[MachineSink] Consider multiple instructions when sinking into cycle #142641
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Instructions are grouped into a sink unit called a "chain" and considered whether it's profitable to sink as a unit by looking at how many liveins and defs the chain has. If the chain has more defs than liveins, then sinking it will (almost always) reduce register pressure by the difference. Multiple chains sharing a livein are considered to not introduce this livein. The max register pressure is estimated in the cycle by estimating cycle liveins and using a reg pressure tracker for each block. Chains are sunk if they positively affect the critical pressure sets until the set is no longer above the max. This approach relies on the fact that values in the preheader will be live throughout the entire loop. So sinking a chain next to a use will almost always reduce pressure.
|
Thank you for submitting a Pull Request (PR) to the LLVM Project! This PR will be automatically labeled and the relevant teams will be notified. If you wish to, you can add reviewers by using the "Reviewers" section on this page. If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers. If you have further questions, they may be answered by the LLVM GitHub User Guide. You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums. |
|
@llvm/pr-subscribers-backend-systemz @llvm/pr-subscribers-backend-amdgpu Author: None (joe-rivos) ChangesInstructions are grouped into a sink unit called a "chain" and The max register pressure is estimated in the cycle by estimating cycle This approach relies on the fact that values in the preheader will be Patch is 182.36 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/142641.diff 3 Files Affected:
diff --git a/llvm/lib/CodeGen/MachineSink.cpp b/llvm/lib/CodeGen/MachineSink.cpp
index e3f6eda8ff065..417e56a908e6b 100644
--- a/llvm/lib/CodeGen/MachineSink.cpp
+++ b/llvm/lib/CodeGen/MachineSink.cpp
@@ -112,6 +112,17 @@ static cl::opt<unsigned> SinkIntoCycleLimit(
"The maximum number of instructions considered for cycle sinking."),
cl::init(50), cl::Hidden);
+static cl::opt<unsigned>
+ SinkMaxFanOut("machine-sink-max-fanout",
+ cl::desc("The maximum fan-out count for a def to consider it "
+ "for sinking into a cycle"),
+ cl::init(2), cl::Hidden);
+
+static cl::opt<unsigned> CloneUsesThreshold(
+ "machine-sink-clone-uses-threshold",
+ cl::desc("Maximum number of uses for a sink candidate to be cloned."),
+ cl::init(8), cl::Hidden);
+
STATISTIC(NumSunk, "Number of machine instructions sunk");
STATISTIC(NumCycleSunk, "Number of machine instructions sunk into a cycle");
STATISTIC(NumSplit, "Number of critical edges split");
@@ -216,6 +227,16 @@ class MachineSinking {
}
private:
+ // Represents a use-def chain of instructions that can be sunk together.
+ struct Chain {
+ SmallVector<MachineInstr *> Instrs;
+ SmallSet<Register, 4> LiveIns;
+ SmallSet<Register, 4> Defs;
+ SmallVector<int, 8> PSetDiff;
+ unsigned TotalLatency;
+ unsigned NumUses;
+ };
+
bool ProcessBlock(MachineBasicBlock &MBB);
void ProcessDbgInst(MachineInstr &MI);
bool isLegalToBreakCriticalEdge(MachineInstr &MI, MachineBasicBlock *From,
@@ -260,9 +281,19 @@ class MachineSinking {
void FindCycleSinkCandidates(MachineCycle *Cycle, MachineBasicBlock *BB,
SmallVectorImpl<MachineInstr *> &Candidates);
- bool
- aggressivelySinkIntoCycle(MachineCycle *Cycle, MachineInstr &I,
- DenseMap<SinkItem, MachineInstr *> &SunkInstrs);
+ MachineInstr *getEarliestSinkPoint(MachineInstr *MI1, MachineInstr *MI2);
