Remove redundant integer suffixes.

[eddyb]

As #21511 has shown, there are lots of cases in which integer suffixes aren't needed at all.
Usually audits are needed to make sure the intended type is inferred, and disabling the fallback to i32 wouldn't really work for tests and whatnot.

It should be easy to modify the compiler to dump the equivalent of --pretty=typed (with pretty-printing showing no suffixes on expressions) into some temporary directory, then compare the results from two make check runs, one with all the suffixes removed.

[gchp]

I'd like to look into this one, though I'm not sure how to modify the compiler in the way that you mention. Any suggestions on where to start?

[eddyb]

The entry-point into the pretty printer seems to be this call in driver::pretty::pretty_print_input.
call_with_pp_support handles --pretty=typed by running analysis passes itself, which isn't exactly what we want.
By combining those two, we get

let src_name = source_name(input);
let src = sess.codemap().get_filemap(&src_name)
                        .src.as_bytes().to_vec();
// For this you'd have to make TypedAnnotation's analysis field public.
let annotation = TypedAnnotation { analysis: analysis };
pprust::print_crate(sess.codemap(),
                    sess.diagnostic(),
                    ast_map.krate(),
                    src_name.to_string(),
                    &mut MemReader::new(src),
                    box old_io::File::create(&pp_out).unwrap(),
                    &annotation
                    true).unwrap();
let analysis = annotation.analysis;

which can be inserted at this point in compile_input.

pp_out_path is still missing, but I think the following method should produce an unique path for each rustc invocation during the build:

let pp_out = outputs.with_extension("rs");
let pp_out_file = pp_out.as_str().expect("non-utf8 pp output file").replace("/", "--");
// If this doesn't work, hardcode an absolute path here.
let pp_out_path = Path::new(concat!(env!(CFG_BUILD_DIR), "int-suffix-pp")).join(pp_out_file);

Aside from that, you would also need to hack the pretty-printer so it doesn't output integer suffixes - should be easy to treat them just like the unsuffixed cases.

[eddyb]

Since I didn't hear back from @gchp, I implemented what I mentioned in a branch of mine - I intend to open a PR soon.
I did go a bit crazy and cleaned up a few things in the process. I test it like so:

rm -rf test-pretty-dump; mkdir test-pretty-dump
# compile stage1 rustc first, with no flags, to avoid errors as stage0 doesn't have them
make x86_64-unknown-linux-gnu/stage1/bin/rustc
make check-stage1 RUSTFLAGS="-Z unstable-options -Z pretty-keep-going -Z pretty-dump-dir=test-pretty-dump --xpretty=typed,unsuffixed_literals"

This gives me a bunch of files (I didn't run this build for long - an almost full one with tests had over 2600):

home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--alloc-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--arena-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--collections-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--core-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--flate-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--getopts-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--graphviz-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--libc-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--log-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--rand-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--rustc_bitflags-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--serialize-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--std-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--term-4e7c5e5c.rs
home--eddy--Projects--rust--x86_64-unknown-linux-gnu--stage1--lib--rustlib--x86_64-unknown-linux-gnu--lib--unicode-4e7c5e5c.rs

An example from libcollections:

// source
static TAG_CONT_U8: u8 = 128u8;
// pretty-printed
static TAG_CONT_U8: u8 = (128: u8);

The flag-based approach has two big problems, though: it messes up pretty tests and it wouldn't work with cargo.
I'll attempt adding support for an env var of the form RUSTC_PRETTY_DUMP=mode:dir that is equivalent to --xpretty=mode -Z pretty-keep-going -Z pretty-dump-dir=dir, would be ignored when --pretty is already passed, and it would also work with cargo.

My invocation would thus become:

# no bootstrapping issues!
make check-stage1 RUSTC_PRETTY_DUMP="typed,unsuffixed_literals:test-pretty-dump"

[eddyb]

I've rebased my branch and added the env var functionality - oh and I think I have an almost-perfect way to remove integer literal suffixes everywhere:

sed -i -r -e 's/\b(0x[0-9a-fA-F_]*[0-9a-fA-F]|0o[0-7_]*[0-7]|0b[01_]*[01]|[0-9]([0-9_]*[0-9])?)_*([ui](8|16|32|64|s(ize)?)?|uint|int)/\1/g' **.rs

[eddyb]

This is working out! I have some stats (cc @huonw @aturon @alexcrichton):

• same type before and after suffix removal: 1120 lines (eddyb/rust@19354d9)
• harmless fallback to i32: 302 lines (eddyb/rust@3586312)
• remaining (dangerous to remove) suffixes: 627 lines

How I did the diffs:

# make a fresh git working dir
cd ..; mkdir rust-unsuffix-data; cd rust-unsuffix-data; git init
# add a commit for the "before" dataset (sed removes comments which are only diff noise)
cp ../rust/before-int-suffix-removal/* .; sed -i -e 's|//.*$||' *.rs; git add *; git commit -m 'before'
# add a commit for the "after" dataset, on top of "before" (same sed to remove comments)
git rm *; cp ../rust/after-int-suffix-removal/* .; sed -i -e 's|//.*$||' *.rs; git add *; git commit -m 'after'

[aturon]

@eddyb Just want to say, this is awesome 👍