline_vector_convert() {
2445 const TypeInt* opr = gvn().type(argument(0))->isa_int();
2446
2447 const TypeInstPtr* vector_klass_from = gvn().type(argument(1))->isa_instptr();
2448 const TypeInstPtr* elem_klass_from = gvn().type(argument(2))->isa_instptr();
2449 const TypeInt* vlen_from = gvn().type(argument(3))->isa_int();
2450
2451 const TypeInstPtr* vector_klass_to = gvn().type(argument(4))->isa_instptr();
2452 const TypeInstPtr* elem_klass_to = gvn().type(argument(5))->isa_instptr();
2453 const TypeInt* vlen_to = gvn().type(argument(6))->isa_int();
2454
2455 if (opr == nullptr ||
2456 vector_klass_from == nullptr || elem_klass_from == nullptr || vlen_from == nullptr ||
2457 vector_klass_to == nullptr || elem_klass_to == nullptr || vlen_to == nullptr) {
2458 return false; // dead code
2459 }
2460 if (!opr->is_con() ||
2461 vector_klass_from->const_oop() == nullptr || elem_klass_from->const_oop() == nullptr || !vlen_from->is_con() ||
2462 vector_klass_to->const_oop() == nullptr || elem_klass_to->const_oop() == nullptr || !vlen_to->is_con()) {
2463 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",
2464 NodeClassNames[argument(0)->Opcode()],
2465 NodeClassNames[argument(1)->Opcode()],
2466 NodeClassNames[argument(2)->Opcode()],
2467 NodeClassNames[argument(3)->Opcode()],
2468 NodeClassNames[argument(4)->Opcode()],
2469 NodeClassNames[argument(5)->Opcode()],
2470 NodeClassNames[argument(6)->Opcode()]);
2471 return false; // not enough info for intrinsification
2472 }
2473 if (!is_klass_initialized(vector_klass_from) || !is_klass_initialized(vector_klass_to)) {
2474 log_if_needed(" ** klass argument not initialized");
2475 return false;
2476 }
2477
2478 assert(opr->get_con() == VectorSupport::VECTOR_OP_CAST ||
2479 opr->get_con() == VectorSupport::VECTOR_OP_UCAST ||
2480 opr->get_con() == VectorSupport::VECTOR_OP_REINTERPRET, "wrong opcode");
2481 bool is_cast = (opr->get_con() == VectorSupport::VECTOR_OP_CAST || opr->get_con() == VectorSupport::VECTOR_OP_UCAST);
2482 bool is_ucast = (opr->get_con() == VectorSupport::VECTOR_OP_UCAST);
2483
2484 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
2485 ciKlass* vbox_klass_to = vector_klass_to->const_oop()->as_instance()->java_lang_Class_klass();
2486 if (is_vector_shuffle(vbox_klass_from)) {
2487 return false; // vector shuffles aren't supported
2488 }
2489 bool is_mask = is_vector_mask(vbox_klass_from);
2490
2491 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
2492 if (!elem_type_from->is_primitive_type()) {
2493 return false; // should be primitive type
2494 }
2495 BasicType elem_bt_from = elem_type_from->basic_type();
2496 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
2497 if (!elem_type_to->is_primitive_type()) {
2498 return false; // should be primitive type
2499 }
2500 BasicType elem_bt_to = elem_type_to->basic_type();
2501
2502 int num_elem_from = vlen_from->get_con();
2503 int num_elem_to = vlen_to->get_con();
2504
2505 // Check whether we can unbox to appropriate size. Even with casting, checking for reinterpret is needed
2506 // since we may need to change size.
2507 if (!arch_supports_vector(Op_VectorReinterpret,
2508 num_elem_from,
2509 elem_bt_from,
2510 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
2511 log_if_needed(" ** not supported: arity=1 op=%s/1 vlen1=%d etype1=%s ismask=%d",
2512 is_cast ? "cast" : "reinterpret",
2513 num_elem_from, type2name(elem_bt_from), is_mask);
2514 return false;
2515 }
2516
2517 // Check whether we can support resizing/reinterpreting to the new size.
