2008-10-03 02:29:27 +08:00
|
|
|
//===------ 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.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
2009-01-08 23:50:22 +08:00
|
|
|
//
|
2008-10-03 02:29:27 +08:00
|
|
|
// 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
|
2009-01-08 23:50:22 +08:00
|
|
|
// code is inserted and the process repeated.
|
2008-10-03 02:29:27 +08:00
|
|
|
//
|
|
|
|
// The PBQP solver (pbqp.c) provided for this allocator uses a heuristic tuned
|
|
|
|
// for register allocation. For more information on PBQP for register
|
2009-01-09 00:40:25 +08:00
|
|
|
// allocation, see the following papers:
|
2008-10-03 02:29:27 +08:00
|
|
|
//
|
|
|
|
// (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.
|
2009-01-08 23:50:22 +08:00
|
|
|
//
|
2008-10-03 02:29:27 +08:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-12-04 00:50:05 +08:00
|
|
|
#include "llvm/CodeGen/RegAllocPBQP.h"
|
2011-06-27 05:41:06 +08:00
|
|
|
#include "RegisterCoalescer.h"
|
2012-12-04 00:50:05 +08:00
|
|
|
#include "Spiller.h"
|
2011-12-06 09:45:57 +08:00
|
|
|
#include "llvm/Analysis/AliasAnalysis.h"
|
2009-12-14 14:49:42 +08:00
|
|
|
#include "llvm/CodeGen/CalcSpillWeights.h"
|
2008-10-03 02:29:27 +08:00
|
|
|
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
|
2012-04-03 06:44:18 +08:00
|
|
|
#include "llvm/CodeGen/LiveRangeEdit.h"
|
2008-11-16 20:12:54 +08:00
|
|
|
#include "llvm/CodeGen/LiveStackAnalysis.h"
|
2013-06-18 03:00:36 +08:00
|
|
|
#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
|
2011-12-06 09:45:57 +08:00
|
|
|
#include "llvm/CodeGen/MachineDominators.h"
|
2009-01-08 23:50:22 +08:00
|
|
|
#include "llvm/CodeGen/MachineFunctionPass.h"
|
2013-07-02 04:47:47 +08:00
|
|
|
#include "llvm/CodeGen/MachineLoopInfo.h"
|
2009-01-08 23:50:22 +08:00
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
|
|
|
#include "llvm/CodeGen/RegAllocRegistry.h"
|
2012-11-29 03:13:06 +08:00
|
|
|
#include "llvm/CodeGen/VirtRegMap.h"
|
2013-01-02 19:36:10 +08:00
|
|
|
#include "llvm/IR/Module.h"
|
2008-10-03 02:29:27 +08:00
|
|
|
#include "llvm/Support/Debug.h"
|
2014-04-30 07:26:49 +08:00
|
|
|
#include "llvm/Support/FileSystem.h"
|
2009-07-25 08:23:56 +08:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2009-01-08 23:50:22 +08:00
|
|
|
#include "llvm/Target/TargetInstrInfo.h"
|
2014-08-05 05:25:23 +08:00
|
|
|
#include "llvm/Target/TargetSubtargetInfo.h"
|
2009-01-08 23:50:22 +08:00
|
|
|
#include <limits>
|
|
|
|
#include <memory>
|
2014-10-19 01:26:07 +08:00
|
|
|
#include <queue>
|
2008-10-03 02:29:27 +08:00
|
|
|
#include <set>
|
2012-03-27 07:07:23 +08:00
|
|
|
#include <sstream>
|
2008-10-03 02:29:27 +08:00
|
|
|
#include <vector>
|
|
|
|
|
2010-09-23 12:28:54 +08:00
|
|
|
using namespace llvm;
|
2010-09-18 17:07:10 +08:00
|
|
|
|
2014-04-22 10:02:50 +08:00
|
|
|
#define DEBUG_TYPE "regalloc"
|
|
|
|
|
2008-10-03 02:29:27 +08:00
|
|
|
static RegisterRegAlloc
|
2014-10-10 02:20:51 +08:00
|
|
|
RegisterPBQPRepAlloc("pbqp", "PBQP register allocator",
|
2010-09-23 12:28:54 +08:00
|
|
|
createDefaultPBQPRegisterAllocator);
|
2008-10-03 02:29:27 +08:00
|
|
|
|
2009-08-19 09:36:14 +08:00
|
|
|
static cl::opt<bool>
|
2014-10-10 02:20:51 +08:00
|
|
|
PBQPCoalescing("pbqp-coalescing",
|
2010-01-26 12:49:58 +08:00
|
|
|
cl::desc("Attempt coalescing during PBQP register allocation."),
|
|
|
|
cl::init(false), cl::Hidden);
|
2009-08-19 09:36:14 +08:00
|
|
|
|
2012-03-27 07:07:23 +08:00
|
|
|
#ifndef NDEBUG
|
|
|
|
static cl::opt<bool>
|
2014-10-10 02:20:51 +08:00
|
|
|
PBQPDumpGraphs("pbqp-dump-graphs",
|
2012-03-27 07:07:23 +08:00
|
|
|
cl::desc("Dump graphs for each function/round in the compilation unit."),
|
|
|
|
cl::init(false), cl::Hidden);
|
|
|
|
#endif
|
|
|
|
|
2010-09-23 12:28:54 +08:00
|
|
|
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.
