Google Git
Sign in
llvm / llvm-project / bf317f66989cac26e17b4cd16ab1c7bdfe73dbe0 / . / llvm / lib / Target / X86 / X86AvoidStoreForwardingBlocks.cpp
blob: 04931afdec51c3e4dee3ee43da7a1a6a0efb116d [file] [log] [blame]
//===- X86AvoidStoreForwardingBlocks.cpp - Avoid HW Store Forward Block ---===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// If a load follows a store and reloads data that the store has written to
// memory, Intel microarchitectures can in many cases forward the data directly
// from the store to the load, This "store forwarding" saves cycles by enabling
// the load to directly obtain the data instead of accessing the data from
// cache or memory.
// A "store forward block" occurs in cases that a store cannot be forwarded to
// the load. The most typical case of store forward block on Intel Core
// microarchitecture that a small store cannot be forwarded to a large load.
// The estimated penalty for a store forward block is ~13 cycles.
//
// This pass tries to recognize and handle cases where "store forward block"
// is created by the compiler when lowering memcpy calls to a sequence
// of a load and a store.
//
// The pass currently only handles cases where memcpy is lowered to
// XMM/YMM registers, it tries to break the memcpy into smaller copies.
// breaking the memcpy should be possible since there is no atomicity
// guarantee for loads and stores to XMM/YMM.
//
// It could be better for performance to solve the problem by loading
// to XMM/YMM then inserting the partial store before storing back from XMM/YMM
// to memory, but this will result in a more conservative optimization since it
// requires we prove that all memory accesses between the blocking store and the
// load must alias/don't alias before we can move the store, whereas the
// transformation done here is correct regardless to other memory accesses.
//===----------------------------------------------------------------------===//
#include "X86.h"
#include "X86InstrInfo.h"
#include "X86Subtarget.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/Function.h"
#include "llvm/InitializePasses.h"
#include "llvm/MC/MCInstrDesc.h"
using namespace llvm;
#define DEBUG_TYPE "x86-avoid-SFB"
static cl::opt<bool> DisableX86AvoidStoreForwardBlocks(
"x86-disable-avoid-SFB", cl::Hidden,
cl::desc("X86: Disable Store Forwarding Blocks fixup."), cl::init(false));
static cl::opt<unsigned> X86AvoidSFBInspectionLimit(
"x86-sfb-inspection-limit",
cl::desc("X86: Number of instructions backward to "
"inspect for store forwarding blocks."),
cl::init(20), cl::Hidden);
namespace {
using DisplacementSizeMap = std::map<int64_t, unsigned>;
class X86AvoidSFBPass : public MachineFunctionPass {
public:
static char ID;
X86AvoidSFBPass() : MachineFunctionPass(ID) { }
StringRef getPassName() const override {
return "X86 Avoid Store Forwarding Blocks";
}
bool runOnMachineFunction(MachineFunction &MF) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
MachineFunctionPass::getAnalysisUsage(AU);
AU.addRequired<AAResultsWrapperPass>();
}
private:
MachineRegisterInfo *MRI = nullptr;
const X86InstrInfo *TII = nullptr;
const X86RegisterInfo *TRI = nullptr;
SmallVector<std::pair<MachineInstr *, MachineInstr *>, 2>
BlockedLoadsStoresPairs;
SmallVector<MachineInstr *, 2> ForRemoval;
AliasAnalysis *AA = nullptr;
/// Returns couples of Load then Store to memory which look
/// like a memcpy.
void findPotentiallylBlockedCopies(MachineFunction &MF);
/// Break the memcpy's load and store into smaller copies
/// such that each memory load that was blocked by a smaller store
/// would now be copied separately.
void breakBlockedCopies(MachineInstr *LoadInst, MachineInstr *StoreInst,
const DisplacementSizeMap &BlockingStoresDispSizeMap);
/// Break a copy of size Size to smaller copies.
