static int genTraceProfileEntry(CompilationUnit *cUnit)
{
intptr_t addr = (intptr_t)dvmJitNextTraceCounter();
assert(__BYTE_ORDER == __LITTLE_ENDIAN);
newLIR1(cUnit, kArm16BitData, addr & 0xffff);
newLIR1(cUnit, kArm16BitData, (addr >> 16) & 0xffff);
cUnit->chainCellOffsetLIR =
(LIR *) newLIR1(cUnit, kArm16BitData, CHAIN_CELL_OFFSET_TAG);
cUnit->headerSize = 6;
if ((gDvmJit.profileMode == kTraceProfilingContinuous) ||
(gDvmJit.profileMode == kTraceProfilingDisabled)) {
/* Thumb[2] instruction used directly here to ensure correct size */
newLIR2(cUnit, kThumb2LdrPcReln12, r0, 8);
newLIR3(cUnit, kThumbLdrRRI5, r1, r0, 0);
newLIR2(cUnit, kThumbAddRI8, r1, 1);
newLIR3(cUnit, kThumbStrRRI5, r1, r0, 0);
return 10;
} else {
int opcode = TEMPLATE_PERIODIC_PROFILING;
newLIR2(cUnit, kThumbBlx1,
(int) gDvmJit.codeCache + templateEntryOffsets[opcode],
(int) gDvmJit.codeCache + templateEntryOffsets[opcode]);
newLIR2(cUnit, kThumbBlx2,
(int) gDvmJit.codeCache + templateEntryOffsets[opcode],
(int) gDvmJit.codeCache + templateEntryOffsets[opcode]);
return 4;
}
}
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 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;
}
/* Export the Dalvik PC assicated with an instruction to the StackSave area */
static ArmLIR *genExportPC(CompilationUnit *cUnit, MIR *mir)
{
ArmLIR *res;
int rDPC = dvmCompilerAllocTemp(cUnit);
int rAddr = dvmCompilerAllocTemp(cUnit);
int offset = offsetof(StackSaveArea, xtra.currentPc);
res = loadConstant(cUnit, rDPC, (int) (cUnit->method->insns + mir->offset));
newLIR2(cUnit, kThumbMovRR, rAddr, rFP);
newLIR2(cUnit, kThumbSubRI8, rAddr, sizeof(StackSaveArea) - offset);
storeWordDisp( cUnit, rAddr, 0, 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;
}
/*
* 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;
}
/*
* Generate a Thumb2 IT instruction, which can nullify up to
* four subsequent instructions based on a condition and its
* inverse. The condition applies to the first instruction, which
* is executed if the condition is met. The string "guide" consists
* of 0 to 3 chars, and applies to the 2nd through 4th instruction.
* A "T" means the instruction is executed if the condition is
* met, and an "E" means the instruction is executed if the condition
* is not met.
*/
static ArmLIR *genIT(CompilationUnit *cUnit, ArmConditionCode code,
const char *guide)
{
int mask;
int condBit = code & 1;
int altBit = condBit ^ 1;
int mask3 = 0;
int mask2 = 0;
int mask1 = 0;
//Note: case fallthroughs intentional
switch(strlen(guide)) {
case 3:
mask1 = (guide[2] == 'T') ? condBit : altBit;
case 2:
mask2 = (guide[1] == 'T') ? condBit : altBit;
case 1:
mask3 = (guide[0] == 'T') ? condBit : altBit;
break;
case 0:
break;
default:
LOGE("Jit: bad case in genIT");
dvmCompilerAbort(cUnit);
}
mask = (mask3 << 3) | (mask2 << 2) | (mask1 << 1) |
(1 << (3 - strlen(guide)));
return newLIR2(cUnit, kThumb2It, code, mask);
}
/*
* 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;
}
/*
* Jump to the out-of-line handler in ARM mode to finish executing the
* remaining of more complex instructions.
*/
static void genDispatchToHandler(CompilationUnit *cUnit, TemplateOpCode opCode)
{
/*
* NOTE - In practice BLX only needs one operand, but since the assembler
* may abort itself and retry due to other out-of-range conditions we
* cannot really use operand[0] to store the absolute target address since
* it may get clobbered by the final relative offset. Therefore,
* we fake BLX_1 is a two operand instruction and the absolute target
* address is stored in operand[1].
