Don't field-project (.0) into SIMD types

[scottmcm]

Part of rust-lang/compiler-team#838, as people have already mentioned in comments like

stdarch/crates/core_arch/src/macros.rs

Lines 112 to 113 in f3be311

	// NOTE: This deliberately doesn't just use `&self.0` because it may soon be banned
	// see https://github.com/rust-lang/compiler-team/issues/838

[rustbot]

r? @Amanieu

rustbot has assigned @Amanieu.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[scottmcm]

@folkertdev, since you've mentioned s390x recently, do you know why I might be getting s390x failures from this?

Seems like they're related to f16 somehow...

rustc-LLVM ERROR: Cannot select: 0x7f52ab2d3bd0: f32 = fp16_to_fp 0x7f52ab2d3cb0, /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/slice/cmp.rs:67:16 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/slice/cmp.rs:15:9 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/array/equality.rs:138:9 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/array/equality.rs:10:9 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/cmp.rs:1922:13 @[ crates/core_arch/src/simd.rs:58:17 ] ] ] ] ]
  0x7f52ab2d3cb0: i32,ch = load<(dereferenceable load (s16) from %ir.1, align 8, !alias.scope !1472, !noalias !1471), zext from i16> 0x7f52ab26b920, 0x7f52ab2d3620, undef:i64, /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/slice/cmp.rs:67:16 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/slice/cmp.rs:15:9 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/array/equality.rs:138:9 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/array/equality.rs:10:9 @[ /rustc/f9e0239a7bc813b4aceffc7f069f4797cde3175c/library/core/src/cmp.rs:1922:13 @[ crates/core_arch/src/simd.rs:58:17 ] ] ] ] ]
    0x7f52ab2d3620: i64,ch = CopyFromReg 0x7f52ab26b920, Register:i64 %2
      0x7f52ab2d3a80: i64 = Register %2
    0x7f52ab2d3230: i64 = undef
In function: _ZN74_$LT$core_arch..core_arch..simd..f16x4$u20$as$u20$core..cmp..PartialEq$GT$2eq17h65faeea59d4b4257E
error: could not compile `core_arch` (lib)

[folkertdev]

it's because f16 is just not really supported on s390x currently, e.g. even a simple f16 comparison fails to compile with a very similar error:

https://godbolt.org/z/xGTEevEjb

edit: seems like we're making progress on this

rust-lang/rust#138022
llvm/llvm-project#109164

[tgross35]

In std/build.rs there is a list of platforms that have problems with f16 and f128. Most of them are pretty quiet and only show up when you try to use specific ops, but s390x stands out since it’s basically an instant crash if half shows up in the IR.

The workaround was using #[inline] on any functions containing these types so we don’t codegen, since rust-lang/rust#133050 that’s automatic but limited. Maybe this PR is hitting a case that isn’t auto-inlined?

edit: oh, yeah, you fixed it already

[folkertdev]

yeah the #[inline] on the manual PartialEq implementation seems to have done the trick

[scottmcm]

Looks like I was overcomplicated things in 77687d6 -- I squashed everything down back to just using the arrays, which seems fine since the == isn't usually actually used, just in tests, so inlining so they don't need to actually codegen on all platforms seems to have "solved" the problems.