https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122967
Bug ID: 122967
Summary: isfinite() assembly for 32-bit float
Product: gcc
Version: 15.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: other
Assignee: unassigned at gcc dot gnu.org
Reporter: Explorer09 at gmail dot com
Target Milestone: ---
This issue was discussed in Clang
(https://github.com/llvm/llvm-project/issues/169270),
and I think it's good to let GCC developers know this and improve the code.
```c
#include <math.h>
#include <stdbool.h>
#include <stdint.h>
bool func(float x) {
return isfinite(x);
}
bool func2(float x) {
// Can't make this transform if issignaling(x)
return isgreaterequal((x - x), (x - x));
}
bool func3(float x) {
union { float f; uint32_t u; } v, inf;
inf.f = INFINITY;
v.f = x;
v.u = (uint32_t)(v.u << 1);
return v.u < inf.u * 2;
}
bool func4(float x) {
union { float f; uint32_t u; } v, inf;
inf.f = INFINITY;
v.f = x;
v.u = ~v.u;
__asm__ ("" : "+r" (v.u));
return (v.u & inf.u) == 0;
}
```
```assembly
# x86-64
func2:
subss %xmm0, %xmm0
ucomiss %xmm0, %xmm0
setnp %al
ret
func3:
movd %xmm0, %eax
addl %eax, %eax
cmpl $-16777217, %eax
setbe %al
ret
func4:
movd %xmm0, %eax
notl %eax
testl $2139095040, %eax
sete %al
ret
```
```assembly
# AArch64
func2:
fsub s0, s0, s0
fcmp s0, s0
cset w0, vc
ret
func3:
fmov w1, s0
mov w0, -16777217
cmp w0, w1, lsl 1
cset w0, cs
ret
func4:
fmov w0, s0
mvn w0, w0
tst w0, 2139095040
cset w0, eq
ret
```
https://godbolt.org/z/1cxc4oWej
The original reporter there (Karl Meakin) suggested the "func2" approach, while
I suggested "func3" and "func4".
"func2" can give the smallest code for '-Os' optimization, but it come with a
caveat that it throws an exception on a signaling NaN (that is, it cannot be
used with '-fsignaling-nans'). "func3" and "func4" are slightly larger but,
AFAIK, do not throw exceptions.
I didn't do performance benchmark on any of these.
