Request for reviews (S): 7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
vladimir.kozlov at oracle.com
Mon May 16 10:10:56 PDT 2011
After thinking more about pre_limit guard in do_range_check() "min(pre_limit,
orig_limit)" I decided to always generate it to be safe. The offset and scale in
RC condition could have such values that X=(low_limit-offset)/scale will be
outside of loop iterations range: max(pre_limit, X) could be > orig_limit.
This change does not invalidate previous testing but I redo it again.
Vladimir Kozlov wrote:
> Fixed 7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index
> out of range: 29488
> The main problem in new RCE code was in case when stride and scale have
> different sign. In that case new main_loop limit was calculated as
> min(old_main_limit, X) where X=(min_int+1 - (offset+1))/scale. With
> offset 0 and scale == -1, X value is min_int since min_int/-1 = min_int.
> So the new main loop limit is min_int. New unrolling code correctly
> handles such case by skipping main loop. Old code did ((limit
> -init)/stride) with stride == 1 and init == 1 (value after increment in
> preloop). So new limit for unrolled loop become (min_int-1) == max_int.
> I solve this by replacing positive (offset+1) with 0 and avoiding
> underflow in X calculation. It is the same approach as one used in
> current code for pre_limit in case scale and stride have the same sign.
> An other problem with new code is it does not generate the adjusted
> pre-loop limit when RangeLimitCheck is off. RCE old code did it before.
> I also factor out common code, replaced (stride*scale) with logical
> expressions to avoid overflow and I removed obsolete code which
> generates additional guard for pre_limit since it was added before
> adjust_limit() starts handle such case.
> Tested with/without UnrollLimitCheck CTW, JPRT, nsk.stress, 5091921 tests.
More information about the hotspot-compiler-dev