[Bug tree-optimization/106458] [12/13 Regression] glibc's malloc/tst-scratch_buffer.c test is miscompiled with gcc-12

[dave.anglin at bell dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106458

--- Comment #12 from dave.anglin at bell dot net ---
On 2022-08-10 9:30 a.m., rguenth at gcc dot gnu.org wrote:
> When I try with a cc1 cross I see
>
>> ./cc1 -quiet t.i -fpreprocessed -O2 -g -std=gnu11 -fgnu89-inline -fmerge-all-constants -frounding-math -fno-stack-protector -fno-common -fmath-errno    -fno-pie -fopt-info-vec
> and nothing actually vectorized.  Besides the PRE there seems to be a missed
> DSE at the end (when vectorization is enabled as after the change):
>
>     # DEBUG BEGIN_STMT
>     memset (&buf.__space.__c, 64, 1024);
>     # DEBUG BEGIN_STMT
> +  sizes[0] = 16;
>     sizes[1] = 1024;
>     sizes[2] = 1040;
>
> otherwise I see nothing suspicious in differences in .optimized when
> comparing -fdump-tree-optimized with/without -fno-tree-vectorize.
I had tried -fno-tree-vectorize and it made no difference to the test.  So, it
seems GCC is miscompiled
with the -ftree-vectorize change.  I thought of building GCC with
-fno-tree-vectorize as a test.

>
> You could try if -fno-tree-pre reproduces it also before the change.
>
Will try.