2518 if (!arch_supports_vector(Op_VectorReinterpret,
2519 num_elem_to,
2520 elem_bt_to,
2521 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
2522 log_if_needed(" ** not supported: arity=1 op=%s/2 vlen2=%d etype2=%s ismask=%d",
2523 is_cast ? "cast" : "reinterpret",
2524 num_elem_to, type2name(elem_bt_to), is_mask);
2525 return false;
2526 }
2527
2528
2529 if (is_vector_shuffle(vbox_klass_to) &&
2530 (!arch_supports_vector(Op_SubVB, num_elem_to, elem_bt_to, VecMaskNotUsed) ||
2531 !arch_supports_vector(Op_VectorBlend, num_elem_to, elem_bt_to, VecMaskNotUsed) ||
2532 !arch_supports_vector(Op_VectorMaskCmp, num_elem_to, elem_bt_to, VecMaskNotUsed) ||
2533 !arch_supports_vector(Op_AndV, num_elem_to, elem_bt_to, VecMaskNotUsed) ||
2534 !arch_supports_vector(Op_Replicate, num_elem_to, elem_bt_to, VecMaskNotUsed))) {
2535 log_if_needed(" ** not supported: arity=1 op=shuffle_index_wrap vlen2=%d etype2=%s",
2536 num_elem_to, type2name(elem_bt_to));
2537 return false;
2538 }
2539
2540 // At this point, we know that both input and output vector registers are supported
2541 // by the architecture. Next check if the casted type is simply to same type - which means
2542 // that it is actually a resize and not a cast.
2543 if (is_cast && elem_bt_from == elem_bt_to) {
2544 is_cast = false;
2545 }
2546
2547 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
2548
2549 Node* opd1 = unbox_vector(argument(7), vbox_type_from, elem_bt_from, num_elem_from);
2550 if (opd1 == nullptr) {
2551 return false;
2552 }
2553
2554 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from, is_mask);
2555 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to, is_mask);
2556
2557 // Safety check to prevent casting if source mask is of type vector
2558 // and destination mask of type predicate vector and vice-versa.
2559 // From X86 standpoint, this case will only arise over KNL target,
2560 // where certain masks (depending on the species) are either propagated
2561 // through a vector or predicate register.
2562 if (is_mask &&
2563 ((src_type->isa_vectmask() == nullptr && dst_type->isa_vectmask()) ||
2564 (dst_type->isa_vectmask() == nullptr && src_type->isa_vectmask()))) {
2565 return false;
2566 }
2567
2568 Node* op = opd1;
2569 if (is_cast) {
2570 assert(!is_mask || num_elem_from == num_elem_to, "vector mask cast needs the same elem num");
2571 int cast_vopc = VectorCastNode::opcode(-1, elem_bt_from, !is_ucast);
2572
2573 // Make sure that vector cast is implemented to particular type/size combination if it is
2574 // not a mask casting.
2575 if (!is_mask && !arch_supports_vector(cast_vopc, num_elem_to, elem_bt_to, VecMaskNotUsed)) {
2576 log_if_needed(" ** not supported: arity=1 op=cast#%d/3 vlen2=%d etype2=%s ismask=%d",
2577 cast_vopc, num_elem_to, type2name(elem_bt_to), is_mask);
2578 return false;
2579 }
2580
2581 if (num_elem_from < num_elem_to) {
2582 // Since input and output number of elements are not consistent, we need to make sure we
2583 // properly size. Thus, first make a cast that retains the number of elements from source.
2584 int num_elem_for_cast = num_elem_from;
2585
2586 // It is possible that arch does not support this intermediate vector size
2587 // TODO More complex logic required here to handle this corner case for the sizes.
2588 if (!arch_supports_vector(cast_vopc, num_elem_for_cast, elem_bt_to, VecMaskNotUsed)) {
2589 log_if_needed(" ** not supported: arity=1 op=cast#%d/4 vlen1=%d etype2=%s ismask=%d",
2590 cast_vopc,
2591 num_elem_for_cast, type2name(elem_bt_to), is_mask);
2592 return false;
2593 }
2594
2595 op = gvn().transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_for_cast));
2596 // Now ensure that the destination gets properly resized to needed size.
2597 op = gvn().transform(new VectorReinterpretNode(op, op->bottom_type()->is_vect(), dst_type));
2598 } else if (num_elem_from > num_elem_to) {
2599 // Since number of elements from input is larger than output, simply reduce size of input
2600 // (we are supposed to drop top elements anyway).
