LLVM  12.0.0git
NVPTXLowerArgs.cpp
Go to the documentation of this file.
1 //===-- NVPTXLowerArgs.cpp - Lower arguments ------------------------------===//
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 //
10 // Arguments to kernel and device functions are passed via param space,
11 // which imposes certain restrictions:
12 // http://docs.nvidia.com/cuda/parallel-thread-execution/#state-spaces
13 //
14 // Kernel parameters are read-only and accessible only via ld.param
15 // instruction, directly or via a pointer. Pointers to kernel
16 // arguments can't be converted to generic address space.
17 //
18 // Device function parameters are directly accessible via
19 // ld.param/st.param, but taking the address of one returns a pointer
20 // to a copy created in local space which *can't* be used with
21 // ld.param/st.param.
22 //
23 // Copying a byval struct into local memory in IR allows us to enforce
24 // the param space restrictions, gives the rest of IR a pointer w/o
25 // param space restrictions, and gives us an opportunity to eliminate
26 // the copy.
27 //
28 // Pointer arguments to kernel functions need more work to be lowered:
29 //
30 // 1. Convert non-byval pointer arguments of CUDA kernels to pointers in the
31 // global address space. This allows later optimizations to emit
32 // ld.global.*/st.global.* for accessing these pointer arguments. For
33 // example,
34 //
35 // define void @foo(float* %input) {
36 // %v = load float, float* %input, align 4
37 // ...
38 // }
39 //
40 // becomes
41 //
42 // define void @foo(float* %input) {
43 // %input2 = addrspacecast float* %input to float addrspace(1)*
44 // %input3 = addrspacecast float addrspace(1)* %input2 to float*
45 // %v = load float, float* %input3, align 4
46 // ...
47 // }
48 //
49 // Later, NVPTXInferAddressSpaces will optimize it to
50 //
51 // define void @foo(float* %input) {
52 // %input2 = addrspacecast float* %input to float addrspace(1)*
53 // %v = load float, float addrspace(1)* %input2, align 4
54 // ...
55 // }
56 //
57 // 2. Convert pointers in a byval kernel parameter to pointers in the global
58 // address space. As #2, it allows NVPTX to emit more ld/st.global. E.g.,
59 //
60 // struct S {
61 // int *x;
62 // int *y;
63 // };
64 // __global__ void foo(S s) {
65 // int *b = s.y;
66 // // use b
67 // }
68 //
69 // "b" points to the global address space. In the IR level,
70 //
71 // define void @foo({i32*, i32*}* byval %input) {
72 // %b_ptr = getelementptr {i32*, i32*}, {i32*, i32*}* %input, i64 0, i32 1
73 // %b = load i32*, i32** %b_ptr
74 // ; use %b
75 // }
76 //
77 // becomes
78 //
79 // define void @foo({i32*, i32*}* byval %input) {
80 // %b_ptr = getelementptr {i32*, i32*}, {i32*, i32*}* %input, i64 0, i32 1
81 // %b = load i32*, i32** %b_ptr
82 // %b_global = addrspacecast i32* %b to i32 addrspace(1)*
83 // %b_generic = addrspacecast i32 addrspace(1)* %b_global to i32*
84 // ; use %b_generic
85 // }
86 //
87 // TODO: merge this pass with NVPTXInferAddressSpaces so that other passes don't
88 // cancel the addrspacecast pair this pass emits.
89 //===----------------------------------------------------------------------===//
90 
91 #include "NVPTX.h"
92 #include "NVPTXTargetMachine.h"
93 #include "NVPTXUtilities.h"
96 #include "llvm/IR/Function.h"
97 #include "llvm/IR/Instructions.h"
98 #include "llvm/IR/Module.h"
99 #include "llvm/IR/Type.h"
100 #include "llvm/Pass.h"
101 
102 using namespace llvm;
103 
104 namespace llvm {
106 }
107 
108 namespace {
109 class NVPTXLowerArgs : public FunctionPass {
110  bool runOnFunction(Function &F) override;
111 
112  bool runOnKernelFunction(Function &F);
113  bool runOnDeviceFunction(Function &F);
114 
115  // handle byval parameters
116  void handleByValParam(Argument *Arg);
117  // Knowing Ptr must point to the global address space, this function
118  // addrspacecasts Ptr to global and then back to generic. This allows
119  // NVPTXInferAddressSpaces to fold the global-to-generic cast into
120  // loads/stores that appear later.
