RPCSX/llvm
1
0
Fork 0
mirror of https://github.com/RPCSX/llvm.git synced 2024-12-05 02:16:46 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity
77bb5f999d
llvm/lib/Target/Hexagon/HexagonSplitDouble.cpp
Duncan P. N. Exon Smith 567409db69 CodeGen: Use MachineInstr& in TargetInstrInfo, NFC 
This is mostly a mechanical change to make TargetInstrInfo API take
MachineInstr& (instead of MachineInstr* or MachineBasicBlock::iterator)
when the argument is expected to be a valid MachineInstr.  This is a
general API improvement.

Although it would be possible to do this one function at a time, that
would demand a quadratic amount of churn since many of these functions
call each other.  Instead I've done everything as a block and just
updated what was necessary.

This is mostly mechanical fixes: adding and removing `*` and `&`
operators.  The only non-mechanical change is to split
ARMBaseInstrInfo::getOperandLatencyImpl out from
ARMBaseInstrInfo::getOperandLatency.  Previously, the latter took a
`MachineInstr*` which it updated to the instruction bundle leader; now,
the latter calls the former either with the same `MachineInstr&` or the
bundle leader.

As a side effect, this removes a bunch of MachineInstr* to
MachineBasicBlock::iterator implicit conversions, a necessary step
toward fixing PR26753.

Note: I updated WebAssembly, Lanai, and AVR (despite being
off-by-default) since it turned out to be easy.  I couldn't run tests
for AVR since llc doesn't link with it turned on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@274189 91177308-0d34-0410-b5e6-96231b3b80d8
2016-06-30 00:01:54 +00:00

1200 lines
36 KiB
C++
Raw Blame History

//===--- HexagonSplitDouble.cpp -------------------------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "hsdr"
#include "HexagonRegisterInfo.h"
#include "HexagonTargetMachine.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Pass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include <map>
#include <set>
#include <vector>
using namespace llvm;
namespace llvm {
FunctionPass *createHexagonSplitDoubleRegs();
void initializeHexagonSplitDoubleRegsPass(PassRegistry&);
}
namespace {
static cl::opt<int> MaxHSDR("max-hsdr", cl::Hidden, cl::init(-1),
cl::desc("Maximum number of split partitions"));
static cl::opt<bool> MemRefsFixed("hsdr-no-mem", cl::Hidden, cl::init(true),
cl::desc("Do not split loads or stores"));
class HexagonSplitDoubleRegs : public MachineFunctionPass {
public:
static char ID;
HexagonSplitDoubleRegs() : MachineFunctionPass(ID), TRI(nullptr),
TII(nullptr) {
initializeHexagonSplitDoubleRegsPass(*PassRegistry::getPassRegistry());
}
const char *getPassName() const override {
return "Hexagon Split Double Registers";
}
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<MachineLoopInfo>();
AU.addPreserved<MachineLoopInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}
bool runOnMachineFunction(MachineFunction &MF) override;
private:
static const TargetRegisterClass *const DoubleRC;
const HexagonRegisterInfo *TRI;
const HexagonInstrInfo *TII;
const MachineLoopInfo *MLI;
MachineRegisterInfo *MRI;
typedef std::set<unsigned> USet;
typedef std::map<unsigned,USet> UUSetMap;
typedef std::pair<unsigned,unsigned> UUPair;
typedef std::map<unsigned,UUPair> UUPairMap;
typedef std::map<const MachineLoop*,USet> LoopRegMap;
bool isInduction(unsigned Reg, LoopRegMap &IRM) const;
bool isVolatileInstr(const MachineInstr *MI) const;
bool isFixedInstr(const MachineInstr *MI) const;
void partitionRegisters(UUSetMap &P2Rs);
int32_t profit(const MachineInstr *MI) const;
bool isProfitable(const USet &Part, LoopRegMap &IRM) const;
void collectIndRegsForLoop(const MachineLoop *L, USet &Rs);
void collectIndRegs(LoopRegMap &IRM);
void createHalfInstr(unsigned Opc, MachineInstr *MI,
const UUPairMap &PairMap, unsigned SubR);
void splitMemRef(MachineInstr *MI, const UUPairMap &PairMap);
void splitImmediate(MachineInstr *MI, const UUPairMap &PairMap);
void splitCombine(MachineInstr *MI, const UUPairMap &PairMap);
void splitExt(MachineInstr *MI, const UUPairMap &PairMap);
void splitShift(MachineInstr *MI, const UUPairMap &PairMap);
void splitAslOr(MachineInstr *MI, const UUPairMap &PairMap);
bool splitInstr(MachineInstr *MI, const UUPairMap &PairMap);
void replaceSubregUses(MachineInstr *MI, const UUPairMap &PairMap);
void collapseRegPairs(MachineInstr *MI, const UUPairMap &PairMap);
bool splitPartition(const USet &Part);
static int Counter;
static void dump_partition(raw_ostream&, const USet&,
const TargetRegisterInfo&);
};
char HexagonSplitDoubleRegs::ID;
int HexagonSplitDoubleRegs::Counter = 0;
const TargetRegisterClass *const HexagonSplitDoubleRegs::DoubleRC
= &Hexagon::DoubleRegsRegClass;
}
INITIALIZE_PASS(HexagonSplitDoubleRegs, "hexagon-split-double",
"Hexagon Split Double Registers", false, false)
void HexagonSplitDoubleRegs::dump_partition(raw_ostream &os,
const USet &Part, const TargetRegisterInfo &TRI) {
dbgs() << '{';
for (auto I : Part)
dbgs() << ' ' << PrintReg(I, &TRI);
dbgs() << " }";
}
bool HexagonSplitDoubleRegs::isInduction(unsigned Reg, LoopRegMap &IRM) const {
for (auto I : IRM) {
const USet &Rs = I.second;
if (Rs.find(Reg) != Rs.end())
return true;
}
return false;
}
bool HexagonSplitDoubleRegs::isVolatileInstr(const MachineInstr *MI) const {
for (auto &I : MI->memoperands())
if (I->isVolatile())
return true;
return false;
}
bool HexagonSplitDoubleRegs::isFixedInstr(const MachineInstr *MI) const {
if (MI->mayLoad() || MI->mayStore())
if (MemRefsFixed || isVolatileInstr(MI))
return true;
if (MI->isDebugValue())
return false;
unsigned Opc = MI->getOpcode();
switch (Opc) {
default:
return true;
case TargetOpcode::PHI:
case TargetOpcode::COPY:
break;
case Hexagon::L2_loadrd_io:
// Not handling stack stores (only reg-based addresses).
if (MI->getOperand(1).isReg())
break;
return true;
case Hexagon::S2_storerd_io:
// Not handling stack stores (only reg-based addresses).
if (MI->getOperand(0).isReg())
break;
return true;
case Hexagon::L2_loadrd_pi:
case Hexagon::S2_storerd_pi:
case Hexagon::A2_tfrpi:
case Hexagon::A2_combineii:
case Hexagon::A4_combineir:
case Hexagon::A4_combineii:
case Hexagon::A4_combineri:
case Hexagon::A2_combinew:
case Hexagon::CONST64_Int_Real:
case Hexagon::A2_sxtw:
case Hexagon::A2_andp:
case Hexagon::A2_orp:
case Hexagon::A2_xorp:
case Hexagon::S2_asl_i_p_or:
case Hexagon::S2_asl_i_p:
case Hexagon::S2_asr_i_p:
case Hexagon::S2_lsr_i_p:
break;
}
for (auto &Op : MI->operands()) {
if (!Op.isReg())
continue;
unsigned R = Op.getReg();
if (!TargetRegisterInfo::isVirtualRegister(R))
return true;
}
return false;
}
void HexagonSplitDoubleRegs::partitionRegisters(UUSetMap &P2Rs) {
typedef std::map<unsigned,unsigned> UUMap;
typedef std::vector<unsigned> UVect;
unsigned NumRegs = MRI->getNumVirtRegs();
BitVector DoubleRegs(NumRegs);
for (unsigned i = 0; i < NumRegs; ++i) {
unsigned R = TargetRegisterInfo::index2VirtReg(i);
if (MRI->getRegClass(R) == DoubleRC)
DoubleRegs.set(i);
}
BitVector FixedRegs(NumRegs);
for (int x = DoubleRegs.find_first(); x >= 0; x = DoubleRegs.find_next(x)) {
unsigned R = TargetRegisterInfo::index2VirtReg(x);
MachineInstr *DefI = MRI->getVRegDef(R);
// In some cases a register may exist, but never be defined or used.
// It should never appear anywhere, but mark it as "fixed", just to be
// safe.
if (!DefI || isFixedInstr(DefI))
FixedRegs.set(x);
}
UUSetMap AssocMap;
for (int x = DoubleRegs.find_first(); x >= 0; x = DoubleRegs.find_next(x)) {
if (FixedRegs[x])
continue;
unsigned R = TargetRegisterInfo::index2VirtReg(x);
DEBUG(dbgs() << PrintReg(R, TRI) << " ~~");
USet &Asc = AssocMap[R];
for (auto U = MRI->use_nodbg_begin(R), Z = MRI->use_nodbg_end();
U != Z; ++U) {
MachineOperand &Op = *U;
MachineInstr *UseI = Op.getParent();
if (isFixedInstr(UseI))
continue;
for (unsigned i = 0, n = UseI->getNumOperands(); i < n; ++i) {
MachineOperand &MO = UseI->getOperand(i);
// Skip non-registers or registers with subregisters.
if (&MO == &Op || !MO.isReg() || MO.getSubReg())
continue;
unsigned T = MO.getReg();
if (!TargetRegisterInfo::isVirtualRegister(T)) {
FixedRegs.set(x);
continue;
}
if (MRI->getRegClass(T) != DoubleRC)
continue;
unsigned u = TargetRegisterInfo::virtReg2Index(T);
if (FixedRegs[u])
continue;
DEBUG(dbgs() << ' ' << PrintReg(T, TRI));
Asc.insert(T);
// Make it symmetric.
AssocMap[T].insert(R);
}
}
DEBUG(dbgs() << '\n');
}
UUMap R2P;
unsigned NextP = 1;
USet Visited;
for (int x = DoubleRegs.find_first(); x >= 0; x = DoubleRegs.find_next(x)) {
unsigned R = TargetRegisterInfo::index2VirtReg(x);
if (Visited.count(R))
continue;
// Create a new partition for R.
unsigned ThisP = FixedRegs[x] ? 0 : NextP++;
UVect WorkQ;
WorkQ.push_back(R);
for (unsigned i = 0; i < WorkQ.size(); ++i) {
unsigned T = WorkQ[i];
if (Visited.count(T))
continue;
R2P[T] = ThisP;
Visited.insert(T);
// Add all registers associated with T.
