1 /*
2 * Copyright (c) 2019, 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 #ifndef CPU_AARCH64_CONTINUATIONFREEZETHAW_AARCH64_INLINE_HPP
26 #define CPU_AARCH64_CONTINUATIONFREEZETHAW_AARCH64_INLINE_HPP
27
28 #include "code/codeBlob.inline.hpp"
29 #include "oops/stackChunkOop.inline.hpp"
30 #include "runtime/frame.hpp"
31 #include "runtime/frame.inline.hpp"
32
33
34 inline void patch_callee_link(const frame& f, intptr_t* fp) {
35 DEBUG_ONLY(intptr_t* orig = *ContinuationHelper::Frame::callee_link_address(f));
36 *ContinuationHelper::Frame::callee_link_address(f) = fp;
37 }
38
39 inline void patch_callee_link_relative(const frame& f, intptr_t* fp) {
40 intptr_t* la = (intptr_t*)ContinuationHelper::Frame::callee_link_address(f);
41 intptr_t new_value = fp - la;
42 *la = new_value;
43 }
44
45 ////// Freeze
46
47 // Fast path
48
49 inline void FreezeBase::patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp) {
50 // copy the spilled fp from the heap to the stack
51 *(frame_sp - frame::sender_sp_offset) = *(heap_sp - frame::sender_sp_offset);
52 }
53
54 // Slow path
55
56 template<typename FKind>
57 inline frame FreezeBase::sender(const frame& f) {
58 assert(FKind::is_instance(f), "");
59 if (FKind::interpreted) {
60 return frame(f.sender_sp(), f.interpreter_frame_sender_sp(), f.link(), f.sender_pc());
61 }
62 intptr_t** link_addr = link_address<FKind>(f);
63
64 intptr_t* sender_sp = (intptr_t*)(link_addr + frame::sender_sp_offset); // f.unextended_sp() + (fsize/wordSize); //
65 address sender_pc = ContinuationHelper::return_address_at(sender_sp - 1);
66 assert(sender_sp != f.sp(), "must have changed");
67
68 int slot = 0;
69 CodeBlob* sender_cb = CodeCache::find_blob_and_oopmap(sender_pc, slot);
70 return sender_cb != nullptr
71 ? frame(sender_sp, sender_sp, *link_addr, sender_pc, sender_cb,
72 slot == -1 ? nullptr : sender_cb->oop_map_for_slot(slot, sender_pc),
73 false /* on_heap ? */)
74 : frame(sender_sp, sender_sp, *link_addr, sender_pc);
75 }
76
77 template<typename FKind>
78 frame FreezeBase::new_heap_frame(frame& f, frame& caller) {
79 assert(FKind::is_instance(f), "");
80 assert(!caller.is_interpreted_frame()
81 || caller.unextended_sp() == (intptr_t*)caller.at(frame::interpreter_frame_last_sp_offset), "");
82
83 intptr_t *sp, *fp; // sp is really our unextended_sp
84 if (FKind::interpreted) {
85 assert((intptr_t*)f.at(frame::interpreter_frame_last_sp_offset) == nullptr
86 || f.unextended_sp() == (intptr_t*)f.at_relative(frame::interpreter_frame_last_sp_offset), "");
87 intptr_t locals_offset = *f.addr_at(frame::interpreter_frame_locals_offset);
88 // If the caller.is_empty(), i.e. we're freezing into an empty chunk, then we set
89 // the chunk's argsize in finalize_freeze and make room for it above the unextended_sp
90 bool overlap_caller = caller.is_interpreted_frame() || caller.is_empty();
91 fp = caller.unextended_sp() - 1 - locals_offset + (overlap_caller ? ContinuationHelper::InterpretedFrame::stack_argsize(f) : 0);
92 sp = fp - (f.fp() - f.unextended_sp());
93 assert(sp <= fp, "");
94 assert(fp <= caller.unextended_sp(), "");
95 caller.set_sp(fp + frame::sender_sp_offset);
96
97 assert(_cont.tail()->is_in_chunk(sp), "");
98
99 frame hf(sp, sp, fp, f.pc(), nullptr, nullptr, true /* on_heap */);
100 // copy relativized locals from the stack frame
101 *hf.addr_at(frame::interpreter_frame_locals_offset) = locals_offset;
102 return hf;
103 } else {
104 // We need to re-read fp out of the frame because it may be an oop and we might have
105 // had a safepoint in finalize_freeze, after constructing f.
