calls to the querySelector method are cut off in case of ADVANCED_OPTIMIZATIONS and turned off checkTypes warnings

[nederland074]

The compilation of a file, which contains only one line:
document.querySelector(".test").some_prop = true;

leads to an empty output when using the compiler with these flags:
java -jar compiler.jar --compilation_level=ADVANCED_OPTIMIZATIONS --jscomp_off=checkTypes --js=in.js --js_output_file=out.js

However, enabling the checkTypes warnings (--jscomp_warning=checkTypes) fixes it and the output is ok: document.querySelector(".test").a=!0;

It doesn't matter if the altered property is a custom "some_prop" or a built-in one like the "className" or "id". Also, the querySelectorAll method doesn't have such an issue and the following code is compiled without problems:
document.querySelectorAll(".test")[0].some_prop = true;

It's reproduced in GCC v20161201.

[MatrixFrog]

Hm. I wonder if we accidentally made an assumption somewhere that anyone using advanced mode would also be using typechecking.

[tonsky]

any progress on this? Ran into it today with both v20170910 and v20171023

[brad4d]

I played around some with reproducing this today.
The --print_source_after_each_pass option seems to indicate that peepholeOptimizations pass is removing the whole statement.

Further note regarding the "it doesn't happen with querySelectorAll()" note in the description.
The important difference there is not the method being called, but the [0] array access before
accessing the property.

[ChadKillingsworth]

In ADVANCED mode, property sets do not have arbitrary side-effects (it is ok to remove property writes if the property is never read), as described here

[brad4d]

@ChadKillingsworth thanks for that info.
It's a bit more than that, though.

Suppose we have code like this:

someExternFunction().someProperty = something;

someExternFunction() might return a value that is part of global state, so setting someProperty on it affects global state and is unsafe to remove. We should assume that this is the case for extern functions unless type info tells us otherwise.
PureFunctionIdentifier is incorrectly assuming the opposite when it has no type information.

I have a change to fix this in internal review.

[ChadKillingsworth]

@brad4d Yeah - that was just a reminder for external viewers.