Skip to contentSkip to conversation

staging: rtl8723bs: Mark ACPI table declaration as maybe unused

PATCH In review Nathan Chancellor 3 participants 7 messages Sep 26, 2018 Staging

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 staging: rtl8723bs: Mark ACPI table declaration as maybe unused

Clang emits the following warning:

drivers/staging/rtl8723bs/os_dep/sdio_intf.c:25:36: warning: variable 'acpi_ids' is not needed and will not be emitted [-Wunneeded-internal-declaration]

static const struct acpi_device_id acpi_ids[] = {
                                   ^

1 warning generated.

Mark the declaration as maybe unused like a few other instances of this construct in the kernel.

---
 drivers/staging/rtl8723bs/os_dep/sdio_intf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +1 -1
index 6d02904de63f..3285bf36291b 100644
@@ -22,7 +22,7 @@ static const struct sdio_device_id sdio_ids[] =
 	{ SDIO_DEVICE(0x024c, 0xb723), },
 	{ /* end: all zeroes */				},
 };
-static const struct acpi_device_id acpi_ids[] = {
+static const struct acpi_device_id acpi_ids[] __maybe_unused = {
 	{"OBDA8723", 0x0000},
 	{}
 };
Signature · 1 line
2.19.0
#1 lore raw
Quoted 23 lines · Nathan Chancellor

Clang emits the following warning:

drivers/staging/rtl8723bs/os_dep/sdio_intf.c:25:36: warning: variable 'acpi_ids' is not needed and will not be emitted [-Wunneeded-internal-declaration]

static const struct acpi_device_id acpi_ids[] = {
                                   ^

1 warning generated.

Mark the declaration as maybe unused like a few other instances of this construct in the kernel.

---
 drivers/staging/rtl8723bs/os_dep/sdio_intf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +0 -0
index 6d02904de63f..3285bf36291b 100644

drivers/staging/rtl8723bs/os_dep/sdio_intf.c static const struct sdio_device_id sdio_ids[] = · line 25

Quoted 5 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +1 -1
@@ -22,7 +22,7 @@ static const struct sdio_device_id sdio_ids[] =
 	{ SDIO_DEVICE(0x024c, 0xb723), },
 	{ /* end: all zeroes */				},
 };
-static const struct acpi_device_id acpi_ids[] = {
+static const struct acpi_device_id acpi_ids[] __maybe_unused = {

But it is used. No "maybe" at all here. The MODULE_DEVICE_TABLE() macro does a functional thing. Why is gcc not reporting an issue with this and clang is?

thanks,

greg k-h

#2 lore raw
Quoted 34 lines · Greg Kroah-Hartman
Quoted 23 lines · Nathan Chancellor

Clang emits the following warning:

drivers/staging/rtl8723bs/os_dep/sdio_intf.c:25:36: warning: variable 'acpi_ids' is not needed and will not be emitted [-Wunneeded-internal-declaration]

static const struct acpi_device_id acpi_ids[] = {
                                   ^

1 warning generated.

Mark the declaration as maybe unused like a few other instances of this construct in the kernel.

---
 drivers/staging/rtl8723bs/os_dep/sdio_intf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +0 -0
index 6d02904de63f..3285bf36291b 100644

drivers/staging/rtl8723bs/os_dep/sdio_intf.c static const struct sdio_device_id sdio_ids[] = · line 25

Quoted 5 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +1 -1
@@ -22,7 +22,7 @@ static const struct sdio_device_id sdio_ids[] =
 	{ SDIO_DEVICE(0x024c, 0xb723), },
 	{ /* end: all zeroes */				},
 };
-static const struct acpi_device_id acpi_ids[] = {
+static const struct acpi_device_id acpi_ids[] __maybe_unused = {

But it is used. No "maybe" at all here. The MODULE_DEVICE_TABLE() macro does a functional thing. Why is gcc not reporting an issue with this and clang is?

Quoted 3 lines

thanks,

greg k-h

I am not entirely sure, I've added Nick to this thread to see what he thinks. I'm by no means a Clang expert at the moment.

Thanks,
Nathan

#3 lore raw
Quoted 21 lines · Nathan Chancellor
Quoted 19 lines · Greg Kroah-Hartman
Quoted 18 lines · Nathan Chancellor

Clang emits the following warning:

drivers/staging/rtl8723bs/os_dep/sdio_intf.c:25:36: warning: variable 'acpi_ids' is not needed and will not be emitted [-Wunneeded-internal-declaration]

static const struct acpi_device_id acpi_ids[] = {
                                   ^

1 warning generated.

Mark the declaration as maybe unused like a few other instances of this construct in the kernel.