void buildCopies(int Size, MachineInstr *LoadInst, int64_t LdDispImm,
MachineInstr *StoreInst, int64_t StDispImm,
int64_t LMMOffset, int64_t SMMOffset);
void buildCopy(MachineInstr *LoadInst, unsigned NLoadOpcode, int64_t LoadDisp,
MachineInstr *StoreInst, unsigned NStoreOpcode,
int64_t StoreDisp, unsigned Size, int64_t LMMOffset,
int64_t SMMOffset);
bool alias(const MachineMemOperand &Op1, const MachineMemOperand &Op2) const;
unsigned getRegSizeInBytes(MachineInstr *Inst);
};
} // end anonymous namespace
char X86AvoidSFBPass::ID = 0;
INITIALIZE_PASS_BEGIN(X86AvoidSFBPass, DEBUG_TYPE, "Machine code sinking",
false, false)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
INITIALIZE_PASS_END(X86AvoidSFBPass, DEBUG_TYPE, "Machine code sinking", false,
false)
FunctionPass *llvm::createX86AvoidStoreForwardingBlocks() {
return new X86AvoidSFBPass();
}
static bool isXMMLoadOpcode(unsigned Opcode) {
return Opcode == X86::MOVUPSrm || Opcode == X86::MOVAPSrm ||
Opcode == X86::VMOVUPSrm || Opcode == X86::VMOVAPSrm ||
Opcode == X86::VMOVUPDrm || Opcode == X86::VMOVAPDrm ||
Opcode == X86::VMOVDQUrm || Opcode == X86::VMOVDQArm ||
Opcode == X86::VMOVUPSZ128rm || Opcode == X86::VMOVAPSZ128rm ||
Opcode == X86::VMOVUPDZ128rm || Opcode == X86::VMOVAPDZ128rm ||
Opcode == X86::VMOVDQU64Z128rm || Opcode == X86::VMOVDQA64Z128rm ||
Opcode == X86::VMOVDQU32Z128rm || Opcode == X86::VMOVDQA32Z128rm;
}
static bool isYMMLoadOpcode(unsigned Opcode) {
return Opcode == X86::VMOVUPSYrm || Opcode == X86::VMOVAPSYrm ||
Opcode == X86::VMOVUPDYrm || Opcode == X86::VMOVAPDYrm ||
Opcode == X86::VMOVDQUYrm || Opcode == X86::VMOVDQAYrm ||
Opcode == X86::VMOVUPSZ256rm || Opcode == X86::VMOVAPSZ256rm ||
Opcode == X86::VMOVUPDZ256rm || Opcode == X86::VMOVAPDZ256rm ||
Opcode == X86::VMOVDQU64Z256rm || Opcode == X86::VMOVDQA64Z256rm ||
Opcode == X86::VMOVDQU32Z256rm || Opcode == X86::VMOVDQA32Z256rm;
}
static bool isPotentialBlockedMemCpyLd(unsigned Opcode) {
return isXMMLoadOpcode(Opcode) || isYMMLoadOpcode(Opcode);
}
static bool isPotentialBlockedMemCpyPair(unsigned LdOpcode, unsigned StOpcode) {
switch (LdOpcode) {
case X86::MOVUPSrm:
case X86::MOVAPSrm:
return StOpcode == X86::MOVUPSmr || StOpcode == X86::MOVAPSmr;
case X86::VMOVUPSrm:
case X86::VMOVAPSrm:
return StOpcode == X86::VMOVUPSmr || StOpcode == X86::VMOVAPSmr;
case X86::VMOVUPDrm:
case X86::VMOVAPDrm:
return StOpcode == X86::VMOVUPDmr || StOpcode == X86::VMOVAPDmr;
case X86::VMOVDQUrm:
case X86::VMOVDQArm:
return StOpcode == X86::VMOVDQUmr || StOpcode == X86::VMOVDQAmr;
case X86::VMOVUPSZ128rm:
case X86::VMOVAPSZ128rm:
return StOpcode == X86::VMOVUPSZ128mr || StOpcode == X86::VMOVAPSZ128mr;
case X86::VMOVUPDZ128rm:
case X86::VMOVAPDZ128rm:
return StOpcode == X86::VMOVUPDZ128mr || StOpcode == X86::VMOVAPDZ128mr;
case X86::VMOVUPSYrm:
case X86::VMOVAPSYrm:
return StOpcode == X86::VMOVUPSYmr || StOpcode == X86::VMOVAPSYmr;
