blob: 88517ee663d1eb97f1d15a7681396aa1bb95d223 [file] [log] [blame]
//===- HexagonConstPropagation.cpp ----------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "HexagonInstrInfo.h"
#include "HexagonRegisterInfo.h"
#include "HexagonSubtarget.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringRef.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/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Type.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include <cassert>
#include <cstdint>
#include <cstring>
#include <iterator>
#include <map>
#include <queue>
#include <set>
#include <utility>
#include <vector>
#define DEBUG_TYPE "hcp"
using namespace llvm;
namespace {
// Properties of a value that are tracked by the propagation.
// A property that is marked as present (i.e. bit is set) dentes that the
// value is known (proven) to have this property. Not all combinations
// of bits make sense, for example Zero and NonZero are mutually exclusive,
// but on the other hand, Zero implies Finite. In this case, whenever
// the Zero property is present, Finite should also be present.
class ConstantProperties {
public:
enum {
Unknown = 0x0000,
Zero = 0x0001,
NonZero = 0x0002,
Finite = 0x0004,
Infinity = 0x0008,
NaN = 0x0010,
SignedZero = 0x0020,
NumericProperties = (Zero|NonZero|Finite|Infinity|NaN|SignedZero),
PosOrZero = 0x0100,
NegOrZero = 0x0200,
SignProperties = (PosOrZero|NegOrZero),
Everything = (NumericProperties|SignProperties)
};
// For a given constant, deduce the set of trackable properties that this
// constant has.
static uint32_t deduce(const Constant *C);
};
// A representation of a register as it can appear in a MachineOperand,
// i.e. a pair register:subregister.
// FIXME: Use TargetInstrInfo::RegSubRegPair. Also duplicated in
// HexagonGenPredicate
struct RegisterSubReg {
Register Reg;
unsigned SubReg;
explicit RegisterSubReg(unsigned R, unsigned SR = 0) : Reg(R), SubReg(SR) {}
explicit RegisterSubReg(const MachineOperand &MO)
: Reg(MO.getReg()), SubReg(MO.getSubReg()) {}
void print(const TargetRegisterInfo *TRI = nullptr) const {
dbgs() << printReg(Reg, TRI, SubReg);
}
bool operator== (const RegisterSubReg &R) const {
return (Reg == R.Reg) && (SubReg == R.SubReg);
}
};
// Lattice cell, based on that was described in the W-Z paper on constant
// propagation.
// Latice cell will be allowed to hold multiple constant values. While
// multiple values would normally indicate "bottom", we can still derive
// some useful information from them. For example, comparison X > 0
// could be folded if all the values in the cell associated with X are
// positive.
class LatticeCell {
private:
enum { Normal, Top, Bottom };
static const unsigned MaxCellSize = 4;
unsigned Kind:2;
unsigned Size:3;
unsigned IsSpecial:1;
unsigned :0;
public:
union {
uint32_t Properties;
const Constant *Value;
const Constant *Values[MaxCellSize];
};
LatticeCell() : Kind(Top), Size(0), IsSpecial(false) {
for (const Constant *&Value : Values)
Value = nullptr;
}
bool meet(const LatticeCell &L);
bool add(const Constant *C);
bool add(uint32_t Property);
uint32_t properties() const;
unsigned size() const { return Size; }
LatticeCell(const LatticeCell &L) {
// This memcpy also copies Properties (when L.Size == 0).
uint32_t N =
L.IsSpecial ? sizeof L.Properties : L.Size * sizeof(const Constant *);
memcpy(Values, L.Values, N);
Kind = L.Kind;
Size = L.Size;
IsSpecial = L.IsSpecial;
}
LatticeCell &operator=(const LatticeCell &L) {
if (this != &L) {
// This memcpy also copies Properties (when L.Size == 0).
uint32_t N = L.IsSpecial ? sizeof L.Properties
: L.Size * sizeof(const Constant *);
memcpy(Values, L.Values, N);
Kind = L.Kind;
Size = L.Size;
IsSpecial = L.IsSpecial;
}
return *this;
}
bool isSingle() const { return size() == 1; }
bool isProperty() const { return IsSpecial; }
bool isTop() const { return Kind == Top; }
bool isBottom() const { return Kind == Bottom; }
bool setBottom() {
bool Changed = (Kind != Bottom);
Kind = Bottom;
Size = 0;
IsSpecial = false;
return Changed;
}
void print(raw_ostream &os) const;
private:
void setProperty() {
IsSpecial = true;
Size = 0;
Kind = Normal;
}
bool convertToProperty();
};
#ifndef NDEBUG
raw_ostream &operator<< (raw_ostream &os, const LatticeCell &L) {
L.print(os);
return os;
}
#endif
class MachineConstEvaluator;
class MachineConstPropagator {
public:
MachineConstPropagator(MachineConstEvaluator &E) : MCE(E) {
Bottom.setBottom();
}
// Mapping: vreg -> cell
// The keys are registers _without_ subregisters. This won't allow
// definitions in the form of "vreg:subreg = ...". Such definitions
// would be questionable from the point of view of SSA, since the "vreg"
// could not be initialized in its entirety (specifically, an instruction
// defining the "other part" of "vreg" would also count as a definition
// of "vreg", which would violate the SSA).
// If a value of a pair vreg:subreg needs to be obtained, the cell for
// "vreg" needs to be looked up, and then the value of subregister "subreg"
// needs to be evaluated.
class CellMap {
public:
CellMap() {
assert(Top.isTop());
Bottom.setBottom();
}
void clear() { Map.clear(); }
bool has(Register R) const {
// All non-virtual registers are considered "bottom".
if (!R.isVirtual())
return true;
MapType::const_iterator F = Map.find(R);
return F != Map.end();
}
const LatticeCell &get(Register R) const {
if (!R.isVirtual())
return Bottom;
MapType::const_iterator F = Map.find(R);
if (F != Map.end())
return F->second;
return Top;
}
// Invalidates any const references.
void update(Register R, const LatticeCell &L) { Map[R] = L; }
void print(raw_ostream &os, const TargetRegisterInfo &TRI) const;
private:
using MapType = std::map<Register, LatticeCell>;
MapType Map;
// To avoid creating "top" entries, return a const reference to
// this cell in "get". Also, have a "Bottom" cell to return from
// get when a value of a physical register is requested.
LatticeCell Top, Bottom;
public:
using const_iterator = MapType::const_iterator;
const_iterator begin() const { return Map.begin(); }
const_iterator end() const { return Map.end(); }
};
bool run(MachineFunction &MF);
private:
void visitPHI(const MachineInstr &PN);
void visitNonBranch(const MachineInstr &MI);
void visitBranchesFrom(const MachineInstr &BrI);
void visitUsesOf(unsigned R);
bool computeBlockSuccessors(const MachineBasicBlock *MB,
SetVector<const MachineBasicBlock*> &Targets);
void removeCFGEdge(MachineBasicBlock *From, MachineBasicBlock *To);
void propagate(MachineFunction &MF);
bool rewrite(MachineFunction &MF);
MachineRegisterInfo *MRI = nullptr;
MachineConstEvaluator &MCE;
using CFGEdge = std::pair<unsigned, unsigned>;
using SetOfCFGEdge = std::set<CFGEdge>;
using SetOfInstr = std::set<const MachineInstr *>;
using QueueOfCFGEdge = std::queue<CFGEdge>;
LatticeCell Bottom;
CellMap Cells;
SetOfCFGEdge EdgeExec;
SetOfInstr InstrExec;
QueueOfCFGEdge FlowQ;
};
// The "evaluator/rewriter" of machine instructions. This is an abstract
// base class that provides the interface that the propagator will use,
// as well as some helper functions that are target-independent.
class MachineConstEvaluator {
public:
MachineConstEvaluator(MachineFunction &Fn)
: TRI(*Fn.getSubtarget().getRegisterInfo()),
MF(Fn), CX(Fn.getFunction().getContext()) {}
virtual ~MachineConstEvaluator() = default;
// The required interface:
// - A set of three "evaluate" functions. Each returns "true" if the
// computation succeeded, "false" otherwise.
// (1) Given an instruction MI, and the map with input values "Inputs",
// compute the set of output values "Outputs". An example of when
// the computation can "fail" is if MI is not an instruction that
// is recognized by the evaluator.
// (2) Given a register R (as reg:subreg), compute the cell that
// corresponds to the "subreg" part of the given register.
// (3) Given a branch instruction BrI, compute the set of target blocks.
// If the branch can fall-through, add null (0) to the list of
// possible targets.
// - A function "rewrite", that given the cell map after propagation,
// could rewrite instruction MI in a more beneficial form. Return
// "true" if a change has been made, "false" otherwise.
using CellMap = MachineConstPropagator::CellMap;
virtual bool evaluate(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs) = 0;
virtual bool evaluate(const RegisterSubReg &R, const LatticeCell &SrcC,
LatticeCell &Result) = 0;
virtual bool evaluate(const MachineInstr &BrI, const CellMap &Inputs,
SetVector<const MachineBasicBlock*> &Targets,
bool &CanFallThru) = 0;
virtual bool rewrite(MachineInstr &MI, const CellMap &Inputs) = 0;
const TargetRegisterInfo &TRI;
protected:
MachineFunction &MF;
LLVMContext &CX;
struct Comparison {
enum {
Unk = 0x00,
EQ = 0x01,
NE = 0x02,
L = 0x04, // Less-than property.
G = 0x08, // Greater-than property.
U = 0x40, // Unsigned property.
LTs = L,
LEs = L | EQ,
GTs = G,
GEs = G | EQ,
LTu = L | U,
LEu = L | EQ | U,
GTu = G | U,
GEu = G | EQ | U
};
static uint32_t negate(uint32_t Cmp) {
if (Cmp == EQ)
return NE;
if (Cmp == NE)
return EQ;
assert((Cmp & (L|G)) != (L|G));
return Cmp ^ (L|G);
}
};
// Helper functions.
bool getCell(const RegisterSubReg &R, const CellMap &Inputs, LatticeCell &RC);
bool constToInt(const Constant *C, APInt &Val) const;
bool constToFloat(const Constant *C, APFloat &Val) const;
const ConstantInt *intToConst(const APInt &Val) const;
// Compares.
bool evaluateCMPrr(uint32_t Cmp, const RegisterSubReg &R1, const RegisterSubReg &R2,
const CellMap &Inputs, bool &Result);
bool evaluateCMPri(uint32_t Cmp, const RegisterSubReg &R1, const APInt &A2,
const CellMap &Inputs, bool &Result);
bool evaluateCMPrp(uint32_t Cmp, const RegisterSubReg &R1, uint64_t Props2,
const CellMap &Inputs, bool &Result);
bool evaluateCMPii(uint32_t Cmp, const APInt &A1, const APInt &A2,
bool &Result);
bool evaluateCMPpi(uint32_t Cmp, uint32_t Props, const APInt &A2,
bool &Result);
bool evaluateCMPpp(uint32_t Cmp, uint32_t Props1, uint32_t Props2,
bool &Result);
bool evaluateCOPY(const RegisterSubReg &R1, const CellMap &Inputs,
LatticeCell &Result);
// Logical operations.
bool evaluateANDrr(const RegisterSubReg &R1, const RegisterSubReg &R2,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateANDri(const RegisterSubReg &R1, const APInt &A2,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateANDii(const APInt &A1, const APInt &A2, APInt &Result);
bool evaluateORrr(const RegisterSubReg &R1, const RegisterSubReg &R2,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateORri(const RegisterSubReg &R1, const APInt &A2,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateORii(const APInt &A1, const APInt &A2, APInt &Result);
bool evaluateXORrr(const RegisterSubReg &R1, const RegisterSubReg &R2,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateXORri(const RegisterSubReg &R1, const APInt &A2,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateXORii(const APInt &A1, const APInt &A2, APInt &Result);
// Extensions.
bool evaluateZEXTr(const RegisterSubReg &R1, unsigned Width, unsigned Bits,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateZEXTi(const APInt &A1, unsigned Width, unsigned Bits,
APInt &Result);
bool evaluateSEXTr(const RegisterSubReg &R1, unsigned Width, unsigned Bits,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateSEXTi(const APInt &A1, unsigned Width, unsigned Bits,
APInt &Result);
// Leading/trailing bits.
bool evaluateCLBr(const RegisterSubReg &R1, bool Zeros, bool Ones,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateCLBi(const APInt &A1, bool Zeros, bool Ones, APInt &Result);
bool evaluateCTBr(const RegisterSubReg &R1, bool Zeros, bool Ones,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateCTBi(const APInt &A1, bool Zeros, bool Ones, APInt &Result);
// Bitfield extract.
bool evaluateEXTRACTr(const RegisterSubReg &R1, unsigned Width, unsigned Bits,
unsigned Offset, bool Signed, const CellMap &Inputs,
LatticeCell &Result);
bool evaluateEXTRACTi(const APInt &A1, unsigned Bits, unsigned Offset,
bool Signed, APInt &Result);
// Vector operations.
bool evaluateSplatr(const RegisterSubReg &R1, unsigned Bits, unsigned Count,
const CellMap &Inputs, LatticeCell &Result);
bool evaluateSplati(const APInt &A1, unsigned Bits, unsigned Count,
APInt &Result);
};
} // end anonymous namespace
uint32_t ConstantProperties::deduce(const Constant *C) {
if (isa<ConstantInt>(C)) {
const ConstantInt *CI = cast<ConstantInt>(C);
if (CI->isZero())
return Zero | PosOrZero | NegOrZero | Finite;
uint32_t Props = (NonZero | Finite);
if (CI->isNegative())
return Props | NegOrZero;
return Props | PosOrZero;
}
if (isa<ConstantFP>(C)) {
const ConstantFP *CF = cast<ConstantFP>(C);
uint32_t Props = CF->isNegative() ? (NegOrZero|NonZero)
: PosOrZero;
if (CF->isZero())
return (Props & ~NumericProperties) | (Zero|Finite);
Props = (Props & ~NumericProperties) | NonZero;
if (CF->isNaN())
return (Props & ~NumericProperties) | NaN;
const APFloat &Val = CF->getValueAPF();
if (Val.isInfinity())
return (Props & ~NumericProperties) | Infinity;
Props |= Finite;
return Props;
}
return Unknown;
}
// Convert a cell from a set of specific values to a cell that tracks
// properties.
bool LatticeCell::convertToProperty() {
if (isProperty())
return false;
// Corner case: converting a fresh (top) cell to "special".
