xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPURewriteOutArguments.cpp (revision c66ec88fed842fbaad62c30d510644ceb7bd2d71)
1 //===- AMDGPURewriteOutArgumentsPass.cpp - Create struct returns ----------===//
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 /// \file This pass attempts to replace out argument usage with a return of a
10 /// struct.
11 ///
12 /// We can support returning a lot of values directly in registers, but
13 /// idiomatic C code frequently uses a pointer argument to return a second value
14 /// rather than returning a struct by value. GPU stack access is also quite
15 /// painful, so we want to avoid that if possible. Passing a stack object
16 /// pointer to a function also requires an additional address expansion code
17 /// sequence to convert the pointer to be relative to the kernel's scratch wave
18 /// offset register since the callee doesn't know what stack frame the incoming
19 /// pointer is relative to.
20 ///
21 /// The goal is to try rewriting code that looks like this:
22 ///
23 ///  int foo(int a, int b, int* out) {
24 ///     *out = bar();
25 ///     return a + b;
26 /// }
27 ///
28 /// into something like this:
29 ///
30 ///  std::pair<int, int> foo(int a, int b) {
31 ///     return std::make_pair(a + b, bar());
32 /// }
33 ///
34 /// Typically the incoming pointer is a simple alloca for a temporary variable
35 /// to use the API, which if replaced with a struct return will be easily SROA'd
36 /// out when the stub function we create is inlined
37 ///
38 /// This pass introduces the struct return, but leaves the unused pointer
39 /// arguments and introduces a new stub function calling the struct returning
40 /// body. DeadArgumentElimination should be run after this to clean these up.
41 //
42 //===----------------------------------------------------------------------===//
43 
44 #include "AMDGPU.h"
45 #include "Utils/AMDGPUBaseInfo.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/STLExtras.h"
48 #include "llvm/ADT/SmallSet.h"
49 #include "llvm/ADT/SmallVector.h"
50 #include "llvm/ADT/Statistic.h"
51 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
52 #include "llvm/Analysis/MemoryLocation.h"
53 #include "llvm/IR/Argument.h"
54 #include "llvm/IR/Attributes.h"
55 #include "llvm/IR/BasicBlock.h"
56 #include "llvm/IR/Constants.h"
57 #include "llvm/IR/DataLayout.h"
58 #include "llvm/IR/DerivedTypes.h"
59 #include "llvm/IR/Function.h"
60 #include "llvm/IR/IRBuilder.h"
61 #include "llvm/IR/Instructions.h"
62 #include "llvm/IR/Module.h"
63 #include "llvm/IR/Type.h"
64 #include "llvm/IR/Use.h"
65 #include "llvm/IR/User.h"
66 #include "llvm/IR/Value.h"
67 #include "llvm/InitializePasses.h"
68 #include "llvm/Pass.h"
69 #include "llvm/Support/Casting.h"
70 #include "llvm/Support/CommandLine.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/raw_ostream.h"
73 #include <cassert>
74 #include <utility>
75 
76 #define DEBUG_TYPE "amdgpu-rewrite-out-arguments"
77 
78 using namespace llvm;
79 
80 static cl::opt<bool> AnyAddressSpace(
81   "amdgpu-any-address-space-out-arguments",
82   cl::desc("Replace pointer out arguments with "
83            "struct returns for non-private address space"),
84   cl::Hidden,
85   cl::init(false));
86 
87 static cl::opt<unsigned> MaxNumRetRegs(
88   "amdgpu-max-return-arg-num-regs",
89   cl::desc("Approximately limit number of return registers for replacing out arguments"),
90   cl::Hidden,
91   cl::init(16));
92 
93 STATISTIC(NumOutArgumentsReplaced,
94           "Number out arguments moved to struct return values");
95 STATISTIC(NumOutArgumentFunctionsReplaced,
