Merge pull request #23153 from seiko2plus/removes_old_cpu_dispatcher

SIMD: Get rid of attribute-based CPU dispatching

13 files changed, 436 insertions, 359 deletions

diff --git a/numpy/core/src/common/simd/avx2/memory.h b/numpy/core/src/common/simd/avx2/memory.h
index 64692b54c..81144a36b 100644
--- a/numpy/core/src/common/simd/avx2/memory.h
+++ b/numpy/core/src/common/simd/avx2/memory.h
@@ -215,7 +215,7 @@ NPY_FINLINE npyv_s64 npyv_load_till_s64(const npy_int64 *ptr, npy_uintp nlane, n
     const __m256i steps = npyv_set_s64(0, 1, 2, 3);
     __m256i vnlane  = npyv_setall_s64(nlane > 4 ? 4 : (int)nlane);
     __m256i mask    = _mm256_cmpgt_epi64(vnlane, steps);
-    __m256i payload = _mm256_maskload_epi64((const void*)ptr, mask);
+    __m256i payload = _mm256_maskload_epi64((const long long*)ptr, mask);
     return _mm256_blendv_epi8(vfill, payload, mask);
 }
 // fill zero to rest lanes
@@ -225,7 +225,7 @@ NPY_FINLINE npyv_s64 npyv_load_tillz_s64(const npy_int64 *ptr, npy_uintp nlane)
     const __m256i steps = npyv_set_s64(0, 1, 2, 3);
     __m256i vnlane  = npyv_setall_s64(nlane > 4 ? 4 : (int)nlane);
     __m256i mask    = _mm256_cmpgt_epi64(vnlane, steps);
-    return _mm256_maskload_epi64((const void*)ptr, mask);
+    return _mm256_maskload_epi64((const long long*)ptr, mask);
 }
 
 //// 64-bit nlane
@@ -240,7 +240,7 @@ NPY_FINLINE npyv_s32 npyv_load2_till_s32(const npy_int32 *ptr, npy_uintp nlane,
     const __m256i steps = npyv_set_s64(0, 1, 2, 3);
     __m256i vnlane  = npyv_setall_s64(nlane > 4 ? 4 : (int)nlane);
     __m256i mask    = _mm256_cmpgt_epi64(vnlane, steps);
-    __m256i payload = _mm256_maskload_epi64((const void*)ptr, mask);
+    __m256i payload = _mm256_maskload_epi64((const long long*)ptr, mask);
     return _mm256_blendv_epi8(vfill, payload, mask);
 }
 // fill zero to rest lanes
@@ -253,7 +253,7 @@ NPY_FINLINE npyv_u64 npyv_load2_tillz_s64(const npy_int64 *ptr, npy_uintp nlane)
     assert(nlane > 0);
     npy_int64 m = -((npy_int64)(nlane > 1));
     __m256i mask = npyv_set_s64(-1, -1, m, m);
-    return _mm256_maskload_epi64((const void*)ptr, mask);
+    return _mm256_maskload_epi64((const long long*)ptr, mask);
 }
 // fill zero to rest lanes
 NPY_FINLINE npyv_u64 npyv_load2_till_s64(const npy_int64 *ptr, npy_uintp nlane,
@@ -262,7 +262,7 @@ NPY_FINLINE npyv_u64 npyv_load2_till_s64(const npy_int64 *ptr, npy_uintp nlane,
     const __m256i vfill = npyv_set_s64(0, 0, fill_lo, fill_hi);
     npy_int64 m = -((npy_int64)(nlane > 1));
     __m256i mask = npyv_set_s64(-1, -1, m, m);
-    __m256i payload = _mm256_maskload_epi64((const void*)ptr, mask);
+    __m256i payload = _mm256_maskload_epi64((const long long*)ptr, mask);
     return _mm256_blendv_epi8(vfill, payload, mask);
 }
 /*********************************
@@ -295,7 +295,7 @@ npyv_loadn_till_s64(const npy_int64 *ptr, npy_intp stride, npy_uintp nlane, npy_
     const __m256i steps = npyv_set_s64(0, 1, 2, 3);
     __m256i vnlane  = npyv_setall_s64(nlane > 4 ? 4 : (int)nlane);
     __m256i mask    = _mm256_cmpgt_epi64(vnlane, steps);
-    return _mm256_mask_i64gather_epi64(vfill, (const void*)ptr, idx, mask, 8);
+    return _mm256_mask_i64gather_epi64(vfill, (const long long*)ptr, idx, mask, 8);
 }
 // fill zero to rest lanes
 NPY_FINLINE npyv_s64
@@ -315,7 +315,7 @@ NPY_FINLINE npyv_s64 npyv_loadn2_till_s32(const npy_int32 *ptr, npy_intp stride,
     const __m256i steps = npyv_set_s64(0, 1, 2, 3);
     __m256i vnlane  = npyv_setall_s64(nlane > 4 ? 4 : (int)nlane);
     __m256i mask    = _mm256_cmpgt_epi64(vnlane, steps);
-    return _mm256_mask_i64gather_epi64(vfill, (const void*)ptr, idx, mask, 4);
+    return _mm256_mask_i64gather_epi64(vfill, (const long long*)ptr, idx, mask, 4);
 }
 // fill zero to rest lanes
 NPY_FINLINE npyv_s32 npyv_loadn2_tillz_s32(const npy_int32 *ptr, npy_intp stride, npy_uintp nlane)
@@ -361,7 +361,7 @@ NPY_FINLINE void npyv_store_till_s64(npy_int64 *ptr, npy_uintp nlane, npyv_s64 a
     const __m256i steps = npyv_set_s64(0, 1, 2, 3);
     __m256i vnlane = npyv_setall_s64(nlane > 8 ? 8 : (int)nlane);
     __m256i mask   = _mm256_cmpgt_epi64(vnlane, steps);
-    _mm256_maskstore_epi64((void*)ptr, mask, a);
+    _mm256_maskstore_epi64((long long*)ptr, mask, a);
 }
 