// This can happen, when adding a property to a top cell.
uint32_t Everything = ConstantProperties::Everything;
uint32_t Ps = !isTop() ? properties()
: Everything;
if (Ps != ConstantProperties::Unknown) {
Properties = Ps;
setProperty();
} else {
setBottom();
}
return true;
}
#ifndef NDEBUG
void LatticeCell::print(raw_ostream &os) const {
if (isProperty()) {
os << "{ ";
uint32_t Ps = properties();
if (Ps & ConstantProperties::Zero)
os << "zero ";
if (Ps & ConstantProperties::NonZero)
os << "nonzero ";
if (Ps & ConstantProperties::Finite)
os << "finite ";
if (Ps & ConstantProperties::Infinity)
os << "infinity ";
if (Ps & ConstantProperties::NaN)
os << "nan ";
if (Ps & ConstantProperties::PosOrZero)
os << "poz ";
if (Ps & ConstantProperties::NegOrZero)
os << "nez ";
os << '}';
return;
}
os << "{ ";
if (isBottom()) {
os << "bottom";
} else if (isTop()) {
os << "top";
} else {
for (unsigned i = 0; i < size(); ++i) {
const Constant *C = Values[i];
if (i != 0)
os << ", ";
C->print(os);
}
}
os << " }";
}
#endif
// "Meet" operation on two cells. This is the key of the propagation
// algorithm.
bool LatticeCell::meet(const LatticeCell &L) {
bool Changed = false;
if (L.isBottom())
Changed = setBottom();
if (isBottom() || L.isTop())
return Changed;
if (isTop()) {
*this = L;
// L can be neither Top nor Bottom, so *this must have changed.
return true;
}
// Top/bottom cases covered. Need to integrate L's set into ours.
if (L.isProperty())
return add(L.properties());
for (unsigned i = 0; i < L.size(); ++i) {
const Constant *LC = L.Values[i];
Changed |= add(LC);
}
return Changed;
}
// Add a new constant to the cell. This is actually where the cell update
// happens. If a cell has room for more constants, the new constant is added.
// Otherwise, the cell is converted to a "property" cell (i.e. a cell that
// will track properties of the associated values, and not the values
// themselves. Care is taken to handle special cases, like "bottom", etc.
bool LatticeCell::add(const Constant *LC) {
assert(LC);
if (isBottom())
return false;
if (!isProperty()) {
// Cell is not special. Try to add the constant here first,
// if there is room.
unsigned Index = 0;
while (Index < Size) {
const Constant *C = Values[Index];
// If the constant is already here, no change is needed.
if (C == LC)
return false;
Index++;
}
if (Index < MaxCellSize) {
Values[Index] = LC;
Kind = Normal;
Size++;
return true;
}
}
bool Changed = false;
// This cell is special, or is not special, but is full. After this
// it will be special.
Changed = convertToProperty();
uint32_t Ps = properties();
uint32_t NewPs = Ps & ConstantProperties::deduce(LC);
if (NewPs == ConstantProperties::Unknown) {
setBottom();
return true;
}
if (Ps != NewPs) {
Properties = NewPs;
Changed = true;
}
return Changed;
}
// Add a property to the cell. This will force the cell to become a property-
// tracking cell.
bool LatticeCell::add(uint32_t Property) {
bool Changed = convertToProperty();
uint32_t Ps = properties();
if (Ps == (Ps & Property))
return Changed;
Properties = Property & Ps;
return true;
}
// Return the properties of the values in the cell. This is valid for any
// cell, and does not alter the cell itself.
uint32_t LatticeCell::properties() const {
if (isProperty())
return Properties;
assert(!isTop() && "Should not call this for a top cell");
if (isBottom())
return ConstantProperties::Unknown;
assert(size() > 0 && "Empty cell");
uint32_t Ps = ConstantProperties::deduce(Values[0]);
for (unsigned i = 1; i < size(); ++i) {
if (Ps == ConstantProperties::Unknown)
break;
Ps &= ConstantProperties::deduce(Values[i]);
}
return Ps;
}
#ifndef NDEBUG
void MachineConstPropagator::CellMap::print(raw_ostream &os,
const TargetRegisterInfo &TRI) const {
for (auto &I : Map)
dbgs() << " " << printReg(I.first, &TRI) << " -> " << I.second << '\n';
}
#endif
void MachineConstPropagator::visitPHI(const MachineInstr &PN) {
const MachineBasicBlock *MB = PN.getParent();
unsigned MBN = MB->getNumber();
LLVM_DEBUG(dbgs() << "Visiting FI(" << printMBBReference(*MB) << "): " << PN);
const MachineOperand &MD = PN.getOperand(0);
RegisterSubReg DefR(MD);
assert(DefR.Reg.isVirtual());
bool Changed = false;
// If the def has a sub-register, set the corresponding cell to "bottom".
if (DefR.SubReg) {
Bottomize:
const LatticeCell &T = Cells.get(DefR.Reg);
Changed = !T.isBottom();
Cells.update(DefR.Reg, Bottom);
if (Changed)
visitUsesOf(DefR.Reg);
return;
}
LatticeCell DefC = Cells.get(DefR.Reg);
for (unsigned i = 1, n = PN.getNumOperands(); i < n; i += 2) {
const MachineBasicBlock *PB = PN.getOperand(i+1).getMBB();
unsigned PBN = PB->getNumber();
if (!EdgeExec.count(CFGEdge(PBN, MBN))) {
LLVM_DEBUG(dbgs() << " edge " << printMBBReference(*PB) << "->"
<< printMBBReference(*MB) << " not executable\n");
continue;
}
const MachineOperand &SO = PN.getOperand(i);
RegisterSubReg UseR(SO);
// If the input is not a virtual register, we don't really know what
// value it holds.
if (!UseR.Reg.isVirtual())
goto Bottomize;
// If there is no cell for an input register, it means top.
if (!Cells.has(UseR.Reg))
continue;
LatticeCell SrcC;
bool Eval = MCE.evaluate(UseR, Cells.get(UseR.Reg), SrcC);
LLVM_DEBUG(dbgs() << " edge from " << printMBBReference(*PB) << ": "
<< printReg(UseR.Reg, &MCE.TRI, UseR.SubReg) << SrcC
<< '\n');
Changed |= Eval ? DefC.meet(SrcC)
: DefC.setBottom();
Cells.update(DefR.Reg, DefC);
if (DefC.isBottom())
break;
}
if (Changed)
visitUsesOf(DefR.Reg);
}
void MachineConstPropagator::visitNonBranch(const MachineInstr &MI) {
LLVM_DEBUG(dbgs() << "Visiting MI(" << printMBBReference(*MI.getParent())
<< "): " << MI);
CellMap Outputs;
bool Eval = MCE.evaluate(MI, Cells, Outputs);
LLVM_DEBUG({
if (Eval) {
dbgs() << " outputs:";
for (auto &I : Outputs)
dbgs() << ' ' << I.second;
dbgs() << '\n';
}
});
// Update outputs. If the value was not computed, set all the
// def cells to bottom.
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || !MO.isDef())
continue;
RegisterSubReg DefR(MO);
// Only track virtual registers.
if (!DefR.Reg.isVirtual())
continue;
bool Changed = false;
// If the evaluation failed, set cells for all output registers to bottom.
if (!Eval) {
const LatticeCell &T = Cells.get(DefR.Reg);
Changed = !T.isBottom();
Cells.update(DefR.Reg, Bottom);
} else {
// Find the corresponding cell in the computed outputs.
// If it's not there, go on to the next def.
if (!Outputs.has(DefR.Reg))
continue;
LatticeCell RC = Cells.get(DefR.Reg);
Changed = RC.meet(Outputs.get(DefR.Reg));
Cells.update(DefR.Reg, RC);
}
if (Changed)
visitUsesOf(DefR.Reg);
}
}
// Starting at a given branch, visit remaining branches in the block.
// Traverse over the subsequent branches for as long as the preceding one
// can fall through. Add all the possible targets to the flow work queue,
// including the potential fall-through to the layout-successor block.
void MachineConstPropagator::visitBranchesFrom(const MachineInstr &BrI) {
const MachineBasicBlock &B = *BrI.getParent();
unsigned MBN = B.getNumber();
MachineBasicBlock::const_iterator It = BrI.getIterator();
MachineBasicBlock::const_iterator End = B.end();
SetVector<const MachineBasicBlock*> Targets;
bool EvalOk = true, FallsThru = true;
while (It != End) {
const MachineInstr &MI = *It;
InstrExec.insert(&MI);
LLVM_DEBUG(dbgs() << "Visiting " << (EvalOk ? "BR" : "br") << "("
<< printMBBReference(B) << "): " << MI);
// Do not evaluate subsequent branches if the evaluation of any of the
// previous branches failed. Keep iterating over the branches only
// to mark them as executable.
EvalOk = EvalOk && MCE.evaluate(MI, Cells, Targets, FallsThru);
if (!EvalOk)
FallsThru = true;
if (!FallsThru)
break;
++It;
}
if (B.mayHaveInlineAsmBr())
EvalOk = false;
if (EvalOk) {
// Need to add all CFG successors that lead to EH landing pads.
// There won't be explicit branches to these blocks, but they must
// be processed.
for (const MachineBasicBlock *SB : B.successors()) {
if (SB->isEHPad())
Targets.insert(SB);
}
if (FallsThru) {
const MachineFunction &MF = *B.getParent();
MachineFunction::const_iterator BI = B.getIterator();
MachineFunction::const_iterator Next = std::next(BI);
if (Next != MF.end())
Targets.insert(&*Next);
}
} else {
// If the evaluation of the branches failed, make "Targets" to be the
// set of all successors of the block from the CFG.
// If the evaluation succeeded for all visited branches, then if the
// last one set "FallsThru", then add an edge to the layout successor
// to the targets.
Targets.clear();
LLVM_DEBUG(dbgs() << " failed to evaluate a branch...adding all CFG "
"successors\n");
for (const MachineBasicBlock *SB : B.successors())
Targets.insert(SB);
}
for (const MachineBasicBlock *TB : Targets) {
unsigned TBN = TB->getNumber();
LLVM_DEBUG(dbgs() << " pushing edge " << printMBBReference(B) << " -> "
<< printMBBReference(*TB) << "\n");
FlowQ.push(CFGEdge(MBN, TBN));
}
}
void MachineConstPropagator::visitUsesOf(unsigned Reg) {
LLVM_DEBUG(dbgs() << "Visiting uses of " << printReg(Reg, &MCE.TRI)
<< Cells.get(Reg) << '\n');
for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) {
// Do not process non-executable instructions. They can become exceutable
// later (via a flow-edge in the work queue). In such case, the instruc-
// tion will be visited at that time.
if (!InstrExec.count(&MI))
continue;
if (MI.isPHI())
visitPHI(MI);
else if (!MI.isBranch())
visitNonBranch(MI);
else
visitBranchesFrom(MI);
}
}
bool MachineConstPropagator::computeBlockSuccessors(const MachineBasicBlock *MB,
SetVector<const MachineBasicBlock*> &Targets) {
Targets.clear();
MachineBasicBlock::const_iterator FirstBr = MB->end();
for (const MachineInstr &MI : *MB) {
if (MI.getOpcode() == TargetOpcode::INLINEASM_BR)
return false;
if (MI.isDebugInstr())
continue;
if (MI.isBranch()) {
FirstBr = MI.getIterator();
break;
}
}
MachineBasicBlock::const_iterator End = MB->end();
bool DoNext = true;
for (MachineBasicBlock::const_iterator I = FirstBr; I != End; ++I) {
const MachineInstr &MI = *I;
// Can there be debug instructions between branches?
if (MI.isDebugInstr())
continue;
if (!InstrExec.count(&MI))
continue;
bool Eval = MCE.evaluate(MI, Cells, Targets, DoNext);
if (!Eval)
return false;
if (!DoNext)
break;
}
// If the last branch could fall-through, add block's layout successor.
if (DoNext) {
MachineFunction::const_iterator BI = MB->getIterator();
MachineFunction::const_iterator NextI = std::next(BI);
if (NextI != MB->getParent()->end())
Targets.insert(&*NextI);
}
// Add all the EH landing pads.
for (const MachineBasicBlock *SB : MB->successors())
if (SB->isEHPad())
Targets.insert(SB);
return true;
}
void MachineConstPropagator::removeCFGEdge(MachineBasicBlock *From,
MachineBasicBlock *To) {
// First, remove the CFG successor/predecessor information.
