@jensimmons It’s crazy that people still try to claim that Safari is extremely far behind other browsers. When the reality is a privacy focused browser from the foundation has to be done correctly not quickly.

Congratulations to the entire Safari team for being the first browser to support this tho. 🎉

There was a paper to fix this. https://wg21.link/p0145

If you see in the last version of that paper (r3, http://wg21.link takes you to last revision), they had to leave open the potential for not doing left-to-right eval, because one set of vendors fought the authors.

The history of this is fascinating, because if you go to a few of the CppCon and C++Now panels with Standards Experts, you can hear them -- on video -- laughing at the guy working on Clang for Windows for asking why they didn't standardize the eval order for parameters in calls.

(Sidenote: I'll never forget this shit because it was pre my involvement in the C++ Committee. I remember hearing that Clang for Windows guy get up there, ask, and then have everybody sitting up there imply you're a dumbass for relying on the order of evaluation.

I felt stupid, myself.)

Fast forward to the publication of this paper, all of the C++ experts that were up there shitting on the Clang for Windows dev had egg on their face because common idiomatic code had unspecified evaluation order and things were genuinely fucked up between compilers!

So, these folks write a paper to fix it, because that's actually kind of fucking embarrassing and stupid. If you read the revisions r0 through r2, you'll see they included left-to-right evaluation for Function Calls as part of the deal.

But it didn't make it.

What happened?

See, the people writing the paper knew it was a steep slope. Old compiler dev traditions die hard; unevaluated order is best so they can swizzle things and Do Cool Optimizations™ or whatever the hell they were saying.

So they made some benchmarks and tests, then put that data in the paper.

Turns out, eval order didn't matter as much as your optimizer just not being dumb as bricks. You could change the ordering and even when you weren't trying it was basically noise which direction it went in: +/-2% either way just by doing left-to-right eval, without cluing the optimizer in to the change of the evaluation order. This was pretty solid evidence that if people actually gave a damn and tried, left-to-right enforced eval would have no performance impact, and the Old Compiler Dev traditions was actually just that:

boring, dead tradition!

Alas, despite this paper first showing up in 2015 and having time for others to validate their claims, rather than present counter-data, do you know what the detractors of this paper did? They just complained.

"You're taking away my implementation freedom!" (Not a direct quote, but that's the gist of it.) And, you know what?

It worked.

By the time this paper was settled for C++17, they chose the most batshit option which is:

• you can eval left to right
• you can eval right to left
• you can start anywhere in the list
• order cannot be interleaved

A clownshoes resolution, but on-brand for C++.