================
Comment at: lib/Headers/arm_acle.h:121
@@ +120,3 @@
+__set_saturation_occurred(int occurred) {
+ __asm__ __volatile__ ("msr cpsr_q, %0" : : "r" (occurred));
+}
----------------
t.p.northover wrote:
> rengolin wrote:
> > t.p.northover wrote:
> > > This also doesn't seem to do what ACLE says: set the Q flag based on the
> > > LSB of the input.
> > >
> > > Also, I don't think "cpsr_q" exists, does it? Both the v7A and v7M
> > > descriptions I've found lump it in with NZCV (i.e. "cpsr_nzcvq"). Which
> > > probably means the assembly should be marked as clobbering CPSR too.
> > I may be wrong, but I think you need to mrs first to get the bit pattern,
> > change this value based on (occurred != 0) shifted to the right place and
> > writing it back with msr, or you'll end up writing zeroes to the other
> > flags. Or is that only the case with registers?
> That would preserve NZCV and allow LLVM more optimisation freedom, but at the
> cost of a more expensive __set_saturation_occurred.
>
> As long as LLVM's aware that we *have* clobbered NZCV (slightly confusingly
> tracked as CPSR in the 32-bit backend) it should work either way.
>
> Of course, we don't actually model Q now, so we're not making the required
> guarantees anyway (I think LLVM would consider itself free to move a QADD
> past these no matter what we do, for example).
Shouldn't we be fixing that in LLVM first, then?
http://reviews.llvm.org/D6107
_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
