Stream: cranelift

Topic: basic blocks


view this post on Zulip Sean Stangl (Dec 18 2019 at 21:24):

@Benjamin Bouvier Are we ready to remove the basic-blocks feature gate and make it default?

view this post on Zulip fitzgen (he/him) (Dec 18 2019 at 21:26):

is the plan to completely remove EBBs, or make them available-but-not-on-by-default? or first disable, but eventually remove?

view this post on Zulip Sean Stangl (Dec 18 2019 at 21:27):

The plan is to remove them in the near-term. I'd still like to keep them as an optimization for one particular pass (I believe wasm IR to CLIF conversion) as an internal detail.

view this post on Zulip Sean Stangl (Dec 18 2019 at 21:28):

The new regallocs that @Benjamin Bouvier and Julian Seward (who's not here!) are writing likely depend on non-ebb representations.

view this post on Zulip iximeow (Dec 18 2019 at 22:19):

Benjamin Bouvier Are we ready to remove the basic-blocks feature gate and make it default?

https://github.com/bytecodealliance/cranelift/commit/decec8bb4e29572c13e6e325c2c2ebb1a6ea7b4f <-- basic-blocks has been default for a few months now, fwiw. as soon as EBBs are removed i plan on closing https://github.com/bytecodealliance/cranelift/issues/1219.

Cranelift code generator. Contribute to bytecodealliance/cranelift development by creating an account on GitHub.
First, some clif: ; guest_func_73448 function u0:73448(i64 vmctx, i32, i32) -> i32 system_v { gv0 = vmctx heap0 = static gv0, min 0x0052_0000, bound 0x0052_0000, offset_guard 0x0040_0000, index_...

view this post on Zulip Benjamin Bouvier (Dec 19 2019 at 08:16):

Yes, the plan is to remove basic blocks entirely. If we wanted to keep them, we'd need either to maintain all the optimization passes so they work with EBBs, or disable them all. So I'm strongly in favor of removing them entirely.

view this post on Zulip Benjamin Bouvier (Dec 19 2019 at 08:17):

For the bikeshedding matters, I think "Block" / "block" is better than "Bb" / "bb" as a replacement for "Ebb" / "ebb". We've had "Value" in the past, same number of letters, and fully-expanded is less ambiguous and more readable than the acronym.


Last updated: Nov 22 2024 at 17:03 UTC