SCons: Disable /EHsc for MSVC, speeds up build significantly

[akien-mga]

As seen in #80513, /EHsc seems responsible for extremely long build times for complex files such as gdscript_vm.cpp and variant_call.cpp, when combined with optimization flags.

On my Windows setup, a clean rebuild of the engine with scons p=windows (thus defaulting to optimize=speed-trace, i.e. /O2 on MSVC) went from 35 min down to 16 min with this patch. Incremental rebuilds that would lead to recompiling either gdscript_vm.cpp or variant_call.cpp will benefit even more.

We were using /EHsc to solve warnings in thirdparty code, so instead we silence that warning specifically.

Fixes #80513.

---

I could use help to properly assess the impact of removing /EHsc and ignoring the C4530 warning. I'm not super familiar with exception handling and stack unwinding. Relevant docs: https://learn.microsoft.com/en-us/cpp/build/reference/eh-exception-handling-model?view=msvc-170

Patching thirdparty libraries so they don't raise this warning is also an option. In theory, we don't want any exception handling in Godot.

[akien-mga]

Interesting error on tests after this change:

Error: .\tests/core/io/test_file_access.h(42): error C2338: static_assert failed: 'Exceptions are disabled! Use DOCTEST_CONFIG_NO_EXCEPTIONS_BUT_WITH_ALL_ASSERTS if you want to compile with exceptions disabled.'

[akien-mga]

Interesting error on tests after this change:

Error: .\tests/core/io/test_file_access.h(42): error C2338: static_assert failed: 'Exceptions are disabled! Use DOCTEST_CONFIG_NO_EXCEPTIONS_BUT_WITH_ALL_ASSERTS if you want to compile with exceptions disabled.'

So Doctest needs exceptions for the REQUIRE (and I think FAIL) macros, i.e. macros which mark the test as failed and abort. It has an option to disable the use of exceptions but that basically downgrades those macros to CHECK level, which mark the test as failed but don't abort.

We could look into whether we can get by with only CHECK level macros, but if so we can't test unrecoverable conditions. I haven't assessed what our current tests need.

So for now I just re-define /EHsc for the tests/SCsub environment.

[ryuukk]

Just randomly passing by

Looked around and seems you can disable EH for doctest:

2 config possible:

DOCTEST_CONFIG_NO_EXCEPTIONS
DOCTEST_CONFIG_NO_EXCEPTIONS_BUT_WITH_ALL_ASSERTS

EH is evil

[akien-mga]

Looked around and seems you can disable EH for doctest:

2 config possible:

DOCTEST_CONFIG_NO_EXCEPTIONS DOCTEST_CONFIG_NO_EXCEPTIONS_BUT_WITH_ALL_ASSERTS

Indeed, but as I pointed out above, I believe this will break the tests by downgrading REQUIRE/FAIL asserts to the non terminating CHECKED level, so some tests might crash if they rely on aborting when a non recoverable error happens.

Could still be worth doing but this will likely require changing a lot of tests.

[mihe]

There's also the quasi-supported (and undocumented) _HAS_EXCEPTIONS=0 definition to consider, which I typically see defined in conjunction with omitting /EH[sc], and supposedly allows the MSVC STL implementation to make different choices in some cases. I'll admit I don't fully understand the pros/cons to defining this versus not defining it, but I would imagine it's safe to define it as long as you don't plan to actually use any exception classes or exception-handling anywhere. Considering how little the STL is used in Godot anyway, I would imagine it's even less of a risk. Unreal (who also uses the STL very sparingly) seems to get away with defining this, just to give one example.

Has there been any consideration towards disabling exception-handling for GCC/Clang as well, through -fno-exception? Disabling exception-handling seems fairly ubiquitous for gamedev stuff and can have significant impacts on binary size, and presumably instruction cache performance, since it can omit all the generated code for the stack-unwinding, which is just dead weight if you never use exceptions anyway. This PR nets a ~15% decrease in binary size for target=editor. I would be surprised if GCC/Clang didn't net at least a couple of percentage points as well.

[akien-mga]

I made #80612 as an alternative to fully disable exception handling. Reviews and tests welcome :)

[akien-mga]

Superseded by #80612.