USet &Asc = AssocMap[T];
for (USet::iterator J = Asc.begin(), F = Asc.end(); J != F; ++J)
WorkQ.push_back(*J);
}
}
for (auto I : R2P)
P2Rs[I.second].insert(I.first);
}
static inline int32_t profitImm(unsigned Lo, unsigned Hi) {
int32_t P = 0;
bool LoZ1 = false, HiZ1 = false;
if (Lo == 0 || Lo == 0xFFFFFFFF)
P += 10, LoZ1 = true;
if (Hi == 0 || Hi == 0xFFFFFFFF)
P += 10, HiZ1 = true;
if (!LoZ1 && !HiZ1 && Lo == Hi)
P += 3;
return P;
}
int32_t HexagonSplitDoubleRegs::profit(const MachineInstr *MI) const {
unsigned ImmX = 0;
unsigned Opc = MI->getOpcode();
switch (Opc) {
case TargetOpcode::PHI:
for (const auto &Op : MI->operands())
if (!Op.getSubReg())
return 0;
return 10;
case TargetOpcode::COPY:
if (MI->getOperand(1).getSubReg() != 0)
return 10;
return 0;
case Hexagon::L2_loadrd_io:
case Hexagon::S2_storerd_io:
return -1;
case Hexagon::L2_loadrd_pi:
case Hexagon::S2_storerd_pi:
return 2;
case Hexagon::A2_tfrpi:
case Hexagon::CONST64_Int_Real: {
uint64_t D = MI->getOperand(1).getImm();
unsigned Lo = D & 0xFFFFFFFFULL;
unsigned Hi = D >> 32;
return profitImm(Lo, Hi);
}
case Hexagon::A2_combineii:
case Hexagon::A4_combineii:
return profitImm(MI->getOperand(1).getImm(),
MI->getOperand(2).getImm());
case Hexagon::A4_combineri:
ImmX++;
case Hexagon::A4_combineir: {
ImmX++;
int64_t V = MI->getOperand(ImmX).getImm();
if (V == 0 || V == -1)
return 10;
// Fall through into A2_combinew.
}
case Hexagon::A2_combinew:
return 2;
case Hexagon::A2_sxtw:
return 3;
case Hexagon::A2_andp:
case Hexagon::A2_orp:
case Hexagon::A2_xorp:
return 1;
case Hexagon::S2_asl_i_p_or: {
unsigned S = MI->getOperand(3).getImm();
if (S == 0 || S == 32)
return 10;
return -1;
}
case Hexagon::S2_asl_i_p:
case Hexagon::S2_asr_i_p:
case Hexagon::S2_lsr_i_p:
unsigned S = MI->getOperand(2).getImm();
if (S == 0 || S == 32)
return 10;
if (S == 16)
return 5;
if (S == 48)
return 7;
return -10;
}
return 0;
}
bool HexagonSplitDoubleRegs::isProfitable(const USet &Part, LoopRegMap &IRM)
const {
unsigned FixedNum = 0, SplitNum = 0, LoopPhiNum = 0;
int32_t TotalP = 0;
for (unsigned DR : Part) {
MachineInstr *DefI = MRI->getVRegDef(DR);
int32_t P = profit(DefI);
if (P == INT_MIN)
return false;
TotalP += P;
// Reduce the profitability of splitting induction registers.
if (isInduction(DR, IRM))
TotalP -= 30;
for (auto U = MRI->use_nodbg_begin(DR), W = MRI->use_nodbg_end();
U != W; ++U) {
MachineInstr *UseI = U->getParent();
if (isFixedInstr(UseI)) {
FixedNum++;
// Calculate the cost of generating REG_SEQUENCE instructions.
for (auto &Op : UseI->operands()) {
if (Op.isReg() && Part.count(Op.getReg()))
if (Op.getSubReg())
TotalP -= 2;
}
continue;
}
// If a register from this partition is used in a fixed instruction,
// and there is also a register in this partition that is used in
// a loop phi node, then decrease the splitting profit as this can
// confuse the modulo scheduler.
if (UseI->isPHI()) {
const MachineBasicBlock *PB = UseI->getParent();
const MachineLoop *L = MLI->getLoopFor(PB);
if (L && L->getHeader() == PB)
LoopPhiNum++;
}
// Splittable instruction.
SplitNum++;
int32_t P = profit(UseI);
if (P == INT_MIN)
return false;
TotalP += P;
}
}
if (FixedNum > 0 && LoopPhiNum > 0)
TotalP -= 20*LoopPhiNum;
DEBUG(dbgs() << "Partition profit: " << TotalP << '\n');
return TotalP > 0;
}
void HexagonSplitDoubleRegs::collectIndRegsForLoop(const MachineLoop *L,
USet &Rs) {
const MachineBasicBlock *HB = L->getHeader();
const MachineBasicBlock *LB = L->getLoopLatch();
if (!HB || !LB)
return;
// Examine the latch branch. Expect it to be a conditional branch to
// the header (either "br-cond header" or "br-cond exit; br header").
MachineBasicBlock *TB = 0, *FB = 0;
MachineBasicBlock *TmpLB = const_cast<MachineBasicBlock*>(LB);
SmallVector<MachineOperand,2> Cond;
bool BadLB = TII->AnalyzeBranch(*TmpLB, TB, FB, Cond, false);
// Only analyzable conditional branches. HII::AnalyzeBranch will put
// the branch opcode as the first element of Cond, and the predicate
// operand as the second.
if (BadLB || Cond.size() != 2)
return;
// Only simple jump-conditional (with or without negation).
if (!TII->PredOpcodeHasJMP_c(Cond[0].getImm()))
return;
// Must go to the header.
if (TB != HB && FB != HB)
return;
assert(Cond[1].isReg() && "Unexpected Cond vector from AnalyzeBranch");
// Expect a predicate register.
unsigned PR = Cond[1].getReg();
assert(MRI->getRegClass(PR) == &Hexagon::PredRegsRegClass);
// Get the registers on which the loop controlling compare instruction
// depends.
