blob: 9b8970258a2c441f54d3ea00aeca4c8359f17d21 [file] [log] [blame]
//===- HexagonInstrInfo.cpp - Hexagon Instruction Information -------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file contains the Hexagon implementation of the TargetInstrInfo class.
//
//===----------------------------------------------------------------------===//
#include "HexagonInstrInfo.h"
#include "Hexagon.h"
#include "HexagonFrameLowering.h"
#include "HexagonHazardRecognizer.h"
#include "HexagonRegisterInfo.h"
#include "HexagonSubtarget.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/CodeGen/DFAPacketizer.h"
#include "llvm/CodeGen/LivePhysRegs.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineInstrBundle.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/MachineValueType.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCInstrDesc.h"
#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/Support/BranchProbability.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOpcodes.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetSubtargetInfo.h"
#include <cassert>
#include <cctype>
#include <cstdint>
#include <cstring>
#include <iterator>
#include <string>
#include <utility>
using namespace llvm;
#define DEBUG_TYPE "hexagon-instrinfo"
#define GET_INSTRINFO_CTOR_DTOR
#define GET_INSTRMAP_INFO
#include "HexagonDepTimingClasses.h"
#include "HexagonGenDFAPacketizer.inc"
#include "HexagonGenInstrInfo.inc"
cl::opt<bool> ScheduleInlineAsm("hexagon-sched-inline-asm", cl::Hidden,
cl::init(false), cl::desc("Do not consider inline-asm a scheduling/"
"packetization boundary."));
static cl::opt<bool> EnableBranchPrediction("hexagon-enable-branch-prediction",
cl::Hidden, cl::init(true), cl::desc("Enable branch prediction"));
static cl::opt<bool> DisableNVSchedule("disable-hexagon-nv-schedule",
cl::Hidden, cl::ZeroOrMore, cl::init(false),
cl::desc("Disable schedule adjustment for new value stores."));
static cl::opt<bool> EnableTimingClassLatency(
"enable-timing-class-latency", cl::Hidden, cl::init(false),
cl::desc("Enable timing class latency"));
static cl::opt<bool> EnableALUForwarding(
"enable-alu-forwarding", cl::Hidden, cl::init(true),
cl::desc("Enable vec alu forwarding"));
static cl::opt<bool> EnableACCForwarding(
"enable-acc-forwarding", cl::Hidden, cl::init(true),
cl::desc("Enable vec acc forwarding"));
static cl::opt<bool> BranchRelaxAsmLarge("branch-relax-asm-large",
cl::init(true), cl::Hidden, cl::ZeroOrMore, cl::desc("branch relax asm"));
static cl::opt<bool> UseDFAHazardRec("dfa-hazard-rec",
cl::init(true), cl::Hidden, cl::ZeroOrMore,
cl::desc("Use the DFA based hazard recognizer."));
/// Constants for Hexagon instructions.
const int Hexagon_MEMW_OFFSET_MAX = 4095;
const int Hexagon_MEMW_OFFSET_MIN = -4096;
const int Hexagon_MEMD_OFFSET_MAX = 8191;
const int Hexagon_MEMD_OFFSET_MIN = -8192;
const int Hexagon_MEMH_OFFSET_MAX = 2047;
const int Hexagon_MEMH_OFFSET_MIN = -2048;
const int Hexagon_MEMB_OFFSET_MAX = 1023;
const int Hexagon_MEMB_OFFSET_MIN = -1024;
const int Hexagon_ADDI_OFFSET_MAX = 32767;
const int Hexagon_ADDI_OFFSET_MIN = -32768;
// Pin the vtable to this file.
void HexagonInstrInfo::anchor() {}
HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST)
: HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP),
Subtarget(ST) {}
static bool isIntRegForSubInst(unsigned Reg) {
return (Reg >= Hexagon::R0 && Reg <= Hexagon::R7) ||
(Reg >= Hexagon::R16 && Reg <= Hexagon::R23);
}
static bool isDblRegForSubInst(unsigned Reg, const HexagonRegisterInfo &HRI) {
return isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_lo)) &&
isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_hi));
}
/// Calculate number of instructions excluding the debug instructions.
static unsigned nonDbgMICount(MachineBasicBlock::const_instr_iterator MIB,
MachineBasicBlock::const_instr_iterator MIE) {
unsigned Count = 0;
for (; MIB != MIE; ++MIB) {
if (!MIB->isDebugValue())
++Count;
}
return Count;
}
/// Find the hardware loop instruction used to set-up the specified loop.
/// On Hexagon, we have two instructions used to set-up the hardware loop
/// (LOOP0, LOOP1) with corresponding endloop (ENDLOOP0, ENDLOOP1) instructions
/// to indicate the end of a loop.
static MachineInstr *findLoopInstr(MachineBasicBlock *BB, unsigned EndLoopOp,
MachineBasicBlock *TargetBB,
SmallPtrSet<MachineBasicBlock *, 8> &Visited) {
unsigned LOOPi;
unsigned LOOPr;
if (EndLoopOp == Hexagon::ENDLOOP0) {
LOOPi = Hexagon::J2_loop0i;
LOOPr = Hexagon::J2_loop0r;
} else { // EndLoopOp == Hexagon::EndLOOP1
LOOPi = Hexagon::J2_loop1i;
LOOPr = Hexagon::J2_loop1r;
}
// The loop set-up instruction will be in a predecessor block
for (MachineBasicBlock *PB : BB->predecessors()) {
// If this has been visited, already skip it.
if (!Visited.insert(PB).second)
continue;
if (PB == BB)
continue;
for (auto I = PB->instr_rbegin(), E = PB->instr_rend(); I != E; ++I) {
unsigned Opc = I->getOpcode();
if (Opc == LOOPi || Opc == LOOPr)
return &*I;
// We've reached a different loop, which means the loop01 has been
// removed.
if (Opc == EndLoopOp && I->getOperand(0).getMBB() != TargetBB)
return nullptr;
}
// Check the predecessors for the LOOP instruction.
if (MachineInstr *Loop = findLoopInstr(PB, EndLoopOp, TargetBB, Visited))
return Loop;
}
return nullptr;
}
/// Gather register def/uses from MI.
/// This treats possible (predicated) defs as actually happening ones
/// (conservatively).
static inline void parseOperands(const MachineInstr &MI,
SmallVector<unsigned, 4> &Defs, SmallVector<unsigned, 8> &Uses) {
Defs.clear();
Uses.clear();
for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
const MachineOperand &MO = MI.getOperand(i);
if (!MO.isReg())
continue;
unsigned Reg = MO.getReg();
if (!Reg)
continue;
if (MO.isUse())
Uses.push_back(MO.getReg());
if (MO.isDef())
Defs.push_back(MO.getReg());
}
}
// Position dependent, so check twice for swap.
static bool isDuplexPairMatch(unsigned Ga, unsigned Gb) {
switch (Ga) {
case HexagonII::HSIG_None:
default:
return false;
case HexagonII::HSIG_L1:
return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_A);
case HexagonII::HSIG_L2:
return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
Gb == HexagonII::HSIG_A);
case HexagonII::HSIG_S1:
return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_A);
case HexagonII::HSIG_S2:
return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_S2 ||
Gb == HexagonII::HSIG_A);
case HexagonII::HSIG_A:
return (Gb == HexagonII::HSIG_A);
case HexagonII::HSIG_Compound:
return (Gb == HexagonII::HSIG_Compound);
}
return false;
}
/// isLoadFromStackSlot - If the specified machine instruction is a direct
/// load from a stack slot, return the virtual or physical register number of
/// the destination along with the FrameIndex of the loaded stack slot. If
/// not, return 0. This predicate must return 0 if the instruction has
/// any side effects other than loading from the stack slot.
unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
int &FrameIndex) const {
switch (MI.getOpcode()) {
default:
break;
case Hexagon::L2_loadri_io:
case Hexagon::L2_loadrd_io:
case Hexagon::V6_vL32b_ai:
case Hexagon::V6_vL32b_nt_ai:
case Hexagon::V6_vL32Ub_ai:
case Hexagon::LDriw_pred:
case Hexagon::LDriw_mod:
case Hexagon::PS_vloadrq_ai:
case Hexagon::PS_vloadrw_ai:
case Hexagon::PS_vloadrw_nt_ai: {
const MachineOperand OpFI = MI.getOperand(1);
if (!OpFI.isFI())
return 0;
const MachineOperand OpOff = MI.getOperand(2);
if (!OpOff.isImm() || OpOff.getImm() != 0)
return 0;
FrameIndex = OpFI.getIndex();
return MI.getOperand(0).getReg();
}
case Hexagon::L2_ploadrit_io:
case Hexagon::L2_ploadrif_io:
case Hexagon::L2_ploadrdt_io:
case Hexagon::L2_ploadrdf_io: {
const MachineOperand OpFI = MI.getOperand(2);
if (!OpFI.isFI())
return 0;
const MachineOperand OpOff = MI.getOperand(3);
if (!OpOff.isImm() || OpOff.getImm() != 0)
return 0;
FrameIndex = OpFI.getIndex();
return MI.getOperand(0).getReg();
}
}
return 0;
}
/// isStoreToStackSlot - If the specified machine instruction is a direct
/// store to a stack slot, return the virtual or physical register number of
/// the source reg along with the FrameIndex of the loaded stack slot. If
/// not, return 0. This predicate must return 0 if the instruction has
/// any side effects other than storing to the stack slot.
unsigned HexagonInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
int &FrameIndex) const {
switch (MI.getOpcode()) {
default:
break;
case Hexagon::S2_storerb_io:
case Hexagon::S2_storerh_io:
case Hexagon::S2_storeri_io:
case Hexagon::S2_storerd_io:
case Hexagon::V6_vS32b_ai:
case Hexagon::V6_vS32Ub_ai:
case Hexagon::STriw_pred:
case Hexagon::STriw_mod:
case Hexagon::PS_vstorerq_ai:
case Hexagon::PS_vstorerw_ai: {
const MachineOperand &OpFI = MI.getOperand(0);
if (!OpFI.isFI())
return 0;
const MachineOperand &OpOff = MI.getOperand(1);
if (!OpOff.isImm() || OpOff.getImm() != 0)
return 0;
FrameIndex = OpFI.getIndex();
return MI.getOperand(2).getReg();
}
case Hexagon::S2_pstorerbt_io:
case Hexagon::S2_pstorerbf_io:
case Hexagon::S2_pstorerht_io:
case Hexagon::S2_pstorerhf_io:
case Hexagon::S2_pstorerit_io:
case Hexagon::S2_pstorerif_io:
case Hexagon::S2_pstorerdt_io:
case Hexagon::S2_pstorerdf_io: {
const MachineOperand &OpFI = MI.getOperand(1);
if (!OpFI.isFI())
return 0;
const MachineOperand &OpOff = MI.getOperand(2);
if (!OpOff.isImm() || OpOff.getImm() != 0)
return 0;
FrameIndex = OpFI.getIndex();
return MI.getOperand(3).getReg();
}
}
return 0;
}
/// This function can analyze one/two way branching only and should (mostly) be
/// called by target independent side.
