• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1================================
2LLVM Block Frequency Terminology
3================================
4
5.. contents::
6   :local:
7
8Introduction
9============
10
11Block Frequency is a metric for estimating the relative frequency of different
12basic blocks.  This document describes the terminology that the
13``BlockFrequencyInfo`` and ``MachineBlockFrequencyInfo`` analysis passes use.
14
15Branch Probability
16==================
17
18Blocks with multiple successors have probabilities associated with each
19outgoing edge.  These are called branch probabilities.  For a given block, the
20sum of its outgoing branch probabilities should be 1.0.
21
22Branch Weight
23=============
24
25Rather than storing fractions on each edge, we store an integer weight.
26Weights are relative to the other edges of a given predecessor block.  The
27branch probability associated with a given edge is its own weight divided by
28the sum of the weights on the predecessor's outgoing edges.
29
30For example, consider this IR:
31
32.. code-block:: llvm
33
34   define void @foo() {
35       ; ...
36       A:
37           br i1 %cond, label %B, label %C, !prof !0
38       ; ...
39   }
40   !0 = metadata !{metadata !"branch_weights", i32 7, i32 8}
41
42and this simple graph representation::
43
44   A -> B  (edge-weight: 7)
45   A -> C  (edge-weight: 8)
46
47The probability of branching from block A to block B is 7/15, and the
48probability of branching from block A to block C is 8/15.
49
50See :doc:`BranchWeightMetadata` for details about the branch weight IR
51representation.
52
53Block Frequency
54===============
55
56Block frequency is a relative metric that represents the number of times a
57block executes.  The ratio of a block frequency to the entry block frequency is
58the expected number of times the block will execute per entry to the function.
59
60Block frequency is the main output of the ``BlockFrequencyInfo`` and
61``MachineBlockFrequencyInfo`` analysis passes.
62
63Implementation: a series of DAGs
64================================
65
66The implementation of the block frequency calculation analyses each loop,
67bottom-up, ignoring backedges; i.e., as a DAG.  After each loop is processed,
68it's packaged up to act as a pseudo-node in its parent loop's (or the
69function's) DAG analysis.
70
71Block Mass
72==========
73
74For each DAG, the entry node is assigned a mass of ``UINT64_MAX`` and mass is
75distributed to successors according to branch weights.  Block Mass uses a
76fixed-point representation where ``UINT64_MAX`` represents ``1.0`` and ``0``
77represents a number just above ``0.0``.
78
79After mass is fully distributed, in any cut of the DAG that separates the exit
80nodes from the entry node, the sum of the block masses of the nodes succeeded
81by a cut edge should equal ``UINT64_MAX``.  In other words, mass is conserved
82as it "falls" through the DAG.
83
84If a function's basic block graph is a DAG, then block masses are valid block
85frequencies.  This works poorly in practice though, since downstream users rely
86on adding block frequencies together without hitting the maximum.
87
88Loop Scale
89==========
90
91Loop scale is a metric that indicates how many times a loop iterates per entry.
92As mass is distributed through the loop's DAG, the (otherwise ignored) backedge
93mass is collected.  This backedge mass is used to compute the exit frequency,
94and thus the loop scale.
95
96Implementation: Getting from mass and scale to frequency
97========================================================
98
99After analysing the complete series of DAGs, each block has a mass (local to
100its containing loop, if any), and each loop pseudo-node has a loop scale and
101its own mass (from its parent's DAG).
102
103We can get an initial frequency assignment (with entry frequency of 1.0) by
104multiplying these masses and loop scales together.  A given block's frequency
105is the product of its mass, the mass of containing loops' pseudo nodes, and the
106containing loops' loop scales.
107
108Since downstream users need integers (not floating point), this initial
109frequency assignment is shifted as necessary into the range of ``uint64_t``.
110
111Block Bias
112==========
113
114Block bias is a proposed *absolute* metric to indicate a bias toward or away
115from a given block during a function's execution.  The idea is that bias can be
116used in isolation to indicate whether a block is relatively hot or cold, or to
117compare two blocks to indicate whether one is hotter or colder than the other.
118
119The proposed calculation involves calculating a *reference* block frequency,
120where:
121
122* every branch weight is assumed to be 1 (i.e., every branch probability
123  distribution is even) and
124
125* loop scales are ignored.
126
127This reference frequency represents what the block frequency would be in an
128unbiased graph.
129
130The bias is the ratio of the block frequency to this reference block frequency.
131