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 | |
43 | StringRef getPassName() const override { return STPSUPPRESS_PASS_NAME; } |
44 | |
45 | bool runOnMachineFunction(MachineFunction &F) override; |
46 | |
47 | private: |
48 | bool shouldAddSTPToBlock(const MachineBasicBlock *BB); |
49 | |
50 | bool isNarrowFPStore(const MachineInstr &MI); |
51 | |
52 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
53 | AU.setPreservesCFG(); |
54 | AU.addRequired<MachineTraceMetricsWrapperPass>(); |
55 | AU.addPreserved<MachineTraceMetricsWrapperPass>(); |
56 | MachineFunctionPass::getAnalysisUsage(AU); |
57 | } |
58 | }; |
59 | char AArch64StorePairSuppress::ID = 0; |
60 | } // anonymous |
61 | |
62 | INITIALIZE_PASS(AArch64StorePairSuppress, "aarch64-stp-suppress" , |
63 | STPSUPPRESS_PASS_NAME, false, false) |
64 | |
65 | FunctionPass *llvm::createAArch64StorePairSuppressPass() { |
66 | return new AArch64StorePairSuppress(); |
67 | } |
68 | |
69 | /// Return true if an STP can be added to this block without increasing the |
70 | /// critical resource height. STP is good to form in Ld/St limited blocks and |
71 | /// bad to form in float-point limited blocks. This is true independent of the |
72 | /// critical path. If the critical path is longer than the resource height, the |
73 | /// extra vector ops can limit physreg renaming. Otherwise, it could simply |
74 | /// oversaturate the vector units. |
75 | bool AArch64StorePairSuppress::shouldAddSTPToBlock(const MachineBasicBlock *BB) { |
76 | if (!MinInstr) |
77 | MinInstr = Traces->getEnsemble(MachineTraceStrategy::TS_MinInstrCount); |
78 | |
79 | MachineTraceMetrics::Trace BBTrace = MinInstr->getTrace(MBB: BB); |
80 | unsigned ResLength = BBTrace.getResourceLength(); |
81 | |
82 | // Get the machine model's scheduling class for STPDi and STRDui. |
83 | // Bypass TargetSchedule's SchedClass resolution since we only have an opcode. |
84 | unsigned SCIdx = TII->get(Opcode: AArch64::STPDi).getSchedClass(); |
85 | const MCSchedClassDesc *PairSCDesc = |
86 | SchedModel.getMCSchedModel()->getSchedClassDesc(SchedClassIdx: SCIdx); |
87 | |
88 | unsigned SCIdx2 = TII->get(Opcode: AArch64::STRDui).getSchedClass(); |
89 | const MCSchedClassDesc *SingleSCDesc = |
90 | SchedModel.getMCSchedModel()->getSchedClassDesc(SchedClassIdx: SCIdx2); |
91 | |
92 | // If a subtarget does not define resources for STPDi, bail here. |
93 | if (PairSCDesc->isValid() && !PairSCDesc->isVariant() && |
94 | SingleSCDesc->isValid() && !SingleSCDesc->isVariant()) { |
95 | // Compute the new critical resource length after replacing 2 separate |
96 | // STRDui with one STPDi. |
97 | unsigned ResLenWithSTP = |
98 | BBTrace.getResourceLength(Extrablocks: {}, ExtraInstrs: PairSCDesc, RemoveInstrs: {SingleSCDesc, SingleSCDesc}); |
99 | if (ResLenWithSTP > ResLength) { |
100 | LLVM_DEBUG(dbgs() << " Suppress STP in BB: " << BB->getNumber() |
101 | << " resources " << ResLength << " -> " << ResLenWithSTP |
102 | << "\n" ); |
103 | return false; |
104 | } |
105 | } |
106 | return true; |
107 | } |
108 | |
109 | /// Return true if this is a floating-point store smaller than the V reg. On |
110 | /// cyclone, these require a vector shuffle before storing a pair. |
111 | /// Ideally we would call getMatchingPairOpcode() and have the machine model |
112 | /// tell us if it's profitable with no cpu knowledge here. |
113 | /// |
114 | /// FIXME: We plan to develop a decent Target abstraction for simple loads and |
115 | /// stores. Until then use a nasty switch similar to AArch64LoadStoreOptimizer. |
116 | bool AArch64StorePairSuppress::isNarrowFPStore(const MachineInstr &MI) { |
117 | switch (MI.getOpcode()) { |
118 | default: |
119 | return false; |
120 | case AArch64::STRSui: |
121 | case AArch64::STRDui: |
122 | case AArch64::STURSi: |
123 | case AArch64::STURDi: |
124 | return true; |
125 | } |
126 | } |
127 | |
128 | bool AArch64StorePairSuppress::runOnMachineFunction(MachineFunction &MF) { |
129 | if (skipFunction(F: MF.getFunction()) || MF.getFunction().hasOptSize()) |
130 | return false; |
131 | |
132 | const AArch64Subtarget &ST = MF.getSubtarget<AArch64Subtarget>(); |
133 | if (!ST.enableStorePairSuppress()) |
134 | return false; |
135 | |
136 | TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo()); |
137 | TRI = ST.getRegisterInfo(); |
138 | MRI = &MF.getRegInfo(); |
139 | SchedModel.init(TSInfo: &ST); |
140 | Traces = &getAnalysis<MachineTraceMetricsWrapperPass>().getMTM(); |
141 | MinInstr = nullptr; |
142 | |
143 | LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << MF.getName() << '\n'); |
144 | |
145 | if (!SchedModel.hasInstrSchedModel()) { |
146 | LLVM_DEBUG(dbgs() << " Skipping pass: no machine model present.\n" ); |
147 | return false; |
148 | } |
149 | |
150 | // Check for a sequence of stores to the same base address. We don't need to |
151 | // precisely determine whether a store pair can be formed. But we do want to |
152 | // filter out most situations where we can't form store pairs to avoid |
153 | // computing trace metrics in those cases. |
154 | for (auto &MBB : MF) { |
155 | bool SuppressSTP = false; |
156 | unsigned PrevBaseReg = 0; |
157 | for (auto &MI : MBB) { |
158 | if (!isNarrowFPStore(MI)) |
159 | continue; |
160 | const MachineOperand *BaseOp; |
161 | int64_t Offset; |
162 | bool OffsetIsScalable; |
163 | if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, |
164 | TRI) && |
165 | BaseOp->isReg()) { |
166 | Register BaseReg = BaseOp->getReg(); |
167 | if (PrevBaseReg == BaseReg) { |
168 | // If this block can take STPs, skip ahead to the next block. |
169 | if (!SuppressSTP && shouldAddSTPToBlock(BB: MI.getParent())) |
170 | break; |
171 | // Otherwise, continue unpairing the stores in this block. |
172 | LLVM_DEBUG(dbgs() << "Unpairing store " << MI << "\n" ); |
173 | SuppressSTP = true; |
174 | TII->suppressLdStPair(MI); |
175 | } |
176 | PrevBaseReg = BaseReg; |
177 | } else |
178 | PrevBaseReg = 0; |
179 | } |
180 | } |
181 | // This pass just sets some internal MachineMemOperand flags. It can't really |
182 | // invalidate anything. |
183 | return false; |
184 | } |
185 | |