2 * Copyright (C) 2008 The Android Open Source Project
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
19 * Target independent portion of Android's Jit
26 #include "dexdump/OpCodeNames.h"
31 #include "compiler/Compiler.h"
32 #include "compiler/CompilerUtility.h"
33 #include "compiler/CompilerIR.h"
36 #if defined(WITH_SELF_VERIFICATION)
37 /* Allocate space for per-thread ShadowSpace data structures */
38 void* dvmSelfVerificationShadowSpaceAlloc(Thread* self)
40 self->shadowSpace = (ShadowSpace*) calloc(1, sizeof(ShadowSpace));
41 if (self->shadowSpace == NULL)
44 self->shadowSpace->registerSpaceSize = REG_SPACE;
45 self->shadowSpace->registerSpace =
46 (int*) calloc(self->shadowSpace->registerSpaceSize, sizeof(int));
48 return self->shadowSpace->registerSpace;
51 /* Free per-thread ShadowSpace data structures */
52 void dvmSelfVerificationShadowSpaceFree(Thread* self)
54 free(self->shadowSpace->registerSpace);
55 free(self->shadowSpace);
59 * Save out PC, FP, InterpState, and registers to shadow space.
60 * Return a pointer to the shadow space for JIT to use.
62 void* dvmSelfVerificationSaveState(const u2* pc, const void* fp,
63 InterpState* interpState, int targetTrace)
65 Thread *self = dvmThreadSelf();
66 ShadowSpace *shadowSpace = self->shadowSpace;
67 unsigned preBytes = interpState->method->outsSize*4 + sizeof(StackSaveArea);
68 unsigned postBytes = interpState->method->registersSize*4;
70 //LOGD("### selfVerificationSaveState(%d) pc: 0x%x fp: 0x%x",
71 // self->threadId, (int)pc, (int)fp);
73 if (shadowSpace->selfVerificationState != kSVSIdle) {
74 LOGD("~~~ Save: INCORRECT PREVIOUS STATE(%d): %d",
75 self->threadId, shadowSpace->selfVerificationState);
76 LOGD("********** SHADOW STATE DUMP **********");
77 LOGD("PC: 0x%x FP: 0x%x", (int)pc, (int)fp);
79 shadowSpace->selfVerificationState = kSVSStart;
81 if (interpState->entryPoint == kInterpEntryResume) {
82 interpState->entryPoint = kInterpEntryInstr;
84 /* Tracking the success rate of resume after single-stepping */
85 if (interpState->jitResumeDPC == pc) {
86 LOGD("SV single step resumed at %p", pc);
89 LOGD("real %p DPC %p NPC %p", pc, interpState->jitResumeDPC,
90 interpState->jitResumeNPC);
95 // Dynamically grow shadow register space if necessary
96 if (preBytes + postBytes > shadowSpace->registerSpaceSize * sizeof(u4)) {
97 free(shadowSpace->registerSpace);
98 shadowSpace->registerSpaceSize = (preBytes + postBytes) / sizeof(u4);
99 shadowSpace->registerSpace =
100 (int*) calloc(shadowSpace->registerSpaceSize, sizeof(u4));
103 // Remember original state
104 shadowSpace->startPC = pc;
105 shadowSpace->fp = fp;
106 shadowSpace->glue = interpState;
108 * Store the original method here in case the trace ends with a
109 * return/invoke, the last method.
111 shadowSpace->method = interpState->method;
112 shadowSpace->shadowFP = shadowSpace->registerSpace +
113 shadowSpace->registerSpaceSize - postBytes/4;
115 // Create a copy of the InterpState
116 memcpy(&(shadowSpace->interpState), interpState, sizeof(InterpState));
117 shadowSpace->interpState.fp = shadowSpace->shadowFP;
118 shadowSpace->interpState.interpStackEnd = (u1*)shadowSpace->registerSpace;
120 // Create a copy of the stack
121 memcpy(((char*)shadowSpace->shadowFP)-preBytes, ((char*)fp)-preBytes,
124 // Setup the shadowed heap space
125 shadowSpace->heapSpaceTail = shadowSpace->heapSpace;
127 // Reset trace length
128 shadowSpace->traceLength = 0;
134 * Save ending PC, FP and compiled code exit point to shadow space.
135 * Return a pointer to the shadow space for JIT to restore state.
