/* Write the numbers in the literal pool to the codegen stream */
static void installDataContent(CompilationUnit *cUnit)
{
int *dataPtr = (int *) ((char *) cUnit->baseAddr + cUnit->dataOffset);
ArmLIR *dataLIR = (ArmLIR *) cUnit->wordList;
while (dataLIR) {
*dataPtr++ = dataLIR->operands[0];
dataLIR = NEXT_LIR(dataLIR);
}
}
/*
* Identify unconditional branches that jump to the immediate successor of the
* branch itself.
*/
static void applyRedundantBranchElimination(CompilationUnit *cUnit)
{
ArmLIR *thisLIR;
for (thisLIR = (ArmLIR *) cUnit->firstLIRInsn;
thisLIR != (ArmLIR *) cUnit->lastLIRInsn;
thisLIR = NEXT_LIR(thisLIR)) {
/* Branch to the next instruction */
if (thisLIR->opcode == kThumbBUncond) {
ArmLIR *nextLIR = thisLIR;
while (true) {
nextLIR = NEXT_LIR(nextLIR);
/*
* Is the branch target the next instruction?
*/
if (nextLIR == (ArmLIR *) thisLIR->generic.target) {
thisLIR->flags.isNop = true;
break;
}
/*
* Found real useful stuff between the branch and the target.
* Need to explicitly check the lastLIRInsn here since with
* method-based JIT the branch might be the last real
* instruction.
*/
if (!isPseudoOpcode(nextLIR->opcode) ||
(nextLIR == (ArmLIR *) cUnit->lastLIRInsn))
break;
}
}
}
}
/* Return TRUE if error happens */
static bool assembleInstructions(CompilationUnit *cUnit, intptr_t startAddr)
{
short *bufferAddr = (short *) cUnit->codeBuffer;
ArmLIR *lir;
for (lir = (ArmLIR *) cUnit->firstLIRInsn; lir; lir = NEXT_LIR(lir)) {
if (lir->opCode < 0) {
if ((lir->opCode == ARM_PSEUDO_ALIGN4) &&
/* 1 means padding is needed */
(lir->operands[0] == 1)) {
*bufferAddr++ = PADDING_MOV_R0_R0;
}
continue;
}
if (lir->isNop) {
continue;
}
if (lir->opCode == THUMB_LDR_PC_REL ||
lir->opCode == THUMB_ADD_PC_REL) {
ArmLIR *lirTarget = (ArmLIR *) lir->generic.target;
intptr_t pc = (lir->generic.offset + 4) & ~3;
/*
* Allow an offset (stored in operands[2] to be added to the
* PC-relative target. Useful to get to a fixed field inside a
* chaining cell.
*/
intptr_t target = lirTarget->generic.offset + lir->operands[2];
int delta = target - pc;
if (delta & 0x3) {
LOGE("PC-rel distance is not multiples of 4: %d\n", delta);
dvmAbort();
}
if (delta > 1023) {
return true;
}
lir->operands[1] = delta >> 2;
} else if (lir->opCode == THUMB2_CBNZ || lir->opCode == THUMB2_CBZ) {
/*
* Perform a pass of top-down walk, from the second-last instruction in the
* superblock, to eliminate redundant loads and stores.
*
* An earlier load can eliminate a later load iff
* 1) They are must-aliases
* 2) The native register is not clobbered in between
* 3) The memory location is not written to in between
*
* An earlier store can eliminate a later load iff
* 1) They are must-aliases
* 2) The native register is not clobbered in between
* 3) The memory location is not written to in between
*
* A later store can be eliminated by an earlier store iff
* 1) They are must-aliases
* 2) The memory location is not written to in between
*/
static void applyLoadStoreElimination(CompilationUnit *cUnit,
MipsLIR *headLIR,
MipsLIR *tailLIR)
{
MipsLIR *thisLIR;
if (headLIR == tailLIR) return;
for (thisLIR = PREV_LIR(tailLIR);
thisLIR != headLIR;
thisLIR = PREV_LIR(thisLIR)) {
int sinkDistance = 0;
/* Skip non-interesting instructions */
if ((thisLIR->flags.isNop == true) ||
isPseudoOpCode(thisLIR->opcode) ||
!(EncodingMap[thisLIR->opcode].flags & (IS_LOAD | IS_STORE))) {
continue;
}
int nativeRegId = thisLIR->operands[0];
bool isThisLIRLoad = EncodingMap[thisLIR->opcode].flags & IS_LOAD;
MipsLIR *checkLIR;
/* Use the mem mask to determine the rough memory location */
u8 thisMemMask = (thisLIR->useMask | thisLIR->defMask) & ENCODE_MEM;
/*
* Currently only eliminate redundant ld/st for constant and Dalvik
* register accesses.
*/
if (!(thisMemMask & (ENCODE_LITERAL | ENCODE_DALVIK_REG))) continue;
/*
* Add r15 (pc) to the resource mask to prevent this instruction
* from sinking past branch instructions. Also take out the memory
* region bits since stopMask is used to check data/control
* dependencies.
