Builtin options not forwarded correctly to subprojects

[bonzini]

Describe the bug
The simplest reproducer is:

diff --git a/test cases/common/281 subproj options/subprojects/sub/meson.build b/test cases/common/281 subproj options/subprojects/sub/meson.build
index 9e3bceacf..b1fb9558a 100644
--- a/test cases/common/281 subproj options/subprojects/sub/meson.build	
+++ b/test cases/common/281 subproj options/subprojects/sub/meson.build	
@@ -1,2 +1,3 @@
-project('subproject')
+project('subproject', 'c')
 assert(get_option('bar') == true)
+assert(get_option('b_lto') == true)
diff --git a/test cases/common/281 subproj options/test.json b/test cases/common/281 subproj options/test.json
new file mode 100644
index 000000000..fcdf4ddd8
--- /dev/null
+++ b/test cases/common/281 subproj options/test.json	
@@ -0,0 +1,7 @@
+{
+  "matrix": {
+    "options": {
+      "b_lto":      [{ "val": "true" }]
+    }
+  }
+}

That is, a superproject is configured with -Db_lto=true but the subproject sees b_lto as false. This breaks QEMU's CFI builds (where the clang flags for CFI are added with add_global_arguments) but is a regression in general.

system parameters

• Is this a cross build or just a plain native build (for the same computer)? no
• what operating system (e.g. MacOS Catalina, Windows 10, CentOS 8.0, Ubuntu 18.04, etc.) Fedora 42
• what Python version are you using e.g. 3.13.2
• what meson --version 1.7.99 commit 02792fa
• what ninja --version if it's a Ninja build N/A

[bonzini]

Some more notes... If (in addition to the above diff) test cases/common/281 subproj options/meson.build is changed to specify project('pkg_opt_test', 'c'), it also doesn't work, but I have the following patch (though I'm not sure if I know what I'm doing):

diff --git a/mesonbuild/coredata.py b/mesonbuild/coredata.py
index 36120ff7e..562c3d3cb 100644
--- a/mesonbuild/coredata.py
+++ b/mesonbuild/coredata.py
@@ -598,6 +598,8 @@ class CoreData:
     def process_compiler_options(self, lang: str, comp: Compiler, env: Environment, subproject: str) -> None:
         self.add_compiler_options(comp.get_options(), lang, comp.for_machine, env, subproject)
 
+        # Separate bug: does not take into account pending_options?
+        print('---', self.optstore.pending_options)
         for key in comp.base_options:
             if subproject:
                 skey = key.evolve(subproject=subproject)
@@ -605,17 +607,15 @@ class CoreData:
                 skey = key
             if skey not in self.optstore:
                 self.optstore.add_system_option(skey, copy.deepcopy(options.COMPILER_BASE_OPTIONS[key]))
-                if skey in env.options:
-                    self.optstore.set_option(skey, env.options[skey])
-                elif subproject and key in env.options:
-                    self.optstore.set_option(skey, env.options[key])
-                # FIXME
-                #if subproject and not self.optstore.has_option(key):
-                #    self.optstore[key] = copy.deepcopy(self.optstore[skey])
-            elif skey in env.options:
+
+            if skey in env.options:
                 self.optstore.set_option(skey, env.options[skey])
             elif subproject and key in env.options:
+                # ??? is this needed at all ???
                 self.optstore.set_option(skey, env.options[key])
+            elif subproject and key in self.optstore.options:
+                self.optstore.set_option(skey, self.optstore.options[key].value)
+
         self.emit_base_options_warnings()
 
     def emit_base_options_warnings(self) -> None:

In the testcase at the top of the issue, OptionStore.pending_options is also involved because the superproject does not enable a language that processes b_lto.

I am not sure if these are two completely separate bugs, or just one on top of the other.

[bonzini]

The remaining issue is that coredata should IMO not be using env.options at all. project() passes its contents are passed to the option store via

self.coredata.optstore.initialize_from_top_level_project_call(self.project_default_options,
                                                              self.user_defined_options.cmd_line_options,
                                                              self.environment.options)

and after this call that should be it. Therefore add_compiler_options()/process_compiler_options() should be entirely based on the OptionStore's pending_options.

[dcbaker]

I kinda suspected this was a problem and really wanted more time to continue reworking the option refactor stuff before we made the 1.8 branch, but I've been busy with the work I'm actually paid for.

[Totto16]

I kinda suspected this was a problem and really wanted more time to continue reworking the option refactor stuff before we made the 1.8 branch, but I've been busy with the work I'm actually paid for.

For that reason, it's only rc1 and not 1.8.0, so no worries there 😄

release candidates are there for that exact reason, find such bugs