96           "Number of functions with out arguments moved to struct return values");
97 
98 namespace {
99 
100 class AMDGPURewriteOutArguments : public FunctionPass {
101 private:
102   const DataLayout *DL = nullptr;
103   MemoryDependenceResults *MDA = nullptr;
104 
105   bool checkArgumentUses(Value &Arg) const;
106   bool isOutArgumentCandidate(Argument &Arg) const;
107 
108 #ifndef NDEBUG
109   bool isVec3ToVec4Shuffle(Type *Ty0, Type* Ty1) const;
110 #endif
111 
112 public:
113   static char ID;
114 
115   AMDGPURewriteOutArguments() : FunctionPass(ID) {}
116 
117   void getAnalysisUsage(AnalysisUsage &AU) const override {
118     AU.addRequired<MemoryDependenceWrapperPass>();
119     FunctionPass::getAnalysisUsage(AU);
120   }
121 
122   bool doInitialization(Module &M) override;
123   bool runOnFunction(Function &F) override;
124 };
125 
126 } // end anonymous namespace
127 
128 INITIALIZE_PASS_BEGIN(AMDGPURewriteOutArguments, DEBUG_TYPE,
129                       "AMDGPU Rewrite Out Arguments", false, false)
130 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
131 INITIALIZE_PASS_END(AMDGPURewriteOutArguments, DEBUG_TYPE,
132                     "AMDGPU Rewrite Out Arguments", false, false)
133 
134 char AMDGPURewriteOutArguments::ID = 0;
135 
136 bool AMDGPURewriteOutArguments::checkArgumentUses(Value &Arg) const {
137   const int MaxUses = 10;
138   int UseCount = 0;
139 
140   for (Use &U : Arg.uses()) {
141     StoreInst *SI = dyn_cast<StoreInst>(U.getUser());
142     if (UseCount > MaxUses)
143       return false;
144 
145     if (!SI) {
146       auto *BCI = dyn_cast<BitCastInst>(U.getUser());
147       if (!BCI || !BCI->hasOneUse())
148         return false;
149 
150       // We don't handle multiple stores currently, so stores to aggregate
151       // pointers aren't worth the trouble since they are canonically split up.
152       Type *DestEltTy = BCI->getType()->getPointerElementType();
153       if (DestEltTy->isAggregateType())
154         return false;
155 
156       // We could handle these if we had a convenient way to bitcast between
157       // them.
158       Type *SrcEltTy = Arg.getType()->getPointerElementType();
159       if (SrcEltTy->isArrayTy())
160         return false;
161 
162       // Special case handle structs with single members. It is useful to handle
163       // some casts between structs and non-structs, but we can't bitcast
164       // directly between them.  directly bitcast between them.  Blender uses
165       // some casts that look like { <3 x float> }* to <4 x float>*
166       if ((SrcEltTy->isStructTy() && (SrcEltTy->getStructNumElements() != 1)))
167         return false;
168 
169       // Clang emits OpenCL 3-vector type accesses with a bitcast to the
170       // equivalent 4-element vector and accesses that, and we're looking for
171       // this pointer cast.
172       if (DL->getTypeAllocSize(SrcEltTy) != DL->getTypeAllocSize(DestEltTy))
173         return false;
174 
175       return checkArgumentUses(*BCI);
176     }
177 
178     if (!SI->isSimple() ||
179         U.getOperandNo() != StoreInst::getPointerOperandIndex())
180       return false;
181 
182     ++UseCount;
183   }
184 
185   // Skip unused arguments.
186   return UseCount > 0;
187 }
188 
189 bool AMDGPURewriteOutArguments::isOutArgumentCandidate(Argument &Arg) const {
190   const unsigned MaxOutArgSizeBytes = 4 * MaxNumRetRegs;
191   PointerType *ArgTy = dyn_cast<PointerType>(Arg.getType());
192 
193   // TODO: It might be useful for any out arguments, not just privates.
