On Wed, 17 Sep 2025, Pali Rohár wrote:
On Wednesday 17 September 2025 11:57:59 Martin Storsjö wrote:
On Sun, 31 Aug 2025, Pali Rohár wrote:
When the msvc setjmpex.h file is included in source file then msvc changes
the behavior of C setjmp() and longjmp() calls to be SEH safe. Without
including the setjmpex.h file, these C calls are not SEH safe by default.
In following table is behavior of the msvc compiler, which symbols are
called in the emitted binary object code (not linked):
without setjmpex.h with setjmpex.h
i386 _setjmp3 + longjmp _setjmp3 + _longjmpex
amd64 _setjmp + longjmp _setjmpex + longjmp
arm32 _setjmp + longjmp _setjmpex + longjmp
arm64 _setjmpex + longjmp _setjmpex + longjmp
This table was produced from results on https://godbolt.org/z/Gabv1qT9c
When the compiled code is linked into final binary with UCRT library, the
_setjmp symbol is resolved to UCRT _intrinsic_setjmp symbol and the
_setjmpex symbol is resolved to UCRT _intrinsic_setjmpex. i386 _setjmp3
stays as without the _intrinsic prefix and the i386 _setjmp is not emitted
at all.
Fix the mingw-w64 setjmpex.h and setjmp.h include files to define setjmp()
and longjmp() macros which expands to functions emitted by msvc compiler to
match the msvc behavior.
Note that msvc setjmpex.h and setjmp.h header files are different than
mingw-w64 and is because it looks like that msvc compiler parses the source
code with setjmp and longjmp keywords differently and is doing some
transformations not specified in header files. Something which gcc is not
doing.
mingw-w64 setjmpex.h and setjmp.h files are defined based on the
observation of output from msvc compiler, not from the msvc header files.
---
mingw-w64-headers/crt/setjmp.h | 104 ++++++++++++++++++-------------
mingw-w64-headers/crt/setjmpex.h | 16 +----
2 files changed, 63 insertions(+), 57 deletions(-)
diff --git a/mingw-w64-headers/crt/setjmp.h b/mingw-w64-headers/crt/setjmp.h
index 54e3ad03b67f..0d85d0e9f9f8 100644
--- a/mingw-w64-headers/crt/setjmp.h
+++ b/mingw-w64-headers/crt/setjmp.h
@@ -26,7 +26,7 @@
extern "C" {
#endif
-#if (defined(_X86_) && !defined(__x86_64))
+#if defined(__i386__)
#define _JBLEN 16
#define _JBTYPE int
@@ -40,6 +40,7 @@ extern "C" {
unsigned long Eip;
unsigned long Registration;
unsigned long TryLevel;
+ /* Following fields are only for new _setjmp3(), the are not for old
_setjmp(). */
unsigned long Cookie;
unsigned long UnwindFunc;
unsigned long UnwindData[6];
@@ -107,7 +108,7 @@ extern "C" {
} _JUMP_BUFFER;
-#elif defined(__x86_64)
+#elif defined(__x86_64__)
These changes to ifdefs (this one and the i386 one above) are ok, but
ideally I wouldn't have them intermixed with other functional changes like
this one. But it's not fatally distracting in the case of this commit
anyway, so it's ok this time.
I tried to make ifdefs consistent with the new changes, so I replaced
them. I do not mind if this change would be split into separate commit.
Same with the added comment - I don't really see if it is logically
connected to the rest of this change, or if it just happened to end up in
this same commit.
// Martin
As these series of commit was fixing the i386 usage of _setjmp3() and
_setjmp() functions, I added the comment to i386 jmp_buf structure which
is important for knowing the difference between those two function.
Also this comment change can be separated into its own commit.
Ok - I split these minor changes to separate commits, and pushed this set
now.
// Martin
_______________________________________________
Mingw-w64-public mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
