401.bzip2 fails with -arch i386 -enable-iv-rewrite=false

[atrick]

Bugzilla Link	10920
Resolution	FIXED
Resolved on	Sep 15, 2011 17:35
Version	trunk
OS	All

Extended Description

Miscompilation caused by r139579: disable IV rewrite.

[atrick]

Temporarily fixed in r139583 by reverting r139579.

[atrick]

This bug isn't caused by any new indvars code. Rather, it was hidden by the old canonical IVs transform. In fact, I can easily reproduce without running indvars at all. Nonetheless, it's blocking my progress.

Two InstCombine transforms are causing problems:
visitICmpInstWithInstAndIntCst: LHS Add -> hasOneUse
visitICmpInstWithCastAndCast: if (CastInst *CI = dyn_cast(ICI.getOperand(1)))

but only when followed by jump threading...

[atrick]

My previous comment was incorrect because I was chasing interactions between passes. This is actually a register coalescing bug that's exposed by an early-cse transform. I'll fill in the details and hopefully provide a fix tomorrow.

[atrick]

We have a loop that looks like this:
a = 0;
while (a < t && g < s-1) {
g++;
a += array[g];
}

     if (g > x
         && expr1 && expr2 && expr3) {
        a -= array[g];
        g--;
     }
     call(a)

Early CSE removes the second load from array[g] and subsequent subtract. Now the value of 'a' when all conditions evaluate true is actually it's old loop pre increment value. Hence we need a copy on this path, or on all other paths to call(a).

With no IV rewrite, a's pre increment value is itself a copy of 'a' from the previous loop iteration. So now we have multiple copies of 'a'. But the pre increment copy is NOT the same value as the critical edge copies. RegistersDefinedFromSameValue does perform a reaching value check, but was performing it incorrectly and removing one of the copies.

In the full test case, phi elimination splits several critical edges at expr1 && expr2... and inserts copies from a. That's not so good, but will be fixed eventually and not the subject of this bug.

Fixed in r139765

[atrick]

The same bug was responsible for miscompilation of:
256.bzip -arch i386 -flto
sqlite3 -arch arm