Push ifs up and fors down: The idiom, its algebra, and its limits | Hacker News Reader

Push ifs up and fors down: The idiom, its algebra, and its limits

debasishg.github.io

171 pointsspeckx80 comments

Are we talking about this from the perspective of CS (algorithm optimization) or SE (code design)?

From an SE perspective, make a flatmap function that explicitly handles Collection<Optional<Walrus>>. The implementation doesn't matter. If your language/framework already has a compatible flatmap function, make a single frobnicate(Optional<Walrus>) function that returns whatever value is necessary for flatmap(frobnicate) to discard them.

From a CS perspective, doing a filter from Collection<Optional<Walrus>> to Collection<Walrus> is probably a bad idea. If your collection is small, nothing matters. If your collection is large, you probably don't want to spend time making a new copy of it. If your filter just returns a view rather than a hard copy, then there is no optimization benefit and you should just do whatever makes the most sense from an SE perspective. If frobnicate is cheap then you're paying the branch prediction failure tax anyway regardless of when you frobnicate, and if frobnicate is more expensive then your should probably parallelize and have each thread handle unpacking the Optional. Either way, you probably don't want to spend time making a copy.

These are all generalizations based on hypotheticals and there are certainly a lot of exceptions, but broadly speaking I don't see a strong argument here. If optimization matters then optimize based on your own profiling of your situation, and if optimization doesn't matter then design your functions based on what features and paradigms are available/common in your area.

Save some time and read the original post instead: https://matklad.github.io/2023/11/15/push-ifs-up-and-fors-do...

I am continually impressed by the ability of LLMs to take trivial ideas and turn them into lengthy and obtuse blog posts with unnecessary analogies.

Yet another encroachment on traditionally human activity.

I am the

  Option<Walrus>

I know we’re not supposed to comment just for that, but this might be my single favorite joke comment I’ve ever read here. Good job.

what is the joke

https://en.wikipedia.org/wiki/I_Am_the_Walrus

goo-goo g'joob

Except that TFA is a bog standard example of traditional human activity and the GP's comment is nonsensical trolling.

Honestly, this just looks like one of those lingo-heavy-but-surface-level blog posts that used to make functional programming spaces so insufferable to everyone on the outside

These things are so divorced from the reality of programming, even when they involve actual code instead of fancy lingo. Like in Scala, not a pure functional language, tutorials used to find the most convoluted higher-order functional way to do simple things.

You want to avoid branches in hot paths. If you branch inside the loop, lots of branches. If you branch outside the loop (into different specialized loops), few branches. Big fucking deal.

https://en.wikipedia.org/wiki/Loop_unswitching

    bool w;
    int x[1000];
    int y[1000];
    for(int i=0;i<1000;i++){
      x[i] += y[i];
    }
    if(w){
      for(int i=0;i<1000;i++){
        y[i] = 0;
      }
    }

Slower but prettier with the if pushed down.

Is that true? The loop is a branch.

Who actually considers a loop a branch? A loop is a chunk of code that will be run N times depending on some evaluation that’s run before or after an iteration. A branch is a single decision which of several pieces of code to run, once.

>Who actually considers a loop a branch?

The machine?

> Who actually considers a loop a branch?

A loop contains a branch - keep looping or break. At least according to structured programming -- https://en.wikipedia.org/wiki/Structured_program_theorem.

> Who actually considers a loop a branch?

interesting question...

> A loop is a chunk of code that will be run N times depending

oh! turns out you do!

branchless for loop for funsies

  for (fn, n) = {
    stop = void (i) {}
    run  = void (i) { fn(i); go(i + 1) }
    go   = void (i) { *((i < n) * &run + (i >= n) * &stop)(i) }
    go(0)
  }

Or you could write all the implementations for each possible n and dispatch through a registry.

https://godbolt.org/z/rz8MxhdGd

Just a simple example, but the conditional branch occurs on line 14 of the generated assembly. It does a comparison (line 13) and then a conditional jump (jl, line 14).

> Who actually considers a loop a branch?

Anyone who started out programming in assembly languages

> Who actually considers a loop a branch?

The CPU.

The "go around" branch is obviously not completely avoidable (though unrolling helps). Branches inside the loop body sometimes are.

And also generate a shorter version.

Semantic compressor and decompressor

Debasish has been writing for a long time. I have one of his books on FP in my office.

No, TFA is standard fare for bloggers who dwell in category theory 24/7.

To be fair, I bought Martin Fowler's Refactoring book expecting to learn a bunch of new stuff and up my game. What I found was a bunch of stuff I knew already just from experience but thought was too obvious to enumerate and write down. This was all written a long time before LLMs, of course.