unsigned CmpR1 = 0, CmpR2 = 0;
const MachineInstr *CmpI = MRI->getVRegDef(PR);
while (CmpI->getOpcode() == Hexagon::C2_not)
CmpI = MRI->getVRegDef(CmpI->getOperand(1).getReg());
int Mask = 0, Val = 0;
bool OkCI = TII->analyzeCompare(*CmpI, CmpR1, CmpR2, Mask, Val);
if (!OkCI)
return;
// Eliminate non-double input registers.
if (CmpR1 && MRI->getRegClass(CmpR1) != DoubleRC)
CmpR1 = 0;
if (CmpR2 && MRI->getRegClass(CmpR2) != DoubleRC)
CmpR2 = 0;
if (!CmpR1 && !CmpR2)
return;
// Now examine the top of the loop: the phi nodes that could poten-
// tially define loop induction registers. The registers defined by
// such a phi node would be used in a 64-bit add, which then would
// be used in the loop compare instruction.
// Get the set of all double registers defined by phi nodes in the
// loop header.
typedef std::vector<unsigned> UVect;
UVect DP;
for (auto &MI : *HB) {
if (!MI.isPHI())
break;
const MachineOperand &MD = MI.getOperand(0);
unsigned R = MD.getReg();
if (MRI->getRegClass(R) == DoubleRC)
DP.push_back(R);
}
if (DP.empty())
return;
auto NoIndOp = [this, CmpR1, CmpR2] (unsigned R) -> bool {
for (auto I = MRI->use_nodbg_begin(R), E = MRI->use_nodbg_end();
I != E; ++I) {
const MachineInstr *UseI = I->getParent();
if (UseI->getOpcode() != Hexagon::A2_addp)
continue;
// Get the output from the add. If it is one of the inputs to the
// loop-controlling compare instruction, then R is likely an induc-
// tion register.
unsigned T = UseI->getOperand(0).getReg();
if (T == CmpR1 || T == CmpR2)
return false;
}
return true;
};
UVect::iterator End = std::remove_if(DP.begin(), DP.end(), NoIndOp);
Rs.insert(DP.begin(), End);
Rs.insert(CmpR1);
Rs.insert(CmpR2);
DEBUG({
dbgs() << "For loop at BB#" << HB->getNumber() << " ind regs: ";
dump_partition(dbgs(), Rs, *TRI);
dbgs() << '\n';
});
}
void HexagonSplitDoubleRegs::collectIndRegs(LoopRegMap &IRM) {
typedef std::vector<MachineLoop*> LoopVector;
LoopVector WorkQ;
for (auto I : *MLI)
WorkQ.push_back(I);
for (unsigned i = 0; i < WorkQ.size(); ++i) {
for (auto I : *WorkQ[i])
WorkQ.push_back(I);
}
USet Rs;
for (unsigned i = 0, n = WorkQ.size(); i < n; ++i) {
MachineLoop *L = WorkQ[i];
Rs.clear();
collectIndRegsForLoop(L, Rs);
if (!Rs.empty())
IRM.insert(std::make_pair(L, Rs));
}
}
void HexagonSplitDoubleRegs::createHalfInstr(unsigned Opc, MachineInstr *MI,
const UUPairMap &PairMap, unsigned SubR) {
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
MachineInstr *NewI = BuildMI(B, MI, DL, TII->get(Opc));
for (auto &Op : MI->operands()) {
if (!Op.isReg()) {
NewI->addOperand(Op);
continue;
}
// For register operands, set the subregister.
unsigned R = Op.getReg();
unsigned SR = Op.getSubReg();
bool isVirtReg = TargetRegisterInfo::isVirtualRegister(R);
bool isKill = Op.isKill();
if (isVirtReg && MRI->getRegClass(R) == DoubleRC) {
isKill = false;
UUPairMap::const_iterator F = PairMap.find(R);
if (F == PairMap.end()) {
SR = SubR;
} else {
const UUPair &P = F->second;
R = (SubR == Hexagon::subreg_loreg) ? P.first : P.second;
SR = 0;
}
}
auto CO = MachineOperand::CreateReg(R, Op.isDef(), Op.isImplicit(), isKill,
Op.isDead(), Op.isUndef(), Op.isEarlyClobber(), SR, Op.isDebug(),
Op.isInternalRead());
NewI->addOperand(CO);
}
}
void HexagonSplitDoubleRegs::splitMemRef(MachineInstr *MI,
const UUPairMap &PairMap) {
bool Load = MI->mayLoad();
unsigned OrigOpc = MI->getOpcode();
bool PostInc = (OrigOpc == Hexagon::L2_loadrd_pi ||
OrigOpc == Hexagon::S2_storerd_pi);
MachineInstr *LowI, *HighI;
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
// Index of the base-address-register operand.
unsigned AdrX = PostInc ? (Load ? 2 : 1)
: (Load ? 1 : 0);
MachineOperand &AdrOp = MI->getOperand(AdrX);
unsigned RSA = getRegState(AdrOp);
MachineOperand &ValOp = Load ? MI->getOperand(0)
: (PostInc ? MI->getOperand(3)
: MI->getOperand(2));
UUPairMap::const_iterator F = PairMap.find(ValOp.getReg());
assert(F != PairMap.end());
if (Load) {
const UUPair &P = F->second;
int64_t Off = PostInc ? 0 : MI->getOperand(2).getImm();
LowI = BuildMI(B, MI, DL, TII->get(Hexagon::L2_loadri_io), P.first)
.addReg(AdrOp.getReg(), RSA & ~RegState::Kill, AdrOp.getSubReg())
.addImm(Off);
HighI = BuildMI(B, MI, DL, TII->get(Hexagon::L2_loadri_io), P.second)
.addReg(AdrOp.getReg(), RSA & ~RegState::Kill, AdrOp.getSubReg())
.addImm(Off+4);
} else {
const UUPair &P = F->second;
int64_t Off = PostInc ? 0 : MI->getOperand(1).getImm();
LowI = BuildMI(B, MI, DL, TII->get(Hexagon::S2_storeri_io))
.addReg(AdrOp.getReg(), RSA & ~RegState::Kill, AdrOp.getSubReg())
.addImm(Off)
.addReg(P.first);
HighI = BuildMI(B, MI, DL, TII->get(Hexagon::S2_storeri_io))
.addReg(AdrOp.getReg(), RSA & ~RegState::Kill, AdrOp.getSubReg())
.addImm(Off+4)
.addReg(P.second);
}
if (PostInc) {
// Create the increment of the address register.
