Optimize shadow stack pass

[dcodeIO]

This is an experiment to reduce overhead of the shadow stack pass. The pass is currently maintaining more stack slots than necessary (anywhere something managed naively goes to the stack) to be sure, but there is room for improvement.

The relatively simple changes in this PR reduce optimized binary size by about 10% and increase throughput by a little less than 10% (testing with the compiler). Optimizations here are risky, though, because any mistake, not only in the shadow stack pass itself but also in compiler code, for instance where an actually necessary (pre-)instrumentation is missing, can lead to unforeseen consequences that only show themselves randomly.

As such it may make sense to test with various programs and various GC configurations first to minimize the risk.

• I've read the contributing guidelines

[MaxGraey]

any updates? I know this is not huge improvement but still improvement =)

[dcodeIO]

I guess we could just give it a try, but there is a potential for unexpected breakage there. On the other hand, without actually using it in the wild, it may be hard to find problems as these can be quite random.

[MaxGraey]

How about hide this changes under some experimental flag? And ask people turn it on for testing on real projects? But by default turn it off

[dcodeIO]

Hmm, in this case this may actually work. Going to check :)

[MaxGraey]

Or just add extra "incremental-experimental" mode for runtime

[dcodeIO]

Together with #1692:

// JS
I/O Read   :    25.798 ms  n=106
I/O Write  :    71.081 ms  n=3
Parse      :   423.531 ms  n=164
Initialize :    36.616 ms  n=1
Compile    :  1858.461 ms  n=1
Emit       :  2184.024 ms  n=3
Validate   :   246.967 ms  n=1
Optimize   : 65314.422 ms  n=1
Transform  :          n/a  n=0

// Wasm
I/O Read   :    20.495 ms  n=106
I/O Write  :    94.715 ms  n=3
Parse      :   284.738 ms  n=164  <
Initialize :    10.400 ms  n=1    <
Compile    :  1339.539 ms  n=1    <
Emit       :  2369.944 ms  n=3
Validate   :   267.235 ms  n=1
Optimize   : 72261.810 ms  n=1
Transform  :          n/a  n=0

[dcodeIO]

Currently running this branch against Rtrace and worst-case config options, similar to what has been done in initial bootstrapping, and so far is looking good.

[MaxGraey]

Could you also test with assemblyscript-regexp? I guess it pretty great battletest example

[dcodeIO]

Did that now (after figuring out how), and it passes the tests.

For everyone curious, this is how one can do it without hitting any of the errors (missing dist files, missing transforms, ...) I've encountered while trying more convenient things. In your local AS fork:

npm run build
node scripts/prepublish
npm pack
node scripts/postpublish

then replace node_modules/assemblyscript with the contents of assemblyscript.tar.gz in the dependency graph of your module.

[dcodeIO]

Unable to provoke problems here, so probably not worth the trouble to introduce a flag. Suggesting to release it into the wild.

[MaxGraey]

Overall LGTM