234 lines · plain
1===========================2LLVM Branch Weight Metadata3===========================4 5.. contents::6 :local:7 8Introduction9============10 11Branch Weight Metadata represents branch weights as its likeliness to be taken12(see :doc:`BlockFrequencyTerminology`). Metadata is assigned to a13terminator ``Instruction`` as an ``MDNode`` of the ``MD_prof`` kind.14The first operand is always an ``MDString`` node with the string15"branch_weights". The number of operands depends on the terminator type.16 17Branch weights might be fetched from the profiling file or generated based on18`__builtin_expect`_ and `__builtin_expect_with_probability`_ instructions.19 20All weights are represented as unsigned 32-bit values, where a higher value21indicates a greater chance of being taken.22 23Supported Instructions24======================25 26``BranchInst``27^^^^^^^^^^^^^^28 29Metadata is only assigned to conditional branches. There are two extra30operands for the true and the false branch.31We optionally track if the metadata was added by ``__builtin_expect`` or32``__builtin_expect_with_probability`` with an optional field ``!"expected"``.33 34.. code-block:: none35 36 !0 = !{37 !"branch_weights",38 [ !"expected", ]39 i32 <TRUE_BRANCH_WEIGHT>,40 i32 <FALSE_BRANCH_WEIGHT>41 }42 43``SwitchInst``44^^^^^^^^^^^^^^45 46Branch weights are assigned to every case (including the ``default`` case, which47is always case #0).48 49.. code-block:: none50 51 !0 = !{52 !"branch_weights",53 [ !"expected", ]54 i32 <DEFAULT_BRANCH_WEIGHT>55 [ , i32 <CASE_BRANCH_WEIGHT> ... ]56 }57 58``IndirectBrInst``59^^^^^^^^^^^^^^^^^^60 61Branch weights are assigned to every destination.62 63.. code-block:: none64 65 !0 = !{66 !"branch_weights",67 [ !"expected", ]68 i32 <LABEL_BRANCH_WEIGHT>69 [ , i32 <LABEL_BRANCH_WEIGHT> ... ]70 }71 72``CallInst``73^^^^^^^^^^^^^^^^^^74 75Calls may have branch weight metadata, containing the execution count of76the call. It is currently used in SamplePGO mode only, to augment the77block and entry counts, which may not be accurate with sampling.78 79.. code-block:: none80 81 !0 = !{82 !"branch_weights",83 [ !"expected", ]84 i32 <CALL_BRANCH_WEIGHT>85 }86 87``InvokeInst``88^^^^^^^^^^^^^^^^^^89 90Invoke instruction may have branch weight metadata with one or two weights.91The second weight is optional and corresponds to the unwind branch.92If only one weight is set, then it contains the execution count of the call93and used in SamplePGO mode only as described for the call instruction. If both94weights are specified then the second weight contains the count of unwind branch95taken and the first weight contains the execution count of the call minus96the count of unwind branch taken. Both weights specified are used to calculate97BranchProbability as for BranchInst and for SamplePGO the sum of both weights98is used.99 100.. code-block:: none101 102 !0 = !{103 !"branch_weights",104 [ !"expected", ]105 i32 <INVOKE_NORMAL_WEIGHT>106 [ , i32 <INVOKE_UNWIND_WEIGHT> ]107 }108 109Other110^^^^^111 112Other terminator instructions are not allowed to contain Branch Weight Metadata.113 114.. _\__builtin_expect:115 116Built-in ``expect`` Instructions117================================118 119``__builtin_expect(long exp, long c)`` instruction provides branch prediction120information. The return value is the value of ``exp``.121 122It is especially useful in conditional statements. Currently Clang supports two123conditional statements:124 125``if`` statement126^^^^^^^^^^^^^^^^127 128The ``exp`` parameter is the condition. The ``c`` parameter is the expected129comparison value. If it is equal to 1 (true), the condition is likely to be130true, in other case condition is likely to be false. For example:131 132.. code-block:: c++133 134 if (__builtin_expect(x > 0, 1)) {135 // This block is likely to be taken.136 }137 138``switch`` statement139^^^^^^^^^^^^^^^^^^^^140 141The ``exp`` parameter is the value. The ``c`` parameter is the expected142value. If the expected value doesn't appear in the cases list, the ``default``143case is assumed to be likely taken.144 145.. code-block:: c++146 147 switch (__builtin_expect(x, 5)) {148 default: break;149 case 0: // ...150 case 3: // ...151 case 5: // This case is likely to be taken.152 }153 154.. _\__builtin_expect_with_probability:155 156Built-in ``expect.with.probability`` Instruction157================================================158 159``__builtin_expect_with_probability(long exp, long c, double probability)`` has160the same semantics as ``__builtin_expect``, but the caller provides the161probability that ``exp == c``. The last argument ``probability`` must be162a constant floating-point expression and be in the range [0.0, 1.0] inclusive.163The usage is also similar as ``__builtin_expect``, for example:164 165``if`` statement166^^^^^^^^^^^^^^^^167 168If the expected comparison value ``c`` is equal to 1(true), and probability169value ``probability`` is set to 0.8, that means the probability of condition170being true is 80% while that of false is 20%.171 172.. code-block:: c++173 174 if (__builtin_expect_with_probability(x > 0, 1, 0.8)) {175 // This block is likely to be taken with probability 80%.176 }177 178``switch`` statement179^^^^^^^^^^^^^^^^^^^^180 181This is similar to the ``switch`` statement in ``__builtin_expect``.182The probability that ``exp`` is equal to the expected value is given in183the third argument ``probability``, while the probability of other value is184the average of remaining probability(``1.0 - probability``). For example:185 186.. code-block:: c++187 188 switch (__builtin_expect_with_probability(x, 5, 0.7)) {189 default: break; // Take this case with probability 10%190 case 0: break; // Take this case with probability 10%191 case 3: break; // Take this case with probability 10%192 case 5: break; // This case is likely to be taken with probability 70%193 }194 195CFG Modifications196=================197 198Branch Weight Metadata is not proof against CFG changes. If terminator operands'199are changed, some action should be taken. Otherwise, misoptimizations may200occur due to incorrect branch prediction information.201 202Function Entry Counts203=====================204 205To allow comparing different functions during inter-procedural analysis and206optimization, ``MD_prof`` nodes can also be assigned to a function definition.207The first operand is a string indicating the name of the associated counter.208 209Currently, one counter is supported: "function_entry_count". The second operand210is a 64-bit counter that indicates the number of times that this function was211invoked (in the case of instrumentation-based profiles). In the case of212sampling-based profiles, this operand is an approximation of how many times213the function was invoked.214 215For example, in the code below, the instrumentation for function ``foo()``216indicates that it was called 2,590 times at runtime.217 218.. code-block:: llvm219 220 define i32 @foo() !prof !1 {221 ret i32 0222 }223 !1 = !{!"function_entry_count", i64 2590}224 225If "function_entry_count" has more than 2 operands, the subsequent operands are226the GUID of the functions that need to be imported by ThinLTO. This is only227set by sampling-based profile. It is needed because the sampling-based profile228was collected on a binary that had already imported and inlined these functions,229and we need to ensure the IR matches in the ThinLTO backends for profile230annotation. The reason why we cannot annotate this on the callsite is that it231can only go down 1 level in the call chain. For the cases where232``foo_in_a_cc()->bar_in_b_cc()->baz_in_c_cc()``, we will need to go down 2 levels233in the call chain to import both ``bar_in_b_cc`` and ``baz_in_c_cc``.234