[mem2reg] Local alloca optimization incorrect for read-before-write case

[llvmbot]

Bugzilla Link	590
Resolution	FIXED
Resolved on	Feb 22, 2010 12:42
Version	1.4
OS	All
Reporter	LLVM Bugzilla Contributor

Extended Description

llvm-gcc miscompiles the following function:

void zero(char *p, int n) {
int i;
int out, undef;
for (i = 0; i < n+1; ++i) {
out = undef;
undef = 0;
if (i > 0)
p[i-1] = out;
}
}

This function zeros out n locations starting at p. (out is undefined on iteration 0 of the loop, but it is
not written there.) llvm-gcc miscompiles this function to ret void. I have narrowed the problem down
to the scalarrepl pass. When I compile with scalarrepl in, it fails, and with scalarrepl out it succeeds.

[llvmbot]

assigned to @lattner

[lattner]

Fixed. Testcase here: Transforms/Mem2Reg/2005-06-30-ReadBeforeWrite.ll

Patch here:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20050627/026866.html

Thanks!

-Chris