1 //===-- NVPTXTargetMachine.cpp - Define TargetMachine for NVPTX -----------===//
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 // Top-level implementation for the NVPTX target.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "NVPTXTargetMachine.h"
15 #include "NVPTX.h"
16 #include "NVPTXAllocaHoisting.h"
17 #include "NVPTXLowerAggrCopies.h"
18 #include "NVPTXTargetObjectFile.h"
19 #include "NVPTXTargetTransformInfo.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Triple.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/CodeGen/Passes.h"
24 #include "llvm/CodeGen/TargetPassConfig.h"
25 #include "llvm/IR/LegacyPassManager.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/TargetRegistry.h"
29 #include "llvm/Target/TargetMachine.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
32 #include "llvm/Transforms/Scalar.h"
33 #include "llvm/Transforms/Scalar/GVN.h"
34 #include "llvm/Transforms/Vectorize.h"
35 #include <cassert>
36 #include <string>
37 
38 using namespace llvm;
39 
40 // LSV is still relatively new; this switch lets us turn it off in case we
41 // encounter (or suspect) a bug.
42 static cl::opt<bool>
43     DisableLoadStoreVectorizer("disable-nvptx-load-store-vectorizer",
44                                cl::desc("Disable load/store vectorizer"),
45                                cl::init(false), cl::Hidden);
46 
47 // TODO: Remove this flag when we are confident with no regressions.
48 static cl::opt<bool> DisableRequireStructuredCFG(
49     "disable-nvptx-require-structured-cfg",
50     cl::desc("Transitional flag to turn off NVPTX's requirement on preserving "
51              "structured CFG. The requirement should be disabled only when "
52              "unexpected regressions happen."),
53     cl::init(false), cl::Hidden);
54 
55 namespace llvm {
56 
57 void initializeNVVMIntrRangePass(PassRegistry&);
58 void initializeNVVMReflectPass(PassRegistry&);
59 void initializeGenericToNVVMPass(PassRegistry&);
60 void initializeNVPTXAllocaHoistingPass(PassRegistry &);
61 void initializeNVPTXAssignValidGlobalNamesPass(PassRegistry&);
62 void initializeNVPTXLowerAggrCopiesPass(PassRegistry &);
63 void initializeNVPTXLowerArgsPass(PassRegistry &);
64 void initializeNVPTXLowerAllocaPass(PassRegistry &);
65 
66 } // end namespace llvm
67 
68 extern "C" void LLVMInitializeNVPTXTarget() {
69   // Register the target.
70   RegisterTargetMachine<NVPTXTargetMachine32> X(getTheNVPTXTarget32());
71   RegisterTargetMachine<NVPTXTargetMachine64> Y(getTheNVPTXTarget64());
72 
73   // FIXME: This pass is really intended to be invoked during IR optimization,
74   // but it's very NVPTX-specific.
75   PassRegistry &PR = *PassRegistry::getPassRegistry();
76   initializeNVVMReflectPass(PR);
77   initializeNVVMIntrRangePass(PR);
78   initializeGenericToNVVMPass(PR);
79   initializeNVPTXAllocaHoistingPass(PR);
80   initializeNVPTXAssignValidGlobalNamesPass(PR);
81   initializeNVPTXLowerArgsPass(PR);
82   initializeNVPTXLowerAllocaPass(PR);
83   initializeNVPTXLowerAggrCopiesPass(PR);
84 }
85 
86 static std::string computeDataLayout(bool is64Bit) {
87   std::string Ret = "e";
88 
89   if (!is64Bit)
90     Ret += "-p:32:32";
91 
92   Ret += "-i64:64-i128:128-v16:16-v32:32-n16:32:64";
93 
94   return Ret;
95 }
96 
97 static CodeModel::Model getEffectiveCodeModel(Optional<CodeModel::Model> CM) {
98   if (CM)
99     return *CM;
100   return CodeModel::Small;
101 }
102 
103 NVPTXTargetMachine::NVPTXTargetMachine(const Target &T, const Triple &TT,
104                                        StringRef CPU, StringRef FS,
105                                        const TargetOptions &Options,
106                                        Optional<Reloc::Model> RM,
107                                        Optional<CodeModel::Model> CM,
108                                        CodeGenOpt::Level OL, bool is64bit)
109     // The pic relocation model is used regardless of what the client has
110     // specified, as it is the only relocation model currently supported.
