[jdk17] RFR: 8269230: C2: main loop in micro benchmark never executed

[Vladimir Kozlov]

On Mon, 28 Jun 2021 08:20:17 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> The benchmark loop is roughly:
> 
> for (int i = 0; i < limit; i++) {
>   if (!(i < max_jint && i > min_jint)) {
>     uncommon_trap();
>   }
> }
> 
> which IfNode::fold_compares() tranforms to:
> 
> for (int i = 0; i < limit; i++) {
>   if (!(i - min_jint - 1 <u -2)) {
>     uncommon_trap();
>   }
> }
> 
> Range elimination then adjusts loop limits so the unsigned comparison
> can be removed. The new limit is:
> 
> min(limit, max(-2 + min_jint + 1, min_jint)) = min(limit, min_jint) = min_jint
> 
> and the main loop is never executed.
> 
> The logic that protects against overflows during range check
> elimination is what causes the main loop to never be executed. As a
> fix, I propose delaying IfNode::fold_compares() if the resulting test
> compares with a potential negative number which could then cause RC
> overflow code to trigger. I think that would preserve the most
> interesting cases for IfNode::fold_compares() when it comes to range
> check elimination:
> 
> i >= 0 && i < array.length

I can't approve this for JDK 17. The issue is performance for very corner case but suggested fix may affect other cases. And this is not a bug IMHO - it is RFE which we should spend more time to investigate. I would suggest to defer it to JDK 18.

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

Changes requested by kvn (Reviewer).

PR: https://git.openjdk.java.net/jdk17/pull/156