[PGO][PGSO] Tune flags for profile guided size optimization.

[hjyamauchi]

Summary

Tune the profile threshold flag value for instrumentation PGO based on internal
benchmarks.

Also, add flags to allow profile guided size optimizations for non-cold code
to be enabled separately for instrumentation and sample PGSO.

Neither changes the default behavior (yet) as it's disabled for non-cold code.

[hjyamauchi]

Rebase. PTAL.

[davidxl]

llvm/include/llvm/Transforms/Utils/SizeOpts.h

58–60

Is this master option needed anymore?

[hjyamauchi]

llvm/include/llvm/Transforms/Utils/SizeOpts.h

58–60

Technically not, but I'd like to keep it for a while (eg. a simpler/shorter way to work around issues) We can clean up flags later.

[davidxl]

lgtm

llvm/include/llvm/Transforms/Utils/SizeOpts.h

 #include "llvm/Analysis/BlockFrequencyInfo.h"
 #include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/Support/CommandLine.h"
 
 extern llvm::cl::opt<bool> EnablePGSO;
 extern llvm::cl::opt<bool> PGSOLargeWorkingSetSizeOnly;
 extern llvm::cl::opt<bool> PGSOIRPassOrTestOnly;
 extern llvm::cl::opt<bool> PGSOColdCodeOnly;
+extern llvm::cl::opt<bool> PGSOColdCodeOnlyForInstrPGO;
+extern llvm::cl::opt<bool> PGSOColdCodeOnlyForSamplePGO;
 extern llvm::cl::opt<bool> ForcePGSO;
 extern llvm::cl::opt<int> PgsoCutoffInstrProf;
 extern llvm::cl::opt<int> PgsoCutoffSampleProf;
 
 namespace llvm {
 
 class BasicBlock;
 class BlockFrequencyInfo;
if (ForcePGSO)
     return true;
   if (!EnablePGSO)
     return false;
   // Temporarily enable size optimizations only for the IR pass or test query
   // sites for gradual commit/rollout. This is to be removed later.
   if (PGSOIRPassOrTestOnly && !(QueryType == PGSOQueryType::IRPass ||
                                 QueryType == PGSOQueryType::Test))
     return false;
   if (PGSOColdCodeOnly ||
+      (PSI->hasInstrumentationProfile() && PGSOColdCodeOnlyForInstrPGO) ||
+      (PSI->hasSampleProfile() && PGSOColdCodeOnlyForSamplePGO) ||
       (PGSOLargeWorkingSetSizeOnly && !PSI->hasLargeWorkingSetSize())) {
     // Even if the working set size isn't large, size-optimize cold code.
     return AdapterT::isFunctionColdInCallGraph(F, PSI, *BFI);
   }
   return !AdapterT::isFunctionHotInCallGraphNthPercentile(
       PSI->hasSampleProfile() ? PgsoCutoffSampleProf : PgsoCutoffInstrProf,
       F, PSI, *BFI);
 }
 
 template<typename AdapterT, typename BlockTOrBlockFreq, typename BFIT>
 bool shouldOptimizeForSizeImpl(BlockTOrBlockFreq BBOrBlockFreq, ProfileSummaryInfo *PSI,
                                BFIT *BFI, PGSOQueryType QueryType) {
   if (!PSI || !BFI || !PSI->hasProfileSummary())
     return false;
   if (ForcePGSO)
     return true;
   if (!EnablePGSO)
     return false;
   // Temporarily enable size optimizations only for the IR pass or test query
   // sites for gradual commit/rollout. This is to be removed later.
   if (PGSOIRPassOrTestOnly && !(QueryType == PGSOQueryType::IRPass ||
                                 QueryType == PGSOQueryType::Test))
     return false;
   if (PGSOColdCodeOnly ||
+      (PSI->hasInstrumentationProfile() && PGSOColdCodeOnlyForInstrPGO) ||
+      (PSI->hasSampleProfile() && PGSOColdCodeOnlyForSamplePGO) ||
       (PGSOLargeWorkingSetSizeOnly && !PSI->hasLargeWorkingSetSize())) {
     // Even if the working set size isn't large, size-optimize cold code.
     return AdapterT::isColdBlock(BBOrBlockFreq, PSI, BFI);
   }
   return !AdapterT::isHotBlockNthPercentile(
       PSI->hasSampleProfile() ? PgsoCutoffSampleProf : PgsoCutoffInstrProf,
       BBOrBlockFreq, PSI, BFI);
 }

llvm/lib/Transforms/Utils/SizeOpts.cpp

cl::opt<bool> PGSOLargeWorkingSetSizeOnly(
     cl::desc("Apply the profile guided size optimizations only "
              "if the working set size is large (except for cold code.)"));
 
 cl::opt<bool> PGSOColdCodeOnly(
     "pgso-cold-code-only", cl::Hidden, cl::init(true),
     cl::desc("Apply the profile guided size optimizations only "
              "to cold code."));
 
+cl::opt<bool> PGSOColdCodeOnlyForInstrPGO(
+    "pgso-cold-code-only-for-instr-pgo", cl::Hidden, cl::init(true),
+    cl::desc("Apply the profile guided size optimizations only "
+             "to cold code under instrumentation PGO."));
+
+cl::opt<bool> PGSOColdCodeOnlyForSamplePGO(
+    "pgso-cold-code-only-for-sample-pgo", cl::Hidden, cl::init(true),
+    cl::desc("Apply the profile guided size optimizations only "
+             "to cold code under sample PGO."));
+
 cl::opt<bool> PGSOIRPassOrTestOnly(
     "pgso-ir-pass-or-test-only", cl::Hidden, cl::init(false),
     cl::desc("Apply the profile guided size optimizations only"
              "to the IR passes or tests."));
 
 cl::opt<bool> ForcePGSO(
     "force-pgso", cl::Hidden, cl::init(false),
     cl::desc("Force the (profiled-guided) size optimizations. "));
 
 cl::opt<int> PgsoCutoffInstrProf(
-    "pgso-cutoff-instr-prof", cl::Hidden, cl::init(250000), cl::ZeroOrMore,
+    "pgso-cutoff-instr-prof", cl::Hidden, cl::init(950000), cl::ZeroOrMore,
     cl::desc("The profile guided size optimization profile summary cutoff "
              "for instrumentation profile."));
 
 cl::opt<int> PgsoCutoffSampleProf(
     "pgso-cutoff-sample-prof", cl::Hidden, cl::init(800000), cl::ZeroOrMore,
     cl::desc("The profile guided size optimization profile summary cutoff "
              "for sample profile."));