137 void* dvmSelfVerificationRestoreState(const u2* pc, const void* fp,
138 SelfVerificationState exitPoint)
140 Thread *self = dvmThreadSelf();
141 ShadowSpace *shadowSpace = self->shadowSpace;
142 // Official InterpState structure
143 InterpState *realGlue = shadowSpace->glue;
144 shadowSpace->endPC = pc;
145 shadowSpace->endShadowFP = fp;
147 //LOGD("### selfVerificationRestoreState(%d) pc: 0x%x fp: 0x%x endPC: 0x%x",
148 // self->threadId, (int)shadowSpace->startPC, (int)shadowSpace->fp,
151 if (shadowSpace->selfVerificationState != kSVSStart) {
152 LOGD("~~~ Restore: INCORRECT PREVIOUS STATE(%d): %d",
153 self->threadId, shadowSpace->selfVerificationState);
154 LOGD("********** SHADOW STATE DUMP **********");
155 LOGD("Dalvik PC: 0x%x endPC: 0x%x", (int)shadowSpace->startPC,
156 (int)shadowSpace->endPC);
157 LOGD("Interp FP: 0x%x", (int)shadowSpace->fp);
158 LOGD("Shadow FP: 0x%x endFP: 0x%x", (int)shadowSpace->shadowFP,
159 (int)shadowSpace->endShadowFP);
162 // Move the resume [ND]PC from the shadow space to the real space so that
163 // the debug interpreter can return to the translation
164 if (exitPoint == kSVSSingleStep) {
165 realGlue->jitResumeNPC = shadowSpace->interpState.jitResumeNPC;
166 realGlue->jitResumeDPC = shadowSpace->interpState.jitResumeDPC;
168 realGlue->jitResumeNPC = NULL;
169 realGlue->jitResumeDPC = NULL;
172 // Special case when punting after a single instruction
173 if (exitPoint == kSVSPunt && pc == shadowSpace->startPC) {
174 shadowSpace->selfVerificationState = kSVSIdle;
176 shadowSpace->selfVerificationState = exitPoint;
182 /* Print contents of virtual registers */
183 static void selfVerificationPrintRegisters(int* addr, int* addrRef,
187 for (i = 0; i < numWords; i++) {
188 LOGD("(v%d) 0x%8x%s", i, addr[i], addr[i] != addrRef[i] ? " X" : "");
192 /* Print values maintained in shadowSpace */
193 static void selfVerificationDumpState(const u2* pc, Thread* self)
195 ShadowSpace* shadowSpace = self->shadowSpace;
196 StackSaveArea* stackSave = SAVEAREA_FROM_FP(self->curFrame);
197 int frameBytes = (int) shadowSpace->registerSpace +
198 shadowSpace->registerSpaceSize*4 -
199 (int) shadowSpace->shadowFP;
202 if (self->curFrame < shadowSpace->fp) {
203 localRegs = (stackSave->method->registersSize -
204 stackSave->method->insSize)*4;
205 frameBytes2 = (int) shadowSpace->fp - (int) self->curFrame - localRegs;
207 LOGD("********** SHADOW STATE DUMP **********");
208 LOGD("CurrentPC: 0x%x, Offset: 0x%04x", (int)pc,
209 (int)(pc - stackSave->method->insns));
210 LOGD("Class: %s", shadowSpace->method->clazz->descriptor);
211 LOGD("Method: %s", shadowSpace->method->name);
212 LOGD("Dalvik PC: 0x%x endPC: 0x%x", (int)shadowSpace->startPC,
213 (int)shadowSpace->endPC);
214 LOGD("Interp FP: 0x%x endFP: 0x%x", (int)shadowSpace->fp,
215 (int)self->curFrame);
216 LOGD("Shadow FP: 0x%x endFP: 0x%x", (int)shadowSpace->shadowFP,
217 (int)shadowSpace->endShadowFP);
218 LOGD("Frame1 Bytes: %d Frame2 Local: %d Bytes: %d", frameBytes,
219 localRegs, frameBytes2);
220 LOGD("Trace length: %d State: %d", shadowSpace->traceLength,
221 shadowSpace->selfVerificationState);
224 /* Print decoded instructions in the current trace */
225 static void selfVerificationDumpTrace(const u2* pc, Thread* self)
227 ShadowSpace* shadowSpace = self->shadowSpace;
228 StackSaveArea* stackSave = SAVEAREA_FROM_FP(self->curFrame);
230 DecodedInstruction *decInsn;
232 LOGD("********** SHADOW TRACE DUMP **********");
233 for (i = 0; i < shadowSpace->traceLength; i++) {
234 addr = shadowSpace->trace[i].addr;
235 offset = (int)((u2*)addr - stackSave->method->insns);
236 decInsn = &(shadowSpace->trace[i].decInsn);
237 /* Not properly decoding instruction, some registers may be garbage */
238 LOGD("0x%x: (0x%04x) %s", addr, offset, getOpcodeName(decInsn->opCode));
242 /* Code is forced into this spin loop when a divergence is detected */
243 static void selfVerificationSpinLoop(ShadowSpace *shadowSpace)
245 const u2 *startPC = shadowSpace->startPC;
246 JitTraceDescription* desc = dvmCopyTraceDescriptor(startPC, NULL);
248 dvmCompilerWorkEnqueue(startPC, kWorkOrderTraceDebug, desc);
250 * This function effectively terminates the VM right here, so not
251 * freeing the desc pointer when the enqueuing fails is acceptable.
254 gDvmJit.selfVerificationSpin = true;
255 while(gDvmJit.selfVerificationSpin) sleep(10);
258 /* Manage self verification while in the debug interpreter */
259 static bool selfVerificationDebugInterp(const u2* pc, Thread* self,
260 InterpState *interpState)
262 ShadowSpace *shadowSpace = self->shadowSpace;
263 SelfVerificationState state = shadowSpace->selfVerificationState;
265 DecodedInstruction decInsn;
266 dexDecodeInstruction(gDvm.instrFormat, pc, &decInsn);
268 //LOGD("### DbgIntp(%d): PC: 0x%x endPC: 0x%x state: %d len: %d %s",
269 // self->threadId, (int)pc, (int)shadowSpace->endPC, state,
270 // shadowSpace->traceLength, getOpcodeName(decInsn.opCode));
272 if (state == kSVSIdle || state == kSVSStart) {
273 LOGD("~~~ DbgIntrp: INCORRECT PREVIOUS STATE(%d): %d",
274 self->threadId, state);
275 selfVerificationDumpState(pc, self);
276 selfVerificationDumpTrace(pc, self);
280 * Skip endPC once when trace has a backward branch. If the SV state is
281 * single step, keep it that way.
