[llvm-exegesis] Opcode stabilization / reclusterization (PR40715)

[lebedev.ri]

Summary

Given an instruction Opcode, we can make benchmarks (measurements) of the
instruction characteristics/performance. Then, to facilitate further analysis
we group the benchmarks with *similar* characteristics into clusters.
Now, this is all not entirely deterministic. Some instructions have variable
characteristics, depending on their arguments. And thus, if we do several
benchmarks of the same instruction Opcode, we may end up with *different*
performance characteristics measurements. And when we then do clustering,
these several benchmarks of the same instruction Opcode may end up being
clustered into *different* clusters. This is not great for further analysis.

We shall find every Opcode with benchmarks not in just one cluster, and move
*all* the benchmarks of said Opcode into one new unstable cluster per Opcode.

I have solved this by making ClusterId a bit field, adding a IsUnstable bit,
and introducing -analysis-display-unstable-clusters switch to toggle between
displaying stable-only clusters and unstable-only clusters.

The reclusterization is deterministically stable, produces identical reports
between runs. (Or at least that is what i'm seeing, maybe it isn't)

Timings/comparisons:
old (current trunk/head)

clusters-old.html6 MBDownload

$ perf stat -r 25 ./bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-old.html
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-old.html'
...
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-old.html'

 Performance counter stats for './bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-old.html' (25 runs):

           6624.73 msec task-clock                #    0.999 CPUs utilized            ( +-  0.53% )
               172      context-switches          #   25.965 M/sec                    ( +- 29.89% )
                 0      cpu-migrations            #    0.042 M/sec                    ( +- 56.54% )
             31073      page-faults               # 4690.754 M/sec                    ( +-  0.08% )
       26538711696      cycles                    # 4006230.292 GHz                   ( +-  0.53% )  (83.31%)
        2017496807      stalled-cycles-frontend   #    7.60% frontend cycles idle     ( +-  0.93% )  (83.32%)
       13403650062      stalled-cycles-backend    #   50.51% backend cycles idle      ( +-  0.33% )  (33.37%)
       19770706799      instructions              #    0.74  insn per cycle         
                                                  #    0.68  stalled cycles per insn  ( +-  0.04% )  (50.04%)
        4419821812      branches                  # 667207369.714 M/sec               ( +-  0.03% )  (66.69%)
         121741669      branch-misses             #    2.75% of all branches          ( +-  0.28% )  (83.34%)

            6.6283 +- 0.0358 seconds time elapsed  ( +-  0.54% )

patch, with reclustering but without filtering (i.e. outputting all the stable *and* unstable clusters)

clusters-new-all.html6 MBDownload

$ perf stat -r 25 ./bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-new-all.html
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-new-all.html'
...
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-new-all.html'

 Performance counter stats for './bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-new-all.html' (25 runs):

           6475.29 msec task-clock                #    0.999 CPUs utilized            ( +-  0.31% )
               213      context-switches          #   32.952 M/sec                    ( +- 23.81% )
                 1      cpu-migrations            #    0.130 M/sec                    ( +- 43.84% )
             31287      page-faults               # 4832.057 M/sec                    ( +-  0.08% )
       25939086577      cycles                    # 4006160.279 GHz                   ( +-  0.31% )  (83.31%)
        1958812858      stalled-cycles-frontend   #    7.55% frontend cycles idle     ( +-  0.68% )  (83.32%)
       13218961512      stalled-cycles-backend    #   50.96% backend cycles idle      ( +-  0.29% )  (33.37%)
       19752995402      instructions              #    0.76  insn per cycle         
                                                  #    0.67  stalled cycles per insn  ( +-  0.04% )  (50.04%)
        4417079244      branches                  # 682195472.305 M/sec               ( +-  0.03% )  (66.70%)
         121510065      branch-misses             #    2.75% of all branches          ( +-  0.19% )  (83.34%)

            6.4832 +- 0.0229 seconds time elapsed  ( +-  0.35% )

Funnily, *this* measurement shows that said reclustering actually improved performance.

patch, with reclustering, only the stable clusters

clusters-new-stable.html6 MBDownload

$ perf stat -r 25 ./bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-new-stable.html
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-new-stable.html'
...
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-new-stable.html'

 Performance counter stats for './bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-new-stable.html' (25 runs):

           6387.71 msec task-clock                #    0.999 CPUs utilized            ( +-  0.13% )
               133      context-switches          #   20.792 M/sec                    ( +- 23.39% )
                 0      cpu-migrations            #    0.063 M/sec                    ( +- 61.24% )
             31318      page-faults               # 4903.256 M/sec                    ( +-  0.08% )
       25591984967      cycles                    # 4006786.266 GHz                   ( +-  0.13% )  (83.31%)
        1881234904      stalled-cycles-frontend   #    7.35% frontend cycles idle     ( +-  0.25% )  (83.33%)
       13209749965      stalled-cycles-backend    #   51.62% backend cycles idle      ( +-  0.16% )  (33.36%)
       19767554347      instructions              #    0.77  insn per cycle         
                                                  #    0.67  stalled cycles per insn  ( +-  0.04% )  (50.03%)
        4417480305      branches                  # 691618858.046 M/sec               ( +-  0.03% )  (66.68%)
         118676358      branch-misses             #    2.69% of all branches          ( +-  0.07% )  (83.33%)

