1 /* 2 * Copyright (c) 2019, 2023, 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 assert((hf.fp() - hf.unextended_sp()) == (f.fp() - f.unextended_sp()), ""); 159 assert(hf.unextended_sp() == (intptr_t*)hf.at(frame::interpreter_frame_last_sp_offset), ""); 160 assert(hf.unextended_sp() <= (intptr_t*)hf.at(frame::interpreter_frame_initial_sp_offset), ""); 161 assert(hf.unextended_sp() > (intptr_t*)hf.at(frame::interpreter_frame_extended_sp_offset), ""); 162 assert(hf.fp() > (intptr_t*)hf.at(frame::interpreter_frame_initial_sp_offset), ""); 163 assert(hf.fp() <= (intptr_t*)hf.at(frame::interpreter_frame_locals_offset), ""); 164 } 165 166 inline void FreezeBase::set_top_frame_metadata_pd(const frame& hf) { 167 stackChunkOop chunk = _cont.tail(); 168 assert(chunk->is_in_chunk(hf.sp() - 1), ""); 169 assert(chunk->is_in_chunk(hf.sp() - frame::sender_sp_offset), ""); 170 171 *(hf.sp() - 1) = (intptr_t)hf.pc(); 172 173 intptr_t* fp_addr = hf.sp() - frame::sender_sp_offset; 174 *fp_addr = hf.is_interpreted_frame() ? (intptr_t)(hf.fp() - fp_addr) 175 : (intptr_t)hf.fp(); 176 } 177 178 inline void FreezeBase::patch_pd(frame& hf, const frame& caller) { 179 if (caller.is_interpreted_frame()) { 180 assert(!caller.is_empty(), ""); 181 patch_callee_link_relative(caller, caller.fp()); 182 } else { 183 // If we're the bottom-most frame frozen in this freeze, the caller might have stayed frozen in the chunk, 184 // and its oop-containing fp fixed. We've now just overwritten it, so we must patch it back to its value 185 // as read from the chunk. 186 patch_callee_link(caller, caller.fp()); 187 } 188 } 189 190 //////// Thaw 191 192 // Fast path 193 194 inline void ThawBase::prefetch_chunk_pd(void* start, int size) { 195 size <<= LogBytesPerWord; 196 Prefetch::read(start, size); 197 Prefetch::read(start, size - 64); 198 } 199 200 template <typename ConfigT> 201 inline void Thaw<ConfigT>::patch_caller_links(intptr_t* sp, intptr_t* bottom) { 202 // Fast path depends on !PreserveFramePointer. See can_thaw_fast(). 203 assert(!PreserveFramePointer, "Frame pointers need to be fixed"); 204 } 205 206 // Slow path 207 208 inline frame ThawBase::new_entry_frame() { 209 intptr_t* sp = _cont.entrySP(); 210 return frame(sp, sp, _cont.entryFP(), _cont.entryPC()); // TODO PERF: This finds code blob and computes deopt state 211 } 212 213 template<typename FKind> frame ThawBase::new_stack_frame(const frame& hf, frame& caller, bool bottom) { 214 assert(FKind::is_instance(hf), ""); 215 // The values in the returned frame object will be written into the callee's stack in patch. 216 217 if (FKind::interpreted) { 218 intptr_t* heap_sp = hf.unextended_sp(); 219 // If caller is interpreted it already made room for the callee arguments 220 int overlap = caller.is_interpreted_frame() ? ContinuationHelper::InterpretedFrame::stack_argsize(hf) : 0; 221 const int fsize = (int)(ContinuationHelper::InterpretedFrame::frame_bottom(hf) - hf.unextended_sp() - overlap); 222 const int locals = hf.interpreter_frame_method()->max_locals(); 223 intptr_t* frame_sp = caller.unextended_sp() - fsize; 224 intptr_t* fp = frame_sp + (hf.fp() - heap_sp); 225 if ((intptr_t)fp % frame::frame_alignment != 0) { 226 fp--; 227 frame_sp--; 228 log_develop_trace(continuations)("Adding internal interpreted frame alignment"); 229 } 230 DEBUG_ONLY(intptr_t* unextended_sp = fp + *hf.addr_at(frame::interpreter_frame_last_sp_offset);) 231 assert(frame_sp == unextended_sp, ""); 232 caller.set_sp(fp + frame::sender_sp_offset); 233 frame f(frame_sp, frame_sp, fp, hf.pc()); 234 // we need to set the locals so that the caller of new_stack_frame() can call 235 // ContinuationHelper::InterpretedFrame::frame_bottom 236 // copy relativized locals from the heap frame 237 *f.addr_at(frame::interpreter_frame_locals_offset) = *hf.addr_at(frame::interpreter_frame_locals_offset); 238 assert((intptr_t)f.fp() % frame::frame_alignment == 0, ""); 239 return f; 240 } else { 241 int fsize = FKind::size(hf); 242 intptr_t* frame_sp = caller.unextended_sp() - fsize; 243 if (bottom || caller.is_interpreted_frame()) { 244 int argsize = FKind::stack_argsize(hf); 245 246 fsize += argsize; 247 frame_sp -= argsize; 248 caller.set_sp(caller.sp() - argsize); 249 assert(caller.sp() == frame_sp + (fsize-argsize), ""); 250 251 frame_sp = align(hf, frame_sp, caller, bottom); 252 } 253 254 assert(hf.cb() != nullptr, ""); 255 assert(hf.oop_map() != nullptr, ""); 256 intptr_t* fp; 257 if (PreserveFramePointer) { 258 // we need to recreate a "real" frame pointer, pointing into the stack 259 fp = frame_sp + FKind::size(hf) - frame::sender_sp_offset; 260 } else { 261 fp = FKind::stub 262 ? frame_sp + fsize - frame::sender_sp_offset // fp always points to the address below the pushed return pc. We need correct address. 