[misexpect] Re-implement MisExpect Diagnostics

Reimplements MisExpect diagnostics from D66324 to reconstruct its
original checking methodology only using MD_prof branch_weights
metadata.

New checks rely on 2 invariants:

1) For frontend instrumentation, MD_prof branch_weights will always be
   populated before llvm.expect intrinsics are lowered.

2) for IR and sample profiling, llvm.expect intrinsics will always be
   lowered before branch_weights are populated from the IR profiles.

These invariants allow the checking to assume how the existing branch
weights are populated depending on the profiling method used, and emit
the correct diagnostics. If these invariants are ever invalidated, the
MisExpect related checks would need to be updated, potentially by
re-introducing MD_misexpect metadata, and ensuring it always will be
transformed the same way as branch_weights in other optimization passes.

Frontend based profiling is now enabled without using LLVM Args, by
introducing a new CodeGen option, and checking if the -Wmisexpect flag
has been passed on the command line.

Reviewed By: tejohnson

Differential Revision: https://reviews.llvm.org/D115907

Author: ilovepi

clang/docs/MisExpect.rst

@@ -0,0 +1,75 @@
+===================
+Misexpect
+===================
+.. contents::
+
+.. toctree::
+   :maxdepth: 1
+
+When developers use ``llvm.expect`` intrinsics, i.e., through use of
+``__builtin_expect(...)``, they are trying to communicate how their code is
+expected to behave at runtime to the optimizer. These annotations, however, can
+be incorrect for a variety of reasons: changes to the code base invalidate them
+silently, the developer mis-annotated them (e.g., using ``LIKELY`` instead of
+``UNLIKELY``), or perhaps they assumed something incorrectly when they wrote
+the annotation. Regardless of why, it is useful to detect these situations so
+that the optimizer can make more useful decisions about the code. 
+
+MisExpect diagnostics are intended to help developers identify and address
+these situations, by comparing the branch weights added by the ``llvm.expect``
+intrinsic to those collected through profiling. Whenever these values are
+mismatched, a diagnostic is surfaced to the user. Details on how the checks
+operate in the LLVM backed can be found in LLVM's documentation.
+
+By default MisExpect checking is quite strict, because the use of the
+``llvm.expect`` intrinsic is designed for specialized cases, where the outcome
+of a condition is severely skewed. As a result, the optimizer can be extremely
+aggressive, which can result in performance degradation if the outcome is less
+predictable than the annotation suggests. Even when the annotation is correct
+90% of the time, it may be beneficial to either remove the annotation or to use
+a different intrinsic that can communicate the probability more directly.
+
+Because this may be too strict, MisExpect diagnostics are not enabled by
+default, and support an additional flag to tolerate some deviation from the
+exact thresholds. The ``-fdiagnostic-misexpect-tolerance=N`` accepts
+deviations when comparing branch weights within ``N%`` of the expected values.
+So passing ``-fdiagnostic-misexpect-tolerance=5`` will not report diagnostic messages
+if the branch weight from the profile is within 5% of the weight added by
+the ``llvm.expect`` intrinsic.
+
+MisExpect diagnostics are also available in the form of optimization remarks,
+which can be serialized and processed through the ``opt-viewer.py``
+scripts in LLVM.
+
+.. option:: -Rpass=misexpect
+
+  Enables optimization remarks for misexpect when profiling data conflicts with
+  use of ``llvm.expect`` intrinsics.
+
+
+.. option:: -Wmisexpect
+
+  Enables misexpect warnings when profiling data conflicts with use of
+  ``llvm.expect`` intrinsics.
+
+.. option:: -fdiagnostic-misexpect-tolerance=N
+
+   Relaxes misexpect checking to tolerate profiling values within N% of the 
+   expected branch weight. e.g., a value of ``N=5`` allows misexpect to check against
+   ``0.95 * Threshold``
+
+LLVM supports 4 types of profile formats: Frontend, IR, CS-IR, and
+Sampling. MisExpect Diagnostics are compatible with all Profiling formats.
+
++----------------+--------------------------------------------------------------------------------------+
+| Profile Type   | Description                                                                          |
++================+======================================================================================+
+| Frontend       | Profiling instrumentation added during compilation by the frontend, i.e. ``clang``   |
++----------------+--------------------------------------------------------------------------------------+
+| IR             | Profiling instrumentation added during by the LLVM backend                           |
++----------------+--------------------------------------------------------------------------------------+
+| CS-IR          | Context Sensitive IR based profiles                                                  |
++----------------+--------------------------------------------------------------------------------------+
+| Sampling       | Profiles collected through sampling with external tools, such as ``perf`` on Linux   |
++----------------+--------------------------------------------------------------------------------------+
+

clang/docs/ReleaseNotes.rst

@@ -157,6 +157,9 @@ Improvements to Clang's diagnostics
   function with a prototype. e.g., ``void f(int); void f() {}`` is now properly
   diagnosed.
 
