1 //===--------------------- TimelineView.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 /// \brief 10 /// 11 /// This file implements a timeline view for the llvm-mca tool. 12 /// 13 /// Class TimelineView observes events generated by the pipeline. For every 14 /// instruction executed by the pipeline, it stores information related to 15 /// state transition. It then plots that information in the form of a table 16 /// as reported by the example below: 17 /// 18 /// Timeline view: 19 /// 0123456 20 /// Index 0123456789 21 /// 22 /// [0,0] DeER . . .. vmovshdup %xmm0, %xmm1 23 /// [0,1] DeER . . .. vpermilpd $1, %xmm0, %xmm2 24 /// [0,2] .DeER. . .. vpermilps $231, %xmm0, %xmm5 25 /// [0,3] .DeeeER . .. vaddss %xmm1, %xmm0, %xmm3 26 /// [0,4] . D==eeeER. .. vaddss %xmm3, %xmm2, %xmm4 27 /// [0,5] . D=====eeeER .. vaddss %xmm4, %xmm5, %xmm6 28 /// 29 /// [1,0] . DeE------R .. vmovshdup %xmm0, %xmm1 30 /// [1,1] . DeE------R .. vpermilpd $1, %xmm0, %xmm2 31 /// [1,2] . DeE-----R .. vpermilps $231, %xmm0, %xmm5 32 /// [1,3] . D=eeeE--R .. vaddss %xmm1, %xmm0, %xmm3 33 /// [1,4] . D===eeeER .. vaddss %xmm3, %xmm2, %xmm4 34 /// [1,5] . D======eeeER vaddss %xmm4, %xmm5, %xmm6 35 /// 36 /// There is an entry for every instruction in the input assembly sequence. 37 /// The first field is a pair of numbers obtained from the instruction index. 38 /// The first element of the pair is the iteration index, while the second 39 /// element of the pair is a sequence number (i.e. a position in the assembly 40 /// sequence). 41 /// The second field of the table is the actual timeline information; each 42 /// column is the information related to a specific cycle of execution. 43 /// The timeline of an instruction is described by a sequence of character 44 /// where each character represents the instruction state at a specific cycle. 45 /// 46 /// Possible instruction states are: 47 /// D: Instruction Dispatched 48 /// e: Instruction Executing 49 /// E: Instruction Executed (write-back stage) 50 /// R: Instruction retired 51 /// =: Instruction waiting in the Scheduler's queue 52 /// -: Instruction executed, waiting to retire in order. 53 /// 54 /// dots ('.') and empty spaces are cycles where the instruction is not 55 /// in-flight. 56 /// 57 /// The last column is the assembly instruction associated to the entry. 58 /// 59 /// Based on the timeline view information from the example, instruction 0 60 /// at iteration 0 was dispatched at cycle 0, and was retired at cycle 3. 61 /// Instruction [0,1] was also dispatched at cycle 0, and it retired at 62 /// the same cycle than instruction [0,0]. 63 /// Instruction [0,4] has been dispatched at cycle 2. However, it had to 64 /// wait for two cycles before being issued. That is because operands 65 /// became ready only at cycle 5. 66 /// 67 /// This view helps further understanding bottlenecks and the impact of 68 /// resource pressure on the code. 69 /// 70 /// To better understand why instructions had to wait for multiple cycles in 71 /// the scheduler's queue, class TimelineView also reports extra timing info 72 /// in another table named "Average Wait times" (see example below). 73 /// 74 /// 75 /// Average Wait times (based on the timeline view): 76 /// [0]: Executions 77 /// [1]: Average time spent waiting in a scheduler's queue 78 /// [2]: Average time spent waiting in a scheduler's queue while ready 79 /// [3]: Average time elapsed from WB until retire stage 80 /// 81 /// [0] [1] [2] [3] 82 /// 0. 2 1.0 1.0 3.0 vmovshdup %xmm0, %xmm1 83 /// 1. 2 1.0 1.0 3.0 vpermilpd $1, %xmm0, %xmm2 84 /// 2. 2 1.0 1.0 2.5 vpermilps $231, %xmm0, %xmm5 85 /// 3. 2 1.5 0.5 1.0 vaddss %xmm1, %xmm0, %xmm3 86 /// 4. 2 3.5 0.0 0.0 vaddss %xmm3, %xmm2, %xmm4 87 /// 5. 