The most obvious change you'll see is the use of SSA, which has become the dominant representation in IR starting 25-30 years ago.
There's also been an increase in the importance of compiler IRs, and especially the concept of code passing through multiple IRs before reaching machine code.
Formal semantics has become more of a thing in the past decade or so. It's now routine that even weak memory models have a detailed formal model of how they work. In LLVM, it's now a requirement that you demonstrate formal correctness of new InstCombine transformations (which are essentially peephole optimizations).
The use of parser generators has really fallen into disrepute; everything has transitioned to handrolled parsers these days. Language standards themselves are starting to rely on context-sensitive keywords, which are hard to implement in a generator-based lexer/parser setup.
Optimizations have generally broadened in scope; we're now seeing whole-function level of optimization being the default way to look at stuff for analysis, and there's been a variety of techniques introduced to make whole-program optimization (aka LTO) much more tractable.
Another subtle but major shift is that compilers are increasingly reliant on inferred analysis from dumber representations over the original declared intent of the code. For example, SROA in LLVM (which breaks up structs so that individual fields can be independently allocated in registers) relies not on looking at the way the struct is declared but the offsets within the struct that various load and store operations use.
A final major shift is the trend towards support for ancillary tooling in the programming language space, so that the compiler isn't merely a tool that goes from source to object code, but is something that can be actively queried for information about the source code. Things like the language server, or smart formatting, or automatic refactoring tooling.
That "middle-pass" approach that will let you address many targets is still valid; the trick is finding a sufficiently robust and flexible internal representation at the right level. You also have to be able to out-guess the chip vendors where before you could go to the architect or a complete "System" book and get the real scoop, including things you shouldn't do. Oddly enough, there is simultaneously useful and completely worthless documentation scattered about the internet.
You might want to take a look at Muchnick and Jones' _Program_Flow_Analysis_ (yes, it's from 1981) but chapters 4-6 can be applied at code-generation time. How that fits modern Intel processors (for example) is unknown. Idealizing your processor as a RISC-V might be a reasonable way to proceed but in the end, you'll have to translate the code for the target -- it will be reasonably straight-forward if you drive it all from tables but it's not trivial.
https://news.ycombinator.com/item?id=40940799
> So what's different about writing a compiler in 2024 than say 10, 20, or 30 years ago?
As far as I can tell, the main difference is that static single assignment (SSA) as an intermediate form was not the norm 30 years ago, but it is nowadays. Also, in newer books, it's more common to go over global register allocation now, whether that's graph coloring or linear scan register allocation. If you read old compiler books, the main optimizations they talk about are use-def chains, moving computations out of loops, and using the local and tree-based Sethi-Ullman register allocation algorithm.
Today most languages are front-ends for LLVM IR, but LLVM is very slow and takes a long time to optimize. Many new languages target x86/arm directly with their own weakly optimized backends, and output an LLVM IR for "release builds".
Not sure about Jai.
Now-a-days, the difference between "big compiler optimized" and "little compiler not optimized" can be quite dramatic; but, is probably no more than 4x — certainly within range of the distinction between "systems programming language" and "high tuned JITted scripting language". I think most people are perfectly fine with the performance of highly-tuned scripting languages. The result is that all of the overhead of "big compiler" is just ... immaterial; overhead. This is especially true for the case of extremely well-tuned code, where the algorithm and — last resort — assembly, will easily beat out the best optimizer by at least an order-of-magnitude, or more.
Repeating others, today’s compilers are really just “optimizing compilers”, there is no room for toying in production environments.
If you find some time to go through gcc bugzilla you'll find shockingly simple snippets of code that miscompiled (often by optimization passes), with fixes never backported to older versions that production environments like RHEL are using.
I still insist that a production grade compiler can’t leave performance on table. Which is where the current battlefield is.
The techniques employed are very similar
in summary, optimizing compilers for c or pascal or zig or rust or whatever can only be used for code where considerations like compatibility, ease of programming, security, and predictability are more important than performance
probably the vast majority of production code is already in python and javascript, which don't even have reasonable compilers at all
Main post is about a C compiler, post I responded is saying
> When your computer was anemic, and could barely do the tasks required for it
which I could only interpret as a machine that can run a single process at a time, so not really about gpus or what.
Were they? GCC abandoned bison in favour of their own parser relatively recently.
so gcc has literally been using a parser-generator-generated parser for c for more than half its existence, at which point it had already become the most popular c compiler across the unix world and basically the only one used for linux, which had already mostly annihilated the proprietary unix market. it was also imposingly dominant in the embedded space
and i think that kind of development path is fairly typical; getting a parser up and running is easier with a parser generator, but it can be tricky to get it to do strange things when that's what you want (especially with lalr, less so with peg)
and, while we're talking about ocaml, ocaml does use ocamllex and ocamlyacc for its own parser
so, while you can certainly do without parser generators, they have very commonly been used for making real-world programming languages. almost every programming language anyone here has ever heard of was first implemented with a parser generator. the main exceptions are probably fortran, cobol, algol, lisps, c, and pascal
I meant to say the idea of a parser generator is a solution to a problem that that real world langs don't really have. When writing a programming language, your issue isnt how much time the parser is going to take to write, or how complex it's going to be. The parser is a relatively trivial part of the problem.
Due to language designers often being taught to develop langauges in this fashion, many have relied on these tools. But the modern view of compliers as "programming langauge UIs" and the focus on DX, i'd argue its actively pathological to use a parser generator.
Much academic work has, til recently, focused on these areas -- whereas today, the bulk of the difficulty is in understanding SSA/LLVM/ARM/Basic Optimizatiosn/etc. details which are "boring, circumstantial" etc. and not really research projects. I was just pointing this out since a lot of people, myself included, go down the EBNF parser-generator rabbit hole and think inventing a langauge is some formal exercise -- when the reality is the opposite: it's standard programming-engineering work.
but no language starts out as a 'real world lang'; every language is initially a toy language, and only becomes a 'real world lang' in the unlikely case that it turns out to be useful. and parser generators are very useful for creating toy languages. that's why virtually every real world lang you've ever used was implemented first using a parser generator, even if the parser you're using for it now is handwritten
having a formally defined grammar is also very helpful for dx features like syntax highlighting and automated refactoring
maybe you're talking about stuff you haven't released?
I will soon likely create a probabilistic programming language and compiler.
But how can you have assurance which grammar it defines, or that it even defines a well-defined grammar?
I’m well aware that some languages don’t bother defining a proper grammar, or define it without having a mechanism to ensure their implementation matches it, but lacking that assurance is exactly the drawback of not using a parser generator.
The f18 compiler’s parser uses parser combinations to construct a backtracking recursive descent parser that builds a parse tree for the whole source file before doing any semantic analysis. This approach allows good error recovery without having to revert any updates to the symbol table.
i assume that by 'parser combinations' you mean parser combinators
what i meant about fortran is that the first fortran compiler didn't use a parser generator