|
2014-10-10 02:20:51 +08:00
|
|
|
RegAllocPBQP(char *cPassID = nullptr)
|
|
|
|
: MachineFunctionPass(ID), customPassID(cPassID) {
|
2010-10-20 01:21:58 +08:00
|
|
|
initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
|
|
|
|
initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
|
|
|
|
initializeLiveStacksPass(*PassRegistry::getPassRegistry());
|
|
|
|
initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
|
|
|
|
}
|
2010-09-23 12:28:54 +08:00
|
|
|
|
|
|
|
/// Return the pass name.
|
2014-03-07 17:26:03 +08:00
|
|
|
const char* getPassName() const override {
|
2010-09-23 12:28:54 +08:00
|
|
|
return "PBQP Register Allocator";
|
|
|
|
}
|
|
|
|
|
|
|
|
/// PBQP analysis usage.
|
2014-03-07 17:26:03 +08:00
|
|
|
void getAnalysisUsage(AnalysisUsage &au) const override;
|
2010-09-23 12:28:54 +08:00
|
|
|
|
|
|
|
/// Perform register allocation
|
2014-03-07 17:26:03 +08:00
|
|
|
bool runOnMachineFunction(MachineFunction &MF) override;
|
2010-09-23 12:28:54 +08:00
|
|
|
|
|
|
|
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::pair<unsigned, unsigned> RegPair;
|
|
|
|
typedef std::map<RegPair, PBQP::PBQPNum> CoalesceMap;
|
|
|
|
typedef std::set<unsigned> RegSet;
|
|
|
|
|
2011-06-17 15:09:01 +08:00
|
|
|
char *customPassID;
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
RegSet VRegsToAlloc, EmptyIntervalVRegs;
|
2010-09-23 12:28:54 +08:00
|
|
|
|
|
|
|
/// \brief Finds the initial set of vreg intervals to allocate.
|
2014-10-10 02:20:51 +08:00
|
|
|
void findVRegIntervalsToAlloc(const MachineFunction &MF, LiveIntervals &LIS);
|
|
|
|
|
|
|
|
/// \brief Constructs an initial graph.
|
|
|
|
void initializeGraph(PBQPRAGraph &G);
|
2010-09-23 12:28:54 +08:00
|
|
|
|
|
|
|
/// \brief Given a solved PBQP problem maps this solution back to a register
|
|
|
|
/// assignment.
|
2014-10-10 02:20:51 +08:00
|
|
|
bool mapPBQPToRegAlloc(const PBQPRAGraph &G,
|
|
|
|
const PBQP::Solution &Solution,
|
|
|
|
VirtRegMap &VRM,
|
|
|
|
Spiller &VRegSpiller);
|
2010-09-23 12:28:54 +08:00
|
|
|
|
|
|
|
/// \brief Postprocessing before final spilling. Sets basic block "live in"
|
|
|
|
/// variables.
|
2014-10-10 02:20:51 +08:00
|
|
|
void finalizeAlloc(MachineFunction &MF, LiveIntervals &LIS,
|
|
|
|
VirtRegMap &VRM) const;
|
2010-09-23 12:28:54 +08:00
|
|
|
|
|
|
|
};
|
|
|
|
|
2010-09-18 17:07:10 +08:00
|
|
|
char RegAllocPBQP::ID = 0;
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
/// @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;
|
|
|
|
|
[PBQP] Tweak spill costs and coalescing benefits
This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
- coalescing (or any other "side effect" of reg alloc) is negative, and
instead of being derived from a spill cost, they use the block
frequency info.
- spill costs are in the [MinSpillCost:+inf( range
- register or interference costs are in [0.0:MinSpillCost( or +inf
The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.
The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.
llvm-svn: 221292
2014-11-05 04:51:24 +08:00
|
|
|
// 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;
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
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();
|
[PBQP] Tweak spill costs and coalescing benefits
This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
- coalescing (or any other "side effect" of reg alloc) is negative, and
instead of being derived from a spill cost, they use the block
frequency info.
- spill costs are in the [MinSpillCost:+inf( range
- register or interference costs are in [0.0:MinSpillCost( or +inf
The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.
