//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "regalloc"
-
-#include "RenderMachineFunction.h"
-#include "Splitter.h"
-#include "VirtRegMap.h"
-#include "VirtRegRewriter.h"
+#include "llvm/CodeGen/RegAllocPBQP.h"
+#include "RegisterCoalescer.h"
+#include "Spiller.h"
+#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/CalcSpillWeights.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/LiveRangeEdit.h"
#include "llvm/CodeGen/LiveStackAnalysis.h"
-#include "llvm/CodeGen/RegAllocPBQP.h"
+#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
+#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/CodeGen/PBQP/HeuristicSolver.h"
-#include "llvm/CodeGen/PBQP/Graph.h"
-#include "llvm/CodeGen/PBQP/Heuristics/Briggs.h"
#include "llvm/CodeGen/RegAllocRegistry.h"
-#include "llvm/CodeGen/RegisterCoalescer.h"
+#include "llvm/CodeGen/VirtRegMap.h"
+#include "llvm/IR/Module.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/FileSystem.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
#include <limits>
#include <memory>
+#include <queue>
#include <set>
+#include <sstream>
#include <vector>
-namespace llvm {
+using namespace llvm;
-using namespace PBQP;
- using namespace PBQP::Heuristics;
+#define DEBUG_TYPE "regalloc"
static RegisterRegAlloc
-registerPBQPRepAlloc("pbqp", "PBQP register allocator",
- llvm::createPBQPRegisterAllocator);
+RegisterPBQPRepAlloc("pbqp", "PBQP register allocator",
+ createDefaultPBQPRegisterAllocator);
static cl::opt<bool>
-pbqpCoalescing("pbqp-coalescing",
+PBQPCoalescing("pbqp-coalescing",
cl::desc("Attempt coalescing during PBQP register allocation."),
cl::init(false), cl::Hidden);
+#ifndef NDEBUG
static cl::opt<bool>
-pbqpBuilder("pbqp-builder",
- cl::desc("Use new builder system."),
- cl::init(false), cl::Hidden);
-
-
-static cl::opt<bool>
-pbqpPreSplitting("pbqp-pre-splitting",
- cl::desc("Pre-splite before PBQP register allocation."),
- cl::init(false), cl::Hidden);
-
-char RegAllocPBQP::ID = 0;
-
-unsigned PBQPRAProblem::getVRegForNode(PBQP::Graph::ConstNodeItr node) const {
- Node2VReg::const_iterator vregItr = node2VReg.find(node);
- assert(vregItr != node2VReg.end() && "No vreg for node.");
- return vregItr->second;
-}
-
-PBQP::Graph::NodeItr PBQPRAProblem::getNodeForVReg(unsigned vreg) const {
- VReg2Node::const_iterator nodeItr = vreg2Node.find(vreg);
- assert(nodeItr != vreg2Node.end() && "No node for vreg.");
- return nodeItr->second;
-
-}
-
-const PBQPRAProblem::AllowedSet&
- PBQPRAProblem::getAllowedSet(unsigned vreg) const {
- AllowedSetMap::const_iterator allowedSetItr = allowedSets.find(vreg);
- assert(allowedSetItr != allowedSets.end() && "No pregs for vreg.");
- const AllowedSet &allowedSet = allowedSetItr->second;
- return allowedSet;
-}
-
-unsigned PBQPRAProblem::getPRegForOption(unsigned vreg, unsigned option) const {
- assert(isPRegOption(vreg, option) && "Not a preg option.");
-
- const AllowedSet& allowedSet = getAllowedSet(vreg);
- assert(option <= allowedSet.size() && "Option outside allowed set.");
- return allowedSet[option - 1];
-}
-
-std::auto_ptr<PBQPRAProblem> PBQPBuilder::build(
- MachineFunction *mf,
- const LiveIntervals *lis,
- const RegSet &vregs) {
-
- typedef std::vector<const LiveInterval*> LIVector;
-
- MachineRegisterInfo *mri = &mf->getRegInfo();
- const TargetRegisterInfo *tri = mf->getTarget().getRegisterInfo();
-
- std::auto_ptr<PBQPRAProblem> p(new PBQPRAProblem());
- PBQP::Graph &g = p->getGraph();
- RegSet pregs;
+PBQPDumpGraphs("pbqp-dump-graphs",
+ cl::desc("Dump graphs for each function/round in the compilation unit."),
+ cl::init(false), cl::Hidden);
+#endif
+
+namespace {
+
+///
+/// PBQP based allocators solve the register allocation problem by mapping
+/// register allocation problems to Partitioned Boolean Quadratic
+/// Programming problems.
+class RegAllocPBQP : public MachineFunctionPass {
+public:
+
+ static char ID;
+
+ /// Construct a PBQP register allocator.
+ RegAllocPBQP(char *cPassID = nullptr)
+ : MachineFunctionPass(ID), customPassID(cPassID) {
+ initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
+ initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
+ initializeLiveStacksPass(*PassRegistry::getPassRegistry());
+ initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
+ }
- // Collect the set of preg intervals, record that they're used in the MF.
- for (LiveIntervals::const_iterator itr = lis->begin(), end = lis->end();
- itr != end; ++itr) {
- if (TargetRegisterInfo::isPhysicalRegister(itr->first)) {
- pregs.insert(itr->first);
- mri->setPhysRegUsed(itr->first);
- }
+ /// Return the pass name.
+ const char* getPassName() const override {
+ return "PBQP Register Allocator";
}
- BitVector reservedRegs = tri->getReservedRegs(*mf);
+ /// PBQP analysis usage.
+ void getAnalysisUsage(AnalysisUsage &au) const override;
- // Iterate over vregs.
- for (RegSet::const_iterator vregItr = vregs.begin(), vregEnd = vregs.end();
- vregItr != vregEnd; ++vregItr) {
- unsigned vreg = *vregItr;
- const TargetRegisterClass *trc = mri->getRegClass(vreg);
- const LiveInterval *vregLI = &lis->getInterval(vreg);
+ /// Perform register allocation
+ bool runOnMachineFunction(MachineFunction &MF) override;
- // Compute an initial allowed set for the current vreg.
- typedef std::vector<unsigned> VRAllowed;
- VRAllowed vrAllowed;
- for (TargetRegisterClass::iterator aoItr = trc->allocation_order_begin(*mf),
- aoEnd = trc->allocation_order_end(*mf);
- aoItr != aoEnd; ++aoItr) {
- unsigned preg = *aoItr;
- if (!reservedRegs.test(preg)) {
- vrAllowed.push_back(preg);
- }
- }
-
- // Remove any physical registers which overlap.
- for (RegSet::const_iterator pregItr = pregs.begin(),
- pregEnd = pregs.end();
- pregItr != pregEnd; ++pregItr) {
- unsigned preg = *pregItr;
- const LiveInterval *pregLI = &lis->getInterval(preg);
+private:
- if (pregLI->empty())
- continue;
+ typedef std::map<const LiveInterval*, unsigned> LI2NodeMap;
+ typedef std::vector<const LiveInterval*> Node2LIMap;
+ typedef std::vector<unsigned> AllowedSet;
+ typedef std::vector<AllowedSet> AllowedSetMap;
+ typedef std::pair<unsigned, unsigned> RegPair;
+ typedef std::map<RegPair, PBQP::PBQPNum> CoalesceMap;
+ typedef std::set<unsigned> RegSet;
- if (!vregLI->overlaps(*pregLI))
- continue;
+ char *customPassID;
- // Remove the register from the allowed set.
- VRAllowed::iterator eraseItr =
- std::find(vrAllowed.begin(), vrAllowed.end(), preg);
+ RegSet VRegsToAlloc, EmptyIntervalVRegs;
- if (eraseItr != vrAllowed.end()) {
- vrAllowed.erase(eraseItr);
- }
+ /// \brief Finds the initial set of vreg intervals to allocate.
+ void findVRegIntervalsToAlloc(const MachineFunction &MF, LiveIntervals &LIS);
- // Also remove any aliases.
- const unsigned *aliasItr = tri->getAliasSet(preg);
- if (aliasItr != 0) {
- for (; *aliasItr != 0; ++aliasItr) {
- VRAllowed::iterator eraseItr =
- std::find(vrAllowed.begin(), vrAllowed.end(), *aliasItr);
+ /// \brief Constructs an initial graph.