/// First entry is always the opcode of the branching instruction, except when
/// the Cond vector is supposed to be empty, e.g., when AnalyzeBranch fails, a
/// BB with only unconditional jump. Subsequent entries depend upon the opcode,
/// e.g. Jump_c p will have
/// Cond[0] = Jump_c
/// Cond[1] = p
/// HW-loop ENDLOOP:
/// Cond[0] = ENDLOOP
/// Cond[1] = MBB
/// New value jump:
/// Cond[0] = Hexagon::CMPEQri_f_Jumpnv_t_V4 -- specific opcode
/// Cond[1] = R
/// Cond[2] = Imm
bool HexagonInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
MachineBasicBlock *&TBB,
MachineBasicBlock *&FBB,
SmallVectorImpl<MachineOperand> &Cond,
bool AllowModify) const {
TBB = nullptr;
FBB = nullptr;
Cond.clear();
// If the block has no terminators, it just falls into the block after it.
MachineBasicBlock::instr_iterator I = MBB.instr_end();
if (I == MBB.instr_begin())
return false;
// A basic block may looks like this:
//
// [ insn
// EH_LABEL
// insn
// insn
// insn
// EH_LABEL
// insn ]
//
// It has two succs but does not have a terminator
// Don't know how to handle it.
do {
--I;
if (I->isEHLabel())
// Don't analyze EH branches.
return true;
} while (I != MBB.instr_begin());
I = MBB.instr_end();
--I;
while (I->isDebugValue()) {
if (I == MBB.instr_begin())
return false;
--I;
}
bool JumpToBlock = I->getOpcode() == Hexagon::J2_jump &&
I->getOperand(0).isMBB();
// Delete the J2_jump if it's equivalent to a fall-through.
if (AllowModify && JumpToBlock &&
MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
DEBUG(dbgs() << "\nErasing the jump to successor block\n";);
I->eraseFromParent();
I = MBB.instr_end();
if (I == MBB.instr_begin())
return false;
--I;
}
if (!isUnpredicatedTerminator(*I))
return false;
// Get the last instruction in the block.
MachineInstr *LastInst = &*I;
MachineInstr *SecondLastInst = nullptr;
// Find one more terminator if present.
while (true) {
if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) {
if (!SecondLastInst)
SecondLastInst = &*I;
else
// This is a third branch.
return true;
}
if (I == MBB.instr_begin())
break;
--I;
}
int LastOpcode = LastInst->getOpcode();
int SecLastOpcode = SecondLastInst ? SecondLastInst->getOpcode() : 0;
// If the branch target is not a basic block, it could be a tail call.
// (It is, if the target is a function.)
if (LastOpcode == Hexagon::J2_jump && !LastInst->getOperand(0).isMBB())
return true;
if (SecLastOpcode == Hexagon::J2_jump &&
!SecondLastInst->getOperand(0).isMBB())
return true;
bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode);
bool LastOpcodeHasNVJump = isNewValueJump(*LastInst);
if (LastOpcodeHasJMP_c && !LastInst->getOperand(1).isMBB())
return true;
// If there is only one terminator instruction, process it.
if (LastInst && !SecondLastInst) {
if (LastOpcode == Hexagon::J2_jump) {
TBB = LastInst->getOperand(0).getMBB();
return false;
}
if (isEndLoopN(LastOpcode)) {
TBB = LastInst->getOperand(0).getMBB();
Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
Cond.push_back(LastInst->getOperand(0));
return false;
}
if (LastOpcodeHasJMP_c) {
TBB = LastInst->getOperand(1).getMBB();
Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
Cond.push_back(LastInst->getOperand(0));
return false;
}
// Only supporting rr/ri versions of new-value jumps.
if (LastOpcodeHasNVJump && (LastInst->getNumExplicitOperands() == 3)) {
TBB = LastInst->getOperand(2).getMBB();
Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
Cond.push_back(LastInst->getOperand(0));
Cond.push_back(LastInst->getOperand(1));
return false;
}
DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber()
<< " with one jump\n";);
// Otherwise, don't know what this is.
return true;
}
bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode);
bool SecLastOpcodeHasNVJump = isNewValueJump(*SecondLastInst);
if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::J2_jump)) {
if (!SecondLastInst->getOperand(1).isMBB())
return true;
TBB = SecondLastInst->getOperand(1).getMBB();
Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
Cond.push_back(SecondLastInst->getOperand(0));
FBB = LastInst->getOperand(0).getMBB();
return false;
}
// Only supporting rr/ri versions of new-value jumps.
if (SecLastOpcodeHasNVJump &&
(SecondLastInst->getNumExplicitOperands() == 3) &&
(LastOpcode == Hexagon::J2_jump)) {
TBB = SecondLastInst->getOperand(2).getMBB();
Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
Cond.push_back(SecondLastInst->getOperand(0));
Cond.push_back(SecondLastInst->getOperand(1));
FBB = LastInst->getOperand(0).getMBB();
return false;
}
// If the block ends with two Hexagon:JMPs, handle it. The second one is not
// executed, so remove it.
if (SecLastOpcode == Hexagon::J2_jump && LastOpcode == Hexagon::J2_jump) {
TBB = SecondLastInst->getOperand(0).getMBB();
I = LastInst->getIterator();
if (AllowModify)
I->eraseFromParent();
return false;
}
// If the block ends with an ENDLOOP, and J2_jump, handle it.
if (isEndLoopN(SecLastOpcode) && LastOpcode == Hexagon::J2_jump) {
TBB = SecondLastInst->getOperand(0).getMBB();
Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
Cond.push_back(SecondLastInst->getOperand(0));
FBB = LastInst->getOperand(0).getMBB();
return false;
}
DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber()
<< " with two jumps";);
// Otherwise, can't handle this.
return true;
}
unsigned HexagonInstrInfo::removeBranch(MachineBasicBlock &MBB,
int *BytesRemoved) const {
assert(!BytesRemoved && "code size not handled");
DEBUG(dbgs() << "\nRemoving branches out of BB#" << MBB.getNumber());
MachineBasicBlock::iterator I = MBB.end();
unsigned Count = 0;
while (I != MBB.begin()) {
--I;
if (I->isDebugValue())
continue;
// Only removing branches from end of MBB.
if (!I->isBranch())
return Count;
if (Count && (I->getOpcode() == Hexagon::J2_jump))
llvm_unreachable("Malformed basic block: unconditional branch not last");
MBB.erase(&MBB.back());
I = MBB.end();
++Count;
}
return Count;
}
unsigned HexagonInstrInfo::insertBranch(MachineBasicBlock &MBB,
MachineBasicBlock *TBB,
MachineBasicBlock *FBB,
ArrayRef<MachineOperand> Cond,
const DebugLoc &DL,
int *BytesAdded) const {
unsigned BOpc = Hexagon::J2_jump;
unsigned BccOpc = Hexagon::J2_jumpt;
assert(validateBranchCond(Cond) && "Invalid branching condition");
assert(TBB && "insertBranch must not be told to insert a fallthrough");
assert(!BytesAdded && "code size not handled");
// Check if reverseBranchCondition has asked to reverse this branch
// If we want to reverse the branch an odd number of times, we want
// J2_jumpf.
if (!Cond.empty() && Cond[0].isImm())
BccOpc = Cond[0].getImm();
if (!FBB) {
if (Cond.empty()) {
// Due to a bug in TailMerging/CFG Optimization, we need to add a
// special case handling of a predicated jump followed by an
// unconditional jump. If not, Tail Merging and CFG Optimization go
// into an infinite loop.
MachineBasicBlock *NewTBB, *NewFBB;
SmallVector<MachineOperand, 4> Cond;
auto Term = MBB.getFirstTerminator();
if (Term != MBB.end() && isPredicated(*Term) &&
!analyzeBranch(MBB, NewTBB, NewFBB, Cond, false) &&
MachineFunction::iterator(NewTBB) == ++MBB.getIterator()) {
reverseBranchCondition(Cond);
removeBranch(MBB);
return insertBranch(MBB, TBB, nullptr, Cond, DL);
}
BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
} else if (isEndLoopN(Cond[0].getImm())) {
int EndLoopOp = Cond[0].getImm();
assert(Cond[1].isMBB());
// Since we're adding an ENDLOOP, there better be a LOOP instruction.
// Check for it, and change the BB target if needed.
SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(),
VisitedBBs);
assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP");
Loop->getOperand(0).setMBB(TBB);
// Add the ENDLOOP after the finding the LOOP0.
BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB);
} else if (isNewValueJump(Cond[0].getImm())) {
assert((Cond.size() == 3) && "Only supporting rr/ri version of nvjump");
// New value jump
// (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset)
// (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset)
unsigned Flags1 = getUndefRegState(Cond[1].isUndef());
DEBUG(dbgs() << "\nInserting NVJump for BB#" << MBB.getNumber(););
if (Cond[2].isReg()) {
unsigned Flags2 = getUndefRegState(Cond[2].isUndef());
BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1).
addReg(Cond[2].getReg(), Flags2).addMBB(TBB);
} else if(Cond[2].isImm()) {
BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1).
addImm(Cond[2].getImm()).addMBB(TBB);
} else
llvm_unreachable("Invalid condition for branching");
} else {
assert((Cond.size() == 2) && "Malformed cond vector");
const MachineOperand &RO = Cond[1];
unsigned Flags = getUndefRegState(RO.isUndef());
BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB);
}
return 1;
}
assert((!Cond.empty()) &&
"Cond. cannot be empty when multiple branchings are required");
assert((!isNewValueJump(Cond[0].getImm())) &&
"NV-jump cannot be inserted with another branch");
// Special case for hardware loops. The condition is a basic block.
if (isEndLoopN(Cond[0].getImm())) {
int EndLoopOp = Cond[0].getImm();
assert(Cond[1].isMBB());
// Since we're adding an ENDLOOP, there better be a LOOP instruction.