            6.3954 +- 0.0118 seconds time elapsed  ( +-  0.18% )

Performance improved even further?! Makes sense i guess, less clusters to print.

patch, with reclustering, only the unstable clusters

clusters-new-unstable.html46 KBDownload

$ perf stat -r 25 ./bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-new-unstable.html -analysis-display-unstable-clusters
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-new-unstable.html'
...
no exegesis target for x86_64-unknown-linux-gnu, using default
Parsed 43970 benchmark points
Printing sched class consistency analysis results to file '/tmp/clusters-new-unstable.html'

 Performance counter stats for './bin/llvm-exegesis -mode=analysis -analysis-epsilon=0.5 -benchmarks-file=/home/lebedevri/PileDriver-Sched/benchmarks-inverse_throughput.yaml -analysis-inconsistencies-output-file=/tmp/clusters-new-unstable.html -analysis-display-unstable-clusters' (25 runs):

           6124.96 msec task-clock                #    1.000 CPUs utilized            ( +-  0.20% )
               194      context-switches          #   31.709 M/sec                    ( +- 20.46% )
                 0      cpu-migrations            #    0.039 M/sec                    ( +- 49.77% )
             31413      page-faults               # 5129.261 M/sec                    ( +-  0.06% )
       24536794267      cycles                    # 4006425.858 GHz                   ( +-  0.19% )  (83.31%)
        1676085087      stalled-cycles-frontend   #    6.83% frontend cycles idle     ( +-  0.46% )  (83.32%)
       13035595603      stalled-cycles-backend    #   53.13% backend cycles idle      ( +-  0.16% )  (33.36%)
       18260877653      instructions              #    0.74  insn per cycle         
                                                  #    0.71  stalled cycles per insn  ( +-  0.05% )  (50.03%)
        4112411983      branches                  # 671484364.603 M/sec               ( +-  0.03% )  (66.68%)
         114066929      branch-misses             #    2.77% of all branches          ( +-  0.11% )  (83.32%)

            6.1278 +- 0.0121 seconds time elapsed  ( +-  0.20% )

This tells us that the actual -analysis-inconsistencies-output-file= outputting only takes ~0.4 sec for 43970 benchmark points (3 whole sweeps)
(Also, wow this is fast, it used to take several minutes originally)

Fixes PR40715.

[courbet]

tools/llvm-exegesis/lib/Clustering.cpp

168 ↗

(On Diff #187245)

"The list of opcodes that have more than one cluster".

172 ↗

(On Diff #187245)

Why not const auto& ?

192 ↗

(On Diff #187245)

for (const size_t UnstableOpcode for safety.

214 ↗

(On Diff #187245)

at least, not at most.

216 ↗

(On Diff #187245)

This for loop + CleanedPointIndices could be removed using std::remove_if:

// Find which points should be moved to the new cluster.
const auto it = std::remove_if(OldCluster.PointIndices.begin(), 
                              OldCluster.PointIndices.end(),
                              [this, UnstableOpcode](size_t P){ return Points_[P].keyInstruction().getOpcode() == UnstableOpcode; });
// Move removed points to the new cluster:
UnstableCluster.PointIndices.insert(it, OldCluster.PointIndices.end());
// Remove points form the old cluster.
OldCluster.PointIndices.erase(it, OldCluster.PointIndices.end());

tools/llvm-exegesis/llvm-exegesis.cpp

444 ↗

(On Diff #187245)

nit: InstrInfo

445 ↗

(On Diff #187245)

Why not: std::unique_ptr<llvm::MCInstrInfo> InstrInfo(TheTarget->createMCInstrInfo()); ?

[lebedev.ri]

Address most of @courbet's review notes.

tools/llvm-exegesis/lib/Clustering.cpp

172 ↗

(On Diff #187245)

Hm, i see all three variants within the codebase.
I guess const auto& is better.

214 ↗

(On Diff #187245)

Sure. I have meant that any other assumption would be too optimistic,
and we would end up reallocating if we guessed wrong.
A guess of one will never be too optimistic, and won't cause reallocations.
At worst, we will over-allocate a bit.

216 ↗

(On Diff #187245)

Hmm, are you sure?

http://cpp.sh/7dfh5
If that is so, then that ^ should have
printed Textwithsomewhitespaces but it prints Textwithsomewhitespacesespaces.

https://en.cppreference.com/w/cpp/algorithm/remove

Iterators pointing to an element between the new logical end and the physical end of the range are still dereferenceable, but the elements themselves have unspecified values (as per MoveAssignable post-condition).

unspecified values is pretty self-explanatory..

