[llvm-dev] [RFC] Polly Status and Integration

[Sebastian Pop via llvm-dev]

Hi Gerolf,

On Tue, Sep 12, 2017 at 10:26 PM, Gerolf Hoflehner via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> Are you saying the LLVM Dependence Analysis is incorrect or do you actually mean less conservative (or "more accurate" or something like that)?
>

Yes, the LLVM dependence analysis is broken from day one, by design,
due to a misunderstanding of the meaning of GEPs:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20130701/179509.html
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20130701/179529.html
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20130701/179570.html

Loop interchange and any other pass that relies on the current llvm
dependence analysis may generate wrong code.
See https://reviews.llvm.org/D35430

Another point, the MIV test in the llvm depednence analysis is not
implemented, and so the scope of the llvm dependence analysis is
rather narrow: i.e., it would not be able to solve the loop
interchange in spec2000/swim.

Sebastian