https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70018
Jan Hubicka <hubicka at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |hubicka at gcc dot
gnu.org
--- Comment #11 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
For
int *a;
bool t()
{
return *a==*a;
}
we get
$ more t.C.*original
;; Function bool t() (null)
;; enabled by -tree-original
return <retval> = 1;
even at -O0 while I think very stupid compiler is allowed to keep the memory
access.
In fact clang does precisely that:
_Z1tv: # @_Z1tv
.cfi_startproc
# BB#0:
pushq %rbp
.Ltmp0:
.cfi_def_cfa_offset 16
.Ltmp1:
.cfi_offset %rbp, -16
movq %rsp, %rbp
.Ltmp2:
.cfi_def_cfa_register %rbp
movq a, %rax
movl (%rax), %ecx
movq a, %rax
cmpl (%rax), %ecx
sete %dl
andb $1, %dl
movzbl %dl, %eax
popq %rbp
retq
.Ltmp3:
I am trying to make conservatively correct fix for this. I suppose we can still
discover functions seen as const/pure as pure and incrementally teach optimizer
that given function does possibly use global memory state but it doesn't depend
on it. It should be still correct to CSE those functions.
For that we finally need to cleanup pure/const/looping flags into sane form
(i.e. clobber memory/has side effects/is CSEable and uses no memory/is CSEable
and uses memory)