[courbet]

tools/llvm-exegesis/lib/Clustering.cpp

216 ↗

(On Diff #187245)

Yes, sorry. std::stable_partition should work.

[lebedev.ri]

Address @courbet's review notes.

tools/llvm-exegesis/lib/Clustering.cpp

216 ↗

(On Diff #187245)

Aha. Not as simple as that snippet, but works.

[lebedev.ri]

tools/llvm-exegesis/lib/Clustering.cpp

216 ↗

(On Diff #187245)

Thinking about this a bit more, do we really care that "Relative order of the elements is preserved."?
I don't think so. Only that the new order is deterministic.
Can we just use [[ https://en.cppreference.com/w/cpp/algorithm/partition | std::partition ]] instead?

[lebedev.ri]

tools/llvm-exegesis/lib/Clustering.cpp

216 ↗

(On Diff #187245)

Actually, hmm, i'm not confident std::partition *is* deterministic.
Never mind.

[lebedev.ri]

Yay, thank you for the review!

llvm/trunk/docs/CommandGuide/llvm-exegesis.rst

  Specify the numPoints parameters to be used for DBSCAN clustering
  (`analysis` mode).
 
 .. option:: -analysis-epsilon=<dbscan epsilon parameter>
 
  Specify the numPoints parameters to be used for DBSCAN clustering
  (`analysis` mode).
 
+.. option:: -analysis-display-unstable-clusters
+
+ If there is more than one benchmark for an opcode, said benchmarks may end up
+ not being clustered into the same cluster if the measured performance
+ characteristics are different. by default all such opcodes are filtered out.
+ This flag will instead show only such unstable opcodes.
+
 .. option:: -ignore-invalid-sched-class=false
 
  If set, ignore instructions that do not have a sched class (class idx = 0).
 
  .. option:: -mcpu=<cpu name>
 
   If set, measure the cpu characteristics using the counters for this CPU. This
   is useful when creating new sched models (the host CPU is unknown to LLVM).
 
 EXIT STATUS
 -----------
 
 :program:`llvm-exegesis` returns 0 on success. Otherwise, an error message is
 printed to standard error, and the tool returns a non 0 value.

llvm/trunk/test/tools/llvm-exegesis/X86/analysis-cluster-stabilization.test

+# RUN: llvm-exegesis -mode=analysis -benchmarks-file=%s -analysis-clusters-output-file=- -analysis-epsilon=0.1 -analysis-numpoints=1 | FileCheck -check-prefixes=CHECK-CLUSTERS %s
+# RUN: llvm-exegesis -mode=analysis -benchmarks-file=%s -analysis-inconsistencies-output-file=- -analysis-epsilon=0.5 -analysis-numpoints=1 | FileCheck -check-prefixes=CHECK-INCONSISTENCIES-ALL,CHECK-INCONSISTENCIES-STABLE %s
+# RUN: llvm-exegesis -mode=analysis -benchmarks-file=%s -analysis-inconsistencies-output-file=- -analysis-epsilon=0.5 -analysis-display-unstable-clusters -analysis-numpoints=1 | FileCheck -check-prefixes=CHECK-INCONSISTENCIES-ALL,CHECK-INCONSISTENCIES-UNSTABLE %s
+
+# We have one ADD32rr measurement, and two measurements for SQRTSSr.
+# The ADD32rr measurement and one of the SQRTSSr measurements are identical,
+# and thus will be be in the same cluster. But the second SQRTSSr measurement
+# is different from the first SQRTSSr measurement, and thus it will be in it's
+# own cluster. We do reclusterization, and thus since there is more than one
+# measurement from SQRTSSr, and they are not in the same cluster, we move
+# all two SQRTSSr measurements into their own cluster, and mark it as unstable.
+# By default, we do not show such unstable clusters.
+# If told to show, we *only* show such unstable clusters.
+
+# CHECK-CLUSTERS: {{^}}cluster_id,opcode_name,config,sched_class,latency{{$}}
+# CHECK-CLUSTERS-NEXT: {{^}}0,
+# CHECK-CLUSTERS-SAME: ,90.00{{$}}
+# CHECK-CLUSTERS: {{^}}3,
+# CHECK-CLUSTERS-SAME: ,90.11{{$}}
+# CHECK-CLUSTERS-NEXT: {{^}}3,
+# CHECK-CLUSTERS-SAME: ,100.00{{$}}
+
+# CHECK-INCONSISTENCIES-STABLE: ADD32rr
+# CHECK-INCONSISTENCIES-STABLE-NOT: ADD32rr
+# CHECK-INCONSISTENCIES-STABLE-NOT: SQRTSSr
+
+# CHECK-INCONSISTENCIES-UNSTABLE: SQRTSSr
+# CHECK-INCONSISTENCIES-UNSTABLE: SQRTSSr
+# CHECK-INCONSISTENCIES-UNSTABLE-NOT: SQRTSSr
+# CHECK-INCONSISTENCIES-UNSTABLE-NOT: ADD32rr
+
+---
+mode:            latency
+key:
+  instructions:
+    - 'ADD32rr EDX EDX EAX'
+  config:          ''
+  register_initial_values:
+    - 'EDX=0x0'
+    - 'EAX=0x0'
+cpu_name:        bdver2
+llvm_triple:     x86_64-unknown-linux-gnu
+num_repetitions: 10000
+measurements:
+  - { key: latency, value: 90.0000, per_snippet_value: 90.0000 }
+error:           ''
+info:            Repeating a single implicitly serial instruction
+assembled_snippet: BA00000000B80000000001C201C201C201C201C201C201C201C201C201C201C201C201C201C201C201C2C3
+---
+mode:            latency
+key:
+  instructions:
+    - 'SQRTSSr XMM11 XMM11'
+  config:          ''
+  register_initial_values:
+    - 'XMM11=0x0'
+cpu_name:        bdver2
+llvm_triple:     x86_64-unknown-linux-gnu
+num_repetitions: 10000
+measurements:
+  - { key: latency, value: 90.1111, per_snippet_value: 90.1111 }
+error:           ''
+info:            Repeating a single explicitly serial instruction
+assembled_snippet: 4883EC10C7042400000000C744240400000000C744240800000000C744240C00000000C57A6F1C244883C410F3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBC3
+...
+---
+mode:            latency
+key:
+  instructions:
+    - 'SQRTSSr XMM11 XMM11'
+  config:          ''
+  register_initial_values:
+    - 'XMM11=0x0'
+cpu_name:        bdver2
+llvm_triple:     x86_64-unknown-linux-gnu
+num_repetitions: 10000
+measurements:
+  - { key: latency, value: 100, per_snippet_value: 100 }
+error:           ''
+info:            Repeating a single explicitly serial instruction
+assembled_snippet: 4883EC10C7042400000000C744240400000000C744240800000000C744240C00000000C57A6F1C244883C410F3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBF3450F51DBC3
+...

