Stream: cranelift

Topic: ✔ Arrow statements

kmeakin (Mar 17 2023 at 19:54):

What do arrow statements in CLIF mean?
eg

function %mul_minus_one_commuted(i32) -> i32 {
block0(v0: i32):
    v1 = iconst.i32 0xffff_ffff ; -1
    v2 = imul v1, v0
    return v2
    ; check: v3 = ineg v0
    ; check: v5 -> v3
    ; check: v6 -> v3
    ; check: return v3
}

Dan Gohman (Mar 17 2023 at 20:03):

Those are value aliases, so v5 is an alias of v3, and so on.

kmeakin (Mar 17 2023 at 20:06):

Ok. But why do the aliases exist in the first place?

Chris Fallin (Mar 17 2023 at 20:07):

they allow for rewrites without editing all use-sites

kmeakin (Mar 17 2023 at 20:09):

I see. So at some point a v5 and v6 were referenced somewhere?

fitzgen (he/him) (Mar 17 2023 at 20:15):

presumably in intermediate forms in the e-graph, but which were ultimately not selected during extraction

kmeakin (Mar 17 2023 at 20:27):

cool

kmeakin (Mar 17 2023 at 20:28):

maybe we could remove dead aliases from IR before its emitted? Would make more concise/readable output

bjorn3 (Mar 17 2023 at 20:36):

That would lose information. It is possible to replace an alias with another alias or a regular instruction again.

Chris Fallin (Mar 17 2023 at 20:48):

yes, it's useful to have a 1:1 dump of the actual IR data structure; if it becomes an issue for readable dumps at various points I suppose we could have simplification passes (DCE would be another useful one when printing) that can run before debug dumps

bjorn3 (Mar 17 2023 at 22:20):

We do have resolve_aliases to replace all alias uses with their definition and we also don't print unused aliases, so just calling resolve_aliases before printing would be enough to remove all aliases from the output.

Notification Bot (Mar 21 2023 at 20:45):

kmeakin has marked this topic as resolved.

Last updated: Dec 23 2024 at 13:07 UTC