1 //===--- AArch64StorePairSuppress.cpp --- Suppress store pair formation ---===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass identifies floating point stores that should not be combined into
10 // store pairs. Later we may do the same for floating point loads.
11 // ===---------------------------------------------------------------------===//
12 
13 #include "AArch64InstrInfo.h"
14 #include "llvm/CodeGen/MachineFunction.h"
15 #include "llvm/CodeGen/MachineFunctionPass.h"
16 #include "llvm/CodeGen/MachineInstr.h"
17 #include "llvm/CodeGen/MachineTraceMetrics.h"
18 #include "llvm/CodeGen/TargetInstrInfo.h"
19 #include "llvm/CodeGen/TargetSchedule.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Support/raw_ostream.h"
22 
23 using namespace llvm;
24 
25 #define DEBUG_TYPE "aarch64-stp-suppress"
26 
27 #define STPSUPPRESS_PASS_NAME "AArch64 Store Pair Suppression"
28 
29 namespace {
30 class AArch64StorePairSuppress : public MachineFunctionPass {
31   const AArch64InstrInfo *TII;
32   const TargetRegisterInfo *TRI;
33   const MachineRegisterInfo *MRI;
34   TargetSchedModel SchedModel;
35   MachineTraceMetrics *Traces;
36   MachineTraceMetrics::Ensemble *MinInstr;
37 
38 public:
39   static char ID;
AArch64StorePairSuppress()40   AArch64StorePairSuppress() : MachineFunctionPass(ID) {
41     initializeAArch64StorePairSuppressPass(*PassRegistry::getPassRegistry());
42   }
43 
getPassName() const44   StringRef getPassName() const override { return STPSUPPRESS_PASS_NAME; }
45 
46   bool runOnMachineFunction(MachineFunction &F) override;
47 
48 private:
49   bool shouldAddSTPToBlock(const MachineBasicBlock *BB);
50 
51   bool isNarrowFPStore(const MachineInstr &MI);
52 
getAnalysisUsage(AnalysisUsage & AU) const53   void getAnalysisUsage(AnalysisUsage &AU) const override {
54     AU.setPreservesCFG();
55     AU.addRequired<MachineTraceMetrics>();
56     AU.addPreserved<MachineTraceMetrics>();
57     MachineFunctionPass::getAnalysisUsage(AU);
58   }
59 };
60 char AArch64StorePairSuppress::ID = 0;
61 } // anonymous
62 
63 INITIALIZE_PASS(AArch64StorePairSuppress, "aarch64-stp-suppress",
64                 STPSUPPRESS_PASS_NAME, false, false)
65 
createAArch64StorePairSuppressPass()66 FunctionPass *llvm::createAArch64StorePairSuppressPass() {
67   return new AArch64StorePairSuppress();
68 }
69 
70 /// Return true if an STP can be added to this block without increasing the
71 /// critical resource height. STP is good to form in Ld/St limited blocks and
72 /// bad to form in float-point limited blocks. This is true independent of the
73 /// critical path. If the critical path is longer than the resource height, the
74 /// extra vector ops can limit physreg renaming. Otherwise, it could simply
75 /// oversaturate the vector units.
shouldAddSTPToBlock(const MachineBasicBlock * BB)76 bool AArch64StorePairSuppress::shouldAddSTPToBlock(const MachineBasicBlock *BB) {
77   if (!MinInstr)
78     MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount);
79 
80   MachineTraceMetrics::Trace BBTrace = MinInstr->getTrace(BB);
81   unsigned ResLength = BBTrace.getResourceLength();
82 
83   // Get the machine model's scheduling class for STPQi.
84   // Bypass TargetSchedule's SchedClass resolution since we only have an opcode.
85   unsigned SCIdx = TII->get(AArch64::STPDi).getSchedClass();
86   const MCSchedClassDesc *SCDesc =
87       SchedModel.getMCSchedModel()->getSchedClassDesc(SCIdx);
88 
89   // If a subtarget does not define resources for STPQi, bail here.
90   if (SCDesc->isValid() && !SCDesc->isVariant()) {
91     unsigned ResLenWithSTP = BBTrace.getResourceLength(None, SCDesc);
92     if (ResLenWithSTP > ResLength) {
93       LLVM_DEBUG(dbgs() << "  Suppress STP in BB: " << BB->getNumber()
94                         << " resources " << ResLength << " -> " << ResLenWithSTP
95                         << "\n");
96       return false;
97     }
98   }
99   return true;
100 }
101 
102 /// Return true if this is a floating-point store smaller than the V reg. On
103 /// cyclone, these require a vector shuffle before storing a pair.
104 /// Ideally we would call getMatchingPairOpcode() and have the machine model
105 /// tell us if it's profitable with no cpu knowledge here.
106 ///
107 /// FIXME: We plan to develop a decent Target abstraction for simple loads and
108 /// stores. Until then use a nasty switch similar to AArch64LoadStoreOptimizer.
isNarrowFPStore(const MachineInstr & MI)109 bool AArch64StorePairSuppress::isNarrowFPStore(const MachineInstr &MI) {
110   switch (MI.getOpcode()) {
111   default:
112     return false;
113   case AArch64::STRSui:
114   case AArch64::STRDui:
115   case AArch64::STURSi:
116   case AArch64::STURDi:
117     return true;
118   }
119 }
120 
runOnMachineFunction(MachineFunction & MF)121 bool AArch64StorePairSuppress::runOnMachineFunction(MachineFunction &MF) {
122   if (skipFunction(MF.getFunction()) || MF.getFunction().hasOptSize())
123     return false;
124 
125   const TargetSubtargetInfo &ST = MF.getSubtarget();
126   TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo());
127   TRI = ST.getRegisterInfo();
128   MRI = &MF.getRegInfo();
129   SchedModel.init(&ST);
130   Traces = &getAnalysis<MachineTraceMetrics>();
131   MinInstr = nullptr;
132 
133   LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << MF.getName() << '\n');
134 
135   if (!SchedModel.hasInstrSchedModel()) {
136     LLVM_DEBUG(dbgs() << "  Skipping pass: no machine model present.\n");
137     return false;
138   }
139 
140   // Check for a sequence of stores to the same base address. We don't need to
141   // precisely determine whether a store pair can be formed. But we do want to
142   // filter out most situations where we can't form store pairs to avoid
143   // computing trace metrics in those cases.
144   for (auto &MBB : MF) {
145     bool SuppressSTP = false;
146     unsigned PrevBaseReg = 0;
147     for (auto &MI : MBB) {
148       if (!isNarrowFPStore(MI))
149         continue;
150       const MachineOperand *BaseOp;
151       int64_t Offset;
152       bool OffsetIsScalable;
153       if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable,
154                                        TRI) &&
155           BaseOp->isReg()) {
156         Register BaseReg = BaseOp->getReg();
157         if (PrevBaseReg == BaseReg) {
158           // If this block can take STPs, skip ahead to the next block.
159           if (!SuppressSTP && shouldAddSTPToBlock(MI.getParent()))
160             break;
161           // Otherwise, continue unpairing the stores in this block.
162           LLVM_DEBUG(dbgs() << "Unpairing store " << MI << "\n");
163           SuppressSTP = true;
164           TII->suppressLdStPair(MI);
165         }
166         PrevBaseReg = BaseReg;
167       } else
168         PrevBaseReg = 0;
169     }
170   }
171   // This pass just sets some internal MachineMemOperand flags. It can't really
172   // invalidate anything.
173   return false;
174 }
175