Quick compiler: restore optimizations
This CL re-enables optizations on the Quick compile path.
Notes:
o Although all optimization are enabled, several are now useless
because of llvm and bitcode constraints:
- Large method de-optimization (i.e. - skipping expensive dataflow
analysis) can't be done because we have to do the analysis to
produce a CFG that makes the bitcode verifier happy.
- Small method pattern matching isn't applicable w/ bitcode (though
I can probably do something similar in the Quick backend, but
looking for bitcode instead of dex patterns).
- Branch fusing doesn't translate to bitcode.
- Bitcode generation has de-optimized code layout. We'll try to
repair the damage in a subsequent CL.
o There is an ugly workaround related to the way we're loading and
unloading the compiler .so containing llvm. [See comment in compiler.cc]
o We're still running single-threaded - need to add the magic to allow
multi-threaded use of llvm.
o With the CL, the phone boots, all target tests pass and all cts VM
tests pass (except those being dealt with via a verifier change).
o Compile time is pretty bad - when flashing it's best to follow
with an adb sync to avoid on-device compilation of system apps.
Change-Id: I1c98f9e64aefbcbd24b957c71544c28450eb2023
diff --git a/src/compiler/codegen/MethodBitcode.cc b/src/compiler/codegen/MethodBitcode.cc
index e7e4e5a..6b78765 100644
--- a/src/compiler/codegen/MethodBitcode.cc
+++ b/src/compiler/codegen/MethodBitcode.cc
@@ -107,17 +107,63 @@
loc.wide = ((ty == cUnit->irb->getInt64Ty()) ||
(ty == cUnit->irb->getDoubleTy()));
loc.defined = true;
- if ((ty == cUnit->irb->getFloatTy()) ||
- (ty == cUnit->irb->getDoubleTy())) {
- loc.fp = true;
- } else if (ty == cUnit->irb->GetJObjectTy()) {
- loc.ref = true;
- } else {
- loc.core = true;
- }
- loc.home = false; // Will change during promotion
+ loc.home = false; // May change during promotion
loc.sRegLow = baseSReg;
loc.origSReg = cUnit->locMap.size();
+ PromotionMap pMap = cUnit->promotionMap[baseSReg];
+ if (ty == cUnit->irb->getFloatTy()) {
+ loc.fp = true;
+ if (pMap.fpLocation == kLocPhysReg) {
+ loc.lowReg = pMap.fpReg;
+ loc.location = kLocPhysReg;
+ loc.home = true;
+ }
+ } else if (ty == cUnit->irb->getDoubleTy()) {
+ loc.fp = true;
+ PromotionMap pMapHigh = cUnit->promotionMap[baseSReg + 1];
+ if ((pMap.fpLocation == kLocPhysReg) &&
+ (pMapHigh.fpLocation == kLocPhysReg) &&
+ ((pMap.fpReg & 0x1) == 0) &&
+ (pMap.fpReg + 1 == pMapHigh.fpReg)) {
+ loc.lowReg = pMap.fpReg;
+ loc.highReg = pMapHigh.fpReg;
+ loc.location = kLocPhysReg;
+ loc.home = true;
+ }
+ } else if (ty == cUnit->irb->GetJObjectTy()) {
+ loc.ref = true;
+ if (pMap.coreLocation == kLocPhysReg) {
+ loc.lowReg = pMap.coreReg;
+ loc.location = kLocPhysReg;
+ loc.home = true;
+ }
+ } else if (ty == cUnit->irb->getInt64Ty()) {
+ loc.core = true;
+ PromotionMap pMapHigh = cUnit->promotionMap[baseSReg + 1];
+ if ((pMap.coreLocation == kLocPhysReg) &&
+ (pMapHigh.coreLocation == kLocPhysReg)) {
+ loc.lowReg = pMap.coreReg;
+ loc.highReg = pMapHigh.coreReg;
+ loc.location = kLocPhysReg;
+ loc.home = true;
+ }
+ } else {
+ loc.core = true;
+ if (pMap.coreLocation == kLocPhysReg) {
+ loc.lowReg = pMap.coreReg;
+ loc.location = kLocPhysReg;
+ loc.home = true;
+ }
+ }
+
+ if (cUnit->printMe && loc.home) {
+ if (loc.wide) {
+ LOG(INFO) << "Promoted wide " << s << " to regs " << loc.lowReg
+ << "/" << loc.highReg;
+ } else {
+ LOG(INFO) << "Promoted " << s << " to reg " << loc.lowReg;
+ }
+ }
cUnit->locMap.Put(val, loc);
}
@@ -2883,12 +2929,19 @@
oatNew(cUnit, sizeof(RegLocation) * cUnit->numIns, true, kAllocMisc);
llvm::Function::arg_iterator it(cUnit->func->arg_begin());
llvm::Function::arg_iterator it_end(cUnit->func->arg_end());
+ // Skip past Method*
+ it++;
for (unsigned i = 0; it != it_end; ++it) {
llvm::Value* val = it;
argLocs[i++] = valToLoc(cUnit, val);
llvm::Type* ty = val->getType();
if ((ty == cUnit->irb->getInt64Ty()) || (ty == cUnit->irb->getDoubleTy())) {
- argLocs[i++].sRegLow = INVALID_SREG;
+ argLocs[i] = argLocs[i-1];
+ argLocs[i].lowReg = argLocs[i].highReg;
+ argLocs[i].origSReg++;
+ argLocs[i].sRegLow = INVALID_SREG;
+ argLocs[i].highWord = true;
+ i++;
}
}
genEntrySequence(cUnit, argLocs, cUnit->methodLoc);
@@ -3365,15 +3418,78 @@
cUnit->numFPSpills = 0;
cUnit->coreVmapTable.clear();
cUnit->fpVmapTable.clear();
- oatAdjustSpillMask(cUnit);
- cUnit->frameSize = oatComputeFrameSize(cUnit);
/*
* At this point, we've lost all knowledge of register promotion.
