Andrea Di Biagio | 3a6b092 | 2018-03-08 13:05:02 +0000 | [diff] [blame] | 1 | //===------------------------- LSUnit.h --------------------------*- C++-*-===// |
| 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 | /// \file |
| 10 | /// |
| 11 | /// A Load/Store unit class that models load/store queues and that implements |
| 12 | /// a simple weak memory consistency model. |
| 13 | /// |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
| 16 | #ifndef LLVM_TOOLS_LLVM_MCA_LSUNIT_H |
| 17 | #define LLVM_TOOLS_LLVM_MCA_LSUNIT_H |
| 18 | |
| 19 | #include "llvm/Support/Debug.h" |
| 20 | #include "llvm/Support/raw_ostream.h" |
| 21 | #include <set> |
| 22 | |
| 23 | #define DEBUG_TYPE "llvm-mca" |
| 24 | |
| 25 | namespace mca { |
| 26 | |
| 27 | /// \brief A Load/Store Unit implementing a load and store queues. |
| 28 | /// |
| 29 | /// This class implements a load queue and a store queue to emulate the |
| 30 | /// out-of-order execution of memory operations. |
| 31 | /// Each load (or store) consumes an entry in the load (or store) queue. |
| 32 | /// |
| 33 | /// Rules are: |
| 34 | /// 1) A younger load is allowed to pass an older load only if there are no |
| 35 | /// stores nor barriers in between the two loads. |
| 36 | /// 2) An younger store is not allowed to pass an older store. |
| 37 | /// 3) A younger store is not allowed to pass an older load. |
| 38 | /// 4) A younger load is allowed to pass an older store only if the load does |
| 39 | /// not alias with the store. |
| 40 | /// |
| 41 | /// This class optimistically assumes that loads don't alias store operations. |
| 42 | /// Under this assumption, younger loads are always allowed to pass older |
| 43 | /// stores (this would only affects rule 4). |
| 44 | /// Essentially, this LSUnit doesn't attempt to run any sort alias analysis to |
| 45 | /// predict when loads and stores don't alias with eachother. |
| 46 | /// |
| 47 | /// To enforce aliasing between loads and stores, flag `AssumeNoAlias` must be |
| 48 | /// set to `false` by the constructor of LSUnit. |
| 49 | /// |
| 50 | /// In the case of write-combining memory, rule 2. could be relaxed to allow |
| 51 | /// reordering of non-aliasing store operations. At the moment, this is not |
| 52 | /// allowed. |
| 53 | /// To put it in another way, there is no option to specify a different memory |
| 54 | /// type for memory operations (example: write-through, write-combining, etc.). |
| 55 | /// Also, there is no way to weaken the memory model, and this unit currently |
| 56 | /// doesn't support write-combining behavior. |
| 57 | /// |
| 58 | /// No assumptions are made on the size of the store buffer. |
| 59 | /// As mentioned before, this class doesn't perform alias analysis. |
| 60 | /// Consequently, LSUnit doesn't know how to identify cases where |
| 61 | /// store-to-load forwarding may occur. |
| 62 | /// |
| 63 | /// LSUnit doesn't attempt to predict whether a load or store hits or misses |
| 64 | /// the L1 cache. To be more specific, LSUnit doesn't know anything about |
| 65 | /// the cache hierarchy and memory types. |
| 66 | /// It only knows if an instruction "mayLoad" and/or "mayStore". For loads, the |
| 67 | /// scheduling model provides an "optimistic" load-to-use latency (which usually |
| 68 | /// matches the load-to-use latency for when there is a hit in the L1D). |
| 69 | /// |
| 70 | /// Class MCInstrDesc in LLVM doesn't know about serializing operations, nor |
| 71 | /// memory-barrier like instructions. |
| 72 | /// LSUnit conservatively assumes that an instruction which `mayLoad` and has |
| 73 | /// `unmodeled side effects` behave like a "soft" load-barrier. That means, it |
| 74 | /// serializes loads without forcing a flush of the load queue. |