case X86::VMOVUPDYrm:
case X86::VMOVAPDYrm:
return StOpcode == X86::VMOVUPDYmr || StOpcode == X86::VMOVAPDYmr;
case X86::VMOVDQUYrm:
case X86::VMOVDQAYrm:
return StOpcode == X86::VMOVDQUYmr || StOpcode == X86::VMOVDQAYmr;
case X86::VMOVUPSZ256rm:
case X86::VMOVAPSZ256rm:
return StOpcode == X86::VMOVUPSZ256mr || StOpcode == X86::VMOVAPSZ256mr;
case X86::VMOVUPDZ256rm:
case X86::VMOVAPDZ256rm:
return StOpcode == X86::VMOVUPDZ256mr || StOpcode == X86::VMOVAPDZ256mr;
case X86::VMOVDQU64Z128rm:
case X86::VMOVDQA64Z128rm:
return StOpcode == X86::VMOVDQU64Z128mr || StOpcode == X86::VMOVDQA64Z128mr;
case X86::VMOVDQU32Z128rm:
case X86::VMOVDQA32Z128rm:
return StOpcode == X86::VMOVDQU32Z128mr || StOpcode == X86::VMOVDQA32Z128mr;
case X86::VMOVDQU64Z256rm:
case X86::VMOVDQA64Z256rm:
return StOpcode == X86::VMOVDQU64Z256mr || StOpcode == X86::VMOVDQA64Z256mr;
case X86::VMOVDQU32Z256rm:
case X86::VMOVDQA32Z256rm:
return StOpcode == X86::VMOVDQU32Z256mr || StOpcode == X86::VMOVDQA32Z256mr;
default:
return false;
}
}
static bool isPotentialBlockingStoreInst(unsigned Opcode, unsigned LoadOpcode) {
bool PBlock = false;
PBlock |= Opcode == X86::MOV64mr || Opcode == X86::MOV64mi32 ||
Opcode == X86::MOV32mr || Opcode == X86::MOV32mi ||
Opcode == X86::MOV16mr || Opcode == X86::MOV16mi ||
Opcode == X86::MOV8mr || Opcode == X86::MOV8mi;
if (isYMMLoadOpcode(LoadOpcode))
PBlock |= Opcode == X86::VMOVUPSmr || Opcode == X86::VMOVAPSmr ||
Opcode == X86::VMOVUPDmr || Opcode == X86::VMOVAPDmr ||
Opcode == X86::VMOVDQUmr || Opcode == X86::VMOVDQAmr ||
Opcode == X86::VMOVUPSZ128mr || Opcode == X86::VMOVAPSZ128mr ||
Opcode == X86::VMOVUPDZ128mr || Opcode == X86::VMOVAPDZ128mr ||
Opcode == X86::VMOVDQU64Z128mr ||
Opcode == X86::VMOVDQA64Z128mr ||
Opcode == X86::VMOVDQU32Z128mr || Opcode == X86::VMOVDQA32Z128mr;
return PBlock;
}
static const int MOV128SZ = 16;
static const int MOV64SZ = 8;
static const int MOV32SZ = 4;
static const int MOV16SZ = 2;
static const int MOV8SZ = 1;
static unsigned getYMMtoXMMLoadOpcode(unsigned LoadOpcode) {
switch (LoadOpcode) {
case X86::VMOVUPSYrm:
case X86::VMOVAPSYrm:
return X86::VMOVUPSrm;
case X86::VMOVUPDYrm:
case X86::VMOVAPDYrm:
return X86::VMOVUPDrm;
case X86::VMOVDQUYrm:
case X86::VMOVDQAYrm:
return X86::VMOVDQUrm;
case X86::VMOVUPSZ256rm:
case X86::VMOVAPSZ256rm:
return X86::VMOVUPSZ128rm;
case X86::VMOVUPDZ256rm:
case X86::VMOVAPDZ256rm:
return X86::VMOVUPDZ128rm;
case X86::VMOVDQU64Z256rm:
case X86::VMOVDQA64Z256rm:
return X86::VMOVDQU64Z128rm;
case X86::VMOVDQU32Z256rm:
case X86::VMOVDQA32Z256rm:
return X86::VMOVDQU32Z128rm;
default:
llvm_unreachable("Unexpected Load Instruction Opcode");
}
return 0;
}
static unsigned getYMMtoXMMStoreOpcode(unsigned StoreOpcode) {
switch (StoreOpcode) {
case X86::VMOVUPSYmr:
case X86::VMOVAPSYmr:
return X86::VMOVUPSmr;
case X86::VMOVUPDYmr:
case X86::VMOVAPDYmr:
return X86::VMOVUPDmr;
