RFC: Make builtin types only valid for some target features

[Kewen.Lin]

Hi Andrew,

on 2022/12/5 18:10, Andrew Pinski wrote:
> On Sun, Dec 4, 2022 at 11:33 PM Richard Sandiford via Gcc
> <gcc@gcc.gnu.org> wrote:
>>
>> "Kewen.Lin" <linkw@linux.ibm.com> writes:
>>> Hi,
>>>
>>> I'm working to find one solution for PR106736, which requires us to
>>> make some built-in types only valid for some target features, and
>>> emit error messages for the types when the condition isn't satisfied.
>>> A straightforward idea is to guard the registry of built-in type under
>>> the corresponding target feature.  But as Peter pointed out in the
>>> PR, it doesn't work, as these built-in types are used by some built-in
>>> functions which are required to be initialized always regardless of
>>> target features, in order to support target pragma/attribute.  For
>>> the validity checking on the built-in functions, it happens during
>>> expanding the built-in functions calls, since till then we already
>>> know the exact information on specific target feature.
>>>
>>> One idea is to support built-in type checking in a similar way, to
>>> check if all used type_decl (built-in type) are valid or not somewhere.
>>> I hacked to simply check currently expanding gimple stmt is gassign
>>> and further check the types of its operands, it did work but checking
>>> gassign isn't enough.  Maybe I missed something, there seems not an
>>> efficient way for a full check IMHO.
>>>
>>> So I tried another direction, which was inspired by the existing
>>> attribute altivec, to introduce an artificial type attribute and the
>>> corresponding macro definition, during attribute handling it can check
>>> target option node about target feature for validity.  The advantage
>>> is that the checking happens in FE, so it reports error early, and it
>>> doesn't need a later full checking on types.  But with some prototyping
>>> work, I found one issue that it doesn't support param decl well, since
>>> the handling on attributes of function decl happens after that on
>>> attributes of param decl, so we aren't able to get exact target feature
>>> information when handling the attributes on param decl.  It requires
>>> front-end needs to change the parsing order, I guess it's not acceptable?
>>> So I planed to give up and return to the previous direction.
>>>
>>> Does the former idea sound good?  Any comments/suggestions, and other
>>> ideas?
>>>
>>> Thanks a lot in advance!
>>
>> FWIW, the aarch64 fp move patterns emit the error directly.  They then
>> expand an integer-mode move, to provide some error recovery.  (The
>> alternative would be to make the error fatal.)
>>
>> (define_expand "mov<mode>"
>>   [(set (match_operand:GPF_TF_F16_MOV 0 "nonimmediate_operand")
>>         (match_operand:GPF_TF_F16_MOV 1 "general_operand"))]
>>   ""
>>   {
>>     if (!TARGET_FLOAT)
>>       {
>>         aarch64_err_no_fpadvsimd (<MODE>mode);
>>         machine_mode intmode
>>           = int_mode_for_size (GET_MODE_BITSIZE (<MODE>mode), 0).require ();
>>         emit_move_insn (gen_lowpart (intmode, operands[0]),
>>                         gen_lowpart (intmode, operands[1]));
>>         DONE;
>>       }
>>
>> This isn't as user-friendly as catching the error directly in the FE,
>> but I think in practice it's going to be very hard to trap all invalid
>> uses of a type there.  Also, the user error in these situations is likely
>> to be forgetting to enable the right architecture feature, rather than
>> accidentally using the wrong type.  So an error about missing architecture
>> features is probably good enough in most cases.
> 
> I did have a patch which improved the situation for the SVE types to
> provide an error message at compile time when SVE is not enabled
> but I didn't get any feedback from either the C or C++ front-end folks.
> https://gcc.gnu.org/pipermail/gcc-patches/2021-November/583786.html
> 

Nice!  Many thanks for providing this new direction.

> I suspect if that patch gets reviewed by the front-end folks, Kewen
> could use the same infrastructure to error out on the types for rs6000
> backend.

Yeah, I just confirmed that on top of your patch introducing function
rs6000_verify_type_context to take care of those MMA types can fix the
issue in PR106736.  TBH, I'm not sure if your patch can cover all
possible places where a built-in type can be used, but I guess it can
cover the most.

BR,
Kewen