[llvm] [ValueTracking] Clarify KnownBits recurrence code (PR #222266)

[Nikita Popov via llvm-commits]

================
@@ -1889,22 +1889,20 @@ static void computeKnownBitsFromOperator(const Operator *I,
         // D69571).
         SimplifyQuery RecQ = Q.getWithoutCondContext();
 
-        unsigned OpNum = P->getOperand(0) == R ? 0 : 1;
-        Instruction *RInst = P->getIncomingBlock(OpNum)->getTerminator();
-        Instruction *LInst = P->getIncomingBlock(1 - OpNum)->getTerminator();
+        unsigned OpNum = P->getOperand(0) == Start ? 0 : 1;
+        Instruction *StartTerm = P->getIncomingBlock(OpNum)->getTerminator();
 
-        // Ok, we have a PHI of the form L op= R. Check for low
+        // Ok, we have a recurrence of the form {Start,op,Step}. Check for low
         // zero bits.
-        RecQ.CxtI = RInst;
-        computeKnownBits(R, DemandedElts, Known2, RecQ, Depth + 1);
+        RecQ.CxtI = StartTerm;
+        computeKnownBits(Start, DemandedElts, Known2, RecQ, Depth + 1);
 
         // We need to take the minimum number of known bits
-        KnownBits Known3(BitWidth);
-        RecQ.CxtI = LInst;
-        computeKnownBits(L, DemandedElts, Known3, RecQ, Depth + 1);
+        KnownBits KnownStep(BitWidth);
+        computeKnownBits(Step, DemandedElts, KnownStep, Q, Depth + 1);
----------------
nikic wrote:

Generally we need to do this dropping of cond context + context instruction adjustment when we inspect phi argument. The step is not a phi argument.

You do make a good point about loop variance though. I think just keeping the original context may mean that we assume a condition on Step that only holds on loop exit. I'll see if I can come up with a test that gets miscompiled...

https://github.com/llvm/llvm-project/pull/222266