llvm/trunk/tools/llvm-exegesis/lib/Analysis.h

 
 namespace llvm {
 namespace exegesis {
 
 // A helper class to analyze benchmark results for a target.
 class Analysis {
 public:
   Analysis(const llvm::Target &Target,
-           const InstructionBenchmarkClustering &Clustering);
+           std::unique_ptr<llvm::MCInstrInfo> InstrInfo,
+           const InstructionBenchmarkClustering &Clustering,
+           bool AnalysisDisplayUnstableOpcodes);
 
   // Prints a csv of instructions for each cluster.
   struct PrintClusters {};
   // Find potential errors in the scheduling information given measurements.
   struct PrintSchedClassInconsistencies {};
 
   template <typename Pass> llvm::Error run(llvm::raw_ostream &OS) const;
 
private:
   llvm::MCObjectFileInfo ObjectFileInfo_;
   std::unique_ptr<llvm::MCContext> Context_;
   std::unique_ptr<llvm::MCSubtargetInfo> SubtargetInfo_;
   std::unique_ptr<llvm::MCInstrInfo> InstrInfo_;
   std::unique_ptr<llvm::MCRegisterInfo> RegInfo_;
   std::unique_ptr<llvm::MCAsmInfo> AsmInfo_;
   std::unique_ptr<llvm::MCInstPrinter> InstPrinter_;
   std::unique_ptr<llvm::MCDisassembler> Disasm_;
+  const bool AnalysisDisplayUnstableOpcodes_;
 };
 
 // Computes the idealized ProcRes Unit pressure. This is the expected
 // distribution if the CPU scheduler can distribute the load as evenly as
 // possible.
 std::vector<std::pair<uint16_t, float>> computeIdealizedProcResPressure(
     const llvm::MCSchedModel &SM,
     llvm::SmallVector<llvm::MCWriteProcResEntry, 8> WPRS);
 
