asm directive: symbolic label syntax

Clang Frontend
#1

Helllo,

I have tried to adapt gnu as code for Apple Silicon, and I have found explanations for most differences. There is one issue I still have:

In inline C assembly, the following works on the GNU toolchain:

asm (
  "MOV X4, %2\n"
  "loop: LDRB W5, [%1], #1\n"
  "CMP W5, #'z'\n"
  "BGT cont\n"
  "CMP W5, #'a'\n"
  "BLT cont\n"
  "SUB W5, W5, #('a'-'A')\n"
  "cont: STRB W5, [%2], #1\n"
  "CMP W5, #0\n"
  "B.NE loop\n“
// …
);

However, on the clang toolchain, I apparently must replace the forward label ‚cont‘ with a numeric value, like this (note: the backward label 'loop‘ still works):

asm
(
  "MOV X4, %2\n"
  "loop: LDRB W5, [%1], #1\n"
  "CMP W5, #'z'\n"
  "BGT 2f\n"
  "CMP W5, #'a'\n"
  "BLT 2f\n"
  "SUB W5, W5, #('a'-'A')\n"
  "2: STRB W5, [%2], #1\n"
  "CMP W5, #0\n"
  "B.NE loop\n“
//…
);

When using as by itself and not in inline assembly, the numeric label works.

Is there a reason for this? Is this a bug? Can I find any sort of documentation on the (inline) assembly syntax for clang?

Thanks a lot!

Alex

https://github.com/below/HelloSilicon#listing-9-8

#2

Do you actually have a complete self-contained example? Try godbolt.org.

Joerg

#3

Thanks a lot for your answer, and yes, I do:

Working code (using numeric label): https://github.com/below/HelloSilicon/blob/main/Chapter%2009/uppertst4.c
Non-working (original) code: https://github.com/below/HelloSilicon/blob/4c4b2911c43644adfd3b78aee093857444f28472/Chapter%209/uppertst4.c

Compile using the makefile in the same folder, or using "clang -o uppertst4 uppertst4.c“

On godbolt, running the original code through armv8-a-clang 11.0.1 or trunk, everything seems to work fine. There is an unrelated warning, but no errors. The output shows both the loop and the cont labels: https://godbolt.org/z/xY5afcYdW

Compiling the code with the symbolic cont label on an M1 using clang-1300.0.29.3 produces the errors:

uppertst4.c:22:4: error: conditional branch requires assembler-local label. 'cont' is external.
                "BGT cont\n"
                 ^
<inline asm>:4:1: note: instantiated into assembly here
BGT cont
^
uppertst4.c:24:4: error: conditional branch requires assembler-local label. 'cont' is external.
                "BLT cont\n"
                 ^
<inline asm>:6:1: note: instantiated into assembly here
BLT cont
^
uppertst4.c:28:4: error: conditional branch requires assembler-local label. 'loop' is external.
                "B.NE loop\n"
                 ^
<inline asm>:10:1: note: instantiated into assembly here
B.NE loop
^

Notably, now for both loop and cont, while it is sufficient to replace cont with a numeric label.

I would be glad to find out if I am doing it wrong, or if this is an issue in clang. And now my curiosity is raised why godbolt appears to differ from clang on macOS …

Thanks again

Alex

#4

Does the behavior change if you declare “cont” as a global symbol? Does the symbol have to be “cont”? Would “.Lcont” suffice? Many targets treat leading “.L” as implicitly local symbols.

#5

Hello,

one thing I should stress: This appears to be a C-Frontend issue (and thus I am posting it here) because when using as directly, this problem does not occur: https://github.com/below/HelloSilicon/blob/main/Chapter%2005/upper.s

I have taken your suggestions: Changing the name of „cont“ to something that does not sound like a keyword does not help.
Adding a ".global _faz\n“ (or ".global faz“) does not change anything, neither does „BGT .Lfaz\n“ seem to have any effect.

Let me know if you have any other ideas

Thanks

Alex

#6

Thanks for everyone on this list!

The big difference between my native assembly code and the inline code is, that the cfe generates the directive: ‘.subsections_via_symbols’ , and I will have to do some searching to understand precisely what this does …

But because of this, indeed, labels must be prefixed with „L“. Like

        b.lt Lcont
        sub w5, w5, #32
Lcont:
        strb w5, [x9], #1

It also keeps my curios why the gcc frontend seems to behave differently here

#7

Can you post a godbolt link? that would also show which command line options you’re using.

So this is with gas and -fno-integrated-as?

