[llvm-branch-commits] [llvm-branch] r85750 - in /llvm/branches/Apple/Leela: include/llvm/Analysis/InlineCost.h include/llvm/Transforms/Utils/BasicInliner.h include/llvm/Transforms/Utils/InlineCost.h lib/Analysis/InlineCost.cpp lib/Transforms/IPO/InlineAlways.cpp lib/Transforms/IPO/InlineSimple.cpp lib/Transforms/IPO/Inliner.cpp lib/Transforms/Utils/InlineCost.cpp
Bill Wendling
isanbard at gmail.com
Sun Nov 1 14:12:57 PST 2009
Author: void
Date: Sun Nov 1 16:12:57 2009
New Revision: 85750
URL: http://llvm.org/viewvc/llvm-project?rev=85750&view=rev
Log:
$ svn merge -c 83997 https://llvm.org/svn/llvm-project/llvm/trunk
--- Merging r83997 into '.':
U include/llvm/Transforms/Utils/InlineCost.h
U lib/Transforms/Utils/InlineCost.cpp
$ svn merge -c 83998 https://llvm.org/svn/llvm-project/llvm/trunk
--- Merging r83998 into '.':
A include/llvm/Analysis/InlineCost.h
U include/llvm/Transforms/Utils/BasicInliner.h
D include/llvm/Transforms/Utils/InlineCost.h
A lib/Analysis/InlineCost.cpp
U lib/Transforms/IPO/InlineAlways.cpp
U lib/Transforms/IPO/Inliner.cpp
U lib/Transforms/IPO/InlineSimple.cpp
$ svn merge -c 83999 https://llvm.org/svn/llvm-project/llvm/trunk
--- Merging r83999 into '.':
D lib/Transforms/Utils/InlineCost.cpp
$ svn merge -c 84013 https://llvm.org/svn/llvm-project/llvm/trunk
--- Merging r84013 into '.':
U include/llvm/Analysis/InlineCost.h
U lib/Analysis/InlineCost.cpp
$ svn merge -c 84015 https://llvm.org/svn/llvm-project/llvm/trunk
--- Merging r84015 into '.':
G lib/Analysis/InlineCost.cpp
Added:
llvm/branches/Apple/Leela/include/llvm/Analysis/InlineCost.h
- copied, changed from r83998, llvm/trunk/include/llvm/Analysis/InlineCost.h
llvm/branches/Apple/Leela/lib/Analysis/InlineCost.cpp
- copied, changed from r83998, llvm/trunk/lib/Analysis/InlineCost.cpp
Removed:
llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/InlineCost.h
llvm/branches/Apple/Leela/lib/Transforms/Utils/InlineCost.cpp
Modified:
llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/BasicInliner.h
llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineAlways.cpp
llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineSimple.cpp
llvm/branches/Apple/Leela/lib/Transforms/IPO/Inliner.cpp
Copied: llvm/branches/Apple/Leela/include/llvm/Analysis/InlineCost.h (from r83998, llvm/trunk/include/llvm/Analysis/InlineCost.h)
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/include/llvm/Analysis/InlineCost.h?p2=llvm/branches/Apple/Leela/include/llvm/Analysis/InlineCost.h&p1=llvm/trunk/include/llvm/Analysis/InlineCost.h&r1=83998&r2=85750&rev=85750&view=diff
==============================================================================
--- llvm/trunk/include/llvm/Analysis/InlineCost.h (original)
+++ llvm/branches/Apple/Leela/include/llvm/Analysis/InlineCost.h Sun Nov 1 16:12:57 2009
@@ -28,6 +28,40 @@
template<class PtrType, unsigned SmallSize>
class SmallPtrSet;
+ // CodeMetrics - Calculate size and a few similar metrics for a set of
+ // basic blocks.
+ struct CodeMetrics {
+ /// NeverInline - True if this callee should never be inlined into a
+ /// caller.
+ bool NeverInline;
+
+ /// usesDynamicAlloca - True if this function calls alloca (in the C sense).
+ bool usesDynamicAlloca;
+
+ /// NumInsts, NumBlocks - Keep track of how large each function is, which
+ /// is used to estimate the code size cost of inlining it.
+ unsigned NumInsts, NumBlocks;
+
+ /// NumVectorInsts - Keep track of how many instructions produce vector
+ /// values. The inliner is being more aggressive with inlining vector
+ /// kernels.
+ unsigned NumVectorInsts;
+
+ /// NumRets - Keep track of how many Ret instructions the block contains.
+ unsigned NumRets;
+
+ CodeMetrics() : NeverInline(false), usesDynamicAlloca(false), NumInsts(0),
+ NumBlocks(0), NumVectorInsts(0), NumRets(0) {}
+
+ /// analyzeBasicBlock - Add information about the specified basic block
+ /// to the current structure.
+ void analyzeBasicBlock(const BasicBlock *BB);
+
+ /// analyzeFunction - Add information about the specified function
+ /// to the current structure.
+ void analyzeFunction(Function *F);
+ };
+
namespace InlineConstants {
// Various magic constants used to adjust heuristics.
const int CallPenalty = 5;
@@ -97,58 +131,32 @@
: ConstantWeight(CWeight), AllocaWeight(AWeight) {}
};
- // RegionInfo - Calculate size and a few related metrics for a set of
- // basic blocks.
- struct RegionInfo {
- /// NeverInline - True if this callee should never be inlined into a
- /// caller.
- bool NeverInline;
-
- /// usesDynamicAlloca - True if this function calls alloca (in the C sense).
- bool usesDynamicAlloca;
-
- /// NumInsts, NumBlocks - Keep track of how large each function is, which
- /// is used to estimate the code size cost of inlining it.
- unsigned NumInsts, NumBlocks;
-
- /// NumVectorInsts - Keep track of how many instructions produce vector
- /// values. The inliner is being more aggressive with inlining vector
- /// kernels.
- unsigned NumVectorInsts;
-
- /// NumRets - Keep track of how many Ret instructions the block contains.
- unsigned NumRets;
+ struct FunctionInfo {
+ CodeMetrics Metrics;
/// ArgumentWeights - Each formal argument of the function is inspected to
/// see if it is used in any contexts where making it a constant or alloca
/// would reduce the code size. If so, we add some value to the argument
/// entry here.
std::vector<ArgInfo> ArgumentWeights;
-
- RegionInfo() : NeverInline(false), usesDynamicAlloca(false), NumInsts(0),
- NumBlocks(0), NumVectorInsts(0), NumRets(0) {}
-
- /// analyzeBasicBlock - Add information about the specified basic block
- /// to the current structure.
- void analyzeBasicBlock(const BasicBlock *BB);
-
- /// analyzeFunction - Add information about the specified function
- /// to the current structure.
- void analyzeFunction(Function *F);
-
+
/// CountCodeReductionForConstant - Figure out an approximation for how
/// many instructions will be constant folded if the specified value is
/// constant.
unsigned CountCodeReductionForConstant(Value *V);
-
+
/// CountCodeReductionForAlloca - Figure out an approximation of how much
/// smaller the function will be if it is inlined into a context where an
/// argument becomes an alloca.
///
unsigned CountCodeReductionForAlloca(Value *V);
+
+ /// analyzeFunction - Add information about the specified function
+ /// to the current structure.
+ void analyzeFunction(Function *F);
};
- std::map<const Function *, RegionInfo> CachedFunctionInfo;
+ std::map<const Function *, FunctionInfo> CachedFunctionInfo;
public:
@@ -164,7 +172,7 @@
/// resetCachedFunctionInfo - erase any cached cost info for this function.
void resetCachedCostInfo(Function* Caller) {
- CachedFunctionInfo[Caller].NumBlocks = 0;
+ CachedFunctionInfo[Caller].Metrics.NumBlocks = 0;
}
};
}
Modified: llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/BasicInliner.h
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/BasicInliner.h?rev=85750&r1=85749&r2=85750&view=diff
==============================================================================
--- llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/BasicInliner.h (original)
+++ llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/BasicInliner.h Sun Nov 1 16:12:57 2009
@@ -15,7 +15,7 @@
#ifndef BASICINLINER_H
#define BASICINLINER_H
-#include "llvm/Transforms/Utils/InlineCost.h"
+#include "llvm/Analysis/InlineCost.h"
namespace llvm {
Removed: llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/InlineCost.h
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/InlineCost.h?rev=85749&view=auto
==============================================================================
--- llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/InlineCost.h (original)
+++ llvm/branches/Apple/Leela/include/llvm/Transforms/Utils/InlineCost.h (removed)
@@ -1,164 +0,0 @@
-//===- InlineCost.cpp - Cost analysis for inliner ---------------*- C++ -*-===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-//
-// This file implements heuristics for inlining decisions.
-//
-//===----------------------------------------------------------------------===//
-
-#ifndef LLVM_TRANSFORMS_UTILS_INLINECOST_H
-#define LLVM_TRANSFORMS_UTILS_INLINECOST_H
-
-#include <cassert>
-#include <climits>
-#include <map>
-#include <vector>
-
-namespace llvm {
-
- class Value;
- class Function;
- class CallSite;
- template<class PtrType, unsigned SmallSize>
- class SmallPtrSet;
-
- namespace InlineConstants {
- // Various magic constants used to adjust heuristics.
- const int CallPenalty = 5;
- const int LastCallToStaticBonus = -15000;
- const int ColdccPenalty = 2000;
- const int NoreturnPenalty = 10000;
- }
-
- /// InlineCost - Represent the cost of inlining a function. This
- /// supports special values for functions which should "always" or
- /// "never" be inlined. Otherwise, the cost represents a unitless
- /// amount; smaller values increase the likelyhood of the function
- /// being inlined.
- class InlineCost {
- enum Kind {
- Value,
- Always,
- Never
- };
-
- // This is a do-it-yourself implementation of
- // int Cost : 30;
- // unsigned Type : 2;
- // We used to use bitfields, but they were sometimes miscompiled (PR3822).
- enum { TYPE_BITS = 2 };
- enum { COST_BITS = unsigned(sizeof(unsigned)) * CHAR_BIT - TYPE_BITS };
- unsigned TypedCost; // int Cost : COST_BITS; unsigned Type : TYPE_BITS;
-
- Kind getType() const {
- return Kind(TypedCost >> COST_BITS);
- }
-
- int getCost() const {
- // Sign-extend the bottom COST_BITS bits.
- return (int(TypedCost << TYPE_BITS)) >> TYPE_BITS;
- }
-
- InlineCost(int C, int T) {
- TypedCost = (unsigned(C << TYPE_BITS) >> TYPE_BITS) | (T << COST_BITS);
- assert(getCost() == C && "Cost exceeds InlineCost precision");
- }
- public:
- static InlineCost get(int Cost) { return InlineCost(Cost, Value); }
- static InlineCost getAlways() { return InlineCost(0, Always); }
- static InlineCost getNever() { return InlineCost(0, Never); }
-
- bool isVariable() const { return getType() == Value; }
- bool isAlways() const { return getType() == Always; }
- bool isNever() const { return getType() == Never; }
-
- /// getValue() - Return a "variable" inline cost's amount. It is
- /// an error to call this on an "always" or "never" InlineCost.
- int getValue() const {
- assert(getType() == Value && "Invalid access of InlineCost");
- return getCost();
- }
- };
-
- /// InlineCostAnalyzer - Cost analyzer used by inliner.
- class InlineCostAnalyzer {
- struct ArgInfo {
- public:
- unsigned ConstantWeight;
- unsigned AllocaWeight;
-
- ArgInfo(unsigned CWeight, unsigned AWeight)
- : ConstantWeight(CWeight), AllocaWeight(AWeight) {}
- };
-
- // FunctionInfo - For each function, calculate the size of it in blocks and
- // instructions.
- struct FunctionInfo {
- /// NeverInline - True if this callee should never be inlined into a
- /// caller.
- bool NeverInline;
-
- /// usesDynamicAlloca - True if this function calls alloca (in the C sense).
- bool usesDynamicAlloca;
-
- /// NumInsts, NumBlocks - Keep track of how large each function is, which
- /// is used to estimate the code size cost of inlining it.
- unsigned NumInsts, NumBlocks;
-
- /// NumVectorInsts - Keep track of how many instructions produce vector
- /// values. The inliner is being more aggressive with inlining vector
- /// kernels.