+ void FilterChains(const MachineCycle *Cycle, std::vector<Chain> &Chains);
+ void SortIntoUseDefChains(SmallVectorImpl<MachineInstr *> &Instrs,
+ std::vector<Chain> &Chains,
+ const DenseSet<Register> &CycleLiveIns);
+ bool SinkChainIntoCycle(MachineCycle *Cycle, Chain &C,
+ std::vector<int> &OverMaxSetPressure);
+ void CalculateMaxSetPressures(const MachineBasicBlock *MBB,
+ const DenseSet<Register> &CycleLiveIns,
+ std::vector<unsigned> &MaxSetPressure);
+ void CalculateMaxSetPressures(const MachineCycle *Cycle,
+ const DenseSet<Register> &CycleLiveIns,
+ std::vector<unsigned> &MaxSetPressure);
bool isProfitableToSinkTo(Register Reg, MachineInstr &MI,
MachineBasicBlock *MBB,
@@ -284,8 +315,6 @@ class MachineSinking {
bool registerPressureSetExceedsLimit(unsigned NRegs,
const TargetRegisterClass *RC,
const MachineBasicBlock &MBB);
-
- bool registerPressureExceedsLimit(const MachineBasicBlock &MBB);
};
class MachineSinkingLegacy : public MachineFunctionPass {
@@ -715,7 +744,14 @@ void MachineSinking::FindCycleSinkCandidates(
MachineCycle *Cycle, MachineBasicBlock *BB,
SmallVectorImpl<MachineInstr *> &Candidates) {
for (auto &MI : *BB) {
+ if (MI.isDebugInstr())
+ continue;
LLVM_DEBUG(dbgs() << "CycleSink: Analysing candidate: " << MI);
+ // TODO: support instructions with multiple defs
+ if (MI.getNumDefs() > 1) {
+ LLVM_DEBUG(dbgs() << "CycleSink: not sinking multi-def instruction\n");
+ continue;
+ }
if (MI.isMetaInstruction()) {
LLVM_DEBUG(dbgs() << "CycleSink: not sinking meta instruction\n");
continue;
@@ -741,10 +777,10 @@ void MachineSinking::FindCycleSinkCandidates(
if (MI.isConvergent())
continue;
- const MachineOperand &MO = MI.getOperand(0);
- if (!MO.isReg() || !MO.getReg() || !MO.isDef())
+ const MachineOperand &DefMO = MI.getOperand(0);
+ if (!DefMO.isReg() || !DefMO.getReg() || !DefMO.isDef())
continue;
- if (!MRI->hasOneDef(MO.getReg()))
+ if (!MRI->hasOneDef(DefMO.getReg()))
continue;
LLVM_DEBUG(dbgs() << "CycleSink: Instruction added as candidate.\n");
@@ -752,6 +788,282 @@ void MachineSinking::FindCycleSinkCandidates(
}
}
+void MachineSinking::SortIntoUseDefChains(
+ SmallVectorImpl<MachineInstr *> &Instrs, std::vector<Chain> &UserChains,
+ const DenseSet<Register> &CycleLiveIns) {
+ DenseMap<Register, MachineInstr *> DefToInstr;
+ DenseMap<Register, unsigned> UseCounts;
+ SmallSet<MachineInstr *, 8> SkipMI;
+
+ // Do not consider defs with high fan-out count, I.E a def with many uses in
+ // the chain. Unrestricting this can lead to huge chains. This is bad for two
+ // reasons. Firstly, one node with a high use-count is likely to have a large
+ // liverange, so sinking it is less beneficial. Secondly, if there is one
+ // instruction in a chain that cannot be sunk, then this prevents the entire
+ // chain to be sunk. Splitting it up allows more granularity to both avoid
+ // this and be better for SinkIntoCycleLimit.
+ for (MachineInstr *MI : Instrs) {
+ for (const MachineOperand &MO : MI->operands()) {
+ if (MO.isReg()) {
+ Register Reg = MO.getReg();
+ if (MO.isDef()) {
+ DefToInstr[Reg] = MI;
+ } else if (MO.isUse()) {
+ UseCounts[Reg]++;
+ if (UseCounts[Reg] > SinkMaxFanOut) {
+ LLVM_DEBUG(dbgs() << "CycleSink: Not considering sinking " << *MI
+ << ". Uses exceeds max fanout.\n");
+ SkipMI.insert(DefToInstr[Reg]);
+ }
+ }
+ }
+ }
+ }
+
+ DenseMap<Register, unsigned> RegToChain;
+ DenseMap<MachineInstr *, unsigned> InstrToChain;
+ SmallVector<Chain, 8> Chains;
+ // Process each instruction, placing it into a chain based on defs/uses.
