[Bug tree-optimization/127166] [17 Regression] Recent saturation/clipping match.pd changes regressing on RISC-V

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127166

--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Kyrylo Tkachov <ktkachov@gcc.gnu.org>:

https://gcc.gnu.org/g:da3b0463f66312c9e93a027bae73bf17c24a8833

commit r17-3845-gda3b0463f66312c9e93a027bae73bf17c24a8833
Author: Kyrylo Tkachov <ktkachov@nvidia.com>
Date:   Tue Sep 1 00:07:44 2026 +0200

    tree-optimization: Canonicalize conditional signed narrow clips [PR127166]

    The PR120378 RISC-V tests use this source form:

      (NT) ((UT) x > NT_MAX ? (-x) >> (PREC - 1) : x)

    For every defined execution, it is an unsigned saturating truncation of
    MAX (x, 0).  The signed negation is undefined at the signed minimum under
    the default overflow rules.

    Loop if-conversion changes the selected negation to unsigned arithmetic.
    This loses the distinction between the signed source and the explicit
    unsigned-negation form from PR126981, which must not become a saturating
    truncation.

    Canonicalize the signed form while its overflow semantics are still
    available before loop if-conversion:

      (NT) MIN ((UT) MAX (x, 0), NT_MAX)

    Require an unsigned comparison, a signed right shift, undefined and
    unsanitized negation overflow, exact constants, and no side effects.  This
    keeps the original form with -fwrapv, -ftrapv, sanitization, volatile
    operands, and explicit unsigned negation.

    With this patch the tests restore the RISC-V saturation patterns.
    Update pr103771.c because the direct form is
    canonicalized before phiopt.

    Bootstrapped and tested on aarch64-none-linux-gnu.

    gcc/ChangeLog:

            PR tree-optimization/127166
            * match.pd: Canonicalize conditional signed narrow clips before
loop
            if-conversion.

    gcc/testsuite/ChangeLog:

            PR tree-optimization/127166
            * gcc.dg/tree-ssa/pr103771.c: Check the earlier canonicalization.
            * gcc.dg/tree-ssa/pr127166.c: New test.
            * gcc.dg/tree-ssa/pr127166-3.c: New test.
            * gcc.dg/vect/pr127166.c: New test.
            * gcc.target/riscv/rvv/autovec/pr120378-1.c: Update expected dump
            counts.
            * gcc.target/riscv/rvv/autovec/pr120378-2.c: Request an optimized
dump.
            * gcc.target/riscv/rvv/autovec/pr120378-3.c: Likewise.  Update
expected
            dump counts.
            * gcc.target/riscv/rvv/autovec/pr120378-4.c: Request an optimized
dump.

    Signed-off-by: Kyrylo Tkachov <ktkachov@nvidia.com>