From->removeSuccessor(To);
// Remove all corresponding PHI operands in the To block.
for (MachineInstr &PN : To->phis()) {
// reg0 = PHI reg1, bb2, reg3, bb4, ...
int N = PN.getNumOperands() - 2;
while (N > 0) {
if (PN.getOperand(N + 1).getMBB() == From) {
PN.removeOperand(N + 1);
PN.removeOperand(N);
}
N -= 2;
}
}
}
void MachineConstPropagator::propagate(MachineFunction &MF) {
MachineBasicBlock *Entry = GraphTraits<MachineFunction*>::getEntryNode(&MF);
unsigned EntryNum = Entry->getNumber();
// Start with a fake edge, just to process the entry node.
FlowQ.push(CFGEdge(EntryNum, EntryNum));
while (!FlowQ.empty()) {
CFGEdge Edge = FlowQ.front();
FlowQ.pop();
LLVM_DEBUG(
dbgs() << "Picked edge "
<< printMBBReference(*MF.getBlockNumbered(Edge.first)) << "->"
<< printMBBReference(*MF.getBlockNumbered(Edge.second)) << '\n');
if (Edge.first != EntryNum)
if (EdgeExec.count(Edge))
continue;
EdgeExec.insert(Edge);
MachineBasicBlock *SB = MF.getBlockNumbered(Edge.second);
// Process the block in three stages:
// - visit all PHI nodes,
// - visit all non-branch instructions,
// - visit block branches.
MachineBasicBlock::const_iterator It = SB->begin(), End = SB->end();
// Visit PHI nodes in the successor block.
while (It != End && It->isPHI()) {
InstrExec.insert(&*It);
visitPHI(*It);
++It;
}
// If the successor block just became executable, visit all instructions.
// To see if this is the first time we're visiting it, check the first
// non-debug instruction to see if it is executable.
while (It != End && It->isDebugInstr())
++It;
assert(It == End || !It->isPHI());
// If this block has been visited, go on to the next one.
if (It != End && InstrExec.count(&*It))
continue;
// For now, scan all non-branch instructions. Branches require different
// processing.
while (It != End && !It->isBranch()) {
if (!It->isDebugInstr()) {
InstrExec.insert(&*It);
visitNonBranch(*It);
}
++It;
}
// Time to process the end of the block. This is different from
// processing regular (non-branch) instructions, because there can
// be multiple branches in a block, and they can cause the block to
// terminate early.
if (It != End) {
visitBranchesFrom(*It);
} else {
// If the block didn't have a branch, add all successor edges to the
// work queue. (There should really be only one successor in such case.)
unsigned SBN = SB->getNumber();
for (const MachineBasicBlock *SSB : SB->successors())
FlowQ.push(CFGEdge(SBN, SSB->getNumber()));
}
} // while (FlowQ)
LLVM_DEBUG({
dbgs() << "Cells after propagation:\n";
Cells.print(dbgs(), MCE.TRI);
dbgs() << "Dead CFG edges:\n";
for (const MachineBasicBlock &B : MF) {
unsigned BN = B.getNumber();
for (const MachineBasicBlock *SB : B.successors()) {
unsigned SN = SB->getNumber();
if (!EdgeExec.count(CFGEdge(BN, SN)))
dbgs() << " " << printMBBReference(B) << " -> "
<< printMBBReference(*SB) << '\n';
}
}
});
}
bool MachineConstPropagator::rewrite(MachineFunction &MF) {
bool Changed = false;
// Rewrite all instructions based on the collected cell information.
//
// Traverse the instructions in a post-order, so that rewriting an
// instruction can make changes "downstream" in terms of control-flow
// without affecting the rewriting process. (We should not change
// instructions that have not yet been visited by the rewriter.)
// The reason for this is that the rewriter can introduce new vregs,
// and replace uses of old vregs (which had corresponding cells
// computed during propagation) with these new vregs (which at this
// point would not have any cells, and would appear to be "top").
// If an attempt was made to evaluate an instruction with a fresh
// "top" vreg, it would cause an error (abend) in the evaluator.
// Collect the post-order-traversal block ordering. The subsequent
// traversal/rewrite will update block successors, so it's safer
// if the visiting order it computed ahead of time.
std::vector<MachineBasicBlock*> POT;
for (MachineBasicBlock *B : post_order(&MF))
if (!B->empty())
POT.push_back(B);
for (MachineBasicBlock *B : POT) {
// Walk the block backwards (which usually begin with the branches).
// If any branch is rewritten, we may need to update the successor
// information for this block. Unless the block's successors can be
// precisely determined (which may not be the case for indirect
// branches), we cannot modify any branch.
// Compute the successor information.
SetVector<const MachineBasicBlock*> Targets;
bool HaveTargets = computeBlockSuccessors(B, Targets);
// Rewrite the executable instructions. Skip branches if we don't
// have block successor information.
for (MachineInstr &MI : llvm::reverse(*B)) {
if (InstrExec.count(&MI)) {
if (MI.isBranch() && !HaveTargets)
continue;
Changed |= MCE.rewrite(MI, Cells);
}
}
// The rewriting could rewrite PHI nodes to non-PHI nodes, causing
// regular instructions to appear in between PHI nodes. Bring all
// the PHI nodes to the beginning of the block.
for (auto I = B->begin(), E = B->end(); I != E; ++I) {
if (I->isPHI())
continue;
// I is not PHI. Find the next PHI node P.
auto P = I;
while (++P != E)
if (P->isPHI())
break;
// Not found.
if (P == E)
break;
// Splice P right before I.
B->splice(I, B, P);
// Reset I to point at the just spliced PHI node.
--I;
}
// Update the block successor information: remove unnecessary successors.
if (HaveTargets) {
SmallVector<MachineBasicBlock*,2> ToRemove;
for (MachineBasicBlock *SB : B->successors()) {
if (!Targets.count(SB))
ToRemove.push_back(const_cast<MachineBasicBlock*>(SB));
Targets.remove(SB);
}
for (MachineBasicBlock *MBB : ToRemove)
removeCFGEdge(B, MBB);
// If there are any blocks left in the computed targets, it means that
// we think that the block could go somewhere, but the CFG does not.
// This could legitimately happen in blocks that have non-returning
// calls---we would think that the execution can continue, but the
// CFG will not have a successor edge.
}
}
// Need to do some final post-processing.
// If a branch was not executable, it will not get rewritten, but should
// be removed (or replaced with something equivalent to a A2_nop). We can't
// erase instructions during rewriting, so this needs to be delayed until
// now.
for (MachineBasicBlock &B : MF) {
for (MachineInstr &MI : llvm::make_early_inc_range(B))
if (MI.isBranch() && !InstrExec.count(&MI))
B.erase(&MI);
}
return Changed;
}
// This is the constant propagation algorithm as described by Wegman-Zadeck.
// Most of the terminology comes from there.
bool MachineConstPropagator::run(MachineFunction &MF) {
LLVM_DEBUG(MF.print(dbgs() << "Starting MachineConstPropagator\n", nullptr));
MRI = &MF.getRegInfo();
Cells.clear();
EdgeExec.clear();
InstrExec.clear();
assert(FlowQ.empty());
propagate(MF);
bool Changed = rewrite(MF);
LLVM_DEBUG({
dbgs() << "End of MachineConstPropagator (Changed=" << Changed << ")\n";
if (Changed)
MF.print(dbgs(), nullptr);
});
return Changed;
}
// --------------------------------------------------------------------
// Machine const evaluator.
bool MachineConstEvaluator::getCell(const RegisterSubReg &R, const CellMap &Inputs,
LatticeCell &RC) {
if (!R.Reg.isVirtual())
return false;
const LatticeCell &L = Inputs.get(R.Reg);
if (!R.SubReg) {
RC = L;
return !RC.isBottom();
}
bool Eval = evaluate(R, L, RC);
return Eval && !RC.isBottom();
}
bool MachineConstEvaluator::constToInt(const Constant *C,
APInt &Val) const {
const ConstantInt *CI = dyn_cast<ConstantInt>(C);
if (!CI)
return false;
Val = CI->getValue();
return true;
}
const ConstantInt *MachineConstEvaluator::intToConst(const APInt &Val) const {
return ConstantInt::get(CX, Val);
}
bool MachineConstEvaluator::evaluateCMPrr(uint32_t Cmp, const RegisterSubReg &R1,
const RegisterSubReg &R2, const CellMap &Inputs, bool &Result) {
assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg));
LatticeCell LS1, LS2;
if (!getCell(R1, Inputs, LS1) || !getCell(R2, Inputs, LS2))
return false;
bool IsProp1 = LS1.isProperty();
bool IsProp2 = LS2.isProperty();
if (IsProp1) {
uint32_t Prop1 = LS1.properties();
if (IsProp2)
return evaluateCMPpp(Cmp, Prop1, LS2.properties(), Result);
uint32_t NegCmp = Comparison::negate(Cmp);
return evaluateCMPrp(NegCmp, R2, Prop1, Inputs, Result);
}
if (IsProp2) {
uint32_t Prop2 = LS2.properties();
return evaluateCMPrp(Cmp, R1, Prop2, Inputs, Result);
}
APInt A;
bool IsTrue = true, IsFalse = true;
for (unsigned i = 0; i < LS2.size(); ++i) {
bool Res;
bool Computed = constToInt(LS2.Values[i], A) &&
evaluateCMPri(Cmp, R1, A, Inputs, Res);
if (!Computed)
return false;
IsTrue &= Res;
IsFalse &= !Res;
}
assert(!IsTrue || !IsFalse);
// The actual logical value of the comparison is same as IsTrue.
Result = IsTrue;
// Return true if the result was proven to be true or proven to be false.
return IsTrue || IsFalse;
}
bool MachineConstEvaluator::evaluateCMPri(uint32_t Cmp, const RegisterSubReg &R1,
const APInt &A2, const CellMap &Inputs, bool &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS;
if (!getCell(R1, Inputs, LS))
return false;
if (LS.isProperty())
return evaluateCMPpi(Cmp, LS.properties(), A2, Result);
APInt A;
bool IsTrue = true, IsFalse = true;
for (unsigned i = 0; i < LS.size(); ++i) {
bool Res;
bool Computed = constToInt(LS.Values[i], A) &&
evaluateCMPii(Cmp, A, A2, Res);
if (!Computed)
return false;
IsTrue &= Res;
IsFalse &= !Res;
}
assert(!IsTrue || !IsFalse);
// The actual logical value of the comparison is same as IsTrue.
Result = IsTrue;
// Return true if the result was proven to be true or proven to be false.
return IsTrue || IsFalse;
}
bool MachineConstEvaluator::evaluateCMPrp(uint32_t Cmp, const RegisterSubReg &R1,
uint64_t Props2, const CellMap &Inputs, bool &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS;
if (!getCell(R1, Inputs, LS))
return false;
if (LS.isProperty())
return evaluateCMPpp(Cmp, LS.properties(), Props2, Result);
APInt A;
uint32_t NegCmp = Comparison::negate(Cmp);
bool IsTrue = true, IsFalse = true;
for (unsigned i = 0; i < LS.size(); ++i) {
bool Res;
bool Computed = constToInt(LS.Values[i], A) &&
evaluateCMPpi(NegCmp, Props2, A, Res);
if (!Computed)
return false;
IsTrue &= Res;
IsFalse &= !Res;
}
assert(!IsTrue || !IsFalse);
Result = IsTrue;
return IsTrue || IsFalse;
}
bool MachineConstEvaluator::evaluateCMPii(uint32_t Cmp, const APInt &A1,
const APInt &A2, bool &Result) {
// NE is a special kind of comparison (not composed of smaller properties).
if (Cmp == Comparison::NE) {
Result = !APInt::isSameValue(A1, A2);
return true;
}
if (Cmp == Comparison::EQ) {
Result = APInt::isSameValue(A1, A2);
return true;
}
if (Cmp & Comparison::EQ) {
if (APInt::isSameValue(A1, A2))
return (Result = true);
}
assert((Cmp & (Comparison::L | Comparison::G)) && "Malformed comparison");
Result = false;
unsigned W1 = A1.getBitWidth();
unsigned W2 = A2.getBitWidth();
unsigned MaxW = (W1 >= W2) ? W1 : W2;
if (Cmp & Comparison::U) {
APInt Zx1 = A1.zext(MaxW);
APInt Zx2 = A2.zext(MaxW);
if (Cmp & Comparison::L)
Result = Zx1.ult(Zx2);
else if (Cmp & Comparison::G)
Result = Zx2.ult(Zx1);
return true;
}
// Signed comparison.
