1 //===-- NVPTXTargetMachine.cpp - Define TargetMachine for NVPTX -----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Top-level implementation for the NVPTX target.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "NVPTXTargetMachine.h"
14 #include "NVPTX.h"
15 #include "NVPTXAllocaHoisting.h"
16 #include "NVPTXLowerAggrCopies.h"
17 #include "NVPTXTargetObjectFile.h"
18 #include "NVPTXTargetTransformInfo.h"
19 #include "TargetInfo/NVPTXTargetInfo.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/Passes/PassBuilder.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/TargetRegistry.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
33 #include "llvm/Transforms/Scalar.h"
34 #include "llvm/Transforms/Scalar/GVN.h"
35 #include "llvm/Transforms/Vectorize.h"
36 #include <cassert>
37 #include <string>
38 
39 using namespace llvm;
40 
41 // LSV is still relatively new; this switch lets us turn it off in case we
42 // encounter (or suspect) a bug.
43 static cl::opt<bool>
44     DisableLoadStoreVectorizer("disable-nvptx-load-store-vectorizer",
45                                cl::desc("Disable load/store vectorizer"),
46                                cl::init(false), cl::Hidden);
47 
48 // TODO: Remove this flag when we are confident with no regressions.
49 static cl::opt<bool> DisableRequireStructuredCFG(
50     "disable-nvptx-require-structured-cfg",
51     cl::desc("Transitional flag to turn off NVPTX's requirement on preserving "
52              "structured CFG. The requirement should be disabled only when "
53              "unexpected regressions happen."),
54     cl::init(false), cl::Hidden);
55 
56 static cl::opt<bool> UseShortPointersOpt(
57     "nvptx-short-ptr",
58     cl::desc(
59         "Use 32-bit pointers for accessing const/local/shared address spaces."),
60     cl::init(false), cl::Hidden);
61 
62 namespace llvm {
63 
64 void initializeNVVMIntrRangePass(PassRegistry&);
65 void initializeNVVMReflectPass(PassRegistry&);
66 void initializeGenericToNVVMPass(PassRegistry&);
67 void initializeNVPTXAllocaHoistingPass(PassRegistry &);
68 void initializeNVPTXAssignValidGlobalNamesPass(PassRegistry&);
69 void initializeNVPTXLowerAggrCopiesPass(PassRegistry &);
70 void initializeNVPTXLowerArgsPass(PassRegistry &);
71 void initializeNVPTXLowerAllocaPass(PassRegistry &);
72 void initializeNVPTXProxyRegErasurePass(PassRegistry &);
73 
74 } // end namespace llvm
75 
76 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeNVPTXTarget() {
77   // Register the target.
78   RegisterTargetMachine<NVPTXTargetMachine32> X(getTheNVPTXTarget32());
79   RegisterTargetMachine<NVPTXTargetMachine64> Y(getTheNVPTXTarget64());
80 
81   // FIXME: This pass is really intended to be invoked during IR optimization,
82   // but it's very NVPTX-specific.
83   PassRegistry &PR = *PassRegistry::getPassRegistry();
84   initializeNVVMReflectPass(PR);
85   initializeNVVMIntrRangePass(PR);
86   initializeGenericToNVVMPass(PR);
87   initializeNVPTXAllocaHoistingPass(PR);
88   initializeNVPTXAssignValidGlobalNamesPass(PR);
89   initializeNVPTXLowerArgsPass(PR);
90   initializeNVPTXLowerAllocaPass(PR);
91   initializeNVPTXLowerAggrCopiesPass(PR);
92   initializeNVPTXProxyRegErasurePass(PR);
93 }
94 
95 static std::string computeDataLayout(bool is64Bit, bool UseShortPointers) {
96   std::string Ret = "e";
97 
98   if (!is64Bit)
99     Ret += "-p:32:32";
100   else if (UseShortPointers)
101     Ret += "-p3:32:32-p4:32:32-p5:32:32";
102 
103   Ret += "-i64:64-i128:128-v16:16-v32:32-n16:32:64";
104 
105   return Ret;
106 }
107 
108 NVPTXTargetMachine::NVPTXTargetMachine(const Target &T, const Triple &TT,
109                                        StringRef CPU, StringRef FS,
110                                        const TargetOptions &Options,
111                                        Optional<Reloc::Model> RM,
112                                        Optional<CodeModel::Model> CM,
113                                        CodeGenOpt::Level OL, bool is64bit)
114     // The pic relocation model is used regardless of what the client has
115     // specified, as it is the only relocation model currently supported.