 //// 64-bit nlane
diff --git a/numpy/core/src/common/simd/sse/memory.h b/numpy/core/src/common/simd/sse/memory.h
index b97f3ec2e..4c8e86a6f 100644
--- a/numpy/core/src/common/simd/sse/memory.h
+++ b/numpy/core/src/common/simd/sse/memory.h
@@ -244,13 +244,14 @@ NPY_FINLINE npyv_s32 npyv_load_tillz_s32(const npy_int32 *ptr, npy_uintp nlane)
         return _mm_cvtsi32_si128(*ptr);
     case 2:
         return _mm_loadl_epi64((const __m128i*)ptr);
-    case 3:;
-        npyv_s32 a = _mm_loadl_epi64((const __m128i*)ptr);
-    #ifdef NPY_HAVE_SSE41
-        return _mm_insert_epi32(a, ptr[2], 2);
-    #else
-        return _mm_unpacklo_epi64(a, _mm_cvtsi32_si128(ptr[2]));
-    #endif
+    case 3: {
+            npyv_s32 a = _mm_loadl_epi64((const __m128i*)ptr);
+        #ifdef NPY_HAVE_SSE41
+            return _mm_insert_epi32(a, ptr[2], 2);
+        #else
+            return _mm_unpacklo_epi64(a, _mm_cvtsi32_si128(ptr[2]));
+        #endif
+        }
     default:
         return npyv_load_s32(ptr);
     }
@@ -371,23 +372,27 @@ npyv_loadn_tillz_s32(const npy_int32 *ptr, npy_intp stride, npy_uintp nlane)
     case 1:
         return _mm_cvtsi32_si128(ptr[0]);
     case 2:;
-        npyv_s32 a = _mm_cvtsi32_si128(ptr[0]);
-#ifdef NPY_HAVE_SSE41
-        return _mm_insert_epi32(a, ptr[stride], 1);
-#else
-        return _mm_unpacklo_epi32(a, _mm_cvtsi32_si128(ptr[stride]));
-#endif // NPY_HAVE_SSE41
-    case 3:;
-        a = _mm_cvtsi32_si128(ptr[0]);
-#ifdef NPY_HAVE_SSE41
-        a = _mm_insert_epi32(a, ptr[stride], 1);
-        a = _mm_insert_epi32(a, ptr[stride*2], 2);
-        return a;
-#else
-        a = _mm_unpacklo_epi32(a, _mm_cvtsi32_si128(ptr[stride]));
-        a = _mm_unpacklo_epi64(a, _mm_cvtsi32_si128(ptr[stride*2]));
-        return a;
-#endif // NPY_HAVE_SSE41
+        {
+            npyv_s32 a = _mm_cvtsi32_si128(ptr[0]);
+    #ifdef NPY_HAVE_SSE41
+            return _mm_insert_epi32(a, ptr[stride], 1);
+    #else
+            return _mm_unpacklo_epi32(a, _mm_cvtsi32_si128(ptr[stride]));
+    #endif // NPY_HAVE_SSE41
+        }
+    case 3:
+        {
+            npyv_s32 a = _mm_cvtsi32_si128(ptr[0]);
+    #ifdef NPY_HAVE_SSE41
+            a = _mm_insert_epi32(a, ptr[stride], 1);
+            a = _mm_insert_epi32(a, ptr[stride*2], 2);
+            return a;
+    #else
+            a = _mm_unpacklo_epi32(a, _mm_cvtsi32_si128(ptr[stride]));
+            a = _mm_unpacklo_epi64(a, _mm_cvtsi32_si128(ptr[stride*2]));
+            return a;
+    #endif // NPY_HAVE_SSE41
+        }
     default:
         return npyv_loadn_s32(ptr, stride);
     }
diff --git a/numpy/core/src/common/simd/vec/arithmetic.h b/numpy/core/src/common/simd/vec/arithmetic.h
index 3c13ab06c..85f4d6b26 100644
--- a/numpy/core/src/common/simd/vec/arithmetic.h
+++ b/numpy/core/src/common/simd/vec/arithmetic.h
@@ -366,6 +366,7 @@ NPY_FINLINE float npyv_sum_f32(npyv_f32 a)
 {
     npyv_f32 sum = vec_add(a, npyv_combineh_f32(a, a));
     return vec_extract(sum, 0) + vec_extract(sum, 1);
+    (void)sum;
 }
 #endif
 
@@ -386,6 +387,7 @@ NPY_FINLINE npy_uint16 npyv_sumup_u8(npyv_u8 a)
     npyv_u32 four = vec_sum4s(a, zero);
     npyv_s32 one  = vec_sums((npyv_s32)four, (npyv_s32)zero);
     return (npy_uint16)vec_extract(one, 3);
+    (void)one;
 #endif
 }
 
@@ -400,6 +402,7 @@ NPY_FINLINE npy_uint32 npyv_sumup_u16(npyv_u16 a)
     npyv_u32   four  = vec_add(eight.val[0], eight.val[1]);
     npyv_s32   one   = vec_sums((npyv_s32)four, zero);
     return (npy_uint32)vec_extract(one, 3);
+    (void)one;
 #endif
 }
 
