Hi,
On Wed, 23 Nov 2016, Segher Boessenkool wrote:
> > Even with non-constant shifts it remains an extract and make_extraction
> > does support variable start positions (which is the shift amount), as does
> > the zero_extract pattern (depending on target of course).
>
> Sure, but the extraction length is non-constant as well then! And not
> just not-constant, with a conditional inside it, even.
Right, that would be the consequence. I don't see much care for this
problem anywhere, which probably hints that there aren't many targets
accepting variant positions for zero_extract :)
> > > if (GET_CODE (inner) == LSHIFTRT
> > > + && CONST_INT_P (XEXP (inner, 1))
> > > && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (inner))
> > > && subreg_lowpart_p (x))
> > > {
> > > new_rtx = make_compound_operation (XEXP (inner, 0), next_code);
> > > + int width = GET_MODE_PRECISION (GET_MODE (inner))
> > > + - INTVAL (XEXP (inner, 1));
> >
> > GET_MODE (new_rtx), because that's the object you're giving to
> > make_extraction, not inner (and not XEXP(inner, 0)).
>
> This is about the *original* rtx, the lshiftrt: how many non-zero bits
> does that leave. The bits zeroed out by the lshiftrt have to be zeroed
> out by the zero_extract as well, to keep the same semantics in the
> resulting rtl.
Of course. I saw it more as a requirements on the inputs of
make_extraction. In the end this is moot, as all three modes (of inner,
of (inner,0) and of new_rtx) are the same :)
Ciao,
Michael.
