makereport didn't catch it because it uses -march=native flags by
default which produces code without fucompp.

It's really the same stuff just before and after running ia32_x87
simulator, there is no real need to have 2 different things.

simplify it by always using .asciz instead of sometimes .string/.ascii.
All new gcc/binutils understand .asciz.

Also improve doku of machine_size element in backend_params struct.

Otherwise the extra effort to get the operands to the right position is not worth it.
Further, this avoids the unnecessary effort when we not going to use fucompp in favor of fucomi anyway.

The last parameter of x87_create_fpush() is used for debugging only.

Do not store up to three registers for an instruction.
Instead only store the explicit register operand and whether the operands are permuted and/or the result is to be placed in the explicit register operand or st0.

Whether the operands are permuted can be deduced by testing op1_idx != 0.

Let a flag and the emitter handle printing the pop variant.

Let a flag and the emitter handle printing the pop variant.

Let a flag and the emitter handle printing the pop variant.

Let the emitter handle printing the r by checking the assigned registers.

out cannot be null.

Simulating it as normal binop does only work in one special case.
It lacks the reverse form and selecting the second operand register.
So in many cases invalid code was generated silently.