2601 int num_elem_for_resize = num_elem_to;
2602
2603 // It is possible that arch does not support this intermediate vector size
2604 // TODO More complex logic required here to handle this corner case for the sizes.
2605 if (!arch_supports_vector(Op_VectorReinterpret,
2606 num_elem_for_resize,
2607 elem_bt_from,
2608 VecMaskNotUsed)) {
2609 log_if_needed(" ** not supported: arity=1 op=cast/5 vlen2=%d etype1=%s ismask=%d",
2610 num_elem_for_resize, type2name(elem_bt_from), is_mask);
2611 return false;
2612 }
2613
2614 const TypeVect* resize_type = TypeVect::make(elem_bt_from, num_elem_for_resize);
2615 op = gvn().transform(new VectorReinterpretNode(op, src_type, resize_type));
2616 op = gvn().transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
2617 } else { // num_elem_from == num_elem_to
2618 if (is_mask) {
2619 // Make sure that cast for vector mask is implemented to particular type/size combination.
2620 if (!arch_supports_vector(Op_VectorMaskCast, num_elem_to, elem_bt_to, VecMaskNotUsed)) {
2621 log_if_needed(" ** not supported: arity=1 op=maskcast vlen2=%d etype2=%s ismask=%d",
2622 num_elem_to, type2name(elem_bt_to), is_mask);
2623 return false;
2624 }
2625 op = gvn().transform(new VectorMaskCastNode(op, dst_type));
2626 } else {
2627 // Since input and output number of elements match, and since we know this vector size is
2628 // supported, simply do a cast with no resize needed.
2629 op = gvn().transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
2630 }
2631 }
2632 } else if (!Type::equals(src_type, dst_type)) {
2633 assert(!is_cast, "must be reinterpret");
2634 op = gvn().transform(new VectorReinterpretNode(op, src_type, dst_type));
2635 }
2636
2637 if (is_vector_shuffle(vbox_klass_to)) {
2638 op = partially_wrap_indexes(op, num_elem_to, elem_bt_to);
2639 }
2640
2641 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
2642 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
2643 set_result(vbox);
2644 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem_to * type2aelembytes(elem_bt_to))));
2645 return true;
2646 }
2647
2648 // public static
2649 // <V extends Vector<E>,
2650 // E>
2651 // V insert(Class<? extends V> vectorClass, Class<E> elementType, int vlen,
2652 // V vec, int ix, long val,
2653 // VecInsertOp<V> defaultImpl)
2654 bool LibraryCallKit::inline_vector_insert() {
2655 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2656 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
2657 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
2658 const TypeInt* idx = gvn().type(argument(4))->isa_int();
2659
2660 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr || idx == nullptr) {
2661 return false; // dead code
2662 }
2663 if (vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con() || !idx->is_con()) {
2664 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s idx=%s",
2665 NodeClassNames[argument(0)->Opcode()],
2666 NodeClassNames[argument(1)->Opcode()],
2667 NodeClassNames[argument(2)->Opcode()],
2668 NodeClassNames[argument(4)->Opcode()]);
2669 return false; // not enough info for intrinsification
2670 }
2671 if (!is_klass_initialized(vector_klass)) {
2672 log_if_needed(" ** klass argument not initialized");
2673 return false;
2674 }
2675 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2676 if (!elem_type->is_primitive_type()) {
2677 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2678 return false; // should be primitive type
2679 }
2680 BasicType elem_bt = elem_type->basic_type();
2681 int num_elem = vlen->get_con();
2682 if (!arch_supports_vector(Op_VectorInsert, num_elem, elem_bt, VecMaskNotUsed)) {
2683 log_if_needed(" ** not supported: arity=1 op=insert vlen=%d etype=%s ismask=no",
2684 num_elem, type2name(elem_bt));
2685 return false; // not supported
2686 }
2687
2688 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2689 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2690
2691 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2692 if (opd == nullptr) {
2693 return false;
2694 }
2695
2696 Node* insert_val = argument(5);
2697 assert(gvn().type(insert_val)->isa_long() != nullptr, "expected to be long");
2698
2699 // Convert insert value back to its appropriate type.
