STATISTIC(NumPreFolded, "Number of pre-index updates folded");
STATISTIC(NumUnscaledPairCreated,
"Number of load/store from unscaled generated");
+STATISTIC(NumSmallTypeMerged, "Number of small type loads merged");
static cl::opt<unsigned> ScanLimit("aarch64-load-store-scan-limit",
cl::init(20), cl::Hidden);
-// Place holder while testing unscaled load/store combining
-static cl::opt<bool> EnableAArch64UnscaledMemOp(
- "aarch64-unscaled-mem-op", cl::Hidden,
- cl::desc("Allow AArch64 unscaled load/store combining"), cl::init(true));
-
namespace llvm {
void initializeAArch64LoadStoreOptPass(PassRegistry &);
}
// pre or post indexed addressing with writeback. Scan forwards.
MachineBasicBlock::iterator
findMatchingUpdateInsnForward(MachineBasicBlock::iterator I, unsigned Limit,
- int Value);
+ int UnscaledOffset);
// Scan the instruction list to find a base register update that can
// be combined with the current instruction (a load or store) using
mergeUpdateInsn(MachineBasicBlock::iterator I,
MachineBasicBlock::iterator Update, bool IsPreIdx);
- bool optimizeBlock(MachineBasicBlock &MBB);
+ // Find and merge foldable ldr/str instructions.
+ bool tryToMergeLdStInst(MachineBasicBlock::iterator &MBBI);
+
+ // Check if converting two narrow loads into a single wider load with
+ // bitfield extracts could be enabled.
+ bool enableNarrowLdMerge(MachineFunction &Fn);
+
+ bool optimizeBlock(MachineBasicBlock &MBB, bool enableNarrowLdOpt);
bool runOnMachineFunction(MachineFunction &Fn) override;
case AArch64::LDURWi:
case AArch64::LDURXi:
case AArch64::LDURSWi:
+ case AArch64::LDURHHi:
return true;
}
}
return isUnscaledLdSt(MI->getOpcode());
}
+static bool isSmallTypeLdMerge(unsigned Opc) {
+ switch (Opc) {
+ default:
+ return false;
+ case AArch64::LDRHHui:
+ case AArch64::LDURHHi:
+ return true;
+ // FIXME: Add other instructions (e.g, LDRBBui, LDURSHWi, LDRSHWui, etc.).
+ }
+}
+static bool isSmallTypeLdMerge(MachineInstr *MI) {
+ return isSmallTypeLdMerge(MI->getOpcode());
+}
+
// Scaling factor for unscaled load or store.
static int getMemScale(MachineInstr *MI) {
switch (MI->getOpcode()) {
case AArch64::STRBBui:
return 1;
case AArch64::LDRHHui:
+ case AArch64::LDURHHi:
case AArch64::STRHHui:
return 2;
case AArch64::LDRSui:
case AArch64::STRWui:
case AArch64::STURWi:
case AArch64::LDPSi:
+ case AArch64::LDPSWi:
case AArch64::LDPWi:
case AArch64::STPSi:
case AArch64::STPWi:
case AArch64::STURSi:
case AArch64::LDRSui:
case AArch64::LDURSi:
+ case AArch64::LDRHHui:
+ case AArch64::LDURHHi:
return Opc;
case AArch64::LDRSWui:
return AArch64::LDRWui;
case AArch64::LDRSWui:
case AArch64::LDURSWi:
return AArch64::LDPSWi;
+ case AArch64::LDRHHui:
+ return AArch64::LDRWui;
+ case AArch64::LDURHHi:
+ return AArch64::LDURWi;
}
}
return AArch64::LDRSWpre;
case AArch64::LDPSi:
return AArch64::LDPSpre;
+ case AArch64::LDPSWi:
+ return AArch64::LDPSWpre;
case AArch64::LDPDi:
return AArch64::LDPDpre;
case AArch64::LDPQi:
return AArch64::LDRSWpost;
case AArch64::LDPSi:
return AArch64::LDPSpost;
+ case AArch64::LDPSWi:
+ return AArch64::LDPSWpost;
case AArch64::LDPDi:
return AArch64::LDPDpost;
case AArch64::LDPQi:
default:
return false;
case AArch64::LDPSi:
+ case AArch64::LDPSWi:
case AArch64::LDPDi:
case AArch64::LDPQi:
case AArch64::LDPWi:
return MI->getOperand(Idx);
}
+// Copy MachineMemOperands from Op0 and Op1 to a new array assigned to MI.
