1 //===--- AArch64StorePairSuppress.cpp --- Suppress store pair formation ---===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass identifies floating point stores that should not be combined into
11 // store pairs. Later we may do the same for floating point loads.
12 // ===---------------------------------------------------------------------===//
13 
14 #include "AArch64InstrInfo.h"
15 #include "llvm/CodeGen/MachineFunction.h"
16 #include "llvm/CodeGen/MachineFunctionPass.h"
17 #include "llvm/CodeGen/MachineInstr.h"
18 #include "llvm/CodeGen/MachineTraceMetrics.h"
19 #include "llvm/CodeGen/TargetInstrInfo.h"
20 #include "llvm/CodeGen/TargetSchedule.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/raw_ostream.h"
23 
24 using namespace llvm;
25 
26 #define DEBUG_TYPE "aarch64-stp-suppress"
27 
28 #define STPSUPPRESS_PASS_NAME "AArch64 Store Pair Suppression"
29 
30 namespace {
31 class AArch64StorePairSuppress : public MachineFunctionPass {
32   const AArch64InstrInfo *TII;
33   const TargetRegisterInfo *TRI;
34   const MachineRegisterInfo *MRI;
35   TargetSchedModel SchedModel;
36   MachineTraceMetrics *Traces;
37   MachineTraceMetrics::Ensemble *MinInstr;
38 
39 public:
40   static char ID;
AArch64StorePairSuppress()41   AArch64StorePairSuppress() : MachineFunctionPass(ID) {
42     initializeAArch64StorePairSuppressPass(*PassRegistry::getPassRegistry());
43   }
44 
getPassName() const45   StringRef getPassName() const override { return STPSUPPRESS_PASS_NAME; }
46 
47   bool runOnMachineFunction(MachineFunction &F) override;
48 
49 private:
50   bool shouldAddSTPToBlock(const MachineBasicBlock *BB);
51 
52   bool isNarrowFPStore(const MachineInstr &MI);
53 
getAnalysisUsage(AnalysisUsage & AU) const54   void getAnalysisUsage(AnalysisUsage &AU) const override {
55     AU.setPreservesCFG();
56     AU.addRequired<MachineTraceMetrics>();
57     AU.addPreserved<MachineTraceMetrics>();
58     MachineFunctionPass::getAnalysisUsage(AU);
59   }
60 };
61 char AArch64StorePairSuppress::ID = 0;
62 } // anonymous
63 
64 INITIALIZE_PASS(AArch64StorePairSuppress, "aarch64-stp-suppress",
65                 STPSUPPRESS_PASS_NAME, false, false)
66 
createAArch64StorePairSuppressPass()67 FunctionPass *llvm::createAArch64StorePairSuppressPass() {
68   return new AArch64StorePairSuppress();
69 }
70 
71 /// Return true if an STP can be added to this block without increasing the
72 /// critical resource height. STP is good to form in Ld/St limited blocks and
73 /// bad to form in float-point limited blocks. This is true independent of the
74 /// critical path. If the critical path is longer than the resource height, the
75 /// extra vector ops can limit physreg renaming. Otherwise, it could simply
76 /// oversaturate the vector units.
shouldAddSTPToBlock(const MachineBasicBlock * BB)77 bool AArch64StorePairSuppress::shouldAddSTPToBlock(const MachineBasicBlock *BB) {
78   if (!MinInstr)
79     MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount);
80 
81   MachineTraceMetrics::Trace BBTrace = MinInstr->getTrace(BB);
82   unsigned ResLength = BBTrace.getResourceLength();
83 
84   // Get the machine model's scheduling class for STPQi.
85   // Bypass TargetSchedule's SchedClass resolution since we only have an opcode.
86   unsigned SCIdx = TII->get(AArch64::STPDi).getSchedClass();
87   const MCSchedClassDesc *SCDesc =
88       SchedModel.getMCSchedModel()->getSchedClassDesc(SCIdx);
89 
90   // If a subtarget does not define resources for STPQi, bail here.
91   if (SCDesc->isValid() && !SCDesc->isVariant()) {
92     unsigned ResLenWithSTP = BBTrace.getResourceLength(None, SCDesc);
93     if (ResLenWithSTP > ResLength) {
94       LLVM_DEBUG(dbgs() << "  Suppress STP in BB: " << BB->getNumber()
95                         << " resources " << ResLength << " -> " << ResLenWithSTP
96                         << "\n");
97       return false;
98     }
99   }
100   return true;
101 }
102 
103 /// Return true if this is a floating-point store smaller than the V reg. On
104 /// cyclone, these require a vector shuffle before storing a pair.
105 /// Ideally we would call getMatchingPairOpcode() and have the machine model
106 /// tell us if it's profitable with no cpu knowledge here.
107 ///
108 /// FIXME: We plan to develop a decent Target abstraction for simple loads and
109 /// stores. Until then use a nasty switch similar to AArch64LoadStoreOptimizer.
isNarrowFPStore(const MachineInstr & MI)110 bool AArch64StorePairSuppress::isNarrowFPStore(const MachineInstr &MI) {
111   switch (MI.getOpcode()) {
112   default:
113     return false;
114   case AArch64::STRSui:
115   case AArch64::STRDui:
116   case AArch64::STURSi:
117   case AArch64::STURDi:
118     return true;
119   }
120 }
121 
runOnMachineFunction(MachineFunction & MF)122 bool AArch64StorePairSuppress::runOnMachineFunction(MachineFunction &MF) {
123   if (skipFunction(MF.getFunction()))
124     return false;
125 
126   const TargetSubtargetInfo &ST = MF.getSubtarget();
127   TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo());
128   TRI = ST.getRegisterInfo();
129   MRI = &MF.getRegInfo();
130   SchedModel.init(&ST);
131   Traces = &getAnalysis<MachineTraceMetrics>();
132   MinInstr = nullptr;
133 
134   LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << MF.getName() << '\n');
135 
136   if (!SchedModel.hasInstrSchedModel()) {
137     LLVM_DEBUG(dbgs() << "  Skipping pass: no machine model present.\n");
138     return false;
139   }
140 
141   // Check for a sequence of stores to the same base address. We don't need to
142   // precisely determine whether a store pair can be formed. But we do want to
143   // filter out most situations where we can't form store pairs to avoid
144   // computing trace metrics in those cases.
145   for (auto &MBB : MF) {
146     bool SuppressSTP = false;
147     unsigned PrevBaseReg = 0;
148     for (auto &MI : MBB) {
149       if (!isNarrowFPStore(MI))
150         continue;
151       MachineOperand *BaseOp;
152       int64_t Offset;
153       if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, TRI) &&
154           BaseOp->isReg()) {
155         unsigned BaseReg = BaseOp->getReg();
156         if (PrevBaseReg == BaseReg) {
157           // If this block can take STPs, skip ahead to the next block.
158           if (!SuppressSTP && shouldAddSTPToBlock(MI.getParent()))
159             break;
160           // Otherwise, continue unpairing the stores in this block.
161           LLVM_DEBUG(dbgs() << "Unpairing store " << MI << "\n");
162           SuppressSTP = true;
163           TII->suppressLdStPair(MI);
164         }
165         PrevBaseReg = BaseReg;
166       } else
167         PrevBaseReg = 0;
168     }
169   }
170   // This pass just sets some internal MachineMemOperand flags. It can't really
171   // invalidate anything.
172   return false;
173 }
174