case X86::VMOVDQUYmr:
case X86::VMOVDQAYmr:
return X86::VMOVDQUmr;
case X86::VMOVUPSZ256mr:
case X86::VMOVAPSZ256mr:
return X86::VMOVUPSZ128mr;
case X86::VMOVUPDZ256mr:
case X86::VMOVAPDZ256mr:
return X86::VMOVUPDZ128mr;
case X86::VMOVDQU64Z256mr:
case X86::VMOVDQA64Z256mr:
return X86::VMOVDQU64Z128mr;
case X86::VMOVDQU32Z256mr:
case X86::VMOVDQA32Z256mr:
return X86::VMOVDQU32Z128mr;
default:
llvm_unreachable("Unexpected Load Instruction Opcode");
}
return 0;
}
static int getAddrOffset(const MachineInstr *MI) {
const MCInstrDesc &Descl = MI->getDesc();
int AddrOffset = X86II::getMemoryOperandNo(Descl.TSFlags);
assert(AddrOffset != -1 && "Expected Memory Operand");
AddrOffset += X86II::getOperandBias(Descl);
return AddrOffset;
}
static MachineOperand &getBaseOperand(MachineInstr *MI) {
int AddrOffset = getAddrOffset(MI);
return MI->getOperand(AddrOffset + X86::AddrBaseReg);
}
static MachineOperand &getDispOperand(MachineInstr *MI) {
int AddrOffset = getAddrOffset(MI);
return MI->getOperand(AddrOffset + X86::AddrDisp);
}
// Relevant addressing modes contain only base register and immediate
// displacement or frameindex and immediate displacement.
// TODO: Consider expanding to other addressing modes in the future
static bool isRelevantAddressingMode(MachineInstr *MI) {
int AddrOffset = getAddrOffset(MI);
const MachineOperand &Base = getBaseOperand(MI);
const MachineOperand &Disp = getDispOperand(MI);
const MachineOperand &Scale = MI->getOperand(AddrOffset + X86::AddrScaleAmt);
const MachineOperand &Index = MI->getOperand(AddrOffset + X86::AddrIndexReg);
const MachineOperand &Segment = MI->getOperand(AddrOffset + X86::AddrSegmentReg);
if (!((Base.isReg() && Base.getReg() != X86::NoRegister) || Base.isFI()))
return false;
if (!Disp.isImm())
return false;
if (Scale.getImm() != 1)
return false;
if (!(Index.isReg() && Index.getReg() == X86::NoRegister))
return false;
if (!(Segment.isReg() && Segment.getReg() == X86::NoRegister))
return false;
return true;
}
// Collect potentially blocking stores.
// Limit the number of instructions backwards we want to inspect
// since the effect of store block won't be visible if the store
// and load instructions have enough instructions in between to
// keep the core busy.
static SmallVector<MachineInstr *, 2>
findPotentialBlockers(MachineInstr *LoadInst) {
SmallVector<MachineInstr *, 2> PotentialBlockers;
unsigned BlockCount = 0;
const unsigned InspectionLimit = X86AvoidSFBInspectionLimit;
for (auto PBInst = std::next(MachineBasicBlock::reverse_iterator(LoadInst)),
E = LoadInst->getParent()->rend();
PBInst != E; ++PBInst) {
if (PBInst->isMetaInstruction())
continue;
BlockCount++;
if (BlockCount >= InspectionLimit)
break;
MachineInstr &MI = *PBInst;
if (MI.getDesc().isCall())
return PotentialBlockers;
PotentialBlockers.push_back(&MI);
}
// If we didn't get to the instructions limit try predecessing blocks.
