fuzzgen: Add scalar float support #4467
Conversation
github-actions
bot
added
the
cranelift
There was a problem hiding this comment.
I think this looks great and I'm excited to get more fuzzing for Cranelift!
I saw one copy-paste bug and sort of expect I might have missed some, because having a separate function for each arity leaves a lot of opportunities for this kind of bug. I don't mind merging this with just the one bug fixed, but if you're willing to take a little time to see if you can refactor that, that would be cool.
I'm trying to decide whether to wait on #4460 before merging this. I don't know how long that will take to get reviewed. If you're currently only seeing bugs in the Fma opcode, how would you feel about commenting out that opcode in this patch? Then there's no question, I'm happy to merge it immediately.
Sure, sounds good. I've commented it out and left a link on it. |
|
Fantastic! We're hitting CI hard today so this probably won't get a chance for the CI checks to run for a while, but I'm setting it to auto-merge once that passes. Thanks for your efforts on this! I think I understand why fuzzing didn't catch the misuse of |
|
Apologies for the CI breakage, but things should be fixed with a rebase. |
Add support for generating floats and some float instructions.
Both IEEE754 and the Wasm spec are somewhat loose about what is allowed to be returned from NaN producing operations. And in practice this changes from X86 to Aarch64 and others. Even in the same host machine, the interpreter may produce a code sequence different from cranelift that generates different NaN's but produces legal results according to the spec. These differences cause spurious failures in the fuzzer. To fix this we enable the NaN Canonicalization pass that replaces any NaN's produced with a single fixed canonical NaN value.
This deduplicates a few inserters!
Head branch was pushed to by a user without write access
👋 Hey,
This PR adds support for generating floats in fuzzgen and adds a bunch of instructions.
I think we have all the float ops, except bit ops that operate on floats.
So far this has found #4446 and #4460. We can merge this before #4460 but it will probably start crashing in oss-fuzz.
cc: @jameysharp