LLVM's maxnum builder is broken in LLVM6

[gnzlbg]

This was fixed in this LLVM commit: llvm-mirror/llvm@e271f8c

Once we upgrade to LLVM 7 a couple of fail tests are going to start passing because of this.

This issue is referenced in all parts of the code that need updating.

[cuviper]

When testing with LLVM master, I just hit a test "failure" on run-fail/simd-intrinsic-float-minmax.rs, because the run actually passed! IMO this is the wrong use of run-fail -- these should be tests that are supposed to fail, not just "known to fail because of a bug."

Wouldn't it be better to make that test a run-pass, marked with min-llvm-version 7.0? Or maybe run-pass with the known-broken parts FIXME-commented out?

[gnzlbg]

@cuviper that makes sense.

Wouldn't it be better to make that test a run-pass, marked with min-llvm-version 7.0?

Now that we know the test passes with master, then yes. But I did not test it with master, so I did not know. I wanted that to be loud the first time somebody did that so that we can move the test to run-pass. I did not thought about somebody testing with master though, I thought this would just happen when we tried to upgrade to LLVM7. Sorry about that.

Now that we know the fix in master really did fix this, moving it to run-pass with min-llvm-version 7.0 makes sense.

[gnzlbg]

@cuviper I'll fix that this weekend if somebody doesn't beat me to it.

[gnzlbg]

Closed in #50521