- unsigned NumVectorInsts;
-
- /// ArgumentWeights - Each formal argument of the function is inspected to
- /// see if it is used in any contexts where making it a constant or alloca
- /// would reduce the code size. If so, we add some value to the argument
- /// entry here.
- std::vector<ArgInfo> ArgumentWeights;
-
- FunctionInfo() : NeverInline(false), usesDynamicAlloca(false), NumInsts(0),
- NumBlocks(0), NumVectorInsts(0) {}
-
- /// analyzeFunction - Fill in the current structure with information
- /// gleaned from the specified function.
- void analyzeFunction(Function *F);
-
- /// CountCodeReductionForConstant - Figure out an approximation for how
- /// many instructions will be constant folded if the specified value is
- /// constant.
- unsigned CountCodeReductionForConstant(Value *V);
-
- /// CountCodeReductionForAlloca - Figure out an approximation of how much
- /// smaller the function will be if it is inlined into a context where an
- /// argument becomes an alloca.
- ///
- unsigned CountCodeReductionForAlloca(Value *V);
- };
-
- std::map<const Function *, FunctionInfo> CachedFunctionInfo;
-
- public:
-
- /// getInlineCost - The heuristic used to determine if we should inline the
- /// function call or not.
- ///
- InlineCost getInlineCost(CallSite CS,
- SmallPtrSet<const Function *, 16> &NeverInline);
-
- /// getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
- /// higher threshold to determine if the function call should be inlined.
- float getInlineFudgeFactor(CallSite CS);
-
- /// resetCachedFunctionInfo - erase any cached cost info for this function.
- void resetCachedCostInfo(Function* Caller) {
- CachedFunctionInfo[Caller].NumBlocks = 0;
- }
- };
-}
-
-#endif
Copied: llvm/branches/Apple/Leela/lib/Analysis/InlineCost.cpp (from r83998, llvm/trunk/lib/Analysis/InlineCost.cpp)
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/lib/Analysis/InlineCost.cpp?p2=llvm/branches/Apple/Leela/lib/Analysis/InlineCost.cpp&p1=llvm/trunk/lib/Analysis/InlineCost.cpp&r1=83998&r2=85750&rev=85750&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/InlineCost.cpp (original)
+++ llvm/branches/Apple/Leela/lib/Analysis/InlineCost.cpp Sun Nov 1 16:12:57 2009
@@ -21,7 +21,7 @@
// CountCodeReductionForConstant - Figure out an approximation for how many
// instructions will be constant folded if the specified value is constant.
//
-unsigned InlineCostAnalyzer::RegionInfo::
+unsigned InlineCostAnalyzer::FunctionInfo::
CountCodeReductionForConstant(Value *V) {
unsigned Reduction = 0;
for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
@@ -77,7 +77,7 @@
// the function will be if it is inlined into a context where an argument
// becomes an alloca.
//
-unsigned InlineCostAnalyzer::RegionInfo::
+unsigned InlineCostAnalyzer::FunctionInfo::
CountCodeReductionForAlloca(Value *V) {
if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
unsigned Reduction = 0;
@@ -101,7 +101,7 @@
/// analyzeBasicBlock - Fill in the current structure with information gleaned
/// from the specified block.
-void InlineCostAnalyzer::RegionInfo::analyzeBasicBlock(const BasicBlock *BB) {
+void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
++NumBlocks;
for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
@@ -121,10 +121,8 @@
// probably won't do this in callers.
if (Function *F = CS.getCalledFunction())
if (F->isDeclaration() &&
- (F->getName() == "setjmp" || F->getName() == "_setjmp")) {
+ (F->getName() == "setjmp" || F->getName() == "_setjmp"))
NeverInline = true;
- return;
- }
// Calls often compile into many machine instructions. Bump up their
// cost to reflect this.
@@ -166,17 +164,22 @@
/// analyzeFunction - Fill in the current structure with information gleaned
/// from the specified function.
-void InlineCostAnalyzer::RegionInfo::analyzeFunction(Function *F) {
- // Look at the size of the callee. Each basic block counts as 20 units, and
- // each instruction counts as 5.
+void CodeMetrics::analyzeFunction(Function *F) {
+ // Look at the size of the callee.
for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
analyzeBasicBlock(&*BB);
+}
+
+/// analyzeFunction - Fill in the current structure with information gleaned
+/// from the specified function.
+void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
+ Metrics.analyzeFunction(F);
// A function with exactly one return has it removed during the inlining
// process (see InlineFunction), so don't count it.
- // FIXME: This knowledge should really be encoded outside of RegionInfo.
- if (NumRets==1)
- --NumInsts;
+ // FIXME: This knowledge should really be encoded outside of FunctionInfo.
+ if (Metrics.NumRets==1)
+ --Metrics.NumInsts;
// 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.
@@ -185,8 +188,6 @@
CountCodeReductionForAlloca(I)));
}
-
-
// getInlineCost - The heuristic used to determine if we should inline the
// function call or not.
//
@@ -229,35 +230,35 @@
InlineCost += InlineConstants::NoreturnPenalty;
// Get information about the callee...
- RegionInfo &CalleeFI = CachedFunctionInfo[Callee];
+ FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
// If we haven't calculated this information yet, do so now.
- if (CalleeFI.NumBlocks == 0)
+ if (CalleeFI.Metrics.NumBlocks == 0)
CalleeFI.analyzeFunction(Callee);
// If we should never inline this, return a huge cost.
- if (CalleeFI.NeverInline)
+ if (CalleeFI.Metrics.NeverInline)
return InlineCost::getNever();
// FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
- // could move this up and avoid computing the RegionInfo for
+ // 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) {
+ if (CalleeFI.Metrics.usesDynamicAlloca) {
// Get infomation about the caller...
- RegionInfo &CallerFI = CachedFunctionInfo[Caller];
+ FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
// If we haven't calculated this information yet, do so now.
- if (CallerFI.NumBlocks == 0)
+ if (CallerFI.Metrics.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)
+ if (!CallerFI.Metrics.usesDynamicAlloca)
return InlineCost::getNever();
}
@@ -305,7 +306,7 @@
InlineCost += Caller->size()/15;
// Look at the size of the callee. Each instruction counts as 5.
- InlineCost += CalleeFI.NumInsts*5;
+ InlineCost += CalleeFI.Metrics.NumInsts*5;
return llvm::InlineCost::get(InlineCost);
}
@@ -316,22 +317,22 @@
Function *Callee = CS.getCalledFunction();
// Get information about the callee...
- RegionInfo &CalleeFI = CachedFunctionInfo[Callee];
+ FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
// If we haven't calculated this information yet, do so now.
- if (CalleeFI.NumBlocks == 0)
+ if (CalleeFI.Metrics.NumBlocks == 0)
CalleeFI.analyzeFunction(Callee);
float Factor = 1.0f;
// Single BB functions are often written to be inlined.
- if (CalleeFI.NumBlocks == 1)
+ if (CalleeFI.Metrics.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)
+ if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
Factor += 2.0f;
- else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)
+ else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
Factor += 1.5f;
return Factor;
}
Modified: llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineAlways.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineAlways.cpp?rev=85750&r1=85749&r2=85750&view=diff
==============================================================================
--- llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineAlways.cpp (original)
+++ llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineAlways.cpp Sun Nov 1 16:12:57 2009
@@ -19,11 +19,11 @@
#include "llvm/Module.h"
#include "llvm/Type.h"
#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Analysis/InlineCost.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Transforms/IPO.h"
#include "llvm/Transforms/IPO/InlinerPass.h"
-#include "llvm/Transforms/Utils/InlineCost.h"
#include "llvm/ADT/SmallPtrSet.h"
using namespace llvm;
Modified: llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineSimple.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineSimple.cpp?rev=85750&r1=85749&r2=85750&view=diff
==============================================================================
--- llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineSimple.cpp (original)
+++ llvm/branches/Apple/Leela/lib/Transforms/IPO/InlineSimple.cpp Sun Nov 1 16:12:57 2009
@@ -18,11 +18,11 @@
#include "llvm/Module.h"
#include "llvm/Type.h"
#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Analysis/InlineCost.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Transforms/IPO.h"
#include "llvm/Transforms/IPO/InlinerPass.h"
-#include "llvm/Transforms/Utils/InlineCost.h"
#include "llvm/ADT/SmallPtrSet.h"
using namespace llvm;
Modified: llvm/branches/Apple/Leela/lib/Transforms/IPO/Inliner.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/lib/Transforms/IPO/Inliner.cpp?rev=85750&r1=85749&r2=85750&view=diff
==============================================================================
--- llvm/branches/Apple/Leela/lib/Transforms/IPO/Inliner.cpp (original)
+++ llvm/branches/Apple/Leela/lib/Transforms/IPO/Inliner.cpp Sun Nov 1 16:12:57 2009
@@ -18,10 +18,10 @@
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Analysis/InlineCost.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Transforms/IPO/InlinerPass.h"
-#include "llvm/Transforms/Utils/InlineCost.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
Removed: llvm/branches/Apple/Leela/lib/Transforms/Utils/InlineCost.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/Apple/Leela/lib/Transforms/Utils/InlineCost.cpp?rev=85749&view=auto
==============================================================================
--- llvm/branches/Apple/Leela/lib/Transforms/Utils/InlineCost.cpp (original)
+++ llvm/branches/Apple/Leela/lib/Transforms/Utils/InlineCost.cpp (removed)
@@ -1,337 +0,0 @@
-//===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-//
-// This file implements inline cost analysis.
-//
-//===----------------------------------------------------------------------===//
-
-#include "llvm/Transforms/Utils/InlineCost.h"
-#include "llvm/Support/CallSite.h"
-#include "llvm/CallingConv.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/ADT/SmallPtrSet.h"
-using namespace llvm;
-
-// CountCodeReductionForConstant - Figure out an approximation for how many
-// instructions will be constant folded if the specified value is constant.
-//
-unsigned InlineCostAnalyzer::FunctionInfo::
- CountCodeReductionForConstant(Value *V) {
- unsigned Reduction = 0;
- for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
- if (isa<BranchInst>(*UI))
- Reduction += 40; // Eliminating a conditional branch is a big win
- else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
- // Eliminating a switch is a big win, proportional to the number of edges
- // deleted.
- Reduction += (SI->getNumSuccessors()-1) * 40;
- else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
- // Turning an indirect call into a direct call is a BIG win
- Reduction += CI->getCalledValue() == V ? 500 : 0;
- } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
- // Turning an indirect call into a direct call is a BIG win
- Reduction += II->getCalledValue() == V ? 500 : 0;
- } else {
- // Figure out if this instruction will be removed due to simple constant
- // propagation.
- Instruction &Inst = cast<Instruction>(**UI);
-
- // We can't constant propagate instructions which have effects or
- // read memory.
- //
- // FIXME: It would be nice to capture the fact that a load from a
- // pointer-to-constant-global is actually a *really* good thing to zap.
- // Unfortunately, we don't know the pointer that may get propagated here,
- // so we can't make this decision.
- if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
- isa<AllocationInst>(Inst))
- continue;
-
- bool AllOperandsConstant = true;
- for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
- if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
- AllOperandsConstant = false;
- break;
- }
-
- if (AllOperandsConstant) {
- // We will get to remove this instruction...
- Reduction += 7;
-
- // And any other instructions that use it which become constants
- // themselves.
- Reduction += CountCodeReductionForConstant(&Inst);
- }
- }
-
- return Reduction;
-}
-
-// CountCodeReductionForAlloca - Figure out an approximation of how much smaller
-// the function will be if it is inlined into a context where an argument
-// becomes an alloca.
-//
-unsigned InlineCostAnalyzer::FunctionInfo::
- CountCodeReductionForAlloca(Value *V) {
- if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
- unsigned Reduction = 0;
- for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
- Instruction *I = cast<Instruction>(*UI);
- if (isa<LoadInst>(I) || isa<StoreInst>(I))
- Reduction += 10;
- else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
- // If the GEP has variable indices, we won't be able to do much with it.
- if (!GEP->hasAllConstantIndices())
- Reduction += CountCodeReductionForAlloca(GEP)+15;
- } else {
- // If there is some other strange instruction, we're not going to be able
- // to do much if we inline this.
