https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86609
Bug ID: 86609
Summary: Reassociate (int) round sequences
Product: gcc
Version: unknown
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Not entirely sure what to call this optimisation.
Consider:
int f(double x, double *p)
{
double r = __builtin_round (x);
*p = r;
return (int) r;
}
For aarch64 GCC with -O2 -fno-math-errno generates:
f:
frinta d0, d0 // 7 [c=8 l=4] rounddf2
str d0, [x0] // 8 [c=4 l=4] *movdf_aarch64/7
fcvtzs w0, d0 // 14 [c=8 l=4] fix_truncdfsi2
ret // 29 [c=0 l=4] *do_return
The problem here is that the two FRINT* operations cannot be done in parallel.
Clang can break the chain like so:
f: // @f
// %bb.0: // %entry
frinta d1, d0
fcvtas w8, d0
str d1, [x0]
mov w0, w8
ret
Note how the two expensive operations are now independent.
I think in C terms this means transforming the above to:
int f2 (double x, double *p)
{
double r = __builtin_round (x);
*p = r;
return (int)__builtin_iround (x);
}
Would this be something for the reassociation pass to do?
