Re: [PATCH] hw/block: nvme: Fix a build error in nvme_process_sq()

[Peter Maydell <peter.maydell@linaro.org>]

On Wed, 10 Feb 2021 at 11:37, Klaus Jensen <its@irrelevant.dk> wrote:
>
> On Feb 10 11:01, Peter Maydell wrote:
> > On Wed, 10 Feb 2021 at 10:31, Klaus Jensen <its@irrelevant.dk> wrote:
> > > On Feb 10 18:24, Bin Meng wrote:
> > > > I am using the default GCC 5.4 on a Ubuntu 16.04 host.
> > > >
> > >
> > > Alright. I'm actually not sure why newer compilers does not report this.
> > > The warning looks reasonable.
> >
> > It's not actually ever possible for nvme_ns() to return
> > NULL in this loop, because nvme_ns() will only return NULL
> > if it is passed an nsid value that is 0 or > n->num_namespaces,
>
> NvmeCtrl.namespaces is an array of pointers and some of those will most
> likely be NULL (those are unallocated namespaces).

Whoops, yes.

> I know that "it's never the compiler". But in this case, wtf? If there
> are no allocated namespaces, then we will actually never hit the
> statement that initializes result. I just confirmed this with a
> configuration without any namespaces.
>
> The patch is good. I wonder why newer GCCs does NOT detect this. Trying
> to use `result` as the first statement in the loop also does not cause a
> warning. Only using the variable just before the loop triggers a
> warning on this.

My new hypothesis is that maybe newer GCCs are more cautious
about when they produce the 'may be used uninitialized' warning,
to avoid having too many false positives.

-- PMM