[BOLT] Control aggregation mode output profile file format

[Amir]

Summary

In perf2bolt and -aggregate-only BOLT mode, the output profile file is written
in fdata format by default. Provide a knob -profile-format=[fdata,yaml] to
control the format.
Note that -w option still dumps in YAML format.

[Amir]

typo

[Amir]

typos

[Amir]

Fix build issue, extend X86/pre-aggregated-perf.test

[Amir]

Use sort for fdata output

[maksfb]

Can you s/ProfileOutput/ProfileFormat/g and s/profile-output/profile-format/g?

bolt/test/X86/pre-aggregated-perf.test

17 ↗

(On Diff #460871)

nit: use double dash for long options.

[Amir]

Address comments

[Amir]

In D133995#3796613, @maksfb wrote:

Can you s/ProfileOutput/ProfileFormat/g and s/profile-output/profile-format/g?

Done

[maksfb]

bolt/test/X86/pre-aggregated-perf.test

22 ↗

(On Diff #460900)

It's super confusing, but -pa is a long option.

[Amir]

Use --pa

[Amir]

bolt/test/X86/pre-aggregated-perf.test

22 ↗

(On Diff #460900)

Sorry, I didn't notice that while copying. Fixed.

[maksfb]

LGTM

bolt/include/bolt/Profile/DataAggregator.h

 
 namespace llvm {
 namespace bolt {
 
 class BinaryFunction;
 class BinaryContext;
 class BoltAddressTranslation;
 
+// The format to use with -o in aggregation mode (perf2bolt)
+enum ProfileOutputKind { PO_Fdata, PO_YAML };
+
 /// DataAggregator inherits all parsing logic from DataReader as well as
 /// its data structures used to represent aggregated profile data in memory.
 ///
 /// The aggregator works by dispatching two separate perf-script jobs that
 /// read perf samples and perf task annotations. Later, we read the output
 /// files to extract information about which PID was used for this binary.
 /// With the PID, we filter the samples and extract all LBR entries.
 ///

bolt/lib/Profile/DataAggregator.cpp

 static cl::opt<unsigned long long>
 MaxSamples("max-samples",
   cl::init(-1ULL),
   cl::desc("maximum number of samples to read from LBR profile"),
   cl::Optional,
   cl::Hidden,
   cl::cat(AggregatorCategory));
 
+extern cl::opt<ProfileOutputKind> ProfileOutput;
+
 cl::opt<bool> ReadPreAggregated(
     "pa", cl::desc("skip perf and read data from a pre-aggregated file format"),
     cl::cat(AggregatorCategory));
 
 static cl::opt<bool>
 TimeAggregator("time-aggr",
   cl::desc("time BOLT aggregator"),
   cl::init(false),

 Error DataAggregator::readProfile(BinaryContext &BC) {
   processProfile(BC);
 
   for (auto &BFI : BC.getBinaryFunctions()) {
     BinaryFunction &Function = BFI.second;
     convertBranchData(Function);
   }
 
-  if (opts::AggregateOnly) {
+  if (opts::AggregateOnly &&
+      opts::ProfileOutput == ProfileOutputKind::PO_Fdata) {
     if (std::error_code EC = writeAggregatedFile(opts::OutputFilename))
       report_error("cannot create output data file", EC);
   }
 
   return Error::success();
 }
 
 bool DataAggregator::mayHaveProfileData(const BinaryFunction &Function) {

bolt/lib/Rewrite/RewriteInstance.cpp

cl::values(clEnumValN(PPP_Probes_Section_Decode, "decode",
                           "decode probes section from binary"),
                clEnumValN(PPP_Probes_Address_Conversion, "address_conversion",
                           "update address2ProbesMap with output block address"),
                clEnumValN(PPP_Encoded_Probes, "encoded_probes",
                           "display the encoded probes in binary section"),
                clEnumValN(PPP_All, "all", "enable all debugging printout")),
     cl::ZeroOrMore, cl::Hidden, cl::cat(BoltCategory));
 
+cl::opt<ProfileOutputKind> ProfileOutput(
+    "profile-output",
+    cl::desc(
+        "format to dump profile output in aggregation mode, default is fdata"),
+    cl::init(PO_Fdata),
+    cl::values(clEnumValN(PO_Fdata, "fdata", "offset-based plaintext format"),
+               clEnumValN(PO_YAML, "yaml", "dense YAML reprensentation")),
+    cl::ZeroOrMore, cl::Hidden, cl::cat(BoltCategory));
+
 static cl::opt<cl::boolOrDefault> RelocationMode(
     "relocs", cl::desc("use relocations in the binary (default=autodetect)"),
     cl::cat(BoltCategory));
 
 static cl::opt<std::string>
 SaveProfile("w",
   cl::desc("save recorded profile to a file"),
   cl::cat(BoltOutputCategory));
void RewriteInstance::processProfileData() {
 
   if (Error E = ProfileReader->readProfile(*BC.get()))
     report_error("cannot read profile", std::move(E));
 
   if (!opts::SaveProfile.empty()) {
     YAMLProfileWriter PW(opts::SaveProfile);
     PW.writeProfile(*this);
   }
+  if (opts::AggregateOnly &&
+      opts::ProfileOutput == ProfileOutputKind::PO_YAML) {
+    YAMLProfileWriter PW(opts::OutputFilename);
+    PW.writeProfile(*this);
+  }
 
   // Release memory used by profile reader.
   ProfileReader.reset();
 
   if (opts::AggregateOnly)
     exit(0);
 }