On Fri, Jul 28, 2023 at 12:01:29PM +0200, Thomas Huth wrote:
> On 28/07/2023 11.50, Thomas Huth wrote:
> > On 28/07/2023 11.32, Marc-André Lureau wrote:
> > > Hi
> > >
> > > On Fri, Jul 28, 2023 at 12:59 PM Daniel P. Berrangé
> > > <berra...@redhat.com> wrote:
> > > >
> > > > On Fri, Jul 28, 2023 at 10:35:35AM +0200, Thomas Huth wrote:
> > > > > On 27/07/2023 12.39, Daniel P. Berrangé wrote:
> > > > > > On Wed, Jul 26, 2023 at 08:21:33PM +0200, Thomas Huth wrote:
> > > > > > > On 26/07/2023 18.19, Daniel P. Berrangé wrote:
> > > > > ...
> > > > > > > Anyway, before we unify the compiler package name suffix between
> > > > > > > the two
> > > > > > > jobs, I really would like to see whether the mingw
> > > > > > > Clang builds QEMU faster
> > > > > > > in the 64-bit job ... but so far I failed to convince meson to
> > > > > > > accept the
> > > > > > > Clang from the mingw package ... does anybody know how to use
> > > > > > > Clang with
> > > > > > > MSYS2 properly?
> > > > > >
> > > > > > AFAIK it shouldn't be anything worse than
> > > > > >
> > > > > > CC=clang ./configure ....
> > > > > >
> > > > > > if that doesn't work then its a bug IMHO
> > > > >
> > > > > No, it's not that easy ... As Marc-André explained to me, MSYS2
> > > > > maintains a
> > > > > completely separate environment for Clang, i.e. you have to select
> > > > > this
> > > > > different environment with $env:MSYSTEM = 'CLANG64' and then install
> > > > > the
> > > > > packages that have the "mingw-w64-clang-x86_64-" prefix.
> > > > >
> > > > > After lots of trial and error, I was able to get a test build here:
> > > > >
> > > > > https://gitlab.com/thuth/qemu/-/jobs/4758605925
> > > > >
> > > > > I had to disable Spice and use --disable-werror in that build to make
> > > > > it
> > > > > succeed, but at least it shows that Clang seems to be a little bit
> > > > > faster -
> > > > > the job finished in 58 minutes. So if we can get the warnings fixed,
> > > > > this
> > > > > might be a solution for the timeouts here...
> > > >
> > > > Those packing warnings look pretty serious
> > > >
> > > > C:/GitLab-Runner/builds/thuth/qemu/include/block/nvme.h:1781:16:
> > > > warning: unknown attribute 'gcc_struct' ignored
> > > > [-Wunknown-attributes]
> > > >
> > > > This means CLang is using the MSVC struct packing ABI for bitfields,
> > > > which is different from the GCC struct packing ABI. If any of those
> > > > structs use bitfields and are exposed as guest hardware ABI, or in
> > > > migration vmstate, then this is potentially broken compilation.
> > > >
> > >
> > > Yes .. gcc >=4.7 and clang >=12 have mms-bitfiles enabled by default,
> > > but we can't undo that MS struct packing on clang apparently:
> > > https://discourse.llvm.org/t/how-to-undo-the-effect-of-mms-bitfields/72271
> >
> > I wonder whether we really still need the gcc_struct in QEMU...
> > As far as I understand, this was mainly required for bitfields in packed
> > structs in the past
>
> Ok, never mind, according to this post:
>
> https://lists.gnu.org/archive/html/qemu-devel/2011-08/msg00964.html
>
> this affects all structs, not only the ones with bitfieds.
>
> And it seems like we also still have packed structs with bitfields in code
> base, see e.g. "struct ip" in net/util.h, so using Clang on Windows likely
> currently can't work?
Just because it has bitfields doesn't mean it will definitely be
different.
I'm not sure if it is an entirely accurate comparison, but I modified the
native linux build to use 'gcc_struct' and again to use 'ms_struct'. Then
fed all the .o files to 'pahole', and compared the output. There was only
a single difference:
union VTD_IR_TableEntry {
struct {
uint32_t present:1; /* 0: 0 4 */
uint32_t fault_disable:1; /* 0: 1 4 */
uint32_t dest_mode:1; /* 0: 2 4 */
uint32_t redir_hint:1; /* 0: 3 4 */
uint32_t trigger_mode:1; /* 0: 4 4 */
uint32_t delivery_mode:3; /* 0: 5 4 */
uint32_t __avail:4; /* 0: 8 4 */
uint32_t __reserved_0:3; /* 0:12 4 */
uint32_t irte_mode:1; /* 0:15 4 */
uint32_t vector:8; /* 0:16 4 */
uint32_t __reserved_1:8; /* 0:24 4 */
uint32_t dest_id; /* 4 4 */
uint16_t source_id; /* 8 2 */
/* Bitfield combined with previous fields */
uint64_t sid_q:2; /* 8:16 8 */
uint64_t sid_vtype:2; /* 8:18 8 */
uint64_t __reserved_2:44; /* 8:20 8 */
- } irte; /* 0 18 */
+ } irte; /* 0 16 */
uint64_t data[2]; /* 0 16 */
};
from the intel_iommu.c file.
IOW, ms_struct added a 2 byte padding after the uint16_t source_id
field despite 'packed' attribute, but gcc_struct collapsed the
uint16_t into the uint64_t bitfield since only 48 bits were consumed.
IIUC, this could be made portable by changing
uint16_t source_id; /* 8 2 */
to
uint64_t source_id:16; /* 8 2 */
NB, this was a --target-list=x86_64-softmmu build only, so hasn't
covered the hole codebase. Still shows the gcc_struct annotation
might not be as critical as we imagined.
NB a limitation of the pahole analysis is that it only reports structs that
are actually declared as variabls somewhere - either stack allocated or
heap allocate is fine, as long as there's a declarion of usage somewhre.
With regards,
Daniel
--
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org -o- https://fstop138.berrange.com :|
|: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
