[PATCH] D99069: [ValueTracking] Improve phi handling in isKnownNonEqual()

[Nikita Popov via Phabricator via llvm-commits]

nikic added inline comments.


================
Comment at: llvm/lib/Analysis/ValueTracking.cpp:2459
       if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
         if (!C->isZero() && !C->isNegative()) {
+          BinaryOperator *BO = nullptr;
----------------
I think the matchSimpleRecurrence() check should be higher, e.g. `PN->getNumIncomingValues() == 2` is already part of it.


================
Comment at: llvm/lib/Analysis/ValueTracking.cpp:2461
+          BinaryOperator *BO = nullptr;
+          Value *R = nullptr, *L = nullptr;
+          if (Q.IIQ.UseInstrInfo && matchSimpleRecurrence(PN, BO, R, L)) {
----------------
I'd rename R/L to Start and Step.


================
Comment at: llvm/lib/Analysis/ValueTracking.cpp:2468
+                return true;
+              if (BO->getOpcode() == Instruction::Mul && !C->isOne() &&
+                  (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) &&
----------------
Why does this exclude C = 1? `{1,*,2}<nuw>` is still non-zero. The one value is important for inequality, but here we're checking for non-zero.


================
Comment at: llvm/lib/Analysis/ValueTracking.cpp:2471
+                  !StepCst->isNegative() && !StepCst->isZero() &&
+                  !StepCst->isOne())
+                return true;
----------------
Again, this isOne() check looks unnecessary to me.


================
Comment at: llvm/test/Analysis/ValueTracking/monotonic-phi.ll:61
+exit:
+  %add = or i8 %A, %r
+  %cmp = icmp eq i8 %add, 0
----------------
Why is this or needed?


================
Comment at: llvm/test/Analysis/ValueTracking/monotonic-phi.ll:79
+  %cmp = icmp eq i8 %add, 0
+  ; CHECK-NOT: ret i1 false
+  ret i1 %cmp
----------------
Why does this return false? Isn't this supposed to be a negative test case, due to lack of `nuw`/`nsw`?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99069/new/

https://reviews.llvm.org/D99069