+static void concatenateMemOperands(MachineInstr *MI, MachineInstr *Op0,
+ MachineInstr *Op1) {
+ assert(MI->memoperands_empty() && "expected a new machineinstr");
+ size_t numMemRefs = (Op0->memoperands_end() - Op0->memoperands_begin()) +
+ (Op1->memoperands_end() - Op1->memoperands_begin());
+
+ MachineFunction *MF = MI->getParent()->getParent();
+ MachineSDNode::mmo_iterator MemBegin = MF->allocateMemRefsArray(numMemRefs);
+ MachineSDNode::mmo_iterator MemEnd =
+ std::copy(Op0->memoperands_begin(), Op0->memoperands_end(), MemBegin);
+ MemEnd = std::copy(Op1->memoperands_begin(), Op1->memoperands_end(), MemEnd);
+ MI->setMemRefs(MemBegin, MemEnd);
+}
+
MachineBasicBlock::iterator
AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
MachineBasicBlock::iterator Paired,
unsigned Opc =
SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode());
bool IsUnscaled = isUnscaledLdSt(Opc);
- int OffsetStride =
- IsUnscaled && EnableAArch64UnscaledMemOp ? getMemScale(I) : 1;
+ int OffsetStride = IsUnscaled ? getMemScale(I) : 1;
bool MergeForward = Flags.getMergeForward();
unsigned NewOpc = getMatchingPairOpcode(Opc);
RtMI = I;
Rt2MI = Paired;
}
- // Handle Unscaled
+
int OffsetImm = getLdStOffsetOp(RtMI).getImm();
- if (IsUnscaled && EnableAArch64UnscaledMemOp)
+
+ if (isSmallTypeLdMerge(Opc)) {
+ // Change the scaled offset from small to large type.
+ if (!IsUnscaled)
+ OffsetImm /= 2;
+ MachineInstr *RtNewDest = MergeForward ? I : Paired;
+ // Construct the new load instruction.
+ // FIXME: currently we support only halfword unsigned load. We need to
+ // handle byte type, signed, and store instructions as well.
+ MachineInstr *NewMemMI, *BitExtMI1, *BitExtMI2;
+ NewMemMI = BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
+ TII->get(NewOpc))
+ .addOperand(getLdStRegOp(RtNewDest))
+ .addOperand(BaseRegOp)
+ .addImm(OffsetImm);
+
+ // Copy MachineMemOperands from the original loads.
+ concatenateMemOperands(NewMemMI, I, Paired);
+
+ DEBUG(
+ dbgs()
+ << "Creating the new load and extract. Replacing instructions:\n ");
+ DEBUG(I->print(dbgs()));
+ DEBUG(dbgs() << " ");
+ DEBUG(Paired->print(dbgs()));
+ DEBUG(dbgs() << " with instructions:\n ");
+ DEBUG((NewMemMI)->print(dbgs()));
+
+ MachineInstr *ExtDestMI = MergeForward ? Paired : I;
+ if (ExtDestMI == Rt2MI) {
+ // Create the bitfield extract for high half.
+ BitExtMI1 = BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
+ TII->get(AArch64::UBFMWri))
+ .addOperand(getLdStRegOp(Rt2MI))
+ .addReg(getLdStRegOp(RtNewDest).getReg())
+ .addImm(16)
+ .addImm(31);
+ // Create the bitfield extract for low half.
+ BitExtMI2 = BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
+ TII->get(AArch64::ANDWri))
+ .addOperand(getLdStRegOp(RtMI))
+ .addReg(getLdStRegOp(RtNewDest).getReg())
+ .addImm(15);
+ } else {
+ // Create the bitfield extract for low half.
+ BitExtMI1 = BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
+ TII->get(AArch64::ANDWri))
+ .addOperand(getLdStRegOp(RtMI))
+ .addReg(getLdStRegOp(RtNewDest).getReg())
+ .addImm(15);
+ // Create the bitfield extract for high half.
+ BitExtMI2 = BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
+ TII->get(AArch64::UBFMWri))
+ .addOperand(getLdStRegOp(Rt2MI))
+ .addReg(getLdStRegOp(RtNewDest).getReg())
+ .addImm(16)
+ .addImm(31);
+ }
+ DEBUG(dbgs() << " ");
+ DEBUG((BitExtMI1)->print(dbgs()));
+ DEBUG(dbgs() << " ");
+ DEBUG((BitExtMI2)->print(dbgs()));
+ DEBUG(dbgs() << "\n");
+
+ // Erase the old instructions.
+ I->eraseFromParent();
+ Paired->eraseFromParent();
+ return NextI;
+ }
+
+ // Handle Unscaled
+ if (IsUnscaled)
OffsetImm /= OffsetStride;
// Construct the new instruction.
/// be combined with the current instruction into a load/store pair.
MachineBasicBlock::iterator
AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I,
- LdStPairFlags &Flags,
- unsigned Limit) {
+ LdStPairFlags &Flags, unsigned Limit) {
MachineBasicBlock::iterator E = I->getParent()->end();
MachineBasicBlock::iterator MBBI = I;
MachineInstr *FirstMI = I;
// Early exit if the offset if not possible to match. (6 bits of positive
// range, plus allow an extra one in case we find a later insn that matches
// with Offset-1)
- int OffsetStride =
- IsUnscaled && EnableAArch64UnscaledMemOp ? getMemScale(FirstMI) : 1;
- if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride))
+ int OffsetStride = IsUnscaled ? getMemScale(FirstMI) : 1;
+ if (!isSmallTypeLdMerge(Opc) &&
+ !inBoundsForPair(IsUnscaled, Offset, OffsetStride))
return E;
// Track which registers have been modified and used between the first insn
// If the resultant immediate offset of merging these instructions
// is out of range for a pairwise instruction, bail and keep looking.
bool MIIsUnscaled = isUnscaledLdSt(MI);
- if (!inBoundsForPair(MIIsUnscaled, MinOffset, OffsetStride)) {
+ bool IsSmallTypeLd = isSmallTypeLdMerge(MI->getOpcode());
+ if (!IsSmallTypeLd &&
+ !inBoundsForPair(MIIsUnscaled, MinOffset, OffsetStride)) {
trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
MemInsns.push_back(MI);
continue;
}
- // If the alignment requirements of the paired (scaled) instruction
- // can't express the offset of the unscaled input, bail and keep
- // looking.
- if (IsUnscaled && EnableAArch64UnscaledMemOp &&
- (alignTo(MinOffset, OffsetStride) != MinOffset)) {
- trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
- MemInsns.push_back(MI);
- continue;
+
+ if (IsSmallTypeLd) {
+ // If the alignment requirements of the larger type scaled load
+ // instruction can't express the scaled offset of the smaller type
+ // input, bail and keep looking.
+ if (!IsUnscaled && alignTo(MinOffset, 2) != MinOffset) {
+ trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
+ MemInsns.push_back(MI);
+ continue;
+ }
+ } else {
+ // If the alignment requirements of the paired (scaled) instruction
+ // can't express the offset of the unscaled input, bail and keep
+ // looking.
+ if (IsUnscaled && (alignTo(MinOffset, OffsetStride) != MinOffset)) {
+ trackRegDefsUses(MI, ModifiedRegs, UsedRegs, TRI);
+ MemInsns.push_back(MI);
+ continue;
+ }
}
// If the destination register of the loads is the same register, bail
// and keep looking. A load-pair instruction with both destination
}
MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward(
- MachineBasicBlock::iterator I, unsigned Limit, int Value) {
+ MachineBasicBlock::iterator I, unsigned Limit, int UnscaledOffset) {
MachineBasicBlock::iterator E = I->getParent()->end();
MachineInstr *MemMI = I;
MachineBasicBlock::iterator MBBI = I;
unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
- int Offset = getLdStOffsetOp(MemMI).getImm() * getMemScale(MemMI);
+ int MIUnscaledOffset = getLdStOffsetOp(MemMI).getImm() * getMemScale(MemMI);
+
+ // Scan forward looking for post-index opportunities. Updating instructions
+ // can't be formed if the memory instruction doesn't have the offset we're
+ // looking for.
