[Bug target/102239] powerpc suboptimal boolean test of contiguous bits

["luoxhu at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102239

luoxhu at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |luoxhu at gcc dot gnu.org

--- Comment #2 from luoxhu at gcc dot gnu.org ---
(In reply to Segher Boessenkool from comment #1)
> Confirmed.
> 
> So the relevant insn
> 
> (parallel [(set (reg:CC 123)
>                 (compare:CC (and:DI (reg:DI 124)
>                                     (const_int 25769803776 [0x600000000]))
>                             (const_int 0 [0])))
>            (clobber (scratch:DI))])
> 
> is matched by *and<mode>3_2insn but not by any pattern that ends up as just
> one insn.  Not *and<mode>3_mask_dot, because that doesn't do a shift first,
> is just an AND and there are no machine insns to do that; but there is no
> pattern for what we can do.
> 
> *rotl<mode>3_mask_dot cannot do this either; the base and the dot2 of that
> cannot be done, they return a shifted result, but that doesn't matter for
> comparing it to 0.  So we should add a specialised version.

Seems different with what you describe, in combine, it was combined to
anddi3_2insn_dot:

(insn 9 8 10 2 (parallel [
            (set (reg:CC 122)
                (compare:CC (and:DI (reg:DI 123)
                        (const_int 25769803776 [0x600000000]))
                    (const_int 0 [0])))
            (clobber (scratch:DI))
        ]) "pr102239.c":3:6 210 {*anddi3_2insn_dot}
     (expr_list:REG_DEAD (reg:DI 123)
        (nil)))
(jump_insn 10 9 11 2 (set (pc)
        (if_then_else (eq (reg:CC 122)
                (const_int 0 [0]))
            (label_ref 15)
            (pc))) "pr102239.c":3:6 868 {*cbranch}
     (expr_list:REG_DEAD (reg:CC 122)
        (int_list:REG_BR_PROB 536870916 (nil)))



Then in pr102239.c.302r.split2, it is split by "*and<mode>3_2insn_dot" to
rotldi3_mask+lshrdi3_dot:

Splitting with gen_split_80 (rs6000.md:3721)

(insn 32 8 33 2 (set (reg:DI 3 3 [124])
        (and:DI (ashift:DI (reg:DI 3 3 [123])
                (const_int 29 [0x1d]))
            (const_int -4611686018427387904 [0xc000000000000000])))
"pr102239.c":3:6 236 {*rotldi3_mask}
     (nil))
(insn 33 32 10 2 (parallel [
            (set (reg:CC 100 0 [122])
                (compare:CC (lshiftrt:DI (reg:DI 3 3 [124])
                        (const_int 29 [0x1d]))
                    (const_int 0 [0])))
            (clobber (reg:DI 3 3 [124]))
        ]) "pr102239.c":3:6 281 {*lshrdi3_dot}
     (nil))


Why this difference happens?

0x600000000 is not a valid mask for anddi3_2insn_dot:


 "(<MODE>mode == Pmode || UINTVAL (operands[2]) <= 0x7fffffff)
   && rs6000_is_valid_2insn_and (operands[2], <MODE>mode)
   && !(rs6000_is_valid_and_mask (operands[2], <MODE>mode)
        || logical_const_operand (operands[2], <MODE>mode))"


(gdb) p UINTVAL (operands[2]) <= 0x7fffffff
$84 = false
(gdb) p rs6000_is_valid_2insn_and (operands[2], E_DImode)
$85 = true
(gdb) p logical_const_operand (operands[2], E_DImode)
$86 = false
(gdb) p rs6000_is_valid_and_mask (operands[2], E_DImode)
$87 = false
(gdb) p Pmode
$88 = DImode