Compiler tuning
Significant reduction in memory usage by the compiler.
o Estimated sizes of growable lists to avoid waste
o Changed basic block predecessor structure from a growable bitmap
to a growable list.
o Conditionalized code which produced disassembly strings.
o Avoided generating some dataflow-related structures when compiling
in dataflow-disabled mode.
o Added memory usage statistics
o Eliminated floating point usage as a barrier to disabling expensive
dataflow analysis for very large init routines.
o Because iterating through sparse bit maps is much less of a concern now,
removed earlier hack that remembered runs of leading and trailing
zeroes.
Also, some general tuning.
o Minor tweaks to register utilties
o Speed up the assembly loop
o Rewrite of the bit vector iterator
Our previous worst-case method originally consumed 360 megabytes, but through
earlier changes was whittled down to 113 megabytes. Now it consumes 12 (which
so far appears to close to the highest compiler heap usage of anything
I've seen).
Post-wipe cold boot time is now less than 7 minutes.
Installation time for our application test cases also shows a large
gain - typically 25% to 40% speedup.
Single-threaded host compilation of core.jar down to <3.0s, boot.oat builds
in 17.2s. Next up: multi-threaded compilation.
Change-Id: I493d0d584c4145a6deccdd9bff344473023deb46
diff --git a/src/compiler/Dataflow.cc b/src/compiler/Dataflow.cc
index 3f88ea5..cd63290 100644
--- a/src/compiler/Dataflow.cc
+++ b/src/compiler/Dataflow.cc
@@ -900,7 +900,7 @@
}
}
int length = strlen(buffer) + 1;
- ret = (char*)oatNew(length, false);
+ ret = (char*)oatNew(length, false, kAllocDFInfo);
memcpy(ret, buffer, length);
return ret;
}
@@ -1036,7 +1036,7 @@
done:
length = strlen(buffer) + 1;
- ret = (char*) oatNew(length, false);
+ ret = (char*) oatNew(length, false, kAllocDFInfo);
memcpy(ret, buffer, length);
return ret;
}
@@ -1078,7 +1078,7 @@
}
int length = strlen(buffer) + 1;
- ret = (char*)oatNew(length, false);
+ ret = (char*)oatNew(length, false, kAllocDFInfo);
memcpy(ret, buffer, length);
return ret;
}
@@ -1111,11 +1111,11 @@
if (bb->dataFlowInfo == NULL) return false;
useV = bb->dataFlowInfo->useV =
- oatAllocBitVector(cUnit->numDalvikRegisters, false);
+ oatAllocBitVector(cUnit->numDalvikRegisters, false, kBitMapUse);
defV = bb->dataFlowInfo->defV =
- oatAllocBitVector(cUnit->numDalvikRegisters, false);
+ oatAllocBitVector(cUnit->numDalvikRegisters, false, kBitMapDef);
liveInV = bb->dataFlowInfo->liveInV =
- oatAllocBitVector(cUnit->numDalvikRegisters, false);
+ oatAllocBitVector(cUnit->numDalvikRegisters, false, kBitMapLiveIn);
for (mir = bb->firstMIRInsn; mir; mir = mir->next) {
int dfAttributes =
@@ -1186,9 +1186,11 @@
int i;
mir->ssaRep->numUses = numUses;
- mir->ssaRep->uses = (int *)oatNew(sizeof(int) * numUses, true);
+ mir->ssaRep->uses = (int *)oatNew(sizeof(int) * numUses, true,
+ kAllocDFInfo);
// NOTE: will be filled in during type & size inference pass
- mir->ssaRep->fpUse = (bool *)oatNew(sizeof(bool) * numUses, true);
+ mir->ssaRep->fpUse = (bool *)oatNew(sizeof(bool) * numUses, true,
+ kAllocDFInfo);
for (i = 0; i < numUses; i++) {
handleSSAUse(cUnit, mir->ssaRep->uses, dInsn->arg[i], i);
@@ -1203,9 +1205,11 @@
int i;
mir->ssaRep->numUses = numUses;
- mir->ssaRep->uses = (int *)oatNew(sizeof(int) * numUses, true);
+ mir->ssaRep->uses = (int *)oatNew(sizeof(int) * numUses, true,
+ kAllocDFInfo);
// NOTE: will be filled in during type & size inference pass
- mir->ssaRep->fpUse = (bool *)oatNew(sizeof(bool) * numUses, true);
+ mir->ssaRep->fpUse = (bool *)oatNew(sizeof(bool) * numUses, true,
+ kAllocDFInfo);
for (i = 0; i < numUses; i++) {
handleSSAUse(cUnit, mir->ssaRep->uses, dInsn->vC+i, i);
@@ -1221,7 +1225,7 @@
for (mir = bb->firstMIRInsn; mir; mir = mir->next) {
mir->ssaRep = (struct SSARepresentation *)
- oatNew(sizeof(SSARepresentation), true);
+ oatNew(sizeof(SSARepresentation), true, kAllocDFInfo);
int dfAttributes =
oatDataFlowAttributes[mir->dalvikInsn.opcode];
@@ -1272,9 +1276,9 @@
if (numUses) {
mir->ssaRep->numUses = numUses;
mir->ssaRep->uses = (int *)oatNew(sizeof(int) * numUses,
- false);
+ false, kAllocDFInfo);
mir->ssaRep->fpUse = (bool *)oatNew(sizeof(bool) * numUses,
- false);
+ false, kAllocDFInfo);
}
int numDefs = 0;
@@ -1289,9 +1293,9 @@
if (numDefs) {
mir->ssaRep->numDefs = numDefs;
mir->ssaRep->defs = (int *)oatNew(sizeof(int) * numDefs,
- false);
+ false, kAllocDFInfo);
mir->ssaRep->fpDef = (bool *)oatNew(sizeof(bool) * numDefs,
- false);
+ false, kAllocDFInfo);
}
DecodedInstruction *dInsn = &mir->dalvikInsn;
@@ -1336,16 +1340,19 @@
}
}
- /*
- * Take a snapshot of Dalvik->SSA mapping at the end of each block. The
- * input to PHI nodes can be derived from the snapshot of all predecessor
- * blocks.
