Merge pull request numba#9336 from stuartarchibald/wip/add_tli_pass

Add TargetLibraryInfo pass to CPU LLVM pipeline.

docs/upcoming_changes/9336.improvement.rst

@@ -0,0 +1,11 @@
+Add ``TargetLibraryInfo`` pass to CPU LLVM pipeline.
+""""""""""""""""""""""""""""""""""""""""""""""""""""
+
+The ``TargetLibraryInfo`` pass makes sure that the optimisations that take place
+during call simplification are appropriate for the target, without this the
+target is assumed to be Linux and code will be optimised to produce e.g. math
+symbols that do not exit on Windows. Historically this issue has been avoided
+through the use of Numba internal libraries carrying wrapped symbols, but doing
+so potentially detriments performance. As a result of this change Numba internal
+libraries are smaller and there is an increase in optimisation opportunity in
+code using ``exp2`` and ``log2`` functions.

numba/_helperlib.c

@@ -122,30 +122,6 @@ numba_ldexpf(float x, int exp)
     return x;
 }
 
-NUMBA_EXPORT_FUNC(double)
-numba_exp2(double x)
-{
-    return exp2(x);
-}
-
-NUMBA_EXPORT_FUNC(float)
-numba_exp2f(float x)
-{
-    return exp2f(x);
-}
-
-NUMBA_EXPORT_FUNC(double)
-numba_log2(double x)
-{
-    return log2(x);
-}
-
-NUMBA_EXPORT_FUNC(float)
-numba_log2f(float x)
-{
-    return log2f(x);
-}
-
 /* provide complex power */
 NUMBA_EXPORT_FUNC(void)
 numba_cpow(Py_complex *a, Py_complex *b, Py_complex *out) {

numba/_helpermod.c

@@ -55,10 +55,6 @@ build_c_helpers_dict(void)
     declmethod(frexpf);
     declmethod(ldexp);
     declmethod(ldexpf);
-    declmethod(exp2);
-    declmethod(exp2f);
-    declmethod(log2);
-    declmethod(log2f);
     declmethod(cpow);
     declmethod(cpowf);
     declmethod(erf);

numba/core/codegen.py

@@ -1225,6 +1225,7 @@ def _create_empty_module(self, name):
 
     def _module_pass_manager(self, **kwargs):
         pm = ll.create_module_pass_manager()
+        pm.add_target_library_info(ll.get_process_triple())
         self._tm.add_analysis_passes(pm)
         cost = kwargs.pop("cost", None)
         with self._pass_manager_builder(**kwargs) as pmb:
@@ -1256,6 +1257,7 @@ def _module_pass_manager(self, **kwargs):
 
     def _function_pass_manager(self, llvm_module, **kwargs):
         pm = ll.create_function_pass_manager(llvm_module)
+        pm.add_target_library_info(llvm_module.triple)
         self._tm.add_analysis_passes(pm)
         with self._pass_manager_builder(**kwargs) as pmb:
             pmb.populate(pm)

numba/np/npyfuncs.py

@@ -648,13 +648,11 @@ def np_complex_exp_impl(context, builder, sig, args):
 def np_real_exp2_impl(context, builder, sig, args):
     _check_arity_and_homogeneity(sig, args, 1)
 
-    dispatch_table = {
-        types.float32: 'numba_exp2f',
-        types.float64: 'numba_exp2',
-    }
-
-    return _dispatch_func_by_name_type(context, builder, sig, args,
-                                       dispatch_table, 'exp2')
+    ll_ty = args[0].type
+    fnty = llvmlite.ir.FunctionType(ll_ty, [ll_ty,])
+    fn = cgutils.insert_pure_function(builder.module, fnty,
+                                      name='llvm.exp2')
+    return builder.call(fn, [args[0]])
 
 
 def np_complex_exp2_impl(context, builder, sig, args):
@@ -687,13 +685,11 @@ def np_complex_log_impl(context, builder, sig, args):
 def np_real_log2_impl(context, builder, sig, args):
     _check_arity_and_homogeneity(sig, args, 1)
 
-    dispatch_table = {
-        types.float32: 'numba_log2f',
-        types.float64: 'numba_log2',
-    }
-
-    return _dispatch_func_by_name_type(context, builder, sig, args,
-                                       dispatch_table, 'log2')
+    ll_ty = args[0].type
+    fnty = llvmlite.ir.FunctionType(ll_ty, [ll_ty,])
+    fn = cgutils.insert_pure_function(builder.module, fnty,
+                                      name='llvm.log2')
+    return builder.call(fn, [args[0]])
 
 def np_complex_log2_impl(context, builder, sig, args):
     _check_arity_and_homogeneity(sig, args, 1)

numba/tests/npyufunc/test_ufunc.py

@@ -145,7 +145,7 @@ class TestUFuncsMisc(TestCase):
     # Test for miscellaneous ufunc issues
 
     def test_exp2(self):
-        # See issue #8898
+        # See issue #8898, and TargetLibraryInfo based fix in #9336
         @njit
         def foo(x):
             return np.exp2(x)
@@ -157,7 +157,7 @@ def foo(x):
             self.assertPreciseEqual(expected, got)
 
     def test_log2(self):
-        # See issue #8898
+        # See issue #8898, and TargetLibraryInfo based fix in #9336
         @njit
         def foo(x):
             return np.log2(x)