Google Git
Sign in
swiftshader / SwiftShader / af6471bdf141a130af8ccc84f7323e2a9df25f16 / . / third_party / llvm-10.0 / llvm / lib / CodeGen / RegisterCoalescer.cpp
blob: a3f75d82d0ecff3b325f86952c51e723a39c5932 [file] [log] [blame]
//===- RegisterCoalescer.cpp - Generic Register Coalescing Interface ------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements the generic RegisterCoalescer interface which
// is used as the common interface used by all clients and
// implementations of register coalescing.
//
//===----------------------------------------------------------------------===//
#include "RegisterCoalescer.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/LiveInterval.h"
#include "llvm/CodeGen/LiveIntervals.h"
#include "llvm/CodeGen/LiveRangeEdit.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/RegisterClassInfo.h"
#include "llvm/CodeGen/SlotIndexes.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/InitializePasses.h"
#include "llvm/MC/LaneBitmask.h"
#include "llvm/MC/MCInstrDesc.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/Pass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <iterator>
#include <limits>
#include <tuple>
#include <utility>
#include <vector>
using namespace llvm;
#define DEBUG_TYPE "regalloc"
STATISTIC(numJoins , "Number of interval joins performed");
STATISTIC(numCrossRCs , "Number of cross class joins performed");
STATISTIC(numCommutes , "Number of instruction commuting performed");
STATISTIC(numExtends , "Number of copies extended");
STATISTIC(NumReMats , "Number of instructions re-materialized");
STATISTIC(NumInflated , "Number of register classes inflated");
STATISTIC(NumLaneConflicts, "Number of dead lane conflicts tested");
STATISTIC(NumLaneResolves, "Number of dead lane conflicts resolved");
STATISTIC(NumShrinkToUses, "Number of shrinkToUses called");
static cl::opt<bool> EnableJoining("join-liveintervals",
cl::desc("Coalesce copies (default=true)"),
cl::init(true), cl::Hidden);
static cl::opt<bool> UseTerminalRule("terminal-rule",
cl::desc("Apply the terminal rule"),
cl::init(false), cl::Hidden);
/// Temporary flag to test critical edge unsplitting.
static cl::opt<bool>
EnableJoinSplits("join-splitedges",
cl::desc("Coalesce copies on split edges (default=subtarget)"), cl::Hidden);
/// Temporary flag to test global copy optimization.
static cl::opt<cl::boolOrDefault>
EnableGlobalCopies("join-globalcopies",
cl::desc("Coalesce copies that span blocks (default=subtarget)"),
cl::init(cl::BOU_UNSET), cl::Hidden);
static cl::opt<bool>
VerifyCoalescing("verify-coalescing",
cl::desc("Verify machine instrs before and after register coalescing"),
cl::Hidden);
static cl::opt<unsigned> LateRematUpdateThreshold(
"late-remat-update-threshold", cl::Hidden,
cl::desc("During rematerialization for a copy, if the def instruction has "
"many other copy uses to be rematerialized, delay the multiple "
"separate live interval update work and do them all at once after "
"all those rematerialization are done. It will save a lot of "
"repeated work. "),
cl::init(100));
static cl::opt<unsigned> LargeIntervalSizeThreshold(
"large-interval-size-threshold", cl::Hidden,
cl::desc("If the valnos size of an interval is larger than the threshold, "
"it is regarded as a large interval. "),
cl::init(100));
static cl::opt<unsigned> LargeIntervalFreqThreshold(
"large-interval-freq-threshold", cl::Hidden,
cl::desc("For a large interval, if it is coalesed with other live "
"intervals many times more than the threshold, stop its "
"coalescing to control the compile time. "),
cl::init(100));
namespace {
class JoinVals;
class RegisterCoalescer : public MachineFunctionPass,
private LiveRangeEdit::Delegate {
MachineFunction* MF = nullptr;
MachineRegisterInfo* MRI = nullptr;
const TargetRegisterInfo* TRI = nullptr;
const TargetInstrInfo* TII = nullptr;
LiveIntervals *LIS = nullptr;
const MachineLoopInfo* Loops = nullptr;
AliasAnalysis *AA = nullptr;
RegisterClassInfo RegClassInfo;
/// Debug variable location tracking -- for each VReg, maintain an
/// ordered-by-slot-index set of DBG_VALUEs, to help quick
/// identification of whether coalescing may change location validity.
using DbgValueLoc = std::pair<SlotIndex, MachineInstr*>;
DenseMap<unsigned, std::vector<DbgValueLoc>> DbgVRegToValues;
/// VRegs may be repeatedly coalesced, and have many DBG_VALUEs attached.
/// To avoid repeatedly merging sets of DbgValueLocs, instead record
/// which vregs have been coalesced, and where to. This map is from
/// vreg => {set of vregs merged in}.
DenseMap<unsigned, SmallVector<unsigned, 4>> DbgMergedVRegNums;
/// A LaneMask to remember on which subregister live ranges we need to call
/// shrinkToUses() later.
LaneBitmask ShrinkMask;
/// True if the main range of the currently coalesced intervals should be
/// checked for smaller live intervals.
bool ShrinkMainRange = false;
/// True if the coalescer should aggressively coalesce global copies
/// in favor of keeping local copies.
bool JoinGlobalCopies = false;
/// True if the coalescer should aggressively coalesce fall-thru
/// blocks exclusively containing copies.
bool JoinSplitEdges = false;
/// Copy instructions yet to be coalesced.
SmallVector<MachineInstr*, 8> WorkList;
SmallVector<MachineInstr*, 8> LocalWorkList;
/// Set of instruction pointers that have been erased, and
/// that may be present in WorkList.
SmallPtrSet<MachineInstr*, 8> ErasedInstrs;
/// Dead instructions that are about to be deleted.
SmallVector<MachineInstr*, 8> DeadDefs;
/// Virtual registers to be considered for register class inflation.
SmallVector<unsigned, 8> InflateRegs;
/// The collection of live intervals which should have been updated
/// immediately after rematerialiation but delayed until
/// lateLiveIntervalUpdate is called.
DenseSet<unsigned> ToBeUpdated;
/// Record how many times the large live interval with many valnos
/// has been tried to join with other live interval.
DenseMap<unsigned, unsigned long> LargeLIVisitCounter;
/// Recursively eliminate dead defs in DeadDefs.
void eliminateDeadDefs();
/// LiveRangeEdit callback for eliminateDeadDefs().
void LRE_WillEraseInstruction(MachineInstr *MI) override;
/// Coalesce the LocalWorkList.
void coalesceLocals();
/// Join compatible live intervals
void joinAllIntervals();
/// Coalesce copies in the specified MBB, putting
/// copies that cannot yet be coalesced into WorkList.
void copyCoalesceInMBB(MachineBasicBlock *MBB);
/// Tries to coalesce all copies in CurrList. Returns true if any progress
/// was made.
bool copyCoalesceWorkList(MutableArrayRef<MachineInstr*> CurrList);
/// If one def has many copy like uses, and those copy uses are all
/// rematerialized, the live interval update needed for those
/// rematerializations will be delayed and done all at once instead
/// of being done multiple times. This is to save compile cost because
/// live interval update is costly.
void lateLiveIntervalUpdate();
/// Attempt to join intervals corresponding to SrcReg/DstReg, which are the
/// src/dst of the copy instruction CopyMI. This returns true if the copy
/// was successfully coalesced away. If it is not currently possible to
/// coalesce this interval, but it may be possible if other things get
/// coalesced, then it returns true by reference in 'Again'.
bool joinCopy(MachineInstr *CopyMI, bool &Again);
/// Attempt to join these two intervals. On failure, this
/// returns false. The output "SrcInt" will not have been modified, so we
/// can use this information below to update aliases.
bool joinIntervals(CoalescerPair &CP);
/// Attempt joining two virtual registers. Return true on success.
bool joinVirtRegs(CoalescerPair &CP);
/// If a live interval has many valnos and is coalesced with other
/// live intervals many times, we regard such live interval as having
/// high compile time cost.
bool isHighCostLiveInterval(LiveInterval &LI);
/// Attempt joining with a reserved physreg.
bool joinReservedPhysReg(CoalescerPair &CP);
/// Add the LiveRange @p ToMerge as a subregister liverange of @p LI.
/// Subranges in @p LI which only partially interfere with the desired
/// LaneMask are split as necessary. @p LaneMask are the lanes that
/// @p ToMerge will occupy in the coalescer register. @p LI has its subrange
/// lanemasks already adjusted to the coalesced register.
void mergeSubRangeInto(LiveInterval &LI, const LiveRange &ToMerge,
LaneBitmask LaneMask, CoalescerPair &CP,
unsigned DstIdx);
/// Join the liveranges of two subregisters. Joins @p RRange into
/// @p LRange, @p RRange may be invalid afterwards.
void joinSubRegRanges(LiveRange &LRange, LiveRange &RRange,
LaneBitmask LaneMask, const CoalescerPair &CP);
/// We found a non-trivially-coalescable copy. If the source value number is
/// defined by a copy from the destination reg see if we can merge these two
/// destination reg valno# into a single value number, eliminating a copy.
/// This returns true if an interval was modified.
bool adjustCopiesBackFrom(const CoalescerPair &CP, MachineInstr *CopyMI);
/// Return true if there are definitions of IntB
/// other than BValNo val# that can reach uses of AValno val# of IntA.
bool hasOtherReachingDefs(LiveInterval &IntA, LiveInterval &IntB,
VNInfo *AValNo, VNInfo *BValNo);
/// We found a non-trivially-coalescable copy.
/// If the source value number is defined by a commutable instruction and
/// its other operand is coalesced to the copy dest register, see if we
/// can transform the copy into a noop by commuting the definition.
/// This returns a pair of two flags:
/// - the first element is true if an interval was modified,
/// - the second element is true if the destination interval needs
/// to be shrunk after deleting the copy.
std::pair<bool,bool> removeCopyByCommutingDef(const CoalescerPair &CP,
MachineInstr *CopyMI);
/// We found a copy which can be moved to its less frequent predecessor.
bool removePartialRedundancy(const CoalescerPair &CP, MachineInstr &CopyMI);
/// If the source of a copy is defined by a
/// trivial computation, replace the copy by rematerialize the definition.
bool reMaterializeTrivialDef(const CoalescerPair &CP, MachineInstr *CopyMI,
bool &IsDefCopy);
/// Return true if a copy involving a physreg should be joined.
bool canJoinPhys(const CoalescerPair &CP);
/// Replace all defs and uses of SrcReg to DstReg and update the subregister
/// number if it is not zero. If DstReg is a physical register and the
/// existing subregister number of the def / use being updated is not zero,
/// make sure to set it to the correct physical subregister.
void updateRegDefsUses(unsigned SrcReg, unsigned DstReg, unsigned SubIdx);
/// If the given machine operand reads only undefined lanes add an undef
/// flag.
/// This can happen when undef uses were previously concealed by a copy
/// which we coalesced. Example:
/// %0:sub0<def,read-undef> = ...
/// %1 = COPY %0 <-- Coalescing COPY reveals undef
/// = use %1:sub1 <-- hidden undef use
void addUndefFlag(const LiveInterval &Int, SlotIndex UseIdx,
MachineOperand &MO, unsigned SubRegIdx);
/// Handle copies of undef values. If the undef value is an incoming
/// PHI value, it will convert @p CopyMI to an IMPLICIT_DEF.
/// Returns nullptr if @p CopyMI was not in any way eliminable. Otherwise,
/// it returns @p CopyMI (which could be an IMPLICIT_DEF at this point).
MachineInstr *eliminateUndefCopy(MachineInstr *CopyMI);
/// Check whether or not we should apply the terminal rule on the
/// destination (Dst) of \p Copy.
/// When the terminal rule applies, Copy is not profitable to
/// coalesce.
/// Dst is terminal if it has exactly one affinity (Dst, Src) and
/// at least one interference (Dst, Dst2). If Dst is terminal, the
/// terminal rule consists in checking that at least one of
/// interfering node, say Dst2, has an affinity of equal or greater
/// weight with Src.
/// In that case, Dst2 and Dst will not be able to be both coalesced
/// with Src. Since Dst2 exposes more coalescing opportunities than
/// Dst, we can drop \p Copy.
bool applyTerminalRule(const MachineInstr &Copy) const;
/// Wrapper method for \see LiveIntervals::shrinkToUses.
/// This method does the proper fixing of the live-ranges when the afore
/// mentioned method returns true.
void shrinkToUses(LiveInterval *LI,
SmallVectorImpl<MachineInstr * > *Dead = nullptr) {
NumShrinkToUses++;
if (LIS->shrinkToUses(LI, Dead)) {
/// Check whether or not \p LI is composed by multiple connected
/// components and if that is the case, fix that.
SmallVector<LiveInterval*, 8> SplitLIs;
LIS->splitSeparateComponents(*LI, SplitLIs);
}
}
/// Wrapper Method to do all the necessary work when an Instruction is
/// deleted.
/// Optimizations should use this to make sure that deleted instructions
/// are always accounted for.
void deleteInstr(MachineInstr* MI) {
ErasedInstrs.insert(MI);
LIS->RemoveMachineInstrFromMaps(*MI);
MI->eraseFromParent();
}
/// Walk over function and initialize the DbgVRegToValues map.
void buildVRegToDbgValueMap(MachineFunction &MF);
/// Test whether, after merging, any DBG_VALUEs would refer to a
/// different value number than before merging, and whether this can
/// be resolved. If not, mark the DBG_VALUE as being undef.
void checkMergingChangesDbgValues(CoalescerPair &CP, LiveRange &LHS,
JoinVals &LHSVals, LiveRange &RHS,
JoinVals &RHSVals);
void checkMergingChangesDbgValuesImpl(unsigned Reg, LiveRange &OtherRange,
LiveRange &RegRange, JoinVals &Vals2);
public:
static char ID; ///< Class identification, replacement for typeinfo
RegisterCoalescer() : MachineFunctionPass(ID) {
initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
}
void getAnalysisUsage(AnalysisUsage &AU) const override;
void releaseMemory() override;
/// This is the pass entry point.
bool runOnMachineFunction(MachineFunction&) override;
/// Implement the dump method.
void print(raw_ostream &O, const Module* = nullptr) const override;
};
} // end anonymous namespace
char RegisterCoalescer::ID = 0;
char &llvm::RegisterCoalescerID = RegisterCoalescer::ID;
INITIALIZE_PASS_BEGIN(RegisterCoalescer, "simple-register-coalescing",
"Simple Register Coalescing", false, false)
INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
INITIALIZE_PASS_END(RegisterCoalescer, "simple-register-coalescing",
"Simple Register Coalescing", false, false)
LLVM_NODISCARD static bool isMoveInstr(const TargetRegisterInfo &tri,
const MachineInstr *MI, unsigned &Src,
unsigned &Dst, unsigned &SrcSub,
unsigned &DstSub) {
if (MI->isCopy()) {
Dst = MI->getOperand(0).getReg();
DstSub = MI->getOperand(0).getSubReg();
Src = MI->getOperand(1).getReg();
SrcSub = MI->getOperand(1).getSubReg();
} else if (MI->isSubregToReg()) {
Dst = MI->getOperand(0).getReg();
DstSub = tri.composeSubRegIndices(MI->getOperand(0).getSubReg(),
MI->getOperand(3).getImm());
Src = MI->getOperand(2).getReg();
SrcSub = MI->getOperand(2).getSubReg();
} else
return false;
return true;
}
/// Return true if this block should be vacated by the coalescer to eliminate
/// branches. The important cases to handle in the coalescer are critical edges
/// split during phi elimination which contain only copies. Simple blocks that
/// contain non-branches should also be vacated, but this can be handled by an
/// earlier pass similar to early if-conversion.
static bool isSplitEdge(const MachineBasicBlock *MBB) {
if (MBB->pred_size() != 1 || MBB->succ_size() != 1)
return false;
for (const auto &MI : *MBB) {
if (!MI.isCopyLike() && !MI.isUnconditionalBranch())
return false;
}
return true;
}
bool CoalescerPair::setRegisters(const MachineInstr *MI) {
SrcReg = DstReg = 0;
SrcIdx = DstIdx = 0;
NewRC = nullptr;
Flipped = CrossClass = false;
unsigned Src, Dst, SrcSub, DstSub;
if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
return false;
Partial = SrcSub || DstSub;
// If one register is a physreg, it must be Dst.
if (Register::isPhysicalRegister(Src)) {
if (Register::isPhysicalRegister(Dst))
return false;
std::swap(Src, Dst);
std::swap(SrcSub, DstSub);
Flipped = true;
}
const MachineRegisterInfo &MRI = MI->getMF()->getRegInfo();
if (Register::isPhysicalRegister(Dst)) {
// Eliminate DstSub on a physreg.
if (DstSub) {
Dst = TRI.getSubReg(Dst, DstSub);
if (!Dst) return false;
DstSub = 0;
}
// Eliminate SrcSub by picking a corresponding Dst superregister.
if (SrcSub) {
Dst = TRI.getMatchingSuperReg(Dst, SrcSub, MRI.getRegClass(Src));
if (!Dst) return false;
} else if (!MRI.getRegClass(Src)->contains(Dst)) {
return false;
}
} else {
// Both registers are virtual.
const TargetRegisterClass *SrcRC = MRI.getRegClass(Src);
const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
// Both registers have subreg indices.
if (SrcSub && DstSub) {
// Copies between different sub-registers are never coalescable.
if (Src == Dst && SrcSub != DstSub)
return false;
NewRC = TRI.getCommonSuperRegClass(SrcRC, SrcSub, DstRC, DstSub,
SrcIdx, DstIdx);
if (!NewRC)
return false;
} else if (DstSub) {
// SrcReg will be merged with a sub-register of DstReg.
SrcIdx = DstSub;
NewRC = TRI.getMatchingSuperRegClass(DstRC, SrcRC, DstSub);
} else if (SrcSub) {
// DstReg will be merged with a sub-register of SrcReg.
DstIdx = SrcSub;
NewRC = TRI.getMatchingSuperRegClass(SrcRC, DstRC, SrcSub);
} else {
// This is a straight copy without sub-registers.
NewRC = TRI.getCommonSubClass(DstRC, SrcRC);
}
// The combined constraint may be impossible to satisfy.
if (!NewRC)
return false;
// Prefer SrcReg to be a sub-register of DstReg.
