https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116199
Jeffrey A. Law <law at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2024-08-02
Ever confirmed|0 |1
--- Comment #3 from Jeffrey A. Law <law at gcc dot gnu.org> ---
So the code coming out of ext-dce looks quite sensible. The key insns:
(insn 7 4 8 2 (set (reg:DI 31 [ _1 ])
(subreg:DI (reg/v:SI 37 [ __major ]) 0)) "j.c":7:32 75 {*m68k.md:1568}
(nil))
(insn 8 7 10 2 (set (reg:DI 39 [ _2 ])
(ashift:DI (reg:DI 31 [ _1 ])
(const_int 8 [0x8]))) "j.c":7:48 322 {*ashldi3}
(expr_list:REG_DEAD (reg:DI 31 [ _1 ])
(nil)))
Insn 7 had previously been a extension, but the upper bits aren't used and we
could safely turn it into a paradoxical subreg.
Combine's actions are sensible, combining insns 7 and 8 into:
(insn 8 7 10 2 (set (reg:DI 39 [ _2 ])
(ashift:DI (subreg:DI (reg/v:SI 37 [ __major ]) 0)
(const_int 8 [0x8]))) "j.c":7:48 322 {*ashldi3}
(nil))
Still quite sensible. IRA allocates (reg:SI 37) to d0. Still quite sensible.
We ultimately have this inside reload:
(insn 8 7 10 2 (set (reg:DI 8 %a0 [orig:39 _2 ] [39])
(ashift:DI (subreg:DI (reg/v:SI 0 %d0 [orig:37 __major ] [37]) 0)
(const_int 8 [0x8]))) "j.c":7:48 322 {*ashldi3}
(nil))
Then reload blows up trying to ensure that op0/op1 match. Looks like a reload
issue to me.