121  void markPointerAsGlobal(Value *Ptr);
122 
123 public:
124  static char ID; // Pass identification, replacement for typeid
125  NVPTXLowerArgs(const NVPTXTargetMachine *TM = nullptr)
126  : FunctionPass(ID), TM(TM) {}
127  StringRef getPassName() const override {
128  return "Lower pointer arguments of CUDA kernels";
129  }
130 
131 private:
132  const NVPTXTargetMachine *TM;
133 };
134 } // namespace
135 
136 char NVPTXLowerArgs::ID = 1;
137 
138 INITIALIZE_PASS(NVPTXLowerArgs, "nvptx-lower-args",
139  "Lower arguments (NVPTX)", false, false)
140 
141 // =============================================================================
142 // If the function had a byval struct ptr arg, say foo(%struct.x* byval %d),
143 // then add the following instructions to the first basic block:
144 //
145 // %temp = alloca %struct.x, align 8
146 // %tempd = addrspacecast %struct.x* %d to %struct.x addrspace(101)*
147 // %tv = load %struct.x addrspace(101)* %tempd
148 // store %struct.x %tv, %struct.x* %temp, align 8
149 //
150 // The above code allocates some space in the stack and copies the incoming
151 // struct from param space to local space.
152 // Then replace all occurrences of %d by %temp.
153 // =============================================================================
154 void NVPTXLowerArgs::handleByValParam(Argument *Arg) {
155  Function *Func = Arg->getParent();
156  Instruction *FirstInst = &(Func->getEntryBlock().front());
157  PointerType *PType = dyn_cast<PointerType>(Arg->getType());
158 
159  assert(PType && "Expecting pointer type in handleByValParam");
160 
161  Type *StructType = PType->getElementType();
162  const DataLayout &DL = Func->getParent()->getDataLayout();
163  unsigned AS = DL.getAllocaAddrSpace();
164  AllocaInst *AllocA = new AllocaInst(StructType, AS, Arg->getName(), FirstInst);
165  // Set the alignment to alignment of the byval parameter. This is because,
166  // later load/stores assume that alignment, and we are going to replace
167  // the use of the byval parameter with this alloca instruction.
168  AllocA->setAlignment(Func->getParamAlign(Arg->getArgNo())
169  .getValueOr(DL.getPrefTypeAlign(StructType)));
170  Arg->replaceAllUsesWith(AllocA);
171 
172  Value *ArgInParam = new AddrSpaceCastInst(
174  FirstInst);
175  // Be sure to propagate alignment to this load; LLVM doesn't know that NVPTX
176  // addrspacecast preserves alignment. Since params are constant, this load is
177  // definitely not volatile.
178  LoadInst *LI =
179  new LoadInst(StructType, ArgInParam, Arg->getName(),
180  /*isVolatile=*/false, AllocA->getAlign(), FirstInst);
181  new StoreInst(LI, AllocA, FirstInst);
182 }
183 
184 void NVPTXLowerArgs::markPointerAsGlobal(Value *Ptr) {
186  return;
187 
188  // Deciding where to emit the addrspacecast pair.
189  BasicBlock::iterator InsertPt;
190  if (Argument *Arg = dyn_cast<Argument>(Ptr)) {
191  // Insert at the functon entry if Ptr is an argument.
192  InsertPt = Arg->getParent()->getEntryBlock().begin();
193  } else {
194  // Insert right after Ptr if Ptr is an instruction.
195  InsertPt = ++cast<Instruction>(Ptr)->getIterator();
196  assert(InsertPt != InsertPt->getParent()->end() &&
197  "We don't call this function with Ptr being a terminator.");
198  }
199 
200  Instruction *PtrInGlobal = new AddrSpaceCastInst(
203  Ptr->getName(), &*InsertPt);
204  Value *PtrInGeneric = new AddrSpaceCastInst(PtrInGlobal, Ptr->getType(),
205  Ptr->getName(), &*InsertPt);
206  // Replace with PtrInGeneric all uses of Ptr except PtrInGlobal.
