error: inlining failed in call to 'always_inline' in some builds

[shr-project]

Not sure if this is a bug in gcc or in libdeflate, but I came across interesting side effect - build failure after -march was dropped and only -mcpu was kept with:

https://git.openembedded.org/openembedded-core/commit/?h=scarthgap&id=e64f0c1b6ac5d598a79a21de5f3060f83cb9523e

The minimal reproducer for libdeflate for me was:

gcc -mcpu=cortex-a76+crypto \
--sysroot=../recipe-sysroot -I. -o lib/crc32.c.o -c lib/crc32.c

Which fails with:

In file included from lib/lib_common.h:40,
                 from lib/crc32.c:171:
lib/../common_defs.h: In function 'crc32_arm_pmullx12_crc_eor3':
lib/../common_defs.h:513:1: error: inlining failed in call to 'always_inline' 'get_unaligned_le64': target specific option mismatch
  513 | get_unaligned_le64(const u8 *p)
      | ^~~~~~~~~~~~~~~~~~
In file included from lib/arm/crc32_impl.h:564,
                 from lib/crc32.c:226:
lib/arm/crc32_pmull_wide.h:199:23: note: called from here
  199 |                 crc = __crc32d(crc, get_unaligned_le64(p + 24));
      |                       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib/../common_defs.h:513:1: error: inlining failed in call to 'always_inline' 'get_unaligned_le64': target specific option mismatch
  513 | get_unaligned_le64(const u8 *p)
      | ^~~~~~~~~~~~~~~~~~
lib/arm/crc32_pmull_wide.h:198:23: note: called from here
  198 |                 crc = __crc32d(crc, get_unaligned_le64(p + 16));
      |                       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib/../common_defs.h:513:1: error: inlining failed in call to 'always_inline' 'get_unaligned_le64': target specific option mismatch
  513 | get_unaligned_le64(const u8 *p)
      | ^~~~~~~~~~~~~~~~~~
...

While restoring -march builds fine with:

gcc -mcpu=cortex-a76+crypto -march=armv8.2-a+crypto \
--sysroot=../recipe-sysroot -I. -o lib/crc32.c.o -c lib/crc32.c

Interestingly if I use -E and compare preprocess output, then it's identical in both cases (__ARM_FEATURE_CRYPTO, HAVE_PMULL, forceinline are the same). This was with gcc-13.2.0.

[ebiggers]

It looks like there's a gcc bug/limitation that got fixed in gcc 13.2, and libdeflate's workaround for that bug/limitation is causing problems on gcc 13.2 and later. #370 should fix this.