This fixes fa66e6b8.

ia32: When emitting an asm memory operand, ignore the size of the referenced value and asm operand size modifier.

The register, which is emitted, contains the address of the value, so it always is the size of a pointer, not whatever the register points to.

It does not provide sufficient information to handle float in asm.

The reloader never is a block.

bespillutil: Inline be_add_reload2() into its only caller be_add_reload(), which in turn only calls the former.

bemain: Centrally call be_gas_end_compilation_unit() and be_emit_exit() instead of doing it per backend.

Also remove the now write-only attribute main_env from struct arch_env_t.

bemain: Centrally call be_emit_init() and be_gas_begin_compilation_unit() instead of doing it per backend.

beemit: Provide be_emit_node(), which merges the (almost) identical node emit functions of the backends.

Also use one common be_emit_nothing().

bestate: Use be_get_end_of_block_insertion_point() instead of implementing something not entierly unlike it.

bespillslots: Use be_get_end_of_block_insertion_point() instead of implementing something not entierly unlike it.

This also handles the latent problem, that nobody checks, whether the user of a tuple is actually a Proj and not the End node.

Projs are not scheduled anymore.

All users check, that the blocks are equal, beforehand.