263 : *(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. 264 } 265 return frame(frame_sp, frame_sp, fp, hf.pc(), hf.cb(), hf.oop_map(), false); // TODO PERF : this computes deopt state; is it necessary? 266 } 267 } 268 269 inline intptr_t* ThawBase::align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom) { 270 #ifdef _LP64 271 if (((intptr_t)frame_sp & 0xf) != 0) { 272 assert(caller.is_interpreted_frame() || (bottom && hf.compiled_frame_stack_argsize() % 2 != 0), ""); 273 frame_sp--; 274 caller.set_sp(caller.sp() - 1); 275 } 276 assert(is_aligned(frame_sp, frame::frame_alignment), ""); 277 #endif 278 279 return frame_sp; 280 } 281 282 inline void ThawBase::patch_pd(frame& f, const frame& caller) { 283 patch_callee_link(caller, caller.fp()); 284 } 285 286 inline void ThawBase::patch_pd(frame& f, intptr_t* caller_sp) { 287 Unimplemented(); 288 } 289 290 inline intptr_t* ThawBase::push_preempt_rerun_adapter(frame top, bool is_interpreted_frame) { 291 intptr_t* sp = top.sp(); 292 CodeBlob* cb = top.cb(); 293 if (!is_interpreted_frame && cb->frame_size() == 2) { 294 // C2 runtime stub case. For aarch64 the real size of the c2 runtime stub is 2 words bigger 295 // than what we think, i.e. size is 4. This is because the _last_Java_sp is not set to the 296 // sp right before making the call to the VM, but rather it is artificially set 2 words above 297 // this real sp so that we can store the return address at last_Java_sp[-1], and keep this 298 // property where we can retrieve the last_Java_pc from the last_Java_sp. But that means that 299 // once we return to the runtime stub, the code will adjust sp according to this real size. 300 // So we must adjust the frame size back here. We just copy lr/rfp again. These 2 top words 301 // will be the ones popped in generate_cont_preempt_rerun_compiler_adapter(). The other 2 words 302 // will just be discarded once back in the runtime stub (add sp, sp, #0x10). 303 sp -= 2; 304 sp[-2] = sp[0]; 305 sp[-1] = sp[1]; 306 } 307 308 intptr_t* fp = sp - frame::sender_sp_offset; 309 address pc = is_interpreted_frame ? Interpreter::cont_preempt_rerun_interpreter_adapter() 310 : StubRoutines::cont_preempt_rerun_compiler_adapter(); 311 312 sp -= frame::metadata_words; 313 *(address*)(sp - frame::sender_sp_ret_address_offset()) = pc; 314 *(intptr_t**)(sp - frame::sender_sp_offset) = fp; 315 316 log_develop_trace(continuations, preempt)("push_preempt_rerun_%s_adapter() initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT " fp: " INTPTR_FORMAT, 317 is_interpreted_frame ? "interpreter" : "safepointblob", p2i(sp + frame::metadata_words), p2i(sp), p2i(fp)); 318 return sp; 319 } 320 321 inline intptr_t* ThawBase::push_preempt_monitorenter_redo(stackChunkOop chunk) { 322 323 // fprintf(stderr, "push_preempt_monitorenter_redo\n"); 324 frame enterSpecial = new_entry_frame(); 325 intptr_t* sp = enterSpecial.sp(); 326 327 // First push the return barrier frame 328 sp -= frame::metadata_words; 329 sp[1] = (intptr_t)StubRoutines::cont_returnBarrier(); 330 sp[0] = (intptr_t)enterSpecial.fp(); 331 332 // Now push the ObjectMonitor* 333 sp -= frame::metadata_words; 334 sp[1] = (intptr_t)chunk->objectMonitor(); // alignment 335 sp[0] = (intptr_t)chunk->objectMonitor(); 336 337 // Finally arrange to return to the monitorenter_redo stub 338 sp[-1] = (intptr_t)StubRoutines::cont_preempt_monitorenter_redo(); 339 sp[-2] = (intptr_t)enterSpecial.fp(); 340 log_develop_trace(continuations, preempt)("push_preempt_monitorenter_redo initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT, p2i(sp + 2 * frame::metadata_words), p2i(sp)); 341 return sp; 342 } 343 344 inline void ThawBase::derelativize_interpreted_frame_metadata(const frame& hf, const frame& f) { 345 // Make sure that last_sp is kept relativized. 346 assert((intptr_t*)f.at_relative(frame::interpreter_frame_last_sp_offset) == f.unextended_sp(), ""); 347 348 // Make sure that monitor_block_top is still relativized. 349 assert(f.at_absolute(frame::interpreter_frame_monitor_block_top_offset) <= frame::interpreter_frame_initial_sp_offset, ""); 350 351 // Make sure that extended_sp is kept relativized. 352 assert((intptr_t*)f.at_relative(frame::interpreter_frame_extended_sp_offset) < f.unextended_sp(), ""); 353 } 354 355 #endif // CPU_AARCH64_CONTINUATIONFREEZETHAW_AARCH64_INLINE_HPP