APInt Sx1 = A1.sext(MaxW);
APInt Sx2 = A2.sext(MaxW);
if (Cmp & Comparison::L)
Result = Sx1.slt(Sx2);
else if (Cmp & Comparison::G)
Result = Sx2.slt(Sx1);
return true;
}
bool MachineConstEvaluator::evaluateCMPpi(uint32_t Cmp, uint32_t Props,
const APInt &A2, bool &Result) {
if (Props == ConstantProperties::Unknown)
return false;
// Should never see NaN here, but check for it for completeness.
if (Props & ConstantProperties::NaN)
return false;
// Infinity could theoretically be compared to a number, but the
// presence of infinity here would be very suspicious. If we don't
// know for sure that the number is finite, bail out.
if (!(Props & ConstantProperties::Finite))
return false;
// Let X be a number that has properties Props.
if (Cmp & Comparison::U) {
// In case of unsigned comparisons, we can only compare against 0.
if (A2 == 0) {
// Any x!=0 will be considered >0 in an unsigned comparison.
if (Props & ConstantProperties::Zero)
Result = (Cmp & Comparison::EQ);
else if (Props & ConstantProperties::NonZero)
Result = (Cmp & Comparison::G) || (Cmp == Comparison::NE);
else
return false;
return true;
}
// A2 is not zero. The only handled case is if X = 0.
if (Props & ConstantProperties::Zero) {
Result = (Cmp & Comparison::L) || (Cmp == Comparison::NE);
return true;
}
return false;
}
// Signed comparisons are different.
if (Props & ConstantProperties::Zero) {
if (A2 == 0)
Result = (Cmp & Comparison::EQ);
else
Result = (Cmp == Comparison::NE) ||
((Cmp & Comparison::L) && !A2.isNegative()) ||
((Cmp & Comparison::G) && A2.isNegative());
return true;
}
if (Props & ConstantProperties::PosOrZero) {
// X >= 0 and !(A2 < 0) => cannot compare
if (!A2.isNegative())
return false;
// X >= 0 and A2 < 0
Result = (Cmp & Comparison::G) || (Cmp == Comparison::NE);
return true;
}
if (Props & ConstantProperties::NegOrZero) {
// X <= 0 and Src1 < 0 => cannot compare
if (A2 == 0 || A2.isNegative())
return false;
// X <= 0 and A2 > 0
Result = (Cmp & Comparison::L) || (Cmp == Comparison::NE);
return true;
}
return false;
}
bool MachineConstEvaluator::evaluateCMPpp(uint32_t Cmp, uint32_t Props1,
uint32_t Props2, bool &Result) {
using P = ConstantProperties;
if ((Props1 & P::NaN) && (Props2 & P::NaN))
return false;
if (!(Props1 & P::Finite) || !(Props2 & P::Finite))
return false;
bool Zero1 = (Props1 & P::Zero), Zero2 = (Props2 & P::Zero);
bool NonZero1 = (Props1 & P::NonZero), NonZero2 = (Props2 & P::NonZero);
if (Zero1 && Zero2) {
Result = (Cmp & Comparison::EQ);
return true;
}
if (Cmp == Comparison::NE) {
if ((Zero1 && NonZero2) || (NonZero1 && Zero2))
return (Result = true);
return false;
}
if (Cmp & Comparison::U) {
// In unsigned comparisons, we can only compare against a known zero,
// or a known non-zero.
if (Zero1 && NonZero2) {
Result = (Cmp & Comparison::L);
return true;
}
if (NonZero1 && Zero2) {
Result = (Cmp & Comparison::G);
return true;
}
return false;
}
// Signed comparison. The comparison is not NE.
bool Poz1 = (Props1 & P::PosOrZero), Poz2 = (Props2 & P::PosOrZero);
bool Nez1 = (Props1 & P::NegOrZero), Nez2 = (Props2 & P::NegOrZero);
if (Nez1 && Poz2) {
if (NonZero1 || NonZero2) {
Result = (Cmp & Comparison::L);
return true;
}
// Either (or both) could be zero. Can only say that X <= Y.
if ((Cmp & Comparison::EQ) && (Cmp & Comparison::L))
return (Result = true);
}
if (Poz1 && Nez2) {
if (NonZero1 || NonZero2) {
Result = (Cmp & Comparison::G);
return true;
}
// Either (or both) could be zero. Can only say that X >= Y.
if ((Cmp & Comparison::EQ) && (Cmp & Comparison::G))
return (Result = true);
}
return false;
}
bool MachineConstEvaluator::evaluateCOPY(const RegisterSubReg &R1,
const CellMap &Inputs, LatticeCell &Result) {
return getCell(R1, Inputs, Result);
}
bool MachineConstEvaluator::evaluateANDrr(const RegisterSubReg &R1,
const RegisterSubReg &R2, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg));
const LatticeCell &L1 = Inputs.get(R2.Reg);
const LatticeCell &L2 = Inputs.get(R2.Reg);
// If both sources are bottom, exit. Otherwise try to evaluate ANDri
// with the non-bottom argument passed as the immediate. This is to
// catch cases of ANDing with 0.
if (L2.isBottom()) {
if (L1.isBottom())
return false;
return evaluateANDrr(R2, R1, Inputs, Result);
}
LatticeCell LS2;
if (!evaluate(R2, L2, LS2))
return false;
if (LS2.isBottom() || LS2.isProperty())
return false;
APInt A;
for (unsigned i = 0; i < LS2.size(); ++i) {
LatticeCell RC;
bool Eval = constToInt(LS2.Values[i], A) &&
evaluateANDri(R1, A, Inputs, RC);
if (!Eval)
return false;
Result.meet(RC);
}
return !Result.isBottom();
}
bool MachineConstEvaluator::evaluateANDri(const RegisterSubReg &R1,
const APInt &A2, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
if (A2 == -1)
return getCell(R1, Inputs, Result);
if (A2 == 0) {
LatticeCell RC;
RC.add(intToConst(A2));
// Overwrite Result.
Result = RC;
return true;
}
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isBottom() || LS1.isProperty())
return false;
APInt A, ResA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateANDii(A, A2, ResA);
if (!Eval)
return false;
const Constant *C = intToConst(ResA);
Result.add(C);
}
return !Result.isBottom();
}
bool MachineConstEvaluator::evaluateANDii(const APInt &A1,
const APInt &A2, APInt &Result) {
Result = A1 & A2;
return true;
}
bool MachineConstEvaluator::evaluateORrr(const RegisterSubReg &R1,
const RegisterSubReg &R2, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg));
const LatticeCell &L1 = Inputs.get(R2.Reg);
const LatticeCell &L2 = Inputs.get(R2.Reg);
// If both sources are bottom, exit. Otherwise try to evaluate ORri
// with the non-bottom argument passed as the immediate. This is to
// catch cases of ORing with -1.
if (L2.isBottom()) {
if (L1.isBottom())
return false;
return evaluateORrr(R2, R1, Inputs, Result);
}
LatticeCell LS2;
if (!evaluate(R2, L2, LS2))
return false;
if (LS2.isBottom() || LS2.isProperty())
return false;
APInt A;
for (unsigned i = 0; i < LS2.size(); ++i) {
LatticeCell RC;
bool Eval = constToInt(LS2.Values[i], A) &&
evaluateORri(R1, A, Inputs, RC);
if (!Eval)
return false;
Result.meet(RC);
}
return !Result.isBottom();
}
bool MachineConstEvaluator::evaluateORri(const RegisterSubReg &R1,
const APInt &A2, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
if (A2 == 0)
return getCell(R1, Inputs, Result);
if (A2 == -1) {
LatticeCell RC;
RC.add(intToConst(A2));
// Overwrite Result.
Result = RC;
return true;
}
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isBottom() || LS1.isProperty())
return false;
APInt A, ResA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateORii(A, A2, ResA);
if (!Eval)
return false;
const Constant *C = intToConst(ResA);
Result.add(C);
}
return !Result.isBottom();
}
bool MachineConstEvaluator::evaluateORii(const APInt &A1,
const APInt &A2, APInt &Result) {
Result = A1 | A2;
return true;
}
bool MachineConstEvaluator::evaluateXORrr(const RegisterSubReg &R1,
const RegisterSubReg &R2, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg));
LatticeCell LS1, LS2;
if (!getCell(R1, Inputs, LS1) || !getCell(R2, Inputs, LS2))
return false;
if (LS1.isProperty()) {
if (LS1.properties() & ConstantProperties::Zero)
return !(Result = LS2).isBottom();
return false;
}
if (LS2.isProperty()) {
if (LS2.properties() & ConstantProperties::Zero)
return !(Result = LS1).isBottom();
return false;
}
APInt A;
for (unsigned i = 0; i < LS2.size(); ++i) {
LatticeCell RC;
bool Eval = constToInt(LS2.Values[i], A) &&
evaluateXORri(R1, A, Inputs, RC);
if (!Eval)
return false;
Result.meet(RC);
}
return !Result.isBottom();
}
bool MachineConstEvaluator::evaluateXORri(const RegisterSubReg &R1,
const APInt &A2, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isProperty()) {
if (LS1.properties() & ConstantProperties::Zero) {
const Constant *C = intToConst(A2);
Result.add(C);
return !Result.isBottom();
}
return false;
}
APInt A, XA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateXORii(A, A2, XA);
if (!Eval)
return false;
const Constant *C = intToConst(XA);
Result.add(C);
}
return !Result.isBottom();
}
bool MachineConstEvaluator::evaluateXORii(const APInt &A1,
const APInt &A2, APInt &Result) {
Result = A1 ^ A2;
return true;
}
bool MachineConstEvaluator::evaluateZEXTr(const RegisterSubReg &R1, unsigned Width,
unsigned Bits, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isProperty())
return false;
APInt A, XA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateZEXTi(A, Width, Bits, XA);
if (!Eval)
return false;
const Constant *C = intToConst(XA);
Result.add(C);
}
return true;
}
bool MachineConstEvaluator::evaluateZEXTi(const APInt &A1, unsigned Width,
unsigned Bits, APInt &Result) {
unsigned BW = A1.getBitWidth();
(void)BW;
assert(Width >= Bits && BW >= Bits);
APInt Mask = APInt::getLowBitsSet(Width, Bits);
Result = A1.zextOrTrunc(Width) & Mask;
return true;
}
bool MachineConstEvaluator::evaluateSEXTr(const RegisterSubReg &R1, unsigned Width,
unsigned Bits, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isBottom() || LS1.isProperty())
return false;
APInt A, XA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateSEXTi(A, Width, Bits, XA);
if (!Eval)
return false;
const Constant *C = intToConst(XA);
Result.add(C);
}
return true;
}
bool MachineConstEvaluator::evaluateSEXTi(const APInt &A1, unsigned Width,
unsigned Bits, APInt &Result) {
unsigned BW = A1.getBitWidth();
assert(Width >= Bits && BW >= Bits);
// Special case to make things faster for smaller source widths.
// Sign extension of 0 bits generates 0 as a result. This is consistent
// with what the HW does.
if (Bits == 0) {
Result = APInt(Width, 0);
return true;
}
// In C, shifts by 64 invoke undefined behavior: handle that case in APInt.
if (BW <= 64 && Bits != 0) {
int64_t V = A1.getSExtValue();
switch (Bits) {
case 8:
V = static_cast<int8_t>(V);
break;
case 16:
V = static_cast<int16_t>(V);
break;
case 32:
V = static_cast<int32_t>(V);
break;
default:
// Shift left to lose all bits except lower "Bits" bits, then shift
// the value back, replicating what was a sign bit after the first
// shift.
V = (V << (64-Bits)) >> (64-Bits);
break;
}
// V is a 64-bit sign-extended value. Convert it to APInt of desired
// width.
