Doug S (foo.delete@this.bar.bar) on June 15, 2022 1:55 pm wrote:
> none (none.delete@this.none.com) on June 15, 2022 9:57 am wrote:
> > dmcq (dmcq.delete@this.fano.co.uk) on June 15, 2022 9:38 am wrote:
> > [...]
> > > The ARM Branch consistent seems the wrong way around to
> > > me. I haven't done this for a while but when I used to
> > > use the rarely pragma it ensured the rarely used code, e.g.
> > > for handling errors, was put out of line at the end
> > > of the code so the non-rarely code wasn't polluted by stuff
> > > that hardly ever ran - good for the cache. This would
> > > need a marker saying the branch hardly ever branched rather than that it prctically always branched.
> >
> > I might have missed something but your branch will consistently fail so what Arm does is
> > fine. By consistent they mean that it is unlikely to change direction so that's OK for your
> > example.
>
>
> Assuming most people don't do PGO, outside of loops (where it can tell it will run for enough iterations)
> are pragmas going to be the only way for a compiler to know where to generate BC instructions?
>
> Theoretically even if you don't explicitly use PGO, if the CPU had an interface to pass branch history
> information to the OS, the OS could use that to modify binaries. But seems problematic in today's world,
> if you have signature checking on binaries, they are installed on read only partitions, etc.

I think you'd be pretty safe assuming that any code that calls a function that doesn't return is rarely executed. A lot of code is just checking for errors.