RFR: 8330274: C2 SuperWord: VPointer invar: same sum with different addition order should be equal

[Emanuel Peter]

On Tue, 16 Apr 2024 12:00:51 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

> This is an enhancement for AutoVectorization.
> 
> I want to improve the detection of `invar`s that are equivalent (guaranteed to compute the same value), but don't have the identical node (the computation is in a different order).
> 
> Note: only about 100 lines are real changes, the rest is tests. These are the first tests that check vectorization for MemorySegments.
> 
> **Solution Sketch: "canonicalize" the invar**
> 
> - Extract all summands of the `invar`: make a list.
>   - Parse through `AddL`, `SubL`, `AddI`, `SubI`, to get summands.
>   - Bypass `CastLL` and `CastII`
>   - Recursively treat `ConvI2L`, `LShiftI` and `LShiftL`: i.e. canonicalize their input.
> 
> - Sort all extracted summands by node idx.
> - Add up all summands in new order.
> 
> If two `invar`s use the same summands, then we know that after canonicalization the new nodes representing the `invar`s must be the same.
> 
> **Example**
> 
> 
> invar1 = b + c + d + a
> invar2 = d + b + a + c
> 
> -> equivalent but not identical nodes
> 
> Sort, and add up again:
> 
> invar1 = a + b + c + d
> invar2 = a + b + c + d
> 
> -> now the nodes are identical
> 
> **Motivation: MemorySegment with invar**
> 
> One might think that this is a big of a special case: why would anybody write indices to an Array or MemorySegment where the invar has a different addition order for its summands?
> 
> This example did not vectorize, even though it should:
> https://github.com/openjdk/jdk/blob/78e42d6e311c33548d16c6c74493388d9850238e/test/hotspot/jtreg/compiler/loopopts/superword/TestEquivalentInvariants.java#L425-L441
> 
> Both the `get` and the `set` look like they have the same address, and the address increases by a byte in each iteration.
> 
> Upon inspection, I saw that the `invar` that `VPointer` produces for the two operations are not identical: the order of addition of the `invar`'s summands is different, and thus the `invar` nodes are different.
> 
> The consequence: Only if we can prove that the two `invar` are identical can we know that the addresses are identical, and that there is no aliasing for loop carried dependencies. Since we have different `invar`, we don't know how the two addresses alias, and that prevents vectorization.
> 
> Why does this happen? After parsing, the graph looks like this:
> ![image](https://github.com/openjdk/jdk/assets/32593061/f768d0b0-0b2f-48f0-bfdc-61e93e62bb4f)
> 
> We already see that the two addresses are different only by a `CastLL`, with type `long:>=0`. Somehow, that was only deduced for the load, and not the store.
> 
> load_adr = base + memory_segment_offs...

src/hotspot/share/opto/vectorization.cpp line 101:

> 99:       tty->print("  lpt->_head %d", _cl->_idx); _cl->dump();
> 100:       _lpt->dump_head();
> 101:       _cl->dump_bfs(100, _cl_exit, "c-");

Note: Simply makes it easier to see what kind of CFG is between the loop head and end. Is it a `RangeCheck`, or some `If`?

src/hotspot/share/opto/vectorization.cpp line 503:

> 501: #ifdef ASSERT
> 502:   // We are changing the invar, and the debug info may no longer be accurate.
> 503:   if (new_invar != _invar) { _debug_invar = NodeSentinel; }

Note: Roland had inserted this `_debug_invar` verification code a year ago. Putting `NodeSentinel` basically just disables the verifiation, he uses that also elsewhere already.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/18795#discussion_r1568819915
PR Review Comment: https://git.openjdk.org/jdk/pull/18795#discussion_r1568821470