On Thu, 01 Jun 2006 08:56:59 BST, Richard Earnshaw wrote:
>
> > is it possible to use the inline assembler of GCC for arm elf target in
> > conjunction with 64 bit variables?
>
> > I tried the following example (arithmetic average), but with little
> > success!
>
> > uint64_t a,b,r; asm(
> > "adds %A0, %A1, %A2" "\n\t"
> > "adds %B0, %B1, %B2" "\n\t"
> > "mov %B0, %B0, rrx" "\n\t"
> > "mov %A0, %A0, rrx"
> > : "=r" (r) : "r" (a), "r" (b) : "cc" );
>
>
> Yes, it is, but you need to use the right operand modifiers (these vary by
> target).
>
> On ARM the modifiers for accessing 64-bit types are
>
> %<n> The lowest numbered register of a pair
> %H<n> The highest numbered register of a pair
> %Q<n> The register containing the least significant part of the 32-bit
> value
> %R<n> The register containing the most significant part of the 32-bit value
>
> Why so many? Well it depends on whether you want your code to compile
> correctly for big-endian as well as little-endian systems.
>
> So your test case above should probably read something like
>
> int64_t a,b,r;
> asm(
> "adds %Q0, %Q1, %Q2" "\n\t"
> "adds %R0, %R1, %R2" "\n\t"
> "mov %R0, %R0, rrx" "\n\t"
> "mov %Q0, %Q0, rrx"
> : "=&r" (r) : "r" (a), "r" (b) : "cc"
> );
>
> Note the use of =&r in the constraint for the variable 'r'. This ensures
> that your input operands won't be corrupted before they have been fully
> read.
>
> R.
>
Incidentally, if you want the arithmetic average, the code you want is
int64_t a,b,r;
asm(
"adds %Q0, %Q1, %Q2" "\n\t"
"adcs %R0, %R1, %R2" "\n\t"
"movs %R0, %R0, rrx" "\n\t"
"mov %Q0, %Q0, rrx"
: "=&r" (r) : "r" (a), "r" (b) : "cc"
);
R.
