Comparison of constant inttoptrs not folded

[efriedma-quic]

Bugzilla Link	2329
Resolution	FIXED
Resolved on	May 25, 2008 15:56
Version	unspecified
OS	Linux
CC	@nlewycky

Extended Description

From GCC Bug 17141:

void abort (void);
struct a
{
int i;
} *a;
int f(void)
{
int *ii = &a->i;
void *l;
a->i = 1;
if (*ii)
l = &&l1;
else
l = &&l2;
goto *l;
l1:
return 0;
l2:
abort ();
}

This should fold to almost nothing, but we end up with a branch on a complicated constant expression:
br i1 icmp eq (i8* inttoptr (i32 1 to i8*), i8* inttoptr (i32 2 to i8*)), label %l2, label %l1

"icmp eq (i8* inttoptr (i32 1 to i8*), i8* inttoptr (i32 2 to i8*))" should fold to false.

[efriedma-quic]

assigned to @nlewycky

[nlewycky]

In general, icmp eq/ne Const, Const can always be folded away.

[nlewycky]

Mmm... maybe that was overzealous of me. Is it really legal to say "icmp eq ConstantInt, ptrtoint(some function)"? Or can we always collapse those into false?

I was thinking that the structural equivalence of Constants that made them pointer-comparable for equality would be sufficient to test for eq/ne, but there's a whole lot of code in ConstantFold::evaluateICmpRelation which makes me think there's some reason that isn't the case.

[nlewycky]

It now optimizes down to:

define i32 @​f() nounwind {
entry:
%tmp1 = load %struct.a** @​a, align 4 ; <%struct.a*> [#uses=1]
%tmp2 = getelementptr %struct.a* %tmp1, i32 0, i32 0 ; <i32*> [#uses=1]
store i32 1, i32* %tmp2, align 4
ret i32 0
}

[nlewycky]

proposed patch
Attached patch fixes the bug, but due to unrelated issues, I can't check whether it breaks other things in llvm-test.

I'll check it in either once I verify that it works, or someone else confirms that it doesn't break llvm-test/MultiSource.

[nlewycky]

My patch (and the testcase it contains) are incorrect.

[nlewycky]

Patch checked in.

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080519/062906.html