PR53914, rs6000 constraints and reload queries

Thanks very much Uli for verifying my conclusions about reload,
operand predicates and constraints, and particularly the general
unusability of the "o" constraint.

Re http://gcc.gnu.org/ml/gcc/2012-07/msg00142.html, this patch adds
the missing secondary reload patterns, corrects constraints I got
wrong ("?*d", not "*?d"), and fixes pr54009.

Uli said:
> An address involving pseudos should be
> considered "legitimate" if there exists an assignment of hard
> registers that makes it strictly legitimate (not if *any* such
> assignment would be strictly legitimate).  [ It might make sense
> in some cases to make the check stricter; for example if we know
> that an address would nearly always require a reload, we might
> choose to completely reject it if that actually increases performance.
> But that would be just performance tuning, not required for
> correctness ... ]
So there is quite a bit more work in rs6000.c to fully implement this.
See ??? comments that I added on code handling lo_sum, and I'll admit
to not even trying to relax rs6000_legitimate_offset_address_p
conditions for e500.  That can wait for another day.  The patch is
large enough already.

Some notes:
- word_offset_memref_operand isn't used as a predicate and as both Uli
and I noted, constraints calling predicates lead to trouble with
reload_legitimize_address output.  So move it out of predicates.md to
rs6000.c (renamed as mem_operand_gpr and without checks more suited to
predicates).
- where I changed a bunch of mode tests to GET_MODE_SIZE checks, the
original mode list missing TImode is irrelevant for 32-bit, since
TImode isn't supported on 32-bit (Why do we have 32-bit TImode insns?)
- reordering insn alternatives in some cases is cosmetic.  As the
comments say, putting r->Y and Y->r before r->r is necessary, but
reordering d->m,m->d,d->d isn't strictly necessary.  I did that for
consistency, and future proofing should the m constraint need to be
changed.  Putting r->Y before Y->r is also cosmetic but I prefer it
that way for insns that land in reload as pseudo->pseudo ie. mem->mem,
where both load and store alternatives match with reloading.  I think
it's nicer to choose input reloads rather than output reloads, so put
the store first.
- I haven't actually seen the 32-bit gpr secondary reload patterns
trigger (it's hard to make a testcase), so that code is largely
untested.  Fortunately the code is very similar to the 64-bit gpr
secondary reload code.
- movdf_hardfloat32 insn lengths looked wrong to me, so I fixed that.
gpr load and store ought to be just two insns, not four.  I also took
out the ?? kludge since the offsettable address problem is now fixed.
- I don't really like disparaging fprs in a number of DImode insns,
but without that reload prefers to reload inputs.  So you get code
like stw 10,xxx(1); stw 11,xxx+4(1); lfd 0,xxx(1); stfd 0,32764(9);
rather than addi 9,9,32764; stw 10,0(9); stw 11,4(9);  The former is
slower and requires a stack frame.

Bootstrapped and regression tested powerpc-linux.  OK to apply?

	PR target/53914
	PR target/54009
	* config/rs6000/constraints.md (Y): Use mem_operand_gpr.
	(wY): New constraint using mem_operand_fpr.
	* config/rs6000/predicates.md (word_offset_memref_operand): Delete.
	Adjust all rs6000_legitimate_offset_address_p calls.
	* config/rs6000/rs6000-protos.h (mem_operand_gpr): Declare.
	(mem_operand_fpr, rs6000_secondary_reload_gpr): Declare.
	(rs6000_legitimate_offset_address_p): Update prototype.
	(rs6000_offsettable_memref_p): Delete.
	(rs6000_secondary_reload_ppc64): Delete.
	* config/rs6000/rs6000.c (address_offset): New function.
	(mem_operand_gpr, mem_operand_fpr): Likewise.
	(rs6000_legitimate_offset_address_p): Add worst_case param.  When
	not worst_case assume class of regs with least restrictive offsets.
	Adjust all calls.
	(legitimate_lo_sum_address_p): Simplify register mode tests.
	(rs6000_legitimize_address): Likewise.  Assume best case offset
	addressing.  Combine ELF and MACHO lo_sum code.
	(rs6000_mode_dependent_address): Correct offset addressing limits.
	(rs6000_offsettable_memref_p): Make static, add reg_mode param.
	Use reg_mode to help rs6000_legitimate_offset_address_p.
	(rs6000_secondary_reload): Use address_offset.  Handle 32-bit multi
	gpr load/store when offset too large.
	(rs6000_secondary_reload_gpr): Renamed rs6000_secondary_reload_ppc64.
	(rs6000_split_multireg_move): Adjust rs6000_offsettable_memref_p calls.
	* config/rs6000/rs6000.md (movdf_hardfloat32): Use 'Y' constraint
	for gpr load/store.  Order alternatives as r->Y,Y->r,r->r and
	d->m,m->d,d->d.  Correct size of gpr load/store.
	(movdf_softfloat32): Use 'Y' constraint for gpr load/store.  Order
	alternatives.
	(movti_ppc64): Likewise.
	(movdi_internal32): Likewise.  Also disparage fprs.
	(movdi_mfpgpr, movdi_internal64): Likewise.
	(movtf_internal): Use 'wY' for fpr load/store.  Order alternatives.
	(movtf_softfloat): Order alternatives.
	(extenddftf2_internal): Use 'wY' and 'Y' for store.
	(movti_power, movti_string): Use 'Y' for gpr load/store.  Order.
	(stack_protect_setdi, stack_protect_testdi): Likewise.
	(movdf_hardfloat64_mfpgpr, movdf_hardfloat64): Order alternatives.
	(movdf_softfloat64): Likewise.
	(reload_<mode>_store): Adjust reload_di_store to provide
	reload_si_store as well.
	(reload_<mode>_load): Likewise.

PR53914, rs6000 constraints and reload queries

Commit Message

Comments

Patch