*/
u8 stopUseRegMask = (ENCODE_REG_PC | thisLIR->useMask) &
~ENCODE_MEM;
u8 stopDefRegMask = thisLIR->defMask & ~ENCODE_MEM;
for (checkLIR = NEXT_LIR(thisLIR);
checkLIR != tailLIR;
checkLIR = NEXT_LIR(checkLIR)) {
/*
* Skip already dead instructions (whose dataflow information is
* outdated and misleading).
*/
if (checkLIR->flags.isNop) continue;
u8 checkMemMask = (checkLIR->useMask | checkLIR->defMask) &
ENCODE_MEM;
u8 aliasCondition = thisMemMask & checkMemMask;
bool stopHere = false;
/*
* Potential aliases seen - check the alias relations
*/
if (checkMemMask != ENCODE_MEM && aliasCondition != 0) {
bool isCheckLIRLoad = EncodingMap[checkLIR->opcode].flags &
IS_LOAD;
if (aliasCondition == ENCODE_LITERAL) {
/*
* Should only see literal loads in the instruction
* stream.
*/
assert(!(EncodingMap[checkLIR->opcode].flags &
IS_STORE));
/* Same value && same register type */
if (checkLIR->aliasInfo == thisLIR->aliasInfo &&
REGTYPE(checkLIR->operands[0]) == REGTYPE(nativeRegId)){
/*
* Different destination register - insert
* a move
*/
if (checkLIR->operands[0] != nativeRegId) {
convertMemOpIntoMove(cUnit, checkLIR,
checkLIR->operands[0],
nativeRegId);
}
checkLIR->flags.isNop = true;
}
} else if (aliasCondition == ENCODE_DALVIK_REG) {
/* Must alias */
if (checkLIR->aliasInfo == thisLIR->aliasInfo) {
/* Only optimize compatible registers */
bool regCompatible =
REGTYPE(checkLIR->operands[0]) ==
REGTYPE(nativeRegId);
if ((isThisLIRLoad && isCheckLIRLoad) ||
(!isThisLIRLoad && isCheckLIRLoad)) {
/* RAR or RAW */
if (regCompatible) {
/*
* Different destination register -
* insert a move
*/
if (checkLIR->operands[0] !=
nativeRegId) {
convertMemOpIntoMove(cUnit,
checkLIR,
checkLIR->operands[0],
nativeRegId);
}
checkLIR->flags.isNop = true;
} else {
/*
* Destinaions are of different types -
* something complicated going on so
* stop looking now.
*/
stopHere = true;
}
} else if (isThisLIRLoad && !isCheckLIRLoad) {
/* WAR - register value is killed */
stopHere = true;
} else if (!isThisLIRLoad && !isCheckLIRLoad) {
/* WAW - nuke the earlier store */
thisLIR->flags.isNop = true;
stopHere = true;
}
/* Partial overlap */
} else if (isDalvikRegisterClobbered(thisLIR, checkLIR)) {
/*
* It is actually ok to continue if checkLIR
* is a read. But it is hard to make a test
* case for this so we just stop here to be
* conservative.
*/
stopHere = true;
}
}
/* Memory content may be updated. Stop looking now. */
if (stopHere) {
break;
/* The checkLIR has been transformed - check the next one */
} else if (checkLIR->flags.isNop) {
continue;
}
}
/*
* this and check LIRs have no memory dependency. Now check if
* their register operands have any RAW, WAR, and WAW
* dependencies. If so, stop looking.
*/
if (stopHere == false) {
stopHere = CHECK_REG_DEP(stopUseRegMask, stopDefRegMask,
checkLIR);
}
if (stopHere == true) {
DEBUG_OPT(dumpDependentInsnPair(thisLIR, checkLIR,
"REG CLOBBERED"));
/* Only sink store instructions */
if (sinkDistance && !isThisLIRLoad) {
MipsLIR *newStoreLIR =
(MipsLIR *) dvmCompilerNew(sizeof(MipsLIR), true);
*newStoreLIR = *thisLIR;
/*
* Stop point found - insert *before* the checkLIR
* since the instruction list is scanned in the
* top-down order.
*/
dvmCompilerInsertLIRBefore((LIR *) checkLIR,
(LIR *) newStoreLIR);
thisLIR->flags.isNop = true;
}
break;
} else if (!checkLIR->flags.isNop) {
sinkDistance++;
}
}
}
}
/*
* Perform a pass of bottom-up walk, from the second instruction in the
* superblock, to try to hoist loads to earlier slots.
*/
static void applyLoadHoisting(CompilationUnit *cUnit,
MipsLIR *headLIR,
MipsLIR *tailLIR)
{
MipsLIR *thisLIR, *checkLIR;
/*
* Store the list of independent instructions that can be hoisted past.
* Will decide the best place to insert later.