+ if (MIUnscaledOffset != UnscaledOffset)
+ return E;
// If the base register overlaps a destination register, we can't
// merge the update.
return E;
}
- // Scan forward looking for post-index opportunities.
- // Updating instructions can't be formed if the memory insn already
- // has an offset other than the value we're looking for.
- if (Offset != Value)
- return E;
-
// Track which registers have been modified and used between the first insn
// (inclusive) and the second insn.
BitVector ModifiedRegs, UsedRegs;
++Count;
// If we found a match, return it.
- if (isMatchingUpdateInsn(I, MI, BaseReg, Value))
+ if (isMatchingUpdateInsn(I, MI, BaseReg, UnscaledOffset))
return MBBI;
// Update the status of what the instruction clobbered and used.
unsigned BaseReg = getLdStBaseOp(MemMI).getReg();
int Offset = getLdStOffsetOp(MemMI).getImm();
- unsigned MemSize = getMemScale(MemMI);
// If the load/store is the first instruction in the block, there's obviously
// not any matching update. Ditto if the memory offset isn't zero.
++Count;
// If we found a match, return it.
- if (isMatchingUpdateInsn(I, MI, BaseReg, MemSize))
+ if (isMatchingUpdateInsn(I, MI, BaseReg, Offset))
return MBBI;
// Update the status of what the instruction clobbered and used.
return E;
}
-bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB) {
+bool AArch64LoadStoreOpt::tryToMergeLdStInst(
+ MachineBasicBlock::iterator &MBBI) {
+ MachineInstr *MI = MBBI;
+ MachineBasicBlock::iterator E = MI->getParent()->end();
+ // If this is a volatile load/store, don't mess with it.
+ if (MI->hasOrderedMemoryRef())
+ return false;
+
+ // Make sure this is a reg+imm (as opposed to an address reloc).
+ if (!getLdStOffsetOp(MI).isImm())
+ return false;
+
+ // Check if this load/store has a hint to avoid pair formation.
+ // MachineMemOperands hints are set by the AArch64StorePairSuppress pass.
+ if (TII->isLdStPairSuppressed(MI))
+ return false;
+
+ // Look ahead up to ScanLimit instructions for a pairable instruction.
+ LdStPairFlags Flags;
+ MachineBasicBlock::iterator Paired = findMatchingInsn(MBBI, Flags, ScanLimit);
+ if (Paired != E) {
+ if (isSmallTypeLdMerge(MI)) {
+ ++NumSmallTypeMerged;
+ } else {
+ ++NumPairCreated;
+ if (isUnscaledLdSt(MI))
+ ++NumUnscaledPairCreated;
+ }
+
+ // Merge the loads into a pair. Keeping the iterator straight is a
+ // pain, so we let the merge routine tell us what the next instruction
+ // is after it's done mucking about.
+ MBBI = mergePairedInsns(MBBI, Paired, Flags);
+ return true;
+ }
+ return false;
+}
+
+bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB,
+ bool enableNarrowLdOpt) {
bool Modified = false;
- // Two tranformations to do here:
- // 1) Find loads and stores that can be merged into a single load or store
+ // Three tranformations to do here:
+ // 1) Find halfword loads that can be merged into a single 32-bit word load
+ // with bitfield extract instructions.
+ // e.g.,
+ // ldrh w0, [x2]
+ // ldrh w1, [x2, #2]
+ // ; becomes
+ // ldr w0, [x2]
+ // ubfx w1, w0, #16, #16
+ // and w0, w0, #ffff
+ // 2) Find loads and stores that can be merged into a single load or store
// pair instruction.
// e.g.,
// ldr x0, [x2]
// ldr x1, [x2, #8]
// ; becomes
// ldp x0, x1, [x2]
- // 2) Find base register updates that can be merged into the load or store
+ // 3) Find base register updates that can be merged into the load or store
// as a base-reg writeback.
// e.g.,
// ldr x0, [x2]
// ; becomes
// ldr x0, [x2], #4
+ for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
+ enableNarrowLdOpt && MBBI != E;) {
+ MachineInstr *MI = MBBI;
+ switch (MI->getOpcode()) {
+ default:
+ // Just move on to the next instruction.
+ ++MBBI;
+ break;
+ // Scaled instructions.
+ case AArch64::LDRHHui:
+ // Unscaled instructions.
