https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98801
Bug ID: 98801
Summary: Request for a conditional move built-in function
Product: gcc
Version: 10.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: jeffhurchalla at gmail dot com
Target Milestone: ---
There are a number of idiomatic ways to hint to gcc to generate a conditional
move, but there is no way that is documented/guaranteed. In practice, they do
not always result in a conditional move. I would like to request a builtin
function that generates a conditional move.
The current situation:
Perhaps the most common hint is to write
a = (cond) ? x : y
Another way is to bit-hack, and hope the compiler transforms it to cmov. E.g.
a = ((-cond) & x) | ((cond-1) & y)
There are more ways still, see
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93165 . None can be relied on for
the purpose.
Certainly a conditional move can be accomplished via inline asm, but this is
non-portable across ISAs, it's bug-prone, and it seems to hinder the optimizer.
The existing function __builtin_expect_with_probability is interesting: in
theory it seems like it might allow a user to get conditional moves via a
probability of 0.5. I'm unsure if this works as desired in practice; its
documentation does not mention anything about conditional moves or
unpredictable branches. (Also note that a probability of 0.5 does not
necessarily mean a condition is unpredictable - e.g. a condition could
alternate between true and false each time it's executed.)
Perhaps also interesting is clang's __builtin_unpredictable, which is
documented as a way to indicate to the compiler that a branch condition is
unpredictable. I'm unsure here too if this affects conditional moves; their
docs don't mention it.
Personally, I'd like conditional moves for performance in situations where I
know a branch is completely unpredictable. However, cmovs appear to also be
useful for security, to avoid timing attacks.