+ void initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM, Spiller &VRegSpiller);
- if (eraseItr != vrAllowed.end()) {
- vrAllowed.erase(eraseItr);
- }
- }
- }
- }
+ /// \brief Spill the given VReg.
+ void spillVReg(unsigned VReg, SmallVectorImpl<unsigned> &NewIntervals,
+ MachineFunction &MF, LiveIntervals &LIS, VirtRegMap &VRM,
+ Spiller &VRegSpiller);
- // Construct the node.
- PBQP::Graph::NodeItr node =
- g.addNode(PBQP::Vector(vrAllowed.size() + 1, 0));
+ /// \brief Given a solved PBQP problem maps this solution back to a register
+ /// assignment.
+ bool mapPBQPToRegAlloc(const PBQPRAGraph &G,
+ const PBQP::Solution &Solution,
+ VirtRegMap &VRM,
+ Spiller &VRegSpiller);
- // Record the mapping and allowed set in the problem.
- p->recordVReg(vreg, node, vrAllowed.begin(), vrAllowed.end());
+ /// \brief Postprocessing before final spilling. Sets basic block "live in"
+ /// variables.
+ void finalizeAlloc(MachineFunction &MF, LiveIntervals &LIS,
+ VirtRegMap &VRM) const;
- PBQP::PBQPNum spillCost = (vregLI->weight != 0.0) ?
- vregLI->weight : std::numeric_limits<PBQP::PBQPNum>::min();
+};
- addSpillCosts(g.getNodeCosts(node), spillCost);
- }
+char RegAllocPBQP::ID = 0;
- for (RegSet::const_iterator vr1Itr = vregs.begin(), vrEnd = vregs.end();
- vr1Itr != vrEnd; ++vr1Itr) {
- unsigned vr1 = *vr1Itr;
- const LiveInterval &l1 = lis->getInterval(vr1);
- const PBQPRAProblem::AllowedSet &vr1Allowed = p->getAllowedSet(vr1);
-
- for (RegSet::iterator vr2Itr = llvm::next(vr1Itr);
- vr2Itr != vrEnd; ++vr2Itr) {
- unsigned vr2 = *vr2Itr;
- const LiveInterval &l2 = lis->getInterval(vr2);
- const PBQPRAProblem::AllowedSet &vr2Allowed = p->getAllowedSet(vr2);
-
- assert(!l2.empty() && "Empty interval in vreg set?");
- if (l1.overlaps(l2)) {
- PBQP::Graph::EdgeItr edge =
- g.addEdge(p->getNodeForVReg(vr1), p->getNodeForVReg(vr2),
- PBQP::Matrix(vr1Allowed.size()+1, vr2Allowed.size()+1, 0));
-
- addInterferenceCosts(g.getEdgeCosts(edge), vr1Allowed, vr2Allowed, tri);
- }
+/// @brief Set spill costs for each node in the PBQP reg-alloc graph.
+class SpillCosts : public PBQPRAConstraint {
+public:
+ void apply(PBQPRAGraph &G) override {
+ LiveIntervals &LIS = G.getMetadata().LIS;
+
+ // A minimum spill costs, so that register constraints can can be set
+ // without normalization in the [0.0:MinSpillCost( interval.
+ const PBQP::PBQPNum MinSpillCost = 10.0;
+
+ for (auto NId : G.nodeIds()) {
+ PBQP::PBQPNum SpillCost =
+ LIS.getInterval(G.getNodeMetadata(NId).getVReg()).weight;
+ if (SpillCost == 0.0)
+ SpillCost = std::numeric_limits<PBQP::PBQPNum>::min();
+ else
+ SpillCost += MinSpillCost;
+ PBQPRAGraph::RawVector NodeCosts(G.getNodeCosts(NId));
+ NodeCosts[PBQP::RegAlloc::getSpillOptionIdx()] = SpillCost;
+ G.setNodeCosts(NId, std::move(NodeCosts));
}
}
+};
- return p;
-}
-
-void PBQPBuilder::addSpillCosts(PBQP::Vector &costVec,
- PBQP::PBQPNum spillCost) {
- costVec[0] = spillCost;
-}
-
-void PBQPBuilder::addInterferenceCosts(PBQP::Matrix &costMat,
- const PBQPRAProblem::AllowedSet &vr1Allowed,
- const PBQPRAProblem::AllowedSet &vr2Allowed,
- const TargetRegisterInfo *tri) {
- assert(costMat.getRows() == vr1Allowed.size() + 1 && "Matrix height mismatch.");
- assert(costMat.getCols() == vr2Allowed.size() + 1 && "Matrix width mismatch.");
+/// @brief Add interference edges between overlapping vregs.
+class Interference : public PBQPRAConstraint {
+private:
- for (unsigned i = 0; i < vr1Allowed.size(); ++i) {
- unsigned preg1 = vr1Allowed[i];
+ typedef const PBQP::RegAlloc::AllowedRegVector* AllowedRegVecPtr;
+ typedef std::pair<AllowedRegVecPtr, AllowedRegVecPtr> IMatrixKey;
+ typedef DenseMap<IMatrixKey, PBQPRAGraph::MatrixPtr> IMatrixCache;
- for (unsigned j = 0; j < vr2Allowed.size(); ++j) {
- unsigned preg2 = vr2Allowed[j];
+ // Holds (Interval, CurrentSegmentID, and NodeId). The first two are required
+ // for the fast interference graph construction algorithm. The last is there
+ // to save us from looking up node ids via the VRegToNode map in the graph
+ // metadata.
+ typedef std::tuple<LiveInterval*, size_t, PBQP::GraphBase::NodeId>
+ IntervalInfo;
- if (tri->regsOverlap(preg1, preg2)) {
- costMat[i + 1][j + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
- }
- }
+ static SlotIndex getStartPoint(const IntervalInfo &I) {
+ return std::get<0>(I)->segments[std::get<1>(I)].start;
}
-}
-
-
-
-void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
- au.addRequired<SlotIndexes>();
- au.addPreserved<SlotIndexes>();
- au.addRequired<LiveIntervals>();
- //au.addRequiredID(SplitCriticalEdgesID);
- au.addRequired<RegisterCoalescer>();
- au.addRequired<CalculateSpillWeights>();
- au.addRequired<LiveStacks>();
- au.addPreserved<LiveStacks>();
- au.addRequired<MachineLoopInfo>();
- au.addPreserved<MachineLoopInfo>();
- if (pbqpPreSplitting)
- au.addRequired<LoopSplitter>();
- au.addRequired<VirtRegMap>();
- au.addRequired<RenderMachineFunction>();
- MachineFunctionPass::getAnalysisUsage(au);
-}
-template <typename RegContainer>
-PBQP::Vector RegAllocPBQP::buildCostVector(unsigned vReg,
- const RegContainer &allowed,
- const CoalesceMap &coalesces,
- PBQP::PBQPNum spillCost) const {
-
- typedef typename RegContainer::const_iterator AllowedItr;
-
- // Allocate vector. Additional element (0th) used for spill option
- PBQP::Vector v(allowed.size() + 1, 0);
+ static SlotIndex getEndPoint(const IntervalInfo &I) {
+ return std::get<0>(I)->segments[std::get<1>(I)].end;
+ }
- v[0] = spillCost;
+ static PBQP::GraphBase::NodeId getNodeId(const IntervalInfo &I) {
+ return std::get<2>(I);
+ }
- // Iterate over the allowed registers inserting coalesce benefits if there
- // are any.
- unsigned ai = 0;
- for (AllowedItr itr = allowed.begin(), end = allowed.end();
- itr != end; ++itr, ++ai) {
+ static bool lowestStartPoint(const IntervalInfo &I1,
+ const IntervalInfo &I2) {
+ // Condition reversed because priority queue has the *highest* element at
+ // the front, rather than the lowest.
+ return getStartPoint(I1) > getStartPoint(I2);
+ }
- unsigned pReg = *itr;
+ static bool lowestEndPoint(const IntervalInfo &I1,
+ const IntervalInfo &I2) {
+ SlotIndex E1 = getEndPoint(I1);
+ SlotIndex E2 = getEndPoint(I2);
- CoalesceMap::const_iterator cmItr =
- coalesces.find(RegPair(vReg, pReg));
+ if (E1 < E2)
+ return true;
- // No coalesce - on to the next preg.
- if (cmItr == coalesces.end())
- continue;
+ if (E1 > E2)
+ return false;
- // We have a coalesce - insert the benefit.