2700 switch (elem_bt) {
2701 case T_BYTE:
2702 insert_val = gvn().transform(new ConvL2INode(insert_val, TypeInt::BYTE));
2703 break;
2704 case T_SHORT:
2705 insert_val = gvn().transform(new ConvL2INode(insert_val, TypeInt::SHORT));
2706 break;
2707 case T_INT:
2708 insert_val = gvn().transform(new ConvL2INode(insert_val));
2709 break;
2710 case T_FLOAT:
2711 insert_val = gvn().transform(new ConvL2INode(insert_val));
2712 insert_val = gvn().transform(new MoveI2FNode(insert_val));
2713 break;
2714 case T_DOUBLE:
2715 insert_val = gvn().transform(new MoveL2DNode(insert_val));
2716 break;
2717 case T_LONG:
2718 // no conversion needed
2719 break;
2720 default: fatal("%s", type2name(elem_bt)); break;
2721 }
2722
2723 Node* operation = gvn().transform(VectorInsertNode::make(opd, insert_val, idx->get_con(), gvn()));
2724
2725 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
2726 set_result(vbox);
2727 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2728 return true;
2729 }
2730
2731 // public static
2732 // <VM extends VectorPayload,
2733 // E>
2734 // long extract(Class<? extends VM> vClass, Class<E> eClass,
2735 // int length,
2736 // VM vm, int i,
2737 // VecExtractOp<VM> defaultImpl)
2738 bool LibraryCallKit::inline_vector_extract() {
2739 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2740 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
2741 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
2742 const TypeInt* idx = gvn().type(argument(4))->isa_int();
2743
2744 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr || idx == nullptr) {
2745 return false; // dead code
2746 }
2747 if (vector_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
2748 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s",
2749 NodeClassNames[argument(0)->Opcode()],
2750 NodeClassNames[argument(1)->Opcode()],
2751 NodeClassNames[argument(2)->Opcode()]);
2752 return false; // not enough info for intrinsification
2753 }
2754 if (!is_klass_initialized(vector_klass)) {
2755 log_if_needed(" ** klass argument not initialized");
2756 return false;
2757 }
2758 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2759 if (!elem_type->is_primitive_type()) {
2760 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2761 return false; // should be primitive type
2762 }
2763 BasicType elem_bt = elem_type->basic_type();
2764 int num_elem = vlen->get_con();
2765
2766 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2767 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2768
2769 Node* opd = nullptr;
2770
2771 if (is_vector_mask(vbox_klass)) {
2772 // vbox_klass is mask. This is used for VectorMask.laneIsSet(int).
2773
2774 Node* pos = argument(4); // can be variable
2775 if (arch_supports_vector(Op_ExtractUB, num_elem, elem_bt, VecMaskUseAll)) {
2776 // Transform mask to vector with type of boolean and utilize ExtractUB node.
2777 opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2778 if (opd == nullptr) {
2779 return false;
2780 }
2781 opd = gvn().transform(VectorStoreMaskNode::make(gvn(), opd, elem_bt, num_elem));
2782 opd = gvn().transform(new ExtractUBNode(opd, pos));
2783 opd = gvn().transform(new ConvI2LNode(opd));
2784 } else if (arch_supports_vector(Op_VectorMaskToLong, num_elem, elem_bt, VecMaskUseLoad)) {
2785 opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2786 if (opd == nullptr) {
2787 return false;
2788 }
2789 // VectorMaskToLongNode requires the input is either a mask or a vector with BOOLEAN type.
2790 if (opd->bottom_type()->isa_vectmask() == nullptr) {
2791 opd = gvn().transform(VectorStoreMaskNode::make(gvn(), opd, elem_bt, num_elem));
2792 }
2793 // ((toLong() >>> pos) & 1L
2794 opd = gvn().transform(new VectorMaskToLongNode(opd, TypeLong::LONG));
2795 opd = gvn().transform(new URShiftLNode(opd, pos));
2796 opd = gvn().transform(new AndLNode(opd, gvn().makecon(TypeLong::ONE)));
2797 } else {
2798 log_if_needed(" ** Rejected mask extraction because architecture does not support it");
2799 return false; // not supported
2800 }
2801 } else {
2802 // vbox_klass is vector. This is used for Vector.lane(int).