int64_t Inc = Load ? MI->getOperand(3).getImm()
: MI->getOperand(2).getImm();
MachineOperand &UpdOp = Load ? MI->getOperand(1) : MI->getOperand(0);
const TargetRegisterClass *RC = MRI->getRegClass(UpdOp.getReg());
unsigned NewR = MRI->createVirtualRegister(RC);
assert(!UpdOp.getSubReg() && "Def operand with subreg");
BuildMI(B, MI, DL, TII->get(Hexagon::A2_addi), NewR)
.addReg(AdrOp.getReg(), RSA)
.addImm(Inc);
MRI->replaceRegWith(UpdOp.getReg(), NewR);
// The original instruction will be deleted later.
}
// Generate a new pair of memory-operands.
MachineFunction &MF = *B.getParent();
for (auto &MO : MI->memoperands()) {
const MachinePointerInfo &Ptr = MO->getPointerInfo();
unsigned F = MO->getFlags();
int A = MO->getAlignment();
auto *Tmp1 = MF.getMachineMemOperand(Ptr, F, 4/*size*/, A);
LowI->addMemOperand(MF, Tmp1);
auto *Tmp2 = MF.getMachineMemOperand(Ptr, F, 4/*size*/, std::min(A, 4));
HighI->addMemOperand(MF, Tmp2);
}
}
void HexagonSplitDoubleRegs::splitImmediate(MachineInstr *MI,
const UUPairMap &PairMap) {
MachineOperand &Op0 = MI->getOperand(0);
MachineOperand &Op1 = MI->getOperand(1);
assert(Op0.isReg() && Op1.isImm());
uint64_t V = Op1.getImm();
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
UUPairMap::const_iterator F = PairMap.find(Op0.getReg());
assert(F != PairMap.end());
const UUPair &P = F->second;
// The operand to A2_tfrsi can only have 32 significant bits. Immediate
// values in MachineOperand are stored as 64-bit integers, and so the
// value -1 may be represented either as 64-bit -1, or 4294967295. Both
// will have the 32 higher bits truncated in the end, but -1 will remain
// as -1, while the latter may appear to be a large unsigned value
// requiring a constant extender. The casting to int32_t will select the
// former representation. (The same reasoning applies to all 32-bit
// values.)
BuildMI(B, MI, DL, TII->get(Hexagon::A2_tfrsi), P.first)
.addImm(int32_t(V & 0xFFFFFFFFULL));
BuildMI(B, MI, DL, TII->get(Hexagon::A2_tfrsi), P.second)
.addImm(int32_t(V >> 32));
}
void HexagonSplitDoubleRegs::splitCombine(MachineInstr *MI,
const UUPairMap &PairMap) {
MachineOperand &Op0 = MI->getOperand(0);
MachineOperand &Op1 = MI->getOperand(1);
MachineOperand &Op2 = MI->getOperand(2);
assert(Op0.isReg());
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
UUPairMap::const_iterator F = PairMap.find(Op0.getReg());
assert(F != PairMap.end());
const UUPair &P = F->second;
if (Op1.isImm()) {
BuildMI(B, MI, DL, TII->get(Hexagon::A2_tfrsi), P.second)
.addImm(Op1.getImm());
} else if (Op1.isReg()) {
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), P.second)
.addReg(Op1.getReg(), getRegState(Op1), Op1.getSubReg());
} else
llvm_unreachable("Unexpected operand");
if (Op2.isImm()) {
BuildMI(B, MI, DL, TII->get(Hexagon::A2_tfrsi), P.first)
.addImm(Op2.getImm());
} else if (Op2.isReg()) {
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), P.first)
.addReg(Op2.getReg(), getRegState(Op2), Op2.getSubReg());
} else
llvm_unreachable("Unexpected operand");
}
void HexagonSplitDoubleRegs::splitExt(MachineInstr *MI,
const UUPairMap &PairMap) {
MachineOperand &Op0 = MI->getOperand(0);
MachineOperand &Op1 = MI->getOperand(1);
assert(Op0.isReg() && Op1.isReg());
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
UUPairMap::const_iterator F = PairMap.find(Op0.getReg());
assert(F != PairMap.end());
const UUPair &P = F->second;
unsigned RS = getRegState(Op1);
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), P.first)
.addReg(Op1.getReg(), RS & ~RegState::Kill, Op1.getSubReg());
BuildMI(B, MI, DL, TII->get(Hexagon::S2_asr_i_r), P.second)
.addReg(Op1.getReg(), RS, Op1.getSubReg())
.addImm(31);
}
void HexagonSplitDoubleRegs::splitShift(MachineInstr *MI,
const UUPairMap &PairMap) {
MachineOperand &Op0 = MI->getOperand(0);
MachineOperand &Op1 = MI->getOperand(1);
MachineOperand &Op2 = MI->getOperand(2);
assert(Op0.isReg() && Op1.isReg() && Op2.isImm());
int64_t Sh64 = Op2.getImm();
assert(Sh64 >= 0 && Sh64 < 64);
unsigned S = Sh64;
UUPairMap::const_iterator F = PairMap.find(Op0.getReg());
assert(F != PairMap.end());
const UUPair &P = F->second;
unsigned LoR = P.first;
unsigned HiR = P.second;
using namespace Hexagon;
unsigned Opc = MI->getOpcode();
bool Right = (Opc == S2_lsr_i_p || Opc == S2_asr_i_p);
bool Left = !Right;
bool Signed = (Opc == S2_asr_i_p);
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
unsigned RS = getRegState(Op1);
unsigned ShiftOpc = Left ? S2_asl_i_r
: (Signed ? S2_asr_i_r : S2_lsr_i_r);
unsigned LoSR = subreg_loreg;
unsigned HiSR = subreg_hireg;
if (S == 0) {
// No shift, subregister copy.
