https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88793
Bug ID: 88793
Summary: Document that __attribute__ ((cold)) is not equivalent
to __builtin_except because of optimization for size
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Keywords: documentation
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: fw at gcc dot gnu.org
Target Milestone: ---
The documentation says this:
'cold'
The 'cold' attribute on functions is used to inform the compiler
that the function is unlikely to be executed. The function is
optimized for size rather than speed and on many targets it is
placed into a special subsection of the text section so all cold
functions appear close together, improving code locality of
non-cold parts of program. The paths leading to calls of cold
functions within code are marked as unlikely by the branch
prediction mechanism. It is thus useful to mark functions used to
handle unlikely conditions, such as 'perror', as cold to improve
optimization of hot functions that do call marked functions in rare
occasions.
However, optimizing for size is a very big hammer and causes substantial
performance issues on i386 and x86-64 due to string function inlining. As a
result, the cold attribute is only suitable for code that is basically never
executed. For other cases, like repeated execution which only happens in an
unlikely configuration, it is inappropriate. The documentation does not make
this clear enough, and programmers use __attribute__ ((cold)) and the
propagation into call sites as a more convenient alternative for
__builtin_expect, leading to performance issues.
The referenced downstream bug concerns some old GCC version where GCC
incorrectly inferred cold code regions from __builtin_expect. I'm including it
here to indicate that the distinction matters in practice.