// Check for it, and change the BB target if needed.
SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(),
VisitedBBs);
assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP");
Loop->getOperand(0).setMBB(TBB);
// Add the ENDLOOP after the finding the LOOP0.
BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB);
} else {
const MachineOperand &RO = Cond[1];
unsigned Flags = getUndefRegState(RO.isUndef());
BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB);
}
BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
return 2;
}
/// Analyze the loop code to find the loop induction variable and compare used
/// to compute the number of iterations. Currently, we analyze loop that are
/// controlled using hardware loops. In this case, the induction variable
/// instruction is null. For all other cases, this function returns true, which
/// means we're unable to analyze it.
bool HexagonInstrInfo::analyzeLoop(MachineLoop &L,
MachineInstr *&IndVarInst,
MachineInstr *&CmpInst) const {
MachineBasicBlock *LoopEnd = L.getBottomBlock();
MachineBasicBlock::iterator I = LoopEnd->getFirstTerminator();
// We really "analyze" only hardware loops right now.
if (I != LoopEnd->end() && isEndLoopN(I->getOpcode())) {
IndVarInst = nullptr;
CmpInst = &*I;
return false;
}
return true;
}
/// Generate code to reduce the loop iteration by one and check if the loop is
/// finished. Return the value/register of the new loop count. this function
/// assumes the nth iteration is peeled first.
unsigned HexagonInstrInfo::reduceLoopCount(MachineBasicBlock &MBB,
MachineInstr *IndVar, MachineInstr &Cmp,
SmallVectorImpl<MachineOperand> &Cond,
SmallVectorImpl<MachineInstr *> &PrevInsts,
unsigned Iter, unsigned MaxIter) const {
// We expect a hardware loop currently. This means that IndVar is set
// to null, and the compare is the ENDLOOP instruction.
assert((!IndVar) && isEndLoopN(Cmp.getOpcode())
&& "Expecting a hardware loop");
MachineFunction *MF = MBB.getParent();
DebugLoc DL = Cmp.getDebugLoc();
SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
MachineInstr *Loop = findLoopInstr(&MBB, Cmp.getOpcode(),
Cmp.getOperand(0).getMBB(), VisitedBBs);
if (!Loop)
return 0;
// If the loop trip count is a compile-time value, then just change the
// value.
if (Loop->getOpcode() == Hexagon::J2_loop0i ||
Loop->getOpcode() == Hexagon::J2_loop1i) {
int64_t Offset = Loop->getOperand(1).getImm();
if (Offset <= 1)
Loop->eraseFromParent();
else
Loop->getOperand(1).setImm(Offset - 1);
return Offset - 1;
}
// The loop trip count is a run-time value. We generate code to subtract
// one from the trip count, and update the loop instruction.
assert(Loop->getOpcode() == Hexagon::J2_loop0r && "Unexpected instruction");
unsigned LoopCount = Loop->getOperand(1).getReg();
// Check if we're done with the loop.
unsigned LoopEnd = createVR(MF, MVT::i1);
MachineInstr *NewCmp = BuildMI(&MBB, DL, get(Hexagon::C2_cmpgtui), LoopEnd).
addReg(LoopCount).addImm(1);
unsigned NewLoopCount = createVR(MF, MVT::i32);
MachineInstr *NewAdd = BuildMI(&MBB, DL, get(Hexagon::A2_addi), NewLoopCount).
addReg(LoopCount).addImm(-1);
const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
// Update the previously generated instructions with the new loop counter.
for (SmallVectorImpl<MachineInstr *>::iterator I = PrevInsts.begin(),
E = PrevInsts.end(); I != E; ++I)
(*I)->substituteRegister(LoopCount, NewLoopCount, 0, HRI);
PrevInsts.clear();
PrevInsts.push_back(NewCmp);
PrevInsts.push_back(NewAdd);
// Insert the new loop instruction if this is the last time the loop is
// decremented.
if (Iter == MaxIter)
BuildMI(&MBB, DL, get(Hexagon::J2_loop0r)).
addMBB(Loop->getOperand(0).getMBB()).addReg(NewLoopCount);
// Delete the old loop instruction.
if (Iter == 0)
Loop->eraseFromParent();
Cond.push_back(MachineOperand::CreateImm(Hexagon::J2_jumpf));
Cond.push_back(NewCmp->getOperand(0));
return NewLoopCount;
}
bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
unsigned NumCycles, unsigned ExtraPredCycles,
BranchProbability Probability) const {
return nonDbgBBSize(&MBB) <= 3;
}
bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
unsigned NumTCycles, unsigned ExtraTCycles, MachineBasicBlock &FMBB,
unsigned NumFCycles, unsigned ExtraFCycles, BranchProbability Probability)
const {
return nonDbgBBSize(&TMBB) <= 3 && nonDbgBBSize(&FMBB) <= 3;
}
bool HexagonInstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
unsigned NumInstrs, BranchProbability Probability) const {
return NumInstrs <= 4;
}
void HexagonInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I,
const DebugLoc &DL, unsigned DestReg,
unsigned SrcReg, bool KillSrc) const {
const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
unsigned KillFlag = getKillRegState(KillSrc);
if (Hexagon::IntRegsRegClass.contains(SrcReg, DestReg)) {
BuildMI(MBB, I, DL, get(Hexagon::A2_tfr), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::DoubleRegsRegClass.contains(SrcReg, DestReg)) {
BuildMI(MBB, I, DL, get(Hexagon::A2_tfrp), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::PredRegsRegClass.contains(SrcReg, DestReg)) {
// Map Pd = Ps to Pd = or(Ps, Ps).
BuildMI(MBB, I, DL, get(Hexagon::C2_or), DestReg)
.addReg(SrcReg).addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::CtrRegsRegClass.contains(DestReg) &&
Hexagon::IntRegsRegClass.contains(SrcReg)) {
BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::IntRegsRegClass.contains(DestReg) &&
Hexagon::CtrRegsRegClass.contains(SrcReg)) {
BuildMI(MBB, I, DL, get(Hexagon::A2_tfrcrr), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::ModRegsRegClass.contains(DestReg) &&
Hexagon::IntRegsRegClass.contains(SrcReg)) {
BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::PredRegsRegClass.contains(SrcReg) &&
Hexagon::IntRegsRegClass.contains(DestReg)) {
BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::IntRegsRegClass.contains(SrcReg) &&
Hexagon::PredRegsRegClass.contains(DestReg)) {
BuildMI(MBB, I, DL, get(Hexagon::C2_tfrrp), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::PredRegsRegClass.contains(SrcReg) &&
Hexagon::IntRegsRegClass.contains(DestReg)) {
BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::HvxVRRegClass.contains(SrcReg, DestReg)) {
BuildMI(MBB, I, DL, get(Hexagon::V6_vassign), DestReg).
addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::HvxWRRegClass.contains(SrcReg, DestReg)) {
unsigned LoSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
unsigned HiSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
BuildMI(MBB, I, DL, get(Hexagon::V6_vcombine), DestReg)
.addReg(HiSrc, KillFlag)
.addReg(LoSrc, KillFlag);
return;
}
if (Hexagon::HvxQRRegClass.contains(SrcReg, DestReg)) {
BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DestReg)
.addReg(SrcReg)
.addReg(SrcReg, KillFlag);
return;
}
if (Hexagon::HvxQRRegClass.contains(SrcReg) &&
Hexagon::HvxVRRegClass.contains(DestReg)) {
llvm_unreachable("Unimplemented pred to vec");
return;
}
if (Hexagon::HvxQRRegClass.contains(DestReg) &&
Hexagon::HvxVRRegClass.contains(SrcReg)) {
llvm_unreachable("Unimplemented vec to pred");
return;
}
#ifndef NDEBUG
// Show the invalid registers to ease debugging.
dbgs() << "Invalid registers for copy in BB#" << MBB.getNumber()
<< ": " << PrintReg(DestReg, &HRI)
<< " = " << PrintReg(SrcReg, &HRI) << '\n';
#endif
llvm_unreachable("Unimplemented");
}
void HexagonInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I, unsigned SrcReg, bool isKill, int FI,
const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const {
DebugLoc DL = MBB.findDebugLoc(I);
MachineFunction &MF = *MBB.getParent();
MachineFrameInfo &MFI = MF.getFrameInfo();
unsigned SlotAlign = MFI.getObjectAlignment(FI);
unsigned RegAlign = TRI->getSpillAlignment(*RC);
unsigned KillFlag = getKillRegState(isKill);
bool HasAlloca = MFI.hasVarSizedObjects();
const HexagonFrameLowering &HFI = *Subtarget.getFrameLowering();
MachineMemOperand *MMO = MF.getMachineMemOperand(
MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
MFI.getObjectSize(FI), SlotAlign);
if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::S2_storeri_io))
.addFrameIndex(FI).addImm(0)
.addReg(SrcReg, KillFlag).addMemOperand(MMO);
} else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::S2_storerd_io))
.addFrameIndex(FI).addImm(0)
.addReg(SrcReg, KillFlag).addMemOperand(MMO);
} else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::STriw_pred))
.addFrameIndex(FI).addImm(0)
.addReg(SrcReg, KillFlag).addMemOperand(MMO);
} else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::STriw_mod))
.addFrameIndex(FI).addImm(0)
.addReg(SrcReg, KillFlag).addMemOperand(MMO);
} else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerq_ai))
.addFrameIndex(FI).addImm(0)
.addReg(SrcReg, KillFlag).addMemOperand(MMO);
} else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) {
// If there are variable-sized objects, spills will not be aligned.
if (HasAlloca)
SlotAlign = HFI.getStackAlignment();
unsigned Opc = SlotAlign < RegAlign ? Hexagon::V6_vS32Ub_ai
: Hexagon::V6_vS32b_ai;
MachineMemOperand *MMOA = MF.getMachineMemOperand(
MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
MFI.getObjectSize(FI), SlotAlign);
BuildMI(MBB, I, DL, get(Opc))
.addFrameIndex(FI).addImm(0)
.addReg(SrcReg, KillFlag).addMemOperand(MMOA);
} else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) {
// If there are variable-sized objects, spills will not be aligned.