// FIXME: Coalescer should support subregs symmetrically.
if (DstIdx && !SrcIdx) {
std::swap(Src, Dst);
std::swap(SrcIdx, DstIdx);
Flipped = !Flipped;
}
CrossClass = NewRC != DstRC || NewRC != SrcRC;
}
// Check our invariants
assert(Register::isVirtualRegister(Src) && "Src must be virtual");
assert(!(Register::isPhysicalRegister(Dst) && DstSub) &&
"Cannot have a physical SubIdx");
SrcReg = Src;
DstReg = Dst;
return true;
}
bool CoalescerPair::flip() {
if (Register::isPhysicalRegister(DstReg))
return false;
std::swap(SrcReg, DstReg);
std::swap(SrcIdx, DstIdx);
Flipped = !Flipped;
return true;
}
bool CoalescerPair::isCoalescable(const MachineInstr *MI) const {
if (!MI)
return false;
unsigned Src, Dst, SrcSub, DstSub;
if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
return false;
// Find the virtual register that is SrcReg.
if (Dst == SrcReg) {
std::swap(Src, Dst);
std::swap(SrcSub, DstSub);
} else if (Src != SrcReg) {
return false;
}
// Now check that Dst matches DstReg.
if (Register::isPhysicalRegister(DstReg)) {
if (!Register::isPhysicalRegister(Dst))
return false;
assert(!DstIdx && !SrcIdx && "Inconsistent CoalescerPair state.");
// DstSub could be set for a physreg from INSERT_SUBREG.
if (DstSub)
Dst = TRI.getSubReg(Dst, DstSub);
// Full copy of Src.
if (!SrcSub)
return DstReg == Dst;
// This is a partial register copy. Check that the parts match.
return TRI.getSubReg(DstReg, SrcSub) == Dst;
} else {
// DstReg is virtual.
if (DstReg != Dst)
return false;
// Registers match, do the subregisters line up?
return TRI.composeSubRegIndices(SrcIdx, SrcSub) ==
TRI.composeSubRegIndices(DstIdx, DstSub);
}
}
void RegisterCoalescer::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
AU.addRequired<AAResultsWrapperPass>();
AU.addRequired<LiveIntervals>();
AU.addPreserved<LiveIntervals>();
AU.addPreserved<SlotIndexes>();
AU.addRequired<MachineLoopInfo>();
AU.addPreserved<MachineLoopInfo>();
AU.addPreservedID(MachineDominatorsID);
MachineFunctionPass::getAnalysisUsage(AU);
}
void RegisterCoalescer::eliminateDeadDefs() {
SmallVector<unsigned, 8> NewRegs;
LiveRangeEdit(nullptr, NewRegs, *MF, *LIS,
nullptr, this).eliminateDeadDefs(DeadDefs);
}
void RegisterCoalescer::LRE_WillEraseInstruction(MachineInstr *MI) {
// MI may be in WorkList. Make sure we don't visit it.
ErasedInstrs.insert(MI);
}
bool RegisterCoalescer::adjustCopiesBackFrom(const CoalescerPair &CP,
MachineInstr *CopyMI) {
assert(!CP.isPartial() && "This doesn't work for partial copies.");
assert(!CP.isPhys() && "This doesn't work for physreg copies.");
LiveInterval &IntA =
LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
LiveInterval &IntB =
LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
SlotIndex CopyIdx = LIS->getInstructionIndex(*CopyMI).getRegSlot();
// We have a non-trivially-coalescable copy with IntA being the source and
// IntB being the dest, thus this defines a value number in IntB. If the
// source value number (in IntA) is defined by a copy from B, see if we can
// merge these two pieces of B into a single value number, eliminating a copy.
// For example:
//
// A3 = B0
// ...
// B1 = A3 <- this copy
//
// In this case, B0 can be extended to where the B1 copy lives, allowing the
// B1 value number to be replaced with B0 (which simplifies the B
// liveinterval).
// BValNo is a value number in B that is defined by a copy from A. 'B1' in
// the example above.
LiveInterval::iterator BS = IntB.FindSegmentContaining(CopyIdx);
if (BS == IntB.end()) return false;
VNInfo *BValNo = BS->valno;
// Get the location that B is defined at. Two options: either this value has
// an unknown definition point or it is defined at CopyIdx. If unknown, we
// can't process it.
if (BValNo->def != CopyIdx) return false;
// AValNo is the value number in A that defines the copy, A3 in the example.
SlotIndex CopyUseIdx = CopyIdx.getRegSlot(true);
LiveInterval::iterator AS = IntA.FindSegmentContaining(CopyUseIdx);
// The live segment might not exist after fun with physreg coalescing.
if (AS == IntA.end()) return false;
VNInfo *AValNo = AS->valno;
// If AValNo is defined as a copy from IntB, we can potentially process this.
// Get the instruction that defines this value number.
MachineInstr *ACopyMI = LIS->getInstructionFromIndex(AValNo->def);
// Don't allow any partial copies, even if isCoalescable() allows them.
if (!CP.isCoalescable(ACopyMI) || !ACopyMI->isFullCopy())
return false;
// Get the Segment in IntB that this value number starts with.
LiveInterval::iterator ValS =
IntB.FindSegmentContaining(AValNo->def.getPrevSlot());
if (ValS == IntB.end())
return false;
// Make sure that the end of the live segment is inside the same block as
// CopyMI.
MachineInstr *ValSEndInst =
LIS->getInstructionFromIndex(ValS->end.getPrevSlot());
if (!ValSEndInst || ValSEndInst->getParent() != CopyMI->getParent())
return false;
// Okay, we now know that ValS ends in the same block that the CopyMI
// live-range starts. If there are no intervening live segments between them
// in IntB, we can merge them.
if (ValS+1 != BS) return false;
LLVM_DEBUG(dbgs() << "Extending: " << printReg(IntB.reg, TRI));
SlotIndex FillerStart = ValS->end, FillerEnd = BS->start;
// We are about to delete CopyMI, so need to remove it as the 'instruction
// that defines this value #'. Update the valnum with the new defining
// instruction #.
BValNo->def = FillerStart;
// Okay, we can merge them. We need to insert a new liverange:
// [ValS.end, BS.begin) of either value number, then we merge the
// two value numbers.
IntB.addSegment(LiveInterval::Segment(FillerStart, FillerEnd, BValNo));
// Okay, merge "B1" into the same value number as "B0".
if (BValNo != ValS->valno)
IntB.MergeValueNumberInto(BValNo, ValS->valno);
// Do the same for the subregister segments.
for (LiveInterval::SubRange &S : IntB.subranges()) {
// Check for SubRange Segments of the form [1234r,1234d:0) which can be
// removed to prevent creating bogus SubRange Segments.
LiveInterval::iterator SS = S.FindSegmentContaining(CopyIdx);
if (SS != S.end() && SlotIndex::isSameInstr(SS->start, SS->end)) {
S.removeSegment(*SS, true);
continue;
}
VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
S.addSegment(LiveInterval::Segment(FillerStart, FillerEnd, SubBValNo));
VNInfo *SubValSNo = S.getVNInfoAt(AValNo->def.getPrevSlot());
if (SubBValNo != SubValSNo)
S.MergeValueNumberInto(SubBValNo, SubValSNo);
}
LLVM_DEBUG(dbgs() << " result = " << IntB << '\n');
// If the source instruction was killing the source register before the
// merge, unset the isKill marker given the live range has been extended.
int UIdx = ValSEndInst->findRegisterUseOperandIdx(IntB.reg, true);
if (UIdx != -1) {
ValSEndInst->getOperand(UIdx).setIsKill(false);
}
// Rewrite the copy.
CopyMI->substituteRegister(IntA.reg, IntB.reg, 0, *TRI);
// If the copy instruction was killing the destination register or any
// subrange before the merge trim the live range.
bool RecomputeLiveRange = AS->end == CopyIdx;
if (!RecomputeLiveRange) {
for (LiveInterval::SubRange &S : IntA.subranges()) {
LiveInterval::iterator SS = S.FindSegmentContaining(CopyUseIdx);
if (SS != S.end() && SS->end == CopyIdx) {
RecomputeLiveRange = true;
break;
}
}
}
if (RecomputeLiveRange)
shrinkToUses(&IntA);
++numExtends;
return true;
}
bool RegisterCoalescer::hasOtherReachingDefs(LiveInterval &IntA,
LiveInterval &IntB,
VNInfo *AValNo,
VNInfo *BValNo) {
// If AValNo has PHI kills, conservatively assume that IntB defs can reach
// the PHI values.
if (LIS->hasPHIKill(IntA, AValNo))
return true;
for (LiveRange::Segment &ASeg : IntA.segments) {
if (ASeg.valno != AValNo) continue;
LiveInterval::iterator BI = llvm::upper_bound(IntB, ASeg.start);
if (BI != IntB.begin())
--BI;
for (; BI != IntB.end() && ASeg.end >= BI->start; ++BI) {
if (BI->valno == BValNo)
continue;
if (BI->start <= ASeg.start && BI->end > ASeg.start)
return true;
if (BI->start > ASeg.start && BI->start < ASeg.end)
return true;
}
}
return false;
}
/// Copy segments with value number @p SrcValNo from liverange @p Src to live
/// range @Dst and use value number @p DstValNo there.
static std::pair<bool,bool>
addSegmentsWithValNo(LiveRange &Dst, VNInfo *DstValNo, const LiveRange &Src,
const VNInfo *SrcValNo) {
bool Changed = false;
bool MergedWithDead = false;
for (const LiveRange::Segment &S : Src.segments) {
if (S.valno != SrcValNo)
continue;
// This is adding a segment from Src that ends in a copy that is about
// to be removed. This segment is going to be merged with a pre-existing
// segment in Dst. This works, except in cases when the corresponding
// segment in Dst is dead. For example: adding [192r,208r:1) from Src
// to [208r,208d:1) in Dst would create [192r,208d:1) in Dst.
// Recognized such cases, so that the segments can be shrunk.
LiveRange::Segment Added = LiveRange::Segment(S.start, S.end, DstValNo);
LiveRange::Segment &Merged = *Dst.addSegment(Added);
if (Merged.end.isDead())
MergedWithDead = true;
Changed = true;
}
return std::make_pair(Changed, MergedWithDead);
}
std::pair<bool,bool>
RegisterCoalescer::removeCopyByCommutingDef(const CoalescerPair &CP,
MachineInstr *CopyMI) {
assert(!CP.isPhys());
LiveInterval &IntA =
LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
LiveInterval &IntB =
LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
// We found a non-trivially-coalescable copy with IntA being the source and
// IntB being the dest, thus this defines a value number in IntB. If the
// source value number (in IntA) is defined by a commutable instruction and
// its other operand is coalesced to the copy dest register, see if we can
// transform the copy into a noop by commuting the definition. For example,
//
// A3 = op A2 killed B0
// ...
// B1 = A3 <- this copy
// ...
// = op A3 <- more uses
//
// ==>
//
// B2 = op B0 killed A2
// ...
// B1 = B2 <- now an identity copy
// ...
// = op B2 <- more uses
// BValNo is a value number in B that is defined by a copy from A. 'B1' in
// the example above.
SlotIndex CopyIdx = LIS->getInstructionIndex(*CopyMI).getRegSlot();
VNInfo *BValNo = IntB.getVNInfoAt(CopyIdx);
assert(BValNo != nullptr && BValNo->def == CopyIdx);
// AValNo is the value number in A that defines the copy, A3 in the example.
VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx.getRegSlot(true));
assert(AValNo && !AValNo->isUnused() && "COPY source not live");
if (AValNo->isPHIDef())
return { false, false };
MachineInstr *DefMI = LIS->getInstructionFromIndex(AValNo->def);
if (!DefMI)
return { false, false };
if (!DefMI->isCommutable())
return { false, false };
// If DefMI is a two-address instruction then commuting it will change the
// destination register.
int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
assert(DefIdx != -1);
unsigned UseOpIdx;
if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
return { false, false };
// FIXME: The code below tries to commute 'UseOpIdx' operand with some other
// commutable operand which is expressed by 'CommuteAnyOperandIndex'value
// passed to the method. That _other_ operand is chosen by
// the findCommutedOpIndices() method.
//
// That is obviously an area for improvement in case of instructions having
// more than 2 operands. For example, if some instruction has 3 commutable
// operands then all possible variants (i.e. op#1<->op#2, op#1<->op#3,
// op#2<->op#3) of commute transformation should be considered/tried here.
unsigned NewDstIdx = TargetInstrInfo::CommuteAnyOperandIndex;
if (!TII->findCommutedOpIndices(*DefMI, UseOpIdx, NewDstIdx))
return { false, false };
MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
Register NewReg = NewDstMO.getReg();
if (NewReg != IntB.reg || !IntB.Query(AValNo->def).isKill())
return { false, false };
// Make sure there are no other definitions of IntB that would reach the
// uses which the new definition can reach.
if (hasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
return { false, false };
// If some of the uses of IntA.reg is already coalesced away, return false.
// It's not possible to determine whether it's safe to perform the coalescing.
for (MachineOperand &MO : MRI->use_nodbg_operands(IntA.reg)) {
MachineInstr *UseMI = MO.getParent();
unsigned OpNo = &MO - &UseMI->getOperand(0);
SlotIndex UseIdx = LIS->getInstructionIndex(*UseMI);
LiveInterval::iterator US = IntA.FindSegmentContaining(UseIdx);
if (US == IntA.end() || US->valno != AValNo)
continue;
// If this use is tied to a def, we can't rewrite the register.
if (UseMI->isRegTiedToDefOperand(OpNo))
return { false, false };
}
LLVM_DEBUG(dbgs() << "\tremoveCopyByCommutingDef: " << AValNo->def << '\t'
<< *DefMI);
// At this point we have decided that it is legal to do this
// transformation. Start by commuting the instruction.
MachineBasicBlock *MBB = DefMI->getParent();
MachineInstr *NewMI =
TII->commuteInstruction(*DefMI, false, UseOpIdx, NewDstIdx);
if (!NewMI)
return { false, false };
if (Register::isVirtualRegister(IntA.reg) &&
Register::isVirtualRegister(IntB.reg) &&
!MRI->constrainRegClass(IntB.reg, MRI->getRegClass(IntA.reg)))
return { false, false };
if (NewMI != DefMI) {
LIS->ReplaceMachineInstrInMaps(*DefMI, *NewMI);
MachineBasicBlock::iterator Pos = DefMI;
MBB->insert(Pos, NewMI);
MBB->erase(DefMI);
}
// If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
// A = or A, B
// ...
// B = A
// ...
// C = killed A
// ...
// = B
// Update uses of IntA of the specific Val# with IntB.
for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(IntA.reg),
UE = MRI->use_end();
UI != UE; /* ++UI is below because of possible MI removal */) {
MachineOperand &UseMO = *UI;
++UI;
if (UseMO.isUndef())
continue;
MachineInstr *UseMI = UseMO.getParent();
if (UseMI->isDebugValue()) {
// FIXME These don't have an instruction index. Not clear we have enough
// info to decide whether to do this replacement or not. For now do it.
UseMO.setReg(NewReg);
continue;
}
SlotIndex UseIdx = LIS->getInstructionIndex(*UseMI).getRegSlot(true);
LiveInterval::iterator US = IntA.FindSegmentContaining(UseIdx);
assert(US != IntA.end() && "Use must be live");
if (US->valno != AValNo)
continue;
// Kill flags are no longer accurate. They are recomputed after RA.
UseMO.setIsKill(false);
if (Register::isPhysicalRegister(NewReg))
UseMO.substPhysReg(NewReg, *TRI);
else
UseMO.setReg(NewReg);
if (UseMI == CopyMI)
continue;
if (!UseMI->isCopy())
continue;
if (UseMI->getOperand(0).getReg() != IntB.reg ||
UseMI->getOperand(0).getSubReg())
continue;
// This copy will become a noop. If it's defining a new val#, merge it into
// BValNo.
SlotIndex DefIdx = UseIdx.getRegSlot();
VNInfo *DVNI = IntB.getVNInfoAt(DefIdx);
if (!DVNI)
continue;
LLVM_DEBUG(dbgs() << "\t\tnoop: " << DefIdx << '\t' << *UseMI);
assert(DVNI->def == DefIdx);
BValNo = IntB.MergeValueNumberInto(DVNI, BValNo);
for (LiveInterval::SubRange &S : IntB.subranges()) {
VNInfo *SubDVNI = S.getVNInfoAt(DefIdx);
if (!SubDVNI)
continue;
VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
assert(SubBValNo->def == CopyIdx);
S.MergeValueNumberInto(SubDVNI, SubBValNo);
}
deleteInstr(UseMI);
}
// Extend BValNo by merging in IntA live segments of AValNo. Val# definition
// is updated.
bool ShrinkB = false;
BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
if (IntA.hasSubRanges() || IntB.hasSubRanges()) {
if (!IntA.hasSubRanges()) {
LaneBitmask Mask = MRI->getMaxLaneMaskForVReg(IntA.reg);
IntA.createSubRangeFrom(Allocator, Mask, IntA);
} else if (!IntB.hasSubRanges()) {
LaneBitmask Mask = MRI->getMaxLaneMaskForVReg(IntB.reg);
IntB.createSubRangeFrom(Allocator, Mask, IntB);
}
SlotIndex AIdx = CopyIdx.getRegSlot(true);
LaneBitmask MaskA;
const SlotIndexes &Indexes = *LIS->getSlotIndexes();
for (LiveInterval::SubRange &SA : IntA.subranges()) {
VNInfo *ASubValNo = SA.getVNInfoAt(AIdx);
// Even if we are dealing with a full copy, some lanes can
// still be undefined.
// E.g.,
// undef A.subLow = ...
// B = COPY A <== A.subHigh is undefined here and does
// not have a value number.
if (!ASubValNo)
continue;
MaskA |= SA.LaneMask;
IntB.refineSubRanges(
Allocator, SA.LaneMask,
[&Allocator, &SA, CopyIdx, ASubValNo,
&ShrinkB](LiveInterval::SubRange &SR) {
VNInfo *BSubValNo = SR.empty() ? SR.getNextValue(CopyIdx, Allocator)
: SR.getVNInfoAt(CopyIdx);
assert(BSubValNo != nullptr);
auto P = addSegmentsWithValNo(SR, BSubValNo, SA, ASubValNo);
ShrinkB |= P.second;
if (P.first)
BSubValNo->def = ASubValNo->def;
},
Indexes, *TRI);
}
// Go over all subranges of IntB that have not been covered by IntA,
// and delete the segments starting at CopyIdx. This can happen if
// IntA has undef lanes that are defined in IntB.
for (LiveInterval::SubRange &SB : IntB.subranges()) {
if ((SB.LaneMask & MaskA).any())
continue;
if (LiveRange::Segment *S = SB.getSegmentContaining(CopyIdx))
if (S->start.getBaseIndex() == CopyIdx.getBaseIndex())
SB.removeSegment(*S, true);
}
}
BValNo->def = AValNo->def;
auto P = addSegmentsWithValNo(IntB, BValNo, IntA, AValNo);
ShrinkB |= P.second;
LLVM_DEBUG(dbgs() << "\t\textended: " << IntB << '\n');
LIS->removeVRegDefAt(IntA, AValNo->def);
LLVM_DEBUG(dbgs() << "\t\ttrimmed: " << IntA << '\n');
++numCommutes;
return { true, ShrinkB };
}
/// For copy B = A in BB2, if A is defined by A = B in BB0 which is a
/// predecessor of BB2, and if B is not redefined on the way from A = B
/// in BB0 to B = A in BB2, B = A in BB2 is partially redundant if the
/// execution goes through the path from BB0 to BB2. We may move B = A
/// to the predecessor without such reversed copy.
/// So we will transform the program from:
/// BB0:
/// A = B; BB1:
/// ... ...
/// / \ /
/// BB2:
/// ...
/// B = A;
///
/// to:
///
/// BB0: BB1:
/// A = B; ...
/// ... B = A;
/// / \ /
/// BB2:
/// ...
