`llvm.memmove` not implemented

[newgpudev]

I am encountering code where Clang generates llvm.memmove, and this intrinsic is not implemented. As a workaround, I've temporarily handled llvm.memmove using the same lowering logic as llvm.memcpy, via COPY_BYTES. This isn't semantically correct for overlapping memory regions, but it unblocks compilation for now.

shady/src/frontend/llvm/l2s_instr.c

Line 523 in d31c4ec

if (shd_string_starts_with(intrinsic, "llvm.memcpy")) {

I added || shd_string_starts_with(intrinsic, "llvm.memmove") inside the if branch.

[Hugobros3]

Can you post a source file where Clang would generate memmove ? I haven't seen it in my testing so far.

[newgpudev]

I am getting memmove from cases like this:

#include <stdint.h>
#include "shady.h"

struct Vec3
{
    // Need at least 12 bytes here.
    float x, y, z;
};

compute_shader local_size(3, 1, 1) void test_memmove(Vec3* in, Vec3* out)
{
    int id = vcc::gl_GlobalInvocationID.x;

    // This generates memmove.
    Vec3* ptr = &in[id];
    Vec3 val = *ptr;
    out[id] = val;
}

I assume Clang somehow can't confirm that the source/dest regions are non-overlapping.
Doing out[id] = in[id] does generate a memcpy instead of a memmove. I am aware that this code can be optimized to that single assignment but I am working with an existing codebase that also contains a lot of generated C++ code like this and I do not want to modify them.

I am using Clang 18.1.3. Thanks a lot!