106 fp = *(intptr_t**)(f.sp() - frame::sender_sp_offset);
107
108 int fsize = FKind::size(f);
109 sp = caller.unextended_sp() - fsize;
110 if (caller.is_interpreted_frame()) {
111 // If the caller is interpreted, our stackargs are not supposed to overlap with it
112 // so we make more room by moving sp down by argsize
113 int argsize = FKind::stack_argsize(f);
114 sp -= argsize;
115 }
116 caller.set_sp(sp + fsize);
117
118 assert(_cont.tail()->is_in_chunk(sp), "");
119
120 return frame(sp, sp, fp, f.pc(), nullptr, nullptr, true /* on_heap */);
121 }
122 }
123
124 void FreezeBase::adjust_interpreted_frame_unextended_sp(frame& f) {
125 assert((f.at(frame::interpreter_frame_last_sp_offset) != 0) || (f.unextended_sp() == f.sp()), "");
126 intptr_t* real_unextended_sp = (intptr_t*)f.at_relative_or_null(frame::interpreter_frame_last_sp_offset);
127 if (real_unextended_sp != nullptr) {
128 f.set_unextended_sp(real_unextended_sp); // can be null at a safepoint
129 }
130 }
131
132 inline void FreezeBase::prepare_freeze_interpreted_top_frame(const frame& f) {
133 assert(*f.addr_at(frame::interpreter_frame_last_sp_offset) == 0, "should be null for top frame");
134 intptr_t* lspp = f.addr_at(frame::interpreter_frame_last_sp_offset);
135 *lspp = f.unextended_sp() - f.fp();
136 }
137
138 inline void FreezeBase::relativize_interpreted_frame_metadata(const frame& f, const frame& hf) {
139 assert(hf.fp() == hf.unextended_sp() + (f.fp() - f.unextended_sp()), "");
140 assert((f.at(frame::interpreter_frame_last_sp_offset) != 0)
141 || (f.unextended_sp() == f.sp()), "");
142 assert(f.fp() > (intptr_t*)f.at_relative(frame::interpreter_frame_initial_sp_offset), "");
143
144 // on AARCH64, we may insert padding between the locals and the rest of the frame
145 // (see TemplateInterpreterGenerator::generate_normal_entry, and AbstractInterpreter::layout_activation)
146 // because we freeze the padding word (see recurse_freeze_interpreted_frame) in order to keep the same relativized
147 // locals value, we don't need to change the locals value here.
148
149 // Make sure that last_sp is already relativized.
150 assert((intptr_t*)hf.at_relative(frame::interpreter_frame_last_sp_offset) == hf.unextended_sp(), "");
151
152 // Make sure that monitor_block_top is already relativized.
153 assert(hf.at_absolute(frame::interpreter_frame_monitor_block_top_offset) <= frame::interpreter_frame_initial_sp_offset, "");
154
155 // extended_sp is already relativized by TemplateInterpreterGenerator::generate_normal_entry or
156 // AbstractInterpreter::layout_activation
157
158 // The interpreter native wrapper code adds space in the stack equal to size_of_parameters()
159 // after the fixed part of the frame. For wait0 this is equal to 3 words (this + long parameter).
160 // We adjust by this size since otherwise the saved last sp will be less than the extended_sp.
161 DEBUG_ONLY(Method* m = hf.interpreter_frame_method();)
162 DEBUG_ONLY(int extra_space = m->is_object_wait0() ? m->size_of_parameters() : 0;)
163
164 assert((hf.fp() - hf.unextended_sp()) == (f.fp() - f.unextended_sp()), "");
165 assert(hf.unextended_sp() == (intptr_t*)hf.at(frame::interpreter_frame_last_sp_offset), "");
166 assert(hf.unextended_sp() <= (intptr_t*)hf.at(frame::interpreter_frame_initial_sp_offset), "");
167 assert(hf.unextended_sp() + extra_space > (intptr_t*)hf.at(frame::interpreter_frame_extended_sp_offset), "");
168 assert(hf.fp() > (intptr_t*)hf.at(frame::interpreter_frame_initial_sp_offset), "");
169 assert(hf.fp() <= (intptr_t*)hf.at(frame::interpreter_frame_locals_offset), "");
170 }
171
172 inline void FreezeBase::set_top_frame_metadata_pd(const frame& hf) {
173 stackChunkOop chunk = _cont.tail();
174 assert(chunk->is_in_chunk(hf.sp() - 1), "");
175 assert(chunk->is_in_chunk(hf.sp() - frame::sender_sp_offset), "");
176
177 *(hf.sp() - 1) = (intptr_t)hf.pc();
178
179 intptr_t* fp_addr = hf.sp() - frame::sender_sp_offset;
180 *fp_addr = hf.is_interpreted_frame() ? (intptr_t)(hf.fp() - fp_addr)
181 : (intptr_t)hf.fp();
182 }
183
184 inline void FreezeBase::patch_pd(frame& hf, const frame& caller) {
185 if (caller.is_interpreted_frame()) {
186 assert(!caller.is_empty(), "");
187 patch_callee_link_relative(caller, caller.fp());
188 } else {
189 // If we're the bottom-most frame frozen in this freeze, the caller might have stayed frozen in the chunk,
190 // and its oop-containing fp fixed. We've now just overwritten it, so we must patch it back to its value
191 // as read from the chunk.
