non-async script needs to be decoded + instantiated on main thread.
currently stencil-XDR decode takes almost same time as JSScript-XDR decode,
and script instantiation takes also takes similar time.
it means if we replace JSScript-XDR decode with stencil-XDR decode + instantiation, it takes twice as longer.

we should reduce the time taken by those steps

benchmark taken from raptor tp6 gmail.
it contains 4 XDR encoded data for with/without stencil XDR.

those 4 scripts are <script> without async, and decoded on main thread.
and the extra time taken by them almost matches the regression.

the result is the following

$ js xdr-test/b.js
[before]
11825 [us]
3403 [us]
4081 [us]
4956 [us]

$ js --no-off-thread-parse-global xdr-test/b.js
[after]
19841 [us]
5201 [us]
8163 [us]
7431 [us]

Updated the testcase to print the best score of 100 runs.

and the output

$ js xdr-test/b.js
[before]
7108
1871
3037
2875

$ js --no-off-thread-parse-global xdr-test/b.js
[after]
14687
3316
5396
5535

Updated bench for 1675804

closing, given the standalone decode+instantiate perf is good enough to flip the stencil-mvp pref.
any new bugs should be filed under bug 1609486.