Slower matrixmultiply in nightly-2017-12-16

[bluss]

• Good rustc 1.24.0-nightly (0077d12 2017-12-14) (rustup nightly-2017-12-15)
• Bad rustc 1.24.0-nightly (77efd68 2017-12-15) (rustup nightly-2017-12-16)

matrixmultiply version 0.1.14. This code is deliberately seeking autovectorization, and it's understood that such things could be fragile; but for a long time it has worked fine.

Before:

RUSTFLAGS=-Ctarget-feature=+avx cargo +nightly-2017-12-15 bench -v m127

test mat_mul_f32::m127             ... bench:     121,631 ns/iter (+/- 10,097)
test mat_mul_f64::m127             ... bench:     198,662 ns/iter (+/- 13,402)

After:

RUSTFLAGS=-Ctarget-feature=+avx cargo +nightly-2017-12-16 bench -v m127

test mat_mul_f32::m127             ... bench:     209,268 ns/iter (+/- 12,809)
test mat_mul_f64::m127             ... bench:     295,810 ns/iter (+/- 2,167)

The PRs in that range are:

git log --oneline --first-parent 0077d128d..77efd6800
77efd68 Auto merge of #46623 - eddyb:issue-46449, r=nagisa
50f6c3e Auto merge of #46745 - steveklabnik:rollup, r=steveklabnik
5a0dc2d Auto merge of #46740 - oli-obk:rls, r=nrc
b4f3232 Auto merge of #46706 - sunjay:gat-lifetimes, r=nikomatsakis
04b2344 Auto merge of #46653 - estebank:str-as-ch, r=petrochenkov
84feab3 Auto merge of #46537 - pnkfelix:two-phase-borrows, r=arielb1
9331031 Auto merge of #46701 - eddyb:vector-newtypes, r=nagisa
2f35141 Auto merge of #46641 - petrochenkov:nohelp2, r=nikomatsakis

[pmarcelll]

This might be the same issue as #46897.

[eddyb]

@pmarcelll I think so too. @bluss As per #46897 (comment), can you retry with the latest nightly?

EDIT: that might've been accidental, given no relevant changes, and the regression is back anyway.

[bluss]

New results (no fix ☹️ )

• Bad rustc 1.24.0-nightly (7eb64b8 2017-12-20) (rustup +nightly-2017-12-21)
• Bad rustc 1.24.0-nightly (250b492 2017-12-21) (rustup +nightly-2017-12-22)
• Bad rustc 1.24.0-nightly (5165ee9 2017-12-22) (rustup +nightly-2017-12-23)

[bluss]

I haven't had time to investigate, but I'll do that over christmas.

[arielb1]

This really looks like #46897

[bluss]

Probably. The reason I discovered it was due to that issue, but when I read it it seemed to talk about a different nightly version as the breaking point, so I reported this too.

[bluss]

Fixed in rustc 1.24.0-nightly (1abeb43 2017-12-27) 🌟 @eddyb 🌟

RUSTFLAGS=-Ctarget-feature=+avx\ -Ccodegen-units=1 cargo +nightly bench -v m127
test mat_mul_f32::m127             ... bench:     123,960 ns/iter (+/- 8,354)
test mat_mul_f64::m127             ... bench:     194,921 ns/iter (+/- 3,862)

So it was resolved like #46897