1 /*
2 * Copyright (c) 2020, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciSymbols.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "opto/castnode.hpp"
28 #include "opto/graphKit.hpp"
29 #include "opto/phaseX.hpp"
30 #include "opto/rootnode.hpp"
31 #include "opto/vector.hpp"
32 #include "utilities/macros.hpp"
33
34 static bool is_vector_mask(ciKlass* klass) {
35 return klass->is_subclass_of(ciEnv::current()->vector_VectorMask_klass());
36 }
37
38 void PhaseVector::optimize_vector_boxes() {
39 Compile::TracePhase tp(_t_vector_elimination);
40
41 // Signal GraphKit it's post-parse phase.
42 assert(C->inlining_incrementally() == false, "sanity");
43 C->set_inlining_incrementally(true);
44
45 C->igvn_worklist()->ensure_empty(); // should be done with igvn
46
47 expand_vunbox_nodes();
48 scalarize_vbox_nodes();
49
50 C->inline_vector_reboxing_calls();
51
52 expand_vbox_nodes();
53 eliminate_vbox_alloc_nodes();
54
55 C->set_inlining_incrementally(false);
56
57 do_cleanup();
58 }
59
60 void PhaseVector::do_cleanup() {
61 if (C->failing()) return;
62 {
63 Compile::TracePhase tp(_t_vector_pru);
64 ResourceMark rm;
65 PhaseRemoveUseless pru(C->initial_gvn(), *C->igvn_worklist());
66 if (C->failing()) return;
67 }
68 {
69 Compile::TracePhase tp(_t_vector_igvn);
70 _igvn.reset_from_gvn(C->initial_gvn());
71 _igvn.optimize();
72 if (C->failing()) return;
73 }
74 C->print_method(PHASE_ITER_GVN_BEFORE_EA, 3);
75 }
76
77 void PhaseVector::scalarize_vbox_nodes() {
78 if (C->failing()) return;
79
80 if (!EnableVectorReboxing) {
81 return; // don't scalarize vector boxes
82 }
83
84 int macro_idx = C->macro_count() - 1;
85 while (macro_idx >= 0) {
86 Node * n = C->macro_node(macro_idx);
87 assert(n->is_macro(), "only macro nodes expected here");
88 if (n->Opcode() == Op_VectorBox) {
89 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(n);
90 scalarize_vbox_node(vbox);
91 if (C->failing()) return;
92 C->print_method(PHASE_SCALARIZE_VBOX, 3, vbox);
93 }
94 if (C->failing()) return;
95 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
96 }
97 }
98
99 void PhaseVector::expand_vbox_nodes() {
100 if (C->failing()) return;
101
102 int macro_idx = C->macro_count() - 1;
103 while (macro_idx >= 0) {
104 Node * n = C->macro_node(macro_idx);
105 assert(n->is_macro(), "only macro nodes expected here");
106 if (n->Opcode() == Op_VectorBox) {
107 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(n);
108 expand_vbox_node(vbox);
109 if (C->failing()) return;
110 }
111 if (C->failing()) return;
112 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
113 }
114 }
115
116 void PhaseVector::expand_vunbox_nodes() {
117 if (C->failing()) return;
118
119 int macro_idx = C->macro_count() - 1;
120 while (macro_idx >= 0) {
121 Node * n = C->macro_node(macro_idx);
122 assert(n->is_macro(), "only macro nodes expected here");
123 if (n->Opcode() == Op_VectorUnbox) {
124 VectorUnboxNode* vec_unbox = static_cast<VectorUnboxNode*>(n);
125 expand_vunbox_node(vec_unbox);
126 if (C->failing()) return;
127 C->print_method(PHASE_EXPAND_VUNBOX, 3, vec_unbox);
128 }
129 if (C->failing()) return;
130 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
131 }
132 }
133
134 void PhaseVector::eliminate_vbox_alloc_nodes() {
135 if (C->failing()) return;
136
137 int macro_idx = C->macro_count() - 1;
138 while (macro_idx >= 0) {
139 Node * n = C->macro_node(macro_idx);
140 assert(n->is_macro(), "only macro nodes expected here");
141 if (n->Opcode() == Op_VectorBoxAllocate) {
142 VectorBoxAllocateNode* vbox_alloc = static_cast<VectorBoxAllocateNode*>(n);
143 eliminate_vbox_alloc_node(vbox_alloc);
144 if (C->failing()) return;
145 C->print_method(PHASE_ELIMINATE_VBOX_ALLOC, 3, vbox_alloc);
146 }
147 if (C->failing()) return;
148 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
149 }
150 }
151
152 static JVMState* clone_jvms(Compile* C, SafePointNode* sfpt) {
153 JVMState* new_jvms = sfpt->jvms()->clone_shallow(C);
154 uint size = sfpt->req();
155 SafePointNode* map = new SafePointNode(size, new_jvms);
156 for (uint i = 0; i < size; i++) {
157 map->init_req(i, sfpt->in(i));
158 }
159 Node* mem = map->memory();
160 if (!mem->is_MergeMem()) {
161 // Since we are not in parsing, the SafePointNode does not guarantee that the memory
162 // input is necessarily a MergeMemNode. But we need to ensure that there is that
163 // MergeMemNode, since the GraphKit assumes the memory input of the map to be a
164 // MergeMemNode, so that it can directly access the memory slices.
