https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97701
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |ktkachov at gcc dot gnu.org,
| |rsandifo at gcc dot gnu.org,
| |vmakarov at gcc dot gnu.org
--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I can still reproduce with current trunk and it seems it was latent before and
is a ra or backend bug.
The problematic insn is created during LRA, and first is:
(insn 2616 2615 0 (set (reg:DI 1467)
(lo_sum:DI (reg:DI 1468)
(symbol_ref/u:DI ("*.LC0") [flags 0x82]))) -1
(nil))
which is fine for:
(set (match_operand:DI 0 ("register_operand") ("=r"))
(lo_sum:DI (match_operand:DI 1 ("register_operand") ("r"))
(match_operand 2 ("aarch64_valid_symref") ("S"))))
insn, but later on LRA changes the destination register to:
(insn 2616 2615 2617 5 (set (reg:DI 32 v0 [1467])
(lo_sum:DI (reg:DI 1468)
(symbol_ref/u:DI ("*.LC0") [flags 0x82]))) 1007 {add_losym_di}
(expr_list:REG_DEAD (reg:DI 1468)
(nil)))
and that doesn't match anymore, as v0 register is not GENERAL_REGS.
