https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93141
Peter Cordes <peter at cordes dot ca> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |peter at cordes dot ca
--- Comment #2 from Peter Cordes <peter at cordes dot ca> ---
gcc doesn't actually *branch* unless you use an if(), it just uses cmp/sbb to
do a 128-bit compare. CMP is like a SUB that only sets flags. The CF result
of SBB is used as an input for ADC.
https://godbolt.org/z/64C4R- of a testcase
GCC also wastes a varying number of MOV instructions beyond the minimum one to
make cmp/sbb work, depending on BMI2 MULX or not, and how the sum is written.
u128 prod = a[i] * (unsigned __int128) b[i];
#if 1
sum += prod;
//if(sum<prod) overflow++; // gcc branches on mov/cmp/sbb
overflow += sum<prod; // gcc uses adc after a mov/cmp/sbb
#else
overflow += __builtin_add_overflow(sum, prod, &sum); // gcc less bad,
fewer MOV
#endif
clang makes efficient asm for all 3 ways, if you stop it from unrolling and
using setc/movzx/add.
