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`_ and `__builtin_expect_with_probability`_ 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.
36 i32 <TRUE_BRANCH_WEIGHT>,
37 i32 <FALSE_BRANCH_WEIGHT>
43 Branch weights are assigned to every case (including the ``default`` case which
50 i32 <DEFAULT_BRANCH_WEIGHT>
51 [ , i32 <CASE_BRANCH_WEIGHT> ... ]
57 Branch weights are assigned to every destination.
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.
78 i32 <CALL_BRANCH_WEIGHT>
84 Invoke instruction may have branch weight metadata with one or two weights.
85 The second weight is optional and corresponds to the unwind branch.
86 If only one weight is set then it contains the execution count of the call
87 and used in SamplePGO mode only as described for the call instruction. If both
88 weights are specified then the second weight contains count of unwind branch
89 taken and the first weights contains the execution count of the call minus
90 the count of unwind branch taken. Both weights specified are used to calculate
91 BranchProbability as for BranchInst and for SamplePGO the sum of both weights
98 i32 <INVOKE_NORMAL_WEIGHT>
99 [ , i32 <INVOKE_UNWIND_WEIGHT> ]
105 Other terminator instructions are not allowed to contain Branch Weight Metadata.
107 .. _\__builtin_expect:
109 Built-in ``expect`` Instructions
110 ================================
112 ``__builtin_expect(long exp, long c)`` instruction provides branch prediction
113 information. The return value is the value of ``exp``.
115 It is especially useful in conditional statements. Currently Clang supports two
116 conditional statements:
121 The ``exp`` parameter is the condition. The ``c`` parameter is the expected
122 comparison value. If it is equal to 1 (true), the condition is likely to be
123 true, in other case condition is likely to be false. For example:
127 if (__builtin_expect(x > 0, 1)) {
128 // This block is likely to be taken.
134 The ``exp`` parameter is the value. The ``c`` parameter is the expected
135 value. If the expected value doesn't show on the cases list, the ``default``
136 case is assumed to be likely taken.
140 switch (__builtin_expect(x, 5)) {
144 case 5: // This case is likely to be taken.
147 .. _\__builtin_expect_with_probability:
149 Built-in ``expect.with.probability`` Instruction
150 ================================================
152 ``__builtin_expect_with_probability(long exp, long c, double probability)`` has
153 the same semantics as ``__builtin_expect``, but the caller provides the
154 probability that ``exp == c``. The last argument ``probability`` must be
155 constant floating-point expression and be in the range [0.0, 1.0] inclusive.
156 The usage is also similar as ``__builtin_expect``, for example:
161 If the expect comparison value ``c`` is equal to 1(true), and probability
162 value ``probability`` is set to 0.8, that means the probability of condition
163 to be true is 80% while that of false is 20%.
167 if (__builtin_expect_with_probability(x > 0, 1, 0.8)) {
168 // This block is likely to be taken with probability 80%.
174 This is basically the same as ``switch`` statement in ``__builtin_expect``.
175 The probability that ``exp`` is equal to the expect value is given in
176 the third argument ``probability``, while the probability of other value is
177 the average of remaining probability(``1.0 - probability``). For example:
181 switch (__builtin_expect_with_probability(x, 5, 0.7)) {
182 default: break; // Take this case with probability 10%
183 case 0: break; // Take this case with probability 10%
184 case 3: break; // Take this case with probability 10%
185 case 5: break; // This case is likely to be taken with probability 70%
191 Branch Weight Metatada is not proof against CFG changes. If terminator operands'
192 are changed some action should be taken. In other case some misoptimizations may
193 occur due to incorrect branch prediction information.
195 Function Entry Counts
196 =====================
198 To allow comparing different functions during inter-procedural analysis and
199 optimization, ``MD_prof`` nodes can also be assigned to a function definition.
200 The first operand is a string indicating the name of the associated counter.
202 Currently, one counter is supported: "function_entry_count". The second operand
203 is a 64-bit counter that indicates the number of times that this function was
204 invoked (in the case of instrumentation-based profiles). In the case of
205 sampling-based profiles, this operand is an approximation of how many times
206 the function was invoked.
208 For example, in the code below, the instrumentation for function foo()
209 indicates that it was called 2,590 times at runtime.
213 define i32 @foo() !prof !1 {
216 !1 = !{!"function_entry_count", i64 2590}
218 If "function_entry_count" has more than 2 operands, the later operands are
219 the GUID of the functions that needs to be imported by ThinLTO. This is only
220 set by sampling based profile. It is needed because the sampling based profile
221 was collected on a binary that had already imported and inlined these functions,
222 and we need to ensure the IR matches in the ThinLTO backends for profile
223 annotation. The reason why we cannot annotate this on the callsite is that it
224 can only goes down 1 level in the call chain. For the cases where
225 foo_in_a_cc()->bar_in_b_cc()->baz_in_c_cc(), we will need to go down 2 levels
226 in the call chain to import both bar_in_b_cc and baz_in_c_cc.