BUG: rolling with MSVC 2017 build (pandas-dev#21813)

* Appveyor 3.7

* ci package list

* change image type

* try hack fix

* lint

* use isnan on problem function

* use numpy compat isnan

* use right isnan

* work around OSX math undefs

* cleanup const

* fix reversion

* ...

(cherry picked from commit 7a2fbce)

Author: chris-b1

appveyor.yml

@@ -20,12 +20,14 @@ environment:
   matrix:
 
     - CONDA_ROOT: "C:\\Miniconda3_64"
+      APPVEYOR_BUILD_WORKER_IMAGE: Visual Studio 2017
       PYTHON_VERSION: "3.6"
       PYTHON_ARCH: "64"
       CONDA_PY: "36"
       CONDA_NPY: "113"
 
     - CONDA_ROOT: "C:\\Miniconda3_64"
+      APPVEYOR_BUILD_WORKER_IMAGE: Visual Studio 2015
       PYTHON_VERSION: "2.7"
       PYTHON_ARCH: "64"
       CONDA_PY: "27"

doc/source/whatsnew/v0.23.4.txt

@@ -16,7 +16,7 @@ and bug fixes. We recommend that all users upgrade to this version.
 Fixed Regressions
 ~~~~~~~~~~~~~~~~~
 
-- 
+- Python 3.7 with Windows gave all missing values for rolling variance calculations (:issue:`21813`)
 -
 
 .. _whatsnew_0234.bug_fixes:

pandas/_libs/src/headers/cmath

@@ -6,6 +6,7 @@
 #if defined(_MSC_VER) && (_MSC_VER < 1800)
 #include <cmath>
 namespace std {
+  __inline int isnan(double x) { return _isnan(x); }
   __inline int signbit(double num) { return _copysign(1.0, num) < 0; }
 }
 #else

pandas/_libs/window.pyx

@@ -14,6 +14,7 @@ cnp.import_array()
 
 
 cdef extern from "../src/headers/cmath" namespace "std":
+    bint isnan(double) nogil
     int signbit(double) nogil
     double sqrt(double x) nogil
 
@@ -654,16 +655,16 @@ cdef inline void add_var(double val, double *nobs, double *mean_x,
                          double *ssqdm_x) nogil:
     """ add a value from the var calc """
     cdef double delta
-
-    # Not NaN
-    if val == val:
-        nobs[0] = nobs[0] + 1
-
-        # a part of Welford's method for the online variance-calculation
-        # https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
-        delta = val - mean_x[0]
-        mean_x[0] = mean_x[0] + delta / nobs[0]
-        ssqdm_x[0] = ssqdm_x[0] + ((nobs[0] - 1) * delta ** 2) / nobs[0]
+    # `isnan` instead of equality as fix for GH-21813, msvc 2017 bug
+    if isnan(val):
+        return
+
+    nobs[0] = nobs[0] + 1
+    # a part of Welford's method for the online variance-calculation
+    # https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
+    delta = val - mean_x[0]
+    mean_x[0] = mean_x[0] + delta / nobs[0]
+    ssqdm_x[0] = ssqdm_x[0] + ((nobs[0] - 1) * delta ** 2) / nobs[0]
 
 
 cdef inline void remove_var(double val, double *nobs, double *mean_x,