https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118580
--- Comment #7 from federico <federico.perini at gmail dot com> ---
Thanks Harald and Steve, great hints. Looking at the `-fdump-tree-original`, I
see:
- the on-the-fly expression seems the difference between a complex(8) and a
complex(4) expressions:
```
__builtin_cabs ( COMPLEX_EXPR <(real(kind=8)) REALPART_EXPR <SAVE_EXPR <a[S.7 +
D.3055]>>,
(real(kind=8)) IMAGPART_EXPR <SAVE_EXPR <a[S.7 +
D.3055]>> > ! complex(8)
-COMPLEX_EXPR <eye[S.7 + D.3056], 0.0>) ! complex(4)
```
double `cabs` is OK here I believe, and it is saved to `real(kind=8) limit.4;`.
- when B is saved:
```
b[S.3 + D.3048] = COMPLEX_EXPR <(real(kind=4)) REALPART_EXPR <SAVE_EXPR
<COMPLEX_EXPR <(real(kind=8)) REALPART_EXPR <SAVE_EXPR <a[S.3 + D.3046]>>,
(real(kind=8)) IMAGPART_EXPR <SAVE_EXPR <a[S.3 + D.3046]>>>
-COMPLEX_EXPR <eye[S.3 + D.3047], 0.0>> >,
(real(kind=4)) IMAGPART_EXPR <SAVE_EXPR
<COMPLEX_EXPR <(real(kind=8)) REALPART_EXPR <SAVE_EXPR <a[S.3 + D.3046]>>,
(real(kind=8)) IMAGPART_EXPR <SAVE_EXPR <a[S.3 + D.3046]>>>
-COMPLEX_EXPR <eye[S.3 + D.3047], 0.0>>>>;
```
The expression seems identical except there is a SAVE_EXPR and is then
downgraded to a real(4) complex expression for saving into b (uses
`__builtin_cabsf`).
