blob: 5deb601822b8c3d6bd98babed2c75de6d74ad90d [file] [log] [blame]
//===--- AArch64StorePairSuppress.cpp --- Suppress store pair formation ---===//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// This pass identifies floating point stores that should not be combined into
// store pairs. Later we may do the same for floating point loads.
// ===---------------------------------------------------------------------===//
#include "AArch64InstrInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineTraceMetrics.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetSchedule.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
#define DEBUG_TYPE "aarch64-stp-suppress"
#define STPSUPPRESS_PASS_NAME "AArch64 Store Pair Suppression"
namespace {
class AArch64StorePairSuppress : public MachineFunctionPass {
const AArch64InstrInfo *TII;
const TargetRegisterInfo *TRI;
const MachineRegisterInfo *MRI;
TargetSchedModel SchedModel;
MachineTraceMetrics *Traces;
MachineTraceMetrics::Ensemble *MinInstr;
static char ID;
AArch64StorePairSuppress() : MachineFunctionPass(ID) {
StringRef getPassName() const override { return STPSUPPRESS_PASS_NAME; }
bool runOnMachineFunction(MachineFunction &F) override;
bool shouldAddSTPToBlock(const MachineBasicBlock *BB);
bool isNarrowFPStore(const MachineInstr &MI);
void getAnalysisUsage(AnalysisUsage &AU) const override {
char AArch64StorePairSuppress::ID = 0;
} // anonymous
INITIALIZE_PASS(AArch64StorePairSuppress, "aarch64-stp-suppress",
FunctionPass *llvm::createAArch64StorePairSuppressPass() {
return new AArch64StorePairSuppress();
/// Return true if an STP can be added to this block without increasing the
/// critical resource height. STP is good to form in Ld/St limited blocks and
/// bad to form in float-point limited blocks. This is true independent of the
/// critical path. If the critical path is longer than the resource height, the
/// extra vector ops can limit physreg renaming. Otherwise, it could simply
/// oversaturate the vector units.
bool AArch64StorePairSuppress::shouldAddSTPToBlock(const MachineBasicBlock *BB) {
if (!MinInstr)
MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount);
MachineTraceMetrics::Trace BBTrace = MinInstr->getTrace(BB);
unsigned ResLength = BBTrace.getResourceLength();
// Get the machine model's scheduling class for STPQi.
// Bypass TargetSchedule's SchedClass resolution since we only have an opcode.
unsigned SCIdx = TII->get(AArch64::STPDi).getSchedClass();
const MCSchedClassDesc *SCDesc =
// If a subtarget does not define resources for STPQi, bail here.
if (SCDesc->isValid() && !SCDesc->isVariant()) {
unsigned ResLenWithSTP = BBTrace.getResourceLength(None, SCDesc);
if (ResLenWithSTP > ResLength) {
LLVM_DEBUG(dbgs() << " Suppress STP in BB: " << BB->getNumber()
<< " resources " << ResLength << " -> " << ResLenWithSTP
<< "\n");
return false;
return true;
/// Return true if this is a floating-point store smaller than the V reg. On
/// cyclone, these require a vector shuffle before storing a pair.
/// Ideally we would call getMatchingPairOpcode() and have the machine model
/// tell us if it's profitable with no cpu knowledge here.
/// FIXME: We plan to develop a decent Target abstraction for simple loads and
/// stores. Until then use a nasty switch similar to AArch64LoadStoreOptimizer.
bool AArch64StorePairSuppress::isNarrowFPStore(const MachineInstr &MI) {
switch (MI.getOpcode()) {
return false;
case AArch64::STRSui:
case AArch64::STRDui:
case AArch64::STURSi:
case AArch64::STURDi:
return true;
bool AArch64StorePairSuppress::runOnMachineFunction(MachineFunction &MF) {
if (skipFunction(MF.getFunction()))
return false;
const TargetSubtargetInfo &ST = MF.getSubtarget();
TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo());
TRI = ST.getRegisterInfo();
MRI = &MF.getRegInfo();
Traces = &getAnalysis<MachineTraceMetrics>();
MinInstr = nullptr;
LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << MF.getName() << '\n');
if (!SchedModel.hasInstrSchedModel()) {
LLVM_DEBUG(dbgs() << " Skipping pass: no machine model present.\n");
return false;
// Check for a sequence of stores to the same base address. We don't need to
// precisely determine whether a store pair can be formed. But we do want to
// filter out most situations where we can't form store pairs to avoid
// computing trace metrics in those cases.
for (auto &MBB : MF) {
bool SuppressSTP = false;
unsigned PrevBaseReg = 0;
for (auto &MI : MBB) {
if (!isNarrowFPStore(MI))
const MachineOperand *BaseOp;
int64_t Offset;
if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, TRI) &&
BaseOp->isReg()) {
Register BaseReg = BaseOp->getReg();
if (PrevBaseReg == BaseReg) {
// If this block can take STPs, skip ahead to the next block.
if (!SuppressSTP && shouldAddSTPToBlock(MI.getParent()))
// Otherwise, continue unpairing the stores in this block.
LLVM_DEBUG(dbgs() << "Unpairing store " << MI << "\n");
SuppressSTP = true;
PrevBaseReg = BaseReg;
} else
PrevBaseReg = 0;
// This pass just sets some internal MachineMemOperand flags. It can't really
// invalidate anything.
return false;