[WebAseembly] Zero-length checks for constant-sized bulk memory ops

[SingleAccretion]

I noticed this a bit surprising code quality regression while trying out the new LLVM that uses bulk memory by default.

Godbolt link.

C source:

void test_fill(int* d, int* s)  {
    __builtin_memset(d, 0, 128);
}

void test_copy(int* d, int* s)  {
    __builtin_memcpy(d, s, 128);
}

Output:

test_fill(int*, int*):
        block           
        i32.const       128
        i32.eqz
        br_if           0
        local.get       0
        i32.const       0
        i32.const       128
        memory.fill     0
        end_block
        end_function

test_copy(int*, int*):
        block           
        i32.const       128
        i32.eqz
        br_if           0
        local.get       0
        local.get       1
        i32.const       128
        memory.copy     0, 0
        end_block
        end_function

Notice the useless checks.

[llvmbot]

@llvm/issue-subscribers-backend-webassembly

Author: None (SingleAccretion)

I noticed this a bit surprising code quality regression while trying out the new LLVM that uses bulk memory by default.

Godbolt link.

C source:

void test_fill(int* d, int* s)  {
    __builtin_memset(d, 0, 128);
}

void test_copy(int* d, int* s)  {
    __builtin_memcpy(d, s, 128);
}

Output:

test_fill(int*, int*):
        block           
        i32.const       128
        i32.eqz
        br_if           0
        local.get       0
        i32.const       0
        i32.const       128
        memory.fill     0
        end_block
        end_function

test_copy(int*, int*):
        block           
        i32.const       128
        i32.eqz
        br_if           0
        local.get       0
        local.get       1
        i32.const       128
        memory.copy     0, 0
        end_block
        end_function

Notice the useless checks.

[dschuff]

Hm yeah the check was added in 1184458 because the Wasm memory instructions trap on out of bounds even when the length is 0, but memcpy itself does not. But clearly we are missing an optimization to remove the checks when they are constant.