https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93946
--- Comment #11 from sandra at gcc dot gnu.org ---
RTL before sched2 does look sane and similar to that generated for x86 with
-m32.
I've been trying to step through sched2. I think that where things get
interesting is the call to true_dependence at sched-deps.c:2663.
Breakpoint 6, true_dependence (mem=0x7ffff742c9a8, mem_mode=E_VOIDmode,
x=0x7ffff742cac8)
at /scratch/sandra/nios2-elf-fsf/src/gcc-mainline/gcc/alias.c:3056
3056 return true_dependence_1 (mem, mem_mode, NULL_RTX,
(gdb) print debug_rtx(mem)
(mem/j:SI (reg/v/f:SI 5 r5 [orig:48 ptr ] [48]) [1 MEM[(struct aa
*)ptr_1(D)].a.u.i+0 S4 A32])
$19 = void
(gdb) print debug_rtx(x)
(mem/j:SI (reg/v/f:SI 4 r4 [orig:47 bv ] [47]) [1 bv_3(D)->b.u.f+0 S4 A32])
$20 = void
This is making it all the way to the end of true_dependence_1, into
rtx_refs_may_alias_p, and into refs_may_alias_p_1, which is returning false,
which gets propagated back up as the result of true_dependence. IIUC, this is
what is allowing sched2 to move the read from "x" ahead of the write to "mem".
Before I spend more time on this, am I on the right track here? And is this
pointing at the problem being in refs_may_alias_p_1 rather than somewhere along
the way e.g. in true_dependence_1?
