GitLab

This resolves problems with hidden dependencies during code selection, which cause dependency cycles and therefore fixes backend/scheduled.c.
Also it grants the code selection slightly more freedom by not arbitrarily restricting the order by stack dependencies.

be: Pass the input requirements directly to the constructors of nodes, which have a variable number of arguments.

Properly use get_irn_pinned/set_irn_pinned as bool value instead of
comparing setting the ir_op enum values.

The generator chooses the right default automatically.

All backends had identical implementations.
Now rematerialization works out of the box for new backends.

This adds two builtin kinds, ir_bk_va_start and ir_bk_va_arg. va_arg Builtins
are lowered in the lower_builtins pass with a function provided by the backend,
whereby most architectures can use be_default_lower_va_args. va_start Builtins
are lowered in the backend.

Note: This commit breaks all backends except for IA32. SPARC and AMD64 are
implemented in later commits. The ARM backend previously miscompiled variadic
functions, and will now throw a proper error.

All backends have similar empty Start nodes, which just provide registers.
So having one generic backend Start node to share some infrastructure is sensible.

It is now done generically in dump_backend_info_hook().

bescripts: Automatically determine the mode of nodes by default and add the option "first" to use the mode of the first output requirement.

The node must use the mode of the floating point register class, irrespective of the mode of the loaded constant.

arm, be, ia32, sparc: Do not go on with peephole optimization of IncSP, when be_peephole_IncSP_IncSP() succeeded.

The current IncSP was removed in this case and we now would operate on its predecessor IncSP.
This is wrong, e.g. when asking for currently free registers.
This corrects backend/incsp_pop.c, which was broken on ia32.
It was wrongly assumed that a certain register was free and could be used to turn an IncSP into a Pop.

Duplicate Projs are a problem when CSE is off, e.g. -O0.

It is like new_r_Proj(), but the mode is determined automatically from the out requirement of the predecessor.
This probably resolves some mode mismatches in untested code paths -- AMD64 backend and SPARC hard float, I'm looking at you!

We use lower_CopyB anyway so there are no copybs left when lowering.
They are all lowered to load/store sequences or memcpy calls.

arm_cmp_attr_t contains an arm_shifter_operand_t as base, but this was not checked.
This should correct langshootout/bintree.c.

One Conv is sufficient.

The modes did not match in several places.

Simplify the creation of instrinsic/runtime function entities a bit and avoid creating duplicate types.

Their values are either unused or trivial, i.e. result k has reg_offset k.

This is simpler, removes some code duplication and makes the backends more robust against duplicate Projs of Start, which can happen if optimisations are disabled.

- Rename other_same field to should_be_same.
- Remove arch_register_req_type_other_same, it's enough to see if the
  should_be_same field is not zero.