diff --git a/numpy/core/src/common/simd/vec/conversion.h b/numpy/core/src/common/simd/vec/conversion.h
index 0c0d5b3ac..922109f7b 100644
--- a/numpy/core/src/common/simd/vec/conversion.h
+++ b/numpy/core/src/common/simd/vec/conversion.h
@@ -96,6 +96,8 @@ npyv_pack_b8_b64(npyv_b64 a, npyv_b64 b, npyv_b64 c, npyv_b64 d,
     #else
         return vec_extract(r, 4);
     #endif
+        // to suppress ambiguous warning: variable `r` but not used [-Wunused-but-set-variable]
+	(void)r;
     }
     NPY_FINLINE npy_uint64 npyv_tobits_b16(npyv_b16 a)
     {
@@ -106,6 +108,8 @@ npyv_pack_b8_b64(npyv_b64 a, npyv_b64 b, npyv_b64 c, npyv_b64 d,
     #else
         return vec_extract(r, 8);
     #endif
+	// to suppress ambiguous warning: variable `r` but not used [-Wunused-but-set-variable]
+        (void)r;
     }
     NPY_FINLINE npy_uint64 npyv_tobits_b32(npyv_b32 a)
     {
@@ -120,6 +124,8 @@ npyv_pack_b8_b64(npyv_b64 a, npyv_b64 b, npyv_b64 c, npyv_b64 d,
     #else
         return vec_extract(r, 8);
     #endif
+	// to suppress ambiguous warning: variable `r` but not used [-Wunused-but-set-variable]
+        (void)r;
     }
     NPY_FINLINE npy_uint64 npyv_tobits_b64(npyv_b64 a)
     {
@@ -134,6 +140,8 @@ npyv_pack_b8_b64(npyv_b64 a, npyv_b64 b, npyv_b64 c, npyv_b64 d,
     #else
         return vec_extract(r, 8);
     #endif
+	// to suppress ambiguous warning: variable `r` but not used [-Wunused-but-set-variable]
+        (void)r;
     }
 #else
     NPY_FINLINE npy_uint64 npyv_tobits_b8(npyv_b8 a)
diff --git a/numpy/core/src/common/simd/vec/math.h b/numpy/core/src/common/simd/vec/math.h
index 95b16fdf7..85690f76c 100644
--- a/numpy/core/src/common/simd/vec/math.h
+++ b/numpy/core/src/common/simd/vec/math.h
@@ -186,6 +186,7 @@ NPY_IMPL_VEC_REDUCE_MINMAX(max, int32, s32)
     {                                                                   \
         npyv_##SFX r = vec_##INTRIN(a, vec_sld(a, a, 8));               \
         return (npy_##STYPE)vec_extract(r, 0);                          \
+    	(void)r;					 	                                \
     }
 NPY_IMPL_VEC_REDUCE_MINMAX(min, uint64, u64)
 NPY_IMPL_VEC_REDUCE_MINMAX(max, uint64, u64)
@@ -225,6 +226,7 @@ NPY_IMPL_VEC_REDUCE_MINMAX(max, int64, s64)
     {                                                             \
         npyv_f64 r = vec_##INTRIN(a, vec_sld(a, a, 8));           \
         return vec_extract(r, 0);                                 \
+        (void)r;                                                  \
     }                                                             \
     NPY_FINLINE double npyv_reduce_##INTRIN##n_f64(npyv_f64 a)    \
     {                                                             \
diff --git a/numpy/core/src/common/simd/vec/memory.h b/numpy/core/src/common/simd/vec/memory.h
index 3c17617b1..de78d02e3 100644
--- a/numpy/core/src/common/simd/vec/memory.h
+++ b/numpy/core/src/common/simd/vec/memory.h
@@ -46,14 +46,8 @@
 #endif
 
 // avoid aliasing rules
-#ifdef __cplusplus
-    template<typename T_PTR>
-    NPY_FINLINE npy_uint64 *npyv__ptr2u64(const T_PTR *ptr)
-    { npy_uint64 *ptr64 = (npy_uint64*)ptr; return ptr64; }
-#else
-    NPY_FINLINE npy_uint64 *npyv__ptr2u64(const void *ptr)
-    { npy_uint64 *ptr64 = (npy_uint64*)ptr; return ptr64; }
-#endif // __cplusplus
+NPY_FINLINE npy_uint64 *npyv__ptr2u64(const void *ptr)
+{ npy_uint64 *ptr64 = (npy_uint64*)ptr; return ptr64; }
 
 // load lower part
 NPY_FINLINE npyv_u64 npyv__loadl(const void *ptr)
diff --git a/numpy/core/src/common/simd/vec/misc.h b/numpy/core/src/common/simd/vec/misc.h
index 7ea0f21f6..79c194d90 100644
--- a/numpy/core/src/common/simd/vec/misc.h
+++ b/numpy/core/src/common/simd/vec/misc.h
@@ -26,28 +26,28 @@
 #define NPYV_IMPL_VEC_SPLTW(T_VEC, V) ((T_VEC){V, V, V, V})
 #define NPYV_IMPL_VEC_SPLTD(T_VEC, V) ((T_VEC){V, V})
 
-#define npyv_setall_u8(VAL)  NPYV_IMPL_VEC_SPLTB(npyv_u8,  (unsigned char)VAL)
-#define npyv_setall_s8(VAL)  NPYV_IMPL_VEC_SPLTB(npyv_s8,  (signed char)VAL)
-#define npyv_setall_u16(VAL) NPYV_IMPL_VEC_SPLTH(npyv_u16, (unsigned short)VAL)
-#define npyv_setall_s16(VAL) NPYV_IMPL_VEC_SPLTH(npyv_s16, (short)VAL)
-#define npyv_setall_u32(VAL) NPYV_IMPL_VEC_SPLTW(npyv_u32, (unsigned int)VAL)
-#define npyv_setall_s32(VAL) NPYV_IMPL_VEC_SPLTW(npyv_s32, (int)VAL)
+#define npyv_setall_u8(VAL)  NPYV_IMPL_VEC_SPLTB(npyv_u8,  (unsigned char)(VAL))
+#define npyv_setall_s8(VAL)  NPYV_IMPL_VEC_SPLTB(npyv_s8,  (signed char)(VAL))
+#define npyv_setall_u16(VAL) NPYV_IMPL_VEC_SPLTH(npyv_u16, (unsigned short)(VAL))
+#define npyv_setall_s16(VAL) NPYV_IMPL_VEC_SPLTH(npyv_s16, (short)(VAL))
+#define npyv_setall_u32(VAL) NPYV_IMPL_VEC_SPLTW(npyv_u32, (unsigned int)(VAL))
+#define npyv_setall_s32(VAL) NPYV_IMPL_VEC_SPLTW(npyv_s32, (int)(VAL))
 #if NPY_SIMD_F32
-    #define npyv_setall_f32(VAL) NPYV_IMPL_VEC_SPLTW(npyv_f32, VAL)
+    #define npyv_setall_f32(VAL) NPYV_IMPL_VEC_SPLTW(npyv_f32, (VAL))
 #endif
-#define npyv_setall_u64(VAL) NPYV_IMPL_VEC_SPLTD(npyv_u64, (npy_uint64)VAL)
-#define npyv_setall_s64(VAL) NPYV_IMPL_VEC_SPLTD(npyv_s64, (npy_int64)VAL)
+#define npyv_setall_u64(VAL) NPYV_IMPL_VEC_SPLTD(npyv_u64, (npy_uint64)(VAL))
+#define npyv_setall_s64(VAL) NPYV_IMPL_VEC_SPLTD(npyv_s64, (npy_int64)(VAL))
 #define npyv_setall_f64(VAL) NPYV_IMPL_VEC_SPLTD(npyv_f64, VAL)
 
