Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | //==-- llvm/CodeGen/ExecutionDomainFix.h - Execution Domain Fix -*- C++ -*--==// |
| 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 | /// \file Execution Domain Fix pass. |
||
| 10 | /// |
||
| 11 | /// Some X86 SSE instructions like mov, and, or, xor are available in different |
||
| 12 | /// variants for different operand types. These variant instructions are |
||
| 13 | /// equivalent, but on Nehalem and newer cpus there is extra latency |
||
| 14 | /// transferring data between integer and floating point domains. ARM cores |
||
| 15 | /// have similar issues when they are configured with both VFP and NEON |
||
| 16 | /// pipelines. |
||
| 17 | /// |
||
| 18 | /// This pass changes the variant instructions to minimize domain crossings. |
||
| 19 | // |
||
| 20 | //===----------------------------------------------------------------------===// |
||
| 21 | |||
| 22 | #ifndef LLVM_CODEGEN_EXECUTIONDOMAINFIX_H |
||
| 23 | #define LLVM_CODEGEN_EXECUTIONDOMAINFIX_H |
||
| 24 | |||
| 25 | #include "llvm/ADT/SmallVector.h" |
||
| 26 | #include "llvm/CodeGen/LoopTraversal.h" |
||
| 27 | #include "llvm/CodeGen/MachineFunctionPass.h" |
||
| 28 | #include "llvm/CodeGen/ReachingDefAnalysis.h" |
||
| 29 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
||
| 30 | |||
| 31 | namespace llvm { |
||
| 32 | |||
| 33 | class MachineInstr; |
||
| 34 | class TargetInstrInfo; |
||
| 35 | |||
| 36 | /// A DomainValue is a bit like LiveIntervals' ValNo, but it also keeps track |
||
| 37 | /// of execution domains. |
||
| 38 | /// |
||
| 39 | /// An open DomainValue represents a set of instructions that can still switch |
||
| 40 | /// execution domain. Multiple registers may refer to the same open |
||
| 41 | /// DomainValue - they will eventually be collapsed to the same execution |
||
| 42 | /// domain. |
||
| 43 | /// |
||
| 44 | /// A collapsed DomainValue represents a single register that has been forced |
||
| 45 | /// into one of more execution domains. There is a separate collapsed |
||
| 46 | /// DomainValue for each register, but it may contain multiple execution |
||
| 47 | /// domains. A register value is initially created in a single execution |
||
| 48 | /// domain, but if we were forced to pay the penalty of a domain crossing, we |
||
| 49 | /// keep track of the fact that the register is now available in multiple |
||
| 50 | /// domains. |
||
| 51 | struct DomainValue { |
||
| 52 | /// Basic reference counting. |
||
| 53 | unsigned Refs = 0; |
||
| 54 | |||
| 55 | /// Bitmask of available domains. For an open DomainValue, it is the still |
||
| 56 | /// possible domains for collapsing. For a collapsed DomainValue it is the |
||
| 57 | /// domains where the register is available for free. |
||
| 58 | unsigned AvailableDomains; |
||
| 59 | |||
| 60 | /// Pointer to the next DomainValue in a chain. When two DomainValues are |
||
| 61 | /// merged, Victim.Next is set to point to Victor, so old DomainValue |
||
| 62 | /// references can be updated by following the chain. |
||
| 63 | DomainValue *Next; |
||
| 64 | |||
| 65 | /// Twiddleable instructions using or defining these registers. |
||
| 66 | SmallVector<MachineInstr *, 8> Instrs; |
||
| 67 | |||
| 68 | DomainValue() { clear(); } |
||
| 69 | |||