///
/// A special case is when BB0 and BB2 are the same BB which is the only
/// BB in a loop:
/// BB1:
/// ...
/// BB0/BB2: ----
/// B = A; |
/// ... |
/// A = B; |
/// |-------
/// |
/// We may hoist B = A from BB0/BB2 to BB1.
///
/// The major preconditions for correctness to remove such partial
/// redundancy include:
/// 1. A in B = A in BB2 is defined by a PHI in BB2, and one operand of
/// the PHI is defined by the reversed copy A = B in BB0.
/// 2. No B is referenced from the start of BB2 to B = A.
/// 3. No B is defined from A = B to the end of BB0.
/// 4. BB1 has only one successor.
///
/// 2 and 4 implicitly ensure B is not live at the end of BB1.
/// 4 guarantees BB2 is hotter than BB1, so we can only move a copy to a
/// colder place, which not only prevent endless loop, but also make sure
/// the movement of copy is beneficial.
bool RegisterCoalescer::removePartialRedundancy(const CoalescerPair &CP,
MachineInstr &CopyMI) {
assert(!CP.isPhys());
if (!CopyMI.isFullCopy())
return false;
MachineBasicBlock &MBB = *CopyMI.getParent();
if (MBB.isEHPad())
return false;
if (MBB.pred_size() != 2)
return false;
LiveInterval &IntA =
LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
LiveInterval &IntB =
LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
// A is defined by PHI at the entry of MBB.
SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot(true);
VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx);
assert(AValNo && !AValNo->isUnused() && "COPY source not live");
if (!AValNo->isPHIDef())
return false;
// No B is referenced before CopyMI in MBB.
if (IntB.overlaps(LIS->getMBBStartIdx(&MBB), CopyIdx))
return false;
// MBB has two predecessors: one contains A = B so no copy will be inserted
// for it. The other one will have a copy moved from MBB.
bool FoundReverseCopy = false;
MachineBasicBlock *CopyLeftBB = nullptr;
for (MachineBasicBlock *Pred : MBB.predecessors()) {
VNInfo *PVal = IntA.getVNInfoBefore(LIS->getMBBEndIdx(Pred));
MachineInstr *DefMI = LIS->getInstructionFromIndex(PVal->def);
if (!DefMI || !DefMI->isFullCopy()) {
CopyLeftBB = Pred;
continue;
}
// Check DefMI is a reverse copy and it is in BB Pred.
if (DefMI->getOperand(0).getReg() != IntA.reg ||
DefMI->getOperand(1).getReg() != IntB.reg ||
DefMI->getParent() != Pred) {
CopyLeftBB = Pred;
continue;
}
// If there is any other def of B after DefMI and before the end of Pred,
// we need to keep the copy of B = A at the end of Pred if we remove
// B = A from MBB.
bool ValB_Changed = false;
for (auto VNI : IntB.valnos) {
if (VNI->isUnused())
continue;
if (PVal->def < VNI->def && VNI->def < LIS->getMBBEndIdx(Pred)) {
ValB_Changed = true;
break;
}
}
if (ValB_Changed) {
CopyLeftBB = Pred;
continue;
}
FoundReverseCopy = true;
}
// If no reverse copy is found in predecessors, nothing to do.
if (!FoundReverseCopy)
return false;
// If CopyLeftBB is nullptr, it means every predecessor of MBB contains
// reverse copy, CopyMI can be removed trivially if only IntA/IntB is updated.
// If CopyLeftBB is not nullptr, move CopyMI from MBB to CopyLeftBB and
// update IntA/IntB.
//
// If CopyLeftBB is not nullptr, ensure CopyLeftBB has a single succ so
// MBB is hotter than CopyLeftBB.
if (CopyLeftBB && CopyLeftBB->succ_size() > 1)
return false;
// Now (almost sure it's) ok to move copy.
if (CopyLeftBB) {
// Position in CopyLeftBB where we should insert new copy.
auto InsPos = CopyLeftBB->getFirstTerminator();
// Make sure that B isn't referenced in the terminators (if any) at the end
// of the predecessor since we're about to insert a new definition of B
// before them.
if (InsPos != CopyLeftBB->end()) {
SlotIndex InsPosIdx = LIS->getInstructionIndex(*InsPos).getRegSlot(true);
if (IntB.overlaps(InsPosIdx, LIS->getMBBEndIdx(CopyLeftBB)))
return false;
}
LLVM_DEBUG(dbgs() << "\tremovePartialRedundancy: Move the copy to "
<< printMBBReference(*CopyLeftBB) << '\t' << CopyMI);
// Insert new copy to CopyLeftBB.
MachineInstr *NewCopyMI = BuildMI(*CopyLeftBB, InsPos, CopyMI.getDebugLoc(),
TII->get(TargetOpcode::COPY), IntB.reg)
.addReg(IntA.reg);
SlotIndex NewCopyIdx =
LIS->InsertMachineInstrInMaps(*NewCopyMI).getRegSlot();
IntB.createDeadDef(NewCopyIdx, LIS->getVNInfoAllocator());
for (LiveInterval::SubRange &SR : IntB.subranges())
SR.createDeadDef(NewCopyIdx, LIS->getVNInfoAllocator());
// If the newly created Instruction has an address of an instruction that was
// deleted before (object recycled by the allocator) it needs to be removed from
// the deleted list.
ErasedInstrs.erase(NewCopyMI);
} else {
LLVM_DEBUG(dbgs() << "\tremovePartialRedundancy: Remove the copy from "
<< printMBBReference(MBB) << '\t' << CopyMI);
}
// Remove CopyMI.
// Note: This is fine to remove the copy before updating the live-ranges.
// While updating the live-ranges, we only look at slot indices and
// never go back to the instruction.
// Mark instructions as deleted.
deleteInstr(&CopyMI);
// Update the liveness.
SmallVector<SlotIndex, 8> EndPoints;
VNInfo *BValNo = IntB.Query(CopyIdx).valueOutOrDead();
LIS->pruneValue(*static_cast<LiveRange *>(&IntB), CopyIdx.getRegSlot(),
&EndPoints);
BValNo->markUnused();
// Extend IntB to the EndPoints of its original live interval.
LIS->extendToIndices(IntB, EndPoints);
// Now, do the same for its subranges.
for (LiveInterval::SubRange &SR : IntB.subranges()) {
EndPoints.clear();
VNInfo *BValNo = SR.Query(CopyIdx).valueOutOrDead();
assert(BValNo && "All sublanes should be live");
LIS->pruneValue(SR, CopyIdx.getRegSlot(), &EndPoints);
BValNo->markUnused();
// We can have a situation where the result of the original copy is live,
// but is immediately dead in this subrange, e.g. [336r,336d:0). That makes
// the copy appear as an endpoint from pruneValue(), but we don't want it
// to because the copy has been removed. We can go ahead and remove that
// endpoint; there is no other situation here that there could be a use at
// the same place as we know that the copy is a full copy.
for (unsigned I = 0; I != EndPoints.size(); ) {
if (SlotIndex::isSameInstr(EndPoints[I], CopyIdx)) {
EndPoints[I] = EndPoints.back();
EndPoints.pop_back();
continue;
}
++I;
}
LIS->extendToIndices(SR, EndPoints);
}
// If any dead defs were extended, truncate them.
shrinkToUses(&IntB);
// Finally, update the live-range of IntA.
shrinkToUses(&IntA);
return true;
}
/// Returns true if @p MI defines the full vreg @p Reg, as opposed to just
/// defining a subregister.
static bool definesFullReg(const MachineInstr &MI, unsigned Reg) {
assert(!Register::isPhysicalRegister(Reg) &&
"This code cannot handle physreg aliasing");
for (const MachineOperand &Op : MI.operands()) {
if (!Op.isReg() || !Op.isDef() || Op.getReg() != Reg)
continue;
// Return true if we define the full register or don't care about the value
// inside other subregisters.
if (Op.getSubReg() == 0 || Op.isUndef())
return true;
}
return false;
}
bool RegisterCoalescer::reMaterializeTrivialDef(const CoalescerPair &CP,
MachineInstr *CopyMI,
bool &IsDefCopy) {
IsDefCopy = false;
unsigned SrcReg = CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg();
unsigned SrcIdx = CP.isFlipped() ? CP.getDstIdx() : CP.getSrcIdx();
unsigned DstReg = CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg();
unsigned DstIdx = CP.isFlipped() ? CP.getSrcIdx() : CP.getDstIdx();
if (Register::isPhysicalRegister(SrcReg))
return false;
LiveInterval &SrcInt = LIS->getInterval(SrcReg);
SlotIndex CopyIdx = LIS->getInstructionIndex(*CopyMI);
VNInfo *ValNo = SrcInt.Query(CopyIdx).valueIn();
if (!ValNo)
return false;
if (ValNo->isPHIDef() || ValNo->isUnused())
return false;
MachineInstr *DefMI = LIS->getInstructionFromIndex(ValNo->def);
if (!DefMI)
return false;
if (DefMI->isCopyLike()) {
IsDefCopy = true;
return false;
}
if (!TII->isAsCheapAsAMove(*DefMI))
return false;
if (!TII->isTriviallyReMaterializable(*DefMI, AA))
return false;
if (!definesFullReg(*DefMI, SrcReg))
return false;
bool SawStore = false;
if (!DefMI->isSafeToMove(AA, SawStore))
return false;
const MCInstrDesc &MCID = DefMI->getDesc();
if (MCID.getNumDefs() != 1)
return false;
// Only support subregister destinations when the def is read-undef.
MachineOperand &DstOperand = CopyMI->getOperand(0);
Register CopyDstReg = DstOperand.getReg();
if (DstOperand.getSubReg() && !DstOperand.isUndef())
return false;
// If both SrcIdx and DstIdx are set, correct rematerialization would widen
// the register substantially (beyond both source and dest size). This is bad
// for performance since it can cascade through a function, introducing many
// extra spills and fills (e.g. ARM can easily end up copying QQQQPR registers
// around after a few subreg copies).
if (SrcIdx && DstIdx)
return false;
const TargetRegisterClass *DefRC = TII->getRegClass(MCID, 0, TRI, *MF);
if (!DefMI->isImplicitDef()) {
if (Register::isPhysicalRegister(DstReg)) {
unsigned NewDstReg = DstReg;
unsigned NewDstIdx = TRI->composeSubRegIndices(CP.getSrcIdx(),
DefMI->getOperand(0).getSubReg());
if (NewDstIdx)
NewDstReg = TRI->getSubReg(DstReg, NewDstIdx);
// Finally, make sure that the physical subregister that will be
// constructed later is permitted for the instruction.
if (!DefRC->contains(NewDstReg))
return false;
} else {
// Theoretically, some stack frame reference could exist. Just make sure
// it hasn't actually happened.
assert(Register::isVirtualRegister(DstReg) &&
"Only expect to deal with virtual or physical registers");
}
}
DebugLoc DL = CopyMI->getDebugLoc();
MachineBasicBlock *MBB = CopyMI->getParent();
MachineBasicBlock::iterator MII =
std::next(MachineBasicBlock::iterator(CopyMI));
TII->reMaterialize(*MBB, MII, DstReg, SrcIdx, *DefMI, *TRI);
MachineInstr &NewMI = *std::prev(MII);
NewMI.setDebugLoc(DL);
// In a situation like the following:
// %0:subreg = instr ; DefMI, subreg = DstIdx
// %1 = copy %0:subreg ; CopyMI, SrcIdx = 0
// instead of widening %1 to the register class of %0 simply do:
// %1 = instr
const TargetRegisterClass *NewRC = CP.getNewRC();
if (DstIdx != 0) {
MachineOperand &DefMO = NewMI.getOperand(0);
if (DefMO.getSubReg() == DstIdx) {
assert(SrcIdx == 0 && CP.isFlipped()
&& "Shouldn't have SrcIdx+DstIdx at this point");
const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
const TargetRegisterClass *CommonRC =
TRI->getCommonSubClass(DefRC, DstRC);
if (CommonRC != nullptr) {
NewRC = CommonRC;
DstIdx = 0;
DefMO.setSubReg(0);
DefMO.setIsUndef(false); // Only subregs can have def+undef.
}
}
}
// CopyMI may have implicit operands, save them so that we can transfer them
// over to the newly materialized instruction after CopyMI is removed.
SmallVector<MachineOperand, 4> ImplicitOps;
ImplicitOps.reserve(CopyMI->getNumOperands() -
CopyMI->getDesc().getNumOperands());
for (unsigned I = CopyMI->getDesc().getNumOperands(),
E = CopyMI->getNumOperands();
I != E; ++I) {
MachineOperand &MO = CopyMI->getOperand(I);
if (MO.isReg()) {
assert(MO.isImplicit() && "No explicit operands after implicit operands.");
// Discard VReg implicit defs.
if (Register::isPhysicalRegister(MO.getReg()))
ImplicitOps.push_back(MO);
}
}
LIS->ReplaceMachineInstrInMaps(*CopyMI, NewMI);
CopyMI->eraseFromParent();
ErasedInstrs.insert(CopyMI);
// NewMI may have dead implicit defs (E.g. EFLAGS for MOV<bits>r0 on X86).
// We need to remember these so we can add intervals once we insert
// NewMI into SlotIndexes.
SmallVector<unsigned, 4> NewMIImplDefs;
for (unsigned i = NewMI.getDesc().getNumOperands(),
e = NewMI.getNumOperands();
i != e; ++i) {
MachineOperand &MO = NewMI.getOperand(i);
if (MO.isReg() && MO.isDef()) {
assert(MO.isImplicit() && MO.isDead() &&
Register::isPhysicalRegister(MO.getReg()));
NewMIImplDefs.push_back(MO.getReg());
}
}
if (Register::isVirtualRegister(DstReg)) {
unsigned NewIdx = NewMI.getOperand(0).getSubReg();
if (DefRC != nullptr) {
if (NewIdx)
NewRC = TRI->getMatchingSuperRegClass(NewRC, DefRC, NewIdx);
else
NewRC = TRI->getCommonSubClass(NewRC, DefRC);
assert(NewRC && "subreg chosen for remat incompatible with instruction");
}
// Remap subranges to new lanemask and change register class.
LiveInterval &DstInt = LIS->getInterval(DstReg);
for (LiveInterval::SubRange &SR : DstInt.subranges()) {
SR.LaneMask = TRI->composeSubRegIndexLaneMask(DstIdx, SR.LaneMask);
}
MRI->setRegClass(DstReg, NewRC);
// Update machine operands and add flags.
updateRegDefsUses(DstReg, DstReg, DstIdx);
NewMI.getOperand(0).setSubReg(NewIdx);
// updateRegDefUses can add an "undef" flag to the definition, since
// it will replace DstReg with DstReg.DstIdx. If NewIdx is 0, make
// sure that "undef" is not set.
if (NewIdx == 0)
NewMI.getOperand(0).setIsUndef(false);
// Add dead subregister definitions if we are defining the whole register
// but only part of it is live.
// This could happen if the rematerialization instruction is rematerializing
// more than actually is used in the register.
// An example would be:
// %1 = LOAD CONSTANTS 5, 8 ; Loading both 5 and 8 in different subregs
// ; Copying only part of the register here, but the rest is undef.
// %2:sub_16bit<def, read-undef> = COPY %1:sub_16bit
// ==>
// ; Materialize all the constants but only using one
// %2 = LOAD_CONSTANTS 5, 8
//
// at this point for the part that wasn't defined before we could have
// subranges missing the definition.
if (NewIdx == 0 && DstInt.hasSubRanges()) {
SlotIndex CurrIdx = LIS->getInstructionIndex(NewMI);
SlotIndex DefIndex =
CurrIdx.getRegSlot(NewMI.getOperand(0).isEarlyClobber());
LaneBitmask MaxMask = MRI->getMaxLaneMaskForVReg(DstReg);
VNInfo::Allocator& Alloc = LIS->getVNInfoAllocator();
for (LiveInterval::SubRange &SR : DstInt.subranges()) {
if (!SR.liveAt(DefIndex))
SR.createDeadDef(DefIndex, Alloc);
MaxMask &= ~SR.LaneMask;
}
if (MaxMask.any()) {
LiveInterval::SubRange *SR = DstInt.createSubRange(Alloc, MaxMask);
SR->createDeadDef(DefIndex, Alloc);
}
}
// Make sure that the subrange for resultant undef is removed
// For example:
// %1:sub1<def,read-undef> = LOAD CONSTANT 1
// %2 = COPY %1
// ==>
// %2:sub1<def, read-undef> = LOAD CONSTANT 1
// ; Correct but need to remove the subrange for %2:sub0
// ; as it is now undef
if (NewIdx != 0 && DstInt.hasSubRanges()) {
// The affected subregister segments can be removed.
SlotIndex CurrIdx = LIS->getInstructionIndex(NewMI);
LaneBitmask DstMask = TRI->getSubRegIndexLaneMask(NewIdx);
bool UpdatedSubRanges = false;
for (LiveInterval::SubRange &SR : DstInt.subranges()) {
if ((SR.LaneMask & DstMask).none()) {
LLVM_DEBUG(dbgs()
<< "Removing undefined SubRange "
<< PrintLaneMask(SR.LaneMask) << " : " << SR << "\n");
// VNI is in ValNo - remove any segments in this SubRange that have this ValNo
if (VNInfo *RmValNo = SR.getVNInfoAt(CurrIdx.getRegSlot())) {
SR.removeValNo(RmValNo);
UpdatedSubRanges = true;
}
}
}
if (UpdatedSubRanges)
DstInt.removeEmptySubRanges();
}
} else if (NewMI.getOperand(0).getReg() != CopyDstReg) {
// The New instruction may be defining a sub-register of what's actually
// been asked for. If so it must implicitly define the whole thing.
assert(Register::isPhysicalRegister(DstReg) &&
"Only expect virtual or physical registers in remat");
NewMI.getOperand(0).setIsDead(true);
NewMI.addOperand(MachineOperand::CreateReg(
CopyDstReg, true /*IsDef*/, true /*IsImp*/, false /*IsKill*/));
// Record small dead def live-ranges for all the subregisters
// of the destination register.
// Otherwise, variables that live through may miss some
// interferences, thus creating invalid allocation.
// E.g., i386 code:
// %1 = somedef ; %1 GR8
// %2 = remat ; %2 GR32
// CL = COPY %2.sub_8bit
// = somedef %1 ; %1 GR8
// =>
// %1 = somedef ; %1 GR8
// dead ECX = remat ; implicit-def CL
// = somedef %1 ; %1 GR8
// %1 will see the interferences with CL but not with CH since
// no live-ranges would have been created for ECX.
// Fix that!
SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
for (MCRegUnitIterator Units(NewMI.getOperand(0).getReg(), TRI);
Units.isValid(); ++Units)
if (LiveRange *LR = LIS->getCachedRegUnit(*Units))
LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
}
if (NewMI.getOperand(0).getSubReg())
NewMI.getOperand(0).setIsUndef();
// Transfer over implicit operands to the rematerialized instruction.
for (MachineOperand &MO : ImplicitOps)
NewMI.addOperand(MO);
SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
for (unsigned i = 0, e = NewMIImplDefs.size(); i != e; ++i) {
unsigned Reg = NewMIImplDefs[i];
for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units)
if (LiveRange *LR = LIS->getCachedRegUnit(*Units))
LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
}
LLVM_DEBUG(dbgs() << "Remat: " << NewMI);
++NumReMats;
// If the virtual SrcReg is completely eliminated, update all DBG_VALUEs
// to describe DstReg instead.
if (MRI->use_nodbg_empty(SrcReg)) {
for (MachineOperand &UseMO : MRI->use_operands(SrcReg)) {
MachineInstr *UseMI = UseMO.getParent();
if (UseMI->isDebugValue()) {
if (Register::isPhysicalRegister(DstReg))
UseMO.substPhysReg(DstReg, *TRI);
else
UseMO.setReg(DstReg);
// Move the debug value directly after the def of the rematerialized
// value in DstReg.
MBB->splice(std::next(NewMI.getIterator()), UseMI->getParent(), UseMI);
LLVM_DEBUG(dbgs() << "\t\tupdated: " << *UseMI);
}
}
}
if (ToBeUpdated.count(SrcReg))
return true;
unsigned NumCopyUses = 0;
for (MachineOperand &UseMO : MRI->use_nodbg_operands(SrcReg)) {
if (UseMO.getParent()->isCopyLike())
NumCopyUses++;
}
if (NumCopyUses < LateRematUpdateThreshold) {
// The source interval can become smaller because we removed a use.
shrinkToUses(&SrcInt, &DeadDefs);
if (!DeadDefs.empty())
eliminateDeadDefs();
} else {
ToBeUpdated.insert(SrcReg);
}
return true;
}
MachineInstr *RegisterCoalescer::eliminateUndefCopy(MachineInstr *CopyMI) {
// ProcessImplicitDefs may leave some copies of <undef> values, it only
// removes local variables. When we have a copy like:
//
// %1 = COPY undef %2
//
// We delete the copy and remove the corresponding value number from %1.
// Any uses of that value number are marked as <undef>.
// Note that we do not query CoalescerPair here but redo isMoveInstr as the
// CoalescerPair may have a new register class with adjusted subreg indices
// at this point.
unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
if(!isMoveInstr(*TRI, CopyMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
return nullptr;
SlotIndex Idx = LIS->getInstructionIndex(*CopyMI);
const LiveInterval &SrcLI = LIS->getInterval(SrcReg);
// CopyMI is undef iff SrcReg is not live before the instruction.
if (SrcSubIdx != 0 && SrcLI.hasSubRanges()) {
LaneBitmask SrcMask = TRI->getSubRegIndexLaneMask(SrcSubIdx);
for (const LiveInterval::SubRange &SR : SrcLI.subranges()) {
if ((SR.LaneMask & SrcMask).none())
continue;
if (SR.liveAt(Idx))
return nullptr;
}
} else if (SrcLI.liveAt(Idx))
return nullptr;
// If the undef copy defines a live-out value (i.e. an input to a PHI def),
// then replace it with an IMPLICIT_DEF.
LiveInterval &DstLI = LIS->getInterval(DstReg);
SlotIndex RegIndex = Idx.getRegSlot();
LiveRange::Segment *Seg = DstLI.getSegmentContaining(RegIndex);
assert(Seg != nullptr && "No segment for defining instruction");
if (VNInfo *V = DstLI.getVNInfoAt(Seg->end)) {
if (V->isPHIDef()) {
CopyMI->setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
for (unsigned i = CopyMI->getNumOperands(); i != 0; --i) {
MachineOperand &MO = CopyMI->getOperand(i-1);
if (MO.isReg() && MO.isUse())
CopyMI->RemoveOperand(i-1);
}
LLVM_DEBUG(dbgs() << "\tReplaced copy of <undef> value with an "
"implicit def\n");
return CopyMI;
}
}
// Remove any DstReg segments starting at the instruction.
LLVM_DEBUG(dbgs() << "\tEliminating copy of <undef> value\n");
// Remove value or merge with previous one in case of a subregister def.
if (VNInfo *PrevVNI = DstLI.getVNInfoAt(Idx)) {
VNInfo *VNI = DstLI.getVNInfoAt(RegIndex);
DstLI.MergeValueNumberInto(VNI, PrevVNI);
// The affected subregister segments can be removed.
LaneBitmask DstMask = TRI->getSubRegIndexLaneMask(DstSubIdx);
for (LiveInterval::SubRange &SR : DstLI.subranges()) {
if ((SR.LaneMask & DstMask).none())
continue;
VNInfo *SVNI = SR.getVNInfoAt(RegIndex);
assert(SVNI != nullptr && SlotIndex::isSameInstr(SVNI->def, RegIndex));
SR.removeValNo(SVNI);
}
DstLI.removeEmptySubRanges();
} else
LIS->removeVRegDefAt(DstLI, RegIndex);
// Mark uses as undef.
for (MachineOperand &MO : MRI->reg_nodbg_operands(DstReg)) {
if (MO.isDef() /*|| MO.isUndef()*/)
continue;
const MachineInstr &MI = *MO.getParent();
SlotIndex UseIdx = LIS->getInstructionIndex(MI);
LaneBitmask UseMask = TRI->getSubRegIndexLaneMask(MO.getSubReg());
bool isLive;
if (!UseMask.all() && DstLI.hasSubRanges()) {
isLive = false;
for (const LiveInterval::SubRange &SR : DstLI.subranges()) {
if ((SR.LaneMask & UseMask).none())
continue;
if (SR.liveAt(UseIdx)) {
isLive = true;
break;
}
}
} else
isLive = DstLI.liveAt(UseIdx);
if (isLive)
continue;
MO.setIsUndef(true);
LLVM_DEBUG(dbgs() << "\tnew undef: " << UseIdx << '\t' << MI);
}
// A def of a subregister may be a use of the other subregisters, so
// deleting a def of a subregister may also remove uses. Since CopyMI
// is still part of the function (but about to be erased), mark all
// defs of DstReg in it as <undef>, so that shrinkToUses would
// ignore them.
for (MachineOperand &MO : CopyMI->operands())
if (MO.isReg() && MO.isDef() && MO.getReg() == DstReg)
MO.setIsUndef(true);
LIS->shrinkToUses(&DstLI);
return CopyMI;
}
void RegisterCoalescer::addUndefFlag(const LiveInterval &Int, SlotIndex UseIdx,
MachineOperand &MO, unsigned SubRegIdx) {
LaneBitmask Mask = TRI->getSubRegIndexLaneMask(SubRegIdx);
if (MO.isDef())
Mask = ~Mask;
bool IsUndef = true;
for (const LiveInterval::SubRange &S : Int.subranges()) {
if ((S.LaneMask & Mask).none())
continue;
if (S.liveAt(UseIdx)) {
IsUndef = false;
break;
}
}
if (IsUndef) {
MO.setIsUndef(true);
// We found out some subregister use is actually reading an undefined
// value. In some cases the whole vreg has become undefined at this
// point so we have to potentially shrink the main range if the
// use was ending a live segment there.
LiveQueryResult Q = Int.Query(UseIdx);
if (Q.valueOut() == nullptr)
ShrinkMainRange = true;
}
}
void RegisterCoalescer::updateRegDefsUses(unsigned SrcReg, unsigned DstReg,
unsigned SubIdx) {
bool DstIsPhys = Register::isPhysicalRegister(DstReg);
LiveInterval *DstInt = DstIsPhys ? nullptr : &LIS->getInterval(DstReg);
if (DstInt && DstInt->hasSubRanges() && DstReg != SrcReg) {
for (MachineOperand &MO : MRI->reg_operands(DstReg)) {
unsigned SubReg = MO.getSubReg();
if (SubReg == 0 || MO.isUndef())
continue;
MachineInstr &MI = *MO.getParent();
if (MI.isDebugValue())
continue;
SlotIndex UseIdx = LIS->getInstructionIndex(MI).getRegSlot(true);
addUndefFlag(*DstInt, UseIdx, MO, SubReg);
}
}
SmallPtrSet<MachineInstr*, 8> Visited;
for (MachineRegisterInfo::reg_instr_iterator
I = MRI->reg_instr_begin(SrcReg), E = MRI->reg_instr_end();
I != E; ) {
MachineInstr *UseMI = &*(I++);
// Each instruction can only be rewritten once because sub-register
// composition is not always idempotent. When SrcReg != DstReg, rewriting
// the UseMI operands removes them from the SrcReg use-def chain, but when
// SrcReg is DstReg we could encounter UseMI twice if it has multiple
// operands mentioning the virtual register.
if (SrcReg == DstReg && !Visited.insert(UseMI).second)
continue;
SmallVector<unsigned,8> Ops;
bool Reads, Writes;
std::tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops);
// If SrcReg wasn't read, it may still be the case that DstReg is live-in
// because SrcReg is a sub-register.
if (DstInt && !Reads && SubIdx && !UseMI->isDebugValue())
Reads = DstInt->liveAt(LIS->getInstructionIndex(*UseMI));
// Replace SrcReg with DstReg in all UseMI operands.
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
MachineOperand &MO = UseMI->getOperand(Ops[i]);
// Adjust <undef> flags in case of sub-register joins. We don't want to
// turn a full def into a read-modify-write sub-register def and vice
// versa.
if (SubIdx && MO.isDef())
MO.setIsUndef(!Reads);
// A subreg use of a partially undef (super) register may be a complete
// undef use now and then has to be marked that way.
if (SubIdx != 0 && MO.isUse() && MRI->shouldTrackSubRegLiveness(DstReg)) {
if (!DstInt->hasSubRanges()) {
BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
LaneBitmask FullMask = MRI->getMaxLaneMaskForVReg(DstInt->reg);
LaneBitmask UsedLanes = TRI->getSubRegIndexLaneMask(SubIdx);
LaneBitmask UnusedLanes = FullMask & ~UsedLanes;
DstInt->createSubRangeFrom(Allocator, UsedLanes, *DstInt);
// The unused lanes are just empty live-ranges at this point.
// It is the caller responsibility to set the proper
// dead segments if there is an actual dead def of the
// unused lanes. This may happen with rematerialization.
DstInt->createSubRange(Allocator, UnusedLanes);
}
SlotIndex MIIdx = UseMI->isDebugValue()
? LIS->getSlotIndexes()->getIndexBefore(*UseMI)
: LIS->getInstructionIndex(*UseMI);
SlotIndex UseIdx = MIIdx.getRegSlot(true);
addUndefFlag(*DstInt, UseIdx, MO, SubIdx);
}
if (DstIsPhys)
MO.substPhysReg(DstReg, *TRI);
else
MO.substVirtReg(DstReg, SubIdx, *TRI);
}
LLVM_DEBUG({
dbgs() << "\t\tupdated: ";
if (!UseMI->isDebugValue())
dbgs() << LIS->getInstructionIndex(*UseMI) << "\t";
dbgs() << *UseMI;
});
}
}
bool RegisterCoalescer::canJoinPhys(const CoalescerPair &CP) {
// Always join simple intervals that are defined by a single copy from a
// reserved register. This doesn't increase register pressure, so it is
// always beneficial.
if (!MRI->isReserved(CP.getDstReg())) {
LLVM_DEBUG(dbgs() << "\tCan only merge into reserved registers.\n");
return false;
}
LiveInterval &JoinVInt = LIS->getInterval(CP.getSrcReg());
if (JoinVInt.containsOneValue())
return true;
LLVM_DEBUG(
dbgs() << "\tCannot join complex intervals into reserved register.\n");
return false;
}
bool RegisterCoalescer::joinCopy(MachineInstr *CopyMI, bool &Again) {
Again = false;
LLVM_DEBUG(dbgs() << LIS->getInstructionIndex(*CopyMI) << '\t' << *CopyMI);
CoalescerPair CP(*TRI);
if (!CP.setRegisters(CopyMI)) {
LLVM_DEBUG(dbgs() << "\tNot coalescable.\n");
return false;
}
if (CP.getNewRC()) {
auto SrcRC = MRI->getRegClass(CP.getSrcReg());
auto DstRC = MRI->getRegClass(CP.getDstReg());
unsigned SrcIdx = CP.getSrcIdx();
unsigned DstIdx = CP.getDstIdx();
if (CP.isFlipped()) {
std::swap(SrcIdx, DstIdx);
std::swap(SrcRC, DstRC);
}
if (!TRI->shouldCoalesce(CopyMI, SrcRC, SrcIdx, DstRC, DstIdx,
CP.getNewRC(), *LIS)) {
LLVM_DEBUG(dbgs() << "\tSubtarget bailed on coalescing.\n");
return false;
}
}
// Dead code elimination. This really should be handled by MachineDCE, but
// sometimes dead copies slip through, and we can't generate invalid live
// ranges.
if (!CP.isPhys() && CopyMI->allDefsAreDead()) {
LLVM_DEBUG(dbgs() << "\tCopy is dead.\n");
DeadDefs.push_back(CopyMI);
eliminateDeadDefs();
return true;
}
// Eliminate undefs.
if (!CP.isPhys()) {
// If this is an IMPLICIT_DEF, leave it alone, but don't try to coalesce.
if (MachineInstr *UndefMI = eliminateUndefCopy(CopyMI)) {
if (UndefMI->isImplicitDef())
return false;
deleteInstr(CopyMI);
return false; // Not coalescable.
}
}
// Coalesced copies are normally removed immediately, but transformations
// like removeCopyByCommutingDef() can inadvertently create identity copies.
// When that happens, just join the values and remove the copy.
if (CP.getSrcReg() == CP.getDstReg()) {
LiveInterval &LI = LIS->getInterval(CP.getSrcReg());
LLVM_DEBUG(dbgs() << "\tCopy already coalesced: " << LI << '\n');
const SlotIndex CopyIdx = LIS->getInstructionIndex(*CopyMI);
LiveQueryResult LRQ = LI.Query(CopyIdx);
if (VNInfo *DefVNI = LRQ.valueDefined()) {
VNInfo *ReadVNI = LRQ.valueIn();
assert(ReadVNI && "No value before copy and no <undef> flag.");
assert(ReadVNI != DefVNI && "Cannot read and define the same value.");
LI.MergeValueNumberInto(DefVNI, ReadVNI);
// Process subregister liveranges.
for (LiveInterval::SubRange &S : LI.subranges()) {
LiveQueryResult SLRQ = S.Query(CopyIdx);
if (VNInfo *SDefVNI = SLRQ.valueDefined()) {
VNInfo *SReadVNI = SLRQ.valueIn();
S.MergeValueNumberInto(SDefVNI, SReadVNI);
}
}
LLVM_DEBUG(dbgs() << "\tMerged values: " << LI << '\n');
}
deleteInstr(CopyMI);
return true;
}
// Enforce policies.
if (CP.isPhys()) {
LLVM_DEBUG(dbgs() << "\tConsidering merging "
<< printReg(CP.getSrcReg(), TRI) << " with "
<< printReg(CP.getDstReg(), TRI, CP.getSrcIdx()) << '\n');
if (!canJoinPhys(CP)) {
// Before giving up coalescing, if definition of source is defined by
// trivial computation, try rematerializing it.
bool IsDefCopy;
if (reMaterializeTrivialDef(CP, CopyMI, IsDefCopy))
return true;
if (IsDefCopy)
Again = true; // May be possible to coalesce later.
return false;
}
} else {
// When possible, let DstReg be the larger interval.
if (!CP.isPartial() && LIS->getInterval(CP.getSrcReg()).size() >
LIS->getInterval(CP.getDstReg()).size())
CP.flip();
LLVM_DEBUG({
dbgs() << "\tConsidering merging to "
<< TRI->getRegClassName(CP.getNewRC()) << " with ";
if (CP.getDstIdx() && CP.getSrcIdx())
dbgs() << printReg(CP.getDstReg()) << " in "
<< TRI->getSubRegIndexName(CP.getDstIdx()) << " and "
<< printReg(CP.getSrcReg()) << " in "
<< TRI->getSubRegIndexName(CP.getSrcIdx()) << '\n';
else
dbgs() << printReg(CP.getSrcReg(), TRI) << " in "
<< printReg(CP.getDstReg(), TRI, CP.getSrcIdx()) << '\n';
});
}
ShrinkMask = LaneBitmask::getNone();
ShrinkMainRange = false;
// Okay, attempt to join these two intervals. On failure, this returns false.
// Otherwise, if one of the intervals being joined is a physreg, this method
// always canonicalizes DstInt to be it. The output "SrcInt" will not have
// been modified, so we can use this information below to update aliases.
if (!joinIntervals(CP)) {
// Coalescing failed.
// If definition of source is defined by trivial computation, try
// rematerializing it.
bool IsDefCopy;
if (reMaterializeTrivialDef(CP, CopyMI, IsDefCopy))
return true;
// If we can eliminate the copy without merging the live segments, do so
// now.
if (!CP.isPartial() && !CP.isPhys()) {
bool Changed = adjustCopiesBackFrom(CP, CopyMI);
bool Shrink = false;
if (!Changed)
std::tie(Changed, Shrink) = removeCopyByCommutingDef(CP, CopyMI);
if (Changed) {
deleteInstr(CopyMI);
if (Shrink) {
unsigned DstReg = CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg();
LiveInterval &DstLI = LIS->getInterval(DstReg);
shrinkToUses(&DstLI);
LLVM_DEBUG(dbgs() << "\t\tshrunk: " << DstLI << '\n');
}
LLVM_DEBUG(dbgs() << "\tTrivial!\n");
return true;
}
}
// Try and see if we can partially eliminate the copy by moving the copy to
// its predecessor.
if (!CP.isPartial() && !CP.isPhys())
if (removePartialRedundancy(CP, *CopyMI))
return true;
// Otherwise, we are unable to join the intervals.
LLVM_DEBUG(dbgs() << "\tInterference!\n");
Again = true; // May be possible to coalesce later.
return false;
}
// Coalescing to a virtual register that is of a sub-register class of the
// other. Make sure the resulting register is set to the right register class.
if (CP.isCrossClass()) {
++numCrossRCs;
MRI->setRegClass(CP.getDstReg(), CP.getNewRC());
}
// Removing sub-register copies can ease the register class constraints.
// Make sure we attempt to inflate the register class of DstReg.
if (!CP.isPhys() && RegClassInfo.isProperSubClass(CP.getNewRC()))
InflateRegs.push_back(CP.getDstReg());
// CopyMI has been erased by joinIntervals at this point. Remove it from
// ErasedInstrs since copyCoalesceWorkList() won't add a successful join back
// to the work list. This keeps ErasedInstrs from growing needlessly.
ErasedInstrs.erase(CopyMI);
// Rewrite all SrcReg operands to DstReg.