165 PhaseGVN& gvn = *C->initial_gvn();
166 Node* mergemem = MergeMemNode::make(mem);
167 gvn.set_type_bottom(mergemem);
168 map->set_memory(mergemem);
169 }
170 new_jvms->set_map(map);
171 return new_jvms;
172 }
173
174 void PhaseVector::scalarize_vbox_node(VectorBoxNode* vec_box) {
175 Node* vec_value = vec_box->in(VectorBoxNode::Value);
176 PhaseGVN& gvn = *C->initial_gvn();
177
178 // Process merged VBAs
179
180 if (EnableVectorAggressiveReboxing) {
181 Unique_Node_List calls;
182 for (DUIterator_Fast imax, i = vec_box->fast_outs(imax); i < imax; i++) {
183 Node* use = vec_box->fast_out(i);
184 if (use->is_CallJava()) {
185 CallJavaNode* call = use->as_CallJava();
186 if (call->has_non_debug_use(vec_box) && vec_box->in(VectorBoxNode::Box)->is_Phi()) {
187 calls.push(call);
188 }
189 }
190 }
191
192 while (calls.size() > 0) {
193 CallJavaNode* call = calls.pop()->as_CallJava();
194 // Attach new VBA to the call and use it instead of Phi (VBA ... VBA).
195
196 JVMState* jvms = clone_jvms(C, call);
197 GraphKit kit(jvms);
198 PhaseGVN& gvn = kit.gvn();
199
200 // Adjust JVMS from post-call to pre-call state: put args on stack
201 uint nargs = call->method()->arg_size();
202 kit.ensure_stack(kit.sp() + nargs);
203 for (uint i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
204 kit.push(call->in(i));
205 }
206 jvms = kit.sync_jvms();
207
208 Node* new_vbox = nullptr;
209 {
210 Node* vect = vec_box->in(VectorBoxNode::Value);
211 const TypeInstPtr* vbox_type = vec_box->box_type();
212 const TypeVect* vt = vec_box->vec_type();
213 BasicType elem_bt = vt->element_basic_type();
214 int num_elem = vt->length();
215
216 new_vbox = kit.box_vector(vect, vbox_type, elem_bt, num_elem, /*deoptimize=*/true);
217
218 kit.replace_in_map(vec_box, new_vbox);
219 }
220
221 kit.dec_sp(nargs);
222 jvms = kit.sync_jvms();
223
224 call->set_req(TypeFunc::Control , kit.control());
225 call->set_req(TypeFunc::I_O , kit.i_o());
226 call->set_req(TypeFunc::Memory , kit.reset_memory());
227 call->set_req(TypeFunc::FramePtr, kit.frameptr());
228 call->replace_edge(vec_box, new_vbox);
229
230 C->record_for_igvn(call);
231 }
232 }
233
234 // Process debug uses at safepoints
235 Unique_Node_List safepoints;
236
237 Unique_Node_List worklist;
238 worklist.push(vec_box);
239 while (worklist.size() > 0) {
240 Node* n = worklist.pop();
241 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
242 Node* use = n->fast_out(i);
243 if (use->is_SafePoint()) {
244 SafePointNode* sfpt = use->as_SafePoint();
245 if (!sfpt->is_Call() || !sfpt->as_Call()->has_non_debug_use(n)) {
246 safepoints.push(sfpt);
247 }
248 } else if (use->is_ConstraintCast()) {
249 worklist.push(use); // reversed version of Node::uncast()
250 }
251 }
252 }
253
254 ciInstanceKlass* iklass = vec_box->box_type()->instance_klass();
255 int n_fields = iklass->nof_nonstatic_fields();
256 assert(n_fields == 1, "sanity");
257
258 // If a mask is feeding into safepoint[s], then its value should be
259 // packed into a boolean/byte vector first, this will simplify the
260 // re-materialization logic for both predicated and non-predicated
261 // targets.
