[LLVMdev] Asm syntax of Mips m[tf]cX coprocessor instructions

[Jeremy Fitzhardinge]

On 02/08/2013 03:47 PM, Carter, Jack wrote:
> Jeremy,
>
> Could you send/attach a small test case that demonstrates the problem?
>
> It doesn't need to go past the stage that creates a .o.
>
> Also, what version of gnu as are you using?
>
> Unless it conflicts with a fundamental llvm/clang philosophy, we are
> trying to keep Mips assembly compatible with AS.
>
> Also, keep in mind that the Mips llvm assembler is current development
> and is not considered "prime time" for production. Thus, a test case
> is always welcome.

Sure!

This file:

$ cat t.c
int foo()
{
	int c0;

	/* OK for clang */
	asm("dmfc2 %0,$1" : "=r" (c0));

	/* OK for gcc */
	asm("dmfc2 %0,1" : "=r" (c0));

	return c0;
}

produces:

$ mips64-octeon-linux-gnu-gcc -c t.c
t.s: Assembler messages:
t.s:19: Error: illegal operands `dmfc2 $1,$1'

$ clang -target mips64 -integrated-as -c t.c
t.c:9:6: error: invalid operand for instruction
        asm("dmfc2 %0,1" : "=r" (c0));
            ^
<inline asm>:1:11: note: instantiated into assembly here
        dmfc2 $1,1
                 ^

The assembler is from the Cavium Octeon SDK:

$ mips64-octeon-linux-gnu-as --version
GNU assembler (Cavium Networks Version: 2_2_0 build 113) 2.19
Copyright 2007 Free Software Foundation, Inc.
This program is free software; you may redistribute it under the terms of
the GNU General Public License version 3 or later.
This program has absolutely no warranty.
This assembler was configured for a target of `mips64-octeon-linux-gnu'.

I have not compared its behaviour with upstream gas (either the
equivalent version or a more recent one).

Thanks,

    J

>
> Thanks,
>
> Jack
>
> On 02/08/2013 10:53 AM, Jeremy Fitzhardinge wrote:> Hi all,
> >
> > I'm experimenting with compiling some Mips code which is normally built
> > with a gcc cross toolchain with clang instead.  This code targets the
> > Cavium Octeon, and uses some of that CPU's crypto engine features which
> > are implemented in coprocessor 2.
> >
> > The inline asm for the crypto code uses instructions of the form "dmtc2
> > %0, 0xNNNN" - that is the coprocessor register is represented as a
> > constant, which is the only syntax that gas seems to accept
> > (specifically, an expression which evaluates to a constant).  Clang's
> > integrated-as, however, only accepts "dmtc2 %0, $0xNNNN", that is, a
> > coprocessor register number.
> >
> > I think clang is correct here, but for cross-tool compatibility it would
> > be useful if it could also accept the same syntax as gas.  I could hack
> > up the Octeon SDK headers to switch syntax depending on the compiler,
> > but I'd really prefer to avoid that (esp since it really depends on the
> > assembler in use, and I'm not sure if there's preprocessor symbols to
> > detect whether integrated-as is enabled).
> >
> > I've had a look at Mips(64)InstrInfo.td which seems to be where the
> > syntax is defined, but since this is my first encounter with
> > clang/llvm's internals, I'm not sure how to go about making the change.
> > Could someone give me some pointers?  Is this something that InstAlias
> > can handle?
> >
> > An alternative workaround would be to just use gas as-is, but it gets
> > upset by the .cfi_section directives that clang emits.  Is there some
> > way to suppress those (while keeping the rest of the cfi directives that
> > this version of gas does understand)?
> >
> > Thanks,
> >      J
> >
>