http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48155
Summary: Reload doesn't handle subreg properly
Product: gcc
Version: 4.6.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: hjl.to...@gmail.com
Given input:
(plus:SI (subreg:SI (plus:DI (reg/f:DI 7 sp)
(const_int 16 [0x10])) 0)
(const_int -1 [0xffffffffffffffff]))
reloads tries to add
(subreg:SI (plus:DI (reg/f:DI 7 sp)
(const_int 16 [0x10])) 0)
to
(reg:SI 1 dx)
This code:
/* How to do this reload can get quite tricky. Normally, we are being
asked to reload a simple operand, such as a MEM, a constant, or a pseudo
register that didn't get a hard register. In that case we can just
call emit_move_insn.
We can also be asked to reload a PLUS that adds a register or a MEM to
another register, constant or MEM. This can occur during frame pointer
elimination and while reloading addresses. This case is handled by
trying to emit a single insn to perform the add. If it is not valid,
we use a two insn sequence.
Or we can be asked to reload an unary operand that was a fragment of
an addressing mode, into a register. If it isn't recognized as-is,
we try making the unop operand and the reload-register the same:
(set reg:X (unop:X expr:Y))
-> (set reg:Y expr:Y) (set reg:X (unop:X reg:Y)).
Finally, we could be called to handle an 'o' constraint by putting
an address into a register. In that case, we first try to do this
with a named pattern of "reload_load_address". If no such pattern
exists, we just emit a SET insn and hope for the best (it will normally
be valid on machines that use 'o').
This entire process is made complex because reload will never
process the insns we generate here and so we must ensure that
they will fit their constraints and also by the fact that parts of
IN might be being reloaded separately and replaced with spill registers.
Because of this, we are, in some sense, just guessing the right approach
here. The one listed above seems to work.
??? At some point, this whole thing needs to be rethought. */
if (GET_CODE (in) == PLUS
&& (REG_P (XEXP (in, 0))
|| GET_CODE (XEXP (in, 0)) == SUBREG
|| MEM_P (XEXP (in, 0)))
&& (REG_P (XEXP (in, 1))
|| GET_CODE (XEXP (in, 1)) == SUBREG
|| CONSTANT_P (XEXP (in, 1))
|| MEM_P (XEXP (in, 1))))
doesn't check if XEXP (in, 0/1) is a SUBREG of REG.