The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.
llvm-svn: 221292
2014-11-05 04:51:24 +08:00
|
|
|
else
|
|
|
|
SpillCost += MinSpillCost;
|
2014-10-10 02:20:51 +08:00
|
|
|
PBQPRAGraph::RawVector NodeCosts(G.getNodeCosts(NId));
|
|
|
|
NodeCosts[PBQP::RegAlloc::getSpillOptionIdx()] = SpillCost;
|
|
|
|
G.setNodeCosts(NId, std::move(NodeCosts));
|
|
|
|
}
|
2010-09-18 17:07:10 +08:00
|
|
|
}
|
2014-10-10 02:20:51 +08:00
|
|
|
};
|
2010-09-18 17:07:10 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
/// @brief Add interference edges between overlapping vregs.
|
|
|
|
class Interference : public PBQPRAConstraint {
|
2014-10-19 01:26:07 +08:00
|
|
|
private:
|
|
|
|
|
2014-10-28 01:44:25 +08:00
|
|
|
private:
|
|
|
|
|
|
|
|
typedef const PBQP::RegAlloc::AllowedRegVector* AllowedRegVecPtr;
|
|
|
|
typedef std::pair<AllowedRegVecPtr, AllowedRegVecPtr> IMatrixKey;
|
|
|
|
typedef DenseMap<IMatrixKey, PBQPRAGraph::MatrixPtr> IMatrixCache;
|
|
|
|
|
2014-10-19 01:26:07 +08:00
|
|
|
// 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;
|
|
|
|
|
|
|
|
static SlotIndex getStartPoint(const IntervalInfo &I) {
|
|
|
|
return std::get<0>(I)->segments[std::get<1>(I)].start;
|
|
|
|
}
|
|
|
|
|
|
|
|
static SlotIndex getEndPoint(const IntervalInfo &I) {
|
|
|
|
return std::get<0>(I)->segments[std::get<1>(I)].end;
|
|
|
|
}
|
|
|
|
|
|
|
|
static PBQP::GraphBase::NodeId getNodeId(const IntervalInfo &I) {
|
|
|
|
return std::get<2>(I);
|
|
|
|
}
|
|
|
|
|
|
|
|
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);
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool lowestEndPoint(const IntervalInfo &I1,
|
|
|
|
const IntervalInfo &I2) {
|
|
|
|
SlotIndex E1 = getEndPoint(I1);
|
|
|
|
SlotIndex E2 = getEndPoint(I2);
|
|
|
|
|
|
|
|
if (E1 < E2)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (E1 > E2)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// 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;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool isAtLastSegment(const IntervalInfo &I) {
|
|
|
|
return std::get<1>(I) == std::get<0>(I)->size() - 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static IntervalInfo nextSegment(const IntervalInfo &I) {
|
|
|
|
return std::make_tuple(std::get<0>(I), std::get<1>(I) + 1, std::get<2>(I));
|
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
public:
|
2012-06-21 06:32:05 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
void apply(PBQPRAGraph &G) override {
|
2014-10-19 01:26:07 +08:00
|
|
|
// 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.
|
2014-10-10 02:20:51 +08:00
|
|
|
LiveIntervals &LIS = G.getMetadata().LIS;
|
2014-10-28 01:44:25 +08:00
|
|
|
|
|
|
|
// 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;
|
2014-10-10 02:20:51 +08:00
|
|
|
|
2014-10-19 01:26:07 +08:00
|
|
|
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));
|
|
|
|
}
|
|
|
|
|
|
|
|
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.
|
2014-10-28 01:44:25 +08:00
|
|
|
if (G.findEdge(NId, MId) != PBQPRAGraph::invalidEdgeId())
|
2014-10-19 01:26:07 +08:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// This is a new edge - add it to the graph.
|
2014-10-28 01:44:25 +08:00
|
|
|
createInterferenceEdge(G, NId, MId, C);
|
2010-09-18 17:07:10 +08:00
|
|
|
}
|
2014-10-19 01:26:07 +08:00
|
|
|
|
|
|
|
// Finally, add Cur to the Active set.