if (HasAlloca)
SlotAlign = HFI.getStackAlignment();
unsigned Opc = SlotAlign < RegAlign ? Hexagon::PS_vstorerwu_ai
: Hexagon::PS_vstorerw_ai;
MachineMemOperand *MMOA = MF.getMachineMemOperand(
MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
MFI.getObjectSize(FI), SlotAlign);
BuildMI(MBB, I, DL, get(Opc))
.addFrameIndex(FI).addImm(0)
.addReg(SrcReg, KillFlag).addMemOperand(MMOA);
} else {
llvm_unreachable("Unimplemented");
}
}
void HexagonInstrInfo::loadRegFromStackSlot(
MachineBasicBlock &MBB, MachineBasicBlock::iterator I, unsigned DestReg,
int FI, const TargetRegisterClass *RC,
const TargetRegisterInfo *TRI) const {
DebugLoc DL = MBB.findDebugLoc(I);
MachineFunction &MF = *MBB.getParent();
MachineFrameInfo &MFI = MF.getFrameInfo();
unsigned SlotAlign = MFI.getObjectAlignment(FI);
unsigned RegAlign = TRI->getSpillAlignment(*RC);
bool HasAlloca = MFI.hasVarSizedObjects();
const HexagonFrameLowering &HFI = *Subtarget.getFrameLowering();
MachineMemOperand *MMO = MF.getMachineMemOperand(
MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
MFI.getObjectSize(FI), SlotAlign);
if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::L2_loadri_io), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO);
} else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::L2_loadrd_io), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO);
} else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::LDriw_pred), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO);
} else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::LDriw_mod), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO);
} else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) {
BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrq_ai), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO);
} else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) {
// If there are variable-sized objects, spills will not be aligned.
if (HasAlloca)
SlotAlign = HFI.getStackAlignment();
unsigned Opc = SlotAlign < RegAlign ? Hexagon::V6_vL32Ub_ai
: Hexagon::V6_vL32b_ai;
MachineMemOperand *MMOA = MF.getMachineMemOperand(
MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
MFI.getObjectSize(FI), SlotAlign);
BuildMI(MBB, I, DL, get(Opc), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMOA);
} else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) {
// If there are variable-sized objects, spills will not be aligned.
if (HasAlloca)
SlotAlign = HFI.getStackAlignment();
unsigned Opc = SlotAlign < RegAlign ? Hexagon::PS_vloadrwu_ai
: Hexagon::PS_vloadrw_ai;
MachineMemOperand *MMOA = MF.getMachineMemOperand(
MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
MFI.getObjectSize(FI), SlotAlign);
BuildMI(MBB, I, DL, get(Opc), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMOA);
} else {
llvm_unreachable("Can't store this register to stack slot");
}
}
static void getLiveRegsAt(LivePhysRegs &Regs, const MachineInstr &MI) {
const MachineBasicBlock &B = *MI.getParent();
Regs.addLiveOuts(B);
auto E = ++MachineBasicBlock::const_iterator(MI.getIterator()).getReverse();
for (auto I = B.rbegin(); I != E; ++I)
Regs.stepBackward(*I);
}
/// expandPostRAPseudo - This function is called for all pseudo instructions
/// that remain after register allocation. Many pseudo instructions are
/// created to help register allocation. This is the place to convert them
/// into real instructions. The target can edit MI in place, or it can insert
/// new instructions and erase MI. The function should return true if
/// anything was changed.
bool HexagonInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
MachineBasicBlock &MBB = *MI.getParent();
MachineFunction &MF = *MBB.getParent();
MachineRegisterInfo &MRI = MF.getRegInfo();
const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
DebugLoc DL = MI.getDebugLoc();
unsigned Opc = MI.getOpcode();
switch (Opc) {
case TargetOpcode::COPY: {
MachineOperand &MD = MI.getOperand(0);
MachineOperand &MS = MI.getOperand(1);
MachineBasicBlock::iterator MBBI = MI.getIterator();
if (MD.getReg() != MS.getReg() && !MS.isUndef()) {
copyPhysReg(MBB, MI, DL, MD.getReg(), MS.getReg(), MS.isKill());
std::prev(MBBI)->copyImplicitOps(*MBB.getParent(), MI);
}
MBB.erase(MBBI);
return true;
}
case Hexagon::PS_aligna:
BuildMI(MBB, MI, DL, get(Hexagon::A2_andir), MI.getOperand(0).getReg())
.addReg(HRI.getFrameRegister())
.addImm(-MI.getOperand(1).getImm());
MBB.erase(MI);
return true;
case Hexagon::V6_vassignp: {
unsigned SrcReg = MI.getOperand(1).getReg();
unsigned DstReg = MI.getOperand(0).getReg();
unsigned Kill = getKillRegState(MI.getOperand(1).isKill());
BuildMI(MBB, MI, DL, get(Hexagon::V6_vcombine), DstReg)
.addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_hi), Kill)
.addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_lo), Kill);
MBB.erase(MI);
return true;
}
case Hexagon::V6_lo: {
unsigned SrcReg = MI.getOperand(1).getReg();
unsigned DstReg = MI.getOperand(0).getReg();
unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
copyPhysReg(MBB, MI, DL, DstReg, SrcSubLo, MI.getOperand(1).isKill());
MBB.erase(MI);
MRI.clearKillFlags(SrcSubLo);
return true;
}
case Hexagon::V6_hi: {
unsigned SrcReg = MI.getOperand(1).getReg();
unsigned DstReg = MI.getOperand(0).getReg();
unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
copyPhysReg(MBB, MI, DL, DstReg, SrcSubHi, MI.getOperand(1).isKill());
MBB.erase(MI);
MRI.clearKillFlags(SrcSubHi);
return true;
}
case Hexagon::PS_vstorerw_ai:
case Hexagon::PS_vstorerwu_ai: {
bool Aligned = Opc == Hexagon::PS_vstorerw_ai;
unsigned SrcReg = MI.getOperand(2).getReg();
unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
unsigned NewOpc = Aligned ? Hexagon::V6_vS32b_ai : Hexagon::V6_vS32Ub_ai;
unsigned Offset = HRI.getSpillSize(Hexagon::HvxVRRegClass);
MachineInstr *MI1New =
BuildMI(MBB, MI, DL, get(NewOpc))
.add(MI.getOperand(0))
.addImm(MI.getOperand(1).getImm())
.addReg(SrcSubLo)
.setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
MI1New->getOperand(0).setIsKill(false);
BuildMI(MBB, MI, DL, get(NewOpc))
.add(MI.getOperand(0))
// The Vectors are indexed in multiples of vector size.
.addImm(MI.getOperand(1).getImm() + Offset)
.addReg(SrcSubHi)
.setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
MBB.erase(MI);
return true;
}
case Hexagon::PS_vloadrw_ai:
case Hexagon::PS_vloadrwu_ai: {
bool Aligned = Opc == Hexagon::PS_vloadrw_ai;
unsigned DstReg = MI.getOperand(0).getReg();
unsigned NewOpc = Aligned ? Hexagon::V6_vL32b_ai : Hexagon::V6_vL32Ub_ai;
unsigned Offset = HRI.getSpillSize(Hexagon::HvxVRRegClass);
MachineInstr *MI1New = BuildMI(MBB, MI, DL, get(NewOpc),
HRI.getSubReg(DstReg, Hexagon::vsub_lo))
.add(MI.getOperand(1))
.addImm(MI.getOperand(2).getImm())
.setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
MI1New->getOperand(1).setIsKill(false);
BuildMI(MBB, MI, DL, get(NewOpc), HRI.getSubReg(DstReg, Hexagon::vsub_hi))
.add(MI.getOperand(1))
// The Vectors are indexed in multiples of vector size.
.addImm(MI.getOperand(2).getImm() + Offset)
.setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
MBB.erase(MI);
return true;
}
case Hexagon::PS_true: {
unsigned Reg = MI.getOperand(0).getReg();
BuildMI(MBB, MI, DL, get(Hexagon::C2_orn), Reg)
.addReg(Reg, RegState::Undef)
.addReg(Reg, RegState::Undef);
MBB.erase(MI);
return true;
}
case Hexagon::PS_false: {
unsigned Reg = MI.getOperand(0).getReg();
BuildMI(MBB, MI, DL, get(Hexagon::C2_andn), Reg)
.addReg(Reg, RegState::Undef)
.addReg(Reg, RegState::Undef);
MBB.erase(MI);
return true;
}
case Hexagon::PS_vmulw: {
// Expand a 64-bit vector multiply into 2 32-bit scalar multiplies.
unsigned DstReg = MI.getOperand(0).getReg();
unsigned Src1Reg = MI.getOperand(1).getReg();
unsigned Src2Reg = MI.getOperand(2).getReg();
unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi);
unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo);
unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi);
unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo);
BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi),
HRI.getSubReg(DstReg, Hexagon::isub_hi))
.addReg(Src1SubHi)
.addReg(Src2SubHi);
BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi),
HRI.getSubReg(DstReg, Hexagon::isub_lo))
.addReg(Src1SubLo)
.addReg(Src2SubLo);
MBB.erase(MI);
MRI.clearKillFlags(Src1SubHi);
MRI.clearKillFlags(Src1SubLo);
MRI.clearKillFlags(Src2SubHi);
MRI.clearKillFlags(Src2SubLo);
return true;
}
case Hexagon::PS_vmulw_acc: {
// Expand 64-bit vector multiply with addition into 2 scalar multiplies.
