compiler-rt: Use FreeBSD's elf_aux_info for detecting ARM HW features

[dim]

Summary

FreeBSD provides 'elf_aux_info' function that does the same as 'getauxval' on Linux.

[arichardson]

Can you upload this with full context (https://llvm.org/docs/Phabricator.html#requesting-a-review-via-the-command-line)?

compiler-rt/lib/builtins/cpu_model.c

807

This means non-linux, non-FreeBSD systems always return false even if they prvide a getauxval() function. I would do something like the following to ensure that it fails to compile on unsupported systems instead of always returning false:

#if defined(__FreeBSD__)
  unsigned long hwcap = elf_aux_info(AT_HWCAP, &hwcap, sizeof(unsigned long));
#else
  unsigned long hwcap = getauxval(AT_HWCAP);
#endif

[MikaelUrankar]

Address comment.

[e-mail] committed a in slightly different version in FreeBSD src tree. Should I use its version instead?

Thanks!

[arichardson]

In D109330#2986440, @MikaelUrankar wrote:

Address comment.

[e-mail] committed a in slightly different version in FreeBSD src tree. Should I use its version instead?

Thanks!

Yes that looks reasonable to me. CC: @dim

[dim]

In D109330#2986445, @arichardson wrote:

In D109330#2986440, @MikaelUrankar wrote:

Address comment.

[e-mail] committed a in slightly different version in FreeBSD src tree. Should I use its version instead?

Thanks!

Yes that looks reasonable to me. CC: @dim

I committed a change to FreeBSD in https://cgit.freebsd.org/src/commit/?id=efe67f33c322265eb303ec0ab40275100795b22a, which is almost identical. However, I also check the return value of elf_aux_info(), since (theoretically at least) it could fail. In that case, I'm assuming that HWCAP is also not working. :)

@MikaelUrankar any objections if I add that to this review?

[MikaelUrankar]

In D109330#2987375, @dim wrote:

In D109330#2986445, @arichardson wrote:

In D109330#2986440, @MikaelUrankar wrote:

Address comment.

[e-mail] committed a in slightly different version in FreeBSD src tree. Should I use its version instead?

Thanks!

Yes that looks reasonable to me. CC: @dim

I committed a change to FreeBSD in https://cgit.freebsd.org/src/commit/?id=efe67f33c322265eb303ec0ab40275100795b22a, which is almost identical. However, I also check the return value of elf_aux_info(), since (theoretically at least) it could fail. In that case, I'm assuming that HWCAP is also not working. :)

@MikaelUrankar any objections if I add that to this review?

No objection from me, feel free to commandeer this review.
Thanks!

[dim]

Use version committed into FreeBSD 14-CURRENT.

compiler-rt/lib/builtins/cpu_model.c

 #include <sys/auxv.h>
 #ifndef AT_HWCAP
 #define AT_HWCAP 16
 #endif
 #ifndef HWCAP_ATOMICS
 #define HWCAP_ATOMICS (1 << 8)
 #endif
 static void CONSTRUCTOR_ATTRIBUTE init_have_lse_atomics(void) {
+#if defined(__FreeBSD__)
+  unsigned long hwcap;
+  int result = elf_aux_info(AT_HWCAP, &hwcap, sizeof hwcap);
+  __aarch64_have_lse_atomics = result == 0 && (hwcap & HWCAP_ATOMICS) != 0;
+#else
   unsigned long hwcap = getauxval(AT_HWCAP);
   __aarch64_have_lse_atomics = (hwcap & HWCAP_ATOMICS) != 0;
+#endif
 }
 #endif // defined(__has_include)
 #endif // __has_include(<sys/auxv.h>)
 #endif // defined(__aarch64__)