static void genInterpSingleStep(CompilationUnit *cUnit, MIR *mir)
{
int flags = dexGetFlagsFromOpcode(mir->dalvikInsn.opcode);
int flagsToCheck = kInstrCanBranch | kInstrCanSwitch | kInstrCanReturn |
kInstrCanThrow;
//If already optimized out, just ignore
if (mir->dalvikInsn.opcode == OP_NOP)
return;
//Ugly, but necessary. Flush all Dalvik regs so Interp can find them
dvmCompilerFlushAllRegs(cUnit);
if ((mir->next == NULL) || (flags & flagsToCheck)) {
genPuntToInterp(cUnit, mir->offset);
return;
}
int entryAddr = offsetof(Thread,
jitToInterpEntries.dvmJitToInterpSingleStep);
loadWordDisp(cUnit, rEBP, 0, rECX); // Get glue
loadWordDisp(cUnit, rECX, entryAddr, rEAX); // rEAX<- entry address
/* rPC = dalvik pc */
loadConstant(cUnit, rPC, (int) (cUnit->method->insns + mir->offset));
/* rECX = dalvik pc of following instruction */
loadConstant(cUnit, rECX, (int) (cUnit->method->insns + mir->next->offset));
/* Pass on the stack */
storeWordDisp(cUnit, rESP, OUT_ARG0, rECX);
opReg(cUnit, kOpCall, rEAX);
}
/*
* Bail to the interpreter. Will not return to this trace.
* On entry, rPC must be set correctly.
*/
static void genPuntToInterp(CompilationUnit *cUnit, unsigned int offset)
{
dvmCompilerFlushAllRegs(cUnit);
loadConstant(cUnit, rPC, (int)(cUnit->method->insns + offset));
loadWordDisp(cUnit, rEBP, 0, rECX); // Get glue
loadWordDisp(cUnit, rECX,
offsetof(Thread, jitToInterpEntries.dvmJitToInterpPunt),
rEAX);
opReg(cUnit, kOpUncondBr, rEAX);
}
/*
* For monitor unlock, we don't have to use ldrex/strex. Once
* we've determined that the lock is thin and that we own it with
* a zero recursion count, it's safe to punch it back to the
* initial, unlock thin state with a store word.
*/
static void genMonitorExit(CompilationUnit *cUnit, MIR *mir)
{
RegLocation rlSrc = dvmCompilerGetSrc(cUnit, mir, 0);
ArmLIR *target;
ArmLIR *branch;
ArmLIR *hopTarget;
ArmLIR *hopBranch;
assert(LW_SHAPE_THIN == 0);
loadValueDirectFixed(cUnit, rlSrc, r1); // Get obj
dvmCompilerLockAllTemps(cUnit); // Prepare for explicit register usage
dvmCompilerFreeTemp(cUnit, r4PC); // Free up r4 for general use
genNullCheck(cUnit, rlSrc.sRegLow, r1, mir->offset, NULL);
loadWordDisp(cUnit, r1, offsetof(Object, lock), r2); // Get object->lock
loadWordDisp(cUnit, r6SELF, offsetof(Thread, threadId), r3); // Get threadId
// Is lock unheld on lock or held by us (==threadId) on unlock?
