blob: 7e9c992031f8d3fc061a35d024da8787be6b86cb [file] [log] [blame]
//===- RegisterCoalescer.cpp - Generic Register Coalescing Interface ------===//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See 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>
cl::desc("Coalesce copies on split edges (default=subtarget)"), cl::Hidden);
/// Temporary flag to test global copy optimization.
static cl::opt<cl::boolOrDefault>
cl::desc("Coalesce copies that span blocks (default=subtarget)"),
cl::init(cl::BOU_UNSET), cl::Hidden);
static cl::opt<bool>
cl::desc("Verify machine instrs before and after register coalescing"),
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. "),
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. "),
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. "),
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;
/// Position and VReg of a PHI instruction during coalescing.
struct PHIValPos {
SlotIndex SI; ///< Slot where this PHI occurs.
Register Reg; ///< VReg the PHI occurs in.
unsigned SubReg; ///< Qualifying subregister for Reg.
/// Map from debug instruction number to PHI position during coalescing.
DenseMap<unsigned, PHIValPos> PHIValToPos;
/// Index of, for each VReg, which debug instruction numbers and
/// corresponding PHIs are sensitive to coalescing. Each VReg may have
/// multiple PHI defs, at different positions.
DenseMap<Register, SmallVector<unsigned, 2>> RegToPHIIdx;
/// 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<Register, std::vector<DbgValueLoc>> DbgVRegToValues;
/// 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<Register, 8> InflateRegs;
/// The collection of live intervals which should have been updated
/// immediately after rematerialiation but delayed until
/// lateLiveIntervalUpdate is called.
DenseSet<Register> ToBeUpdated;
/// Record how many times the large live interval with many valnos
/// has been tried to join with other live interval.
DenseMap<Register, unsigned long> LargeLIVisitCounter;
/// Recursively eliminate dead defs in DeadDefs.
void eliminateDeadDefs(LiveRangeEdit *Edit = nullptr);
/// 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();
/// Check if the incoming value defined by a COPY at \p SLRQ in the subrange
/// has no value defined in the predecessors. If the incoming value is the
/// same as defined by the copy itself, the value is considered undefined.
bool copyValueUndefInPredecessors(LiveRange &S,
const MachineBasicBlock *MBB,
LiveQueryResult SLRQ);
/// Set necessary undef flags on subregister uses after pruning out undef
/// lane segments from the subrange.
void setUndefOnPrunedSubRegUses(LiveInterval &LI, Register Reg,
LaneBitmask PrunedLanes);
/// 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,
SmallPtrSetImpl<MachineInstr *> &CurrentErasedInstrs);
/// 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(Register SrcReg, Register 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) {
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) {
/// 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(Register Reg, LiveRange &OtherRange,
LiveRange &RegRange, JoinVals &Vals2);
static char ID; ///< Class identification, replacement for typeinfo
RegisterCoalescer() : MachineFunctionPass(ID) {
void getAnalysisUsage(AnalysisUsage &AU) const override;
MachineFunctionProperties getClearedProperties() const override {
return MachineFunctionProperties().set(
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, "register-coalescer",
"Register Coalescer", false, false)
INITIALIZE_PASS_END(RegisterCoalescer, "register-coalescer",
"Register Coalescer", false, false)
[[nodiscard]] static bool isMoveInstr(const TargetRegisterInfo &tri,
const MachineInstr *MI, Register &Src,
Register &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(),
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 = Register();
SrcIdx = DstIdx = 0;
NewRC = nullptr;
Flipped = CrossClass = false;
Register Src, Dst;
unsigned SrcSub = 0, DstSub = 0;
if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
return false;
Partial = SrcSub || DstSub;
// If one register is a physreg, it must be Dst.
if (Src.isPhysical()) {
if (Dst.isPhysical())
return false;
std::swap(Src, Dst);
std::swap(SrcSub, DstSub);
Flipped = true;
const MachineRegisterInfo &MRI = MI->getMF()->getRegInfo();
if (Dst.isPhysical()) {
// 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(Src.isVirtual() && "Src must be virtual");
assert(!(Dst.isPhysical() && DstSub) && "Cannot have a physical SubIdx");
SrcReg = Src;
DstReg = Dst;
return true;
bool CoalescerPair::flip() {
if (DstReg.isPhysical())
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;
Register Src, Dst;
unsigned SrcSub = 0, DstSub = 0;
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 (DstReg.isPhysical()) {
if (!Dst.isPhysical())
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 Register(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 {
void RegisterCoalescer::eliminateDeadDefs(LiveRangeEdit *Edit) {
if (Edit) {
SmallVector<Register, 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.
