1 //===-- NVPTXTargetMachine.cpp - Define TargetMachine for NVPTX -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Top-level implementation for the NVPTX target.
12 //===----------------------------------------------------------------------===//
14 #include "NVPTXTargetMachine.h"
15 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
17 #include "NVPTXAllocaHoisting.h"
18 #include "NVPTXLowerAggrCopies.h"
19 #include "llvm/Analysis/Passes.h"
20 #include "llvm/CodeGen/AsmPrinter.h"
21 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
22 #include "llvm/CodeGen/MachineModuleInfo.h"
23 #include "llvm/CodeGen/Passes.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/IRPrintingPasses.h"
26 #include "llvm/IR/Verifier.h"
27 #include "llvm/MC/MCAsmInfo.h"
28 #include "llvm/MC/MCInstrInfo.h"
29 #include "llvm/MC/MCStreamer.h"
30 #include "llvm/MC/MCSubtargetInfo.h"
31 #include "llvm/PassManager.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/FormattedStream.h"
35 #include "llvm/Support/TargetRegistry.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetInstrInfo.h"
38 #include "llvm/Target/TargetLowering.h"
39 #include "llvm/Target/TargetLoweringObjectFile.h"
40 #include "llvm/Target/TargetMachine.h"
41 #include "llvm/Target/TargetOptions.h"
42 #include "llvm/Target/TargetRegisterInfo.h"
43 #include "llvm/Target/TargetSubtargetInfo.h"
44 #include "llvm/Transforms/Scalar.h"
49 void initializeNVVMReflectPass(PassRegistry&);
50 void initializeGenericToNVVMPass(PassRegistry&);
51 void initializeNVPTXAssignValidGlobalNamesPass(PassRegistry&);
52 void initializeNVPTXFavorNonGenericAddrSpacesPass(PassRegistry &);
53 void initializeNVPTXLowerStructArgsPass(PassRegistry &);
56 extern "C" void LLVMInitializeNVPTXTarget() {
57 // Register the target.
58 RegisterTargetMachine<NVPTXTargetMachine32> X(TheNVPTXTarget32);
59 RegisterTargetMachine<NVPTXTargetMachine64> Y(TheNVPTXTarget64);
61 // FIXME: This pass is really intended to be invoked during IR optimization,
62 // but it's very NVPTX-specific.
63 initializeNVVMReflectPass(*PassRegistry::getPassRegistry());
64 initializeGenericToNVVMPass(*PassRegistry::getPassRegistry());
65 initializeNVPTXAssignValidGlobalNamesPass(*PassRegistry::getPassRegistry());
66 initializeNVPTXFavorNonGenericAddrSpacesPass(
67 *PassRegistry::getPassRegistry());
68 initializeNVPTXLowerStructArgsPass(*PassRegistry::getPassRegistry());
71 NVPTXTargetMachine::NVPTXTargetMachine(const Target &T, StringRef TT,
72 StringRef CPU, StringRef FS,
73 const TargetOptions &Options,
74 Reloc::Model RM, CodeModel::Model CM,
75 CodeGenOpt::Level OL, bool is64bit)
76 : LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL),
77 Subtarget(TT, CPU, FS, *this, is64bit) {
81 void NVPTXTargetMachine32::anchor() {}
83 NVPTXTargetMachine32::NVPTXTargetMachine32(
84 const Target &T, StringRef TT, StringRef CPU, StringRef FS,
85 const TargetOptions &Options, Reloc::Model RM, CodeModel::Model CM,
87 : NVPTXTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
89 void NVPTXTargetMachine64::anchor() {}
91 NVPTXTargetMachine64::NVPTXTargetMachine64(
92 const Target &T, StringRef TT, StringRef CPU, StringRef FS,
93 const TargetOptions &Options, Reloc::Model RM, CodeModel::Model CM,
95 : NVPTXTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
98 class NVPTXPassConfig : public TargetPassConfig {
100 NVPTXPassConfig(NVPTXTargetMachine *TM, PassManagerBase &PM)
101 : TargetPassConfig(TM, PM) {}
103 NVPTXTargetMachine &getNVPTXTargetMachine() const {
104 return getTM<NVPTXTargetMachine>();
107 void addIRPasses() override;
108 bool addInstSelector() override;
109 bool addPreRegAlloc() override;
110 bool addPostRegAlloc() override;
111 void addMachineSSAOptimization() override;
113 FunctionPass *createTargetRegisterAllocator(bool) override;
114 void addFastRegAlloc(FunctionPass *RegAllocPass) override;
115 void addOptimizedRegAlloc(FunctionPass *RegAllocPass) override;
117 } // end anonymous namespace
119 TargetPassConfig *NVPTXTargetMachine::createPassConfig(PassManagerBase &PM) {
120 NVPTXPassConfig *PassConfig = new NVPTXPassConfig(this, PM);
124 void NVPTXTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
125 // Add first the target-independent BasicTTI pass, then our NVPTX pass. This
126 // allows the NVPTX pass to delegate to the target independent layer when
128 PM.add(createBasicTargetTransformInfoPass(this));
129 PM.add(createNVPTXTargetTransformInfoPass(this));
132 void NVPTXPassConfig::addIRPasses() {
133 // The following passes are known to not play well with virtual regs hanging
134 // around after register allocation (which in our case, is *all* registers).
