[LLVMdev] [RFC] Overhauling Attributes

[Bill Wendling]

On Dec 31, 2012, at 4:37 AM, Duncan Sands <baldrick at free.fr> wrote:

> Hi Bill,
> 
> On 30/12/12 03:21, Bill Wendling wrote:
>> Hi Rafael,
>> 
>> Sorry, I forgot to respond to this. They can be arbitrary strings that are known only to the specific back-end. It may be beneficial to define them inside of the LangRef document though.
> 
> this sounds so much like metadata...  What was the reason for not enhancing
> metadata to cover this use case?  I'm sure you explained but I've completely
> forgotten...
> 
Hi Duncan,

There are a couple of reasons why I don't think it's a good idea to use metadata (or more specifically, the module-level flags). Firstly, everything would have to be specified as a strings: "noinline", "ssp", etc., because a metadata object can hold only a Value type. Secondly, I don't want to have a "loop" in the attributes:

  !1 = metadata !{ !"noinline", !2 }
  !2 = metadata !{ !1, !"ssp" }

This makes uniquifying the attributes that much harder.

We also want to be able to intelligently merge the attribute groups. These two groups are identical:

  #1 = attributes { noredzone noinline sspreq "mcpu"="cortex-a8" }
  #2 = attributes { "mcpu"="cortex-a8" sspreq noinline noredzone }

Which brings up the '<kind>=<value>' syntax. The ability to have (possibly multiple) values assigned to an attribute kind isn't something easily modeled in metadata, as far as I can tell.

These (and maybe a few more, it's late) are some of the reasons why I feel a new first-class IR construct is needed. It's possible to modify metadata to handle these. But I feel that it's much more cumbersome to do that.

-bw