// Ideally we should traverse the predecessor blocks in depth with some
// coloring algorithm, but for now let's just look at the first order
// predecessors.
if (BlockCount < InspectionLimit) {
MachineBasicBlock *MBB = LoadInst->getParent();
int LimitLeft = InspectionLimit - BlockCount;
for (MachineBasicBlock *PMBB : MBB->predecessors()) {
int PredCount = 0;
for (MachineInstr &PBInst : llvm::reverse(*PMBB)) {
if (PBInst.isMetaInstruction())
continue;
PredCount++;
if (PredCount >= LimitLeft)
break;
if (PBInst.getDesc().isCall())
break;
PotentialBlockers.push_back(&PBInst);
}
}
}
return PotentialBlockers;
}
void X86AvoidSFBPass::buildCopy(MachineInstr *LoadInst, unsigned NLoadOpcode,
int64_t LoadDisp, MachineInstr *StoreInst,
unsigned NStoreOpcode, int64_t StoreDisp,
unsigned Size, int64_t LMMOffset,
int64_t SMMOffset) {
MachineOperand &LoadBase = getBaseOperand(LoadInst);
MachineOperand &StoreBase = getBaseOperand(StoreInst);
MachineBasicBlock *MBB = LoadInst->getParent();
MachineMemOperand *LMMO = *LoadInst->memoperands_begin();
MachineMemOperand *SMMO = *StoreInst->memoperands_begin();
Register Reg1 = MRI->createVirtualRegister(
TII->getRegClass(TII->get(NLoadOpcode), 0, TRI, *(MBB->getParent())));
MachineInstr *NewLoad =
BuildMI(*MBB, LoadInst, LoadInst->getDebugLoc(), TII->get(NLoadOpcode),
Reg1)
.add(LoadBase)
.addImm(1)
.addReg(X86::NoRegister)
.addImm(LoadDisp)
.addReg(X86::NoRegister)
.addMemOperand(
MBB->getParent()->getMachineMemOperand(LMMO, LMMOffset, Size));
if (LoadBase.isReg())
getBaseOperand(NewLoad).setIsKill(false);
LLVM_DEBUG(NewLoad->dump());
// If the load and store are consecutive, use the loadInst location to
// reduce register pressure.
MachineInstr *StInst = StoreInst;
auto PrevInstrIt = prev_nodbg(MachineBasicBlock::instr_iterator(StoreInst),
MBB->instr_begin());
if (PrevInstrIt.getNodePtr() == LoadInst)
StInst = LoadInst;
MachineInstr *NewStore =
BuildMI(*MBB, StInst, StInst->getDebugLoc(), TII->get(NStoreOpcode))
.add(StoreBase)
.addImm(1)
.addReg(X86::NoRegister)
.addImm(StoreDisp)
.addReg(X86::NoRegister)
.addReg(Reg1)
.addMemOperand(
MBB->getParent()->getMachineMemOperand(SMMO, SMMOffset, Size));
if (StoreBase.isReg())
getBaseOperand(NewStore).setIsKill(false);
MachineOperand &StoreSrcVReg = StoreInst->getOperand(X86::AddrNumOperands);
assert(StoreSrcVReg.isReg() && "Expected virtual register");
NewStore->getOperand(X86::AddrNumOperands).setIsKill(StoreSrcVReg.isKill());
LLVM_DEBUG(NewStore->dump());
}
void X86AvoidSFBPass::buildCopies(int Size, MachineInstr *LoadInst,
int64_t LdDispImm, MachineInstr *StoreInst,
int64_t StDispImm, int64_t LMMOffset,
int64_t SMMOffset) {
int LdDisp = LdDispImm;
int StDisp = StDispImm;
while (Size > 0) {
if ((Size - MOV128SZ >= 0) && isYMMLoadOpcode(LoadInst->getOpcode())) {
Size = Size - MOV128SZ;
buildCopy(LoadInst, getYMMtoXMMLoadOpcode(LoadInst->getOpcode()), LdDisp,
StoreInst, getYMMtoXMMStoreOpcode(StoreInst->getOpcode()),
StDisp, MOV128SZ, LMMOffset, SMMOffset);
LdDisp += MOV128SZ;
