[llvm-commits] [llvm] r81928 - /llvm/trunk/lib/CodeGen/DwarfEHPrepare.cpp

Bill Wendling wendling at apple.com
Tue Sep 22 13:31:01 CDT 2009


On Sep 16, 2009, at 4:38 AM, Duncan Sands wrote:

> Hi Eric,
>
>>> not sure how this is possible, got an example?  In practice I only  
>>> ever
>>> saw abnormal landing pads due to critical edge splitting.
>>
>> With the change to using the unwind inst i'll create one every time  
>> you
>> inline something with an unwind afaict since it turns the unwind  
>> into a
>> branch to the invoke's landing pad.
>
> that makes sense - thanks for the explanation!
>
Hi Duncan,

Here's a testcase that makes this concrete:

$ cat testcase.ii
struct Foo {
   Foo();
};

template <class Ty>
class Bar {
   Ty &f;
public:
   Bar(Ty &x) : f(x) {}
};

template <class Ty>
class Baz {
   void *pointer;
   void *create(void *(*make)());
public:
   Ty &operator() () {
     create(make);
   }
   static void *make() {
     return new Ty;
   }
};

class Qux {
   char *f;
public:
   ~Qux() {
     if (f) delete f;
   }
};

class Zork {
   struct Bork : public Qux, public Foo {};
   static Baz<Bork> sessionMap;
public:
   static void m() {
     Bar<Foo> _(sessionMap());
   }
};

void Func1() {
   Zork::m();
}

$ llvm-g++ -O3 -S -dA testcase.ii -emit-llvm -o - 2>&1 | grep br |  
grep lpad
   br i1 %4, label %lpad, label %bb.i.i
   br label %lpad

(Most of this stuff comes from hash_map/vector code.)

-bw




More information about the llvm-commits mailing list