Warnings (was: Subsurface libdivecomputer branch merged with upstream..)

[Berthold Stoeger]

On Sonntag, 17. Dezember 2017 22:00:43 CET Linus Torvalds wrote:
> On Sun, Dec 17, 2017 at 4:38 AM, Berthold Stoeger
> 
> <bstoeger at mail.tuwien.ac.at> wrote:
> > I tried to look at the assembler output and the first thing I noted is
> > that we don't compile with -O2? Is that on purpose?
> 
> I do think we should use -O2. In fact, we *do* use it for release builds:
> 
>   CMAKE_CXX_FLAGS_RELEASE:STRING=-O2 -DNDEBUG
>   CMAKE_C_FLAGS_RELEASE:STRING=-O2 -DNDEBUG
> 
> but the DEBUG builds are done with just "-g". That's a mistake, in my
> opinion.

Ok, that explains it. When I find time, I will try to find out how to change 
that.

> Sure, -O2 *can* make debugging a bit harder in that it can remove some
> trivial variables, but honestly, that's very rare, and it can go the
> other way too.
> 
> For example, for the kernel, unoptimized assembly is *so* stupid that
> it's actually hard to debug, and we've never even supported it. We
> don't have those kinds of low-level issues with subsurface, though.

Indeed, the assembler output for a simple if-statement was absolute SIMD 
gibberish. This must be on purpose to help the optimizer, but it makes -O0 
useless even for debug mode.

> > Anyway, with -O3, the first case in plannernotes.h produces better code
> > without casting to float. You could of course get the same result by
> > replacing 1000.0 with 1000.0f, but this seems rather obscure.
> 
> I would suggest strongly against -O3.
> 
> It can improve things. It can also make things _much_ worse, and it's
> generally not nearly as well tested as -O2. It tends to be where gcc
> starts doing "unproven" optimizations.

OT: Some compilers do funky things even with -O2. Infamous example:
https://kristerw.blogspot.co.at/2017/09/why-undefined-behavior-may-call-never.html

> Like "short", the only time "float" makes lots of sense is in
> datastructures where size really really matters.

Thanks, noted.

Berthold