*/
MipsLIR *prevInstList[MAX_HOIST_DISTANCE];
/* Empty block */
if (headLIR == tailLIR) return;
/* Start from the second instruction */
for (thisLIR = NEXT_LIR(headLIR);
thisLIR != tailLIR;
thisLIR = NEXT_LIR(thisLIR)) {
/* Skip non-interesting instructions */
if ((thisLIR->flags.isNop == true) ||
isPseudoOpCode(thisLIR->opcode) ||
!(EncodingMap[thisLIR->opcode].flags & IS_LOAD)) {
continue;
}
u8 stopUseAllMask = thisLIR->useMask;
/*
* Branches for null/range checks are marked with the true resource
* bits, and loads to Dalvik registers, constant pools, and non-alias
* locations are safe to be hoisted. So only mark the heap references
* conservatively here.
*/
if (stopUseAllMask & ENCODE_HEAP_REF) {
stopUseAllMask |= ENCODE_REG_PC;
}
/* Similar as above, but just check for pure register dependency */
u8 stopUseRegMask = stopUseAllMask & ~ENCODE_MEM;
u8 stopDefRegMask = thisLIR->defMask & ~ENCODE_MEM;
int nextSlot = 0;
bool stopHere = false;
/* Try to hoist the load to a good spot */
for (checkLIR = PREV_LIR(thisLIR);
checkLIR != headLIR;
checkLIR = PREV_LIR(checkLIR)) {
/*
* Skip already dead instructions (whose dataflow information is
* outdated and misleading).
*/
if (checkLIR->flags.isNop) continue;
u8 checkMemMask = checkLIR->defMask & ENCODE_MEM;
u8 aliasCondition = stopUseAllMask & checkMemMask;
stopHere = false;
/* Potential WAR alias seen - check the exact relation */
if (checkMemMask != ENCODE_MEM && aliasCondition != 0) {
/* We can fully disambiguate Dalvik references */
if (aliasCondition == ENCODE_DALVIK_REG) {
/* Must alias or partually overlap */
if ((checkLIR->aliasInfo == thisLIR->aliasInfo) ||
isDalvikRegisterClobbered(thisLIR, checkLIR)) {
stopHere = true;
}
/* Conservatively treat all heap refs as may-alias */
} else {
assert(aliasCondition == ENCODE_HEAP_REF);
stopHere = true;
}
/* Memory content may be updated. Stop looking now. */
if (stopHere) {
prevInstList[nextSlot++] = checkLIR;
break;
}
}
if (stopHere == false) {
stopHere = CHECK_REG_DEP(stopUseRegMask, stopDefRegMask,
checkLIR);
}
/*
* Store the dependent or non-pseudo/indepedent instruction to the
* list.
*/
if (stopHere || !isPseudoOpCode(checkLIR->opcode)) {
prevInstList[nextSlot++] = checkLIR;
if (nextSlot == MAX_HOIST_DISTANCE) break;
}
/* Found a new place to put the load - move it here */
if (stopHere == true) {
DEBUG_OPT(dumpDependentInsnPair(checkLIR, thisLIR
"HOIST STOP"));
break;
}
}
/*
* Reached the top - use headLIR as the dependent marker as all labels
* are barriers.
*/
if (stopHere == false && nextSlot < MAX_HOIST_DISTANCE) {
prevInstList[nextSlot++] = headLIR;
}
/*
* At least one independent instruction is found. Scan in the reversed
* direction to find a beneficial slot.
*/
if (nextSlot >= 2) {
int firstSlot = nextSlot - 2;
int slot;
MipsLIR *depLIR = prevInstList[nextSlot-1];
/* If there is ld-ld dependency, wait LDLD_DISTANCE cycles */
if (!isPseudoOpCode(depLIR->opcode) &&
(EncodingMap[depLIR->opcode].flags & IS_LOAD)) {
firstSlot -= LDLD_DISTANCE;
}
/*
* Make sure we check slot >= 0 since firstSlot may be negative
* when the loop is first entered.
*/
for (slot = firstSlot; slot >= 0; slot--) {
MipsLIR *curLIR = prevInstList[slot];
MipsLIR *prevLIR = prevInstList[slot+1];
/*
* Check the highest instruction.
* ENCODE_ALL represents a scheduling barrier.
*/
if (prevLIR->defMask == ENCODE_ALL) {
/*
* If the first instruction is a load, don't hoist anything
* above it since it is unlikely to be beneficial.
*/
if (EncodingMap[curLIR->opcode].flags & IS_LOAD) continue;
/*
* Need to unconditionally break here even if the hoisted
* distance is greater than LD_LATENCY (ie more than enough
* cycles are inserted to hide the load latency) since theu
* subsequent code doesn't expect to compare against a
* pseudo opcode (whose opcode value is negative).
*/
break;
}
/*
* NOTE: now prevLIR is guaranteed to be a non-pseudo
* instruction (ie accessing EncodingMap[prevLIR->opcode] is
* safe).
*
* Try to find two instructions with load/use dependency until
* the remaining instructions are less than LD_LATENCY.
