A Partitioned Boolean Quadratic Programming (PBQP) based register allocator.
Contributed by Lang Hames.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56959 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/RegAllocPBQP.cpp b/lib/CodeGen/RegAllocPBQP.cpp
new file mode 100644
index 0000000..ee82f63
--- /dev/null
+++ b/lib/CodeGen/RegAllocPBQP.cpp
@@ -0,0 +1,529 @@
+//===------ RegAllocPBQP.cpp ---- PBQP Register Allocator -------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains a Partitioned Boolean Quadratic Programming (PBQP) based
+// register allocator for LLVM. This allocator works by constructing a PBQP
+// problem representing the register allocation problem under consideration,
+// solving this using a PBQP solver, and mapping the solution back to a
+// register assignment. If any variables are selected for spilling then spill
+// code is inserted and the process repeated.
+//
+// The PBQP solver (pbqp.c) provided for this allocator uses a heuristic tuned
+// for register allocation. For more information on PBQP for register
+// allocation see the following papers:
+//
+// (1) Hames, L. and Scholz, B. 2006. Nearly optimal register allocation with
+// PBQP. In Proceedings of the 7th Joint Modular Languages Conference
+// (JMLC'06). LNCS, vol. 4228. Springer, New York, NY, USA. 346-361.
+//
+// (2) Scholz, B., Eckstein, E. 2002. Register allocation for irregular
+// architectures. In Proceedings of the Joint Conference on Languages,
+// Compilers and Tools for Embedded Systems (LCTES'02), ACM Press, New York,
+// NY, USA, 139-148.
+//
+// Author: Lang Hames
+// Email: lhames@gmail.com
+//
+//===----------------------------------------------------------------------===//
+
+// TODO:
+//
+// * Use of std::set in constructPBQPProblem destroys allocation order preference.
+// Switch to an order preserving container.
+//
+// * Coalescing support.
+
+#define DEBUG_TYPE "regalloc"
+
+#include "PBQP.h"
+#include "VirtRegMap.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/RegAllocRegistry.h"
+#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Support/Debug.h"
+#include <memory>
+#include <map>
+#include <set>
+#include <vector>
+#include <limits>
+
+using namespace llvm;
+
+static RegisterRegAlloc
+registerPBQPRepAlloc("pbqp", " PBQP register allocator",
+ createPBQPRegisterAllocator);
+
+
+namespace {
+
+ //!
+ //! PBQP based allocators solve the register allocation problem by mapping
+ //! register allocation problems to Partitioned Boolean Quadratic
+ //! Programming problems.
+ class VISIBILITY_HIDDEN PBQPRegAlloc : public MachineFunctionPass {
+ public:
+
+ static char ID;
+
+ //! Construct a PBQP register allocator.
+ PBQPRegAlloc() : MachineFunctionPass((intptr_t)&ID) {}
+
+ //! Return the pass name.
+ virtual const char* getPassName() const throw() {
+ return "PBQP Register Allocator";
+ }
+
+ //! PBQP analysis usage.
+ virtual void getAnalysisUsage(AnalysisUsage &au) const {
+ au.addRequired<LiveIntervals>();
+ au.addRequired<MachineLoopInfo>();
+ MachineFunctionPass::getAnalysisUsage(au);
+ }
+
+ //! Perform register allocation
+ virtual bool runOnMachineFunction(MachineFunction &MF);
+
+ private:
+ 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::set<unsigned> IgnoreSet;
+
+ MachineFunction *mf;
+ const TargetMachine *tm;
+ const TargetRegisterInfo *tri;
+ const TargetInstrInfo *tii;
+ const MachineLoopInfo *loopInfo;
+ MachineRegisterInfo *mri;
+
+ LiveIntervals *li;
+ VirtRegMap *vrm;
+
+ LI2NodeMap li2Node;
+ Node2LIMap node2LI;
+ AllowedSetMap allowedSets;
+ IgnoreSet ignoreSet;
+
+ //! Builds a PBQP cost vector.