---
 drivers/staging/rtl8723bs/os_dep/sdio_intf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 14 lines
Quoted 6 lines
Quoted 6 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +0 -0
index 6d02904de63f..3285bf36291b 100644
@@ -22,7 +22,7 @@ static const struct sdio_device_id sdio_ids[] =
    { SDIO_DEVICE(0x024c, 0xb723), },
Quoted 8 lines

drivers/staging/rtl8723bs/os_dep/sdio_intf.c static const struct sdio_device_id sdio_ids[] = · line 25

Quoted 4 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +1 -1
    { /* end: all zeroes */                         },
 };
-static const struct acpi_device_id acpi_ids[] = {
+static const struct acpi_device_id acpi_ids[] __maybe_unused = {

But it is used. No "maybe" at all here. The MODULE_DEVICE_TABLE() macro does a functional thing. Why is gcc not reporting an issue with this and clang is?

Looks like GCC and Clang handle __attribute__((alias)) differently when optimizations are enabled. Clang has -Wunneeded-internal-declaration (GCC doesn't have that specific flag, which is why GCC doesn't report, but its behavior is also different), as part of -Wall, which warns that it thinks the static var is dead, then removes it from -O1 and up.

https://godbolt.org/z/Dpxnbp

I consider this a bug in Clang: https://bugs.llvm.org/show_bug.cgi?id=39088

As Nathan notes in this comment in V1:
https://lore.kernel.org/lkml/20180926064437.GA29417@flashbox/ having additional references to the static array is enough to keep it around. When there are no other references, then it should not be marked static, in order to still be emitted.

We can work around this by removing static from struct definitions that no other references than being passed to the MODULE_DEVICE_TABLE macro. GCC and Clang will then both emit references to both identifiers.

Signature · 2 lines
Thanks,
~Nick Desaulniers
#4 lore raw
Quoted 56 lines · Nick Desaulniers
Quoted 21 lines · Nathan Chancellor
Quoted 19 lines · Greg Kroah-Hartman
Quoted 18 lines · Nathan Chancellor

Clang emits the following warning:

drivers/staging/rtl8723bs/os_dep/sdio_intf.c:25:36: warning: variable 'acpi_ids' is not needed and will not be emitted [-Wunneeded-internal-declaration]

static const struct acpi_device_id acpi_ids[] = {
                                   ^

1 warning generated.

Mark the declaration as maybe unused like a few other instances of this construct in the kernel.

---
 drivers/staging/rtl8723bs/os_dep/sdio_intf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 14 lines
Quoted 6 lines
Quoted 6 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +0 -0
index 6d02904de63f..3285bf36291b 100644
@@ -22,7 +22,7 @@ static const struct sdio_device_id sdio_ids[] =
    { SDIO_DEVICE(0x024c, 0xb723), },
Quoted 8 lines

drivers/staging/rtl8723bs/os_dep/sdio_intf.c static const struct sdio_device_id sdio_ids[] = · line 25

Quoted 4 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +1 -1
    { /* end: all zeroes */                         },
 };
-static const struct acpi_device_id acpi_ids[] = {
+static const struct acpi_device_id acpi_ids[] __maybe_unused = {

But it is used. No "maybe" at all here. The MODULE_DEVICE_TABLE() macro does a functional thing. Why is gcc not reporting an issue with this and clang is?

Looks like GCC and Clang handle __attribute__((alias)) differently when optimizations are enabled. Clang has -Wunneeded-internal-declaration (GCC doesn't have that specific flag, which is why GCC doesn't report, but its behavior is also different), as part of -Wall, which warns that it thinks the static var is dead, then removes it from -O1 and up.

https://godbolt.org/z/Dpxnbp

I consider this a bug in Clang: https://bugs.llvm.org/show_bug.cgi?id=39088

As Nathan notes in this comment in V1:
https://lore.kernel.org/lkml/20180926064437.GA29417@flashbox/ having additional references to the static array is enough to keep it around. When there are no other references, then it should not be marked static, in order to still be emitted.

Quoted 4 lines

We can work around this by removing static from struct definitions that no other references than being passed to the MODULE_DEVICE_TABLE macro. GCC and Clang will then both emit references to both identifiers.

Sorry, after thinking more about this and discussing more with a coworker, I think Clang is doing the right thing, and that static should be removed from declarations passed to MODULE_DEVICE_TABLE without other references. Clang's Dead Code Elimination (DCE) here seems to be more aggressive, but it still looks correct to me.

int foo [] = { 42, 0xDEAD }; // extern
extern typeof(foo) bar __attribute__((unused, alias("foo")));

GCC and Clang at -O2 both produce references to foo and bar.

