Do not run loop-aligner at -fast (e.g. -O0).
[oota-llvm.git] / lib / CodeGen / LLVMTargetMachine.cpp
1 //===-- LLVMTargetMachine.cpp - Implement the LLVMTargetMachine class -----===//
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 the LLVMTargetMachine class.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Target/TargetMachine.h"
15 #include "llvm/PassManager.h"
16 #include "llvm/Pass.h"
17 #include "llvm/Assembly/PrintModulePass.h"
18 #include "llvm/Analysis/LoopPass.h"
19 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/CodeGen/Collector.h"
21 #include "llvm/Target/TargetOptions.h"
22 #include "llvm/Target/TargetAsmInfo.h"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Support/CommandLine.h"
25 using namespace llvm;
26
27 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden,
28     cl::desc("Print LLVM IR produced by the loop-reduce pass"));
29 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden,
30     cl::desc("Print LLVM IR input to isel pass"));
31 static cl::opt<bool> PrintEmittedAsm("print-emitted-asm", cl::Hidden,
32     cl::desc("Dump emitter generated instructions as assembly"));
33 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden,
34     cl::desc("Dump garbage collector data"));
35
36 // Hidden options to help debugging
37 static cl::opt<bool>
38 EnableSinking("enable-sinking", cl::init(false), cl::Hidden,
39               cl::desc("Perform sinking on machine code"));
40 static cl::opt<bool>
41 AlignLoops("align-loops", cl::init(true), cl::Hidden,
42            cl::desc("Align loop headers"));
43 static cl::opt<bool>
44 PerformLICM("machine-licm",
45             cl::init(false), cl::Hidden,
46             cl::desc("Perform loop-invariant code motion on machine code"));
47
48 // When this works it will be on by default.
49 static cl::opt<bool>
50 DisablePostRAScheduler("disable-post-RA-scheduler",
51                        cl::desc("Disable scheduling after register allocation"),
52                        cl::init(true));
53
54 FileModel::Model
55 LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
56                                        std::ostream &Out,
57                                        CodeGenFileType FileType,
58                                        bool Fast) {
59   // Standard LLVM-Level Passes.
60   
61   // Run loop strength reduction before anything else.
62   if (!Fast) {
63     PM.add(createLoopStrengthReducePass(getTargetLowering()));
64     if (PrintLSR)
65       PM.add(new PrintFunctionPass("\n\n*** Code after LSR ***\n", &cerr));
66   }
67   
68   PM.add(createGCLoweringPass());
69
70   if (!getTargetAsmInfo()->doesSupportExceptionHandling())
71     PM.add(createLowerInvokePass(getTargetLowering()));
72
73   // Make sure that no unreachable blocks are instruction selected.
74   PM.add(createUnreachableBlockEliminationPass());
75
76   if (!Fast)
77     PM.add(createCodeGenPreparePass(getTargetLowering()));
78
79   if (PrintISelInput)
80     PM.add(new PrintFunctionPass("\n\n*** Final LLVM Code input to ISel ***\n",
81                                  &cerr));
82   
83   // Ask the target for an isel.
84   if (addInstSelector(PM, Fast))
85     return FileModel::Error;
86
87   // Print the instruction selected machine code...
88   if (PrintMachineCode)
89     PM.add(createMachineFunctionPrinterPass(cerr));
90
91   if (PerformLICM)
92     PM.add(createMachineLICMPass());
93   
94   if (EnableSinking)
95     PM.add(createMachineSinkingPass());
96
97   // Run pre-ra passes.
98   if (addPreRegAlloc(PM, Fast) && PrintMachineCode)
99     PM.add(createMachineFunctionPrinterPass(cerr));
100
101   // Perform register allocation to convert to a concrete x86 representation
102   PM.add(createRegisterAllocator());
103   
104   if (PrintMachineCode)
105     PM.add(createMachineFunctionPrinterPass(cerr));
106     
107   PM.add(createLowerSubregsPass());
108   
109   if (PrintMachineCode)  // Print the subreg lowered code
110     PM.add(createMachineFunctionPrinterPass(cerr));
111
112   // Run post-ra passes.
113   if (addPostRegAlloc(PM, Fast) && PrintMachineCode)
114     PM.add(createMachineFunctionPrinterPass(cerr));
115
116   // Insert prolog/epilog code.  Eliminate abstract frame index references...
117   PM.add(createPrologEpilogCodeInserter());
118   
119   // Second pass scheduler.
120   if (!Fast && !DisablePostRAScheduler)
121     PM.add(createPostRAScheduler());
122
123   // Branch folding must be run after regalloc and prolog/epilog insertion.
124   if (!Fast)
125     PM.add(createBranchFoldingPass(getEnableTailMergeDefault()));
126
127   PM.add(createGCMachineCodeAnalysisPass());
128   if (PrintMachineCode)
129     PM.add(createMachineFunctionPrinterPass(cerr));
130   
131   if (PrintGCInfo)
132     PM.add(createCollectorMetadataPrinter(*cerr));
133   
134   // Fold redundant debug labels.