283 if ((state == kSVSBackwardBranch && pc == shadowSpace->endPC) ||
284 (state != kSVSBackwardBranch && state != kSVSSingleStep)) {
285 shadowSpace->selfVerificationState = kSVSDebugInterp;
288 /* Check that the current pc is the end of the trace */
289 if ((state == kSVSDebugInterp || state == kSVSSingleStep) &&
290 pc == shadowSpace->endPC) {
292 shadowSpace->selfVerificationState = kSVSIdle;
294 /* Check register space */
295 int frameBytes = (int) shadowSpace->registerSpace +
296 shadowSpace->registerSpaceSize*4 -
297 (int) shadowSpace->shadowFP;
298 if (memcmp(shadowSpace->fp, shadowSpace->shadowFP, frameBytes)) {
299 LOGD("~~~ DbgIntp(%d): REGISTERS DIVERGENCE!", self->threadId);
300 selfVerificationDumpState(pc, self);
301 selfVerificationDumpTrace(pc, self);
302 LOGD("*** Interp Registers: addr: 0x%x bytes: %d",
303 (int)shadowSpace->fp, frameBytes);
304 selfVerificationPrintRegisters((int*)shadowSpace->fp,
305 (int*)shadowSpace->shadowFP,
307 LOGD("*** Shadow Registers: addr: 0x%x bytes: %d",
308 (int)shadowSpace->shadowFP, frameBytes);
309 selfVerificationPrintRegisters((int*)shadowSpace->shadowFP,
310 (int*)shadowSpace->fp,
312 selfVerificationSpinLoop(shadowSpace);
314 /* Check new frame if it exists (invokes only) */
315 if (self->curFrame < shadowSpace->fp) {
316 StackSaveArea* stackSave = SAVEAREA_FROM_FP(self->curFrame);
317 int localRegs = (stackSave->method->registersSize -
318 stackSave->method->insSize)*4;
319 int frameBytes2 = (int) shadowSpace->fp -
320 (int) self->curFrame - localRegs;
321 if (memcmp(((char*)self->curFrame)+localRegs,
322 ((char*)shadowSpace->endShadowFP)+localRegs, frameBytes2)) {
323 LOGD("~~~ DbgIntp(%d): REGISTERS (FRAME2) DIVERGENCE!",
325 selfVerificationDumpState(pc, self);
326 selfVerificationDumpTrace(pc, self);
327 LOGD("*** Interp Registers: addr: 0x%x l: %d bytes: %d",
328 (int)self->curFrame, localRegs, frameBytes2);
329 selfVerificationPrintRegisters((int*)self->curFrame,
330 (int*)shadowSpace->endShadowFP,
331 (frameBytes2+localRegs)/4);
332 LOGD("*** Shadow Registers: addr: 0x%x l: %d bytes: %d",
333 (int)shadowSpace->endShadowFP, localRegs, frameBytes2);
334 selfVerificationPrintRegisters((int*)shadowSpace->endShadowFP,
335 (int*)self->curFrame,
336 (frameBytes2+localRegs)/4);
337 selfVerificationSpinLoop(shadowSpace);
341 /* Check memory space */
342 bool memDiff = false;
343 ShadowHeap* heapSpacePtr;
344 for (heapSpacePtr = shadowSpace->heapSpace;
345 heapSpacePtr != shadowSpace->heapSpaceTail; heapSpacePtr++) {
346 int memData = *((unsigned int*) heapSpacePtr->addr);
347 if (heapSpacePtr->data != memData) {
348 LOGD("~~~ DbgIntp(%d): MEMORY DIVERGENCE!", self->threadId);
349 LOGD("Addr: 0x%x Intrp Data: 0x%x Jit Data: 0x%x",
350 heapSpacePtr->addr, memData, heapSpacePtr->data);
351 selfVerificationDumpState(pc, self);
352 selfVerificationDumpTrace(pc, self);
356 if (memDiff) selfVerificationSpinLoop(shadowSpace);
359 * Switch to JIT single step mode to stay in the debug interpreter for
360 * one more instruction
362 if (state == kSVSSingleStep) {
363 interpState->jitState = kJitSingleStepEnd;
367 /* If end not been reached, make sure max length not exceeded */
368 } else if (shadowSpace->traceLength >= JIT_MAX_TRACE_LEN) {
369 LOGD("~~~ DbgIntp(%d): CONTROL DIVERGENCE!", self->threadId);
370 LOGD("startPC: 0x%x endPC: 0x%x currPC: 0x%x",
371 (int)shadowSpace->startPC, (int)shadowSpace->endPC, (int)pc);
372 selfVerificationDumpState(pc, self);
373 selfVerificationDumpTrace(pc, self);
374 selfVerificationSpinLoop(shadowSpace);
378 /* Log the instruction address and decoded instruction for debug */
379 shadowSpace->trace[shadowSpace->traceLength].addr = (int)pc;
380 shadowSpace->trace[shadowSpace->traceLength].decInsn = decInsn;
381 shadowSpace->traceLength++;
388 * If one of our fixed tables or the translation buffer fills up,
389 * call this routine to avoid wasting cycles on future translation requests.