 } // namespace exegesis
 } // namespace llvm
 
 #endif // LLVM_TOOLS_LLVM_EXEGESIS_CLUSTERING_H

llvm/trunk/tools/llvm-exegesis/lib/Analysis.cpp

void Analysis::printInstructionRowCsv(const size_t PointId,
   const InstructionBenchmark &Point = Clustering_.getPoints()[PointId];
   writeClusterId<kEscapeCsv>(OS, Clustering_.getClusterIdForPoint(PointId));
   OS << kCsvSep;
   writeSnippet<EscapeTag, kEscapeCsv>(OS, Point.AssembledSnippet, "; ");
   OS << kCsvSep;
   writeEscaped<kEscapeCsv>(OS, Point.Key.Config);
   OS << kCsvSep;
   assert(!Point.Key.Instructions.empty());
-  const llvm::MCInst &MCI = Point.Key.Instructions[0];
+  const llvm::MCInst &MCI = Point.keyInstruction();
   const unsigned SchedClassId = resolveSchedClassId(
       *SubtargetInfo_, InstrInfo_->get(MCI.getOpcode()).getSchedClass(), MCI);
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   const llvm::MCSchedClassDesc *const SCDesc =
       SubtargetInfo_->getSchedModel().getSchedClassDesc(SchedClassId);
   writeEscaped<kEscapeCsv>(OS, SCDesc->Name);
 #else
   OS << SchedClassId;
 #endif
   for (const auto &Measurement : Point.Measurements) {
     OS << kCsvSep;
     writeMeasurementValue<kEscapeCsv>(OS, Measurement.PerInstructionValue);
   }
   OS << "\n";
 }
 
 Analysis::Analysis(const llvm::Target &Target,
-                   const InstructionBenchmarkClustering &Clustering)
-    : Clustering_(Clustering) {
+                   std::unique_ptr<llvm::MCInstrInfo> InstrInfo,
+                   const InstructionBenchmarkClustering &Clustering,
+                   bool AnalysisDisplayUnstableOpcodes)
+    : Clustering_(Clustering), InstrInfo_(std::move(InstrInfo)),
+      AnalysisDisplayUnstableOpcodes_(AnalysisDisplayUnstableOpcodes) {
   if (Clustering.getPoints().empty())
     return;
 
   const InstructionBenchmark &FirstPoint = Clustering.getPoints().front();
-  InstrInfo_.reset(Target.createMCInstrInfo());
   RegInfo_.reset(Target.createMCRegInfo(FirstPoint.LLVMTriple));
   AsmInfo_.reset(Target.createMCAsmInfo(*RegInfo_, FirstPoint.LLVMTriple));
   SubtargetInfo_.reset(Target.createMCSubtargetInfo(FirstPoint.LLVMTriple,
                                                     FirstPoint.CpuName, ""));
   InstPrinter_.reset(Target.createMCInstPrinter(
       llvm::Triple(FirstPoint.LLVMTriple), 0 /*default variant*/, *AsmInfo_,
       *InstrInfo_, *RegInfo_));
 
Analysis::makePointsPerSchedClass() const {
   const auto &Points = Clustering_.getPoints();
   for (size_t PointId = 0, E = Points.size(); PointId < E; ++PointId) {
     const InstructionBenchmark &Point = Points[PointId];
     if (!Point.Error.empty())
       continue;
     assert(!Point.Key.Instructions.empty());
     // FIXME: we should be using the tuple of classes for instructions in the
     // snippet as key.
-    const llvm::MCInst &MCI = Point.Key.Instructions[0];
+    const llvm::MCInst &MCI = Point.keyInstruction();
     unsigned SchedClassId = InstrInfo_->get(MCI.getOpcode()).getSchedClass();
     const bool WasVariant = SchedClassId && SubtargetInfo_->getSchedModel()
                                                 .getSchedClassDesc(SchedClassId)
                                                 ->isVariant();
     SchedClassId = resolveSchedClassId(*SubtargetInfo_, SchedClassId, MCI);
     const auto IndexIt = SchedClassIdToIndex.find(SchedClassId);
     if (IndexIt == SchedClassIdToIndex.end()) {
       // Create a new entry.
for (const auto &RSCAndPoints : makePointsPerSchedClass()) {
     if (!RSCAndPoints.RSC.SCDesc)
       continue;
     // Bucket sched class points into sched class clusters.
     std::vector<SchedClassCluster> SchedClassClusters;
     for (const size_t PointId : RSCAndPoints.PointIds) {
       const auto &ClusterId = Clustering_.getClusterIdForPoint(PointId);
       if (!ClusterId.isValid())
         continue; // Ignore noise and errors. FIXME: take noise into account ?
+      if (ClusterId.isUnstable() ^ AnalysisDisplayUnstableOpcodes_)
+        continue; // Either display stable or unstable clusters only.
       auto SchedClassClusterIt =
           std::find_if(SchedClassClusters.begin(), SchedClassClusters.end(),
                        [ClusterId](const SchedClassCluster &C) {
                          return C.id() == ClusterId;
                        });
       if (SchedClassClusterIt == SchedClassClusters.end()) {
         SchedClassClusters.emplace_back();
         SchedClassClusterIt = std::prev(SchedClassClusters.end());

llvm/trunk/tools/llvm-exegesis/lib/BenchmarkResult.h

 
 // The result of an instruction benchmark.
 struct InstructionBenchmark {
   InstructionBenchmarkKey Key;
   enum ModeE { Unknown, Latency, Uops, InverseThroughput };
   ModeE Mode;
   std::string CpuName;
   std::string LLVMTriple;
+  // Which instruction is being benchmarked here?
+  const llvm::MCInst &keyInstruction() const { return Key.Instructions[0]; }
   // The number of instructions inside the repeated snippet. For example, if a
   // snippet of 3 instructions is repeated 4 times, this is 12.
   int NumRepetitions = 0;
   // Note that measurements are per instruction.
   std::vector<BenchmarkMeasure> Measurements;
   std::string Error;
   std::string Info;
   std::vector<uint8_t> AssembledSnippet;

llvm/trunk/tools/llvm-exegesis/lib/Clustering.h

 /// Utilities to compute benchmark result clusters.
 ///
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_TOOLS_LLVM_EXEGESIS_CLUSTERING_H
 #define LLVM_TOOLS_LLVM_EXEGESIS_CLUSTERING_H
 
