RFA: ARM: Implement __builtin_trap

[Mikael Pettersson]

Richard Earnshaw writes:
 > 
 > On Thu, 2010-09-09 at 14:20 -0700, Chris Lattner wrote:
 > > On Sep 9, 2010, at 11:44 AM, Mark Mitchell wrote:
 > > 
 > > > On 9/9/2010 11:34 AM, Chris Lattner wrote:
 > > > 
 > > >> FWIW, the LLVM ARM backend compiles __builtin_trap into "trap" (aka .word 0xe7ffdefe).
 > > > 
 > > > Is "trap" a standard pseudo-op?  I didn't see it in either the ARM ARM
 > > > or in ARM's assembler manual, but that doesn't mean it isn't there.
 > > > ARM's assembler has an "UND" pseudo-op, which takes an argument to say
 > > > exactly which undefined instruction you want.  From the documentation,
 > > > it looks like the default, if you don't specify an argument is 0xe7f000f0.
 > > 
 > > Specifically, the LLVM backend emits:
 > > 
 > > _t:                                     @ @t
 > > 	.long 0xe7ffdefe @ trap
 > > 
 > > 
 > > So no, I don't think that trap is a standard mnemonic.  I have a vague memory of ARM not having an official trap instruction, but then retroactively deciding that some invalid encodings would never be defined.
 > > 
 > > -Chris
 > 
 > You shouldn't use ".long" for instructions, use ".inst".  ".long"
 > implies a word of data while ".inst" implies an instruction: in
 > big-endian mode the difference is significant.

It also affects objdump, which in recent binutils won't disassemble
code in .text entered with pseudo-ops like .long or .word. ".inst" works.

/Mikael