+ template <typename Container>
+ PBQPVector* buildCostVector(const Container &allowed,
+ PBQPNum spillCost) const;
+
+ //! \brief Builds a PBQP interfernce matrix.
+ //!
+ //! @return Either a pointer to a non-zero PBQP matrix representing the
+ //! allocation option costs, or a null pointer for a zero matrix.
+ //!
+ //! Expects allowed sets for two interfering LiveIntervals. These allowed
+ //! sets should contain only allocable registers from the LiveInterval's
+ //! register class, with any interfering pre-colored registers removed.
+ template <typename Container>
+ PBQPMatrix* buildInterferenceMatrix(const Container &allowed1,
+ const Container &allowed2) const;
+
+ //!
+ //! Expects allowed sets for two potentially coalescable LiveIntervals,
+ //! and an estimated benefit due to coalescing. The allowed sets should
+ //! contain only allocable registers from the LiveInterval's register
+ //! classes, with any interfering pre-colored registers removed.
+ template <typename Container>
+ PBQPMatrix* buildCoalescingMatrix(const Container &allowed1,
+ const Container &allowed2,
+ PBQPNum cBenefit) const;
+
+ //! \brief Helper functior for constructInitialPBQPProblem().
+ //!
+ //! This function iterates over the Function we are about to allocate for
+ //! and computes spill costs.
+ void calcSpillCosts();
+
+ //! \brief Scans the MachineFunction being allocated to find coalescing
+ // opportunities.
+ void findCoalescingOpportunities();
+
+ //! \brief Constructs a PBQP problem representation of the register
+ //! allocation problem for this function.
+ //!
+ //! @return a PBQP solver object for the register allocation problem.
+ pbqp* constructPBQPProblem();
+
+ //! \brief Given a solved PBQP problem maps this solution back to a register
+ //! assignment.
+ bool mapPBQPToRegAlloc(pbqp *problem);
+
+ };
+
+ char PBQPRegAlloc::ID = 0;
+}
+
+
+template <typename Container>
+PBQPVector* PBQPRegAlloc::buildCostVector(const Container &allowed,
+ PBQPNum spillCost) const {
+
+ // Allocate vector. Additional element (0th) used for spill option
+ PBQPVector *v = new PBQPVector(allowed.size() + 1);
+
+ (*v)[0] = spillCost;
+
+ return v;
+}
+
+template <typename Container>
+PBQPMatrix* PBQPRegAlloc::buildInterferenceMatrix(
+ const Container &allowed1, const Container &allowed2) const {
+
+ typedef typename Container::const_iterator ContainerIterator;
+
+ // 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).
+ PBQPMatrix *m = new PBQPMatrix(allowed1.size() + 1, allowed2.size() + 1);
+
+ // 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 (ContainerIterator 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 (ContainerIterator 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 ((reg1 == reg2) || tri->areAliases(reg1, reg2)) {
+ (*m)[ri][ci] = std::numeric_limits<PBQPNum>::infinity();
+ isZeroMatrix = false;
+ }
+
+ ++ci;
+ }
+
+ ++ri;
+ }
+
+ // If this turns out to be a zero matrix...
+ if (isZeroMatrix) {
+ // free it and return null.
+ delete m;
+ return 0;
+ }
+
+ // ...otherwise return the cost matrix.
+ return m;
+}
+
+void PBQPRegAlloc::calcSpillCosts() {
+
+ // Calculate the spill cost for each live interval by iterating over the
+ // function counting loads and stores, with loop depth taken into account.
+ for (MachineFunction::const_iterator bbItr = mf->begin(), bbEnd = mf->end();
+ bbItr != bbEnd; ++bbItr) {
+
+ const MachineBasicBlock *mbb = &*bbItr;
+ float loopDepth = loopInfo->getLoopDepth(mbb);
+
+ for (MachineBasicBlock::const_iterator
+ iItr = mbb->begin(), iEnd = mbb->end(); iItr != iEnd; ++iItr) {
+
+ const MachineInstr *instr = &*iItr;
+
+ for (unsigned opNo = 0; opNo < instr->getNumOperands(); ++opNo) {
+
+ const MachineOperand &mo = instr->getOperand(opNo);
+
+ // We're not interested in non-registers...
