llvm-project/llvm/lib/Target/ARM/ARMBaseInstrInfo.h

558 lines
24 KiB
C
Raw Normal View History

//===-- ARMBaseInstrInfo.h - ARM Base Instruction Information ---*- C++ -*-===//
//
// 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 Base ARM implementation of the TargetInstrInfo class.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H
#define LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H
#include "MCTargetDesc/ARMBaseInfo.h"
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/Target/TargetInstrInfo.h"
#include <array>
#include <cstdint>
#define GET_INSTRINFO_HEADER
#include "ARMGenInstrInfo.inc"
namespace llvm {
class ARMBaseRegisterInfo;
class ARMSubtarget;
class ARMBaseInstrInfo : public ARMGenInstrInfo {
const ARMSubtarget &Subtarget;
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
protected:
// Can be only subclassed.
explicit ARMBaseInstrInfo(const ARMSubtarget &STI);
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
void expandLoadStackGuardBase(MachineBasicBlock::iterator MI,
unsigned LoadImmOpc, unsigned LoadOpc) const;
/// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI
/// and \p DefIdx.
/// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of
/// the list is modeled as <Reg:SubReg, SubIdx>.
/// E.g., REG_SEQUENCE vreg1:sub1, sub0, vreg2, sub1 would produce
/// two elements:
/// - vreg1:sub1, sub0
/// - vreg2<:0>, sub1
///
/// \returns true if it is possible to build such an input sequence
/// with the pair \p MI, \p DefIdx. False otherwise.
///
/// \pre MI.isRegSequenceLike().
bool getRegSequenceLikeInputs(
const MachineInstr &MI, unsigned DefIdx,
SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const override;
/// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI
/// and \p DefIdx.
/// \p [out] InputReg of the equivalent EXTRACT_SUBREG.
/// E.g., EXTRACT_SUBREG vreg1:sub1, sub0, sub1 would produce:
/// - vreg1:sub1, sub0
///
/// \returns true if it is possible to build such an input sequence
/// with the pair \p MI, \p DefIdx. False otherwise.
///
/// \pre MI.isExtractSubregLike().
bool getExtractSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
RegSubRegPairAndIdx &InputReg) const override;
/// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI
/// and \p DefIdx.
/// \p [out] BaseReg and \p [out] InsertedReg contain
/// the equivalent inputs of INSERT_SUBREG.
/// E.g., INSERT_SUBREG vreg0:sub0, vreg1:sub1, sub3 would produce:
/// - BaseReg: vreg0:sub0
/// - InsertedReg: vreg1:sub1, sub3
///
/// \returns true if it is possible to build such an input sequence
/// with the pair \p MI, \p DefIdx. False otherwise.
///
/// \pre MI.isInsertSubregLike().
bool
getInsertSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
RegSubRegPair &BaseReg,
RegSubRegPairAndIdx &InsertedReg) const override;
/// Commutes the operands in the given instruction.
/// The commutable operands are specified by their indices OpIdx1 and OpIdx2.
///
/// Do not call this method for a non-commutable instruction or for
/// non-commutable pair of operand indices OpIdx1 and OpIdx2.
/// Even though the instruction is commutable, the method may still
/// fail to commute the operands, null pointer is returned in such cases.
MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
unsigned OpIdx1,
unsigned OpIdx2) const override;
public:
// Return whether the target has an explicit NOP encoding.
bool hasNOP() const;
// Return the non-pre/post incrementing version of 'Opc'. Return 0
// if there is not such an opcode.
virtual unsigned getUnindexedOpcode(unsigned Opc) const = 0;
MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
MachineInstr &MI,
LiveVariables *LV) const override;
virtual const ARMBaseRegisterInfo &getRegisterInfo() const = 0;
const ARMSubtarget &getSubtarget() const { return Subtarget; }
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
ScheduleHazardRecognizer *
CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
const ScheduleDAG *DAG) const override;
ScheduleHazardRecognizer *
CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
const ScheduleDAG *DAG) const override;
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
// Branch analysis.
bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
MachineBasicBlock *&FBB,
SmallVectorImpl<MachineOperand> &Cond,
bool AllowModify = false) const override;
unsigned removeBranch(MachineBasicBlock &MBB,
int *BytesRemoved = nullptr) const override;
unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
const DebugLoc &DL,
int *BytesAdded = nullptr) const override;
bool
reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const override;
// Predication support.
bool isPredicated(const MachineInstr &MI) const override;
ARMCC::CondCodes getPredicate(const MachineInstr &MI) const {
int PIdx = MI.findFirstPredOperandIdx();
return PIdx != -1 ? (ARMCC::CondCodes)MI.getOperand(PIdx).getImm()
: ARMCC::AL;
}
bool PredicateInstruction(MachineInstr &MI,
ArrayRef<MachineOperand> Pred) const override;
bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
ArrayRef<MachineOperand> Pred2) const override;
bool DefinesPredicate(MachineInstr &MI,
std::vector<MachineOperand> &Pred) const override;
bool isPredicable(const MachineInstr &MI) const override;
// CPSR defined in instruction
static bool isCPSRDefined(const MachineInstr &MI);
bool isAddrMode3OpImm(const MachineInstr &MI, unsigned Op) const;
bool isAddrMode3OpMinusReg(const MachineInstr &MI, unsigned Op) const;
// Load, scaled register offset
bool isLdstScaledReg(const MachineInstr &MI, unsigned Op) const;
// Load, scaled register offset, not plus LSL2
bool isLdstScaledRegNotPlusLsl2(const MachineInstr &MI, unsigned Op) const;
// Minus reg for ldstso addr mode
bool isLdstSoMinusReg(const MachineInstr &MI, unsigned Op) const;
// Scaled register offset in address mode 2
bool isAm2ScaledReg(const MachineInstr &MI, unsigned Op) const;
// Load multiple, base reg in list
bool isLDMBaseRegInList(const MachineInstr &MI) const;
// get LDM variable defs size
unsigned getLDMVariableDefsSize(const MachineInstr &MI) const;
/// GetInstSize - Returns the size of the specified MachineInstr.
///
unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
unsigned isLoadFromStackSlot(const MachineInstr &MI,
int &FrameIndex) const override;
unsigned isStoreToStackSlot(const MachineInstr &MI,
int &FrameIndex) const override;
unsigned isLoadFromStackSlotPostFE(const MachineInstr &MI,
int &FrameIndex) const override;
unsigned isStoreToStackSlotPostFE(const MachineInstr &MI,
int &FrameIndex) const override;
void copyToCPSR(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
unsigned SrcReg, bool KillSrc,
const ARMSubtarget &Subtarget) const;
void copyFromCPSR(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
unsigned DestReg, bool KillSrc,
const ARMSubtarget &Subtarget) const;
void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
const DebugLoc &DL, unsigned DestReg, unsigned SrcReg,
bool KillSrc) const override;
void storeRegToStackSlot(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI,
unsigned SrcReg, bool isKill, int FrameIndex,
const TargetRegisterClass *RC,
const TargetRegisterInfo *TRI) const override;
void loadRegFromStackSlot(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI,
unsigned DestReg, int FrameIndex,
const TargetRegisterClass *RC,
const TargetRegisterInfo *TRI) const override;
bool expandPostRAPseudo(MachineInstr &MI) const override;
void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
unsigned DestReg, unsigned SubIdx,
const MachineInstr &Orig,
const TargetRegisterInfo &TRI) const override;
2009-11-08 08:15:23 +08:00
MachineInstr *duplicate(MachineInstr &Orig,
MachineFunction &MF) const override;
const MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
unsigned SubIdx, unsigned State,
const TargetRegisterInfo *TRI) const;
bool produceSameValue(const MachineInstr &MI0, const MachineInstr &MI1,
const MachineRegisterInfo *MRI) const override;
/// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
/// determine if two loads are loading from the same base address. It should
/// only return true if the base pointers are the same and the only
/// differences between the two addresses is the offset. It also returns the
/// offsets by reference.
bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, int64_t &Offset1,
int64_t &Offset2) const override;
/// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
/// determine (in conjunction with areLoadsFromSameBasePtr) if two loads
/// should be scheduled togther. On some targets if two loads are loading from
/// addresses in the same cache line, it's better if they are scheduled
/// together. This function takes two integers that represent the load offsets
/// from the common base address. It returns true if it decides it's desirable
/// to schedule the two loads together. "NumLoads" is the number of loads that
/// have already been scheduled after Load1.
bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
int64_t Offset1, int64_t Offset2,
unsigned NumLoads) const override;
bool isSchedulingBoundary(const MachineInstr &MI,
const MachineBasicBlock *MBB,
const MachineFunction &MF) const override;
bool isProfitableToIfCvt(MachineBasicBlock &MBB,
unsigned NumCycles, unsigned ExtraPredCycles,
BranchProbability Probability) const override;
bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
unsigned ExtraT, MachineBasicBlock &FMBB,
unsigned NumF, unsigned ExtraF,
BranchProbability Probability) const override;
bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
BranchProbability Probability) const override {
2011-04-13 14:39:16 +08:00
return NumCycles == 1;
}
bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
MachineBasicBlock &FMBB) const override;
/// analyzeCompare - For a comparison instruction, return the source registers
/// in SrcReg and SrcReg2 if having two register operands, and the value it
/// compares against in CmpValue. Return true if the comparison instruction
/// can be analyzed.
bool analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
unsigned &SrcReg2, int &CmpMask,
int &CmpValue) const override;
/// optimizeCompareInstr - Convert the instruction to set the zero flag so
/// that we can remove a "comparison with zero"; Remove a redundant CMP
/// instruction if the flags can be updated in the same way by an earlier
/// instruction such as SUB.
bool optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
unsigned SrcReg2, int CmpMask, int CmpValue,
const MachineRegisterInfo *MRI) const override;
bool analyzeSelect(const MachineInstr &MI,
SmallVectorImpl<MachineOperand> &Cond, unsigned &TrueOp,
unsigned &FalseOp, bool &Optimizable) const override;
MachineInstr *optimizeSelect(MachineInstr &MI,
SmallPtrSetImpl<MachineInstr *> &SeenMIs,
bool) const override;
/// FoldImmediate - 'Reg' is known to be defined by a move immediate
/// instruction, try to fold the immediate into the use instruction.
bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, unsigned Reg,
MachineRegisterInfo *MRI) const override;
unsigned getNumMicroOps(const InstrItineraryData *ItinData,
const MachineInstr &MI) const override;
int getOperandLatency(const InstrItineraryData *ItinData,
const MachineInstr &DefMI, unsigned DefIdx,
const MachineInstr &UseMI,
unsigned UseIdx) const override;
int getOperandLatency(const InstrItineraryData *ItinData,
SDNode *DefNode, unsigned DefIdx,
SDNode *UseNode, unsigned UseIdx) const override;
/// VFP/NEON execution domains.
std::pair<uint16_t, uint16_t>
getExecutionDomain(const MachineInstr &MI) const override;
void setExecutionDomain(MachineInstr &MI, unsigned Domain) const override;
unsigned
getPartialRegUpdateClearance(const MachineInstr &, unsigned,
const TargetRegisterInfo *) const override;
void breakPartialRegDependency(MachineInstr &, unsigned,
const TargetRegisterInfo *TRI) const override;
/// Get the number of addresses by LDM or VLDM or zero for unknown.
