Re: [PATCH 7/9] infiniband: uverbs: avoid out-of-range warnings

[Arnd Bergmann]

On Wed, Apr 3, 2024, at 17:45, Jason Gunthorpe wrote:
> On Thu, Mar 28, 2024 at 03:30:45PM +0100, Arnd Bergmann wrote:
>> From: Arnd Bergmann <arnd@xxxxxxxx>
>> 
>> clang warns for comparisons that are always true, which is the case
>> for these two page size checks on architectures with 64KB pages:
>> 
>> drivers/infiniband/core/uverbs_ioctl.c:90:39: error: result of comparison of constant 65536 with expression of type 'u16' (aka 'unsigned short') is always false [-Werror,-Wtautological-constant-out-of-range-compare]
>>         WARN_ON_ONCE(method_elm->bundle_size > PAGE_SIZE);
>>         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~
>> include/asm-generic/bug.h:104:25: note: expanded from macro 'WARN_ON_ONCE'
>>         int __ret_warn_on = !!(condition);                      \
>>                                ^~~~~~~~~
>> drivers/infiniband/core/uverbs_ioctl.c:621:17: error: result of comparison of constant 65536 with expression of type '__u16' (aka 'unsigned short') is always false [-Werror,-Wtautological-constant-out-of-range-compare]
>>         if (hdr.length > PAGE_SIZE ||
>>             ~~~~~~~~~~ ^ ~~~~~~~~~
>> 
>> Add a cast to u32 in both cases, so it never warns about this.
>
> But doesn't that hurt the codegen?

I just double-checked in the compiler explorer to confirm that
this works as I expected: both gcc and clang are still able
to optimize out the comparison for 64K pages, but clang no
longer complains after my change that this is an obvious
case.

I also see that gcc still produces a -Wtype-limits warning, but that
likely has to stay disabled because it produces too much output
elsewhere and I don't see an easy way to shut it up.

     Arnd