Compiling "Hello World" can take a prohibitive amount of time

[bstrie]

I've noticed recently that the compilation times of even trivial programs has become relatively large. This is most noticeable in the context of my sort-of-REPL, which now exhibits noticeable pauses between entering commands due to constant recompiles. On my beefy desktop, compiling the standard HW:

fn main() {
    io::println("hello, world!");
}

takes around 0.4 seconds. I've spoken with @pcwalton briefly about this and he's acknowledged the issue and mentioned that he has ideas to mitigate it.

Here's a more severe case: a fellow on IRC (MattCoder) has noted that on a Windows XP virtual machine, Hello World takes more than a minute to compile. The machine is an Intel Core 2 Quad CPU Q8400 @ 2.66Ghz with 4GB RAM, and the virtual machine itself has 2GB. It should also be noted that the virtual machine runs VS2008 comfortably and compiles C and C# programs in the negligible time you would expect: http://postimage.org/image/o9y6d1znl/

[graydon]

Yeah. This is something we need to spend some time on before 0.4; it's sufficiently bad that we're seeing people landing a lot of code (myself included) without completing a full testsuite run. Because it takes too long. That's not cool.

(Not to say that the various improvements in the compiler that have cost it performance were bad ideas! Just that it seems like there's some cost to mop up)

[bblum]

oh, wow, i thought the increased times to build libcore and libstd were my fault for adding so much code to them.

fwiw, last time i ran -Z time-passes building one of these i remember LLVM being slower than any other pass by a factor of 20 or so.

[graydon]

non-critical for 0.6, de-milestoning

[pcwalton]

Yes, there is still too much metadata reading.

[alexcrichton]

Today, this takes ~250ms to compile. The worst offenders are:

time: 0.064 s   external crate/lib resolution
time: 0.056 s   resolution
time: 0.046 s   coherence checking
time: 0.021 s   linking

Note that this is also an incredible pain point for rusti. Currently rusti feeds input 3 times to the compiler per line of input, so it's very very slow as a result of this. While rusti has its own class of performance issues, speeding this up would certainly help speeding that up.

Nominating for production-ready

[graydon]

accepted for production-ready milestone

[pcwalton]

This is each_path badness. Those 3 passes are the ones that call each_path. We need to remove all uses of it.

[graydon]

Dupe of #4572