https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88236
Senthil Kumar Selvaraj <saaadhu at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |saaadhu at gcc dot
gnu.org
--- Comment #2 from Senthil Kumar Selvaraj <saaadhu at gcc dot gnu.org> ---
Created attachment 45119
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45119&action=edit
Bug fix patch
If reload choses r30 (or r31) as operand 0 of xload<mode>_8, the lpmx
constraint
alternative ends up generating wrong code, as avrt_out_xload would emit
lpm r30, Z
sbrc <operand1>, 7
ld r30, Z
r30 gets clobbered by LPM, and LD with Z will therefore not work correctly.
This
happens even though operand 0 is marked as early clobber in the insn,
presumably
because of the hard coded (reg:HI REG_Z).
If (reg:HI REG_Z) is replaced with
(match_operand:HI 2 "register_operand" "z,z")
then reload does respect the early clobber, but fails with "unable to find
register to spill in pointer class POINTER_REGS_Z", probably because the
xload8<mode>_A hardcodes REG_Z as the third operand.
Adding a clobber for REG_Z prevents reload from assigning r30 as the output
operand, even though REG_Z isn't technically clobbered. One other way to fix
this would be to add a new register constraint that tells reload to use any
register other than REG_Z. It likely won't make any difference in code gen, as
REG_Z would have been clobbered by xload8<mode>_A anyway.
