RE: [PATCH 1/2]middle-end Fold BIT_FIELD_REF and Shifts into BIT_FIELD_REFs alone

[Richard Biener <rguenther@suse.de>]

On Mon, 26 Sep 2022, Tamar Christina wrote:

> > -----Original Message-----
> > From: Andrew Pinski <pinskia@gmail.com>
> > Sent: Saturday, September 24, 2022 8:57 PM
> > To: Tamar Christina <Tamar.Christina@arm.com>
> > Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>; rguenther@suse.de
> > Subject: Re: [PATCH 1/2]middle-end Fold BIT_FIELD_REF and Shifts into
> > BIT_FIELD_REFs alone
> > 
> > On Fri, Sep 23, 2022 at 4:43 AM Tamar Christina via Gcc-patches <gcc-
> > patches@gcc.gnu.org> wrote:
> > >
> > > Hi All,
> > >
> > > This adds a match.pd rule that can fold right shifts and
> > > bit_field_refs of integers into just a bit_field_ref by adjusting the
> > > offset and the size of the extract and adds an extend to the previous size.
> > >
> > > Concretely turns:
> > >
> > > #include <arm_neon.h>
> > >
> > > unsigned int foor (uint32x4_t x)
> > > {
> > >     return x[1] >> 16;
> > > }
> > >
> > > which used to generate:
> > >
> > >   _1 = BIT_FIELD_REF <x_2(D), 32, 32>;
> > >   _3 = _1 >> 16;
> > >
> > > into
> > >
> > >   _4 = BIT_FIELD_REF <x_1(D), 16, 48>;
> > >   _2 = (unsigned int) _4;
> > >
> > > I currently limit the rewrite to only doing it if the resulting
> > > extract is in a mode the target supports. i.e. it won't rewrite it to
> > > extract say 13-bits because I worry that for targets that won't have a
> > > bitfield extract instruction this may be a de-optimization.
> > 
> > It is only a de-optimization for the following case:
> > * vector extraction
> > 
> > All other cases should be handled correctly in the middle-end when
> > expanding to RTL because they need to be handled for bit-fields anyways.
> > Plus SIGN_EXTRACT and ZERO_EXTRACT would be used in the integer case
> > for the RTL.
> > Getting SIGN_EXTRACT/ZERO_EXTRACT early on in the RTL is better than
> > waiting until combine really.
> > 
> 
> Fair enough, I've dropped the constraint.
> 
> > 
> > >
> > > Bootstrapped Regtested on aarch64-none-linux-gnu, x86_64-pc-linux-gnu
> > > and no issues.
> > >
> > > Testcase are added in patch 2/2.
> > >
> > > Ok for master?
> > >
> > > Thanks,
> > > Tamar
> > >
> > > gcc/ChangeLog:
> > >
> > >         * match.pd: Add bitfield and shift folding.
> > >
> > > --- inline copy of patch --
> > > diff --git a/gcc/match.pd b/gcc/match.pd index
> > >
> > 1d407414bee278c64c00d425d9f025c1c58d853d..b225d36dc758f1581502c8d03
> > 761
> > > 544bfd499c01 100644
> > > --- a/gcc/match.pd
> > > +++ b/gcc/match.pd
> > > @@ -7245,6 +7245,23 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > >        && ANY_INTEGRAL_TYPE_P (type) && ANY_INTEGRAL_TYPE_P
> > (TREE_TYPE(@0)))
> > >    (IFN_REDUC_PLUS_WIDEN @0)))
> > >
> > > +/* Canonicalize BIT_FIELD_REFS and shifts to BIT_FIELD_REFS.  */
> > > (for
> > > +shift (rshift)
> > > +     op (plus)

why have a for when you only iterate over a single operation?!  And 'op'
seems unused?

> > > + (simplify
> > > +  (shift (BIT_FIELD_REF @0 @1 @2) integer_pow2p@3)
> > > +  (if (INTEGRAL_TYPE_P (type))
> > > +   (with { /* Can't use wide-int here as the precision differs between
> > > +             @1 and @3.  */
> > > +          unsigned HOST_WIDE_INT size = tree_to_uhwi (@1);
> > > +          unsigned HOST_WIDE_INT shiftc = tree_to_uhwi (@3);

But you should then test tree_fits_uhwi_p.

> > > +          unsigned HOST_WIDE_INT newsize = size - shiftc;
> > > +          tree nsize = wide_int_to_tree (bitsizetype, newsize);
> > > +          tree ntype
> > > +            = build_nonstandard_integer_type (newsize, 1); }

build_nonstandard_integer_type never fails so I don't see how
you "limit" this to extractions fitting a mode.

I'm quite sure this breaks with BYTES_BIG_ENDIAN.  Please try
BIT_FIELD_REF _offsets_ that make the extraction cross byte
boundaries.

Also I'm missing a testcase?

Thanks,
Richard.

> > Maybe use `build_nonstandard_integer_type (newsize, /* unsignedp = */
> > true);` or better yet `build_nonstandard_integer_type (newsize,
> > UNSIGNED);`
> 
> Ah, will do,
> Tamar.
> 
> > 
> > I had started to convert some of the unsignedp into enum signop but I never
> > finished or submitted the patch.
> > 
> > Thanks,
> > Andrew Pinski
> > 
> > 
> > > +    (if (ntype)
> > > +     (convert:type (BIT_FIELD_REF:ntype @0 { nsize; } (op @2
> > > + @3))))))))
> > > +
> > >  (simplify
> > >   (BIT_FIELD_REF (BIT_FIELD_REF @0 @1 @2) @3 @4)
> > >   (BIT_FIELD_REF @0 @3 { const_binop (PLUS_EXPR, bitsizetype, @2, @4);
> > > }))
> > >
> > >
> > >
> > >
> > > --
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)