RFR: JDK-8289748: C2 compiled code crashes with SIGFPE with -XX:+StressLCM and -XX:+StressGCM [v3]

[Tobias Hartmann]

On Thu, 19 Jan 2023 08:39:03 GMT, Tobias Holenstein <tholenstein at openjdk.org> wrote:

>> # Problem 
>> 
>> We have an outer strip-mined loop `207`/`205` with inner strip-mined loop `283`/`203`. `262` is the zero-trip guard and `157 ModI` has the trip count `284 Phi` in the divisor. The corresponding zero guard of `157 ModI` was removed because the divisor is guaranteed to always be non-zero.
>> 
>> `PhaseIdealLoop::cast_incr_before_loop(..)` inserts a CastII to pin `284 Phi` to `262 if` but the following `igvn` phase removes the CastII again because it becomes a constant. Therefore, `284 Phi` loses its pin to `262 if`.
>> ![fail before](https://user-images.githubusercontent.com/71546117/212355062-feaa78a9-b1f4-4bf6-b70b-03a134b2de9d.png)
>> 
>> `IdealLoopTree::do_remove_empty_loop(..)` removes the loop `283`/`203` and replaces `284 Phi` with the final value of the last iteration. The problem is that the computation of final value (exact_limit - stride) floats above `262 if` - so now `157 ModI` is not guarded by `262 if` anymore: in our attached test case we get an `SIGFPE` because the modulo divides by zero at runtime. 
>> 
>> ![fail after](https://user-images.githubusercontent.com/71546117/212355115-e8015569-0d1a-450e-bd09-40e1f8999bdd.png)
>> 
>> # Solution
>> The solution is to insert a `CastII` in `IdealLoopTree::do_remove_empty_loop(..)` to prevent final_iv (exact_limit - stride) to float above the zero trip guard. 
>> 
>> ![fix after](https://user-images.githubusercontent.com/71546117/212355168-a0a86420-dbcc-48c7-9466-a6bf2db8b13b.png)
>
> Tobias Holenstein has updated the pull request incrementally with one additional commit since the last revision:
> 
>   update comment

Marked as reviewed by thartmann (Reviewer).

-------------

PR: https://git.openjdk.org/jdk/pull/11990