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 =
if (ValS == IntB.end())
return false;
// Make sure that the end of the live segment is inside the same block as
// CopyMI.
MachineInstr *ValSEndInst =
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);
// The subrange may have ended before FillerStart. If so, extend it.
if (!S.getVNInfoAt(FillerStart)) {
SlotIndex BBStart =
S.extendInBlock(BBStart, FillerStart);
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) {
// 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;
if (RecomputeLiveRange)
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())
for (; BI != IntB.end() && ASeg.end >= BI->start; ++BI) {
if (BI->valno == BValNo)
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)
// 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);
RegisterCoalescer::removeCopyByCommutingDef(const CoalescerPair &CP,
MachineInstr *CopyMI) {
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)
// 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 (IntA.reg().isVirtual() && IntB.reg().isVirtual() &&
!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);
// 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 (MachineOperand &UseMO :
llvm::make_early_inc_range(MRI->use_operands(IntA.reg()))) {
if (UseMO.isUndef())
MachineInstr *UseMI = UseMO.getParent();
if (UseMI->isDebugInstr()) {
// 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.
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)
// Kill flags are no longer accurate. They are recomputed after RA.
if (NewReg.isPhysical())
UseMO.substPhysReg(NewReg, *TRI);
if (UseMI == CopyMI)
if (!UseMI->isCopy())
if (UseMI->getOperand(0).getReg() != IntB.reg() ||
// 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)
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)
VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
assert(SubBValNo->def == CopyIdx);
S.MergeValueNumberInto(SubDVNI, SubBValNo);
// 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)
MaskA |= SA.LaneMask;
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())
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');
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) {
if (!CopyMI.isFullCopy())
return false;
MachineBasicBlock &MBB = *CopyMI.getParent();
// If this block is the target of an invoke/inlineasm_br, moving the copy into
// the predecessor is tricker, and we don't handle it.
if (MBB.isEHPad() || MBB.isInlineAsmBrIndirectTarget())
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;
// 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;
// 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())
if (PVal->def < VNI->def && VNI->def < LIS->getMBBEndIdx(Pred)) {
ValB_Changed = true;
if (ValB_Changed) {
CopyLeftBB = Pred;
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())
SlotIndex NewCopyIdx =
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.
} else {
LLVM_DEBUG(dbgs() << "\tremovePartialRedundancy: Remove the copy from "
<< printMBBReference(MBB) << '\t' << CopyMI);
const bool IsUndefCopy = CopyMI.getOperand(1).isUndef();
// 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.
// Update the liveness.
SmallVector<SlotIndex, 8> EndPoints;
VNInfo *BValNo = IntB.Query(CopyIdx).valueOutOrDead();
LIS->pruneValue(*static_cast<LiveRange *>(&IntB), CopyIdx.getRegSlot(),
if (IsUndefCopy) {
// We're introducing an undef phi def, and need to set undef on any users of
// the previously local def to avoid artifically extending the lifetime
// through the block.
for (MachineOperand &MO : MRI->use_nodbg_operands(IntB.reg())) {
const MachineInstr &MI = *MO.getParent();
SlotIndex UseIdx = LIS->getInstructionIndex(MI);
if (!IntB.liveAt(UseIdx))
// 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()) {
VNInfo *BValNo = SR.Query(CopyIdx).valueOutOrDead();
assert(BValNo && "All sublanes should be live");
LIS->pruneValue(SR, CopyIdx.getRegSlot(), &EndPoints);
// 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();
SmallVector<SlotIndex, 8> Undefs;
IntB.computeSubRangeUndefs(Undefs, SR.LaneMask, *MRI,
LIS->extendToIndices(SR, EndPoints, Undefs);
// If any dead defs were extended, truncate them.
