https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128845
Bug ID: 128845
Summary: postreload deletes a volatile store whose source is a
plain load of the same location (mem-to-mem move) at
-O1 / -fno-strict-aliasing
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: lzy20000419 at outlook dot com
Target Milestone: ---
Fails with an m68k-linux-gnu cross trunk build (gcc 17.0.0 20261118,
1190f702416e) and with 13.1.0 through 16.1.0.
Target: m68k-linux-gnu; msp430-elf and vax-netbsdelf delete the store too.
void
touch (int *p)
{
*(volatile int *) p = *p;
}
m68k-linux-gnu-gcc -O1 -S t.c
Observed:
touch:
rts
Expected (what -O2 gives on 16.1 and trunk, and what -O1
-fdisable-rtl-postreload gives):
touch:
move.l 4(%sp),%a0
move.l (%a0),(%a0)
rts
On trunk and 16.1 the store is deleted at -O1 and at -O2 -fno-strict-aliasing.
On 15 and older it is also deleted at -O2 and -Os. At run time, calling touch
on a read-only page does not fault with the -O1 build, while the -O0 and -O2
builds fault as they should (checked under qemu-m68k with the trunk build).
Godbolt (16.1.0): https://godbolt.org/z/ePsqKnf69
Cause: cselib_redundant_set_p (cselib.cc), as changed by r16-6229 for PR122343:
rtx src = SET_SRC (set);
if ((MEM_P (src) && MEM_VOLATILE_P (src))
|| !rtx_equal_for_cselib_p (dest, src))
return false;
...strip SUBREG/ZERO_EXTRACT/STRICT_LOW_PART...
if (!flag_strict_aliasing || !MEM_P (dest))
return true;
if (MEM_VOLATILE_P (dest))
return false;
The insn after reload is (set (mem/v:SI (reg %a0)) (mem:SI (reg %a0))).
rtx_equal_for_cselib_p (dest, src) is true, and with -fno-strict-aliasing the
function returns true before it checks MEM_VOLATILE_P (dest), so
reload_cse_simplify deletes the store. Before r16-6229 there was no volatile
check at all.
Possible fix: check MEM_P (dest) && MEM_VOLATILE_P (dest) before the
!flag_strict_aliasing early return. With that change on trunk, touch keeps
move.l (%a0),(%a0) at -O1 and -O2 -fno-strict-aliasing, and the read-only page
test faults as expected.
Related: PR122343 (r16-6229).
Disclaimer: This test case was found with LLM-assisted targeted testing, and
the report was written with LLM assistance.