Commit 3802df8
kbuild: Add '-fno-builtin-wcslen'
commit 84ffc79 upstream.
A recent optimization change in LLVM [1] aims to transform certain loop
idioms into calls to strlen() or wcslen(). This change transforms the
first while loop in UniStrcat() into a call to wcslen(), breaking the
build when UniStrcat() gets inlined into alloc_path_with_tree_prefix():
ld.lld: error: undefined symbol: wcslen
>>> referenced by nls_ucs2_utils.h:54 (fs/smb/client/../../nls/nls_ucs2_utils.h:54)
>>> vmlinux.o:(alloc_path_with_tree_prefix)
>>> referenced by nls_ucs2_utils.h:54 (fs/smb/client/../../nls/nls_ucs2_utils.h:54)
>>> vmlinux.o:(alloc_path_with_tree_prefix)
Disable this optimization with '-fno-builtin-wcslen', which prevents the
compiler from assuming that wcslen() is available in the kernel's C
library.
[ More to the point - it's not that we couldn't implement wcslen(), it's
that this isn't an optimization at all in the context of the kernel.
Replacing a simple inlined loop with a function call to the same loop
is just stupid and pointless if you don't have long strings and fancy
libraries with vectorization support etc.
For the regular 'strlen()' cases, we want the compiler to do this in
order to handle the trivial case of constant strings. And we do have
optimized versions of 'strlen()' on some architectures. But for
wcslen? Just no. - Linus ]
Cc: [email protected]
Link: llvm/llvm-project@9694844 [1]
Signed-off-by: Nathan Chancellor <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
[nathan: Resolve small conflict in older trees]
Signed-off-by: Nathan Chancellor <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>1 parent 0f1ee79 commit 3802df8
1 file changed
+3
-0
lines changed| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
1015 | 1015 | | |
1016 | 1016 | | |
1017 | 1017 | | |
| 1018 | + | |
| 1019 | + | |
| 1020 | + | |
1018 | 1021 | | |
1019 | 1022 | | |
1020 | 1023 | | |
| |||
0 commit comments