Make X86 intrinsics const

[sayantn]

Tracking issue: rust-lang/rust#149298
closes: #1076

(This uses stdarch_const_x86 for the intrinsics only, and stdarch_const_helpers for everything else, so we can go for aarch64 next)

Ik the diffcount is a lot, but most of it is just boilerplate. It is easier to review on a per-commit basis, and the following commits contain the "interesting" bits (others are just soulless adding of const keyword everywhere)

• 7134a49 (modifies the integer reduce_add and reduce_mul intrinsics to use simd_reduce_add_ordered and simd_ordered_mul_ordered, respectively. As integer algebra is associative, it doesn't matter for correctness, and LLVM will use the most performant arrangement. The reason is that rustc_const_eval only implements these ordered reduction intrinsics, unordered reduction seems to be tricky to implement in const_eval)
• 1f94820 (rewrites some test helper functions for const)
• 90ca47b (modifies #[simd_test] quite a bit, adds a new "no arguments" mode to ease testing of intrinsics that require no target features and adds support for const testing)
• 962e4ab and 36905a2 (contains some modifications to avoid pointer-to-usize casts and for loops)

[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

[Amanieu]

Is this actually needed? These functions aren't exported and shouldn't need stability attributes.

[sayantn]

Yes, unfortunately const-stability attributes are needed even for private functions

[Amanieu]

Since this is only used in tests, could we always us the ct version here? That would simplify the code.

[sayantn]

The ct version doesn't print the left and right arguments, because debug-printing is not const yet. I wanted to print the values in rt at least, so had 2 separate impls.

Also one more problem is the ct version uses simple float equality, which differs from intel's definition for denormals and nans. In the functions we actually call the const version, we don't have such problems because those have standard float conventions.