1 ================================
2 LLVM Block Frequency Terminology
3 ================================
11 Block Frequency is a metric for estimating the relative frequency of different
12 basic blocks. This document describes the terminology that the
13 ``BlockFrequencyInfo`` and ``MachineBlockFrequencyInfo`` analysis passes use.
18 Blocks with multiple successors have probabilities associated with each
19 outgoing edge. These are called branch probabilities. For a given block, the
20 sum of its outgoing branch probabilities should be 1.0.
25 Rather than storing fractions on each edge, we store an integer weight.
26 Weights are relative to the other edges of a given predecessor block. The
27 branch probability associated with a given edge is its own weight divided by
28 the sum of the weights on the predecessor's outgoing edges.
30 For example, consider this IR:
37 br i1 %cond, label %B, label %C, !prof !0
40 !0 = metadata !{metadata !"branch_weights", i32 7, i32 8}
42 and this simple graph representation::
44 A -> B (edge-weight: 7)
45 A -> C (edge-weight: 8)
47 The probability of branching from block A to block B is 7/15, and the
48 probability of branching from block A to block C is 8/15.
50 See :doc:`BranchWeightMetadata` for details about the branch weight IR
56 Block frequency is a relative metric that represents the number of times a
57 block executes. The ratio of a block frequency to the entry block frequency is
58 the expected number of times the block will execute per entry to the function.
60 Block frequency is the main output of the ``BlockFrequencyInfo`` and
61 ``MachineBlockFrequencyInfo`` analysis passes.
63 Implementation: a series of DAGs
64 ================================
66 The implementation of the block frequency calculation analyses each loop,
67 bottom-up, ignoring backedges; i.e., as a DAG. After each loop is processed,
68 it's packaged up to act as a pseudo-node in its parent loop's (or the
69 function's) DAG analysis.
74 For each DAG, the entry node is assigned a mass of ``UINT64_MAX`` and mass is
75 distributed to successors according to branch weights. Block Mass uses a
76 fixed-point representation where ``UINT64_MAX`` represents ``1.0`` and ``0``
77 represents a number just above ``0.0``.
79 After mass is fully distributed, in any cut of the DAG that separates the exit
80 nodes from the entry node, the sum of the block masses of the nodes succeeded
81 by a cut edge should equal ``UINT64_MAX``. In other words, mass is conserved
82 as it "falls" through the DAG.
84 If a function's basic block graph is a DAG, then block masses are valid block
85 frequencies. This works poorly in practise though, since downstream users rely
86 on adding block frequencies together without hitting the maximum.
91 Loop scale is a metric that indicates how many times a loop iterates per entry.
92 As mass is distributed through the loop's DAG, the (otherwise ignored) backedge
93 mass is collected. This backedge mass is used to compute the exit frequency,
94 and thus the loop scale.
96 Implementation: Getting from mass and scale to frequency
97 ========================================================
99 After analysing the complete series of DAGs, each block has a mass (local to
100 its containing loop, if any), and each loop pseudo-node has a loop scale and
101 its own mass (from its parent's DAG).
103 We can get an initial frequency assignment (with entry frequency of 1.0) by
104 multiplying these masses and loop scales together. A given block's frequency
105 is the product of its mass, the mass of containing loops' pseudo nodes, and the
106 containing loops' loop scales.
108 Since downstream users need integers (not floating point), this initial
109 frequency assignment is shifted as necessary into the range of ``uint64_t``.
114 Block bias is a proposed *absolute* metric to indicate a bias toward or away
115 from a given block during a function's execution. The idea is that bias can be
116 used in isolation to indicate whether a block is relatively hot or cold, or to
117 compare two blocks to indicate whether one is hotter or colder than the other.
119 The proposed calculation involves calculating a *reference* block frequency,
122 * every branch weight is assumed to be 1 (i.e., every branch probability
123 distribution is even) and
125 * loop scales are ignored.
127 This reference frequency represents what the block frequency would be in an
130 The bias is the ratio of the block frequency to this reference block frequency.