Alex Turjan wrote:
> Dear all,
> Im writing to you regarding the dead store elimination (dse) which runs after
> register allocation. Apparently dse removes wrongly the following store
> (present in bb2):
>
> (insn 374 47 52 2 test.c:107 (set (mem/c:SI (plus:PSI (reg/f:PSI 55 ptr15)
> (const_int 96 [0x60])) [19 fac_iter+0 S4 A32])
> (reg/v:SI 16 r16 [orig:161 step109 ] [161])) 48
> {si_indexed_store_incl_ra} (nil))
>
> despite being consumed (in bb3) by the following 2 loads:
> (insn 380 71 64 3 test.c:112 (set (reg:HI 1 r1)
> (mem:HI (plus:PSI (reg/f:PSI 55 ptr15)
> (const_int 96 [0x60])) [0 S2 A16])) 12 {load} (nil))
>
> (insn 382 346 65 3 test.c:112 (set (reg:HI 5 r5)
> (mem:HI (plus:PSI (reg/f:PSI 55 ptr15)
> (const_int 98 [0x62])) [0 S2 A16])) 12 {load} (nil))
>
>
> Can anyone point what may be the problem?
>
> As you can see the store is SI while the loads are HI. While looking to the
> comments from dse.c I get to the following remark:
>
> " There are three cases where dse falls short:
> a) Reload sometimes creates the slot for one mode of access, and
> then inserts loads and/or stores for a smaller mode. "
>
> Does it mean that such cases are not treated properly by dse?
>
> I observed that if I run with the flag -fno-strict-aliasing the wrongly
> removed store is no longer removed and the code is runs correctly.
> Im wondering does the dse after register allocation make use of type based
> alias analysis?
Here's part of the comment in alias.c:
/* The alias sets assigned to MEMs assist the back-end in determining
which MEMs can alias which other MEMs. In general, two MEMs in
different alias sets cannot alias each other ...
There's a lot more information in the comments there.
Andrew.
