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 "AArch64Subtarget.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; |
41 | AArch64StorePairSuppress() : MachineFunctionPass(ID) { |
42 | initializeAArch64StorePairSuppressPass(*PassRegistry::getPassRegistry()); |
43 | } |
44 | |
45 | 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 | |
54 | 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 | |
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. |
77 | bool AArch64StorePairSuppress::shouldAddSTPToBlock(const MachineBasicBlock *BB) { |
78 | if (!MinInstr) |
79 | MinInstr = Traces->getEnsemble(MachineTraceStrategy::TS_MinInstrCount); |
80 | |
81 | MachineTraceMetrics::Trace BBTrace = MinInstr->getTrace(MBB: BB); |
82 | unsigned ResLength = BBTrace.getResourceLength(); |
83 | |
84 | // Get the machine model's scheduling class for STPDi and STRDui. |
85 | // Bypass TargetSchedule's SchedClass resolution since we only have an opcode. |
86 | unsigned SCIdx = TII->get(Opcode: AArch64::STPDi).getSchedClass(); |
87 | const MCSchedClassDesc *PairSCDesc = |
88 | SchedModel.getMCSchedModel()->getSchedClassDesc(SchedClassIdx: SCIdx); |
89 | |
90 | unsigned SCIdx2 = TII->get(Opcode: AArch64::STRDui).getSchedClass(); |
91 | const MCSchedClassDesc *SingleSCDesc = |
92 | SchedModel.getMCSchedModel()->getSchedClassDesc(SchedClassIdx: SCIdx2); |
93 | |
94 | // If a subtarget does not define resources for STPDi, bail here. |
95 | if (PairSCDesc->isValid() && !PairSCDesc->isVariant() && |
96 | SingleSCDesc->isValid() && !SingleSCDesc->isVariant()) { |
97 | // Compute the new critical resource length after replacing 2 separate |
98 | // STRDui with one STPDi. |
99 | unsigned ResLenWithSTP = BBTrace.getResourceLength( |
100 | Extrablocks: std::nullopt, ExtraInstrs: PairSCDesc, RemoveInstrs: {SingleSCDesc, SingleSCDesc}); |
101 | if (ResLenWithSTP > ResLength) { |
102 | LLVM_DEBUG(dbgs() << " Suppress STP in BB: " << BB->getNumber() |
103 | << " resources " << ResLength << " -> " << ResLenWithSTP |
104 | << "\n" ); |
105 | return false; |
106 | } |
107 | } |
108 | return true; |
109 | } |
110 | |
111 | /// Return true if this is a floating-point store smaller than the V reg. On |
112 | /// cyclone, these require a vector shuffle before storing a pair. |
113 | /// Ideally we would call getMatchingPairOpcode() and have the machine model |
114 | /// tell us if it's profitable with no cpu knowledge here. |
115 | /// |
116 | /// FIXME: We plan to develop a decent Target abstraction for simple loads and |
117 | /// stores. Until then use a nasty switch similar to AArch64LoadStoreOptimizer. |
118 | bool AArch64StorePairSuppress::isNarrowFPStore(const MachineInstr &MI) { |
119 | switch (MI.getOpcode()) { |
120 | default: |
121 | return false; |
122 | case AArch64::STRSui: |
123 | case AArch64::STRDui: |
124 | case AArch64::STURSi: |
125 | case AArch64::STURDi: |
126 | return true; |
127 | } |
128 | } |
129 | |
130 | bool AArch64StorePairSuppress::runOnMachineFunction(MachineFunction &MF) { |
131 | if (skipFunction(F: MF.getFunction()) || MF.getFunction().hasOptSize()) |
132 | return false; |
133 | |
134 | const AArch64Subtarget &ST = MF.getSubtarget<AArch64Subtarget>(); |
135 | if (!ST.enableStorePairSuppress()) |
136 | return false; |
137 | |
138 | TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo()); |
139 | TRI = ST.getRegisterInfo(); |
140 | MRI = &MF.getRegInfo(); |
141 | SchedModel.init(TSInfo: &ST); |
142 | Traces = &getAnalysis<MachineTraceMetrics>(); |
143 | MinInstr = nullptr; |
144 | |
145 | LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << MF.getName() << '\n'); |
146 | |
147 | if (!SchedModel.hasInstrSchedModel()) { |
148 | LLVM_DEBUG(dbgs() << " Skipping pass: no machine model present.\n" ); |
149 | return false; |
150 | } |
151 | |
152 | // Check for a sequence of stores to the same base address. We don't need to |
153 | // precisely determine whether a store pair can be formed. But we do want to |
154 | // filter out most situations where we can't form store pairs to avoid |
155 | // computing trace metrics in those cases. |
156 | for (auto &MBB : MF) { |
157 | bool SuppressSTP = false; |
158 | unsigned PrevBaseReg = 0; |
159 | for (auto &MI : MBB) { |
160 | if (!isNarrowFPStore(MI)) |
161 | continue; |
162 | const MachineOperand *BaseOp; |
163 | int64_t Offset; |
164 | bool OffsetIsScalable; |
165 | if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, |
166 | TRI) && |
167 | BaseOp->isReg()) { |
168 | Register BaseReg = BaseOp->getReg(); |
169 | if (PrevBaseReg == BaseReg) { |
170 | // If this block can take STPs, skip ahead to the next block. |
171 | if (!SuppressSTP && shouldAddSTPToBlock(BB: MI.getParent())) |
172 | break; |
173 | // Otherwise, continue unpairing the stores in this block. |
174 | LLVM_DEBUG(dbgs() << "Unpairing store " << MI << "\n" ); |
175 | SuppressSTP = true; |
176 | TII->suppressLdStPair(MI); |
177 | } |
178 | PrevBaseReg = BaseReg; |
179 | } else |
180 | PrevBaseReg = 0; |
181 | } |
182 | } |
183 | // This pass just sets some internal MachineMemOperand flags. It can't really |
184 | // invalidate anything. |
185 | return false; |
186 | } |
187 | |