// Also update DstReg operands to include DstIdx if it is set.
if (CP.getDstIdx())
updateRegDefsUses(CP.getDstReg(), CP.getDstReg(), CP.getDstIdx());
updateRegDefsUses(CP.getSrcReg(), CP.getDstReg(), CP.getSrcIdx());
// Shrink subregister ranges if necessary.
if (ShrinkMask.any()) {
LiveInterval &LI = LIS->getInterval(CP.getDstReg());
for (LiveInterval::SubRange &S : LI.subranges()) {
if ((S.LaneMask & ShrinkMask).none())
continue;
LLVM_DEBUG(dbgs() << "Shrink LaneUses (Lane " << PrintLaneMask(S.LaneMask)
<< ")\n");
LIS->shrinkToUses(S, LI.reg);
}
LI.removeEmptySubRanges();
}
// CP.getSrcReg()'s live interval has been merged into CP.getDstReg's live
// interval. Since CP.getSrcReg() is in ToBeUpdated set and its live interval
// is not up-to-date, need to update the merged live interval here.
if (ToBeUpdated.count(CP.getSrcReg()))
ShrinkMainRange = true;
if (ShrinkMainRange) {
LiveInterval &LI = LIS->getInterval(CP.getDstReg());
shrinkToUses(&LI);
}
// SrcReg is guaranteed to be the register whose live interval that is
// being merged.
LIS->removeInterval(CP.getSrcReg());
// Update regalloc hint.
TRI->updateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *MF);
LLVM_DEBUG({
dbgs() << "\tSuccess: " << printReg(CP.getSrcReg(), TRI, CP.getSrcIdx())
<< " -> " << printReg(CP.getDstReg(), TRI, CP.getDstIdx()) << '\n';
dbgs() << "\tResult = ";
if (CP.isPhys())
dbgs() << printReg(CP.getDstReg(), TRI);
else
dbgs() << LIS->getInterval(CP.getDstReg());
dbgs() << '\n';
});
++numJoins;
return true;
}
bool RegisterCoalescer::joinReservedPhysReg(CoalescerPair &CP) {
unsigned DstReg = CP.getDstReg();
unsigned SrcReg = CP.getSrcReg();
assert(CP.isPhys() && "Must be a physreg copy");
assert(MRI->isReserved(DstReg) && "Not a reserved register");
LiveInterval &RHS = LIS->getInterval(SrcReg);
LLVM_DEBUG(dbgs() << "\t\tRHS = " << RHS << '\n');
assert(RHS.containsOneValue() && "Invalid join with reserved register");
// Optimization for reserved registers like ESP. We can only merge with a
// reserved physreg if RHS has a single value that is a copy of DstReg.
// The live range of the reserved register will look like a set of dead defs
// - we don't properly track the live range of reserved registers.
// Deny any overlapping intervals. This depends on all the reserved
// register live ranges to look like dead defs.
if (!MRI->isConstantPhysReg(DstReg)) {
for (MCRegUnitIterator UI(DstReg, TRI); UI.isValid(); ++UI) {
// Abort if not all the regunits are reserved.
for (MCRegUnitRootIterator RI(*UI, TRI); RI.isValid(); ++RI) {
if (!MRI->isReserved(*RI))
return false;
}
if (RHS.overlaps(LIS->getRegUnit(*UI))) {
LLVM_DEBUG(dbgs() << "\t\tInterference: " << printRegUnit(*UI, TRI)
<< '\n');
return false;
}
}
// We must also check for overlaps with regmask clobbers.
BitVector RegMaskUsable;
if (LIS->checkRegMaskInterference(RHS, RegMaskUsable) &&
!RegMaskUsable.test(DstReg)) {
LLVM_DEBUG(dbgs() << "\t\tRegMask interference\n");
return false;
}
}
// Skip any value computations, we are not adding new values to the
// reserved register. Also skip merging the live ranges, the reserved
// register live range doesn't need to be accurate as long as all the
// defs are there.
// Delete the identity copy.
MachineInstr *CopyMI;
if (CP.isFlipped()) {
// Physreg is copied into vreg
// %y = COPY %physreg_x
// ... //< no other def of %physreg_x here
// use %y
// =>
// ...
// use %physreg_x
CopyMI = MRI->getVRegDef(SrcReg);
} else {
// VReg is copied into physreg:
// %y = def
// ... //< no other def or use of %physreg_x here
// %physreg_x = COPY %y
// =>
// %physreg_x = def
// ...
if (!MRI->hasOneNonDBGUse(SrcReg)) {
LLVM_DEBUG(dbgs() << "\t\tMultiple vreg uses!\n");
return false;
}
if (!LIS->intervalIsInOneMBB(RHS)) {
LLVM_DEBUG(dbgs() << "\t\tComplex control flow!\n");
return false;
}
MachineInstr &DestMI = *MRI->getVRegDef(SrcReg);
CopyMI = &*MRI->use_instr_nodbg_begin(SrcReg);
SlotIndex CopyRegIdx = LIS->getInstructionIndex(*CopyMI).getRegSlot();
SlotIndex DestRegIdx = LIS->getInstructionIndex(DestMI).getRegSlot();
if (!MRI->isConstantPhysReg(DstReg)) {
// We checked above that there are no interfering defs of the physical
// register. However, for this case, where we intend to move up the def of
// the physical register, we also need to check for interfering uses.
SlotIndexes *Indexes = LIS->getSlotIndexes();
for (SlotIndex SI = Indexes->getNextNonNullIndex(DestRegIdx);
SI != CopyRegIdx; SI = Indexes->getNextNonNullIndex(SI)) {
MachineInstr *MI = LIS->getInstructionFromIndex(SI);
if (MI->readsRegister(DstReg, TRI)) {
LLVM_DEBUG(dbgs() << "\t\tInterference (read): " << *MI);
return false;
}
}
}
// We're going to remove the copy which defines a physical reserved
// register, so remove its valno, etc.
LLVM_DEBUG(dbgs() << "\t\tRemoving phys reg def of "
<< printReg(DstReg, TRI) << " at " << CopyRegIdx << "\n");
LIS->removePhysRegDefAt(DstReg, CopyRegIdx);
// Create a new dead def at the new def location.
for (MCRegUnitIterator UI(DstReg, TRI); UI.isValid(); ++UI) {
LiveRange &LR = LIS->getRegUnit(*UI);
LR.createDeadDef(DestRegIdx, LIS->getVNInfoAllocator());
}
}
deleteInstr(CopyMI);
// We don't track kills for reserved registers.
MRI->clearKillFlags(CP.getSrcReg());
return true;
}
//===----------------------------------------------------------------------===//
// Interference checking and interval joining
//===----------------------------------------------------------------------===//
//
// In the easiest case, the two live ranges being joined are disjoint, and
// there is no interference to consider. It is quite common, though, to have
// overlapping live ranges, and we need to check if the interference can be
// resolved.
//
// The live range of a single SSA value forms a sub-tree of the dominator tree.
// This means that two SSA values overlap if and only if the def of one value
// is contained in the live range of the other value. As a special case, the
// overlapping values can be defined at the same index.
//
// The interference from an overlapping def can be resolved in these cases:
//
// 1. Coalescable copies. The value is defined by a copy that would become an
// identity copy after joining SrcReg and DstReg. The copy instruction will
// be removed, and the value will be merged with the source value.
//
// There can be several copies back and forth, causing many values to be
// merged into one. We compute a list of ultimate values in the joined live
// range as well as a mappings from the old value numbers.
//
// 2. IMPLICIT_DEF. This instruction is only inserted to ensure all PHI
// predecessors have a live out value. It doesn't cause real interference,
// and can be merged into the value it overlaps. Like a coalescable copy, it
// can be erased after joining.
//
// 3. Copy of external value. The overlapping def may be a copy of a value that
// is already in the other register. This is like a coalescable copy, but
// the live range of the source register must be trimmed after erasing the
// copy instruction:
//
// %src = COPY %ext
// %dst = COPY %ext <-- Remove this COPY, trim the live range of %ext.
//
// 4. Clobbering undefined lanes. Vector registers are sometimes built by
// defining one lane at a time:
//
// %dst:ssub0<def,read-undef> = FOO
// %src = BAR
// %dst:ssub1 = COPY %src
//
// The live range of %src overlaps the %dst value defined by FOO, but
// merging %src into %dst:ssub1 is only going to clobber the ssub1 lane
// which was undef anyway.
//
// The value mapping is more complicated in this case. The final live range
// will have different value numbers for both FOO and BAR, but there is no
// simple mapping from old to new values. It may even be necessary to add
// new PHI values.
//
// 5. Clobbering dead lanes. A def may clobber a lane of a vector register that
// is live, but never read. This can happen because we don't compute
// individual live ranges per lane.
//
// %dst = FOO
// %src = BAR
// %dst:ssub1 = COPY %src
//
// This kind of interference is only resolved locally. If the clobbered
// lane value escapes the block, the join is aborted.
namespace {
/// Track information about values in a single virtual register about to be
/// joined. Objects of this class are always created in pairs - one for each
/// side of the CoalescerPair (or one for each lane of a side of the coalescer
/// pair)
class JoinVals {
/// Live range we work on.
LiveRange &LR;
/// (Main) register we work on.
const unsigned Reg;
/// Reg (and therefore the values in this liverange) will end up as
/// subregister SubIdx in the coalesced register. Either CP.DstIdx or
/// CP.SrcIdx.
const unsigned SubIdx;
/// The LaneMask that this liverange will occupy the coalesced register. May
/// be smaller than the lanemask produced by SubIdx when merging subranges.
const LaneBitmask LaneMask;
/// This is true when joining sub register ranges, false when joining main
/// ranges.
const bool SubRangeJoin;
/// Whether the current LiveInterval tracks subregister liveness.
const bool TrackSubRegLiveness;
/// Values that will be present in the final live range.
SmallVectorImpl<VNInfo*> &NewVNInfo;
const CoalescerPair &CP;
LiveIntervals *LIS;
SlotIndexes *Indexes;
const TargetRegisterInfo *TRI;
/// Value number assignments. Maps value numbers in LI to entries in
/// NewVNInfo. This is suitable for passing to LiveInterval::join().
SmallVector<int, 8> Assignments;
public:
/// Conflict resolution for overlapping values.
enum ConflictResolution {
/// No overlap, simply keep this value.
CR_Keep,
/// Merge this value into OtherVNI and erase the defining instruction.
/// Used for IMPLICIT_DEF, coalescable copies, and copies from external
/// values.
CR_Erase,
/// Merge this value into OtherVNI but keep the defining instruction.
/// This is for the special case where OtherVNI is defined by the same
/// instruction.
CR_Merge,
/// Keep this value, and have it replace OtherVNI where possible. This
/// complicates value mapping since OtherVNI maps to two different values
/// before and after this def.
/// Used when clobbering undefined or dead lanes.
CR_Replace,
/// Unresolved conflict. Visit later when all values have been mapped.
CR_Unresolved,
/// Unresolvable conflict. Abort the join.
CR_Impossible
};
private:
/// Per-value info for LI. The lane bit masks are all relative to the final
/// joined register, so they can be compared directly between SrcReg and
/// DstReg.
struct Val {
ConflictResolution Resolution = CR_Keep;
/// Lanes written by this def, 0 for unanalyzed values.
LaneBitmask WriteLanes;
/// Lanes with defined values in this register. Other lanes are undef and
/// safe to clobber.
LaneBitmask ValidLanes;
/// Value in LI being redefined by this def.
VNInfo *RedefVNI = nullptr;
/// Value in the other live range that overlaps this def, if any.
VNInfo *OtherVNI = nullptr;
/// Is this value an IMPLICIT_DEF that can be erased?
///
/// IMPLICIT_DEF values should only exist at the end of a basic block that
/// is a predecessor to a phi-value. These IMPLICIT_DEF instructions can be
/// safely erased if they are overlapping a live value in the other live
/// interval.
///
/// Weird control flow graphs and incomplete PHI handling in
/// ProcessImplicitDefs can very rarely create IMPLICIT_DEF values with
/// longer live ranges. Such IMPLICIT_DEF values should be treated like
/// normal values.
bool ErasableImplicitDef = false;
/// True when the live range of this value will be pruned because of an
/// overlapping CR_Replace value in the other live range.
bool Pruned = false;
/// True once Pruned above has been computed.
bool PrunedComputed = false;
/// True if this value is determined to be identical to OtherVNI
/// (in valuesIdentical). This is used with CR_Erase where the erased
/// copy is redundant, i.e. the source value is already the same as
/// the destination. In such cases the subranges need to be updated
/// properly. See comment at pruneSubRegValues for more info.
bool Identical = false;
Val() = default;
bool isAnalyzed() const { return WriteLanes.any(); }
};
/// One entry per value number in LI.
SmallVector<Val, 8> Vals;
/// Compute the bitmask of lanes actually written by DefMI.
/// Set Redef if there are any partial register definitions that depend on the
/// previous value of the register.
LaneBitmask computeWriteLanes(const MachineInstr *DefMI, bool &Redef) const;
/// Find the ultimate value that VNI was copied from.
std::pair<const VNInfo*,unsigned> followCopyChain(const VNInfo *VNI) const;
bool valuesIdentical(VNInfo *Value0, VNInfo *Value1, const JoinVals &Other) const;
/// Analyze ValNo in this live range, and set all fields of Vals[ValNo].
/// Return a conflict resolution when possible, but leave the hard cases as
/// CR_Unresolved.
/// Recursively calls computeAssignment() on this and Other, guaranteeing that
/// both OtherVNI and RedefVNI have been analyzed and mapped before returning.
/// The recursion always goes upwards in the dominator tree, making loops
/// impossible.
ConflictResolution analyzeValue(unsigned ValNo, JoinVals &Other);
/// Compute the value assignment for ValNo in RI.
/// This may be called recursively by analyzeValue(), but never for a ValNo on
/// the stack.
void computeAssignment(unsigned ValNo, JoinVals &Other);
/// Assuming ValNo is going to clobber some valid lanes in Other.LR, compute
/// the extent of the tainted lanes in the block.
///
/// Multiple values in Other.LR can be affected since partial redefinitions
/// can preserve previously tainted lanes.
///
/// 1 %dst = VLOAD <-- Define all lanes in %dst
/// 2 %src = FOO <-- ValNo to be joined with %dst:ssub0
/// 3 %dst:ssub1 = BAR <-- Partial redef doesn't clear taint in ssub0
/// 4 %dst:ssub0 = COPY %src <-- Conflict resolved, ssub0 wasn't read
///
/// For each ValNo in Other that is affected, add an (EndIndex, TaintedLanes)
/// entry to TaintedVals.
///
/// Returns false if the tainted lanes extend beyond the basic block.
bool
taintExtent(unsigned ValNo, LaneBitmask TaintedLanes, JoinVals &Other,
SmallVectorImpl<std::pair<SlotIndex, LaneBitmask>> &TaintExtent);
/// Return true if MI uses any of the given Lanes from Reg.
/// This does not include partial redefinitions of Reg.
bool usesLanes(const MachineInstr &MI, unsigned, unsigned, LaneBitmask) const;
/// Determine if ValNo is a copy of a value number in LR or Other.LR that will
/// be pruned:
///
/// %dst = COPY %src
/// %src = COPY %dst <-- This value to be pruned.
/// %dst = COPY %src <-- This value is a copy of a pruned value.
bool isPrunedValue(unsigned ValNo, JoinVals &Other);
public:
JoinVals(LiveRange &LR, unsigned Reg, unsigned SubIdx, LaneBitmask LaneMask,
SmallVectorImpl<VNInfo*> &newVNInfo, const CoalescerPair &cp,
LiveIntervals *lis, const TargetRegisterInfo *TRI, bool SubRangeJoin,
bool TrackSubRegLiveness)
: LR(LR), Reg(Reg), SubIdx(SubIdx), LaneMask(LaneMask),
SubRangeJoin(SubRangeJoin), TrackSubRegLiveness(TrackSubRegLiveness),
NewVNInfo(newVNInfo), CP(cp), LIS(lis), Indexes(LIS->getSlotIndexes()),
TRI(TRI), Assignments(LR.getNumValNums(), -1), Vals(LR.getNumValNums()) {}
/// Analyze defs in LR and compute a value mapping in NewVNInfo.
/// Returns false if any conflicts were impossible to resolve.
bool mapValues(JoinVals &Other);
/// Try to resolve conflicts that require all values to be mapped.
/// Returns false if any conflicts were impossible to resolve.
bool resolveConflicts(JoinVals &Other);
/// Prune the live range of values in Other.LR where they would conflict with
/// CR_Replace values in LR. Collect end points for restoring the live range
/// after joining.
void pruneValues(JoinVals &Other, SmallVectorImpl<SlotIndex> &EndPoints,
bool changeInstrs);
/// Removes subranges starting at copies that get removed. This sometimes
/// happens when undefined subranges are copied around. These ranges contain
/// no useful information and can be removed.
void pruneSubRegValues(LiveInterval &LI, LaneBitmask &ShrinkMask);
/// Pruning values in subranges can lead to removing segments in these
/// subranges started by IMPLICIT_DEFs. The corresponding segments in
/// the main range also need to be removed. This function will mark
/// the corresponding values in the main range as pruned, so that
/// eraseInstrs can do the final cleanup.
/// The parameter @p LI must be the interval whose main range is the
/// live range LR.
void pruneMainSegments(LiveInterval &LI, bool &ShrinkMainRange);
/// Erase any machine instructions that have been coalesced away.
/// Add erased instructions to ErasedInstrs.
/// Add foreign virtual registers to ShrinkRegs if their live range ended at
/// the erased instrs.
void eraseInstrs(SmallPtrSetImpl<MachineInstr*> &ErasedInstrs,
SmallVectorImpl<unsigned> &ShrinkRegs,
LiveInterval *LI = nullptr);
/// Remove liverange defs at places where implicit defs will be removed.
void removeImplicitDefs();
/// Get the value assignments suitable for passing to LiveInterval::join.
const int *getAssignments() const { return Assignments.data(); }
/// Get the conflict resolution for a value number.
ConflictResolution getResolution(unsigned Num) const {
return Vals[Num].Resolution;
}
};
} // end anonymous namespace
LaneBitmask JoinVals::computeWriteLanes(const MachineInstr *DefMI, bool &Redef)
const {
LaneBitmask L;
for (const MachineOperand &MO : DefMI->operands()) {
if (!MO.isReg() || MO.getReg() != Reg || !MO.isDef())
continue;
L |= TRI->getSubRegIndexLaneMask(
TRI->composeSubRegIndices(SubIdx, MO.getSubReg()));
if (MO.readsReg())
Redef = true;
}
return L;
}
std::pair<const VNInfo*, unsigned> JoinVals::followCopyChain(
const VNInfo *VNI) const {
unsigned TrackReg = Reg;
while (!VNI->isPHIDef()) {
SlotIndex Def = VNI->def;
MachineInstr *MI = Indexes->getInstructionFromIndex(Def);
assert(MI && "No defining instruction");
if (!MI->isFullCopy())
return std::make_pair(VNI, TrackReg);
Register SrcReg = MI->getOperand(1).getReg();
if (!Register::isVirtualRegister(SrcReg))
return std::make_pair(VNI, TrackReg);
const LiveInterval &LI = LIS->getInterval(SrcReg);
const VNInfo *ValueIn;
// No subrange involved.
if (!SubRangeJoin || !LI.hasSubRanges()) {
LiveQueryResult LRQ = LI.Query(Def);
ValueIn = LRQ.valueIn();
} else {
// Query subranges. Ensure that all matching ones take us to the same def
// (allowing some of them to be undef).
ValueIn = nullptr;
for (const LiveInterval::SubRange &S : LI.subranges()) {
// Transform lanemask to a mask in the joined live interval.