2803 if (!idx->is_con()) {
2804 log_if_needed(" ** missing constant: idx=%s", NodeClassNames[argument(4)->Opcode()]);
2805 return false; // not enough info for intrinsification
2806 }
2807
2808 int vopc = ExtractNode::opcode(elem_bt);
2809 if (!arch_supports_vector(vopc, num_elem, elem_bt, VecMaskNotUsed)) {
2810 log_if_needed(" ** not supported: arity=1 op=extract vlen=%d etype=%s ismask=no",
2811 num_elem, type2name(elem_bt));
2812 return false; // not supported
2813 }
2814
2815 opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
2816 if (opd == nullptr) {
2817 return false;
2818 }
2819 ConINode* idx_con = gvn().intcon(idx->get_con())->as_ConI();
2820
2821 opd = gvn().transform(ExtractNode::make(opd, idx_con, elem_bt));
2822 switch (elem_bt) {
2823 case T_BYTE:
2824 case T_SHORT:
2825 case T_INT: {
2826 opd = gvn().transform(new ConvI2LNode(opd));
2827 break;
2828 }
2829 case T_FLOAT: {
2830 opd = gvn().transform(new MoveF2INode(opd));
2831 opd = gvn().transform(new ConvI2LNode(opd));
2832 break;
2833 }
2834 case T_DOUBLE: {
2835 opd = gvn().transform(new MoveD2LNode(opd));
2836 break;
2837 }
2838 case T_LONG: {
2839 // no conversion needed
2840 break;
2841 }
2842 default: fatal("%s", type2name(elem_bt));
2843 }
2844 }
2845 set_result(opd);
2846 return true;
2847 }
2848
2849 // public static
2850 // <V extends Vector<E>,
2851 // M extends VectorMask<E>,
2852 // E>
2853 // V compressExpandOp(int opr,
2854 // Class<? extends V> vClass, Class<? extends M> mClass, Class<E> eClass,
2855 // int length, V v, M m,
2856 // CompressExpandOperation<V, M> defaultImpl)
2857 bool LibraryCallKit::inline_vector_compress_expand() {
2858 const TypeInt* opr = gvn().type(argument(0))->isa_int();
2859 const TypeInstPtr* vector_klass = gvn().type(argument(1))->isa_instptr();
2860 const TypeInstPtr* mask_klass = gvn().type(argument(2))->isa_instptr();
2861 const TypeInstPtr* elem_klass = gvn().type(argument(3))->isa_instptr();
2862 const TypeInt* vlen = gvn().type(argument(4))->isa_int();
2863
2864 if (vector_klass == nullptr || elem_klass == nullptr || mask_klass == nullptr || vlen == nullptr ||
2865 vector_klass->const_oop() == nullptr || mask_klass->const_oop() == nullptr ||
2866 elem_klass->const_oop() == nullptr || !vlen->is_con() || !opr->is_con()) {
2867 log_if_needed(" ** missing constant: opr=%s vclass=%s mclass=%s etype=%s vlen=%s",
2868 NodeClassNames[argument(0)->Opcode()],
2869 NodeClassNames[argument(1)->Opcode()],
2870 NodeClassNames[argument(2)->Opcode()],
2871 NodeClassNames[argument(3)->Opcode()],
2872 NodeClassNames[argument(4)->Opcode()]);
2873 return false; // not enough info for intrinsification
2874 }
2875
2876 if (!is_klass_initialized(vector_klass) || !is_klass_initialized(mask_klass)) {
2877 log_if_needed(" ** klass argument not initialized");
2878 return false;
2879 }
2880
2881 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2882 if (!elem_type->is_primitive_type()) {
2883 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2884 return false; // should be primitive type
2885 }
2886
2887 int num_elem = vlen->get_con();
2888 BasicType elem_bt = elem_type->basic_type();
2889 int opc = VectorSupport::vop2ideal(opr->get_con(), elem_bt);
2890
2891 if (!arch_supports_vector(opc, num_elem, elem_bt, VecMaskUseLoad)) {
2892 log_if_needed(" ** not supported: opc=%d vlen=%d etype=%s ismask=useload",
2893 opc, num_elem, type2name(elem_bt));
2894 return false; // not supported
2895 }
2896
2897 Node* opd1 = nullptr;
2898 const TypeInstPtr* vbox_type = nullptr;
2899 if (opc != Op_CompressM) {
2900 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
2901 vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
2902 opd1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
2903 if (opd1 == nullptr) {
2904 log_if_needed(" ** unbox failed vector=%s",
2905 NodeClassNames[argument(5)->Opcode()]);
2906 return false;
2907 }
2908 }
2909
2910 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
2911 assert(is_vector_mask(mbox_klass), "argument(6) should be a mask class");
2912 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
2913
2914 Node* mask = unbox_vector(argument(6), mbox_type, elem_bt, num_elem);
2915 if (mask == nullptr) {
2916 log_if_needed(" ** unbox failed mask=%s",
2917 NodeClassNames[argument(6)->Opcode()]);
2918 return false;
2919 }
2920
2921 const TypeVect* vt = TypeVect::make(elem_bt, num_elem, opc == Op_CompressM);
2922 Node* operation = gvn().transform(VectorNode::make(opc, opd1, mask, vt));
2923
2924 // Wrap it up in VectorBox to keep object type information.
