[Inlining] Teach shouldBeDeferred to take the total cost into account

[kazu]

Summary

This patch teaches shouldBeDeferred to take into account the total
cost of inlining.

Suppose we have a call hierarchy {A1,A2,A3,...}->B->C. (Each of A1,
A2, A3, ... calls B, which in turn calls C.)

Without this patch, shouldBeDeferred essentially returns true if

TotalSecondaryCost < IC.getCost()

where TotalSecondaryCost is the total cost of inlining B into As.
This means that if B is a small wraper function, for example, it would
get inlined into all of As. In turn, C gets inlined into all of As.
In other words, shouldBeDeferred ignores the cost of inlining C into
each of As.

This patch replaces the expression above with:

TotalCost < Allowance

where

• TotalCost is TotalSecondaryCost + IC.getCost() * # of As, and
• Allowance is IC.getCost() * Scale

For now, Scale defaults to 2, which essentially limits the number of
As to 1 for shouldBeDeferred to return true.

With this patch, Clang PGO bootstrap results in a 0.33% smaller .text*
sections. Compiling the 10 largest preprocessed files of Clang with
the PGO bootstrapped clang takes:

• 69.677 seconds on average of five runs without the patch, and
• 68.939 seconds on average of five runs with the patch.

[davidxl]

llvm/lib/Transforms/IPO/Inliner.cpp

346

NumCallerUsers may be more explicit.

[kazu]

I've renamed SecondaryUsers to NumCallerUsers.

[davidxl]

This is a conceptually a very good change. I think it should be split into two stages. The first is to make the changes but with default settings as NFC. The second stage is to reset the parameter with some benchmark number (e.g, spec perf should not regress, also code size impact).

One way to to the first stage is to not consider NumCaller adjustment if the deferral scale is a special value (such as -1).

[kazu]

I've updated the patch to turn off the new cost calculation by default.

[davidxl]

lgtm

llvm/lib/Transforms/IPO/Inliner.cpp

 /// prior to LLVM's code generator having support for stack coloring based on
 /// lifetime markers. It is now in the process of being removed. To experiment
 /// with disabling it and relying fully on lifetime marker based stack
 /// coloring, you can pass this flag to LLVM.
 static cl::opt<bool>
     DisableInlinedAllocaMerging("disable-inlined-alloca-merging",
                                 cl::init(false), cl::Hidden);
 
+static cl::opt<int>
+    InlineDeferralScale("inline-deferral-scale",
+                        cl::desc("Scale to limit the cost of inline deferral"),
+                        cl::init(2), cl::Hidden);
+
 namespace {
 
 enum class InlinerFunctionImportStatsOpts {
   No = 0,
   Basic = 1,
   Verbose = 2,
 };
 
shouldBeDeferred(Function *Caller, InlineCost IC, int &TotalSecondaryCost,
   TotalSecondaryCost = 0;
   // The candidate cost to be imposed upon the current function.
   int CandidateCost = IC.getCost() - 1;
   // If the caller has local linkage and can be inlined to all its callers, we
   // can apply a huge negative bonus to TotalSecondaryCost.
   bool ApplyLastCallBonus = Caller->hasLocalLinkage() && !Caller->hasOneUse();
   // This bool tracks what happens if we DO inline C into B.
   bool InliningPreventsSomeOuterInline = false;
+  unsigned NumCallerUsers = 0;
   for (User *U : Caller->users()) {
-    // If the caller will not be removed (either because it does not have a
-    // local linkage or because the LastCallToStaticBonus has been already
-    // applied), then we can exit the loop early.
-    if (!ApplyLastCallBonus && TotalSecondaryCost >= IC.getCost())
-      return false;
     CallBase *CS2 = dyn_cast<CallBase>(U);
 
     // If this isn't a call to Caller (it could be some other sort
     // of reference) skip it.  Such references will prevent the caller
     // from being removed.
     if (!CS2 || CS2->getCalledFunction() != Caller) {
       ApplyLastCallBonus = false;
       continue;
if (IC2.isAlways())
       continue;
 
     // See if inlining of the original callsite would erase the cost delta of
     // this callsite. We subtract off the penalty for the call instruction,
     // which we would be deleting.
     if (IC2.getCostDelta() <= CandidateCost) {
       InliningPreventsSomeOuterInline = true;
       TotalSecondaryCost += IC2.getCost();
+      NumCallerUsers++;
     }
   }
+
+  if (!InliningPreventsSomeOuterInline)
+    return false;
+
   // If all outer calls to Caller would get inlined, the cost for the last
   // one is set very low by getInlineCost, in anticipation that Caller will
   // be removed entirely.  We did not account for this above unless there
   // is only one caller of Caller.
   if (ApplyLastCallBonus)
     TotalSecondaryCost -= InlineConstants::LastCallToStaticBonus;
 
-  return InliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost();
+  int TotalCost = TotalSecondaryCost + IC.getCost() * NumCallerUsers;
+  int Allowance = IC.getCost() * InlineDeferralScale;
+  return TotalCost < Allowance;
 }
 
 static std::basic_ostream<char> &operator<<(std::basic_ostream<char> &R,
                                             const ore::NV &Arg) {
   return R << Arg.Val;
 }
 
 template <class RemarkT>