llvm dev - Aug 2012 - [LLVMdev] Generate data16 assembly instruction for TLS with PIC

Hi all,

System: x86-64 ubuntu 11.04
LLVM: 3.0
gcc: 4.5.2

I declare a thread_local global variable and access it in a function in
llvm IR.
for example,

@my_value = linkonce thread_local global %dummy* null

define void @test1()  {
entry:
    %load_my_value = load %dummy** @my_value
    ....
}

After that, I use the following command:

bash$ llc -relocation-model=pic test.ll

And I have test.s which is an assembly file.

I tried to build it with gcc, but I have the following warning

bash$ gcc -c test.s
test.s: Assembler messages:
test.s:43: Warning: stand-alone `data16' prefix

I just wonder why I have data16 instruction in my assembly file.
(ps. If I build it with "llc test.ll", the assembly file will not have
data16 instruction)

I search the llvm project and find some clues in this
file: lib/Target/X86/X86MCInstLower.cpp
static void LowerTlsAddr(...)
{
    bool is64Bits = MI.getOpcode() == X86::TLS_addr64;
    ....
    if (is64Bits) {
        MCInst prefix;
        prefix.setOpcode(X86::DATA16_PREFIX);
        ....
    }
    ....
}

I think it is the logic to generate data16 instruction, but I wonder why
gcc report warning. Is it a LLVM's bug?
Thanks for your time in advance :D
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120818/d2f80ca6/attachment.html>

Hey Wei-zhi,

I also had this issue with the X86 backend. And, unfortunately, searching
for the warning did not turn up too much information. My solution was to
emit a raw 0x66 byte in place of the data16 prefix. I believe those are
equivalent and the assembler is much happier about it...
> #if LLVM_31_MERGE
>    SmallString<11> byte;
>    byte = "\t.byte 0x66";
>    OutStreamer.EmitRawText(byte);
>#else /* LLVM_31_MERGE */
>    prefix.setOpcode(X86::DATA16_PREFIX);
>    OutStreamer.EmitInstruction(prefix);
>#endif /* LLVM_31_MERGE */
-Cameron

On Fri, Aug 17, 2012 at 12:27 PM, Wei-zhi Liao <giggle at zillians.com>
wrote:
...
>
> I think it is the logic to generate data16 instruction, but I wonder why
> gcc report warning. Is it a LLVM's bug?
> Thanks for your time in advance :D
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120817/93d75f02/attachment.html>