Yes the cost of the jo is not what people seem to conventionally think and teach. I don’t have an explanation yet. Maybe the scheduling is much better without it as it’s trying to run the overflow checks as fixed control flow otherwise?
I was also expecting to be writing about how removing overflow checks enables multiple arithmetic operations to be selected into single LEA instructions using addressing modes, as that’s the other common wisdom, but I couldn’t get the instruction selector to do that. Don’t know why yet either.
And the injection isn’t the proof - the mask was the proof. The injection is to see if we can do it without the actual mask instruction if we cared about that. You’re right it’s unsafe - I have an example of the failure mode of abusing it in the article. Turns out we didn’t need it in this case anyway as it’s the same performance.
I was also expecting to be writing about how removing overflow checks enables multiple arithmetic operations to be selected into single LEA instructions using addressing modes, as that’s the other common wisdom, but I couldn’t get the instruction selector to do that. Don’t know why yet either.
And the injection isn’t the proof - the mask was the proof. The injection is to see if we can do it without the actual mask instruction if we cared about that. You’re right it’s unsafe - I have an example of the failure mode of abusing it in the article. Turns out we didn’t need it in this case anyway as it’s the same performance.