unsigned getNumLDMAddresses(const MachineInstr &MI) const;
private:
unsigned getInstBundleLength(const MachineInstr &MI) const;
2010-10-08 07:12:15 +08:00
int getVLDMDefCycle(const InstrItineraryData *ItinData,
const MCInstrDesc &DefMCID,
2010-10-08 07:12:15 +08:00
unsigned DefClass,
unsigned DefIdx, unsigned DefAlign) const;
int getLDMDefCycle(const InstrItineraryData *ItinData,
const MCInstrDesc &DefMCID,
2010-10-08 07:12:15 +08:00
unsigned DefClass,
unsigned DefIdx, unsigned DefAlign) const;
int getVSTMUseCycle(const InstrItineraryData *ItinData,
const MCInstrDesc &UseMCID,
2010-10-08 07:12:15 +08:00
unsigned UseClass,
unsigned UseIdx, unsigned UseAlign) const;
int getSTMUseCycle(const InstrItineraryData *ItinData,
const MCInstrDesc &UseMCID,
2010-10-08 07:12:15 +08:00
unsigned UseClass,
unsigned UseIdx, unsigned UseAlign) const;
int getOperandLatency(const InstrItineraryData *ItinData,
const MCInstrDesc &DefMCID,
unsigned DefIdx, unsigned DefAlign,
const MCInstrDesc &UseMCID,
unsigned UseIdx, unsigned UseAlign) const;
int getOperandLatencyImpl(const InstrItineraryData *ItinData,
const MachineInstr &DefMI, unsigned DefIdx,
const MCInstrDesc &DefMCID, unsigned DefAdj,
const MachineOperand &DefMO, unsigned Reg,
const MachineInstr &UseMI, unsigned UseIdx,
const MCInstrDesc &UseMCID, unsigned UseAdj) const;
unsigned getPredicationCost(const MachineInstr &MI) const override;
unsigned getInstrLatency(const InstrItineraryData *ItinData,
const MachineInstr &MI,
2014-04-28 12:05:08 +08:00
unsigned *PredCost = nullptr) const override;
int getInstrLatency(const InstrItineraryData *ItinData,
SDNode *Node) const override;
bool hasHighOperandLatency(const TargetSchedModel &SchedModel,
const MachineRegisterInfo *MRI,
const MachineInstr &DefMI, unsigned DefIdx,
const MachineInstr &UseMI,
unsigned UseIdx) const override;
bool hasLowDefLatency(const TargetSchedModel &SchedModel,
const MachineInstr &DefMI,
unsigned DefIdx) const override;
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
/// verifyInstruction - Perform target specific instruction verification.
bool verifyInstruction(const MachineInstr &MI,
StringRef &ErrInfo) const override;
virtual void expandLoadStackGuard(MachineBasicBlock::iterator MI) const = 0;
void expandMEMCPY(MachineBasicBlock::iterator) const;
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
private:
/// Modeling special VFP / NEON fp MLA / MLS hazards.
/// MLxEntryMap - Map fp MLA / MLS to the corresponding entry in the internal
/// MLx table.
DenseMap<unsigned, unsigned> MLxEntryMap;
/// MLxHazardOpcodes - Set of add / sub and multiply opcodes that would cause
/// stalls when scheduled together with fp MLA / MLS opcodes.
SmallSet<unsigned, 16> MLxHazardOpcodes;
public:
/// isFpMLxInstruction - Return true if the specified opcode is a fp MLA / MLS
/// instruction.
bool isFpMLxInstruction(unsigned Opcode) const {
return MLxEntryMap.count(Opcode);
}
/// isFpMLxInstruction - This version also returns the multiply opcode and the
/// addition / subtraction opcode to expand to. Return true for 'HasLane' for
/// the MLX instructions with an extra lane operand.
bool isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
unsigned &AddSubOpc, bool &NegAcc,
bool &HasLane) const;
/// canCauseFpMLxStall - Return true if an instruction of the specified opcode
/// will cause stalls when scheduled after (within 4-cycle window) a fp
/// MLA / MLS instruction.
bool canCauseFpMLxStall(unsigned Opcode) const {
return MLxHazardOpcodes.count(Opcode);
}
/// Returns true if the instruction has a shift by immediate that can be
/// executed in one cycle less.
bool isSwiftFastImmShift(const MachineInstr *MI) const;
/// Returns predicate register associated with the given frame instruction.