*/
if (((curLIR->useMask & prevLIR->defMask) &&
(EncodingMap[prevLIR->opcode].flags & IS_LOAD)) ||
(slot < LD_LATENCY)) {
break;
}
}
/* Found a slot to hoist to */
if (slot >= 0) {
MipsLIR *curLIR = prevInstList[slot];
MipsLIR *newLoadLIR = (MipsLIR *) dvmCompilerNew(sizeof(MipsLIR),
true);
*newLoadLIR = *thisLIR;
/*
* Insertion is guaranteed to succeed since checkLIR
* is never the first LIR on the list
*/
dvmCompilerInsertLIRBefore((LIR *) curLIR,
(LIR *) newLoadLIR);
thisLIR->flags.isNop = true;
}
}
}
}
/*
* Perform a pass of top-down walk to
* 1) Eliminate redundant loads and stores
* 2) Sink stores to latest possible slot
*/
static void applyLoadStoreElimination(CompilationUnit *cUnit,
ArmLIR *headLIR,
ArmLIR *tailLIR)
{
ArmLIR *thisLIR;
cUnit->optRound++;
for (thisLIR = headLIR;
thisLIR != tailLIR;
thisLIR = NEXT_LIR(thisLIR)) {
/* Skip newly added instructions */
if (thisLIR->age >= cUnit->optRound) {
continue;
}
if (isDalvikStore(thisLIR)) {
int dRegId = DECODE_ALIAS_INFO_REG(thisLIR->aliasInfo);
int dRegIdHi = dRegId + DECODE_ALIAS_INFO_WIDE(thisLIR->aliasInfo);
int nativeRegId = thisLIR->operands[0];
ArmLIR *checkLIR;
int sinkDistance = 0;
/*
* Add r15 (pc) to the mask to prevent this instruction
* from sinking past branch instructions. Unset the Dalvik register
* bit when checking with native resource constraints.
*/
u8 stopMask = (ENCODE_REG_PC | thisLIR->useMask) &
~ENCODE_DALVIK_REG;
for (checkLIR = NEXT_LIR(thisLIR);
checkLIR != tailLIR;
checkLIR = NEXT_LIR(checkLIR)) {
/* Check if a Dalvik register load is redundant */
if (isDalvikLoad(checkLIR) &&
(checkLIR->aliasInfo == thisLIR->aliasInfo) &&
(REGTYPE(checkLIR->operands[0]) == REGTYPE(nativeRegId))) {
/* Insert a move to replace the load */
if (checkLIR->operands[0] != nativeRegId) {
ArmLIR *moveLIR;
moveLIR = dvmCompilerRegCopyNoInsert(
cUnit, checkLIR->operands[0], nativeRegId);
/*
* Insert the converted checkLIR instruction after the
* the original checkLIR since the optimization is
* scannng in the top-down order and the new instruction
* will need to be checked.
*/
dvmCompilerInsertLIRAfter((LIR *) checkLIR,
(LIR *) moveLIR);
}
checkLIR->isNop = true;
continue;
/*
* Found a true output dependency - nuke the previous store.
* The register type doesn't matter here.
*/
} else if (isDalvikStore(checkLIR) &&
(checkLIR->aliasInfo == thisLIR->aliasInfo)) {
thisLIR->isNop = true;
break;
/* Find out the latest slot that the store can be sunk into */
} else {
/* Last instruction reached */
bool stopHere = (NEXT_LIR(checkLIR) == tailLIR);
/* Store data is clobbered */
stopHere |= ((stopMask & checkLIR->defMask) != 0);
/* Store data partially clobbers the Dalvik register */
if (stopHere == false &&
((checkLIR->useMask | checkLIR->defMask) &
ENCODE_DALVIK_REG)) {
stopHere = isDalvikRegisterClobbered(thisLIR, checkLIR);
}
/* Found a new place to put the store - move it here */
if (stopHere == true) {
DEBUG_OPT(dumpDependentInsnPair(thisLIR, checkLIR,
"SINK STORE"));
/* The store can be sunk for at least one cycle */
if (sinkDistance != 0) {
ArmLIR *newStoreLIR =
dvmCompilerNew(sizeof(ArmLIR), true);
*newStoreLIR = *thisLIR;
newStoreLIR->age = cUnit->optRound;
/*
* Stop point found - insert *before* the checkLIR
* since the instruction list is scanned in the
* top-down order.
*/
dvmCompilerInsertLIRBefore((LIR *) checkLIR,
(LIR *) newStoreLIR);
thisLIR->isNop = true;
}
break;
}
/*
* Saw a real instruction that the store can be sunk after
*/
if (!isPseudoOpCode(checkLIR->opCode)) {
sinkDistance++;
}
}
}
}
}
}
static void applyLoadHoisting(CompilationUnit *cUnit,
ArmLIR *headLIR,
ArmLIR *tailLIR)
{
ArmLIR *thisLIR;
/*
* Don't want to hoist in front of first load following a barrier (or
* first instruction of the block.
*/
bool firstLoad = true;
int maxHoist = dvmCompilerTargetOptHint(kMaxHoistDistance);
cUnit->optRound++;
for (thisLIR = headLIR;
thisLIR != tailLIR;
thisLIR = NEXT_LIR(thisLIR)) {
/* Skip newly added instructions */
if (thisLIR->age >= cUnit->optRound ||
thisLIR->isNop == true) {
continue;
}
if (firstLoad && (EncodingMap[thisLIR->opCode].flags & IS_LOAD)) {
/*
* Ensure nothing will be hoisted in front of this load because
* it's result will likely be needed soon.