+ case AArch64::LDURHHi: {
+ if (tryToMergeLdStInst(MBBI)) {
+ Modified = true;
+ break;
+ }
+ ++MBBI;
+ break;
+ }
+ // FIXME: Do the other instructions.
+ }
+ }
+
for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
MBBI != E;) {
MachineInstr *MI = MBBI;
case AArch64::LDURWi:
case AArch64::LDURXi:
case AArch64::LDURSWi: {
- // If this is a volatile load/store, don't mess with it.
- if (MI->hasOrderedMemoryRef()) {
- ++MBBI;
- break;
- }
- // Make sure this is a reg+imm (as opposed to an address reloc).
- if (!getLdStOffsetOp(MI).isImm()) {
- ++MBBI;
- break;
- }
- // Check if this load/store has a hint to avoid pair formation.
- // MachineMemOperands hints are set by the AArch64StorePairSuppress pass.
- if (TII->isLdStPairSuppressed(MI)) {
- ++MBBI;
- break;
- }
- // Look ahead up to ScanLimit instructions for a pairable instruction.
- LdStPairFlags Flags;
- MachineBasicBlock::iterator Paired =
- findMatchingInsn(MBBI, Flags, ScanLimit);
- if (Paired != E) {
- ++NumPairCreated;
- if (isUnscaledLdSt(MI))
- ++NumUnscaledPairCreated;
-
- // Merge the loads into a pair. Keeping the iterator straight is a
- // pain, so we let the merge routine tell us what the next instruction
- // is after it's done mucking about.
- MBBI = mergePairedInsns(MBBI, Paired, Flags);
+ if (tryToMergeLdStInst(MBBI)) {
Modified = true;
break;
}
case AArch64::LDURXi:
// Paired instructions.
case AArch64::LDPSi:
+ case AArch64::LDPSWi:
case AArch64::LDPDi:
case AArch64::LDPQi:
case AArch64::LDPWi:
++NumPreFolded;
break;
}
- // The immediate in the load/store is scaled by the size of the register
- // being loaded. The immediate in the add we're looking for,
+ // The immediate in the load/store is scaled by the size of the memory
+ // operation. The immediate in the add we're looking for,
// however, is not, so adjust here.
- int Value =
- MI->getOperand(isPairedLdSt(MI) ? 3 : 2).getImm() * getMemScale(MI);
-
- // FIXME: The immediate in the load/store should be scaled by the size of
- // the memory operation, not the size of the register being loaded/stored.
- // This works in general, but does not work for the LDPSW instruction,
- // which defines two 64-bit registers, but loads 32-bit values.
+ int UnscaledOffset = getLdStOffsetOp(MI).getImm() * getMemScale(MI);
// Look forward to try to find a post-index instruction. For example,
// ldr x1, [x0, #64]
// add x0, x0, #64
// merged into:
// ldr x1, [x0, #64]!
- Update = findMatchingUpdateInsnForward(MBBI, ScanLimit, Value);
+ Update = findMatchingUpdateInsnForward(MBBI, ScanLimit, UnscaledOffset);
if (Update != E) {
// Merge the update into the ld/st.
MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true);
return Modified;
}
+bool AArch64LoadStoreOpt::enableNarrowLdMerge(MachineFunction &Fn) {
+ const AArch64Subtarget *SubTarget =
+ &static_cast<const AArch64Subtarget &>(Fn.getSubtarget());
+ bool ProfitableArch = SubTarget->isCortexA57();
+ // FIXME: The benefit from converting narrow loads into a wider load could be
+ // microarchitectural as it assumes that a single load with two bitfield
+ // extracts is cheaper than two narrow loads. Currently, this conversion is
+ // enabled only in cortex-a57 on which performance benefits were verified.
+ return ProfitableArch & (!SubTarget->requiresStrictAlign());
+}
+
bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
TII = static_cast<const AArch64InstrInfo *>(Fn.getSubtarget().getInstrInfo());
TRI = Fn.getSubtarget().getRegisterInfo();
bool Modified = false;
+ bool enableNarrowLdOpt = enableNarrowLdMerge(Fn);
for (auto &MBB : Fn)
- Modified |= optimizeBlock(MBB);
+ Modified |= optimizeBlock(MBB, enableNarrowLdOpt);
return Modified;
}
projects / oota-llvm.git / blobdiff