116     : LLVMTargetMachine(T, computeDataLayout(is64bit, UseShortPointersOpt), TT,
117                         CPU, FS, Options, Reloc::PIC_,
118                         getEffectiveCodeModel(CM, CodeModel::Small), OL),
119       is64bit(is64bit), UseShortPointers(UseShortPointersOpt),
120       TLOF(std::make_unique<NVPTXTargetObjectFile>()),
121       Subtarget(TT, std::string(CPU), std::string(FS), *this) {
122   if (TT.getOS() == Triple::NVCL)
123     drvInterface = NVPTX::NVCL;
124   else
125     drvInterface = NVPTX::CUDA;
126   if (!DisableRequireStructuredCFG)
127     setRequiresStructuredCFG(true);
128   initAsmInfo();
129 }
130 
131 NVPTXTargetMachine::~NVPTXTargetMachine() = default;
132 
133 void NVPTXTargetMachine32::anchor() {}
134 
135 NVPTXTargetMachine32::NVPTXTargetMachine32(const Target &T, const Triple &TT,
136                                            StringRef CPU, StringRef FS,
137                                            const TargetOptions &Options,
138                                            Optional<Reloc::Model> RM,
139                                            Optional<CodeModel::Model> CM,
140                                            CodeGenOpt::Level OL, bool JIT)
141     : NVPTXTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
142 
143 void NVPTXTargetMachine64::anchor() {}
144 
145 NVPTXTargetMachine64::NVPTXTargetMachine64(const Target &T, const Triple &TT,
146                                            StringRef CPU, StringRef FS,
147                                            const TargetOptions &Options,
148                                            Optional<Reloc::Model> RM,
149                                            Optional<CodeModel::Model> CM,
150                                            CodeGenOpt::Level OL, bool JIT)
151     : NVPTXTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
152 
153 namespace {
154 
155 class NVPTXPassConfig : public TargetPassConfig {
156 public:
157   NVPTXPassConfig(NVPTXTargetMachine &TM, PassManagerBase &PM)
158       : TargetPassConfig(TM, PM) {}
159 
160   NVPTXTargetMachine &getNVPTXTargetMachine() const {
161     return getTM<NVPTXTargetMachine>();
162   }
163 
164   void addIRPasses() override;
165   bool addInstSelector() override;
166   void addPreRegAlloc() override;
167   void addPostRegAlloc() override;
168   void addMachineSSAOptimization() override;
169 
170   FunctionPass *createTargetRegisterAllocator(bool) override;
171   void addFastRegAlloc() override;
172   void addOptimizedRegAlloc() override;
173 
174   bool addRegAssignAndRewriteFast() override {
175     llvm_unreachable("should not be used");
176   }
177 
178   bool addRegAssignAndRewriteOptimized() override {
179     llvm_unreachable("should not be used");
180   }
181 
182 private:
183   // If the opt level is aggressive, add GVN; otherwise, add EarlyCSE. This
184   // function is only called in opt mode.
185   void addEarlyCSEOrGVNPass();
186 
187   // Add passes that propagate special memory spaces.
188   void addAddressSpaceInferencePasses();
189 
190   // Add passes that perform straight-line scalar optimizations.