135 // We explicitly disable them here. We do, however, need some functionality
136 // of the PrologEpilogCodeInserter pass, so we emulate that behavior in the
137 // NVPTXPrologEpilog pass (see NVPTXPrologEpilogPass.cpp).
138 disablePass(&PrologEpilogCodeInserterID);
139 disablePass(&MachineCopyPropagationID);
140 disablePass(&BranchFolderPassID);
141 disablePass(&TailDuplicateID);
143 addPass(createNVPTXImageOptimizerPass());
144 TargetPassConfig::addIRPasses();
145 addPass(createNVPTXAssignValidGlobalNamesPass());
146 addPass(createGenericToNVVMPass());
147 addPass(createNVPTXFavorNonGenericAddrSpacesPass());
148 addPass(createSeparateConstOffsetFromGEPPass());
149 // The SeparateConstOffsetFromGEP pass creates variadic bases that can be used
150 // by multiple GEPs. Run GVN or EarlyCSE to really reuse them. GVN generates
151 // significantly better code than EarlyCSE for some of our benchmarks.
152 if (getOptLevel() == CodeGenOpt::Aggressive)
153 addPass(createGVNPass());
155 addPass(createEarlyCSEPass());
156 // Both FavorNonGenericAddrSpaces and SeparateConstOffsetFromGEP may leave
157 // some dead code. We could remove dead code in an ad-hoc manner, but that
158 // requires manual work and might be error-prone.
160 // The FavorNonGenericAddrSpaces pass shortcuts unnecessary addrspacecasts,
161 // and leave them unused.
163 // SeparateConstOffsetFromGEP rebuilds a new index from the old index, and the
164 // old index and some of its intermediate results may become unused.
165 addPass(createDeadCodeEliminationPass());
168 bool NVPTXPassConfig::addInstSelector() {
169 const NVPTXSubtarget &ST =
170 getTM<NVPTXTargetMachine>().getSubtarget<NVPTXSubtarget>();
172 addPass(createLowerAggrCopies());
173 addPass(createAllocaHoisting());
174 addPass(createNVPTXISelDag(getNVPTXTargetMachine(), getOptLevel()));
176 if (!ST.hasImageHandles())
177 addPass(createNVPTXReplaceImageHandlesPass());
182 bool NVPTXPassConfig::addPreRegAlloc() { return false; }
183 bool NVPTXPassConfig::addPostRegAlloc() {
184 addPass(createNVPTXPrologEpilogPass());
188 FunctionPass *NVPTXPassConfig::createTargetRegisterAllocator(bool) {
189 return nullptr; // No reg alloc
192 void NVPTXPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
193 assert(!RegAllocPass && "NVPTX uses no regalloc!");
194 addPass(&PHIEliminationID);
195 addPass(&TwoAddressInstructionPassID);
198 void NVPTXPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
199 assert(!RegAllocPass && "NVPTX uses no regalloc!");
201 addPass(&ProcessImplicitDefsID);
202 addPass(&LiveVariablesID);
203 addPass(&MachineLoopInfoID);
204 addPass(&PHIEliminationID);
206 addPass(&TwoAddressInstructionPassID);
207 addPass(&RegisterCoalescerID);
209 // PreRA instruction scheduling.
210 if (addPass(&MachineSchedulerID))
211 printAndVerify("After Machine Scheduling");
214 addPass(&StackSlotColoringID);
216 // FIXME: Needs physical registers
217 //addPass(&PostRAMachineLICMID);
219 printAndVerify("After StackSlotColoring");
222 void NVPTXPassConfig::addMachineSSAOptimization() {
223 // Pre-ra tail duplication.
224 if (addPass(&EarlyTailDuplicateID))
225 printAndVerify("After Pre-RegAlloc TailDuplicate");
227 // Optimize PHIs before DCE: removing dead PHI cycles may make more
228 // instructions dead.
229 addPass(&OptimizePHIsID);
231 // This pass merges large allocas. StackSlotColoring is a different pass
232 // which merges spill slots.
233 addPass(&StackColoringID);
235 // If the target requests it, assign local variables to stack slots relative
236 // to one another and simplify frame index references where possible.
237 addPass(&LocalStackSlotAllocationID);
239 // With optimization, dead code should already be eliminated. However
240 // there is one known exception: lowered code for arguments that are only
241 // used by tail calls, where the tail calls reuse the incoming stack
242 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
243 addPass(&DeadMachineInstructionElimID);
244 printAndVerify("After codegen DCE pass");
246 // Allow targets to insert passes that improve instruction level parallelism,
247 // like if-conversion. Such passes will typically need dominator trees and
248 // loop info, just like LICM and CSE below.
250 printAndVerify("After ILP optimizations");
252 addPass(&MachineLICMID);
253 addPass(&MachineCSEID);
255 addPass(&MachineSinkingID);
256 printAndVerify("After Machine LICM, CSE and Sinking passes");
258 addPass(&PeepholeOptimizerID);
259 printAndVerify("After codegen peephole optimization pass");