// Finally, update the live-range of 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, Register Reg) {
assert(!Reg.isPhysical() && "This code cannot handle physreg aliasing");
for (const MachineOperand &Op : MI.all_defs()) {
if (Op.getReg() != Reg)
// 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;
Register SrcReg = CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg();
unsigned SrcIdx = CP.isFlipped() ? CP.getDstIdx() : CP.getSrcIdx();
Register DstReg = CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg();
unsigned DstIdx = CP.isFlipped() ? CP.getSrcIdx() : CP.getDstIdx();
if (SrcReg.isPhysical())
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;
SmallVector<Register, 8> NewRegs;
LiveRangeEdit Edit(&SrcInt, NewRegs, *MF, *LIS, nullptr, this);
if (!Edit.checkRematerializable(ValNo, DefMI))
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;
[[maybe_unused]] const unsigned DefSubIdx = DefMI->getOperand(0).getSubReg();
const TargetRegisterClass *DefRC = TII->getRegClass(MCID, 0, TRI, *MF);
if (!DefMI->isImplicitDef()) {
if (DstReg.isPhysical()) {
Register NewDstReg = DstReg;
unsigned NewDstIdx = TRI->composeSubRegIndices(CP.getSrcIdx(),
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(DstReg.isVirtual() &&
"Only expect to deal with virtual or physical registers");
LiveRangeEdit::Remat RM(ValNo);
RM.OrigMI = DefMI;
if (!Edit.canRematerializeAt(RM, ValNo, CopyIdx, true))
return false;
DebugLoc DL = CopyMI->getDebugLoc();
MachineBasicBlock *MBB = CopyMI->getParent();
MachineBasicBlock::iterator MII =
Edit.rematerializeAt(*MBB, MII, DstReg, RM, *TRI, false, SrcIdx, CopyMI);
MachineInstr &NewMI = *std::prev(MII);
// 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;
// Instruction might contain "undef %0:subreg" as use operand:
// %0:subreg = instr op_1, ..., op_N, undef %0:subreg, op_N+2, ...
// Need to check all operands.
for (MachineOperand &MO : NewMI.operands()) {
if (MO.isReg() && MO.getReg() == DstReg && MO.getSubReg() == DstIdx) {
DstIdx = 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() -
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.");
assert((MO.getReg().isPhysical() ||
(MO.getSubReg() == 0 && MO.getReg() == DstOperand.getReg())) &&
"unexpected implicit virtual register def");
// 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.
// We also expect to have tied implicit-defs of super registers originating
// from SUBREG_TO_REG, such as:
// $edi = MOV32r0 implicit-def dead $eflags, implicit-def $rdi
// undef %0.sub_32bit = MOV32r0 implicit-def dead $eflags, implicit-def %0
// The implicit-def of the super register may have been reduced to
// subregisters depending on the uses.
bool NewMIDefinesFullReg = false;
SmallVector<MCRegister, 4> NewMIImplDefs;
for (unsigned i = NewMI.getDesc().getNumOperands(),
e = NewMI.getNumOperands();
i != e; ++i) {
MachineOperand &MO = NewMI.getOperand(i);
if (MO.isReg() && MO.isDef()) {
if (MO.getReg().isPhysical()) {
if (MO.getReg() == DstReg)
NewMIDefinesFullReg = true;
assert(MO.isImplicit() && MO.getReg().isPhysical() &&
(MO.isDead() ||
(DefSubIdx &&
((TRI->getSubReg(MO.getReg(), DefSubIdx) ==
MCRegister((unsigned)NewMI.getOperand(0).getReg())) ||
} else {
assert(MO.getReg() == NewMI.getOperand(0).getReg());
// We're only expecting another def of the main output, so the range
// should get updated with the regular output range.
