https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117650
Bug ID: 117650
Summary: __glibcxx_assert_fail should be marked
__attribute__((cold))
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: redbeard0531 at gmail dot com
Target Milestone: ---
The cold attribute marks incoming edges much more strongly cold than
__builtin_expect (which IIRC currently assumes only a 90/10 bias). It even
appears to be stronger than __builtin_expect_with_probability when passed
0/100, and is treated closer to how a thrown exception is, eg by splitting the
function into hot and cold parts, so that the cold part doesn't waste icache
and iTLB space that could go to another function. Since __glibcxx_assert_fail
should be even less frequent than a thrown exception (ideally never, if
software didn't have all those pesky bugs...), it should be optimized against
at least as strongly.
https://godbolt.org/z/svjPzrf5z Shows the impact on codegen by making
__GLIBCXX_NORETURN imply cold in addition to noreturn. This seemed like the
easiest way to demonstrate this. It also might be worth considering since
noreturn functions tend to be cold ones. We have noreturn imply cold in our
codebase, FWIW.
PS- ABI concerns may make it too late to change this, but it may also be worth
considering having the fail function take a single pointer to a statically
allocated struct with all of the arguments, rather than generating them as
separate arguments. As you can see, there are 4 instructions filling in the
arguments where it could just be 1. Ideally there would be some way to encode
the string addresses such that they don't need runtime relocations at startup,
eg, by encoding them as the offset from the static struct pointer to the
string, then decoding them only in the fail function.