391 void dvmJitStopTranslationRequests()
394 * Note 1: This won't necessarily stop all translation requests, and
395 * operates on a delayed mechanism. Running threads look to the copy
396 * of this value in their private InterpState structures and won't see
397 * this change until it is refreshed (which happens on interpreter
399 * Note 2: This is a one-shot memory leak on this table. Because this is a
400 * permanent off switch for Jit profiling, it is a one-time leak of 1K
401 * bytes, and no further attempt will be made to re-allocate it. Can't
402 * free it because some thread may be holding a reference.
404 gDvmJit.pProfTable = NULL;
407 #if defined(JIT_STATS)
408 /* Convenience function to increment counter from assembly code */
409 void dvmBumpNoChain(int from)
411 gDvmJit.noChainExit[from]++;
414 /* Convenience function to increment counter from assembly code */
417 gDvmJit.normalExit++;
420 /* Convenience function to increment counter from assembly code */
421 void dvmBumpPunt(int from)
427 /* Dumps debugging & tuning stats to the log */
435 if (gDvmJit.pJitEntryTable) {
436 for (i=0, stubs=chains=hit=not_hit=0;
437 i < (int) gDvmJit.jitTableSize;
439 if (gDvmJit.pJitEntryTable[i].dPC != 0) {
441 if (gDvmJit.pJitEntryTable[i].codeAddress ==
442 gDvmJit.interpretTemplate)
446 if (gDvmJit.pJitEntryTable[i].u.info.chain != gDvmJit.jitTableSize)
449 LOGD("JIT: table size is %d, entries used is %d",
450 gDvmJit.jitTableSize, gDvmJit.jitTableEntriesUsed);
451 LOGD("JIT: %d traces, %d slots, %d chains, %d thresh, %s",
452 hit, not_hit + hit, chains, gDvmJit.threshold,
453 gDvmJit.blockingMode ? "Blocking" : "Non-blocking");
455 #if defined(JIT_STATS)
456 LOGD("JIT: Lookups: %d hits, %d misses; %d normal, %d punt",
457 gDvmJit.addrLookupsFound, gDvmJit.addrLookupsNotFound,
458 gDvmJit.normalExit, gDvmJit.puntExit);
459 LOGD("JIT: noChainExit: %d IC miss, %d interp callsite, "
460 "%d switch overflow",
461 gDvmJit.noChainExit[kInlineCacheMiss],
462 gDvmJit.noChainExit[kCallsiteInterpreted],
463 gDvmJit.noChainExit[kSwitchOverflow]);
465 LOGD("JIT: Invoke: %d mono, %d poly, %d native, %d return",
466 gDvmJit.invokeMonomorphic, gDvmJit.invokePolymorphic,
467 gDvmJit.invokeNative, gDvmJit.returnOp);
468 LOGD("JIT: Total compilation time: %llu ms", gDvmJit.jitTime / 1000);
469 LOGD("JIT: Avg unit compilation time: %llu us",
470 gDvmJit.jitTime / gDvmJit.numCompilations);
473 LOGD("JIT: %d Translation chains, %d interp stubs",
474 gDvmJit.translationChains, stubs);
475 if (gDvmJit.profile) {
476 dvmCompilerSortAndPrintTraceProfiles();
482 void setTraceConstruction(JitEntry *slot, bool value)
485 JitEntryInfoUnion oldValue;
486 JitEntryInfoUnion newValue;
490 newValue.info.traceConstruction = value;
491 } while (!ATOMIC_CMP_SWAP( &slot->u.infoWord,
492 oldValue.infoWord, newValue.infoWord));
495 void resetTracehead(InterpState* interpState, JitEntry *slot)
497 slot->codeAddress = gDvmJit.interpretTemplate;
498 setTraceConstruction(slot, false);
501 /* Clean up any pending trace builds */
502 void dvmJitAbortTraceSelect(InterpState* interpState)
504 if (interpState->jitState == kJitTSelect)
505 interpState->jitState = kJitDone;
509 * Find an entry in the JitTable, creating if necessary.
510 * Returns null if table is full.
512 static JitEntry *lookupAndAdd(const u2* dPC, bool callerLocked)
514 u4 chainEndMarker = gDvmJit.jitTableSize;
515 u4 idx = dvmJitHash(dPC);
517 /* Walk the bucket chain to find an exact match for our PC */
518 while ((gDvmJit.pJitEntryTable[idx].u.info.chain != chainEndMarker) &&
519 (gDvmJit.pJitEntryTable[idx].dPC != dPC)) {
520 idx = gDvmJit.pJitEntryTable[idx].u.info.chain;
523 if (gDvmJit.pJitEntryTable[idx].dPC != dPC) {
525 * No match. Aquire jitTableLock and find the last
526 * slot in the chain. Possibly continue the chain walk in case
527 * some other thread allocated the slot we were looking
528 * at previuosly (perhaps even the dPC we're trying to enter).
