dwp should use post-bolt external debug information

Summary: We are passing in pre-bolt dwo files to `llvm-dwp`. This is not a problem if `llvm-bolt` and `llvm-dwp` both run locally, but causes issue when we attempt to run them remotely to cache the binaries.

Reviewed By: keshavdv

Differential Revision: D68116838

fbshipit-source-id: e54b5894c31e68e65bab57b5965244659622c7b6

prelude/cxx/cxx_bolt.bzl

@@ -15,13 +15,19 @@ load(
 )
 load(":cxx_context.bzl", "get_cxx_toolchain_info")
 
+CxxBoltOutput = record(
+    output = field(Artifact),
+    dwo_output = field(Artifact | None),
+)
+
 def cxx_use_bolt(ctx: AnalysisContext) -> bool:
     cxx_toolchain_info = get_cxx_toolchain_info(ctx)
     return cxx_toolchain_info.bolt_enabled and ctx.attrs.bolt_profile != None
 
-def bolt(ctx: AnalysisContext, prebolt_output: Artifact, external_debug_info: ArtifactTSet, identifier: [str, None]) -> Artifact:
+def bolt(ctx: AnalysisContext, prebolt_output: Artifact, external_debug_info: ArtifactTSet, identifier: [str, None], generate_dwp: bool) -> CxxBoltOutput:
     output_name = prebolt_output.short_path.removesuffix("-wrapper")
     postbolt_output = ctx.actions.declare_output(output_name)
+    dwo_output = None
     bolt_msdk = get_cxx_toolchain_info(ctx).binary_utilities_info.bolt_msdk
 
     if not bolt_msdk or not cxx_use_bolt(ctx):
@@ -41,6 +47,17 @@ def bolt(ctx: AnalysisContext, prebolt_output: Artifact, external_debug_info: Ar
         hidden = materialized_external_debug_info,
     )
 
+    if generate_dwp:
+        dwo_output = ctx.actions.declare_output(output_name + ".dwo.d", dir = True)
+        args.add(cmd_args(dwo_output.as_output(), format = "--dwarf-output-path={}"))
+        args = cmd_args(
+            "/bin/sh",
+            "-c",
+            cmd_args(dwo_output.as_output(), format = 'mkdir -p {}; "$@"'),
+            '""',
+            args,
+        )
+
     ctx.actions.run(
         args,
         category = "bolt",
@@ -67,4 +84,4 @@ def bolt(ctx: AnalysisContext, prebolt_output: Artifact, external_debug_info: Ar
         )
         output = stripped_postbolt_output
 
-    return output
+    return CxxBoltOutput(output = output, dwo_output = dwo_output)

prelude/cxx/dist_lto/dist_lto.bzl

@@ -626,12 +626,22 @@ def cxx_gnu_dist_link(
         ],
     )
 
-    final_output = output if not (executable_link and cxx_use_bolt(ctx)) else bolt(ctx, output, external_debug_info, identifier)
+    if (executable_link and cxx_use_bolt(ctx)):
+        bolt_output = bolt(ctx, output, external_debug_info, identifier, generate_dwp)
+        final_output = bolt_output.output
+        split_debug_output = bolt_output.dwo_output
+    else:
+        final_output = output
+        split_debug_output = None
+
     dwp_output = ctx.actions.declare_output(output.short_path.removesuffix("-wrapper") + ".dwp") if generate_dwp else None
 
     if generate_dwp:
-        materialized_external_debug_info = project_artifacts(ctx.actions, [external_debug_info])
-        referenced_objects = final_link_inputs + materialized_external_debug_info
+        if split_debug_output:
+            referenced_objects = final_link_inputs + [split_debug_output]
+        else:
+            materialized_external_debug_info = project_artifacts(ctx.actions, [external_debug_info])
+            referenced_objects = final_link_inputs + materialized_external_debug_info
         run_dwp_action(
             ctx = ctx,
             toolchain = cxx_toolchain,

prelude/cxx/link.bzl

@@ -331,13 +331,19 @@ def cxx_link_into(
         strip_args = opts.strip_args_factory(ctx) if opts.strip_args_factory else cmd_args()
         output = strip_object(ctx, cxx_toolchain_info, output, strip_args, opts.category_suffix)
 
-    final_output = output if not (is_result_executable and cxx_use_bolt(ctx)) else bolt(ctx, output, external_debug_info, opts.identifier)
+    use_bolt = (is_result_executable and cxx_use_bolt(ctx))
+    if use_bolt:
+        bolt_output = bolt(ctx, output, external_debug_info, opts.identifier, should_generate_dwp)
+        output = bolt_output.output
+        split_debug_output = bolt_output.dwo_output
 
     dwp_artifact = None
     if should_generate_dwp:
         dwp_inputs = cmd_args()
         dwp_from_dwo = getattr(ctx.attrs, "separate_debug_info", False) and cxx_toolchain_info.split_debug_mode == SplitDebugMode("split")
-        if dwp_from_dwo:
+        if use_bolt:
+            dwp_inputs.add([split_debug_output])
+        elif dwp_from_dwo:
             dwp_inputs.add(project_identified_artifacts(ctx.actions, [external_debug_info]))
         else:
             for link in opts.links:
@@ -347,7 +353,7 @@ def cxx_link_into(
         dwp_artifact = dwp(
             ctx,
             cxx_toolchain_info,
-            final_output,
+            output,
             identifier = opts.identifier,
             category_suffix = opts.category_suffix,
             # TODO(T110378142): Ideally, referenced objects are a list of
@@ -358,10 +364,11 @@ def cxx_link_into(
             from_exe = not dwp_from_dwo,
         )
 
-    final_output = stamp_build_info(ctx, final_output) if is_result_executable else final_output
+    if is_result_executable:
+        output = stamp_build_info(ctx, output)
 
     linked_object = LinkedObject(
-        output = final_output,
+        output = output,
         link_args = opts.links,
         bitcode_bundle = bitcode_artifact.artifact if bitcode_artifact else None,
         prebolt_output = output,