On 2024-09-27 19:23, Linus Torvalds wrote:
On Fri, 27 Sept 2024 at 10:17, Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
The barrier() is ineffective at fixing the issue.
It does not prevent the compiler CSE from losing the
address dependency:
Ok. Thanks for actually specifying code.
That needs to be
(a) in a comment
OK. I'll add the code/asm examples to the comment above ADDRESS_EQ().
(b) the value barrier needs to be on *both* values so that the order
of the equality testing doesn't matter.
If we use OPTIMIZER_HIDE_VAR() on both parameters, it indeed minimizes
the odds that someone get the order wrong, but it disallows using
ADDRESS_EQ() with a constant parameter
(e.g. ADDRESS_EQ(ptr, &mystruct)), which would be nice to cover. It
works fine with using OPTIMIZER_HIDE_VAR() on the first argument,
but it opens the door to misuses.
Perhaps there is a trick with compiler builtins we could do to only
use OPTIMIZER_HIDE_VAR() on non-constant arguments, but I can't get
it to work so far.
I'm preparing a small series that aims to show how a minimal
hazard pointer implementation can help improve common scenarios:
I want actual numbers on real loads. Just so you know. Not "this can
help". But "this actually really _does_ help".
Noted, thanks!
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
