RFR: 8313672: C2: PhaseCCP does not correctly track analysis dependencies involving shift, convert, and mask [v3]

[Emanuel Peter]

On Tue, 31 Oct 2023 10:19:02 GMT, Daniel Lundén <duke at openjdk.org> wrote:

>> This changeset fixes an issue where the `PhaseCCP` analysis stopped before fixpoint due to `Value`-updates for `AndL` nodes involving shift, convert, and mask (`MulNode::AndIL_shift_and_mask_is_always_zero`). Such updates did not previously repopulate the `PhaseCCP` worklist correctly.
>> 
>> Changes:
>> - Fix the issue by adding the missing pattern to `PhaseCCP::push_and`.
>> - Update the `assert` error message that reported the issue to better describe the problem. The `PhaseCCP` analysis is optimistic, and the previous error message seemed to indicate the opposite ("Missed optimization opportunity"). If the assert triggers, it could be that the analysis result leads to unsound transformations later on.
>> 
>> Result after `PhaseCCP` analysis (without fix, note `long:0` for node 116):
>> ![before](https://github.com/openjdk/jdk/assets/4222397/3f2343da-8573-464d-9352-a147d0faeab8)
>> 
>> Result after `PhaseCCP` analysis (with fix, note `long` for node 116):
>> ![after](https://github.com/openjdk/jdk/assets/4222397/6bf44964-cff8-49db-9825-4a4667f4e769)
>> 
>> ### Testing (on all Oracle-supported platforms)
>> - `tier1`
>> - HotSpot parts of `tier2` and `tier3`
>
> Daniel Lundén has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Fix push_and to step through all ConvI2L children

src/hotspot/share/opto/phaseX.cpp line 1998:

> 1996:       }
> 1997:     };
> 1998:     ConstraintCastNode::visit_uncasted_uses(use, push_and_uses_to_worklist_skip_conv);

Maybe it would be nicer to generalize `visit_uncasted_uses`, and allow it to also peek through `ConvI2L` (optionally enabled with flag, false by default). Something like this:


        if (internal_use->is_ConstraintCast()) {
          internals.push(internal_use); // traverse this cast also
        } else if (peek_thru_convi2l && internal_use->Opcode() == Op_ConvI2L) {
          internals.push(internal_use); // traverse this cast also
        } else {
          callback(internal_use);
        }


Maybe you'd have to rename `visit_uncasted_uses` a bit. Or just create a common utility method and then call that one from `visit_uncasted_uses` with `peek_thru_convi2l = false`.

Maybe you have a better idea. But I have a sense that this may have to be done again in the future.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16429#discussion_r1377379218