invalid instruction mnemonic with Clang-Cl and FA flag

Hello Clang-Devs,

I run into a strange error with Clang-Cl - I hope I can explain it. We use Clang-Cl via the LLVM extension in Visual Studio and use it to compile our source code. One of the flags we set for the compiler is “/FA” which gives us assembly output. Everything fine so far, no problems.

I then add a simple copy loop to our code, that is copying int32 values from one address to another one. When compiling this construct Clang-Cl suddenly exited with: “error: invalid instruction mnemonic ‘movsdi’”.

However, when I remove the “/FA” flag everything compiles fine and runs fine in the end… Strangely, I couldn’t figure out what the “movsdi” mnemonic is…

The Assembly code looks like this:

.LBB15_31: # in Loop: Header=BB15_17 Depth=1

mov ecx, 32

mov edx, 4

call calloc

mov qword ptr [r12 + rbp + 96], rax

mov qword ptr [r12 + rbp + 88], rax

mov ecx, 32

mov rdi, rax

rep movsdi es:[rdi], [rsi]

jmp .LBB15_33

My Clang-Cl version is:

clang version 11.0.0 (https://github.com/llvm/llvm-project.git 176249bd6732a8044d457092ed932768724a6f06)

Target: x86_64-pc-windows-msvc

Thread model: posix

The target is the same as in the version (x86_64-pc-windows-msvc).

Does anyone know what is going on?

Kind greetings

Björn

Clang’s Intel assembler parser isn’t as robustly tested as the GNU assembler parser. This has been a recurring issue ever since we switched clang-cl’s assembly output to the Intel dialect:
https://bugs.llvm.org/show_bug.cgi?id=36676

I see the string repmovsdi here:
https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/X86/X86InstrCompiler.td#L383

I think it’s just a typo, let me see if I can fix it…