StDisp += MOV128SZ;
LMMOffset += MOV128SZ;
SMMOffset += MOV128SZ;
continue;
}
if (Size - MOV64SZ >= 0) {
Size = Size - MOV64SZ;
buildCopy(LoadInst, X86::MOV64rm, LdDisp, StoreInst, X86::MOV64mr, StDisp,
MOV64SZ, LMMOffset, SMMOffset);
LdDisp += MOV64SZ;
StDisp += MOV64SZ;
LMMOffset += MOV64SZ;
SMMOffset += MOV64SZ;
continue;
}
if (Size - MOV32SZ >= 0) {
Size = Size - MOV32SZ;
buildCopy(LoadInst, X86::MOV32rm, LdDisp, StoreInst, X86::MOV32mr, StDisp,
MOV32SZ, LMMOffset, SMMOffset);
LdDisp += MOV32SZ;
StDisp += MOV32SZ;
LMMOffset += MOV32SZ;
SMMOffset += MOV32SZ;
continue;
}
if (Size - MOV16SZ >= 0) {
Size = Size - MOV16SZ;
buildCopy(LoadInst, X86::MOV16rm, LdDisp, StoreInst, X86::MOV16mr, StDisp,
MOV16SZ, LMMOffset, SMMOffset);
LdDisp += MOV16SZ;
StDisp += MOV16SZ;
LMMOffset += MOV16SZ;
SMMOffset += MOV16SZ;
continue;
}
if (Size - MOV8SZ >= 0) {
Size = Size - MOV8SZ;
buildCopy(LoadInst, X86::MOV8rm, LdDisp, StoreInst, X86::MOV8mr, StDisp,
MOV8SZ, LMMOffset, SMMOffset);
LdDisp += MOV8SZ;
StDisp += MOV8SZ;
LMMOffset += MOV8SZ;
SMMOffset += MOV8SZ;
continue;
}
}
assert(Size == 0 && "Wrong size division");
}
static void updateKillStatus(MachineInstr *LoadInst, MachineInstr *StoreInst) {
MachineOperand &LoadBase = getBaseOperand(LoadInst);
MachineOperand &StoreBase = getBaseOperand(StoreInst);
auto *StorePrevNonDbgInstr =
prev_nodbg(MachineBasicBlock::instr_iterator(StoreInst),
LoadInst->getParent()->instr_begin())
.getNodePtr();
if (LoadBase.isReg()) {
MachineInstr *LastLoad = LoadInst->getPrevNode();
// If the original load and store to xmm/ymm were consecutive
// then the partial copies were also created in
// a consecutive order to reduce register pressure,
// and the location of the last load is before the last store.
if (StorePrevNonDbgInstr == LoadInst)
LastLoad = LoadInst->getPrevNode()->getPrevNode();
getBaseOperand(LastLoad).setIsKill(LoadBase.isKill());
}
if (StoreBase.isReg()) {
MachineInstr *StInst = StoreInst;
if (StorePrevNonDbgInstr == LoadInst)
StInst = LoadInst;
getBaseOperand(StInst->getPrevNode()).setIsKill(StoreBase.isKill());
}
}
bool X86AvoidSFBPass::alias(const MachineMemOperand &Op1,
const MachineMemOperand &Op2) const {
if (!Op1.getValue() || !Op2.getValue())
return true;
int64_t MinOffset = std::min(Op1.getOffset(), Op2.getOffset());
int64_t Overlapa = Op1.getSize() + Op1.getOffset() - MinOffset;
int64_t Overlapb = Op2.getSize() + Op2.getOffset() - MinOffset;
return !AA->isNoAlias(
MemoryLocation(Op1.getValue(), Overlapa, Op1.getAAInfo()),
MemoryLocation(Op2.getValue(), Overlapb, Op2.getAAInfo()));
}
void X86AvoidSFBPass::findPotentiallylBlockedCopies(MachineFunction &MF) {
for (auto &MBB : MF)
for (auto &MI : MBB) {
if (!isPotentialBlockedMemCpyLd(MI.getOpcode()))
continue;
int DefVR = MI.getOperand(0).getReg();
if (!MRI->hasOneNonDBGUse(DefVR))
continue;
for (MachineOperand &StoreMO :
llvm::make_early_inc_range(MRI->use_nodbg_operands(DefVR))) {
MachineInstr &StoreMI = *StoreMO.getParent();
// Skip cases where the memcpy may overlap.
