[llvm-commits] [llvm] r106662 - /llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp

[Bob Wilson]

On Jun 24, 2010, at 8:27 AM, Rafael Espindola wrote:

>> Wow, this is a pretty bad limitation.  Does GCC have the same limitation?
> 
> Not sure if I am testing the right thing, but compiling
> 
> int f(void);
> int g(void) {
>  asm ("" : : : "r7");
>  return f();
> }
> 
> I get
> 
> g:
> 	@ args = 0, pretend = 0, frame = 0
> 	@ frame_needed = 0, uses_anonymous_args = 0
> 	stmfd	sp!, {r7, lr}
> 	ldmfd	sp!, {r7, lr}
> 	b	f
> 	.size	g, .-g
> 
> I don't know much about thumb2, why can't the linker make this work?

The B instruction does not support ARM/Thumb interworking, so if functions f and g are compiled with different instruction sets, the linker would need to insert a stub.  The stub would need to load the address of the target into a register and do a BX to it.  That would certainly negate any advantage from the tail call transformation, although hopefully the common case would not require interworking.  Does the GNU linker do that?