opRegRegImm(cUnit, kOpAnd, r7, r2,
(LW_HASH_STATE_MASK << LW_HASH_STATE_SHIFT));
opRegImm(cUnit, kOpLsl, r3, LW_LOCK_OWNER_SHIFT); // Align owner
newLIR3(cUnit, kThumb2Bfc, r2, LW_HASH_STATE_SHIFT,
LW_LOCK_OWNER_SHIFT - 1);
opRegReg(cUnit, kOpSub, r2, r3);
hopBranch = opCondBranch(cUnit, kArmCondNe);
dvmCompilerGenMemBarrier(cUnit, kSY);
storeWordDisp(cUnit, r1, offsetof(Object, lock), r7);
branch = opNone(cUnit, kOpUncondBr);
hopTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
hopTarget->defMask = ENCODE_ALL;
hopBranch->generic.target = (LIR *)hopTarget;
// Export PC (part 1)
loadConstant(cUnit, r3, (int) (cUnit->method->insns + mir->offset));
LOAD_FUNC_ADDR(cUnit, r7, (int)dvmUnlockObject);
genRegCopy(cUnit, r0, r6SELF);
// Export PC (part 2)
newLIR3(cUnit, kThumb2StrRRI8Predec, r3, r5FP,
sizeof(StackSaveArea) -
offsetof(StackSaveArea, xtra.currentPc));
opReg(cUnit, kOpBlx, r7);
/* Did we throw? */
ArmLIR *branchOver = genCmpImmBranch(cUnit, kArmCondNe, r0, 0);
loadConstant(cUnit, r0,
(int) (cUnit->method->insns + mir->offset +
dexGetWidthFromOpcode(OP_MONITOR_EXIT)));
genDispatchToHandler(cUnit, TEMPLATE_THROW_EXCEPTION_COMMON);
// Resume here
target = newLIR0(cUnit, kArmPseudoTargetLabel);
target->defMask = ENCODE_ALL;
branch->generic.target = (LIR *)target;
branchOver->generic.target = (LIR *) target;
}
/*
* Handle simple case (thin lock) inline. If it's complicated, bail
* out to the heavyweight lock/unlock routines. We'll use dedicated
* registers here in order to be in the right position in case we
* to bail to dvm[Lock/Unlock]Object(self, object)
*
* r0 -> self pointer [arg0 for dvm[Lock/Unlock]Object
* r1 -> object [arg1 for dvm[Lock/Unlock]Object
* r2 -> intial contents of object->lock, later result of strex
* r3 -> self->threadId
* r7 -> temp to hold new lock value [unlock only]
* r4 -> allow to be used by utilities as general temp
*
* The result of the strex is 0 if we acquire the lock.
*
* See comments in Sync.c for the layout of the lock word.
* Of particular interest to this code is the test for the
* simple case - which we handle inline. For monitor enter, the
* simple case is thin lock, held by no-one. For monitor exit,
* the simple case is thin lock, held by the unlocking thread with
* a recurse count of 0.
*
* A minor complication is that there is a field in the lock word
* unrelated to locking: the hash state. This field must be ignored, but
* preserved.
*
*/
static void genMonitorEnter(CompilationUnit *cUnit, MIR *mir)
{
RegLocation rlSrc = dvmCompilerGetSrc(cUnit, mir, 0);
bool enter = (mir->dalvikInsn.opCode == OP_MONITOR_ENTER);
ArmLIR *target;
ArmLIR *hopTarget;
ArmLIR *branch;
ArmLIR *hopBranch;
assert(LW_SHAPE_THIN == 0);
loadValueDirectFixed(cUnit, rlSrc, r1); // Get obj
dvmCompilerLockAllTemps(cUnit); // Prepare for explicit register usage
dvmCompilerFreeTemp(cUnit, r4PC); // Free up r4 for general use
loadWordDisp(cUnit, rGLUE, offsetof(InterpState, self), r0); // Get self
genNullCheck(cUnit, rlSrc.sRegLow, r1, mir->offset, NULL);
loadWordDisp(cUnit, r0, offsetof(Thread, threadId), r3); // Get threadId
newLIR3(cUnit, kThumb2Ldrex, r2, r1,
offsetof(Object, lock) >> 2); // Get object->lock
opRegImm(cUnit, kOpLsl, r3, LW_LOCK_OWNER_SHIFT); // Align owner
// Is lock unheld on lock or held by us (==threadId) on unlock?