Result = APInt(Width, V, true);
return true;
}
// Slow case: the value doesn't fit in int64_t.
if (Bits < BW)
Result = A1.trunc(Bits).sext(Width);
else // Bits == BW
Result = A1.sext(Width);
return true;
}
bool MachineConstEvaluator::evaluateCLBr(const RegisterSubReg &R1, bool Zeros,
bool Ones, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isBottom() || LS1.isProperty())
return false;
APInt A, CA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateCLBi(A, Zeros, Ones, CA);
if (!Eval)
return false;
const Constant *C = intToConst(CA);
Result.add(C);
}
return true;
}
bool MachineConstEvaluator::evaluateCLBi(const APInt &A1, bool Zeros,
bool Ones, APInt &Result) {
unsigned BW = A1.getBitWidth();
if (!Zeros && !Ones)
return false;
unsigned Count = 0;
if (Zeros && (Count == 0))
Count = A1.countLeadingZeros();
if (Ones && (Count == 0))
Count = A1.countLeadingOnes();
Result = APInt(BW, static_cast<uint64_t>(Count), false);
return true;
}
bool MachineConstEvaluator::evaluateCTBr(const RegisterSubReg &R1, bool Zeros,
bool Ones, const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isBottom() || LS1.isProperty())
return false;
APInt A, CA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateCTBi(A, Zeros, Ones, CA);
if (!Eval)
return false;
const Constant *C = intToConst(CA);
Result.add(C);
}
return true;
}
bool MachineConstEvaluator::evaluateCTBi(const APInt &A1, bool Zeros,
bool Ones, APInt &Result) {
unsigned BW = A1.getBitWidth();
if (!Zeros && !Ones)
return false;
unsigned Count = 0;
if (Zeros && (Count == 0))
Count = A1.countTrailingZeros();
if (Ones && (Count == 0))
Count = A1.countTrailingOnes();
Result = APInt(BW, static_cast<uint64_t>(Count), false);
return true;
}
bool MachineConstEvaluator::evaluateEXTRACTr(const RegisterSubReg &R1,
unsigned Width, unsigned Bits, unsigned Offset, bool Signed,
const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
assert(Bits+Offset <= Width);
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isBottom())
return false;
if (LS1.isProperty()) {
uint32_t Ps = LS1.properties();
if (Ps & ConstantProperties::Zero) {
const Constant *C = intToConst(APInt(Width, 0, false));
Result.add(C);
return true;
}
return false;
}
APInt A, CA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateEXTRACTi(A, Bits, Offset, Signed, CA);
if (!Eval)
return false;
const Constant *C = intToConst(CA);
Result.add(C);
}
return true;
}
bool MachineConstEvaluator::evaluateEXTRACTi(const APInt &A1, unsigned Bits,
unsigned Offset, bool Signed, APInt &Result) {
unsigned BW = A1.getBitWidth();
assert(Bits+Offset <= BW);
// Extracting 0 bits generates 0 as a result (as indicated by the HW people).
if (Bits == 0) {
Result = APInt(BW, 0);
return true;
}
if (BW <= 64) {
int64_t V = A1.getZExtValue();
V <<= (64-Bits-Offset);
if (Signed)
V >>= (64-Bits);
else
V = static_cast<uint64_t>(V) >> (64-Bits);
Result = APInt(BW, V, Signed);
return true;
}
if (Signed)
Result = A1.shl(BW-Bits-Offset).ashr(BW-Bits);
else
Result = A1.shl(BW-Bits-Offset).lshr(BW-Bits);
return true;
}
bool MachineConstEvaluator::evaluateSplatr(const RegisterSubReg &R1,
unsigned Bits, unsigned Count, const CellMap &Inputs,
LatticeCell &Result) {
assert(Inputs.has(R1.Reg));
LatticeCell LS1;
if (!getCell(R1, Inputs, LS1))
return false;
if (LS1.isBottom() || LS1.isProperty())
return false;
APInt A, SA;
for (unsigned i = 0; i < LS1.size(); ++i) {
bool Eval = constToInt(LS1.Values[i], A) &&
evaluateSplati(A, Bits, Count, SA);
if (!Eval)
return false;
const Constant *C = intToConst(SA);
Result.add(C);
}
return true;
}
bool MachineConstEvaluator::evaluateSplati(const APInt &A1, unsigned Bits,
unsigned Count, APInt &Result) {
assert(Count > 0);
unsigned BW = A1.getBitWidth(), SW = Count*Bits;
APInt LoBits = (Bits < BW) ? A1.trunc(Bits) : A1.zext(Bits);
if (Count > 1)
LoBits = LoBits.zext(SW);
APInt Res(SW, 0, false);
for (unsigned i = 0; i < Count; ++i) {
Res <<= Bits;
Res |= LoBits;
}
Result = Res;
return true;
}
// ----------------------------------------------------------------------
// Hexagon-specific code.
namespace llvm {
FunctionPass *createHexagonConstPropagationPass();
void initializeHexagonConstPropagationPass(PassRegistry &Registry);
} // end namespace llvm
namespace {
class HexagonConstEvaluator : public MachineConstEvaluator {
public:
HexagonConstEvaluator(MachineFunction &Fn);
bool evaluate(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs) override;
bool evaluate(const RegisterSubReg &R, const LatticeCell &SrcC,
LatticeCell &Result) override;
bool evaluate(const MachineInstr &BrI, const CellMap &Inputs,
SetVector<const MachineBasicBlock*> &Targets, bool &FallsThru)
override;
bool rewrite(MachineInstr &MI, const CellMap &Inputs) override;
private:
unsigned getRegBitWidth(unsigned Reg) const;
static uint32_t getCmp(unsigned Opc);
static APInt getCmpImm(unsigned Opc, unsigned OpX,
const MachineOperand &MO);
void replaceWithNop(MachineInstr &MI);
bool evaluateHexRSEQ32(RegisterSubReg RL, RegisterSubReg RH, const CellMap &Inputs,
LatticeCell &Result);
bool evaluateHexCompare(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs);
// This is suitable to be called for compare-and-jump instructions.
bool evaluateHexCompare2(uint32_t Cmp, const MachineOperand &Src1,
const MachineOperand &Src2, const CellMap &Inputs, bool &Result);
bool evaluateHexLogical(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs);
bool evaluateHexCondMove(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs);
bool evaluateHexExt(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs);
bool evaluateHexVector1(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs);
bool evaluateHexVector2(const MachineInstr &MI, const CellMap &Inputs,
CellMap &Outputs);
void replaceAllRegUsesWith(Register FromReg, Register ToReg);
bool rewriteHexBranch(MachineInstr &BrI, const CellMap &Inputs);
bool rewriteHexConstDefs(MachineInstr &MI, const CellMap &Inputs,
bool &AllDefs);
bool rewriteHexConstUses(MachineInstr &MI, const CellMap &Inputs);
MachineRegisterInfo *MRI;
const HexagonInstrInfo &HII;
const HexagonRegisterInfo &HRI;
};
class HexagonConstPropagation : public MachineFunctionPass {
public:
static char ID;
HexagonConstPropagation() : MachineFunctionPass(ID) {}
StringRef getPassName() const override {
return "Hexagon Constant Propagation";
}
bool runOnMachineFunction(MachineFunction &MF) override {
const Function &F = MF.getFunction();
if (skipFunction(F))
return false;
HexagonConstEvaluator HCE(MF);
return MachineConstPropagator(HCE).run(MF);
}
};
} // end anonymous namespace
char HexagonConstPropagation::ID = 0;
INITIALIZE_PASS(HexagonConstPropagation, "hexagon-constp",
"Hexagon Constant Propagation", false, false)
HexagonConstEvaluator::HexagonConstEvaluator(MachineFunction &Fn)
: MachineConstEvaluator(Fn),
HII(*Fn.getSubtarget<HexagonSubtarget>().getInstrInfo()),
HRI(*Fn.getSubtarget<HexagonSubtarget>().getRegisterInfo()) {
MRI = &Fn.getRegInfo();
}
bool HexagonConstEvaluator::evaluate(const MachineInstr &MI,
const CellMap &Inputs, CellMap &Outputs) {
if (MI.isCall())
return false;
if (MI.getNumOperands() == 0 || !MI.getOperand(0).isReg())
return false;
const MachineOperand &MD = MI.getOperand(0);
if (!MD.isDef())
return false;
unsigned Opc = MI.getOpcode();
RegisterSubReg DefR(MD);
assert(!DefR.SubReg);
if (!DefR.Reg.isVirtual())
return false;
if (MI.isCopy()) {
LatticeCell RC;
RegisterSubReg SrcR(MI.getOperand(1));
bool Eval = evaluateCOPY(SrcR, Inputs, RC);
if (!Eval)
return false;
Outputs.update(DefR.Reg, RC);
return true;
}
if (MI.isRegSequence()) {
unsigned Sub1 = MI.getOperand(2).getImm();
unsigned Sub2 = MI.getOperand(4).getImm();
const TargetRegisterClass &DefRC = *MRI->getRegClass(DefR.Reg);
unsigned SubLo = HRI.getHexagonSubRegIndex(DefRC, Hexagon::ps_sub_lo);
unsigned SubHi = HRI.getHexagonSubRegIndex(DefRC, Hexagon::ps_sub_hi);
if (Sub1 != SubLo && Sub1 != SubHi)
return false;
if (Sub2 != SubLo && Sub2 != SubHi)
return false;
assert(Sub1 != Sub2);
bool LoIs1 = (Sub1 == SubLo);
const MachineOperand &OpLo = LoIs1 ? MI.getOperand(1) : MI.getOperand(3);
const MachineOperand &OpHi = LoIs1 ? MI.getOperand(3) : MI.getOperand(1);
LatticeCell RC;
RegisterSubReg SrcRL(OpLo), SrcRH(OpHi);
bool Eval = evaluateHexRSEQ32(SrcRL, SrcRH, Inputs, RC);
if (!Eval)
return false;
Outputs.update(DefR.Reg, RC);
return true;
}
if (MI.isCompare()) {
bool Eval = evaluateHexCompare(MI, Inputs, Outputs);
return Eval;
}
switch (Opc) {
default:
return false;
case Hexagon::A2_tfrsi:
case Hexagon::A2_tfrpi:
case Hexagon::CONST32:
case Hexagon::CONST64:
{
const MachineOperand &VO = MI.getOperand(1);
// The operand of CONST32 can be a blockaddress, e.g.
// %0 = CONST32 <blockaddress(@eat, %l)>
// Do this check for all instructions for safety.
if (!VO.isImm())
return false;
int64_t V = MI.getOperand(1).getImm();
unsigned W = getRegBitWidth(DefR.Reg);
if (W != 32 && W != 64)
return false;
IntegerType *Ty = (W == 32) ? Type::getInt32Ty(CX)
: Type::getInt64Ty(CX);
const ConstantInt *CI = ConstantInt::get(Ty, V, true);
LatticeCell RC = Outputs.get(DefR.Reg);
RC.add(CI);
Outputs.update(DefR.Reg, RC);
break;
}
case Hexagon::PS_true:
case Hexagon::PS_false:
{
LatticeCell RC = Outputs.get(DefR.Reg);
bool NonZero = (Opc == Hexagon::PS_true);
uint32_t P = NonZero ? ConstantProperties::NonZero
: ConstantProperties::Zero;
RC.add(P);
Outputs.update(DefR.Reg, RC);
break;
}
case Hexagon::A2_and:
case Hexagon::A2_andir:
case Hexagon::A2_andp:
case Hexagon::A2_or:
case Hexagon::A2_orir:
case Hexagon::A2_orp:
case Hexagon::A2_xor:
case Hexagon::A2_xorp:
{
bool Eval = evaluateHexLogical(MI, Inputs, Outputs);
if (!Eval)
return false;
break;
}
case Hexagon::A2_combineii: // combine(#s8Ext, #s8)
case Hexagon::A4_combineii: // combine(#s8, #u6Ext)
{
if (!MI.getOperand(1).isImm() || !MI.getOperand(2).isImm())
return false;
uint64_t Hi = MI.getOperand(1).getImm();
uint64_t Lo = MI.getOperand(2).getImm();
uint64_t Res = (Hi << 32) | (Lo & 0xFFFFFFFF);
IntegerType *Ty = Type::getInt64Ty(CX);
const ConstantInt *CI = ConstantInt::get(Ty, Res, false);
LatticeCell RC = Outputs.get(DefR.Reg);
RC.add(CI);
Outputs.update(DefR.Reg, RC);
break;
}
case Hexagon::S2_setbit_i:
{
int64_t B = MI.getOperand(2).getImm();
assert(B >=0 && B < 32);
APInt A(32, (1ull << B), false);
RegisterSubReg R(MI.getOperand(1));
LatticeCell RC = Outputs.get(DefR.Reg);
bool Eval = evaluateORri(R, A, Inputs, RC);
if (!Eval)
return false;
Outputs.update(DefR.Reg, RC);
break;
}
case Hexagon::C2_mux:
case Hexagon::C2_muxir:
case Hexagon::C2_muxri:
case Hexagon::C2_muxii:
{
bool Eval = evaluateHexCondMove(MI, Inputs, Outputs);
if (!Eval)
return false;
break;
}
case Hexagon::A2_sxtb:
case Hexagon::A2_sxth:
case Hexagon::A2_sxtw:
case Hexagon::A2_zxtb:
case Hexagon::A2_zxth:
{
bool Eval = evaluateHexExt(MI, Inputs, Outputs);
if (!Eval)
return false;
break;
}
case Hexagon::S2_ct0:
case Hexagon::S2_ct0p:
case Hexagon::S2_ct1:
case Hexagon::S2_ct1p:
{
using namespace Hexagon;
bool Ones = (Opc == S2_ct1) || (Opc == S2_ct1p);
RegisterSubReg R1(MI.getOperand(1));
assert(Inputs.has(R1.Reg));
LatticeCell T;
bool Eval = evaluateCTBr(R1, !Ones, Ones, Inputs, T);
if (!Eval)
return false;
// All of these instructions return a 32-bit value. The evaluate
// will generate the same type as the operand, so truncate the
// result if necessary.