531 dvmLockMutex(&gDvmJit.tableLock);
533 * At this point, if .dPC is NULL, then the slot we're
534 * looking at is the target slot from the primary hash
535 * (the simple, and common case). Otherwise we're going
536 * to have to find a free slot and chain it.
538 MEM_BARRIER(); /* Make sure we reload [].dPC after lock */
539 if (gDvmJit.pJitEntryTable[idx].dPC != NULL) {
541 while (gDvmJit.pJitEntryTable[idx].u.info.chain != chainEndMarker) {
542 if (gDvmJit.pJitEntryTable[idx].dPC == dPC) {
543 /* Another thread got there first for this dPC */
545 dvmUnlockMutex(&gDvmJit.tableLock);
546 return &gDvmJit.pJitEntryTable[idx];
548 idx = gDvmJit.pJitEntryTable[idx].u.info.chain;
550 /* Here, idx should be pointing to the last cell of an
551 * active chain whose last member contains a valid dPC */
552 assert(gDvmJit.pJitEntryTable[idx].dPC != NULL);
553 /* Linear walk to find a free cell and add it to the end */
557 if (idx == chainEndMarker)
558 idx = 0; /* Wraparound */
559 if ((gDvmJit.pJitEntryTable[idx].dPC == NULL) ||
564 JitEntryInfoUnion oldValue;
565 JitEntryInfoUnion newValue;
567 * Although we hold the lock so that noone else will
568 * be trying to update a chain field, the other fields
569 * packed into the word may be in use by other threads.
572 oldValue = gDvmJit.pJitEntryTable[prev].u;
574 newValue.info.chain = idx;
575 } while (!ATOMIC_CMP_SWAP(
576 &gDvmJit.pJitEntryTable[prev].u.infoWord,
577 oldValue.infoWord, newValue.infoWord));
580 if (gDvmJit.pJitEntryTable[idx].dPC == NULL) {
582 * Initialize codeAddress and allocate the slot. Must
583 * happen in this order (since dPC is set, the entry is live.
585 gDvmJit.pJitEntryTable[idx].dPC = dPC;
586 gDvmJit.jitTableEntriesUsed++;
589 idx = chainEndMarker;
592 dvmUnlockMutex(&gDvmJit.tableLock);
594 return (idx == chainEndMarker) ? NULL : &gDvmJit.pJitEntryTable[idx];
598 * Adds to the current trace request one instruction at a time, just
599 * before that instruction is interpreted. This is the primary trace
600 * selection function. NOTE: return instruction are handled a little
601 * differently. In general, instructions are "proposed" to be added
602 * to the current trace prior to interpretation. If the interpreter
603 * then successfully completes the instruction, is will be considered
604 * part of the request. This allows us to examine machine state prior
605 * to interpretation, and also abort the trace request if the instruction
606 * throws or does something unexpected. However, return instructions
607 * will cause an immediate end to the translation request - which will
608 * be passed to the compiler before the return completes. This is done
609 * in response to special handling of returns by the interpreter (and
610 * because returns cannot throw in a way that causes problems for the
613 int dvmCheckJit(const u2* pc, Thread* self, InterpState* interpState)
616 int switchInterp = false;
617 bool debugOrProfile = dvmDebuggerOrProfilerActive();
619 /* Prepare to handle last PC and stage the current PC */
620 const u2 *lastPC = interpState->lastPC;
621 interpState->lastPC = pc;
623 switch (interpState->jitState) {
627 DecodedInstruction decInsn;
629 /* First instruction - just remember the PC and exit */
630 if (lastPC == NULL) break;
631 /* Grow the trace around the last PC if jitState is kJitTSelect */
632 dexDecodeInstruction(gDvm.instrFormat, lastPC, &decInsn);
635 * Treat {PACKED,SPARSE}_SWITCH as trace-ending instructions due
636 * to the amount of space it takes to generate the chaining
639 if (interpState->totalTraceLen != 0 &&
640 (decInsn.opCode == OP_PACKED_SWITCH ||
641 decInsn.opCode == OP_SPARSE_SWITCH)) {
642 interpState->jitState = kJitTSelectEnd;
647 #if defined(SHOW_TRACE)
648 LOGD("TraceGen: adding %s",getOpcodeName(decInsn.opCode));
650 flags = dexGetInstrFlags(gDvm.instrFlags, decInsn.opCode);
651 len = dexGetInstrOrTableWidthAbs(gDvm.instrWidth, lastPC);
652 offset = lastPC - interpState->method->insns;
653 assert((unsigned) offset <
654 dvmGetMethodInsnsSize(interpState->method));
655 if (lastPC != interpState->currRunHead + interpState->currRunLen) {
657 /* We need to start a new trace run */
658 currTraceRun = ++interpState->currTraceRun;
659 interpState->currRunLen = 0;
660 interpState->currRunHead = (u2*)lastPC;
661 interpState->trace[currTraceRun].frag.startOffset = offset;
662 interpState->trace[currTraceRun].frag.numInsts = 0;