LaneBitmask SMask = TRI->composeSubRegIndexLaneMask(SubIdx, S.LaneMask);
if ((SMask & LaneMask).none())
continue;
LiveQueryResult LRQ = S.Query(Def);
if (!ValueIn) {
ValueIn = LRQ.valueIn();
continue;
}
if (LRQ.valueIn() && ValueIn != LRQ.valueIn())
return std::make_pair(VNI, TrackReg);
}
}
if (ValueIn == nullptr) {
// Reaching an undefined value is legitimate, for example:
//
// 1 undef %0.sub1 = ... ;; %0.sub0 == undef
// 2 %1 = COPY %0 ;; %1 is defined here.
// 3 %0 = COPY %1 ;; Now %0.sub0 has a definition,
// ;; but it's equivalent to "undef".
return std::make_pair(nullptr, SrcReg);
}
VNI = ValueIn;
TrackReg = SrcReg;
}
return std::make_pair(VNI, TrackReg);
}
bool JoinVals::valuesIdentical(VNInfo *Value0, VNInfo *Value1,
const JoinVals &Other) const {
const VNInfo *Orig0;
unsigned Reg0;
std::tie(Orig0, Reg0) = followCopyChain(Value0);
if (Orig0 == Value1 && Reg0 == Other.Reg)
return true;
const VNInfo *Orig1;
unsigned Reg1;
std::tie(Orig1, Reg1) = Other.followCopyChain(Value1);
// If both values are undefined, and the source registers are the same
// register, the values are identical. Filter out cases where only one
// value is defined.
if (Orig0 == nullptr || Orig1 == nullptr)
return Orig0 == Orig1 && Reg0 == Reg1;
// The values are equal if they are defined at the same place and use the
// same register. Note that we cannot compare VNInfos directly as some of
// them might be from a copy created in mergeSubRangeInto() while the other
// is from the original LiveInterval.
return Orig0->def == Orig1->def && Reg0 == Reg1;
}
JoinVals::ConflictResolution
JoinVals::analyzeValue(unsigned ValNo, JoinVals &Other) {
Val &V = Vals[ValNo];
assert(!V.isAnalyzed() && "Value has already been analyzed!");
VNInfo *VNI = LR.getValNumInfo(ValNo);
if (VNI->isUnused()) {
V.WriteLanes = LaneBitmask::getAll();
return CR_Keep;
}
// Get the instruction defining this value, compute the lanes written.
const MachineInstr *DefMI = nullptr;
if (VNI->isPHIDef()) {
// Conservatively assume that all lanes in a PHI are valid.
LaneBitmask Lanes = SubRangeJoin ? LaneBitmask::getLane(0)
: TRI->getSubRegIndexLaneMask(SubIdx);
V.ValidLanes = V.WriteLanes = Lanes;
} else {
DefMI = Indexes->getInstructionFromIndex(VNI->def);
assert(DefMI != nullptr);
if (SubRangeJoin) {
// We don't care about the lanes when joining subregister ranges.
V.WriteLanes = V.ValidLanes = LaneBitmask::getLane(0);
if (DefMI->isImplicitDef()) {
V.ValidLanes = LaneBitmask::getNone();
V.ErasableImplicitDef = true;
}
} else {
bool Redef = false;
V.ValidLanes = V.WriteLanes = computeWriteLanes(DefMI, Redef);
// If this is a read-modify-write instruction, there may be more valid
// lanes than the ones written by this instruction.
// This only covers partial redef operands. DefMI may have normal use
// operands reading the register. They don't contribute valid lanes.
//
// This adds ssub1 to the set of valid lanes in %src:
//
// %src:ssub1 = FOO
//
// This leaves only ssub1 valid, making any other lanes undef:
//
// %src:ssub1<def,read-undef> = FOO %src:ssub2
//
// The <read-undef> flag on the def operand means that old lane values are
// not important.
if (Redef) {
V.RedefVNI = LR.Query(VNI->def).valueIn();
assert((TrackSubRegLiveness || V.RedefVNI) &&
"Instruction is reading nonexistent value");
if (V.RedefVNI != nullptr) {
computeAssignment(V.RedefVNI->id, Other);
V.ValidLanes |= Vals[V.RedefVNI->id].ValidLanes;
}
}
// An IMPLICIT_DEF writes undef values.
if (DefMI->isImplicitDef()) {
// We normally expect IMPLICIT_DEF values to be live only until the end
// of their block. If the value is really live longer and gets pruned in
// another block, this flag is cleared again.
//
// Clearing the valid lanes is deferred until it is sure this can be
// erased.
V.ErasableImplicitDef = true;
}
}
}
// Find the value in Other that overlaps VNI->def, if any.
LiveQueryResult OtherLRQ = Other.LR.Query(VNI->def);
// It is possible that both values are defined by the same instruction, or
// the values are PHIs defined in the same block. When that happens, the two
// values should be merged into one, but not into any preceding value.
// The first value defined or visited gets CR_Keep, the other gets CR_Merge.
if (VNInfo *OtherVNI = OtherLRQ.valueDefined()) {
assert(SlotIndex::isSameInstr(VNI->def, OtherVNI->def) && "Broken LRQ");
// One value stays, the other is merged. Keep the earlier one, or the first
// one we see.
if (OtherVNI->def < VNI->def)
Other.computeAssignment(OtherVNI->id, *this);
else if (VNI->def < OtherVNI->def && OtherLRQ.valueIn()) {
// This is an early-clobber def overlapping a live-in value in the other
// register. Not mergeable.
V.OtherVNI = OtherLRQ.valueIn();
return CR_Impossible;
}
V.OtherVNI = OtherVNI;
Val &OtherV = Other.Vals[OtherVNI->id];
// Keep this value, check for conflicts when analyzing OtherVNI.
if (!OtherV.isAnalyzed())
return CR_Keep;
// Both sides have been analyzed now.
// Allow overlapping PHI values. Any real interference would show up in a
// predecessor, the PHI itself can't introduce any conflicts.
if (VNI->isPHIDef())
return CR_Merge;
if ((V.ValidLanes & OtherV.ValidLanes).any())
// Overlapping lanes can't be resolved.
return CR_Impossible;
else
return CR_Merge;
}
// No simultaneous def. Is Other live at the def?
V.OtherVNI = OtherLRQ.valueIn();
if (!V.OtherVNI)
// No overlap, no conflict.
return CR_Keep;
assert(!SlotIndex::isSameInstr(VNI->def, V.OtherVNI->def) && "Broken LRQ");
// We have overlapping values, or possibly a kill of Other.
// Recursively compute assignments up the dominator tree.
Other.computeAssignment(V.OtherVNI->id, *this);
Val &OtherV = Other.Vals[V.OtherVNI->id];
if (OtherV.ErasableImplicitDef) {
// Check if OtherV is an IMPLICIT_DEF that extends beyond its basic block.
// This shouldn't normally happen, but ProcessImplicitDefs can leave such
// IMPLICIT_DEF instructions behind, and there is nothing wrong with it
// technically.
//
// When it happens, treat that IMPLICIT_DEF as a normal value, and don't try
// to erase the IMPLICIT_DEF instruction.
if (DefMI &&
DefMI->getParent() != Indexes->getMBBFromIndex(V.OtherVNI->def)) {
LLVM_DEBUG(dbgs() << "IMPLICIT_DEF defined at " << V.OtherVNI->def
<< " extends into "
<< printMBBReference(*DefMI->getParent())
<< ", keeping it.\n");
OtherV.ErasableImplicitDef = false;
} else {
// We deferred clearing these lanes in case we needed to save them
OtherV.ValidLanes &= ~OtherV.WriteLanes;
}
}
// Allow overlapping PHI values. Any real interference would show up in a
// predecessor, the PHI itself can't introduce any conflicts.
if (VNI->isPHIDef())
return CR_Replace;
// Check for simple erasable conflicts.
if (DefMI->isImplicitDef()) {
// We need the def for the subregister if there is nothing else live at the
// subrange at this point.
if (TrackSubRegLiveness
&& (V.WriteLanes & (OtherV.ValidLanes | OtherV.WriteLanes)).none())
return CR_Replace;
return CR_Erase;
}
// Include the non-conflict where DefMI is a coalescable copy that kills
// OtherVNI. We still want the copy erased and value numbers merged.
if (CP.isCoalescable(DefMI)) {
// Some of the lanes copied from OtherVNI may be undef, making them undef
// here too.
V.ValidLanes &= ~V.WriteLanes | OtherV.ValidLanes;
return CR_Erase;
}
// This may not be a real conflict if DefMI simply kills Other and defines
// VNI.
if (OtherLRQ.isKill() && OtherLRQ.endPoint() <= VNI->def)
return CR_Keep;
// Handle the case where VNI and OtherVNI can be proven to be identical:
//
// %other = COPY %ext
// %this = COPY %ext <-- Erase this copy
//
if (DefMI->isFullCopy() && !CP.isPartial() &&
valuesIdentical(VNI, V.OtherVNI, Other)) {
V.Identical = true;
return CR_Erase;
}
// The remaining checks apply to the lanes, which aren't tracked here. This
// was already decided to be OK via the following CR_Replace condition.
// CR_Replace.
if (SubRangeJoin)
return CR_Replace;
// If the lanes written by this instruction were all undef in OtherVNI, it is
// still safe to join the live ranges. This can't be done with a simple value
// mapping, though - OtherVNI will map to multiple values:
//
// 1 %dst:ssub0 = FOO <-- OtherVNI
// 2 %src = BAR <-- VNI
// 3 %dst:ssub1 = COPY killed %src <-- Eliminate this copy.
// 4 BAZ killed %dst
// 5 QUUX killed %src
//
// Here OtherVNI will map to itself in [1;2), but to VNI in [2;5). CR_Replace
// handles this complex value mapping.
if ((V.WriteLanes & OtherV.ValidLanes).none())
return CR_Replace;
// If the other live range is killed by DefMI and the live ranges are still
// overlapping, it must be because we're looking at an early clobber def:
//
// %dst<def,early-clobber> = ASM killed %src
//
// In this case, it is illegal to merge the two live ranges since the early
// clobber def would clobber %src before it was read.
if (OtherLRQ.isKill()) {
// This case where the def doesn't overlap the kill is handled above.
assert(VNI->def.isEarlyClobber() &&
"Only early clobber defs can overlap a kill");
return CR_Impossible;
}
// VNI is clobbering live lanes in OtherVNI, but there is still the
// possibility that no instructions actually read the clobbered lanes.
// If we're clobbering all the lanes in OtherVNI, at least one must be read.
// Otherwise Other.RI wouldn't be live here.
if ((TRI->getSubRegIndexLaneMask(Other.SubIdx) & ~V.WriteLanes).none())
return CR_Impossible;
// We need to verify that no instructions are reading the clobbered lanes. To
// save compile time, we'll only check that locally. Don't allow the tainted
// value to escape the basic block.
MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
if (OtherLRQ.endPoint() >= Indexes->getMBBEndIdx(MBB))
return CR_Impossible;
// There are still some things that could go wrong besides clobbered lanes
// being read, for example OtherVNI may be only partially redefined in MBB,
// and some clobbered lanes could escape the block. Save this analysis for
// resolveConflicts() when all values have been mapped. We need to know
// RedefVNI and WriteLanes for any later defs in MBB, and we can't compute
// that now - the recursive analyzeValue() calls must go upwards in the
// dominator tree.
return CR_Unresolved;
}
void JoinVals::computeAssignment(unsigned ValNo, JoinVals &Other) {
Val &V = Vals[ValNo];
if (V.isAnalyzed()) {
// Recursion should always move up the dominator tree, so ValNo is not
// supposed to reappear before it has been assigned.
assert(Assignments[ValNo] != -1 && "Bad recursion?");
return;
}
switch ((V.Resolution = analyzeValue(ValNo, Other))) {
case CR_Erase:
case CR_Merge:
// Merge this ValNo into OtherVNI.
assert(V.OtherVNI && "OtherVNI not assigned, can't merge.");
assert(Other.Vals[V.OtherVNI->id].isAnalyzed() && "Missing recursion");
Assignments[ValNo] = Other.Assignments[V.OtherVNI->id];
LLVM_DEBUG(dbgs() << "\t\tmerge " << printReg(Reg) << ':' << ValNo << '@'
<< LR.getValNumInfo(ValNo)->def << " into "
<< printReg(Other.Reg) << ':' << V.OtherVNI->id << '@'
<< V.OtherVNI->def << " --> @"
<< NewVNInfo[Assignments[ValNo]]->def << '\n');
break;
case CR_Replace:
case CR_Unresolved: {
// The other value is going to be pruned if this join is successful.
assert(V.OtherVNI && "OtherVNI not assigned, can't prune");
Val &OtherV = Other.Vals[V.OtherVNI->id];
// We cannot erase an IMPLICIT_DEF if we don't have valid values for all
// its lanes.
if (OtherV.ErasableImplicitDef &&
TrackSubRegLiveness &&
(OtherV.WriteLanes & ~V.ValidLanes).any()) {
LLVM_DEBUG(dbgs() << "Cannot erase implicit_def with missing values\n");
OtherV.ErasableImplicitDef = false;
// The valid lanes written by the implicit_def were speculatively cleared
// before, so make this more conservative. It may be better to track this,
// I haven't found a testcase where it matters.
OtherV.ValidLanes = LaneBitmask::getAll();
}
OtherV.Pruned = true;
LLVM_FALLTHROUGH;
}
default:
// This value number needs to go in the final joined live range.
Assignments[ValNo] = NewVNInfo.size();
NewVNInfo.push_back(LR.getValNumInfo(ValNo));
break;
}
}
bool JoinVals::mapValues(JoinVals &Other) {
for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
computeAssignment(i, Other);
if (Vals[i].Resolution == CR_Impossible) {
LLVM_DEBUG(dbgs() << "\t\tinterference at " << printReg(Reg) << ':' << i
<< '@' << LR.getValNumInfo(i)->def << '\n');
return false;
}
}
return true;
}
bool JoinVals::
taintExtent(unsigned ValNo, LaneBitmask TaintedLanes, JoinVals &Other,
SmallVectorImpl<std::pair<SlotIndex, LaneBitmask>> &TaintExtent) {
VNInfo *VNI = LR.getValNumInfo(ValNo);
MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
SlotIndex MBBEnd = Indexes->getMBBEndIdx(MBB);
// Scan Other.LR from VNI.def to MBBEnd.
LiveInterval::iterator OtherI = Other.LR.find(VNI->def);
assert(OtherI != Other.LR.end() && "No conflict?");
do {
// OtherI is pointing to a tainted value. Abort the join if the tainted
// lanes escape the block.
SlotIndex End = OtherI->end;
if (End >= MBBEnd) {
LLVM_DEBUG(dbgs() << "\t\ttaints global " << printReg(Other.Reg) << ':'
<< OtherI->valno->id << '@' << OtherI->start << '\n');
return false;
}
LLVM_DEBUG(dbgs() << "\t\ttaints local " << printReg(Other.Reg) << ':'
<< OtherI->valno->id << '@' << OtherI->start << " to "
<< End << '\n');
// A dead def is not a problem.
if (End.isDead())
break;
TaintExtent.push_back(std::make_pair(End, TaintedLanes));
// Check for another def in the MBB.
if (++OtherI == Other.LR.end() || OtherI->start >= MBBEnd)
break;
// Lanes written by the new def are no longer tainted.
const Val &OV = Other.Vals[OtherI->valno->id];
TaintedLanes &= ~OV.WriteLanes;
if (!OV.RedefVNI)
break;
} while (TaintedLanes.any());
return true;
}
bool JoinVals::usesLanes(const MachineInstr &MI, unsigned Reg, unsigned SubIdx,
LaneBitmask Lanes) const {
if (MI.isDebugInstr())
return false;
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || MO.isDef() || MO.getReg() != Reg)
continue;
if (!MO.readsReg())
continue;
unsigned S = TRI->composeSubRegIndices(SubIdx, MO.getSubReg());
if ((Lanes & TRI->getSubRegIndexLaneMask(S)).any())
return true;
}
return false;
}
bool JoinVals::resolveConflicts(JoinVals &Other) {
for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
Val &V = Vals[i];
assert(V.Resolution != CR_Impossible && "Unresolvable conflict");
if (V.Resolution != CR_Unresolved)
continue;
LLVM_DEBUG(dbgs() << "\t\tconflict at " << printReg(Reg) << ':' << i << '@'
<< LR.getValNumInfo(i)->def << '\n');
if (SubRangeJoin)
return false;
++NumLaneConflicts;
assert(V.OtherVNI && "Inconsistent conflict resolution.");
VNInfo *VNI = LR.getValNumInfo(i);
const Val &OtherV = Other.Vals[V.OtherVNI->id];
// VNI is known to clobber some lanes in OtherVNI. If we go ahead with the
// join, those lanes will be tainted with a wrong value. Get the extent of
// the tainted lanes.
LaneBitmask TaintedLanes = V.WriteLanes & OtherV.ValidLanes;
SmallVector<std::pair<SlotIndex, LaneBitmask>, 8> TaintExtent;
if (!taintExtent(i, TaintedLanes, Other, TaintExtent))
// Tainted lanes would extend beyond the basic block.
return false;
assert(!TaintExtent.empty() && "There should be at least one conflict.");
// Now look at the instructions from VNI->def to TaintExtent (inclusive).
MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
MachineBasicBlock::iterator MI = MBB->begin();
if (!VNI->isPHIDef()) {
MI = Indexes->getInstructionFromIndex(VNI->def);
// No need to check the instruction defining VNI for reads.
++MI;
}
assert(!SlotIndex::isSameInstr(VNI->def, TaintExtent.front().first) &&
"Interference ends on VNI->def. Should have been handled earlier");
MachineInstr *LastMI =
Indexes->getInstructionFromIndex(TaintExtent.front().first);
assert(LastMI && "Range must end at a proper instruction");
unsigned TaintNum = 0;
while (true) {
assert(MI != MBB->end() && "Bad LastMI");
if (usesLanes(*MI, Other.Reg, Other.SubIdx, TaintedLanes)) {
LLVM_DEBUG(dbgs() << "\t\ttainted lanes used by: " << *MI);
return false;
}
// LastMI is the last instruction to use the current value.
if (&*MI == LastMI) {
if (++TaintNum == TaintExtent.size())
break;
LastMI = Indexes->getInstructionFromIndex(TaintExtent[TaintNum].first);
assert(LastMI && "Range must end at a proper instruction");
TaintedLanes = TaintExtent[TaintNum].second;
}
++MI;
}
// The tainted lanes are unused.
V.Resolution = CR_Replace;
++NumLaneResolves;
}
return true;
}
bool JoinVals::isPrunedValue(unsigned ValNo, JoinVals &Other) {
Val &V = Vals[ValNo];
if (V.Pruned || V.PrunedComputed)
return V.Pruned;
if (V.Resolution != CR_Erase && V.Resolution != CR_Merge)
return V.Pruned;
// Follow copies up the dominator tree and check if any intermediate value
// has been pruned.