*/
dvmCompilerClobberHandlerRegs(cUnit);
newLIR2(cUnit, kThumbBlx1,
(int) gDvmJit.codeCache + templateEntryOffsets[opCode],
(int) gDvmJit.codeCache + templateEntryOffsets[opCode]);
newLIR2(cUnit, kThumbBlx2,
(int) gDvmJit.codeCache + templateEntryOffsets[opCode],
(int) gDvmJit.codeCache + templateEntryOffsets[opCode]);
}
static void genNegFloat(CompilationUnit *cUnit, RegLocation rlDest,
RegLocation rlSrc)
{
RegLocation rlResult;
rlSrc = loadValue(cUnit, rlSrc, kFPReg);
rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
newLIR2(cUnit, kThumb2Vnegs, rlResult.lowReg, rlSrc.lowReg);
storeValue(cUnit, rlDest, rlResult);
}
/* No select in thumb, so we need to branch. Thumb2 will do better */
static bool genInlinedMinMaxInt(CompilationUnit *cUnit, MIR *mir, bool isMin)
{
int offset = offsetof(InterpState, retval);
RegLocation rlSrc1 = dvmCompilerGetSrc(cUnit, mir, 0);
RegLocation rlSrc2 = dvmCompilerGetSrc(cUnit, mir, 1);
int reg0 = loadValue(cUnit, rlSrc1, kCoreReg).lowReg;
int reg1 = loadValue(cUnit, rlSrc2, kCoreReg).lowReg;
newLIR2(cUnit, kThumbCmpRR, reg0, reg1);
ArmLIR *branch1 = newLIR2(cUnit, kThumbBCond, 2,
isMin ? kArmCondLt : kArmCondGt);
newLIR2(cUnit, kThumbMovRR, reg0, reg1);
ArmLIR *target = newLIR0(cUnit, kArmPseudoTargetLabel);
target->defMask = ENCODE_ALL;
newLIR3(cUnit, kThumbStrRRI5, reg0, rGLUE, offset >> 2);
branch1->generic.target = (LIR *)target;
//TUNING: rewrite this to not clobber
dvmCompilerClobber(cUnit,reg0);
return false;
}
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 bool genInlinedAbsFloat(CompilationUnit *cUnit, MIR *mir)
{
RegLocation rlSrc = dvmCompilerGetSrc(cUnit, mir, 0);
RegLocation rlDest = inlinedTarget(cUnit, mir, true);
rlSrc = loadValue(cUnit, rlSrc, kFPReg);
RegLocation rlResult = dvmCompilerEvalLoc(cUnit, rlDest, kFPReg, true);
newLIR2(cUnit, kThumb2Vabss, rlResult.lowReg, rlSrc.lowReg);
storeValue(cUnit, rlDest, rlResult);
return false;
}
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;
}
static bool genInlineSqrt(CompilationUnit *cUnit, MIR *mir)
{
int offset = offsetof(InterpState, retval);
int vSrc = mir->dalvikInsn.vA;
loadDouble(cUnit, vSrc, dr1);
newLIR2(cUnit, THUMB2_VSQRTD, dr0, dr1);
assert(offset & 0x3 == 0); /* Must be word aligned */
assert(offset < 1024);
newLIR3(cUnit, THUMB2_VSTRD, dr0, rGLUE, offset >> 2);
return true;
}
/*
* Jump to the out-of-line handler in ARM mode to finish executing the
* remaining of more complex instructions.
*/
static void genDispatchToHandler(CompilationUnit *cUnit, TemplateOpCode opCode)
{
#if USE_IN_CACHE_HANDLER
/*
* NOTE - In practice BLX only needs one operand, but since the assembler
* may abort itself and retry due to other out-of-range conditions we
* cannot really use operand[0] to store the absolute target address since
* it may get clobbered by the final relative offset. Therefore,
* we fake BLX_1 is a two operand instruction and the absolute target
* address is stored in operand[1].
*/
newLIR2(cUnit, THUMB_BLX_1,
(int) gDvmJit.codeCache + templateEntryOffsets[opCode],
(int) gDvmJit.codeCache + templateEntryOffsets[opCode]);
newLIR2(cUnit, THUMB_BLX_2,
(int) gDvmJit.codeCache + templateEntryOffsets[opCode],
(int) gDvmJit.codeCache + templateEntryOffsets[opCode]);
#else
/*
* In case we want to access the statically compiled handlers for
* debugging purposes, define USE_IN_CACHE_HANDLER to 0
*/
void *templatePtr;
#define JIT_TEMPLATE(X) extern void dvmCompiler_TEMPLATE_##X();
#include "../../../template/armv5te-vfp/TemplateOpList.h"
#undef JIT_TEMPLATE
switch (opCode) {
#define JIT_TEMPLATE(X) \
case TEMPLATE_##X: { templatePtr = dvmCompiler_TEMPLATE_##X; break; }
#include "../../../template/armv5te-vfp/TemplateOpList.h"
#undef JIT_TEMPLATE
default: templatePtr = NULL;
}
loadConstant(cUnit, r7, (int) templatePtr);
newLIR1(cUnit, THUMB_BLX_R, r7);
#endif
}
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;
}
/*
* Jump to the out-of-line handler to finish executing the
* remaining of more complex instructions.
*/
static void genDispatchToHandler(CompilationUnit *cUnit, TemplateOpcode opCode)
{
/*
* We're jumping from a trace to a template. Using jal is preferable to jalr,
* but we need to ensure source and target addresses allow the use of jal.
* This should almost always be the case, but if source and target are in
* different 256mb regions then use jalr. The test below is very conservative
* since we don't have a source address yet, but this is ok for now given that
* we expect this case to be very rare. The test can be made less conservative
* as needed in the future in coordination with address assignment during
* the assembly process.
*/
dvmCompilerClobberHandlerRegs(cUnit);
int targetAddr = (int) gDvmJit.codeCache + templateEntryOffsets[opCode];
int maxSourceAddr = (int) gDvmJit.codeCache + gDvmJit.codeCacheSize;
if ((targetAddr & 0xF0000000) == (maxSourceAddr & 0xF0000000)) {
newLIR1(cUnit, kMipsJal, targetAddr);
} else {
loadConstant(cUnit, r_T9, targetAddr);
newLIR2(cUnit, kMipsJalr, r_RA, r_T9);
}
}
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;
}
static ArmLIR *opCondBranch(CompilationUnit *cUnit, ArmConditionCode cc)
{
return newLIR2(cUnit, kThumbBCond, 0 /* offset to be patched */, cc);
}
/* Handle the content in each basic block */
static bool methodBlockCodeGen(CompilationUnit *cUnit, BasicBlock *bb)
{
MIR *mir;
ArmLIR *labelList = (ArmLIR *) cUnit->blockLabelList;
int blockId = bb->id;
cUnit->curBlock = bb;
labelList[blockId].operands[0] = bb->startOffset;
/* Insert the block label */
labelList[blockId].opcode = kArmPseudoNormalBlockLabel;
dvmCompilerAppendLIR(cUnit, (LIR *) &labelList[blockId]);
dvmCompilerClobberAllRegs(cUnit);
dvmCompilerResetNullCheck(cUnit);
ArmLIR *headLIR = NULL;
if (bb->blockType == kEntryBlock) {
/* r0 = callsitePC */
opImm(cUnit, kOpPush, (1 << r0 | 1 << r1 | 1 << r5FP | 1 << r14lr));
opRegImm(cUnit, kOpSub, r5FP,
sizeof(StackSaveArea) + cUnit->method->registersSize * 4);
} else if (bb->blockType == kExitBlock) {
/* No need to pop r0 and r1 */
opRegImm(cUnit, kOpAdd, r13sp, 8);
opImm(cUnit, kOpPop, (1 << r5FP | 1 << r15pc));
}
for (mir = bb->firstMIRInsn; mir; mir = mir->next) {
dvmCompilerResetRegPool(cUnit);
if (gDvmJit.disableOpt & (1 << kTrackLiveTemps)) {
dvmCompilerClobberAllRegs(cUnit);
}
if (gDvmJit.disableOpt & (1 << kSuppressLoads)) {
dvmCompilerResetDefTracking(cUnit);
}
Opcode dalvikOpcode = mir->dalvikInsn.opcode;
InstructionFormat dalvikFormat =
dexGetFormatFromOpcode(dalvikOpcode);
ArmLIR *boundaryLIR;
/*
* Don't generate the boundary LIR unless we are debugging this
* trace or we need a scheduling barrier.