663 interpState->trace[currTraceRun].frag.runEnd = false;
664 interpState->trace[currTraceRun].frag.hint = kJitHintNone;
666 interpState->trace[interpState->currTraceRun].frag.numInsts++;
667 interpState->totalTraceLen++;
668 interpState->currRunLen += len;
670 /* Will probably never hit this with the current trace buildier */
671 if (interpState->currTraceRun == (MAX_JIT_RUN_LEN - 1)) {
672 interpState->jitState = kJitTSelectEnd;
675 if ( ((flags & kInstrUnconditional) == 0) &&
676 /* don't end trace on INVOKE_DIRECT_EMPTY */
677 (decInsn.opCode != OP_INVOKE_DIRECT_EMPTY) &&
678 ((flags & (kInstrCanBranch |
681 kInstrInvoke)) != 0)) {
682 interpState->jitState = kJitTSelectEnd;
683 #if defined(SHOW_TRACE)
684 LOGD("TraceGen: ending on %s, basic block end",
685 getOpcodeName(decInsn.opCode));
688 /* Break on throw or self-loop */
689 if ((decInsn.opCode == OP_THROW) || (lastPC == pc)){
690 interpState->jitState = kJitTSelectEnd;
692 if (interpState->totalTraceLen >= JIT_MAX_TRACE_LEN) {
693 interpState->jitState = kJitTSelectEnd;
695 /* Abandon the trace request if debugger/profiler is attached */
696 if (debugOrProfile) {
697 interpState->jitState = kJitDone;
700 if ((flags & kInstrCanReturn) != kInstrCanReturn) {
703 /* NOTE: intentional fallthrough for returns */
707 if (interpState->totalTraceLen == 0) {
708 /* Bad trace - mark as untranslatable */
709 interpState->jitState = kJitDone;
713 JitTraceDescription* desc =
714 (JitTraceDescription*)malloc(sizeof(JitTraceDescription) +
715 sizeof(JitTraceRun) * (interpState->currTraceRun+1));
717 LOGE("Out of memory in trace selection");
718 dvmJitStopTranslationRequests();
719 interpState->jitState = kJitDone;
723 interpState->trace[interpState->currTraceRun].frag.runEnd =
725 desc->method = interpState->method;
726 memcpy((char*)&(desc->trace[0]),
727 (char*)&(interpState->trace[0]),
728 sizeof(JitTraceRun) * (interpState->currTraceRun+1));
729 #if defined(SHOW_TRACE)
730 LOGD("TraceGen: trace done, adding to queue");
732 if (dvmCompilerWorkEnqueue(
733 interpState->currTraceHead,kWorkOrderTrace,desc)) {
734 /* Work order successfully enqueued */
735 if (gDvmJit.blockingMode) {
736 dvmCompilerDrainQueue();
740 * Make sure the descriptor for the abandoned work order is
746 * Reset "trace in progress" flag whether or not we
747 * successfully entered a work order.
750 lookupAndAdd(interpState->currTraceHead, false);
752 setTraceConstruction(jitEntry, false);
754 interpState->jitState = kJitDone;
759 interpState->jitState = kJitSingleStepEnd;
761 case kJitSingleStepEnd:
762 interpState->entryPoint = kInterpEntryResume;
763 interpState->jitState = kJitDone;
769 #if defined(WITH_SELF_VERIFICATION)
770 case kJitSelfVerification:
771 if (selfVerificationDebugInterp(pc, self, interpState)) {
773 * If the next state is not single-step end, we can switch
776 if (interpState->jitState != kJitSingleStepEnd) {
777 interpState->jitState = kJitDone;
784 * If the debug interpreter was entered for non-JIT reasons, check if
785 * the original reason still holds. If not, we have to force the
786 * interpreter switch here and use dvmDebuggerOrProfilerActive instead
787 * of dvmJitDebuggerOrProfilerActive since the latter will alwasy
788 * return true when the debugger/profiler is already detached and the
789 * JIT profiling table is restored.
792 switchInterp = !dvmDebuggerOrProfilerActive();
795 LOGE("Unexpected JIT state: %d entry point: %d",
796 interpState->jitState, interpState->entryPoint);
801 * Final check to see if we can really switch the interpreter. Make sure
802 * the jitState is kJitDone or kJitNot when switchInterp is set to true.
804 assert(switchInterp == false || interpState->jitState == kJitDone ||
805 interpState->jitState == kJitNot);
806 return switchInterp && !debugOrProfile;
809 JitEntry *dvmFindJitEntry(const u2* pc)
811 int idx = dvmJitHash(pc);
813 /* Expect a high hit rate on 1st shot */
814 if (gDvmJit.pJitEntryTable[idx].dPC == pc)
815 return &gDvmJit.pJitEntryTable[idx];
817 int chainEndMarker = gDvmJit.jitTableSize;
818 while (gDvmJit.pJitEntryTable[idx].u.info.chain != chainEndMarker) {
819 idx = gDvmJit.pJitEntryTable[idx].u.info.chain;
820 if (gDvmJit.pJitEntryTable[idx].dPC == pc)
821 return &gDvmJit.pJitEntryTable[idx];
828 * If a translated code address exists for the davik byte code
829 * pointer return it. This routine needs to be fast.
