PokhodenkoSA · Jun 18, 2021 · Jun 18, 2021 · Jun 11, 2021 · Jun 15, 2021
diff --git a/numba/core/cgutils.py b/numba/core/cgutils.py
@@ -370,13 +370,20 @@ def alloca_once(builder, ty, size=None, name='', zfill=False):
     """
     if isinstance(size, int):
         size = ir.Constant(intp_t, size)
+    # suspend dwarf emission else it links up python source lines with alloca in
+    # the entry block as well as their actual location and it makes the debug
+    # info "jump about".
+    ref = builder.debug_metadata
+    builder.debug_metadata = None
     with builder.goto_entry_block():
         ptr = builder.alloca(ty, size=size, name=name)
         # Always zero-fill at init-site.  This is safe.
         builder.store(ptr.type.pointee(None), ptr)
     # Also zero-fill at the use-site
     if zfill:
         builder.store(ptr.type.pointee(None), ptr)
+    # continue with dwarf emission
+    builder.debug_metadata = ref
     return ptr
 
 

diff --git a/numba/core/debuginfo.py b/numba/core/debuginfo.py
@@ -215,7 +215,7 @@ def _di_file(self):
 
     def _di_compile_unit(self):
         return self.module.add_debug_info('DICompileUnit', {
-            'language': ir.DIToken('DW_LANG_Python'),
+            'language': ir.DIToken('DW_LANG_C_plus_plus'),
             'file': self.difile,
             'producer': 'Numba',
             'runtimeVersion': 0,

diff --git a/numba/core/lowering.py b/numba/core/lowering.py
@@ -1302,11 +1302,15 @@ def alloca_lltype(self, name, lltype):
         aptr = cgutils.alloca_once(self.builder, lltype,
                                    name=name, zfill=False)
         if is_uservar:
+            # If it's an arg set the location as the function definition line
+            if name in self.func_ir.arg_names:
+                loc = self.loc.with_lineno(self.func_ir.loc.line)
+            else:
+                loc = self.loc
             # Emit debug info for user variable
             sizeof = self.context.get_abi_sizeof(lltype)
             self.debuginfo.mark_variable(self.builder, aptr, name=name,
-                                         lltype=lltype, size=sizeof,
-                                         loc=self.loc)
+                                         lltype=lltype, size=sizeof, loc=loc)
         return aptr
 
     def incref(self, typ, val):

diff --git a/numba/tests/test_debuginfo.py b/numba/tests/test_debuginfo.py
@@ -1,15 +1,18 @@
 import re
+import inspect
 
 from numba.tests.support import TestCase, override_config
-from numba import jit
+from numba import jit, njit
 from numba.core import types
 import unittest
+import llvmlite.binding as llvm
 
 
 class TestDebugInfo(TestCase):
     """
     These tests only checks the compiled assembly for debuginfo.
     """
+
     def _getasm(self, fn, sig):
         fn.compile(sig)
         return fn.inspect_asm(sig)
@@ -49,5 +52,136 @@ def bar(x):
             self._check(bar, sig=(types.int32,), expect=False)
 