+ unsigned NextChainID = 0;
+ for (MachineInstr *MI : Instrs) {
+ if (SkipMI.contains(MI))
+ continue;
+ SmallSet<unsigned, 4> RelatedChains;
+ SmallSet<Register, 4> LiveIns;
+
+ // Check uses to see if this instruction interacts with a chain.
+ for (const MachineOperand &MO : MI->operands()) {
+ if (!MO.isReg())
+ continue;
+ Register Reg = MO.getReg();
+
+ // Skip physical regs as they are known to be cycle invariant.
+ if (!Reg.isVirtual())
+ continue;
+
+ auto It = RegToChain.find(Reg);
+ if (It != RegToChain.end())
+ RelatedChains.insert(It->second);
+ else if (MO.isUse() && !CycleLiveIns.contains(Reg))
+ LiveIns.insert(Reg);
+ }
+
+ unsigned ChainID;
+ if (RelatedChains.empty()) {
+ // No related chain — create a new one.
+ ChainID = NextChainID++;
+ Chains.resize(NextChainID);
+ } else if (RelatedChains.size() == 1) {
+ // Single related chain — join it.
+ ChainID = *RelatedChains.begin();
+ } else {
+ // Multiple related chains — merge together.
+ ChainID = *RelatedChains.begin();
+ Chain &MainChain = Chains[ChainID];
+
+ for (auto It = ++RelatedChains.begin(); It != RelatedChains.end(); ++It) {
+ Chain &Other = Chains[*It];
+ MainChain.Instrs.append(Other.Instrs.begin(), Other.Instrs.end());
+ for (Register R : Other.LiveIns)
+ MainChain.LiveIns.insert(R);
+
+ // Update mappings
+ for (MachineInstr *MovedMI : Other.Instrs)
+ InstrToChain[MovedMI] = ChainID;
+
+ for (const MachineInstr *MovedMI : Other.Instrs) {
+ for (const MachineOperand &MO : MovedMI->operands()) {
+ if (!MO.isReg())
+ continue;
+ Register Reg = MO.getReg();
+ if (Reg && RegToChain.contains(Reg))
+ RegToChain[Reg] = ChainID;
+ }
+ }
+ Other.Instrs.clear();
+ Other.LiveIns.clear();
+ }
+ }
+
+ // Place this instruction into the chosen chain.
+ Chain &C = Chains[ChainID];
+ C.Instrs.push_back(MI);
+ for (Register R : LiveIns)
+ C.LiveIns.insert(R);
+ InstrToChain[MI] = ChainID;
+
+ // Update reg -> chain mappings for all defs.
+ for (const MachineOperand &MO : MI->operands()) {
+ if (MO.isReg() && MO.isDef())
+ RegToChain[MO.getReg()] = ChainID;
+ }
+ }
+
+ // Now calculate defs of each chain by traversing our list of instructions
+ // bottom-up, keeping a list of visited so that we don't count them as an
+ // external use.
+ for (Chain &C : Chains) {
+ SmallPtrSet<MachineInstr *, 4> VisitedMI;
+ unsigned TotalLatency = 0;
+ unsigned NumUses = 0;
+ for (MachineInstr *MI : llvm::reverse(C.Instrs)) {
+ TotalLatency += SchedModel.computeInstrLatency(MI, true);
+ VisitedMI.insert(MI);
+ for (MachineOperand &MO : MI->defs()) {
+ for (MachineInstr &UseMI : MRI->use_instructions(MO.getReg())) {
+ if (VisitedMI.count(&UseMI))
+ continue;
+ NumUses++;
+ C.Defs.insert(MO.getReg());
+ }
+ }
+ }
+ C.TotalLatency = TotalLatency;
+ C.NumUses = NumUses;
+ // No def chains can exist due to earlier passes not cleaning up after
+ // themselves. Ignore them as they will be cleaned up later.
+ if (C.Defs.empty())
+ C.Instrs.clear();
+ }
+
+ // Copy all our working chains into the user-supplied list
+ for (Chain &C : Chains)
+ if (!C.Instrs.empty())
+ UserChains.push_back(C);
+}
+
+/// Helper function to calculate cycle liveins.
+static void calculateCycleLiveIns(const MachineCycle *Cycle,
+ DenseSet<Register> &CycleLiveIns) {
+ DenseSet<Register> Defs;
+
+ // Collect defs.
