[RISCV] Unsigned atomics use signed load + zero-extension instead of unsigned loads

[tom-rein]

The following code (godbolt):

#include <atomic>
unsigned char load(std::atomic_uchar &x) {
    return x.load(std::memory_order_relaxed);
}

Compiles into:

load(std::atomic<unsigned char>&):
        lb      a0, 0(a0)
        andi    a0, a0, 255
        ret

Instead of:

load(std::atomic<unsigned char>&):
        lbu     a0, 0(a0)
        ret

Same is the case for unsigned short and signed atomics casted to unsigned.
For unsigned int signed load is correct because of the ABI, but when it is zero-extended, llvm fails to combine the zero-extension with the load.
Non-atomic loads correctly result in unsigned loads.

[topperc]

This has been fixed on trunk and will be in LLVM 21. https://godbolt.org/z/9qExz3ccv

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Tom Reinhardt (tom-rein)

The following code ([godbolt](https://godbolt.org/z/raqnhrGxa)): ```C++ #include <atomic> unsigned char load(std::atomic_uchar &x) { return x.load(std::memory_order_relaxed); } ```

Compiles into:

load(std::atomic<unsigned char>&):
        lb      a0, 0(a0)
        andi    a0, a0, 255
        ret

Instead of:

load(std::atomic<unsigned char>&):
        lbu     a0, 0(a0)
        ret

Same is the case for unsigned short and signed atomics casted to unsigned.
For unsigned int signed load is correct because of the ABI, but when it is zero-extended, llvm fails to combine the zero-extension with the load.
Non-atomic loads correctly result in unsigned loads.