191   void addStraightLineScalarOptimizationPasses();
192 };
193 
194 } // end anonymous namespace
195 
196 TargetPassConfig *NVPTXTargetMachine::createPassConfig(PassManagerBase &PM) {
197   return new NVPTXPassConfig(*this, PM);
198 }
199 
200 void NVPTXTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
201   Builder.addExtension(
202     PassManagerBuilder::EP_EarlyAsPossible,
203     [&](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
204       PM.add(createNVVMReflectPass(Subtarget.getSmVersion()));
205       PM.add(createNVVMIntrRangePass(Subtarget.getSmVersion()));
206     });
207 }
208 
209 void NVPTXTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB,
210                                                       bool DebugPassManager) {
211   PB.registerPipelineParsingCallback(
212       [](StringRef PassName, FunctionPassManager &PM,
213          ArrayRef<PassBuilder::PipelineElement>) {
214         if (PassName == "nvvm-reflect") {
215           PM.addPass(NVVMReflectPass());
216           return true;
217         }
218         if (PassName == "nvvm-intr-range") {
219           PM.addPass(NVVMIntrRangePass());
220           return true;
221         }
222         return false;
223       });
224 
225   PB.registerPipelineStartEPCallback(
226       [this, DebugPassManager](ModulePassManager &PM,
227                                PassBuilder::OptimizationLevel Level) {
228         FunctionPassManager FPM(DebugPassManager);
229         FPM.addPass(NVVMReflectPass(Subtarget.getSmVersion()));
230         FPM.addPass(NVVMIntrRangePass(Subtarget.getSmVersion()));
231         PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
232       });
233 }
234 
235 TargetTransformInfo
236 NVPTXTargetMachine::getTargetTransformInfo(const Function &F) {
237   return TargetTransformInfo(NVPTXTTIImpl(this, F));
238 }
239 
240 void NVPTXPassConfig::addEarlyCSEOrGVNPass() {
241   if (getOptLevel() == CodeGenOpt::Aggressive)
242     addPass(createGVNPass());
243   else
244     addPass(createEarlyCSEPass());
245 }
246 
247 void NVPTXPassConfig::addAddressSpaceInferencePasses() {
248   // NVPTXLowerArgs emits alloca for byval parameters which can often
249   // be eliminated by SROA.
250   addPass(createSROAPass());
251   addPass(createNVPTXLowerAllocaPass());
252   addPass(createInferAddressSpacesPass());
253 }
254 
255 void NVPTXPassConfig::addStraightLineScalarOptimizationPasses() {
256   addPass(createSeparateConstOffsetFromGEPPass());
257   addPass(createSpeculativeExecutionPass());
258   // ReassociateGEPs exposes more opportunites for SLSR. See
259   // the example in reassociate-geps-and-slsr.ll.
260   addPass(createStraightLineStrengthReducePass());
261   // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
262   // EarlyCSE can reuse. GVN generates significantly better code than EarlyCSE
263   // for some of our benchmarks.
264   addEarlyCSEOrGVNPass();
265   // Run NaryReassociate after EarlyCSE/GVN to be more effective.
266   addPass(createNaryReassociatePass());
267   // NaryReassociate on GEPs creates redundant common expressions, so run
268   // EarlyCSE after it.
269   addPass(createEarlyCSEPass());
270 }
271 
272 void NVPTXPassConfig::addIRPasses() {
273   // The following passes are known to not play well with virtual regs hanging
274   // around after register allocation (which in our case, is *all* registers).
275   // We explicitly disable them here.  We do, however, need some functionality
276   // of the PrologEpilogCodeInserter pass, so we emulate that behavior in the
277   // NVPTXPrologEpilog pass (see NVPTXPrologEpilogPass.cpp).
278   disablePass(&PrologEpilogCodeInserterID);
279   disablePass(&MachineCopyPropagationID);
280   disablePass(&TailDuplicateID);
281   disablePass(&StackMapLivenessID);
282   disablePass(&LiveDebugValuesID);
283   disablePass(&PostRAMachineSinkingID);
284   disablePass(&PostRASchedulerID);
285   disablePass(&FuncletLayoutID);
286   disablePass(&PatchableFunctionID);
287   disablePass(&ShrinkWrapID);
288 
289   // NVVMReflectPass is added in addEarlyAsPossiblePasses, so hopefully running
290   // it here does nothing.  But since we need it for correctness when lowering
291   // to NVPTX, run it here too, in case whoever built our pass pipeline didn't
292   // call addEarlyAsPossiblePasses.
293   const NVPTXSubtarget &ST = *getTM<NVPTXTargetMachine>().getSubtargetImpl();
294   addPass(createNVVMReflectPass(ST.getSmVersion()));
295 
296   if (getOptLevel() != CodeGenOpt::None)
297     addPass(createNVPTXImageOptimizerPass());
298   addPass(createNVPTXAssignValidGlobalNamesPass());
299   addPass(createGenericToNVVMPass());
300 
301   // NVPTXLowerArgs is required for correctness and should be run right
302   // before the address space inference passes.
303   addPass(createNVPTXLowerArgsPass(&getNVPTXTargetMachine()));
304   if (getOptLevel() != CodeGenOpt::None) {
305     addAddressSpaceInferencePasses();
306     addStraightLineScalarOptimizationPasses();
307   }
308 
309   // === LSR and other generic IR passes ===
310   TargetPassConfig::addIRPasses();
311   // EarlyCSE is not always strong enough to clean up what LSR produces. For
312   // example, GVN can combine
313   //
314   //   %0 = add %a, %b
315   //   %1 = add %b, %a
316   //
317   // and
318   //
319   //   %0 = shl nsw %a, 2
320   //   %1 = shl %a, 2
321   //
322   // but EarlyCSE can do neither of them.