 // vector with specific values set to each lane and
 // set a specific value to all remained lanes
-#define npyv_setf_u8(FILL, ...)  ((npyv_u8){NPYV__SET_FILL_16(char, FILL, __VA_ARGS__)})
-#define npyv_setf_s8(FILL, ...)  ((npyv_s8){NPYV__SET_FILL_16(char, FILL, __VA_ARGS__)})
-#define npyv_setf_u16(FILL, ...) ((npyv_u16){NPYV__SET_FILL_8(short, FILL, __VA_ARGS__)})
+#define npyv_setf_u8(FILL, ...)  ((npyv_u8){NPYV__SET_FILL_16(unsigned char, FILL, __VA_ARGS__)})
+#define npyv_setf_s8(FILL, ...)  ((npyv_s8){NPYV__SET_FILL_16(signed char, FILL, __VA_ARGS__)})
+#define npyv_setf_u16(FILL, ...) ((npyv_u16){NPYV__SET_FILL_8(unsigned short, FILL, __VA_ARGS__)})
 #define npyv_setf_s16(FILL, ...) ((npyv_s16){NPYV__SET_FILL_8(short, FILL, __VA_ARGS__)})
-#define npyv_setf_u32(FILL, ...) ((npyv_u32){NPYV__SET_FILL_4(int, FILL, __VA_ARGS__)})
+#define npyv_setf_u32(FILL, ...) ((npyv_u32){NPYV__SET_FILL_4(unsigned int, FILL, __VA_ARGS__)})
 #define npyv_setf_s32(FILL, ...) ((npyv_s32){NPYV__SET_FILL_4(int, FILL, __VA_ARGS__)})
-#define npyv_setf_u64(FILL, ...) ((npyv_u64){NPYV__SET_FILL_2(npy_int64, FILL, __VA_ARGS__)})
+#define npyv_setf_u64(FILL, ...) ((npyv_u64){NPYV__SET_FILL_2(npy_uint64, FILL, __VA_ARGS__)})
 #define npyv_setf_s64(FILL, ...) ((npyv_s64){NPYV__SET_FILL_2(npy_int64, FILL, __VA_ARGS__)})
 #if NPY_SIMD_F32
     #define npyv_setf_f32(FILL, ...) ((npyv_f32){NPYV__SET_FILL_4(float, FILL, __VA_ARGS__)})
diff --git a/numpy/core/src/npysort/mergesort.cpp b/numpy/core/src/npysort/mergesort.cpp
index 60d89ddb7..f53feb320 100644
--- a/numpy/core/src/npysort/mergesort.cpp
+++ b/numpy/core/src/npysort/mergesort.cpp
@@ -171,6 +171,7 @@ amergesort_(type *v, npy_intp *tosort, npy_intp num)
 }
 
 /*
+ 
  *****************************************************************************
  **                             STRING SORTS                                **
  *****************************************************************************
diff --git a/numpy/core/src/umath/fast_loop_macros.h b/numpy/core/src/umath/fast_loop_macros.h
index cade26db4..b8c1926b2 100644
--- a/numpy/core/src/umath/fast_loop_macros.h
+++ b/numpy/core/src/umath/fast_loop_macros.h
@@ -12,6 +12,19 @@
 
 #include <assert.h>
 
+#include "simd/simd.h"
+
+/*
+ * largest simd vector size in bytes numpy supports
+ * it is currently a extremely large value as it is only used for memory
+ * overlap checks
+ */
+#if NPY_SIMD > 0
+    // Enough for compiler unroll
+    #define AUTOVEC_OVERLAP_SIZE NPY_SIMD_WIDTH*4
+#else
+    #define AUTOVEC_OVERLAP_SIZE 1024
+#endif
 /*
  * MAX_STEP_SIZE is used to determine if we need to use SIMD version of the ufunc.
  * Very large step size can be as slow as processing it using scalar. The
@@ -219,11 +232,11 @@ abs_ptrdiff(char *a, char *b)
         /* condition allows compiler to optimize the generic macro */ \
         if (IS_BINARY_CONT(tin, tout)) { \
             if (abs_ptrdiff(args[2], args[0]) == 0 && \
-                    abs_ptrdiff(args[2], args[1]) >= NPY_MAX_SIMD_SIZE) { \
+                    abs_ptrdiff(args[2], args[1]) >= AUTOVEC_OVERLAP_SIZE) { \
                 BASE_BINARY_LOOP_INP(tin, tout, op) \
             } \
             else if (abs_ptrdiff(args[2], args[1]) == 0 && \
-                         abs_ptrdiff(args[2], args[0]) >= NPY_MAX_SIMD_SIZE) { \
+                         abs_ptrdiff(args[2], args[0]) >= AUTOVEC_OVERLAP_SIZE) { \
                 BASE_BINARY_LOOP_INP(tin, tout, op) \
             } \
             else { \
diff --git a/numpy/core/src/umath/loops.c.src b/numpy/core/src/umath/loops.c.src
index 5684f26a5..397ebaca2 100644
--- a/numpy/core/src/umath/loops.c.src
+++ b/numpy/core/src/umath/loops.c.src
@@ -32,16 +32,6 @@
  */
 #define PW_BLOCKSIZE    128
 
-
-/*
- * largest simd vector size in bytes numpy supports
- * it is currently a extremely large value as it is only used for memory
- * overlap checks
- */
-#ifndef NPY_MAX_SIMD_SIZE
-#define NPY_MAX_SIMD_SIZE 1024
-#endif
-
 /** Provides the various *_LOOP macros */
 #include "fast_loop_macros.h"
 
@@ -417,24 +407,6 @@ BOOL__ones_like(char **args, npy_intp const *dimensions, npy_intp const *steps,
     }
 }
 