newLIR4(cUnit, kThumb2Bfi, r3, r2, 0, LW_LOCK_OWNER_SHIFT - 1);
newLIR3(cUnit, kThumb2Bfc, r2, LW_HASH_STATE_SHIFT,
LW_LOCK_OWNER_SHIFT - 1);
hopBranch = newLIR2(cUnit, kThumb2Cbnz, r2, 0);
newLIR4(cUnit, kThumb2Strex, r2, r3, r1, offsetof(Object, lock) >> 2);
branch = newLIR2(cUnit, kThumb2Cbz, r2, 0);
hopTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
hopTarget->defMask = ENCODE_ALL;
hopBranch->generic.target = (LIR *)hopTarget;
// Clear the lock
ArmLIR *inst = newLIR0(cUnit, kThumb2Clrex);
// ...and make it a scheduling barrier
inst->defMask = ENCODE_ALL;
// Export PC (part 1)
loadConstant(cUnit, r3, (int) (cUnit->method->insns + mir->offset));
/* Get dPC of next insn */
loadConstant(cUnit, r4PC, (int)(cUnit->method->insns + mir->offset +
dexGetInstrWidthAbs(gDvm.instrWidth, OP_MONITOR_ENTER)));
// Export PC (part 2)
newLIR3(cUnit, kThumb2StrRRI8Predec, r3, rFP,
sizeof(StackSaveArea) -
offsetof(StackSaveArea, xtra.currentPc));
/* Call template, and don't return */
genDispatchToHandler(cUnit, TEMPLATE_MONITOR_ENTER);
// Resume here
target = newLIR0(cUnit, kArmPseudoTargetLabel);
target->defMask = ENCODE_ALL;
branch->generic.target = (LIR *)target;
}
/*
* Load a Dalvik register into a physical register. Take care when
* using this routine, as it doesn't perform any bookkeeping regarding
* register liveness. That is the responsibility of the caller.
*/
static void loadValueDirect(CompilationUnit *cUnit, RegLocation rlSrc,
int reg1)
{
rlSrc = dvmCompilerUpdateLoc(cUnit, rlSrc);
if (rlSrc.location == kLocPhysReg) {
genRegCopy(cUnit, reg1, rlSrc.lowReg);
} else if (rlSrc.location == kLocRetval) {
loadWordDisp(cUnit, rGLUE, offsetof(InterpState, retval), reg1);
} else {
assert(rlSrc.location == kLocDalvikFrame);
loadWordDisp(cUnit, rFP, dvmCompilerS2VReg(cUnit, rlSrc.sRegLow) << 2,
reg1);
}
}
/*
* Rebuild the interpreter frame then punt to the interpreter to execute
* instruction at specified PC.
*
* Currently parameters are passed to the current frame, so we just need to
* grow the stack save area above it, fill certain fields in StackSaveArea and
* Thread that are skipped during whole-method invocation (specified below),
* then return to the interpreter.
*
* StackSaveArea:
* - prevSave
* - prevFrame
* - savedPc
* - returnAddr
* - method
*
* Thread:
* - method
* - methodClassDex
* - curFrame
*/
static void genMethodInflateAndPunt(CompilationUnit *cUnit, MIR *mir,
BasicBlock *bb)
{
int oldStackSave = r0;
int newStackSave = r1;
int oldFP = r2;
int savedPC = r3;
int currentPC = r4PC;
int returnAddr = r7;
int method = r8;
int pDvmDex = r9;
/*
* TODO: check whether to raise the stack overflow exception when growing
* the stack save area.