 #include "BenchmarkResult.h"
+#include "llvm/ADT/Optional.h"
 #include "llvm/Support/Error.h"
+#include <limits>
 #include <vector>
 
 namespace llvm {
 namespace exegesis {
 
 class InstructionBenchmarkClustering {
 public:
   // Clusters `Points` using DBSCAN with the given parameters. See the cc file
   // for more explanations on the algorithm.
   static llvm::Expected<InstructionBenchmarkClustering>
   create(const std::vector<InstructionBenchmark> &Points, size_t MinPts,
-         double Epsilon);
+         double Epsilon, llvm::Optional<unsigned> NumOpcodes = llvm::None);
 
   class ClusterId {
   public:
     static ClusterId noise() { return ClusterId(kNoise); }
     static ClusterId error() { return ClusterId(kError); }
     static ClusterId makeValid(size_t Id) { return ClusterId(Id); }
-    ClusterId() : Id_(kUndef) {}
+    static ClusterId makeValidUnstable(size_t Id) {
+      return ClusterId(Id, /*IsUnstable=*/true);
+    }
+
+    ClusterId() : Id_(kUndef), IsUnstable_(false) {}
+
+    // Compare id's, ignoring the 'unstability' bit.
     bool operator==(const ClusterId &O) const { return Id_ == O.Id_; }
     bool operator<(const ClusterId &O) const { return Id_ < O.Id_; }
 
     bool isValid() const { return Id_ <= kMaxValid; }
-    bool isUndef() const { return Id_ == kUndef; }
+    bool isUnstable() const { return IsUnstable_; }
     bool isNoise() const { return Id_ == kNoise; }
     bool isError() const { return Id_ == kError; }
+    bool isUndef() const { return Id_ == kUndef; }
 
     // Precondition: isValid().
     size_t getId() const {
       assert(isValid());
       return Id_;
     }
 
   private:
-    explicit ClusterId(size_t Id) : Id_(Id) {}
+    ClusterId(size_t Id, bool IsUnstable = false)
+        : Id_(Id), IsUnstable_(IsUnstable) {}
+
     static constexpr const size_t kMaxValid =
-        std::numeric_limits<size_t>::max() - 4;
+        (std::numeric_limits<size_t>::max() >> 1) - 4;
     static constexpr const size_t kNoise = kMaxValid + 1;
     static constexpr const size_t kError = kMaxValid + 2;
     static constexpr const size_t kUndef = kMaxValid + 3;
-    size_t Id_;
+
+    size_t Id_ : (std::numeric_limits<size_t>::digits - 1);
+    size_t IsUnstable_ : 1;
   };
+  static_assert(sizeof(ClusterId) == sizeof(size_t), "should be a bit field.");
 
   struct Cluster {
     Cluster() = delete;
     explicit Cluster(const ClusterId &Id) : Id(Id) {}
 
     const ClusterId Id;
     // Indices of benchmarks within the cluster.
     std::vector<int> PointIndices;
for (size_t I = 0, E = P.size(); I < E; ++I) {
       DistanceSquared += Diff * Diff;
     }
     return DistanceSquared <= EpsilonSquared_;
   }
 
 private:
   InstructionBenchmarkClustering(
       const std::vector<InstructionBenchmark> &Points, double EpsilonSquared);
+
   llvm::Error validateAndSetup();
   void dbScan(size_t MinPts);
+  void stabilize(unsigned NumOpcodes);
   void rangeQuery(size_t Q, std::vector<size_t> &Scratchpad) const;
 
   const std::vector<InstructionBenchmark> &Points_;
   const double EpsilonSquared_;
   int NumDimensions_ = 0;
   // ClusterForPoint_[P] is the cluster id for Points[P].
   std::vector<ClusterId> ClusterIdForPoint_;
   std::vector<Cluster> Clusters_;
   Cluster NoiseCluster_;
   Cluster ErrorCluster_;
 };
 
