On 9/6/19 8:44 AM, Eric Blake wrote:
> On 9/6/19 8:24 AM, Philippe Mathieu-Daudé wrote:
>
>>>>> static const MemoryRegionOps notdirty_mem_ops = {
>>>>> .write = notdirty_mem_write,
>>>>> - .valid.accepts = notdirty_mem_accepts,
>>>>> .endianness = DEVICE_NATIVE_ENDIAN,
>>>>> .valid = {
>>>>> .min_access_size = 1,
>>>>> .max_access_size = 8,
>>>>> .unaligned = false,
>>>>> + .accepts = notdirty_mem_accepts,
>>>>
>>>> I'm surprised the compiler doesn't emit any warning...
>>>
>>> Same here.
Actually, I just played with -Woverride-init in gcc 9.2.1 (and clang's
comparable -Winitializer-overrides, which we intentionally disable
during configure), and they come pretty close - both compilers DO flag
when an implicit zero-initialization due to partial ={} overrides an
earlier initialization. But sadly, they also warn when one specific
init of a smaller subobject overrides another earlier specific init of a
larger subobject such as an array range operator. So
qobject/json-lexer.c and others fail to compile under the existing
warning option, which is why we disable it during configure (clang has
it as part of -Wall; gcc only has it as part of -Wextra which we do not
use).
In researching further, I see:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=24010#c4
which explains why -Woverride-init is NOT part of gcc's -Wall, precisely
because of our range pre-initialization usage.
So I filed:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91688
seeing if the gcc devs would consider splitting into
-Woverride-init=[12], where 1 only flags a larger subobject overriding
an earlier smaller one (would have caught our bug) and 2 flags an
equal-size or smaller subobject overriding an earlier large one (which
we would not use, because we rely on that for range pre-initialization).
--
Eric Blake, Principal Software Engineer
Red Hat, Inc. +1-919-301-3226
Virtualization: qemu.org | libvirt.org
