https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106594
--- Comment #15 from Roger Sayle <roger at nextmovesoftware dot com> ---
An example of combine's temporary lapses of sanity can be seen on powerpc:
Trying 14 -> 15:
14: %3:SI=sign_extend(r128:SI#2)*sign_extend(r127:SI#2)
REG_DEAD r128:SI
REG_DEAD r127:SI
15: use %3:SI
Failed to match this instruction:
(parallel [
(use (mult:SI (ashiftrt:SI (ashift:SI (reg:SI 128)
(const_int 16 [0x10]))
(const_int 16 [0x10]))
(ashiftrt:SI (ashift:SI (reg:SI 127)
(const_int 16 [0x10]))
(const_int 16 [0x10]))))
(set (reg/i:SI 3 3)
(mult:SI (sign_extend:SI (subreg:HI (reg:SI 128) 2))
(sign_extend:SI (subreg:HI (reg:SI 127) 2))))
])
with the proposed patch, we now see:
Trying 14 -> 15:
14: %3:SI=sign_extend(r128:SI#2)*sign_extend(r127:SI#2)
REG_DEAD r128:SI
REG_DEAD r127:SI
15: use %3:SI
Failed to match this instruction:
(parallel [
(use (mult:SI (sign_extend:SI (subreg:HI (reg:SI 128) 2))
(sign_extend:SI (subreg:HI (reg:SI 127) 2))))
(set (reg/i:SI 3 3)
(mult:SI (sign_extend:SI (subreg:HI (reg:SI 128) 2))
(sign_extend:SI (subreg:HI (reg:SI 127) 2))))
])
i.e. the "canonical" form of the expression actually matches the RTL pattern
used/expected by the rs6000 backend.