-
-/**begin repeat
- * #kind = isnan, isinf, isfinite#
- * #func = npy_isnan, npy_isinf, npy_isfinite#
- * #val = NPY_FALSE, NPY_FALSE, NPY_TRUE#
- **/
-NPY_NO_EXPORT void
-BOOL_@kind@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    /*
-     * The (void)in; suppresses an unused variable warning raised by gcc and allows
-     * us to re-use this macro even though we do not depend on in
-     */
-    UNARY_LOOP_FAST(npy_bool, npy_bool, (void)in; *out = @val@);
-}
-
-/**end repeat**/
-
 /*
  *****************************************************************************
  **                           INTEGER LOOPS
@@ -466,82 +438,16 @@ NPY_NO_EXPORT void
         *((@type@ *)op1) = 1;
     }
 }
-
-NPY_NO_EXPORT void
-@TYPE@_positive(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = +in);
-}
-
 /**begin repeat1
- * #isa = , _avx2#
- * #CHK = 1, defined(HAVE_ATTRIBUTE_TARGET_AVX2)#
- * #ATTR = , NPY_GCC_TARGET_AVX2#
- */
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_square@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = in * in);
-}
-#endif
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_reciprocal@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = 1.0 / in);
-}
-#endif
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_conjugate@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = in);
-}
-#endif
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_logical_not@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, npy_bool, *out = !in);
-}
-#endif
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_invert@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = ~in);
-}
-#endif
-
-/**begin repeat2
  * Arithmetic
  * #kind = add, subtract, multiply, bitwise_and, bitwise_or, bitwise_xor#
  * #OP = +, -, *, &, |, ^#
  */
 
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_@kind@@isa@(char **args, npy_intp const *dimensions,
-                   npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    if (IS_BINARY_REDUCE) {
-        BINARY_REDUCE_LOOP_FAST(@type@, io1 @OP@= in2);
-    }
-    else {
-        BINARY_LOOP_FAST(@type@, @type@, *out = in1 @OP@ in2);
-    }
-}
-
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ int
-@TYPE@_@kind@@isa@_indexed(PyArrayMethod_Context *NPY_UNUSED(context),
-         char **args, npy_intp const *dimensions, npy_intp const *steps,
-         void *NPY_UNUSED(func))
+NPY_NO_EXPORT NPY_GCC_OPT_3 int
+@TYPE@_@kind@_indexed(PyArrayMethod_Context *NPY_UNUSED(context),
+                      char **args, npy_intp const *dimensions, npy_intp const *steps,
+                      void *NPY_UNUSED(func))
 {
     char *ip1 = args[0];
     char *indx = args[1];
@@ -556,86 +462,6 @@ NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ int
     }
     return 0;
 }
-
-#endif
-
-/**end repeat2**/
-
-/*
- * Arithmetic bit shift operations.
- *
- * Intel hardware masks bit shift values, so large shifts wrap around
- * and can produce surprising results. The special handling ensures that
- * behavior is independent of compiler or hardware.
- * TODO: We could implement consistent behavior for negative shifts,
- *       which is undefined in C.
- */
-
-#define INT_left_shift_needs_clear_floatstatus
-#define UINT_left_shift_needs_clear_floatstatus
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 void
-@TYPE@_left_shift@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps,
-                  void *NPY_UNUSED(func))
-{
-    BINARY_LOOP_FAST(@type@, @type@, *out = npy_lshift@c@(in1, in2));
-
-#ifdef @TYPE@_left_shift_needs_clear_floatstatus
-    // For some reason, our macOS CI sets an "invalid" flag here, but only
-    // for some types.
-    npy_clear_floatstatus_barrier((char*)dimensions);
-#endif
-}
-#endif
-
-#undef INT_left_shift_needs_clear_floatstatus
-#undef UINT_left_shift_needs_clear_floatstatus
-
-#if @CHK@
-NPY_NO_EXPORT
-#ifndef NPY_DO_NOT_OPTIMIZE_@TYPE@_right_shift
-NPY_GCC_OPT_3
-#endif
-void
-@TYPE@_right_shift@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps,
-                   void *NPY_UNUSED(func))
-{
-    BINARY_LOOP_FAST(@type@, @type@, *out = npy_rshift@c@(in1, in2));
-}
-#endif
-
-/**begin repeat2
- * #kind = logical_and, logical_or#
- * #OP = &&, ||#
- */
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_@kind@@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    /*
-     * gcc vectorization of this is not good (PR60575) but manual integer
-     * vectorization is too tedious to be worthwhile
-     */
-    BINARY_LOOP_FAST(@type@, npy_bool, *out = in1 @OP@ in2);
-}
-#endif
-
-/**end repeat2**/
-
-#if @CHK@
-NPY_NO_EXPORT NPY_GCC_OPT_3 @ATTR@ void
-@TYPE@_logical_xor@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    BINARY_LOOP {
-        const int t1 = !!*(@type@ *)ip1;
-        const int t2 = !!*(@type@ *)ip2;
-        *((npy_bool *)op1) = (t1 != t2);
-    }
-}
-#endif
-
 /**end repeat1**/
 
 NPY_NO_EXPORT void
@@ -677,23 +503,6 @@ NPY_NO_EXPORT void
         *((@type@ *) op1) = out;
     }
 }
-
-/**begin repeat1
- * #kind = isnan, isinf, isfinite#
- * #func = npy_isnan, npy_isinf, npy_isfinite#
- * #val = NPY_FALSE, NPY_FALSE, NPY_TRUE#
- **/
-NPY_NO_EXPORT void
-@TYPE@_@kind@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    /*
-     * The (void)in; suppresses an unused variable warning raised by gcc and allows
-     * us to re-use this macro even though we do not depend on in
-     */
-    UNARY_LOOP_FAST(@type@, npy_bool, (void)in; *out = @val@);
-}
-/**end repeat1**/
-
 /**end repeat**/
 
 /**begin repeat
@@ -701,19 +510,6 @@ NPY_NO_EXPORT void
  * #type = npy_byte, npy_short, npy_int, npy_long, npy_longlong#
  * #c    = ,,,l,ll#
  */
-
-NPY_NO_EXPORT NPY_GCC_OPT_3 void
-@TYPE@_absolute(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = (in >= 0) ? in : -in);
-}
-
-NPY_NO_EXPORT NPY_GCC_OPT_3 void
-@TYPE@_sign(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = in > 0 ? 1 : (in < 0 ? -1 : 0));
-}
-
 /**begin repeat1
  * #kind = gcd, lcm#
  **/
@@ -727,7 +523,6 @@ NPY_NO_EXPORT void
     }
 }
 /**end repeat1**/
-
 /**end repeat**/
 
