[11] RFR(M): 8205033: [REDO] Induction variable of over-unrolled loop conflicts with range checks

[Tobias Hartmann]

Hi,

since JDK-8203197 was just pushed, I had to merge my changes again:
http://cr.openjdk.java.net/~thartmann/8205033/webrev.01/

I will run some sanity testing and then push this.

Thanks,
Tobias


On 15.06.2018 14:07, Tobias Hartmann wrote:
> Hi,
> 
> my initial patch for JDK-8203915 [1] failed because the computation of the value of the loop
> induction variable at the end of the first iteration of the unrolled loop was wrong. As a result,
> the predicate was incorrectly found to always fail and we hit a HaltNode (SIGILL) in compiled code.
> 
> The problem is very easy to reproduce with the jdk/sun/security/tools/keytool/ tests, I therefore
> did not add an additional regression test. The problematic loop looks like this:
> 
> for (int i = 20; i > 0; i -= 4) {
>   array[i-1] = ...
>   array[i-2] = ...
>   array[i-3] = ...
>   array[i-4] = ...
> }
> 
> To make sure that the last array access is in bounds, we add a skeleton predicate of the form
> 
> max_i - 4 = (init + stride + 1) - 4 <u array.length
> 
> before the main loop and update the stride during unrolling. After unrolling 4 times, we know that
> init <= 16 (because the pre-loop is executed at least once) and the stride is 4*-4 = -16. The
> predicate is updated to:
> 
> ( (int,16] - 16 + 1) - 4 = (int,-3] <u array.length
> 
> which is always false. The computation of max_i is incorrect because the (fully) unrolled loop looks
> like this:
> 
>   // i <= 16
>   array[i-1] = ...
>   array[i-2] = ...
>   array[i-3] = ...
>   array[i-4] = ...
>   i -= 4;
>   array[i-1] = ...
>   array[i-2] = ...
>   array[i-3] = ...
>   array[i-4] = ...
>   i -= 4;
>   array[i-1] = ...
>   array[i-2] = ...
>   array[i-3] = ...
>   array[i-4] = ...
>   i -= 4;
>   array[i-1] = ...
>   array[i-2] = ...
>   array[i-3] = ...
>   array[i-4] = ...
>   // i <= 4
> 
> And therefore the value of i at the end of the iteration is max_i = init + stride - init_inc (where
> 'init_inc' is the initial loop increment). The correct predicate looks like this:
> 
> max_i - 4 = (init + stride - init_inc) - 4 = (16 - 16 + 4) -4  = 0 <u array.length
> 
> Incremental changes:
> http://cr.openjdk.java.net/~thartmann/8205033/webrev.inc/
> 
> Full webrev:
> http://cr.openjdk.java.net/~thartmann/8205033/webrev.00/
> 
> The failing tests pass. I'm now testing with hs-tier1-3, jdk-tier1-3 and hs-precheckin-comp. I will
> also rerun the benchmarks.
> 
> Thanks,
> Tobias
> 
> [1]
> https://bugs.openjdk.java.net/browse/JDK-8203915
> http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/2018-June/029250.html
>