https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79671
--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Bernd Edlinger from comment #12)
> Hi,
>
> I don't know if it helps, but on the assembler level there are only two
> instructions that need to be moved to make the test case pass:
That of course helps a lot, that is what I was trying to narrow through the
hacks.
Before sched1 there is:
(insn 598 595 1707 67 (set (mem/c:QI (plus:SI (reg/f:SI 102 sfp)
(const_int -19 [0xffffffffffffffed])) [98 MEM[(struct
parser_binder *)&D.616009 + 5B]+0 S1 A8])
(subreg:QI (reg:SI 505) 0)) 192 {*arm_movqi_insn}
(expr_list:REG_DEAD (reg:SI 505)
(nil)))
...
(insn 604 603 605 67 (set (reg/f:SI 692)
(plus:SI (reg/f:SI 102 sfp)
(const_int -20 [0xffffffffffffffec])))
"/usr/include/boost/function/function_template.hpp":998 4 {*arm_addsi3}
(nil))
(insn 605 604 606 67 (parallel [
(set (reg:SI 0 r0)
(mem/c:SI (reg/f:SI 692) [19 MEM[(struct function4
*)&D.616009].D.542442.functor+0 S4 A32]))
(set (reg:SI 1 r1)
(mem/c:SI (plus:SI (reg/f:SI 692)
(const_int 4 [0x4])) [19 MEM[(struct function4
*)&D.616009].D.542442.functor+4 S4 A32]))
(set (reg:SI 2 r2)
(mem/c:SI (plus:SI (reg/f:SI 692)
(const_int 8 [0x8])) [19 MEM[(struct function4
*)&D.616009].D.542442.functor+8 S4 A32]))
]) "/usr/include/boost/function/function_template.hpp":998 364 {*ldm3_}
(nil))
(current trunk, the above command line options, no patches), so ignoring strict
aliasing, there is a must alias in between the middle mem in insn 605 and the
mem insn 598 stores to.
Now if I do:
break sched_analyze_2 if x->code == MEM && insn->u2.insn_uid == 605
run
continue
then x is:
(mem/c:SI (plus:SI (reg/f:SI 692)
(const_int 4 [0x4])) [19 MEM[(struct function4
*)&D.616009].D.542442.functor+4 S4 A32])
and deps->pending_write_mems is:
(expr_list:REG_DEP_TRUE (mem/f/c:SI (plus:SI (reg/f:SI 102 sfp)
(const_int -264 [0xfffffffffffffef8])) [18 tmp.D.542442.vtable+0 S4
A64])
(expr_list:REG_DEP_TRUE (mem/c:QI (plus:SI (reg/f:SI 102 sfp)
(const_int -19 [0xffffffffffffffed])) [98 MEM[(struct
parser_binder *)&D.616009 + 5B]+0 S1 A8])
(nil)))
This then calls true_dependence with mem:
(mem/c:QI (plus:SI (reg/f:SI 102 sfp)
(const_int -19 [0xffffffffffffffed])) [98 MEM[(struct parser_binder
*)&D.616009 + 5B]+0 S1 A8])
x:
(mem/c:SI (plus:SI (reg/f:SI 692)
(const_int 4 [0x4])) [19 MEM[(struct function4
*)&D.616009].D.542442.functor+4 S4 A32])
and we return 0 because of:
2931 if (mems_in_disjoint_alias_sets_p (x, mem))
2932 return 0;
In *.optimized dump I believe this is:
MEM[(struct parser_binder *)&D.616009 + 5B] = 93;
value_283 = (size_t) &stored_vtable.base;
value_284 = value_283 | 1;
value.75_285 = (struct vtable_base *) value_284;
MEM[(struct function4 *)&D.616009].D.542442.vtable = value.75_285;
MEM[(struct &)&tmp] ={v} {CLOBBER};
tmp.D.542442.vtable = value.75_285;
tmp.D.542442.functor = MEM[(struct function4 *)&D.616009].D.542442.functor;
insn 598 is that MEM[(struct parser_binder *)&D.616009 + 5B] = 93; and insn 605
(ldm3_) is the load for the structure assignment
tmp.D.542442.functor = MEM[(struct function4 *)&D.616009].D.542442.functor;
Richard, could you please have a look from the alias oracle POV?
