[llvm-dev] Request suggestions about how to remove redundencies caused by SCEV expansion fundementally

[Wei Mi via llvm-dev]

On Wed, Aug 24, 2016 at 3:07 PM, Daniel Berlin <dberlin at dberlin.org> wrote:
>
>
> On Fri, Aug 19, 2016 at 3:57 PM, Wei Mi via llvm-dev
> <llvm-dev at lists.llvm.org> wrote:
>>
>> SCEV expansion sometimes generates redundent expr even if there is an
>> available expr which can be reused. The redundent exprs can be a lot
>> different from existing exprs so that existing cleanup passes cannot
>> remove them.
>>
>> https://llvm.org/bugs/show_bug.cgi?id=24920
>> https://llvm.org/bugs/show_bug.cgi?id=24442
>>
>> https://reviews.llvm.org/D12090 and https://reviews.llvm.org/D21313
>> already relieved the problem somewhat, but it is far from being solved
>> fundamentally. Recently we found another problematic case described in
>> https://llvm.org/bugs/show_bug.cgi?id=29065. And I believe I can
>> create more tests revealing similar problems.  In PR29065, I explained
>> why it is difficult to fix the problem by extending D12090 and D21313.
>> It is possible but still not very easy to fix the problem by enhancing
>> existing cleanup passes. So I am soliciting ideas here about how to
>> solve the problem in a more fundamental way.
>>
>> One thing I am always wondering is GCC also uses SCEV to calculate the
>> loop iteration number and why it doesn't have such problem. From my
>> limited knowledge about GCC's SCEV I guess it is because gcc only has
>> the basic function of SCEV which can represent loop recursive expr
>> like llvm SCEVAddRecExpr part, SCEV operands are still gimple
>> variables and it doesn't have such complex transformations on SCEV as
>> llvm does.
>
>
> This is not really correct :)
>
> In fact, at one point, GCC had *all of the possible scev* (sebastian worked
> a lot on it) ,including mixers, etc.
>
>
> The real thing here is that  GCC just doesn't do a lot of scev expansion. It
> mainly uses it as an analysis, not as a code generation mechanism.
>
>

You are right, Thanks! I just looked at loop iteration number
computation in GCC. It only uses SCEV to identify simple induction
variables, represent them as affine expr, and then use those affine
exprs and their value ranges in loop to compute loop iteration number,
so it doesn't involve SCEV expansion.

Wei.