wrong codegen


the current release_20 branch seems to miscompile the following reduced
testcase (not target architecture specific):

#define UInt16 unsigned short
#define UInt8 unsigned char
UInt8 foo(UInt16 a) {
  return (UInt8)(((a >> 10) & 1) << 1);

it - misleadingly - optimizes the expression to something like (undef &
2). I guess the problem is related to the DAG combiner but I haven't
done a thorough analysis yet.

a quick search didn't bring up existing bug reports. is this a known
problem? anybody else seeing this?

Verified, I filed this bug, I'm investigating. Thanks!