[InstructionSimplify] Limit threadCmpOverPHI recursion depth to 1

[0xdc03]

Summary

Other places in the compiler (such as ValueTracking) only allow
recursing one level into phi nodes, however in InstructionSimplify the
limit is 3.

By reducing it to 1, there is a slight reduction in compile times (0.1%):
https://llvm-compile-time-tracker.com/compare.php?from=47a4331cd79ff63b13e369e28ccca268773cf9f2&to=20d987ab937caf22929f0c5790f0a8197b53b447&stat=instructions:u

Depends on D155718.

[0xdc03]

• Invert the patch stack

[0xdc03]

• Change the patch order

[0xdc03]

• Fix some broken logic

[0xdc03]

This patch is causing a regression that is proving too hard to track down, and for a 0.1% improvement its not really worth it.

llvm/lib/Analysis/InstructionSimplify.cpp

 /// yields the same result every time. If so returns the common result,
 /// otherwise returns null.
 static Value *threadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
                                const SimplifyQuery &Q, unsigned MaxRecurse) {
   // Recursion is always used, so bail out at once if we already hit the limit.
   if (!MaxRecurse--)
     return nullptr;
 
+  // Prevent this function from being recursively called more than once.
+  MaxRecurse = 0;
+
   // Make sure the phi is on the LHS.
   if (!isa<PHINode>(LHS)) {
     std::swap(LHS, RHS);
     Pred = CmpInst::getSwappedPredicate(Pred);
   }
   assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
   PHINode *PI = cast<PHINode>(LHS);