6x slower compilation for "simple" file than C

[alexcrichton]

In attempting to find some local hot spots in rustc I've been playing around with various benchmarks. If you use this script to generate a Rust and a C file (which should be equivalent)

N=200000

echo > lots-of-used.rs
for i in `seq 1 $N`; do
  echo "pub fn foo$i() {}" >> lots-of-used.rs
done
echo 'fn main() {' >> lots-of-used.rs
for i in `seq 1 $N`; do
  echo "foo$i();" >> lots-of-used.rs
done
echo '}' >> lots-of-used.rs

echo > lots-of-used.c
for i in `seq 1 $N`; do
  echo "static void foo$i() {}" >> lots-of-used.c
done
echo 'void foo() {' >> lots-of-used.c
for i in `seq 1 $N`; do
  echo "foo$i();" >> lots-of-used.c
done
echo '}' >> lots-of-used.c

I've benchmarked with:

$ rustc +beta -V
rustc 1.20.0-beta.1 (e93aa3aa8 2017-07-18)
$ sh foo.sh
$ time clang -c lots-of-used.c
clang -c lots-of-used.c  14.90s user 0.36s system 99% cpu 15.259 total
$ time rustc +beta lots-of-used.rs --emit obj --crate-type rlib
rustc +beta lots-of-used.rs --emit obj --crate-type rlib  49.62s user 37.03s system 99% cpu 1:26.96 total

The -Z time-passes output is particularly illuminating, the highest portions being:

time: 57.439; rss: 2890MB translation
  time: 51.813; rss: 2497MB translation item collection
time: 10.094; rss: 534MB LLVM passes
  time: 9.238; rss: 536MB       codegen passes [1]
time: 5.100; rss: 2405MB borrow checking
time: 4.029; rss: 1656MB        item-bodies checking
time: 2.166; rss: 1781MB const checking
time: 1.543; rss: 1069MB wf checking

It looks like the main slowdown of the the translation item collection is related to the ElaborateDrops pass?

[alexcrichton]

I fixed an "obvious omission" in #43581 but I'm not sure why we're creating such massive bit sets in the first place

[alexcrichton]

This also peaks at using 32GB (!) of memory, a massif dump here, the peak of which looks like this

[michaelwoerister]

cc @arielb1 because of ElaborateDrops.

Oh interesting, optimizing MIR now shows up as part of item collection because of on-demand.