On Tue, Jun 27, 2023 at 12:07 PM Krister Walfridsson via Gcc
<gcc@gcc.gnu.org> wrote:
>
> I'm working on an updated version of my translation validator [*], and I
> have some problems with CLOBBER(eol).
>
> I currently treat CLOBBER(eol) as making the memory invalid (i.e. all
> subsequent accesses is undefined behavior), but the tool reports
> miscompilation for a few cases where the tree-nrv pass makes the IR
> clobber <retval> by changing code such as [**]
>
> union bu o;
> ...
> o = i;
> MEM[(union *)&o].b18 = _11;
> MEM[(union *)&o].b20 = _11;
> <retval> = o;
> o ={v} {CLOBBER(eol)};
> return <retval>;
>
> to use <retval> instead of o
>
> union bu o [value-expr: <retval>];
> ...
> <retval> = i;
> MEM[(union *)&<retval>].b18 = _11;
> MEM[(union *)&<retval>].b20 = _11;
> <retval> ={v} {CLOBBER(eol)};
> return <retval>;
>
> As a result, the tool therefore thinks the returned value is unavailable.
>
> Is my understanding of CLOBBER(eol) correct? (I.e., is this a bug in
> tree-nrv?) Or is <retval> special so that I should just ignore
> CLOBBER(eol) for this case?
I think this is a bug in NRV, yes, <retval> is special but the above would
allow to DSE the three stores.
Can you open a bugreport?
Thanks,
Richard.
>
> /Krister
>
>
> [*] I'm planning to release the updated version in a few weeks. Meanwhile,
> you can find the old version at https://github.com/kristerw/pysmtgcc
>
> [**] This example originates from gcc.c-torture/execute/921204-1.c
> compiled for x86 using the flags "-O -m32".
