-//===-- MLxExpansionPass.cpp - Expand MLx instrs to avoid hazards ----------=//
+//===-- MLxExpansionPass.cpp - Expand MLx instrs to avoid hazards ---------===//
//
// The LLVM Compiler Infrastructure
//
#define DEBUG_TYPE "mlx-expansion"
#include "ARM.h"
#include "ARMBaseInstrInfo.h"
+#include "ARMSubtarget.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
const TargetRegisterInfo *TRI;
MachineRegisterInfo *MRI;
+ bool isLikeA9;
+ bool isSwift;
unsigned MIIdx;
MachineInstr* LastMIs[4];
+ SmallPtrSet<MachineInstr*, 4> IgnoreStall;
void clearStack();
void pushStack(MachineInstr *MI);
MachineInstr *getAccDefMI(MachineInstr *MI) const;
unsigned getDefReg(MachineInstr *MI) const;
+ bool hasLoopHazard(MachineInstr *MI) const;
bool hasRAWHazard(unsigned Reg, MachineInstr *MI) const;
- bool FindMLxHazard(MachineInstr *MI) const;
+ bool FindMLxHazard(MachineInstr *MI);
void ExpandFPMLxInstruction(MachineBasicBlock &MBB, MachineInstr *MI,
unsigned MulOpc, unsigned AddSubOpc,
bool NegAcc, bool HasLane);
return Reg;
}
-bool MLxExpansion::hasRAWHazard(unsigned Reg, MachineInstr *MI) const {
- const TargetInstrDesc &TID = MI->getDesc();
- // FIXME: Detect integer instructions properly.
- unsigned Domain = TID.TSFlags & ARMII::DomainMask;
- if (Domain == ARMII::DomainVFP) {
- unsigned Opcode = TID.getOpcode();
- if (Opcode == ARM::VSTRS || Opcode == ARM::VSTRD ||
- Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD)
- return false;
- } else if (Domain == ARMII::DomainNEON) {
- if (TID.mayStore() || TID.mayLoad())
- return false;
- } else {
+/// hasLoopHazard - Check whether an MLx instruction is chained to itself across
+/// a single-MBB loop.
+bool MLxExpansion::hasLoopHazard(MachineInstr *MI) const {
+ unsigned Reg = MI->getOperand(1).getReg();
+ if (TargetRegisterInfo::isPhysicalRegister(Reg))
return false;
+
+ MachineBasicBlock *MBB = MI->getParent();
+ MachineInstr *DefMI = MRI->getVRegDef(Reg);
+ while (true) {
+outer_continue:
+ if (DefMI->getParent() != MBB)
+ break;
+
+ if (DefMI->isPHI()) {
+ for (unsigned i = 1, e = DefMI->getNumOperands(); i < e; i += 2) {
+ if (DefMI->getOperand(i + 1).getMBB() == MBB) {
+ unsigned SrcReg = DefMI->getOperand(i).getReg();
+ if (TargetRegisterInfo::isVirtualRegister(SrcReg)) {
+ DefMI = MRI->getVRegDef(SrcReg);
+ goto outer_continue;
+ }
+ }
+ }
+ } else if (DefMI->isCopyLike()) {
+ Reg = DefMI->getOperand(1).getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ DefMI = MRI->getVRegDef(Reg);
+ continue;
+ }
+ } else if (DefMI->isInsertSubreg()) {
+ Reg = DefMI->getOperand(2).getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ DefMI = MRI->getVRegDef(Reg);
+ continue;
+ }
+ }
+
+ break;
}
- return MI->readsRegister(Reg, TRI);
+ return DefMI == MI;
+}
+
+bool MLxExpansion::hasRAWHazard(unsigned Reg, MachineInstr *MI) const {
+ // FIXME: Detect integer instructions properly.
+ const MCInstrDesc &MCID = MI->getDesc();
+ unsigned Domain = MCID.TSFlags & ARMII::DomainMask;
+ if (MI->mayStore())
+ return false;
+ unsigned Opcode = MCID.getOpcode();
+ if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD)
+ return false;
+ if ((Domain & ARMII::DomainVFP) || (Domain & ARMII::DomainNEON))
+ return MI->readsRegister(Reg, TRI);
return false;
}
+static bool isFpMulInstruction(unsigned Opcode) {
+ switch (Opcode) {
+ case ARM::VMULS:
+ case ARM::VMULfd:
+ case ARM::VMULfq:
+ case ARM::VMULD:
+ case ARM::VMULslfd:
+ case ARM::VMULslfq:
+ return true;
+ default:
+ return false;
+ }
+}
-bool MLxExpansion::FindMLxHazard(MachineInstr *MI) const {
+bool MLxExpansion::FindMLxHazard(MachineInstr *MI) {
if (NumExpand >= ExpandLimit)
return false;
return true;
MachineInstr *DefMI = getAccDefMI(MI);
- if (TII->isFpMLxInstruction(DefMI->getOpcode()))
+ if (TII->isFpMLxInstruction(DefMI->getOpcode())) {
// r0 = vmla
// r3 = vmla r0, r1, r2
// takes 16 - 17 cycles
// r4 = vmul r1, r2
// r3 = vadd r0, r4
// takes about 14 - 15 cycles even with vmul stalling for 4 cycles.