- v[ai + 1] = -cmItr->second;
+ // If two intervals end at the same point, we need a way to break the tie or
+ // the set will assume they're actually equal and refuse to insert a
+ // "duplicate". Just compare the vregs - fast and guaranteed unique.
+ return std::get<0>(I1)->reg < std::get<0>(I2)->reg;
}
- return v;
-}
-
-template <typename RegContainer>
-PBQP::Matrix* RegAllocPBQP::buildInterferenceMatrix(
- const RegContainer &allowed1, const RegContainer &allowed2) const {
-
- typedef typename RegContainer::const_iterator RegContainerIterator;
-
- // Construct a PBQP matrix representing the cost of allocation options. The
- // rows and columns correspond to the allocation options for the two live
- // intervals. Elements will be infinite where corresponding registers alias,
- // since we cannot allocate aliasing registers to interfering live intervals.
- // All other elements (non-aliasing combinations) will have zero cost. Note
- // that the spill option (element 0,0) has zero cost, since we can allocate
- // both intervals to memory safely (the cost for each individual allocation
- // to memory is accounted for by the cost vectors for each live interval).
- PBQP::Matrix *m =
- new PBQP::Matrix(allowed1.size() + 1, allowed2.size() + 1, 0);
-
- // Assume this is a zero matrix until proven otherwise. Zero matrices occur
- // between interfering live ranges with non-overlapping register sets (e.g.
- // non-overlapping reg classes, or disjoint sets of allowed regs within the
- // same class). The term "overlapping" is used advisedly: sets which do not
- // intersect, but contain registers which alias, will have non-zero matrices.
- // We optimize zero matrices away to improve solver speed.
- bool isZeroMatrix = true;
-
-
- // Row index. Starts at 1, since the 0th row is for the spill option, which
- // is always zero.
- unsigned ri = 1;
-
- // Iterate over allowed sets, insert infinities where required.
- for (RegContainerIterator a1Itr = allowed1.begin(), a1End = allowed1.end();
- a1Itr != a1End; ++a1Itr) {
-
- // Column index, starts at 1 as for row index.
- unsigned ci = 1;
- unsigned reg1 = *a1Itr;
-
- for (RegContainerIterator a2Itr = allowed2.begin(), a2End = allowed2.end();
- a2Itr != a2End; ++a2Itr) {
-
- unsigned reg2 = *a2Itr;
-
- // If the row/column regs are identical or alias insert an infinity.
- if (tri->regsOverlap(reg1, reg2)) {
- (*m)[ri][ci] = std::numeric_limits<PBQP::PBQPNum>::infinity();
- isZeroMatrix = false;
- }
-
- ++ci;
- }
-
- ++ri;
+ static bool isAtLastSegment(const IntervalInfo &I) {
+ return std::get<1>(I) == std::get<0>(I)->size() - 1;
}
- // If this turns out to be a zero matrix...
- if (isZeroMatrix) {
- // free it and return null.
- delete m;
- return 0;
+ static IntervalInfo nextSegment(const IntervalInfo &I) {
+ return std::make_tuple(std::get<0>(I), std::get<1>(I) + 1, std::get<2>(I));
}
- // ...otherwise return the cost matrix.
- return m;
-}
+public:
+
+ void apply(PBQPRAGraph &G) override {
+ // The following is loosely based on the linear scan algorithm introduced in
+ // "Linear Scan Register Allocation" by Poletto and Sarkar. This version
+ // isn't linear, because the size of the active set isn't bound by the
+ // number of registers, but rather the size of the largest clique in the
+ // graph. Still, we expect this to be better than N^2.
+ LiveIntervals &LIS = G.getMetadata().LIS;
+
+ // Interferenc matrices are incredibly regular - they're only a function of
+ // the allowed sets, so we cache them to avoid the overhead of constructing
+ // and uniquing them.
+ IMatrixCache C;
+
+ typedef std::set<IntervalInfo, decltype(&lowestEndPoint)> IntervalSet;
+ typedef std::priority_queue<IntervalInfo, std::vector<IntervalInfo>,
+ decltype(&lowestStartPoint)> IntervalQueue;
+ IntervalSet Active(lowestEndPoint);
+ IntervalQueue Inactive(lowestStartPoint);
+
+ // Start by building the inactive set.
+ for (auto NId : G.nodeIds()) {
+ unsigned VReg = G.getNodeMetadata(NId).getVReg();
+ LiveInterval &LI = LIS.getInterval(VReg);
+ assert(!LI.empty() && "PBQP graph contains node for empty interval");
+ Inactive.push(std::make_tuple(&LI, 0, NId));
+ }
-template <typename RegContainer>
-PBQP::Matrix* RegAllocPBQP::buildCoalescingMatrix(
- const RegContainer &allowed1, const RegContainer &allowed2,
- PBQP::PBQPNum cBenefit) const {
-
- typedef typename RegContainer::const_iterator RegContainerIterator;
-
- // Construct a PBQP Matrix representing the benefits of coalescing. As with
- // interference matrices the rows and columns represent allowed registers
- // for the LiveIntervals which are (potentially) to be coalesced. The amount
- // -cBenefit will be placed in any element representing the same register
- // for both intervals.
- PBQP::Matrix *m =
- new PBQP::Matrix(allowed1.size() + 1, allowed2.size() + 1, 0);
-
- // Reset costs to zero.
- m->reset(0);
-
- // Assume the matrix is zero till proven otherwise. Zero matrices will be
- // optimized away as in the interference case.
- bool isZeroMatrix = true;
-
- // Row index. Starts at 1, since the 0th row is for the spill option, which
- // is always zero.
- unsigned ri = 1;
-
- // Iterate over the allowed sets, insert coalescing benefits where
- // appropriate.
- for (RegContainerIterator a1Itr = allowed1.begin(), a1End = allowed1.end();
- a1Itr != a1End; ++a1Itr) {
-
- // Column index, starts at 1 as for row index.
- unsigned ci = 1;
- unsigned reg1 = *a1Itr;
-
- for (RegContainerIterator a2Itr = allowed2.begin(), a2End = allowed2.end();
- a2Itr != a2End; ++a2Itr) {
-
- // If the row and column represent the same register insert a beneficial
- // cost to preference this allocation - it would allow us to eliminate a
- // move instruction.
- if (reg1 == *a2Itr) {
- (*m)[ri][ci] = -cBenefit;
- isZeroMatrix = false;
+ while (!Inactive.empty()) {
+ // Tentatively grab the "next" interval - this choice may be overriden
+ // below.
+ IntervalInfo Cur = Inactive.top();
+
+ // Retire any active intervals that end before Cur starts.
+ IntervalSet::iterator RetireItr = Active.begin();
+ while (RetireItr != Active.end() &&
+ (getEndPoint(*RetireItr) <= getStartPoint(Cur))) {
+ // If this interval has subsequent segments, add the next one to the
+ // inactive list.
+ if (!isAtLastSegment(*RetireItr))
+ Inactive.push(nextSegment(*RetireItr));
+
+ ++RetireItr;
}
+ Active.erase(Active.begin(), RetireItr);
+
+ // One of the newly retired segments may actually start before the
+ // Cur segment, so re-grab the front of the inactive list.
+ Cur = Inactive.top();
+ Inactive.pop();
+
+ // At this point we know that Cur overlaps all active intervals. Add the
+ // interference edges.
+ PBQP::GraphBase::NodeId NId = getNodeId(Cur);
+ for (const auto &A : Active) {
+ PBQP::GraphBase::NodeId MId = getNodeId(A);
+
+ // Check that we haven't already added this edge
+ // FIXME: findEdge is expensive in the worst case (O(max_clique(G))).
+ // It might be better to replace this with a local bit-matrix.
+ if (G.findEdge(NId, MId) != PBQPRAGraph::invalidEdgeId())
+ continue;
- ++ci;
- }
-
- ++ri;
- }
+ // This is a new edge - add it to the graph.
+ createInterferenceEdge(G, NId, MId, C);
+ }
- // If this turns out to be a zero matrix...
- if (isZeroMatrix) {
- // ...free it and return null.
- delete m;
- return 0;
+ // Finally, add Cur to the Active set.
+ Active.insert(Cur);
+ }
}
- return m;
-}
-
-RegAllocPBQP::CoalesceMap RegAllocPBQP::findCoalesces() {
-
- typedef MachineFunction::const_iterator MFIterator;
- typedef MachineBasicBlock::const_iterator MBBIterator;
- typedef LiveInterval::const_vni_iterator VNIIterator;
-
- CoalesceMap coalescesFound;
+private:
- // To find coalesces we need to iterate over the function looking for
- // copy instructions.
