Richard Henderson <r...@redhat.com> writes:
> On 2012-06-16 00:45, Richard Sandiford wrote:
>> [(mem:GPR (match_operand:P 1 "register_operand" "d"))]
>>
>> Instead, we should define a new memory predicate/constraint pair
>> for memories that only accept register addresses. I.e. there
>> should be a new predicate to go alongside things like
>> memory_operand and stack_operand, except that the new one would
>> be even more restrictive in the set of addresses that it allows.
>> mem_reg_operand seems as good a name as any, but I'm not wedded
>> to a particular name.
>
> C.f. mem_noofs_operand inthe ARM port, and its uses in sync.md.
>
>> The atomic_exchange and atomic_fetch_add expanders should use
>> the code I quoted in the earlier message to force the original
>> memory_operand into this more restrictive form:
>>
>> if (!mem_reg_operand (operands[1], <MODE>mode))
>> {
>> addr = force_reg (Pmode, XEXP (operands[1], 0));
>> operands[1] = replace_equiv_address (operands[1], addr);
>> }
>
> Not required if you use the proper predicate in the expander.
> The middle-end will take care of this for you.
I might be misunderstanding, sorry, but this expander is shared with
the normal LL/SC path, which can accept plain memory_operands.
I was thinking we'd want to keep the expander predicates the same
and apply these SWAP-style restrictions only when needed.
I suppose we could define a predicate that's equivalent to
mem_reg_operand when ISA_HAS_SWAP and to memory_operand otherwise
(and maybe do the same for arith_operand vs. register_operand
in point (2)). That might get confusing though. E.g. Maxim
has kept the ISA flags for SWAP and LDADD separate, which seems
like a good thing. We'd then need another pair of predicates
for the LDADD case.
Richard
