Re: alpha: Checking source code positions for the setting of error codes

From: Al Viro
Date: Wed Jan 18 2017 - 12:44:33 EST


On Wed, Jan 18, 2017 at 04:41:10PM +0100, SF Markus Elfring wrote:
> >> A local variable was set to an error code in two cases before a concrete
> >> error situation was detected. Thus move the corresponding assignment into
> >> an if branch to indicate a software failure there.
> >>
> >> This issue was detected by using the Coccinelle software.
> >
> > Why the hell is that an issue?
>
> * Can misplaced variable assignments result in unwanted run time consequences
> because of the previous approach for a control flow specification?

More like the opposite.
load constant to register
test
branch usually not taken
is considerably cheaper than
test
branch usually taken

Something like
if (unlikely(foo)) {
err = -ESOMETHING;
goto sod_off;
}
would be more or less on par (and quite possibly would be compiled into
the same code - depends upon the scheduling details for processor,
but speculative load of constant can be an optimization). However, that
has an effect of splattering the source with tons of those unlikely() *and*
visually cluttering the common path.

> * How do you think about to achieve that error codes will only be set
> after a specific software failure was detected?

Sounds like an arbitrary requirement, TBH...

Again, loading a constant into register tends to be cheap and easy to
combine with other instructions at CPU pipeline level. If anything, this
pattern is a microoptimization, often in spots that are not on hotpaths
by any stretch of imagination. But estimating whether a given place is
on a hot path takes a lot more delicate analysis than feasible for
cocci scripts. And visual cluttering of the common execution path remains -
it doesn't matter for compiler, but it can matter a lot for human readers.