*/
thisLIR->defMask |= ENCODE_MEM_USE;
firstLoad = false;
}
firstLoad |= (thisLIR->defMask == ENCODE_ALL);
if (isDalvikLoad(thisLIR)) {
int dRegId = DECODE_ALIAS_INFO_REG(thisLIR->aliasInfo);
int dRegIdHi = dRegId + DECODE_ALIAS_INFO_WIDE(thisLIR->aliasInfo);
int nativeRegId = thisLIR->operands[0];
ArmLIR *checkLIR;
int hoistDistance = 0;
u8 stopUseMask = (ENCODE_REG_PC | thisLIR->useMask);
u8 stopDefMask = thisLIR->defMask;
u8 checkResult;
/* First check if the load can be completely elinimated */
for (checkLIR = PREV_LIR(thisLIR);
checkLIR != headLIR;
checkLIR = PREV_LIR(checkLIR)) {
if (checkLIR->isNop) continue;
/*
* Check if the Dalvik register is previously accessed
* with exactly the same type.
*/
if ((isDalvikLoad(checkLIR) || isDalvikStore(checkLIR)) &&
(checkLIR->aliasInfo == thisLIR->aliasInfo) &&
(checkLIR->operands[0] == nativeRegId)) {
/*
* If it is previously accessed but with a different type,
* the search will terminate later at the point checking
* for partially overlapping stores.
*/
thisLIR->isNop = true;
break;
}
/*
* No earlier use/def can reach this load if:
* 1) Head instruction is reached
*/
if (checkLIR == headLIR) {
break;
}
checkResult = (stopUseMask | stopDefMask) & checkLIR->defMask;
/*
* If both instructions are verified Dalvik accesses, clear the
* may- and must-alias bits to detect true resource
* dependencies.
*/
if (checkResult & ENCODE_DALVIK_REG) {
checkResult &= ~(ENCODE_DALVIK_REG | ENCODE_FRAME_REF);
}
/*
* 2) load target register is clobbered
* 3) A branch is seen (stopUseMask has the PC bit set).
*/
if (checkResult) {
break;
}
/* Store data partially clobbers the Dalvik register */
if (isDalvikStore(checkLIR) &&
isDalvikRegisterClobbered(thisLIR, checkLIR)) {
break;
}
}
/* The load has been eliminated */
if (thisLIR->isNop) continue;
/*
* The load cannot be eliminated. See if it can be hoisted to an
* earlier spot.
*/
for (checkLIR = PREV_LIR(thisLIR);
/* empty by intention */;
checkLIR = PREV_LIR(checkLIR)) {
if (checkLIR->isNop) continue;
/*
* Check if the "thisLIR" load is redundant
* NOTE: At one point, we also triggered if the checkLIR
* instruction was a load. However, that tended to insert
* a load/use dependency because the full scheduler is
* not yet complete. When it is, we chould also trigger
* on loads.
*/
if (isDalvikStore(checkLIR) &&
(checkLIR->aliasInfo == thisLIR->aliasInfo) &&
(REGTYPE(checkLIR->operands[0]) == REGTYPE(nativeRegId))) {
/* Insert a move to replace the load */
if (checkLIR->operands[0] != nativeRegId) {
ArmLIR *moveLIR;
moveLIR = dvmCompilerRegCopyNoInsert(
cUnit, nativeRegId, checkLIR->operands[0]);
/*
* Convert *thisLIR* load into a move
*/
dvmCompilerInsertLIRAfter((LIR *) checkLIR,
(LIR *) moveLIR);
}
thisLIR->isNop = true;
break;
/* Find out if the load can be yanked past the checkLIR */
} else {
/* Last instruction reached */
bool stopHere = (checkLIR == headLIR);
/* Base address is clobbered by checkLIR */
checkResult = stopUseMask & checkLIR->defMask;
if (checkResult & ENCODE_DALVIK_REG) {
checkResult &= ~(ENCODE_DALVIK_REG | ENCODE_FRAME_REF);
}
stopHere |= (checkResult != 0);
/* Load target clobbers use/def in checkLIR */
checkResult = stopDefMask &
(checkLIR->useMask | checkLIR->defMask);
if (checkResult & ENCODE_DALVIK_REG) {
checkResult &= ~(ENCODE_DALVIK_REG | ENCODE_FRAME_REF);
}
stopHere |= (checkResult != 0);
/* Store data partially clobbers the Dalvik register */
if (stopHere == false &&
(checkLIR->defMask & ENCODE_DALVIK_REG)) {
stopHere = isDalvikRegisterClobbered(thisLIR, checkLIR);
}
/*
* Stop at an earlier Dalvik load if the offset of checkLIR
* is not less than thisLIR
*
* Experiments show that doing
*
* ldr r1, [r5, #16]
* ldr r0, [r5, #20]
*
* is much faster than
*
* ldr r0, [r5, #20]
* ldr r1, [r5, #16]
*/
if (isDalvikLoad(checkLIR)) {
int dRegId2 =
DECODE_ALIAS_INFO_REG(checkLIR->aliasInfo);
if (dRegId2 <= dRegId) {
stopHere = true;
}
}
/* Don't go too far */
stopHere |= (hoistDistance >= maxHoist);
/* Found a new place to put the load - move it here */
if (stopHere == true) {
DEBUG_OPT(dumpDependentInsnPair(thisLIR, checkLIR,
"HOIST LOAD"));
/* The load can be hoisted for at least one cycle */
if (hoistDistance != 0) {
ArmLIR *newLoadLIR =
dvmCompilerNew(sizeof(ArmLIR), true);
*newLoadLIR = *thisLIR;
newLoadLIR->age = cUnit->optRound;
/*
* Stop point found - insert *after* the checkLIR
* since the instruction list is scanned in the
* bottom-up order.