323   if (getOptLevel() != CodeGenOpt::None) {
324     addEarlyCSEOrGVNPass();
325     if (!DisableLoadStoreVectorizer)
326       addPass(createLoadStoreVectorizerPass());
327   }
328 }
329 
330 bool NVPTXPassConfig::addInstSelector() {
331   const NVPTXSubtarget &ST = *getTM<NVPTXTargetMachine>().getSubtargetImpl();
332 
333   addPass(createLowerAggrCopies());
334   addPass(createAllocaHoisting());
335   addPass(createNVPTXISelDag(getNVPTXTargetMachine(), getOptLevel()));
336 
337   if (!ST.hasImageHandles())
338     addPass(createNVPTXReplaceImageHandlesPass());
339 
340   return false;
341 }
342 
343 void NVPTXPassConfig::addPreRegAlloc() {
344   // Remove Proxy Register pseudo instructions used to keep `callseq_end` alive.
345   addPass(createNVPTXProxyRegErasurePass());
346 }
347 
348 void NVPTXPassConfig::addPostRegAlloc() {
349   addPass(createNVPTXPrologEpilogPass(), false);
350   if (getOptLevel() != CodeGenOpt::None) {
351     // NVPTXPrologEpilogPass calculates frame object offset and replace frame
352     // index with VRFrame register. NVPTXPeephole need to be run after that and
353     // will replace VRFrame with VRFrameLocal when possible.
354     addPass(createNVPTXPeephole());
355   }
356 }
357 
358 FunctionPass *NVPTXPassConfig::createTargetRegisterAllocator(bool) {
359   return nullptr; // No reg alloc
360 }
361 
362 void NVPTXPassConfig::addFastRegAlloc() {
363   addPass(&PHIEliminationID);
364   addPass(&TwoAddressInstructionPassID);
365 }
366 
367 void NVPTXPassConfig::addOptimizedRegAlloc() {
368   addPass(&ProcessImplicitDefsID);
369   addPass(&LiveVariablesID);
370   addPass(&MachineLoopInfoID);
371   addPass(&PHIEliminationID);
372 
373   addPass(&TwoAddressInstructionPassID);
374   addPass(&RegisterCoalescerID);
375 
376   // PreRA instruction scheduling.
377   if (addPass(&MachineSchedulerID))
378     printAndVerify("After Machine Scheduling");
379 
380 
381   addPass(&StackSlotColoringID);
382 
383   // FIXME: Needs physical registers
384   //addPass(&MachineLICMID);
385 
386   printAndVerify("After StackSlotColoring");
387 }
388 
389 void NVPTXPassConfig::addMachineSSAOptimization() {
390   // Pre-ra tail duplication.
391   if (addPass(&EarlyTailDuplicateID))
392     printAndVerify("After Pre-RegAlloc TailDuplicate");
393 
394   // Optimize PHIs before DCE: removing dead PHI cycles may make more
395   // instructions dead.
396   addPass(&OptimizePHIsID);
397 
398   // This pass merges large allocas. StackSlotColoring is a different pass
399   // which merges spill slots.
400   addPass(&StackColoringID);
401 
402   // If the target requests it, assign local variables to stack slots relative
403   // to one another and simplify frame index references where possible.
404   addPass(&LocalStackSlotAllocationID);
405 
406   // With optimization, dead code should already be eliminated. However
407   // there is one known exception: lowered code for arguments that are only
408   // used by tail calls, where the tail calls reuse the incoming stack
409   // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
410   addPass(&DeadMachineInstructionElimID);
411   printAndVerify("After codegen DCE pass");
412 
413   // Allow targets to insert passes that improve instruction level parallelism,
414   // like if-conversion. Such passes will typically need dominator trees and
415   // loop info, just like LICM and CSE below.
416   if (addILPOpts())
417     printAndVerify("After ILP optimizations");
418 
419   addPass(&EarlyMachineLICMID);
420   addPass(&MachineCSEID);
421 
422   addPass(&MachineSinkingID);
423   printAndVerify("After Machine LICM, CSE and Sinking passes");
424 
425   addPass(&PeepholeOptimizerID);
426   printAndVerify("After codegen peephole optimization pass");
427 }
428