Fix Windows LTO: preserve required procedures with llvm.used

On Windows with LTO, required procedures with external linkage need to
be added to @llvm.used to survive linker-level dead code elimination.

LLVM may generate implicit calls to runtime builtins (e.g., __extendhfsf2
for f16 conversions) during instruction lowering, after the IR is
finalized. Without @llvm.used, the linker discards these procedures
before the implicit calls are generated.

This adds required procedures to @llvm.used at creation time. The fix
is Windows-specific; other platforms handle this correctly.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

2 files changed, 33 insertions, 9 deletions

diff --git a/src/llvm_backend_opt.cpp b/src/llvm_backend_opt.cpp
index 4131f32bf..cb7fe1c75 100644
--- a/src/llvm_backend_opt.cpp
+++ b/src/llvm_backend_opt.cpp
@@ -512,8 +512,9 @@ gb_internal void llvm_delete_function(LLVMValueRef func) {
 	LLVMDeleteFunction(func);
 }
 
-gb_internal void lb_append_to_compiler_used(lbModule *m, LLVMValueRef value) {
-	LLVMValueRef global = LLVMGetNamedGlobal(m->mod, "llvm.compiler.used");
+// Helper to append a value to an llvm metadata array global (llvm.used or llvm.compiler.used)
+gb_internal void lb_append_to_llvm_used_list(lbModule *m, LLVMValueRef value, char const *list_name) {
+	LLVMValueRef global = LLVMGetNamedGlobal(m->mod, list_name);
 
 	LLVMValueRef *constants;
 	int operands = 1;
@@ -543,33 +544,43 @@ gb_internal void lb_append_to_compiler_used(lbModule *m, LLVMValueRef value) {
 	constants[operands - 1] = LLVMConstBitCast(value, Int8PtrTy);
 	LLVMValueRef initializer = LLVMConstArray(Int8PtrTy, constants, operands);
 
-	global = LLVMAddGlobal(m->mod, ATy, "llvm.compiler.used");
+	global = LLVMAddGlobal(m->mod, ATy, list_name);
 	LLVMSetLinkage(global, LLVMAppendingLinkage);
 	LLVMSetSection(global, "llvm.metadata");
 	LLVMSetInitializer(global, initializer);
 }
 
+gb_internal void lb_append_to_compiler_used(lbModule *m, LLVMValueRef value) {
+	lb_append_to_llvm_used_list(m, value, "llvm.compiler.used");
+}
+
+// llvm.used survives LTO linker optimizations (unlike llvm.compiler.used)
+gb_internal void lb_append_to_used(lbModule *m, LLVMValueRef value) {
+	lb_append_to_llvm_used_list(m, value, "llvm.used");
+}
+
 gb_internal void lb_run_remove_unused_function_pass(lbModule *m) {
 	isize removal_count = 0;
 	isize pass_count = 0;
 	isize const max_pass_count = 10;
-	// Custom remove dead function pass
+
+	// Custom remove dead function pass (for internal linkage functions)
 	for (; pass_count < max_pass_count; pass_count++) {
-		bool was_dead = false;	
+		bool was_dead = false;
 		for (LLVMValueRef func = LLVMGetFirstFunction(m->mod);
 		     func != nullptr;
 		     /**/
 		     ) {
 		     	LLVMValueRef curr_func = func;
 		     	func = LLVMGetNextFunction(func);
-		     	
+
 			LLVMUseRef first_use = LLVMGetFirstUse(curr_func);
 			if (first_use != nullptr)  {
 				continue;
 			}
 			String name = {};
 			name.text = cast(u8 *)LLVMGetValueName2(curr_func, cast(size_t *)&name.len);
-						
+
 			if (LLVMIsDeclaration(curr_func)) {
 				// Ignore for the time being
 				continue;
@@ -578,7 +589,7 @@ gb_internal void lb_run_remove_unused_function_pass(lbModule *m) {
 			if (linkage != LLVMInternalLinkage) {
 				continue;
 			}
-			
+
 			Entity **found = map_get(&m->procedure_values, curr_func);
 			if (found && *found) {
 				Entity *e = *found;
@@ -588,7 +599,7 @@ gb_internal void lb_run_remove_unused_function_pass(lbModule *m) {
 					continue;
 				}
 			}
-			
+
 			llvm_delete_function(curr_func);
 			was_dead = true;
 			removal_count += 1;
diff --git a/src/llvm_backend_proc.cpp b/src/llvm_backend_proc.cpp
index 837d7ce48..cb8ffcf91 100644
--- a/src/llvm_backend_proc.cpp
+++ b/src/llvm_backend_proc.cpp
@@ -289,6 +289,19 @@ gb_internal lbProcedure *lb_create_procedure(lbModule *m, Entity *entity, bool i
 
 	lb_set_linkage_from_entity_flags(p->module, p->value, entity->flags);
 
+	// With LTO on Windows, required procedures with external linkage need to be added to
+	// llvm.used to survive linker-level dead code elimination. This is necessary because
+	// LLVM may generate implicit calls to runtime builtins (e.g., __extendhfsf2 for f16
+	// conversions) during instruction lowering, after the IR is finalized.
+	if (build_context.lto_kind != LTO_None && build_context.metrics.os == TargetOs_windows) {
+		if (entity->flags & EntityFlag_Require) {
+			LLVMLinkage linkage = LLVMGetLinkage(p->value);
+			if (linkage != LLVMInternalLinkage) {
+				lb_append_to_used(m, p->value);
+			}
+		}
+	}
+
 	if (m->debug_builder) { // Debug Information
 		Type *bt = base_type(p->type);