Fix fastmath propagation in SVML.

There's still remaining problem with compile_internal that is not considering the flags.
sklam · Jun 25, 2021 · May 24, 2021 · May 25, 2021 · May 25, 2021 · May 25, 2021
commit 4188fe5ca8656bde018c0b2f71df809fa6aae510
diff --git a/numba/core/cpu.py b/numba/core/cpu.py
@@ -162,6 +162,8 @@ def build_map(self, builder, dict_type, item_types, items):
 
 
     def post_lowering(self, mod, library):
+        if self.fastmath:
+            fastmathpass.rewrite_module(mod, self.fastmath)
 
         if self.is32bit:
             # 32-bit machine needs to replace all 64-bit div/rem to avoid

diff --git a/numba/core/fastmathpass.py b/numba/core/fastmathpass.py
@@ -39,6 +39,6 @@ def rewrite_module(mod, options):
     Rewrite the given LLVM module to use fastmath everywhere.
     """
     flags = options.flags
-    FastFloatBinOpVisitor(flags).visit(mod)
+    # FastFloatBinOpVisitor(flags).visit(mod)
     FastFloatCallVisitor(flags).visit(mod)
 
diff --git a/numba/np/ufunc/array_exprs.py b/numba/np/ufunc/array_exprs.py
@@ -376,9 +376,10 @@ def _lower_array_expr(lowerer, expr):
             inner_sig_args.append(argty)
     inner_sig = outer_sig.return_type.dtype(*inner_sig_args)
 
+    flags = utils.ConfigStack().top_or_none() or compiler.Flags()
+    flags = flags.copy() # make sure it's a clone or a fresh instance
     # Follow the Numpy error model.  Note this also allows e.g. vectorizing
     # division (issue #1223).
-    flags = compiler.Flags()
     flags.error_model = 'numpy'
     cres = context.compile_subroutine(builder, impl, inner_sig, flags=flags,
                                       caching=False)