https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84251
Bug ID: 84251
Summary: Performance regression in gcc 8 when comparing
floating point numbers
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: mikulas at artax dot karlin.mff.cuni.cz
Target Milestone: ---
See this simple function and compile it with -O2 on x86-64:
#include <stdlib.h>
#include <math.h>
int cmp(double a, double b)
{
if (isnan(a) || isnan(b))
abort();
return a == b;
}
On gcc-7, we get optimal code:
0000000000000000 <cmp>:
0: 66 0f 2e c1 ucomisd %xmm1,%xmm0
4: 7a 07 jp d <cmp+0xd>
6: 0f 94 c0 sete %al
9: 0f b6 c0 movzbl %al,%eax
c: c3 retq
d: 48 83 ec 08 sub $0x8,%rsp
11: e8 00 00 00 00 callq 16 <cmp+0x16>
On gcc-8, we get inoptimal code. See the nonsensical "setnp" instruction in a
block of code where it is known that the "P" flag must be clear.
0000000000000000 <cmp>:
0: 48 83 ec 08 sub $0x8,%rsp
4: 66 0f 2e c1 ucomisd %xmm1,%xmm0
8: 0f 8a 00 00 00 00 jp e <cmp+0xe>
e: 0f 9b c0 setnp %al
11: ba 00 00 00 00 mov $0x0,%edx
16: 0f b6 c0 movzbl %al,%eax
19: 0f 45 c2 cmovne %edx,%eax
1c: 48 83 c4 08 add $0x8,%rsp
20: c3 retq
Disassembly of section .text.unlikely:
0000000000000000 <cmp.cold.0>:
0: e8 00 00 00 00 callq 5 <cmp.cold.0+0x5>
