I put here a simple example to illustrate what looks like a bug to me #include<iostream> inline unsigned next_exp2(unsigned x) { unsigned n; asm volatile("bsr %0, %1" : "=r" (n) : "r" (x)); // most significant bit return((x == (1u << n)) ? n : n); }; inline unsigned next_exp2_(unsigned x) { unsigned n; asm volatile("bsr %0, %1" : "=r" (n) : "r" (x)); // most significant bit return((x == (1u << n)) ? n : n+1); }; int main() { unsigned x = 1025; std::cout << "n1 = " << next_exp2(x) << std::endl; std::cout << "n2 = " << next_exp2_(x) << std::endl; return(0); } The difference between two function is n+1 returned from the underscored version. Compiled with g++ -O2 ./foo.cpp, my output is: n1 = 10 n2 = 3078596801 while I would expect n2 = 11. If I compile it with g++ -g foo.cpp, everything works as expected.
The problem you are seeing is Intel vs AT&T asm formats. GNU as defaults to AT&T format in that it is src, dst. So you have the operands swapped. Note also you also don't clobber the flags register as bsr sets the Zero flag. asm volatile("bsr %1, %0" : "=r" (n) : "r" (x) : "flags"); Is the correct code you want. It just happened to work at -O0 because the register allocator used the same registers for the input and output.
*** Bug 43263 has been marked as a duplicate of this bug. ***
You're absolutely right. Sorry for the false alarm.