APInt C;
LatticeCell RC = Outputs.get(DefR.Reg);
for (unsigned i = 0; i < T.size(); ++i) {
const Constant *CI = T.Values[i];
if (constToInt(CI, C) && C.getBitWidth() > 32)
CI = intToConst(C.trunc(32));
RC.add(CI);
}
Outputs.update(DefR.Reg, RC);
break;
}
case Hexagon::S2_cl0:
case Hexagon::S2_cl0p:
case Hexagon::S2_cl1:
case Hexagon::S2_cl1p:
case Hexagon::S2_clb:
case Hexagon::S2_clbp:
{
using namespace Hexagon;
bool OnlyZeros = (Opc == S2_cl0) || (Opc == S2_cl0p);
bool OnlyOnes = (Opc == S2_cl1) || (Opc == S2_cl1p);
RegisterSubReg R1(MI.getOperand(1));
assert(Inputs.has(R1.Reg));
LatticeCell T;
bool Eval = evaluateCLBr(R1, !OnlyOnes, !OnlyZeros, Inputs, T);
if (!Eval)
return false;
// All of these instructions return a 32-bit value. The evaluate
// will generate the same type as the operand, so truncate the
// result if necessary.
APInt C;
LatticeCell RC = Outputs.get(DefR.Reg);
for (unsigned i = 0; i < T.size(); ++i) {
const Constant *CI = T.Values[i];
if (constToInt(CI, C) && C.getBitWidth() > 32)
CI = intToConst(C.trunc(32));
RC.add(CI);
}
Outputs.update(DefR.Reg, RC);
break;
}
case Hexagon::S4_extract:
case Hexagon::S4_extractp:
case Hexagon::S2_extractu:
case Hexagon::S2_extractup:
{
bool Signed = (Opc == Hexagon::S4_extract) ||
(Opc == Hexagon::S4_extractp);
RegisterSubReg R1(MI.getOperand(1));
unsigned BW = getRegBitWidth(R1.Reg);
unsigned Bits = MI.getOperand(2).getImm();
unsigned Offset = MI.getOperand(3).getImm();
LatticeCell RC = Outputs.get(DefR.Reg);
if (Offset >= BW) {
APInt Zero(BW, 0, false);
RC.add(intToConst(Zero));
break;
}
if (Offset+Bits > BW) {
// If the requested bitfield extends beyond the most significant bit,
// the extra bits are treated as 0s. To emulate this behavior, reduce
// the number of requested bits, and make the extract unsigned.
Bits = BW-Offset;
Signed = false;
}
bool Eval = evaluateEXTRACTr(R1, BW, Bits, Offset, Signed, Inputs, RC);
if (!Eval)
return false;
Outputs.update(DefR.Reg, RC);
break;
}
case Hexagon::S2_vsplatrb:
case Hexagon::S2_vsplatrh:
// vabsh, vabsh:sat
// vabsw, vabsw:sat
// vconj:sat
// vrndwh, vrndwh:sat
// vsathb, vsathub, vsatwuh
// vsxtbh, vsxthw
// vtrunehb, vtrunohb
// vzxtbh, vzxthw
{
bool Eval = evaluateHexVector1(MI, Inputs, Outputs);
if (!Eval)
return false;
break;
}
// TODO:
// A2_vaddh
// A2_vaddhs
// A2_vaddw
// A2_vaddws
}
return true;
}
bool HexagonConstEvaluator::evaluate(const RegisterSubReg &R,
const LatticeCell &Input, LatticeCell &Result) {
if (!R.SubReg) {
Result = Input;
return true;
}
const TargetRegisterClass *RC = MRI->getRegClass(R.Reg);
if (RC != &Hexagon::DoubleRegsRegClass)
return false;
if (R.SubReg != Hexagon::isub_lo && R.SubReg != Hexagon::isub_hi)
return false;
assert(!Input.isTop());
if (Input.isBottom())
return false;
using P = ConstantProperties;
if (Input.isProperty()) {
uint32_t Ps = Input.properties();
if (Ps & (P::Zero|P::NaN)) {
uint32_t Ns = (Ps & (P::Zero|P::NaN|P::SignProperties));
Result.add(Ns);
return true;
}
if (R.SubReg == Hexagon::isub_hi) {
uint32_t Ns = (Ps & P::SignProperties);
Result.add(Ns);
return true;
}
return false;
}
// The Input cell contains some known values. Pick the word corresponding
// to the subregister.
APInt A;
for (unsigned i = 0; i < Input.size(); ++i) {
const Constant *C = Input.Values[i];
if (!constToInt(C, A))
return false;
if (!A.isIntN(64))
return false;
uint64_t U = A.getZExtValue();
if (R.SubReg == Hexagon::isub_hi)
U >>= 32;
U &= 0xFFFFFFFFULL;
uint32_t U32 = Lo_32(U);
int32_t V32;
memcpy(&V32, &U32, sizeof V32);
IntegerType *Ty = Type::getInt32Ty(CX);
const ConstantInt *C32 = ConstantInt::get(Ty, static_cast<int64_t>(V32));
Result.add(C32);
}
return true;
}
bool HexagonConstEvaluator::evaluate(const MachineInstr &BrI,
const CellMap &Inputs, SetVector<const MachineBasicBlock*> &Targets,
bool &FallsThru) {
// We need to evaluate one branch at a time. TII::analyzeBranch checks
// all the branches in a basic block at once, so we cannot use it.
unsigned Opc = BrI.getOpcode();
bool SimpleBranch = false;
bool Negated = false;
switch (Opc) {
case Hexagon::J2_jumpf:
case Hexagon::J2_jumpfnew:
case Hexagon::J2_jumpfnewpt:
Negated = true;
[[fallthrough]];
case Hexagon::J2_jumpt:
case Hexagon::J2_jumptnew:
case Hexagon::J2_jumptnewpt:
// Simple branch: if([!]Pn) jump ...
// i.e. Op0 = predicate, Op1 = branch target.
SimpleBranch = true;
break;
case Hexagon::J2_jump:
Targets.insert(BrI.getOperand(0).getMBB());
FallsThru = false;
return true;
default:
Undetermined:
// If the branch is of unknown type, assume that all successors are
// executable.
FallsThru = !BrI.isUnconditionalBranch();
return false;
}
if (SimpleBranch) {
const MachineOperand &MD = BrI.getOperand(0);
RegisterSubReg PR(MD);
// If the condition operand has a subregister, this is not something
// we currently recognize.
if (PR.SubReg)
goto Undetermined;
assert(Inputs.has(PR.Reg));
const LatticeCell &PredC = Inputs.get(PR.Reg);
if (PredC.isBottom())
goto Undetermined;
uint32_t Props = PredC.properties();
bool CTrue = false, CFalse = false;
if (Props & ConstantProperties::Zero)
CFalse = true;
else if (Props & ConstantProperties::NonZero)
CTrue = true;
// If the condition is not known to be either, bail out.
if (!CTrue && !CFalse)
goto Undetermined;
const MachineBasicBlock *BranchTarget = BrI.getOperand(1).getMBB();
FallsThru = false;
if ((!Negated && CTrue) || (Negated && CFalse))
Targets.insert(BranchTarget);
else if ((!Negated && CFalse) || (Negated && CTrue))
FallsThru = true;
else
goto Undetermined;
}
return true;
}
bool HexagonConstEvaluator::rewrite(MachineInstr &MI, const CellMap &Inputs) {
if (MI.isBranch())
return rewriteHexBranch(MI, Inputs);
unsigned Opc = MI.getOpcode();
switch (Opc) {
default:
break;
case Hexagon::A2_tfrsi:
case Hexagon::A2_tfrpi:
case Hexagon::CONST32:
case Hexagon::CONST64:
case Hexagon::PS_true:
case Hexagon::PS_false:
return false;
}
unsigned NumOp = MI.getNumOperands();
if (NumOp == 0)
return false;
bool AllDefs, Changed;
Changed = rewriteHexConstDefs(MI, Inputs, AllDefs);
// If not all defs have been rewritten (i.e. the instruction defines
// a register that is not compile-time constant), then try to rewrite
// register operands that are known to be constant with immediates.
if (!AllDefs)
Changed |= rewriteHexConstUses(MI, Inputs);
return Changed;
}
unsigned HexagonConstEvaluator::getRegBitWidth(unsigned Reg) const {
const TargetRegisterClass *RC = MRI->getRegClass(Reg);
if (Hexagon::IntRegsRegClass.hasSubClassEq(RC))
return 32;
if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC))
return 64;
if (Hexagon::PredRegsRegClass.hasSubClassEq(RC))
return 8;
llvm_unreachable("Invalid register");
return 0;
}
uint32_t HexagonConstEvaluator::getCmp(unsigned Opc) {
switch (Opc) {
case Hexagon::C2_cmpeq:
case Hexagon::C2_cmpeqp:
case Hexagon::A4_cmpbeq:
case Hexagon::A4_cmpheq:
case Hexagon::A4_cmpbeqi:
case Hexagon::A4_cmpheqi:
case Hexagon::C2_cmpeqi:
case Hexagon::J4_cmpeqn1_t_jumpnv_nt:
case Hexagon::J4_cmpeqn1_t_jumpnv_t:
case Hexagon::J4_cmpeqi_t_jumpnv_nt:
case Hexagon::J4_cmpeqi_t_jumpnv_t:
case Hexagon::J4_cmpeq_t_jumpnv_nt:
case Hexagon::J4_cmpeq_t_jumpnv_t:
return Comparison::EQ;
case Hexagon::C4_cmpneq:
case Hexagon::C4_cmpneqi:
case Hexagon::J4_cmpeqn1_f_jumpnv_nt:
case Hexagon::J4_cmpeqn1_f_jumpnv_t:
case Hexagon::J4_cmpeqi_f_jumpnv_nt:
case Hexagon::J4_cmpeqi_f_jumpnv_t:
case Hexagon::J4_cmpeq_f_jumpnv_nt:
case Hexagon::J4_cmpeq_f_jumpnv_t:
return Comparison::NE;
case Hexagon::C2_cmpgt:
case Hexagon::C2_cmpgtp:
case Hexagon::A4_cmpbgt:
case Hexagon::A4_cmphgt:
case Hexagon::A4_cmpbgti:
case Hexagon::A4_cmphgti:
case Hexagon::C2_cmpgti:
case Hexagon::J4_cmpgtn1_t_jumpnv_nt:
case Hexagon::J4_cmpgtn1_t_jumpnv_t:
case Hexagon::J4_cmpgti_t_jumpnv_nt:
case Hexagon::J4_cmpgti_t_jumpnv_t:
case Hexagon::J4_cmpgt_t_jumpnv_nt:
case Hexagon::J4_cmpgt_t_jumpnv_t:
return Comparison::GTs;
case Hexagon::C4_cmplte:
case Hexagon::C4_cmpltei:
case Hexagon::J4_cmpgtn1_f_jumpnv_nt:
case Hexagon::J4_cmpgtn1_f_jumpnv_t:
case Hexagon::J4_cmpgti_f_jumpnv_nt:
case Hexagon::J4_cmpgti_f_jumpnv_t:
case Hexagon::J4_cmpgt_f_jumpnv_nt:
case Hexagon::J4_cmpgt_f_jumpnv_t:
return Comparison::LEs;
case Hexagon::C2_cmpgtu:
case Hexagon::C2_cmpgtup:
case Hexagon::A4_cmpbgtu:
case Hexagon::A4_cmpbgtui:
case Hexagon::A4_cmphgtu:
case Hexagon::A4_cmphgtui:
case Hexagon::C2_cmpgtui:
case Hexagon::J4_cmpgtui_t_jumpnv_nt:
case Hexagon::J4_cmpgtui_t_jumpnv_t:
case Hexagon::J4_cmpgtu_t_jumpnv_nt:
case Hexagon::J4_cmpgtu_t_jumpnv_t:
return Comparison::GTu;
case Hexagon::J4_cmpltu_f_jumpnv_nt:
case Hexagon::J4_cmpltu_f_jumpnv_t:
return Comparison::GEu;
case Hexagon::J4_cmpltu_t_jumpnv_nt:
case Hexagon::J4_cmpltu_t_jumpnv_t:
return Comparison::LTu;
case Hexagon::J4_cmplt_f_jumpnv_nt:
case Hexagon::J4_cmplt_f_jumpnv_t:
return Comparison::GEs;
case Hexagon::C4_cmplteu:
case Hexagon::C4_cmplteui:
case Hexagon::J4_cmpgtui_f_jumpnv_nt:
case Hexagon::J4_cmpgtui_f_jumpnv_t:
case Hexagon::J4_cmpgtu_f_jumpnv_nt:
case Hexagon::J4_cmpgtu_f_jumpnv_t:
return Comparison::LEu;
case Hexagon::J4_cmplt_t_jumpnv_nt:
case Hexagon::J4_cmplt_t_jumpnv_t:
return Comparison::LTs;
default:
break;
}
return Comparison::Unk;
}
APInt HexagonConstEvaluator::getCmpImm(unsigned Opc, unsigned OpX,