194   if (!ArgTy || (ArgTy->getAddressSpace() != DL->getAllocaAddrSpace() &&
195                  !AnyAddressSpace) ||
196       Arg.hasByValAttr() || Arg.hasStructRetAttr() ||
197       DL->getTypeStoreSize(ArgTy->getPointerElementType()) > MaxOutArgSizeBytes) {
198     return false;
199   }
200 
201   return checkArgumentUses(Arg);
202 }
203 
204 bool AMDGPURewriteOutArguments::doInitialization(Module &M) {
205   DL = &M.getDataLayout();
206   return false;
207 }
208 
209 #ifndef NDEBUG
210 bool AMDGPURewriteOutArguments::isVec3ToVec4Shuffle(Type *Ty0, Type* Ty1) const {
211   auto *VT0 = dyn_cast<FixedVectorType>(Ty0);
212   auto *VT1 = dyn_cast<FixedVectorType>(Ty1);
213   if (!VT0 || !VT1)
214     return false;
215 
216   if (VT0->getNumElements() != 3 ||
217       VT1->getNumElements() != 4)
218     return false;
219 
220   return DL->getTypeSizeInBits(VT0->getElementType()) ==
221          DL->getTypeSizeInBits(VT1->getElementType());
222 }
223 #endif
224 
225 bool AMDGPURewriteOutArguments::runOnFunction(Function &F) {
226   if (skipFunction(F))
227     return false;
228 
229   // TODO: Could probably handle variadic functions.
230   if (F.isVarArg() || F.hasStructRetAttr() ||
231       AMDGPU::isEntryFunctionCC(F.getCallingConv()))
232     return false;
233 
234   MDA = &getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
235 
236   unsigned ReturnNumRegs = 0;
237   SmallSet<int, 4> OutArgIndexes;
238   SmallVector<Type *, 4> ReturnTypes;
239   Type *RetTy = F.getReturnType();
240   if (!RetTy->isVoidTy()) {
241     ReturnNumRegs = DL->getTypeStoreSize(RetTy) / 4;
242 
243     if (ReturnNumRegs >= MaxNumRetRegs)
244       return false;
245 
246     ReturnTypes.push_back(RetTy);
247   }
248 
249   SmallVector<Argument *, 4> OutArgs;
250   for (Argument &Arg : F.args()) {
251     if (isOutArgumentCandidate(Arg)) {
252       LLVM_DEBUG(dbgs() << "Found possible out argument " << Arg
253                         << " in function " << F.getName() << '\n');
254       OutArgs.push_back(&Arg);
255     }
256   }
257 
258   if (OutArgs.empty())
259     return false;
260 
261   using ReplacementVec = SmallVector<std::pair<Argument *, Value *>, 4>;
262 
263   DenseMap<ReturnInst *, ReplacementVec> Replacements;
264 
265   SmallVector<ReturnInst *, 4> Returns;
266   for (BasicBlock &BB : F) {
267     if (ReturnInst *RI = dyn_cast<ReturnInst>(&BB.back()))
268       Returns.push_back(RI);
269   }
270 
271   if (Returns.empty())
272     return false;
273 
274   bool Changing;
275 
276   do {
277     Changing = false;
278 
279     // Keep retrying if we are able to successfully eliminate an argument. This
280     // helps with cases with multiple arguments which may alias, such as in a
281     // sincos implemntation. If we have 2 stores to arguments, on the first
282     // attempt the MDA query will succeed for the second store but not the
283     // first. On the second iteration we've removed that out clobbering argument
284     // (by effectively moving it into another function) and will find the second
285     // argument is OK to move.
286     for (Argument *OutArg : OutArgs) {
287       bool ThisReplaceable = true;
288       SmallVector<std::pair<ReturnInst *, StoreInst *>, 4> ReplaceableStores;
289 
290       Type *ArgTy = OutArg->getType()->getPointerElementType();
291 
292       // Skip this argument if converting it will push us over the register
293       // count to return limit.
294 
295       // TODO: This is an approximation. When legalized this could be more. We
296       // can ask TLI for exactly how many.
297       unsigned ArgNumRegs = DL->getTypeStoreSize(ArgTy) / 4;
298       if (ArgNumRegs + ReturnNumRegs > MaxNumRetRegs)
299         continue;
300 
301       // An argument is convertible only if all exit blocks are able to replace
302       // it.
303       for (ReturnInst *RI : Returns) {
304         BasicBlock *BB = RI->getParent();
305 
306         MemDepResult Q = MDA->getPointerDependencyFrom(MemoryLocation(OutArg),
307                                                        true, BB->end(), BB, RI);
308         StoreInst *SI = nullptr;
309         if (Q.isDef())
310           SI = dyn_cast<StoreInst>(Q.getInst());
311 
312         if (SI) {
313           LLVM_DEBUG(dbgs() << "Found out argument store: " << *SI << '\n');
314           ReplaceableStores.emplace_back(RI, SI);
315         } else {
316           ThisReplaceable = false;
317           break;
318         }
319       }
320 
321       if (!ThisReplaceable)
322         continue; // Try the next argument candidate.
323 
324       for (std::pair<ReturnInst *, StoreInst *> Store : ReplaceableStores) {
325         Value *ReplVal = Store.second->getValueOperand();
326 
327         auto &ValVec = Replacements[Store.first];
328         if (llvm::find_if(ValVec,
329               [OutArg](const std::pair<Argument *, Value *> &Entry) {
330                  return Entry.first == OutArg;}) != ValVec.end()) {
331           LLVM_DEBUG(dbgs()
332                      << "Saw multiple out arg stores" << *OutArg << '\n');
333           // It is possible to see stores to the same argument multiple times,
334           // but we expect these would have been optimized out already.
335           ThisReplaceable = false;
336           break;
337         }
338 
339         ValVec.emplace_back(OutArg, ReplVal);
340         Store.second->eraseFromParent();
341       }
342 
343       if (ThisReplaceable) {
344         ReturnTypes.push_back(ArgTy);
345         OutArgIndexes.insert(OutArg->getArgNo());
346         ++NumOutArgumentsReplaced;
347         Changing = true;
348       }
349     }
350   } while (Changing);
351 
352   if (Replacements.empty())
353     return false;
354 
355   LLVMContext &Ctx = F.getParent()->getContext();
356   StructType *NewRetTy = StructType::create(Ctx, ReturnTypes, F.getName());
357 
358   FunctionType *NewFuncTy = FunctionType::get(NewRetTy,
359                                               F.getFunctionType()->params(),
360                                               F.isVarArg());
361 
362   LLVM_DEBUG(dbgs() << "Computed new return type: " << *NewRetTy << '\n');
363 
364   Function *NewFunc = Function::Create(NewFuncTy, Function::PrivateLinkage,
365                                        F.getName() + ".body");
366   F.getParent()->getFunctionList().insert(F.getIterator(), NewFunc);
367   NewFunc->copyAttributesFrom(&F);
368   NewFunc->setComdat(F.getComdat());
369 
370   // We want to preserve the function and param attributes, but need to strip
371   // off any return attributes, e.g. zeroext doesn't make sense with a struct.
372   NewFunc->stealArgumentListFrom(F);
373 
374   AttrBuilder RetAttrs;
375   RetAttrs.addAttribute(Attribute::SExt);
376   RetAttrs.addAttribute(Attribute::ZExt);
377   RetAttrs.addAttribute(Attribute::NoAlias);
378   NewFunc->removeAttributes(AttributeList::ReturnIndex, RetAttrs);
379   // TODO: How to preserve metadata?
380 
381   // Move the body of the function into the new rewritten function, and replace
382   // this function with a stub.