* Rebuild that info from the MethodInfo intrinsic (if it
- * exists - not required for correctness).
+ * exists - not required for correctness). Normally, this will
+ * be the first instruction we encounter, so we won't have to iterate
+ * through everything.
*/
- // TODO: find and recover MethodInfo.
+ for (llvm::inst_iterator i = llvm::inst_begin(func),
+ e = llvm::inst_end(func); i != e; ++i) {
+ llvm::CallInst* callInst = llvm::dyn_cast<llvm::CallInst>(&*i);
+ if (callInst != NULL) {
+ llvm::Function* callee = callInst->getCalledFunction();
+ greenland::IntrinsicHelper::IntrinsicId id =
+ cUnit->intrinsic_helper->GetIntrinsicId(callee);
+ if (id == greenland::IntrinsicHelper::MethodInfo) {
+ if (cUnit->printMe) {
+ LOG(INFO) << "Found MethodInfo";
+ }
+ llvm::MDNode* regInfoNode = callInst->getMetadata("RegInfo");
+ if (regInfoNode != NULL) {
+ llvm::ConstantInt* numInsValue =
+ static_cast<llvm::ConstantInt*>(regInfoNode->getOperand(0));
+ llvm::ConstantInt* numRegsValue =
+ static_cast<llvm::ConstantInt*>(regInfoNode->getOperand(1));
+ llvm::ConstantInt* numOutsValue =
+ static_cast<llvm::ConstantInt*>(regInfoNode->getOperand(2));
+ llvm::ConstantInt* numCompilerTempsValue =
+ static_cast<llvm::ConstantInt*>(regInfoNode->getOperand(3));
+ llvm::ConstantInt* numSSARegsValue =
+ static_cast<llvm::ConstantInt*>(regInfoNode->getOperand(4));
+ if (cUnit->printMe) {
+ LOG(INFO) << "RegInfo - Ins:" << numInsValue->getZExtValue()
+ << ", Regs:" << numRegsValue->getZExtValue()
+ << ", Outs:" << numOutsValue->getZExtValue()
+ << ", CTemps:" << numCompilerTempsValue->getZExtValue()
+ << ", SSARegs:" << numSSARegsValue->getZExtValue();
+ }
+ }
+ llvm::MDNode* pmapInfoNode = callInst->getMetadata("PromotionMap");
+ if (pmapInfoNode != NULL) {
+ int elems = pmapInfoNode->getNumOperands();
+ if (cUnit->printMe) {
+ LOG(INFO) << "PMap size: " << elems;
+ }
+ for (int i = 0; i < elems; i++) {
+ llvm::ConstantInt* rawMapData =
+ static_cast<llvm::ConstantInt*>(pmapInfoNode->getOperand(i));
+ uint32_t mapData = rawMapData->getZExtValue();
+ PromotionMap* p = &cUnit->promotionMap[i];
+ p->firstInPair = (mapData >> 24) & 0xff;
+ p->fpReg = (mapData >> 16) & 0xff;
+ p->coreReg = (mapData >> 8) & 0xff;
+ p->fpLocation = static_cast<RegLocationType>((mapData >> 4) & 0xf);
+ if (p->fpLocation == kLocPhysReg) {
+ oatRecordFpPromotion(cUnit, p->fpReg, i);
+ }
+ p->coreLocation = static_cast<RegLocationType>(mapData & 0xf);
+ if (p->coreLocation == kLocPhysReg) {
+ oatRecordCorePromotion(cUnit, p->coreReg, i);
+ }
+ }
+ if (cUnit->printMe) {
+ oatDumpPromotionMap(cUnit);
+ }
+ }
+ break;
+ }
+ }
+ }
+ oatAdjustSpillMask(cUnit);
+ cUnit->frameSize = oatComputeFrameSize(cUnit);
// Create RegLocations for arguments
llvm::Function::arg_iterator it(cUnit->func->arg_begin());