Re: [PATCH] rs6000: Split movsf_from_si from high word before reload[PR89310]

[Segher Boessenkool <segher@kernel.crashing.org>]

Hi!

On Tue, Jul 07, 2020 at 04:39:58PM +0800, luoxhu wrote:
> > Lots of questions, sorry!
> 
> Thanks for the nice suggestions of the initial patch contains many issues:),

Pretty much all of it should *work*, it just can be improved and
simplified quite a bit :-)

> For this case, %1:SF matches with "=wa"?  And how to construct cases to
> match("=?r", "wa") and ("=!r", "r") combinations, please?

operands[0], not operands[1]?

Simple testcases will not put the output into a GPR, unless you force
the compiler to do that, because of the ? and !.

Often you can just do

  asm("#" : "+r"(x));

to force "x" into a GPR at that point of the program.  But there is
nothing stopping the compiler from copying it back to a VSR where it
thinks that is cheaper ;-)

So maybe this pattern should just have the GPR-to-VSR alternative?  It
does not look like the GPR destination variants are useful?

> +  rtx op0 = operands[0];
> +  rtx op1 = operands[1];
> +  rtx op2 = operands[2];

(Please just write out operands[N] everywhere).

> +  if (GET_CODE (operands[2]) == SCRATCH)
> +    op2 = gen_reg_rtx (DImode);
> +
> +  rtx mask = GEN_INT (HOST_WIDE_INT_M1U << 32);
> +  emit_insn (gen_anddi3 (op2, op1, mask));

Groovy :-)

So, it looks like you can remove the ? and ! alternatives, leaving just
the first alternative?


Segher