2925 const TypeInstPtr* box_type = opc == Op_CompressM ? mbox_type : vbox_type;
2926 Node* vbox = box_vector(operation, box_type, elem_bt, num_elem);
2927 set_result(vbox);
2928 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
2929 return true;
2930 }
2931
2932 // public static
2933 // <V extends Vector<E>,
2934 // E,
2935 // S extends VectorSpecies<E>>
2936 // V indexVector(Class<? extends V> vClass, Class<E> eClass,
2937 // int length,
2938 // V v, int step, S s,
2939 // IndexOperation<V, S> defaultImpl)
2940 bool LibraryCallKit::inline_index_vector() {
2941 const TypeInstPtr* vector_klass = gvn().type(argument(0))->isa_instptr();
2942 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
2943 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
2944
2945 if (vector_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
2946 vector_klass->const_oop() == nullptr || !vlen->is_con() ||
2947 elem_klass->const_oop() == nullptr) {
2948 log_if_needed(" ** missing constant: vclass=%s etype=%s vlen=%s",
2949 NodeClassNames[argument(0)->Opcode()],
2950 NodeClassNames[argument(1)->Opcode()],
2951 NodeClassNames[argument(2)->Opcode()]);
2952 return false; // not enough info for intrinsification
2953 }
2954
2955 if (!is_klass_initialized(vector_klass)) {
2956 log_if_needed(" ** klass argument not initialized");
2957 return false;
2958 }
2959
2960 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
2961 if (!elem_type->is_primitive_type()) {
2962 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
2963 return false; // should be primitive type
2964 }
2965
2966 int num_elem = vlen->get_con();
2967 BasicType elem_bt = elem_type->basic_type();
2968
2969 // Check whether the iota index generation op is supported by the current hardware
2970 if (!arch_supports_vector(Op_VectorLoadConst, num_elem, elem_bt, VecMaskNotUsed)) {
2971 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
2972 return false; // not supported
2973 }
2974
2975 int mul_op = VectorSupport::vop2ideal(VectorSupport::VECTOR_OP_MUL, elem_bt);
2976 int vmul_op = VectorNode::opcode(mul_op, elem_bt);
2977 bool needs_mul = true;
2978 Node* scale = argument(4);
2979 const TypeInt* scale_type = gvn().type(scale)->isa_int();
2980 // Multiply is not needed if the scale is a constant "1".