192 patch_callee_link(caller, caller.fp());
193 }
194 }
195
196 //////// Thaw
197
198 // Fast path
199
200 inline void ThawBase::prefetch_chunk_pd(void* start, int size) {
201 size <<= LogBytesPerWord;
202 Prefetch::read(start, size);
203 Prefetch::read(start, size - 64);
204 }
205
206 template <typename ConfigT>
207 inline void Thaw<ConfigT>::patch_caller_links(intptr_t* sp, intptr_t* bottom) {
208 // Fast path depends on !PreserveFramePointer. See can_thaw_fast().
209 assert(!PreserveFramePointer, "Frame pointers need to be fixed");
210 }
211
212 // Slow path
213
214 inline frame ThawBase::new_entry_frame() {
215 intptr_t* sp = _cont.entrySP();
216 return frame(sp, sp, _cont.entryFP(), _cont.entryPC()); // TODO PERF: This finds code blob and computes deopt state
217 }
218
219 template<typename FKind> frame ThawBase::new_stack_frame(const frame& hf, frame& caller, bool bottom) {
220 assert(FKind::is_instance(hf), "");
221 // The values in the returned frame object will be written into the callee's stack in patch.
222
223 if (FKind::interpreted) {
224 intptr_t* heap_sp = hf.unextended_sp();
225 // If caller is interpreted it already made room for the callee arguments
226 int overlap = caller.is_interpreted_frame() ? ContinuationHelper::InterpretedFrame::stack_argsize(hf) : 0;
227 const int fsize = (int)(ContinuationHelper::InterpretedFrame::frame_bottom(hf) - hf.unextended_sp() - overlap);
228 intptr_t* frame_sp = caller.unextended_sp() - fsize;
229 intptr_t* fp = frame_sp + (hf.fp() - heap_sp);
230 if ((intptr_t)fp % frame::frame_alignment != 0) {
231 fp--;
232 frame_sp--;
233 log_develop_trace(continuations)("Adding internal interpreted frame alignment");
234 }
235 DEBUG_ONLY(intptr_t* unextended_sp = fp + *hf.addr_at(frame::interpreter_frame_last_sp_offset);)
236 assert(frame_sp == unextended_sp, "");
237 caller.set_sp(fp + frame::sender_sp_offset);
238 frame f(frame_sp, frame_sp, fp, hf.pc());
239 // we need to set the locals so that the caller of new_stack_frame() can call
240 // ContinuationHelper::InterpretedFrame::frame_bottom
241 // copy relativized locals from the heap frame
242 *f.addr_at(frame::interpreter_frame_locals_offset) = *hf.addr_at(frame::interpreter_frame_locals_offset);
243 assert((intptr_t)f.fp() % frame::frame_alignment == 0, "");
244 return f;
245 } else {
246 int fsize = FKind::size(hf);
247 intptr_t* frame_sp = caller.unextended_sp() - fsize;
248 if (bottom || caller.is_interpreted_frame()) {
249 int argsize = FKind::stack_argsize(hf);
250
251 fsize += argsize;
252 frame_sp -= argsize;
253 caller.set_sp(caller.sp() - argsize);
254 assert(caller.sp() == frame_sp + (fsize-argsize), "");
255
256 frame_sp = align(hf, frame_sp, caller, bottom);
257 }
258
259 assert(hf.cb() != nullptr, "");
260 assert(hf.oop_map() != nullptr, "");
261 intptr_t* fp;
262 if (PreserveFramePointer) {
263 // we need to recreate a "real" frame pointer, pointing into the stack
264 fp = frame_sp + FKind::size(hf) - frame::sender_sp_offset;
265 } else {
266 fp = FKind::stub || FKind::native
267 ? frame_sp + fsize - frame::sender_sp_offset // fp always points to the address below the pushed return pc. We need correct address.
268 : *(intptr_t**)(hf.sp() - frame::sender_sp_offset); // we need to re-read fp because it may be an oop and we might have fixed the frame.
269 }
270 return frame(frame_sp, frame_sp, fp, hf.pc(), hf.cb(), hf.oop_map(), false); // TODO PERF : this computes deopt state; is it necessary?
