Comment by almostgotcaught | Hacker News Reader

Comment by almostgotcaught | Hacker News Reader

Parent Full thread

almostgotcaught

> A good IR means that the only contract between compiler passes is the IR itself.

this is like impossible (or at least np-hard). phase ordering exists and will always exist - SLP vectorizer will dramatically affect passes after it and sometimes even passes before it!

> Writing a cool new pass that obeys the IR’s laws is not going to break the other passes.

I don't really get how you can fail to obey the IR's laws other than just emitting unparseable IR? alternatively if you like have uses without defs or flip-flopping types, the verifier needs to catch that after the guilty pass and just bail (not put the onus on the subsequent passes to fix).

Anyway IMHO (even though no one asked me), IR needs to have enough structure/info that you can quickly reproduce whatever metadata you need for XYZ arbitrary analysis. Meaning it should represent all of the useful/necessary things (don't make me infer non-negativity structurally if I can just annotate) and it should propagate those things across rewrites.

> this is like impossible (or at least np-hard). phase ordering exists and will always exist - SLP vectorizer will dramatically affect passes after it and sometimes even passes before it!

It's not impossible.

SLP vectorizer will affect the performance upside of passes before/after it. It won't affect their correctness.

> I don't really get how you can fail to obey the IR's laws other than just emitting unparseable IR? alternatively if you like have uses without defs or flip-flopping types, the verifier needs to catch that after the guilty pass and just bail (not put the onus on the subsequent passes to fix).

Lots of compilers - not LLVM IR - have a totally different kind of phase ordering problem than what you are thinking of: if you reorder the phases then the compiler crashes and burns. Some of these compiler are in production, sadly. I've heard some crazy horror stories. And it's all too easy to find yourself in that situation if you're not careful.

LLVM IR is an example of being careful.

> SLP vectorizer will affect the performance upside of passes before/after it. It won't affect their correctness.

maybe. it's late so i can't think of a counter-example off the top of my head where someone downstream of SLP somehow took something for granted and then it blew up in their faces. even if it is somehow magically not true for SLP, it's definitely true for other sequences of passes, where subsequent passes depend on prior passes.

> if you reorder the phases then the compiler crashes and burns. Some of these compiler are in production, sadly. I've heard some crazy horror stories. And it's all too easy to find yourself in that situation if you're not careful.

i mean i have witnessed this in a prod MLIR compiler i have personally worked on - it's exactly what i had in mind above. did it crash because it tried to deref some pointer to some attribute/node/whatever? no that's true - it crashed via llvm::report_fatal_error (or assert) - but it crashed nonetheless.

> it's definitely true for other sequences of passes, where subsequent passes depend on prior passes.

I don’t think it’s true for optimization passes in llvm and that’s what makes llvm kinda great.

I would love to hear of a counter example.

But say that you found a counterexample. That would still not be as bad as what happens in a lot of compilers out there where being able to reorder passes (or even run them a second time) is the exception rather than the norm, and our debate would be about whether there exists even a single pair of passes that could be safely reordered.

Sure - I agree that LLVM is mostly reasonable - I'm just saying I don't think has to do with the design of the IR itself and is primarily because of convention/code quality/stewardship. Like to ask my question more precisely: can you give me an example of an IR design that didn't "isolate ugliness"?

Edit: I'm being rate limited so here's what I wanted to post in response:

You kept talking about LLVM and I just assumed LLVM IR but this thing (different rules before and after regalloc) is exactly the case for LLVM's MIR (even going so far as no longer being SSA after a certain part of the pipeline). And in fact I have a PR hanging out right now for adding a legalization for GISel (which also operates on MIR) that crashes in exactly the way you say it shouldn't - I run my legalizer and it's fine and dandy and then somewhere down the line (another legalizer) there's a null dereference. So I guess LLVM isn't reasonable everywhere! And also I should take a think about the differences between LLVMIR and MIR.

In my experience with compilers where phase ordering causes fires, it's exactly to do with the IR.

Because I don't want to pick on others (as much fun as that would be), I'll just pick on compilers I wrote that have this property and how much I regret everything I did there.

- The first compiler IR I designed was called Fiji C1, and it had both an SSA form and a not-SSA form, and it had multiple stages of lowering within the same IR. You absolutely could not breathe on the phase order without destroying everything.

- I'm maybe like mostly responsible for JavaScriptCore's DFG IR, which has a LoadStore form, a ThreadedCPS form, and an SSA form, plus maybe other forms too. And there are many phases in there that serve the purpose of carefully lowering some aspect of type speculation. Foolish is the man who tries to mess with the order of those phases.

- The Assembly IR in JavaScriptCore's top tier JIT has different laws before and after regalloc, and before and after stack alloc. That's maybe inevitable and like not too bad since it also has a small number of phases, but still.

Note that in each of these cases, it absolutely is an IR issue. It's not like there were some rules that I failed to follow. It's that I designed the compiler specifically to have an IR that follows different laws in different parts of the pipeline.

The "different laws" thing really should be handled by changing the types of all the nodes around the critical passes. Unfortunately, every language I know sucks at typing if you want to avoid all the copying that the naive approach gives you. (remember especially that some types will only exist during certain phases. The most basic of these is: which token types are valid in a parse tree and which aren't? E.g. binary operator tokens usually remain valid, but parentheses disappear, whereas many new types appear. For later phases this only gets more complicated.)

Many passes in LLVM tend to assume that the IR is in some sort of poorly-defined canonical form. For example a loop transformation might bother to only handle loops where the latch and exiting block are one and the same, or assume the existence of a dedicated loop preheader. In principle, this dependency on canonical form should only be one where code is less optimal if the code isn't canonical, but compilers are big, complex things, so it should be no surprise that sometimes the noncanonical IR ends up causing crashes or miscompiles.

Do you remember a case where it did cause a crash or a miscompile, and it wasn't also a case where that same crash/miscompile could have happened under default/optimal phase order?

I can easily imagine that reordering phases increases overall likelihood of surfacing bugs that are there anyway.

I have a harder time imagining a dependence on canonical form that leads to a crash under phase reordering but is totally sound otherwise.

Offhand, I don't have any cases.

The cases where they're the most likely to occur is in the codegen pipeline, but codegen is already acknowledged to be a situation where you have non-optional passes that have to occur in a particular order, so I suspect you're already discounting those scenarios.

As a fun note, here's a talk of automatically reordering LLVM passes to optimize for code size, which seemingly worked enough to make a talk/paper about: https://www.youtube.com/watch?v=_SqWd74zG2Y

Reply on news.ycombinator.com