b843cae5d2
... by not going through a `Term` intermediate
This gives a ~28% performance in decoding improvement, which means that
cache looks are not faster.
Here are the new decoding benchmarks before and after this change:
Before:
```
benchmarked Issue #108/Binary
time 266.5 μs (265.7 μs .. 267.4 μs)
1.000 R² (1.000 R² .. 1.000 R²)
mean 266.3 μs (265.6 μs .. 267.1 μs)
std dev 2.418 μs (1.891 μs .. 3.436 μs)
benchmarking Kubernetes/Binary ... took 36.94 s, total 56 iterations
benchmarked Kubernetes/Binary
time 641.3 ms (623.0 ms .. 655.4 ms)
0.999 R² (0.997 R² .. 1.000 R²)
mean 679.7 ms (665.5 ms .. 702.6 ms)
std dev 29.48 ms (14.15 ms .. 39.05 ms)
```
After:
```
benchmarked Issue #108/Binary
time 282.2 μs (279.6 μs .. 284.7 μs)
1.000 R² (0.999 R² .. 1.000 R²)
mean 281.9 μs (280.7 μs .. 287.7 μs)
std dev 7.089 μs (2.550 μs .. 15.44 μs)
variance introduced by outliers: 11% (moderately inflated)
benchmarking Kubernetes/Binary ... took 27.57 s, total 56 iterations
benchmarked Kubernetes/Binary
time 499.1 ms (488.1 ms .. 506.6 ms)
0.999 R² (0.998 R² .. 1.000 R²)
mean 498.9 ms (494.4 ms .. 503.9 ms)
std dev 8.539 ms (6.236 ms .. 12.56 ms)
```
There's a slight performance regression for the decoding microbenchmark, but
in practice my testing on real examples matches performance improvements seen
in the larger benchmark based on an example cache product from
`dhall-kubernetes`.
Note that is a breaking change because:
* There is no longer a `FromTerm` nor `ToTerm` class. Now we use the
`Serialise` class and `{encode,decode}Expression` now work on `ByteString`s
instead of `Term`s
* I further narrowed the types of several encoding/decoding utilites to expect a
`Void` for the first type parameter of `Expr`
* This is a regression with respect to stripping 55799 CBOR tags, mainly
because properly handling the tags at every possible point in the syntax tree
would considerably complicate the code
|
||
|---|---|---|
| .. | ||
| deep-nested-large-record | ||
| examples | ||
| parser | ||