 /**begin repeat
@@ -735,19 +530,6 @@ NPY_NO_EXPORT void
  * #type = npy_ubyte, npy_ushort, npy_uint, npy_ulong, npy_ulonglong#
  * #c    = u,u,u,ul,ull#
  */
-
-NPY_NO_EXPORT NPY_GCC_OPT_3 void
-@TYPE@_absolute(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = in);
-}
-
-NPY_NO_EXPORT NPY_GCC_OPT_3 void
-@TYPE@_sign(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(@type@, @type@, *out = in > 0 ? 1 : 0);
-}
-
 /**begin repeat1
  * #kind = gcd, lcm#
  **/
@@ -761,7 +543,6 @@ NPY_NO_EXPORT void
     }
 }
 /**end repeat1**/
-
 /**end repeat**/
 
 /*
@@ -840,12 +621,6 @@ NPY_NO_EXPORT void
 }
 
 NPY_NO_EXPORT void
-@TYPE@_isinf(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(npy_bool, npy_bool, (void)in; *out = NPY_FALSE);
-}
-
-NPY_NO_EXPORT void
 @TYPE@__ones_like(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data))
 {
     OUTPUT_LOOP {
@@ -1714,7 +1489,7 @@ HALF_@kind@_indexed(void *NPY_UNUSED(context),
         const float v = npy_half_to_float(*(npy_half *)value);
         *indexed = npy_float_to_half(npy_half_to_float(*indexed) @OP@ v);
     }
-    return 0; 
+    return 0;
 }
 /**end repeat**/
 
@@ -1974,12 +1749,6 @@ HALF_conjugate(char **args, npy_intp const *dimensions, npy_intp const *steps, v
     }
 }
 
-NPY_NO_EXPORT NPY_GCC_OPT_3 void
-HALF_absolute(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
-{
-    UNARY_LOOP_FAST(npy_half, npy_half, *out = in&0x7fffu);
-}
-
 NPY_NO_EXPORT void
 HALF_negative(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
 {
diff --git a/numpy/core/src/umath/loops.h.src b/numpy/core/src/umath/loops.h.src
index e393b8310..e36067822 100644
--- a/numpy/core/src/umath/loops.h.src
+++ b/numpy/core/src/umath/loops.h.src
@@ -68,6 +68,17 @@ NPY_NO_EXPORT void
 BOOL_@kind@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 /**end repeat**/
 
+
+#ifndef NPY_DISABLE_OPTIMIZATION
+    #include "loops_autovec.dispatch.h"
+#endif
+/**begin repeat
+ * #kind = isnan, isinf, isfinite#
+ */
+NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void BOOL_@kind@,
+    (char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)))
+/**end repeat**/
+
 /*
  *****************************************************************************
  **                           INTEGER LOOPS
@@ -123,16 +134,34 @@ NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@kind@,
 /**end repeat1**/
 /**end repeat**/
 
+
+#ifndef NPY_DISABLE_OPTIMIZATION
+    #include "loops_autovec.dispatch.h"
+#endif
 /**begin repeat
- * #TYPE = BYTE, SHORT, INT, LONG, LONGLONG#
+ * #TYPE = UBYTE, USHORT, UINT, ULONG, ULONGLONG,
+           BYTE,  SHORT,  INT,  LONG,  LONGLONG#
+ */
+/**begin repeat1
+ * #kind = invert, logical_not, conjugate, reciprocal, square, add,
+ *         subtract, multiply, bitwise_and, bitwise_or, bitwise_xor,
+ *         left_shift, right_shift, logical_and, logical_or,
+ *         logical_xor, isnan, isinf, isfinite,
+ *         absolute, sign#
  */
+NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@kind@,
+    (char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)))
+/**end repeat1**/
+/**end repeat**/
 
+/**begin repeat
+ * #TYPE = BYTE, SHORT, INT, LONG, LONGLONG#
+ */
 /**begin repeat1
  * both signed and unsigned integer types
  * #s = , u#
  * #S = , U#
  */
-
 #define @S@@TYPE@_floor_divide @S@@TYPE@_divide
 #define @S@@TYPE@_floor_divide_indexed @S@@TYPE@_divide_indexed
 #define @S@@TYPE@_fmax @S@@TYPE@_maximum
@@ -147,49 +176,15 @@ NPY_NO_EXPORT void
 @S@@TYPE@_positive(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 
 /**begin repeat2
- * #isa = , _avx2#
- */
-
-NPY_NO_EXPORT void
-@S@@TYPE@_square@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data));
-
-NPY_NO_EXPORT void
-@S@@TYPE@_reciprocal@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data));
-
-NPY_NO_EXPORT void
-@S@@TYPE@_conjugate@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-NPY_NO_EXPORT void
-@S@@TYPE@_logical_not@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-NPY_NO_EXPORT void
-@S@@TYPE@_invert@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-/**begin repeat3
  * Arithmetic
  * #kind = add, subtract, multiply, bitwise_and, bitwise_or, bitwise_xor,
  *          left_shift, right_shift#
  * #OP = +, -,*, &, |, ^, <<, >>#
  */
-NPY_NO_EXPORT void
-@S@@TYPE@_@kind@@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
 NPY_NO_EXPORT int
-@S@@TYPE@_@kind@@isa@_indexed(PyArrayMethod_Context *NPY_UNUSED(context), char *const *args, npy_intp const *dimensions, npy_intp const *steps, NpyAuxData *NPY_UNUSED(func));
-
-/**end repeat3**/
-
-/**begin repeat3
- * #kind = logical_and, logical_or#
- * #OP = &&, ||#
- */
-NPY_NO_EXPORT void
-@S@@TYPE@_@kind@@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-/**end repeat3**/
+@S@@TYPE@_@kind@_indexed(PyArrayMethod_Context *NPY_UNUSED(context), char *const *args,
+                         npy_intp const *dimensions, npy_intp const *steps, NpyAuxData *NPY_UNUSED(func));
 