*/
/* Send everything to home location */
dvmCompilerFlushAllRegs(cUnit);
/* oldStackSave = r5FP + sizeof(current frame) */
opRegRegImm(cUnit, kOpAdd, oldStackSave, r5FP,
cUnit->method->registersSize * 4);
/* oldFP = oldStackSave + sizeof(stackSaveArea) */
opRegRegImm(cUnit, kOpAdd, oldFP, oldStackSave, sizeof(StackSaveArea));
/* newStackSave = r5FP - sizeof(StackSaveArea) */
opRegRegImm(cUnit, kOpSub, newStackSave, r5FP, sizeof(StackSaveArea));
loadWordDisp(cUnit, r13sp, 0, savedPC);
loadConstant(cUnit, currentPC, (int) (cUnit->method->insns + mir->offset));
loadConstant(cUnit, method, (int) cUnit->method);
loadConstant(cUnit, pDvmDex, (int) cUnit->method->clazz->pDvmDex);
#ifdef EASY_GDB
/* newStackSave->prevSave = oldStackSave */
storeWordDisp(cUnit, newStackSave, offsetof(StackSaveArea, prevSave),
oldStackSave);
#endif
/* newStackSave->prevSave = oldStackSave */
storeWordDisp(cUnit, newStackSave, offsetof(StackSaveArea, prevFrame),
oldFP);
/* newStackSave->savedPc = savedPC */
storeWordDisp(cUnit, newStackSave, offsetof(StackSaveArea, savedPc),
savedPC);
/* return address */
loadConstant(cUnit, returnAddr, 0);
storeWordDisp(cUnit, newStackSave, offsetof(StackSaveArea, returnAddr),
returnAddr);
/* newStackSave->method = method */
storeWordDisp(cUnit, newStackSave, offsetof(StackSaveArea, method), method);
/* thread->method = method */
storeWordDisp(cUnit, r6SELF, offsetof(InterpSaveState, method), method);
/* thread->interpSave.curFrame = current FP */
storeWordDisp(cUnit, r6SELF, offsetof(Thread, interpSave.curFrame), r5FP);
/* thread->methodClassDex = pDvmDex */
storeWordDisp(cUnit, r6SELF, offsetof(InterpSaveState, methodClassDex),
pDvmDex);
/* Restore the stack pointer */
opRegImm(cUnit, kOpAdd, r13sp, 16);
genPuntToInterp(cUnit, mir->offset);
}
static RegLocation loadValue(CompilationUnit *cUnit, RegLocation rlSrc,
RegisterClass opKind)
{
rlSrc = dvmCompilerEvalLoc(cUnit, rlSrc, opKind, false);
if (rlSrc.location == kLocDalvikFrame) {
loadValueDirect(cUnit, rlSrc, rlSrc.lowReg);
rlSrc.location = kLocPhysReg;
dvmCompilerMarkLive(cUnit, rlSrc.lowReg, rlSrc.sRegLow);
} else if (rlSrc.location == kLocRetval) {
loadWordDisp(cUnit, rGLUE, offsetof(InterpState, retval), rlSrc.lowReg);
rlSrc.location = kLocPhysReg;
dvmCompilerClobber(cUnit, rlSrc.lowReg);
}
return rlSrc;
}
/*
* Generate array store
*
*/
static void genArrayPut(CompilationUnit *cUnit, MIR *mir, OpSize size,
RegLocation rlArray, RegLocation rlIndex,
RegLocation rlSrc, int scale)
{
RegisterClass regClass = dvmCompilerRegClassBySize(size);
int lenOffset = OFFSETOF_MEMBER(ArrayObject, length);
int dataOffset = OFFSETOF_MEMBER(ArrayObject, contents);
int regPtr;
rlArray = loadValue(cUnit, rlArray, kCoreReg);
rlIndex = loadValue(cUnit, rlIndex, kCoreReg);
if (dvmCompilerIsTemp(cUnit, rlArray.lowReg)) {
dvmCompilerClobber(cUnit, rlArray.lowReg);
regPtr = rlArray.lowReg;
} else {
regPtr = dvmCompilerAllocTemp(cUnit);
genRegCopy(cUnit, regPtr, rlArray.lowReg);
}
/* null object? */
ArmLIR * pcrLabel = NULL;
if (!(mir->OptimizationFlags & MIR_IGNORE_NULL_CHECK)) {
pcrLabel = genNullCheck(cUnit, rlArray.sRegLow, rlArray.lowReg,
mir->offset, NULL);
}
if (!(mir->OptimizationFlags & MIR_IGNORE_RANGE_CHECK)) {
int regLen = dvmCompilerAllocTemp(cUnit);
//NOTE: max live temps(4) here.
