[Bug rtl-optimization/83565] RTL combine pass breaks shift result (at least on ia64)

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83565

--- Comment #7 from Segher Boessenkool <segher at gcc dot gnu.org> ---
(In reply to Sergei Trofimovich from comment #5)
> gcc's try_combine() scares me lots :)

It's an acquired taste.

> Segher, can you guide me here what changes size of zero_extract in here from
> 0x1f to 0x20?

Trying 23 -> 24:
   23: r359:DI=zero_extract(r358:DI,0x1f,0x1)
      REG_DEAD r358:DI
   24: r356:SI#0=r359:DI
      REG_DEAD r359:DI
Successfully matched this instruction:
(set (subreg:DI (reg:SI 356 [ off ]) 0)
    (lshiftrt:DI (reg:DI 358)
        (const_int 1 [0x1])))
allowing combination of insns 23 and 24
original costs 4 + 4 = 8
replacement cost 4
deferring deletion of insn with uid = 23.
modifying insn i3    24: r356:SI#0=r358:DI 0>>0x1
      REG_DEAD r358:DI
deferring rescan insn with uid = 24.

[ ... ]

Trying 24 -> 29:
   24: r356:SI#0=r358:DI 0>>0x1
      REG_DEAD r358:DI
   29: r8:SI=r356:SI
      REG_DEAD r356:SI
Successfully matched this instruction:
(set (reg:DI 8 r8)
    (zero_extract:DI (reg:DI 358)
        (const_int 32 [0x20])
        (const_int 1 [0x1])))
allowing combination of insns 24 and 29
original costs 4 + 4 = 8
replacement cost 4
deferring deletion of insn with uid = 24.
modifying insn i3    29: r8:DI=zero_extract(r358:DI,0x20,0x1)
      REG_DEAD r358:DI
deferring rescan insn with uid = 29.



r358 is known to have the high half zero:
   22: r358:DI=r357:SI#0^r341:SI#0

so this is correct.


The zero_extract(31,1) is transformed to a shift right, and that then
later is transformed to a zero_extract(32,1); there is no transform from
31 to 32 :-)

James' patch looks fine fwiw; it's a backend problem.