unsigned getFramePred(const MachineInstr &MI) const {
assert(isFrameInstr(MI));
Add extra operand to CALLSEQ_START to keep frame part set up previously Using arguments with attribute inalloca creates problems for verification of machine representation. This attribute instructs the backend that the argument is prepared in stack prior to CALLSEQ_START..CALLSEQ_END sequence (see http://llvm.org/docs/InAlloca.htm for details). Frame size stored in CALLSEQ_START in this case does not count the size of this argument. However CALLSEQ_END still keeps total frame size, as caller can be responsible for cleanup of entire frame. So CALLSEQ_START and CALLSEQ_END keep different frame size and the difference is treated by MachineVerifier as stack error. Currently there is no way to distinguish this case from actual errors. This patch adds additional argument to CALLSEQ_START and its target-specific counterparts to keep size of stack that is set up prior to the call frame sequence. This argument allows MachineVerifier to calculate actual frame size associated with frame setup instruction and correctly process the case of inalloca arguments. The changes made by the patch are: - Frame setup instructions get the second mandatory argument. It affects all targets that use frame pseudo instructions and touched many files although the changes are uniform. - Access to frame properties are implemented using special instructions rather than calls getOperand(N).getImm(). For X86 and ARM such replacement was made previously. - Changes that reflect appearance of additional argument of frame setup instruction. These involve proper instruction initialization and methods that access instruction arguments. - MachineVerifier retrieves frame size using method, which reports sum of frame parts initialized inside frame instruction pair and outside it. The patch implements approach proposed by Quentin Colombet in https://bugs.llvm.org/show_bug.cgi?id=27481#c1. It fixes 9 tests failed with machine verifier enabled and listed in PR27481. Differential Revision: https://reviews.llvm.org/D32394 llvm-svn: 302527
2017-05-09 21:35:13 +08:00
// Operands of ADJCALLSTACKDOWN/ADJCALLSTACKUP:
// - argument declared in the pattern:
// 0 - frame size
Add extra operand to CALLSEQ_START to keep frame part set up previously Using arguments with attribute inalloca creates problems for verification of machine representation. This attribute instructs the backend that the argument is prepared in stack prior to CALLSEQ_START..CALLSEQ_END sequence (see http://llvm.org/docs/InAlloca.htm for details). Frame size stored in CALLSEQ_START in this case does not count the size of this argument. However CALLSEQ_END still keeps total frame size, as caller can be responsible for cleanup of entire frame. So CALLSEQ_START and CALLSEQ_END keep different frame size and the difference is treated by MachineVerifier as stack error. Currently there is no way to distinguish this case from actual errors. This patch adds additional argument to CALLSEQ_START and its target-specific counterparts to keep size of stack that is set up prior to the call frame sequence. This argument allows MachineVerifier to calculate actual frame size associated with frame setup instruction and correctly process the case of inalloca arguments. The changes made by the patch are: - Frame setup instructions get the second mandatory argument. It affects all targets that use frame pseudo instructions and touched many files although the changes are uniform. - Access to frame properties are implemented using special instructions rather than calls getOperand(N).getImm(). For X86 and ARM such replacement was made previously. - Changes that reflect appearance of additional argument of frame setup instruction. These involve proper instruction initialization and methods that access instruction arguments. - MachineVerifier retrieves frame size using method, which reports sum of frame parts initialized inside frame instruction pair and outside it. The patch implements approach proposed by Quentin Colombet in https://bugs.llvm.org/show_bug.cgi?id=27481#c1. It fixes 9 tests failed with machine verifier enabled and listed in PR27481. Differential Revision: https://reviews.llvm.org/D32394 llvm-svn: 302527
2017-05-09 21:35:13 +08:00
// 1 - arg of CALLSEQ_START/CALLSEQ_END
// 2 - predicate code (like ARMCC::AL)
// - added by predOps:
// 3 - predicate reg
return MI.getOperand(3).getReg();
}
};
/// Get the operands corresponding to the given \p Pred value. By default, the
/// predicate register is assumed to be 0 (no register), but you can pass in a
/// \p PredReg if that is not the case.
static inline std::array<MachineOperand, 2> predOps(ARMCC::CondCodes Pred,
unsigned PredReg = 0) {
return {{MachineOperand::CreateImm(static_cast<int64_t>(Pred)),
MachineOperand::CreateReg(PredReg, false)}};
}
/// Get the operand corresponding to the conditional code result. By default,
/// this is 0 (no register).
static inline MachineOperand condCodeOp(unsigned CCReg = 0) {
return MachineOperand::CreateReg(CCReg, false);
}
2009-07-28 02:25:24 +08:00
/// Get the operand corresponding to the conditional code result for Thumb1.
