The following code produces different output between ICC16.2 and GCC5.3.
#include <stdint.h>
#include <stdlib.h>
#include <iostream>
using namespace std;
int main(){
uint64_t a = 123456789123456789;
uint64_t b = 987654321987654321;
uint64_t L, H;
__asm__(
"mulx %, %, %;"
: "=r" (L), "=r" (H)
: "r" (a), "d" (b)
);
cout << "low = " << L << endl;
cout << "high = " << H << endl;
system("pause");
}
On Windows/ICC16, it outputs:
low = 6609981190679600
high = 14369616054794401669
On Linux/GCC5.3, it outputs:
low = 14369616054794401669
high = 6609981190679600
The order of the output operands on the MULX instruction are swapped. I believe GCC is correct since the inline assembly maps 1-to-1 to the AT&T GCC syntax, whereas ICC converts it to Intel syntax and in the process switches the order of the operands.