- for (MFIterator bbItr = mf->begin(), bbEnd = mf->end();
- bbItr != bbEnd; ++bbItr) {
+ void createInterferenceEdge(PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
+ PBQPRAGraph::NodeId MId, IMatrixCache &C) {
- const MachineBasicBlock *mbb = &*bbItr;
+ const TargetRegisterInfo &TRI =
+ *G.getMetadata().MF.getSubtarget().getRegisterInfo();
- for (MBBIterator iItr = mbb->begin(), iEnd = mbb->end();
- iItr != iEnd; ++iItr) {
+ const auto &NRegs = G.getNodeMetadata(NId).getAllowedRegs();
+ const auto &MRegs = G.getNodeMetadata(MId).getAllowedRegs();
- const MachineInstr *instr = &*iItr;
+ // Try looking the edge costs up in the IMatrixCache first.
+ IMatrixKey K(&NRegs, &MRegs);
+ IMatrixCache::iterator I = C.find(K);
+ if (I != C.end()) {
+ G.addEdgeBypassingCostAllocator(NId, MId, I->second);
+ return;
+ }
- // If this isn't a copy then continue to the next instruction.
- if (!instr->isCopy())
- continue;
+ PBQPRAGraph::RawMatrix M(NRegs.size() + 1, MRegs.size() + 1, 0);
+ for (unsigned I = 0; I != NRegs.size(); ++I) {
+ unsigned PRegN = NRegs[I];
+ for (unsigned J = 0; J != MRegs.size(); ++J) {
+ unsigned PRegM = MRegs[J];
+ if (TRI.regsOverlap(PRegN, PRegM))
+ M[I + 1][J + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
+ }
+ }
- unsigned srcReg = instr->getOperand(1).getReg();
- unsigned dstReg = instr->getOperand(0).getReg();
+ PBQPRAGraph::EdgeId EId = G.addEdge(NId, MId, std::move(M));
+ C[K] = G.getEdgeCostsPtr(EId);
+ }
+};
- // If the registers are already the same our job is nice and easy.
- if (dstReg == srcReg)
- continue;
- bool srcRegIsPhysical = TargetRegisterInfo::isPhysicalRegister(srcReg),
- dstRegIsPhysical = TargetRegisterInfo::isPhysicalRegister(dstReg);
+class Coalescing : public PBQPRAConstraint {
+public:
+ void apply(PBQPRAGraph &G) override {
+ MachineFunction &MF = G.getMetadata().MF;
+ MachineBlockFrequencyInfo &MBFI = G.getMetadata().MBFI;
+ CoalescerPair CP(*MF.getSubtarget().getRegisterInfo());
- // If both registers are physical then we can't coalesce.
- if (srcRegIsPhysical && dstRegIsPhysical)
- continue;
+ // Scan the machine function and add a coalescing cost whenever CoalescerPair
+ // gives the Ok.
+ for (const auto &MBB : MF) {
+ for (const auto &MI : MBB) {
- // If it's a copy that includes two virtual register but the source and
- // destination classes differ then we can't coalesce.
- if (!srcRegIsPhysical && !dstRegIsPhysical &&
- mri->getRegClass(srcReg) != mri->getRegClass(dstReg))
- continue;
-
- // If one is physical and one is virtual, check that the physical is
- // allocatable in the class of the virtual.
- if (srcRegIsPhysical && !dstRegIsPhysical) {
- const TargetRegisterClass *dstRegClass = mri->getRegClass(dstReg);
- if (std::find(dstRegClass->allocation_order_begin(*mf),
- dstRegClass->allocation_order_end(*mf), srcReg) ==
- dstRegClass->allocation_order_end(*mf))
- continue;
- }
- if (!srcRegIsPhysical && dstRegIsPhysical) {
- const TargetRegisterClass *srcRegClass = mri->getRegClass(srcReg);
- if (std::find(srcRegClass->allocation_order_begin(*mf),
- srcRegClass->allocation_order_end(*mf), dstReg) ==
- srcRegClass->allocation_order_end(*mf))
+ // Skip not-coalescable or already coalesced copies.
+ if (!CP.setRegisters(&MI) || CP.getSrcReg() == CP.getDstReg())
continue;
- }
-
- // If we've made it here we have a copy with compatible register classes.
- // We can probably coalesce, but we need to consider overlap.
- const LiveInterval *srcLI = &lis->getInterval(srcReg),
- *dstLI = &lis->getInterval(dstReg);
- if (srcLI->overlaps(*dstLI)) {
- // Even in the case of an overlap we might still be able to coalesce,
- // but we need to make sure that no definition of either range occurs
- // while the other range is live.
+ unsigned DstReg = CP.getDstReg();
+ unsigned SrcReg = CP.getSrcReg();
- // Otherwise start by assuming we're ok.
- bool badDef = false;
+ const float Scale = 1.0f / MBFI.getEntryFreq();
+ PBQP::PBQPNum CBenefit = MBFI.getBlockFreq(&MBB).getFrequency() * Scale;
- // Test all defs of the source range.
- for (VNIIterator
- vniItr = srcLI->vni_begin(), vniEnd = srcLI->vni_end();
- vniItr != vniEnd; ++vniItr) {
-
- // If we find a poorly defined def we err on the side of caution.
- if (!(*vniItr)->def.isValid()) {
- badDef = true;
- break;
- }
+ if (CP.isPhys()) {
+ if (!MF.getRegInfo().isAllocatable(DstReg))
+ continue;
- // If we find a def that kills the coalescing opportunity then
- // record it and break from the loop.
- if (dstLI->liveAt((*vniItr)->def)) {
- badDef = true;
- break;
- }
- }
+ PBQPRAGraph::NodeId NId = G.getMetadata().getNodeIdForVReg(SrcReg);
- // If we have a bad def give up, continue to the next instruction.
- if (badDef)
- continue;
+ const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed =
+ G.getNodeMetadata(NId).getAllowedRegs();
- // Otherwise test definitions of the destination range.
- for (VNIIterator
- vniItr = dstLI->vni_begin(), vniEnd = dstLI->vni_end();
- vniItr != vniEnd; ++vniItr) {
+ unsigned PRegOpt = 0;
+ while (PRegOpt < Allowed.size() && Allowed[PRegOpt] != DstReg)
+ ++PRegOpt;
- // We want to make sure we skip the copy instruction itself.
- if ((*vniItr)->getCopy() == instr)
- continue;
-
- if (!(*vniItr)->def.isValid()) {
- badDef = true;
- break;
+ if (PRegOpt < Allowed.size()) {
+ PBQPRAGraph::RawVector NewCosts(G.getNodeCosts(NId));
+ NewCosts[PRegOpt + 1] -= CBenefit;
+ G.setNodeCosts(NId, std::move(NewCosts));
}
-
- if (srcLI->liveAt((*vniItr)->def)) {
- badDef = true;
- break;
+ } else {
+ PBQPRAGraph::NodeId N1Id = G.getMetadata().getNodeIdForVReg(DstReg);
+ PBQPRAGraph::NodeId N2Id = G.getMetadata().getNodeIdForVReg(SrcReg);
+ const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed1 =
+ &G.getNodeMetadata(N1Id).getAllowedRegs();
+ const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed2 =
+ &G.getNodeMetadata(N2Id).getAllowedRegs();
+
+ PBQPRAGraph::EdgeId EId = G.findEdge(N1Id, N2Id);
+ if (EId == G.invalidEdgeId()) {
+ PBQPRAGraph::RawMatrix Costs(Allowed1->size() + 1,
+ Allowed2->size() + 1, 0);
+ addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
+ G.addEdge(N1Id, N2Id, std::move(Costs));
+ } else {
+ if (G.getEdgeNode1Id(EId) == N2Id) {
+ std::swap(N1Id, N2Id);
+ std::swap(Allowed1, Allowed2);
+ }
+ PBQPRAGraph::RawMatrix Costs(G.getEdgeCosts(EId));
+ addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
+ G.setEdgeCosts(EId, std::move(Costs));
}
}
+ }
+ }
+ }
- // As before a bad def we give up and continue to the next instr.
