Modify the implementation of d_s64 suffix instructions

[SparrowLii]

This PR revised the implementation of the d_s64 suffix neon instructions to be consistent with Clang

[rust-highfive]

r? @Amanieu

(rust-highfive has picked a reviewer for you, use r? to override)

[SparrowLii]

I think vshl may have a bug in llvm. It is compiled correctly in godbolt but it failed in CI:
LLVM ERROR: Cannot select: 0x7fb987899d00: i64 = AArch64ISD::VSHL Constant:i64<1>, Constant:i32<2>, crates/core_arch/src/aarch64/neon/generated.rs:7874:5 @[ crates/core_arch/src/aarch64/neon/generated.rs:15577:32 @[ crates/core_arch/src/aarch64/neon/generated.rs:15572:5 @[ crates/core_arch/src/aarch64/neon/generated.rs:15572:5 @[ /rustc/fe72845f7bb6a77b9e671e6a4f32fe714962cec4/library/core/src/ops/function.rs:227:5 ] ] ] ] 0x7fb985815750: i64 = Constant<1> 0x7fb9d0d1e410: i32 = Constant<2> In function: _ZN4core3ops8function6FnOnce9call_once17h1592c9ba5b3050ecE

[Amanieu]

This should probably use wrapping_add

[Amanieu]

You can reproduce the error in godbolt by adding a test case which uses constant inputs: https://rust.godbolt.org/z/nPq57xxso

I can think of 2 ways to fix this:

• Just use normal shift instructions instead of intrinsics.
• Use int64x1_t for the intrinsic arguments instead of i64.

[SparrowLii]

Yes, it seems that both i64 and int64x1_t will get the same assembly instructions for vshl. My thought is to use the same link as Clang, so that llvm can be optimized to get the desired result. (for example, we would like srsra instruction in vrsrad_n_s64). And it does work

[SparrowLii]

It seems that there is something wrong with Cirrus CI, can we fix it?

error: error: 'sysinfo not supported on this platform'
info: installing component 'clippy'
error: error: 'sysinfo not supported on this platform'
info: installing component 'rust-std'
error: error: 'sysinfo not supported on this platform'