+ if (!mo.isRegister())
+ continue;
+
+ unsigned moReg = mo.getReg();
+
+ // ...Or invalid registers...
+ if (moReg == 0)
+ continue;
+
+ // ...Or physical registers...
+ if (TargetRegisterInfo::isPhysicalRegister(moReg))
+ continue;
+
+ assert ((mo.isUse() || mo.isDef()) &&
+ "Not a use, not a def, what is it?");
+
+ //... Just the virtual registers. We treat loads and stores as equal.
+ li->getInterval(moReg).weight += powf(10.0f, loopDepth);
+ }
+
+ }
+
+ }
+
+}
+
+pbqp* PBQPRegAlloc::constructPBQPProblem() {
+
+ typedef std::vector<const LiveInterval*> LIVector;
+ typedef std::set<unsigned> RegSet;
+
+ // These will store the physical & virtual intervals, respectively.
+ LIVector physIntervals, virtIntervals;
+
+ // Start by clearing the old node <-> live interval mappings & allowed sets
+ li2Node.clear();
+ node2LI.clear();
+ allowedSets.clear();
+
+ // Iterate over intervals classifying them as physical or virtual, and
+ // constructing live interval <-> node number mappings.
+ for (LiveIntervals::iterator itr = li->begin(), end = li->end();
+ itr != end; ++itr) {
+
+ if (itr->second->getNumValNums() != 0) {
+ DOUT << "Live range has " << itr->second->getNumValNums() << ": " << itr->second << "\n";
+ }
+
+ if (TargetRegisterInfo::isPhysicalRegister(itr->first)) {
+ physIntervals.push_back(itr->second);
+ mri->setPhysRegUsed(itr->second->reg);
+ }
+ else {
+
+ // If we've allocated this virtual register interval a stack slot on a
+ // previous round then it's not an allocation candidate
+ if (ignoreSet.find(itr->first) != ignoreSet.end())
+ continue;
+
+ li2Node[itr->second] = node2LI.size();
+ node2LI.push_back(itr->second);
+ virtIntervals.push_back(itr->second);
+ }
+ }
+
+ // Early out if there's no regs to allocate for.
+ if (virtIntervals.empty())
+ return 0;
+
+ // Construct a PBQP solver for this problem
+ pbqp *solver = alloc_pbqp(virtIntervals.size());
+
+ // Resize allowedSets container appropriately.
+ allowedSets.resize(virtIntervals.size());
+
+ // Iterate over virtual register intervals to compute allowed sets...
+ for (unsigned node = 0; node < node2LI.size(); ++node) {
+
+ // Grab pointers to the interval and its register class.
+ const LiveInterval *li = node2LI[node];
+ const TargetRegisterClass *liRC = mri->getRegClass(li->reg);
+
+ // Start by assuming all allocable registers in the class are allowed...
+ RegSet liAllowed(liRC->allocation_order_begin(*mf),
+ liRC->allocation_order_end(*mf));
+
+ // If this range is non-empty then eliminate the physical registers which
+ // overlap with this range, along with all their aliases.
+ if (!li->empty()) {
+ for (LIVector::iterator pItr = physIntervals.begin(),
+ pEnd = physIntervals.end(); pItr != pEnd; ++pItr) {
+
+ if (li->overlaps(**pItr)) {
+
+ unsigned pReg = (*pItr)->reg;
+
+ // Remove the overlapping reg...
+ liAllowed.erase(pReg);
+
+ const unsigned *aliasItr = tri->getAliasSet(pReg);
+
+ if (aliasItr != 0) {
+ // ...and its aliases.
+ for (; *aliasItr != 0; ++aliasItr) {
+ liAllowed.erase(*aliasItr);
+ }
+
+ }
+
+ }
+
+ }
+
+ }
+
+ // 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
+ PBQPNum spillCost = (li->weight != 0.0) ?
