LLVM does not accept generated inline assembly with variable+offset

[JohanEngelen]

Testcase:

void foo()
{
    uint[4] a;
    asm pure nothrow @nogc {
        mov a, EAX;
        mov a+4, EBX;
    }
}

When compiled with ldc2 -c -fsanitize=address test.d, we get the error:

<inline asm>:1:16: error: unknown token in expression
        movl %ebx, 4+(%rcx)
                      ^
LLVM ERROR: Error parsing inline asm

Strangely, when I compile manually to LLVM IR and then use llc to obtain assembly output, everything works fine. So I can't reproduce it from LLVM IR onwards.

The inline asm looks like this in LLVM IR:

  call void asm sideeffect "movl %eax, $0\0A\09movl %ebx, 4+$1", "=*m,=*m,~{memory}"([4 x i32]* %3, [4 x i32]* %3)

without -fsanitize=address, the $1 resolves to ebp plus an offset; with -fsanitize=address, the $1 resolves to a register without any offset and that's invalid (i.e. "4+(%rcx)" is not valid but "4+32(%rcx)" is). At least that's what I think is what's going on.

[JohanEngelen]

(btw, this prevents compilation of druntime with -fsanitize=address :()

[JohanEngelen]

Note: to get equal assembly output from llc as from ldc2, llc must be called with -O0 explicit.

Edit: and with the explicit -O0, I can reproduce it with llc. Perhaps an LLVM bug.
Edit: no, bug in our code: See https://lists.llvm.org/pipermail/llvm-dev/2017-August/116158.html

[JohanEngelen]

Testcase without ASan:

void foo()
{
    align(32) uint[4] a;
    asm pure nothrow @nogc {
        mov a+4, EBX;
    }
}

[JohanEngelen]

LLVM needs some work to support it: https://lists.llvm.org/pipermail/llvm-dev/2017-August/116244.html