Home | History | Annotate | Download | only in NVPTX 
      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::nvvm_read_ptx_sreg_tid_x:
     89         Changed |= addRangeMetadata(0, MaxBlockSize.x, Call);
     90         break;
     91       case Intrinsic::nvvm_read_ptx_sreg_tid_y:
     92         Changed |= addRangeMetadata(0, MaxBlockSize.y, Call);
     93         break;
     94       case Intrinsic::nvvm_read_ptx_sreg_tid_z:
     95         Changed |= addRangeMetadata(0, MaxBlockSize.z, Call);
     96         break;
     97 
     98       // Block size
     99       case Intrinsic::nvvm_read_ptx_sreg_ntid_x:
    100         Changed |= addRangeMetadata(1, MaxBlockSize.x+1, Call);
    101         break;
    102       case Intrinsic::nvvm_read_ptx_sreg_ntid_y:
    103         Changed |= addRangeMetadata(1, MaxBlockSize.y+1, Call);
    104         break;
    105       case Intrinsic::nvvm_read_ptx_sreg_ntid_z:
    106         Changed |= addRangeMetadata(1, MaxBlockSize.z+1, Call);
    107         break;
    108 
    109       // Index within grid
    110       case Intrinsic::nvvm_read_ptx_sreg_ctaid_x:
    111         Changed |= addRangeMetadata(0, MaxGridSize.x, Call);
    112         break;
    113       case Intrinsic::nvvm_read_ptx_sreg_ctaid_y:
    114         Changed |= addRangeMetadata(0, MaxGridSize.y, Call);
    115         break;
    116       case Intrinsic::nvvm_read_ptx_sreg_ctaid_z:
    117         Changed |= addRangeMetadata(0, MaxGridSize.z, Call);
    118         break;
    119 
    120       // Grid size
    121       case Intrinsic::nvvm_read_ptx_sreg_nctaid_x:
    122         Changed |= addRangeMetadata(1, MaxGridSize.x+1, Call);
    123         break;
    124       case Intrinsic::nvvm_read_ptx_sreg_nctaid_y:
    125         Changed |= addRangeMetadata(1, MaxGridSize.y+1, Call);
    126         break;
    127       case Intrinsic::nvvm_read_ptx_sreg_nctaid_z:
    128         Changed |= addRangeMetadata(1, MaxGridSize.z+1, Call);
    129         break;
    130 
    131       // warp size is constant 32.
    132       case Intrinsic::nvvm_read_ptx_sreg_warpsize:
    133         Changed |= addRangeMetadata(32, 32+1, Call);
    134         break;
    135 
    136       // Lane ID is [0..warpsize)
    137       case Intrinsic::nvvm_read_ptx_sreg_laneid:
    138         Changed |= addRangeMetadata(0, 32, Call);
    139         break;
    140 
    141       default:
    142         break;
    143       }
    144     }
    145   }
    146 
    147   return Changed;
    148 }
    149