1 ===========================
2 LLVM Branch Weight Metadata
3 ===========================
11 Branch Weight Metadata represents branch weights as its likeliness to be taken
12 (see :doc:`BlockFrequencyTerminology`). Metadata is assigned to an
13 ``Instruction`` that is a terminator as a ``MDNode`` of the ``MD_prof`` kind.
14 The first operator is always a ``MDString`` node with the string
15 "branch_weights". Number of operators depends on the terminator type.
17 Branch weights might be fetch from the profiling file, or generated based on
18 `__builtin_expect`_ instruction.
20 All weights are represented as an unsigned 32-bit values, where higher value
21 indicates greater chance to be taken.
23 Supported Instructions
24 ======================
29 Metadata is only assigned to the conditional branches. There are two extra
30 operands for the true and the false branch.
35 metadata !"branch_weights",
36 i32 <TRUE_BRANCH_WEIGHT>,
37 i32 <FALSE_BRANCH_WEIGHT>
43 Branch weights are assigned to every case (including the ``default`` case which
49 metadata !"branch_weights",
50 i32 <DEFAULT_BRANCH_WEIGHT>
51 [ , i32 <CASE_BRANCH_WEIGHT> ... ]
57 Branch weights are assigned to every destination.
62 metadata !"branch_weights",
63 i32 <LABEL_BRANCH_WEIGHT>
64 [ , i32 <LABEL_BRANCH_WEIGHT> ... ]
70 Calls may have branch weight metadata, containing the execution count of
71 the call. It is currently used in SamplePGO mode only, to augment the
72 block and entry counts which may not be accurate with sampling.
77 metadata !"branch_weights",
78 i32 <CALL_BRANCH_WEIGHT>
84 Other terminator instructions are not allowed to contain Branch Weight Metadata.
86 .. _\__builtin_expect:
88 Built-in ``expect`` Instructions
89 ================================
91 ``__builtin_expect(long exp, long c)`` instruction provides branch prediction
92 information. The return value is the value of ``exp``.
94 It is especially useful in conditional statements. Currently Clang supports two
95 conditional statements:
100 The ``exp`` parameter is the condition. The ``c`` parameter is the expected
101 comparison value. If it is equal to 1 (true), the condition is likely to be
102 true, in other case condition is likely to be false. For example:
106 if (__builtin_expect(x > 0, 1)) {
107 // This block is likely to be taken.
113 The ``exp`` parameter is the value. The ``c`` parameter is the expected
114 value. If the expected value doesn't show on the cases list, the ``default``
115 case is assumed to be likely taken.
119 switch (__builtin_expect(x, 5)) {
123 case 5: // This case is likely to be taken.
129 Branch Weight Metatada is not proof against CFG changes. If terminator operands'
130 are changed some action should be taken. In other case some misoptimizations may
131 occur due to incorrect branch prediction information.
133 Function Entry Counts
134 =====================
136 To allow comparing different functions during inter-procedural analysis and
137 optimization, ``MD_prof`` nodes can also be assigned to a function definition.
138 The first operand is a string indicating the name of the associated counter.
140 Currently, one counter is supported: "function_entry_count". The second operand
141 is a 64-bit counter that indicates the number of times that this function was
142 invoked (in the case of instrumentation-based profiles). In the case of
143 sampling-based profiles, this operand is an approximation of how many times
144 the function was invoked.
146 For example, in the code below, the instrumentation for function foo()
147 indicates that it was called 2,590 times at runtime.
151 define i32 @foo() !prof !1 {
154 !1 = !{!"function_entry_count", i64 2590}
156 If "function_entry_count" has more than 2 operands, the later operands are
157 the GUID of the functions that needs to be imported by ThinLTO. This is only
158 set by sampling based profile. It is needed because the sampling based profile
159 was collected on a binary that had already imported and inlined these functions,
160 and we need to ensure the IR matches in the ThinLTO backends for profile
161 annotation. The reason why we cannot annotate this on the callsite is that it
162 can only goes down 1 level in the call chain. For the cases where
163 foo_in_a_cc()->bar_in_b_cc()->baz_in_c_cc(), we will need to go down 2 levels
164 in the call chain to import both bar_in_b_cc and baz_in_c_cc.