1 //===- NVVMIntrRange.cpp - Set !range metadata for NVVM intrinsics --------===//
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 // This pass adds appropriate !range metadata for calls to NVVM
11 // intrinsics that return a limited range of values.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "NVPTX.h"
16 #include "llvm/IR/Constants.h"
17 #include "llvm/IR/InstIterator.h"
18 #include "llvm/IR/Intrinsics.h"
19 #include "llvm/IR/Instructions.h"
20 
21 using namespace llvm;
22 
23 #define DEBUG_TYPE "nvvm-intr-range"
24 
25 namespace llvm { void initializeNVVMIntrRangePass(PassRegistry &); }
26 
27 // Add !range metadata based on limits of given SM variant.
28 static cl::opt<unsigned> NVVMIntrRangeSM("nvvm-intr-range-sm", cl::init(20),
29                                          cl::Hidden, cl::desc("SM variant"));
30 
31 namespace {
32 class NVVMIntrRange : public FunctionPass {
33  private:
34    struct {
35      unsigned x, y, z;
36    } MaxBlockSize, MaxGridSize;
37 
38  public:
39    static char ID;
40    NVVMIntrRange() : NVVMIntrRange(NVVMIntrRangeSM) {}
41    NVVMIntrRange(unsigned int SmVersion) : FunctionPass(ID) {
42      MaxBlockSize.x = 1024;
43      MaxBlockSize.y = 1024;
44      MaxBlockSize.z = 64;
45 
46      MaxGridSize.x = SmVersion >= 30 ? 0x7fffffff : 0xffff;
47      MaxGridSize.y = 0xffff;
48      MaxGridSize.z = 0xffff;
49 
50      initializeNVVMIntrRangePass(*PassRegistry::getPassRegistry());
51    }
52 
53    bool runOnFunction(Function &) override;
54 };
55 }
56 
57 FunctionPass *llvm::createNVVMIntrRangePass(unsigned int SmVersion) {
58   return new NVVMIntrRange(SmVersion);
59 }
60 
61 char NVVMIntrRange::ID = 0;
62 INITIALIZE_PASS(NVVMIntrRange, "nvvm-intr-range",
63                 "Add !range metadata to NVVM intrinsics.", false, false)
64 
65 // Adds the passed-in [Low,High) range information as metadata to the
66 // passed-in call instruction.
67 static bool addRangeMetadata(uint64_t Low, uint64_t High, CallInst *C) {
68   LLVMContext &Context = C->getParent()->getContext();
69   IntegerType *Int32Ty = Type::getInt32Ty(Context);
70   Metadata *LowAndHigh[] = {
71       ConstantAsMetadata::get(ConstantInt::get(Int32Ty, Low)),
72       ConstantAsMetadata::get(ConstantInt::get(Int32Ty, High))};
73   C->setMetadata(LLVMContext::MD_range, MDNode::get(Context, LowAndHigh));
74   return true;
75 }
76 
77 bool NVVMIntrRange::runOnFunction(Function &F) {
78   // Go through the calls in this function.
79   bool Changed = false;
80   for (Instruction &I : instructions(F)) {
81     CallInst *Call = dyn_cast<CallInst>(&I);
82     if (!Call)
83       continue;
84 
85     if (Function *Callee = Call->getCalledFunction()) {
86       switch (Callee->getIntrinsicID()) {
87       // Index within block
88       case Intrinsic::ptx_read_tid_x:
89       case Intrinsic::nvvm_read_ptx_sreg_tid_x:
90         Changed |= addRangeMetadata(0, MaxBlockSize.x, Call);
91         break;
92       case Intrinsic::ptx_read_tid_y:
93       case Intrinsic::nvvm_read_ptx_sreg_tid_y:
94         Changed |= addRangeMetadata(0, MaxBlockSize.y, Call);
95         break;
96       case Intrinsic::ptx_read_tid_z:
97       case Intrinsic::nvvm_read_ptx_sreg_tid_z:
98         Changed |= addRangeMetadata(0, MaxBlockSize.z, Call);
99         break;
100 
101       // Block size
102       case Intrinsic::ptx_read_ntid_x:
103       case Intrinsic::nvvm_read_ptx_sreg_ntid_x:
104         Changed |= addRangeMetadata(1, MaxBlockSize.x+1, Call);
105         break;
106       case Intrinsic::ptx_read_ntid_y:
107       case Intrinsic::nvvm_read_ptx_sreg_ntid_y:
108         Changed |= addRangeMetadata(1, MaxBlockSize.y+1, Call);
109         break;
110       case Intrinsic::ptx_read_ntid_z:
111       case Intrinsic::nvvm_read_ptx_sreg_ntid_z:
112         Changed |= addRangeMetadata(1, MaxBlockSize.z+1, Call);
113         break;
114 
115       // Index within grid
116       case Intrinsic::ptx_read_ctaid_x:
117       case Intrinsic::nvvm_read_ptx_sreg_ctaid_x:
118         Changed |= addRangeMetadata(0, MaxGridSize.x, Call);
119         break;
120       case Intrinsic::ptx_read_ctaid_y:
121       case Intrinsic::nvvm_read_ptx_sreg_ctaid_y:
122         Changed |= addRangeMetadata(0, MaxGridSize.y, Call);
123         break;
124       case Intrinsic::ptx_read_ctaid_z:
125       case Intrinsic::nvvm_read_ptx_sreg_ctaid_z:
126         Changed |= addRangeMetadata(0, MaxGridSize.z, Call);
127         break;
128 
129       // Grid size
130       case Intrinsic::ptx_read_nctaid_x:
131       case Intrinsic::nvvm_read_ptx_sreg_nctaid_x:
132         Changed |= addRangeMetadata(1, MaxGridSize.x+1, Call);
133         break;
134       case Intrinsic::ptx_read_nctaid_y:
135       case Intrinsic::nvvm_read_ptx_sreg_nctaid_y:
136         Changed |= addRangeMetadata(1, MaxGridSize.y+1, Call);
137         break;
138       case Intrinsic::ptx_read_nctaid_z:
139       case Intrinsic::nvvm_read_ptx_sreg_nctaid_z:
140         Changed |= addRangeMetadata(1, MaxGridSize.z+1, Call);
141         break;
142 
143       // warp size is constant 32.
144       case Intrinsic::nvvm_read_ptx_sreg_warpsize:
145         Changed |= addRangeMetadata(32, 32+1, Call);
146         break;
147 
148       // Lane ID is [0..warpsize)
149       case Intrinsic::ptx_read_laneid:
150         Changed |= addRangeMetadata(0, 32, Call);
151         break;
152 
153       default:
154         break;
155       }
156     }
157   }
158 
159   return Changed;
160 }
161