https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123626

--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jeff Law <[email protected]>:

https://gcc.gnu.org/g:6b79f1a5b8b62ce711d590a7f3e6ecb40b465464

commit r16-6939-g6b79f1a5b8b62ce711d590a7f3e6ecb40b465464
Author: Jeff Law <[email protected]>
Date:   Tue Jan 20 13:57:10 2026 -0700

    [RISC-V][PR target/123626] Fix VXRM state after calls

    This is a partial fix for a long standing issue that Richard S. raised
about a
    year ago.

    Specifically he indicated that he believed our handling of VXRM mode
switching
    was wrong and could lead to incorrect code, particularly WRT handling of
calls.

    Without rehashing everything related to VXRM, its sufficient to say that it
has
    no known value at function entry or upon returning from a call.

    If we look at the main scan loop in mode-switching we have:

    >           FOR_BB_INSNS (bb, insn)
    >             {
    >               if (NONDEBUG_INSN_P (insn))
    >                 {
    >                   int mode = targetm.mode_switching.needed (e, insn,
live_now);
    >                   rtx link;
    >
    >                   if (mode != no_mode && mode != last_mode)
    >                     {
    >                       ptr = new_seginfo (last_mode, mode, insn,
live_now);
    >                       add_seginfo (&tail_ptr, ptr);
    >                       bitmap_clear_bit (transp_all, bb->index);
    >                       any_set_required = true;
    >                       last_mode = mode;
    >                     }

    The way to think about this is if INSN requests a mode and it is not the
same
    as the last mode, then we've got a new point where we need to logically
insert
    a mode switch and we clear TRANSP.

    A CALL_INSN in the RISC-V backend produces NO_MODE (it doesn't need VXRM
    state).  So we never get into the then clause of that inner if statement
and
    TRANSP stays on.

    The fix is quite simple.  We need one more state in the VXRM mode switching
    that indicates we don't know VXRM's state after a call. While I could have
    hacked up the various hooks to special case CALLs, it was just as easy to
    adjust the attribute's generic handling so that any CALL_P is given the
    VXRM_MODE_CLOBBER state.

    Out of an abundance of caution if we'll filter out any actual code
generation
    setting it to CLOBBER state.

    This is enough to make the testcase pass for rv64.  It's still failing
rv32,
    but likely for completely different reasons.   It obviously doesn't cause
any
    regressions on riscv{32,64}-elf and bootstraps will fire up later today on
the
    Pioneer and BPI.

            PR target/123626
    gcc/
            * config/riscv/vector.md (vxrm_mode): Handle CALL_INSNs, which set
            the attribute to the new VXRM_MODE_CLOBBER state.
            * config/riscv/riscv.cc (riscv_emit_mode_set): Don't emit code when
            VXRM's state changes to VXRM_MODE_CLOBBER.

    gcc/testsuite
            * gcc.target/riscv/rvv/base/pr123626.c: New test.

Reply via email to