- */
- bb->dataFlowInfo->dalvikToSSAMap =
- (int *)oatNew(sizeof(int) * cUnit->numDalvikRegisters, false);
+ if (!cUnit->disableDataflow) {
+ /*
+ * Take a snapshot of Dalvik->SSA mapping at the end of each block. The
+ * input to PHI nodes can be derived from the snapshot of all
+ * predecessor blocks.
+ */
+ bb->dataFlowInfo->dalvikToSSAMap =
+ (int *)oatNew(sizeof(int) * cUnit->numDalvikRegisters, false,
+ kAllocDFInfo);
- memcpy(bb->dataFlowInfo->dalvikToSSAMap, cUnit->dalvikToSSAMap,
- sizeof(int) * cUnit->numDalvikRegisters);
+ memcpy(bb->dataFlowInfo->dalvikToSSAMap, cUnit->dalvikToSSAMap,
+ sizeof(int) * cUnit->numDalvikRegisters);
+ }
return true;
}
@@ -1436,9 +1443,11 @@
int numDalvikReg = cUnit->numDalvikRegisters;
cUnit->ssaToDalvikMap = (GrowableList *)oatNew(sizeof(GrowableList),
- false);
- oatInitGrowableList(cUnit->ssaToDalvikMap, numDalvikReg);
-
+ false, kAllocDFInfo);
+ // Create the SSAtoDalvikMap, estimating the max size
+ oatInitGrowableList(cUnit->ssaToDalvikMap,
+ numDalvikReg + cUnit->defCount + 128,
+ kListSSAtoDalvikMap);
/*
* Initial number of SSA registers is equal to the number of Dalvik
* registers.
@@ -1460,10 +1469,10 @@
* register N is mapped to SSA register N with subscript 0.
*/
cUnit->dalvikToSSAMap = (int *)oatNew(sizeof(int) * numDalvikReg,
- false);
+ false, kAllocDFInfo);
/* Keep track of the higest def for each dalvik reg */
cUnit->SSALastDefs = (int *)oatNew(sizeof(int) * numDalvikReg,
- false);
+ false, kAllocDFInfo);
for (i = 0; i < numDalvikReg; i++) {
cUnit->dalvikToSSAMap[i] = i;
@@ -1486,7 +1495,7 @@
bb->blockType == kExitBlock) {
bb->dataFlowInfo = (BasicBlockDataFlow *)
oatNew(sizeof(BasicBlockDataFlow),
- true);
+ true, kAllocDFInfo);
}
}
}
@@ -1618,7 +1627,7 @@
{
if (bb->dataFlowInfo == NULL) return false;
bb->dataFlowInfo->endingNullCheckV =
- oatAllocBitVector(cUnit->numSSARegs, false);
+ oatAllocBitVector(cUnit->numSSARegs, false, kBitMapNullCheck);
oatClearAllBits(bb->dataFlowInfo->endingNullCheckV);
return true;
}
@@ -1628,12 +1637,12 @@
struct BasicBlock* bb)
{
if (bb->dataFlowInfo == NULL) return false;
+
/*
* Set initial state. Be conservative with catch
* blocks and start with no assumptions about null check
* status (except for "this").
*/
-
if ((bb->blockType == kEntryBlock) | bb->catchEntry) {
oatClearAllBits(cUnit->tempSSARegisterV);
if ((cUnit->access_flags & kAccStatic) == 0) {
@@ -1643,20 +1652,15 @@
}
} else {
// Starting state is intesection of all incoming arcs
- GrowableList* blockList = &cUnit->blockList;
- ArenaBitVectorIterator bvIterator;
- oatBitVectorIteratorInit(bb->predecessors, &bvIterator);
- int predBBIdx = oatBitVectorIteratorNext(&bvIterator);
- DCHECK_NE(predBBIdx, -1);
- BasicBlock* predBB = (BasicBlock*)oatGrowableListGetElement(
- blockList, predBBIdx);
+ GrowableListIterator iter;
+ oatGrowableListIteratorInit(bb->predecessors, &iter);
+ BasicBlock* predBB = (BasicBlock*)oatGrowableListIteratorNext(&iter);
+ DCHECK(predBB != NULL);
oatCopyBitVector(cUnit->tempSSARegisterV,
predBB->dataFlowInfo->endingNullCheckV);
while (true) {
- predBBIdx = oatBitVectorIteratorNext(&bvIterator);
- if (predBBIdx == -1) break;
- predBB = (BasicBlock*)oatGrowableListGetElement(
- blockList, predBBIdx);
+ predBB = (BasicBlock*)oatGrowableListIteratorNext(&iter);
+ if (!predBB) break;
if ((predBB->dataFlowInfo == NULL) ||
(predBB->dataFlowInfo->endingNullCheckV == NULL)) {
continue;