*/
dvmCompilerInsertLIRAfter((LIR *) checkLIR,
(LIR *) newLoadLIR);
thisLIR->isNop = true;
}
break;
}
/*
* Saw a real instruction that hosting the load is
* beneficial
*/
if (!isPseudoOpCode(checkLIR->opCode)) {
hoistDistance++;
}
}
}
} else if (isLiteralLoad(thisLIR)) {
int litVal = thisLIR->aliasInfo;
int nativeRegId = thisLIR->operands[0];
ArmLIR *checkLIR;
int hoistDistance = 0;
u8 stopUseMask = (ENCODE_REG_PC | thisLIR->useMask) &
~ENCODE_LITPOOL_REF;
u8 stopDefMask = thisLIR->defMask & ~ENCODE_LITPOOL_REF;
/* First check if the load can be completely elinimated */
for (checkLIR = PREV_LIR(thisLIR);
checkLIR != headLIR;
checkLIR = PREV_LIR(checkLIR)) {
if (checkLIR->isNop) continue;
/* Reloading same literal into same tgt reg? Eliminate if so */
if (isLiteralLoad(checkLIR) &&
(checkLIR->aliasInfo == litVal) &&
(checkLIR->operands[0] == nativeRegId)) {
thisLIR->isNop = true;
break;
}
/*
* No earlier use/def can reach this load if:
* 1) Head instruction is reached
* 2) load target register is clobbered
* 3) A branch is seen (stopUseMask has the PC bit set).
*/
if ((checkLIR == headLIR) ||
(stopUseMask | stopDefMask) & checkLIR->defMask) {
break;
}
}
/* The load has been eliminated */
if (thisLIR->isNop) continue;
/*
* The load cannot be eliminated. See if it can be hoisted to an
* earlier spot.
*/
for (checkLIR = PREV_LIR(thisLIR);
/* empty by intention */;
checkLIR = PREV_LIR(checkLIR)) {
if (checkLIR->isNop) continue;
/*
* TUNING: once a full scheduler exists, check here
* for conversion of a redundant load into a copy similar
* to the way redundant loads are handled above.
*/
/* Find out if the load can be yanked past the checkLIR */
/* Last instruction reached */
bool stopHere = (checkLIR == headLIR);
/* Base address is clobbered by checkLIR */
stopHere |= ((stopUseMask & checkLIR->defMask) != 0);
/* Load target clobbers use/def in checkLIR */
stopHere |= ((stopDefMask &
(checkLIR->useMask | checkLIR->defMask)) != 0);
/* Avoid re-ordering literal pool loads */
stopHere |= isLiteralLoad(checkLIR);
/* Don't go too far */
stopHere |= (hoistDistance >= maxHoist);
/* Found a new place to put the load - move it here */
if (stopHere == true) {
DEBUG_OPT(dumpDependentInsnPair(thisLIR, checkLIR,
"HOIST LOAD"));
/* The store can be hoisted for at least one cycle */
if (hoistDistance != 0) {
ArmLIR *newLoadLIR =
dvmCompilerNew(sizeof(ArmLIR), true);
*newLoadLIR = *thisLIR;
newLoadLIR->age = cUnit->optRound;
/*
* Insertion is guaranteed to succeed since checkLIR
* is never the first LIR on the list
*/
dvmCompilerInsertLIRAfter((LIR *) checkLIR,
(LIR *) newLoadLIR);
thisLIR->isNop = true;
}
break;
}
/*
* Saw a real instruction that hosting the load is
* beneficial
*/
if (!isPseudoOpCode(checkLIR->opCode)) {
hoistDistance++;
}
}
}
}
}
/*
* Find all lsl/lsr and add that can be replaced with a
* combined lsl/lsr + add
*/
static void applyShiftArithmeticOpts(CompilationUnit *cUnit,
ArmLIR *headLIR,
