Add a matching call to llvm.va_end when llvm.va_start was called.

Resolves #1744

gen/functions.cpp

@@ -1011,6 +1011,17 @@ void DtoDefineFunction(FuncDeclaration *fd, bool linkageAvailableExternally) {
     // copy _arguments to a memory location
     irFunc->_arguments =
         DtoAllocaDump(irFunc->_arguments, 0, "_arguments_mem");
+
+    // Push cleanup block that calls va_end to match the va_start call.
+    {
+      auto *vaendBB =
+          llvm::BasicBlock::Create(gIR->context(), "vaend", gIR->topfunc());
+      IRScope saveScope = gIR->scope();
+      gIR->scope() = IRScope(vaendBB);
+      gIR->ir->CreateCall(GET_INTRINSIC_DECL(vaend), llAp);
+      funcGen.scopes.pushCleanup(vaendBB, gIR->scopebb());
+      gIR->scope() = saveScope;
+    }
   }
 
   funcGen.pgo.emitCounterIncrement(fd->fbody);
@@ -1019,6 +1030,17 @@ void DtoDefineFunction(FuncDeclaration *fd, bool linkageAvailableExternally) {
   // output function body
   Statement_toIR(fd->fbody, gIR);
 
+  // D varargs: emit the cleanup block that calls va_end.
+  if (f->linkage == LINKd && f->varargs == 1) {
+    if (!gIR->scopereturned()) {
+      if (!funcGen.retBlock)
+        funcGen.retBlock = gIR->insertBB("return");
+      funcGen.scopes.runCleanups(0, funcGen.retBlock);
+      gIR->scope() = IRScope(funcGen.retBlock);
+    }
+    funcGen.scopes.popCleanups(0);
+  }
+
   llvm::BasicBlock *bb = gIR->scopebb();
   if (pred_begin(bb) == pred_end(bb) &&
       bb != &bb->getParent()->getEntryBlock()) {

tests/codegen/vastart_vaend_gh1744.d

@@ -0,0 +1,50 @@
+// Test that va_end is called when returning from a variadic function.
+
+// Because the IR is kind of ugly, testing at -O0 is very brittle. Instead we
+// test that at -O3, LLVM was able to analyse the function correctly and
+// optimize-out the va_start and va_end calls and remove the call to
+// return_two (Github #1744).
+// The optimization (removing back-to-back calls to va_start and va_end) only
+// happens from LLVM >= 3.9.
+// REQUIRES: atleast_llvm309
+// RUN: %ldc %s -c -output-ll -O3 -of=%t.O3.ll \
+// RUN:   && FileCheck %s --check-prefix OPT3 < %t.O3.ll
+
+module mod;
+
+// OPT3-LABEL: define {{.*}} @{{.*}}void_three_return_paths
+void void_three_return_paths(int a, ...)
+{
+    // OPT3: call void @llvm.va_start({{.*}} %[[VA:[0-9]+]])
+    // OPT3-NOT: return_two
+    return_two();
+
+    if (a > 0)
+    {
+        throw new Exception("");
+        return;
+    }
+
+    // There are two control paths (normal return, exception resume) that
+    // should call va_end.
+    // OPT3: call void @llvm.va_end({{.*}} %[[VA]])
+    // OPT3: call void @llvm.va_end({{.*}} %[[VA]])
+}
+
+// OPT3-LABEL: define {{.*}} @{{.*}}return_two
+int return_two(...)
+{
+    // OPT3-NOT: va_start
+    // OPT3-NOT: va_end
+    // OPT3: ret i32 2
+    return 2;
+}
+
+// Github #1744:
+// OPT3-LABEL: define {{.*}} @{{.*}}two_four
+int two_four()
+{
+    // OPT3-NOT: return_two
+    // OPT3: ret i32 8
+    return return_two() * 4;
+}