Seems like a poor decision to use whitespace as a factor in estimating code size. Some people like to space out their code with lots of comments even if it is only a few lines long. This is my naive view, though - does anyone out there have any insight as to why semicolons alone aren't sufficient to estimate code size?
That heuristic is actually usually a very good one. However, in this case, the inline assembly contains a lot of computation in directives, so it consists of several lines to compute one small instruction.
To answer the question you actually asked rather than what you meant to ask (because the question you asked is interesting in and of itself): compilers throw away the actual source code very quickly. At the stage that inlining is considered, the program is going to consist of a sequence of instructions not unlike the C abstract machine as opposed to a more direct representation of code. The simplest way to estimate a program size is to simply count these instructions, but some instructions stand a good chance of being eliminated (say, a bitcast instruction that changes type without adjusting value), so there's usually some adjustment for "this is really expensive on this machine" or "this instruction is free."
Take for example auto-vecorization. If we have a loop that just calls a function which increments a value in an array, we may very well have a great candidate to vectorize. However, if the function isn't inlined, the auto vectorization optimization will just see a function call as opposed to an increment and won't be able to do anything with it.
Similarly, the article talks about how the LLVM inliner clearly is smarter since it seems likely they ask their built-in assembler to provide an estimate of the code size rather than using more coarse heuristics. It seems like there's clearly a better path forward for GCC.
Compile time. Compiler optimizations are a tradeoff between code size, running time, and compile time, complicated by the fact that that sneezing at your code can change performance by 40% or more and the fact that correctness comes with steep performance costs and often isn't actually desired (until it really is).
Of all the optimizations, inlining is probably the worst for the difficulty of balancing the triad of objectives and the black magic of its heuristics.
HP Labs' Dynamo project (~20 years ago) was a JIT for PA-RISC on PA-RISC, and it made most programs run faster, in large part by being able to inline functions at runtime even across library boundaries.
Was there something impractical about this strategy that caused it to end up in the dustbin of computer science history?
Latency vs Throughput. And cost of optimizations vs win of optimization.
I don't think I fully agree with this. The amount of information available to the JIT compared to an AoT compiler also depends on what input a JIT works on. In the case of Java, I'd imagine that the JVM has a strictly greater amount of information to work with than an AoT compiler because JVM bytecode carries basically all the information available in Java source code. However, a Rust JIT would have semantic information that a lower-level JIT like Dynamo wouldn't (assuming I'm understanding Dynamo right), like aliasing information. Dynamo could infer that a particular memory store probably doesn't affect a later memory load, and optimize based on that, but a Rust JIT could guarantee it, allowing for the same potential optimizations without having to insert deoptimization checks.
This is where profile-guided optimization comes into play, I think. Assuming a representative profile (and that's admittedly a big if), you can get much of the same information that a JIT gets while also getting to work with a potentially semantically-richer source.
> HP Labs' Dynamo project (~20 years ago) was a JIT for PA-RISC on PA-RISC, and it made most programs run faster, in large part by being able to inline functions at runtime even across library boundaries.
I'm curious what kind of code Dynamo was tested on, as the effectiveness of a JIT depends on the particular code patterns of the software being run. For example, JITs can help a lot with code with lots of virtual function calls or indirections that traditional compilers tend to stumble on, but I don't think a JIT would help as much for something like a cache-optimized computational kernel.
Admittedly, my point is weakened by the graph shown in an old Ars Technica article on Dynamo [0], where Dynamo showed significant improvements on the SPEC benchmark suite, which I assume doesn't rely on shared libraries. Now I'm curious whether similar differences can be shown given the nearly 20 years C/C++/Fortran compilers have had to improve.
> Was there something impractical about this strategy that caused it to end up in the dustbin of computer science history?
I'd hardly say that JITs ended up in the dustbin of computer science history; on the contrary, they're quite common. The JVM and the various Javascript JITs are probably the most well-known ones; you also have Pypy (and everything derived from RPython), LuaJIT, Julia, MATLAB, the new JITs showing up in Ruby and PostgreSQL, and the more recent push to develop new JITs (like Graal). The JVM got its JIT in 1998, and the JIT was made the default in 1.3 in 2000, so it's not like all these are recent developments.
Now, between all that, I don't know why a low-level JIT like Dynamo (if I understood the article I found [1] right) didn't become more popular. It certainly seemed like promising technology at the time. I'd love to know why it seemed to die out and whether it's worth pursuing today.
[0]: http://archive.arstechnica.com/reviews/1q00/dynamo/dynamo-3.... [1]: http://archive.arstechnica.com/reviews/1q00/dynamo/dynamo-1....
It's a heuristic. It's always going to be more poor in some respect that doing something precise. You could probably make it better in a million different ways, but it's suppose to be quick and simple.
Or is the problem very asymmetric, that inlining too-large pieces has dramatic downsides that failing to inline small pieces does not? Or is it just that the magic number was chosen decades ago, and the tradeoffs (perhaps re memory) are different now?
I guess ideally you'd run a sophisticated machine learning algorithm across a huge corpus of real-world code, on real-world devices to discover the right number... but in reality people just don't have time for that so they pick a number and move on. Can't really blame them.
Therein lies the bulk of the problem, gcc doesn't parse the assembly code. The assembly is just copied with some minor string replacements for the registry names, and it's up to the assembler to deal with it and parse it.
Yes... that's the point. For your metric counting AST nodes you need to have parsed the method. The other heuristic works on unparsed source-code. So you don't waste time parsing a method that you won't inline.
> Counting AST nodes is very quick and simple
But parsing is neither. And you need to have done that first. I think (there's a paper I'm sure but can't find it) that parse time for JS applications is very substantial.
It's not hard to think of worst-cases. Think about an enormous function that you would not want to inline. It's been JITed, or it's running in a byte code interpreter, and the AST has been discarded. In your scheme you'll have to parse it to find it's too big. Huge waste of time, if you can see it's too large from the source code.
So you can always use the size of your bytecode as an inlining heuristic (which is what Chakra does).
Of course that wouldn't work before v8 had an interpreter, but they could have used other heuristics. For example, they could have used the size of the AST as a heuristic and saved that even after the AST gets thrown away.
Edit: Actually, on some lesser-used architectures it does support using ; to start a comment, e.g. z80 [1], but not the most common ones.
[1] https://sourceware.org/binutils/docs-2.26/as/Z80_002dChars.h...