- if (badDef)
- continue;
+private:
+
+ void addVirtRegCoalesce(
+ PBQPRAGraph::RawMatrix &CostMat,
+ const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed1,
+ const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed2,
+ PBQP::PBQPNum Benefit) {
+ assert(CostMat.getRows() == Allowed1.size() + 1 && "Size mismatch.");
+ assert(CostMat.getCols() == Allowed2.size() + 1 && "Size mismatch.");
+ for (unsigned I = 0; I != Allowed1.size(); ++I) {
+ unsigned PReg1 = Allowed1[I];
+ for (unsigned J = 0; J != Allowed2.size(); ++J) {
+ unsigned PReg2 = Allowed2[J];
+ if (PReg1 == PReg2)
+ CostMat[I + 1][J + 1] -= Benefit;
}
+ }
+ }
- // If we make it to here then either the ranges didn't overlap, or they
- // did, but none of their definitions would prevent us from coalescing.
- // We're good to go with the coalesce.
+};
- float cBenefit = std::pow(10.0f, (float)loopInfo->getLoopDepth(mbb)) / 5.0;
+} // End anonymous namespace.
- coalescesFound[RegPair(srcReg, dstReg)] = cBenefit;
- coalescesFound[RegPair(dstReg, srcReg)] = cBenefit;
- }
+// Out-of-line destructor/anchor for PBQPRAConstraint.
+PBQPRAConstraint::~PBQPRAConstraint() {}
+void PBQPRAConstraint::anchor() {}
+void PBQPRAConstraintList::anchor() {}
- }
-
- return coalescesFound;
+void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
+ au.setPreservesCFG();
+ au.addRequired<AliasAnalysis>();
+ au.addPreserved<AliasAnalysis>();
+ au.addRequired<SlotIndexes>();
+ au.addPreserved<SlotIndexes>();
+ au.addRequired<LiveIntervals>();
+ au.addPreserved<LiveIntervals>();
+ //au.addRequiredID(SplitCriticalEdgesID);
+ if (customPassID)
+ au.addRequiredID(*customPassID);
+ au.addRequired<LiveStacks>();
+ au.addPreserved<LiveStacks>();
+ au.addRequired<MachineBlockFrequencyInfo>();
+ au.addPreserved<MachineBlockFrequencyInfo>();
+ au.addRequired<MachineLoopInfo>();
+ au.addPreserved<MachineLoopInfo>();
+ au.addRequired<MachineDominatorTree>();
+ au.addPreserved<MachineDominatorTree>();
+ au.addRequired<VirtRegMap>();
+ au.addPreserved<VirtRegMap>();
+ MachineFunctionPass::getAnalysisUsage(au);
}
-void RegAllocPBQP::findVRegIntervalsToAlloc() {
+void RegAllocPBQP::findVRegIntervalsToAlloc(const MachineFunction &MF,
+ LiveIntervals &LIS) {
+ const MachineRegisterInfo &MRI = MF.getRegInfo();
// Iterate over all live ranges.
- for (LiveIntervals::iterator itr = lis->begin(), end = lis->end();
- itr != end; ++itr) {
-
- // Ignore physical ones.
- if (TargetRegisterInfo::isPhysicalRegister(itr->first))
+ for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) {
+ unsigned Reg = TargetRegisterInfo::index2VirtReg(I);
+ if (MRI.reg_nodbg_empty(Reg))
continue;
-
- LiveInterval *li = itr->second;
+ LiveInterval &LI = LIS.getInterval(Reg);
// If this live interval is non-empty we will use pbqp to allocate it.
// Empty intervals we allocate in a simple post-processing stage in
// finalizeAlloc.
- if (!li->empty()) {
- vregsToAlloc.insert(li->reg);
- }
- else {
- emptyIntervalVRegs.insert(li->reg);
+ if (!LI.empty()) {
+ VRegsToAlloc.insert(LI.reg);
+ } else {
+ EmptyIntervalVRegs.insert(LI.reg);
}
}
}
-PBQP::Graph RegAllocPBQP::constructPBQPProblem() {
-
- typedef std::vector<const LiveInterval*> LIVector;
- typedef std::vector<unsigned> RegVector;
-
- // This will store the physical intervals for easy reference.
- LIVector physIntervals;
-
- // Start by clearing the old node <-> live interval mappings & allowed sets
- li2Node.clear();
- node2LI.clear();
- allowedSets.clear();
-
- // Populate physIntervals, update preg use:
- for (LiveIntervals::iterator itr = lis->begin(), end = lis->end();
- itr != end; ++itr) {
-
- if (TargetRegisterInfo::isPhysicalRegister(itr->first)) {
- physIntervals.push_back(itr->second);
- mri->setPhysRegUsed(itr->second->reg);
- }
- }
-
- // Iterate over vreg intervals, construct live interval <-> node number
- // mappings.
- for (RegSet::const_iterator itr = vregsToAlloc.begin(),
- end = vregsToAlloc.end();
- itr != end; ++itr) {
- const LiveInterval *li = &lis->getInterval(*itr);
-
- li2Node[li] = node2LI.size();
- node2LI.push_back(li);
- }
-
- // Get the set of potential coalesces.
- CoalesceMap coalesces;
-
- if (pbqpCoalescing) {
- coalesces = findCoalesces();
- }
-
- // Construct a PBQP solver for this problem
- PBQP::Graph problem;
- problemNodes.resize(vregsToAlloc.size());
+static bool isACalleeSavedRegister(unsigned reg, const TargetRegisterInfo &TRI,
+ const MachineFunction &MF) {
+ const MCPhysReg *CSR = TRI.getCalleeSavedRegs(&MF);
+ for (unsigned i = 0; CSR[i] != 0; ++i)
+ if (TRI.regsOverlap(reg, CSR[i]))
+ return true;
+ return false;
+}
- // Resize allowedSets container appropriately.
- allowedSets.resize(vregsToAlloc.size());
+void RegAllocPBQP::initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM,
+ Spiller &VRegSpiller) {
+ MachineFunction &MF = G.getMetadata().MF;
- BitVector ReservedRegs = tri->getReservedRegs(*mf);
+ LiveIntervals &LIS = G.getMetadata().LIS;
+ const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
+ const TargetRegisterInfo &TRI =
+ *G.getMetadata().MF.getSubtarget().getRegisterInfo();
- // Iterate over virtual register intervals to compute allowed sets...
- for (unsigned node = 0; node < node2LI.size(); ++node) {
+ std::vector<unsigned> Worklist(VRegsToAlloc.begin(), VRegsToAlloc.end());
- // Grab pointers to the interval and its register class.
- const LiveInterval *li = node2LI[node];
- const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
+ while (!Worklist.empty()) {
+ unsigned VReg = Worklist.back();
+ Worklist.pop_back();
- // Start by assuming all allocable registers in the class are allowed...
- RegVector liAllowed;
- TargetRegisterClass::iterator aob = liRC->allocation_order_begin(*mf);
- TargetRegisterClass::iterator aoe = liRC->allocation_order_end(*mf);
- for (TargetRegisterClass::iterator it = aob; it != aoe; ++it)
- if (!ReservedRegs.test(*it))
- liAllowed.push_back(*it);
+ const TargetRegisterClass *TRC = MRI.getRegClass(VReg);
+ LiveInterval &VRegLI = LIS.getInterval(VReg);
- // Eliminate the physical registers which overlap with this range, along
- // with all their aliases.
- for (LIVector::iterator pItr = physIntervals.begin(),
- pEnd = physIntervals.end(); pItr != pEnd; ++pItr) {
+ // Record any overlaps with regmask operands.
+ BitVector RegMaskOverlaps;
+ LIS.checkRegMaskInterference(VRegLI, RegMaskOverlaps);
- if (!li->overlaps(**pItr))
+ // Compute an initial allowed set for the current vreg.
+ std::vector<unsigned> VRegAllowed;
+ ArrayRef<MCPhysReg> RawPRegOrder = TRC->getRawAllocationOrder(MF);
+ for (unsigned I = 0; I != RawPRegOrder.size(); ++I) {
+ unsigned PReg = RawPRegOrder[I];
+ if (MRI.isReserved(PReg))
continue;
- unsigned pReg = (*pItr)->reg;
-
- // If we get here then the live intervals overlap, but we're still ok
- // if they're coalescable.
- if (coalesces.find(RegPair(li->reg, pReg)) != coalesces.end()) {
- DEBUG(dbgs() << "CoalescingOverride: (" << li->reg << ", " << pReg << ")\n");
+ // vregLI crosses a regmask operand that clobbers preg.
