RFR: 8298935: fix cyclic dependency bug in create_pack logic in SuperWord::find_adjacent_refs [v22]

Wed Mar 8 15:24:41 UTC 2023

The message from this sender included one or more files
which could not be scanned for virus detection; do not
open these files unless you are certain of the sender's intent.

----------------------------------------------------------------------
On Wed, 8 Mar 2023 10:33:03 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> Cyclic dependencies are not handled correctly in all cases. Three examples:
>> 
>> https://github.com/openjdk/jdk/blob/0a834cd991a2f94b784ee4abde06825486fcb97f/test/hotspot/jtreg/compiler/loopopts/superword/TestCyclicDependency.java#L270-L277
>> 
>> And this, compiled with `-XX:CompileCommand=option,compiler.vectorization.TestOptionVectorizeIR::test*,Vectorize`:
>> https://github.com/openjdk/jdk/blob/0a834cd991a2f94b784ee4abde06825486fcb97f/test/hotspot/jtreg/compiler/vectorization/TestOptionVectorizeIR.java#L173-L180
>> 
>> And for `vmIntrinsics::_forEachRemaining` compile option `Vectorize` is always enabled:
>> https://github.com/openjdk/jdk/blob/0a834cd991a2f94b784ee4abde06825486fcb97f/test/hotspot/jtreg/compiler/vectorization/TestForEachRem.java#L69-L73
>> 
>> All of these examples are vectorized, despite the cyclic dependency of distance 2. The cyclic dependency is dropped, instead the emitted vector code implements a shift by 2, instead of repeating the same 2 values.
>> 
>> **Analysis**
>> 
>> The `create_pack` logic in `SuperWord::find_adjacent_refs` is broken in two ways:
>> 
>> - When the compile directive `Vectorize` is on, or we compile `vmIntrinsics::_forEachRemaining` we have `_do_vector_loop == true`. When that is the case, we blindly trust that there is no cyclic dependency larger than distance 1. Distance 1 would already be detected by the `independence(s1, s2)` checks we do for all adjacent memops. But for larger distances, we rely on `memory_alignment == 0`. But the compile directive avoids these checks.
>> - If `best_align_to_mem_ref` is of a different type, and we have `memory_alignment(mem_ref, best_align_to_mem_ref) == 0`, we do not check if `mem_ref` has `memory_alignment == 0` for all other refs of the same type. In the example `TestCyclicDependency::test2`, we have `best_align_to_mem_ref` as the `StoreF`. Then we assess the `StoreI`, which is not aligned with it, but it is of a different type, so we accept it too. Finally, we look at `LoadI`, which has perfect alignment with the `StoreF`, so we accept it too (even though it is in conflict with the `StoreI`).
>> 
>> Generally, the nested if-statements are confusing and buggy. I propose to fix and refactor the code.
>> 
>> I also propose to only allow the compile directive `Vectorize` only if `vectors_should_be_aligned() == false`. If all vector operations have to be `vector_width` aligned, then they also have to be mutually aligned, and we cannot have patterns like `v[i] = v[i] + v[i+1]` for which the compile directive was introduced in the first place https://github.com/openjdk/jdk/commit/c7d33de202203b6da544f2e0f9a13952381b32dd.
>> **Update**: I found a **Test.java** that lead to a crash (`SIGBUS`) on a ARM32 on master. The example bypassed the alignment requirement because of `_do_vector_loop`, and allowed unaligned vector loads to be generated, on a platform that requires alignment. Thanks @fg1417 for running that test for me!
>> 
>> **Solution**
>> 
>> First, I implemented `SuperWord::verify_packs` which catches cyclic dependencies just before scheduling. The idea is to reassess every pack, and check if all memops in it are mutually independent. Turns out that per vector pack, it suffices to do a single BFS over the nodes in the block (see `SuperWord::find_dependence`). With this verification in place we at least get an assert instead of wrong execution.
>> 
>> I then refactored and fixed the `create_pack` code, and put the logic all in `SuperWord::is_mem_ref_alignment_ok`. With the added comments, I hope the logic is more straight forward and readable. If `_do_vector_loop == true`, then I filter the vector packs again in `SuperWord::combine_packs`, since we are at that point not sure that the packs are actually independent, we only know that adjacient memops are independent.
>> 
>> Another change I have made:
>> Disallow `extend_packlist` from adding `MemNodes` back in. Because if we have rejected some memops, we do not want them to be added back in later.
>> 
>> **Testing**
>> 
>> I added a few more regression tests, and am running tier1-3, plus some stress testing.
>> 
>> However, I need help from someone who can test this on **ARM32** and **PPC**, basically machines that have `vectors_should_be_aligned() == true`. I would love to have additional testing on those machine, and some reviews.
>> **Update:** @fg1417 did testing on ARM32, @reinrich did testing on PPC.
>> 
>> **Discussion / Future Work**
>> 
>> I wonder if we should have `_do_vector_loop == true` by default, since it allows more vectorization. With the added filtering, we are sure that we do not schedule packs with cyclic dependencies. We would have to evaluate performance and other side-effects of course. What do you think? [JDK-8303113](https://bugs.openjdk.org/browse/JDK-8303113)
>
> Emanuel Peter has updated the pull request incrementally with one additional commit since the last revision:
> 
>   TestDependencyOffsets.java: add vanilla run

Since this thread is so long, I summarize what I have done again:

**Bug 1: bad nested if**
In pseudo-code we did basically this in `find_adjacent_refs`, when checking that we have alignment with all other packs of the same `velt_type / memory_slice`:

if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
    // go ahead and vectorize!
    // PROBLEM: what if "best" was from a different velt_type / memory slice?
    //          We may have alignment with "best", but we ignore misalignment with other packs!
} else {
    if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
        // misaligned to same velt_type / memory_slice -> no vectorization
    } else {
        // for all other packs with the same velt_type / memory_slice, check if we have alignment -> only then vectorize
    }
}

Fix (swap if conditions, refactor):
https://github.com/openjdk/jdk/blob/a44082b61f22dcdee115697f34d39c1d8382a15d/src/hotspot/share/opto/superword.cpp#L811-L815

**Bug 2: rejected memops should not be resurrected**
If memops were rejected during `find_adjacent_refs`, we sometimes resurrected them again during `extend_packlist`. Memops may have been rejected because they had a misalignment (which could imply dependence). So we should not blindly resurrect such memops.

Fix (only extend to non-memop): 
https://github.com/openjdk/jdk/blob/a44082b61f22dcdee115697f34d39c1d8382a15d/src/hotspot/share/opto/superword.cpp#L1562

However, there is a downside to this fix: there seemed to have been many "happy accidents", where resurrecting memops lead to vectorization that was correct.

For example under `-XX:+AlignVector`, we require all `mem_refs` to align with `best`, and have a `vector_width` at most as large as that of `best`. This is to ensure that all `mem_refs` are aligned to memory, modulo their `vector_width` (once we memory align `best` to memory modulo its `vector_width` by adjusting the pre-loop limit).

    @Test
    @IR(counts = {IRNode.LOAD_VECTOR, ">0",
                  IRNode.VECTOR_CAST_I2X, ">0",
                  IRNode.STORE_VECTOR, ">0"})
    private static void testConvI2D(double[] d, int[] a) {
        for(int i = 0; i < d.length; i++) {
            d[i] = (double) (a[i]);
        }
    }

In `TestVectorizeTypeConversion.testConvI2D`, `best_align_to_mem_ref` is `StoreD`, which on some machines may have a smaller `vector_width` than the `LoadI`, and hence gets rejected. This would prevent vectorization. The "happy accident" was that it was resurrected during `extend_packlist`. This still leads to correct results "by accident", since most machines at most require 8-byte alignment, and not `vector_width` alignment.

This performance regression can be fixed by a follow-up RFE. For now, we should prefer correctness over performance.

**Bug 3: _do_vector_loop should not ignore dependencies**

TODO write

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

PR: https://git.openjdk.org/jdk/pull/12350