I think the main reason I don't write more is I think once I've thought through something it's too obvious to write down. I consider it a defect of mine and sometimes have to force myself to write.

Maybe dialogue would be a more fruitful approach. Then one would know what the other already knows and the participants could only communicate the "deltas".

Dialogue is super effective for this, but it's also a skill: you're suddenly a teacher. You should have no judgement of the other not knowing something, and be able to effectively build up that delta in a way that is understandable and interesting to the other. This can be more difficult than it sounds, but it's also rewarding.

I've always believed the opposite: get conditionals deep in your code so that the higher level control flow is regular.

But I suppose my greater philosophy for making code that avoids bugs is that you have a couple things that are done when dealing with data:

- distribution

- deciding

And you want to avoid distribution and deciding being mixed together in the same spot.

"Distribution" can be for loops but also breaking up some data based on some key into N bistinct buckets

"Deciding" is where you're looking at the data more closely to make some decision (like "is this a big customer or a small customer")

Distribution often involves decision making, but if you mix them all in one spot you can obfuscate your decision points. Splitting it up just makes things "obviously" right or "obviously" wrong. Perf stuff is another discussion of course, but in practice most things are not at a scale where it matters.

    by_category = defaultdict(list)
    for d in data:
      by_category[category(d)].append(d)

    for category, per_category_data in by_category.items():
       do_thing(category, per_category_data)

I really value code patterns that make mistakes obvious, or at least makes it harder to stuff a mistake in somewhere. Some patterns are harder to describe in this model though.

(I do like the advice of having a consistent vocabulary for working on collections as a principle though, I just find that top-level conditional use tends to quickly get you into "... why is this method not called" territory, which is a more annoying problem than "why is this slow")

That's another good way to look at it - sometimes the "base" is more like a physics substrate. Physics doesn't care about semantics, it just is. Putting semantics first would be weird.

I guess it's a case of perspective

I think compilers can push out ifs inside for to be one if with two fors.

If the data don't need to be processed in batches by category, and if the category is derivable from an data item alone, I don't really see the benefit you're proposing.

Even worse, by splitting one state (and one derivable category from that state) into two separate arguments for do_thing, something can be off rather badly. I'd then feel the need to design an assertion of the relation of the arguments in order to make things bearable again:

  def do_thing(category, data):
    # to avoid shadowing lets rename your function `category` to `category_from_item`
    assert all(category == category_from_item(item) for item in data)
    ...

But that would add a third loop to your two loops, and would duplicate the computation of a category.

Instead, if category would be a property of data item:

  class DataItem:
    
    @property
    def category(self):
       ...

...and the do_thing function would work on a single data item, then it'd just become a simple matter of one for loop and one match/case:

  def do_thing(item):
    match item.category:
      case ...:
        ...

One thing I noticed is that the principle (better demonstrated in the linked matklad article) seems to make sense at the "a bunch of related functions" level, but it also goes against one of the main principles of OOP, which is not to prescribe behaviors to others.

You don't want to go "woof if you're a dog, meow if you're a cat", you want to go "make a sound, whoever you are". If you're a fish, you can no-op or not "accept" the call, depending on the language and convention. Obviously there is a lot of nuance, but the general idea is to push responsibilities and logic (which includes branching) down, to the level/entity/object that's best equipped to handle it.

If you apply this to the "a bunch of functions" example, you get low-level functions with branching, but you also get clean callers.

I feel like this isn't such a huge problem IRL, it's usually quite apparent where a decision should be made, but both of these "schools of thought" seem equally valid to me.

But OOP’s insistence on dispatching "at the leaves" is mostly understood as misguided or overly dogmatic these days. In particular, tagged unions and pattern matching are incredibly useful tools that go right against OO dogma. And they’ve always been so, since their introduction in the early 70s. The entire idea of "objects combine behavior and data" is problematic in many ways, not least because it can be very suboptimal on modern hardware compared to data-oriented approaches.

here we go again.

https://en.wikipedia.org/wiki/Expression_problem

What is missing here is any benchmarks backing up this argument for code structure.

Of note, as of C#9 (and maybe prior), the dotnet runtime does this automatically whenever it is deemed safe. https://devblogs.microsoft.com/dotnet/performance-improvemen...

The same technique is applied as an optimization, when deemed safe, in all current gen c compilers (gcc, llvm, etc).

I'm very confused why neither measurements nor references to when this is done automatically in most modern languages is included in the article.