262 bool is_mask = is_vector_mask(iklass);
263 if (is_mask && vec_value->Opcode() != Op_VectorStoreMask) {
264 const TypeVect* vt = vec_value->bottom_type()->is_vect();
265 BasicType bt = vt->element_basic_type();
266 vec_value = gvn.transform(VectorStoreMaskNode::make(gvn, vec_value, bt, vt->length()));
267 }
268
269 while (safepoints.size() > 0) {
270 SafePointNode* sfpt = safepoints.pop()->as_SafePoint();
271
272 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
273 Node* sobj = new SafePointScalarObjectNode(vec_box->box_type(), vec_box, first_ind, sfpt->jvms()->depth(), n_fields);
274 sobj->init_req(0, C->root());
275 sfpt->add_req(vec_value);
276
277 sobj = gvn.transform(sobj);
278
279 JVMState *jvms = sfpt->jvms();
280
281 jvms->set_endoff(sfpt->req());
282 // Now make a pass over the debug information replacing any references
283 // to the allocated object with vector value.
284 for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
285 Node* debug = sfpt->in(i);
286 if (debug != nullptr && debug->uncast(/*keep_deps*/false) == vec_box) {
287 sfpt->set_req(i, sobj);
288 }
289 }
290 C->record_for_igvn(sfpt);
291 }
292 }
293
294 void PhaseVector::expand_vbox_node(VectorBoxNode* vec_box) {
295 if (vec_box->outcnt() > 0) {
296 VectorSet visited;
297 Node* vbox = vec_box->in(VectorBoxNode::Box);
298 Node* vect = vec_box->in(VectorBoxNode::Value);
299 Node* result = expand_vbox_node_helper(vbox, vect, vec_box->box_type(),
300 vec_box->vec_type(), visited);
301 C->gvn_replace_by(vec_box, result);
302 C->print_method(PHASE_EXPAND_VBOX, 3, vec_box);
303 }
304 C->remove_macro_node(vec_box);
305 }
306
307 Node* PhaseVector::expand_vbox_node_helper(Node* vbox,
308 Node* vect,
309 const TypeInstPtr* box_type,
310 const TypeVect* vect_type,
311 VectorSet &visited) {
312 // JDK-8304948 shows an example that there may be a cycle in the graph.
313 if (visited.test_set(vbox->_idx)) {
314 assert(vbox->is_Phi(), "should be phi");
315 return vbox; // already visited
316 }
317
318 // Handle the case when the allocation input to VectorBoxNode is a Proj.
319 // This is the normal case before expanding.
320 if (vbox->is_Proj() && vbox->in(0)->Opcode() == Op_VectorBoxAllocate) {
321 VectorBoxAllocateNode* vbox_alloc = static_cast<VectorBoxAllocateNode*>(vbox->in(0));
322 return expand_vbox_alloc_node(vbox_alloc, vect, box_type, vect_type);
323 }
324
325 // Handle the case when both the allocation input and vector input to
326 // VectorBoxNode are Phi. This case is generated after the transformation of
327 // Phi: Phi (VectorBox1 VectorBox2) => VectorBox (Phi1 Phi2).
328 // With this optimization, the relative two allocation inputs of VectorBox1 and
329 // VectorBox2 are gathered into Phi1 now. Similarly, the original vector
330 // inputs of two VectorBox nodes are in Phi2.
331 //
332 // See PhiNode::merge_through_phi in cfg.cpp for more details.
