1 //===--------------------- Support.cpp --------------------------*- C++ -*-===//
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 /// \file
10 ///
11 /// This file implements a few helper functions used by various pipeline
12 /// components.
13 ///
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/MCA/Support.h"
17 #include "llvm/MC/MCSchedule.h"
18 
19 namespace llvm {
20 namespace mca {
21 
22 void computeProcResourceMasks(const MCSchedModel &SM,
23                               SmallVectorImpl<uint64_t> &Masks) {
24   unsigned ProcResourceID = 0;
25 
26   // Create a unique bitmask for every processor resource unit.
27   // Skip resource at index 0, since it always references 'InvalidUnit'.
28   Masks.resize(SM.getNumProcResourceKinds());
29   for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) {
30     const MCProcResourceDesc &Desc = *SM.getProcResource(I);
31     if (Desc.SubUnitsIdxBegin)
32       continue;
33     Masks[I] = 1ULL << ProcResourceID;
34     ProcResourceID++;
35   }
36 
37   // Create a unique bitmask for every processor resource group.
38   for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) {
39     const MCProcResourceDesc &Desc = *SM.getProcResource(I);
40     if (!Desc.SubUnitsIdxBegin)
41       continue;
42     Masks[I] = 1ULL << ProcResourceID;
43     for (unsigned U = 0; U < Desc.NumUnits; ++U) {
44       uint64_t OtherMask = Masks[Desc.SubUnitsIdxBegin[U]];
45       Masks[I] |= OtherMask;
46     }
47     ProcResourceID++;
48   }
49 }
50 
51 double computeBlockRThroughput(const MCSchedModel &SM, unsigned DispatchWidth,
52                                unsigned NumMicroOps,
53                                ArrayRef<unsigned> ProcResourceUsage) {
54   // The block throughput is bounded from above by the hardware dispatch
55   // throughput. That is because the DispatchWidth is an upper bound on the
56   // number of opcodes that can be part of a single dispatch group.
57   double Max = static_cast<double>(NumMicroOps) / DispatchWidth;
58 
59   // The block throughput is also limited by the amount of hardware parallelism.
60   // The number of available resource units affects the resource pressure
61   // distribution, as well as how many blocks can be executed every cycle.
62   for (unsigned I = 0, E = SM.getNumProcResourceKinds(); I < E; ++I) {
63     unsigned ResourceCycles = ProcResourceUsage[I];
64     if (!ResourceCycles)
65       continue;
66 
67     const MCProcResourceDesc &MCDesc = *SM.getProcResource(I);
68     double Throughput = static_cast<double>(ResourceCycles) / MCDesc.NumUnits;
69     Max = std::max(Max, Throughput);
70   }
71 
72   // The block reciprocal throughput is computed as the MAX of:
73   //  - (NumMicroOps / DispatchWidth)
74   //  - (NumUnits / ResourceCycles)   for every consumed processor resource.
75   return Max;
76 }
77 
78 } // namespace mca
79 } // namespace llvm
80