At least in JVM land, it's pretty easy to thwart that optimization. Particularly if the condition is on a mutable yet unchanged in the loop value.

For example:

    var map = new HashMap<String, String>();
    map.put("foo", "bar");
    for (var i : items) {
      if ("bar".equals(map.get("foo")) {
        doStuff(i);
      }
    }

Even though `map` isn't mutated, it's hard enough for the JVM to detect and the underlying `get` functions are complex enough that it'll run the `get("foo")` every time, which can be quite expensive.

Can't speak for C# but in C/C++ the optimization can rarely be applied safely due to aliasing. If any part of the data you're working with involves a char* then C/C++ optimizers refrain from doing these kinds of optimizations because of how difficult it is to guarantee the absence of mutability.

I would rather the developer do it, not the compiler. I don't, as a matter of course, regularly _read_ compiler output. I do, however, read developer output.

I've done this for years. Not every time of course but where it makes the code easier to understand and maintain.

Speed was almost never the reason.

I take it you never rewrote a Matlab for loop as a vector/matrix op for insane speedups then :)

I've never used Matlab, so no.

A better formulation I read in a magazine back in the 90s is: "push mechanism down and policy up".

Erm, no? You write f(w: Walrus) -> Walrus and then let the caller handle Walrus|None and Iterable[Walrus] however they wish!

And if someone decides the codebase needs an abstraction over (and therefore specific functions to handle) Iterable[Walrus|None] then you check the weather and suggest they take a break and go for a stroll. (You check the weather to see if you should lend them your brolly.)

What am I missing?

Didn’t see it mentioned in the article but isn’t leading with if-statement called a “guard clause”. I like that pattern but it’s just general best practice I thought.

Swift explicitly has a guard statement for this. Rust's let .. else { ... } is also very similar.

https://docs.swift.org/latest/documentation/the-swift-progra...

Guard clauses are things that return early for trivial or problematic cases. Like https://en.wikipedia.org/wiki/Guard_(computer_science)#Flatt...

They're one of those good practices that look like bad practice to everyone who just got a CS degree. Seems ex-students are unsettled by asymmetry or want to minimize the number of return statements.

Particularly for high performance code when branch prediction is taken to account

Why am I even in this function if it shouldn't happen?

If the criteria for it not happening is too complicated to expose to the caller

Related: don't write if(arg == null) throw ArgumentNullError; at the top of every function. If it's not supposed to happen then just let it throw the error naturally when you dereference it. In C it's even worse because you turned an easily caught segfault into a silent nop.

Because a lot of languages can't easily express the semantics of "you're not allowed to call the function this way". I mean, suppose you have a function that takes two integers, but the second must be greater than the first. Most type systems don't have any way to represent that at all, let alone conveniently.

Oh, guard clauses might also use continue statements in loops

Idiomatic Elixir does this with pattern matching on function parameters so you end up with things like the following, raw if statements are discouraged because of this:

  def classify(:ok)
  def classify({:error, reason})
  def classify([first | rest])
  def classify(%{name: name, age: age}) when age >= 18
  def classify(%{name: name, age: age}) when age < 18

I have always phrased this as "Never do one of something".

“Batch is the primitive”

Push lists of 0 or 1 up, but push lists of 0, 1, 2, 3, or etc, down?

Just the branch predictor gains are probably worth it.

Is the idea that "accidental casework" should be moved up, whereas the "reusable bulk ontology" should be moved down?

There's very high-leverage abstractions that completely constrain a space. An example is a good definition - you can't think of something outside to compare it to, it just is. These things survive for a long time since they define it.

But if you're trying to do that philosophy super deep into a program, you're probably violating a bunch of invariants subtly.

Of course, there is no good separation at the end of the day as we all know from spaghetti codebases :)

this is just a guard on the function definition?

I like it, but to do fizzbuzz in this way, you'd have to separate what's inside the loop into a reused function.

I think it's sort of obvious that the limit to this general rule is when data dependencies between fors and ifs forbid you from pushing things further up/down.

Or just use lazy list operations with a single if test at the end:

https://github.com/taolson/Admiran/blob/main/examples/fizzBu...

TL;DR in one sentence:

"the loop runs without a branch, and is a candidate for vectorization".

That's it, that's the article. This matters a lot in huge-scale / scientific computing / HPF, where if you can express something as an operation on vectors on matrices, you win big (those ops parallelize well, can be run on GPUs, clusters, what have you).

For the 99% of developers who are shuffling data around constrained by I/O, you win small.

Reply on news.ycombinator.com