unsigned DstReg = MI.getOperand(0).getReg();
unsigned Src1Reg = MI.getOperand(1).getReg();
unsigned Src2Reg = MI.getOperand(2).getReg();
unsigned Src3Reg = MI.getOperand(3).getReg();
unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi);
unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo);
unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi);
unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo);
unsigned Src3SubHi = HRI.getSubReg(Src3Reg, Hexagon::isub_hi);
unsigned Src3SubLo = HRI.getSubReg(Src3Reg, Hexagon::isub_lo);
BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci),
HRI.getSubReg(DstReg, Hexagon::isub_hi))
.addReg(Src1SubHi)
.addReg(Src2SubHi)
.addReg(Src3SubHi);
BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci),
HRI.getSubReg(DstReg, Hexagon::isub_lo))
.addReg(Src1SubLo)
.addReg(Src2SubLo)
.addReg(Src3SubLo);
MBB.erase(MI);
MRI.clearKillFlags(Src1SubHi);
MRI.clearKillFlags(Src1SubLo);
MRI.clearKillFlags(Src2SubHi);
MRI.clearKillFlags(Src2SubLo);
MRI.clearKillFlags(Src3SubHi);
MRI.clearKillFlags(Src3SubLo);
return true;
}
case Hexagon::PS_pselect: {
const MachineOperand &Op0 = MI.getOperand(0);
const MachineOperand &Op1 = MI.getOperand(1);
const MachineOperand &Op2 = MI.getOperand(2);
const MachineOperand &Op3 = MI.getOperand(3);
unsigned Rd = Op0.getReg();
unsigned Pu = Op1.getReg();
unsigned Rs = Op2.getReg();
unsigned Rt = Op3.getReg();
DebugLoc DL = MI.getDebugLoc();
unsigned K1 = getKillRegState(Op1.isKill());
unsigned K2 = getKillRegState(Op2.isKill());
unsigned K3 = getKillRegState(Op3.isKill());
if (Rd != Rs)
BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpt), Rd)
.addReg(Pu, (Rd == Rt) ? K1 : 0)
.addReg(Rs, K2);
if (Rd != Rt)
BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpf), Rd)
.addReg(Pu, K1)
.addReg(Rt, K3);
MBB.erase(MI);
return true;
}
case Hexagon::PS_vselect: {
const MachineOperand &Op0 = MI.getOperand(0);
const MachineOperand &Op1 = MI.getOperand(1);
const MachineOperand &Op2 = MI.getOperand(2);
const MachineOperand &Op3 = MI.getOperand(3);
LivePhysRegs LiveAtMI(HRI);
getLiveRegsAt(LiveAtMI, MI);
bool IsDestLive = !LiveAtMI.available(MRI, Op0.getReg());
unsigned PReg = Op1.getReg();
assert(Op1.getSubReg() == 0);
unsigned PState = getRegState(Op1);
if (Op0.getReg() != Op2.getReg()) {
unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill
: PState;
auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vcmov))
.add(Op0)
.addReg(PReg, S)
.add(Op2);
if (IsDestLive)
T.addReg(Op0.getReg(), RegState::Implicit);
IsDestLive = true;
}
if (Op0.getReg() != Op3.getReg()) {
auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vncmov))
.add(Op0)
.addReg(PReg, PState)
.add(Op3);
if (IsDestLive)
T.addReg(Op0.getReg(), RegState::Implicit);
}
MBB.erase(MI);
return true;
}
case Hexagon::PS_wselect: {
MachineOperand &Op0 = MI.getOperand(0);
MachineOperand &Op1 = MI.getOperand(1);
MachineOperand &Op2 = MI.getOperand(2);
MachineOperand &Op3 = MI.getOperand(3);
LivePhysRegs LiveAtMI(HRI);
getLiveRegsAt(LiveAtMI, MI);
bool IsDestLive = !LiveAtMI.available(MRI, Op0.getReg());
unsigned PReg = Op1.getReg();
assert(Op1.getSubReg() == 0);
unsigned PState = getRegState(Op1);
if (Op0.getReg() != Op2.getReg()) {
unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill
: PState;
unsigned SrcLo = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_lo);
unsigned SrcHi = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_hi);
auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vccombine))
.add(Op0)
.addReg(PReg, S)
.add(Op1)
.addReg(SrcHi)
.addReg(SrcLo);
if (IsDestLive)
T.addReg(Op0.getReg(), RegState::Implicit);
IsDestLive = true;
}
if (Op0.getReg() != Op3.getReg()) {
unsigned SrcLo = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_lo);
unsigned SrcHi = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_hi);
auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vnccombine))
.add(Op0)
.addReg(PReg, PState)
.addReg(SrcHi)
.addReg(SrcLo);
if (IsDestLive)
T.addReg(Op0.getReg(), RegState::Implicit);
}
MBB.erase(MI);
return true;
}
case Hexagon::PS_tailcall_i:
MI.setDesc(get(Hexagon::J2_jump));
return true;
case Hexagon::PS_tailcall_r:
case Hexagon::PS_jmpret:
MI.setDesc(get(Hexagon::J2_jumpr));
return true;
case Hexagon::PS_jmprett:
MI.setDesc(get(Hexagon::J2_jumprt));
return true;
case Hexagon::PS_jmpretf:
MI.setDesc(get(Hexagon::J2_jumprf));
return true;
case Hexagon::PS_jmprettnewpt:
MI.setDesc(get(Hexagon::J2_jumprtnewpt));
return true;
case Hexagon::PS_jmpretfnewpt:
MI.setDesc(get(Hexagon::J2_jumprfnewpt));
return true;
case Hexagon::PS_jmprettnew:
MI.setDesc(get(Hexagon::J2_jumprtnew));
return true;
case Hexagon::PS_jmpretfnew:
MI.setDesc(get(Hexagon::J2_jumprfnew));
return true;
}
return false;
}
// We indicate that we want to reverse the branch by
// inserting the reversed branching opcode.
bool HexagonInstrInfo::reverseBranchCondition(
SmallVectorImpl<MachineOperand> &Cond) const {
if (Cond.empty())
return true;
assert(Cond[0].isImm() && "First entry in the cond vector not imm-val");
unsigned opcode = Cond[0].getImm();
//unsigned temp;
assert(get(opcode).isBranch() && "Should be a branching condition.");
if (isEndLoopN(opcode))
return true;
unsigned NewOpcode = getInvertedPredicatedOpcode(opcode);
Cond[0].setImm(NewOpcode);
return false;
}
void HexagonInstrInfo::insertNoop(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI) const {
DebugLoc DL;
BuildMI(MBB, MI, DL, get(Hexagon::A2_nop));
}
bool HexagonInstrInfo::isPostIncrement(const MachineInstr &MI) const {
return getAddrMode(MI) == HexagonII::PostInc;
}
// Returns true if an instruction is predicated irrespective of the predicate
// sense. For example, all of the following will return true.
// if (p0) R1 = add(R2, R3)
// if (!p0) R1 = add(R2, R3)
// if (p0.new) R1 = add(R2, R3)
// if (!p0.new) R1 = add(R2, R3)
// Note: New-value stores are not included here as in the current
// implementation, we don't need to check their predicate sense.
bool HexagonInstrInfo::isPredicated(const MachineInstr &MI) const {
const uint64_t F = MI.getDesc().TSFlags;
return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask;
}
bool HexagonInstrInfo::PredicateInstruction(
MachineInstr &MI, ArrayRef<MachineOperand> Cond) const {
if (Cond.empty() || isNewValueJump(Cond[0].getImm()) ||
isEndLoopN(Cond[0].getImm())) {
DEBUG(dbgs() << "\nCannot predicate:"; MI.dump(););
return false;
}
int Opc = MI.getOpcode();
assert (isPredicable(MI) && "Expected predicable instruction");
bool invertJump = predOpcodeHasNot(Cond);
// We have to predicate MI "in place", i.e. after this function returns,
// MI will need to be transformed into a predicated form. To avoid com-
// plicated manipulations with the operands (handling tied operands,
// etc.), build a new temporary instruction, then overwrite MI with it.
MachineBasicBlock &B = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
unsigned PredOpc = getCondOpcode(Opc, invertJump);
MachineInstrBuilder T = BuildMI(B, MI, DL, get(PredOpc));
unsigned NOp = 0, NumOps = MI.getNumOperands();
while (NOp < NumOps) {
MachineOperand &Op = MI.getOperand(NOp);
if (!Op.isReg() || !Op.isDef() || Op.isImplicit())
break;
T.add(Op);
NOp++;
}
unsigned PredReg, PredRegPos, PredRegFlags;
bool GotPredReg = getPredReg(Cond, PredReg, PredRegPos, PredRegFlags);
(void)GotPredReg;
assert(GotPredReg);
T.addReg(PredReg, PredRegFlags);
while (NOp < NumOps)
T.add(MI.getOperand(NOp++));
MI.setDesc(get(PredOpc));
while (unsigned n = MI.getNumOperands())
MI.RemoveOperand(n-1);
for (unsigned i = 0, n = T->getNumOperands(); i < n; ++i)
MI.addOperand(T->getOperand(i));
MachineBasicBlock::instr_iterator TI = T->getIterator();
B.erase(TI);
MachineRegisterInfo &MRI = B.getParent()->getRegInfo();
MRI.clearKillFlags(PredReg);
return true;
}
bool HexagonInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
ArrayRef<MachineOperand> Pred2) const {
// TODO: Fix this
return false;
}
bool HexagonInstrInfo::DefinesPredicate(MachineInstr &MI,
std::vector<MachineOperand> &Pred) const {
const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
for (unsigned oper = 0; oper < MI.getNumOperands(); ++oper) {
MachineOperand MO = MI.getOperand(oper);
if (MO.isReg()) {
if (!MO.isDef())
continue;
const TargetRegisterClass* RC = HRI.getMinimalPhysRegClass(MO.getReg());
if (RC == &Hexagon::PredRegsRegClass) {
Pred.push_back(MO);
return true;
}
continue;
} else if (MO.isRegMask()) {
for (unsigned PR : Hexagon::PredRegsRegClass) {
if (!MI.modifiesRegister(PR, &HRI))
continue;
Pred.push_back(MO);
return true;
}
}
}
return false;
}
bool HexagonInstrInfo::isPredicable(const MachineInstr &MI) const {
if (!MI.getDesc().isPredicable())
return false;
if (MI.isCall() || isTailCall(MI)) {
if (!Subtarget.usePredicatedCalls())
return false;
}
// HVX loads are not predicable on v60, but are on v62.
if (!Subtarget.hasV62TOps()) {
switch (MI.getOpcode()) {
case Hexagon::V6_vL32b_ai:
case Hexagon::V6_vL32b_pi:
case Hexagon::V6_vL32b_ppu:
case Hexagon::V6_vL32b_cur_ai:
case Hexagon::V6_vL32b_cur_pi:
case Hexagon::V6_vL32b_cur_ppu:
case Hexagon::V6_vL32b_nt_ai:
case Hexagon::V6_vL32b_nt_pi:
case Hexagon::V6_vL32b_nt_ppu:
case Hexagon::V6_vL32b_tmp_ai:
case Hexagon::V6_vL32b_tmp_pi:
case Hexagon::V6_vL32b_tmp_ppu:
case Hexagon::V6_vL32b_nt_cur_ai:
case Hexagon::V6_vL32b_nt_cur_pi:
case Hexagon::V6_vL32b_nt_cur_ppu:
case Hexagon::V6_vL32b_nt_tmp_ai:
case Hexagon::V6_vL32b_nt_tmp_pi:
case Hexagon::V6_vL32b_nt_tmp_ppu:
return false;
}
}
return true;
}
bool HexagonInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
const MachineBasicBlock *MBB,
const MachineFunction &MF) const {
// Debug info is never a scheduling boundary. It's necessary to be explicit
// due to the special treatment of IT instructions below, otherwise a
// dbg_value followed by an IT will result in the IT instruction being
// considered a scheduling hazard, which is wrong. It should be the actual
// instruction preceding the dbg_value instruction(s), just like it is
// when debug info is not present.
if (MI.isDebugValue())
return false;
// Throwing call is a boundary.
if (MI.isCall()) {
// Don't mess around with no return calls.
if (doesNotReturn(MI))
return true;
// If any of the block's successors is a landing pad, this could be a
// throwing call.
for (auto I : MBB->successors())
if (I->isEHPad())
return true;
}
// Terminators and labels can't be scheduled around.
if (MI.getDesc().isTerminator() || MI.isPosition())
return true;
if (MI.isInlineAsm() && !ScheduleInlineAsm)
return true;
return false;
}
/// Measure the specified inline asm to determine an approximation of its
/// length.
