/*
* 64-bit 3way compare function.
* mov r7, #-1
* cmp op1hi, op2hi
* blt done
* bgt flip
* sub r7, op1lo, op2lo (treat as unsigned)
* beq done
* ite hi
* mov(hi) r7, #-1
* mov(!hi) r7, #1
* flip:
* neg r7
* done:
*/
static void genCmpLong(CompilationUnit *cUnit, MIR *mir,
RegLocation rlDest, RegLocation rlSrc1,
RegLocation rlSrc2)
{
RegLocation rlTemp = LOC_C_RETURN; // Just using as template, will change
ArmLIR *target1;
ArmLIR *target2;
rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg);
rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg);
rlTemp.lowReg = dvmCompilerAllocTemp(cUnit);
loadConstant(cUnit, rlTemp.lowReg, -1);
opRegReg(cUnit, kOpCmp, rlSrc1.highReg, rlSrc2.highReg);
ArmLIR *branch1 = opCondBranch(cUnit, kArmCondLt);
ArmLIR *branch2 = opCondBranch(cUnit, kArmCondGt);
opRegRegReg(cUnit, kOpSub, rlTemp.lowReg, rlSrc1.lowReg, rlSrc2.lowReg);
ArmLIR *branch3 = opCondBranch(cUnit, kArmCondEq);
genIT(cUnit, kArmCondHi, "E");
newLIR2(cUnit, kThumb2MovImmShift, rlTemp.lowReg, modifiedImmediate(-1));
loadConstant(cUnit, rlTemp.lowReg, 1);
genBarrier(cUnit);
target2 = newLIR0(cUnit, kArmPseudoTargetLabel);
target2->defMask = -1;
opRegReg(cUnit, kOpNeg, rlTemp.lowReg, rlTemp.lowReg);
target1 = newLIR0(cUnit, kArmPseudoTargetLabel);
target1->defMask = -1;
storeValue(cUnit, rlDest, rlTemp);
branch1->generic.target = (LIR *)target1;
branch2->generic.target = (LIR *)target2;
branch3->generic.target = branch1->generic.target;
}
static bool genCmpFP(CompilationUnit *cUnit, MIR *mir, RegLocation rlDest,
RegLocation rlSrc1, RegLocation rlSrc2)
{
bool isDouble;
int defaultResult;
bool ltNaNBias;
RegLocation rlResult;
switch(mir->dalvikInsn.opCode) {
case OP_CMPL_FLOAT:
isDouble = false;
defaultResult = -1;
break;
case OP_CMPG_FLOAT:
isDouble = false;
defaultResult = 1;
break;
case OP_CMPL_DOUBLE:
isDouble = true;
defaultResult = -1;
break;
case OP_CMPG_DOUBLE:
isDouble = true;
defaultResult = 1;
break;
default:
return true;
}
if (isDouble) {
rlSrc1 = loadValueWide(cUnit, rlSrc1, kFPReg);
rlSrc2 = loadValueWide(cUnit, rlSrc2, kFPReg);
dvmCompilerClobberSReg(cUnit, rlDest.sRegLow);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kCoreReg, true);
loadConstant(cUnit, rlResult.lowReg, defaultResult);
newLIR2(cUnit, kThumb2Vcmpd, S2D(rlSrc1.lowReg, r1Src2.highReg),
S2D(rlSrc2.lowReg, rlSrc2.highReg));
} else {
rlSrc1 = loadValue(cUnit, rlSrc1, kFPReg);
rlSrc2 = loadValue(cUnit, rlSrc2, kFPReg);
dvmCompilerClobberSReg(cUnit, rlDest.sRegLow);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kCoreReg, true);
loadConstant(cUnit, rlResult.lowReg, defaultResult);
newLIR2(cUnit, kThumb2Vcmps, rlSrc1.lowReg, rlSrc2.lowReg);
}
assert(!FPREG(rlResult.lowReg));
newLIR0(cUnit, kThumb2Fmstat);
genIT(cUnit, (defaultResult == -1) ? kArmCondGt : kArmCondMi, "");
newLIR2(cUnit, kThumb2MovImmShift, rlResult.lowReg,
modifiedImmediate(-defaultResult)); // Must not alter ccodes
genIT(cUnit, kArmCondEq, "");
loadConstant(cUnit, rlResult.lowReg, 0);
storeValue(cUnit, rlDest, rlResult);
return false;
}
static void genLong3Addr(CompilationUnit *cUnit, MIR *mir, OpKind firstOp,
OpKind secondOp, RegLocation rlDest,
RegLocation rlSrc1, RegLocation rlSrc2)
{
RegLocation rlResult;
rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg);
rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kCoreReg, true);
opRegRegReg(cUnit, firstOp, rlResult.lowReg, rlSrc1.lowReg, rlSrc2.lowReg);
opRegRegReg(cUnit, secondOp, rlResult.highReg, rlSrc1.highReg,
rlSrc2.highReg);
storeValueWide(cUnit, rlDest, rlResult);
}
static bool genInlineSqrt(CompilationUnit *cUnit, MIR *mir)
{
ArmLIR *branch;