2981 if (scale_type && scale_type->is_con() && scale_type->get_con() == 1) {
2982 needs_mul = false;
2983 } else {
2984 // Check whether the vector multiply op is supported by the current hardware
2985 if (!arch_supports_vector(vmul_op, num_elem, elem_bt, VecMaskNotUsed)) {
2986 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
2987 return false; // not supported
2988 }
2989
2990 // Check whether the scalar cast op is supported by the current hardware
2991 if (is_floating_point_type(elem_bt) || elem_bt == T_LONG) {
2992 int cast_op = elem_bt == T_LONG ? Op_ConvI2L :
2993 elem_bt == T_FLOAT? Op_ConvI2F : Op_ConvI2D;
2994 if (!Matcher::match_rule_supported(cast_op)) {
2995 log_if_needed(" ** Rejected op (%s) because architecture does not support it",
2996 NodeClassNames[cast_op]);
2997 return false; // not supported
2998 }
2999 }
3000 }
3001
3002 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
3003 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
3004 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
3005 if (opd == nullptr) {
3006 log_if_needed(" ** unbox failed vector=%s",
3007 NodeClassNames[argument(3)->Opcode()]);
3008 return false;
3009 }
3010
3011 int add_op = VectorSupport::vop2ideal(VectorSupport::VECTOR_OP_ADD, elem_bt);
3012 int vadd_op = VectorNode::opcode(add_op, elem_bt);
3013 bool needs_add = true;
3014 // The addition is not needed if all the element values of "opd" are zero
3015 if (VectorNode::is_all_zeros_vector(opd)) {
3016 needs_add = false;
3017 } else {
3018 // Check whether the vector addition op is supported by the current hardware
3019 if (!arch_supports_vector(vadd_op, num_elem, elem_bt, VecMaskNotUsed)) {
3020 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
3021 return false; // not supported
3022 }
3023 }
3024
3025 // Compute the iota indice vector
3026 const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
3027 Node* index = gvn().transform(new VectorLoadConstNode(gvn().makecon(TypeInt::ZERO), vt));
3028
3029 // Broadcast the "scale" to a vector, and multiply the "scale" with iota indice vector.
3030 if (needs_mul) {
3031 switch (elem_bt) {
3032 case T_BOOLEAN: // fall-through
3033 case T_BYTE: // fall-through
3034 case T_SHORT: // fall-through
3035 case T_CHAR: // fall-through
3036 case T_INT: {
3037 // no conversion needed
3038 break;
3039 }
3040 case T_LONG: {
3041 scale = gvn().transform(new ConvI2LNode(scale));
3042 break;
3043 }
3044 case T_FLOAT: {
3045 scale = gvn().transform(new ConvI2FNode(scale));
3046 break;
3047 }
3048 case T_DOUBLE: {
3049 scale = gvn().transform(new ConvI2DNode(scale));
3050 break;
3051 }
3052 default: fatal("%s", type2name(elem_bt));
3053 }
3054 scale = gvn().transform(VectorNode::scalar2vector(scale, num_elem, Type::get_const_basic_type(elem_bt)));
3055 index = gvn().transform(VectorNode::make(vmul_op, index, scale, vt));
3056 }
3057
3058 // Add "opd" if addition is needed.
3059 if (needs_add) {
3060 index = gvn().transform(VectorNode::make(vadd_op, opd, index, vt));
3061 }
3062 Node* vbox = box_vector(index, vbox_type, elem_bt, num_elem);
3063 set_result(vbox);
3064 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
3065 return true;
3066 }
3067
3068 // public static
3069 // <E,
3070 // M extends VectorMask<E>>
3071 // M indexPartiallyInUpperRange(Class<? extends M> mClass, Class<E> eClass, int length,
3072 // long offset, long limit,
3073 // IndexPartiallyInUpperRangeOperation<E, M> defaultImpl)
3074 bool LibraryCallKit::inline_index_partially_in_upper_range() {
3075 const TypeInstPtr* mask_klass = gvn().type(argument(0))->isa_instptr();
3076 const TypeInstPtr* elem_klass = gvn().type(argument(1))->isa_instptr();
3077 const TypeInt* vlen = gvn().type(argument(2))->isa_int();
3078
3079 if (mask_klass == nullptr || elem_klass == nullptr || vlen == nullptr ||
3080 mask_klass->const_oop() == nullptr || elem_klass->const_oop() == nullptr || !vlen->is_con()) {
3081 log_if_needed(" ** missing constant: mclass=%s etype=%s vlen=%s",
3082 NodeClassNames[argument(0)->Opcode()],
3083 NodeClassNames[argument(1)->Opcode()],
3084 NodeClassNames[argument(2)->Opcode()]);
3085 return false; // not enough info for intrinsification
3086 }
3087
3088 if (!is_klass_initialized(mask_klass)) {
3089 log_if_needed(" ** klass argument not initialized");
3090 return false;
3091 }
3092
3093 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
3094 if (!elem_type->is_primitive_type()) {
3095 log_if_needed(" ** not a primitive bt=%d", elem_type->basic_type());
3096 return false; // should be primitive type
3097 }
3098
3099 int num_elem = vlen->get_con();
3100 BasicType elem_bt = elem_type->basic_type();
3101
3102 // Check whether the necessary ops are supported by current hardware.