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), LoR)
.addReg(Op1.getReg(), RS & ~RegState::Kill, LoSR);
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), HiR)
.addReg(Op1.getReg(), RS, HiSR);
} else if (S < 32) {
const TargetRegisterClass *IntRC = &IntRegsRegClass;
unsigned TmpR = MRI->createVirtualRegister(IntRC);
// Expansion:
// Shift left: DR = shl R, #s
// LoR = shl R.lo, #s
// TmpR = extractu R.lo, #s, #32-s
// HiR = or (TmpR, asl(R.hi, #s))
// Shift right: DR = shr R, #s
// HiR = shr R.hi, #s
// TmpR = shr R.lo, #s
// LoR = insert TmpR, R.hi, #s, #32-s
// Shift left:
// LoR = shl R.lo, #s
// Shift right:
// TmpR = shr R.lo, #s
// Make a special case for A2_aslh and A2_asrh (they are predicable as
// opposed to S2_asl_i_r/S2_asr_i_r).
if (S == 16 && Left)
BuildMI(B, MI, DL, TII->get(A2_aslh), LoR)
.addReg(Op1.getReg(), RS & ~RegState::Kill, LoSR);
else if (S == 16 && Signed)
BuildMI(B, MI, DL, TII->get(A2_asrh), TmpR)
.addReg(Op1.getReg(), RS & ~RegState::Kill, LoSR);
else
BuildMI(B, MI, DL, TII->get(ShiftOpc), (Left ? LoR : TmpR))
.addReg(Op1.getReg(), RS & ~RegState::Kill, LoSR)
.addImm(S);
if (Left) {
// TmpR = extractu R.lo, #s, #32-s
BuildMI(B, MI, DL, TII->get(S2_extractu), TmpR)
.addReg(Op1.getReg(), RS & ~RegState::Kill, LoSR)
.addImm(S)
.addImm(32-S);
// HiR = or (TmpR, asl(R.hi, #s))
BuildMI(B, MI, DL, TII->get(S2_asl_i_r_or), HiR)
.addReg(TmpR)
.addReg(Op1.getReg(), RS, HiSR)
.addImm(S);
} else {
// HiR = shr R.hi, #s
BuildMI(B, MI, DL, TII->get(ShiftOpc), HiR)
.addReg(Op1.getReg(), RS & ~RegState::Kill, HiSR)
.addImm(S);
// LoR = insert TmpR, R.hi, #s, #32-s
BuildMI(B, MI, DL, TII->get(S2_insert), LoR)
.addReg(TmpR)
.addReg(Op1.getReg(), RS, HiSR)
.addImm(S)
.addImm(32-S);
}
} else if (S == 32) {
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), (Left ? HiR : LoR))
.addReg(Op1.getReg(), RS & ~RegState::Kill, (Left ? LoSR : HiSR));
if (!Signed)
BuildMI(B, MI, DL, TII->get(A2_tfrsi), (Left ? LoR : HiR))
.addImm(0);
else // Must be right shift.
BuildMI(B, MI, DL, TII->get(S2_asr_i_r), HiR)
.addReg(Op1.getReg(), RS, HiSR)
.addImm(31);
} else if (S < 64) {
S -= 32;
if (S == 16 && Left)
BuildMI(B, MI, DL, TII->get(A2_aslh), HiR)
.addReg(Op1.getReg(), RS & ~RegState::Kill, LoSR);
else if (S == 16 && Signed)
BuildMI(B, MI, DL, TII->get(A2_asrh), LoR)
.addReg(Op1.getReg(), RS & ~RegState::Kill, HiSR);
else
BuildMI(B, MI, DL, TII->get(ShiftOpc), (Left ? HiR : LoR))
.addReg(Op1.getReg(), RS & ~RegState::Kill, (Left ? LoSR : HiSR))
.addImm(S);
if (Signed)
BuildMI(B, MI, DL, TII->get(S2_asr_i_r), HiR)
.addReg(Op1.getReg(), RS, HiSR)
.addImm(31);
else
BuildMI(B, MI, DL, TII->get(A2_tfrsi), (Left ? LoR : HiR))
.addImm(0);
}
}
void HexagonSplitDoubleRegs::splitAslOr(MachineInstr *MI,
const UUPairMap &PairMap) {
MachineOperand &Op0 = MI->getOperand(0);
MachineOperand &Op1 = MI->getOperand(1);
MachineOperand &Op2 = MI->getOperand(2);
MachineOperand &Op3 = MI->getOperand(3);
assert(Op0.isReg() && Op1.isReg() && Op2.isReg() && Op3.isImm());
int64_t Sh64 = Op3.getImm();
assert(Sh64 >= 0 && Sh64 < 64);
unsigned S = Sh64;
UUPairMap::const_iterator F = PairMap.find(Op0.getReg());
assert(F != PairMap.end());
const UUPair &P = F->second;
unsigned LoR = P.first;
unsigned HiR = P.second;
using namespace Hexagon;
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
unsigned RS1 = getRegState(Op1);
unsigned RS2 = getRegState(Op2);
const TargetRegisterClass *IntRC = &IntRegsRegClass;
unsigned LoSR = subreg_loreg;
unsigned HiSR = subreg_hireg;
// Op0 = S2_asl_i_p_or Op1, Op2, Op3
// means: Op0 = or (Op1, asl(Op2, Op3))
// Expansion of
// DR = or (R1, asl(R2, #s))
//
// LoR = or (R1.lo, asl(R2.lo, #s))
// Tmp1 = extractu R2.lo, #s, #32-s
