[llvm-dev] Function attributes for memory side-effects

Ejjeh, Adel via llvm-dev llvm-dev at lists.llvm.org
Fri May 8 13:49:26 PDT 2020


Johannes, Reid

Thanks for this input. I somehow missed this entire chain earlier and only just saw your replies! I tried with the fast-math flag and it indeed did produce the readnone attribute for the sqrt function (the attribute was not generated without fast-math).

Regards
-Adel

On 4/30/20, 10:36 AM, "Johannes Doerfert" <johannesdoerfert at gmail.com> wrote:


    On 4/29/20 4:12 PM, Reid Kleckner via llvm-dev wrote:
     > On Tue, Apr 28, 2020 at 12:58 PM Ejjeh, Adel via llvm-dev <
     > llvm-dev at lists.llvm.org> wrote:
     >
     >> Specifically, I want to be able to know when a called function does not
     >> have any side effects (e.g. math library functions like sqrt)
     >>
     >
     > Apologies for the pedantry, but I believe sqrt may set errno, so it
     > actually can have side effects. :( See -fno-math-errno and the
     > documentation around it.
     >
     >
     >> , and was wondering if there are attributes that specify this 
    behavior (I
     >> know there is the ‘noread’ attribute but wasn’t sure if there’s 
    something
     >> similar for writes)? Also, how can I tell clang to generate those
     >> attributes at the function declaration? Any information would be 
    helpful.
     >>
     >
     > Yep, I believe the IR attributes are `readonly` and `readnone`. Readonly
     > implies no side effects (no writes), and readnone implies no memory
     > dependencies at all. The return value is a pure function of the 
    arguments.
     > Two calls with the same arguments can be folded together.
     >
     > There are a couple of passes (Attributor, FunctionAttrs, I'm not
     > up-to-date) that will infer these attributes if they can see a precise
     > definition of the function body. Learning anything interesting usually
     > requires expanding the scope of analysis with LTO, so these passes 
    can look
     > across translation unit boundaries.

    Often true. In case of library functions we actually "know" the side
    effects and will add the appropriate attributes. As you said, fast math
    flags are needed for math library functions that may otherwise write
    errno.


    The full list of attributes we have so far is:

        access locations: `readnone`, `inaccessiblememonly`, `argmemonly`, 
    and `inaccessiblemem_or_argmemonly`
      and access "kinds": `readonly` and `writeonly`

    Except for `readnone` you can combine a location attribute with a "kind"
    or have one of either alone. The Attributor does internally derive more
    "locations", basically any combination of:
       local memory
       constant memory
       internal global memory
       external global memory
       argument memory
       inaccessible memory
       malloced memory (returned by a function with the `noalias` return 
    attribute)
       unknown memory
    I want to add some/all of these as attributes but didn't find the time
    yet.


    Cheers,

       Johannes



     >
     > HTH
     >
     >
     > _______________________________________________
     > LLVM Developers mailing list
     > llvm-dev at lists.llvm.org
     > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev




More information about the llvm-dev mailing list