Update codegen for simd wasm intrinsics with LLVM 13

[alexcrichton]

Since LLVM 12 a number of simd wasm intrinsics have been updated to no
longer use custom intrinsics in LLVM but rather pattern-match other
types of IR. This commit updates our support in std::arch to use the
patterns the LLVM recognizes, removing usage of custom intrinsics where
possible.

Affected intrinsics are:

• v128_load{32,64}_zero
• v128_load{8,16,32,64}_lane
• v128_store{8,16,32,64}_lane
• i8x16_popcnt
• f{32x4,64x2}_p{min,max}
• *_extmul_*

Documentation around the load/store intrinsics have been updated to
indicate that the safe alternatives are suitable replacements.

Closes #1202

[rust-highfive]

r? @Amanieu

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

[alexcrichton]

Looks like armv7 is also failing with:

failures:

---- core_arch::arm_shared::neon::assert_vgetq_lane_p64_vmov_32 stdout ----
disassembly for stdarch_test_shim_vgetq_lane_p64_vmov_32: 
	 0: vmov r0, r1, d0
	 1: ldr r2, [pc, #16] ; 35a38 <stdarch_test_shim_vgetq_lane_p64_vmov_32+0x1c>
	 2: ldr r3, [pc, #16] ; 35a3c <stdarch_test_shim_vgetq_lane_p64_vmov_32+0x20>
	 3: add r2, pc, r2
	 4: ldr r3, [pc, r3]
	 5: str r2, [r3]
	 6: bx lr
	 7: .word 0x000e52b4
	 8: .word 0x0019a568
thread 'core_arch::arm_shared::neon::assert_vgetq_lane_p64_vmov_32' panicked at 'failed to find instruction `vmov.32` in the disassembly', crates/stdarch-test/src/lib.rs:154:9
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

---- core_arch::arm_shared::neon::assert_vgetq_lane_s64_vmov_32 stdout ----
disassembly for stdarch_test_shim_vgetq_lane_s64_vmov_32: 
	 0: vmov r0, r1, d0
	 1: ldr r2, [pc, #16] ; 35a84 <stdarch_test_shim_vgetq_lane_s64_vmov_32+0x1c>
	 2: ldr r3, [pc, #16] ; 35a88 <stdarch_test_shim_vgetq_lane_s64_vmov_32+0x20>
	 3: add r2, pc, r2
	 4: ldr r3, [pc, r3]
	 5: str r2, [r3]
	 6: bx lr
	 7: .word 0x000e533f
	 8: .word 0x0019a51c
thread 'core_arch::arm_shared::neon::assert_vgetq_lane_s64_vmov_32' panicked at 'failed to find instruction `vmov.32` in the disassembly', crates/stdarch-test/src/lib.rs:154:9

---- core_arch::arm_shared::neon::assert_vgetq_lane_u64_vmov_32 stdout ----
disassembly for stdarch_test_shim_vgetq_lane_u64_vmov_32: 
	 0: vmov r0, r1, d1
	 1: ldr r2, [pc, #16] ; 358c8 <stdarch_test_shim_vgetq_lane_u64_vmov_32+0x1c>
	 2: ldr r3, [pc, #16] ; 358cc <stdarch_test_shim_vgetq_lane_u64_vmov_32+0x20>
	 3: add r2, pc, r2
	 4: ldr r3, [pc, r3]
	 5: str r2, [r3]
	 6: bx lr
	 7: .word 0x000e4feb
	 8: .word 0x0019a6d8
thread 'core_arch::arm_shared::neon::assert_vgetq_lane_u64_vmov_32' panicked at 'failed to find instruction `vmov.32` in the disassembly', crates/stdarch-test/src/lib.rs:154:9


failures:
    core_arch::arm_shared::neon::assert_vgetq_lane_p64_vmov_32
    core_arch::arm_shared::neon::assert_vgetq_lane_s64_vmov_32
    core_arch::arm_shared::neon::assert_vgetq_lane_u64_vmov_32

I've pushed a commit to disable fail-fast to let it get to finishing wasm intrinsics on CI, and I'll open an issue when CI finishes.

[alexcrichton]

Ok the run finished at https://github.com/rust-lang/stdarch/actions/runs/1156300010 and looks like armv7 is the only one that's failing. I've reverted the fail-fast bit but I believe this otherwise fixes the wasm bits for llvm 11

[Amanieu]

I'll look into the armv7 issues separately, it seems to just be changed codegen from the new LLVM.