This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [MIPS] Add sbasic supoert ffor MSA (SIMD)


Matthew Fortune <Matthew.Fortune@imgtec.com> writes:
> Mike Stump <mikestump@comcast.net> writes:
>> On May 28, 2014, at 7:27 AM, Richard Earnshaw <rearnsha@arm.com> wrote:
>> >
>> > Speed of implementation.  We're gradually replacing these with proper
>> > builtins, but that takes a lot more work.
>> 
>> As an owner of a port with more builtins that yours, I can offer a
>> technological solution to reduce the cost of builtins to:
>> 
>> (define_builtin "my_stop"
>>   [
>>     (define_outputs [(void_operand 0)])
>>     (define_rtl_pattern "my_stop" [])
>>   ]
>> )
>> 
>> (define_insn "my_stop"
>>   [(unspec_volatile [(const_int 0)]
>>                     UNSPECV_STOP)]
>>   ""
>>   "stop")
>> 
>> for example.  This creates the builtins, allows overloading, allows
>> input/output parameters, can reorder operands, allows for complex types,
>> allows memory reference parameters, allows pure markings, does vectors,
>> conditional availability, generates documentation, creates test suites and
>> more.  If you wire up a speaker it even sings.
>> 
>> Someone would have have to step forward with a need and some time to port
>> their port over to the new scheme and help with the reason for why the
>> technology should go in.  It is mostly contained in 5600 lines of self
>> contained python code, and is built to solve the problem generally.  It adds
>> about 800 lines to builtins.c.  It has a macro system that is more powerful
>> than the macro system .md files use, so one gets to share and collapse
>> builtins rather nicely.  It is known to work for C and C++.  Other languages
>> may need extending; C for example cost is around 250 lines to support.
>
> Myself and others at IMG would be interested in reviewing/evaluating the
> implementation and assuming it looks useful then we would of course help to
> get it in shape for submission.
>  
>> One promise, you will never have to create an argument list, or a type, for
>> example here is a two output, type input functional instruction with some
>> doc content:
>> 
>> (define_mode_iterator MYTYPE
>>         [V8QI V4HI V2SI DI ...])
>> 
>> (define_builtin "my_foo" "my_foo2_<type>"
>>   [
>>     (define_desc    "Doc string for operation")
>>     (define_outputs [(var_operand:T_MYTYPE 0)
>>                      (var_operand:T_MYTYPE 1)])
>>     (define_inputs  [(var_operand:T_MYTYPE 2)
>>                      (var_operand:T_MYTYPE 3)])
>>     (define_rtl_pattern "my_foo2_<mode>" [0 2 1 3])
>>     (attributes [pure])
>>   ]
>> )
>> 
>> I stripped it so you can't know what the instruction was, but you get a
>> flavor of multiple outputs, doc bits, pure, overloading, arguments and
>> argument rearranging.
>
> Can you post the implementation as an RFC? I suspect the python aspect
> will cause the most trouble as GCC builds do not currently require python
> I guess that could change depending on the value added. Otherwise it would
> be a rewrite I guess.
>
> Before digging in too deep though it would be useful to know if RichardS
> would be willing to consider this kind of thing for the MIPS port?

Yeah, it definitely sounds good in principle.

Richard


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]