Adding static to foo, then GCC produces both references, while Clang moves the data to bar and removes foo. This is safe because foo has no other references, and bar is what file2alias.c cares about in this case.

Signature · 2 lines
Thanks,
~Nick Desaulniers
maintainer
#5 lore raw
Quoted 72 lines · Nick Desaulniers
Quoted 56 lines · Nick Desaulniers
Quoted 21 lines · Nathan Chancellor
Quoted 19 lines · Greg Kroah-Hartman
Quoted 18 lines · Nathan Chancellor

Clang emits the following warning:

drivers/staging/rtl8723bs/os_dep/sdio_intf.c:25:36: warning: variable 'acpi_ids' is not needed and will not be emitted [-Wunneeded-internal-declaration]

static const struct acpi_device_id acpi_ids[] = {
                                   ^

1 warning generated.

Mark the declaration as maybe unused like a few other instances of this construct in the kernel.

---
 drivers/staging/rtl8723bs/os_dep/sdio_intf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 14 lines
Quoted 6 lines
Quoted 6 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +0 -0
index 6d02904de63f..3285bf36291b 100644
@@ -22,7 +22,7 @@ static const struct sdio_device_id sdio_ids[] =
    { SDIO_DEVICE(0x024c, 0xb723), },
Quoted 8 lines

drivers/staging/rtl8723bs/os_dep/sdio_intf.c static const struct sdio_device_id sdio_ids[] = · line 25

Quoted 4 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +1 -1
    { /* end: all zeroes */                         },
 };
-static const struct acpi_device_id acpi_ids[] = {
+static const struct acpi_device_id acpi_ids[] __maybe_unused = {

But it is used. No "maybe" at all here. The MODULE_DEVICE_TABLE() macro does a functional thing. Why is gcc not reporting an issue with this and clang is?

Looks like GCC and Clang handle __attribute__((alias)) differently when optimizations are enabled. Clang has -Wunneeded-internal-declaration (GCC doesn't have that specific flag, which is why GCC doesn't report, but its behavior is also different), as part of -Wall, which warns that it thinks the static var is dead, then removes it from -O1 and up.

https://godbolt.org/z/Dpxnbp

I consider this a bug in Clang: https://bugs.llvm.org/show_bug.cgi?id=39088

As Nathan notes in this comment in V1:
https://lore.kernel.org/lkml/20180926064437.GA29417@flashbox/ having additional references to the static array is enough to keep it around. When there are no other references, then it should not be marked static, in order to still be emitted.

Quoted 4 lines

We can work around this by removing static from struct definitions that no other references than being passed to the MODULE_DEVICE_TABLE macro. GCC and Clang will then both emit references to both identifiers.

Sorry, after thinking more about this and discussing more with a coworker, I think Clang is doing the right thing, and that static should be removed from declarations passed to MODULE_DEVICE_TABLE without other references. Clang's Dead Code Elimination (DCE) here seems to be more aggressive, but it still looks correct to me.

int foo [] = { 42, 0xDEAD }; // extern
extern typeof(foo) bar __attribute__((unused, alias("foo")));

Quoted 6 lines

GCC and Clang at -O2 both produce references to foo and bar.

Adding static to foo, then GCC produces both references, while Clang moves the data to bar and removes foo. This is safe because foo has no other references, and bar is what file2alias.c cares about in this case.

But we want the variable to be static, and not part of any namespace outside of this file. If we start changing them all, we will start running into namespace collisions.

These go into a separate part of the linker table, shouldn't that be enough to ensure that the compiler keeps it around? Or is that after the compiler processes things?

Anything we can do to add to the MODULE_DEVICE_TABLE() macro instead?

thanks,

greg k-h

#6 lore raw
Quoted 76 lines · Greg KH
Quoted 74 lines · Nick Desaulniers
Quoted 56 lines · Nick Desaulniers
Quoted 21 lines · Nathan Chancellor
Quoted 19 lines · Greg Kroah-Hartman
Quoted 18 lines · Nathan Chancellor

Clang emits the following warning:

drivers/staging/rtl8723bs/os_dep/sdio_intf.c:25:36: warning: variable 'acpi_ids' is not needed and will not be emitted [-Wunneeded-internal-declaration]

static const struct acpi_device_id acpi_ids[] = {
                                   ^

1 warning generated.

Mark the declaration as maybe unused like a few other instances of this construct in the kernel.

