#977 Tcl wrong int64 parameter

[Vera]

I wrote a C function, expecting "long long" as parameter:
int c_my (long long big_offset);

With SWIG I have an interface to Tcl.

In Tcl I call:
c_my 3000000000

in C (c_my) I got:
big_offset = -1294967296.

There is an bug in SWIG_AsVal_long_SS_long.

The whole discussion is in comp.lang.tcl
Subj:"64 bit integers (long long)"

Here is my (working) hotfix. Replace all the code in SWIG_AsVal_long_SS_long with:

SWIGINTERN int
SWIG_AsVal_long_SS_long SWIG_TCL_DECL_ARGS_2(Tcl_Obj *obj, long long *val)
{
if (Tcl_GetWideIntFromObj(0, obj, val) != TCL_OK) {
return SWIG_TypeError;
}
return SWIG_OK;
}

[Kjell Wooding]

This is still a bug in 2.0.1.
The problem is in tclprimtypes.swg:

SWIGINTERN int
SWIG_AsVal_dec(long long)(Tcl_Obj *obj, long long *val)
{
long v;
if (Tcl_GetLongFromObj(0,obj, &v) == TCL_OK) {
if (val) *val = v;
return SWIG_OK;
} else {
...

When trying to convert a signed int64 (long long), this code attempts to convert first to a signed int32 (long). If the input is positive, but between 2^31 and 2^32, the conversion attempt will appear to succeed (at least, with the supplied test), but will cast the result to a negative quantity)

The solution is to not do this silly optimization at all, or to require Tcl >=8.4, and change this call to Tcl_GetWideIntFromObj

Patch attached. Please fix this, as spending a day chasing a bug that was identified back in 2009 is no kind of fun.

[Kjell Wooding]

--- tclprimtypes.swg.orig 2010-12-01 12:11:23.000000000 -0700
+++ tclprimtypes.swg 2010-12-01 12:30:40.000000000 -0700
@@ -133,11 +133,11 @@
SWIGINTERN int
SWIG_AsVal_dec(long long)(Tcl_Obj *obj, long long *val)
{
- long v;
- if (Tcl_GetLongFromObj(0,obj, &v) == TCL_OK) {
- if (val) *val = v;
- return SWIG_OK;
- } else {
+ /*
+ * XXX Bug - remove incorrect optimization.
+ * Optimized fix needs GetWideIntFromObj (tcl >=8.4), so
+ * just remove it for now
+ */
int len = 0;
const char *nptr = Tcl_GetStringFromObj(obj, &len);
if (nptr && len > 0) {
@@ -157,7 +157,6 @@
}
}
}
- }
return SWIG_TypeError;
}
}

[Kjell Wooding]

Here's a regression test for this issue. Add the following as Examples/test-suite/ll_ofl.i:

%module ll_ofl
%inline {
#include <string.h>
#include <stdio.h>
#include <err.h>

void
llofltest(long long a, char *as)
{
char s[22];
snprintf(s, sizeof(s), "%lld", a);
if (strcmp(s, as))
errx(1, "long long convert fail: %s != %s\n", s, as);
}
}

void llofltest(long long a, char *as);

---
Add the test to Examples/test-suite/tcl/Makefile.in:

--- Makefile.in.orig 2010-12-06 16:08:41.000000000 -0700
+++ Makefile.in 2010-12-06 15:30:36.000000000 -0700
@@ -15,6 +15,7 @@
li_cwstring

C_TEST_CASES += \
+ ll_ofl \
li_cstring \
li_cwstring

And add a test to Examples/test-suite/tcl/ll_ofl_runme.tcl

if [ catch { load ./ll_ofl[info sharedlibextension] ll_ofl} err_msg ] {
puts stderr "Could not load shared object:\n$err_msg"
}

llofltest 300000000 300000000
llofltest 3000000000 3000000000
llofltest 30000000000 30000000000

[Olly Betts]

It looks like this is probably still present (at least the code of SWIG_AsVal_dec(long long) is still as pre the fix here) but I think for linux at least it only affects 32 bit architectures, as otherwise sizeof(long) == sizeof(long long).

Here's a regression test just adding a _runme.tcl for an existing testcase:

diff --git a/Examples/test-suite/tcl/integers_runme.tcl b/Examples/test-suite/tcl/integers_runme.tcl
new file mode 100644
index 00000000..c04dd1e9
--- /dev/null
+++ b/Examples/test-suite/tcl/integers_runme.tcl
@@ -0,0 +1,24 @@
+if [ catch { load ./integers[info sharedlibextension] integers} err_msg ] {

+   puts stderr "Could not load shared object:\n$err_msg"
+}
+
+set val 3902408827
+if {[signed_long_identity $val] != $val} {
+    puts stderr "Runtime test 1 failed"
+    exit 1
+}
+
+if {[unsigned_long_identity $val] != $val} {
+    puts stderr "Runtime test 2 failed"
+    exit 1
+}
+
+if {[signed_long_long_identity $val] != $val} {
+    puts stderr "Runtime test 3 failed"
+    exit 1
+}
+
+if {[unsigned_long_long_identity $val] != $val} {
+    puts stderr "Runtime test 4 failed"
+    exit 1
+}

I'll find a 32 bit machine to test on.

Tcl 8.4.0 was tagged 2002-09-10 so we could just require that now.

[Olly Betts]

Reproduced on x86 - the first case above returns a negative value, as expected from the original report.

[Olly Betts]

I've opened a PR with a fix: https://github.com/swig/swig/pull/2331

[Olly Betts]

Merged (https://github.com/swig/swig/commit/06d375cdab63842b223ea5b9091ad0336144e899) - fix should be in SWIG 4.1.0.