 } // namespace exegesis
 } // namespace llvm
 
 #endif // LLVM_TOOLS_LLVM_EXEGESIS_CLUSTERING_H

llvm/trunk/tools/llvm-exegesis/lib/Clustering.cpp

 //===-- Clustering.cpp ------------------------------------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "Clustering.h"
 #include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallVector.h"
+#include <algorithm>
 #include <string>
+#include <vector>
 
 namespace llvm {
 namespace exegesis {
 
 // The clustering problem has the following characteristics:
 //  (A) - Low dimension (dimensions are typically proc resource units,
 //    typically < 10).
 //  (B) - Number of points : ~thousands (points are measurements of an MCInst)
void InstructionBenchmarkClustering::dbScan(const size_t MinPts) {
   // Add noisy points to noise cluster.
   for (size_t P = 0, NumPoints = Points_.size(); P < NumPoints; ++P) {
     if (ClusterIdForPoint_[P].isNoise()) {
       NoiseCluster_.PointIndices.push_back(P);
     }
   }
 }
 
+// Given an instruction Opcode, we can make benchmarks (measurements) of the
+// instruction characteristics/performance. Then, to facilitate further analysis
+// we group the benchmarks with *similar* characteristics into clusters.
+// Now, this is all not entirely deterministic. Some instructions have variable
+// characteristics, depending on their arguments. And thus, if we do several
+// benchmarks of the same instruction Opcode, we may end up with *different*
+// performance characteristics measurements. And when we then do clustering,
+// these several benchmarks of the same instruction Opcode may end up being
+// clustered into *different* clusters. This is not great for further analysis.
+// We shall find every opcode with benchmarks not in just one cluster, and move
+// *all* the benchmarks of said Opcode into one new unstable cluster per Opcode.
+void InstructionBenchmarkClustering::stabilize(unsigned NumOpcodes) {
+  // Given an instruction Opcode, in which clusters do benchmarks of this
+  // instruction lie? Normally, they all should be in the same cluster.
+  std::vector<llvm::SmallSet<ClusterId, 1>> OpcodeToClusterIDs;
+  OpcodeToClusterIDs.resize(NumOpcodes);
+  // The list of opcodes that have more than one cluster.
+  llvm::SetVector<size_t> UnstableOpcodes;
+  // Populate OpcodeToClusterIDs and UnstableOpcodes data structures.
+  assert(ClusterIdForPoint_.size() == Points_.size() && "size mismatch");
+  for (const auto &Point : zip(Points_, ClusterIdForPoint_)) {
+    const ClusterId &ClusterIdOfPoint = std::get<1>(Point);
+    if (!ClusterIdOfPoint.isValid())
+      continue; // Only process fully valid clusters.
+    const unsigned Opcode = std::get<0>(Point).keyInstruction().getOpcode();
+    assert(Opcode < NumOpcodes && "NumOpcodes is incorrect (too small)");
+    llvm::SmallSet<ClusterId, 1> &ClusterIDsOfOpcode =
+        OpcodeToClusterIDs[Opcode];
+    ClusterIDsOfOpcode.insert(ClusterIdOfPoint);
+    // Is there more than one ClusterID for this opcode?.
+    if (ClusterIDsOfOpcode.size() < 2)
+      continue; // If not, then at this moment this Opcode is stable.
+    // Else let's record this unstable opcode for future use.
+    UnstableOpcodes.insert(Opcode);
+  }
+  assert(OpcodeToClusterIDs.size() == NumOpcodes && "sanity check");
+
+  // We know with how many [new] clusters we will end up with.
+  const auto NewTotalClusterCount = Clusters_.size() + UnstableOpcodes.size();
+  Clusters_.reserve(NewTotalClusterCount);
+  for (const size_t UnstableOpcode : UnstableOpcodes.getArrayRef()) {
+    const llvm::SmallSet<ClusterId, 1> &ClusterIDs =
+        OpcodeToClusterIDs[UnstableOpcode];
+    assert(ClusterIDs.size() > 1 &&
+           "Should only have Opcodes with more than one cluster.");
+
+    // Create a new unstable cluster, one per Opcode.
+    Clusters_.emplace_back(ClusterId::makeValidUnstable(Clusters_.size()));
+    Cluster &UnstableCluster = Clusters_.back();
+    // We will find *at least* one point in each of these clusters.
+    UnstableCluster.PointIndices.reserve(ClusterIDs.size());
+
+    // Go through every cluster which we recorded as containing benchmarks
+    // of this UnstableOpcode. NOTE: we only recorded valid clusters.
+    for (const ClusterId &CID : ClusterIDs) {
+      assert(CID.isValid() &&
+             "We only recorded valid clusters, not noise/error clusters.");
+      Cluster &OldCluster = Clusters_[CID.getId()]; // Valid clusters storage.
+      // Within each cluster, go through each point, and either move it to the
+      // new unstable cluster, or 'keep' it.
+      // In this case, we'll reshuffle OldCluster.PointIndices vector
+      // so that all the points that are *not* for UnstableOpcode are first,
+      // and the rest of the points is for the UnstableOpcode.
+      const auto it = std::stable_partition(
+          OldCluster.PointIndices.begin(), OldCluster.PointIndices.end(),
+          [this, UnstableOpcode](size_t P) {
+            return Points_[P].keyInstruction().getOpcode() != UnstableOpcode;
+          });
+      assert(std::distance(it, OldCluster.PointIndices.end()) > 0 &&
+             "Should have found at least one bad point");
+      // Mark to-be-moved points as belonging to the new cluster.
+      std::for_each(it, OldCluster.PointIndices.end(),
+                    [this, &UnstableCluster](size_t P) {
+                      ClusterIdForPoint_[P] = UnstableCluster.Id;
+                    });
+      // Actually append to-be-moved points to the new cluster.
+      UnstableCluster.PointIndices.insert(UnstableCluster.PointIndices.cend(),
+                                          it, OldCluster.PointIndices.end());
+      // And finally, remove "to-be-moved" points form the old cluster.
+      OldCluster.PointIndices.erase(it, OldCluster.PointIndices.cend());
+      // Now, the old cluster may end up being empty, but let's just keep it
+      // in whatever state it ended up. Purging empty clusters isn't worth it.
+    };
+    assert(UnstableCluster.PointIndices.size() > 1 &&
+           "New unstable cluster should end up with more than one point.");
+    assert(UnstableCluster.PointIndices.size() >= ClusterIDs.size() &&
+           "New unstable cluster should end up with no less points than there "
+           "was clusters");
+  }
+  assert(Clusters_.size() == NewTotalClusterCount && "sanity check");
+}
+
 llvm::Expected<InstructionBenchmarkClustering>
 InstructionBenchmarkClustering::create(
     const std::vector<InstructionBenchmark> &Points, const size_t MinPts,
-    const double Epsilon) {
+    const double Epsilon, llvm::Optional<unsigned> NumOpcodes) {
   InstructionBenchmarkClustering Clustering(Points, Epsilon * Epsilon);
   if (auto Error = Clustering.validateAndSetup()) {
     return std::move(Error);
   }
   if (Clustering.ErrorCluster_.PointIndices.size() == Points.size()) {
     return Clustering; // Nothing to cluster.
   }
 