V.PrunedComputed = true;
V.Pruned = Other.isPrunedValue(V.OtherVNI->id, *this);
return V.Pruned;
}
void JoinVals::pruneValues(JoinVals &Other,
SmallVectorImpl<SlotIndex> &EndPoints,
bool changeInstrs) {
for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
SlotIndex Def = LR.getValNumInfo(i)->def;
switch (Vals[i].Resolution) {
case CR_Keep:
break;
case CR_Replace: {
// This value takes precedence over the value in Other.LR.
LIS->pruneValue(Other.LR, Def, &EndPoints);
// Check if we're replacing an IMPLICIT_DEF value. The IMPLICIT_DEF
// instructions are only inserted to provide a live-out value for PHI
// predecessors, so the instruction should simply go away once its value
// has been replaced.
Val &OtherV = Other.Vals[Vals[i].OtherVNI->id];
bool EraseImpDef = OtherV.ErasableImplicitDef &&
OtherV.Resolution == CR_Keep;
if (!Def.isBlock()) {
if (changeInstrs) {
// Remove <def,read-undef> flags. This def is now a partial redef.
// Also remove dead flags since the joined live range will
// continue past this instruction.
for (MachineOperand &MO :
Indexes->getInstructionFromIndex(Def)->operands()) {
if (MO.isReg() && MO.isDef() && MO.getReg() == Reg) {
if (MO.getSubReg() != 0 && MO.isUndef() && !EraseImpDef)
MO.setIsUndef(false);
MO.setIsDead(false);
}
}
}
// This value will reach instructions below, but we need to make sure
// the live range also reaches the instruction at Def.
if (!EraseImpDef)
EndPoints.push_back(Def);
}
LLVM_DEBUG(dbgs() << "\t\tpruned " << printReg(Other.Reg) << " at " << Def
<< ": " << Other.LR << '\n');
break;
}
case CR_Erase:
case CR_Merge:
if (isPrunedValue(i, Other)) {
// This value is ultimately a copy of a pruned value in LR or Other.LR.
// We can no longer trust the value mapping computed by
// computeAssignment(), the value that was originally copied could have
// been replaced.
LIS->pruneValue(LR, Def, &EndPoints);
LLVM_DEBUG(dbgs() << "\t\tpruned all of " << printReg(Reg) << " at "
<< Def << ": " << LR << '\n');
}
break;
case CR_Unresolved:
case CR_Impossible:
llvm_unreachable("Unresolved conflicts");
}
}
}
/// Consider the following situation when coalescing the copy between
/// %31 and %45 at 800. (The vertical lines represent live range segments.)
///
/// Main range Subrange 0004 (sub2)
/// %31 %45 %31 %45
/// 544 %45 = COPY %28 + +
/// | v1 | v1
/// 560B bb.1: + +
/// 624 = %45.sub2 | v2 | v2
/// 800 %31 = COPY %45 + + + +
/// | v0 | v0
/// 816 %31.sub1 = ... + |
/// 880 %30 = COPY %31 | v1 +
/// 928 %45 = COPY %30 | + +
/// | | v0 | v0 <--+
/// 992B ; backedge -> bb.1 | + + |
/// 1040 = %31.sub0 + |
/// This value must remain
/// live-out!
///
/// Assuming that %31 is coalesced into %45, the copy at 928 becomes
/// redundant, since it copies the value from %45 back into it. The
/// conflict resolution for the main range determines that %45.v0 is
/// to be erased, which is ok since %31.v1 is identical to it.
/// The problem happens with the subrange for sub2: it has to be live
/// on exit from the block, but since 928 was actually a point of
/// definition of %45.sub2, %45.sub2 was not live immediately prior
/// to that definition. As a result, when 928 was erased, the value v0
/// for %45.sub2 was pruned in pruneSubRegValues. Consequently, an
/// IMPLICIT_DEF was inserted as a "backedge" definition for %45.sub2,
/// providing an incorrect value to the use at 624.
///
/// Since the main-range values %31.v1 and %45.v0 were proved to be
/// identical, the corresponding values in subranges must also be the
/// same. A redundant copy is removed because it's not needed, and not
/// because it copied an undefined value, so any liveness that originated
/// from that copy cannot disappear. When pruning a value that started
/// at the removed copy, the corresponding identical value must be
/// extended to replace it.
void JoinVals::pruneSubRegValues(LiveInterval &LI, LaneBitmask &ShrinkMask) {
// Look for values being erased.
bool DidPrune = false;
for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
Val &V = Vals[i];
// We should trigger in all cases in which eraseInstrs() does something.
// match what eraseInstrs() is doing, print a message so
if (V.Resolution != CR_Erase &&
(V.Resolution != CR_Keep || !V.ErasableImplicitDef || !V.Pruned))
continue;
// Check subranges at the point where the copy will be removed.
SlotIndex Def = LR.getValNumInfo(i)->def;
SlotIndex OtherDef;
if (V.Identical)
OtherDef = V.OtherVNI->def;
// Print message so mismatches with eraseInstrs() can be diagnosed.
LLVM_DEBUG(dbgs() << "\t\tExpecting instruction removal at " << Def
<< '\n');
for (LiveInterval::SubRange &S : LI.subranges()) {
LiveQueryResult Q = S.Query(Def);
// If a subrange starts at the copy then an undefined value has been
// copied and we must remove that subrange value as well.
VNInfo *ValueOut = Q.valueOutOrDead();
if (ValueOut != nullptr && (Q.valueIn() == nullptr ||
(V.Identical && V.Resolution == CR_Erase &&
ValueOut->def == Def))) {
LLVM_DEBUG(dbgs() << "\t\tPrune sublane " << PrintLaneMask(S.LaneMask)
<< " at " << Def << "\n");
SmallVector<SlotIndex,8> EndPoints;
LIS->pruneValue(S, Def, &EndPoints);
DidPrune = true;
// Mark value number as unused.
ValueOut->markUnused();
if (V.Identical && S.Query(OtherDef).valueOutOrDead()) {
// If V is identical to V.OtherVNI (and S was live at OtherDef),
// then we can't simply prune V from S. V needs to be replaced
// with V.OtherVNI.
LIS->extendToIndices(S, EndPoints);
}
continue;
}
// If a subrange ends at the copy, then a value was copied but only
// partially used later. Shrink the subregister range appropriately.
if (Q.valueIn() != nullptr && Q.valueOut() == nullptr) {
LLVM_DEBUG(dbgs() << "\t\tDead uses at sublane "
<< PrintLaneMask(S.LaneMask) << " at " << Def
<< "\n");
ShrinkMask |= S.LaneMask;
}
}
}
if (DidPrune)
LI.removeEmptySubRanges();
}
/// Check if any of the subranges of @p LI contain a definition at @p Def.
static bool isDefInSubRange(LiveInterval &LI, SlotIndex Def) {
for (LiveInterval::SubRange &SR : LI.subranges()) {
if (VNInfo *VNI = SR.Query(Def).valueOutOrDead())
if (VNI->def == Def)
return true;
}
return false;
}
void JoinVals::pruneMainSegments(LiveInterval &LI, bool &ShrinkMainRange) {
assert(&static_cast<LiveRange&>(LI) == &LR);
for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
if (Vals[i].Resolution != CR_Keep)
continue;
VNInfo *VNI = LR.getValNumInfo(i);
if (VNI->isUnused() || VNI->isPHIDef() || isDefInSubRange(LI, VNI->def))
continue;
Vals[i].Pruned = true;
ShrinkMainRange = true;
}
}
void JoinVals::removeImplicitDefs() {
for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
Val &V = Vals[i];
if (V.Resolution != CR_Keep || !V.ErasableImplicitDef || !V.Pruned)
continue;
VNInfo *VNI = LR.getValNumInfo(i);
VNI->markUnused();
LR.removeValNo(VNI);
}
}
void JoinVals::eraseInstrs(SmallPtrSetImpl<MachineInstr*> &ErasedInstrs,
SmallVectorImpl<unsigned> &ShrinkRegs,
LiveInterval *LI) {
for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
// Get the def location before markUnused() below invalidates it.
VNInfo *VNI = LR.getValNumInfo(i);
SlotIndex Def = VNI->def;
switch (Vals[i].Resolution) {
case CR_Keep: {
// If an IMPLICIT_DEF value is pruned, it doesn't serve a purpose any
// longer. The IMPLICIT_DEF instructions are only inserted by
// PHIElimination to guarantee that all PHI predecessors have a value.
if (!Vals[i].ErasableImplicitDef || !Vals[i].Pruned)
break;
// Remove value number i from LR.
// For intervals with subranges, removing a segment from the main range
// may require extending the previous segment: for each definition of
// a subregister, there will be a corresponding def in the main range.
// That def may fall in the middle of a segment from another subrange.
// In such cases, removing this def from the main range must be
// complemented by extending the main range to account for the liveness
// of the other subrange.
// The new end point of the main range segment to be extended.
SlotIndex NewEnd;
if (LI != nullptr) {
LiveRange::iterator I = LR.FindSegmentContaining(Def);
assert(I != LR.end());
// Do not extend beyond the end of the segment being removed.
// The segment may have been pruned in preparation for joining
// live ranges.
NewEnd = I->end;
}
LR.removeValNo(VNI);
// Note that this VNInfo is reused and still referenced in NewVNInfo,
// make it appear like an unused value number.
VNI->markUnused();
if (LI != nullptr && LI->hasSubRanges()) {
assert(static_cast<LiveRange*>(LI) == &LR);
// Determine the end point based on the subrange information:
// minimum of (earliest def of next segment,
// latest end point of containing segment)
SlotIndex ED, LE;
for (LiveInterval::SubRange &SR : LI->subranges()) {
LiveRange::iterator I = SR.find(Def);
if (I == SR.end())
continue;
if (I->start > Def)
ED = ED.isValid() ? std::min(ED, I->start) : I->start;
else
LE = LE.isValid() ? std::max(LE, I->end) : I->end;
}
if (LE.isValid())
NewEnd = std::min(NewEnd, LE);
if (ED.isValid())
NewEnd = std::min(NewEnd, ED);
// We only want to do the extension if there was a subrange that
// was live across Def.
if (LE.isValid()) {
LiveRange::iterator S = LR.find(Def);
if (S != LR.begin())
std::prev(S)->end = NewEnd;
}
}
LLVM_DEBUG({
dbgs() << "\t\tremoved " << i << '@' << Def << ": " << LR << '\n';
if (LI != nullptr)
dbgs() << "\t\t LHS = " << *LI << '\n';
});
LLVM_FALLTHROUGH;
}
case CR_Erase: {
MachineInstr *MI = Indexes->getInstructionFromIndex(Def);
assert(MI && "No instruction to erase");
if (MI->isCopy()) {
Register Reg = MI->getOperand(1).getReg();
if (Register::isVirtualRegister(Reg) && Reg != CP.getSrcReg() &&
Reg != CP.getDstReg())
ShrinkRegs.push_back(Reg);
}
ErasedInstrs.insert(MI);
LLVM_DEBUG(dbgs() << "\t\terased:\t" << Def << '\t' << *MI);
LIS->RemoveMachineInstrFromMaps(*MI);
MI->eraseFromParent();
break;
}
default:
break;
}
}
}
void RegisterCoalescer::joinSubRegRanges(LiveRange &LRange, LiveRange &RRange,
LaneBitmask LaneMask,
const CoalescerPair &CP) {
SmallVector<VNInfo*, 16> NewVNInfo;
JoinVals RHSVals(RRange, CP.getSrcReg(), CP.getSrcIdx(), LaneMask,
NewVNInfo, CP, LIS, TRI, true, true);
JoinVals LHSVals(LRange, CP.getDstReg(), CP.getDstIdx(), LaneMask,
NewVNInfo, CP, LIS, TRI, true, true);
// Compute NewVNInfo and resolve conflicts (see also joinVirtRegs())
// We should be able to resolve all conflicts here as we could successfully do
// it on the mainrange already. There is however a problem when multiple
// ranges get mapped to the "overflow" lane mask bit which creates unexpected
// interferences.
if (!LHSVals.mapValues(RHSVals) || !RHSVals.mapValues(LHSVals)) {
// We already determined that it is legal to merge the intervals, so this
// should never fail.
llvm_unreachable("*** Couldn't join subrange!\n");
}
if (!LHSVals.resolveConflicts(RHSVals) ||
!RHSVals.resolveConflicts(LHSVals)) {
// We already determined that it is legal to merge the intervals, so this
// should never fail.
llvm_unreachable("*** Couldn't join subrange!\n");
}
// The merging algorithm in LiveInterval::join() can't handle conflicting
// value mappings, so we need to remove any live ranges that overlap a
// CR_Replace resolution. Collect a set of end points that can be used to
// restore the live range after joining.
SmallVector<SlotIndex, 8> EndPoints;
LHSVals.pruneValues(RHSVals, EndPoints, false);
RHSVals.pruneValues(LHSVals, EndPoints, false);
LHSVals.removeImplicitDefs();
RHSVals.removeImplicitDefs();
LRange.verify();
RRange.verify();
// Join RRange into LHS.
LRange.join(RRange, LHSVals.getAssignments(), RHSVals.getAssignments(),
NewVNInfo);
LLVM_DEBUG(dbgs() << "\t\tjoined lanes: " << PrintLaneMask(LaneMask)
<< ' ' << LRange << "\n");
if (EndPoints.empty())
return;
// Recompute the parts of the live range we had to remove because of
// CR_Replace conflicts.
LLVM_DEBUG({
dbgs() << "\t\trestoring liveness to " << EndPoints.size() << " points: ";
for (unsigned i = 0, n = EndPoints.size(); i != n; ++i) {
dbgs() << EndPoints[i];
if (i != n-1)
dbgs() << ',';
}
dbgs() << ": " << LRange << '\n';
});
LIS->extendToIndices(LRange, EndPoints);
}
void RegisterCoalescer::mergeSubRangeInto(LiveInterval &LI,
const LiveRange &ToMerge,
LaneBitmask LaneMask,
CoalescerPair &CP,
unsigned ComposeSubRegIdx) {
BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
LI.refineSubRanges(
Allocator, LaneMask,
[this, &Allocator, &ToMerge, &CP](LiveInterval::SubRange &SR) {
if (SR.empty()) {
SR.assign(ToMerge, Allocator);
} else {
// joinSubRegRange() destroys the merged range, so we need a copy.
LiveRange RangeCopy(ToMerge, Allocator);
joinSubRegRanges(SR, RangeCopy, SR.LaneMask, CP);
}
},
*LIS->getSlotIndexes(), *TRI, ComposeSubRegIdx);
}
bool RegisterCoalescer::isHighCostLiveInterval(LiveInterval &LI) {
if (LI.valnos.size() < LargeIntervalSizeThreshold)
return false;
auto &Counter = LargeLIVisitCounter[LI.reg];
if (Counter < LargeIntervalFreqThreshold) {
Counter++;
return false;
}
return true;
}
bool RegisterCoalescer::joinVirtRegs(CoalescerPair &CP) {
SmallVector<VNInfo*, 16> NewVNInfo;
LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
LiveInterval &LHS = LIS->getInterval(CP.getDstReg());
bool TrackSubRegLiveness = MRI->shouldTrackSubRegLiveness(*CP.getNewRC());
JoinVals RHSVals(RHS, CP.getSrcReg(), CP.getSrcIdx(), LaneBitmask::getNone(),
NewVNInfo, CP, LIS, TRI, false, TrackSubRegLiveness);
JoinVals LHSVals(LHS, CP.getDstReg(), CP.getDstIdx(), LaneBitmask::getNone(),
NewVNInfo, CP, LIS, TRI, false, TrackSubRegLiveness);
LLVM_DEBUG(dbgs() << "\t\tRHS = " << RHS << "\n\t\tLHS = " << LHS << '\n');
if (isHighCostLiveInterval(LHS) || isHighCostLiveInterval(RHS))
return false;
// First compute NewVNInfo and the simple value mappings.
// Detect impossible conflicts early.
if (!LHSVals.mapValues(RHSVals) || !RHSVals.mapValues(LHSVals))
return false;
// Some conflicts can only be resolved after all values have been mapped.
if (!LHSVals.resolveConflicts(RHSVals) || !RHSVals.resolveConflicts(LHSVals))
return false;
// All clear, the live ranges can be merged.
if (RHS.hasSubRanges() || LHS.hasSubRanges()) {
BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
// Transform lanemasks from the LHS to masks in the coalesced register and
// create initial subranges if necessary.
unsigned DstIdx = CP.getDstIdx();
if (!LHS.hasSubRanges()) {
LaneBitmask Mask = DstIdx == 0 ? CP.getNewRC()->getLaneMask()
: TRI->getSubRegIndexLaneMask(DstIdx);
// LHS must support subregs or we wouldn't be in this codepath.
assert(Mask.any());
LHS.createSubRangeFrom(Allocator, Mask, LHS);
} else if (DstIdx != 0) {
// Transform LHS lanemasks to new register class if necessary.
for (LiveInterval::SubRange &R : LHS.subranges()) {
LaneBitmask Mask = TRI->composeSubRegIndexLaneMask(DstIdx, R.LaneMask);
R.LaneMask = Mask;
}
}
LLVM_DEBUG(dbgs() << "\t\tLHST = " << printReg(CP.getDstReg()) << ' ' << LHS
<< '\n');
// Determine lanemasks of RHS in the coalesced register and merge subranges.
unsigned SrcIdx = CP.getSrcIdx();
if (!RHS.hasSubRanges()) {
LaneBitmask Mask = SrcIdx == 0 ? CP.getNewRC()->getLaneMask()
: TRI->getSubRegIndexLaneMask(SrcIdx);
mergeSubRangeInto(LHS, RHS, Mask, CP, DstIdx);
} else {
// Pair up subranges and merge.
for (LiveInterval::SubRange &R : RHS.subranges()) {
LaneBitmask Mask = TRI->composeSubRegIndexLaneMask(SrcIdx, R.LaneMask);
mergeSubRangeInto(LHS, R, Mask, CP, DstIdx);
}
}
LLVM_DEBUG(dbgs() << "\tJoined SubRanges " << LHS << "\n");
// Pruning implicit defs from subranges may result in the main range
// having stale segments.
LHSVals.pruneMainSegments(LHS, ShrinkMainRange);
LHSVals.pruneSubRegValues(LHS, ShrinkMask);
RHSVals.pruneSubRegValues(LHS, ShrinkMask);
}
// The merging algorithm in LiveInterval::join() can't handle conflicting
// value mappings, so we need to remove any live ranges that overlap a
// CR_Replace resolution. Collect a set of end points that can be used to
// restore the live range after joining.
SmallVector<SlotIndex, 8> EndPoints;
LHSVals.pruneValues(RHSVals, EndPoints, true);
RHSVals.pruneValues(LHSVals, EndPoints, true);
// Erase COPY and IMPLICIT_DEF instructions. This may cause some external
// registers to require trimming.
SmallVector<unsigned, 8> ShrinkRegs;
LHSVals.eraseInstrs(ErasedInstrs, ShrinkRegs, &LHS);
RHSVals.eraseInstrs(ErasedInstrs, ShrinkRegs);
while (!ShrinkRegs.empty())
shrinkToUses(&LIS->getInterval(ShrinkRegs.pop_back_val()));
// Scan and mark undef any DBG_VALUEs that would refer to a different value.
checkMergingChangesDbgValues(CP, LHS, LHSVals, RHS, RHSVals);
// Join RHS into LHS.
