Agenda for sync meeting 2/5/21

[tlively]

The next meeting will be Friday, February 5 at 9:00AM - 10:00AM PST/ 6:00PM - 7:00PM CET. Please respond with agenda items you would like to discuss.

If this meeting doesn't already appear on your calendar, or you are a new attendee, please fill out this form to attend.

Agenda so far:

• Revisit signed i64x2 comparisons after further discussion (i64x2.gt_s, i64x2.lt_s, i64x2.ge_s, and i64x2.le_s instructions #412)
• Unsigned i64x2 comparisons (i64x2.gt_u, i64x2.lt_u, i64x2.ge_u, and i64x2.le_u instructions #414)
• Sign select (Sign Select instructions #124)
• i64x2.abs (i64x2.abs instruction #413)

[zeux]

To make it easier to track the remaining proposals it would be great to close the PRs that we voted against or mark PRs that are known to be post-MVP with the relevant tag. I'm not sure if there are other instructions beyond the 4 above that still need discussion before phase 4 or not?

[dtig]

I've added a needs discussion label to the outstanding PRs that do not have a resolution. The issues #247 and #374 haven't seen much discussion of late, and the lowering for #395 is not convincing for late stage inclusion though prototyping and benchmarking efforts are in progress. I'll be adding votes to the PRs not included in the agenda shortly, and we can ratify them in the meeting if necessary.

[omnisip]

@dtig I'm working on the testing and benchmarking now with respect to #395, but am not sure how the lowering is not convincing if the benchmarks pan out. On x86/x64, it's single instruction for all of the unsigned variants, and 2 instructions for all of the signed variants. The 2 instruction variants perform as well as the single instruction ones (according to LLVM-MCA) because the arithmetic shift isn't competitive with the shuffle ops.

I'm definitely curious to see how it performs on A64, but want to finish the x64 testing first.

[omnisip]

Link for implemented prototype lowerings is here: https://chromium-review.googlesource.com/c/v8/v8/+/2617389/9/src/compiler/backend/x64/code-generator-x64.cc

[dtig]

@omnisip Sorry, my comments above are not very clear, I've followed up on #395. I'm aware of the code-gen lowerings, I was a reviewer for the patch that introduced them.

[tlively]

Thanks everyone! Here are the notes. Here are the decisions we came to:

We will merge:

• i64x2 signed comparisons (i64x2.gt_s, i64x2.lt_s, i64x2.ge_s, and i64x2.le_s instructions #412)
• i64x2.abs (i64x2.abs instruction #413)

We will not merge:

• i64x2 unsigned comparisons (i64x2.gt_u, i64x2.lt_u, i64x2.ge_u, and i64x2.le_u instructions #414)
• Sign select (Sign Select instructions #124)
• i32x4.nearest_sat_f32x4_{s,u} (Floating-Point to Nearest Integer Conversions #247)
• 8->32 bit extends (#372 Integer Sign/Zero Extension for {8,16}->{32,64} #395)
• Move zero (move{32,64}_zero_{r,v} instructions #374)