shared: replace packed-struct get/put_unaligned with memcpy

[rawrmonster17]

Fixes #338.

Problem

The current get_unaligned/put_unaligned macros in shared/util.h use a packed-struct pointer-cast pattern:

#define get_unaligned(ptr)                       \
	({                                       \
		struct __attribute__((packed)) { \
			typeof(*(ptr)) __v;      \
		} *__p = (typeof(__p))(ptr);     \
		__p->__v;                        \
	})

This is not guaranteed portable. With GCC 14 at -O2 on RISC-V (and historically on MIPS/SPARC strict-alignment targets), the auto-vectoriser can emit aligned vector loads for the surrounding loop in hash_superfast(), causing SIGBUS on hardware that enforces natural alignment (reported in #338).

The bug affects shared/hash.c (4 call sites in the main and fall-through loops) and libkmod/libkmod-index.c (1 call site reading a uint32_t from a potentially-unaligned address). put_unaligned is defined but currently unused.

Fix

Replace both macros with memcpy-based versions:

#define get_unaligned(ptr)                                               \
	__extension__({                                                  \
		typeof(*(ptr)) __v;                                      \
		memcpy(&__v, (ptr), sizeof(__v));                        \
		__v;                                                     \
	})

#define put_unaligned(val, ptr)                                          \
	do {                                                             \
		typeof(*(ptr)) __v = (val);                              \
		memcpy((ptr), &__v, sizeof(__v));                        \
	} while (0)

memcpy with a compile-time-constant size is recognised by GCC, clang, and ICC and lowered to the same efficient inline sequence as the pointer cast on x86/x86-64, while remaining correct and standards-compliant on all architectures. This is the same approach used by the Linux kernel (include/linux/unaligned.h).

<string.h> is already included by shared/util.h. The macro signatures are unchanged, so all call sites in hash.c and libkmod-index.c are unaffected.

Testing

• Built with meson setup build --buildtype=plain && ninja -C build — zero errors, zero warnings on all changed translation units.
• Ran a dedicated self-check covering all 5 call-site patterns (uint16_t at +1 offset, uint16_t at +3 offset, uint32_t at +1 offset, put uint16_t round-trip, put uint32_t round-trip) with intentionally unaligned pointers — all pass.

[lucasdemarchi]

This is the same approach used by the Linux kernel (include/linux/unaligned.h).

Relevant kernel commit: a339671db64b ("vdso: Switch get/put_unaligned() from packed struct to memcpy()")

[lucasdemarchi]

Please use a real user for commit author and add your s-o-b.

It doesn't pass a warning-free build:

../shared/util.h:78:24: error: passing argument 1 of ‘memcpy’ discards ‘const’ qualifier from pointer target type [-Werror=discarded-qualifiers]
   78 |                 memcpy(&__v, (ptr), sizeof(__v));                        \
      |                        ^~~~

and also triggers array out of bounds:

In function ‘memcpy’,
    inlined from ‘read_u32_mm’ at ../libkmod/libkmod-index.c:662:6,
    inlined from ‘index_mm_open’ at ../libkmod/libkmod-index.c:791:16:
/usr/include/x86_64-linux-gnu/bits/string_fortified.h:29:10: error: array subscript 0 is outside array bounds of ‘const uint8_t[0]’ {aka ‘const unsigned char[]’} [-Werror=array-bounds=]
   29 |   return __builtin___memcpy_chk (__dest, __src, __len,
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
   30 |                                  __glibc_objsize0 (__dest));
      |                                  ~~~~~~~~~~~~~~~~~~~~~~~~~~
In function ‘index_mm_open’:
cc1: note: source object is likely at address zero
cc1: all warnings being treated as errors
[25/34] Compiling C object kmod.p/tools_depmod.c.o
ninja: build stopped: subcommand failed.

I think adopting a exact same macro as the kernel uses should be fine. This simplified version doesn't work as expected.

[rawrmonster17]

Updated the PR branch.

• Rewrote the commit with a real author and Signed-off-by: Chris Beck <beck.chris88@outlook.com>.
• Switched the helpers to the kernel-style pattern using __typeof_unqual__ plus __builtin_memcpy() with void * casts, rather than the simplified memcpy() form.
• Rebuilt on Linux with gcc using -O2 -Werror; the build now passes warning-free.
• Also rechecked with a 32-bit build and a small targeted self-check for the affected unaligned read/write patterns.

No call sites changed; the update is still localized to shared/util.h.

[evelikov]

Thanks for the PR @rawrmonster17 - few questions/suggestions, in random order:

• Please double-check our readme and contributing docs - namely the "Hacking" and "Commit trailers" sections.

• Do share the self-check program and methodology used

• Are the new extensions/built-ins available on our minimum gcc/clang versions, should we update the readme? I think we can omit the __extensions__ keyword.

• In the commit message let's reference the kernel commit and highlight on the strict-aliasing issue, listing the RISC-V improvements as a "bonus".

[lucasdemarchi]

typeof_unqual is gcc >= 14 only. Also, even after that, it doesn't seem the generated code is really the same.