+ if (!RegMaskOverlaps.empty() && !RegMaskOverlaps.test(PReg))
continue;
- }
-
- // If we get here then we have a genuine exclusion.
-
- // Remove the overlapping reg...
- RegVector::iterator eraseItr =
- std::find(liAllowed.begin(), liAllowed.end(), pReg);
-
- if (eraseItr != liAllowed.end())
- liAllowed.erase(eraseItr);
- const unsigned *aliasItr = tri->getAliasSet(pReg);
-
- if (aliasItr != 0) {
- // ...and its aliases.
- for (; *aliasItr != 0; ++aliasItr) {
- RegVector::iterator eraseItr =
- std::find(liAllowed.begin(), liAllowed.end(), *aliasItr);
-
- if (eraseItr != liAllowed.end()) {
- liAllowed.erase(eraseItr);
- }
+ // vregLI overlaps fixed regunit interference.
+ bool Interference = false;
+ for (MCRegUnitIterator Units(PReg, &TRI); Units.isValid(); ++Units) {
+ if (VRegLI.overlaps(LIS.getRegUnit(*Units))) {
+ Interference = true;
+ break;
}
}
- }
-
- // Copy the allowed set into a member vector for use when constructing cost
- // vectors & matrices, and mapping PBQP solutions back to assignments.
- allowedSets[node] = AllowedSet(liAllowed.begin(), liAllowed.end());
-
- // Set the spill cost to the interval weight, or epsilon if the
- // interval weight is zero
- PBQP::PBQPNum spillCost = (li->weight != 0.0) ?
- li->weight : std::numeric_limits<PBQP::PBQPNum>::min();
-
- // Build a cost vector for this interval.
- problemNodes[node] =
- problem.addNode(
- buildCostVector(li->reg, allowedSets[node], coalesces, spillCost));
-
- }
-
-
- // Now add the cost matrices...
- for (unsigned node1 = 0; node1 < node2LI.size(); ++node1) {
- const LiveInterval *li = node2LI[node1];
-
- // Test for live range overlaps and insert interference matrices.
- for (unsigned node2 = node1 + 1; node2 < node2LI.size(); ++node2) {
- const LiveInterval *li2 = node2LI[node2];
-
- CoalesceMap::const_iterator cmItr =
- coalesces.find(RegPair(li->reg, li2->reg));
-
- PBQP::Matrix *m = 0;
-
- if (cmItr != coalesces.end()) {
- m = buildCoalescingMatrix(allowedSets[node1], allowedSets[node2],
- cmItr->second);
- }
- else if (li->overlaps(*li2)) {
- m = buildInterferenceMatrix(allowedSets[node1], allowedSets[node2]);
- }
-
- if (m != 0) {
- problem.addEdge(problemNodes[node1],
- problemNodes[node2],
- *m);
+ if (Interference)
+ continue;
- delete m;
- }
+ // preg is usable for this virtual register.
+ VRegAllowed.push_back(PReg);
}
- }
- assert(problem.getNumNodes() == allowedSets.size());
-/*
- std::cerr << "Allocating for " << problem.getNumNodes() << " nodes, "
- << problem.getNumEdges() << " edges.\n";
-
- problem.printDot(std::cerr);
-*/
- // We're done, PBQP problem constructed - return it.
- return problem;
-}
-
-void RegAllocPBQP::addStackInterval(const LiveInterval *spilled,
- MachineRegisterInfo* mri) {
- int stackSlot = vrm->getStackSlot(spilled->reg);
-
- if (stackSlot == VirtRegMap::NO_STACK_SLOT)
- return;
+ // Check for vregs that have no allowed registers. These should be
+ // pre-spilled and the new vregs added to the worklist.
+ if (VRegAllowed.empty()) {
+ SmallVector<unsigned, 8> NewVRegs;
+ spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
+ for (auto NewVReg : NewVRegs)
+ Worklist.push_back(NewVReg);
+ continue;
+ }
- const TargetRegisterClass *RC = mri->getRegClass(spilled->reg);
- LiveInterval &stackInterval = lss->getOrCreateInterval(stackSlot, RC);
+ PBQPRAGraph::RawVector NodeCosts(VRegAllowed.size() + 1, 0);
- VNInfo *vni;
- if (stackInterval.getNumValNums() != 0)
- vni = stackInterval.getValNumInfo(0);
- else
- vni = stackInterval.getNextValue(
- SlotIndex(), 0, false, lss->getVNInfoAllocator());
+ // Tweak cost of callee saved registers, as using then force spilling and
+ // restoring them. This would only happen in the prologue / epilogue though.
+ for (unsigned i = 0; i != VRegAllowed.size(); ++i)
+ if (isACalleeSavedRegister(VRegAllowed[i], TRI, MF))
+ NodeCosts[1 + i] += 1.0;
- LiveInterval &rhsInterval = lis->getInterval(spilled->reg);
- stackInterval.MergeRangesInAsValue(rhsInterval, vni);
+ PBQPRAGraph::NodeId NId = G.addNode(std::move(NodeCosts));
+ G.getNodeMetadata(NId).setVReg(VReg);
+ G.getNodeMetadata(NId).setAllowedRegs(
+ G.getMetadata().getAllowedRegs(std::move(VRegAllowed)));
+ G.getMetadata().setNodeIdForVReg(VReg, NId);
+ }
}
-bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQP::Solution &solution) {
-
- // Set to true if we have any spills
- bool anotherRoundNeeded = false;
-
- // Clear the existing allocation.
- vrm->clearAllVirt();
-
- // Iterate over the nodes mapping the PBQP solution to a register assignment.
- for (unsigned node = 0; node < node2LI.size(); ++node) {
- unsigned virtReg = node2LI[node]->reg,
- allocSelection = solution.getSelection(problemNodes[node]);
-
-
- // If the PBQP solution is non-zero it's a physical register...
- if (allocSelection != 0) {
- // Get the physical reg, subtracting 1 to account for the spill option.
- unsigned physReg = allowedSets[node][allocSelection - 1];
-
- DEBUG(dbgs() << "VREG " << virtReg << " -> "
- << tri->getName(physReg) << " (Option: " << allocSelection << ")\n");
-
- assert(physReg != 0);
-
- // Add to the virt reg map and update the used phys regs.
- vrm->assignVirt2Phys(virtReg, physReg);
- }
- // ...Otherwise it's a spill.
- else {
-
- // Make sure we ignore this virtual reg on the next round
- // of allocation
- vregsToAlloc.erase(virtReg);
-
- // Insert spill ranges for this live range
- const LiveInterval *spillInterval = node2LI[node];
- double oldSpillWeight = spillInterval->weight;
- SmallVector<LiveInterval*, 8> spillIs;
- rmf->rememberUseDefs(spillInterval);
- std::vector<LiveInterval*> newSpills =
- lis->addIntervalsForSpills(*spillInterval, spillIs, loopInfo, *vrm);
- addStackInterval(spillInterval, mri);
- rmf->rememberSpills(spillInterval, newSpills);
-
- (void) oldSpillWeight;
- DEBUG(dbgs() << "VREG " << virtReg << " -> SPILLED (Option: 0, Cost: "
- << oldSpillWeight << ", New vregs: ");
-
- // Copy any newly inserted live intervals into the list of regs to
- // allocate.
- for (std::vector<LiveInterval*>::const_iterator
- itr = newSpills.begin(), end = newSpills.end();
- itr != end; ++itr) {
-
- assert(!(*itr)->empty() && "Empty spill range.");
-
- DEBUG(dbgs() << (*itr)->reg << " ");
-
- vregsToAlloc.insert((*itr)->reg);
- }
-
- DEBUG(dbgs() << ")\n");
-
- // We need another round if spill intervals were added.
- anotherRoundNeeded |= !newSpills.empty();
- }
+void RegAllocPBQP::spillVReg(unsigned VReg,
+ SmallVectorImpl<unsigned> &NewIntervals,
+ MachineFunction &MF, LiveIntervals &LIS,
+ VirtRegMap &VRM, Spiller &VRegSpiller) {
+
+ VRegsToAlloc.erase(VReg);
+ LiveRangeEdit LRE(&LIS.getInterval(VReg), NewIntervals, MF, LIS, &VRM);
+ VRegSpiller.spill(LRE);
+
+ const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
+ (void)TRI;
+ DEBUG(dbgs() << "VREG " << PrintReg(VReg, &TRI) << " -> SPILLED (Cost: "
+ << LRE.getParent().weight << ", New vregs: ");
+
+ // Copy any newly inserted live intervals into the list of regs to
+ // allocate.
