Enable threaded Whole Module Optimization
* This seems to fix an issue with dsymutil and lldb not getting correct debug maps out of binaries built with WMO enabled.
--
PiperOrigin-RevId: 142578572
MOS_MIGRATED_REVID=142578572
diff --git a/tools/build_defs/apple/swift.bzl b/tools/build_defs/apple/swift.bzl
index d45a447..fcbdf87 100644
--- a/tools/build_defs/apple/swift.bzl
+++ b/tools/build_defs/apple/swift.bzl
@@ -351,21 +351,24 @@
# This is highly experimental and tracked in b/29465250
has_wmo = ("-wmo" in ctx.attr.copts) or ("-whole-module-optimization" in ctx.attr.copts)
- if not has_wmo:
- for source in ctx.files.srcs:
- basename = source.basename
- obj = ctx.new_file(objs_outputs_path + basename + ".o")
+ for source in ctx.files.srcs:
+ basename = source.basename
+ output_map_entry = {}
+
+ # Output an object file
+ obj = ctx.new_file(objs_outputs_path + basename + ".o")
+ output_objs.append(obj)
+ output_map_entry["object"] = obj.path
+
+ # Output a partial module file, unless WMO is enabled in which case only
+ # the final, complete module will be generated.
+ if not has_wmo:
partial_module = ctx.new_file(objs_outputs_path + basename +
".partial_swiftmodule")
- output_objs.append(obj)
swiftc_outputs.append(partial_module)
+ output_map_entry["swiftmodule"] = partial_module.path
- swiftc_output_map += struct(**{
- source.path: struct(object=obj.path, swiftmodule=partial_module.path)})
- else:
- # When WMO is enabled, there is only one output - the module object file,
- # no need to deal with an output map
- output_objs.append(ctx.new_file(objs_outputs_path + module_name + ".o"))
+ swiftc_output_map += struct(**{source.path: struct(**output_map_entry)})
# Write down the intermediate outputs map for this compilation, to be used
# with -output-file-map flag.
@@ -389,10 +392,13 @@
swiftc_output_map_file.path,
]
- # When WMO is enabled, there's only one output, the module object file, which
- # location should be explicitly given to the driver.
if has_wmo:
- args.extend(["-o", output_objs[0].path])
+ # WMO has two modes: threaded and not. We want the threaded mode because it
+ # will use the output map we generate. This leads to a better debug
+ # experience in lldb and Xcode.
+ # TODO(b/32571265): 12 has been chosen as the best option for a Mac Pro,
+ # we should get an interface in Bazel to get core count.
+ args.extend(["-num-threads", "12"])
xcrun_action(
ctx,
