Stream: git-wasmtime

Topic: wasmtime / PR #1783 Improve bitselect codegen with knowle...

Wasmtime GitHub notifications bot (May 28 2020 at 22:43):

teapotd opened PR #1783 from bitselect-blend to master:

This PR fixes #1187 and adds:

• encoding of vselect on x86 using BLEND* instructions
• legalization of vselect to bitselect
• simple_preopt transformation that replaces bitselect with vselect if possible

r? @abrown

Wasmtime GitHub notifications bot (May 28 2020 at 23:38):

teapotd updated PR #1783 from bitselect-blend to master:

This PR fixes #1187 and adds:

• encoding of vselect on x86 using BLEND* instructions
• legalization of vselect to bitselect
• simple_preopt transformation that replaces bitselect with vselect if possible

r? @abrown

Wasmtime GitHub notifications bot (May 29 2020 at 20:54):

abrown submitted PR Review.

Wasmtime GitHub notifications bot (May 29 2020 at 20:54):

abrown submitted PR Review.

Wasmtime GitHub notifications bot (May 29 2020 at 20:54):

abrown created PR Review Comment:

This and below can now use better syntax:

    v4 = vselect v1, v2, v3
    return v4
}
; run: %vselect_i8x16() == [200 101 202 103 204 105 106 107 108 109 210 211 212 213 214 215]

And, even better, could pass in the vectors in the run statement:

; run: %vselect_i8x16([...], [...], [...]) == [...]
~~~

Wasmtime GitHub notifications bot (May 29 2020 at 20:54):

abrown created PR Review Comment:

nit: we might want to note here that using bitselect works because the entire lane is filled with 1s or 0s

Wasmtime GitHub notifications bot (May 29 2020 at 20:54):

abrown created PR Review Comment:

It might be good to use a couple of run tests as well just to make CI assure us that the optimization is running as expected.

Wasmtime GitHub notifications bot (May 29 2020 at 20:54):

abrown created PR Review Comment:

I was puzzling over this for a while and I do understand now why this works; since bitcast was selecting at the bit level, we can safely select at the byte level in this optimization. In this code path we are going to bitcast our way to vselect(B8x16, *8x16, *8x16) every time, right?

Wasmtime GitHub notifications bot (May 29 2020 at 21:16):

teapotd submitted PR Review.

Wasmtime GitHub notifications bot (May 29 2020 at 21:16):

teapotd created PR Review Comment:

That's right, I'll add a comment with explanation to make it more clear.

Wasmtime GitHub notifications bot (May 29 2020 at 22:57):

teapotd updated PR #1783 from bitselect-blend to master:

This PR fixes #1187 and adds:

• encoding of vselect on x86 using BLEND* instructions
• legalization of vselect to bitselect
• simple_preopt transformation that replaces bitselect with vselect if possible

r? @abrown

Wasmtime GitHub notifications bot (May 29 2020 at 23:11):

teapotd updated PR #1783 from bitselect-blend to master:

This PR fixes #1187 and adds:

• encoding of vselect on x86 using BLEND* instructions
• legalization of vselect to bitselect
• simple_preopt transformation that replaces bitselect with vselect if possible

r? @abrown

Wasmtime GitHub notifications bot (May 29 2020 at 23:17):

teapotd requested abrown for a review on PR #1783.

Wasmtime GitHub notifications bot (May 30 2020 at 02:52):

abrown submitted PR Review.

Wasmtime GitHub notifications bot (May 30 2020 at 02:53):

abrown merged PR #1783.

Last updated: Nov 22 2024 at 16:03 UTC