|
|
|
|
Active.insert(Cur);
|
2010-09-18 17:07:10 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
private:
|
2010-07-17 14:31:41 +08:00
|
|
|
|
2014-10-28 01:44:25 +08:00
|
|
|
void createInterferenceEdge(PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
|
|
|
|
PBQPRAGraph::NodeId MId, IMatrixCache &C) {
|
|
|
|
|
|
|
|
const TargetRegisterInfo &TRI =
|
|
|
|
*G.getMetadata().MF.getTarget().getSubtargetImpl()->getRegisterInfo();
|
|
|
|
|
|
|
|
const auto &NRegs = G.getNodeMetadata(NId).getAllowedRegs();
|
|
|
|
const auto &MRegs = G.getNodeMetadata(MId).getAllowedRegs();
|
|
|
|
|
|
|
|
// 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;
|
|
|
|
}
|
|
|
|
|
|
|
|
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];
|
2014-10-10 02:20:51 +08:00
|
|
|
if (TRI.regsOverlap(PRegN, PRegM))
|
|
|
|
M[I + 1][J + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
|
2010-07-17 14:31:41 +08:00
|
|
|
}
|
2010-09-18 17:07:10 +08:00
|
|
|
}
|
2014-10-10 02:20:51 +08:00
|
|
|
|
2014-10-28 01:44:25 +08:00
|
|
|
PBQPRAGraph::EdgeId EId = G.addEdge(NId, MId, std::move(M));
|
|
|
|
C[K] = G.getEdgeCostsPtr(EId);
|
2010-09-18 17:07:10 +08:00
|
|
|
}
|
2014-10-10 02:20:51 +08:00
|
|
|
};
|
2008-10-03 02:29:27 +08:00
|
|
|
|
2010-09-21 21:19:36 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
class Coalescing : public PBQPRAConstraint {
|
|
|
|
public:
|
|
|
|
void apply(PBQPRAGraph &G) override {
|
|
|
|
MachineFunction &MF = G.getMetadata().MF;
|
|
|
|
MachineBlockFrequencyInfo &MBFI = G.getMetadata().MBFI;
|
|
|
|
CoalescerPair CP(*MF.getTarget().getSubtargetImpl()->getRegisterInfo());
|
|
|
|
|
|
|
|
// Scan the machine function and add a coalescing cost whenever CoalescerPair
|
|
|
|
// gives the Ok.
|
|
|
|
for (const auto &MBB : MF) {
|
|
|
|
for (const auto &MI : MBB) {
|
|
|
|
|
|
|
|
// Skip not-coalescable or already coalesced copies.
|
|
|
|
if (!CP.setRegisters(&MI) || CP.getSrcReg() == CP.getDstReg())
|
|
|
|
continue;
|
2010-09-21 21:19:36 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
unsigned DstReg = CP.getDstReg();
|
|
|
|
unsigned SrcReg = CP.getSrcReg();
|
2010-09-21 21:19:36 +08:00
|
|
|
|
[PBQP] Tweak spill costs and coalescing benefits
This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
- coalescing (or any other "side effect" of reg alloc) is negative, and
instead of being derived from a spill cost, they use the block
frequency info.
- spill costs are in the [MinSpillCost:+inf( range
- register or interference costs are in [0.0:MinSpillCost( or +inf
The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.
The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.
llvm-svn: 221292
2014-11-05 04:51:24 +08:00
|
|
|
const float Scale = 1.0f / MBFI.getEntryFreq();
|
|
|
|
PBQP::PBQPNum CBenefit = MBFI.getBlockFreq(&MBB).getFrequency() * Scale;
|
2010-09-21 21:19:36 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
if (CP.isPhys()) {
|
|
|
|
if (!MF.getRegInfo().isAllocatable(DstReg))
|
|
|
|
continue;
|
2010-09-21 21:19:36 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
PBQPRAGraph::NodeId NId = G.getMetadata().getNodeIdForVReg(SrcReg);
|
2012-02-10 12:10:26 +08:00
|
|
|
|
2014-10-28 01:44:25 +08:00
|
|
|
const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed =
|
|
|
|
G.getNodeMetadata(NId).getAllowedRegs();
|
2010-09-21 21:19:36 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
unsigned PRegOpt = 0;
|
|
|
|
while (PRegOpt < Allowed.size() && Allowed[PRegOpt] != DstReg)
|
|
|
|
++PRegOpt;
|
2010-09-21 21:19:36 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
if (PRegOpt < Allowed.size()) {
|
|
|
|
PBQPRAGraph::RawVector NewCosts(G.getNodeCosts(NId));
|
2014-10-22 00:24:15 +08:00
|
|
|
NewCosts[PRegOpt + 1] -= CBenefit;
|
2014-10-10 02:20:51 +08:00
|
|
|
G.setNodeCosts(NId, std::move(NewCosts));
|
|
|
|
}
|
2010-09-21 21:19:36 +08:00
|
|
|
} else {
|
2014-10-10 02:20:51 +08:00
|
|
|
PBQPRAGraph::NodeId N1Id = G.getMetadata().getNodeIdForVReg(DstReg);
|
|
|
|
PBQPRAGraph::NodeId N2Id = G.getMetadata().getNodeIdForVReg(SrcReg);
|
2014-10-28 01:44:25 +08:00
|
|
|
const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed1 =
|
|
|
|
&G.getNodeMetadata(N1Id).getAllowedRegs();
|
|
|
|
const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed2 =
|
|
|
|
&G.getNodeMetadata(N2Id).getAllowedRegs();
|
2014-10-10 02:20:51 +08:00
|
|
|
|
|
|
|
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));
|
2010-09-21 21:19:36 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
private:
|
2010-09-21 21:19:36 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
void addVirtRegCoalesce(
|
2014-10-28 01:44:25 +08:00
|
|
|
PBQPRAGraph::RawMatrix &CostMat,
|
|
|
|
const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed1,
|
|
|
|
const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed2,
|
|
|
|
PBQP::PBQPNum Benefit) {
|
2014-10-10 02:20:51 +08:00
|
|
|
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)
|
2014-10-22 00:24:15 +08:00
|
|
|
CostMat[I + 1][J + 1] -= Benefit;
|
2012-02-10 12:10:26 +08:00
|
|
|
}
|
2010-09-21 21:19:36 +08:00
|
|
|
}
|
|
|
|
}
|
2008-10-03 02:29:27 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
} // End anonymous namespace.
