[PATCH] D72231: [Sema] Adds the pointer-to-int-cast diagnostic

[Nathan Chancellor via Phabricator via cfe-commits]

nathanchance added a comment.

In D72231#1881855 <https://reviews.llvm.org/D72231#1881855>, @rjmccall wrote:

> In D72231#1881797 <https://reviews.llvm.org/D72231#1881797>, @nickdesaulniers wrote:
>
> > In D72231#1881784 <https://reviews.llvm.org/D72231#1881784>, @rjmccall wrote:
> >
> > > In D72231#1881760 <https://reviews.llvm.org/D72231#1881760>, @nickdesaulniers wrote:
> > >
> > > > In D72231#1879347 <https://reviews.llvm.org/D72231#1879347>, @rjmccall wrote:
> > > >
> > > > > In D72231#1878528 <https://reviews.llvm.org/D72231#1878528>, @nathanchance wrote:
> > > > >
> > > > > > There appear to a be semantic difference between GCC and clang with the current version of this patch which results in a lot of additional warnings in the Linux kernel: https://godbolt.org/z/eHFJd8
> > > > >
> > > > >
> > > > > Warning about casting to an enum seems clearly correct and in scope for this warning.  Warning about casting to `_Bool` seems clearly incorrect and should not be warned about at all.
> > > >
> > > >
> > > > Maybe we should only warn if the size of the `void*` is smaller than the size of the `enum`? (32b `void*`, 64b `enum`)? https://godbolt.org/z/oAts-u
> > > >
> > > > Otherwise this warning creates a massive mess for us to clean up, and I suspect Linux kernel developers will just end up disabling the warning.
> > >
> > >
> > > If deployment is easier if we split out a subgroup that we can turn off, that seems fine.  But I don't see any good abstract justification for warning about a cast to `int` and not a cast to an `int`-sized `enum`.  What would the reasoning be, just that the latter "couldn't possibly" be intended to preserve the original pointer value, so it must be an opaque value being represented as a `void*`?  That seems pretty weak to me.
> >
> >
> > Less about enums, more about casts to/from void*, since you might use that in place of a union that would be too large to describe.  Specifically, this `struct` is used throughout the kernel for most drivers: https://elixir.bootlin.com/linux/v5.5.4/source/include/linux/mod_devicetable.h#L260  It is exceedingly common to stuff whatever data in there: https://elixir.bootlin.com/linux/v5.5.4/source/drivers/ata/ahci_brcm.c#L428 so long as the driver is careful not to reinterpret the data as the incorrect type.  Describing such a union for ever possible enum packed in there would not be fun.
>
>
> No, I understand the pattern, but they must have already done some sort of pass over the code to make it warning-clean when they're working with a smaller integer type.  Or do they just in practice never store smaller integers in there, whereas it's hard to control size with an enum?


Yes, if the data is a regular `int`, rather than an `enum`, all of the callsites either cast to `long` or `uintptr_t` (which is typedef'd in the kernel to `unsigned long`). There are a lot fewer of those spots in the kernel (at least from my super quick `rg` search), most of the spots use an `enum`, maybe to purposefully avoid this warning? Most, if not all the sites, only store a number that is less than 5 because they use that number to determine exactly which device is present from the match data so the driver can handle different quirks with things like case statements.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D72231/new/

https://reviews.llvm.org/D72231