if (StoreMI.getParent() == MI.getParent() &&
isPotentialBlockedMemCpyPair(MI.getOpcode(), StoreMI.getOpcode()) &&
isRelevantAddressingMode(&MI) &&
isRelevantAddressingMode(&StoreMI) &&
MI.hasOneMemOperand() && StoreMI.hasOneMemOperand()) {
if (!alias(**MI.memoperands_begin(), **StoreMI.memoperands_begin()))
BlockedLoadsStoresPairs.push_back(std::make_pair(&MI, &StoreMI));
}
}
}
}
unsigned X86AvoidSFBPass::getRegSizeInBytes(MachineInstr *LoadInst) {
const auto *TRC = TII->getRegClass(TII->get(LoadInst->getOpcode()), 0, TRI,
*LoadInst->getParent()->getParent());
return TRI->getRegSizeInBits(*TRC) / 8;
}
void X86AvoidSFBPass::breakBlockedCopies(
MachineInstr *LoadInst, MachineInstr *StoreInst,
const DisplacementSizeMap &BlockingStoresDispSizeMap) {
int64_t LdDispImm = getDispOperand(LoadInst).getImm();
int64_t StDispImm = getDispOperand(StoreInst).getImm();
int64_t LMMOffset = 0;
int64_t SMMOffset = 0;
int64_t LdDisp1 = LdDispImm;
int64_t LdDisp2 = 0;
int64_t StDisp1 = StDispImm;
int64_t StDisp2 = 0;
unsigned Size1 = 0;
unsigned Size2 = 0;
int64_t LdStDelta = StDispImm - LdDispImm;
for (auto DispSizePair : BlockingStoresDispSizeMap) {
LdDisp2 = DispSizePair.first;
StDisp2 = DispSizePair.first + LdStDelta;
Size2 = DispSizePair.second;
// Avoid copying overlapping areas.
if (LdDisp2 < LdDisp1) {
int OverlapDelta = LdDisp1 - LdDisp2;
LdDisp2 += OverlapDelta;
StDisp2 += OverlapDelta;
Size2 -= OverlapDelta;
}
Size1 = LdDisp2 - LdDisp1;
// Build a copy for the point until the current blocking store's
// displacement.
buildCopies(Size1, LoadInst, LdDisp1, StoreInst, StDisp1, LMMOffset,
SMMOffset);
// Build a copy for the current blocking store.
buildCopies(Size2, LoadInst, LdDisp2, StoreInst, StDisp2, LMMOffset + Size1,
SMMOffset + Size1);
LdDisp1 = LdDisp2 + Size2;
StDisp1 = StDisp2 + Size2;
LMMOffset += Size1 + Size2;
SMMOffset += Size1 + Size2;
}
unsigned Size3 = (LdDispImm + getRegSizeInBytes(LoadInst)) - LdDisp1;
buildCopies(Size3, LoadInst, LdDisp1, StoreInst, StDisp1, LMMOffset,
LMMOffset);
}
static bool hasSameBaseOpValue(MachineInstr *LoadInst,
MachineInstr *StoreInst) {
const MachineOperand &LoadBase = getBaseOperand(LoadInst);
const MachineOperand &StoreBase = getBaseOperand(StoreInst);
if (LoadBase.isReg() != StoreBase.isReg())
return false;
if (LoadBase.isReg())
return LoadBase.getReg() == StoreBase.getReg();
return LoadBase.getIndex() == StoreBase.getIndex();
}
static bool isBlockingStore(int64_t LoadDispImm, unsigned LoadSize,
int64_t StoreDispImm, unsigned StoreSize) {
return ((StoreDispImm >= LoadDispImm) &&
(StoreDispImm <= LoadDispImm + (LoadSize - StoreSize)));
}
// Keep track of all stores blocking a load
static void
updateBlockingStoresDispSizeMap(DisplacementSizeMap &BlockingStoresDispSizeMap,
int64_t DispImm, unsigned Size) {
if (BlockingStoresDispSizeMap.count(DispImm)) {
// Choose the smallest blocking store starting at this displacement.
if (BlockingStoresDispSizeMap[DispImm] > Size)
BlockingStoresDispSizeMap[DispImm] = Size;
} else
BlockingStoresDispSizeMap[DispImm] = Size;
}
// Remove blocking stores contained in each other.
