[Bug tree-optimization/86532] [9 Regression] Wrong code due to a wrong strlen folding starting with r262522

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86532

--- Comment #27 from Martin Sebor <msebor at gcc dot gnu.org> ---
I don't think it would be appropriate to introduce dependencies on the
sanitizer for the same reason we can't do that for warnings.  But as I
mentioned in comment 16, I think performing these sorts of simplifications this
early (this one it happens during parsing) prevents subsequent analysis from
making use of the information that can be derived from in the original code. 
Deferring this non-constant folding until later (see bug 86434) would make it
possible not only to detect some of these bugs but also transform them into
traps/unreachable without any runtime overhead (when we know the non-constant
offset's range is out-of-bounds).

I did the comparison in signed because that's what the function returns but I
do agree that folding to zero rather than arbitrary value.  I opened bug 86572
for this idea.  Let me propose making the change separately (here I just want
to fix the wrong code without preventing  existing optimizations).  I would
also be open to emitting __builtin_unreachable().

The chartype loop needs to change to avoid assuming the element type is
necessarily an integer type.  I can reproduce the problem on my local machine
but I don't know why the full regression test run that I run on my team's
server didn't expose this or the previous ICE.  I'll look into it.

I very much appreciate your testing and feedback.  I would just ask you to
comment on the patch on the list so we don't clutter with details that are not
important to the record of the bug.