/// This operand will always refer to CPSR and it will have the Define flag set.
/// You can optionally set the Dead flag by means of \p isDead.
static inline MachineOperand t1CondCodeOp(bool isDead = false) {
return MachineOperand::CreateReg(ARM::CPSR,
/*Define*/ true, /*Implicit*/ false,
/*Kill*/ false, isDead);
}
static inline
bool isUncondBranchOpcode(int Opc) {
return Opc == ARM::B || Opc == ARM::tB || Opc == ARM::t2B;
}
static inline
bool isCondBranchOpcode(int Opc) {
return Opc == ARM::Bcc || Opc == ARM::tBcc || Opc == ARM::t2Bcc;
}
static inline
bool isJumpTableBranchOpcode(int Opc) {
return Opc == ARM::BR_JTr || Opc == ARM::BR_JTm || Opc == ARM::BR_JTadd ||
Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT;
}
static inline
bool isIndirectBranchOpcode(int Opc) {
return Opc == ARM::BX || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND;
}
static inline bool isPopOpcode(int Opc) {
return Opc == ARM::tPOP_RET || Opc == ARM::LDMIA_RET ||
Opc == ARM::t2LDMIA_RET || Opc == ARM::tPOP || Opc == ARM::LDMIA_UPD ||
Opc == ARM::t2LDMIA_UPD || Opc == ARM::VLDMDIA_UPD;
}
static inline bool isPushOpcode(int Opc) {
return Opc == ARM::tPUSH || Opc == ARM::t2STMDB_UPD ||
Opc == ARM::STMDB_UPD || Opc == ARM::VSTMDDB_UPD;
}
/// getInstrPredicate - If instruction is predicated, returns its predicate
/// condition, otherwise returns AL. It also returns the condition code
/// register by reference.
ARMCC::CondCodes getInstrPredicate(const MachineInstr &MI, unsigned &PredReg);
unsigned getMatchingCondBranchOpcode(unsigned Opc);
/// Determine if MI can be folded into an ARM MOVCC instruction, and return the
/// opcode of the SSA instruction representing the conditional MI.
unsigned canFoldARMInstrIntoMOVCC(unsigned Reg,
MachineInstr *&MI,
const MachineRegisterInfo &MRI);
/// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether
/// the instruction is encoded with an 'S' bit is determined by the optional
/// CPSR def operand.
unsigned convertAddSubFlagsOpcode(unsigned OldOpc);
/// emitARMRegPlusImmediate / emitT2RegPlusImmediate - Emits a series of
/// instructions to materializea destreg = basereg + immediate in ARM / Thumb2
/// code.
void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI,
const DebugLoc &dl, unsigned DestReg,
unsigned BaseReg, int NumBytes,
ARMCC::CondCodes Pred, unsigned PredReg,
const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
void emitT2RegPlusImmediate(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI,
const DebugLoc &dl, unsigned DestReg,
unsigned BaseReg, int NumBytes,
ARMCC::CondCodes Pred, unsigned PredReg,
const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI,
const DebugLoc &dl, unsigned DestReg,
unsigned BaseReg, int NumBytes,
const TargetInstrInfo &TII,
const ARMBaseRegisterInfo &MRI,
unsigned MIFlags = 0);
/// Tries to add registers to the reglist of a given base-updating
/// push/pop instruction to adjust the stack by an additional
/// NumBytes. This can save a few bytes per function in code-size, but
/// obviously generates more memory traffic. As such, it only takes
/// effect in functions being optimised for size.
bool tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget,
MachineFunction &MF, MachineInstr *MI,
unsigned NumBytes);
/// rewriteARMFrameIndex / rewriteT2FrameIndex -
/// Rewrite MI to access 'Offset' bytes from the FP. Return false if the
/// offset could not be handled directly in MI, and return the left-over
/// portion by reference.
bool rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
unsigned FrameReg, int &Offset,
const ARMBaseInstrInfo &TII);
bool rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
unsigned FrameReg, int &Offset,
const ARMBaseInstrInfo &TII);
} // end namespace llvm
#endif // LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H