[parallel codegen] Windows linker failure

[est31]

I can reproduce this on current master (7537f95), as well as a branch I have for the i128 PR (so its not a recent regression from the last 1-2 days...).

When I set codegen-units to 0 in config.toml, and start the build with ./x.py build -j 4, the stage 0 compiler artifact build fails with:

build\x86_64-pc-windows-gnu\stage0-rustc\x86_64-pc-windows-gnu\release\deps\rustc_metadata-f72209e13ac87845.0.o:(.text+0x1491a): undefined reference to `rustc_metadata::index::_$LT$impl$u20$rustc_metadata..schema..LazySeq$LT$rustc_metadata..index..Index$GT$$GT$::iter_enumerated::_$u7b$$u7b$closure$u7d$$u7d$::hf7d447e1d3033d13'
collect2.exe: error: ld returned 1 exit status


error: aborting due to previous error

Build failed, waiting for other jobs to finish...
error: Could not compile `rustc_metadata`.

The VM has 4 threads. The error is not spurious, I could reproduce it for three times, without any try where I couldn't reproduce it.

When I set codegen-units back to the default of 1 it compiles successfully.

[est31]

cc @nrc I think I read somewhere that you are interested in parallel codegen regressions.

[alexcrichton]

cc @rust-lang/compiler

[michaelwoerister]

@alexcrichton Is codegen-units=0 a meaningful value in Cargo?

[est31]

@michaelwoerister quoting config.toml.example:

# Number of codegen units to use for each compiler invocation. A value of 0
# means "the number of cores on this machine", and 1+ is passed through to the
# compiler.

[michaelwoerister]

@est31 Thanks for the info! I'll try to reproduce the problem. Did you encounter this on other platforms too?

[est31]

No, did the same on linux, not reproducible there.

[michaelwoerister]

I can reproduce this with MinGW64. With MSVC, I don't get an error, so this is probably related to the MinGW linker somehow.

[Manishearth]

Similar link issue caused by codegen units in #38938 . On linux.

[michaelwoerister]

I cannot reproduce this anymore with current master (possibly because of #38822).

[Manishearth]

@bholley does bumping the nightly date on servo fix this?

[michaelwoerister]

I'm not sure that change has made it into the latest nightly yet.

[SimonSapin]

@Manishearth Yes.