------- Comment #15 from hp at gcc dot gnu dot org 2008-05-10 06:59 -------
(In reply to comment #14)
> If there is no CONST wrapping the difference of symbols, does the CRIS port
> continue to believe the address is legitimate?
As I said, for the failure on CRIS, it's not passed as an
*address*, just an ordinary operand.
> In other words, Paolo's patch
> without the CONST?
The CONST changes the world, so hopefully the right thing
would happen.
> Separate from the specific grouping satisfying legitimate_address, the RTL
> does
> contain a difference of SYMBOL_REFs from different sections. If it is not a
> valid address, valid instruction, valid relocation, GCC will generate the
> difference as an explicit subtraction across multiple instructions because it
> is trying to calculate that operation.
Correct.
> I do not understand how the operation
> GCC is trying to perform in the CRIS port is valid using multiple
> instructions.
Maybe it'll help if I just quote the faulty assembly code:
...
move.d $r12,$r13
add.d .LC6-__ZN...mangling elided...rep_storageE-12,$r13
ba .L98
nop
...
(The add.d does what you think; adds the offset to register $r13).
The ".LC6-__ZN...mangling elided...rep_storageE-12" (wrapped in a CONST) is
just passed to LEGITIMATE_CONSTANT_P at replacement time. No MEM in sight,
hence no GO_IF_LEGITIMATE_ADDRESS call.
brgds, H-P
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36182
