core_arch::x86 : Fix the implementation of _kshift instructions

[madhav-madhusoodanan]

Summary

The _kshiftri_mask32, _kshiftri_mask64, _kshiftli_mask32 and _kshiftli_mask64 intrinsics in core_arch::x86 module do not handle a critical edge case when the shift amount exceeds the bit-length of the input argument.

EDIT: updated the 8-bit and 16-bit versions of the same.

Current behaviour

For _kshiftri_mask32 and _kshiftli_mask32 intrinsics, when the shift amount (passed as a const-generic argument) exceeds 32, the shift amount applied to the argument becomes shift % 32.

Similar is the case with the 8-bit, 16-bit and 64-bit variants of the same.

Godbolt link for the 32-bit variant with minimal replication.
Godbolt link for the 64-bit variant with minimal replication.

Expected behaviour

When the shift amount exceeds 32 (for the 32-bit versions) or 64 (for the 64-bit versions) the result becomes zero, in line with Intel's documentation on the same.

Similar is the case with 8-bit and 16-bit versions.

Fix

Use unbounded_shr() and unbounded_shl() functions instead of >> and << operations.

Godbolt link that shows the corrected version of the 32-bit implementation.
Godbolt link that shows the corrected version of the 64-bit implementation

r? @sayantn
cc: @folkertdev

[sayantn]

There are 8- and 16-bit versions too in avx512dq, check them too pls

[madhav-madhusoodanan]

Done @sayantn

[sayantn]

Thanks, I have opened #1931 to fix the other shift-related bugs