BUG: core: fix ilp64 blas dot/vdot/... for strides > int32 max

Fix overlooked int cast when HAVE_BLAS_ILP64 is defined.
It was supposed to cast to CBLAS_INT, not int.
Also add a regression test.

Move blas_stride() to npy_cblas.h
Replace npy_is_aligned by modulo; we're going to call BLAS so no need to
micro-optimize integer division here.

3 files changed, 40 insertions, 25 deletions

diff --git a/numpy/core/src/common/npy_cblas.h b/numpy/core/src/common/npy_cblas.h
index 97308238a..c0441e81e 100644
--- a/numpy/core/src/common/npy_cblas.h
+++ b/numpy/core/src/common/npy_cblas.h
@@ -59,6 +59,32 @@ enum CBLAS_SIDE {CblasLeft=141, CblasRight=142};
 #undef BLASINT
 #undef BLASNAME
 
+
+/*
+ * Convert NumPy stride to BLAS stride. Returns 0 if conversion cannot be done
+ * (BLAS won't handle negative or zero strides the way we want).
+ */
+static NPY_INLINE CBLAS_INT
+blas_stride(npy_intp stride, unsigned itemsize)
+{
+    /*
+     * Should probably check pointer alignment also, but this may cause
+     * problems if we require complex to be 16 byte aligned.
+     */
+    if (stride > 0 && (stride % itemsize) == 0) {
+        stride /= itemsize;
+#ifndef HAVE_BLAS_ILP64
+        if (stride <= INT_MAX) {
+#else
+        if (stride <= NPY_MAX_INT64) {
+#endif
+            return stride;
+        }
+    }
+    return 0;
+}
+
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/numpy/core/src/multiarray/common.h b/numpy/core/src/multiarray/common.h
index 793cefaf8..e0b143604 100644
--- a/numpy/core/src/multiarray/common.h
+++ b/numpy/core/src/multiarray/common.h
@@ -293,30 +293,6 @@ npy_memchr(char * haystack, char needle,
 }
 
 /*
- * Convert NumPy stride to BLAS stride. Returns 0 if conversion cannot be done
- * (BLAS won't handle negative or zero strides the way we want).
- */
-static NPY_INLINE int
-blas_stride(npy_intp stride, unsigned itemsize)
-{
-    /*
-     * Should probably check pointer alignment also, but this may cause
-     * problems if we require complex to be 16 byte aligned.
-     */
-    if (stride > 0 && npy_is_aligned((void *)stride, itemsize)) {
-        stride /= itemsize;
-#ifndef HAVE_BLAS_ILP64
-        if (stride <= INT_MAX) {
-#else
-        if (stride <= NPY_MAX_INT64) {
-#endif
-            return stride;
-        }
-    }
-    return 0;
-}
-
-/*
  * Define a chunksize for CBLAS. CBLAS counts in integers.
  */
 #if NPY_MAX_INTP > INT_MAX
diff --git a/numpy/core/tests/test_regression.py b/numpy/core/tests/test_regression.py
index f778d4d7c..a97198076 100644
--- a/numpy/core/tests/test_regression.py
+++ b/numpy/core/tests/test_regression.py
@@ -14,7 +14,7 @@ from numpy.testing import (
         assert_raises_regex, assert_warns, suppress_warnings,
         _assert_valid_refcount, HAS_REFCOUNT,
         )
-from numpy.testing._private.utils import _no_tracing
+from numpy.testing._private.utils import _no_tracing, requires_memory
 from numpy.compat import asbytes, asunicode, pickle
 
 try:
@@ -2501,3 +2501,16 @@ class TestRegression:
                            formats=[np.int64, np.int64]))
         descr = np.array((1, 1), dtype=dt).__array_interface__['descr']
         assert descr == [('', '|V8')]  # instead of [(b'', '|V8')]
+
+    @pytest.mark.skipif(sys.maxsize < 2 ** 31 + 1, reason='overflows 32-bit python')
+    @requires_memory(free_bytes=9e9)
+    def test_dot_big_stride(self):
+        # gh-17111
+        # blas stride = stride//itemsize > int32 max
+        int32_max = np.iinfo(np.int32).max
+        n = int32_max + 3
+        a = np.empty([n], dtype=np.float32)
+        b = a[::n-1]
+        b[...] = 1
+        assert b.strides[0] > int32_max * b.dtype.itemsize
+        assert np.dot(b, b) == 2.0