207  Ptr->replaceAllUsesWith(PtrInGeneric);
208  PtrInGlobal->setOperand(0, Ptr);
209 }
210 
211 // =============================================================================
212 // Main function for this pass.
213 // =============================================================================
214 bool NVPTXLowerArgs::runOnKernelFunction(Function &F) {
215  if (TM && TM->getDrvInterface() == NVPTX::CUDA) {
216  // Mark pointers in byval structs as global.
217  for (auto &B : F) {
218  for (auto &I : B) {
219  if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
220  if (LI->getType()->isPointerTy()) {
221  Value *UO = getUnderlyingObject(LI->getPointerOperand());
222  if (Argument *Arg = dyn_cast<Argument>(UO)) {
223  if (Arg->hasByValAttr()) {
224  // LI is a load from a pointer within a byval kernel parameter.
225  markPointerAsGlobal(LI);
226  }
227  }
228  }
229  }
230  }
231  }
232  }
233 
234  for (Argument &Arg : F.args()) {
235  if (Arg.getType()->isPointerTy()) {
236  if (Arg.hasByValAttr())
237  handleByValParam(&Arg);
238  else if (TM && TM->getDrvInterface() == NVPTX::CUDA)
239  markPointerAsGlobal(&Arg);
240  }
241  }
242  return true;
243 }
244 
245 // Device functions only need to copy byval args into local memory.
246 bool NVPTXLowerArgs::runOnDeviceFunction(Function &F) {
247  for (Argument &Arg : F.args())
248  if (Arg.getType()->isPointerTy() && Arg.hasByValAttr())
249  handleByValParam(&Arg);
250  return true;
251 }
252 
254  return isKernelFunction(F) ? runOnKernelFunction(F) : runOnDeviceFunction(F);
255 }
256 
257 FunctionPass *
259  return new NVPTXLowerArgs(TM);
260 }
void setAlignment(Align Align)
Definition: Instructions.h:123
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:111
void initializeNVPTXLowerArgsPass(PassRegistry &)
This class represents an incoming formal argument to a Function.
Definition: Argument.h:29
This class represents lattice values for constants.
Definition: AllocatorList.h:23
static PointerType * get(Type *ElementType, unsigned AddressSpace)
This constructs a pointer to an object of the specified type in a numbered address space.
Definition: Type.cpp:669
F(f)
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:711
An instruction for reading from memory.
Definition: Instructions.h:174
This class represents a conversion between pointers from one address space to another.
Type * getPointerElementType() const
Definition: Type.h:374
Class to represent struct types.
Definition: DerivedTypes.h:212
Align getAlign() const
Return the alignment of the memory that is being allocated by the instruction.
Definition: Instructions.h:119
bool isKernelFunction(const Function &F)
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:246
An instruction for storing to memory.
Definition: Instructions.h:303
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:523
Class to represent pointers.
Definition: DerivedTypes.h:655
static bool runOnFunction(Function &F, bool PostInlining)
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:46
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:298
amdgpu Simplify well known AMD library false FunctionCallee Value * Arg
Iterator for intrusive lists based on ilist_node.
Module.h This file contains the declarations for the Module class.
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
FunctionPass * createNVPTXLowerArgsPass(const NVPTXTargetMachine *TM)
INITIALIZE_PASS(NVPTXLowerArgs, "nvptx-lower-args", "Lower arguments (NVPTX)", false, false) void NVPTXLowerArgs
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:295
#define I(x, y, z)
Definition: MD5.cpp:59
Value * getUnderlyingObject(Value *V, unsigned MaxLookup=6)
This method strips off any GEP address adjustments and pointer casts from the specified value,...
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
LLVM Value Representation.
Definition: Value.h:75
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:57
PassRegistry - This class manages the registration and intitialization of the pass subsystem as appli...
Definition: PassRegistry.h:38
Type * getElementType() const
Definition: DerivedTypes.h:674
an instruction to allocate memory on the stack
Definition: Instructions.h:61
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL