Array addition and array sum Panama benchmarks

[Vladimir Ivanov]

On 3/22/24 03:57, Roland Westrelin wrote:
> 
>> One solution would be for c2 to transform the long memory load + long to
>> to double move into double memory load (an Ideal transformation). The
>> code would then vectorize with no change to the vectorizer
>> required. That seems fairly straightforward as a change.
> 
> Actually, there's already support for that. Vladimir added it with
> 8253734 but the transformation is delayed until after loop optimizations
> so vectorization doesn't benefit from it. This:
> 
> diff --git a/src/hotspot/share/opto/movenode.cpp b/src/hotspot/share/opto/movenode.cpp
> index bfa30f02ada..75c647193a8 100644
> --- a/src/hotspot/share/opto/movenode.cpp
> +++ b/src/hotspot/share/opto/movenode.cpp
> @@ -374,7 +374,7 @@ Node* MoveNode::Ideal(PhaseGVN* phase, bool can_reshape) {
>       if (ld != nullptr && (ld->outcnt() == 1)) { // replace only
>         const Type* rt = bottom_type();
>         if (ld->has_reinterpret_variant(rt)) {
> -        if (phase->C->post_loop_opts_phase()) {
> +        if (phase->C->post_loop_opts_phase() || UseNewCode) {
>             return ld->convert_to_reinterpret_load(*phase, rt);
>           } else {
>             phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
> 
> and then running with -XX:+UnlockDiagnosticVMOptions -XX:+UseNewCode
> shows the code vectorizes now for scalarSegmentArray and performs much
> better.
> 
> I don't see a discussion of the reason for delaying the transformation
> in the PR for 8253734. Vladimir, can you comment on that?

I had concerns about mismatched accesses some of the transformations in 
8253734 introduce. (There's a comment in the description [1].)

Personally, I prefer to see vectorizer handling "MoveX2Y (LoadX mem)" 
=> "VectorReinterpret (LoadVector mem)" well and then introduce rules to 
strength-reduce it to mismatched access.

Best regards,
Vladimir Ivanov

[1] https://github.com/openjdk/jdk/pull/826