https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93112
Bug ID: 93112
Summary: Incorrect rounding for float to uint64 on x86 (32bit)
with -fexcess-precision=standard
Product: gcc
Version: 9.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: stefan.bru...@rwth-aachen.de
Target Milestone: ---
The following program yields 1034567 for all 4 cases on x86_64, arm64, armv7,
... and x86 with -fexcess-precision=fast, but yields 1034566 for the 1st case
on x86 with -fexcess-precision=standard:
---
#include <stdint.h>
#include <stdio.h>
#include <inttypes.h>
#define MHZ(x) ((x) * UINT64_C(1000000))
uint64_t t1(uint64_t x)
{
printf("r: %" PRIu64 "\n", x);
}
void t2(float x)
{
printf("r: %f\n", x);
}
int main(int argc, char** argv)
{
if (argc < 2) {
t1(MHZ(1.034567)); // 1034567 or 1034566
t1(1034567.0); // always 1034567
t2(MHZ(1.034567)); // always 1034567.0
t2(1034567.0); // always 1034567.0
}
return 0;
}
---
1'034'567 can be represented exactly as float value. 1.034567 can't be
represented exactly, but rounding towards nearest should yield the correct
result (the error is about 2^-20, float precision is about 2^-23).
The error only happens if the multiplication result is cast to int.
Error happens with both GCC 9.2.1 and 7.5.0
- gcc-7 (SUSE Linux) 7.5.0
- gcc-9 (SUSE Linux) 9.2.1 20191209 [gcc-9-branch revision 279114]