// FIXME: The range updating below probably needs updating to look at
// the super register if subranges are tracked.
assert(!MRI->shouldTrackSubRegLiveness(DstReg) &&
"subrange update for implicit-def of super register may not be "
"properly handled");
if (DstReg.isVirtual()) {
unsigned NewIdx = NewMI.getOperand(0).getSubReg();
if (DefRC != nullptr) {
if (NewIdx)
NewRC = TRI->getMatchingSuperRegClass(NewRC, DefRC, NewIdx);
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);
// 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)
// 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 =
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;
SlotIndex DefIndex =
VNInfo::Allocator &Alloc = LIS->getVNInfoAllocator();
for (LiveInterval::SubRange &SR : DstInt.subranges()) {
if ((SR.LaneMask & DstMask).none()) {
<< "Removing undefined SubRange "
<< PrintLaneMask(SR.LaneMask) << " : " << SR << "\n");
if (VNInfo *RmValNo = SR.getVNInfoAt(CurrIdx.getRegSlot())) {
// VNI is in ValNo - remove any segments in this SubRange that have
// this ValNo
// We may not have a defined value at this point, but still need to
// clear out any empty subranges tentatively created by
// updateRegDefUses. The original subrange def may have only undefed
// some lanes.
UpdatedSubRanges = true;
} else {
// We know that this lane is defined by this instruction,
// but at this point it may be empty because it is not used by
// anything. This happens when updateRegDefUses adds the missing
// lanes. Assign that lane a dead def so that the interferences
// are properly modeled.
if (SR.empty())
SR.createDeadDef(DefIndex, Alloc);
if (UpdatedSubRanges)
} 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(DstReg.isPhysical() &&
"Only expect virtual or physical registers in remat");
if (!NewMIDefinesFullReg) {
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 (MCRegUnit Unit : TRI->regunits(NewMI.getOperand(0).getReg()))
if (LiveRange *LR = LIS->getCachedRegUnit(Unit))
LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
// Transfer over implicit operands to the rematerialized instruction.
for (MachineOperand &MO : ImplicitOps)
SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
for (MCRegister Reg : NewMIImplDefs) {
for (MCRegUnit Unit : TRI->regunits(Reg))
if (LiveRange *LR = LIS->getCachedRegUnit(Unit))
LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
LLVM_DEBUG(dbgs() << "Remat: " << NewMI);
// If the virtual SrcReg is completely eliminated, update all DBG_VALUEs
// to describe DstReg instead.
if (MRI->use_nodbg_empty(SrcReg)) {
for (MachineOperand &UseMO :
llvm::make_early_inc_range(MRI->use_operands(SrcReg))) {
MachineInstr *UseMI = UseMO.getParent();
if (UseMI->isDebugInstr()) {
if (DstReg.isPhysical())
UseMO.substPhysReg(DstReg, *TRI);
// 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())
if (NumCopyUses < LateRematUpdateThreshold) {
// The source interval can become smaller because we removed a use.
shrinkToUses(&SrcInt, &DeadDefs);
if (!DeadDefs.empty())
} else {
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.
Register SrcReg, DstReg;
unsigned SrcSubIdx = 0, DstSubIdx = 0;
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())
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");
VNInfo *V = DstLI.getVNInfoAt(Seg->end);
// The source interval may also have been on an undef use, in which case the
// copy introduced a live value.
if (((V && V->isPHIDef()) || (!V && !DstLI.liveAt(Idx)))) {
for (unsigned i = CopyMI->getNumOperands(); i != 0; --i) {
MachineOperand &MO = CopyMI->getOperand(i-1);
if (MO.isReg()) {
if (MO.isUse())
CopyMI->removeOperand(i - 1);
} else {
assert(MO.isImm() &&
CopyMI->getOpcode() == TargetOpcode::SUBREG_TO_REG);
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())
VNInfo *SVNI = SR.getVNInfoAt(RegIndex);
assert(SVNI != nullptr && SlotIndex::isSameInstr(SVNI->def, RegIndex));
} else
LIS->removeVRegDefAt(DstLI, RegIndex);
// Mark uses as undef.
