Crystal 0.27.1 built with --release crashes immediately with OverflowError

[ysbaddaden]

If I compile 0.27.1 (or master) with --release -Dpreview_overflow, then I can't compile anything anymore, not even Crystal itself. The compiler crashes immediately:

$ make -B
CRYSTAL_CONFIG_BUILD_COMMIT="0ef090a" ./bin/crystal build -D preview_overflow -D compiler_rt --release  -o .build/crystal src/compiler/crystal.cr -D without_openssl -D without_zlib
Using compiled compiler at `.build/crystal'
Unhandled exception: Overflow (OverflowError)
Failed to raise an exception: END_OF_STACK
[0xd30ca6] ???
[0x417cfb] __crystal_raise +43
[0x429963] ???
[0x422979] ???
[0x41f388] ???
[0x4522bd] ???
[0x41bdbd] main +45
[0x2acaf91c0f45] __libc_start_main +245
[0x41416f] ???
[0x0] ???
make: *** [.build/crystal] Erreur 5

Note that I can build a working compiler with -Dpreview_overflow only (without --release).

All my Travis jobs are currently broken, since the Linux packages of 0.27.1 have been built with these options.

OS: Ubuntu 14.04
Target: x86_64-linux-gnu

[ysbaddaden]

When I say "immediately" I mean it: if I pass --stats then I don't get anything printed.

[oprypin]

Here are cases of (what I think is) this problem caught in CI:
https://travis-ci.org/oprypin/crsfml/jobs/488077739
https://travis-ci.org/myst-lang/myst/jobs/487864236
https://travis-ci.org/anykeyh/clear/jobs/488396058
Ideally we'd catch such things before release

[ysbaddaden]

@oprypin this should be fixed by #7373.

[ysbaddaden]

This is an unfortunate course of events. An unexpected bogus value with RLIMIT_STACK on Linux when run inside make led to an integer overflow in Fiber.new (main thread's main fiber). In addition, the Crystal Makefile already turns on the integer overflow checks.

Weirdly, it affects a simple hello world program, but doesn't affect the compiler built in non-release mode. It only affects the compiler when built with make release=1, which cause the compiler to crash immediately, also leading to the END_OF_STACK —maybe because the crystal runtime isn't fully initialized.

That could have been found out sooner, but compiling crystal in release mode takes a long time, so I usually only use a manually built compiler in non-release mode —I guess nightly does the same?

[bcardiff]

@oprypin catching before release today is only available as docker images which none of those projects use unfortunately.

As @ysbaddaden pointed, checks in nightly can be be improved. The compiler used during the specs is not built in release mode, and the distributable which is compiled in release mode is not actually used.

[bcardiff]

Using the compiled packages to run specs would have caught #7289 also.