On Fri, Jul 10, 2026 at 12:08:25AM +0530, jeevitha wrote:
> Hi All,
>
> The following patch has been bootstrapped and regtested on powerpc64le-linux.
>
> Introduce the new opaque mode `TDOmode` for the `__dmr1024` type, which
> may be supported on a future Power processor, and treat it similarly
> to the existing MMA opaque types when diagnosing invalid function
> parameter and return values. Update the diagnostics to distinguish
> between Dense Math and MMA types in error messages.
>
As it is, it is ok, but you really need to add move patterns
(i.e. movtdo) in mma.md. In adding the move patterns, you will also
need to add support for allocating TDOmode. I'm not sure it makes
sense to add the TDOmode, but not have any way to actually use it.
This will cause a cascade of other things that you will need to do.
Now at this point, until we can get dense math registers added, TDOmode
would have to live in only the VSX registers. In my original patches,
I did not enable __dmr1024 unless we had dense math registers, but we
may want to soften that to just be VSX (similar that we may want XOmode
and OOmode not tied to MMA).
You should add movtdo as a define_insn_and_split. In the split part
after register allocation, you would need to call
rs6000_split_multireg_move.
You will also need to add the reload hooks reload_tdo_from_memory and
reload_tdo_to_memory.
In rs6000-builtin.cc, you would need to setup the TDOmode type and the
__dmr1024 keyword.
Your changes in rs6000-call.cc and rs6000-modes.def are ok.
In rs6000.cc, you have to add support for the TDOmode type. This
includes:
1: In rs6000_modes_tieable_p, you need to add mode tie support
(typically don't allow TDOmode to tie with other modes, but
possibly at some point we want to go back and look at whether
TDOmode, XOmode, OOmode, and the V*modes want to be tied.
2: I find it useful for debugging to add new modes to the
rs6000_debug_reg_global function so -mdebug=reg can print a
summary of the address types and registers it can be allocated to.
3: In rs6000_setup_reg_addr_masks, add support for TDOmode. It
should allow offsettable vector addresses, but not indexed loads.
4: In rs6000_init_hard_regno_mode_ok, you ned to set up the basic
information for TDOmode, i.e. rs6000_vector_unit,
rs6000_vector_mem, and rs6000_vector_align. You need to set up
the reload load/store hooks.
5: In reg_offset_addressing_ok_p, it needs to return true when the
mode is available.
6: In rs6000_preferred_reload_class, you need to prefer VSX
registers.
7: In rs6000_mangle_type, you need to provide a mangling for TDOmode.
8: In rs6000_split_multireg_move, you need to split TDOmodes to
OOmode (vector pair) or V1TImode (normal vector).
9: In rs6000_invalid_conversion, you should add an error, similar to
that for OOmode and XOmode.
In rs6000.h you need to flesh out the support.
1: In VECTOR_ALIGNMENT_P, you need to add TDOmode.
2: In rs6000_builtin_type_index, you need to add TDOmode.
3: Then you need to add a defined for dm1024_type_node to the index
added in step #2.
> 2026-07-10 Jeevitha Palanisamy <[email protected]>
>
> gcc/
> * config/rs6000/rs6000-call.cc (rs6000_return_in_memory): Handle
> TDOmode and
> emit "Dense Math" diagnostics.
> (rs6000_function_arg): Reject TDOmode function parameters and
> emit "Dense Math" diagnostics.
> * config/rs6000/rs6000-modes.def: Add TDOmode.
>
> diff --git a/gcc/config/rs6000/rs6000-call.cc
> b/gcc/config/rs6000/rs6000-call.cc
> index b9b791bfe8a..0da2c1f6071 100644
> --- a/gcc/config/rs6000/rs6000-call.cc
> +++ b/gcc/config/rs6000/rs6000-call.cc
> @@ -432,19 +432,24 @@ rs6000_discover_homogeneous_aggregate (machine_mode
> mode, const_tree type,
> bool
> rs6000_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
> {
> - /* We do not allow MMA types being used as return values. Only report
> - the invalid return value usage the first time we encounter it. */
> + /* We do not allow Dense Math/MMA types being used as return values. Only
> + report the invalid return value usage the first time we encounter it.
> */
> if (cfun
> && !cfun->machine->mma_return_type_error
> && TREE_TYPE (cfun->decl) == fntype
> - && (TYPE_MODE (type) == OOmode || TYPE_MODE (type) == XOmode))
> + && (TYPE_MODE (type) == OOmode
> + || TYPE_MODE (type) == XOmode
> + || TYPE_MODE (type) == TDOmode))
> {
> /* Record we have now handled function CFUN, so the next time we
> are called, we do not re-report the same error. */
> cfun->machine->mma_return_type_error = true;
> if (TYPE_CANONICAL (type) != NULL_TREE)
> type = TYPE_CANONICAL (type);
> - error ("invalid use of MMA type %qs as a function return value",
> + const char *type_class =
> + (TYPE_MODE (type) == TDOmode) ? "Dense Math" : "MMA";
> + error ("invalid use of %s type %qs as a function return value",
> + type_class,
> IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (type))));
> }
>
> @@ -1631,12 +1636,15 @@ rs6000_function_arg (cumulative_args_t cum_v, const
> function_arg_info &arg)
> machine_mode elt_mode;
> int n_elts;
>
> - /* We do not allow MMA types being used as function arguments. */
> - if (mode == OOmode || mode == XOmode)
> + /* We do not allow Dense Math/MMA types being used as function arguments.
> */
> + if (mode == OOmode || mode == XOmode || mode == TDOmode)
> {
> if (TYPE_CANONICAL (type) != NULL_TREE)
> type = TYPE_CANONICAL (type);
> - error ("invalid use of MMA operand of type %qs as a function
> parameter",
> + const char *type_class =
> + (mode == TDOmode) ? "Dense Math" : "MMA";
> + error ("invalid use of %s operand of type %qs as a function parameter",
> + type_class,
> IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (type))));
> return NULL_RTX;
> }
> diff --git a/gcc/config/rs6000/rs6000-modes.def
> b/gcc/config/rs6000/rs6000-modes.def
> index 7140b634c41..0ad326751c4 100644
> --- a/gcc/config/rs6000/rs6000-modes.def
> +++ b/gcc/config/rs6000/rs6000-modes.def
> @@ -76,6 +76,7 @@ VECTOR_MODE (INT, SI, 2); /* V2SI */
> combination. */
> PARTIAL_INT_MODE (TI, 128, PTI);
>
> -/* Modes used by __vector_pair and __vector_quad. */
> +/* Modes used by __vector_pair, __vector_quad and __dmr1024. */
> OPAQUE_MODE (OO, 32);
> OPAQUE_MODE (XO, 64);
> +OPAQUE_MODE (TDO, 128);
>
>
--
Michael Meissner, IBM
PO Box 98, Ayer, Massachusetts, USA, 01432
email: [email protected]