LHS.join(RHS, LHSVals.getAssignments(), RHSVals.getAssignments(), NewVNInfo);
// Kill flags are going to be wrong if the live ranges were overlapping.
// Eventually, we should simply clear all kill flags when computing live
// ranges. They are reinserted after register allocation.
MRI->clearKillFlags(LHS.reg);
MRI->clearKillFlags(RHS.reg);
if (!EndPoints.empty()) {
// Recompute the parts of the live range we had to remove because of
// CR_Replace conflicts.
LLVM_DEBUG({
dbgs() << "\t\trestoring liveness to " << EndPoints.size() << " points: ";
for (unsigned i = 0, n = EndPoints.size(); i != n; ++i) {
dbgs() << EndPoints[i];
if (i != n-1)
dbgs() << ',';
}
dbgs() << ": " << LHS << '\n';
});
LIS->extendToIndices((LiveRange&)LHS, EndPoints);
}
return true;
}
bool RegisterCoalescer::joinIntervals(CoalescerPair &CP) {
return CP.isPhys() ? joinReservedPhysReg(CP) : joinVirtRegs(CP);
}
void RegisterCoalescer::buildVRegToDbgValueMap(MachineFunction &MF)
{
const SlotIndexes &Slots = *LIS->getSlotIndexes();
SmallVector<MachineInstr *, 8> ToInsert;
// After collecting a block of DBG_VALUEs into ToInsert, enter them into the
// vreg => DbgValueLoc map.
auto CloseNewDVRange = [this, &ToInsert](SlotIndex Slot) {
for (auto *X : ToInsert)
DbgVRegToValues[X->getOperand(0).getReg()].push_back({Slot, X});
ToInsert.clear();
};
// Iterate over all instructions, collecting them into the ToInsert vector.
// Once a non-debug instruction is found, record the slot index of the
// collected DBG_VALUEs.
for (auto &MBB : MF) {
SlotIndex CurrentSlot = Slots.getMBBStartIdx(&MBB);
for (auto &MI : MBB) {
if (MI.isDebugValue() && MI.getOperand(0).isReg() &&
MI.getOperand(0).getReg().isVirtual()) {
ToInsert.push_back(&MI);
} else if (!MI.isDebugInstr()) {
CurrentSlot = Slots.getInstructionIndex(MI);
CloseNewDVRange(CurrentSlot);
}
}
// Close range of DBG_VALUEs at the end of blocks.
CloseNewDVRange(Slots.getMBBEndIdx(&MBB));
}
// Sort all DBG_VALUEs we've seen by slot number.
for (auto &Pair : DbgVRegToValues)
llvm::sort(Pair.second);
}
void RegisterCoalescer::checkMergingChangesDbgValues(CoalescerPair &CP,
LiveRange &LHS,
JoinVals &LHSVals,
LiveRange &RHS,
JoinVals &RHSVals) {
auto ScanForDstReg = [&](unsigned Reg) {
checkMergingChangesDbgValuesImpl(Reg, RHS, LHS, LHSVals);
};
auto ScanForSrcReg = [&](unsigned Reg) {
checkMergingChangesDbgValuesImpl(Reg, LHS, RHS, RHSVals);
};
// Scan for potentially unsound DBG_VALUEs: examine first the register number
// Reg, and then any other vregs that may have been merged into it.
auto PerformScan = [this](unsigned Reg, std::function<void(unsigned)> Func) {
Func(Reg);
if (DbgMergedVRegNums.count(Reg))
for (unsigned X : DbgMergedVRegNums[Reg])
Func(X);
};
// Scan for unsound updates of both the source and destination register.
PerformScan(CP.getSrcReg(), ScanForSrcReg);
PerformScan(CP.getDstReg(), ScanForDstReg);
}
void RegisterCoalescer::checkMergingChangesDbgValuesImpl(unsigned Reg,
LiveRange &OtherLR,
LiveRange &RegLR,
JoinVals &RegVals) {
// Are there any DBG_VALUEs to examine?
auto VRegMapIt = DbgVRegToValues.find(Reg);
if (VRegMapIt == DbgVRegToValues.end())
return;
auto &DbgValueSet = VRegMapIt->second;
auto DbgValueSetIt = DbgValueSet.begin();
auto SegmentIt = OtherLR.begin();
bool LastUndefResult = false;
SlotIndex LastUndefIdx;
// If the "Other" register is live at a slot Idx, test whether Reg can
// safely be merged with it, or should be marked undef.
auto ShouldUndef = [&RegVals, &RegLR, &LastUndefResult,
&LastUndefIdx](SlotIndex Idx) -> bool {
// Our worst-case performance typically happens with asan, causing very
// many DBG_VALUEs of the same location. Cache a copy of the most recent
// result for this edge-case.
if (LastUndefIdx == Idx)
return LastUndefResult;
// If the other range was live, and Reg's was not, the register coalescer
// will not have tried to resolve any conflicts. We don't know whether
// the DBG_VALUE will refer to the same value number, so it must be made
// undef.
auto OtherIt = RegLR.find(Idx);
if (OtherIt == RegLR.end())
return true;
// Both the registers were live: examine the conflict resolution record for
// the value number Reg refers to. CR_Keep meant that this value number
// "won" and the merged register definitely refers to that value. CR_Erase
// means the value number was a redundant copy of the other value, which
// was coalesced and Reg deleted. It's safe to refer to the other register
// (which will be the source of the copy).
auto Resolution = RegVals.getResolution(OtherIt->valno->id);
LastUndefResult = Resolution != JoinVals::CR_Keep &&
Resolution != JoinVals::CR_Erase;
LastUndefIdx = Idx;
return LastUndefResult;
};
// Iterate over both the live-range of the "Other" register, and the set of
// DBG_VALUEs for Reg at the same time. Advance whichever one has the lowest
// slot index. This relies on the DbgValueSet being ordered.
while (DbgValueSetIt != DbgValueSet.end() && SegmentIt != OtherLR.end()) {
if (DbgValueSetIt->first < SegmentIt->end) {
// "Other" is live and there is a DBG_VALUE of Reg: test if we should
// set it undef.
if (DbgValueSetIt->first >= SegmentIt->start &&
DbgValueSetIt->second->getOperand(0).getReg() != 0 &&
ShouldUndef(DbgValueSetIt->first)) {
// Mark undef, erase record of this DBG_VALUE to avoid revisiting.
DbgValueSetIt->second->getOperand(0).setReg(0);
continue;
}
++DbgValueSetIt;
} else {
++SegmentIt;
}
}
}
namespace {
/// Information concerning MBB coalescing priority.
struct MBBPriorityInfo {
MachineBasicBlock *MBB;
unsigned Depth;
bool IsSplit;
MBBPriorityInfo(MachineBasicBlock *mbb, unsigned depth, bool issplit)
: MBB(mbb), Depth(depth), IsSplit(issplit) {}
};
} // end anonymous namespace
/// C-style comparator that sorts first based on the loop depth of the basic
/// block (the unsigned), and then on the MBB number.
///
/// EnableGlobalCopies assumes that the primary sort key is loop depth.
static int compareMBBPriority(const MBBPriorityInfo *LHS,
const MBBPriorityInfo *RHS) {
// Deeper loops first
if (LHS->Depth != RHS->Depth)
return LHS->Depth > RHS->Depth ? -1 : 1;
// Try to unsplit critical edges next.
if (LHS->IsSplit != RHS->IsSplit)
return LHS->IsSplit ? -1 : 1;
// Prefer blocks that are more connected in the CFG. This takes care of
// the most difficult copies first while intervals are short.
unsigned cl = LHS->MBB->pred_size() + LHS->MBB->succ_size();
unsigned cr = RHS->MBB->pred_size() + RHS->MBB->succ_size();
if (cl != cr)
return cl > cr ? -1 : 1;
// As a last resort, sort by block number.
return LHS->MBB->getNumber() < RHS->MBB->getNumber() ? -1 : 1;
}
/// \returns true if the given copy uses or defines a local live range.
static bool isLocalCopy(MachineInstr *Copy, const LiveIntervals *LIS) {
if (!Copy->isCopy())
return false;
if (Copy->getOperand(1).isUndef())
return false;
Register SrcReg = Copy->getOperand(1).getReg();
Register DstReg = Copy->getOperand(0).getReg();
if (Register::isPhysicalRegister(SrcReg) ||
Register::isPhysicalRegister(DstReg))
return false;
return LIS->intervalIsInOneMBB(LIS->getInterval(SrcReg))
|| LIS->intervalIsInOneMBB(LIS->getInterval(DstReg));
}
void RegisterCoalescer::lateLiveIntervalUpdate() {
for (unsigned reg : ToBeUpdated) {
if (!LIS->hasInterval(reg))
continue;
LiveInterval &LI = LIS->getInterval(reg);
shrinkToUses(&LI, &DeadDefs);
if (!DeadDefs.empty())
eliminateDeadDefs();
}
ToBeUpdated.clear();
}
bool RegisterCoalescer::
copyCoalesceWorkList(MutableArrayRef<MachineInstr*> CurrList) {
bool Progress = false;
for (unsigned i = 0, e = CurrList.size(); i != e; ++i) {
if (!CurrList[i])
continue;
// Skip instruction pointers that have already been erased, for example by
// dead code elimination.
if (ErasedInstrs.count(CurrList[i])) {
CurrList[i] = nullptr;
continue;
}
bool Again = false;
bool Success = joinCopy(CurrList[i], Again);
Progress |= Success;
if (Success || !Again)
CurrList[i] = nullptr;
}
return Progress;
}
/// Check if DstReg is a terminal node.
/// I.e., it does not have any affinity other than \p Copy.
static bool isTerminalReg(unsigned DstReg, const MachineInstr &Copy,
const MachineRegisterInfo *MRI) {
assert(Copy.isCopyLike());
// Check if the destination of this copy as any other affinity.
for (const MachineInstr &MI : MRI->reg_nodbg_instructions(DstReg))
if (&MI != &Copy && MI.isCopyLike())
return false;
return true;
}
bool RegisterCoalescer::applyTerminalRule(const MachineInstr &Copy) const {
assert(Copy.isCopyLike());
if (!UseTerminalRule)
return false;
unsigned DstReg, DstSubReg, SrcReg, SrcSubReg;
if (!isMoveInstr(*TRI, &Copy, SrcReg, DstReg, SrcSubReg, DstSubReg))
return false;
// Check if the destination of this copy has any other affinity.
if (Register::isPhysicalRegister(DstReg) ||
// If SrcReg is a physical register, the copy won't be coalesced.
// Ignoring it may have other side effect (like missing
// rematerialization). So keep it.
Register::isPhysicalRegister(SrcReg) || !isTerminalReg(DstReg, Copy, MRI))
return false;
// DstReg is a terminal node. Check if it interferes with any other
// copy involving SrcReg.
const MachineBasicBlock *OrigBB = Copy.getParent();
const LiveInterval &DstLI = LIS->getInterval(DstReg);
for (const MachineInstr &MI : MRI->reg_nodbg_instructions(SrcReg)) {
// Technically we should check if the weight of the new copy is
// interesting compared to the other one and update the weight
// of the copies accordingly. However, this would only work if
// we would gather all the copies first then coalesce, whereas
// right now we interleave both actions.
// For now, just consider the copies that are in the same block.
if (&MI == &Copy || !MI.isCopyLike() || MI.getParent() != OrigBB)
continue;
unsigned OtherReg, OtherSubReg, OtherSrcReg, OtherSrcSubReg;
if (!isMoveInstr(*TRI, &Copy, OtherSrcReg, OtherReg, OtherSrcSubReg,
OtherSubReg))
return false;
if (OtherReg == SrcReg)
OtherReg = OtherSrcReg;
// Check if OtherReg is a non-terminal.
if (Register::isPhysicalRegister(OtherReg) ||
isTerminalReg(OtherReg, MI, MRI))
continue;
// Check that OtherReg interfere with DstReg.
if (LIS->getInterval(OtherReg).overlaps(DstLI)) {
LLVM_DEBUG(dbgs() << "Apply terminal rule for: " << printReg(DstReg)
<< '\n');
return true;
}
}
return false;
}
void
RegisterCoalescer::copyCoalesceInMBB(MachineBasicBlock *MBB) {
LLVM_DEBUG(dbgs() << MBB->getName() << ":\n");
// Collect all copy-like instructions in MBB. Don't start coalescing anything
// yet, it might invalidate the iterator.
const unsigned PrevSize = WorkList.size();
if (JoinGlobalCopies) {
SmallVector<MachineInstr*, 2> LocalTerminals;
SmallVector<MachineInstr*, 2> GlobalTerminals;
// Coalesce copies bottom-up to coalesce local defs before local uses. They
// are not inherently easier to resolve, but slightly preferable until we
// have local live range splitting. In particular this is required by
// cmp+jmp macro fusion.
for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
MII != E; ++MII) {
if (!MII->isCopyLike())
continue;
bool ApplyTerminalRule = applyTerminalRule(*MII);
if (isLocalCopy(&(*MII), LIS)) {
if (ApplyTerminalRule)
LocalTerminals.push_back(&(*MII));
else
LocalWorkList.push_back(&(*MII));
} else {
if (ApplyTerminalRule)
GlobalTerminals.push_back(&(*MII));
else
WorkList.push_back(&(*MII));
}
}
// Append the copies evicted by the terminal rule at the end of the list.
LocalWorkList.append(LocalTerminals.begin(), LocalTerminals.end());
WorkList.append(GlobalTerminals.begin(), GlobalTerminals.end());
}
else {
SmallVector<MachineInstr*, 2> Terminals;
for (MachineInstr &MII : *MBB)
if (MII.isCopyLike()) {
if (applyTerminalRule(MII))
Terminals.push_back(&MII);
else
WorkList.push_back(&MII);
}
// Append the copies evicted by the terminal rule at the end of the list.
WorkList.append(Terminals.begin(), Terminals.end());
}
// Try coalescing the collected copies immediately, and remove the nulls.
// This prevents the WorkList from getting too large since most copies are
// joinable on the first attempt.
MutableArrayRef<MachineInstr*>
CurrList(WorkList.begin() + PrevSize, WorkList.end());
if (copyCoalesceWorkList(CurrList))
WorkList.erase(std::remove(WorkList.begin() + PrevSize, WorkList.end(),
nullptr), WorkList.end());
}
void RegisterCoalescer::coalesceLocals() {
copyCoalesceWorkList(LocalWorkList);
for (unsigned j = 0, je = LocalWorkList.size(); j != je; ++j) {
if (LocalWorkList[j])
WorkList.push_back(LocalWorkList[j]);
}
LocalWorkList.clear();
}
void RegisterCoalescer::joinAllIntervals() {
LLVM_DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
assert(WorkList.empty() && LocalWorkList.empty() && "Old data still around.");
std::vector<MBBPriorityInfo> MBBs;
MBBs.reserve(MF->size());
for (MachineFunction::iterator I = MF->begin(), E = MF->end(); I != E; ++I) {
MachineBasicBlock *MBB = &*I;
MBBs.push_back(MBBPriorityInfo(MBB, Loops->getLoopDepth(MBB),
JoinSplitEdges && isSplitEdge(MBB)));
}
array_pod_sort(MBBs.begin(), MBBs.end(), compareMBBPriority);
// Coalesce intervals in MBB priority order.
unsigned CurrDepth = std::numeric_limits<unsigned>::max();
for (unsigned i = 0, e = MBBs.size(); i != e; ++i) {
// Try coalescing the collected local copies for deeper loops.
if (JoinGlobalCopies && MBBs[i].Depth < CurrDepth) {
coalesceLocals();
CurrDepth = MBBs[i].Depth;
}
copyCoalesceInMBB(MBBs[i].MBB);
}
lateLiveIntervalUpdate();
coalesceLocals();
// Joining intervals can allow other intervals to be joined. Iteratively join
// until we make no progress.
while (copyCoalesceWorkList(WorkList))
/* empty */ ;
lateLiveIntervalUpdate();
}
void RegisterCoalescer::releaseMemory() {
ErasedInstrs.clear();
WorkList.clear();
DeadDefs.clear();
InflateRegs.clear();
LargeLIVisitCounter.clear();
}
bool RegisterCoalescer::runOnMachineFunction(MachineFunction &fn) {
MF = &fn;
MRI = &fn.getRegInfo();
const TargetSubtargetInfo &STI = fn.getSubtarget();
TRI = STI.getRegisterInfo();
TII = STI.getInstrInfo();
LIS = &getAnalysis<LiveIntervals>();
AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
Loops = &getAnalysis<MachineLoopInfo>();
if (EnableGlobalCopies == cl::BOU_UNSET)
JoinGlobalCopies = STI.enableJoinGlobalCopies();
else
JoinGlobalCopies = (EnableGlobalCopies == cl::BOU_TRUE);
// The MachineScheduler does not currently require JoinSplitEdges. This will
// either be enabled unconditionally or replaced by a more general live range
// splitting optimization.
JoinSplitEdges = EnableJoinSplits;
LLVM_DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
<< "********** Function: " << MF->getName() << '\n');
if (VerifyCoalescing)
MF->verify(this, "Before register coalescing");
DbgVRegToValues.clear();
DbgMergedVRegNums.clear();
buildVRegToDbgValueMap(fn);
RegClassInfo.runOnMachineFunction(fn);
// Join (coalesce) intervals if requested.
if (EnableJoining)
joinAllIntervals();
// After deleting a lot of copies, register classes may be less constrained.
// Removing sub-register operands may allow GR32_ABCD -> GR32 and DPR_VFP2 ->
// DPR inflation.
array_pod_sort(InflateRegs.begin(), InflateRegs.end());
InflateRegs.erase(std::unique(InflateRegs.begin(), InflateRegs.end()),
InflateRegs.end());
LLVM_DEBUG(dbgs() << "Trying to inflate " << InflateRegs.size()
<< " regs.\n");
for (unsigned i = 0, e = InflateRegs.size(); i != e; ++i) {
unsigned Reg = InflateRegs[i];
if (MRI->reg_nodbg_empty(Reg))
continue;
if (MRI->recomputeRegClass(Reg)) {
LLVM_DEBUG(dbgs() << printReg(Reg) << " inflated to "
<< TRI->getRegClassName(MRI->getRegClass(Reg)) << '\n');
++NumInflated;
LiveInterval &LI = LIS->getInterval(Reg);
if (LI.hasSubRanges()) {
// If the inflated register class does not support subregisters anymore
// remove the subranges.
if (!MRI->shouldTrackSubRegLiveness(Reg)) {
LI.clearSubRanges();
} else {
#ifndef NDEBUG
LaneBitmask MaxMask = MRI->getMaxLaneMaskForVReg(Reg);
// If subranges are still supported, then the same subregs
// should still be supported.
for (LiveInterval::SubRange &S : LI.subranges()) {
assert((S.LaneMask & ~MaxMask).none());
}
#endif
}
}
}
}
LLVM_DEBUG(dump());
if (VerifyCoalescing)
MF->verify(this, "After register coalescing");
return true;
}
void RegisterCoalescer::print(raw_ostream &O, const Module* m) const {
LIS->print(O, m);
}