Re: Register constraint 'not RAX or RDX'

[Andrew Haley <aph@redhat.com>]

tom@womack.net writes:
 > What's the right way of writing the 'compute a*b%c' intrinsic?
 > 
 > I can do
 > 
 > #define mulmod(a,b,c,d,e) { asm ("mul %%rbx; div %%rsi;" : "=d" (d), 
 > "=a" (e) : "a" (a), "b" (b), "S" (c));}
 > 
 > but that causes quite a lot of ugly register-shuffling
 > 
 > If I do
 > 
 > #define mulmod(a,b,c,d,e) { asm ("mul %0; div %1;" : "=d" (d), "=a" 
 > (e) : "a" (a), "r" (b), "r" (c));}
 > 
 > then the program doesn't work; disassembling
 > 
 > u64 zul(u64 a, u64 b, u64 p)
 > {
 >   u64 x,y;
 >   mulmod(a,b,p,x,y);
 >   return x;
 > }
 > 
 > gives
 > 
 > 00000000004005b0 <_Z3zulyyy>:
 >   4005b0:       48 89 f8                mov    %rdi,%rax
 >   4005b3:       48 f7 e2                mul    %rdx
 >   4005b6:       48 f7 f0                div    %rax
 >   4005b9:       48 89 d0                mov    %rdx,%rax
 >   4005bc:       c3                      retq
 > 
 > which is clearly crazy since %rax from 4005b0 will have been 
 > overwritten by the mul command before it's used.

#define mulmod(a,b,c,d,e)			\
{						\
  asm ("mul %3\n\tdiv %4\n"			\
       : "=&d" (d), "=a" (e)			\
       : "1" (a), "g" (b), "g" (c));		\
}

Andrew.