+ li->weight : std::numeric_limits<PBQPNum>::min();
+
+ // Build a cost vector for this interval.
+ add_pbqp_nodecosts(solver, node,
+ buildCostVector(allowedSets[node], spillCost));
+
+ }
+
+ // Now add the cost matrices...
+ for (unsigned node1 = 0; node1 < node2LI.size(); ++node1) {
+
+ const LiveInterval *li = node2LI[node1];
+
+ if (li->empty())
+ continue;
+
+ // Test for live range overlaps and insert interference matrices.
+ for (unsigned node2 = node1 + 1; node2 < node2LI.size(); ++node2) {
+ const LiveInterval *li2 = node2LI[node2];
+
+ if (li2->empty())
+ continue;
+
+ if (li->overlaps(*li2)) {
+ PBQPMatrix *m =
+ buildInterferenceMatrix(allowedSets[node1], allowedSets[node2]);
+
+ if (m != 0) {
+ add_pbqp_edgecosts(solver, node1, node2, m);
+ delete m;
+ }
+ }
+ }
+ }
+
+ // We're done, PBQP problem constructed - return it.
+ return solver;
+}
+
+bool PBQPRegAlloc::mapPBQPToRegAlloc(pbqp *problem) {
+
+ // 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 symReg = node2LI[node]->reg,
+ allocSelection = get_pbqp_solution(problem, 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];
+
+ // Add to the virt reg map and update the used phys regs.
+ vrm->assignVirt2Phys(symReg, physReg);
+ mri->setPhysRegUsed(physReg);
+ }
+ // ...Otherwise it's a spill.
+ else {
+
+ // Make sure we ignore this virtual reg on the next round
+ // of allocation
+ ignoreSet.insert(node2LI[node]->reg);
+
+ float SSWeight;
+
+ // Insert spill ranges for this live range
+ SmallVector<LiveInterval*, 8> spillIs;
+ std::vector<LiveInterval*> newSpills =
+ li->addIntervalsForSpills(*node2LI[node], spillIs, loopInfo, *vrm,
+ SSWeight);
+
+ // We need another round if spill intervals were added.
+ anotherRoundNeeded |= !newSpills.empty();
+ }
+ }
+
+ return !anotherRoundNeeded;
+}
+
+bool PBQPRegAlloc::runOnMachineFunction(MachineFunction &MF) {
+
+ mf = &MF;
+ tm = &mf->getTarget();
+ tri = tm->getRegisterInfo();
+ mri = &mf->getRegInfo();
+
+ li = &getAnalysis<LiveIntervals>();
+ loopInfo = &getAnalysis<MachineLoopInfo>();
+
+ std::auto_ptr<VirtRegMap> vrmAutoPtr(new VirtRegMap(*mf));
+ vrm = vrmAutoPtr.get();
+
+ // Allocator main loop:
+ //
+ // * Map current regalloc problem to a PBQP problem
+ // * Solve the PBQP problem
+ // * Map the solution back to a register allocation
+ // * Spill if necessary
+ //
+ // This process is continued till no more spills are generated.
+
+ bool regallocComplete = false;
+
+ // Calculate spill costs for intervals
+ calcSpillCosts();
+
+ while (!regallocComplete) {
+ pbqp *problem = constructPBQPProblem();
+
+ // Fast out if there's no problem to solve.
+ if (problem == 0)
+ return true;
+
+ solve_pbqp(problem);
+
+ regallocComplete = mapPBQPToRegAlloc(problem);
+
+ free_pbqp(problem);
+ }
+
+ ignoreSet.clear();
+
+ std::auto_ptr<Spiller> spiller(createSpiller());
+
+ spiller->runOnMachineFunction(*mf, *vrm);
+
+ return true;
+}
+
+FunctionPass* llvm::createPBQPRegisterAllocator() {
+ return new PBQPRegAlloc();
+}
+
+
+#undef DEBUG_TYPE