static void
removeRedundantBlockingStores(DisplacementSizeMap &BlockingStoresDispSizeMap) {
if (BlockingStoresDispSizeMap.size() <= 1)
return;
SmallVector<std::pair<int64_t, unsigned>, 0> DispSizeStack;
for (auto DispSizePair : BlockingStoresDispSizeMap) {
int64_t CurrDisp = DispSizePair.first;
unsigned CurrSize = DispSizePair.second;
while (DispSizeStack.size()) {
int64_t PrevDisp = DispSizeStack.back().first;
unsigned PrevSize = DispSizeStack.back().second;
if (CurrDisp + CurrSize > PrevDisp + PrevSize)
break;
DispSizeStack.pop_back();
}
DispSizeStack.push_back(DispSizePair);
}
BlockingStoresDispSizeMap.clear();
for (auto Disp : DispSizeStack)
BlockingStoresDispSizeMap.insert(Disp);
}
bool X86AvoidSFBPass::runOnMachineFunction(MachineFunction &MF) {
bool Changed = false;
if (DisableX86AvoidStoreForwardBlocks || skipFunction(MF.getFunction()) ||
!MF.getSubtarget<X86Subtarget>().is64Bit())
return false;
MRI = &MF.getRegInfo();
assert(MRI->isSSA() && "Expected MIR to be in SSA form");
TII = MF.getSubtarget<X86Subtarget>().getInstrInfo();
TRI = MF.getSubtarget<X86Subtarget>().getRegisterInfo();
AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
LLVM_DEBUG(dbgs() << "Start X86AvoidStoreForwardBlocks\n";);
// Look for a load then a store to XMM/YMM which look like a memcpy
findPotentiallylBlockedCopies(MF);
for (auto LoadStoreInstPair : BlockedLoadsStoresPairs) {
MachineInstr *LoadInst = LoadStoreInstPair.first;
int64_t LdDispImm = getDispOperand(LoadInst).getImm();
DisplacementSizeMap BlockingStoresDispSizeMap;
SmallVector<MachineInstr *, 2> PotentialBlockers =
findPotentialBlockers(LoadInst);
for (auto *PBInst : PotentialBlockers) {
if (!isPotentialBlockingStoreInst(PBInst->getOpcode(),
LoadInst->getOpcode()) ||
!isRelevantAddressingMode(PBInst) || !PBInst->hasOneMemOperand())
continue;
int64_t PBstDispImm = getDispOperand(PBInst).getImm();
unsigned PBstSize = (*PBInst->memoperands_begin())->getSize();
// This check doesn't cover all cases, but it will suffice for now.
// TODO: take branch probability into consideration, if the blocking
// store is in an unreached block, breaking the memcopy could lose
// performance.
if (hasSameBaseOpValue(LoadInst, PBInst) &&
isBlockingStore(LdDispImm, getRegSizeInBytes(LoadInst), PBstDispImm,
PBstSize))
updateBlockingStoresDispSizeMap(BlockingStoresDispSizeMap, PBstDispImm,
PBstSize);
}
if (BlockingStoresDispSizeMap.empty())
continue;
// We found a store forward block, break the memcpy's load and store
// into smaller copies such that each smaller store that was causing
// a store block would now be copied separately.
MachineInstr *StoreInst = LoadStoreInstPair.second;
LLVM_DEBUG(dbgs() << "Blocked load and store instructions: \n");
LLVM_DEBUG(LoadInst->dump());
LLVM_DEBUG(StoreInst->dump());
LLVM_DEBUG(dbgs() << "Replaced with:\n");
removeRedundantBlockingStores(BlockingStoresDispSizeMap);
breakBlockedCopies(LoadInst, StoreInst, BlockingStoresDispSizeMap);
updateKillStatus(LoadInst, StoreInst);
ForRemoval.push_back(LoadInst);
ForRemoval.push_back(StoreInst);
}
for (auto *RemovedInst : ForRemoval) {
RemovedInst->eraseFromParent();
}
ForRemoval.clear();
BlockedLoadsStoresPairs.clear();
LLVM_DEBUG(dbgs() << "End X86AvoidStoreForwardBlocks\n";);
return Changed;
}