[Bug middle-end/99578] [11/12 Regression] gcc-11 -Warray-bounds or -Wstringop-overread warning when accessing a pointer from integer literal

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99578

--- Comment #35 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 52643
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=52643&action=edit
gcc12-pr99578-wip.patch

What we could do is differentiate between the invalid constant addresses
from pointer arithmetics on NULL or propagation of NULL into &ptr->field
and valid (type *)constant used heavily in real-world code and accepted by all
major compilers by representing the former internally as
&MEM_REF[(void *)0B + offset] while the user (type *)constant as INTEGER_CSTs.
That in full seems like a stage1 task because we'd need to ensure we don't
regress much by doing that.  Some things would be unavoidable (like avoiding
SCCVN of those pointer INTEGER_CSTs vs. the &MEM_REF[(void *)0B + offset]
form),
e.g. equality comparison of &MEM_REF[(void *)0B + CST] vs. (void *)CST should
be
optimized etc., and we would need to force that form even on say (int *)0 + 24
etc. in the FEs.
But we badly need to fix this for GCC 12 and backport to GCC 11.
So this WIP patch treats addresses in the first page (a param defaulting to
4KB)
temporarily as emitting the warnings as before in GCC 11/12, and only creates
the &MEM_REF[(void *)0B + offset] form for larger offsets which will make them
very rare.
Some further tweaks will be needed on the gimple-array-bounds* etc. side so
that
we treat those &MEM_REF[(void *)0B + offset] as equivalent to the GCC 11 up to
current trunk handling of (void *)offset, so that say on:
struct S { int a, b[4]; };
struct T { int a, b[8192], c[4]; };

void
foo (struct S *p)
{
  if (p) return;
  __builtin_memset (p->b, 0, sizeof p->b);
}

void
bar (struct T *p)
{
  if (p) return;
  __builtin_memset (p->c, 0, sizeof p->c);
}
one gets the same warnings rather than different.
On IRC Richi suggested just the params.opt and pointer-query.cc
(compute_objsize_r) hunks which would only warn above in foo and not in bar.