https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121591
Bug ID: 121591
Summary: x86 optimization: isless doesn't reuse EFLAGS result
of other floating point comparisons with same operands
Product: gcc
Version: 15.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: Explorer09 at gmail dot com
Target Milestone: ---
This issue seems to be architecture specific. I can reproduce it for x86-64
target but not for ARM64 target.
Test code (same as in bug 121590):
```c
#include <math.h>
#include <stdlib.h>
int float_compare(double a, double b) {
if (isless(a, b))
return 1;
if (isgreater(a, b))
return 3;
if (isunordered(a, b))
return 4;
if (a == b)
return 2;
// Should never reach here
abort();
}
```
x86_64 gcc 15.1 with '-O2 -Wno-float-equal' options produces:
```assembly
float_compare:
ucomisd %xmm0, %xmm1
ja .L14
ucomisd %xmm1, %xmm0
ja .L15
jp .L8
jne .L13
movl $2, %eax
ret
.L14:
movl $1, %eax
ret
.L8:
movl $4, %eax
ret
.L15:
movl $3, %eax
ret
float_compare.cold:
.L13:
pushq %rax
call abort
```
(I tested this in Compiler Explorer (godbolt.org))
(The abort() function call is redundant. I've reported that part of the issue
as bug 121590.)
The point to note in this generated assembly is that isless() produces an
"ucomisd %xmm0, %xmm1" instruction on its own. Ideally only one "ucomisd"
instruction is needed, as the EFLAGS result can be reused for different jumps.
