Errno fix

[rmn30]

Here is a fix for #461 and the associated test. I've done separate commits for the test and the bug fix. Not sure if you would prefer them to be squashed?

[rmn30]

After a bit of tinkering and some more thought I am withdrawing this PR because I think this is not the correct solution: snmalloc should not be setting errno except for functions that are defined to do so. Probably the override wrapper is the correct place for this (at the cost of a nullptr check).

[mjp41]

I don't want a branch on the fast path of malloc, and setting errno at the outer level will lead to this. It has to be buried deep inside.

[rmn30]

In that case setting errno should become part of snmalloc's API and we will have to save and restore errno in the posix_memalign implementation because it explicitly does not set it. Implementations of the other functions already check for alloc returning null and set errno, although the realloc implementation is a bit broken and I will fix it.

[mjp41]

Does posix_memalign require it not to be set, or that you can't depend on it being set?

In particular, I read
https://sourceware.org/bugzilla/show_bug.cgi?id=83#c1

Which implies we can set it.

[davidchisnall]

POSIX2018 is pretty liberal WRT errno:

• It is never set to 0 by any POSIX function.
• It is undefined after any call to a POSIX function that is not explicitly stated to set it.
• It is undefined after a successful call to a POSIX function that is documented to set errno on failure.

posix_memalign is not defined to set errno, so we're free to set or not set it as we desire on failure. On success we may set it to whatever we want as long as we do not transition it from a non-zero to a zero state.

[mjp41]

To get this right, some of the other pals should probably set it. Like

snmalloc/src/pal/pal_windows.h

Lines 163 to 196 in af8ab2d

	# ifdef PLATFORM_HAS_VIRTUALALLOC2
	template<bool state_using>
	static void* reserve_aligned(size_t size) noexcept
	{
	SNMALLOC_ASSERT(bits::is_pow2(size));
	SNMALLOC_ASSERT(size >= minimum_alloc_size);
	DWORD flags = MEM_RESERVE;
	if (state_using)
	flags |= MEM_COMMIT;
	// If we're on Windows 10 or newer, we can use the VirtualAlloc2
	// function. The FromApp variant is useable by UWP applications and
	// cannot allocate executable memory.
	MEM_ADDRESS_REQUIREMENTS addressReqs = {NULL, NULL, size};
	MEM_EXTENDED_PARAMETER param = {
	{MemExtendedParameterAddressRequirements, 0}, {0}};
	// Separate assignment as MSVC doesn't support .Pointer in the
	// initialisation list.
	param.Pointer = &addressReqs;
	void* ret = VirtualAlloc2FromApp(
	nullptr, nullptr, size, flags, PAGE_READWRITE, &param, 1);
	return ret;
	}
	# endif
	static void* reserve(size_t size) noexcept
	{
	return VirtualAlloc(nullptr, size, MEM_RESERVE, PAGE_READWRITE);
	}

Windows Pal should probably set errno when VirtualAlloc* returns nullptr.

[rmn30]

I think this version avoids extra overhead on the fast path except in the case of calloc which would previously attempt to zero a null pointer if allocation failed. Looks like Qemu is crashing on arm build :-(

[mjp41]

LGTM, thanks for fixing