On Wed, Jan 21, 2026 at 03:52:53PM +0000, Wilco Dijkstra wrote:
> Hi Alice,
> �
> > We can't use the cntd instruction in non-streaming mode if SVE is not
> > available, so instead use __arm_get_current_vg to get the value for the
> > VG save slot.� This is more expensive, so continue using cntd if we know
> > we're
> > in streaming mode or have +sve enabled.
> >
> > This is required to support configurations with SME and no (non-streaming)
> > SVE,
> > and also to support calling a non-streaming function from a
> > streaming-compatible function when SVE and SME are both disabled.
>
> This is OK.
There's actually a small bug in this version - see comment below. I'l fix that
in the next version (and make sure that the tests in later patches reflect that
change).
>
> Note this code is starting to look extremely messy and risky - in the future
> all
> this should be cleaned up and simplified into separate functions since the VG
> is
> not a callee-save. Also if do we want to keep this as a call right in the
> middle of
> the prolog, we should add a variant that doesn't corrupt x0.
Hmm, good point - since we were already specifying a custom ABI in the ACLE
spec, maybe we should have considered using different return registers in the
first place. This applies to the helpers for streaming-compatible code as
well.
>
> Cheers,
> Wilco
>
>
> gcc/ChangeLog:
>
> ������� * config/aarch64/aarch64-sme.md
> (UNSPEC_GET_CURRENT_VG): New
> ������� enum value.
> ������� (aarch64_get_current_vg): New insn.
> ������� * config/aarch64/aarch64.cc
> (aarch64_save_callee_saves): Use
> ������� __arm_get_current_vg if cntd is unavailable.
>
>
> diff --git a/gcc/config/aarch64/aarch64-sme.md
> b/gcc/config/aarch64/aarch64-sme.md
> index
> ca3ed47165700dcc488cff55a6a788bdd98ddde6..5e64712e6bec4ae1621f2660f562585cee4c8724
> 100644
> --- a/gcc/config/aarch64/aarch64-sme.md
> +++ b/gcc/config/aarch64/aarch64-sme.md
> @@ -76,6 +76,7 @@
> �(define_c_enum "unspec" [
> �� UNSPEC_OLD_VG_SAVED
> �� UNSPEC_UPDATE_VG
> +� UNSPEC_GET_CURRENT_VG
> �� UNSPEC_GET_SME_STATE
> �� UNSPEC_READ_SVCR
> �])
> @@ -101,6 +102,19 @@
> �� [(set_attr "type" "no_insn")]
> �)
> �
> +(define_insn "aarch64_get_current_vg"
> +� [(set (reg:DI R0_REGNUM)
> +������ (unspec_volatile:DI [(const_int 0)]
> UNSPEC_GET_CURRENT_VG))
> +�� (clobber (reg:DI R16_REGNUM))
> +�� (clobber (reg:DI R17_REGNUM))
> +�� (clobber (reg:DI R18_REGNUM))
> +�� (clobber (reg:DI R30_REGNUM))
> +�� (clobber (reg:CC CC_REGNUM))]
> +� ""
> +� "bl\t__arm_get_current_vg"
> +� [(set_attr "is_call" "yes")]
> +)
> +
> �(define_insn "aarch64_get_sme_state"
> �� [(set (reg:TI R0_REGNUM)
> �������� (unspec_volatile:TI [(const_int 0)]
> UNSPEC_GET_SME_STATE))
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index
> b7f56c8d560fc3e3ea2a403ae8cba3735770fdc4..f150fccbe1b140b7f2f67310190ce17f532ed846
> 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -9387,11 +9387,26 @@ aarch64_save_callee_saves (poly_int64 bytes_below_sp,
> ������ machine_mode mode = aarch64_reg_save_mode (regno);
> ������ rtx reg = gen_rtx_REG (mode, regno);
> ������ rtx move_src = reg;
> +����� rtx old_r0 = NULL_RTX;
> ������ offset = frame.reg_offset[regno] - bytes_below_sp;
> ������ if (regno == VG_REGNUM)
> �������� {
> -�������� move_src = gen_rtx_REG (DImode, IP0_REGNUM);
> -�������� emit_move_insn (move_src, gen_int_mode
> (aarch64_sve_vg, DImode));
> +�������� if (TARGET_SVE || TARGET_STREAMING)
This checks feature availability in the function body, which differs from
availability in the prologue before a locally streaming function enables
streaming mode.
> +���������� {
> +������������ move_src = gen_rtx_REG (DImode,
> IP0_REGNUM);
> +������������ emit_move_insn (move_src, gen_int_mode
> (aarch64_sve_vg, DImode));
> +���������� }
> +�������� else
> +���������� {
> +������������ auto &args = crtl->args.info;
> +������������ if (args.aapcs_ncrn > 0)
> +�������������� {
> +���������������� old_r0 = gen_rtx_REG
> (DImode, PROBE_STACK_FIRST_REGNUM);
> +���������������� emit_move_insn (old_r0,
> gen_rtx_REG (DImode, R0_REGNUM));
> +�������������� }
> +������������ emit_insn (gen_aarch64_get_current_vg
> ());
> +������������ move_src = gen_rtx_REG (DImode,
> R0_REGNUM);
> +���������� }
> �������� }
> ������ rtx base_rtx = stack_pointer_rtx;
> ������ poly_int64 sp_offset = offset;
> @@ -9482,9 +9497,13 @@ aarch64_save_callee_saves (poly_int64 bytes_below_sp,
> ������ RTX_FRAME_RELATED_P (insn) = frame_related_p;
> �
> ������ /* Emit a fake instruction to indicate that the VG save
> slot has
> -������� been initialized.� */
> +������� been initialized, and then restore R0 if necessary.�
> */
> ������ if (regno == VG_REGNUM)
> -������ emit_insn (gen_aarch64_old_vg_saved (move_src, mem));
> +������ {
> +�������� emit_insn (gen_aarch64_old_vg_saved (move_src,
> mem));
> +�������� if (old_r0)
> +���������� emit_move_insn (gen_rtx_REG (DImode,
> R0_REGNUM), old_r0);
> +������ }
> ���� }
> �}
> �
