-//===- InlineCoast.cpp - Cost analysis for inliner ------------------------===//
+//===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
//
// The LLVM Compiler Infrastructure
//
/// analyzeFunction - Fill in the current structure with information gleaned
/// from the specified function.
void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
- unsigned NumInsts = 0, NumBlocks = 0;
+ unsigned NumInsts = 0, NumBlocks = 0, NumVectorInsts = 0;
// Look at the size of the callee. Each basic block counts as 20 units, and
// each instruction counts as 5.
for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
II != E; ++II) {
- if (isa<DbgInfoIntrinsic>(II)) continue; // Debug intrinsics don't count.
+ if (isa<PHINode>(II)) continue; // PHI nodes don't count.
+
+ // Special handling for calls.
+ if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
+ if (isa<DbgInfoIntrinsic>(II))
+ continue; // Debug intrinsics don't count as size.
+
+ CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
+
+ // If this function contains a call to setjmp or _setjmp, never inline
+ // it. This is a hack because we depend on the user marking their local
+ // variables as volatile if they are live across a setjmp call, and they
+ // probably won't do this in callers.
+ if (Function *F = CS.getCalledFunction())
+ if (F->isDeclaration() &&
+ (F->isName("setjmp") || F->isName("_setjmp"))) {
+ NeverInline = true;
+ return;
+ }
+
+ // Calls often compile into many machine instructions. Bump up their
+ // cost to reflect this.
+ if (!isa<IntrinsicInst>(II))
+ NumInsts += 5;
+ }
+
+ if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
+ if (!AI->isStaticAlloca())
+ this->usesDynamicAlloca = true;
+ }
+
+ if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
+ ++NumVectorInsts;
// Noop casts, including ptr <-> int, don't count.
if (const CastInst *CI = dyn_cast<CastInst>(II)) {
isa<PtrToIntInst>(CI))
continue;
} else if (const GetElementPtrInst *GEPI =
- dyn_cast<GetElementPtrInst>(II)) {
+ dyn_cast<GetElementPtrInst>(II)) {
// If a GEP has all constant indices, it will probably be folded with
// a load/store.
bool AllConstant = true;
++NumBlocks;
}
- this->NumBlocks = NumBlocks;
- this->NumInsts = NumInsts;
+ this->NumBlocks = NumBlocks;
+ this->NumInsts = NumInsts;
+ this->NumVectorInsts = NumVectorInsts;
// Check out all of the arguments to the function, figuring out how much
// code can be eliminated if one of the arguments is a constant.
// getInlineCost - The heuristic used to determine if we should inline the
// function call or not.
//
-int InlineCostAnalyzer::getInlineCost(CallSite CS,
+InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
SmallPtrSet<const Function *, 16> &NeverInline) {
Instruction *TheCall = CS.getInstruction();
Function *Callee = CS.getCalledFunction();
- const Function *Caller = TheCall->getParent()->getParent();
-
- // Don't inline a directly recursive call.
- if (Caller == Callee ||
+ Function *Caller = TheCall->getParent()->getParent();
+
// Don't inline functions which can be redefined at link-time to mean
- // something else. link-once linkage is ok though.
- Callee->hasWeakLinkage() ||
-
- // Don't inline functions marked noinline.
- NeverInline.count(Callee))
- return 2000000000;
-
+ // something else.
+ if (Callee->mayBeOverridden() ||
+ // Don't inline functions marked noinline.
+ Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
+ return llvm::InlineCost::getNever();
+
// InlineCost - This value measures how good of an inline candidate this call
// site is to inline. A lower inline cost make is more likely for the call to
// be inlined. This value may go negative.
// If there is only one call of the function, and it has internal linkage,
// make it almost guaranteed to be inlined.
//
- if (Callee->hasInternalLinkage() && Callee->hasOneUse())
- InlineCost -= 30000;
+ if (Callee->hasLocalLinkage() && Callee->hasOneUse())
+ InlineCost -= 15000;
// If this function uses the coldcc calling convention, prefer not to inline
// it.
// If we haven't calculated this information yet, do so now.
if (CalleeFI.NumBlocks == 0)
CalleeFI.analyzeFunction(Callee);
+
+ // If we should never inline this, return a huge cost.
+ if (CalleeFI.NeverInline)
+ return InlineCost::getNever();
+
+ // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
+ // could move this up and avoid computing the FunctionInfo for
+ // things we are going to just return always inline for. This
+ // requires handling setjmp somewhere else, however.
+ if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
+ return InlineCost::getAlways();
+ if (CalleeFI.usesDynamicAlloca) {
+ // Get infomation about the caller...
+ FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
+
+ // If we haven't calculated this information yet, do so now.
+ if (CallerFI.NumBlocks == 0)
+ CallerFI.analyzeFunction(Caller);
+
+ // Don't inline a callee with dynamic alloca into a caller without them.
+ // Functions containing dynamic alloca's are inefficient in various ways;
+ // don't create more inefficiency.
+ if (!CallerFI.usesDynamicAlloca)
+ return InlineCost::getNever();
+ }
+
// Add to the inline quality for properties that make the call valuable to
// inline. This includes factors that indicate that the result of inlining
// the function will be optimizable. Currently this just looks at arguments
// Now that we have considered all of the factors that make the call site more
// likely to be inlined, look at factors that make us not want to inline it.
- // Don't inline into something too big, which would make it bigger. Here, we
- // count each basic block as a single unit.
+ // Don't inline into something too big, which would make it bigger.
//
- InlineCost += Caller->size()/20;
-
+ InlineCost += Caller->size()/15;
- // Look at the size of the callee. Each basic block counts as 20 units, and
- // each instruction counts as 5.
- InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;
- return InlineCost;
+ // Look at the size of the callee. Each instruction counts as 5.
+ InlineCost += CalleeFI.NumInsts*5;
+
+ return llvm::InlineCost::get(InlineCost);
}
+// getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
+// higher threshold to determine if the function call should be inlined.
+float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
+ Function *Callee = CS.getCalledFunction();
+
+ // Get information about the callee...
+ FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
+
+ // If we haven't calculated this information yet, do so now.
+ if (CalleeFI.NumBlocks == 0)
+ CalleeFI.analyzeFunction(Callee);
+
+ float Factor = 1.0f;
+ // Single BB functions are often written to be inlined.
+ if (CalleeFI.NumBlocks == 1)
+ Factor += 0.5f;
+
+ // Be more aggressive if the function contains a good chunk (if it mades up
+ // at least 10% of the instructions) of vector instructions.
+ if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/2)
+ Factor += 2.0f;
+ else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)
+ Factor += 1.5f;
+ return Factor;
+}