271 }
272 }
273
274 inline intptr_t* ThawBase::align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom) {
275 #ifdef _LP64
276 if (((intptr_t)frame_sp & 0xf) != 0) {
277 assert(caller.is_interpreted_frame() || (bottom && hf.compiled_frame_stack_argsize() % 2 != 0), "");
278 frame_sp--;
279 caller.set_sp(caller.sp() - 1);
280 }
281 assert(is_aligned(frame_sp, frame::frame_alignment), "");
282 #endif
283
284 return frame_sp;
285 }
286
287 inline void ThawBase::patch_pd(frame& f, const frame& caller) {
288 patch_callee_link(caller, caller.fp());
289 }
290
291 inline void ThawBase::patch_pd(frame& f, intptr_t* caller_sp) {
292 intptr_t* fp = caller_sp - frame::sender_sp_offset;
293 patch_callee_link(f, fp);
294 }
295
296 inline void ThawBase::fix_native_wrapper_return_pc_pd(frame& top) {
297 // Nothing to do since the last pc saved before making the call to
298 // JVM_MonitorWait() was already set to the correct resume pc. Just
299 // do some sanity check.
300 #ifdef ASSERT
301 Method* method = top.is_interpreted_frame() ? top.interpreter_frame_method() : CodeCache::find_blob(top.pc())->as_nmethod()->method();
302 assert(method->is_object_wait0(), "");
303 #endif
304 }
305
306 inline intptr_t* ThawBase::push_resume_adapter(frame& top) {
307 intptr_t* sp = top.sp();
308 CodeBlob* cb = top.cb();
309
310 #ifdef ASSERT
311 RegisterMap map(JavaThread::current(),
312 RegisterMap::UpdateMap::skip,
313 RegisterMap::ProcessFrames::skip,
314 RegisterMap::WalkContinuation::skip);
315 frame caller = top.sender(&map);
316 intptr_t link_addr = (intptr_t)ContinuationHelper::Frame::callee_link_address(caller);
317 assert(sp[-2] == link_addr, "wrong link address: " INTPTR_FORMAT " != " INTPTR_FORMAT, sp[-2], link_addr);
318 #endif
319
320 if (top.is_interpreted_frame()) {
321 intptr_t* fp = sp - frame::sender_sp_offset;
322 address pc = Interpreter::cont_resume_interpreter_adapter();
323
324 sp -= frame::metadata_words;
325 *(address*)(sp - frame::sender_sp_ret_address_offset()) = pc;
326 *(intptr_t**)(sp - frame::sender_sp_offset) = fp;
327
328 log_develop_trace(continuations, preempt)("push_resume_adapter(): initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT
329 " fp: " INTPTR_FORMAT, p2i(sp + frame::metadata_words), p2i(sp), p2i(fp));
330 } else if (cb->frame_size() == 2) {
331 // C2 runtime stub case. For aarch64 the real size of the c2 runtime stub is 2 words bigger
332 // than what we think, i.e. size is 4. This is because the _last_Java_sp is not set to the
333 // sp right before making the call to the VM, but rather it is artificially set 2 words above
334 // this real sp so that we can store the return address at last_Java_sp[-1], and keep this
335 // property where we can retrieve the last_Java_pc from the last_Java_sp. But that means that
336 // once we return to the runtime stub, the code will adjust sp according to this real size.
337 // So we must adjust the frame size back here and we copy lr/rfp again.
338 sp -= 2;
339 sp[-2] = sp[0];
340 sp[-1] = sp[1];
341
342 log_develop_trace(continuations, preempt)("adjusted sp for c2 runtime stub, initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT
343 " fp: " INTPTR_FORMAT, p2i(sp + frame::metadata_words), p2i(sp), sp[-2]);
344 }
345 return sp;
346 }
347
348 inline intptr_t* ThawBase::push_cleanup_continuation() {
349 frame enterSpecial = new_entry_frame();
350 intptr_t* sp = enterSpecial.sp();
351
352 sp[-1] = (intptr_t)ContinuationEntry::cleanup_pc();
353 sp[-2] = (intptr_t)enterSpecial.fp();
354
355 log_develop_trace(continuations, preempt)("push_cleanup_continuation initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT, p2i(sp + 2 * frame::metadata_words), p2i(sp));
356 return sp;
357 }
358
359 inline void ThawBase::derelativize_interpreted_frame_metadata(const frame& hf, const frame& f) {
360 // Make sure that last_sp is kept relativized.
361 assert((intptr_t*)f.at_relative(frame::interpreter_frame_last_sp_offset) == f.unextended_sp(), "");
362
363 // Make sure that monitor_block_top is still relativized.
364 assert(f.at_absolute(frame::interpreter_frame_monitor_block_top_offset) <= frame::interpreter_frame_initial_sp_offset, "");
365
366 // Make sure that extended_sp is kept relativized.
367 DEBUG_ONLY(Method* m = hf.interpreter_frame_method();)
368 DEBUG_ONLY(int extra_space = m->is_object_wait0() ? m->size_of_parameters() : 0;) // see comment in relativize_interpreted_frame_metadata()
369 assert((intptr_t*)f.at_relative(frame::interpreter_frame_extended_sp_offset) < f.unextended_sp() + extra_space, "");
370 }
371
372 #endif // CPU_AARCH64_CONTINUATIONFREEZETHAW_AARCH64_INLINE_HPP