fix type instability in exp exp1m for Complex{Int} #18786
Conversation
| @@ -424,7 +424,8 @@ function log2(z::Complex) | |||
| a/log(oftype(real(a),2)) | |||
| end | |||
|
|
|||
| function exp(z::Complex) | |||
| exp(z::Complex) = exp(float(z)) | |||
There was a problem hiding this comment.
Why can't type inference figure this out given isnan(::Int) and isfinite(::Int) are both constants (edit: pure)? I was thinking about doing this a while ago (before we had constant propagation in type inference) but this adds branches (isfinite one in particular) that seems hard for LLVM to eliminate.
| @@ -424,7 +424,8 @@ function log2(z::Complex) | |||
| a/log(oftype(real(a),2)) | |||
| end | |||
|
|
|||
| function exp(z::Complex) | |||
| exp(z::Complex) = exp(float(z)) | |||
There was a problem hiding this comment.
Ref #13539 for a better workaround to the same issue.
There was a problem hiding this comment.
And also the (probably) new preferred way of adding similar test one line above the one you added.
There was a problem hiding this comment.
I think I missunderstand you but the test above the one I added looks pretty similar to how I did it?
There was a problem hiding this comment.
I think you need to rebase
Line 945 in fe29cb5
There was a problem hiding this comment.
Oh yeah, of course. Thanks.
|
Subsumed by #18793 |
Fixes #18785