Stream: cranelift

Topic: blog post on regalloc checker

Chris Fallin (Mar 15 2021 at 19:06):

Hi all -- I've just posted blog post 3 of 3 on Cranelift: this one about the regalloc checker: https://cfallin.org/blog/2021/03/15/cranelift-isel-3/

Chris Fallin (Mar 15 2021 at 19:07):

Please feel free to share and/or give feedback and/or ask questions :-)

bjorn3 (Mar 15 2021 at 21:20):

Is this a hard problem, and if so, why? In fact, it is NP-complete

This reads a bit awkward.

bjorn3 (Mar 15 2021 at 21:25):

In the "Checking the Register Allocator" section it would be nice to consistently place "machine code" relative to "register allocator" in the diagram.

bjorn3 (Mar 15 2021 at 21:39):

Nice post!

Chris Fallin (Mar 16 2021 at 04:07):

Thanks! I tweaked the wording there; also added some more detail about lattices, and other bits, from some feedback on Reddit.

fitzgen (he/him) (Mar 16 2021 at 17:21):

lol the footnotes have footnotes :joy:

fitzgen (he/him) (Mar 16 2021 at 17:22):

tweeted it out for you, as usual: https://twitter.com/fitzgen/status/1371872683772628994 :stuck_out_tongue_wink:

Correctness in Register Allocation new blog post by Chris Fallin on the symbolic checker and fuzzing set up we have for Cranelift's regalloc https://cfallin.org/blog/2021/03/15/cranelift-isel-3/

- fitzgen (@fitzgen)

Chris Fallin (Mar 16 2021 at 17:38):

Thanks!

Chris Fallin (Mar 16 2021 at 17:38):

I may have been a bit exuberant with the footnotes, but there's just so much interesting stuff to say...

fitzgen (he/him) (Mar 16 2021 at 17:58):

you didn't even touch on the fact that the fuzz inputs are driven by libfuzzer and the arbitrary crate, which let us get coverage guided fuzzing with structured inputs :-p

fitzgen (he/him) (Mar 16 2021 at 17:59):

it just adds one more layer of niceness onto the whole approach

Chris Fallin (Mar 16 2021 at 18:01):

Oh, yes, the whole "fuzzing is fantastic" discussion could fill a blog post of its own

Ian Rogers (Mar 20 2021 at 03:49):

Nice post! Tbh, register allocators have never been the greatest source of bugs for doing the correctness analysis on for me - that's if you dealt with swaps and the lost copy problems properly. It feels like the meet condition is, "don't allocate the same register twice," which is pretty common-sensical.

Ian Rogers (Mar 20 2021 at 03:59):

I should say, don't allocate two values to the same register if they are live at the same time.

Last updated: Nov 22 2024 at 16:03 UTC