Re: [PATCH] tree-optimization: [PR103855] Fold (type)X / (type)Y

[Richard Biener <richard.guenther@gmail.com>]

On Tue, Feb 22, 2022 at 5:01 AM Zhao Wei Liew via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Tue, 22 Feb 2022 at 11:57, Zhao Wei Liew <zhaoweiliew@gmail.com> wrote:
> >
> > Hi,
> >
> > This is a partial optimization for PR103855.
> >
> > Initially, I looked into optimizing RTL generation or a more complex
> > GIMPLE transformation so that we could optimize other cases as well,
> > such as ((unsigned long long) short / int).
> >
> > However, that is a bit too complex for now. While I continue to look
> > into that change, I've decided to implement this simpler match.pd
> > transformation.
> >
> > Greatly appreciate any feedback on this patch or guidance for
> > implementing the more advanced optimizations!
> >
> > Thanks,
> > Zhao Wei
>
> Sorry, the original patch wasn't recognized as a text file. I've added
> a .txt extension to make it explicit.

A few comments - note the change is only appropriate for next stage1.
Since we're
currently in a regression fixing period reviews can be slow - do not
hesitate to ping
forgotten patches when stage1 opens again.

+/* (type)X / (type)Y -> (type)(X / Y)
+   when the resulting type is at least precise as the original types
+   and when all the types are unsigned integral types. */
+(simplify
+ (trunc_div (convert @0) (convert @1))
+ (if (INTEGRAL_TYPE_P (type)
+      && INTEGRAL_TYPE_P (TREE_TYPE (@0))
+      && INTEGRAL_TYPE_P (TREE_TYPE (@1))
+      && TYPE_UNSIGNED (type)
+      && TYPE_UNSIGNED (TREE_TYPE (@0))
+      && TYPE_UNSIGNED (TREE_TYPE (@1))
+      && TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1))
+      && TYPE_PRECISION (type) >= TYPE_PRECISION (TREE_TYPE (@0)))
+  (convert (trunc_div @0 @1))))

since you are requiring the types of @0 and @1 to match it's easier to write

     && types_match (TREE_TYPE(@0), TREE_TYPE (@1))

that allows you to elide checks on either @0 or @1.  I suppose the transform
does not work for signed types because of the -INT_MIN / -1 overflow case?
It might be possible to use expr_not_equal_to (@0, -INT_MIN) ||
expr_not_equal_to (@1, -1)
(correctly written, lookup the existing examples in match.pd for the X
% -Y transform)

I'll note that as written the transform will not catch CST / (T)x or
(T)x / CST since
you'll not see conversions around constants.  I'm not sure whether
using (convert[12]? ...)
or writing special patterns with INTEGER_CST operands is more convenient.
There is int_fits_type_p to check whether a constant will fit in a
type without truncation
or sign change.

When @0 and @1 do not have the same type there might still be a common type
that can be used and is smaller than 'type', it might be as simple as using
build_nonstandard_integer_type (MIN (@0-prec, @1-prec), 1 /*unsigned_p*/).

In the past there have been attempts to more globally narrow operations using
a new pass rather than using individual patterns.  So for more complicated cases
that might be the way to go.  There's now also the ISEL pass which does
pre-RTL expansion transforms that need some global context and for example
can look at SSA uses.

Richard.