+ for (const MachineBasicBlock *MBB : Cycle->blocks()) {
+ for (const MachineInstr &MI : *MBB) {
+ if (MI.isDebugInstr())
+ continue;
+ for (const MachineOperand &MO : MI.operands()) {
+ if (MO.isReg() && MO.isDef() &&
+ Register::isVirtualRegister(MO.getReg()))
+ Defs.insert(MO.getReg());
+ }
+ }
+ }
+
+ // A register that is used but never defined in the cycle is considered a
+ // live-in.
+ for (const MachineBasicBlock *MBB : Cycle->blocks()) {
+ for (const MachineInstr &MI : *MBB) {
+ if (MI.isDebugInstr())
+ continue;
+ for (const MachineOperand &MO : MI.operands()) {
+ if (MO.isReg() && MO.isUse() &&
+ Register::isVirtualRegister(MO.getReg())) {
+ Register Reg = MO.getReg();
+ if (!Defs.count(Reg)) {
+ CycleLiveIns.insert(Reg);
+ }
+ }
+ }
+ }
+ }
+}
+
+/// If sinking a chain or series of chains introduces fewer liveins than defs,
+/// it is considered profitable. Remove any chains that are unprofitable.
+void MachineSinking::FilterChains(const MachineCycle *Cycle,
+ std::vector<Chain> &Chains) {
+ std::vector<Chain> FilteredChains;
+
+ // Record chain live-in frequency.
+ SmallDenseMap<Register, unsigned> RegToLiveInCount;
+ for (Chain &C : Chains) {
+ for (const Register &LiveIn : C.LiveIns) {
+ RegToLiveInCount[LiveIn]++;
+ }
+ }
+ // Add a chain to the list of FilteredChains if it has more defs than new
+ // liveins. Sharing a livein with another chain is considered as not adding a
+ // new livein.
+ for (Chain &C : Chains) {
+ assert(C.Defs.size() > 0 && "Chain with no defs?");
+ // Create a pressure set diff to see how sinking this chain would affect
+ // register pressure.
+ C.PSetDiff = SmallVector<int, 8>(TRI->getNumRegPressureSets(), 0);
+ unsigned NumNewLiveIns = 0;
+ for (const Register &LiveIn : C.LiveIns) {
+ // If exactly one chain has this register as a livein and it is not
+ // already live in the cycle, consider this as adding a new livein if
+ // sunk.
+ if (RegToLiveInCount[LiveIn] == 1) {
+ NumNewLiveIns++;
+ // Increase register pressure for this new livein.
+ const auto *RC = MRI->getRegClass(LiveIn);
+ const int *PSetID = TRI->getRegClassPressureSets(RC);
+ for (; *PSetID != -1; ++PSetID)
+ C.PSetDiff[*PSetID] += TRI->getRegClassWeight(RC).RegWeight;
+ }
+ }
+ for (const Register &Def : C.Defs) {
+ // Decrease register pressure for this def.
+ const auto *RC = MRI->getRegClass(Def);
+ const int *PSetID = TRI->getRegClassPressureSets(RC);
+ for (; *PSetID != -1; ++PSetID)
+ C.PSetDiff[*PSetID] -= TRI->getRegClassWeight(RC).RegWeight;
+ }
+ if (C.Defs.size() > NumNewLiveIns)
+ FilteredChains.push_back(C);
+ }
+ Chains = FilteredChains;
+}
+
+void MachineSinking::CalculateMaxSetPressures(
+ const MachineBasicBlock *MBB, const DenseSet<Register> &LiveIns,
+ std::vector<unsigned> &MaxSetPressure) {
+ assert(MaxSetPressure.size() == TRI->getNumRegPressureSets());
+ // Create a list of live regs to pre-populate the RPTracker.
