llvm dev - May 2016 - basic blocks with input and output parameters instead of phis

I always found the phi instruction pretty unintuitive, because

  1. I thought that due to the SSA form the register uniquely determines the 
basic block and thus the phi instruction should not need the information of a 
"predecessor block".

  2. I remember to have seen valid LL files where the predecessor block in a phi
instruction was not an immediate predecessor.


I thought that it would be more intuitive to have input and output parameters 
in basic blocks like so:

start:  param (i32 %0, i1 %1)
    ...
    br i1 %1, label start (%2, %3), label stop (%4)

stop: param (i16 %5)
    ...


Has anybody thought about such a structure, yet?

Swift's IR does this;
https://gist.github.com/modocache/0b688175ca77cebc46f8#file-fibonacci-sil-ll

Designed and implemented by some of the same people that designed llvm &
clang.


On Sun, May 1, 2016 at 9:51 PM, Henning Thielemann via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
>
> I always found the phi instruction pretty unintuitive, because
>
>  1. I thought that due to the SSA form the register uniquely determines the
> basic block and thus the phi instruction should not need the information of
> a "predecessor block".
>
>  2. I remember to have seen valid LL files where the predecessor block in a
> phi instruction was not an immediate predecessor.
>
>
> I thought that it would be more intuitive to have input and output
> parameters in basic blocks like so:
>
> start:  param (i32 %0, i1 %1)
>    ...
>    br i1 %1, label start (%2, %3), label stop (%4)
>
> stop: param (i16 %5)
>    ...
>
>
> Has anybody thought about such a structure, yet?
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev