Introduce parser memoization to avoid exponentional behavior #1799
Conversation
|
!bench |
|
Here are the benchmark results for commit cf3b001. Benchmark Metric Change
=====================================================
- stdlib compilation new 2% (49.1 σ)
- stdlib task-clock 1% (214.0 σ)
- workspaceSymbols task-clock 4% (12.3 σ)
- workspaceSymbols wall-clock 4% (12.4 σ) |
|
Instruction counts suggest these are random fluctuations. Which seem to have become more frequent recently. |
|
!bench |
|
Here are the benchmark results for commit 2ee1eb9. |
|
Comparison with previous benchmark: http://speedcenter.informatik.kit.edu/velcom/compare/6b556c5e-941a-48e3-8f57-5457ee362ab8/to/30d3b0d6-b0c2-411e-af98-c57c0f36db58 +14% parsing, though variance is high |
|
!bench |
|
Here are the benchmark results for commit 17dd678. Benchmark Metric Change
=====================================================
- stdlib compilation new 2% (54.4 σ)
- stdlib task-clock 2% (208.5 σ)
- workspaceSymbols task-clock 3% (10.0 σ) |
|
+71% parsing compared to previous run http://speedcenter.informatik.kit.edu/velcom/compare/30d3b0d6-b0c2-411e-af98-c57c0f36db58/to/33767806-2bc2-4029-ad5f-f09a55692675 |
|
!bench |
|
Here are the benchmark results for commit a74b707. Benchmark Metric Change
========================================
- liasolver task-clock 1% (10.8 σ)
- liasolver wall-clock 1% (11.0 σ)
- stdlib task-clock 2% (99.1 σ) |
Kha
force-pushed
the
packrat
branch
3 times, most recently
from
468b341 to
915bd1e
Compare
|
!bench |
|
Here are the benchmark results for commit 915bd1e. |
Kha
changed the title
We better hope the `leading_parser`s are closed terms
…ken table divergence
…ext constructor The context is now manipulated using `adaptCacheableContext` and `adaptUncacheableContext` and created using `ParserFn.run`.
|
!bench |
|
Here are the benchmark results for commit d0f2fe4. Benchmark Metric Change
=========================================
+ stdlib task-clock -2% (-32.2 σ)
+ stdlib wall-clock -2% (-68.0 σ) |
compiler.ir.result reports 0 allocations on happy path
| As this excludes trailing parsers from being cached, we also reset `lhsPrec`, which is not read but set by leading parsers, to 0 | ||
| in order to increase cache hits. Finally, `errorMsg` is also reset to `none` as a leading parser should not be called in the first | ||
| place if there was an error. |
There was a problem hiding this comment.
Sounds reasonable? Test suite says "yes". All ParserState fields are accounted for now.
There was a problem hiding this comment.
I think so. It is great you are documenting it.
Kha
changed the title
Cache invocations of category,
leading_parserandsyntax :=parsers using a mapping from parser name and current context & state to parse result. Abstract parser context and syntax stack to enforce correct use.