ArmLIR *tailLIR) {
ArmLIR *thisLIR = NULL;
for (thisLIR = headLIR;
thisLIR != tailLIR;
thisLIR = NEXT_LIR(thisLIR)) {
if(thisLIR->flags.isNop) {
continue;
}
if(thisLIR->opcode == kThumb2LslRRI5 || thisLIR->opcode == kThumb2LsrRRI5 ||
thisLIR->opcode == kThumbLslRRI5 || thisLIR->opcode == kThumbLsrRRI5) {
/* Find next that is not nop and not pseudo code */
ArmLIR *nextLIR = NULL;
for(nextLIR = NEXT_LIR(thisLIR);
nextLIR != tailLIR;
nextLIR = NEXT_LIR(nextLIR)) {
if (!nextLIR->flags.isNop && !isPseudoOpcode(nextLIR->opcode)) {
break;
}
}
if(nextLIR == tailLIR) {
return;
}
if(nextLIR->opcode == kThumb2AddRRR &&
nextLIR->operands[3] == 0 &&
(nextLIR->operands[1] == thisLIR->operands[0] ||
nextLIR->operands[2] == thisLIR->operands[0])) {
bool applyOpt = true;
if(!(thisLIR->operands[0] == nextLIR->operands[0])) {
/* Check that shift dest reg is not used after
* the addition. */
ArmLIR* tmpLIR = NULL;
for(tmpLIR = NEXT_LIR(nextLIR);
tmpLIR != tailLIR;
tmpLIR = NEXT_LIR(tmpLIR)) {
if (!tmpLIR->flags.isNop &&
!(EncodingMap[tmpLIR->opcode].flags & IS_BRANCH) &&
(tmpLIR->defMask | tmpLIR->useMask) & thisLIR->defMask) {
if(tmpLIR->useMask & thisLIR->defMask) {
/* Shift dest reg is used for src, skip opt. */
applyOpt = false;
}
break;
}
}
}
if(applyOpt) {
/*
* Found lsl/lsr & add, use barrel shifter for add instead
*
* (1) Normal case
* [lsl/lsr] r9, r1, #x
* [add] r0, r2, r9
*
* (2) Changing place of args for add
* [lsl/lsr] r9, r1, #x
* [add] r0, r9, r2
*
* (3) Using r1 and r1 shifted as args for add
* [lsl/lsr] r9, r1, #x
* [add] r0, r1, r9
*
* (4) Using r1 and r1 shifted as args for add, variant 2
* [lsl/lsr] r9, r1, #x
* [add] r0, r9, r1
*
* Result:
* [add] rDest, rSrc1, rSrc2, [lsl/lsr] x
*/
int type = kArmLsl;
if(thisLIR->opcode == kThumb2LsrRRI5 || thisLIR->opcode == kThumbLsrRRI5) {
type = kArmLsr;
}
/* For most cases keep original rSrc1 */
int rSrc1 = nextLIR->operands[1];
if(thisLIR->operands[0] == nextLIR->operands[1]) {
/* Case 2 & 4: move original rSrc2 to rScr1 since
reg to be shifted need to be in rSrc2 */
rSrc1 = nextLIR->operands[2];
}
/* Reg to be shifted need to be in rSrc2 */
int rSrc2 = thisLIR->operands[1];
/* Encode type of shift and amount */
int shift = ((thisLIR->operands[2] & 0x1f) << 2) | type;
/* Keep rDest, but change rSrc1, rSrc2 and use shift */
ArmLIR* newLIR = (ArmLIR *)dvmCompilerNew(sizeof(ArmLIR), true);
newLIR->opcode = nextLIR->opcode;
newLIR->operands[0] = nextLIR->operands[0];
newLIR->operands[1] = rSrc1;
newLIR->operands[2] = rSrc2;
newLIR->operands[3] = shift;
dvmCompilerSetupResourceMasks(newLIR);
dvmCompilerInsertLIRBefore((LIR *) nextLIR, (LIR *) newLIR);
thisLIR->flags.isNop = true;
nextLIR->flags.isNop = true;
}
/*
* Avoid looping through nops already identified.
* Continue directly after the updated instruction
* instead.
*/
thisLIR = nextLIR;
}
}
}
}
/*
* Perform a pass to hoist all frame pointer load instructions that
* are independent, outside the loop.