+ for (LiveRangeEdit::iterator I = LRE.begin(), E = LRE.end();
+ I != E; ++I) {
+ const LiveInterval &LI = LIS.getInterval(*I);
+ assert(!LI.empty() && "Empty spill range.");
+ DEBUG(dbgs() << PrintReg(LI.reg, &TRI) << " ");
+ VRegsToAlloc.insert(LI.reg);
}
- return !anotherRoundNeeded;
+ DEBUG(dbgs() << ")\n");
}
-bool RegAllocPBQP::mapPBQPToRegAlloc2(const PBQPRAProblem &problem,
- const PBQP::Solution &solution) {
+bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAGraph &G,
+ const PBQP::Solution &Solution,
+ VirtRegMap &VRM,
+ Spiller &VRegSpiller) {
+ MachineFunction &MF = G.getMetadata().MF;
+ LiveIntervals &LIS = G.getMetadata().LIS;
+ const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
+ (void)TRI;
+
// Set to true if we have any spills
- bool anotherRoundNeeded = false;
+ bool AnotherRoundNeeded = false;
// Clear the existing allocation.
- vrm->clearAllVirt();
+ VRM.clearAllVirt();
- const PBQP::Graph &g = problem.getGraph();
// Iterate over the nodes mapping the PBQP solution to a register
// assignment.
- for (PBQP::Graph::ConstNodeItr node = g.nodesBegin(),
- nodeEnd = g.nodesEnd();
- node != nodeEnd; ++node) {
- unsigned vreg = problem.getVRegForNode(node);
- unsigned alloc = solution.getSelection(node);
-
- if (problem.isPRegOption(vreg, alloc)) {
- unsigned preg = problem.getPRegForOption(vreg, alloc);
- DEBUG(dbgs() << "VREG " << vreg << " -> " << tri->getName(preg) << "\n");
- assert(preg != 0 && "Invalid preg selected.");
- vrm->assignVirt2Phys(vreg, preg);
- } else if (problem.isSpillOption(vreg, alloc)) {
- vregsToAlloc.erase(vreg);
- const LiveInterval* spillInterval = &lis->getInterval(vreg);
- double oldWeight = spillInterval->weight;
- SmallVector<LiveInterval*, 8> spillIs;
- rmf->rememberUseDefs(spillInterval);
- std::vector<LiveInterval*> newSpills =
- lis->addIntervalsForSpills(*spillInterval, spillIs, loopInfo, *vrm);
- addStackInterval(spillInterval, mri);
- rmf->rememberSpills(spillInterval, newSpills);
-
- (void) oldWeight;
- DEBUG(dbgs() << "VREG " << vreg << " -> SPILLED (Cost: "
- << oldWeight << ", New vregs: ");
-
- // Copy any newly inserted live intervals into the list of regs to
- // allocate.
- for (std::vector<LiveInterval*>::const_iterator
- itr = newSpills.begin(), end = newSpills.end();
- itr != end; ++itr) {
- assert(!(*itr)->empty() && "Empty spill range.");
- DEBUG(dbgs() << (*itr)->reg << " ");
- vregsToAlloc.insert((*itr)->reg);
- }
-
- DEBUG(dbgs() << ")\n");
-
- // We need another round if spill intervals were added.
- anotherRoundNeeded |= !newSpills.empty();
+ for (auto NId : G.nodeIds()) {
+ unsigned VReg = G.getNodeMetadata(NId).getVReg();
+ unsigned AllocOption = Solution.getSelection(NId);
+
+ if (AllocOption != PBQP::RegAlloc::getSpillOptionIdx()) {
+ unsigned PReg = G.getNodeMetadata(NId).getAllowedRegs()[AllocOption - 1];
+ DEBUG(dbgs() << "VREG " << PrintReg(VReg, &TRI) << " -> "
+ << TRI.getName(PReg) << "\n");
+ assert(PReg != 0 && "Invalid preg selected.");
+ VRM.assignVirt2Phys(VReg, PReg);
} else {
- assert(false && "Unknown allocation option.");
+ // Spill VReg. If this introduces new intervals we'll need another round
+ // of allocation.
+ SmallVector<unsigned, 8> NewVRegs;
+ spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
+ AnotherRoundNeeded |= !NewVRegs.empty();
}
}
- return !anotherRoundNeeded;
+ return !AnotherRoundNeeded;
}
-
-void RegAllocPBQP::finalizeAlloc() const {
- typedef LiveIntervals::iterator LIIterator;
- typedef LiveInterval::Ranges::const_iterator LRIterator;
+void RegAllocPBQP::finalizeAlloc(MachineFunction &MF,
+ LiveIntervals &LIS,
+ VirtRegMap &VRM) const {
+ MachineRegisterInfo &MRI = MF.getRegInfo();
// First allocate registers for the empty intervals.
for (RegSet::const_iterator
- itr = emptyIntervalVRegs.begin(), end = emptyIntervalVRegs.end();
- itr != end; ++itr) {
- LiveInterval *li = &lis->getInterval(*itr);
+ I = EmptyIntervalVRegs.begin(), E = EmptyIntervalVRegs.end();
+ I != E; ++I) {
+ LiveInterval &LI = LIS.getInterval(*I);
- unsigned physReg = vrm->getRegAllocPref(li->reg);
+ unsigned PReg = MRI.getSimpleHint(LI.reg);
- if (physReg == 0) {
- const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
- physReg = *liRC->allocation_order_begin(*mf);
+ if (PReg == 0) {
+ const TargetRegisterClass &RC = *MRI.getRegClass(LI.reg);
+ PReg = RC.getRawAllocationOrder(MF).front();
}
- vrm->assignVirt2Phys(li->reg, physReg);
- }
-
- // Finally iterate over the basic blocks to compute and set the live-in sets.
- SmallVector<MachineBasicBlock*, 8> liveInMBBs;
- MachineBasicBlock *entryMBB = &*mf->begin();
-
- for (LIIterator liItr = lis->begin(), liEnd = lis->end();
- liItr != liEnd; ++liItr) {
-
- const LiveInterval *li = liItr->second;
- unsigned reg = 0;
-
- // Get the physical register for this interval
- if (TargetRegisterInfo::isPhysicalRegister(li->reg)) {
- reg = li->reg;
- }
- else if (vrm->isAssignedReg(li->reg)) {
- reg = vrm->getPhys(li->reg);
- }
- else {
- // Ranges which are assigned a stack slot only are ignored.
- continue;
- }
-
- if (reg == 0) {
- // Filter out zero regs - they're for intervals that were spilled.
- continue;
- }
-
- // Iterate over the ranges of the current interval...
- for (LRIterator lrItr = li->begin(), lrEnd = li->end();
- lrItr != lrEnd; ++lrItr) {
-
- // Find the set of basic blocks which this range is live into...
- if (lis->findLiveInMBBs(lrItr->start, lrItr->end, liveInMBBs)) {
- // And add the physreg for this interval to their live-in sets.
- for (unsigned i = 0; i < liveInMBBs.size(); ++i) {
- if (liveInMBBs[i] != entryMBB) {
- if (!liveInMBBs[i]->isLiveIn(reg)) {
- liveInMBBs[i]->addLiveIn(reg);
- }
- }
- }
- liveInMBBs.clear();
- }
- }
+ VRM.assignVirt2Phys(LI.reg, PReg);
}
+}
+static inline float normalizePBQPSpillWeight(float UseDefFreq, unsigned Size,
+ unsigned NumInstr) {
+ // All intervals have a spill weight that is mostly proportional to the number
+ // of uses, with uses in loops having a bigger weight.
+ return NumInstr * normalizeSpillWeight(UseDefFreq, Size, 1);
}
bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
+ LiveIntervals &LIS = getAnalysis<LiveIntervals>();
+ MachineBlockFrequencyInfo &MBFI =
+ getAnalysis<MachineBlockFrequencyInfo>();
- mf = &MF;
- tm = &mf->getTarget();
- tri = tm->getRegisterInfo();
- tii = tm->getInstrInfo();
- mri = &mf->getRegInfo();
+ calculateSpillWeightsAndHints(LIS, MF, getAnalysis<MachineLoopInfo>(), MBFI,
+ normalizePBQPSpillWeight);
- lis = &getAnalysis<LiveIntervals>();
- lss = &getAnalysis<LiveStacks>();
- loopInfo = &getAnalysis<MachineLoopInfo>();
- rmf = &getAnalysis<RenderMachineFunction>();
+ VirtRegMap &VRM = getAnalysis<VirtRegMap>();
- vrm = &getAnalysis<VirtRegMap>();
+ std::unique_ptr<Spiller> VRegSpiller(createInlineSpiller(*this, MF, VRM));
+ MF.getRegInfo().freezeReservedRegs(MF);
- DEBUG(dbgs() << "PBQP Register Allocating for " << mf->getFunction()->getName() << "\n");
+ DEBUG(dbgs() << "PBQP Register Allocating for " << MF.getName() << "\n");
// Allocator main loop:
//
// This process is continued till no more spills are generated.