| 70 | /// A collapsed DomainValue has no instructions to twiddle - it simply keeps |
||
| 71 | /// track of the domains where the registers are already available. |
||
| 72 | bool isCollapsed() const { return Instrs.empty(); } |
||
| 73 | |||
| 74 | /// Is domain available? |
||
| 75 | bool hasDomain(unsigned domain) const { |
||
| 76 | assert(domain < |
||
| 77 | static_cast<unsigned>(std::numeric_limits<unsigned>::digits) && |
||
| 78 | "undefined behavior"); |
||
| 79 | return AvailableDomains & (1u << domain); |
||
| 80 | } |
||
| 81 | |||
| 82 | /// Mark domain as available. |
||
| 83 | void addDomain(unsigned domain) { |
||
| 84 | assert(domain < |
||
| 85 | static_cast<unsigned>(std::numeric_limits<unsigned>::digits) && |
||
| 86 | "undefined behavior"); |
||
| 87 | AvailableDomains |= 1u << domain; |
||
| 88 | } |
||
| 89 | |||
| 90 | // Restrict to a single domain available. |
||
| 91 | void setSingleDomain(unsigned domain) { |
||
| 92 | assert(domain < |
||
| 93 | static_cast<unsigned>(std::numeric_limits<unsigned>::digits) && |
||
| 94 | "undefined behavior"); |
||
| 95 | AvailableDomains = 1u << domain; |
||
| 96 | } |
||
| 97 | |||
| 98 | /// Return bitmask of domains that are available and in mask. |
||
| 99 | unsigned getCommonDomains(unsigned mask) const { |
||
| 100 | return AvailableDomains & mask; |
||
| 101 | } |
||
| 102 | |||
| 103 | /// First domain available. |
||
| 104 | unsigned getFirstDomain() const { |
||
| 105 | return countTrailingZeros(AvailableDomains); |
||
| 106 | } |
||
| 107 | |||
| 108 | /// Clear this DomainValue and point to next which has all its data. |
||
| 109 | void clear() { |
||
| 110 | AvailableDomains = 0; |
||
| 111 | Next = nullptr; |
||
| 112 | Instrs.clear(); |
||
| 113 | } |
||
| 114 | }; |
||
| 115 | |||
| 116 | class ExecutionDomainFix : public MachineFunctionPass { |
||
| 117 | SpecificBumpPtrAllocator<DomainValue> Allocator; |
||
| 118 | SmallVector<DomainValue *, 16> Avail; |
||
| 119 | |||
| 120 | const TargetRegisterClass *const RC; |
||
| 121 | MachineFunction *MF; |
||
| 122 | const TargetInstrInfo *TII; |
||
| 123 | const TargetRegisterInfo *TRI; |
||
| 124 | std::vector<SmallVector<int, 1>> AliasMap; |
||
| 125 | const unsigned NumRegs; |
||
| 126 | /// Value currently in each register, or NULL when no value is being tracked. |
||
| 127 | /// This counts as a DomainValue reference. |
||
| 128 | using LiveRegsDVInfo = std::vector<DomainValue *>; |
||
| 129 | LiveRegsDVInfo LiveRegs; |
||
| 130 | /// Keeps domain information for all registers. Note that this |
||
| 131 | /// is different from the usual definition notion of liveness. The CPU |
||
| 132 | /// doesn't care whether or not we consider a register killed. |
||
| 133 | using OutRegsInfoMap = SmallVector<LiveRegsDVInfo, 4>; |
||
| 134 | OutRegsInfoMap MBBOutRegsInfos; |
||
| 135 | |||
| 136 | ReachingDefAnalysis *RDA; |
||
| 137 | |||
| 138 | public: |
||
| 139 | ExecutionDomainFix(char &PassID, const TargetRegisterClass &RC) |
||
| 140 | : MachineFunctionPass(PassID), RC(&RC), NumRegs(RC.getNumRegs()) {} |
||
| 141 | |||
| 142 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
||
| 143 | AU.setPreservesAll(); |
||
| 144 | AU.addRequired<ReachingDefAnalysis>(); |
||
| 145 | MachineFunctionPass::getAnalysisUsage(AU); |
||