|
|
|
|
|
|
|
|
// Out-of-line destructor/anchor for PBQPRAConstraint.
|
|
|
|
PBQPRAConstraint::~PBQPRAConstraint() {}
|
|
|
|
void PBQPRAConstraint::anchor() {}
|
|
|
|
void PBQPRAConstraintList::anchor() {}
|
2010-09-18 17:07:10 +08:00
|
|
|
|
|
|
|
void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
|
2011-12-06 09:45:57 +08:00
|
|
|
au.setPreservesCFG();
|
|
|
|
au.addRequired<AliasAnalysis>();
|
|
|
|
au.addPreserved<AliasAnalysis>();
|
2010-09-18 17:07:10 +08:00
|
|
|
au.addRequired<SlotIndexes>();
|
|
|
|
au.addPreserved<SlotIndexes>();
|
|
|
|
au.addRequired<LiveIntervals>();
|
2012-10-04 12:50:53 +08:00
|
|
|
au.addPreserved<LiveIntervals>();
|
2010-09-18 17:07:10 +08:00
|
|
|
//au.addRequiredID(SplitCriticalEdgesID);
|
2011-06-17 15:09:01 +08:00
|
|
|
if (customPassID)
|
|
|
|
au.addRequiredID(*customPassID);
|
2010-09-18 17:07:10 +08:00
|
|
|
au.addRequired<LiveStacks>();
|
|
|
|
au.addPreserved<LiveStacks>();
|
2013-06-18 03:00:36 +08:00
|
|
|
au.addRequired<MachineBlockFrequencyInfo>();
|
|
|
|
au.addPreserved<MachineBlockFrequencyInfo>();
|
2013-07-02 04:47:47 +08:00
|
|
|
au.addRequired<MachineLoopInfo>();
|
|
|
|
au.addPreserved<MachineLoopInfo>();
|
2011-12-06 09:45:57 +08:00
|
|
|
au.addRequired<MachineDominatorTree>();
|
|
|
|
au.addPreserved<MachineDominatorTree>();
|
2010-09-18 17:07:10 +08:00
|
|
|
au.addRequired<VirtRegMap>();
|
2012-10-04 12:50:53 +08:00
|
|
|
au.addPreserved<VirtRegMap>();
|
2010-09-18 17:07:10 +08:00
|
|
|
MachineFunctionPass::getAnalysisUsage(au);
|
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
void RegAllocPBQP::findVRegIntervalsToAlloc(const MachineFunction &MF,
|
|
|
|
LiveIntervals &LIS) {
|
|
|
|
const MachineRegisterInfo &MRI = MF.getRegInfo();
|
2008-11-16 20:12:54 +08:00
|
|
|
|
|
|
|
// Iterate over all live ranges.
|
2014-10-10 02:20:51 +08:00
|
|
|
for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) {
|
|
|
|
unsigned Reg = TargetRegisterInfo::index2VirtReg(I);
|
|
|
|
if (MRI.reg_nodbg_empty(Reg))
|
2008-11-16 20:12:54 +08:00
|
|
|
continue;
|
2014-10-10 02:20:51 +08:00
|
|
|
LiveInterval &LI = LIS.getInterval(Reg);
|
2008-11-16 20:12:54 +08:00
|
|
|
|
|
|
|
// 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.
