[llvm-dev] LLVM SCEV isAddRecNeverPoison and strength reduction
Sanjoy Das via llvm-dev
llvm-dev at lists.llvm.org
Wed May 9 20:03:38 PDT 2018
+CC llvm-dev
On Tue, May 8, 2018 at 2:34 AM, Gal Zohar <Gal.Zohar at ceva-dsp.com> wrote:
> I noticed that SCEV, when trying to perform strength reduction, doesn’t use
> the ability to prove an induction variable does not signed/unsigned wrap due
> to infinite loops.
>
> Is there an easy way to use the isAddRecNeverPoison function when
> determining if strength reduction is possible? In getZeroExtendExpr.
>
> Is there a reason why this doesn’t happen?
I guess your point is that in
int foo(int a) {
int sum = 0;
for (short i = 0; i < a; i++) {
sum++;
}
return sum;
}
either the loop is finite (and i <= SHORT_MAX) or the program has UB
(since sum overflows), so we can assume i<=SHORT_MAX and compute the
trip count accordingly?
In LLVM the fix isn't as simple unfortunately because signed integer
overflow is not UB, but it produces a "poison value" that causes UB
(roughly) if consumed by some side effecting operation.
It should still be possible to do this optimization -- the return
value is either poison or i <= SHORT_MAX. Because it is legal to
replace poison with whatever value we want, we can just pretend i <=
SHORT_MAX, compute the exit value under that assumption and delete the
loop. However, I suspect this will be a fair amount of work.
Thanks!
-- Sanjoy
PS: this is the original email for llvm-dev:
I noticed that SCEV, when trying to perform strength reduction,
doesn’t use the ability to prove an induction variable does not
signed/unsigned wrap due to infinite loops.
Is there an easy way to use the isAddRecNeverPoison function when
determining if strength reduction is possible? In getZeroExtendExpr.
Is there a reason why this doesn’t happen?
This simple example is not optimized due to this:
int foo(int a)
{
int sum = 0;
for (short i = 0; i < a; i++)
{
sum++;
}
return sum;
}
Thanks,
More information about the llvm-dev
mailing list