// Find the vreg intervals in need of allocation.
- findVRegIntervalsToAlloc();
-
- // If there are non-empty intervals allocate them using pbqp.
- if (!vregsToAlloc.empty()) {
+ findVRegIntervalsToAlloc(MF, LIS);
- bool pbqpAllocComplete = false;
- unsigned round = 0;
+#ifndef NDEBUG
+ const Function &F = *MF.getFunction();
+ std::string FullyQualifiedName =
+ F.getParent()->getModuleIdentifier() + "." + F.getName().str();
+#endif
- if (!pbqpBuilder) {
- while (!pbqpAllocComplete) {
- DEBUG(dbgs() << " PBQP Regalloc round " << round << ":\n");
-
- PBQP::Graph problem = constructPBQPProblem();
- PBQP::Solution solution =
- PBQP::HeuristicSolver<PBQP::Heuristics::Briggs>::solve(problem);
-
- pbqpAllocComplete = mapPBQPToRegAlloc(solution);
-
- ++round;
+ // If there are non-empty intervals allocate them using pbqp.
+ if (!VRegsToAlloc.empty()) {
+
+ const TargetSubtargetInfo &Subtarget = MF.getSubtarget();
+ std::unique_ptr<PBQPRAConstraintList> ConstraintsRoot =
+ llvm::make_unique<PBQPRAConstraintList>();
+ ConstraintsRoot->addConstraint(llvm::make_unique<SpillCosts>());
+ ConstraintsRoot->addConstraint(llvm::make_unique<Interference>());
+ if (PBQPCoalescing)
+ ConstraintsRoot->addConstraint(llvm::make_unique<Coalescing>());
+ ConstraintsRoot->addConstraint(Subtarget.getCustomPBQPConstraints());
+
+ bool PBQPAllocComplete = false;
+ unsigned Round = 0;
+
+ while (!PBQPAllocComplete) {
+ DEBUG(dbgs() << " PBQP Regalloc round " << Round << ":\n");
+
+ PBQPRAGraph G(PBQPRAGraph::GraphMetadata(MF, LIS, MBFI));
+ initializeGraph(G, VRM, *VRegSpiller);
+ ConstraintsRoot->apply(G);
+
+#ifndef NDEBUG
+ if (PBQPDumpGraphs) {
+ std::ostringstream RS;
+ RS << Round;
+ std::string GraphFileName = FullyQualifiedName + "." + RS.str() +
+ ".pbqpgraph";
+ std::error_code EC;
+ raw_fd_ostream OS(GraphFileName, EC, sys::fs::F_Text);
+ DEBUG(dbgs() << "Dumping graph for round " << Round << " to \""
+ << GraphFileName << "\"\n");
+ G.dump(OS);
}
- } else {
- while (!pbqpAllocComplete) {
- DEBUG(dbgs() << " PBQP Regalloc round " << round << ":\n");
-
- std::auto_ptr<PBQPRAProblem> problem =
- builder->build(mf, lis, vregsToAlloc);
- PBQP::Solution solution =
- HeuristicSolver<Briggs>::solve(problem->getGraph());
-
- pbqpAllocComplete = mapPBQPToRegAlloc2(*problem, solution);
+#endif
- ++round;
- }
+ PBQP::Solution Solution = PBQP::RegAlloc::solve(G);
+ PBQPAllocComplete = mapPBQPToRegAlloc(G, Solution, VRM, *VRegSpiller);
+ ++Round;
}
}
// Finalise allocation, allocate empty ranges.
- finalizeAlloc();
+ finalizeAlloc(MF, LIS, VRM);
+ VRegsToAlloc.clear();
+ EmptyIntervalVRegs.clear();
- rmf->renderMachineFunction("After PBQP register allocation.", vrm);
+ DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << VRM << "\n");
- vregsToAlloc.clear();
- emptyIntervalVRegs.clear();
- li2Node.clear();
- node2LI.clear();
- allowedSets.clear();
- problemNodes.clear();
+ return true;
+}
- DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << *vrm << "\n");
+namespace {
+// A helper class for printing node and register info in a consistent way
+class PrintNodeInfo {
+public:
+ typedef PBQP::RegAlloc::PBQPRAGraph Graph;
+ typedef PBQP::RegAlloc::PBQPRAGraph::NodeId NodeId;
+
+ PrintNodeInfo(NodeId NId, const Graph &G) : G(G), NId(NId) {}
+
+ void print(raw_ostream &OS) const {
+ const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
+ const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo();
+ unsigned VReg = G.getNodeMetadata(NId).getVReg();
+ const char *RegClassName = TRI->getRegClassName(MRI.getRegClass(VReg));
+ OS << NId << " (" << RegClassName << ':' << PrintReg(VReg, TRI) << ')';
+ }
- // Run rewriter
- std::auto_ptr<VirtRegRewriter> rewriter(createVirtRegRewriter());
+private:
+ const Graph &G;
+ NodeId NId;
+};
- rewriter->runOnMachineFunction(*mf, *vrm, lis);
+inline raw_ostream &operator<<(raw_ostream &OS, const PrintNodeInfo &PR) {
+ PR.print(OS);
+ return OS;
+}
+} // anonymous namespace
- return true;
+void PBQP::RegAlloc::PBQPRAGraph::dump(raw_ostream &OS) const {
+ for (auto NId : nodeIds()) {
+ const Vector &Costs = getNodeCosts(NId);
+ assert(Costs.getLength() != 0 && "Empty vector in graph.");
+ OS << PrintNodeInfo(NId, *this) << ": " << Costs << '\n';
+ }
+ OS << '\n';
+
+ for (auto EId : edgeIds()) {
+ NodeId N1Id = getEdgeNode1Id(EId);
+ NodeId N2Id = getEdgeNode2Id(EId);
+ assert(N1Id != N2Id && "PBQP graphs should not have self-edges.");
+ const Matrix &M = getEdgeCosts(EId);
+ assert(M.getRows() != 0 && "No rows in matrix.");
+ assert(M.getCols() != 0 && "No cols in matrix.");
+ OS << PrintNodeInfo(N1Id, *this) << ' ' << M.getRows() << " rows / ";
+ OS << PrintNodeInfo(N2Id, *this) << ' ' << M.getCols() << " cols:\n";
+ OS << M << '\n';
+ }
+}
+
+void PBQP::RegAlloc::PBQPRAGraph::dump() const { dump(dbgs()); }
+
+void PBQP::RegAlloc::PBQPRAGraph::printDot(raw_ostream &OS) const {
+ OS << "graph {\n";
+ for (auto NId : nodeIds()) {
+ OS << " node" << NId << " [ label=\""
+ << PrintNodeInfo(NId, *this) << "\\n"
+ << getNodeCosts(NId) << "\" ]\n";
+ }
+
+ OS << " edge [ len=" << nodeIds().size() << " ]\n";
+ for (auto EId : edgeIds()) {
+ OS << " node" << getEdgeNode1Id(EId)
+ << " -- node" << getEdgeNode2Id(EId)
+ << " [ label=\"";
+ const Matrix &EdgeCosts = getEdgeCosts(EId);
+ for (unsigned i = 0; i < EdgeCosts.getRows(); ++i) {
+ OS << EdgeCosts.getRowAsVector(i) << "\\n";
+ }
+ OS << "\" ]\n";
+ }
+ OS << "}\n";
}
-FunctionPass* createPBQPRegisterAllocator() {
- return new RegAllocPBQP(std::auto_ptr<PBQPBuilder>(new PBQPBuilder()));
+FunctionPass *llvm::createPBQPRegisterAllocator(char *customPassID) {
+ return new RegAllocPBQP(customPassID);
}
+FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
+ return createPBQPRegisterAllocator();
}
#undef DEBUG_TYPE