On Mon, 19 Aug 2024, Pali Rohár wrote:
What about attribute naked? For I386 it is supported since gcc 8.
I think it may be ok. I have very little experience with this attribute
myself though.
It seems like Wine are using this attribute, in some setups. And either
they use a non-static function declared entirely in assembly, or static
functions with the naked attribute.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=25967
https://gcc.gnu.org/onlinedocs/gcc/x86-Function-Attributes.html#index-naked-function-attribute_002c-x86
Per gcc documentation assembler instructions without operands are
supported in naked functions, so something like this could be ok.
What do you think?
diff --git a/mingw-w64-crt/stdio/_scprintf.c b/mingw-w64-crt/stdio/_scprintf.c
index 42f2e5f231a4..fca6f325e13c 100644
--- a/mingw-w64-crt/stdio/_scprintf.c
+++ b/mingw-w64-crt/stdio/_scprintf.c
@@ -51,9 +51,17 @@ static void resolve_scprintf(void)
/* gcc does not provide an easy way to call another variadic function with
reusing current arguments
* this source file is used only on i386, so do this function redirect via
inline i386 assembly */
#define ASM_SYM(sym) __MINGW64_STRINGIFY(__MINGW_USYMBOL(sym))
+
+#if !defined(__GNUC__) || __GNUC__ >= 8
Hmm - the !defined(__GNUC__) is a bit odd here... I think the other,
existing definition might be more safe if we don't know the compiler. (Our
files don't really build with any random compiler anyway, as we pretty
much require GCC or Clang in MinGW mode.)
We could check for Clang too (Clang identifies as GCC 4.2, but the Clang
version can also be inspected) - it should support the naked attribute as
well. I'm not exactly sure for how long it has supported it, but it might
be ok to not check (we don't in general support building with as wide
range of Clang versions as we do for GCC, I think).
+ __attribute__((naked))
+static int __cdecl init_scprintf(const char * __restrict__
__UNUSED_PARAM(format), ...)
+{
+asm (
+#else
asm (
".def\t" ASM_SYM(init_scprintf) ";\t.scl\t3;\t.type\t32;\t.endef\n"
ASM_SYM(init_scprintf) ":\n\t"
+#endif
"pushal\n\t"
"call\t" ASM_SYM(resolve_scprintf) "\n\t"
"popal\n\t"
@@ -63,5 +71,9 @@ ASM_SYM(init_scprintf) ":\n\t"
ASM_SYM(_scprintf) ":\n\t"
"jmp\t*" ASM_SYM(__MINGW_IMP_SYMBOL(_scprintf))
);
+#if !defined(__GNUC__) || __GNUC__ >= 8
+__builtin_unreachable();
+}
+#endif
#endif
I'm just not sure if is safe to mark format argument as __UNUSED_PARAM,
but without it gcc generates warning:
I think it should be safe. There's nothing in that declaration that should
be able to affect the code generation here anyway, I think.
mingw-w64-crt/stdio/_scprintf.c: In function ‘init_scprintf’:
mingw-w64-crt/stdio/_scprintf.c:57:60: warning: unused parameter ‘format’
[-Wunused-parameter]
static int __cdecl init_scprintf(const char * __restrict__ format, ...)
~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~
I guess the alternative of using a void cast within the function also
might work, but that's probalby a bit ugly in a naked function?
How are supposed to be defined naked functions with parameters which are
used in inline assembly by inspecting stack itself (without referencing
C parameters)?
Not sure, to be honest.
// Martin
_______________________________________________
Mingw-w64-public mailing list
Mingw-w64-public@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
