/*
* Load a immediate using a shortcut if possible; otherwise
* grab from the per-translation literal pool. If target is
* a high register, build constant into a low register and copy.
*
* No additional register clobbering operation performed. Use this version when
* 1) rDest is freshly returned from dvmCompilerAllocTemp or
* 2) The codegen is under fixed register usage
*/
static ArmLIR *loadConstantNoClobber(CompilationUnit *cUnit, int rDest,
int value)
{
ArmLIR *res;
int tDest = LOWREG(rDest) ? rDest : dvmCompilerAllocTemp(cUnit);
/* See if the value can be constructed cheaply */
if ((value >= 0) && (value <= 255)) {
res = newLIR2(cUnit, kThumbMovImm, tDest, value);
if (rDest != tDest) {
opRegReg(cUnit, kOpMov, rDest, tDest);
dvmCompilerFreeTemp(cUnit, tDest);
}
return res;
} else if ((value & 0xFFFFFF00) == 0xFFFFFF00) {
res = newLIR2(cUnit, kThumbMovImm, tDest, ~value);
newLIR2(cUnit, kThumbMvn, tDest, tDest);
if (rDest != tDest) {
opRegReg(cUnit, kOpMov, rDest, tDest);
dvmCompilerFreeTemp(cUnit, tDest);
}
return res;
}
/* No shortcut - go ahead and use literal pool */
ArmLIR *dataTarget = scanLiteralPool(cUnit->literalList, value, 255);
if (dataTarget == NULL) {
dataTarget = addWordData(cUnit, &cUnit->literalList, value);
}
ArmLIR *loadPcRel = (ArmLIR *) dvmCompilerNew(sizeof(ArmLIR), true);
loadPcRel->opcode = kThumbLdrPcRel;
loadPcRel->generic.target = (LIR *) dataTarget;
loadPcRel->operands[0] = tDest;
setupResourceMasks(loadPcRel);
setMemRefType(loadPcRel, true, kLiteral);
loadPcRel->aliasInfo = dataTarget->operands[0];
res = loadPcRel;
dvmCompilerAppendLIR(cUnit, (LIR *) loadPcRel);
/*
* To save space in the constant pool, we use the ADD_RRI8 instruction to
* add up to 255 to an existing constant value.
*/
if (dataTarget->operands[0] != value) {
newLIR2(cUnit, kThumbAddRI8, tDest, value - dataTarget->operands[0]);
}
if (rDest != tDest) {
opRegReg(cUnit, kOpMov, rDest, tDest);
dvmCompilerFreeTemp(cUnit, tDest);
}
return res;
}
/*
* 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;
}
/* Export the Dalvik PC assicated with an instruction to the StackSave area */
static ArmLIR *genExportPC(CompilationUnit *cUnit, MIR *mir)
{
ArmLIR *res;
int offset = offsetof(StackSaveArea, xtra.currentPc);
int rDPC = dvmCompilerAllocTemp(cUnit);
res = loadConstant(cUnit, rDPC, (int) (cUnit->method->insns + mir->offset));
newLIR3(cUnit, kThumb2StrRRI8Predec, rDPC, r5FP,
sizeof(StackSaveArea) - offset);
dvmCompilerFreeTemp(cUnit, rDPC);
return res;
}
/*
* 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;
}
static bool genInlinedAbsFloat(CompilationUnit *cUnit, MIR *mir)
{
int offset = offsetof(InterpState, retval);
RegLocation rlSrc = dvmCompilerGetSrc(cUnit, mir, 0);
int reg0 = loadValue(cUnit, rlSrc, kCoreReg).lowReg;
int signMask = dvmCompilerAllocTemp(cUnit);
loadConstant(cUnit, signMask, 0x7fffffff);
newLIR2(cUnit, kThumbAndRR, reg0, signMask);
dvmCompilerFreeTemp(cUnit, signMask);
storeWordDisp(cUnit, rGLUE, offset, reg0);
//TUNING: rewrite this to not clobber
dvmCompilerClobber(cUnit, reg0);
return true;
}
static bool genInlinedAbsDouble(CompilationUnit *cUnit, MIR *mir)
{
int offset = offsetof(Thread, interpSave.retval);
RegLocation rlSrc = dvmCompilerGetSrcWide(cUnit, mir, 0, 1);
RegLocation regSrc = loadValueWide(cUnit, rlSrc, kCoreReg);
int reglo = regSrc.lowReg;
int reghi = regSrc.highReg;
int signMask = dvmCompilerAllocTemp(cUnit);
loadConstant(cUnit, signMask, 0x7fffffff);
storeWordDisp(cUnit, r6SELF, offset, reglo);
newLIR2(cUnit, kThumbAndRR, reghi, signMask);
dvmCompilerFreeTemp(cUnit, signMask);
storeWordDisp(cUnit, r6SELF, offset + 4, reghi);
//TUNING: rewrite this to not clobber
dvmCompilerClobber(cUnit, reghi);
return false;
}
/*
* To avoid possible conflicts, we use a lot of temps here. Note that
* our usage of Thumb2 instruction forms avoids the problems with register
* reuse for multiply instructions prior to arm6.
*/
static void genMulLong(CompilationUnit *cUnit, RegLocation rlDest,
RegLocation rlSrc1, RegLocation rlSrc2)
{
RegLocation rlResult;
int resLo = dvmCompilerAllocTemp(cUnit);
int resHi = dvmCompilerAllocTemp(cUnit);
int tmp1 = dvmCompilerAllocTemp(cUnit);
rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg);
rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg);
newLIR3(cUnit, kThumb2MulRRR, tmp1, rlSrc2.lowReg, rlSrc1.highReg);
newLIR4(cUnit, kThumb2Umull, resLo, resHi, rlSrc2.lowReg, rlSrc1.lowReg);
newLIR4(cUnit, kThumb2Mla, tmp1, rlSrc1.lowReg, rlSrc2.highReg, tmp1);
newLIR4(cUnit, kThumb2AddRRR, resHi, tmp1, resHi, 0);
dvmCompilerFreeTemp(cUnit, tmp1);
rlResult = dvmCompilerGetReturnWide(cUnit); // Just as a template, will patch
rlResult.lowReg = resLo;
rlResult.highReg = resHi;
storeValueWide(cUnit, rlDest, rlResult);
}
/*
* 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);
}
}