+ SmallVector<VRegMaskOrUnit, 8> LiveInRegs;
+ // TODO: Track lanes of liveins
+ for (unsigned Reg : LiveIns)
+ LiveInRegs.emplace_back(Reg, LaneBitmask::getAll());
+ RegionPressure Pressure;
+ RegPressureTracker RPTracker(Pressure);
+ RPTracker.init(MBB->getParent(), &RegClassInfo, nullptr, MBB, MBB->end(),
+ /*TrackLaneMasks*/ false, /*TrackUntiedDefs=*/true);
+ RPTracker.addLiveRegs(LiveInRegs);
+ for (auto MII = MBB->instr_end(), MIE = MBB->instr_begin(); MII != MIE;
+ --MII) {
+ const MachineInstr &MI = *std::prev(MII);
+ if (MI.isDebugInstr() || MI.isPseudoProbe())
+ continue;
+ RegisterOperands RegOpers;
+ RegOpers.collect(MI, *TRI, *MRI, false, false);
+ RPTracker.recedeSkipDebugValues();
+ assert(&*RPTracker.getPos() == &MI && "RPTracker sync error!");
+ RPTracker.recede(RegOpers);
+ }
+ RPTracker.closeRegion();
+ for (unsigned I = 0; I < TRI->getNumRegPressureSets(); I++)
+ MaxSetPressure[I] =
+ std::max(MaxSetPressure[I], RPTracker.getPressure().MaxSetPressure[I]);
+}
+
+/// Estimate max register pressure in the MachineCycle by traversing each
+/// MachineBasicBlock with a RegPressureTracker, prepopulating the tracker with
+/// pre-calculated LiveIns.
+void MachineSinking::CalculateMaxSetPressures(
+ const MachineCycle *Cycle, const DenseSet<Register> &CycleLiveIns,
+ std::vector<unsigned> &MaxSetPressure) {
+
+ assert(MaxSetPressure.size() == TRI->getNumRegPressureSets());
+ for (const MachineBasicBlock *MBB : Cycle->blocks())
+ CalculateMaxSetPressures(MBB, CycleLiveIns, MaxSetPressure);
+
+ const DenseSet<Register> NoLiveIns;
+ CalculateMaxSetPressures(Cycle->getCyclePreheader(), NoLiveIns,
+ MaxSetPressure);
+}
+
PreservedAnalyses
MachineSinkingPass::run(MachineFunction &MF,
MachineFunctionAnalysisManager &MFAM) {
@@ -873,63 +1185,77 @@ bool MachineSinking::run(MachineFunction &MF) {
}
if (SinkInstsIntoCycle) {
- SmallVector<MachineCycle *, 8> Cycles(CI->toplevel_cycles());
SchedModel.init(STI);
- bool HasHighPressure;
-
- DenseMap<SinkItem, MachineInstr *> SunkInstrs;
-
- enum CycleSinkStage { COPY, LOW_LATENCY, AGGRESSIVE, END };
- for (unsigned Stage = CycleSinkStage::COPY; Stage != CycleSinkStage::END;
- ++Stage, SunkInstrs.clear()) {
- HasHighPressure = false;
-
- for (auto *Cycle : Cycles) {
- MachineBasicBlock *Preheader = Cycle->getCyclePreheader();
- if (!Preheader) {
- LLVM_DEBUG(dbgs() << "CycleSink: Can't find preheader\n");
- continue;
+ SmallVector<MachineCycle *, 8> Cycles(CI->toplevel_begin(),
+ CI->toplevel_end());
+ for (auto *Cycle : Cycles) {
+ MachineBasicBlock *Preheader = Cycle->getCyclePreheader();
+ if (!Preheader) {
+ LLVM_DEBUG(dbgs() << "CycleSink: Can't find preheader\n");
+ continue;
+ }
+ SmallVector<MachineInstr *, 8> Candidates;
+ FindCycleSinkCandidates(Cycle, Preheader, Candidates);
+
+ // Pre-calculate liveins to the cycle. This is used to calculate the max
+ // reg pressure in a loop, and determining how many liveins sinking a
+ // chain would introduce.
+ DenseSet<Register> CycleLiveIns;
+ calculateCycleLiveIns(Cycle, CycleLiveIns);
+
+ // Initalise MaxPressureSet accross the cycle
+ std::vector<unsigned> CycleMaxSetPressure;
+ for (unsigned I = 0; I < TRI->getNumRegPressureSets(); I++)
+ CycleMaxSetPressure.push_back(0);
+ CalculateMaxSetPressures(Cycle, CycleLiveIns, CycleMaxSetPressure);
+
+ std::vector<Chain> Chains;
+ SortIntoUseDefChains(Candidates, Chains, CycleLiveIns);
+ FilterChains(Cycle, Chains);
+
+ // Calculate how much over the max set pressure we are for each set.
