Re: [PATCH v2] tree-optimization/94899: Remove "+ 0x80000000" in int comparisons

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

On Thu, Feb 3, 2022 at 4:50 AM Arjun Shankar <arjun@redhat.com> wrote:
>
> Expressions of the form "X + CST < Y + CST" where:
>
> * CST is an unsigned integer constant with only the MSB set, and
> * X and Y's types have integer conversion ranks <= CST's
>
> can be simplified to "(signed) X < (signed) Y".
>
> This is because, assuming a 32-bit signed numbers,
> (unsigned) INT_MIN + 0x80000000 is 0, and
> (unsigned) INT_MAX + 0x80000000 is UINT_MAX.
>
> i.e. the result increases monotonically with signed input.
>
> This means:
> ((signed) X < (signed) Y) iff (X + 0x80000000 < Y + 0x80000000)

+  (op (plus:c INTEGER_CST@0 @1) (plus:c INTEGER_CST@0 @2))

INTEGER_CST are put last by canonicalization so you should write

  (op (plus @1 INTEGER_CST@0) (plus @2 INTEGER_CST@0))

and thus omit the :c.

+#define MAGIC 0x80000000

I _think_ writing the constant this way requires

/* { dg-require-effective-target int32plus } */

as on a target where int is 16 bits it possibly yields a diagnostic?
I've been using explicit int32_t types w/o such effective target
in the past but I'm not sure how to write an explicit precision
constant to avoid diagnostics.

Note I didn't actually check we get a diagnostic ...

Slapping on the dg-requires-effective-target should be safe though.

Otherwise looks good now - note this has to wait for stage1 of GCC 13,
so make sure to ping then or commit then in case you have git access.

Thanks,
Richard.

> gcc/
>         * match.pd (X + C < Y + C -> (signed) X < (signed) Y, if C is
>         0x80000000): New simplification.
> gcc/testsuite/
>         * gcc.dg/pr94899.c: New test.
> ---
>  gcc/match.pd                   | 13 +++++++++
>  gcc/testsuite/gcc.dg/pr94899.c | 48 ++++++++++++++++++++++++++++++++++
>  2 files changed, 61 insertions(+)
>  create mode 100644 gcc/testsuite/gcc.dg/pr94899.c
> ---
> v1: https://gcc.gnu.org/pipermail/gcc-patches/2022-February/589557.html
>
> Notes on v2, based on Richard's review comments:
>
> 1. I removed matching on "convert", and therefore also replaced the
> removal of convert upon simplification with an explicit cast to
> signed. I originally thought this simplification only applies to
> signed operands that have been cast to unsigned, but thinking about
> it, it became clear that they do not necessarily have to be signed
> originally. The simplification is now a bit more general.
>
> 2. Removed checks for operands' types as it seems to be unnecessary. I
> hope this is correct.
>
> 3. Added unsigned types and mismatched sizes of operands to the test.
> These are now simplified.