*/
if (headLIR == NULL || cUnit->printMe == true) {
boundaryLIR =
newLIR2(cUnit, kArmPseudoDalvikByteCodeBoundary,
mir->offset,
(int) dvmCompilerGetDalvikDisassembly(
&mir->dalvikInsn, ""));
/* Remember the first LIR for this block */
if (headLIR == NULL) {
headLIR = boundaryLIR;
/* Set the first boundaryLIR as a scheduling barrier */
headLIR->defMask = ENCODE_ALL;
}
}
/* Don't generate the SSA annotation unless verbose mode is on */
if (cUnit->printMe && mir->ssaRep) {
char *ssaString = dvmCompilerGetSSAString(cUnit, mir->ssaRep);
newLIR1(cUnit, kArmPseudoSSARep, (int) ssaString);
}
bool notHandled;
switch (dalvikFormat) {
case kFmt10t:
case kFmt20t:
case kFmt30t:
notHandled = handleMethodFmt10t_Fmt20t_Fmt30t(cUnit, mir, bb,
labelList);
break;
case kFmt10x:
notHandled = handleMethodFmt10x(cUnit, mir);
break;
case kFmt11n:
case kFmt31i:
notHandled = handleMethodFmt11n_Fmt31i(cUnit, mir);
break;
case kFmt11x:
notHandled = handleMethodFmt11x(cUnit, mir, bb, labelList);
break;
case kFmt12x:
notHandled = handleMethodFmt12x(cUnit, mir);
break;
case kFmt20bc:
notHandled = handleMethodFmt20bc(cUnit, mir);
break;
case kFmt21c:
case kFmt31c:
notHandled = handleMethodFmt21c_Fmt31c(cUnit, mir);
break;
case kFmt21h:
notHandled = handleMethodFmt21h(cUnit, mir);
break;
case kFmt21s:
notHandled = handleMethodFmt21s(cUnit, mir);
break;
case kFmt21t:
notHandled = handleMethodFmt21t(cUnit, mir, bb, labelList);
break;
case kFmt22b:
case kFmt22s:
notHandled = handleMethodFmt22b_Fmt22s(cUnit, mir);
break;
case kFmt22c:
notHandled = handleMethodFmt22c(cUnit, mir);
break;
case kFmt22cs:
notHandled = handleMethodFmt22cs(cUnit, mir);
break;
case kFmt22t:
notHandled = handleMethodFmt22t(cUnit, mir, bb, labelList);
break;
case kFmt22x:
case kFmt32x:
notHandled = handleMethodFmt22x_Fmt32x(cUnit, mir);
break;
case kFmt23x:
notHandled = handleMethodFmt23x(cUnit, mir);
break;
case kFmt31t:
notHandled = handleMethodFmt31t(cUnit, mir);
break;
case kFmt3rc:
case kFmt35c:
notHandled = handleMethodFmt35c_3rc(cUnit, mir, bb, labelList);
break;
case kFmt3rms:
case kFmt35ms:
notHandled = handleMethodFmt35ms_3rms(cUnit, mir, bb,
labelList);
break;
case kFmt35mi:
case kFmt3rmi:
notHandled = handleMethodExecuteInline(cUnit, mir);
break;
case kFmt51l:
notHandled = handleMethodFmt51l(cUnit, mir);
break;
default:
notHandled = true;
break;
}
/* FIXME - to be implemented */
if (notHandled == true && dalvikOpcode >= kNumPackedOpcodes) {
notHandled = false;
}
if (notHandled) {
ALOGE("%#06x: Opcode %#x (%s) / Fmt %d not handled",
mir->offset,
dalvikOpcode, dexGetOpcodeName(dalvikOpcode),
dalvikFormat);
dvmCompilerAbort(cUnit);
break;
}
}
if (headLIR) {
/*
* Eliminate redundant loads/stores and delay stores into later
* slots
*/
dvmCompilerApplyLocalOptimizations(cUnit, (LIR *) headLIR,
cUnit->lastLIRInsn);
/*
* Generate an unconditional branch to the fallthrough block.
*/
if (bb->fallThrough) {
genUnconditionalBranch(cUnit,
&labelList[bb->fallThrough->id]);
}
}
return false;
}
static bool genConversion(CompilationUnit *cUnit, MIR *mir)
{
OpCode opCode = mir->dalvikInsn.opCode;
int vSrc1Dest = mir->dalvikInsn.vA;
int vSrc2 = mir->dalvikInsn.vB;
int op = THUMB_BKPT;
bool longSrc = false;
bool longDest = false;
int srcReg;
int tgtReg;
switch (opCode) {
case OP_INT_TO_FLOAT:
longSrc = false;
longDest = false;
op = THUMB2_VCVTIF;
break;
case OP_FLOAT_TO_INT:
longSrc = false;
longDest = false;
op = THUMB2_VCVTFI;
break;
case OP_DOUBLE_TO_FLOAT:
longSrc = true;
longDest = false;
op = THUMB2_VCVTDF;
break;
case OP_FLOAT_TO_DOUBLE:
longSrc = false;
longDest = true;
op = THUMB2_VCVTFD;
break;
case OP_INT_TO_DOUBLE:
longSrc = false;
longDest = true;
op = THUMB2_VCVTID;
break;
case OP_DOUBLE_TO_INT:
longSrc = true;
longDest = false;
op = THUMB2_VCVTDI;
break;
case OP_FLOAT_TO_LONG:
case OP_LONG_TO_FLOAT:
case OP_DOUBLE_TO_LONG:
case OP_LONG_TO_DOUBLE:
return genConversionPortable(cUnit, mir);
default:
return true;
}
if (longSrc) {
srcReg = dr1;
loadDouble(cUnit, vSrc2, srcReg);
} else {
srcReg = fr2;
loadFloat(cUnit, vSrc2, srcReg);
}
if (longDest) {
newLIR2(cUnit, op, dr0, srcReg);
storeDouble(cUnit, dr0, vSrc1Dest, 0);
} else {
newLIR2(cUnit, op, fr0, srcReg);
storeFloat(cUnit, fr0, vSrc1Dest, 0);
}
return false;
}