Home | History | Annotate | Download | only in docs
      1 ===========================
      2 LLVM Branch Weight Metadata
      3 ===========================
      4 
      5 .. contents::
      6    :local:
      7 
      8 Introduction
      9 ============
     10 
     11 Branch Weight Metadata represents branch weights as its likeliness to be taken
     12 (see :doc:`BlockFrequencyTerminology`). Metadata is assigned to the
     13 ``TerminatorInst`` as a ``MDNode`` of the ``MD_prof`` kind. The first operator
     14 is always a ``MDString`` node with the string "branch_weights".  Number of
     15 operators depends on the terminator type.
     16 
     17 Branch weights might be fetch from the profiling file, or generated based on
     18 `__builtin_expect`_ instruction.
     19 
     20 All weights are represented as an unsigned 32-bit values, where higher value
     21 indicates greater chance to be taken.
     22 
     23 Supported Instructions
     24 ======================
     25 
     26 ``BranchInst``
     27 ^^^^^^^^^^^^^^
     28 
     29 Metadata is only assigned to the conditional branches. There are two extra
     30 operands for the true and the false branch.
     31 
     32 .. code-block:: llvm
     33 
     34   !0 = metadata !{
     35     metadata !"branch_weights",
     36     i32 <TRUE_BRANCH_WEIGHT>,
     37     i32 <FALSE_BRANCH_WEIGHT>
     38   }
     39 
     40 ``SwitchInst``
     41 ^^^^^^^^^^^^^^
     42 
     43 Branch weights are assigned to every case (including the ``default`` case which
     44 is always case #0).
     45 
     46 .. code-block:: llvm
     47 
     48   !0 = metadata !{
     49     metadata !"branch_weights",
     50     i32 <DEFAULT_BRANCH_WEIGHT>
     51     [ , i32 <CASE_BRANCH_WEIGHT> ... ]
     52   }
     53 
     54 ``IndirectBrInst``
     55 ^^^^^^^^^^^^^^^^^^
     56 
     57 Branch weights are assigned to every destination.
     58 
     59 .. code-block:: llvm
     60 
     61   !0 = metadata !{
     62     metadata !"branch_weights",
     63     i32 <LABEL_BRANCH_WEIGHT>
     64     [ , i32 <LABEL_BRANCH_WEIGHT> ... ]
     65   }
     66 
     67 Other
     68 ^^^^^
     69 
     70 Other terminator instructions are not allowed to contain Branch Weight Metadata.
     71 
     72 .. _\__builtin_expect:
     73 
     74 Built-in ``expect`` Instructions
     75 ================================
     76 
     77 ``__builtin_expect(long exp, long c)`` instruction provides branch prediction
     78 information. The return value is the value of ``exp``.
     79 
     80 It is especially useful in conditional statements. Currently Clang supports two
     81 conditional statements:
     82 
     83 ``if`` statement
     84 ^^^^^^^^^^^^^^^^
     85 
     86 The ``exp`` parameter is the condition. The ``c`` parameter is the expected
     87 comparison value. If it is equal to 1 (true), the condition is likely to be
     88 true, in other case condition is likely to be false. For example:
     89 
     90 .. code-block:: c++
     91 
     92   if (__builtin_expect(x > 0, 1)) {
     93     // This block is likely to be taken.
     94   }
     95 
     96 ``switch`` statement
     97 ^^^^^^^^^^^^^^^^^^^^
     98 
     99 The ``exp`` parameter is the value. The ``c`` parameter is the expected
    100 value. If the expected value doesn't show on the cases list, the ``default``
    101 case is assumed to be likely taken.
    102 
    103 .. code-block:: c++
    104 
    105   switch (__builtin_expect(x, 5)) {
    106   default: break;
    107   case 0:  // ...
    108   case 3:  // ...
    109   case 5:  // This case is likely to be taken.
    110   }
    111 
    112 CFG Modifications
    113 =================
    114 
    115 Branch Weight Metatada is not proof against CFG changes. If terminator operands'
    116 are changed some action should be taken. In other case some misoptimizations may
    117 occur due to incorrect branch prediction information.
    118 
    119 Function Entry Counts
    120 =====================
    121 
    122 To allow comparing different functions during inter-procedural analysis and
    123 optimization, ``MD_prof`` nodes can also be assigned to a function definition.
    124 The first operand is a string indicating the name of the associated counter.
    125 
    126 Currently, one counter is supported: "function_entry_count". This is a 64-bit
    127 counter that indicates the number of times that this function was invoked (in
    128 the case of instrumentation-based profiles). In the case of sampling-based
    129 profiles, this counter is an approximation of how many times the function was
    130 invoked.
    131 
    132 For example, in the code below, the instrumentation for function foo()
    133 indicates that it was called 2,590 times at runtime.
    134 
    135 .. code-block:: llvm
    136 
    137   define i32 @foo() !prof !1 {
    138     ret i32 0
    139   }
    140   !1 = !{!"function_entry_count", i64 2590}
    141