135   PM.add(createDebugLabelFoldingPass());
136   
137   if (PrintMachineCode)  // Print the register-allocated code
138     PM.add(createMachineFunctionPrinterPass(cerr));
139
140   if (addPreEmitPass(PM, Fast) && PrintMachineCode)
141     PM.add(createMachineFunctionPrinterPass(cerr));
142
143   if (AlignLoops && !Fast && !OptimizeForSize)
144     PM.add(createLoopAlignerPass());
145
146   switch (FileType) {
147   default:
148     break;
149   case TargetMachine::AssemblyFile:
150     if (addAssemblyEmitter(PM, Fast, Out))
151       return FileModel::Error;
152     return FileModel::AsmFile;
153   case TargetMachine::ObjectFile:
154     if (getMachOWriterInfo())
155       return FileModel::MachOFile;
156     else if (getELFWriterInfo())
157       return FileModel::ElfFile;
158   }
159
160   return FileModel::Error;
161 }
162  
163 /// addPassesToEmitFileFinish - If the passes to emit the specified file had to
164 /// be split up (e.g., to add an object writer pass), this method can be used to
165 /// finish up adding passes to emit the file, if necessary.
166 bool LLVMTargetMachine::addPassesToEmitFileFinish(PassManagerBase &PM,
167                                                   MachineCodeEmitter *MCE,
168                                                   bool Fast) {
169   if (MCE)
170     addSimpleCodeEmitter(PM, Fast, PrintEmittedAsm, *MCE);
171     
172   PM.add(createCollectorMetadataDeleter());
173
174   // Delete machine code for this function
175   PM.add(createMachineCodeDeleter());
176
177   return false; // success!
178 }
179
180 /// addPassesToEmitMachineCode - Add passes to the specified pass manager to
181 /// get machine code emitted.  This uses a MachineCodeEmitter object to handle
182 /// actually outputting the machine code and resolving things like the address
183 /// of functions.  This method should returns true if machine code emission is
184 /// not supported.
185 ///
186 bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
187                                                    MachineCodeEmitter &MCE,
188                                                    bool Fast) {
189   // Standard LLVM-Level Passes.
190   
191   // Run loop strength reduction before anything else.
192   if (!Fast) {
193     PM.add(createLoopStrengthReducePass(getTargetLowering()));
194     if (PrintLSR)
195       PM.add(new PrintFunctionPass("\n\n*** Code after LSR ***\n", &cerr));
196   }
197   
198   PM.add(createGCLoweringPass());
199   
200   if (!getTargetAsmInfo()->doesSupportExceptionHandling())
201     PM.add(createLowerInvokePass(getTargetLowering()));
202   
203   // Make sure that no unreachable blocks are instruction selected.
204   PM.add(createUnreachableBlockEliminationPass());
205
206   if (!Fast)
207     PM.add(createCodeGenPreparePass(getTargetLowering()));
208
209   if (PrintISelInput)
210     PM.add(new PrintFunctionPass("\n\n*** Final LLVM Code input to ISel ***\n",
211                                  &cerr));
212
213   // Ask the target for an isel.
214   if (addInstSelector(PM, Fast))
215     return true;
216
217   // Print the instruction selected machine code...
218   if (PrintMachineCode)
219     PM.add(createMachineFunctionPrinterPass(cerr));
220
221   if (PerformLICM)
222     PM.add(createMachineLICMPass());
223   
224   if (EnableSinking)
225     PM.add(createMachineSinkingPass());
226
227   // Run pre-ra passes.
228   if (addPreRegAlloc(PM, Fast) && PrintMachineCode)
229     PM.add(createMachineFunctionPrinterPass(cerr));
230
231   // Perform register allocation to convert to a concrete x86 representation
232   PM.add(createRegisterAllocator());
233   
234   if (PrintMachineCode)
235     PM.add(createMachineFunctionPrinterPass(cerr));
236     
237   PM.add(createLowerSubregsPass());
238   
239   if (PrintMachineCode)  // Print the subreg lowered code
240     PM.add(createMachineFunctionPrinterPass(cerr));
241
242   // Run post-ra passes.
243   if (addPostRegAlloc(PM, Fast) && PrintMachineCode)
244     PM.add(createMachineFunctionPrinterPass(cerr));
245
246   // Insert prolog/epilog code.  Eliminate abstract frame index references...
247   PM.add(createPrologEpilogCodeInserter());
248   
249   if (PrintMachineCode)  // Print the register-allocated code
250     PM.add(createMachineFunctionPrinterPass(cerr));
251   
252   // Second pass scheduler.
253   if (!Fast)
254     PM.add(createPostRAScheduler());
255
256   // Branch folding must be run after regalloc and prolog/epilog insertion.
257   if (!Fast)
258     PM.add(createBranchFoldingPass(getEnableTailMergeDefault()));
259
260   PM.add(createGCMachineCodeAnalysisPass());
261   if (PrintMachineCode)
262     PM.add(createMachineFunctionPrinterPass(cerr));
263   
264   if (PrintGCInfo)
265     PM.add(createCollectorMetadataPrinter(*cerr));
266   
267   if (addPreEmitPass(PM, Fast) && PrintMachineCode)
268     PM.add(createMachineFunctionPrinterPass(cerr));
269
270   addCodeEmitter(PM, Fast, PrintEmittedAsm, MCE);
271   
272   PM.add(createCollectorMetadataDeleter());
273   
274   // Delete machine code for this function
275   PM.add(createMachineCodeDeleter());
276   
277   return false; // success!
278 }