https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116630
Bug ID: 116630
Summary: Implement spaceshipm3 optab for aarch64
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
For the testcase:
#include <compare>
auto cmp4way(double a, double b)
{
return a <=> b;
}
auto cmp4wayf(float a, float b)
{
return a <=> b;
}
with -O2 -std=c++20 we can generate better code for aarch64 than we do now:
cmp4way(double, double):
fcmp d0, d1
mov w0, 0
beq .L2
fcmpe d0, d1
mov w0, -1
bmi .L2
cset w0, gt
mov w1, 2
sub w0, w1, w0
.L2:
ret
cmp4wayf(float, float):
fcmp s0, s1
mov w0, 0
beq .L8
fcmpe s0, s1
mov w0, -1
bmi .L8
cset w0, gt
mov w1, 2
sub w0, w1, w0
.L8:
ret
LLVM generates:
cmp4way(double, double):
fcmp d0, d1
cset w8, ne
lsl w8, w8, #1
csinc w8, w8, wzr, le
csinv w0, w8, wzr, pl
ret
cmp4wayf(float, float):
fcmp s0, s1
cset w8, ne
lsl w8, w8, #1
csinc w8, w8, wzr, le
csinv w0, w8, wzr, pl
ret
I guess the LLVM sequence executes more instructions on average but it avoids
conditional branches that risk being mispredicted and is lower codesize overall
so should at least be preferred for -Os.
It also has only one fcmp instruction that tends to be lower throughput than
the other cheap GP insns whereas GCC can end up executing two fcmp.
