On Fri, 9 Dec 2016, Thomas Gleixner wrote: > On Tue, 6 Dec 2016, Kyle Huey wrote: > > I propose the following patch, which results in GCC generating this code: > > > > mov (%rsi),%rax > > xor (%rdi),%rax > > mov %rdi,%rbx > > mov %rsi,%r12 > > test $0x2000000,%eax > > mov %rax,%rdx > > jne 3f > > 1: test $0x10000,%edx > > je 2f > > /* do some more work */ > > 2: /* the rest of the function */ > > ret > > 3: /* do work */ > > mov (%r12),%rdx > > xor (%rbx),%rdx > > jmp 1b > > > > When we are not doing the work controlled by these flags this is > > significantly > > better. We only have to load the flags from memory once, and we do a single > > XOR that we can test multiple times instead of repeatedly isolating bits. > > You rely on the compiler doing the right thing. And it's non obvious from > the code that this gets optimized as above.
On i386 this creates worse code due to register pressure it seems.
