https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128690

--- Comment #5 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <[email protected]>:

https://gcc.gnu.org/g:535e0759e5b7153470d4d2bf91c8d1a04747948f

commit r17-6380-g535e0759e5b7153470d4d2bf91c8d1a04747948f
Author: liuhongt <[email protected]>
Date:   Mon Nov 16 21:51:26 2026 -0800

    x86_cse: Drop REG_EQUAL/REG_EQUIV notes on constant integer loads
[PR128690]

    When replace_vector_const converts a CONST_VECTOR load no larger than
    an integer register into a constant integer load, it rewrites the
    destination with gen_lowpart.  For

    (set (subreg:V2QI (reg:HI 126) 0) (plus:V2QI ...))

    with a REG_EQUAL (const_vector:V2QI [5 5]) note, the result is

    (set (reg:HI 126) (const_int 1285))

    with the V2QImode REG_EQUAL note left in place.  Since the destination
    is now a bare REG, IRA records the note as an equivalence of the
    HImode pseudo.  The CONST_VECTOR isn't a legitimate HImode constant, so
    IRA calls force_const_mem (HImode, const_vector:V2QI), which creates a
    constant pool entry whose mode doesn't match its constant.  When LRA
    spills the pseudo and uses that equivalence, optimize_constant_pool
    hits the mode check in native_encode_rtx.

    The SET_SRC is now the constant itself, so the note is redundant.
    Remove REG_EQUAL and REG_EQUIV notes when the source is replaced with
    a CONST_INT.  IRA still records the equivalence from the CONST_INT
    source.

    gcc/ChangeLog:

            PR target/128690
            * config/i386/i386-features.cc (replace_vector_const): Remove
            REG_EQUAL and REG_EQUIV notes when replacing the source with a
            CONST_INT.

    gcc/testsuite/ChangeLog:

            PR target/128690
            * gcc.target/i386/pr128690.c: New test.

Reply via email to