*/
static void applyLoopLoadHoisting(CompilationUnit *cUnit)
{
ArmLIR *thisLIR, *labelLIR, *lastLIR, *insertLIR;
ArmLIR *loadLIR[MAX_LOAD_HOISTS];
ArmLIR *regLIR[MAX_REGISTER_OPS];
u8 defLoadMask = 0;
u8 defMask = 0;
u8 masterDefMask = ~((1ULL << kRegEnd) - 1);
bool isValidLoop = false;
int loadCount = 0;
int regCount = 0;
int loadindex;
int hoistCount = 0;
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("GlobalOpts LoopLoadHoisting applied on '%s'", cUnit->method->name);
cUnit->printMe = true;
#endif
labelLIR = (ArmLIR *) cUnit->firstLIRInsn;
lastLIR = (ArmLIR *) cUnit->lastLIRInsn;
insertLIR = NULL;
/* Find the insert point */
while ((labelLIR != lastLIR) &&
(labelLIR->flags.isNop ||
(labelLIR->opcode != kArmPseudoNormalBlockLabel))) {
if ((cUnit->loopAnalysis->branchesAdded) &&
(labelLIR->opcode == kThumbBUncond) &&
(insertLIR == NULL) && (!labelLIR->flags.isNop))
insertLIR = labelLIR;
labelLIR = NEXT_LIR(labelLIR);
}
if (labelLIR->opcode != kArmPseudoNormalBlockLabel) {
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Can't hoist - no loop label found!");
#endif
return;
}
if (insertLIR == NULL) {
insertLIR = labelLIR;
}
else if ((ArmLIR *) insertLIR->generic.target != labelLIR) {
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Can't hoist - branch target does not match!");
#endif
return;
}
/* Search for eligible load instructions to hoist */
for (thisLIR = labelLIR;
thisLIR != lastLIR;
thisLIR = NEXT_LIR(thisLIR)) {
bool handled = false;
int flags;
/* Skip non-interesting instructions */
if (thisLIR->flags.isNop || isPseudoOpcode(thisLIR->opcode))
continue;
flags = EncodingMap[thisLIR->opcode].flags;
/* If it's a load instruction, check if it's a hoist candidate. */
if (((flags & IS_LOAD) != 0) &&
((thisLIR->useMask & ENCODE_DALVIK_REG) != 0)) {
if (regCount >= MAX_REGISTER_OPS) {
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Out of register list space!");
#endif
return;
}
regLIR[regCount++] = thisLIR;
if ((((defLoadMask | defMask) & thisLIR->defMask) == 0) &&
(loadCount < MAX_LOAD_HOISTS)) {
defLoadMask |= thisLIR->defMask;
loadLIR[loadCount++] = thisLIR;
handled = true;
} else {
masterDefMask |= thisLIR->defMask;
}
/* If it's a store instruction, check if it matches a previous
hoistable load instruction. If so, reset the global def-flag to
indicate that the load is still hoistable. */
} else if (((flags & IS_STORE) != 0) &&
((thisLIR->defMask & ENCODE_DALVIK_REG) != 0)) {
if (regCount >= MAX_REGISTER_OPS) {
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Out of register list space!");
#endif
return;
}
regLIR[regCount++] = thisLIR;
if ((thisLIR->useMask & defLoadMask) != 0) {
handled = true;
for (int i = loadCount - 1; i >= 0; i--) {
if ((thisLIR->aliasInfo == loadLIR[i]->aliasInfo) &&
(thisLIR->operands[0] == loadLIR[i]->operands[0])) {
defMask &= ~(loadLIR[i]->defMask);
break;
}
}
}
/* If it's a branch instruction, check if it's the loop branch.
If it matches the label, mark it as a valid loop. */
} else if ((flags & IS_BRANCH) != 0) {
handled = true;
if (labelLIR == (ArmLIR *) thisLIR->generic.target) {
isValidLoop = true;
} else if ((thisLIR->opcode >= kThumbBlx1) &&
(thisLIR->opcode <= kThumbBlxR))
{
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Trace contains branch to subroutine!");
#endif
return;
}
} else if (thisLIR->opcode == kThumbUndefined) {
break;
}
/* If it's not a 'special' instruction, accumulate into def-flags. */
if (!handled)
defMask |= thisLIR->defMask;
}
defLoadMask &= ~(defMask | masterDefMask);
if (!isValidLoop || (defLoadMask == 0)) {
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Loop not valid, or defLoadMask (0x%llx) was zero!", defLoadMask);
#endif
return;
}
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Masks: masterDef: 0x%llx, def: 0x%llx, final defLoad: 0x%llx",
masterDefMask, defMask, defLoadMask);
#endif
/* Try to hoist the load operations */
for (loadindex = 0; loadindex < loadCount; loadindex++) {
thisLIR = loadLIR[loadindex];
/* Host this load? */
if ((thisLIR->defMask & defLoadMask) == thisLIR->defMask) {
int i;
bool foundAlias = false;
for (i = 0; i < regCount; i++) {
if ((thisLIR->aliasInfo == regLIR[i]->aliasInfo) &&
(thisLIR->operands[0] != regLIR[i]->operands[0])) {
foundAlias = true;
for (int k = loadindex; k < loadCount; k++) {
if (loadLIR[k] == regLIR[i]) {
loadLIR[k]->defMask = -1;
break;
}
}
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Register alias found between these two load ops:");
dvmDumpLIRInsn((LIR*)thisLIR, NULL);
dvmDumpLIRInsn((LIR*)regLIR[i], NULL);
#endif
break;
}
}
if (!foundAlias) {
#ifdef LOOP_LOAD_HOIST_VERBOSE
ALOGD("Hoisting this load op:");
dvmDumpLIRInsn((LIR*)thisLIR, NULL);
#endif
ArmLIR *newLoadLIR = (ArmLIR *) dvmCompilerNew(sizeof(ArmLIR),
true);
*newLoadLIR = *thisLIR;
dvmCompilerInsertLIRBefore((LIR *) insertLIR,
(LIR *) newLoadLIR);
thisLIR->flags.isNop = true;
hoistCount++;
}
}
}
if (cUnit->printMe)
ALOGD("GlobalOpt LoopLoadHoist hoisted %d load ops.", hoistCount);
}