2 6.5 0.0 0.0 vaddss %xmm4, %xmm5, %xmm6 88 /// 89 /// By comparing column [2] with column [1], we get an idea about how many 90 /// cycles were spent in the scheduler's queue due to data dependencies. 91 /// 92 /// In this example, instruction 5 spent an average of ~6 cycles in the 93 /// scheduler's queue. As soon as operands became ready, the instruction 94 /// was immediately issued to the pipeline(s). 95 /// That is expected because instruction 5 cannot transition to the "ready" 96 /// state until %xmm4 is written by instruction 4. 97 /// 98 //===----------------------------------------------------------------------===// 99 100 #ifndef LLVM_TOOLS_LLVM_MCA_TIMELINEVIEW_H 101 #define LLVM_TOOLS_LLVM_MCA_TIMELINEVIEW_H 102 103 #include "SourceMgr.h" 104 #include "View.h" 105 #include "llvm/MC/MCInstPrinter.h" 106 #include "llvm/MC/MCSubtargetInfo.h" 107 #include "llvm/Support/FormattedStream.h" 108 #include "llvm/Support/raw_ostream.h" 109 #include <map> 110 111 namespace mca { 112 113 /// This class listens to instruction state transition events 114 /// in order to construct a timeline information. 115 /// 116 /// For every instruction executed by the Pipeline, this class constructs 117 /// a TimelineViewEntry object. TimelineViewEntry objects are then used 118 /// to print the timeline information, as well as the "average wait times" 119 /// for every instruction in the input assembly sequence. 120 class TimelineView : public View { 121 const llvm::MCSubtargetInfo &STI; 122 llvm::MCInstPrinter &MCIP; 123 const SourceMgr &AsmSequence; 124 125 unsigned CurrentCycle; 126 unsigned MaxCycle; 127 unsigned LastCycle; 128 129 struct TimelineViewEntry { 130 unsigned CycleDispatched; 131 unsigned CycleReady; 132 unsigned CycleIssued; 133 unsigned CycleExecuted; 134 unsigned CycleRetired; 135 }; 136 std::vector<TimelineViewEntry> Timeline; 137 138 struct WaitTimeEntry { 139 unsigned Executions; 140 unsigned CyclesSpentInSchedulerQueue; 141 unsigned CyclesSpentInSQWhileReady; 142 unsigned CyclesSpentAfterWBAndBeforeRetire; 143 }; 144 std::vector<WaitTimeEntry> WaitTime; 145 146 void printTimelineViewEntry(llvm::formatted_raw_ostream &OS, 147 const TimelineViewEntry &E, unsigned Iteration, 148 unsigned SourceIndex) const; 149 void printWaitTimeEntry(llvm::formatted_raw_ostream &OS, 150 const WaitTimeEntry &E, unsigned Index) const; 151 152 const unsigned DEFAULT_ITERATIONS = 10; 153 154 void initialize(unsigned MaxIterations); 155 156 // Display characters for the TimelineView report output. 157 struct DisplayChar { 158 static const char Dispatched = 'D'; 159 static const char Executed = 'E'; 160 static const char Retired = 'R'; 161 static const char Waiting = '='; // Instruction is waiting in the scheduler. 162 static const char Executing = 'e'; 163 static const char RetireLag = '-'; // The instruction is waiting to retire. 164 }; 165 166 public: TimelineView(const llvm::MCSubtargetInfo & sti,llvm::MCInstPrinter & Printer,const SourceMgr & Sequence,unsigned MaxIterations,unsigned Cycles)167 TimelineView(const llvm::MCSubtargetInfo &sti, llvm::MCInstPrinter &Printer, 168 const SourceMgr &Sequence, unsigned MaxIterations, 169 unsigned Cycles) 170 : STI(sti), MCIP(Printer), AsmSequence(Sequence), CurrentCycle(0), 171 MaxCycle(Cycles == 0 ? 80 : Cycles), LastCycle(0) { 172 initialize(MaxIterations); 173 } 174 175 // Event handlers. onCycleEnd()176 void onCycleEnd() override { ++CurrentCycle; } 177 void onEvent(const HWInstructionEvent &Event) override; 178 179 // print functionalities. 180 void printTimeline(llvm::raw_ostream &OS) const; 181 void printAverageWaitTimes(llvm::raw_ostream &OS) const; printView(llvm::raw_ostream & OS)182 void printView(llvm::raw_ostream &OS) const override { 183 printTimeline(OS); 184 printAverageWaitTimes(OS); 185 } 186 }; 187 } // namespace mca 188 189 #endif 190