https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126763
Bug ID: 126763
Summary: a?__builtin_clz(a):__builtin_clz(31) -> clz(a|1)
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: pinskia at gcc dot gnu.org
Target Milestone: ---
Take:
```
int f(unsigned a)
{
if (a)
return __builtin_clz(a);
return __builtin_clz(1);
}
int f0(unsigned a)
{
return __builtin_clz(a|1);
}
int n(unsigned a)
{
return __builtin_clzg(a,__builtin_clz(1)+1);
}
```
This all have well defined code and all return the same thing. f0 produces the
best results; 31 (assuming `sizeof(int)==4`).
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114032#c3 mentions we could add
support for CLZ for other than zero values than CLZ_VALUE_DEFINED_AT_ZERO, that
might be a good idea. For 31 (or 63) we could special case that in expand as
being `a|1` which should be able to support the value always.