/// Comments (which run till the next SeparatorString or newline) do not
/// count as an instruction.
/// Any other non-whitespace text is considered an instruction, with
/// multiple instructions separated by SeparatorString or newlines.
/// Variable-length instructions are not handled here; this function
/// may be overloaded in the target code to do that.
/// Hexagon counts the number of ##'s and adjust for that many
/// constant exenders.
unsigned HexagonInstrInfo::getInlineAsmLength(const char *Str,
const MCAsmInfo &MAI) const {
StringRef AStr(Str);
// Count the number of instructions in the asm.
bool atInsnStart = true;
unsigned Length = 0;
for (; *Str; ++Str) {
if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
strlen(MAI.getSeparatorString())) == 0)
atInsnStart = true;
if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
Length += MAI.getMaxInstLength();
atInsnStart = false;
}
if (atInsnStart && strncmp(Str, MAI.getCommentString().data(),
MAI.getCommentString().size()) == 0)
atInsnStart = false;
}
// Add to size number of constant extenders seen * 4.
StringRef Occ("##");
Length += AStr.count(Occ)*4;
return Length;
}
ScheduleHazardRecognizer*
HexagonInstrInfo::CreateTargetPostRAHazardRecognizer(
const InstrItineraryData *II, const ScheduleDAG *DAG) const {
if (UseDFAHazardRec)
return new HexagonHazardRecognizer(II, this, Subtarget);
return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
}
/// \brief For a comparison instruction, return the source registers in
/// \p SrcReg and \p SrcReg2 if having two register operands, and the value it
/// compares against in CmpValue. Return true if the comparison instruction
/// can be analyzed.
bool HexagonInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
unsigned &SrcReg2, int &Mask,
int &Value) const {
unsigned Opc = MI.getOpcode();
// Set mask and the first source register.
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::C4_cmpneq:
case Hexagon::C4_cmplte:
case Hexagon::C4_cmplteu:
case Hexagon::C2_cmpeqi:
case Hexagon::C2_cmpgti:
case Hexagon::C2_cmpgtui:
case Hexagon::C4_cmpneqi:
case Hexagon::C4_cmplteui:
case Hexagon::C4_cmpltei:
SrcReg = MI.getOperand(1).getReg();
Mask = ~0;
break;
case Hexagon::A4_cmpbeq:
case Hexagon::A4_cmpbgt:
case Hexagon::A4_cmpbgtu:
case Hexagon::A4_cmpbeqi:
case Hexagon::A4_cmpbgti:
case Hexagon::A4_cmpbgtui:
SrcReg = MI.getOperand(1).getReg();
Mask = 0xFF;
break;
case Hexagon::A4_cmpheq:
case Hexagon::A4_cmphgt:
case Hexagon::A4_cmphgtu:
case Hexagon::A4_cmpheqi:
case Hexagon::A4_cmphgti:
case Hexagon::A4_cmphgtui:
SrcReg = MI.getOperand(1).getReg();
Mask = 0xFFFF;
break;
}
// Set the value/second source register.
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::A4_cmpbeq:
case Hexagon::A4_cmpbgt:
case Hexagon::A4_cmpbgtu:
case Hexagon::A4_cmpheq:
case Hexagon::A4_cmphgt:
case Hexagon::A4_cmphgtu:
case Hexagon::C4_cmpneq:
case Hexagon::C4_cmplte:
case Hexagon::C4_cmplteu:
SrcReg2 = MI.getOperand(2).getReg();
return true;
case Hexagon::C2_cmpeqi:
case Hexagon::C2_cmpgtui:
case Hexagon::C2_cmpgti:
case Hexagon::C4_cmpneqi:
case Hexagon::C4_cmplteui:
case Hexagon::C4_cmpltei:
case Hexagon::A4_cmpbeqi:
case Hexagon::A4_cmpbgti:
case Hexagon::A4_cmpbgtui:
case Hexagon::A4_cmpheqi:
case Hexagon::A4_cmphgti:
case Hexagon::A4_cmphgtui: {
SrcReg2 = 0;
const MachineOperand &Op2 = MI.getOperand(2);
if (!Op2.isImm())
return false;
Value = MI.getOperand(2).getImm();
return true;
}
}
return false;
}
unsigned HexagonInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
const MachineInstr &MI,
unsigned *PredCost) const {
return getInstrTimingClassLatency(ItinData, MI);
}
DFAPacketizer *HexagonInstrInfo::CreateTargetScheduleState(
const TargetSubtargetInfo &STI) const {
const InstrItineraryData *II = STI.getInstrItineraryData();
return static_cast<const HexagonSubtarget&>(STI).createDFAPacketizer(II);
}
// Inspired by this pair:
// %R13<def> = L2_loadri_io %R29, 136; mem:LD4[FixedStack0]
// S2_storeri_io %R29, 132, %R1<kill>; flags: mem:ST4[FixedStack1]
// Currently AA considers the addresses in these instructions to be aliasing.
bool HexagonInstrInfo::areMemAccessesTriviallyDisjoint(
MachineInstr &MIa, MachineInstr &MIb, AliasAnalysis *AA) const {
if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() ||
MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
return false;
// Instructions that are pure loads, not loads and stores like memops are not
// dependent.
if (MIa.mayLoad() && !isMemOp(MIa) && MIb.mayLoad() && !isMemOp(MIb))
return true;
// Get the base register in MIa.
unsigned BasePosA, OffsetPosA;
if (!getBaseAndOffsetPosition(MIa, BasePosA, OffsetPosA))
return false;
const MachineOperand &BaseA = MIa.getOperand(BasePosA);
unsigned BaseRegA = BaseA.getReg();
unsigned BaseSubA = BaseA.getSubReg();
// Get the base register in MIb.
unsigned BasePosB, OffsetPosB;
if (!getBaseAndOffsetPosition(MIb, BasePosB, OffsetPosB))
return false;
const MachineOperand &BaseB = MIb.getOperand(BasePosB);
unsigned BaseRegB = BaseB.getReg();
unsigned BaseSubB = BaseB.getSubReg();
if (BaseRegA != BaseRegB || BaseSubA != BaseSubB)
return false;
// Get the access sizes.
unsigned SizeA = getMemAccessSize(MIa);
unsigned SizeB = getMemAccessSize(MIb);
// Get the offsets. Handle immediates only for now.
const MachineOperand &OffA = MIa.getOperand(OffsetPosA);
const MachineOperand &OffB = MIb.getOperand(OffsetPosB);
if (!MIa.getOperand(OffsetPosA).isImm() ||
!MIb.getOperand(OffsetPosB).isImm())
return false;
int OffsetA = isPostIncrement(MIa) ? 0 : OffA.getImm();
int OffsetB = isPostIncrement(MIb) ? 0 : OffB.getImm();
// This is a mem access with the same base register and known offsets from it.
// Reason about it.
if (OffsetA > OffsetB) {
uint64_t OffDiff = (uint64_t)((int64_t)OffsetA - (int64_t)OffsetB);
return SizeB <= OffDiff;
}
if (OffsetA < OffsetB) {
uint64_t OffDiff = (uint64_t)((int64_t)OffsetB - (int64_t)OffsetA);
return SizeA <= OffDiff;
}
return false;
}
/// If the instruction is an increment of a constant value, return the amount.
bool HexagonInstrInfo::getIncrementValue(const MachineInstr &MI,
int &Value) const {
if (isPostIncrement(MI)) {
unsigned BasePos = 0, OffsetPos = 0;
if (!getBaseAndOffsetPosition(MI, BasePos, OffsetPos))
return false;
const MachineOperand &OffsetOp = MI.getOperand(OffsetPos);
if (OffsetOp.isImm()) {
Value = OffsetOp.getImm();
return true;
}
} else if (MI.getOpcode() == Hexagon::A2_addi) {
const MachineOperand &AddOp = MI.getOperand(2);
if (AddOp.isImm()) {
Value = AddOp.getImm();
return true;
}
}
return false;
}
std::pair<unsigned, unsigned>
HexagonInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
return std::make_pair(TF & ~HexagonII::MO_Bitmasks,
TF & HexagonII::MO_Bitmasks);
}
ArrayRef<std::pair<unsigned, const char*>>
HexagonInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
using namespace HexagonII;
static const std::pair<unsigned, const char*> Flags[] = {
{MO_PCREL, "hexagon-pcrel"},
{MO_GOT, "hexagon-got"},
{MO_LO16, "hexagon-lo16"},
{MO_HI16, "hexagon-hi16"},
{MO_GPREL, "hexagon-gprel"},
{MO_GDGOT, "hexagon-gdgot"},
{MO_GDPLT, "hexagon-gdplt"},
{MO_IE, "hexagon-ie"},
{MO_IEGOT, "hexagon-iegot"},
{MO_TPREL, "hexagon-tprel"}
};
return makeArrayRef(Flags);
}
ArrayRef<std::pair<unsigned, const char*>>
HexagonInstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const {
using namespace HexagonII;
static const std::pair<unsigned, const char*> Flags[] = {
{HMOTF_ConstExtended, "hexagon-ext"}
};
return makeArrayRef(Flags);
}
unsigned HexagonInstrInfo::createVR(MachineFunction *MF, MVT VT) const {
MachineRegisterInfo &MRI = MF->getRegInfo();
const TargetRegisterClass *TRC;
if (VT == MVT::i1) {
TRC = &Hexagon::PredRegsRegClass;
} else if (VT == MVT::i32 || VT == MVT::f32) {
TRC = &Hexagon::IntRegsRegClass;
} else if (VT == MVT::i64 || VT == MVT::f64) {
TRC = &Hexagon::DoubleRegsRegClass;
} else {
llvm_unreachable("Cannot handle this register class");
}
unsigned NewReg = MRI.createVirtualRegister(TRC);
return NewReg;
}
bool HexagonInstrInfo::isAbsoluteSet(const MachineInstr &MI) const {
return (getAddrMode(MI) == HexagonII::AbsoluteSet);
}
bool HexagonInstrInfo::isAccumulator(const MachineInstr &MI) const {
const uint64_t F = MI.getDesc().TSFlags;
return((F >> HexagonII::AccumulatorPos) & HexagonII::AccumulatorMask);
}
bool HexagonInstrInfo::isComplex(const MachineInstr &MI) const {
return !isTC1(MI) && !isTC2Early(MI) && !MI.getDesc().mayLoad() &&
!MI.getDesc().mayStore() &&
MI.getDesc().getOpcode() != Hexagon::S2_allocframe &&
MI.getDesc().getOpcode() != Hexagon::L2_deallocframe &&
!isMemOp(MI) && !MI.isBranch() && !MI.isReturn() && !MI.isCall();
}
// Return true if the instruction is a compund branch instruction.