111     : LLVMTargetMachine(T, computeDataLayout(is64bit), TT, CPU, FS, Options,
112                         Reloc::PIC_, getEffectiveCodeModel(CM), OL),
113       is64bit(is64bit), TLOF(llvm::make_unique<NVPTXTargetObjectFile>()),
114       Subtarget(TT, CPU, FS, *this) {
115   if (TT.getOS() == Triple::NVCL)
116     drvInterface = NVPTX::NVCL;
117   else
118     drvInterface = NVPTX::CUDA;
119   if (!DisableRequireStructuredCFG)
120     setRequiresStructuredCFG(true);
121   initAsmInfo();
122 }
123 
124 NVPTXTargetMachine::~NVPTXTargetMachine() = default;
125 
126 void NVPTXTargetMachine32::anchor() {}
127 
128 NVPTXTargetMachine32::NVPTXTargetMachine32(const Target &T, const Triple &TT,
129                                            StringRef CPU, StringRef FS,
130                                            const TargetOptions &Options,
131                                            Optional<Reloc::Model> RM,
132                                            Optional<CodeModel::Model> CM,
133                                            CodeGenOpt::Level OL, bool JIT)
134     : NVPTXTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
135 
136 void NVPTXTargetMachine64::anchor() {}
137 
138 NVPTXTargetMachine64::NVPTXTargetMachine64(const Target &T, const Triple &TT,
139                                            StringRef CPU, StringRef FS,
140                                            const TargetOptions &Options,
141                                            Optional<Reloc::Model> RM,
142                                            Optional<CodeModel::Model> CM,
143                                            CodeGenOpt::Level OL, bool JIT)
144     : NVPTXTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
145 
146 namespace {
147 
148 class NVPTXPassConfig : public TargetPassConfig {
149 public:
150   NVPTXPassConfig(NVPTXTargetMachine &TM, PassManagerBase &PM)
151       : TargetPassConfig(TM, PM) {}
152 
153   NVPTXTargetMachine &getNVPTXTargetMachine() const {
154     return getTM<NVPTXTargetMachine>();
155   }
156 
157   void addIRPasses() override;
158   bool addInstSelector() override;
159   void addPostRegAlloc() override;
160   void addMachineSSAOptimization() override;
161 
162   FunctionPass *createTargetRegisterAllocator(bool) override;
163   void addFastRegAlloc(FunctionPass *RegAllocPass) override;
164   void addOptimizedRegAlloc(FunctionPass *RegAllocPass) override;
165 
166 private:
167   // If the opt level is aggressive, add GVN; otherwise, add EarlyCSE. This
168   // function is only called in opt mode.
169   void addEarlyCSEOrGVNPass();
170 
171   // Add passes that propagate special memory spaces.
172   void addAddressSpaceInferencePasses();
173 
174   // Add passes that perform straight-line scalar optimizations.
175   void addStraightLineScalarOptimizationPasses();
176 };
177 
178 } // end anonymous namespace
179 
180 TargetPassConfig *NVPTXTargetMachine::createPassConfig(PassManagerBase &PM) {
181   return new NVPTXPassConfig(*this, PM);
182 }
183 
184 void NVPTXTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
185   Builder.addExtension(
186     PassManagerBuilder::EP_EarlyAsPossible,
187     [&](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
188       PM.add(createNVVMReflectPass());
189       PM.add(createNVVMIntrRangePass(Subtarget.getSmVersion()));
190     });
191 }
192 
193 TargetTransformInfo
194 NVPTXTargetMachine::getTargetTransformInfo(const Function &F) {
195   return TargetTransformInfo(NVPTXTTIImpl(this, F));
196 }
197 
198 void NVPTXPassConfig::addEarlyCSEOrGVNPass() {
199   if (getOptLevel() == CodeGenOpt::Aggressive)
200     addPass(createGVNPass());
201   else
202     addPass(createEarlyCSEPass());
203 }
204 
205 void NVPTXPassConfig::addAddressSpaceInferencePasses() {
206   // NVPTXLowerArgs emits alloca for byval parameters which can often
207   // be eliminated by SROA.
