• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1//=- X86SchedHaswell.td - X86 Haswell Scheduling -------------*- tablegen -*-=//
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 file defines the machine model for Haswell to support instruction
11// scheduling and other instruction cost heuristics.
12//
13//===----------------------------------------------------------------------===//
14
15def HaswellModel : SchedMachineModel {
16  // All x86 instructions are modeled as a single micro-op, and HW can decode 4
17  // instructions per cycle.
18  let IssueWidth = 4;
19  let MicroOpBufferSize = 192; // Based on the reorder buffer.
20  let LoadLatency = 4;
21  let MispredictPenalty = 16;
22}
23
24let SchedModel = HaswellModel in {
25
26// Haswell can issue micro-ops to 8 different ports in one cycle.
27
28// Ports 0, 1, 5, 6 and 7 handle all computation.
29// Port 4 gets the data half of stores. Store data can be available later than
30// the store address, but since we don't model the latency of stores, we can
31// ignore that.
32// Ports 2 and 3 are identical. They handle loads and the address half of
33// stores. Port 7 can handle address calculations.
34def HWPort0 : ProcResource<1>;
35def HWPort1 : ProcResource<1>;
36def HWPort2 : ProcResource<1>;
37def HWPort3 : ProcResource<1>;
38def HWPort4 : ProcResource<1>;
39def HWPort5 : ProcResource<1>;
40def HWPort6 : ProcResource<1>;
41def HWPort7 : ProcResource<1>;
42
43// Many micro-ops are capable of issuing on multiple ports.
44def HWPort23  : ProcResGroup<[HWPort2, HWPort3]>;
45def HWPort237 : ProcResGroup<[HWPort2, HWPort3, HWPort7]>;
46def HWPort05  : ProcResGroup<[HWPort0, HWPort5]>;
47def HWPort056 : ProcResGroup<[HWPort0, HWPort5, HWPort6]>;
48def HWPort15  : ProcResGroup<[HWPort1, HWPort5]>;
49def HWPort015 : ProcResGroup<[HWPort0, HWPort1, HWPort5]>;
50def HWPort0156: ProcResGroup<[HWPort0, HWPort1, HWPort5, HWPort6]>;
51
52// 60 Entry Unified Scheduler
53def HWPortAny : ProcResGroup<[HWPort0, HWPort1, HWPort2, HWPort3, HWPort4,
54                              HWPort5, HWPort6, HWPort7]> {
55  let BufferSize=60;
56}
57
58// Integer division issued on port 0.
59def HWDivider : ProcResource<1>;
60
61// Loads are 4 cycles, so ReadAfterLd registers needn't be available until 4
62// cycles after the memory operand.
63def : ReadAdvance<ReadAfterLd, 4>;
64
65// Many SchedWrites are defined in pairs with and without a folded load.
66// Instructions with folded loads are usually micro-fused, so they only appear
67// as two micro-ops when queued in the reservation station.
68// This multiclass defines the resource usage for variants with and without
69// folded loads.
70multiclass HWWriteResPair<X86FoldableSchedWrite SchedRW,
71                          ProcResourceKind ExePort,
72                          int Lat> {
73  // Register variant is using a single cycle on ExePort.
74  def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }
75
76  // Memory variant also uses a cycle on port 2/3 and adds 4 cycles to the
77  // latency.
78  def : WriteRes<SchedRW.Folded, [HWPort23, ExePort]> {
79     let Latency = !add(Lat, 4);
80  }
81}
82
83// A folded store needs a cycle on port 4 for the store data, but it does not
84// need an extra port 2/3 cycle to recompute the address.
85def : WriteRes<WriteRMW, [HWPort4]>;
86
87def : WriteRes<WriteStore, [HWPort237, HWPort4]>;
88def : WriteRes<WriteLoad,  [HWPort23]> { let Latency = 4; }
89def : WriteRes<WriteMove,  [HWPort0156]>;
90def : WriteRes<WriteZero,  []>;
91
92defm : HWWriteResPair<WriteALU,   HWPort0156, 1>;
93defm : HWWriteResPair<WriteIMul,  HWPort1,   3>;
94def  : WriteRes<WriteIMulH, []> { let Latency = 3; }
95defm : HWWriteResPair<WriteShift, HWPort056,  1>;
96defm : HWWriteResPair<WriteJump,  HWPort5,   1>;
97
98// This is for simple LEAs with one or two input operands.
99// The complex ones can only execute on port 1, and they require two cycles on
100// the port to read all inputs. We don't model that.
101def : WriteRes<WriteLEA, [HWPort15]>;
102
103// This is quite rough, latency depends on the dividend.
104def : WriteRes<WriteIDiv, [HWPort0, HWDivider]> {
105  let Latency = 25;
106  let ResourceCycles = [1, 10];
107}
108def : WriteRes<WriteIDivLd, [HWPort23, HWPort0, HWDivider]> {
109  let Latency = 29;
110  let ResourceCycles = [1, 1, 10];
111}
112
113// Scalar and vector floating point.
114defm : HWWriteResPair<WriteFAdd,   HWPort1, 3>;
115defm : HWWriteResPair<WriteFMul,   HWPort0, 5>;
116defm : HWWriteResPair<WriteFDiv,   HWPort0, 12>; // 10-14 cycles.
117defm : HWWriteResPair<WriteFRcp,   HWPort0, 5>;
118defm : HWWriteResPair<WriteFSqrt,  HWPort0, 15>;
119defm : HWWriteResPair<WriteCvtF2I, HWPort1, 3>;
120defm : HWWriteResPair<WriteCvtI2F, HWPort1, 4>;
121defm : HWWriteResPair<WriteCvtF2F, HWPort1, 3>;
122
123// Vector integer operations.
124defm : HWWriteResPair<WriteVecShift, HWPort05,  1>;
125defm : HWWriteResPair<WriteVecLogic, HWPort015, 1>;
126defm : HWWriteResPair<WriteVecALU,   HWPort15,  1>;
127defm : HWWriteResPair<WriteVecIMul,  HWPort0,   5>;
128defm : HWWriteResPair<WriteShuffle,  HWPort15,  1>;
129
130def : WriteRes<WriteSystem,     [HWPort0156]> { let Latency = 100; }
131def : WriteRes<WriteMicrocoded, [HWPort0156]> { let Latency = 100; }
132} // SchedModel
133