This will crash if the blocks runtime is not linked, which is not very friendly to callers. It would be nice if it either:

• Gracefully exited with ENOSYS or similar if a blocks runtime is not present (and, ideally, documented this in the man page), or
• Skipped this call if the block's isa pointer is &_NSConcreteGlobalBlock (i.e. the block has no captures).

The cast should not be required here. The Block_copy macro should cast the result to typeof the argument.

As above, this will crash if the blocks runtime isn't implemented, though (with the current code) we'll crash on the copy here first.

I don't see how compar->isa == &_NSConcreteGlobalBlock saves me from having to copy the block. The block descriptor itself could still go out of scope if the caller returns before calling fts_close().

If the block is a global block then the block structure is a global and is never deallocated. The block *descriptor* is always a global (and is never copied by _Block_copy.

something like this?

@@ -59,6 +59,7 @@
 /* only present if linked with blocks runtime */
 void *_Block_copy(const void *) __weak_symbol;
 void _Block_release(const void *) __weak_symbol;
+extern void *_NSConcreteGlobalBlock[] __weak_symbol;
 
 static FTSENT  *fts_alloc(FTS *, char *, size_t);
 static FTSENT  *fts_build(FTS *, int);
@@ -274,6 +275,7 @@
 #ifdef __BLOCKS__
        compar = Block_copy(compar);
 #else
+       if (compar->isa != &_NSConcreteGlobalBlock)
                compar = _Block_copy(compar);
 #endif /* __BLOCKS__ */
        if (compar == NULL) {
@@ -287,6 +289,7 @@
 #ifdef __BLOCKS__
                Block_release(compar);
 #else
+               if (compar->isa != &_NSConcreteGlobalBlock)
                        _Block_release(compar);
 #endif /* __BLOCKS__ */
        }

This needs to be scoped to COMPILER_TYPE == clang

Maybe worth a COMPILER_FEATURES entry now we're going from n=1 in lib/libc/tests/stdlib/Makefile to n=2? (Especially with my CheriBSD hat on as we currently don't implement -fblocks for CHERI so have to disable using it even for Clang)

I'm not opposed, but that's out of scope for this review.

If people think it's not worth it then that's fine, but I don't agree about it being out of scope. The point at which one starts copying code or patterns is a justifiable point at which that should be extracted out, just as how you extracted __fts_open out rather than duplicate the implementation (and had you duplicated it then it would be completely reasonable to object to that). The only difference here is the code in question is a single if so there's less perceived benefit to extracting it out, it's only really us downstream that would benefit (and even then, it's not a huge issue).