Loop Unrolling thwarts dead stores elimination of aggregate elements

[nlewycky]

Bugzilla Link	1310
Resolution	DUPLICATE
Resolved on	Feb 22, 2010 12:54
Version	trunk
OS	All

Extended Description

This program is a return undef:

int main(void) {
int foo[12][8];
foo[0][0] = 1;
for (int i = 0; i < 10; ++i) {
foo[i][5] = i;
}
foo[0][1] = 2;
foo[1][1] = 3;
foo[2][1] = 4;
foo[3][1] = 5;

return foo[4][4];
}

but LLVM has unrolled the loop and otherwise left in all of the stores and the
sole load.

[nlewycky]

The optimized code becomes "ret int 0" after running it through an additional
-scalarrepl pass. The problem seems to be that the early scalarrepl can't
analyze the loop structure. After the loop is unrolled, it works fine.

Adding a scalarrepl late in the compilation would fix this example, but only
because the loop was fully unrolled.

[lattner]

The link-time optimizer should clean this up, because it does run scalarrepl after loop unroll. However,
what we really need is a smarter DSE pass. I'm going to mark this as a dup of 1309.

*** This bug has been marked as a duplicate of 1309 ***