333 if (vbox->is_Phi() && vect->is_Phi()) {
334 assert(vbox->as_Phi()->region() == vect->as_Phi()->region(), "");
335 for (uint i = 1; i < vbox->req(); i++) {
336 Node* new_box = expand_vbox_node_helper(vbox->in(i), vect->in(i),
337 box_type, vect_type, visited);
338 if (!new_box->is_Phi()) {
339 C->initial_gvn()->hash_delete(vbox);
340 vbox->set_req(i, new_box);
341 }
342 }
343 return C->initial_gvn()->transform(vbox);
344 }
345
346 // Handle the case when the allocation input to VectorBoxNode is a phi
347 // but the vector input is not, which can definitely be the case if the
348 // vector input has been value-numbered. It seems to be safe to do by
349 // construction because VectorBoxNode and VectorBoxAllocate come in a
350 // specific order as a result of expanding an intrinsic call. After that, if
351 // any of the inputs to VectorBoxNode are value-numbered they can only
352 // move up and are guaranteed to dominate.
353 if (vbox->is_Phi() && (vect->is_Vector() || vect->is_LoadVector())) {
354 for (uint i = 1; i < vbox->req(); i++) {
355 Node* new_box = expand_vbox_node_helper(vbox->in(i), vect,
356 box_type, vect_type, visited);
357 if (!new_box->is_Phi()) {
358 C->initial_gvn()->hash_delete(vbox);
359 vbox->set_req(i, new_box);
360 }
361 }
362 return C->initial_gvn()->transform(vbox);
363 }
364
365 assert(!vbox->is_Phi(), "should be expanded");
366 // TODO: assert that expanded vbox is initialized with the same value (vect).
367 return vbox; // already expanded
368 }
369
370 Node* PhaseVector::expand_vbox_alloc_node(VectorBoxAllocateNode* vbox_alloc,
371 Node* value,
372 const TypeInstPtr* box_type,
373 const TypeVect* vect_type) {
374 JVMState* jvms = clone_jvms(C, vbox_alloc);
375 GraphKit kit(jvms);
376 PhaseGVN& gvn = kit.gvn();
377
378 ciInstanceKlass* box_klass = box_type->instance_klass();
379 BasicType bt = vect_type->element_basic_type();
380 int num_elem = vect_type->length();
381
382 bool is_mask = is_vector_mask(box_klass);
383 // If boxed mask value is present in a predicate register, it must be
384 // spilled to a vector though a VectorStoreMaskOperation before actual StoreVector
385 // operation to vector payload field.
386 if (is_mask && (value->bottom_type()->isa_vectmask() || bt != T_BOOLEAN)) {
387 value = gvn.transform(VectorStoreMaskNode::make(gvn, value, bt, num_elem));
388 // Although type of mask depends on its definition, in terms of storage everything is stored in boolean array.
389 bt = T_BOOLEAN;
390 assert(value->bottom_type()->is_vect()->element_basic_type() == bt,
391 "must be consistent with mask representation");
392 }
393
394 // Generate array allocation for the field which holds the values.
395 const TypeKlassPtr* array_klass = TypeKlassPtr::make(ciTypeArrayKlass::make(bt));
396 Node* arr = kit.new_array(kit.makecon(array_klass), kit.intcon(num_elem), 1);
397
398 // Store the vector value into the array.
399 // (The store should be captured by InitializeNode and turned into initialized store later.)
400 Node* arr_adr = kit.array_element_address(arr, kit.intcon(0), bt);
401 const TypePtr* arr_adr_type = arr_adr->bottom_type()->is_ptr();
402 Node* arr_mem = kit.memory(arr_adr);
403 Node* vstore = gvn.transform(StoreVectorNode::make(0,
404 kit.control(),
405 arr_mem,
406 arr_adr,
407 arr_adr_type,
408 value,
409 num_elem));
410 kit.set_memory(vstore, arr_adr_type);
411
412 C->set_max_vector_size(MAX2(C->max_vector_size(), vect_type->length_in_bytes()));
413
414 // Generate the allocate for the Vector object.
415 const TypeKlassPtr* klass_type = box_type->as_klass_type();
416 Node* klass_node = kit.makecon(klass_type);
417 Node* vec_obj = kit.new_instance(klass_node);
418
419 // Store the allocated array into object.
420 ciField* field = ciEnv::current()->vector_VectorPayload_klass()->get_field_by_name(ciSymbols::payload_name(),
421 ciSymbols::object_signature(),
422 false);
423 assert(field != nullptr, "");
424 Node* vec_field = kit.basic_plus_adr(vec_obj, field->offset_in_bytes());
425 const TypePtr* vec_adr_type = vec_field->bottom_type()->is_ptr();
426
427 // The store should be captured by InitializeNode and turned into initialized store later.