- return 0;
- }
- }
-
- return Reduction;
-}
-
-/// analyzeFunction - Fill in the current structure with information gleaned
-/// from the specified function.
-void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
- unsigned NumInsts = 0, NumBlocks = 0, NumVectorInsts = 0, NumRets = 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<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->getName() == "setjmp" || F->getName() == "_setjmp")) {
- NeverInline = true;
- return;
- }
-
- // Calls often compile into many machine instructions. Bump up their
- // cost to reflect this.
- if (!isa<IntrinsicInst>(II))
- NumInsts += InlineConstants::CallPenalty;
- }
-
- // These, too, are calls.
- if (isa<MallocInst>(II) || isa<FreeInst>(II))
- NumInsts += InlineConstants::CallPenalty;
-
- 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)) {
- if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
- isa<PtrToIntInst>(CI))
- continue;
- } else if (const GetElementPtrInst *GEPI =
- dyn_cast<GetElementPtrInst>(II)) {
- // If a GEP has all constant indices, it will probably be folded with
- // a load/store.
- if (GEPI->hasAllConstantIndices())
- continue;
- }
-
- if (isa<ReturnInst>(II))
- ++NumRets;
-
- ++NumInsts;
- }
-
- ++NumBlocks;
- }
-
- // A function with exactly one return has it removed during the inlining
- // process (see InlineFunction), so don't count it.
- if (NumRets==1)
- --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.
- for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
- ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
- CountCodeReductionForAlloca(I)));
-}
-
-
-
-// getInlineCost - The heuristic used to determine if we should inline the
-// function call or not.
-//
-InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
- SmallPtrSet<const Function *, 16> &NeverInline) {
- Instruction *TheCall = CS.getInstruction();
- Function *Callee = CS.getCalledFunction();
- Function *Caller = TheCall->getParent()->getParent();
-
- // Don't inline functions which can be redefined at link-time to mean
- // something else. Don't inline functions marked noinline.
- if (Callee->mayBeOverridden() ||
- 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.
- //
- int InlineCost = 0;
-
- // If there is only one call of the function, and it has internal linkage,
- // make it almost guaranteed to be inlined.
- //
- if (Callee->hasLocalLinkage() && Callee->hasOneUse())
- InlineCost += InlineConstants::LastCallToStaticBonus;
-
- // If this function uses the coldcc calling convention, prefer not to inline
- // it.
- if (Callee->getCallingConv() == CallingConv::Cold)
- InlineCost += InlineConstants::ColdccPenalty;
-
- // If the instruction after the call, or if the normal destination of the
- // invoke is an unreachable instruction, the function is noreturn. As such,
- // there is little point in inlining this.
- if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
- if (isa<UnreachableInst>(II->getNormalDest()->begin()))
- InlineCost += InlineConstants::NoreturnPenalty;
- } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
- InlineCost += InlineConstants::NoreturnPenalty;
-
- // 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);
-
- // 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
- // passed into the function.
- //
- unsigned ArgNo = 0;
- for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
- I != E; ++I, ++ArgNo) {
- // Each argument passed in has a cost at both the caller and the callee
- // sides. This favors functions that take many arguments over functions
- // that take few arguments.
- InlineCost -= 20;
-
- // If this is a function being passed in, it is very likely that we will be
- // able to turn an indirect function call into a direct function call.
- if (isa<Function>(I))
- InlineCost -= 100;
-
- // If an alloca is passed in, inlining this function is likely to allow
- // significant future optimization possibilities (like scalar promotion, and
- // scalarization), so encourage the inlining of the function.
- //
- else if (isa<AllocaInst>(I)) {
- if (ArgNo < CalleeFI.ArgumentWeights.size())
- InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
-
- // If this is a constant being passed into the function, use the argument
- // weights calculated for the callee to determine how much will be folded
- // away with this information.
- } else if (isa<Constant>(I)) {
- if (ArgNo < CalleeFI.ArgumentWeights.size())
- InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
- }
- }
-
- // 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.
- // "size" here is the number of basic blocks, not instructions.
- //
- InlineCost += Caller->size()/15;
-
- // 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;
-}
More information about the llvm-branch-commits
mailing list