Patch to BitcodeReader

[Tobias Grosser]

On 09/12/2013 10:05 PM, Tom Roeder wrote:
> Performing the select as a regular instruction (%a = select ...) doesn't
> trigger the bug. It has to end up with a CE_SELECT in the CONSTANTS block
> of the bitcode file. However, I'm far from an expert in LLVM bitcode, so if
> there's a smaller test case, I'd be happy to switch to it.

Hi Tom,

just to explain why I even ask for this. Every unneeded code makes me 
start thinking for some reason it is needed to trigger the test case. So 
by removing any unneeded code, it is a lot more clear why something 
happened. As you have seen just know, this was even necessary to make me 
realize we only work on the constant selects (even though it was in fact 
clear in the patch);

What about this test case:

define <2 x i32> @main() {
  ret <2 x i32> select (<2 x i1> <i1 false, i1 undef>, <2 x i32> <i32 0, 
i32 0>, <2 x i32> <i32 0, i32 undef>)
}

Does this still trigger the bug?

Also it would be nice to add a small comment that this explicitly tests 
constant expressions.

Ah, and as said before, it would be nice to check that the very same IR 
is actually regenerated.

Cheers,
Tobias