bool HexagonInstrInfo::isCompoundBranchInstr(const MachineInstr &MI) const {
return getType(MI) == HexagonII::TypeCJ && MI.isBranch();
}
// TODO: In order to have isExtendable for fpimm/f32Ext, we need to handle
// isFPImm and later getFPImm as well.
bool HexagonInstrInfo::isConstExtended(const MachineInstr &MI) const {
const uint64_t F = MI.getDesc().TSFlags;
unsigned isExtended = (F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask;
if (isExtended) // Instruction must be extended.
return true;
unsigned isExtendable =
(F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask;
if (!isExtendable)
return false;
if (MI.isCall())
return false;
short ExtOpNum = getCExtOpNum(MI);
const MachineOperand &MO = MI.getOperand(ExtOpNum);
// Use MO operand flags to determine if MO
// has the HMOTF_ConstExtended flag set.
if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended)
return true;
// If this is a Machine BB address we are talking about, and it is
// not marked as extended, say so.
if (MO.isMBB())
return false;
// We could be using an instruction with an extendable immediate and shoehorn
// a global address into it. If it is a global address it will be constant
// extended. We do this for COMBINE.
if (MO.isGlobal() || MO.isSymbol() || MO.isBlockAddress() ||
MO.isJTI() || MO.isCPI() || MO.isFPImm())
return true;
// If the extendable operand is not 'Immediate' type, the instruction should
// have 'isExtended' flag set.
assert(MO.isImm() && "Extendable operand must be Immediate type");
int MinValue = getMinValue(MI);
int MaxValue = getMaxValue(MI);
int ImmValue = MO.getImm();
return (ImmValue < MinValue || ImmValue > MaxValue);
}
bool HexagonInstrInfo::isDeallocRet(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
case Hexagon::L4_return:
case Hexagon::L4_return_t:
case Hexagon::L4_return_f:
case Hexagon::L4_return_tnew_pnt:
case Hexagon::L4_return_fnew_pnt:
case Hexagon::L4_return_tnew_pt:
case Hexagon::L4_return_fnew_pt:
return true;
}
return false;
}
// Return true when ConsMI uses a register defined by ProdMI.
bool HexagonInstrInfo::isDependent(const MachineInstr &ProdMI,
const MachineInstr &ConsMI) const {
if (!ProdMI.getDesc().getNumDefs())
return false;
const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
SmallVector<unsigned, 4> DefsA;
SmallVector<unsigned, 4> DefsB;
SmallVector<unsigned, 8> UsesA;
SmallVector<unsigned, 8> UsesB;
parseOperands(ProdMI, DefsA, UsesA);
parseOperands(ConsMI, DefsB, UsesB);
for (auto &RegA : DefsA)
for (auto &RegB : UsesB) {
// True data dependency.
if (RegA == RegB)
return true;
if (TargetRegisterInfo::isPhysicalRegister(RegA))
for (MCSubRegIterator SubRegs(RegA, &HRI); SubRegs.isValid(); ++SubRegs)
if (RegB == *SubRegs)
return true;
if (TargetRegisterInfo::isPhysicalRegister(RegB))
for (MCSubRegIterator SubRegs(RegB, &HRI); SubRegs.isValid(); ++SubRegs)
if (RegA == *SubRegs)
return true;
}
return false;
}
// Returns true if the instruction is alread a .cur.
bool HexagonInstrInfo::isDotCurInst(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
case Hexagon::V6_vL32b_cur_pi:
case Hexagon::V6_vL32b_cur_ai:
return true;
}
return false;
}
// Returns true, if any one of the operands is a dot new
// insn, whether it is predicated dot new or register dot new.
bool HexagonInstrInfo::isDotNewInst(const MachineInstr &MI) const {
if (isNewValueInst(MI) || (isPredicated(MI) && isPredicatedNew(MI)))
return true;
return false;
}
/// Symmetrical. See if these two instructions are fit for duplex pair.
bool HexagonInstrInfo::isDuplexPair(const MachineInstr &MIa,
const MachineInstr &MIb) const {
HexagonII::SubInstructionGroup MIaG = getDuplexCandidateGroup(MIa);
HexagonII::SubInstructionGroup MIbG = getDuplexCandidateGroup(MIb);
return (isDuplexPairMatch(MIaG, MIbG) || isDuplexPairMatch(MIbG, MIaG));
}
bool HexagonInstrInfo::isEarlySourceInstr(const MachineInstr &MI) const {
if (MI.mayLoad() || MI.mayStore() || MI.isCompare())
return true;
// Multiply
unsigned SchedClass = MI.getDesc().getSchedClass();
return is_TC4x(SchedClass) || is_TC3x(SchedClass);
}
bool HexagonInstrInfo::isEndLoopN(unsigned Opcode) const {
return (Opcode == Hexagon::ENDLOOP0 ||
Opcode == Hexagon::ENDLOOP1);
}
bool HexagonInstrInfo::isExpr(unsigned OpType) const {
switch(OpType) {
case MachineOperand::MO_MachineBasicBlock:
case MachineOperand::MO_GlobalAddress:
case MachineOperand::MO_ExternalSymbol:
case MachineOperand::MO_JumpTableIndex:
case MachineOperand::MO_ConstantPoolIndex:
case MachineOperand::MO_BlockAddress:
return true;
default:
return false;
}
}
bool HexagonInstrInfo::isExtendable(const MachineInstr &MI) const {
const MCInstrDesc &MID = MI.getDesc();
const uint64_t F = MID.TSFlags;
if ((F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask)
return true;
// TODO: This is largely obsolete now. Will need to be removed
// in consecutive patches.
switch (MI.getOpcode()) {
// PS_fi and PS_fia remain special cases.
case Hexagon::PS_fi:
case Hexagon::PS_fia:
return true;
default:
return false;
}
return false;
}
// This returns true in two cases:
// - The OP code itself indicates that this is an extended instruction.
// - One of MOs has been marked with HMOTF_ConstExtended flag.
bool HexagonInstrInfo::isExtended(const MachineInstr &MI) const {
// First check if this is permanently extended op code.
const uint64_t F = MI.getDesc().TSFlags;
if ((F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask)
return true;
// Use MO operand flags to determine if one of MI's operands
// has HMOTF_ConstExtended flag set.
for (const MachineOperand &MO : MI.operands())
if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended)
return true;
return false;
}
bool HexagonInstrInfo::isFloat(const MachineInstr &MI) const {
unsigned Opcode = MI.getOpcode();
const uint64_t F = get(Opcode).TSFlags;
return (F >> HexagonII::FPPos) & HexagonII::FPMask;
}
// No V60 HVX VMEM with A_INDIRECT.
bool HexagonInstrInfo::isHVXMemWithAIndirect(const MachineInstr &I,
const MachineInstr &J) const {
if (!isHVXVec(I))
return false;
if (!I.mayLoad() && !I.mayStore())
return false;
return J.isIndirectBranch() || isIndirectCall(J) || isIndirectL4Return(J);
}
bool HexagonInstrInfo::isIndirectCall(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
case Hexagon::J2_callr:
case Hexagon::J2_callrf:
case Hexagon::J2_callrt:
case Hexagon::PS_call_nr:
return true;
}
return false;
}
bool HexagonInstrInfo::isIndirectL4Return(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
case Hexagon::L4_return:
case Hexagon::L4_return_t:
case Hexagon::L4_return_f:
case Hexagon::L4_return_fnew_pnt:
case Hexagon::L4_return_fnew_pt:
case Hexagon::L4_return_tnew_pnt:
case Hexagon::L4_return_tnew_pt:
return true;
}
return false;
}
bool HexagonInstrInfo::isJumpR(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
case Hexagon::J2_jumpr:
case Hexagon::J2_jumprt:
case Hexagon::J2_jumprf:
case Hexagon::J2_jumprtnewpt:
case Hexagon::J2_jumprfnewpt:
case Hexagon::J2_jumprtnew:
case Hexagon::J2_jumprfnew:
return true;
}
return false;
}
// Return true if a given MI can accommodate given offset.
// Use abs estimate as oppose to the exact number.