+ std::vector<int> OverMaxSetPressure;
+ for (unsigned I = 0; I < TRI->getNumRegPressureSets(); I++) {
+ unsigned Limit = TRI->getRegPressureSetLimit(MF, I);
+ unsigned MaxPressure = CycleMaxSetPressure[I];
+ OverMaxSetPressure.push_back(((int)MaxPressure) - ((int)Limit));
+ LLVM_DEBUG(dbgs() << "CycleSink: Pressure set " << I << ". Limit: "
+ << Limit << ". MaxPressure: " << MaxPressure << "\n");
+ }
+ // Prioritise chains that introduce the smallest latency in the cycle.
+ llvm::sort(Chains, [](const Chain &A, const Chain &B) {
+ return (A.TotalLatency * A.NumUses) < (B.TotalLatency * B.NumUses);
+ });
+ // Helper function to determine if sinking a chain will reduce a PSet
+ // that is over the max.
+ auto ReducesCritialPSet = [&](Chain &C) {
+ for (unsigned PSetID = 0; PSetID < TRI->getNumRegPressureSets();
+ PSetID++) {
+ if (OverMaxSetPressure[PSetID] > 0 && C.PSetDiff[PSetID] < 0)
+ return true;
}
- SmallVector<MachineInstr *, 8> Candidates;
- FindCycleSinkCandidates(Cycle, Preheader, Candidates);
-
- unsigned i = 0;
-
- // Walk the candidates in reverse order so that we start with the use
- // of a def-use chain, if there is any.
- // TODO: Sort the candidates using a cost-model.
- for (MachineInstr *I : llvm::reverse(Candidates)) {
- // CycleSinkStage::COPY: Sink a limited number of copies
- if (Stage == CycleSinkStage::COPY) {
- if (i++ == SinkIntoCycleLimit) {
- LLVM_DEBUG(dbgs()
- << "CycleSink: Limit reached of instructions to "
- "be analyzed.");
- break;
- }
-
- if (!I->isCopy())
- continue;
- }
-
- // CycleSinkStage::LOW_LATENCY: sink unlimited number of instructions
- // which the target specifies as low-latency
- if (Stage == CycleSinkStage::LOW_LATENCY &&
- !TII->hasLowDefLatency(SchedModel, *I, 0))
- continue;
-
- if (!aggressivelySinkIntoCycle(Cycle, *I, SunkInstrs))
- continue;
+ return false;
+ };
+ auto IsCopyOrLowLatency = [&](const Chain &C) {
+ for (const MachineInstr *MI : C.Instrs)
+ if (!MI->isCopy() && !TII->hasLowDefLatency(SchedModel, *MI, 0))
+ return false;
+ return true;
+ };
+ // Sink the chains if they reduce a critical pressure set.
+ for (auto &C : Chains) {
+ // If the instruction is low latency, sink anyway.
+ if (!IsCopyOrLowLatency(C) && !ReducesCritialPSet(C))
+ continue;
+ if (NumCycleSunk + C.Instrs.size() >= SinkIntoCycleLimit) {
+ LLVM_DEBUG(dbgs() << "CycleSink: Not sinking Chain as will exceed "
+ "Instruction Sink limit");
+ break;
+ } else if (SinkChainIntoCycle(Cycle, C, OverMaxSetPressure)) {
+ NumCycleSunk += C.Instrs.size();
EverMadeChange = true;
- ++NumCycleSunk;
}
-
- // Recalculate the pressure after sinking
- if (!HasHighPressure)
- HasHighPressure = registerPressureExceedsLimit(*Preheader);
}
- if (!HasHighPressure)
- break;
}
}
@@ -1242,21 +1568,6 @@ bool MachineSinking::r...
[truncated]
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
RFC link
Instructions are grouped into a sink unit called a "chain" and
considered whether it's profitable to sink as a unit by looking at how
many liveins and defs the chain has. If the chain has more defs than
liveins, then sinking it will (almost always) reduce register pressure
by the difference. Multiple chains sharing a livein are considered to
not introduce this livein.
The max register pressure is estimated in the cycle by estimating cycle
liveins and using a reg pressure tracker for each block. Chains are sunk
if they positively affect the critical pressure sets until the set is no
longer above the max.
This approach relies on the fact that values in the preheader will be
live throughout the entire loop. So sinking a chain next to a use will
almost always reduce pressure.