831 void* dvmJitGetCodeAddr(const u2* dPC)
833 int idx = dvmJitHash(dPC);
834 const u2* npc = gDvmJit.pJitEntryTable[idx].dPC;
836 bool hideTranslation = (gDvm.sumThreadSuspendCount != 0) ||
837 (gDvmJit.codeCacheFull == true) ||
838 (gDvmJit.pProfTable == NULL);
841 #if defined(JIT_STATS)
842 gDvmJit.addrLookupsFound++;
844 return hideTranslation ?
845 NULL : gDvmJit.pJitEntryTable[idx].codeAddress;
847 int chainEndMarker = gDvmJit.jitTableSize;
848 while (gDvmJit.pJitEntryTable[idx].u.info.chain != chainEndMarker) {
849 idx = gDvmJit.pJitEntryTable[idx].u.info.chain;
850 if (gDvmJit.pJitEntryTable[idx].dPC == dPC) {
851 #if defined(JIT_STATS)
852 gDvmJit.addrLookupsFound++;
854 return hideTranslation ?
855 NULL : gDvmJit.pJitEntryTable[idx].codeAddress;
860 #if defined(JIT_STATS)
861 gDvmJit.addrLookupsNotFound++;
867 * Register the translated code pointer into the JitTable.
868 * NOTE: Once a codeAddress field transitions from initial state to
869 * JIT'd code, it must not be altered without first halting all
870 * threads. This routine should only be called by the compiler
873 void dvmJitSetCodeAddr(const u2* dPC, void *nPC, JitInstructionSetType set) {
874 JitEntryInfoUnion oldValue;
875 JitEntryInfoUnion newValue;
876 JitEntry *jitEntry = lookupAndAdd(dPC, false);
878 /* Note: order of update is important */
880 oldValue = jitEntry->u;
882 newValue.info.instructionSet = set;
883 } while (!ATOMIC_CMP_SWAP(
884 &jitEntry->u.infoWord,
885 oldValue.infoWord, newValue.infoWord));
886 jitEntry->codeAddress = nPC;
890 * Determine if valid trace-bulding request is active. Return true
891 * if we need to abort and switch back to the fast interpreter, false
894 bool dvmJitCheckTraceRequest(Thread* self, InterpState* interpState)
896 bool switchInterp = false; /* Assume success */
898 intptr_t filterKey = ((intptr_t) interpState->pc) >>
899 JIT_TRACE_THRESH_FILTER_GRAN_LOG2;
900 bool debugOrProfile = dvmDebuggerOrProfilerActive();
902 /* Check if the JIT request can be handled now */
903 if (gDvmJit.pJitEntryTable != NULL && debugOrProfile == false) {
904 /* Bypass the filter for hot trace requests or during stress mode */
905 if (interpState->jitState == kJitTSelectRequest &&
906 gDvmJit.threshold > 6) {
907 /* Two-level filtering scheme */
908 for (i=0; i< JIT_TRACE_THRESH_FILTER_SIZE; i++) {
909 if (filterKey == interpState->threshFilter[i]) {
913 if (i == JIT_TRACE_THRESH_FILTER_SIZE) {
915 * Use random replacement policy - otherwise we could miss a
916 * large loop that contains more traces than the size of our
919 i = rand() % JIT_TRACE_THRESH_FILTER_SIZE;
920 interpState->threshFilter[i] = filterKey;
921 interpState->jitState = kJitDone;
925 /* If the compiler is backlogged, cancel any JIT actions */
926 if (gDvmJit.compilerQueueLength >= gDvmJit.compilerHighWater) {
927 interpState->jitState = kJitDone;
931 * Check for additional reasons that might force the trace select
932 * request to be dropped
934 if (interpState->jitState == kJitTSelectRequest ||
935 interpState->jitState == kJitTSelectRequestHot) {
936 JitEntry *slot = lookupAndAdd(interpState->pc, false);
939 * Table is full. This should have been
940 * detected by the compiler thread and the table
941 * resized before we run into it here. Assume bad things
942 * are afoot and disable profiling.
944 interpState->jitState = kJitDone;
945 LOGD("JIT: JitTable full, disabling profiling");
946 dvmJitStopTranslationRequests();
947 } else if (slot->u.info.traceConstruction) {
949 * Trace request already in progress, but most likely it
950 * aborted without cleaning up. Assume the worst and
951 * mark trace head as untranslatable. If we're wrong,
952 * the compiler thread will correct the entry when the
953 * translation is completed. The downside here is that
954 * some existing translation may chain to the interpret-only
955 * template instead of the real translation during this
956 * window. Performance, but not correctness, issue.
958 interpState->jitState = kJitDone;
959 resetTracehead(interpState, slot);
960 } else if (slot->codeAddress) {
961 /* Nothing to do here - just return */
962 interpState->jitState = kJitDone;
965 * Mark request. Note, we are not guaranteed exclusivity
966 * here. A window exists for another thread to be
967 * attempting to build this same trace. Rather than
968 * bear the cost of locking, we'll just allow that to
969 * happen. The compiler thread, if it chooses, can
970 * discard redundant requests.
