[LLVMdev] choice between SSAPRE and bitvector aporach

Bill Wendling isanbard at gmail.com
Fri Apr 4 17:29:20 PDT 2008


On Fri, Apr 4, 2008 at 2:58 PM, Vikram S. Adve <vadve at cs.uiuc.edu> wrote:
> On Apr 4, 2008, at 4:51 PM, Daniel Berlin wrote:
>
>  > On Fri, Apr 4, 2008 at 2:38 AM, Bill Wendling <isanbard at gmail.com>
>  > wrote:
>  >> On Apr 2, 2008, at 10:11 PM, Xuehai Qian wrote:
>  >>> Hi LLVMers,
>  >>>   I am a PHD student in CS dept in UIUC, I am doing a project for
>  >>> Vikram's course, it is about PRE. I would like to know why you
>  >>> didn't
>  >>> choose SSAPRE in LLVM, since it seems to be more suitable for LLVM
>  >>> (it
>  >>> can operate directly on SSA form and avoid the conversion between
>  >>> SSA
>  >>> and bit-vector). Can anyone tell me the reason?
>  >>
>  >> Hi Xuehai,
>  >>>
>  >>
>  >> If I remember correctly, there were several details that the paper
>  >> assumed that made adapting it to work in LLVM very difficult.
>  >
>  > It would at least require side-data structures to store the
>  > occurrences and expression phis.
>
>
>  Dan,
>
>  Doesn't the paper also assume the invariant that phi operands are
>  effectively dead after the Phi, which is true right after SSA is
>  constructed, but potentially not after transformations?
>
Yes, I think that that was the major problem with it.

-bw



More information about the llvm-dev mailing list