Merge pull request #5908 from slowhei/master

Fix allocation of anonymous globals

2 files changed, 68 insertions, 0 deletions

diff --git a/src/llvm_backend.cpp b/src/llvm_backend.cpp
index d6bd7d72d..71bca42ab 100644
--- a/src/llvm_backend.cpp
+++ b/src/llvm_backend.cpp
@@ -2090,6 +2090,7 @@ gb_internal void lb_create_startup_runtime_generate_body(lbModule *m, lbProcedur
 			name = gb_string_append_length(name, ename.text, ename.len);
 
 			lbProcedure *dummy = lb_create_dummy_procedure(m, make_string_c(name), dummy_type);
+			dummy->is_startup = true;
 			LLVMSetVisibility(dummy->value, LLVMHiddenVisibility);
 			LLVMSetLinkage(dummy->value, LLVMWeakAnyLinkage);
 
diff --git a/tests/internal/test_anonymous_globals.odin b/tests/internal/test_anonymous_globals.odin
new file mode 100644
index 000000000..fc214ebde
--- /dev/null
+++ b/tests/internal/test_anonymous_globals.odin
@@ -0,0 +1,67 @@
+package test_internal
+
+import "core:testing"
+
+
+
+// https://github.com/odin-lang/Odin/pull/5908
+@(test)
+test_address_of_anonymous_global :: proc(t: ^testing.T) {
+	// This loop exists so that we do more computation with stack memory
+	// This increases the likelihood of catching a bug where anonymous globals are incorrectly allocated on the stack
+	// instead of the data segment
+	for _ in 0..<10 {
+		testing.expect(t, global_variable.inner.field == 0xDEADBEEF)
+	}
+}
+
+global_variable := Outer_Struct{
+	inner = &Inner_Struct{
+		field = 0xDEADBEEF,
+	},
+}
+Outer_Struct :: struct{
+	inner: ^Inner_Struct,
+
+	// Must have a second field to prevent the compiler from simplifying the `Outer_Struct` type to `^Inner_Struct`
+	// ...I think? In any case, don't remove this field
+	_: int,
+}
+Inner_Struct :: struct{
+	field: int,
+}
+
+
+
+// https://github.com/odin-lang/Odin/pull/5908
+//
+// Regression test for commit f1e3977cf94dfc0457f05d499cc280d8e1329086 where a larger anonymous global is needed to trigger
+// the bug
+@(test)
+test_address_of_large_anonymous_global :: proc(t: ^testing.T) {
+	// This loop exists so that we do more computation with stack memory
+	// This increases the likelihood of catching a bug where anonymous globals are incorrectly allocated on the stack
+	// instead of the data segment
+	for _ in 0..<10 {
+		for i in 0..<8 {
+			testing.expect(t, global_variable_64.inner.field[i] == i)
+		}
+	}
+}
+
+#assert(size_of(Inner_Struct_64) == 64)
+global_variable_64 := Outer_Struct_64{
+	inner = &Inner_Struct_64{
+		field = [8]int{0, 1, 2, 3, 4, 5, 6, 7},
+	},
+}
+Outer_Struct_64 :: struct{
+	inner: ^Inner_Struct_64,
+
+	// Must have a second field to prevent the compiler from simplifying the `Outer_Struct` type to `^Inner_Struct`
+	// ...I think? In any case, don't remove this field
+	_: int,
+}
+Inner_Struct_64 :: struct{
+	field: [8]int,
+}