|
2014-10-10 02:20:51 +08:00
|
|
|
if (!LI.empty()) {
|
|
|
|
VRegsToAlloc.insert(LI.reg);
|
2010-11-12 13:47:21 +08:00
|
|
|
} else {
|
2014-10-10 02:20:51 +08:00
|
|
|
EmptyIntervalVRegs.insert(LI.reg);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-11-05 04:51:29 +08:00
|
|
|
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;
|
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
void RegAllocPBQP::initializeGraph(PBQPRAGraph &G) {
|
|
|
|
MachineFunction &MF = G.getMetadata().MF;
|
|
|
|
|
|
|
|
LiveIntervals &LIS = G.getMetadata().LIS;
|
|
|
|
const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
|
|
|
|
const TargetRegisterInfo &TRI =
|
|
|
|
*G.getMetadata().MF.getTarget().getSubtargetImpl()->getRegisterInfo();
|
|
|
|
|
|
|
|
for (auto VReg : VRegsToAlloc) {
|
|
|
|
const TargetRegisterClass *TRC = MRI.getRegClass(VReg);
|
|
|
|
LiveInterval &VRegLI = LIS.getInterval(VReg);
|
|
|
|
|
|
|
|
// Record any overlaps with regmask operands.
|
|
|
|
BitVector RegMaskOverlaps;
|
|
|
|
LIS.checkRegMaskInterference(VRegLI, RegMaskOverlaps);
|
|
|
|
|
|
|
|
// 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;
|
|
|
|
|
|
|
|
// vregLI crosses a regmask operand that clobbers preg.
|
|
|
|
if (!RegMaskOverlaps.empty() && !RegMaskOverlaps.test(PReg))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// 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;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (Interference)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// preg is usable for this virtual register.
|
|
|
|
VRegAllowed.push_back(PReg);
|
2008-11-16 20:12:54 +08:00
|
|
|
}
|
2014-10-10 02:20:51 +08:00
|
|
|
|
|
|
|
PBQPRAGraph::RawVector NodeCosts(VRegAllowed.size() + 1, 0);
|
2014-11-05 04:51:29 +08:00
|
|
|
|
|
|
|
// 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;
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
PBQPRAGraph::NodeId NId = G.addNode(std::move(NodeCosts));
|
|
|
|
G.getNodeMetadata(NId).setVReg(VReg);
|
2014-10-28 01:44:25 +08:00
|
|
|
G.getNodeMetadata(NId).setAllowedRegs(
|
|
|
|
G.getMetadata().getAllowedRegs(std::move(VRegAllowed)));
|
2014-10-10 02:20:51 +08:00
|
|
|
G.getMetadata().setNodeIdForVReg(VReg, NId);
|
2008-11-16 20:12:54 +08:00
|
|
|
}
|
2008-10-03 02:29:27 +08:00
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
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.getTarget().getSubtargetImpl()->getRegisterInfo();
|
|
|
|
(void)TRI;
|
|
|
|
|
2010-09-18 17:07:10 +08:00
|
|
|
// Set to true if we have any spills
|
2014-10-10 02:20:51 +08:00
|
|
|
bool AnotherRoundNeeded = false;
|
2010-09-18 17:07:10 +08:00
|
|
|
|
|
|
|
// Clear the existing allocation.
|
2014-10-10 02:20:51 +08:00
|
|
|
VRM.clearAllVirt();
|
2010-09-18 17:07:10 +08:00
|
|
|
|
|
|
|
// Iterate over the nodes mapping the PBQP solution to a register
|
|
|
|
// assignment.
|
2014-10-10 02:20:51 +08:00
|
|
|
for (auto NId : G.nodeIds()) {
|
|
|
|
unsigned VReg = G.getNodeMetadata(NId).getVReg();
|
|
|
|
unsigned AllocOption = Solution.getSelection(NId);
|
|
|
|
|
|
|
|
if (AllocOption != PBQP::RegAlloc::getSpillOptionIdx()) {
|
2014-10-28 01:44:25 +08:00
|
|
|
unsigned PReg = G.getNodeMetadata(NId).getAllowedRegs()[AllocOption - 1];
|
2014-10-10 02:20:51 +08:00
|
|
|
DEBUG(dbgs() << "VREG " << PrintReg(VReg, &TRI) << " -> "
|
|
|
|
<< TRI.getName(PReg) << "\n");
|
|
|
|
assert(PReg != 0 && "Invalid preg selected.");
|
|
|
|
VRM.assignVirt2Phys(VReg, PReg);
|
|
|
|
} else {
|
|
|
|
VRegsToAlloc.erase(VReg);
|
|
|
|
SmallVector<unsigned, 8> NewSpills;
|
|
|
|
LiveRangeEdit LRE(&LIS.getInterval(VReg), NewSpills, MF, LIS, &VRM);
|
|
|
|
VRegSpiller.spill(LRE);
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "VREG " << PrintReg(VReg, &TRI) << " -> SPILLED (Cost: "
|
2011-11-13 07:17:52 +08:00
|
|
|
<< LRE.getParent().weight << ", New vregs: ");
|
2010-09-18 17:07:10 +08:00
|
|
|
|
|
|
|
// Copy any newly inserted live intervals into the list of regs to
|
|
|
|
// allocate.