// Tmp2 = or R1.hi, Tmp1
// HiR = or (Tmp2, asl(R2.hi, #s))
if (S == 0) {
// DR = or (R1, asl(R2, #0))
// -> or (R1, R2)
// i.e. LoR = or R1.lo, R2.lo
// HiR = or R1.hi, R2.hi
BuildMI(B, MI, DL, TII->get(A2_or), LoR)
.addReg(Op1.getReg(), RS1 & ~RegState::Kill, LoSR)
.addReg(Op2.getReg(), RS2 & ~RegState::Kill, LoSR);
BuildMI(B, MI, DL, TII->get(A2_or), HiR)
.addReg(Op1.getReg(), RS1, HiSR)
.addReg(Op2.getReg(), RS2, HiSR);
} else if (S < 32) {
BuildMI(B, MI, DL, TII->get(S2_asl_i_r_or), LoR)
.addReg(Op1.getReg(), RS1 & ~RegState::Kill, LoSR)
.addReg(Op2.getReg(), RS2 & ~RegState::Kill, LoSR)
.addImm(S);
unsigned TmpR1 = MRI->createVirtualRegister(IntRC);
BuildMI(B, MI, DL, TII->get(S2_extractu), TmpR1)
.addReg(Op2.getReg(), RS2 & ~RegState::Kill, LoSR)
.addImm(S)
.addImm(32-S);
unsigned TmpR2 = MRI->createVirtualRegister(IntRC);
BuildMI(B, MI, DL, TII->get(A2_or), TmpR2)
.addReg(Op1.getReg(), RS1, HiSR)
.addReg(TmpR1);
BuildMI(B, MI, DL, TII->get(S2_asl_i_r_or), HiR)
.addReg(TmpR2)
.addReg(Op2.getReg(), RS2, HiSR)
.addImm(S);
} else if (S == 32) {
// DR = or (R1, asl(R2, #32))
// -> or R1, R2.lo
// LoR = R1.lo
// HiR = or R1.hi, R2.lo
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), LoR)
.addReg(Op1.getReg(), RS1 & ~RegState::Kill, LoSR);
BuildMI(B, MI, DL, TII->get(A2_or), HiR)
.addReg(Op1.getReg(), RS1, HiSR)
.addReg(Op2.getReg(), RS2, LoSR);
} else if (S < 64) {
// DR = or (R1, asl(R2, #s))
//
// LoR = R1:lo
// HiR = or (R1:hi, asl(R2:lo, #s-32))
S -= 32;
BuildMI(B, MI, DL, TII->get(TargetOpcode::COPY), LoR)
.addReg(Op1.getReg(), RS1 & ~RegState::Kill, LoSR);
BuildMI(B, MI, DL, TII->get(S2_asl_i_r_or), HiR)
.addReg(Op1.getReg(), RS1, HiSR)
.addReg(Op2.getReg(), RS2, LoSR)
.addImm(S);
}
}
bool HexagonSplitDoubleRegs::splitInstr(MachineInstr *MI,
const UUPairMap &PairMap) {
DEBUG(dbgs() << "Splitting: " << *MI);
bool Split = false;
unsigned Opc = MI->getOpcode();
using namespace Hexagon;
switch (Opc) {
case TargetOpcode::PHI:
case TargetOpcode::COPY: {
unsigned DstR = MI->getOperand(0).getReg();
if (MRI->getRegClass(DstR) == DoubleRC) {
createHalfInstr(Opc, MI, PairMap, subreg_loreg);
createHalfInstr(Opc, MI, PairMap, subreg_hireg);
Split = true;
}
break;
}
case A2_andp:
createHalfInstr(A2_and, MI, PairMap, subreg_loreg);
createHalfInstr(A2_and, MI, PairMap, subreg_hireg);
Split = true;
break;
case A2_orp:
createHalfInstr(A2_or, MI, PairMap, subreg_loreg);
createHalfInstr(A2_or, MI, PairMap, subreg_hireg);
Split = true;
break;
case A2_xorp:
createHalfInstr(A2_xor, MI, PairMap, subreg_loreg);
createHalfInstr(A2_xor, MI, PairMap, subreg_hireg);
Split = true;
break;
case L2_loadrd_io:
case L2_loadrd_pi:
case S2_storerd_io:
case S2_storerd_pi:
splitMemRef(MI, PairMap);
Split = true;
break;
case A2_tfrpi:
case CONST64_Int_Real:
splitImmediate(MI, PairMap);
Split = true;
break;
case A2_combineii:
case A4_combineir:
case A4_combineii:
case A4_combineri:
case A2_combinew:
splitCombine(MI, PairMap);
Split = true;
break;
case A2_sxtw:
splitExt(MI, PairMap);
Split = true;
break;
case S2_asl_i_p:
case S2_asr_i_p:
case S2_lsr_i_p:
splitShift(MI, PairMap);
Split = true;
break;
case S2_asl_i_p_or:
splitAslOr(MI, PairMap);
Split = true;
break;
default:
llvm_unreachable("Instruction not splitable");
return false;
}
return Split;
}
void HexagonSplitDoubleRegs::replaceSubregUses(MachineInstr *MI,
const UUPairMap &PairMap) {
for (auto &Op : MI->operands()) {
if (!Op.isReg() || !Op.isUse() || !Op.getSubReg())
continue;
unsigned R = Op.getReg();
UUPairMap::const_iterator F = PairMap.find(R);
if (F == PairMap.end())
continue;
const UUPair &P = F->second;
switch (Op.getSubReg()) {
case Hexagon::subreg_loreg:
Op.setReg(P.first);