-NPY_NO_EXPORT void
-@S@@TYPE@_logical_xor@isa@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 /**end repeat2**/
 
 /**begin repeat2
@@ -207,26 +202,13 @@ NPY_NO_EXPORT void
 @S@@TYPE@_power(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 
 NPY_NO_EXPORT void
-@S@@TYPE@_absolute(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-NPY_NO_EXPORT void
-@S@@TYPE@_sign(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-NPY_NO_EXPORT void
 @S@@TYPE@_gcd(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 
 NPY_NO_EXPORT void
 @S@@TYPE@_lcm(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-/**begin repeat2
- * #kind = isnan, isinf, isfinite#
- **/
-NPY_NO_EXPORT void
-@S@@TYPE@_@kind@(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 /**end repeat2**/
 
 /**end repeat1**/
-
 /**end repeat**/
 
 
@@ -375,18 +357,6 @@ NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@func@,
 /**end repeat1**/
 /**end repeat**/
 
-/**begin repeat
- *  #TYPE = FLOAT, DOUBLE#
- */
-/**begin repeat1
- * #func = maximum, minimum#
- */
-NPY_NO_EXPORT void
-@TYPE@_@func@_avx512f(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
-/**end repeat1**/
-/**end repeat**/
-
 #ifndef NPY_DISABLE_OPTIMIZATION
     #include "loops_trigonometric.dispatch.h"
 #endif
@@ -562,6 +532,19 @@ NPY_NO_EXPORT void
 /**end repeat1**/
 /**end repeat**/
 
+#ifndef NPY_DISABLE_OPTIMIZATION
+    #include "loops_autovec.dispatch.h"
+#endif
+/**begin repeat
+ * #TYPE = HALF#
+ */
+/**begin repeat1
+ * #kind = absolute#
+ */
+NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@kind@,
+    (char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)))
+/**end repeat1**/
+/**end repeat**/
 /*
  *****************************************************************************
  **                           COMPLEX LOOPS                                 **
@@ -728,9 +711,6 @@ NPY_NO_EXPORT void
 NPY_NO_EXPORT void
 @TYPE@_isfinite(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
 
-NPY_NO_EXPORT void
-@TYPE@_isinf(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func));
-
 #define @TYPE@_isnan @TYPE@_isnat
 
 NPY_NO_EXPORT void
@@ -806,6 +786,21 @@ TIMEDELTA_mm_qm_divmod(char **args, npy_intp const *dimensions, npy_intp const *
 #define TIMEDELTA_md_m_floor_divide TIMEDELTA_md_m_divide
 /* #define TIMEDELTA_mm_d_floor_divide TIMEDELTA_mm_d_divide */
 