RegLocation rlSrc = dvmCompilerGetSrcWide(cUnit, mir, 0, 1);
RegLocation rlDest = inlinedTargetWide(cUnit, mir, true);
rlSrc = loadValueWide(cUnit, rlSrc, kFPReg);
RegLocation rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
newLIR2(cUnit, kThumb2Vsqrtd, S2D(rlResult.lowReg, rlResult.highReg),
S2D(rlSrc.lowReg, rlSrc.highReg));
newLIR2(cUnit, kThumb2Vcmpd, S2D(rlResult.lowReg, rlResult.highReg),
S2D(rlResult.lowReg, rlResult.highReg));
newLIR0(cUnit, kThumb2Fmstat);
branch = newLIR2(cUnit, kThumbBCond, 0, kArmCondEq);
dvmCompilerClobberCallRegs(cUnit);
LOAD_FUNC_ADDR(cUnit, r2, (int) (double (*)(double)) sqrt);
newLIR3(cUnit, kThumb2Fmrrd, r0, r1, S2D(rlSrc.lowReg, rlSrc.highReg));
newLIR1(cUnit, kThumbBlxR, r2);
newLIR3(cUnit, kThumb2Fmdrr, S2D(rlResult.lowReg, rlResult.highReg),
r0, r1);
ArmLIR *label = newLIR0(cUnit, kArmPseudoTargetLabel);
label->defMask = ENCODE_ALL;
branch->generic.target = (LIR *)label;
storeValueWide(cUnit, rlDest, rlResult);
return false;
}
static bool genArithOpDouble(CompilationUnit *cUnit, MIR *mir,
RegLocation rlDest, RegLocation rlSrc1,
RegLocation rlSrc2)
{
int op = kThumbBkpt;
RegLocation rlResult;
switch (mir->dalvikInsn.opcode) {
case OP_ADD_DOUBLE_2ADDR:
case OP_ADD_DOUBLE:
op = kThumb2Vaddd;
break;
case OP_SUB_DOUBLE_2ADDR:
case OP_SUB_DOUBLE:
op = kThumb2Vsubd;
break;
case OP_DIV_DOUBLE_2ADDR:
case OP_DIV_DOUBLE:
op = kThumb2Vdivd;
break;
case OP_MUL_DOUBLE_2ADDR:
case OP_MUL_DOUBLE:
op = kThumb2Vmuld;
break;
case OP_REM_DOUBLE_2ADDR:
case OP_REM_DOUBLE:
case OP_NEG_DOUBLE: {
return genArithOpDoublePortable(cUnit, mir, rlDest, rlSrc1,
rlSrc2);
}
default:
return true;
}
rlSrc1 = loadValueWide(cUnit, rlSrc1, kFPReg);
assert(rlSrc1.wide);
rlSrc2 = loadValueWide(cUnit, rlSrc2, kFPReg);
assert(rlSrc2.wide);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
assert(rlDest.wide);
assert(rlResult.wide);
newLIR3(cUnit, (ArmOpcode)op, S2D(rlResult.lowReg, rlResult.highReg),
S2D(rlSrc1.lowReg, rlSrc1.highReg),
S2D(rlSrc2.lowReg, rlSrc2.highReg));
storeValueWide(cUnit, rlDest, rlResult);
return false;
}
static void genLong3Addr(CompilationUnit *cUnit, MIR *mir, OpKind firstOp,
OpKind secondOp, RegLocation rlDest,
RegLocation rlSrc1, RegLocation rlSrc2)
{
RegLocation rlResult;
if (partialOverlap(rlSrc1.sRegLow,rlSrc2.sRegLow) ||
partialOverlap(rlSrc1.sRegLow,rlDest.sRegLow) ||
partialOverlap(rlSrc2.sRegLow,rlDest.sRegLow)) {
// Rare case - not enough registers to properly handle
genInterpSingleStep(cUnit, mir);
} else if (rlDest.sRegLow == rlSrc1.sRegLow) {
// Already 2-operand
rlResult = loadValueWide(cUnit, rlDest, kCoreReg);
rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg);
opRegReg(cUnit, firstOp, rlResult.lowReg, rlSrc2.lowReg);
opRegReg(cUnit, secondOp, rlResult.highReg, rlSrc2.highReg);
storeValueWide(cUnit, rlDest, rlResult);
} else if (rlDest.sRegLow == rlSrc2.sRegLow) {
// Bad case - must use/clobber Src1 and reassign Dest
rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg);
rlResult = loadValueWide(cUnit, rlDest, kCoreReg);
opRegReg(cUnit, firstOp, rlSrc1.lowReg, rlResult.lowReg);
opRegReg(cUnit, secondOp, rlSrc1.highReg, rlResult.highReg);
// Old reg assignments are now invalid
dvmCompilerClobber(cUnit, rlResult.lowReg);
dvmCompilerClobber(cUnit, rlResult.highReg);
dvmCompilerClobber(cUnit, rlSrc1.lowReg);
dvmCompilerClobber(cUnit, rlSrc1.highReg);
rlDest.location = kLocDalvikFrame;
assert(rlSrc1.location == kLocPhysReg);
// Reassign registers - rlDest will now get rlSrc1's old regs
storeValueWide(cUnit, rlDest, rlSrc1);
} else {
// Copy Src1 to Dest
rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kCoreReg, false);
