|
| 1 | +=================== |
| 2 | +Misexpect |
| 3 | +=================== |
| 4 | +.. contents:: |
| 5 | + |
| 6 | +.. toctree:: |
| 7 | + :maxdepth: 1 |
| 8 | + |
| 9 | +When developers use ``llvm.expect`` intrinsics, i.e., through use of |
| 10 | +``__builtin_expect(...)``, they are trying to communicate how their code is |
| 11 | +expected to behave at runtime to the optimizer. These annotations, however, can |
| 12 | +be incorrect for a variety of reasons: changes to the code base invalidate them |
| 13 | +silently, the developer mis-annotated them (e.g., using ``LIKELY`` instead of |
| 14 | +``UNLIKELY``), or perhaps they assumed something incorrectly when they wrote |
| 15 | +the annotation. Regardless of why, it is useful to detect these situations so |
| 16 | +that the optimizer can make more useful decisions about the code. |
| 17 | + |
| 18 | +MisExpect diagnostics are intended to help developers identify and address |
| 19 | +these situations, by comparing the branch weights added by the ``llvm.expect`` |
| 20 | +intrinsic to those collected through profiling. Whenever these values are |
| 21 | +mismatched, a diagnostic is surfaced to the user. Details on how the checks |
| 22 | +operate in the LLVM backed can be found in LLVM's documentation. |
| 23 | + |
| 24 | +By default MisExpect checking is quite strict, because the use of the |
| 25 | +``llvm.expect`` intrinsic is designed for specialized cases, where the outcome |
| 26 | +of a condition is severely skewed. As a result, the optimizer can be extremely |
| 27 | +aggressive, which can result in performance degradation if the outcome is less |
| 28 | +predictable than the annotation suggests. Even when the annotation is correct |
| 29 | +90% of the time, it may be beneficial to either remove the annotation or to use |
| 30 | +a different intrinsic that can communicate the probability more directly. |
| 31 | + |
| 32 | +Because this may be too strict, MisExpect diagnostics are not enabled by |
| 33 | +default, and support an additional flag to tolerate some deviation from the |
| 34 | +exact thresholds. The ``-fdiagnostic-misexpect-tolerance=N`` accepts |
| 35 | +deviations when comparing branch weights within ``N%`` of the expected values. |
| 36 | +So passing ``-fdiagnostic-misexpect-tolerance=5`` will not report diagnostic messages |
| 37 | +if the branch weight from the profile is within 5% of the weight added by |
| 38 | +the ``llvm.expect`` intrinsic. |
| 39 | + |
| 40 | +MisExpect diagnostics are also available in the form of optimization remarks, |
| 41 | +which can be serialized and processed through the ``opt-viewer.py`` |
| 42 | +scripts in LLVM. |
| 43 | + |
| 44 | +.. option:: -Rpass=misexpect |
| 45 | + |
| 46 | + Enables optimization remarks for misexpect when profiling data conflicts with |
| 47 | + use of ``llvm.expect`` intrinsics. |
| 48 | + |
| 49 | + |
| 50 | +.. option:: -Wmisexpect |
| 51 | + |
| 52 | + Enables misexpect warnings when profiling data conflicts with use of |
| 53 | + ``llvm.expect`` intrinsics. |
| 54 | + |
| 55 | +.. option:: -fdiagnostic-misexpect-tolerance=N |
| 56 | + |
| 57 | + Relaxes misexpect checking to tolerate profiling values within N% of the |
| 58 | + expected branch weight. e.g., a value of ``N=5`` allows misexpect to check against |
| 59 | + ``0.95 * Threshold`` |
| 60 | + |
| 61 | +LLVM supports 4 types of profile formats: Frontend, IR, CS-IR, and |
| 62 | +Sampling. MisExpect Diagnostics are compatible with all Profiling formats. |
| 63 | + |
| 64 | ++----------------+--------------------------------------------------------------------------------------+ |
| 65 | +| Profile Type | Description | |
| 66 | ++================+======================================================================================+ |
| 67 | +| Frontend | Profiling instrumentation added during compilation by the frontend, i.e. ``clang`` | |
| 68 | ++----------------+--------------------------------------------------------------------------------------+ |
| 69 | +| IR | Profiling instrumentation added during by the LLVM backend | |
| 70 | ++----------------+--------------------------------------------------------------------------------------+ |
| 71 | +| CS-IR | Context Sensitive IR based profiles | |
| 72 | ++----------------+--------------------------------------------------------------------------------------+ |
| 73 | +| Sampling | Profiles collected through sampling with external tools, such as ``perf`` on Linux | |
| 74 | ++----------------+--------------------------------------------------------------------------------------+ |
| 75 | + |
0 commit comments