Cache most calls to isRelatedTo

[weswigham]

This lifts the cache used within recursiveTypeRelatedTo up to isRelatedTo as a whole; enabling the results of comparing unions and intersections of types to be cached.

As a side effect of caching the results, this causes us to no longer duplicate elaborations for the newly cached types across multiple diagnostic messages, similarly to what we were already doing for structured types.

Combined with #17947, @sandersn and I saw significant improvements in the compilation time of some type-heavy compilations.

[ahejlsberg]

I recall that we previously rejected caching union and intersection comparison results because of the increased memory consumption and increased GC churn (from allocating key strings for cache lookups). We definitely need some numbers to determine that there's no negative impact on large codebases that otherwise aren't affected by the problem we're trying to solve.

[weswigham]

Metric	master	cached-relationships	Delta	Best	Worst
Monaco - node (v8.3.0, x64)
Memory used	365,086k (± 0.02%)	369,672k (± 0.01%)	+4,586k (+ 1.26%)	369,613k	369,742k
Parse Time	1.84s (± 2.81%)	1.80s (± 0.42%)	-0.04s (- 2.29%)	1.78s	1.81s
Bind Time	0.75s (± 2.26%)	0.74s (± 1.36%)	-0.01s (- 1.73%)	0.72s	0.77s
Check Time	3.71s (± 1.19%)	3.71s (± 0.38%)	+0.00s (+ 0.08%)	3.68s	3.74s
Emit Time	2.39s (± 1.23%)	2.38s (± 1.49%)	-0.02s (- 0.63%)	2.34s	2.51s
Total Time	8.69s (± 1.23%)	8.62s (± 0.47%)	-0.07s (- 0.77%)	8.55s	8.73s
TFS - node (v8.3.0, x64)
Memory used	311,718k (± 0.02%)	314,294k (± 0.02%)	+2,577k (+ 0.83%)	314,192k	314,466k
Parse Time	1.28s (± 1.75%)	1.27s (± 0.54%)	-0.01s (- 1.17%)	1.25s	1.28s
Bind Time	0.77s (± 1.52%)	0.75s (± 0.91%)	-0.01s (- 1.82%)	0.74s	0.77s
Check Time	3.24s (± 1.77%)	3.26s (± 1.03%)	+0.02s (+ 0.59%)	3.14s	3.32s
Emit Time	2.18s (± 2.03%)	2.13s (± 1.44%)	-0.05s (- 2.38%)	2.09s	2.24s
Total Time	7.48s (± 0.87%)	7.42s (± 0.68%)	-0.06s (- 0.82%)	7.25s	7.52s

Perf-wise, there's no difference on our perf projects. Memory-wise, it uses ~1% more memory.

[weswigham]

This change cuts 27s off the wall-clock time that jake local on PR #18285 takes on my machine - I think the goodness of this change is heavily magnified by the presence and usage of discriminated unions and control flow analysis.

[ahejlsberg]

@weswigham What is the corresponding change in memory consumption?

[weswigham]

Perf data for master@ 1b5a0ae vs this branch compiling the version of our compiler which uses unions in tsperf:

Metric	master	cached-relationships	Delta	Best	Worst
Compiler - Unions - node (v8.3.0, x64)
Memory used	212,209k (± 0.02%)	232,950k (± 3.36%)	+20,741k (+ 9.77%)	229,356k	264,566k
Parse Time	0.77s (± 1.72%)	0.81s (± 3.10%)	+0.04s (+ 4.79%)	0.78s	0.88s
Bind Time	0.65s (± 1.91%)	0.65s (± 1.26%)	-0.00s (- 0.46%)	0.63s	0.66s
Check Time	9.65s (± 1.61%)	8.61s (± 1.44%)	-1.04s (- 10.74%)	8.34s	8.89s
Emit Time	1.91s (± 1.21%)	2.00s (± 3.01%)	+0.09s (+ 4.50%)	1.93s	2.22s
Total Time	12.98s (± 1.28%)	12.07s (± 1.28%)	-0.91s (- 7.05%)	11.69s	12.38s
Compiler - Unions - tsc (x86)
Parse Time	0.53s (± 2.71%)	0.55s (± 6.46%)	+0.02s (+ 4.14%)	0.52s	0.68s
Bind Time	0.21s (± 4.16%)	0.23s (± 8.20%)	+0.02s (+ 8.41%)	0.20s	0.29s
Check Time	13.07s (± 1.91%)	12.22s (± 1.70%)	-0.85s (- 6.50%)	11.75s	12.65s
Emit Time	1.61s (± 2.04%)	1.67s (± 3.32%)	+0.06s (+ 3.66%)	1.60s	1.88s
Total Time	15.42s (± 1.82%)	14.67s (± 1.82%)	-0.75s (- 4.89%)	14.09s	15.25s

Significant improvements to check time in exchange for higher memory usage, as expected.

[weswigham]

@ahejlsberg perf data (including memory) collected.

[typescript-bot]

Thanks for your contribution. This PR has not been updated in a while and cannot be automatically merged at the time being. For housekeeping purposes we are closing stale PRs. If you'd still like to continue working on this PR, please leave a message and one of the maintainers can reopen it.