| 146 | } |
||
| 147 | |||
| 148 | bool runOnMachineFunction(MachineFunction &MF) override; |
||
| 149 | |||
| 150 | MachineFunctionProperties getRequiredProperties() const override { |
||
| 151 | return MachineFunctionProperties().set( |
||
| 152 | MachineFunctionProperties::Property::NoVRegs); |
||
| 153 | } |
||
| 154 | |||
| 155 | private: |
||
| 156 | /// Translate TRI register number to a list of indices into our smaller tables |
||
| 157 | /// of interesting registers. |
||
| 158 | iterator_range<SmallVectorImpl<int>::const_iterator> |
||
| 159 | regIndices(unsigned Reg) const; |
||
| 160 | |||
| 161 | /// DomainValue allocation. |
||
| 162 | DomainValue *alloc(int domain = -1); |
||
| 163 | |||
| 164 | /// Add reference to DV. |
||
| 165 | DomainValue *retain(DomainValue *DV) { |
||
| 166 | if (DV) |
||
| 167 | ++DV->Refs; |
||
| 168 | return DV; |
||
| 169 | } |
||
| 170 | |||
| 171 | /// Release a reference to DV. When the last reference is released, |
||
| 172 | /// collapse if needed. |
||
| 173 | void release(DomainValue *); |
||
| 174 | |||
| 175 | /// Follow the chain of dead DomainValues until a live DomainValue is reached. |
||
| 176 | /// Update the referenced pointer when necessary. |
||
| 177 | DomainValue *resolve(DomainValue *&); |
||
| 178 | |||
| 179 | /// Set LiveRegs[rx] = dv, updating reference counts. |
||
| 180 | void setLiveReg(int rx, DomainValue *DV); |
||
| 181 | |||
| 182 | /// Kill register rx, recycle or collapse any DomainValue. |
||
| 183 | void kill(int rx); |
||
| 184 | |||
| 185 | /// Force register rx into domain. |
||
| 186 | void force(int rx, unsigned domain); |
||
| 187 | |||
| 188 | /// Collapse open DomainValue into given domain. If there are multiple |
||
| 189 | /// registers using dv, they each get a unique collapsed DomainValue. |
||
| 190 | void collapse(DomainValue *dv, unsigned domain); |
||
| 191 | |||
| 192 | /// All instructions and registers in B are moved to A, and B is released. |
||
| 193 | bool merge(DomainValue *A, DomainValue *B); |
||
| 194 | |||
| 195 | /// Set up LiveRegs by merging predecessor live-out values. |
||
| 196 | void enterBasicBlock(const LoopTraversal::TraversedMBBInfo &TraversedMBB); |
||
| 197 | |||
| 198 | /// Update live-out values. |
||
| 199 | void leaveBasicBlock(const LoopTraversal::TraversedMBBInfo &TraversedMBB); |
||
| 200 | |||
| 201 | /// Process he given basic block. |
||
| 202 | void processBasicBlock(const LoopTraversal::TraversedMBBInfo &TraversedMBB); |
||
| 203 | |||
| 204 | /// Visit given insturcion. |
||
| 205 | bool visitInstr(MachineInstr *); |
||
| 206 | |||
| 207 | /// Update def-ages for registers defined by MI. |
||
| 208 | /// If Kill is set, also kill off DomainValues clobbered by the defs. |
||
| 209 | void processDefs(MachineInstr *, bool Kill); |
||
| 210 | |||
| 211 | /// A soft instruction can be changed to work in other domains given by mask. |
||
| 212 | void visitSoftInstr(MachineInstr *, unsigned mask); |
||
| 213 | |||
| 214 | /// A hard instruction only works in one domain. All input registers will be |
||
| 215 | /// forced into that domain. |
||
| 216 | void visitHardInstr(MachineInstr *, unsigned domain); |
||
| 217 | }; |
||
| 218 | |||
| 219 | } // namespace llvm |
||
| 220 | |||
| 221 | #endif // LLVM_CODEGEN_EXECUTIONDOMAINFIX_H |