208   addPass(createSROAPass());
209   addPass(createNVPTXLowerAllocaPass());
210   addPass(createInferAddressSpacesPass());
211 }
212 
213 void NVPTXPassConfig::addStraightLineScalarOptimizationPasses() {
214   addPass(createSeparateConstOffsetFromGEPPass());
215   addPass(createSpeculativeExecutionPass());
216   // ReassociateGEPs exposes more opportunites for SLSR. See
217   // the example in reassociate-geps-and-slsr.ll.
218   addPass(createStraightLineStrengthReducePass());
219   // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
220   // EarlyCSE can reuse. GVN generates significantly better code than EarlyCSE
221   // for some of our benchmarks.
222   addEarlyCSEOrGVNPass();
223   // Run NaryReassociate after EarlyCSE/GVN to be more effective.
224   addPass(createNaryReassociatePass());
225   // NaryReassociate on GEPs creates redundant common expressions, so run
226   // EarlyCSE after it.
227   addPass(createEarlyCSEPass());
228 }
229 
230 void NVPTXPassConfig::addIRPasses() {
231   // The following passes are known to not play well with virtual regs hanging
232   // around after register allocation (which in our case, is *all* registers).
233   // We explicitly disable them here.  We do, however, need some functionality
234   // of the PrologEpilogCodeInserter pass, so we emulate that behavior in the
235   // NVPTXPrologEpilog pass (see NVPTXPrologEpilogPass.cpp).
236   disablePass(&PrologEpilogCodeInserterID);
237   disablePass(&MachineCopyPropagationID);
238   disablePass(&TailDuplicateID);
239   disablePass(&StackMapLivenessID);
240   disablePass(&LiveDebugValuesID);
241   disablePass(&PostRAMachineSinkingID);
242   disablePass(&PostRASchedulerID);
243   disablePass(&FuncletLayoutID);
244   disablePass(&PatchableFunctionID);
245   disablePass(&ShrinkWrapID);
246 
247   // NVVMReflectPass is added in addEarlyAsPossiblePasses, so hopefully running
248   // it here does nothing.  But since we need it for correctness when lowering
249   // to NVPTX, run it here too, in case whoever built our pass pipeline didn't
250   // call addEarlyAsPossiblePasses.
251   addPass(createNVVMReflectPass());
252 
253   if (getOptLevel() != CodeGenOpt::None)
254     addPass(createNVPTXImageOptimizerPass());
255   addPass(createNVPTXAssignValidGlobalNamesPass());
256   addPass(createGenericToNVVMPass());
257 
258   // NVPTXLowerArgs is required for correctness and should be run right
259   // before the address space inference passes.
260   addPass(createNVPTXLowerArgsPass(&getNVPTXTargetMachine()));
261   if (getOptLevel() != CodeGenOpt::None) {
262     addAddressSpaceInferencePasses();
263     if (!DisableLoadStoreVectorizer)
264       addPass(createLoadStoreVectorizerPass());
265     addStraightLineScalarOptimizationPasses();
266   }
267 
268   // === LSR and other generic IR passes ===
269   TargetPassConfig::addIRPasses();
270   // EarlyCSE is not always strong enough to clean up what LSR produces. For
271   // example, GVN can combine
272   //
273   //   %0 = add %a, %b
274   //   %1 = add %b, %a
275   //
276   // and
277   //
278   //   %0 = shl nsw %a, 2
279   //   %1 = shl %a, 2
280   //
281   // but EarlyCSE can do neither of them.
282   if (getOptLevel() != CodeGenOpt::None)
283     addEarlyCSEOrGVNPass();
284 }
285 
286 bool NVPTXPassConfig::addInstSelector() {
287   const NVPTXSubtarget &ST = *getTM<NVPTXTargetMachine>().getSubtargetImpl();
288 
289   addPass(createLowerAggrCopies());
290   addPass(createAllocaHoisting());
291   addPass(createNVPTXISelDag(getNVPTXTargetMachine(), getOptLevel()));
292 
293   if (!ST.hasImageHandles())
294     addPass(createNVPTXReplaceImageHandlesPass());
295 
296   return false;
297 }
298 
299 void NVPTXPassConfig::addPostRegAlloc() {
300   addPass(createNVPTXPrologEpilogPass(), false);
301   if (getOptLevel() != CodeGenOpt::None) {
302     // NVPTXPrologEpilogPass calculates frame object offset and replace frame
303     // index with VRFrame register. NVPTXPeephole need to be run after that and
304     // will replace VRFrame with VRFrameLocal when possible.
305     addPass(createNVPTXPeephole());
306   }
307 }
308 
309 FunctionPass *NVPTXPassConfig::createTargetRegisterAllocator(bool) {
310   return nullptr; // No reg alloc
311 }
312 
313 void NVPTXPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
314   assert(!RegAllocPass && "NVPTX uses no regalloc!");
315   addPass(&PHIEliminationID);
316   addPass(&TwoAddressInstructionPassID);
317 }
318 
319 void NVPTXPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
320   assert(!RegAllocPass && "NVPTX uses no regalloc!");
321 
322   addPass(&ProcessImplicitDefsID);
323   addPass(&LiveVariablesID);
324   addPass(&MachineLoopInfoID);
325   addPass(&PHIEliminationID);
326 
327   addPass(&TwoAddressInstructionPassID);
328   addPass(&RegisterCoalescerID);
329 
330   // PreRA instruction scheduling.
331   if (addPass(&MachineSchedulerID))
332     printAndVerify("After Machine Scheduling");
333 
334 
335   addPass(&StackSlotColoringID);
336 
337   // FIXME: Needs physical registers
338   //addPass(&MachineLICMID);
339 
340   printAndVerify("After StackSlotColoring");
341 }
342 
343 void NVPTXPassConfig::addMachineSSAOptimization() {
344   // Pre-ra tail duplication.
345   if (addPass(&EarlyTailDuplicateID))
346     printAndVerify("After Pre-RegAlloc TailDuplicate");
347 
348   // Optimize PHIs before DCE: removing dead PHI cycles may make more
349   // instructions dead.
350   addPass(&OptimizePHIsID);
351 
352   // This pass merges large allocas. StackSlotColoring is a different pass
353   // which merges spill slots.
354   addPass(&StackColoringID);
355 
356   // If the target requests it, assign local variables to stack slots relative
357   // to one another and simplify frame index references where possible.
358   addPass(&LocalStackSlotAllocationID);
359 
360   // With optimization, dead code should already be eliminated. However
361   // there is one known exception: lowered code for arguments that are only
362   // used by tail calls, where the tail calls reuse the incoming stack
363   // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
364   addPass(&DeadMachineInstructionElimID);
365   printAndVerify("After codegen DCE pass");
366 
367   // Allow targets to insert passes that improve instruction level parallelism,
368   // like if-conversion. Such passes will typically need dominator trees and
369   // loop info, just like LICM and CSE below.
370   if (addILPOpts())
371     printAndVerify("After ILP optimizations");
372 
373   addPass(&EarlyMachineLICMID);
374   addPass(&MachineCSEID);
375 
376   addPass(&MachineSinkingID);
377   printAndVerify("After Machine LICM, CSE and Sinking passes");
378 
379   addPass(&PeepholeOptimizerID);
380   printAndVerify("After codegen peephole optimization pass");
381 }
382