loadValueDirectWide(cUnit, rlSrc1, rlResult.lowReg,
rlResult.highReg);
rlResult.location = kLocPhysReg;
opRegReg(cUnit, firstOp, rlResult.lowReg, rlSrc2.lowReg);
opRegReg(cUnit, secondOp, rlResult.highReg, rlSrc2.highReg);
storeValueWide(cUnit, rlDest, rlResult);
}
}
static void genNegDouble(CompilationUnit *cUnit, RegLocation rlDest,
RegLocation rlSrc)
{
RegLocation rlResult;
rlSrc = loadValueWide(cUnit, rlSrc, kFPReg);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
newLIR2(cUnit, kThumb2Vnegd, S2D(rlResult.lowReg, rlResult.highReg),
S2D(rlSrc.lowReg, rlSrc.highReg));
storeValueWide(cUnit, rlDest, rlResult);
}
static void genNegDouble(CompilationUnit *cUnit, RegLocation rlDest,
RegLocation rlSrc)
{
RegLocation rlResult;
rlSrc = loadValueWide(cUnit, rlSrc, kCoreReg);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kCoreReg, true);
opRegRegImm(cUnit, kOpAdd, rlResult.highReg, rlSrc.highReg,
0x80000000);
genRegCopy(cUnit, rlResult.lowReg, rlSrc.lowReg);
storeValueWide(cUnit, rlDest, rlResult);
}
static bool genInlinedAbsDouble(CompilationUnit *cUnit, MIR *mir)
{
RegLocation rlSrc = dvmCompilerGetSrcWide(cUnit, mir, 0, 1);
RegLocation rlDest = inlinedTargetWide(cUnit, mir, true);
rlSrc = loadValueWide(cUnit, rlSrc, kFPReg);
RegLocation rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
newLIR2(cUnit, kThumb2Vabsd, S2D(rlResult.lowReg, rlResult.highReg),
S2D(rlSrc.lowReg, rlSrc.highReg));
storeValueWide(cUnit, rlDest, rlResult);
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);
}
static bool genInlinedAbsDouble(CompilationUnit *cUnit, MIR *mir)
{
int offset = offsetof(InterpState, 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, rGLUE, offset, reglo);
newLIR2(cUnit, kThumbAndRR, reghi, signMask);
dvmCompilerFreeTemp(cUnit, signMask);
storeWordDisp(cUnit, rGLUE, offset + 4, reghi);
//TUNING: rewrite this to not clobber
dvmCompilerClobber(cUnit, reghi);
return true;
}
/*
* 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);
}
}
static bool genConversion(CompilationUnit *cUnit, MIR *mir)
{
Opcode opcode = mir->dalvikInsn.opcode;
int op = kThumbBkpt;
bool longSrc = false;
bool longDest = false;
int srcReg;
RegLocation rlSrc;
RegLocation rlDest;
RegLocation rlResult;
switch (opcode) {
case OP_INT_TO_FLOAT:
longSrc = false;
longDest = false;
op = kThumb2VcvtIF;
break;
case OP_FLOAT_TO_INT:
longSrc = false;
longDest = false;
op = kThumb2VcvtFI;
break;
case OP_DOUBLE_TO_FLOAT:
longSrc = true;
longDest = false;
op = kThumb2VcvtDF;
break;
case OP_FLOAT_TO_DOUBLE:
longSrc = false;
longDest = true;
op = kThumb2VcvtFd;
break;
case OP_INT_TO_DOUBLE:
longSrc = false;
longDest = true;
op = kThumb2VcvtID;
break;
case OP_DOUBLE_TO_INT:
longSrc = true;
longDest = false;
op = kThumb2VcvtDI;
break;
case OP_LONG_TO_DOUBLE:
case OP_FLOAT_TO_LONG:
case OP_LONG_TO_FLOAT:
case OP_DOUBLE_TO_LONG:
return genConversionPortable(cUnit, mir);
default:
return true;
}
if (longSrc) {
rlSrc = dvmCompilerGetSrcWide(cUnit, mir, 0, 1);
rlSrc = loadValueWide(cUnit, rlSrc, kFPReg);
srcReg = S2D(rlSrc.lowReg, rlSrc.highReg);
} else {
rlSrc = dvmCompilerGetSrc(cUnit, mir, 0);
rlSrc = loadValue(cUnit, rlSrc, kFPReg);
srcReg = rlSrc.lowReg;
}
if (longDest) {
rlDest = dvmCompilerGetDestWide(cUnit, mir, 0, 1);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
newLIR2(cUnit, (ArmOpcode)op, S2D(rlResult.lowReg, rlResult.highReg),
srcReg);
storeValueWide(cUnit, rlDest, rlResult);
} else {
rlDest = dvmCompilerGetDest(cUnit, mir, 0);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
newLIR2(cUnit, (ArmOpcode)op, rlResult.lowReg, srcReg);
storeValue(cUnit, rlDest, rlResult);
}
return false;
}