972 setTraceConstruction(slot, true);
976 switch (interpState->jitState) {
977 case kJitTSelectRequest:
978 case kJitTSelectRequestHot:
979 interpState->jitState = kJitTSelect;
980 interpState->currTraceHead = interpState->pc;
981 interpState->currTraceRun = 0;
982 interpState->totalTraceLen = 0;
983 interpState->currRunHead = interpState->pc;
984 interpState->currRunLen = 0;
985 interpState->trace[0].frag.startOffset =
986 interpState->pc - interpState->method->insns;
987 interpState->trace[0].frag.numInsts = 0;
988 interpState->trace[0].frag.runEnd = false;
989 interpState->trace[0].frag.hint = kJitHintNone;
990 interpState->lastPC = 0;
993 * For JIT's perspective there is no need to stay in the debug
994 * interpreter unless debugger/profiler is attached.
1000 LOGE("Unexpected JIT state: %d entry point: %d",
1001 interpState->jitState, interpState->entryPoint);
1006 * Cannot build trace this time - ready to leave the dbg interpreter
1008 interpState->jitState = kJitDone;
1009 switchInterp = true;
1013 * Final check to see if we can really switch the interpreter. Make sure
1014 * the jitState is kJitDone when switchInterp is set to true.
1016 assert(switchInterp == false || interpState->jitState == kJitDone);
1017 return switchInterp && !debugOrProfile;
1021 * Resizes the JitTable. Must be a power of 2, and returns true on failure.
1022 * Stops all threads, and thus is a heavyweight operation. May only be called
1023 * by the compiler thread.
1025 bool dvmJitResizeJitTable( unsigned int size )
1027 JitEntry *pNewTable;
1028 JitEntry *pOldTable;
1031 unsigned int oldSize;
1034 assert(gDvmJit.pJitEntryTable != NULL);
1035 assert(size && !(size & (size - 1))); /* Is power of 2? */
1037 LOGI("Jit: resizing JitTable from %d to %d", gDvmJit.jitTableSize, size);
1041 if (size <= gDvmJit.jitTableSize) {
1045 /* Make sure requested size is compatible with chain field width */
1046 tempEntry.u.info.chain = size;
1047 if (tempEntry.u.info.chain != size) {
1048 LOGD("Jit: JitTable request of %d too big", size);
1052 pNewTable = (JitEntry*)calloc(size, sizeof(*pNewTable));
1053 if (pNewTable == NULL) {
1056 for (i=0; i< size; i++) {
1057 pNewTable[i].u.info.chain = size; /* Initialize chain termination */
1060 /* Stop all other interpreting/jit'ng threads */
1061 dvmSuspendAllThreads(SUSPEND_FOR_TBL_RESIZE);
1063 pOldTable = gDvmJit.pJitEntryTable;
1064 oldSize = gDvmJit.jitTableSize;
1066 dvmLockMutex(&gDvmJit.tableLock);
1067 gDvmJit.pJitEntryTable = pNewTable;
1068 gDvmJit.jitTableSize = size;
1069 gDvmJit.jitTableMask = size - 1;
1070 gDvmJit.jitTableEntriesUsed = 0;
1072 for (i=0; i < oldSize; i++) {
1073 if (pOldTable[i].dPC) {
1076 p = lookupAndAdd(pOldTable[i].dPC, true /* holds tableLock*/ );
1077 p->codeAddress = pOldTable[i].codeAddress;
1078 /* We need to preserve the new chain field, but copy the rest */
1079 chain = p->u.info.chain;
1080 p->u = pOldTable[i].u;
1081 p->u.info.chain = chain;
1084 dvmUnlockMutex(&gDvmJit.tableLock);
1088 /* Restart the world */
1089 dvmResumeAllThreads(SUSPEND_FOR_TBL_RESIZE);
1095 * Reset the JitTable to the initial clean state.
1097 void dvmJitResetTable(void)
1099 JitEntry *jitEntry = gDvmJit.pJitEntryTable;
1100 unsigned int size = gDvmJit.jitTableSize;
1103 dvmLockMutex(&gDvmJit.tableLock);
1104 memset((void *) jitEntry, 0, sizeof(JitEntry) * size);
1105 for (i=0; i< size; i++) {
1106 jitEntry[i].u.info.chain = size; /* Initialize chain termination */
1108 gDvmJit.jitTableEntriesUsed = 0;
1109 dvmUnlockMutex(&gDvmJit.tableLock);
1113 * Float/double conversion requires clamping to min and max of integer form. If
1114 * target doesn't support this normally, use these.
1116 s8 dvmJitd2l(double d)
1118 static const double kMaxLong = (double)(s8)0x7fffffffffffffffULL;
1119 static const double kMinLong = (double)(s8)0x8000000000000000ULL;
1121 return (s8)0x7fffffffffffffffULL;
1122 else if (d <= kMinLong)
1123 return (s8)0x8000000000000000ULL;
1124 else if (d != d) // NaN case
1130 s8 dvmJitf2l(float f)
1132 static const float kMaxLong = (float)(s8)0x7fffffffffffffffULL;
1133 static const float kMinLong = (float)(s8)0x8000000000000000ULL;
1135 return (s8)0x7fffffffffffffffULL;
1136 else if (f <= kMinLong)
1137 return (s8)0x8000000000000000ULL;
1138 else if (f != f) // NaN case
1144 #endif /* WITH_JIT */