[Loop Vectorizer] Loop not vectorized when accessing the same pointer with different offsets. #87336
Comments
dtcxzyw
added
missed-optimization
vectorization
llvm:analysis
and removed
new issue
labels
|
gcc support this: https://godbolt.org/z/zGjhha5qx |
|
Looks like there are a couple of improvements needed, looking now |
fhahn
added a commit
that referenced
this issue
Apr 8, 2024
Add a number of LAA test cases with both forward and backward dependences with non-constant strides and dependence distances. This includes test coverage for #87336 Also includes a LoopLoadElimination test to make sure the pass does not crash on non-constant dependence distances.
fhahn
added a commit
to fhahn/llvm-project
that referenced
this issue
Apr 8, 2024
Extend LoopAccessAnalysis to support different strides and as a consequence non-constant distances between dependences using SCEV to reason about the direction of the dependence. In multiple places, logic to rule out dependences using the stride has been updated to only be used if StrideA == StrideB, i.e. there's a common stride. We now also may bail out at multiple places where we may have to set FoundNonConstantDistanceDependence. This is done when we need to bail out and the distance is not constant to preserve original behavior. I'd like to call out the changes in global_alias.ll in particular. In the modified mayAlias01, and mayAlias02 tests, there should be no aliasing in the original versions of the test, as they are accessing 2 different 100 element fields in a loop with 100 iterations. I moved the original tests to noAlias15 and noAlias16 respectively, while updating the original tests to use a variable trip count. In some cases, like different_non_constant_strides_known_backward_min_distance_3, we now vectorize with runtime checks, even though the runtime checks will always be false. I'll also share a follow-up patch, that also uses SCEV to more accurately identify backwards dependences with non-constant distances. Fixes llvm#87336
fhahn
added a commit
that referenced
this issue
Apr 25, 2024
…EV. (#88039) Extend LoopAccessAnalysis to support different strides and as a consequence non-constant distances between dependences using SCEV to reason about the direction of the dependence. In multiple places, logic to rule out dependences using the stride has been updated to only be used if StrideA == StrideB, i.e. there's a common stride. We now also may bail out at multiple places where we may have to set FoundNonConstantDistanceDependence. This is done when we need to bail out and the distance is not constant to preserve original behavior. Fixes #87336 PR: #88039
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
See reproducer in https://godbolt.org/z/v7fqorb1h.
a2 = a1 + 256;a2has an interleaved store access, anda1has a simple linear access.There are no actual dependencies between them when VF is 4, and it looks safe for the loop to be vectorized.
BTW, changing
a2[2*i]toa2[i]makes vectorization safe.The text was updated successfully, but these errors were encountered: