SILOptimizer: fix a compile time problem in the inliner

When inlining many functions in a very large basic block, the splitting of the block at the call sites is quadratic, when traversing in forward order.
Traversing backwards, fixes the problem.

rdar://problem/56268570

Author: eeckstein

lib/SILOptimizer/Transforms/PerformanceInliner.cpp

@@ -915,7 +915,9 @@ bool SILPerformanceInliner::inlineCallsIntoFunction(SILFunction *Caller) {
     return false;
 
   // Second step: do the actual inlining.
-  for (auto AI : AppliesToInline) {
+  // We inline in reverse order, because for very large blocks with many applies
+  // to inline, splitting the block at every apply would be quadratic.
+  for (auto AI : llvm::reverse(AppliesToInline)) {
     SILFunction *Callee = AI.getReferencedFunctionOrNull();
     assert(Callee && "apply_inst does not have a direct callee anymore");