This patch is the second part of a fix for PR tree-optimization/71343,
that implements Richard Biener's suggestion of using tree-ssa's value
numbering instead of match.pd.  The change is that when assigning a
value number for the expression X<<C, we actually look-up or insert
the value number for the multiplication X*(1<<C).  This elegantly
handles the fact that we (intentionally) don't canonicalize these as
equivalent in GIMPLE, and the optimization/equivalence in PR 71343 now
happens by (tree-ssa SCCVN) magic.

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check, both with and without --target_board=unix{-32},
with no new failures.  Ok for mainline?


2022-09-13  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
        PR tree-optimization/71343
        * tree-ssa-sccvn.cc (visit_nary_op) <case LSHIFT_EXPR>: Make
        the value number of the expression X << C the same as the value
        number for the multiplication X * (1<<C).

gcc/testsuite/ChangeLog
        PR tree-optimization/71343
        * gcc.dg/pr71343-2.c: New test case.


Thanks in advance,
Roger
--

> -----Original Message-----
> From: Richard Biener <richard.guenther@gmail.com>
> Sent: 08 August 2022 12:42
> To: Roger Sayle <roger@nextmovesoftware.com>
> Cc: GCC Patches <gcc-patches@gcc.gnu.org>
> Subject: Re: [PATCH] PR tree-optimization/71343: Optimize (X<<C)&(Y<<C) as
> (X&Y)<<C.
> 
> On Mon, Aug 8, 2022 at 10:07 AM Roger Sayle
> <roger@nextmovesoftware.com> wrote:
> >
> > This patch resolves PR tree-optimization/71343, a missed-optimization
> > enhancement request where GCC fails to see that (a<<2)+(b<<2) == a*4+b*4.
> > This requires two related (sets of) optimizations to be added to match.pd.
> >
> > The first is that (X<<C) op (Y<<C) can be simplified to (X op Y) << C,
> > for many binary operators, including AND, IOR, XOR, and (if overflow
> > isn't an issue) PLUS and MINUS.  Likewise, the right shifts (both
> > logical and arithmetic) and bit-wise logical operators can be
> > simplified in a similar fashion.  These all reduce the number of
> > GIMPLE binary operations from 3 to 2, by combining/eliminating a shift
> operation.
> >
> > The second optimization reflects that the middle-end doesn't impose a
> > canonical form on multiplications by powers of two, vs. left shifts,
> > instead leaving these operations as specified by the programmer unless
> > there's a good reason to change them.  Hence, GIMPLE code may contain
> > the expressions "X * 8" and "X << 3" even though these represent the
> > same value/computation.  The tweak to match.pd is that comparison
> > operations whose operands are equivalent non-canonical expressions can
> > be taught their equivalence.  Hence "(X * 8) == (X << 3)" will always
> > evaluate to true, and "(X<<2) > 4*X" will always evaluate to false.
> >
> > This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> > and make -k check, both with and without --target_board=unix{-m32},
> > with no new failures.  Ok for mainline?
> 
> +/* Shifts by constants distribute over several binary operations,
> +   hence (X << C) + (Y << C) can be simplified to (X + Y) << C.  */
> +(for op (plus minus)
> +  (simplify
> +    (op (lshift:s @0 INTEGER_CST@1) (lshift:s @2 INTEGER_CST@1))
> +    (if (INTEGRAL_TYPE_P (type)
> +        && TYPE_OVERFLOW_WRAPS (type)
> +        && !TYPE_SATURATING (type)
> +        && tree_fits_shwi_p (@1)
> +        && tree_to_shwi (@1) > 0
> +        && tree_to_shwi (@1) < TYPE_PRECISION (type))
> 
> I do wonder why we need to restrict this to shifts by constants?
> Any out-of-bound shift was already there, no?
> 
> +/* Some tree expressions are intentionally non-canonical.
> +   We handle the comparison of the equivalent forms here.  */ (for cmp
> +(eq le ge)
> +  (simplify
> +    (cmp:c (lshift @0 INTEGER_CST@1) (mult @0 integer_pow2p@2))
> +    (if (INTEGRAL_TYPE_P (TREE_TYPE (@0))
> +        && tree_fits_shwi_p (@1)
> +        && tree_to_shwi (@1) > 0
> +        && tree_to_shwi (@1) < TYPE_PRECISION  (TREE_TYPE (@0))
> +        && wi::to_wide (@1) == wi::exact_log2 (wi::to_wide (@2)))
> +      { constant_boolean_node (true, type); })))
> +
> +(for cmp (ne lt gt)
> +  (simplify
> +    (cmp:c (lshift @0 INTEGER_CST@1) (mult @0 integer_pow2p@2))
> +    (if (INTEGRAL_TYPE_P (TREE_TYPE (@0))
> +        && tree_fits_shwi_p (@1)
> +        && tree_to_shwi (@1) > 0
> +        && tree_to_shwi (@1) < TYPE_PRECISION  (TREE_TYPE (@0))
> +        && wi::to_wide (@1) == wi::exact_log2 (wi::to_wide (@2)))
> +      { constant_boolean_node (false, type); })))
> 
> hmm.  I wonder if it makes more sense to handle this in value-numbering.
> tree-ssa-sccvn.cc:visit_nary_op handles some cases that are not exactly
> canonicalization issues but the shift vs mult could be handled there by just
> performing the alternate lookup.  That would also enable CSE and by means of
> that of course the comparisons you do above.
> 
> Thanks,
> Richard.
> 
> >
> > 2022-08-08  Roger Sayle  <roger@nextmovesoftware.com>
> >
> > gcc/ChangeLog
> >         PR tree-optimization/71343
> >         * match.pd (op (lshift @0 @1) (lshift @2 @1)): Optimize the
> >         expression (X<<C) + (Y<<C) to (X+Y)<<C for multiple operators.
> >         (op (rshift @0 @1) (rshift @2 @1)): Likwise, simplify (X>>C)^(Y>>C)
> >         to (X^Y)>>C for binary logical operators, AND, IOR and XOR.
> >         (cmp:c (lshift @0) (mult @1)): Optimize comparisons between
> >         shifts by integer constants and multiplications by powers of 2.
> >
> > gcc/testsuite/ChangeLog
> >         PR tree-optimization/71343
> >         * gcc.dg/pr71343-1.c: New test case.
> >         * gcc.dg/pr71343-2.c: Likewise.
> >
> >
> > Thanks in advance,
> > Roger
> > --