Disable pathologically expensive SimplifySelectOps optimization
#14
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
`SimplifySelectOps` is a late optimization in LLVM that attempts to translate `select(C, load(A), load(B))` into `load(select(C, A, B))`. However, in order for it to do this optimization, it needs to check that `C` does not depend on the result of `load(A)` or `load(B)`. Unfortunately (unlikely Julia and LLVM at the IR level), LLVM does not have a topological order of statements computed at this stage of the compiler, so LLVM needs to iterate through all statements in the function in order to perform this legality check. For large functions, this is extremely expensive, accounting for the majority of all compilation time for such functions. On the other hand, the optimization itself is minor, allowing at most the elision of one additional load (and doesn't fire particularly often, because the middle end can perform similar optimizations). Until there is a proper solution in LLVM, simply disable this optimizations, making LLVM several orders of magnitude faster on real world benchmarks. X-ref: llvm#60132
staticfloat
force-pushed
the
sf/llvm_sso_patch
branch
from
77b871f
to
2105735
Compare
|
Merged into 14.x as 5c82f53 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
SimplifySelectOpsis a late optimization in LLVM that attempts to translateselect(C, load(A), load(B))intoload(select(C, A, B)). However, in order for it to do this optimization, it needs to check thatCdoes not depend on the result ofload(A)orload(B). Unfortunately (unlikely Julia and LLVM at the IR level), LLVM does not have a topological order of statements computed at this stage of the compiler, so LLVM needs to iterate through all statements in the function in order to perform this legality check. For large functions, this is extremely expensive, accounting for the majority of all compilation time for such functions. On the other hand, the optimization itself is minor, allowing at most the elision of one additional load (and doesn't fire particularly often, because the middle end can perform similar optimizations). Until there is a proper solution in LLVM, simply disable this optimizations, making LLVM several orders of magnitude faster on real world benchmarks.