[X86-64] 32-bit registers store variables of unsigned char type without zero extension

[zhaojiangkun-1]

I found that the 32-bit register stores unsigned char type variables without zero extension in x86.
However, in arm and gcc x86, there are zero extension instructions.
In gcc x86, movzx edi and dil are used.
In ARM, and x8, x0, #0xff are used to implement this function.

Why does x86 not have zero extension instructions? Is this a bug?
The related pass is as follows:
https://godbolt.org/z/qT73j94We

[Zhenhang1213]

Reduced testcase:： https://godbolt.org/z/KhqT9h5sq @phoebewang

[phoebewang]

LLVM ABI assumes caller clears the upper bits of i8/i16, so that callee doesn't need to zero extend it again. It doesn't have problem if both caller and callee compiled by Clang, but would have problem if caller by GCC.

We have tried to solve it, but not made consensus yet: https://reviews.llvm.org/D124435

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (zhaojiangkun-1)

I found that the 32-bit register stores unsigned char type variables without zero extension in x86. However, in arm and gcc x86, there are zero extension instructions. In gcc x86, `movzx edi and dil `are used. In ARM, `and x8, x0, #0xff` are used to implement this function.

Why does x86 not have zero extension instructions? Is this a bug?
The related pass is as follows:
https://godbolt.org/z/qT73j94We