+ IgnoreStall.insert(DefMI);
return true;
+ }
+
+ // On Swift, we mostly care about hazards from multiplication instructions
+ // writing the accumulator and the pipelining of loop iterations by out-of-
+ // order execution.
+ if (isSwift)
+ return isFpMulInstruction(DefMI->getOpcode()) || hasLoopHazard(MI);
+
+ if (IgnoreStall.count(MI))
+ return false;
// If a VMLA.F is followed by an VADD.F or VMUL.F with no RAW hazard, the
// VADD.F or VMUL.F will stall 4 cycles before issue. The 4 cycle stall
// preserves the in-order retirement of the instructions.
// Look at the next few instructions, if *most* of them can cause hazards,
// then the scheduler can't *fix* this, we'd better break up the VMLA.
+ unsigned Limit1 = isLikeA9 ? 1 : 4;
+ unsigned Limit2 = isLikeA9 ? 1 : 4;
for (unsigned i = 1; i <= 4; ++i) {
int Idx = ((int)MIIdx - i + 4) % 4;
MachineInstr *NextMI = LastMIs[Idx];
if (!NextMI)
continue;
- if (TII->canCauseFpMLxStall(NextMI->getOpcode()))
- return true;
+ if (TII->canCauseFpMLxStall(NextMI->getOpcode())) {
+ if (i <= Limit1)
+ return true;
+ }
// Look for VMLx RAW hazard.
- if (hasRAWHazard(getDefReg(MI), NextMI))
+ if (i <= Limit2 && hasRAWHazard(getDefReg(MI), NextMI))
return true;
}
ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NextOp).getImm();
unsigned PredReg = MI->getOperand(++NextOp).getReg();
- const TargetInstrDesc &TID1 = TII->get(MulOpc);
- const TargetInstrDesc &TID2 = TII->get(AddSubOpc);
- unsigned TmpReg = MRI->createVirtualRegister(TID1.getRegClass(0, TRI));
+ const MCInstrDesc &MCID1 = TII->get(MulOpc);
+ const MCInstrDesc &MCID2 = TII->get(AddSubOpc);
+ const MachineFunction &MF = *MI->getParent()->getParent();
+ unsigned TmpReg = MRI->createVirtualRegister(
+ TII->getRegClass(MCID1, 0, TRI, MF));
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, MI->getDebugLoc(), TID1, TmpReg)
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, MI->getDebugLoc(), MCID1, TmpReg)
.addReg(Src1Reg, getKillRegState(Src1Kill))
.addReg(Src2Reg, getKillRegState(Src2Kill));
if (HasLane)
MIB.addImm(LaneImm);
MIB.addImm(Pred).addReg(PredReg);
- MIB = BuildMI(MBB, *MI, MI->getDebugLoc(), TID2)
+ MIB = BuildMI(MBB, MI, MI->getDebugLoc(), MCID2)
.addReg(DstReg, getDefRegState(true) | getDeadRegState(DstDead));
if (NegAcc) {
bool Changed = false;
clearStack();
+ IgnoreStall.clear();
unsigned Skip = 0;
MachineBasicBlock::reverse_iterator MII = MBB.rbegin(), E = MBB.rend();
continue;
}
- const TargetInstrDesc &TID = MI->getDesc();
- if (TID.isBarrier()) {
+ const MCInstrDesc &MCID = MI->getDesc();
+ if (MI->isBarrier()) {
clearStack();
Skip = 0;
++MII;
continue;
}
- unsigned Domain = TID.TSFlags & ARMII::DomainMask;
+ unsigned Domain = MCID.TSFlags & ARMII::DomainMask;
if (Domain == ARMII::DomainGeneral) {
if (++Skip == 2)
// Assume dual issues of non-VFP / NEON instructions.
unsigned MulOpc, AddSubOpc;
bool NegAcc, HasLane;
- if (!TII->isFpMLxInstruction(TID.getOpcode(),
+ if (!TII->isFpMLxInstruction(MCID.getOpcode(),
MulOpc, AddSubOpc, NegAcc, HasLane) ||
!FindMLxHazard(MI))
pushStack(MI);
TII = static_cast<const ARMBaseInstrInfo*>(Fn.getTarget().getInstrInfo());
TRI = Fn.getTarget().getRegisterInfo();
MRI = &Fn.getRegInfo();
+ const ARMSubtarget *STI = &Fn.getTarget().getSubtarget<ARMSubtarget>();
+ isLikeA9 = STI->isLikeA9() || STI->isSwift();
+ isSwift = STI->isSwift();
bool Modified = false;
for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;