#8

Hello Brian (and everyone who is reading this),

allow me to be very verbose (-vvvv) so that you, and others, can understand what I did, why I did it, and what I found:

## Motivation

The main objective was to provide an "Apple Companion Guide" to the book „Programming with 64-Bit ARM Assembly Language“ by Stephen Smith. The book uses the RaspberryPi 4, Linux and accordingly the GNU toolchain and GNU syntax. I wanted to provide both information on how Apple aarch64 devices differ, as well as modified, runnable source code. You can see all of what I did here: https://github.com/below/HelloSilicon

## The Problem

I ran into an issue when it came to an example using inline-assembly in C. Here is the original code, which works on ARM64 Linux, simply by invoking gcc on the file without any flags: https://github.com/below/HelloSilicon/blob/4c4b2911c43644adfd3b78aee093857444f28472/Chapter%209/uppertst4.c

When compiling the same exact file with clang on an M1 Mac (again, no flags), there is a trivial warning that we will ignore, and the following errors:

uppertst4.c:22:4: error: conditional branch requires assembler-local label. 'cont' is external.
               "BGT    cont\n"
                ^
<inline asm>:4:1: note: instantiated into assembly here
BGT     cont
^
uppertst4.c:24:4: error: conditional branch requires assembler-local label. 'cont' is external.
               "BLT    cont\n"
                ^
<inline asm>:6:1: note: instantiated into assembly here
BLT     cont
^
uppertst4.c:28:4: error: conditional branch requires assembler-local label. 'loop' is external.
               "B.NE   loop\n"
                ^
<inline asm>:10:1: note: instantiated into assembly here
B.NE    loop
^

## The Provisional Solution

I was able to solve the issue by replacing the `cont` label with a numeric value, 2, and the branches with `BGT 2f` and `BLT 2f`. Curiously, after this change, `loop` could stay like it was.

## Finding The Real Problem

Other than the trivial warnings, godbolt is absolutely happy with the file: https://godbolt.org/z/6s8bs4rW9

Next, I let clang create assembly output using `clang -S uppertst4.c` (on the M1 Mac) on the original, GNU, sourcefile. This also produces no errors, however trying to invoke the clang assembler `as` did:

% as uppertst4.s      
uppertst4.s:33:2: error: conditional branch requires assembler-local label. 'cont' is external.
        b.gt    cont
        ^
uppertst4.s:35:2: error: conditional branch requires assembler-local label. 'cont' is external.
        b.lt    cont
        ^
uppertst4.s:40:2: error: conditional branch requires assembler-local label. 'loop' is external.
        b.ne    loop
        ^

As I had a [working, standalone assembly file](https://github.com/below/HelloSilicon/blob/main/Chapter%2005/upper.s) which uses both `cont` and `loop`, I took a „divide and conquer“ approach: I deleted and added lines to the two files, until I would know what precisely caused the issue.

It turned out to be on the very last line, clang adds the compiler directive: `.subsections_via_symbols`.

llvm does it whenever it outputs assembly for a MachO binary, the source code tells us:


  if (TT.isOSBinFormatMachO()) {
    // Funny Darwin hack: This flag tells the linker that no global symbols
    // contain code that falls through to other global symbols (e.g. the obvious
    // implementation of multiple entry points).  If this doesn't occur, the
    // linker can safely perform dead code stripping.  Since LLVM never
    // generates code that does this, it is always safe to set.
    OutStreamer->emitAssemblerFlag(MCAF_SubsectionsViaSymbols);
  }

https://github.com/llvm/llvm-project/blob/89b57061f7b769e9ea9bf6ed686e284f3e55affe/llvm/lib/Target/AArch64/AArch64AsmPrinter.cpp#L568

The part about "Since LLVM never generates code that does this“ of course leaves out code that llvm did not generate, such as inline-assembly.

## The Real Solution

Because the C-Frontend created its own symbolic forward label, I learned that to remedy my issue, I needed to prefix `cont` with `L`, whereever using it. Here is the final file for ARM64 on Mac: https://github.com/below/HelloSilicon/blob/main/Chapter%2009/uppertst4.c

## What I still don’t know

While I did not take the time to fully understand what `. subsections_via_symbols` does, and still being a learner about these things, three questions remain:

1) What precisely was the violation of the inline code? Are there "global symbols which contain code that falls through to other global symbols“?

2) Why is the `loop` label apparently not a problem? Just to be sure, I changed it’s name to something not beginning with `l`, but that did not cause any issue either

3) Is there a substantial issue in my the standalone assembly code that I should know about?

So thank you for bearing with me. If you have any questions of input, please let me know!

Alex