for (MachineOperand &MO : MRI->reg_nodbg_operands(DstReg)) {
if (MO.isDef() /*|| MO.isUndef()*/)
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())
if (SR.liveAt(UseIdx)) {
isLive = true;
} else
isLive = DstLI.liveAt(UseIdx);
if (isLive)
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->all_defs())
if (MO.getReg() == DstReg)
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())
if (S.liveAt(UseIdx)) {
IsUndef = false;
if (IsUndef) {
// 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(Register SrcReg, Register DstReg,
unsigned SubIdx) {
bool DstIsPhys = DstReg.isPhysical();
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())
MachineInstr &MI = *MO.getParent();
if (MI.isDebugInstr())
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)
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->isDebugInstr())
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())
// 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 (MO.isUse() && !DstIsPhys) {
unsigned SubUseIdx = TRI->composeSubRegIndices(SubIdx, MO.getSubReg());
if (SubUseIdx != 0 && 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->isDebugInstr()
? LIS->getSlotIndexes()->getIndexBefore(*UseMI)
: LIS->getInstructionIndex(*UseMI);
SlotIndex UseIdx = MIIdx.getRegSlot(true);
addUndefFlag(*DstInt, UseIdx, MO, SubUseIdx);
if (DstIsPhys)
MO.substPhysReg(DstReg, *TRI);
MO.substVirtReg(DstReg, SubIdx, *TRI);
dbgs() << "\t\tupdated: ";
if (!UseMI->isDebugInstr())
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;
dbgs() << "\tCannot join complex intervals into reserved register.\n");
return false;
bool RegisterCoalescer::copyValueUndefInPredecessors(
LiveRange &S, const MachineBasicBlock *MBB, LiveQueryResult SLRQ) {
for (const MachineBasicBlock *Pred : MBB->predecessors()) {
SlotIndex PredEnd = LIS->getMBBEndIdx(Pred);
if (VNInfo *V = S.getVNInfoAt(PredEnd.getPrevSlot())) {
// If this is a self loop, we may be reading the same value.
if (V->id != SLRQ.valueOutOrDead()->id)
return false;
return true;
void RegisterCoalescer::setUndefOnPrunedSubRegUses(LiveInterval &LI,
Register Reg,
LaneBitmask PrunedLanes) {
// If we had other instructions in the segment reading the undef sublane
// value, we need to mark them with undef.
for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) {
unsigned SubRegIdx = MO.getSubReg();
if (SubRegIdx == 0 || MO.isUndef())
LaneBitmask SubRegMask = TRI->getSubRegIndexLaneMask(SubRegIdx);
SlotIndex Pos = LIS->getInstructionIndex(*MO.getParent());
for (LiveInterval::SubRange &S : LI.subranges()) {
if (!S.liveAt(Pos) && (PrunedLanes & SubRegMask).any()) {
// A def of a subregister may be a use of other register lanes. Replacing
// such a def with a def of a different register will eliminate the use,
// and may cause the recorded live range to be larger than the actual
// liveness in the program IR.
bool RegisterCoalescer::joinCopy(
MachineInstr *CopyMI, bool &Again,
SmallPtrSetImpl<MachineInstr *> &CurrentErasedInstrs) {
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");
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;
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.");
// Track incoming undef lanes we need to eliminate from the subrange.
LaneBitmask PrunedLanes;
MachineBasicBlock *MBB = CopyMI->getParent();
// Process subregister liveranges.
for (LiveInterval::SubRange &S : LI.subranges()) {
LiveQueryResult SLRQ = S.Query(CopyIdx);
if (VNInfo *SDefVNI = SLRQ.valueDefined()) {
if (VNInfo *SReadVNI = SLRQ.valueIn())
SDefVNI = S.MergeValueNumberInto(SDefVNI, SReadVNI);
// If this copy introduced an undef subrange from an incoming value,
// we need to eliminate the undef live in values from the subrange.
