On linux/i386 and using CWriter, the program attached below gets translated into
wrong source code and after compilation with gcc the

  assertion failed in line 16: 'u.l == 128'

triggers.

Created an attachment (id=193) [details]
Original program.

Created an attachment (id=194) [details]
Output from CWriter which triggers the bug.

Update: this might be some subtle aliasing problem, and not all combinations of
gcc versions and options seem to trigger it. Here are some that do:

Stock GNU gcc-2.95.3
  gcc-2.95.3 -O0 -fno-strict-aliasing bug491_llvm.c 

gcc version 3.4.3 (Mandrakelinux 10.2 3.4.3-1mdk)
  gcc -O2 -fno-strict-aliasing bug491_llvm.c

Here is a reduced testcase.  Run it though DSE and the first store is deleted,
as DSE thinks that the second store kills it.  This is clearly incorrect.

int %test() {
        %V = alloca int
        store int 1234567, int* %V
        %V2 = cast int* %V to sbyte*
        store sbyte 0, sbyte* %V2
        %X = load int* %V 
        ret int %X
}

Fixed.  Testcase here:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20041227/022738.html

Patch here:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20041227/022740.html

Thanks a lot for the bug report!!

-Chris

Update: this program (attachment "Original program") fails with CVS 20051027
when using the i386 code generator.

Nice catch.  Fixed, patch here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20051024/029098.html

-Chris

Note: I checked the testcase into llvm-test so we don't regress on this again.

-Chris