// TODO: This will need to be changed to use MC level
// definition of instruction extendable field size.
bool HexagonInstrInfo::isJumpWithinBranchRange(const MachineInstr &MI,
unsigned offset) const {
// This selection of jump instructions matches to that what
// analyzeBranch can parse, plus NVJ.
if (isNewValueJump(MI)) // r9:2
return isInt<11>(offset);
switch (MI.getOpcode()) {
// Still missing Jump to address condition on register value.
default:
return false;
case Hexagon::J2_jump: // bits<24> dst; // r22:2
case Hexagon::J2_call:
case Hexagon::PS_call_nr:
return isInt<24>(offset);
case Hexagon::J2_jumpt: //bits<17> dst; // r15:2
case Hexagon::J2_jumpf:
case Hexagon::J2_jumptnew:
case Hexagon::J2_jumptnewpt:
case Hexagon::J2_jumpfnew:
case Hexagon::J2_jumpfnewpt:
case Hexagon::J2_callt:
case Hexagon::J2_callf:
return isInt<17>(offset);
case Hexagon::J2_loop0i:
case Hexagon::J2_loop0iext:
case Hexagon::J2_loop0r:
case Hexagon::J2_loop0rext:
case Hexagon::J2_loop1i:
case Hexagon::J2_loop1iext:
case Hexagon::J2_loop1r:
case Hexagon::J2_loop1rext:
return isInt<9>(offset);
// TODO: Add all the compound branches here. Can we do this in Relation model?
case Hexagon::J4_cmpeqi_tp0_jump_nt:
case Hexagon::J4_cmpeqi_tp1_jump_nt:
return isInt<11>(offset);
}
}
bool HexagonInstrInfo::isLateInstrFeedsEarlyInstr(const MachineInstr &LRMI,
const MachineInstr &ESMI) const {
bool isLate = isLateResultInstr(LRMI);
bool isEarly = isEarlySourceInstr(ESMI);
DEBUG(dbgs() << "V60" << (isLate ? "-LR " : " -- "));
DEBUG(LRMI.dump());
DEBUG(dbgs() << "V60" << (isEarly ? "-ES " : " -- "));
DEBUG(ESMI.dump());
if (isLate && isEarly) {
DEBUG(dbgs() << "++Is Late Result feeding Early Source\n");
return true;
}
return false;
}
bool HexagonInstrInfo::isLateResultInstr(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
case TargetOpcode::EXTRACT_SUBREG:
case TargetOpcode::INSERT_SUBREG:
case TargetOpcode::SUBREG_TO_REG:
case TargetOpcode::REG_SEQUENCE:
case TargetOpcode::IMPLICIT_DEF:
case TargetOpcode::COPY:
case TargetOpcode::INLINEASM:
case TargetOpcode::PHI:
return false;
default:
break;
}
unsigned SchedClass = MI.getDesc().getSchedClass();
return !is_TC1(SchedClass);
}
bool HexagonInstrInfo::isLateSourceInstr(const MachineInstr &MI) const {
// Instructions with iclass A_CVI_VX and attribute A_CVI_LATE uses a multiply
// resource, but all operands can be received late like an ALU instruction.
return getType(MI) == HexagonII::TypeCVI_VX_LATE;
}
bool HexagonInstrInfo::isLoopN(const MachineInstr &MI) const {
unsigned Opcode = MI.getOpcode();
return Opcode == Hexagon::J2_loop0i ||
Opcode == Hexagon::J2_loop0r ||
Opcode == Hexagon::J2_loop0iext ||
Opcode == Hexagon::J2_loop0rext ||
Opcode == Hexagon::J2_loop1i ||
Opcode == Hexagon::J2_loop1r ||
Opcode == Hexagon::J2_loop1iext ||
Opcode == Hexagon::J2_loop1rext;
}
bool HexagonInstrInfo::isMemOp(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
default: return false;
case Hexagon::L4_iadd_memopw_io:
case Hexagon::L4_isub_memopw_io:
case Hexagon::L4_add_memopw_io:
case Hexagon::L4_sub_memopw_io:
case Hexagon::L4_and_memopw_io:
case Hexagon::L4_or_memopw_io:
case Hexagon::L4_iadd_memoph_io:
case Hexagon::L4_isub_memoph_io:
case Hexagon::L4_add_memoph_io:
case Hexagon::L4_sub_memoph_io:
case Hexagon::L4_and_memoph_io:
case Hexagon::L4_or_memoph_io:
case Hexagon::L4_iadd_memopb_io:
case Hexagon::L4_isub_memopb_io:
case Hexagon::L4_add_memopb_io:
case Hexagon::L4_sub_memopb_io:
case Hexagon::L4_and_memopb_io:
case Hexagon::L4_or_memopb_io:
case Hexagon::L4_ior_memopb_io:
case Hexagon::L4_ior_memoph_io:
case Hexagon::L4_ior_memopw_io:
case Hexagon::L4_iand_memopb_io:
case Hexagon::L4_iand_memoph_io:
case Hexagon::L4_iand_memopw_io:
return true;
}
return false;
}
bool HexagonInstrInfo::isNewValue(const MachineInstr &MI) const {
const uint64_t F = MI.getDesc().TSFlags;
return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask;
}
bool HexagonInstrInfo::isNewValue(unsigned Opcode) const {
const uint64_t F = get(Opcode).TSFlags;
return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask;
}
bool HexagonInstrInfo::isNewValueInst(const MachineInstr &MI) const {
return isNewValueJump(MI) || isNewValueStore(MI);
}
bool HexagonInstrInfo::isNewValueJump(const MachineInstr &MI) const {
return isNewValue(MI) && MI.isBranch();
}
bool HexagonInstrInfo::isNewValueJump(unsigned Opcode) const {
return isNewValue(Opcode) && get(Opcode).isBranch() && isPredicated(Opcode);
}
bool HexagonInstrInfo::isNewValueStore(const MachineInstr &MI) const {
const uint64_t F = MI.getDesc().TSFlags;
return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask;
}
bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const {
const uint64_t F = get(Opcode).TSFlags;
return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask;
}
// Returns true if a particular operand is extendable for an instruction.
bool HexagonInstrInfo::isOperandExtended(const MachineInstr &MI,
unsigned OperandNum) const {
const uint64_t F = MI.getDesc().TSFlags;
return ((F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask)
== OperandNum;
}
bool HexagonInstrInfo::isPredicatedNew(const MachineInstr &MI) const {
const uint64_t F = MI.getDesc().TSFlags;
assert(isPredicated(MI));
return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask;
}
bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const {
const uint64_t F = get(Opcode).TSFlags;
assert(isPredicated(Opcode));
return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask;
}
bool HexagonInstrInfo::isPredicatedTrue(const MachineInstr &MI) const {
const uint64_t F = MI.getDesc().TSFlags;
return !((F >> HexagonII::PredicatedFalsePos) &
HexagonII::PredicatedFalseMask);
}
bool HexagonInstrInfo::isPredicatedTrue(unsigned Opcode) const {
const uint64_t F = get(Opcode).TSFlags;
// Make sure that the instruction is predicated.
assert((F>> HexagonII::PredicatedPos) & HexagonII::PredicatedMask);
return !((F >> HexagonII::PredicatedFalsePos) &
HexagonII::PredicatedFalseMask);
}
bool HexagonInstrInfo::isPredicated(unsigned Opcode) const {
const uint64_t F = get(Opcode).TSFlags;
return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask;
}
bool HexagonInstrInfo::isPredicateLate(unsigned Opcode) const {
const uint64_t F = get(Opcode).TSFlags;
return ~(F >> HexagonII::PredicateLatePos) & HexagonII::PredicateLateMask;
}
bool HexagonInstrInfo::isPredictedTaken(unsigned Opcode) const {
const uint64_t F = get(Opcode).TSFlags;
assert(get(Opcode).isBranch() &&
(isPredicatedNew(Opcode) || isNewValue(Opcode)));
return (F >> HexagonII::TakenPos) & HexagonII::TakenMask;
}
bool HexagonInstrInfo::isSaveCalleeSavedRegsCall(const MachineInstr &MI) const {
return MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4 ||
MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT ||
MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_PIC ||
MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT_PIC;
}
bool HexagonInstrInfo::isSignExtendingLoad(const MachineInstr &MI) const {
switch (MI.getOpcode()) {
// Byte
case Hexagon::L2_loadrb_io:
case Hexagon::L4_loadrb_ur:
case Hexagon::L4_loadrb_ap:
case Hexagon::L2_loadrb_pr:
case Hexagon::L2_loadrb_pbr:
case Hexagon::L2_loadrb_pi:
case Hexagon::L2_loadrb_pci:
case Hexagon::L2_loadrb_pcr:
case Hexagon::L2_loadbsw2_io:
case Hexagon::L4_loadbsw2_ur:
case Hexagon::L4_loadbsw2_ap:
case Hexagon::L2_loadbsw2_pr:
case Hexagon::L2_loadbsw2_pbr:
case Hexagon::L2_loadbsw2_pi:
case Hexagon::L2_loadbsw2_pci:
case Hexagon::L2_loadbsw2_pcr:
case Hexagon::L2_loadbsw4_io:
case Hexagon::L4_loadbsw4_ur:
case Hexagon::L4_loadbsw4_ap:
case Hexagon::L2_loadbsw4_pr:
case Hexagon::L2_loadbsw4_pbr:
case Hexagon::L2_loadbsw4_pi:
case Hexagon::L2_loadbsw4_pci:
case Hexagon::L2_loadbsw4_pcr:
case Hexagon::L4_loadrb_rr:
case Hexagon::L2_ploadrbt_io:
case Hexagon::L2_ploadrbt_pi:
case Hexagon::L2_ploadrbf_io:
case Hexagon::L2_ploadrbf_pi:
case Hexagon::L2_ploadrbtnew_io:
case Hexagon::L2_ploadrbfnew_io:
case Hexagon::L4_ploadrbt_rr:
case Hexagon::L4_ploadrbf_rr:
case Hexagon::L4_ploadrbtnew_rr:
case Hexagon::L4_ploadrbfnew_rr:
case Hexagon::L2_ploadrbtnew_pi:
case Hexagon::L2_ploadrbfnew_pi:
case Hexagon::L4_ploadrbt_abs:
case Hexagon::L4_ploadrbf_abs:
case Hexagon::L4_ploadrbtnew_abs:
case Hexagon::L4_ploadrbfnew_abs:
case Hexagon::L2_loadrbgp:
// Half
case Hexagon::L2_loadrh_io:
case Hexagon::L4_loadrh_ur:
case Hexagon::L4_loadrh_ap:
case Hexagon::L2_loadrh_pr:
case Hexagon::L2_loadrh_pbr:
case Hexagon::L2_loadrh_pi:
case Hexagon::L2_loadrh_pci:
case Hexagon::L2_loadrh_pcr:
case Hexagon::L4_loadrh_rr:
case Hexagon::L2_ploadrht_io:
case Hexagon::L2_ploadrht_pi:
case Hexagon::L2_ploadrhf_io:
case Hexagon::L2_ploadrhf_pi:
case Hexagon::L2_ploadrhtnew_io:
case Hexagon::L2_ploadrhfnew_io:
case Hexagon::L4_ploadrht_rr:
case Hexagon::L4_ploadrhf_rr:
case Hexagon::L4_ploadrhtnew_rr: