Analyze and improve the performance of AliasAnalysis compiler pass #9235
Comments
|
Unless there is some low hanging fruit we might use this opportunity to rewrite
|
`scopeFor` appears to be a hotspot of the compiler. By choosing a more suitable data structure that indexes on the occurrence's id we seem to gain about 25% on some benchmarks. Quick win. Related to #9235. # Important Notes Local benchmark runs of `org.enso.compiler.benchmarks.module.ManyLocalVarsBenchmark.longMethodWithLotOfLocalVars ` Before ``` [info] # Warmup Iteration 1: 61.638 ms/op [info] # Warmup Iteration 2: 49.224 ms/op [info] # Warmup Iteration 3: 47.341 ms/op [info] # Warmup Iteration 4: 46.946 ms/op [info] # Warmup Iteration 5: 46.901 ms/op [info] # Warmup Iteration 6: 49.536 ms/op [info] Iteration 1: 50.438 ms/op [info] Iteration 2: 47.326 ms/op [info] Iteration 3: 46.917 ms/op [info] Iteration 4: 45.824 ms/op ``` After ``` [info] # Warmup Iteration 1: 86.493 ms/op [info] # Warmup Iteration 2: 36.084 ms/op [info] # Warmup Iteration 3: 32.588 ms/op [info] # Warmup Iteration 4: 33.895 ms/op [info] # Warmup Iteration 5: 31.986 ms/op [info] # Warmup Iteration 6: 31.236 ms/op [info] Iteration 1: 31.258 ms/op [info] Iteration 2: 31.673 ms/op [info] Iteration 3: 30.931 ms/op [info] Iteration 4: 30.902 ms/op ```
|
This likely seems obsolete. I'm not seeing much time spent in AliasAnalysis phase anymore. |
@hubertp You mean you are not seeing much time spent in |
Related: The |
|
@JaroslavTulach This is expected in this benchmark. This benchmark compiles a method that has a lot of local variables and so, we expect the alias analysis to take a lot of time. That is fine - this is an artificial benchmark that tries to measure the performance of AliasAnalysis. The important question is: "In an average user workflow project, how much time is spent in AliasAnalysis?" To answer the question myself, I would say that this issue is important. As the |
Compiler benchmarks were added in #9158.
The most appropriate benchmark for measuring alias analysis compiler pass is
org.enso.compiler.benchmarks.module.ManyLocalVarsBenchmark.longMethodWithLotOfLocalVars. Run it with:Note that you can fine-tune benchmark-related options with, for example,
(This increases the number of warmup iterations).
See https://github.com/enso-org/enso/blob/develop/docs/infrastructure/benchmarks.md#running-the-benchmarks
The following picture from VisualVM CPU profiler shows that the vast majority of time is spent in AliasAnalysis:
The text was updated successfully, but these errors were encountered: