Re: __builtin_assume warnings

[Paul Eggert]

On 8/18/20 12:11 AM, Mattias Engdegård wrote:

> Wouldn't -Wno-assume, either locally (clang pragmas around __builtin_assume) or 
> globally (configure?) be more effective?

No, it's the other way around at least for me: having 'assume' use Clang's 
__builtin_assume makes 'assume' slower. Without __builtin_assume, 'assume' falls 
back on __builtin_unreachable, and Clang generates better code for 
__builtin_unreachable than it does __builtin_assume. For the following code:

int x;
static int f (void) { return x; }
int g (void) { __builtin_assume (!f ()); return f (); }
int h (void) { if (f ()) __builtin_unreachable (); return f (); }

clang -O2 generates suboptimal machine code for g (the generated code loads from 
'x') and better machine code for h (the generated code returns 0 without loading 
from 'x'). This is clang version 9.0.1 (Fedora 9.0.1-2.fc31) on 
x86_64-unknown-linux-gnu.

Perhaps someday the Clang folks will get their act together in this department, 
but in the meantime __builtin_unreachable is a perfectly adequate substitute.