   Clustering.dbScan(MinPts);
+
+  if (NumOpcodes.hasValue())
+    Clustering.stabilize(NumOpcodes.getValue());
+
   return Clustering;
 }
 
 } // namespace exegesis
 } // namespace llvm

llvm/trunk/tools/llvm-exegesis/llvm-exegesis.cpp

 
 static cl::opt<std::string>
     AnalysisClustersOutputFile("analysis-clusters-output-file", cl::desc(""),
                                cl::init(""));
 static cl::opt<std::string>
     AnalysisInconsistenciesOutputFile("analysis-inconsistencies-output-file",
                                       cl::desc(""), cl::init(""));
 
+static cl::opt<bool> AnalysisDisplayUnstableOpcodes(
+    "analysis-display-unstable-clusters",
+    cl::desc("if there is more than one benchmark for an opcode, said "
+             "benchmarks may end up not being clustered into the same cluster "
+             "if the measured performance characteristics are different. by "
+             "default all such opcodes are filtered out. this flag will "
+             "instead show only such unstable opcodes"),
+    cl::init(false));
+
 static cl::opt<std::string>
     CpuName("mcpu",
             cl::desc(
                 "cpu name to use for pfm counters, leave empty to autodetect"),
             cl::init(""));
 
-
 static ExitOnError ExitOnErr;
 
 #ifdef LLVM_EXEGESIS_INITIALIZE_NATIVE_TARGET
 void LLVM_EXEGESIS_INITIALIZE_NATIVE_TARGET();
 #endif
 
 // Checks that only one of OpcodeNames, OpcodeIndex or SnippetsFile is provided,
 // and returns the opcode indices or {} if snippets should be read from
static void analysisMain() {
 
   std::string Error;
   const auto *TheTarget =
       llvm::TargetRegistry::lookupTarget(Points[0].LLVMTriple, Error);
   if (!TheTarget) {
     llvm::errs() << "unknown target '" << Points[0].LLVMTriple << "'\n";
     return;
   }
+
+  std::unique_ptr<llvm::MCInstrInfo> InstrInfo(TheTarget->createMCInstrInfo());
+
   const auto Clustering = ExitOnErr(InstructionBenchmarkClustering::create(
-      Points, AnalysisNumPoints, AnalysisEpsilon));
+      Points, AnalysisNumPoints, AnalysisEpsilon, InstrInfo->getNumOpcodes()));
 
-  const Analysis Analyzer(*TheTarget, Clustering);
+  const Analysis Analyzer(*TheTarget, std::move(InstrInfo), Clustering,
+                          AnalysisDisplayUnstableOpcodes);
 
   maybeRunAnalysis<Analysis::PrintClusters>(Analyzer, "analysis clusters",
                                             AnalysisClustersOutputFile);
   maybeRunAnalysis<Analysis::PrintSchedClassInconsistencies>(
       Analyzer, "sched class consistency analysis",
       AnalysisInconsistenciesOutputFile);
 }