[llvm-commits] CVS: llvm/lib/Transforms/Scalar/LowerInvoke.cpp

[Chris Lattner]

Changes in directory llvm/lib/Transforms/Scalar:

LowerInvoke.cpp updated: 1.30 -> 1.31
---
Log message:

add a note about a way to improve this code further, that I won't be getting
to right now.


---
Diffs of the changes:  (+8 -0)

 LowerInvoke.cpp |    8 ++++++++
 1 files changed, 8 insertions(+)


Index: llvm/lib/Transforms/Scalar/LowerInvoke.cpp
diff -u llvm/lib/Transforms/Scalar/LowerInvoke.cpp:1.30 llvm/lib/Transforms/Scalar/LowerInvoke.cpp:1.31
--- llvm/lib/Transforms/Scalar/LowerInvoke.cpp:1.30	Tue Sep 27 16:33:12 2005
+++ llvm/lib/Transforms/Scalar/LowerInvoke.cpp	Tue Sep 27 17:44:59 2005
@@ -412,6 +412,14 @@
 
   NumInvokes += Invokes.size();
   NumUnwinds += Unwinds.size();
+  
+  // TODO: This is not an optimal way to do this.  In particular, this always
+  // inserts setjmp calls into the entries of functions with invoke instructions
+  // even though there are possibly paths through the function that do not
+  // execute any invokes.  In particular, for functions with early exits, e.g.
+  // the 'addMove' method in hexxagon, it would be nice to not have to do the
+  // setjmp stuff on the early exit path.  This requires a bit of dataflow, but
+  // would not be too hard to do.
 
   // If we have an invoke instruction, insert a setjmp that dominates all
   // invokes.  After the setjmp, use a cond branch that goes to the original