Rename no_split_stack to no_stack_check, and add a -C option

[kmcallister]

r? @brson

Fixes #16980.

[alexchandel]

Test failures:

...
test result: FAILED. 1242 passed; 7 failed; 23 ignored; 0 measured

[kmcallister]

Apparently the LLVM 3.4 tests always fail on Travis and it isn't an issue. But the 3.3 tests only ran for 36 seconds...

[alexchandel]

My bad, I suppose it's passing then.

[huonw]

Needs a rebase.

[kmcallister]

Rebased.

[brson]

@alexcrichton Are you ok with this?

[alexcrichton]

Looks good to me!

[alexchandel]

Bors test failures:

maketest: no-stack-check
----- /Users/rustbuild/src/rust-buildbot/slave/auto-mac-64-nopt-t/build/src/test/run-make/no-stack-check/ --------------------
------ stdout ---------------------------------------------
DYLD_LIBRARY_PATH="/Users/rustbuild/src/rust-buildbot/slave/auto-mac-64-nopt-t/build/obj/x86_64-apple-darwin/test/run-make/no-stack-check:/Users/rustbuild/src/rust-buildbot/slave/auto-mac-64-nopt-t/build/obj/x86_64-apple-darwin/stage2/lib:" /Users/rustbuild/src/rust-buildbot/slave/auto-mac-64-nopt-t/build/obj/x86_64-apple-darwin/stage2/bin/rustc --out-dir /Users/rustbuild/src/rust-buildbot/slave/auto-mac-64-nopt-t/build/obj/x86_64-apple-darwin/test/run-make/no-stack-check -L /Users/rustbuild/src/rust-buildbot/slave/auto-mac-64-nopt-t/build/obj/x86_64-apple-darwin/test/run-make/no-stack-check --emit asm attr.rs
! grep -q morestack /Users/rustbuild/src/rust-buildbot/slave/auto-mac-64-nopt-t/build/obj/x86_64-apple-darwin/test/run-make/no-stack-check/attr.s

------ stderr ---------------------------------------------
make[1]: *** [all] Error 1

[alexchandel]

@kmcallister Any ideas?

[alexcrichton]

ping @kmcallister, any updates on this?

[alexchandel]

I rebased @kmcallister's branch locally and ran the tests (OS X). The problem is that, in src/test/run-make/no-stack-check, the no_stack_check function calls test::black_box, and test::black_box naturally still has a call to morestack. This call is detected by ! grep -q morestack $(TMPDIR)/attr.s which assumes the call came from the no_stack_check function and thus was erroneously generated, and the test fails.

Changing the call in attr.rs and flag.rs from test::black_box(x.as_slice()); to unsafe {asm!("" : : "r"(&x.as_slice()))} and adding #![feature(asm)], i.e. inlining test::black_box by hand, fixes this. @alexcrichton not sure if there is a better approach?

But this could be merged after a rebase and the aforementioned fix.

[kmcallister]

Wow, thanks for tracking that down! 💚

Since the code doesn't need to pass the linker, I'll go with this solution:

#![crate_type="lib"]

extern {
    fn black_box(ptr: *const u8);
}

pub unsafe fn foo() {
    // Make sure we use the stack
    let x: [u8, ..50] = [0, ..50];
    black_box(x.as_ptr());
}

[alexchandel]

No problem, I'm excited for this! :)

[kmcallister]

I fixed the test issue and rebased. I also changed the flag description to "disable checks for stack exhaustion" to avoid confusion with features like -fstack-protect that check for overflows of buffers on the stack.

[kmcallister]

Some kind of network failure on Windows? I don't even know