Fixed a couple of places where 'int' was used where 'long'

should have been used.
This commit is contained in:
Marc-André Lemburg 2000-07-06 19:43:31 +00:00
parent 49b0c3bafe
commit 85cc4d8940
1 changed files with 7 additions and 7 deletions

View File

@ -1063,7 +1063,7 @@ PyObject *PyUnicode_DecodeUnicodeEscape(const char *s,
end = s + size;
while (s < end) {
unsigned char c;
unsigned int x;
unsigned long x;
int i;
/* Non-escape characters are interpreted as Unicode ordinals */
@ -1372,7 +1372,7 @@ PyObject *PyUnicode_DecodeRawUnicodeEscape(const char *s,
end = s + size;
while (s < end) {
unsigned char c;
unsigned int x;
unsigned long x;
int i;
/* Non-escape characters are interpreted as Unicode ordinals */
@ -1852,7 +1852,7 @@ PyObject *PyUnicode_DecodeCharmap(const char *s,
/* Apply mapping */
if (PyInt_Check(x)) {
int value = PyInt_AS_LONG(x);
long value = PyInt_AS_LONG(x);
if (value < 0 || value > 65535) {
PyErr_SetString(PyExc_TypeError,
"character mapping must be in range(65536)");
@ -1971,7 +1971,7 @@ PyObject *PyUnicode_EncodeCharmap(const Py_UNICODE *p,
/* Apply mapping */
if (PyInt_Check(x)) {
int value = PyInt_AS_LONG(x);
long value = PyInt_AS_LONG(x);
if (value < 0 || value > 255) {
PyErr_SetString(PyExc_TypeError,
"character mapping must be in range(256)");
@ -3070,7 +3070,7 @@ unicode_compare(PyUnicodeObject *str1, PyUnicodeObject *str2)
while (len1 > 0 && len2 > 0) {
unsigned short c1, c2; /* 16 bits */
int diff; /* 32 bits */
long diff; /* >=32 bits */
c1 = *s1++;
c2 = *s2++;
@ -3080,7 +3080,7 @@ unicode_compare(PyUnicodeObject *str1, PyUnicodeObject *str2)
c2 += utf16Fixup[c2>>11];
/* now c1 and c2 are in UTF-32-compatible order */
diff = (int)c1 - (int)c2;
diff = (long)c1 - (long)c2;
if (diff)
return (diff < 0) ? -1 : (diff != 0);
len1--; len2--;