const MachineOperand &MO) {
bool Signed = false;
switch (Opc) {
case Hexagon::A4_cmpbgtui: // u7
case Hexagon::A4_cmphgtui: // u7
break;
case Hexagon::A4_cmpheqi: // s8
case Hexagon::C4_cmpneqi: // s8
Signed = true;
break;
case Hexagon::A4_cmpbeqi: // u8
break;
case Hexagon::C2_cmpgtui: // u9
case Hexagon::C4_cmplteui: // u9
break;
case Hexagon::C2_cmpeqi: // s10
case Hexagon::C2_cmpgti: // s10
case Hexagon::C4_cmpltei: // s10
Signed = true;
break;
case Hexagon::J4_cmpeqi_f_jumpnv_nt: // u5
case Hexagon::J4_cmpeqi_f_jumpnv_t: // u5
case Hexagon::J4_cmpeqi_t_jumpnv_nt: // u5
case Hexagon::J4_cmpeqi_t_jumpnv_t: // u5
case Hexagon::J4_cmpgti_f_jumpnv_nt: // u5
case Hexagon::J4_cmpgti_f_jumpnv_t: // u5
case Hexagon::J4_cmpgti_t_jumpnv_nt: // u5
case Hexagon::J4_cmpgti_t_jumpnv_t: // u5
case Hexagon::J4_cmpgtui_f_jumpnv_nt: // u5
case Hexagon::J4_cmpgtui_f_jumpnv_t: // u5
case Hexagon::J4_cmpgtui_t_jumpnv_nt: // u5
case Hexagon::J4_cmpgtui_t_jumpnv_t: // u5
break;
default:
llvm_unreachable("Unhandled instruction");
break;
}
uint64_t Val = MO.getImm();
return APInt(32, Val, Signed);
}
void HexagonConstEvaluator::replaceWithNop(MachineInstr &MI) {
MI.setDesc(HII.get(Hexagon::A2_nop));
while (MI.getNumOperands() > 0)
MI.removeOperand(0);
}
bool HexagonConstEvaluator::evaluateHexRSEQ32(RegisterSubReg RL, RegisterSubReg RH,
const CellMap &Inputs, LatticeCell &Result) {
assert(Inputs.has(RL.Reg) && Inputs.has(RH.Reg));
LatticeCell LSL, LSH;
if (!getCell(RL, Inputs, LSL) || !getCell(RH, Inputs, LSH))
return false;
if (LSL.isProperty() || LSH.isProperty())
return false;
unsigned LN = LSL.size(), HN = LSH.size();
SmallVector<APInt,4> LoVs(LN), HiVs(HN);
for (unsigned i = 0; i < LN; ++i) {
bool Eval = constToInt(LSL.Values[i], LoVs[i]);
if (!Eval)
return false;
assert(LoVs[i].getBitWidth() == 32);
}
for (unsigned i = 0; i < HN; ++i) {
bool Eval = constToInt(LSH.Values[i], HiVs[i]);
if (!Eval)
return false;
assert(HiVs[i].getBitWidth() == 32);
}
for (unsigned i = 0; i < HiVs.size(); ++i) {
APInt HV = HiVs[i].zext(64) << 32;
for (unsigned j = 0; j < LoVs.size(); ++j) {
APInt LV = LoVs[j].zext(64);
const Constant *C = intToConst(HV | LV);
Result.add(C);
if (Result.isBottom())
return false;
}
}
return !Result.isBottom();
}
bool HexagonConstEvaluator::evaluateHexCompare(const MachineInstr &MI,
const CellMap &Inputs, CellMap &Outputs) {
unsigned Opc = MI.getOpcode();
bool Classic = false;
switch (Opc) {
case Hexagon::C2_cmpeq:
case Hexagon::C2_cmpeqp:
case Hexagon::C2_cmpgt:
case Hexagon::C2_cmpgtp:
case Hexagon::C2_cmpgtu:
case Hexagon::C2_cmpgtup:
case Hexagon::C2_cmpeqi:
case Hexagon::C2_cmpgti:
case Hexagon::C2_cmpgtui:
// Classic compare: Dst0 = CMP Src1, Src2
Classic = true;
break;
default:
// Not handling other compare instructions now.
return false;
}
if (Classic) {
const MachineOperand &Src1 = MI.getOperand(1);
const MachineOperand &Src2 = MI.getOperand(2);
bool Result;
unsigned Opc = MI.getOpcode();
bool Computed = evaluateHexCompare2(Opc, Src1, Src2, Inputs, Result);
if (Computed) {
// Only create a zero/non-zero cell. At this time there isn't really
// much need for specific values.
RegisterSubReg DefR(MI.getOperand(0));
LatticeCell L = Outputs.get(DefR.Reg);
uint32_t P = Result ? ConstantProperties::NonZero
: ConstantProperties::Zero;
L.add(P);
Outputs.update(DefR.Reg, L);
return true;
}
}
return false;
}
bool HexagonConstEvaluator::evaluateHexCompare2(unsigned Opc,
const MachineOperand &Src1, const MachineOperand &Src2,
const CellMap &Inputs, bool &Result) {
uint32_t Cmp = getCmp(Opc);
bool Reg1 = Src1.isReg(), Reg2 = Src2.isReg();
bool Imm1 = Src1.isImm(), Imm2 = Src2.isImm();
if (Reg1) {
RegisterSubReg R1(Src1);
if (Reg2) {
RegisterSubReg R2(Src2);
return evaluateCMPrr(Cmp, R1, R2, Inputs, Result);
} else if (Imm2) {
APInt A2 = getCmpImm(Opc, 2, Src2);
return evaluateCMPri(Cmp, R1, A2, Inputs, Result);
}
} else if (Imm1) {
APInt A1 = getCmpImm(Opc, 1, Src1);
if (Reg2) {
RegisterSubReg R2(Src2);
uint32_t NegCmp = Comparison::negate(Cmp);
return evaluateCMPri(NegCmp, R2, A1, Inputs, Result);
} else if (Imm2) {
APInt A2 = getCmpImm(Opc, 2, Src2);
return evaluateCMPii(Cmp, A1, A2, Result);
}
}
// Unknown kind of comparison.
return false;
}
bool HexagonConstEvaluator::evaluateHexLogical(const MachineInstr &MI,
const CellMap &Inputs, CellMap &Outputs) {
unsigned Opc = MI.getOpcode();
if (MI.getNumOperands() != 3)
return false;
const MachineOperand &Src1 = MI.getOperand(1);
const MachineOperand &Src2 = MI.getOperand(2);
RegisterSubReg R1(Src1);
bool Eval = false;
LatticeCell RC;
switch (Opc) {
default:
return false;
case Hexagon::A2_and:
case Hexagon::A2_andp:
Eval = evaluateANDrr(R1, RegisterSubReg(Src2), Inputs, RC);
break;
case Hexagon::A2_andir: {
if (!Src2.isImm())
return false;
APInt A(32, Src2.getImm(), true);
Eval = evaluateANDri(R1, A, Inputs, RC);
break;
}
case Hexagon::A2_or:
case Hexagon::A2_orp:
Eval = evaluateORrr(R1, RegisterSubReg(Src2), Inputs, RC);
break;
case Hexagon::A2_orir: {
if (!Src2.isImm())
return false;
APInt A(32, Src2.getImm(), true);
Eval = evaluateORri(R1, A, Inputs, RC);
break;
}
case Hexagon::A2_xor:
case Hexagon::A2_xorp:
Eval = evaluateXORrr(R1, RegisterSubReg(Src2), Inputs, RC);
break;
}
if (Eval) {
RegisterSubReg DefR(MI.getOperand(0));
Outputs.update(DefR.Reg, RC);
}
return Eval;
}
bool HexagonConstEvaluator::evaluateHexCondMove(const MachineInstr &MI,
const CellMap &Inputs, CellMap &Outputs) {
// Dst0 = Cond1 ? Src2 : Src3
RegisterSubReg CR(MI.getOperand(1));
assert(Inputs.has(CR.Reg));
LatticeCell LS;
if (!getCell(CR, Inputs, LS))
return false;
uint32_t Ps = LS.properties();
unsigned TakeOp;
if (Ps & ConstantProperties::Zero)
TakeOp = 3;
else if (Ps & ConstantProperties::NonZero)
TakeOp = 2;
else
return false;
const MachineOperand &ValOp = MI.getOperand(TakeOp);
RegisterSubReg DefR(MI.getOperand(0));
LatticeCell RC = Outputs.get(DefR.Reg);
if (ValOp.isImm()) {
int64_t V = ValOp.getImm();
unsigned W = getRegBitWidth(DefR.Reg);
APInt A(W, V, true);
const Constant *C = intToConst(A);
RC.add(C);
Outputs.update(DefR.Reg, RC);
return true;
}
if (ValOp.isReg()) {
RegisterSubReg R(ValOp);
const LatticeCell &LR = Inputs.get(R.Reg);
LatticeCell LSR;
if (!evaluate(R, LR, LSR))
return false;
RC.meet(LSR);
Outputs.update(DefR.Reg, RC);
return true;
}
return false;
}
bool HexagonConstEvaluator::evaluateHexExt(const MachineInstr &MI,
const CellMap &Inputs, CellMap &Outputs) {
// Dst0 = ext R1
RegisterSubReg R1(MI.getOperand(1));
assert(Inputs.has(R1.Reg));
unsigned Opc = MI.getOpcode();
unsigned Bits;
switch (Opc) {
case Hexagon::A2_sxtb:
case Hexagon::A2_zxtb:
Bits = 8;
break;
case Hexagon::A2_sxth:
case Hexagon::A2_zxth:
Bits = 16;
break;
case Hexagon::A2_sxtw:
Bits = 32;
break;
default:
llvm_unreachable("Unhandled extension opcode");
}
bool Signed = false;
switch (Opc) {
case Hexagon::A2_sxtb:
case Hexagon::A2_sxth:
case Hexagon::A2_sxtw:
Signed = true;
break;
}
RegisterSubReg DefR(MI.getOperand(0));
unsigned BW = getRegBitWidth(DefR.Reg);
LatticeCell RC = Outputs.get(DefR.Reg);
bool Eval = Signed ? evaluateSEXTr(R1, BW, Bits, Inputs, RC)
: evaluateZEXTr(R1, BW, Bits, Inputs, RC);
if (!Eval)
return false;
Outputs.update(DefR.Reg, RC);
return true;
}
bool HexagonConstEvaluator::evaluateHexVector1(const MachineInstr &MI,
const CellMap &Inputs, CellMap &Outputs) {
// DefR = op R1
RegisterSubReg DefR(MI.getOperand(0));
RegisterSubReg R1(MI.getOperand(1));
assert(Inputs.has(R1.Reg));
LatticeCell RC = Outputs.get(DefR.Reg);
bool Eval;
unsigned Opc = MI.getOpcode();
switch (Opc) {
case Hexagon::S2_vsplatrb:
// Rd = 4 times Rs:0..7
Eval = evaluateSplatr(R1, 8, 4, Inputs, RC);
break;
case Hexagon::S2_vsplatrh:
// Rdd = 4 times Rs:0..15
Eval = evaluateSplatr(R1, 16, 4, Inputs, RC);
break;
default:
return false;
}
if (!Eval)
return false;
Outputs.update(DefR.Reg, RC);
return true;
}
bool HexagonConstEvaluator::rewriteHexConstDefs(MachineInstr &MI,
const CellMap &Inputs, bool &AllDefs) {
AllDefs = false;
// Some diagnostics.
// LLVM_DEBUG({...}) gets confused with all this code as an argument.
#ifndef NDEBUG
bool Debugging = DebugFlag && isCurrentDebugType(DEBUG_TYPE);
if (Debugging) {
bool Const = true, HasUse = false;
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || !MO.isUse() || MO.isImplicit())
continue;
RegisterSubReg R(MO);
if (!R.Reg.isVirtual())
continue;
HasUse = true;
// PHIs can legitimately have "top" cells after propagation.
if (!MI.isPHI() && !Inputs.has(R.Reg)) {
dbgs() << "Top " << printReg(R.Reg, &HRI, R.SubReg)
<< " in MI: " << MI;
continue;
}
const LatticeCell &L = Inputs.get(R.Reg);
Const &= L.isSingle();
if (!Const)
break;
}
if (HasUse && Const) {
if (!MI.isCopy()) {
dbgs() << "CONST: " << MI;
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || !MO.isUse() || MO.isImplicit())
continue;
Register R = MO.getReg();
dbgs() << printReg(R, &TRI) << ": " << Inputs.get(R) << "\n";
}
}
}
}
#endif
// Avoid generating TFRIs for register transfers---this will keep the
// coalescing opportunities.
if (MI.isCopy())
return false;
MachineFunction *MF = MI.getParent()->getParent();
auto &HST = MF->getSubtarget<HexagonSubtarget>();
// Collect all virtual register-def operands.