383   NewFunc->getBasicBlockList().splice(NewFunc->begin(), F.getBasicBlockList());
384 
385   for (std::pair<ReturnInst *, ReplacementVec> &Replacement : Replacements) {
386     ReturnInst *RI = Replacement.first;
387     IRBuilder<> B(RI);
388     B.SetCurrentDebugLocation(RI->getDebugLoc());
389 
390     int RetIdx = 0;
391     Value *NewRetVal = UndefValue::get(NewRetTy);
392 
393     Value *RetVal = RI->getReturnValue();
394     if (RetVal)
395       NewRetVal = B.CreateInsertValue(NewRetVal, RetVal, RetIdx++);
396 
397     for (std::pair<Argument *, Value *> ReturnPoint : Replacement.second) {
398       Argument *Arg = ReturnPoint.first;
399       Value *Val = ReturnPoint.second;
400       Type *EltTy = Arg->getType()->getPointerElementType();
401       if (Val->getType() != EltTy) {
402         Type *EffectiveEltTy = EltTy;
403         if (StructType *CT = dyn_cast<StructType>(EltTy)) {
404           assert(CT->getNumElements() == 1);
405           EffectiveEltTy = CT->getElementType(0);
406         }
407 
408         if (DL->getTypeSizeInBits(EffectiveEltTy) !=
409             DL->getTypeSizeInBits(Val->getType())) {
410           assert(isVec3ToVec4Shuffle(EffectiveEltTy, Val->getType()));
411           Val = B.CreateShuffleVector(Val, UndefValue::get(Val->getType()),
412                                       ArrayRef<int>{0, 1, 2});
413         }
414 
415         Val = B.CreateBitCast(Val, EffectiveEltTy);
416 
417         // Re-create single element composite.
418         if (EltTy != EffectiveEltTy)
419           Val = B.CreateInsertValue(UndefValue::get(EltTy), Val, 0);
420       }
421 
422       NewRetVal = B.CreateInsertValue(NewRetVal, Val, RetIdx++);
423     }
424 
425     if (RetVal)
426       RI->setOperand(0, NewRetVal);
427     else {
428       B.CreateRet(NewRetVal);
429       RI->eraseFromParent();
430     }
431   }
432 
433   SmallVector<Value *, 16> StubCallArgs;
434   for (Argument &Arg : F.args()) {
435     if (OutArgIndexes.count(Arg.getArgNo())) {
436       // It's easier to preserve the type of the argument list. We rely on
437       // DeadArgumentElimination to take care of these.
438       StubCallArgs.push_back(UndefValue::get(Arg.getType()));
439     } else {
440       StubCallArgs.push_back(&Arg);
441     }
442   }
443 
444   BasicBlock *StubBB = BasicBlock::Create(Ctx, "", &F);
445   IRBuilder<> B(StubBB);
446   CallInst *StubCall = B.CreateCall(NewFunc, StubCallArgs);
447 
448   int RetIdx = RetTy->isVoidTy() ? 0 : 1;
449   for (Argument &Arg : F.args()) {
450     if (!OutArgIndexes.count(Arg.getArgNo()))
451       continue;
452 
453     PointerType *ArgType = cast<PointerType>(Arg.getType());
454 
455     auto *EltTy = ArgType->getElementType();
456     const auto Align =
457         DL->getValueOrABITypeAlignment(Arg.getParamAlign(), EltTy);
458 
459     Value *Val = B.CreateExtractValue(StubCall, RetIdx++);
460     Type *PtrTy = Val->getType()->getPointerTo(ArgType->getAddressSpace());
461 
462     // We can peek through bitcasts, so the type may not match.
463     Value *PtrVal = B.CreateBitCast(&Arg, PtrTy);
464 
465     B.CreateAlignedStore(Val, PtrVal, Align);
466   }
467 
468   if (!RetTy->isVoidTy()) {
469     B.CreateRet(B.CreateExtractValue(StubCall, 0));
470   } else {
471     B.CreateRetVoid();
472   }
473 
474   // The function is now a stub we want to inline.
475   F.addFnAttr(Attribute::AlwaysInline);
476 
477   ++NumOutArgumentFunctionsReplaced;
478   return true;
479 }
480 
481 FunctionPass *llvm::createAMDGPURewriteOutArgumentsPass() {
482   return new AMDGPURewriteOutArguments();
483 }
484