Comment by gumby
2 years ago
IMHO the best way to think of it (well, it's how I have long thought of it) is two lemmas:
1 - The compiler has more "fingers" than a human does. Back when I wrote programs in assembly I would use printout to keep track of what I was doing and for debugging, and so would often put a finger on the paper to mark where a jump was and then go to the destination to see if it matched up, etc. This process isn't at all scalable, so the compiler is inevitably going to do better than I ever could on anything more than something trivial.
2 - I know more about the program constraints than I could ever express (much less be bothered expressing) in a HLL. "All the values will be less than 16" or "I can use this memory location for something else after I've read from it" or who knows what. So sometimes I can chop stuff out or even rewrite the compiler output locally to speed it up.
Also I can only do this for a few architectures...and with modern x86 there are so many variants with all sorts of microoptimization affordances that I can rarely beat the compiler (this is a corollary of lemma 1)
A compiler is a combinatorial optimizer (think bin-packing). In general, optimizers/solvers basically search for the best solution. Most production compilers don't have solvers in them, they use heuristics instead, but even the best solvers use tons of heuristics. Naturally a computer will search/try heuristics faster and more thoroughly than you but sometimes you can do better because performant searching is all about "knowing where to look".
Modern compilers are not doing much searching in general. It's mostly apply some feed-forward heuristic to determine whether to apply a transformation or not.
I think a slower, search based compiler could have a lot of potential for the hottest parts you're willing to spend exorbitant time on a search.
I have heard search based compiler optimization called "superoptimization"[1]. It seems interesting, but as far as I know has not seen much industrial use.
1. https://blog.regehr.org/archives/2578
3 replies →
I understand the compiler Microsoft uses to build release versions of Windows, Office etc is like this and can take days to run.
3 replies →
> Modern compilers are not doing much searching in general.
This is false. Any compiler that does register allocation and instruction scheduling (all of them) is searching for an optimal (or just good enough) solution to an optimization problem.
9 replies →
I am going to be the token programming language researcher and say that what you really want is a dependently typed assembly language that your dependently typed higher level language lowers to. One school of thought that has yet to bear fruit in “mainstream“ programming, but gives tantalizing hints of what is possible, is expressing increasing amounts of your programs constraints in the type system, thereby informing the compiler precisely about your intent.
> want is a dependently typed assembly language
Doesn't make any sense. The "type constraints" on assembly operands (registers and numbers) is the ISA and thus those constraints are combinatorial not logical
You misunderstand. If I, at the assembly level, could express things like "this register ranges from 0 to 4"—I might be able to provide ever-lower levels of the stack with more information about my intent. Maybe the ISA can now do something clever to further optimize my program. Now, there will likely always be a time when all types are erased, but if we can push that erasure lower and lower in the stack, then compilers have more and more information about what we actually want them to do, provided our type systems are rich enough to express that intent.
2 replies →
There's the reason "token programming language researchers" are incapable of understanding modern computer architectures: a huge gap where would have been EE education.
1 reply →
While I do think there are a lot of reasons to why one should not write in assembly, time-critical base routines can often benefit a lot from being written in assembly (take GMP's mpn-module for instance). However, this puts some constraint on how the code should be formatted (in GMP's case, this means that everything the mpz-module does is derived from the mpn-module), which cannot be said for all applications.