RFR: 8333393: PhaseCFG::insert_anti_dependences can fail to raise LCAs and to add necessary anti-dependence edges [v2]

[Emanuel Peter]

On Tue, 7 Jan 2025 18:03:21 GMT, Daniel Lundén <dlunden at openjdk.org> wrote:

>> When searching for load anti dependences in GCM, it is not always sufficient to just search starting at the direct initial memory input to the load. Specifically, there are cases when we must also search for anti dependences starting at relevant Phi memory nodes in between the load's early block and the initial memory input's block. Here, "in between" refers to blocks in the dominator tree in between the early and initial memory blocks.
>> 
>> #### Example 1
>> 
>> Consider the ideal graph below. The initial memory for 183 loadI is 107 Phi and there is an important anti dependency for node 64 membar_release. To discover this anti dependency, we must rather search from 119 Phi which contains overlapping memory slices with 107 Phi. Looking at the ideal graph block view, we see that both 107 Phi and 119 Phi are in the initial memory block (B7) and thus dominate the early block (B20). If we only search from 107 Phi, we fail to add the anti dependency to 64 membar_release and do not force the load to schedule before 64 membar_release as we should. In the block view, we see that the load is actually scheduled in B24 _after_ a number of anti-dependent stores, the first of which is in block B20 (corresponding to the anti dependency on 64 membar_release). The result is the failure we see in this issue (we load the wrong value).
>> 
>> ![failure-graph-1](https://github.com/user-attachments/assets/e5458646-7a5c-40e1-b1d8-e3f101e29b73)
>> ![failure-blocks-1](https://github.com/user-attachments/assets/a0b1f724-0809-4b2f-9feb-93e9c59a5d6a)
>> 
>> #### Example 2
>> 
>> There are also situations when we need to start searching from Phis that are strictly in between the initial memory block and early block. Consider the ideal graph below. The initial memory for 100 loadI is 18 MachProj, but we also need to search from 76 Phi to find that we must raise the LCA to the last block on the path between 76 Phi and 75 Phi: B9 (= the load's early block). If we do not search from 76 Phi, the load is again likely scheduled too late (in B11 in the example) after anti-dependent stores (the first of which corresponds to 58 membar_release in B10). Note that the block B6 for 76 Phi is strictly dominated by the initial memory block B2 and also strictly dominates the early block B9.
>> 
>> ![failure-graph-2](https://github.com/user-attachments/assets/ede0c299-6251-4ff8-8b84-af40a1ee9e8c)
>> ![failure-blocks-2](https://github.com/user-attachments/assets/e5a87e43-b6fe-4fa3-8961-54752f63633e)
>> 
>> ### Cha...
>
> Daniel Lundén has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Updates after comments

src/hotspot/share/opto/gcm.cpp line 757:

> 755:   // In some cases, there are other relevant initial memory states besides
> 756:   // initial_mem. In such cases, we are rather dealing with multiple trees and
> 757:   // their fringes.

If I look at these comments here (I reviewed a change by Roland a few months back, so my memory is coming back)...
I see that the load is supposed to be scheduled before any `Memory state modifying nodes include Store and Phi` that is (transitively via any MergeMem) below the `initial_mem`.

In you example1, why do we therefore not put an anti-dependency edge betweeen the `183 load`, and the `106 Phi`? Would that not be enough to ensure the load is scheduled before the other memory affecting nodes further below `106 Phi`?

Or is the issue that this traversal is somehow restricted to blocks - I don't remember that from last time...
I'll keep reading the changes now.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/22852#discussion_r1907230672