RFR: 8252372: Check if cloning is required to move loads out of loops in PhaseIdealLoop::split_if_with_blocks_post()

[Roland Westrelin]

On Mon, 26 Apr 2021 09:46:18 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> Sinking data nodes out of a loop when all uses are out of a loop has
> several issues that this attempts to fix.
> 
> 1- Only non control uses are considered which makes little sense (why
>   not sink if the data node is an argument to a call or a returned
>   value?)
> 
> 2- Sinking of Loads is broken because of the handling of
>   anti-dependence: the get_late_ctrl(n, n_ctrl) call returns a control
>   in the loop because it takes all uses into account.
> 
> 3- For data nodes for which a control edge can't be set, commoning of
>   clones back in the loop is prevented with:
>   _igvn._worklist.yank(x);
>   which gives no guarantee
> 
> This patch tries to address all issues:
> 
> 1- it looks at all uses, not only non control uses
> 
> 2- anti-dependences are computed for each use independently
> 
> 3- Cast nodes are used to pin clones out of loop
> 
> 
> 2- requires refactoring of the PhaseIdealLoop::get_late_ctrl()
> logic. While working on this, I noticed a bug in anti-dependence
> analysis: when the use is a cfg node, the code sometimes looks at uses
> of the memory state of the cfg. The logic uses the use of the cfg
> which is a projection of adr_type identical to the cfg. It should
> instead look at the use of the memory projection.
> 
> The existing logic for sinking loads calls clear_dom_lca_tags() for
> every load which seems like quite a waste. I added a
> _dom_lca_tags_round variable that's or'ed with the tag_node's _idx. By
> incrementing _dom_lca_tags_round, new tags that don't conflict with
> existing ones are produced and there's no need for
> clear_dom_lca_tags().
> 
> For anti-dependence analysis to return a correct result, early control
> of the load is needed. The only way to get it at this stage, AFAICT,
> is to compute it by following the load's input until a pinned node is
> reached.
> 
> The existing logic pins cloned nodes next to their use. The logic I
> propose pins them right out of the loop. This could possibly avoid
> some redundant clones. It also makes some special handling for corner
> cases with loop strip mining useless.
> 
> For 3-, I added extra Cast nodes for float types. If a chain of data
> nodes are sunk, the new logic tries to keep a single Cast for the
> entire chain rather than one Cast per node.

That issue should be fixed. I added a CastVV node for vectors.

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

PR: https://git.openjdk.java.net/jdk/pull/3689