Remove Legacy ABI Compatibility #298
Conversation
Yeah but I was not sure how much to delete. Do we want to leave the gsv1 upgrade logic in place or remove it as well? |
|
And this would fall under |
I'm not sure. We currently don't support the new ABI on Mach-O in clang and it's quite useful to be able to run tests on macOS so that we can compare their behaviour against the Apple runtime. I've not tried WebAssembly, but it probably also needs some custom work. The main reason I'd drop the GCC ABI is that anyone using GCC should probably use their runtime. Ideally we'd get every target supported by the v1 GNUstep ABI onto v2 and then drop v1 as well. |
I did not know that we support v1 on Macho-O. It would be very interesting to support v2 as well just for benchmarking purposes. |
|
I'll remove the GCC upgrade logic then. |
hmelder
force-pushed
the
cleanups
branch
2 times, most recently
from
6b50aaf
to
f37d890
Compare
|
Google recently updated the clang prebuilds for the Android SDK. The fast path test now executes and fails instead of being skipped. |
This commit removes the GCC ABI (gcc_abi = 8 in abi_version.c) including the class and protocol upgrade logic and other quirks related to the GCC ABI. This should speed up selector type comparison significantly.
|
@davidchisnall can you take another look at it? I've rebased it onto the latest changes in master. |
| // We can't use all of the values in the type encoding for the hash, | ||
| // because our equality test is a bit more complex than simple string | ||
| // encoding (for example, * and ^C have to be considered equivalent, since | ||
| // they are both used as encodings for C strings in different situations) |
There was a problem hiding this comment.
Are you sure clang never does this? I vaguely remember things like char* and [un]signed char* encoded differently.
| switch (c) | ||
| { | ||
| case '@': case 'i': case 'I': case 'l': case 'L': | ||
| case 'q': case 'Q': case 's': case 'S': | ||
| hash = hash * 33 + c; | ||
| } |
There was a problem hiding this comment.
I'm not sure if this change is correct with respect to structs and so on.
We are carrying a lot of dead-weight just by supporting the old GCC ABI which probably no one uses in combination with libobjc2.
There are still lot more ugly hacks which can be removed.