Compiler hangs with 2.5.0-dev.20170613

[falsandtru]

It was working until 2.4.0.

https://travis-ci.org/falsandtru/spica/jobs/242650734

TypeScript Version: 2.5.0-dev.20170613

Code

https://github.com/falsandtru/spica

Expected behavior:

pass

Actual behavior:

$ gulp test
[09:29:50] Using gulpfile ...\spica\gulpfile.js
[09:29:50] Starting 'clean'...
[09:29:50] Finished 'clean' after 10 ms
[09:29:50] Starting 'test'...
[09:29:50] Starting 'ts:test'...

<--- Last few GCs --->

[8108:000001D865C4C640]   191743 ms: Mark-sweep 1401.1 (1504.4) -> 1401.1 (1504.4) MB, 5382.0 / 0.0 ms  allocation failure GC in old space requested
[8108:000001D865C4C640]   196651 ms: Mark-sweep 1401.1 (1504.4) -> 1401.1 (1473.4) MB, 4907.0 / 0.0 ms  last resort
[8108:000001D865C4C640]   201277 ms: Mark-sweep 1401.1 (1473.4) -> 1401.0 (1473.4) MB, 4624.0 / 0.0 ms  last resort


<--- JS stacktrace --->

==== JS stack trace =========================================

Security context: 000002707D029891 <JS Object>
    1: set [native collection.js:~247] [pc=00000272AD3ED052](this=000000DA342CCA89 <a Map with map 00000362EC684AE9>,p=0000038C322A0099 <String[9]: 1653672:2>,x=0000038C322A00C1 <a TypeObject with map 00000362EC6E9249>)
    2: getTypeWithThisArgument [...\spica\node_modules\typescript\lib\typescript.js:~31548] [pc=00000272AD3BC751](this=000001933E08A311 <JS Global Ob...

FATAL ERROR: CALL_AND_RETRY_LAST Allocation failed - JavaScript heap out of memory

[falsandtru]

@sandersn Any updates?

[sandersn]

When I updated spica to master, it compiles in 14.6 seconds on my machine. Before #17984, it compiles in 26.92 seconds. Are you still seeing oomemory crashes?

[falsandtru]

Yep, I got crashes when repeating code compiling, using gulp watch. Anyways, a compiler is still too heavy and slow (needs 2.2 min).

[sandersn]

When does the 2.2 minute compile time happen? Is it gulp watch? (I just tried tsc in the project root).

[falsandtru]

Here is my log:

$ gulp watch
[06:51:03] Using gulpfile ...\gulpfile.js
[06:51:03] Starting 'clean'...
[06:51:03] Finished 'clean' after 11 ms
[06:51:03] Starting 'watch'...
[06:51:03] Starting 'ts:test'...
[06:51:05] Finished 'watch' after 1.79 s
[06:53:15] Finished 'ts:test' after 2.2 min
[06:53:15] Starting 'ts:watch'...
[06:53:15] Finished 'ts:watch' after 664 ms
[06:53:15] Starting 'karma:watch'...

You may not have compiled .test.ts files.

[sandersn]

Yes, when I did try gulp I ran ts:dist. When I run gulp test, I get the following output:

[08:30:20] Using gulpfile ~/src/spica/gulpfile.js
[08:30:20] Starting 'clean'...
[08:30:20] Finished 'clean' after 5.3 ms
[08:30:20] Starting 'test'...
[08:30:20] Starting 'ts:test'...
[08:30:56] Finished 'ts:test' after 35 s
[08:30:56] Starting 'karma:test'...

So it looks like the performance improves considerably with master, even when building tests.

[falsandtru]

I ran CI tests with latest nightly build.

https://travis-ci.org/falsandtru/spica/builds/268032017

On Linux, performance has improved but not enough as follows.

[15:56:22] Starting 'ts:test'...
[15:56:49] Finished 'ts:test' after 27 s

On OSX, improved but takes a long time, about 200%.

[17:08:35] Starting 'ts:test'...
[17:09:15] Finished 'ts:test' after 40 s

On Win locally, improved but takes a long time over 200%.

[00:49:35] Starting 'ts:test'...
[00:50:43] Finished 'ts:test' after 1.13 min

With v2.4.0, just a 9.6 sec.

https://travis-ci.org/falsandtru/spica/jobs/264343733

[12:46:14] Starting 'ts:test'...
[12:46:24] Finished 'ts:test' after 9.6 s

Compile time is increased to over 700%.

[sandersn]

I forgot to mention this earlier, but have you tried --noStrictGenericChecks ? If it's slow only 2.4.1 and after, then skipping the improving checking should speed things up considerably. There is a 7x speedup when running bare tsc on spica -- around 14 seconds down to 2 or so. Of course you probably want to compile with the improved checks on sometimes, but for incremental compilation it might be worthwhile to turn them off.

[falsandtru]

I don't want to weaken type checks. And this package is also compiled from other projects like https://github.com/falsandtru/typed-dom and https://github.com/falsandtru/pjax-api. So you are requesting to use --noStrictGenericChecks for all projects using this package. It is impractical. Additionally, I believe you shouldn't spoil the usability on Win and Mac machines.

[falsandtru]

I have read your edited comment. Incremental compilation is nice to me but seems like not implemented. And performance gap on Win and Mac is a different problem. Can you reduce that?

[sandersn]

Sorry, I meant gulp watch compilation. I believe the performance gap is due to differences in node itself on the different platforms -- what are your 2.4.0 numbers on different machines?

I'm looking at compilation times for complex code right now, but I haven't looked at platform-caused differences.

[falsandtru]

due to differences in node itself on the different platform -- what are your 2.4.0 numbers on different machines?

I've compared that on Mac. So I don't think it is a result of platform difference.

On OSX with TypeScript 2.6.0-dev.20170824:

40 sec

https://travis-ci.org/falsandtru/spica/jobs/268032020

[17:08:35] Starting 'ts:test'...
[17:09:15] Finished 'ts:test' after 40 s

On OSX with TypeScript 2.4.0:

9.6 sec

https://travis-ci.org/falsandtru/spica/jobs/264343733

[12:46:14] Starting 'ts:test'...
[12:46:24] Finished 'ts:test' after 9.6 s

This difference is made on the same platform. Probably Windows has a similar performance problem too.

[falsandtru]

Seems like platform-based performance problem is resolved too by #18231. Thanks @sandersn @ahejlsberg .