SmallVector<unsigned,2> DefRegs;
for (const MachineOperand &MO : MI.operands()) {
if (!MO.isReg() || !MO.isDef())
continue;
Register R = MO.getReg();
if (!R.isVirtual())
continue;
assert(!MO.getSubReg());
assert(Inputs.has(R));
DefRegs.push_back(R);
}
MachineBasicBlock &B = *MI.getParent();
const DebugLoc &DL = MI.getDebugLoc();
unsigned ChangedNum = 0;
#ifndef NDEBUG
SmallVector<const MachineInstr*,4> NewInstrs;
#endif
// For each defined register, if it is a constant, create an instruction
// NewR = const
// and replace all uses of the defined register with NewR.
for (unsigned i = 0, n = DefRegs.size(); i < n; ++i) {
unsigned R = DefRegs[i];
const LatticeCell &L = Inputs.get(R);
if (L.isBottom())
continue;
const TargetRegisterClass *RC = MRI->getRegClass(R);
MachineBasicBlock::iterator At = MI.getIterator();
if (!L.isSingle()) {
// If this a zero/non-zero cell, we can fold a definition
// of a predicate register.
using P = ConstantProperties;
uint64_t Ps = L.properties();
if (!(Ps & (P::Zero|P::NonZero)))
continue;
const TargetRegisterClass *PredRC = &Hexagon::PredRegsRegClass;
if (RC != PredRC)
continue;
const MCInstrDesc *NewD = (Ps & P::Zero) ?
&HII.get(Hexagon::PS_false) :
&HII.get(Hexagon::PS_true);
Register NewR = MRI->createVirtualRegister(PredRC);
const MachineInstrBuilder &MIB = BuildMI(B, At, DL, *NewD, NewR);
(void)MIB;
#ifndef NDEBUG
NewInstrs.push_back(&*MIB);
#endif
replaceAllRegUsesWith(R, NewR);
} else {
// This cell has a single value.
APInt A;
if (!constToInt(L.Value, A) || !A.isSignedIntN(64))
continue;
const TargetRegisterClass *NewRC;
const MCInstrDesc *NewD;
unsigned W = getRegBitWidth(R);
int64_t V = A.getSExtValue();
assert(W == 32 || W == 64);
if (W == 32)
NewRC = &Hexagon::IntRegsRegClass;
else
NewRC = &Hexagon::DoubleRegsRegClass;
Register NewR = MRI->createVirtualRegister(NewRC);
const MachineInstr *NewMI;
if (W == 32) {
NewD = &HII.get(Hexagon::A2_tfrsi);
NewMI = BuildMI(B, At, DL, *NewD, NewR)
.addImm(V);
} else {
if (A.isSignedIntN(8)) {
NewD = &HII.get(Hexagon::A2_tfrpi);
NewMI = BuildMI(B, At, DL, *NewD, NewR)
.addImm(V);
} else {
int32_t Hi = V >> 32;
int32_t Lo = V & 0xFFFFFFFFLL;
if (isInt<8>(Hi) && isInt<8>(Lo)) {
NewD = &HII.get(Hexagon::A2_combineii);
NewMI = BuildMI(B, At, DL, *NewD, NewR)
.addImm(Hi)
.addImm(Lo);
} else if (MF->getFunction().hasOptSize() || !HST.isTinyCore()) {
// Disable CONST64 for tiny core since it takes a LD resource.
NewD = &HII.get(Hexagon::CONST64);
NewMI = BuildMI(B, At, DL, *NewD, NewR)
.addImm(V);
} else
return false;
}
}
(void)NewMI;
#ifndef NDEBUG
NewInstrs.push_back(NewMI);
#endif
replaceAllRegUsesWith(R, NewR);
}
ChangedNum++;
}
LLVM_DEBUG({
if (!NewInstrs.empty()) {
MachineFunction &MF = *MI.getParent()->getParent();
dbgs() << "In function: " << MF.getName() << "\n";
dbgs() << "Rewrite: for " << MI << " created " << *NewInstrs[0];
for (unsigned i = 1; i < NewInstrs.size(); ++i)
dbgs() << " " << *NewInstrs[i];
}
});
AllDefs = (ChangedNum == DefRegs.size());
return ChangedNum > 0;
}
bool HexagonConstEvaluator::rewriteHexConstUses(MachineInstr &MI,
const CellMap &Inputs) {
bool Changed = false;
unsigned Opc = MI.getOpcode();
MachineBasicBlock &B = *MI.getParent();
const DebugLoc &DL = MI.getDebugLoc();
MachineBasicBlock::iterator At = MI.getIterator();
MachineInstr *NewMI = nullptr;
switch (Opc) {
case Hexagon::M2_maci:
// Convert DefR += mpyi(R2, R3)
// to DefR += mpyi(R, #imm),
// or DefR -= mpyi(R, #imm).
{
RegisterSubReg DefR(MI.getOperand(0));
assert(!DefR.SubReg);
RegisterSubReg R2(MI.getOperand(2));
RegisterSubReg R3(MI.getOperand(3));
assert(Inputs.has(R2.Reg) && Inputs.has(R3.Reg));
LatticeCell LS2, LS3;
// It is enough to get one of the input cells, since we will only try
// to replace one argument---whichever happens to be a single constant.
bool HasC2 = getCell(R2, Inputs, LS2), HasC3 = getCell(R3, Inputs, LS3);
if (!HasC2 && !HasC3)
return false;
bool Zero = ((HasC2 && (LS2.properties() & ConstantProperties::Zero)) ||
(HasC3 && (LS3.properties() & ConstantProperties::Zero)));
// If one of the operands is zero, eliminate the multiplication.
if (Zero) {
// DefR == R1 (tied operands).
MachineOperand &Acc = MI.getOperand(1);
RegisterSubReg R1(Acc);
unsigned NewR = R1.Reg;
if (R1.SubReg) {
// Generate COPY. FIXME: Replace with the register:subregister.
const TargetRegisterClass *RC = MRI->getRegClass(DefR.Reg);
NewR = MRI->createVirtualRegister(RC);
NewMI = BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR)
.addReg(R1.Reg, getRegState(Acc), R1.SubReg);
}
replaceAllRegUsesWith(DefR.Reg, NewR);
MRI->clearKillFlags(NewR);
Changed = true;
break;
}
bool Swap = false;
if (!LS3.isSingle()) {
if (!LS2.isSingle())
return false;
Swap = true;
}
const LatticeCell &LI = Swap ? LS2 : LS3;
const MachineOperand &OpR2 = Swap ? MI.getOperand(3)
: MI.getOperand(2);
// LI is single here.
APInt A;
if (!constToInt(LI.Value, A) || !A.isSignedIntN(8))
return false;
int64_t V = A.getSExtValue();
const MCInstrDesc &D = (V >= 0) ? HII.get(Hexagon::M2_macsip)
: HII.get(Hexagon::M2_macsin);
if (V < 0)
V = -V;
const TargetRegisterClass *RC = MRI->getRegClass(DefR.Reg);
Register NewR = MRI->createVirtualRegister(RC);
const MachineOperand &Src1 = MI.getOperand(1);
NewMI = BuildMI(B, At, DL, D, NewR)
.addReg(Src1.getReg(), getRegState(Src1), Src1.getSubReg())
.addReg(OpR2.getReg(), getRegState(OpR2), OpR2.getSubReg())
.addImm(V);
replaceAllRegUsesWith(DefR.Reg, NewR);
Changed = true;
break;
}
case Hexagon::A2_and:
{
RegisterSubReg R1(MI.getOperand(1));
RegisterSubReg R2(MI.getOperand(2));
assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg));
LatticeCell LS1, LS2;
unsigned CopyOf = 0;
// Check if any of the operands is -1 (i.e. all bits set).
if (getCell(R1, Inputs, LS1) && LS1.isSingle()) {
APInt M1;
if (constToInt(LS1.Value, M1) && !~M1)
CopyOf = 2;
}
else if (getCell(R2, Inputs, LS2) && LS2.isSingle()) {
APInt M1;
if (constToInt(LS2.Value, M1) && !~M1)
CopyOf = 1;
}
if (!CopyOf)
return false;
MachineOperand &SO = MI.getOperand(CopyOf);
RegisterSubReg SR(SO);
RegisterSubReg DefR(MI.getOperand(0));
unsigned NewR = SR.Reg;
if (SR.SubReg) {
const TargetRegisterClass *RC = MRI->getRegClass(DefR.Reg);
NewR = MRI->createVirtualRegister(RC);
NewMI = BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR)
.addReg(SR.Reg, getRegState(SO), SR.SubReg);
}
replaceAllRegUsesWith(DefR.Reg, NewR);
MRI->clearKillFlags(NewR);
Changed = true;
}
break;
case Hexagon::A2_or:
{
RegisterSubReg R1(MI.getOperand(1));
RegisterSubReg R2(MI.getOperand(2));
assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg));
LatticeCell LS1, LS2;
unsigned CopyOf = 0;
using P = ConstantProperties;
if (getCell(R1, Inputs, LS1) && (LS1.properties() & P::Zero))
CopyOf = 2;
else if (getCell(R2, Inputs, LS2) && (LS2.properties() & P::Zero))
CopyOf = 1;
if (!CopyOf)
return false;
MachineOperand &SO = MI.getOperand(CopyOf);
RegisterSubReg SR(SO);
RegisterSubReg DefR(MI.getOperand(0));
unsigned NewR = SR.Reg;
if (SR.SubReg) {
const TargetRegisterClass *RC = MRI->getRegClass(DefR.Reg);
NewR = MRI->createVirtualRegister(RC);
NewMI = BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR)
.addReg(SR.Reg, getRegState(SO), SR.SubReg);
}
replaceAllRegUsesWith(DefR.Reg, NewR);
MRI->clearKillFlags(NewR);
Changed = true;
}
break;
}
if (NewMI) {
// clear all the kill flags of this new instruction.
for (MachineOperand &MO : NewMI->operands())
if (MO.isReg() && MO.isUse())
MO.setIsKill(false);
}
LLVM_DEBUG({
if (NewMI) {
dbgs() << "Rewrite: for " << MI;
if (NewMI != &MI)
dbgs() << " created " << *NewMI;
else
dbgs() << " modified the instruction itself and created:" << *NewMI;
}
});
return Changed;
}
void HexagonConstEvaluator::replaceAllRegUsesWith(Register FromReg,
Register ToReg) {
assert(FromReg.isVirtual());
assert(ToReg.isVirtual());
for (MachineOperand &O :
llvm::make_early_inc_range(MRI->use_operands(FromReg)))
O.setReg(ToReg);
}
bool HexagonConstEvaluator::rewriteHexBranch(MachineInstr &BrI,
const CellMap &Inputs) {
MachineBasicBlock &B = *BrI.getParent();
unsigned NumOp = BrI.getNumOperands();
if (!NumOp)
return false;
bool FallsThru;
SetVector<const MachineBasicBlock*> Targets;
bool Eval = evaluate(BrI, Inputs, Targets, FallsThru);
unsigned NumTargets = Targets.size();
if (!Eval || NumTargets > 1 || (NumTargets == 1 && FallsThru))
return false;
if (BrI.getOpcode() == Hexagon::J2_jump)
return false;
LLVM_DEBUG(dbgs() << "Rewrite(" << printMBBReference(B) << "):" << BrI);
bool Rewritten = false;
if (NumTargets > 0) {
assert(!FallsThru && "This should have been checked before");
// MIB.addMBB needs non-const pointer.
MachineBasicBlock *TargetB = const_cast<MachineBasicBlock*>(Targets[0]);
bool Moot = B.isLayoutSuccessor(TargetB);
if (!Moot) {
// If we build a branch here, we must make sure that it won't be
// erased as "non-executable". We can't mark any new instructions
// as executable here, so we need to overwrite the BrI, which we
// know is executable.
const MCInstrDesc &JD = HII.get(Hexagon::J2_jump);
auto NI = BuildMI(B, BrI.getIterator(), BrI.getDebugLoc(), JD)
.addMBB(TargetB);
BrI.setDesc(JD);
while (BrI.getNumOperands() > 0)
BrI.removeOperand(0);
// This ensures that all implicit operands (e.g. implicit-def %r31, etc)
// are present in the rewritten branch.
for (auto &Op : NI->operands())
BrI.addOperand(Op);
NI->eraseFromParent();
Rewritten = true;
}
}
// Do not erase instructions. A newly created instruction could get
// the same address as an instruction marked as executable during the
// propagation.
if (!Rewritten)
replaceWithNop(BrI);
return true;
}
FunctionPass *llvm::createHexagonConstPropagationPass() {
return new HexagonConstPropagation();
}