if (copyValueUndefInPredecessors(S, MBB, SLRQ)) {
LLVM_DEBUG(dbgs() << "Incoming sublane value is undef at copy\n");
PrunedLanes |= S.LaneMask;
LI.MergeValueNumberInto(DefVNI, ReadVNI);
if (PrunedLanes.any()) {
LLVM_DEBUG(dbgs() << "Pruning undef incoming lanes: "
<< PrunedLanes << '\n');
setUndefOnPrunedSubRegUses(LI, CP.getSrcReg(), PrunedLanes);
LLVM_DEBUG(dbgs() << "\tMerged values: " << LI << '\n');
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 = false;
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() >
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';
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 = false;
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) {
if (Shrink) {
Register DstReg = CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg();
LiveInterval &DstLI = LIS->getInterval(DstReg);
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()) {
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()))
// 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.
if (ErasedInstrs.erase(CopyMI))
// But we may encounter the instruction again in this iteration.
// 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())
LLVM_DEBUG(dbgs() << "Shrink LaneUses (Lane " << PrintLaneMask(S.LaneMask)
<< ")\n");
LIS->shrinkToUses(S, LI.reg());
ShrinkMainRange = true;
// 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());
// SrcReg is guaranteed to be the register whose live interval that is
// being merged.
// Update regalloc hint.
TRI->updateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *MF);
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);
dbgs() << LIS->getInterval(CP.getDstReg());
dbgs() << '\n';
return true;
bool RegisterCoalescer::joinReservedPhysReg(CoalescerPair &CP) {
Register DstReg = CP.getDstReg();
Register 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 (MCRegUnit Unit : TRI->regunits(DstReg)) {
// Abort if not all the regunits are reserved.
for (MCRegUnitRootIterator RI(Unit, TRI); RI.isValid(); ++RI) {
if (!MRI->isReserved(*RI))
return false;
if (RHS.overlaps(LIS->getRegUnit(Unit))) {
LLVM_DEBUG(dbgs() << "\t\tInterference: " << printRegUnit(Unit, 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.asMCReg(), CopyRegIdx);
// Create a new dead def at the new def location.
for (MCRegUnit Unit : TRI->regunits(DstReg)) {
LiveRange &LR = LIS->getRegUnit(Unit);
LR.createDeadDef(DestRegIdx, LIS->getVNInfoAllocator());
// We don't track kills for reserved registers.
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 Register 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;
/// Conflict resolution for overlapping values.
enum ConflictResolution {
/// No overlap, simply keep this value.
/// Merge this value into OtherVNI and erase the defining instruction.
/// Used for IMPLICIT_DEF, coalescable copies, and copies from external
/// values.
/// Merge this value into OtherVNI but keep the defining instruction.
/// This is for the special case where OtherVNI is defined by the same
/// instruction.
/// 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.
/// Unresolved conflict. Visit later when all values have been mapped.
/// Unresolvable conflict. Abort the join.
/// 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(); }
/// Mark this value as an IMPLICIT_DEF which must be kept as if it were an
/// ordinary value.
void mustKeepImplicitDef(const TargetRegisterInfo &TRI,
const MachineInstr &ImpDef) {
ErasableImplicitDef = false;
ValidLanes = TRI.getSubRegIndexLaneMask(ImpDef.getOperand(0).getSubReg());
/// 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 *, Register> 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.
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, Register, 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);
JoinVals(LiveRange &LR, Register 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<Register> &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; }
/// 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->all_defs()) {
if (MO.getReg() != Reg)
L |= TRI->getSubRegIndexLaneMask(
TRI->composeSubRegIndices(SubIdx, MO.getSubReg()));
if (MO.readsReg())
Redef = true;
return L;
std::pair<const VNInfo *, Register>
JoinVals::followCopyChain(const VNInfo *VNI) const {
Register 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 (!SrcReg.isVirtual())
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())
LiveQueryResult LRQ = S.Query(Def);
if (!ValueIn) {
ValueIn = LRQ.valueIn();
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;
Register Reg0;
std::tie(Orig0, Reg0) = followCopyChain(Value0);
if (Orig0 == Value1 && Reg0 == Other.Reg)
return true;
const VNInfo *Orig1;
Register 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::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.