LLVM -O2 compile time regression

[seanmonstar]

rustc 1.1.0-nightly (9ecc9896d 2015-05-10) (built 2015-05-11)

time: 58.110 LLVM passes

vs

rustc 1.1.0-nightly (c2b30b86d 2015-05-12) (built 2015-05-13)

time: 785.935 LLVM passes

running cargo rustc --release -- -Z time-passes --test on hyper@0.5.0

Full output at https://gist.github.com/seanmonstar/390e04d63ad28f234050

[huonw]

cc @nrc (There's a slight suspicion that this may be caused by #24619, since that's touching trans/codegen.)

[seanmonstar]

ran with rustc 1.1.0-nightly (4b88e8f63 2015-05-11) (built 2015-05-12), got time: 58.253 LLVM passes

[nrc]

cc @eddyb

[eddyb]

@seanmonstar Could you provide .ll files for before and after? Hopefully cargo rustc --emit=llvm-ir should work now.

[seanmonstar]

@eddyb https://gist.github.com/seanmonstar/b302a74d1e0aea3a59e2

[dotdash]

Looks like this is caused by lilyball@e1e34e9 -- the assume intrinsics no longer get simplified into nonnull metadata on load instructions and cause massive slowdowns.