5acd6aa5e1276befef13c4bb42539404125be963
[oota-llvm.git] / lib / Transforms / Scalar / SimplifyHalfPowrLibCalls.cpp
1 //===- SimplifyHalfPowrLibCalls.cpp - Optimize specific half_powr calls ---===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements a simple pass that applies an experimental
11 // transformation on calls to specific functions.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #define DEBUG_TYPE "simplify-libcalls-halfpowr"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/Intrinsics.h"
19 #include "llvm/Module.h"
20 #include "llvm/Pass.h"
21 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
22 #include "llvm/Transforms/Utils/Cloning.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Support/Debug.h"
26 using namespace llvm;
27
28 namespace {
29   /// This pass optimizes well half_powr function calls.
30   ///
31   class SimplifyHalfPowrLibCalls : public FunctionPass {
32     const TargetData *TD;
33   public:
34     static char ID; // Pass identification
35     SimplifyHalfPowrLibCalls() : FunctionPass(&ID) {}
36
37     bool runOnFunction(Function &F);
38
39     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
40     }
41
42     Instruction *
43     InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs,
44                     Instruction *InsertPt);
45   };
46   char SimplifyHalfPowrLibCalls::ID = 0;
47 } // end anonymous namespace.
48
49 static RegisterPass<SimplifyHalfPowrLibCalls>
50 X("simplify-libcalls-halfpowr", "Simplify half_powr library calls");
51
52 // Public interface to the Simplify HalfPowr LibCalls pass.
53 FunctionPass *llvm::createSimplifyHalfPowrLibCallsPass() {
54   return new SimplifyHalfPowrLibCalls(); 
55 }
56
57 /// InlineHalfPowrs - Inline a sequence of adjacent half_powr calls, rearranging
58 /// their control flow to better facilitate subsequent optimization.
59 Instruction *
60 SimplifyHalfPowrLibCalls::
61 InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs,
62                 Instruction *InsertPt) {
63   std::vector<BasicBlock *> Bodies;
64   BasicBlock *NewBlock = 0;
65
66   for (unsigned i = 0, e = HalfPowrs.size(); i != e; ++i) {
67     CallInst *Call = cast<CallInst>(HalfPowrs[i]);
68     Function *Callee = Call->getCalledFunction();
69
70     // Minimally sanity-check the CFG of half_powr to ensure that it contains
71     // the the kind of code we expect.  If we're running this pass, we have
72     // reason to believe it will be what we expect.
73     Function::iterator I = Callee->begin();
74     BasicBlock *Prologue = I++;
75     if (I == Callee->end()) break;
76     BasicBlock *SubnormalHandling = I++;
77     if (I == Callee->end()) break;
78     BasicBlock *Body = I++;
79     if (I != Callee->end()) break;
80     if (SubnormalHandling->getSinglePredecessor() != Prologue)
81       break;
82     BranchInst *PBI = dyn_cast<BranchInst>(Prologue->getTerminator());
83     if (!PBI || !PBI->isConditional())
84       break;
85     BranchInst *SNBI = dyn_cast<BranchInst>(SubnormalHandling->getTerminator());
86     if (!SNBI || SNBI->isConditional())
87       break;
88     if (!isa<ReturnInst>(Body->getTerminator()))
89       break;
90
91     Instruction *NextInst = llvm::next(BasicBlock::iterator(Call));
92
93     // Inline the call, taking care of what code ends up where.
94     NewBlock = SplitBlock(NextInst->getParent(), NextInst, this);
95
96     bool B = InlineFunction(Call, 0, TD);
97     assert(B && "half_powr didn't inline?"); B=B;
98
99     BasicBlock *NewBody = NewBlock->getSinglePredecessor();
100     assert(NewBody);
101     Bodies.push_back(NewBody);
102   }
103
104   if (!NewBlock)
105     return InsertPt;
106
107   // Put the code for all the bodies into one block, to facilitate
108   // subsequent optimization.
109   (void)SplitEdge(NewBlock->getSinglePredecessor(), NewBlock, this);
110   for (unsigned i = 0, e = Bodies.size(); i != e; ++i) {
111     BasicBlock *Body = Bodies[i];
112     Instruction *FNP = Body->getFirstNonPHI();
113     // Splice the insts from body into NewBlock.
114     NewBlock->getInstList().splice(NewBlock->begin(), Body->getInstList(),
115                                    FNP, Body->getTerminator());
116   }
117
118   return NewBlock->begin();
119 }
120
121 /// runOnFunction - Top level algorithm.
122 ///
123 bool SimplifyHalfPowrLibCalls::runOnFunction(Function &F) {
124   TD = getAnalysisIfAvailable<TargetData>();
125   
126   bool Changed = false;
127   std::vector<Instruction *> HalfPowrs;
128   for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
129     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
130       // Look for calls.
131       bool IsHalfPowr = false;
132       if (CallInst *CI = dyn_cast<CallInst>(I)) {
133         // Look for direct calls and calls to non-external functions.
134         Function *Callee = CI->getCalledFunction();
135         if (Callee && Callee->hasExternalLinkage()) {
136           // Look for calls with well-known names.
137           if (Callee->getName() == "__half_powrf4")
138             IsHalfPowr = true;
139         }
140       }
141       if (IsHalfPowr)
142         HalfPowrs.push_back(I);
143       // We're looking for sequences of up to three such calls, which we'll
144       // simplify as a group.
145       if ((!IsHalfPowr && !HalfPowrs.empty()) || HalfPowrs.size() == 3) {
146         I = InlineHalfPowrs(HalfPowrs, I);
147         E = I->getParent()->end();
148         HalfPowrs.clear();
149         Changed = true;
150       }
151     }
152     assert(HalfPowrs.empty() && "Block had no terminator!");
153   }
154
155   return Changed;
156 }