Forcing certain compiler-generated variables into specific ELF sections (with gcc)

Question

I'll start with the ultimate question: In C with gcc, is it possible to get the value(s) of __func__ (or equivalently, __FUNCTION__) stored in a section other than .rodata (or wherever -mrodata= points) or subsection thereof?

The full explanation:

Say I have a logging macro:

#define LOG(fmt, ...) log_internal(__FILE__, __LINE__, __func__, fmt, ##__VA_ARGS__)

(The string concatenation operator ## used in that unary context consumes the preceding comma if and only if the __VA_ARGS__ list is empty, thereby allowing use of a format string with or without arguments.)

I can then use the macro normally:

void my_function(void) {
    LOG("foo!");
    LOG("bar: %p", &bar);
}

might print (obviously depending on the implementation of log_internal):

foo.c:201(my_function) foo!
foo.c:202(my_function) bar: 0x12345678

In this case, the format strings ("foo" and "bar: %p") and the preprocessor strings ("foo.c" and "my_function") are anonymous read only data, and they get placed into the .rodata section automatically.

But say I want them to go to a different place (I'm on an embedded platform running almost everything from RAM for speed, but memory constraints are pushing for moving some things into ROM). It's "easy" to move __FILE__ and the format string:

#define ROM_STR(str) (__extension__({static const __attribute__((__section__(".rom_data"))) char __c[] = (str); (const char *)&__c;}))
#define LOG(fmt, ...) log_internal(ROM_STR(__FILE__), __LINE__, __func__, ROM_STR(fmt), ##__VA_ARGS__)

You can't put an __attribute__ on an anonymous string, so the ROM_STR macro gives it a transient name, affixes it to a specific section, then evaluates to the starting address, so it can substitute cleanly. This doesn't work if you try to pass a char * variable to LOG as your format string, but I'm willing to exclude that use case.

Normally, anonymous strings that happen to be identical get combined by the compiler into a single storage location, so every instance of __FILE__ in one file would share the same runtime address. With the explicit naming in ROM_STR, each instance will get its own storage location, so it probably doesn't actually make sense to use it on __FILE__.

However, I would like to use it on __func__. The problem is that __func__ is not the same kind of magic as __FILE__. From the gcc manual, "Function Names as Strings":

The identifier __func__ is implicitly declared by the translator as if, immediately following the opening brace of each function definition, the declaration

static const char __func__[] = "function-name";

appeared, where function-name is the name of the lexically-enclosing function. This name is the unadorned name of the function. ... These identifiers are not preprocessor macros. In GCC 3.3 and earlier, and in C only, __FUNCTION__ and __PRETTY_FUNCTION__ were treated as string literals; they could be used to initialize char arrays, and they could be concatenated with other string literas. GCC 3.4 and later treat them as variables, like __func__.

Thus, if you wrap __func__ with ROM_STR, you get

error: invalid initializer

and if you try to put a section attribute before or after the use of __func__, you get

error: expected expression before ‘__attribute__’

or

error: expected ‘)’ before ‘__attribute__’

And thus we loop back to the opening question: Is it possible to get __func__ stored in a section of my choosing? Maybe I can use -fdata-sections and do some linker script magic to get .rodata.__func__.* excluded from the rest of .rodata? If so, what is the syntax for globbing with exclusion in a linker script? In other words, somewhere you have a *(.rodata*) - I could put a *(.rodata.__func__*) somewhere else, but I would need to modify the original glob to exclude it so I don't get two copies.

Bagwig

Answer 1

One approach, which may be hackier than you'd like, would be to interpose a script to change the section names between compilation and assembly. Eg:

gcc -fdata-sections -S -o test.s test.c
sed 's/^\t.section\t\.rodata\.__func__\.[0-9]*/\t.section .rom_data/' -i test.s
gcc -c test.s

You could also try writing a clang transformation pass to place the __func__ declarations in a section of your choosing, or writing an object file manipulation program using libbfd.

Benediction

Answer 2

It looks like I answered my own question at the end with the -fdata-sections business, I just didn't understand the GNU Linker enough to see it. I don't actually need the globbing with exclusion as long as I specify the *(.rodata.__func__*) bit first. Any sections that glob matches will get marked as used, so a later glob for *(.rodata*) won't double count them and copy them somewhere else. I don't need to tag them with ROM_STR at all. Cool!

It's important to note that -fdata-sections does actually put each function string into its own .rodata.__func__.1234 section (I'm not sure what pattern the numbers follow). I don't know if anonymous strings also get their own sections; if so, I could use the same linker tricks to capture all of the anonymous strings instead of the ROM_STR section attribute macro, but it would probably be a bad idea. ROM_STR gets used in the LOG macro, so it's guaranteed only to be applied to logging format strings. If I forced all anonymous strings into ROM with a linker trick, that would include normal message data, and I would pay a runtime performance penalty to access it from flash. So I don't know if it's even possible, but its advisability would depend on your specific system requirements.

Bagwig

As of 2015, gcc no longer does this. See: gcc.gnu.org/bugzilla/show_bug.cgi?id=192 – Tempestuous