| 75 | /// Similarly, instructions that both `mayStore` and have `unmodeled side |
| 76 | /// effects` are treated like store barriers. A full memory |
| 77 | /// barrier is a 'mayLoad' and 'mayStore' instruction with unmodeled side |
| 78 | /// effects. This is obviously inaccurate, but this is the best that we can do |
| 79 | /// at the moment. |
| 80 | /// |
| 81 | /// Each load/store barrier consumes one entry in the load/store queue. A |
| 82 | /// load/store barrier enforces ordering of loads/stores: |
| 83 | /// - A younger load cannot pass a load barrier. |
| 84 | /// - A younger store cannot pass a store barrier. |
| 85 | /// |
| 86 | /// A younger load has to wait for the memory load barrier to execute. |
| 87 | /// A load/store barrier is "executed" when it becomes the oldest entry in |
| 88 | /// the load/store queue(s). That also means, all the older loads/stores have |
| 89 | /// already been executed. |
| 90 | class LSUnit { |
| 91 | // Load queue size. |
| 92 | // LQ_Size == 0 means that there are infinite slots in the load queue. |
| 93 | unsigned LQ_Size; |
| 94 | |
| 95 | // Store queue size. |
| 96 | // SQ_Size == 0 means that there are infinite slots in the store queue. |
| 97 | unsigned SQ_Size; |
| 98 | |
| 99 | // If true, loads will never alias with stores. This is the default. |
| 100 | bool NoAlias; |
| 101 | |
| 102 | std::set<unsigned> LoadQueue; |
| 103 | std::set<unsigned> StoreQueue; |
| 104 | |
| 105 | void assignLQSlot(unsigned Index); |
| 106 | void assignSQSlot(unsigned Index); |
| 107 | bool isReadyNoAlias(unsigned Index) const; |
| 108 | |
| 109 | // An instruction that both 'mayStore' and 'HasUnmodeledSideEffects' is |
| 110 | // conservatively treated as a store barrier. It forces older store to be |
| 111 | // executed before newer stores are issued. |
| 112 | std::set<unsigned> StoreBarriers; |
| 113 | |
| 114 | // An instruction that both 'MayLoad' and 'HasUnmodeledSideEffects' is |
| 115 | // conservatively treated as a load barrier. It forces older loads to execute |
| 116 | // before newer loads are issued. |
| 117 | std::set<unsigned> LoadBarriers; |
| 118 | |
| 119 | public: |
| 120 | LSUnit(unsigned LQ = 0, unsigned SQ = 0, bool AssumeNoAlias = false) |
| 121 | : LQ_Size(LQ), SQ_Size(SQ), NoAlias(AssumeNoAlias) {} |
| 122 | |
| 123 | #ifndef NDEBUG |
| 124 | void dump() const; |
| 125 | #endif |
| 126 | |
| 127 | bool isSQEmpty() const { return StoreQueue.empty(); } |
| 128 | bool isLQEmpty() const { return LoadQueue.empty(); } |
| 129 | bool isSQFull() const { return SQ_Size != 0 && StoreQueue.size() == SQ_Size; } |
| 130 | bool isLQFull() const { return LQ_Size != 0 && LoadQueue.size() == LQ_Size; } |
| 131 | |
| 132 | void reserve(unsigned Index, bool MayLoad, bool MayStore, bool IsMemBarrier) { |
| 133 | if (!MayLoad && !MayStore) |
| 134 | return; |
| 135 | if (MayLoad) { |
| 136 | if (IsMemBarrier) |
| 137 | LoadBarriers.insert(Index); |
| 138 | assignLQSlot(Index); |
| 139 | } |
| 140 | if (MayStore) { |
| 141 | if (IsMemBarrier) |
| 142 | StoreBarriers.insert(Index); |
| 143 | assignSQSlot(Index); |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | // The rules are: |
| 148 | // 1. A store may not pass a previous store. |
| 149 | // 2. A load may not pass a previous store unless flag 'NoAlias' is set. |
| 150 | // 3. A load may pass a previous load. |
| 151 | // 4. A store may not pass a previous load (regardless of flag 'NoAlias'). |
| 152 | // 5. A load has to wait until an older load barrier is fully executed. |
| 153 | // 6. A store has to wait until an older store barrier is fully executed. |
| 154 | bool isReady(unsigned Index) const; |
| 155 | void onInstructionExecuted(unsigned Index); |
| 156 | }; |
| 157 | |
| 158 | } // namespace mca |
| 159 | |
| 160 | #endif |