[fifilura]

I am not familiar with the Itanium story and I don't know who to ask.

But it seems to me like this would be a safe space to experiment. With heuristics and pragmas as a fallback. Because with the right approach solutions would mostly be better than not doing anything.

And you could do it in runtime when you know the size of the input.

And what about applying the logic to places where you can see that the loop will end?

I believe query planners in for example Trino/BigQuery do this already?

[vlovich123]

> I am not familiar with the Itanium story and I don't know who to ask.

https://en.wikipedia.org/wiki/Itanium#Market_reception

> A compiler must attempt to find valid combinations of instructions that can be executed at the same time, effectively performing the instruction scheduling that conventional superscalar processors must do in hardware at runtime.

> In an interview, Donald Knuth said "The Itanium approach...was supposed to be so terrific—until it turned out that the wished-for compilers were basically impossible to write."[222]

There were lots of people but the gist of it is that scheduling auto-optimization is not really possible to do statically at compile time and too much overhead to attempt to do generically at run time. In this case it was instruction parallelism, but the same fundamental issues apply to this kind of auto loop idea.

> I believe query planners in for example Trino/BigQuery do this already?

That is completely different from trying it on every for loop in a program. They do it at the broader task level to determine how much parallelism will help.

[fifilura]

Well. History is full of good ideas executed with the wrong timing.

I am not convinced that this is not possible, just because a project from the early 2000s failed, and Knuth said it was a bad idea.

I am not talking about a general optimizer of poorly written arbitrary programs. Rather an optimizer for parts of the code written with a certain paradigm. Just like BigQuery does with SQL.

(Thank you for sticking with the thread)

[vlovich123]

Sure, such "optimizers" exist already and are well deployed in the field but I've not seen anything that makes runtime decisions about every for loops within a program; these are developers applying their domain knowledge to say "it's important to parallelize here" and the framework has capabilities to do so efficiently because the workload can leverage it.

I'm not saying it'll never happen, but to apply it to every for loop is very different from a framework that speeds up a kind of pattern (which as I said is already a thing). The vast majority of loops don't benefit from this which is why you don't see #openmp pragmas on top of every for loop. The gains are likely minimal from doing it for every loop, you can introduce serious performance regressions when you guess wrong because you don't know about the data pattern within the loop (e.g. you introduce cache contention because the multiple threads end up writing to the same cache line), & most loops don't benefit from parallelization in the first place meaning you're doing a lot of work for no benefit. Maybe bad timing but my hunch is that I'm unlikely to see something like this in my lifetime.

[fifilura]

FWIW, The whole thread was about ditching generic for loops...