break;
case Hexagon::subreg_hireg:
Op.setReg(P.second);
break;
}
Op.setSubReg(0);
}
}
void HexagonSplitDoubleRegs::collapseRegPairs(MachineInstr *MI,
const UUPairMap &PairMap) {
MachineBasicBlock &B = *MI->getParent();
DebugLoc DL = MI->getDebugLoc();
for (auto &Op : MI->operands()) {
if (!Op.isReg() || !Op.isUse())
continue;
unsigned R = Op.getReg();
if (!TargetRegisterInfo::isVirtualRegister(R))
continue;
if (MRI->getRegClass(R) != DoubleRC || Op.getSubReg())
continue;
UUPairMap::const_iterator F = PairMap.find(R);
if (F == PairMap.end())
continue;
const UUPair &Pr = F->second;
unsigned NewDR = MRI->createVirtualRegister(DoubleRC);
BuildMI(B, MI, DL, TII->get(TargetOpcode::REG_SEQUENCE), NewDR)
.addReg(Pr.first)
.addImm(Hexagon::subreg_loreg)
.addReg(Pr.second)
.addImm(Hexagon::subreg_hireg);
Op.setReg(NewDR);
}
}
bool HexagonSplitDoubleRegs::splitPartition(const USet &Part) {
const TargetRegisterClass *IntRC = &Hexagon::IntRegsRegClass;
typedef std::set<MachineInstr*> MISet;
bool Changed = false;
DEBUG(dbgs() << "Splitting partition: "; dump_partition(dbgs(), Part, *TRI);
dbgs() << '\n');
UUPairMap PairMap;
MISet SplitIns;
for (unsigned DR : Part) {
MachineInstr *DefI = MRI->getVRegDef(DR);
SplitIns.insert(DefI);
// Collect all instructions, including fixed ones. We won't split them,
// but we need to visit them again to insert the REG_SEQUENCE instructions.
for (auto U = MRI->use_nodbg_begin(DR), W = MRI->use_nodbg_end();
U != W; ++U)
SplitIns.insert(U->getParent());
unsigned LoR = MRI->createVirtualRegister(IntRC);
unsigned HiR = MRI->createVirtualRegister(IntRC);
DEBUG(dbgs() << "Created mapping: " << PrintReg(DR, TRI) << " -> "
<< PrintReg(HiR, TRI) << ':' << PrintReg(LoR, TRI) << '\n');
PairMap.insert(std::make_pair(DR, UUPair(LoR, HiR)));
}
MISet Erase;
for (auto MI : SplitIns) {
if (isFixedInstr(MI)) {
collapseRegPairs(MI, PairMap);
} else {
bool Done = splitInstr(MI, PairMap);
if (Done)
Erase.insert(MI);
Changed |= Done;
}
}
for (unsigned DR : Part) {
// Before erasing "double" instructions, revisit all uses of the double
// registers in this partition, and replace all uses of them with subre-
// gisters, with the corresponding single registers.
MISet Uses;
for (auto U = MRI->use_nodbg_begin(DR), W = MRI->use_nodbg_end();
U != W; ++U)
Uses.insert(U->getParent());
for (auto M : Uses)
replaceSubregUses(M, PairMap);
}
for (auto MI : Erase) {
MachineBasicBlock *B = MI->getParent();
B->erase(MI);
}
return Changed;
}
bool HexagonSplitDoubleRegs::runOnMachineFunction(MachineFunction &MF) {
DEBUG(dbgs() << "Splitting double registers in function: "
<< MF.getName() << '\n');
if (skipFunction(*MF.getFunction()))
return false;
auto &ST = MF.getSubtarget<HexagonSubtarget>();
TRI = ST.getRegisterInfo();
TII = ST.getInstrInfo();
MRI = &MF.getRegInfo();
MLI = &getAnalysis<MachineLoopInfo>();
UUSetMap P2Rs;
LoopRegMap IRM;
collectIndRegs(IRM);
partitionRegisters(P2Rs);
DEBUG({
dbgs() << "Register partitioning: (partition #0 is fixed)\n";
for (UUSetMap::iterator I = P2Rs.begin(), E = P2Rs.end(); I != E; ++I) {
dbgs() << '#' << I->first << " -> ";
dump_partition(dbgs(), I->second, *TRI);
dbgs() << '\n';
}
});
bool Changed = false;
int Limit = MaxHSDR;
for (UUSetMap::iterator I = P2Rs.begin(), E = P2Rs.end(); I != E; ++I) {
if (I->first == 0)
continue;
if (Limit >= 0 && Counter >= Limit)
break;
USet &Part = I->second;
DEBUG(dbgs() << "Calculating profit for partition #" << I->first << '\n');
if (!isProfitable(Part, IRM))
continue;
Counter++;
Changed |= splitPartition(Part);
}
return Changed;
}
FunctionPass *llvm::createHexagonSplitDoubleRegs() {
return new HexagonSplitDoubleRegs();
}
Reference in New Issue View Git Blame Copy Permalink