/*
* 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 void genMultiplyByShiftAndReverseSubtract(CompilationUnit *cUnit,
RegLocation rlSrc, RegLocation rlResult, int lit)
{
int tReg = dvmCompilerAllocTemp(cUnit);
opRegRegImm(cUnit, kOpLsl, tReg, rlSrc.lowReg, lit);
opRegRegReg(cUnit, kOpSub, rlResult.lowReg, tReg, rlSrc.lowReg);
}
static void genMultiplyByTwoBitMultiplier(CompilationUnit *cUnit,
RegLocation rlSrc, RegLocation rlResult, int lit,
int firstBit, int secondBit)
{
// We can't implement "add src, src, src, lsl#shift" on Thumb, so we have
// to do a regular multiply.
opRegRegImm(cUnit, kOpMul, rlResult.lowReg, rlSrc.lowReg, lit);
}
/*
* Perform a "reg cmp imm" operation and jump to the PCR region if condition
* satisfies.
*/
static void genNegFloat(CompilationUnit *cUnit, RegLocation rlDest,
RegLocation rlSrc)
{
RegLocation rlResult;
rlSrc = loadValue(cUnit, rlSrc, kCoreReg);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kCoreReg, true);
opRegRegImm(cUnit, kOpAdd, rlResult.lowReg,
rlSrc.lowReg, 0x80000000);
storeValue(cUnit, rlDest, rlResult);
}
static void genMultiplyByTwoBitMultiplier(CompilationUnit *cUnit,
RegLocation rlSrc, RegLocation rlResult, int lit,
int firstBit, int secondBit)
{
opRegRegRegShift(cUnit, kOpAdd, rlResult.lowReg, rlSrc.lowReg, rlSrc.lowReg,
encodeShift(kArmLsl, secondBit - firstBit));
if (firstBit != 0) {
opRegRegImm(cUnit, kOpLsl, rlResult.lowReg, rlResult.lowReg, firstBit);
}
}
/*
* 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;
}
/*
* Take the address of a Dalvik register and store it into rDest.
* Clobber any live values associated either with the Dalvik value
* or the target register and lock the target fixed register.
*/
static void loadValueAddressDirect(CompilationUnit *cUnit, RegLocation rlSrc,
int rDest)
{
rlSrc = rlSrc.wide ? dvmCompilerUpdateLocWide(cUnit, rlSrc) :
dvmCompilerUpdateLoc(cUnit, rlSrc);
if (rlSrc.location == kLocPhysReg) {
if (rlSrc.wide) {
dvmCompilerFlushRegWideForV5TEVFP(cUnit, rlSrc.lowReg,
rlSrc.highReg);
} else {
dvmCompilerFlushRegForV5TEVFP(cUnit, rlSrc.lowReg);
}
}
dvmCompilerClobber(cUnit, rDest);
dvmCompilerLockTemp(cUnit, rDest);
opRegRegImm(cUnit, kOpAdd, rDest, rFP,
dvmCompilerS2VReg(cUnit, rlSrc.sRegLow) << 2);
}
/*
* 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);
}
}