---
 drivers/staging/rtl8723bs/os_dep/sdio_intf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 14 lines
Quoted 6 lines
Quoted 6 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +0 -0
index 6d02904de63f..3285bf36291b 100644
@@ -22,7 +22,7 @@ static const struct sdio_device_id sdio_ids[] =
    { SDIO_DEVICE(0x024c, 0xb723), },
Quoted 8 lines

drivers/staging/rtl8723bs/os_dep/sdio_intf.c static const struct sdio_device_id sdio_ids[] = · line 25

Quoted 4 lines
drivers/staging/rtl8723bs/os_dep/sdio_intf.c +1 -1
    { /* end: all zeroes */                         },
 };
-static const struct acpi_device_id acpi_ids[] = {
+static const struct acpi_device_id acpi_ids[] __maybe_unused = {

But it is used. No "maybe" at all here. The MODULE_DEVICE_TABLE() macro does a functional thing. Why is gcc not reporting an issue with this and clang is?

Looks like GCC and Clang handle __attribute__((alias)) differently when optimizations are enabled. Clang has -Wunneeded-internal-declaration (GCC doesn't have that specific flag, which is why GCC doesn't report, but its behavior is also different), as part of -Wall, which warns that it thinks the static var is dead, then removes it from -O1 and up.

https://godbolt.org/z/Dpxnbp

I consider this a bug in Clang: https://bugs.llvm.org/show_bug.cgi?id=39088

As Nathan notes in this comment in V1:
https://lore.kernel.org/lkml/20180926064437.GA29417@flashbox/ having additional references to the static array is enough to keep it around. When there are no other references, then it should not be marked static, in order to still be emitted.

Quoted 4 lines

We can work around this by removing static from struct definitions that no other references than being passed to the MODULE_DEVICE_TABLE macro. GCC and Clang will then both emit references to both identifiers.

Sorry, after thinking more about this and discussing more with a coworker, I think Clang is doing the right thing, and that static should be removed from declarations passed to MODULE_DEVICE_TABLE without other references. Clang's Dead Code Elimination (DCE) here seems to be more aggressive, but it still looks correct to me.

int foo [] = { 42, 0xDEAD }; // extern
extern typeof(foo) bar __attribute__((unused, alias("foo")));

GCC and Clang at -O2 both produce references to foo and bar.

Quoted 8 lines
Quoted 4 lines

Adding static to foo, then GCC produces both references, while Clang moves the data to bar and removes foo. This is safe because foo has no other references, and bar is what file2alias.c cares about in this case.

But we want the variable to be static, and not part of any namespace outside of this file. If we start changing them all, we will start running into namespace collisions.

That's a fair point. __attribute__((used)) (in the kernel, we use the macro __used) will silence this warning in Clang for this particular instance while being no functional change for GCC or Clang.

Nathan's v2 solution was very close! Either attribute silences the warning, but __used makes it clearer that we don't want the warning, without inhibiting dead code elimination from pruning the unreferenced internal symbol while moving its data to the external/globally reference-able symbol.

Exploring more of what GCC is doing (feel free to skip this, as it's minutia):
GCC emits BOTH symbols foo and bar (from my above simplified case, when foo is static), but only bar (the alias) is in the global namespace (extern linkage). I missed in the disassembly that only bar had the .globl assembler directive and assumed labels were by default globl (their not, I've discovered from testing. Today I learned).

More concretely, looking at the symbols of the translation unit in question (rather than my simplified toy):

$ nm -S drivers/staging/rtl8723bs/os_dep/sdio_intf.o | grep acpi
00000000000000e0 0000000000000040 r acpi_ids
00000000000000e0 0000000000000040 R __mod_acpi__acpi_ids_device_tabl

Shows that both symbols are created, man nm says about the capitalization and r vs R:
"if uppercase, the symbol is global (external)."
"R"
"r" The symbol is in a read only data section.

The disassembler only shows the reference to __mod_acpi__acpi_ids_device_tabl at file offset 0x15a0:

$ objdump -DF drivers/staging/rtl8723bs/os_dep/sdio_intf.o |less
00000000000000e0 <__mod_acpi__acpi_ids_device_table> (File Offset: 0x15a0):

And a dump of the binary shows the data from acpi_ids at that offset:

$ hexdump -C drivers/staging/rtl8723bs/os_dep/sdio_intf.o| less
000015a0  4f 42 44 41 38 37 32 33  00 00 00 00 00 00 00 00  |OBDA8723........|

^ should look familiar based on the
declaration in question.

The same can be verified with Clang, which wont produce the internal only symbol (that has zero references) acpi_ids, and will place __mod_acpi__acpi_ids_device_tabl at file offset 0x1610 for that translation unit.

Quoted 7 lines

These go into a separate part of the linker table, shouldn't that be
enough to ensure that the compiler keeps it around? Or is that after the compiler processes things?

Anything we can do to add to the MODULE_DEVICE_TABLE() macro instead?

Signature · 2 lines
Thanks,
~Nick Desaulniers