+
+#ifndef NPY_DISABLE_OPTIMIZATION
+    #include "loops_autovec.dispatch.h"
+#endif
+/**begin repeat
+ * #TYPE = TIMEDELTA, DATETIME#
+ */
+/**begin repeat1
+ * #kind = isinf#
+ */
+NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@kind@,
+    (char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)))
+/**end repeat1**/
+/**end repeat**/
+
 /*
  *****************************************************************************
  **                            OBJECT LOOPS                                 **
diff --git a/numpy/core/src/umath/loops_autovec.dispatch.c.src b/numpy/core/src/umath/loops_autovec.dispatch.c.src
new file mode 100644
index 000000000..bdbfa0f86
--- /dev/null
+++ b/numpy/core/src/umath/loops_autovec.dispatch.c.src
@@ -0,0 +1,287 @@
+/*@targets
+ ** $maxopt $autovec baseline
+ ** sse2 avx2
+ ** neon
+ ** vsx2
+ ** vx
+ **/
+#define _UMATHMODULE
+#define _MULTIARRAYMODULE
+#define NPY_NO_DEPRECATED_API NPY_API_VERSION
+
+#include "simd/simd.h"
+#include "loops_utils.h"
+#include "loops.h"
+// Provides the various *_LOOP macros
+#include "fast_loop_macros.h"
+
+/*
+ *****************************************************************************
+ **                           INTEGER LOOPS
+ *****************************************************************************
+ */
+/*
+ * Arithmetic bit shift operations.
+ *
+ * Intel hardware masks bit shift values, so large shifts wrap around
+ * and can produce surprising results. The special handling ensures that
+ * behavior is independent of compiler or hardware.
+ * TODO: We could implement consistent behavior for negative shifts,
+ *       which is undefined in C.
+ */
+#define INT_left_shift_needs_clear_floatstatus
+#define UINT_left_shift_needs_clear_floatstatus
+
+/**begin repeat
+ * #TYPE = BYTE, UBYTE, SHORT, USHORT, INT, UINT,
+ *         LONG, ULONG, LONGLONG, ULONGLONG#
+ * #type = npy_byte, npy_ubyte, npy_short, npy_ushort, npy_int, npy_uint,
+ *         npy_long, npy_ulong, npy_longlong, npy_ulonglong#
+ * #ftype = npy_float, npy_float, npy_float, npy_float, npy_double, npy_double,
+ *          npy_double, npy_double, npy_double, npy_double#
+ * #SIGNED = 1, 0, 1, 0, 1, 0, 1, 0, 1, 0#
+ * #c = hh,uhh,h,uh,,u,l,ul,ll,ull#
+ */
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_positive)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = +in);
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_square)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = in * in);
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_reciprocal)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = 1.0 / in);
+}
+
+/**begin repeat1
+ * Arithmetic
+ * #kind = add, subtract, multiply#
+ * #OP = +, -, *#
+ */
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    if (IS_BINARY_REDUCE) {
+        BINARY_REDUCE_LOOP_FAST(@type@, io1 @OP@= in2);
+    }
+    else {
+        BINARY_LOOP_FAST(@type@, @type@, *out = in1 @OP@ in2);
+    }
+}
+/**end repeat1**/
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_left_shift)
+(char **args, npy_intp const *dimensions, npy_intp const *steps,
+                  void *NPY_UNUSED(func))
+{
+    BINARY_LOOP_FAST(@type@, @type@, *out = npy_lshift@c@(in1, in2));
+#ifdef @TYPE@_left_shift_needs_clear_floatstatus
+    // For some reason, our macOS CI sets an "invalid" flag here, but only
+    // for some types.
+    npy_clear_floatstatus_barrier((char*)dimensions);
+#endif
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_right_shift)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+#ifndef NPY_DO_NOT_OPTIMIZE_@TYPE@_right_shift
+    BINARY_LOOP_FAST(@type@, @type@, *out = npy_rshift@c@(in1, in2));
+#else
+    BINARY_LOOP {
+        @type@ in1 = *(@type@ *)ip1;
+        @type@ in2 = *(@type@ *)ip2;
+        *(@type@ *)op1 = npy_rshift@c@(in1, in2);
+    }
+#endif
+}
+/**end repeat**/
+
+/*
+ *****************************************************************************
+ **                         UNSIGNED INTEGER LOOPS
+ *****************************************************************************
+ */
+/**begin repeat
+ * #TYPE = UBYTE, USHORT, UINT, ULONG, ULONGLONG#
+ * #type = npy_ubyte, npy_ushort, npy_uint, npy_ulong, npy_ulonglong#
+ * #c    = u,u,u,ul,ull#
+ */
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_absolute)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = in);
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_sign)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = in > 0 ? 1 : 0);
+}
+
+/**begin repeat1
+ * Arithmetic
+ * #kind = bitwise_and, bitwise_or, bitwise_xor#
+ * #OP = &, |, ^#
+ */
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    if (IS_BINARY_REDUCE) {
+        BINARY_REDUCE_LOOP_FAST(@type@, io1 @OP@= in2);
+    }
+    else {
+        BINARY_LOOP_FAST(@type@, @type@, *out = in1 @OP@ in2);
+    }
+}
+/**end repeat1**/
+
+/**begin repeat1
+ * #kind = logical_and, logical_or#
+ * #OP = &&, ||#
+ */
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    /*
+     * gcc vectorization of this is not good (PR60575) but manual integer
+     * vectorization is too tedious to be worthwhile
+     */
+    BINARY_LOOP_FAST(@type@, npy_bool, *out = in1 @OP@ in2);
+}
+/**end repeat1**/
+
+NPY_FINLINE npy_bool @TYPE@_logical_xor_(@type@ in1, @type@ in2)
+{ return (!!in1) != (!!in2); }
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_logical_xor)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    BINARY_LOOP_FAST(@type@, npy_bool, *out = @TYPE@_logical_xor_(in1, in2));
+}
+
+/**begin repeat1
+ * #kind = isnan, isinf, isfinite#
+ * #func = npy_isnan, npy_isinf, npy_isfinite#
+ * #val = NPY_FALSE, NPY_FALSE, NPY_TRUE#
+ **/
+NPY_NO_EXPORT void  NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    /*
+     * The (void)in; suppresses an unused variable warning raised by gcc and allows
+     * us to re-use this macro even though we do not depend on in
+     */
+    UNARY_LOOP_FAST(@type@, npy_bool, (void)in; *out = @val@);
+}
+/**end repeat1**/
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_conjugate)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = in);
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_logical_not)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, npy_bool, *out = !in);
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_invert)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = ~in);
+}
+/**end repeat**/
+
+/*
+ *****************************************************************************
+ **                         SIGNED! INTEGER LOOPS
+ *****************************************************************************
+ */
+
+/**begin repeat
+ * #TYPE = BYTE, SHORT, INT, LONG, LONGLONG#
+ * #type = npy_byte, npy_short, npy_int, npy_long, npy_longlong#
+ * #c    = ,,,l,ll#
+ */
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_absolute)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = (in >= 0) ? in : -in);
+}
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_sign)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(@type@, @type@, *out = in > 0 ? 1 : (in < 0 ? -1 : 0));
+}
+
+/**begin repeat1
+ * #kind = conjugate, invert, isnan, isinf, isfinite,
+ *         logical_and, logical_or, logical_xor, logical_not,
+ *         bitwise_and, bitwise_or, bitwise_xor#
+ **/
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *func)
+{
+    NPY_CPU_DISPATCH_CURFX(U@TYPE@_@kind@)(args, dimensions, steps, func);
+}
+/**end repeat1**/
+/**end repeat**/
+
+/*
+ *****************************************************************************
+ **                             BOOLEAN LOOPS                               **
+ *****************************************************************************
+ */
+/**begin repeat
+ * #kind = isnan, isinf, isfinite#
+ * #func = npy_isnan, npy_isinf, npy_isfinite#
+ * #val = NPY_FALSE, NPY_FALSE, NPY_TRUE#
+ **/
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(BOOL_@kind@)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *func)
+{
+    NPY_CPU_DISPATCH_CURFX(UBYTE_@kind@)(args, dimensions, steps, func);
+}
+/**end repeat**/
+
+/*
+ *****************************************************************************
+ **                          HALF-FLOAT LOOPS                               **
+ *****************************************************************************
+ */
+
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(HALF_absolute)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
+{
+    UNARY_LOOP_FAST(npy_half, npy_half, *out = in&0x7fffu);
+}
+
+/*
+ *****************************************************************************
+ **                           DATETIME LOOPS                                **
+ *****************************************************************************
+ */
+
+/**begin repeat
+ * #type = npy_datetime, npy_timedelta#
+ * #TYPE = DATETIME, TIMEDELTA#
+ */
+NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_isinf)
+(char **args, npy_intp const *dimensions, npy_intp const *steps, void *func)
+{
+    NPY_CPU_DISPATCH_CURFX(ULONGLONG_isinf)(args, dimensions, steps, func);
+}
+/**end repeat**/
diff --git a/numpy/core/src/umath/loops_exponent_log.dispatch.c.src b/numpy/core/src/umath/loops_exponent_log.dispatch.c.src
index 182c57b01..1fac3c150 100644
--- a/numpy/core/src/umath/loops_exponent_log.dispatch.c.src
+++ b/numpy/core/src/umath/loops_exponent_log.dispatch.c.src
@@ -239,7 +239,7 @@ fma_scalef_ps(__m256 poly, __m256 quadrant)
 
 #ifdef SIMD_AVX512F
 
-NPY_FINLINE NPY_GCC_OPT_3 NPY_GCC_TARGET_AVX512F __mmask16
+NPY_FINLINE __mmask16
 avx512_get_full_load_mask_ps(void)
 {
     return 0xFFFF;