https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123482
Bug ID: 123482
Summary: in -frounding-math, GCC moves floating-point operation
across fesetround call
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: pascal_cuoq at hotmail dot com
Target Milestone: ---
This issue is pretty close to
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=34678 and GCC developers may
consider it a duplicate as the same fix may fix both issues, but it seems just
distinct enough to report separately:
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <fenv.h>
#ifdef __TRUSTINSOFT_ANALYZER__
#include <tis_builtin.h>
// display values even if they are not singletons:
#define printf tis_show_each
#endif
#pragma STDC FENV_ACCESS ON
int main(void)
{
/* IEEE-754 rounding modes required by the C standard */
const int rounding_modes[] = {
FE_TONEAREST,
FE_DOWNWARD,
FE_UPWARD,
FE_TOWARDZERO
};
const char *rounding_names[] = {
"FE_TONEAREST",
"FE_DOWNWARD",
"FE_UPWARD",
"FE_TOWARDZERO"
};
const size_t num_modes = sizeof(rounding_modes) / sizeof(int);
/* Pick a random rounding mode */
// srand((unsigned)time(NULL));
// size_t idx = (size_t)(rand() % num_modes);
size_t idx = 2;
int mode = rounding_modes[idx];
/* Set rounding mode */
fesetround(mode);
/* Operands */
double a = 0.1;
double b = 0.3;
/* Multiplication performed under selected rounding mode */
double result = a * b;
/* Set mode back to NE for printing */
fesetround(FE_TONEAREST);
printf("Rounding mode: %s\n", rounding_names[idx]);
printf("a = %.17g\n", a);
printf("b = %.17g\n", b);
printf("a * b = %.17g\n", result);
}
CE link: https://godbolt.org/z/TKse9MKs3
GCC 15.2 prints:
Rounding mode: FE_UPWARD
a = 0.10000000000000001
b = 0.29999999999999999
a * b = 0.029999999999999999
I expected:
Rounding mode: FE_UPWARD
a = 0.10000000000000001
b = 0.29999999999999999
a * b = 0.030000000000000002
Looking at the assembly code, the multiplication has been re-ordered after the
second call to fesetround. Unfortunately, in the context of this program, the
second call is necessary, as the rounding mode influences the binary-to-decimal
conversion behavior of printf:
main:
subq $8, %rsp
movl $2048, %edi
call fesetround ;;;;;;;;;;;;;; HERE
xorl %edi, %edi
call fesetround ;;;;;;;;;;;;;; HERE
movl $.LC0, %esi
movl $.LC1, %edi
xorl %eax, %eax
call printf
movl $.LC3, %edi
movl $1, %eax
movsd .LC2(%rip), %xmm0
call printf
movl $.LC5, %edi
movl $1, %eax
movsd .LC4(%rip), %xmm0
call printf
movl $.LC6, %edi
movl $1, %eax
movsd .LC4(%rip), %xmm0
mulsd .LC2(%rip), %xmm0 ;;;;;;;;;;;;;; HERE
call printf
xorl %eax, %eax
addq $8, %rsp
ret