|
2014-10-10 02:20:51 +08:00
|
|
|
for (LiveRangeEdit::iterator I = LRE.begin(), E = LRE.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
LiveInterval &LI = LIS.getInterval(*I);
|
|
|
|
assert(!LI.empty() && "Empty spill range.");
|
|
|
|
DEBUG(dbgs() << PrintReg(LI.reg, &TRI) << " ");
|
|
|
|
VRegsToAlloc.insert(LI.reg);
|
2008-11-16 20:12:54 +08:00
|
|
|
}
|
|
|
|
|
2010-01-05 09:25:43 +08:00
|
|
|
DEBUG(dbgs() << ")\n");
|
2008-10-03 02:29:27 +08:00
|
|
|
|
|
|
|
// We need another round if spill intervals were added.
|
2014-10-10 02:20:51 +08:00
|
|
|
AnotherRoundNeeded |= !LRE.empty();
|
2008-10-03 02:29:27 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
return !AnotherRoundNeeded;
|
2008-10-03 02:29:27 +08:00
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
void RegAllocPBQP::finalizeAlloc(MachineFunction &MF,
|
|
|
|
LiveIntervals &LIS,
|
|
|
|
VirtRegMap &VRM) const {
|
|
|
|
MachineRegisterInfo &MRI = MF.getRegInfo();
|
|
|
|
|
2008-11-16 20:12:54 +08:00
|
|
|
// First allocate registers for the empty intervals.
|
2010-09-18 17:07:10 +08:00
|
|
|
for (RegSet::const_iterator
|
2014-10-10 02:20:51 +08:00
|
|
|
I = EmptyIntervalVRegs.begin(), E = EmptyIntervalVRegs.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
LiveInterval &LI = LIS.getInterval(*I);
|
2008-11-16 20:12:54 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
unsigned PReg = MRI.getSimpleHint(LI.reg);
|
2009-08-07 07:32:48 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
if (PReg == 0) {
|
|
|
|
const TargetRegisterClass &RC = *MRI.getRegClass(LI.reg);
|
|
|
|
PReg = RC.getRawAllocationOrder(MF).front();
|
2008-11-16 20:12:54 +08:00
|
|
|
}
|
2009-01-08 23:50:22 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
VRM.assignVirt2Phys(LI.reg, PReg);
|
2008-11-16 20:12:54 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
[PBQP] Tweak spill costs and coalescing benefits
This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
- coalescing (or any other "side effect" of reg alloc) is negative, and
instead of being derived from a spill cost, they use the block
frequency info.
- spill costs are in the [MinSpillCost:+inf( range
- register or interference costs are in [0.0:MinSpillCost( or +inf
The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.
The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.
llvm-svn: 221292
2014-11-05 04:51:24 +08:00
|
|
|
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);
|
|
|
|
}
|
|
|
|
|
2010-09-18 17:07:10 +08:00
|
|
|
bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
|
2014-10-10 02:20:51 +08:00
|
|
|
LiveIntervals &LIS = getAnalysis<LiveIntervals>();
|
|
|
|
MachineBlockFrequencyInfo &MBFI =
|
|
|
|
getAnalysis<MachineBlockFrequencyInfo>();
|
2008-11-16 20:12:54 +08:00
|
|
|
|
[PBQP] Tweak spill costs and coalescing benefits
This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
- coalescing (or any other "side effect" of reg alloc) is negative, and
instead of being derived from a spill cost, they use the block
frequency info.
- spill costs are in the [MinSpillCost:+inf( range
- register or interference costs are in [0.0:MinSpillCost( or +inf
The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.
The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.
llvm-svn: 221292
2014-11-05 04:51:24 +08:00
|
|
|
calculateSpillWeightsAndHints(LIS, MF, getAnalysis<MachineLoopInfo>(), MBFI,
|
|
|
|
normalizePBQPSpillWeight);
|
2008-10-03 02:29:27 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
VirtRegMap &VRM = getAnalysis<VirtRegMap>();
|
2013-11-11 01:46:31 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
std::unique_ptr<Spiller> VRegSpiller(createInlineSpiller(*this, MF, VRM));
|
2008-10-03 02:29:27 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
MF.getRegInfo().freezeReservedRegs(MF);
|
2012-11-28 08:21:29 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
DEBUG(dbgs() << "PBQP Register Allocating for " << MF.getName() << "\n");
|
2008-11-16 20:12:54 +08:00
|
|
|
|
2008-10-03 02:29:27 +08:00
|
|
|
// Allocator main loop:
|
2009-01-08 23:50:22 +08:00
|
|
|
//
|
2008-10-03 02:29:27 +08:00
|
|
|
// * Map current regalloc problem to a PBQP problem
|
|
|
|
// * Solve the PBQP problem
|
|
|
|
// * Map the solution back to a register allocation
|
|
|
|
// * Spill if necessary
|
2009-01-08 23:50:22 +08:00
|
|
|
//
|
2008-10-03 02:29:27 +08:00
|
|
|
// This process is continued till no more spills are generated.