/* Get len */
loadWordDisp(cUnit, rlArray.lowReg, lenOffset, regLen);
/* regPtr -> array data */
opRegImm(cUnit, kOpAdd, regPtr, dataOffset);
genBoundsCheck(cUnit, rlIndex.lowReg, regLen, mir->offset,
pcrLabel);
dvmCompilerFreeTemp(cUnit, regLen);
} else {
/* regPtr -> array data */
opRegImm(cUnit, kOpAdd, regPtr, dataOffset);
}
/* at this point, regPtr points to array, 2 live temps */
if ((size == kLong) || (size == kDouble)) {
//TODO: need specific wide routine that can handle fp regs
if (scale) {
int rNewIndex = dvmCompilerAllocTemp(cUnit);
opRegRegImm(cUnit, kOpLsl, rNewIndex, rlIndex.lowReg, scale);
opRegReg(cUnit, kOpAdd, regPtr, rNewIndex);
} else {
opRegReg(cUnit, kOpAdd, regPtr, rlIndex.lowReg);
}
rlSrc = loadValueWide(cUnit, rlSrc, regClass);
HEAP_ACCESS_SHADOW(true);
storePair(cUnit, regPtr, rlSrc.lowReg, rlSrc.highReg);
HEAP_ACCESS_SHADOW(false);
dvmCompilerFreeTemp(cUnit, regPtr);
} else {
rlSrc = loadValue(cUnit, rlSrc, regClass);
HEAP_ACCESS_SHADOW(true);
storeBaseIndexed(cUnit, regPtr, rlIndex.lowReg, rlSrc.lowReg,
scale, size);
HEAP_ACCESS_SHADOW(false);
}
}
/*
* Generate array load
*/
static void genArrayGet(CompilationUnit *cUnit, MIR *mir, OpSize size,
RegLocation rlArray, RegLocation rlIndex,
RegLocation rlDest, int scale)
{
RegisterClass regClass = dvmCompilerRegClassBySize(size);
int lenOffset = OFFSETOF_MEMBER(ArrayObject, length);
int dataOffset = OFFSETOF_MEMBER(ArrayObject, contents);
RegLocation rlResult;
rlArray = loadValue(cUnit, rlArray, kCoreReg);
rlIndex = loadValue(cUnit, rlIndex, kCoreReg);
int regPtr;
/* null object? */
ArmLIR * pcrLabel = NULL;
if (!(mir->OptimizationFlags & MIR_IGNORE_NULL_CHECK)) {
pcrLabel = genNullCheck(cUnit, rlArray.sRegLow,
rlArray.lowReg, mir->offset, NULL);
}
regPtr = dvmCompilerAllocTemp(cUnit);
if (!(mir->OptimizationFlags & MIR_IGNORE_RANGE_CHECK)) {
int regLen = dvmCompilerAllocTemp(cUnit);
/* Get len */
loadWordDisp(cUnit, rlArray.lowReg, lenOffset, regLen);
/* regPtr -> array data */
opRegRegImm(cUnit, kOpAdd, regPtr, rlArray.lowReg, dataOffset);
genBoundsCheck(cUnit, rlIndex.lowReg, regLen, mir->offset,
pcrLabel);
dvmCompilerFreeTemp(cUnit, regLen);
} else {
/* regPtr -> array data */
opRegRegImm(cUnit, kOpAdd, regPtr, rlArray.lowReg, dataOffset);
}
if ((size == kLong) || (size == kDouble)) {
if (scale) {
int rNewIndex = dvmCompilerAllocTemp(cUnit);
opRegRegImm(cUnit, kOpLsl, rNewIndex, rlIndex.lowReg, scale);
opRegReg(cUnit, kOpAdd, regPtr, rNewIndex);
dvmCompilerFreeTemp(cUnit, rNewIndex);
} else {
opRegReg(cUnit, kOpAdd, regPtr, rlIndex.lowReg);
}
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, regClass, true);
HEAP_ACCESS_SHADOW(true);
loadPair(cUnit, regPtr, rlResult.lowReg, rlResult.highReg);
HEAP_ACCESS_SHADOW(false);
dvmCompilerFreeTemp(cUnit, regPtr);
storeValueWide(cUnit, rlDest, rlResult);
} else {
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, regClass, true);
HEAP_ACCESS_SHADOW(true);
loadBaseIndexed(cUnit, regPtr, rlIndex.lowReg, rlResult.lowReg,
scale, size);
HEAP_ACCESS_SHADOW(false);
dvmCompilerFreeTemp(cUnit, regPtr);
storeValue(cUnit, rlDest, rlResult);
}
}