428 Node* field_store = gvn.transform(kit.access_store_at(vec_obj,
429 vec_field,
430 vec_adr_type,
431 arr,
432 TypeOopPtr::make_from_klass(field->type()->as_klass()),
433 T_OBJECT,
434 IN_HEAP));
435 kit.set_memory(field_store, vec_adr_type);
436
437 kit.replace_call(vbox_alloc, vec_obj, true);
438 C->remove_macro_node(vbox_alloc);
439
440 return vec_obj;
441 }
442
443 void PhaseVector::expand_vunbox_node(VectorUnboxNode* vec_unbox) {
444 if (vec_unbox->outcnt() > 0) {
445 GraphKit kit;
446 PhaseGVN& gvn = kit.gvn();
447
448 Node* obj = vec_unbox->obj();
449 const TypeInstPtr* tinst = gvn.type(obj)->isa_instptr();
450 ciInstanceKlass* from_kls = tinst->instance_klass();
451 const TypeVect* vt = vec_unbox->bottom_type()->is_vect();
452 BasicType bt = vt->element_basic_type();
453 BasicType masktype = bt;
454
455 if (is_vector_mask(from_kls)) {
456 bt = T_BOOLEAN;
457 }
458
459 ciField* field = ciEnv::current()->vector_VectorPayload_klass()->get_field_by_name(ciSymbols::payload_name(),
460 ciSymbols::object_signature(),
461 false);
462 assert(field != nullptr, "");
463 int offset = field->offset_in_bytes();
464 Node* vec_adr = kit.basic_plus_adr(obj, offset);
465
466 Node* mem = vec_unbox->mem();
467 Node* ctrl = vec_unbox->in(0);
468 Node* vec_field_ld;
469 {
470 DecoratorSet decorators = MO_UNORDERED | IN_HEAP;
471 C2AccessValuePtr addr(vec_adr, vec_adr->bottom_type()->is_ptr());
472 MergeMemNode* local_mem = MergeMemNode::make(mem);
473 gvn.record_for_igvn(local_mem);
474 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
475 C2OptAccess access(gvn, ctrl, local_mem, decorators, T_OBJECT, obj, addr);
476 const Type* type = TypeOopPtr::make_from_klass(field->type()->as_klass());
477 vec_field_ld = bs->load_at(access, type);
478 }
479
480 // For proper aliasing, attach concrete payload type.
481 ciKlass* payload_klass = ciTypeArrayKlass::make(bt);
482 const Type* payload_type = TypeAryPtr::make_from_klass(payload_klass)->cast_to_ptr_type(TypePtr::NotNull);
483 vec_field_ld = gvn.transform(new CastPPNode(nullptr, vec_field_ld, payload_type));
484
485 Node* adr = kit.array_element_address(vec_field_ld, gvn.intcon(0), bt);
486 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
487 int num_elem = vt->length();
488 Node* vec_val_load = LoadVectorNode::make(0,
489 ctrl,
490 mem,
491 adr,
492 adr_type,
493 num_elem,
494 bt);
495 vec_val_load = gvn.transform(vec_val_load);
496
497 C->set_max_vector_size(MAX2(C->max_vector_size(), vt->length_in_bytes()));
498
499 if (is_vector_mask(from_kls)) {
500 vec_val_load = gvn.transform(new VectorLoadMaskNode(vec_val_load, TypeVect::makemask(masktype, num_elem)));
501 }
502
503 gvn.hash_delete(vec_unbox);
504 vec_unbox->disconnect_inputs(C);
505 C->gvn_replace_by(vec_unbox, vec_val_load);
506 }
507 C->remove_macro_node(vec_unbox);
508 }
509
510 void PhaseVector::eliminate_vbox_alloc_node(VectorBoxAllocateNode* vbox_alloc) {
511 JVMState* jvms = clone_jvms(C, vbox_alloc);
512 GraphKit kit(jvms);
513 // Remove VBA, but leave a safepoint behind.
514 // Otherwise, it may end up with a loop without any safepoint polls.
515 kit.replace_call(vbox_alloc, kit.map(), true);
516 C->remove_macro_node(vbox_alloc);
517 }