llvm dev - Mar 2012 - [LLVMdev] llvm::Type as llvm::Value in metadata

On Tue, Mar 27, 2012 at 01:55 AM, Eli Friedman <eli.friedman at gmail.com>
wrote:
> On Mon, Mar 26, 2012 at 3:08 PM, Marcus Frenkel
> <marcus.frenkel at fernuni-hagen.de> wrote:
>> Hi all,
>>
>> I've a question about something that's bothering me for quite
some time now: why is a llvm::Function seen as a llvm::Value (in a sense that
llvm::Function is a subclass of llvm::Value), but llvm::Type is not? Should they
not - technically seen - be equal?
>
> llvm::Function is a subclass of llvm::Value because you can use a
> llvm::Function as a value (of type pointer-to-function).  The same is
> not true for llvm::Type.
Yeah, I know that (in fact, I hoped to express exactly that ;) ). The question
is, why is that so? And how can you store references to llvm::Types in the
metadata section of a module?


Greetings,

Marcus

On Mon, Mar 26, 2012 at 11:48 PM, Marcus Frenkel
<marcus.frenkel at fernuni-hagen.de> wrote:
> On Tue, Mar 27, 2012 at 01:55 AM, Eli Friedman <eli.friedman at
gmail.com>
> wrote:
>>
>> On Mon, Mar 26, 2012 at 3:08 PM, Marcus Frenkel
>> <marcus.frenkel at fernuni-hagen.de> wrote:
>>>
>>> Hi all,
>>>
>>> I've a question about something that's bothering me for
quite some time
>>> now: why is a llvm::Function seen as a llvm::Value (in a sense that
>>> llvm::Function is a subclass of llvm::Value), but llvm::Type is
not? Should
>>> they not - technically seen - be equal?
>>
>>
>> llvm::Function is a subclass of llvm::Value because you can use a
>> llvm::Function as a value (of type pointer-to-function).  The same is
>> not true for llvm::Type.
>
>
> Yeah, I know that (in fact, I hoped to express exactly that ;) ). The
> question is, why is that so?
Types are purely a compile-time concept in LLVM IR, just like in C;
there isn't any meta-type or reflection data associated with an LLVM
type.
> And how can you store references to llvm::Types
> in the metadata section of a module?
Essentially, you don't; you can store generic data in a string, though.

-Eli

On Tue, Mar 27, 2012, 10:11 AM, Eli Friedman <eli.friedman at gmail.com>
wrote:
> On Mon, Mar 26, 2012 at 11:48 PM, Marcus Frenkel
> <marcus.frenkel at fernuni-hagen.de> wrote:
>> On Tue, Mar 27, 2012 at 01:55 AM, Eli Friedman <eli.friedman at
gmail.com>
>> wrote:
>>>
>>> On Mon, Mar 26, 2012 at 3:08 PM, Marcus Frenkel
>>> <marcus.frenkel at fernuni-hagen.de> wrote:
>>>>
>>>> Hi all,
>>>>
>>>> I've a question about something that's bothering me for
quite some time
>>>> now: why is a llvm::Function seen as a llvm::Value (in a sense
that
>>>> llvm::Function is a subclass of llvm::Value), but llvm::Type is
not? Should
>>>> they not - technically seen - be equal?
>>>
>>>
>>> llvm::Function is a subclass of llvm::Value because you can use a
>>> llvm::Function as a value (of type pointer-to-function).  The same
is
>>> not true for llvm::Type.
>>
>>
>> Yeah, I know that (in fact, I hoped to express exactly that ;) ). The
>> question is, why is that so?
>
> Types are purely a compile-time concept in LLVM IR, just like in C;
> there isn't any meta-type or reflection data associated with an LLVM
> type.
>
>> And how can you store references to llvm::Types
>> in the metadata section of a module?
>
> Essentially, you don't; you can store generic data in a string, though.
>
> -Eli
>
That's quite a pity, since unnamed types cannot be referenced by their name
nor by their index or any other thing, but only by the pointer you get from
llvm::StructType::get/llvm::ArrayType::get/..., am I right? Furthermore, one has
to know the exact meaning of each string in the metadata (though it would
probably be possible to do some name trickery to encode type names). However
thanks for the answer.

Still, and just for the sake of interest, it would be nice to know about the
reasons for this design decision (that means, why are llvm::Types not considered
llvm::Values while llvm::Functions are).


Greetings

Marcus