Yeah, you're right. It's one second out of 2 billion, I don't know if we care, but no I didn't mean to affect that behavior.

This would be easier if we had a TIME_MAX/MIN or weren't in a system header.

I tried a version of convincing gcc with something like if (sizeof(_ts.tv_sec) * __CHAR_BIT > 32), but it wasn't enough for it to bite. Let me take another attempt at something like that.

It's just really aggressive. Neither of these silence the warning:

if (0 && _ts.tv_sec > SBT_MAX >> 32)

if ((int64_t)_ts.tv_sec > SBT_MAX >> 32)

Frankly the former not silencing the warning seems like a gcc bug.

The least tortured thing I have come up with is

#ifndef __i386__

Ugly, but straightforward. Thoughts?

The underlying issue here is that time_t is 32-bits on x86 and arm platforms where __LP64__ is not defined. See sys/{x86,arm}/include/_types.h.

We could do something like:

static __inline sbintime_t
tstosbt_sat(struct timespec _ts)
{
#ifdef __LP64__
	if (_ts.tv_sec > SBT_MAX >> 32)
		return (SBT_MAX);
	if (_ts.tv_sec < -(SBT_MAX >> 32) - 1)
		return (-SBT_MAX - 1);
#endif
	return (tstosbt(_ts));
}

But this skips the check for other platforms (arm64, powerpc64, riscv64) that don't define __LP64__ but have a 64-bit time_t.

I think we might need to do something like:

static __inline sbintime_t
tstosbt_sat(struct timespec _ts)
{
	if (sizeof(_ts.tv_sec) > sizeof(int32_t)) {
		if (_ts.tv_sec > SBT_MAX >> 32)
			return (SBT_MAX);
		if (_ts.tv_sec < -(SBT_MAX >> 32) - 1)
			return (-SBT_MAX - 1);
	}
	return (tstosbt(_ts));
}

Which is ugly, but seems to cover all cases...

But this skips the check for other platforms (arm64, powerpc64, riscv64) that don't define __LP64__ but have a 64-bit time_t.

Oops, minor correction here. I didn't mean to include 64 when I gave the list of platforms. arm64, powerpc64, riscv64 should define __LP64__. The 32-bit variants of these architectures would be the ones to define __LP64__ but have a 64-bit time_t.

But this skips the check for other platforms (arm64, powerpc64, riscv64) that don't define __LP64__ but have a 64-bit time_t.

Oops, minor correction here. I didn't mean to include 64 when I gave the list of platforms. arm64, powerpc64, riscv64 should define __LP64__. The 32-bit variants of these architectures would be the ones to define __LP64__ but have a 64-bit time_t.

Sigh, I really need to check my comments before I press submit. A correction to the previous correction:

The 32-bit variants of these architectures would be the ones to NOT define __LP64__ but have a 64-bit time_t.

I'm not sure LP64 is the right check. It looks to me like i386 is the only arch left that defines a 32-bit time_t. arm, arm64, powerpc, and riscv all unconditionally typedef int64_t time_t, it's only x86 that conditions it on LP64. I'm not sure which of those arch types we still build for 32 bits though.

The sizeof-based check unfortunately doesn't work. gcc doesn't inhibit the warning even though it should be able to tell that the -Wtype-limits comparison it is complaining about would not be evaluated in that case. And you can't use sizeof in a preprocessor condition to just ifdef it out that way.