unreachable reduces performance


Bugzilla Link	PR25952
Status	NEW
Importance	P normal
Reported by	Gunnar Bergmann (g.bergmann.official@gmail.com)
Reported on	2015-12-27 17:39:51 -0800
Last modified on	2015-12-27 17:39:51 -0800
Version	3.7
Hardware	PC Linux
CC	llvm-bugs@lists.llvm.org
Fixed by commit(s)
Attachments	`unreachable-example.tar.gz` (20480 bytes, application/x-gzip)
Blocks
Blocked by
See also

Created attachment 15526
brainfuck interpreter and example data

Delaring cases in a switch statement like below as unreachable may reduce
performance:
switch x {
    case 0: // ...
        break;
    case 1: // ...
        break;
 // ....
    case 7: // ...
        break;

    default:
        __builtin_unreachable();
}

Removing the __builtin_unreachable() improves the performance. On my PC the
attached example needs about 25 with it and 17 without.
The attached code is a simple brainfuck interpreter that can be tested by
calling "brainfuck-optimized mandel.brainfuck".
It was test with options -O3 and -O2 as 64 bit build on a haswell processor.
The problem is reproducible in Rust.

Profiling shows a suspiciously high number of branch mispredictions. From what
I have seen of the assembler code, I think that nested conditional branches
like a binary search are faster than a branch table by causing fewer
mispredictions.

Quuxplusone / LLVMBugzillaTest

unreachable reduces performance #25951