[11] RFR(XS) 8202948: C2: assert(init_offset >= 0) failed: positive offset from object start

[Tobias Hartmann]

Hi Vladimir,

looks good to me but shouldn't we add the test (looks like you came up with a very simple version)?

Thanks,
Tobias

On 13.06.2018 19:14, Vladimir Kozlov wrote:
> https://bugs.openjdk.java.net/browse/JDK-8202948
> 
> Predicates and Range checks before main loop will not help since they are runtime checks and limit
> is not constant. The problem happens during compilation because vectorization is attempted for
> over-unrolling loop which will not be executed in runtime (dead loop). Stride is negative (-3) value
> and initial is 2 - loop will be executed at least once (or never) because of other conditions in
> code. C2 did not recognize this because the test fails only with -Xcomp when profiling data is no
> available (it passed without -Xcomp).
> 
> Simplest fix is to convert the assert into compilation check which will skip superword optimization
> in such case.
> 
> diff -r 0d47e89382ed src/hotspot/share/opto/superword.cpp
> --- a/src/hotspot/share/opto/superword.cpp
> +++ b/src/hotspot/share/opto/superword.cpp
> @@ -887,7 +887,9 @@
>    if (init_nd->is_Con() && p.invar() == NULL) {
>      int init = init_nd->bottom_type()->is_int()->get_con();
>      int init_offset = init * p.scale_in_bytes() + offset;
> -    assert(init_offset >= 0, "positive offset from object start");
> +    if (init_offset < 0) { // negative offset from object start?
> +      return false;        // may happen in dead loop
> +    }
>      if (vw % span == 0) {
>        // If vm is a multiple of span, we use formula (1).
>        if (span > 0) {
> 
> Test passed with and without Predicates with this fix.
> 
> Tested with tier1, tier2, precheckin-xcomp
>