|
|
|
|
|
2008-11-16 20:12:54 +08:00
|
|
|
// Find the vreg intervals in need of allocation.
|
2014-10-10 02:20:51 +08:00
|
|
|
findVRegIntervalsToAlloc(MF, LIS);
|
2009-01-08 23:50:22 +08:00
|
|
|
|
2012-08-22 14:07:19 +08:00
|
|
|
#ifndef NDEBUG
|
2014-10-10 02:20:51 +08:00
|
|
|
const Function &F = *MF.getFunction();
|
|
|
|
std::string FullyQualifiedName =
|
|
|
|
F.getParent()->getModuleIdentifier() + "." + F.getName().str();
|
2012-08-22 14:07:19 +08:00
|
|
|
#endif
|
2012-03-27 07:07:23 +08:00
|
|
|
|
2008-11-16 20:12:54 +08:00
|
|
|
// If there are non-empty intervals allocate them using pbqp.
|
2014-10-10 02:20:51 +08:00
|
|
|
if (!VRegsToAlloc.empty()) {
|
2008-11-16 20:12:54 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
const TargetSubtargetInfo &Subtarget = *MF.getTarget().getSubtargetImpl();
|
|
|
|
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());
|
2008-11-16 20:12:54 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
bool PBQPAllocComplete = false;
|
|
|
|
unsigned Round = 0;
|
2009-08-07 07:32:48 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
while (!PBQPAllocComplete) {
|
|
|
|
DEBUG(dbgs() << " PBQP Regalloc round " << Round << ":\n");
|
|
|
|
|
|
|
|
PBQPRAGraph G(PBQPRAGraph::GraphMetadata(MF, LIS, MBFI));
|
|
|
|
initializeGraph(G);
|
|
|
|
ConstraintsRoot->apply(G);
|
2012-03-27 07:07:23 +08:00
|
|
|
|
|
|
|
#ifndef NDEBUG
|
2014-10-10 02:20:51 +08:00
|
|
|
if (PBQPDumpGraphs) {
|
|
|
|
std::ostringstream RS;
|
|
|
|
RS << Round;
|
|
|
|
std::string GraphFileName = FullyQualifiedName + "." + RS.str() +
|
|
|
|
".pbqpgraph";
|
2014-08-26 02:16:47 +08:00
|
|
|
std::error_code EC;
|
2014-10-10 02:20:51 +08:00
|
|
|
raw_fd_ostream OS(GraphFileName, EC, sys::fs::F_Text);
|
|
|
|
DEBUG(dbgs() << "Dumping graph for round " << Round << " to \""
|
|
|
|
<< GraphFileName << "\"\n");
|
|
|
|
G.dumpToStream(OS);
|
2012-03-27 07:07:23 +08:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
PBQP::Solution Solution = PBQP::RegAlloc::solve(G);
|
|
|
|
PBQPAllocComplete = mapPBQPToRegAlloc(G, Solution, VRM, *VRegSpiller);
|
|
|
|
++Round;
|
2008-11-16 20:12:54 +08:00
|
|
|
}
|
2008-10-03 02:29:27 +08:00
|
|
|
}
|
|
|
|
|
2008-11-16 20:12:54 +08:00
|
|
|
// Finalise allocation, allocate empty ranges.
|
2014-10-10 02:20:51 +08:00
|
|
|
finalizeAlloc(MF, LIS, VRM);
|
|
|
|
VRegsToAlloc.clear();
|
|
|
|
EmptyIntervalVRegs.clear();
|
2008-10-03 02:29:27 +08:00
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << VRM << "\n");
|
2008-11-16 20:12:54 +08:00
|
|
|
|
2009-01-08 23:50:22 +08:00
|
|
|
return true;
|
2008-10-03 02:29:27 +08:00
|
|
|
}
|
|
|
|
|
2014-10-10 02:20:51 +08:00
|
|
|
FunctionPass *llvm::createPBQPRegisterAllocator(char *customPassID) {
|
|
|
|
return new RegAllocPBQP(customPassID);
|
2008-10-03 02:29:27 +08:00
|
|
|
}
|
|
|
|
|
2010-09-23 12:28:54 +08:00
|
|
|
FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
|
2014-10-10 02:20:51 +08:00
|
|
|
return createPBQPRegisterAllocator();
|
2010-09-18 17:07:10 +08:00
|
|
|
}
|
2008-10-03 02:29:27 +08:00
|
|
|
|
|
|
|
#undef DEBUG_TYPE
|