3103 bool supports_mask_gen = arch_supports_vector(Op_VectorMaskGen, num_elem, elem_bt, VecMaskUseStore);
3104 if (!supports_mask_gen) {
3105 if (!arch_supports_vector(Op_VectorLoadConst, num_elem, elem_bt, VecMaskNotUsed) ||
3106 !arch_supports_vector(Op_Replicate, num_elem, elem_bt, VecMaskNotUsed) ||
3107 !arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
3108 log_if_needed(" ** not supported: vlen=%d etype=%s", num_elem, type2name(elem_bt));
3109 return false; // not supported
3110 }
3111
3112 // Check whether the scalar cast operation is supported by current hardware.
3113 if (elem_bt != T_LONG) {
3114 int cast_op = is_integral_type(elem_bt) ? Op_ConvL2I
3115 : (elem_bt == T_FLOAT ? Op_ConvL2F : Op_ConvL2D);
3116 if (!Matcher::match_rule_supported(cast_op)) {
3117 log_if_needed(" ** Rejected op (%s) because architecture does not support it",
3118 NodeClassNames[cast_op]);
3119 return false; // not supported
3120 }
3121 }
3122 }
3123
3124 Node* offset = argument(3);
3125 Node* limit = argument(5);
3126 if (offset == nullptr || limit == nullptr) {
3127 log_if_needed(" ** offset or limit argument is null");
3128 return false; // not supported
3129 }
3130
3131 ciKlass* box_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
3132 assert(is_vector_mask(box_klass), "argument(0) should be a mask class");
3133 const TypeInstPtr* box_type = TypeInstPtr::make_exact(TypePtr::NotNull, box_klass);
3134
3135 // We assume "offset > 0 && limit >= offset && limit - offset < num_elem".
3136 // So directly get indexLimit with "indexLimit = limit - offset".
3137 Node* indexLimit = gvn().transform(new SubLNode(limit, offset));
3138 Node* mask = nullptr;
3139 if (supports_mask_gen) {
3140 mask = gvn().transform(VectorMaskGenNode::make(indexLimit, elem_bt, num_elem));
3141 } else {
3142 // Generate the vector mask based on "mask = iota < indexLimit".
3143 // Broadcast "indexLimit" to a vector.
3144 switch (elem_bt) {
3145 case T_BOOLEAN: // fall-through
3146 case T_BYTE: // fall-through
3147 case T_SHORT: // fall-through
3148 case T_CHAR: // fall-through
3149 case T_INT: {
3150 indexLimit = gvn().transform(new ConvL2INode(indexLimit));
3151 break;
3152 }
3153 case T_DOUBLE: {
3154 indexLimit = gvn().transform(new ConvL2DNode(indexLimit));
3155 break;
3156 }
3157 case T_FLOAT: {
3158 indexLimit = gvn().transform(new ConvL2FNode(indexLimit));
3159 break;
3160 }
3161 case T_LONG: {
3162 // no conversion needed
3163 break;
3164 }
3165 default: fatal("%s", type2name(elem_bt));
3166 }
3167 indexLimit = gvn().transform(VectorNode::scalar2vector(indexLimit, num_elem, Type::get_const_basic_type(elem_bt)));
3168
3169 // Load the "iota" vector.
3170 const TypeVect* vt = TypeVect::make(elem_bt, num_elem);
3171 Node* iota = gvn().transform(new VectorLoadConstNode(gvn().makecon(TypeInt::ZERO), vt));
3172
3173 // Compute the vector mask with "mask = iota < indexLimit".
3174 ConINode* pred_node = (ConINode*)gvn().makecon(TypeInt::make(BoolTest::lt));
3175 const TypeVect* vmask_type = TypeVect::makemask(elem_bt, num_elem);
3176 mask = gvn().transform(new VectorMaskCmpNode(BoolTest::lt, iota, indexLimit, pred_node, vmask_type));
3177 }
3178 Node* vbox = box_vector(mask, box_type, elem_bt, num_elem);
3179 set_result(vbox);
3180 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
3181 return true;
3182 }
3183
3184 #undef non_product_log_if_needed
3185 #undef log_if_needed