+- ``-Wmisexpect`` warns when the branch weights collected during profiling
+  conflict with those added by ``llvm.expect``.
+
 Non-comprehensive list of changes in this release
 -------------------------------------------------
 - Improve __builtin_dump_struct:

clang/docs/index.rst

@@ -42,6 +42,7 @@ Using Clang as a Compiler
    SourceBasedCodeCoverage
    Modules
    MSVCCompatibility
+   MisExpect
    OpenCLSupport
    OpenMPSupport
    SYCLSupport

clang/include/clang/Basic/CodeGenOptions.def

@@ -174,6 +174,7 @@ CODEGENOPT(NoExecStack       , 1, 0) ///< Set when -Wa,--noexecstack is enabled.
 CODEGENOPT(FatalWarnings     , 1, 0) ///< Set when -Wa,--fatal-warnings is
                                      ///< enabled.
 CODEGENOPT(NoWarn            , 1, 0) ///< Set when -Wa,--no-warn is enabled.
+CODEGENOPT(MisExpect         , 1, 0) ///< Set when -Wmisexpect is enabled
 CODEGENOPT(EnableSegmentedStacks , 1, 0) ///< Set when -fsplit-stack is enabled.
 CODEGENOPT(NoInlineLineTables, 1, 0) ///< Whether debug info should contain
                                      ///< inline line tables.

clang/include/clang/Basic/CodeGenOptions.h

@@ -423,6 +423,10 @@ class CodeGenOptions : public CodeGenOptionsBase {
   /// If threshold option is not specified, it is disabled by default.
   Optional<uint64_t> DiagnosticsHotnessThreshold = 0;
 
+  /// The maximum percentage profiling weights can deviate from the expected
+  /// values in order to be included in misexpect diagnostics.
+  Optional<uint64_t> DiagnosticsMisExpectTolerance = 0;
+
 public:
   // Define accessors/mutators for code generation options of enumeration type.
 #define CODEGENOPT(Name, Bits, Default)

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -149,6 +149,8 @@ def err_drv_invalid_darwin_version : Error<
   "invalid Darwin version number: %0">;
 def err_drv_invalid_diagnotics_hotness_threshold : Error<
   "invalid argument in '%0', only integer or 'auto' is supported">;
+def err_drv_invalid_diagnotics_misexpect_tolerance : Error<
+  "invalid argument in '%0', only integers are supported">;
 def err_drv_missing_argument : Error<
   "argument to '%0' is missing (expected %1 value%s1)">;
 def err_drv_invalid_Xarch_argument_with_args : Error<
@@ -378,6 +380,9 @@ def warn_drv_empty_joined_argument : Warning<
 def warn_drv_diagnostics_hotness_requires_pgo : Warning<
   "argument '%0' requires profile-guided optimization information">,
   InGroup<UnusedCommandLineArgument>;
+def warn_drv_diagnostics_misexpect_requires_pgo : Warning<
+  "argument '%0' requires profile-guided optimization information">,
+  InGroup<UnusedCommandLineArgument>;
 def warn_drv_clang_unsupported : Warning<
   "the clang compiler does not support '%0'">;
 def warn_drv_deprecated_arg : Warning<

clang/include/clang/Basic/DiagnosticFrontendKinds.td

@@ -315,6 +315,11 @@ def warn_profile_data_missing : Warning<
 def warn_profile_data_unprofiled : Warning<
   "no profile data available for file \"%0\"">,
   InGroup<ProfileInstrUnprofiled>;
+def warn_profile_data_misexpect : Warning<
+  "Potential performance regression from use of __builtin_expect(): "
+  "Annotation was correct on %0 of profiled executions.">,
+  BackendInfo,
+  InGroup<MisExpect>;
 } // end of instrumentation issue category
 
 }

clang/include/clang/Basic/DiagnosticGroups.td

@@ -1256,6 +1256,7 @@ def BackendWarningAttributes : DiagGroup<"attribute-warning">;
 def ProfileInstrMissing : DiagGroup<"profile-instr-missing">;
 def ProfileInstrOutOfDate : DiagGroup<"profile-instr-out-of-date">;
 def ProfileInstrUnprofiled : DiagGroup<"profile-instr-unprofiled">;
+def MisExpect : DiagGroup<"misexpect">;
 
 // AddressSanitizer frontend instrumentation remarks.
 def SanitizeAddressRemarks : DiagGroup<"sanitize-address">;

clang/include/clang/Driver/Options.td

@@ -1430,6 +1430,9 @@ def fdiagnostics_hotness_threshold_EQ : Joined<["-"], "fdiagnostics-hotness-thre
     Group<f_Group>, Flags<[CC1Option]>, MetaVarName<"<value>">,
     HelpText<"Prevent optimization remarks from being output if they do not have at least this profile count. "
     "Use 'auto' to apply the threshold from profile summary">;
+def fdiagnostics_misexpect_tolerance_EQ : Joined<["-"], "fdiagnostics-misexpect-tolerance=">,
+    Group<f_Group>, Flags<[CC1Option]>, MetaVarName<"<value>">,
+    HelpText<"Prevent misexpect diagnostics from being output if the profile counts are within N% of the expected. ">;
 defm diagnostics_show_option : BoolFOption<"diagnostics-show-option",
     DiagnosticOpts<"ShowOptionNames">, DefaultTrue,
     NegFlag<SetFalse, [CC1Option]>, PosFlag<SetTrue, [], "Print option name with mappable diagnostics">>;

clang/lib/CodeGen/BackendUtil.cpp

@@ -474,6 +474,7 @@ static bool initTargetOptions(DiagnosticsEngine &Diags,
           Entry.IgnoreSysRoot ? Entry.Path : HSOpts.Sysroot + Entry.Path);
   Options.MCOptions.Argv0 = CodeGenOpts.Argv0;
   Options.MCOptions.CommandLineArgs = CodeGenOpts.CommandLineArgs;
+  Options.MisExpect = CodeGenOpts.MisExpect;
 
   return true;
 }