 
+class TestDebugInfoEmission(TestCase):
+    """ Tests that debug info is emitted correctly.
+    """
+
+    def _get_llvmir(self, fn, sig):
+        with override_config('OPT', 0):
+            fn.compile(sig)
+            return fn.inspect_llvm(sig)
+
+    def _get_metadata(self, fn, sig):
+        ll = self._get_llvmir(fn, sig).splitlines()
+        meta_re = re.compile(r'![0-9]+ =.*')
+        metadata = []
+        for line in ll:
+            if meta_re.match(line):
+                metadata.append(line)
+        return metadata
+
+    def test_DW_LANG(self):
+
+        @njit(debug=True)
+        def foo():
+            pass
+
+        metadata = self._get_metadata(foo, sig=())
+        DICompileUnit = metadata[0]
+        self.assertEqual('!0', DICompileUnit[:2])
+        self.assertIn('!DICompileUnit(language: DW_LANG_C_plus_plus',
+                      DICompileUnit)
+        self.assertIn('producer: "Numba"', DICompileUnit)
+
+    def test_DILocation(self):
+        """ Tests that DILocation information is reasonable.
+        """
+
+        @njit(debug=True, error_model='numpy')
+        def foo(a):
+            b = a + 1.23
+            c = a * 2.34
+            d = b / c
+            print(d)
+            return d
+
+        # the above produces LLVM like:
+        # define function() {
+        # entry:
+        #   alloca
+        #   store 0 to alloca
+        #   <arithmetic for doing the operations on b, c, d>
+        #   setup for print
+        #   branch
+        # other_labels:
+        # ... <elided>
+        # }
+        #
+        # The following checks that:
+        # * the alloca and store have no !dbg
+        # * the arithmetic occurs in the order defined and with !dbg
+        # * that the !dbg entries are monotonically increasing in value with
+        #   source line number
+
+        metadata = self._get_metadata(foo, sig=(types.float64,))
+        full_ir = self._get_llvmir(foo, sig=(types.float64,))
+
+        llvm.initialize()
+        llvm.initialize_native_target()
+        llvm.initialize_native_asmprinter()
+
+        module = llvm.parse_assembly(full_ir)
+
+        name = foo.overloads[foo.signatures[0]].fndesc.mangled_name
+        funcs = [x for x in module.functions if x.name == name]
+        self.assertEqual(len(funcs), 1)
+        func = funcs[0]
+        blocks = [x for x in func.blocks]
+        self.assertGreater(len(blocks), 1)
+        block = blocks[0]
+
+        # Find non-call instr and check the sequence is as expected
+        instrs = [x for x in block.instructions if x.opcode != 'call']
+        op_seq = [x.opcode for x in instrs]
+        op_expect = ('fadd', 'fmul', 'fdiv')
+        self.assertIn(''.join(op_expect), ''.join(op_seq))
+
+        # Parse out metadata from end of each line, check it monotonically
+        # ascends with LLVM source line. Also store all the dbg references,
+        # these will be checked later.
+        line2dbg = set()
+        re_dbg_ref = re.compile(r'.*!dbg (![0-9]+).*$')
+        found = -1
+        for instr in instrs:
+            inst_as_str = str(instr)
+            matched = re_dbg_ref.match(inst_as_str)
+            if not matched:
+                # if there's no match, ensure it is one of alloca or store,
+                # it's important that the zero init/alloca instructions have
+                # no dbg data
+                accepted = ('alloca ', 'store ')
+                self.assertTrue(any([x in inst_as_str for x in accepted]))
+                continue
+            groups = matched.groups()
+            self.assertEqual(len(groups), 1)
+            dbg_val = groups[0]
+            int_dbg_val = int(dbg_val[1:])
+            if found >= 0:
+                self.assertTrue(int_dbg_val >= found)
+            found = int_dbg_val
+            # some lines will alias dbg info, this is fine, it's only used to
+            # make sure that the line numbers are correct WRT python
+            line2dbg.add(dbg_val)
+
+        pysrc, pysrc_line_start = inspect.getsourcelines(foo)
+
+        # build a map of dbg reference to DI* information
+        metadata_definition_map = dict()
+        meta_definition_split = re.compile(r'(![0-9]+) = (.*)')
+        for line in metadata:
+            matched = meta_definition_split.match(line)
+            if matched:
+                dbg_val, info = matched.groups()
+                metadata_definition_map[dbg_val] = info
+
+        # Pull out metadata entries referred to by the llvm line end !dbg
+        # check they match the python source, the +2 is for the @njit decorator
+        # and the function definition line.
+        pyln_range = range(pysrc_line_start + 2, pysrc_line_start + len(pysrc))
+        for (k, line_no) in zip(sorted(line2dbg), pyln_range):
+            dilocation_info = metadata_definition_map[k]
+            self.assertIn(f'line: {line_no}', dilocation_info)
+
+
 if __name__ == '__main__':
     unittest.main()