4 ms·
Compilers are not really complicated compared to plenty of other hard topics. They're large, but they always separate out into different passes that you just ru
by astrange 11y ago
Compilers are not really complicated compared to plenty of other hard topics. They're large, but they always separate out into different passes that you just run one after the other. You don't have to manage the program as a whole.
And optimizations don't even matter anyway. x86 runs your -O1 program as fast as your -O3, and nobody has put a real accurate model of an x86 into a production compiler for the last decade.
- mafribe 11y agonobody has put a real accurate model of an x86 That's a pretty good indication that there is a big problem. Deep at the heart of it is that processor designers don't quite understand the performance characteristics of their processors any more, so they cannot produce e.g. an abstract memory model that compiler writers could base their optimisations on. Optimisations for parallel processing, e.g. auto-parallelisation are not well developed either.
- astrange 11y agoOh, I think Intel knows it, but they consider it too much of a trade secret to even tell the compiler team about it. There are actually numerous performance counters exposed by the CPU too, which are good for dynamic benchmarking, but hardly anyone cares to use them. BTW, even though the ALU part of the CPU is extremely wide and dynamic, you still have to fit your instructions through the x86 decoder at the front, and it's worth writing a scheduler just to model that part.
- mafribe 11y agoWhy would they keep it secret? I understand that they need to keep the low level details of their CPU secret, but a memory model? All it says is when/how reads and write commute. If a CPU manufacturer were to provide a nice, high-level memory model, this would make the CPU more sellable. Why? Because compiler writers could more aggressively optimise when compiling for the CPU. I suspect that the reason they don't give out memory models is that they don't know how to do it. That's also the vibe I've been getting when I talked to people who would know.
- sklogic 11y agoTbh, it is really hard to have an "accurate" model of a very wide OoO core. It is far too dynamic, you instruction scheduler cannot predict the occupancy of the pipelines.
- mafribe 11y agoIn order to determine the meaning of a programming language, you need a platform independent memory model. This is increasingly important as CPUs are more and more concurrent and programming languages are exposing some of that concurrency to the programmer. Of course the language designer can always play it safe and choose a ridiculous memory model like sequential consistency, but that prevents many useful compiler optimisations. At the other extreme you could explicitly state that the memory model is a specific processor, but then (a) the language becomes too platform dependent and program semantics will change with every processor upgrade. (b) nobody really knows what exactly the memory model is in any meaningful sense because even the manufacturers don't know (other than saying: it is what the CPU does). That means the semantics of the language is unclear. What you want is something between the Scylla of inefficiency and the Charybdis of platform dependence. This is a difficult problem, and unsolved even for as simple a language as C, see e.g. [1]. Indeed it is unclear if memory models possessing the required attributes even exist, maybe it's contradictory to assume that they can. [1] M. Batty et al., The Problem of Programming Language Concurrency Semantics. https://www.cl.cam.ac.uk/~jp622/the_problem_of_programming_language_concurrency_semantics.pdf https://www.cl.cam.ac.uk/~jp622/the_problem_of_programming_l...