3 files changed, 128 insertions, 1 deletions
diff --git a/lib/Target/X86/X86.td b/lib/Target/X86/X86.td
index bf095017f8..ec9a8230bd 100644
--- a/lib/Target/X86/X86.td
+++ b/lib/Target/X86/X86.td
@@ -215,7 +215,7 @@ def : ProcessorModel<"core-avx-i", SandyBridgeModel,
                       FeatureF16C, FeatureFSGSBase]>;
 
 // Haswell
-def : ProcessorModel<"core-avx2", SandyBridgeModel,
+def : ProcessorModel<"core-avx2", HaswellModel,
                      [FeatureAVX2, FeatureCMPXCHG16B, FeatureFastUAMem,
                       FeaturePOPCNT, FeatureAES, FeaturePCLMUL, FeatureRDRAND,
                       FeatureF16C, FeatureFSGSBase, FeatureMOVBE, FeatureLZCNT,
diff --git a/lib/Target/X86/X86SchedHaswell.td b/lib/Target/X86/X86SchedHaswell.td
new file mode 100644
index 0000000000..b3eb460d3c
--- /dev/null
+++ b/lib/Target/X86/X86SchedHaswell.td
@@ -0,0 +1,126 @@
+//=- X86SchedHaswell.td - X86 Haswell Scheduling -------------*- tablegen -*-=//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the machine model for Haswell to support instruction
+// scheduling and other instruction cost heuristics.
+//
+//===----------------------------------------------------------------------===//
+
+def HaswellModel : SchedMachineModel {
+  // All x86 instructions are modeled as a single micro-op, and HW can decode 4
+  // instructions per cycle.
+  let IssueWidth = 4;
+  let MinLatency = 0; // 0 = Out-of-order execution.
+  let LoadLatency = 4;
+  let ILPWindow = 40;
+  let MispredictPenalty = 16;
+}
+
+let SchedModel = HaswellModel in {
+
+// Haswell can issue micro-ops to 8 different ports in one cycle.
+
+// Ports 0, 1, 5, 6 and 7 handle all computation.
+// Port 4 gets the data half of stores. Store data can be available later than
+// the store address, but since we don't model the latency of stores, we can
+// ignore that.
+// Ports 2 and 3 are identical. They handle loads and the address half of
+// stores. Port 7 can handle address calculations.
+def HWPort0 : ProcResource<1>;
+def HWPort1 : ProcResource<1>;
+def HWPort2 : ProcResource<1>;
+def HWPort3 : ProcResource<1>;
+def HWPort4 : ProcResource<1>;
+def HWPort5 : ProcResource<1>;
+def HWPort6 : ProcResource<1>;
+def HWPort7 : ProcResource<1>;
+
+// Many micro-ops are capable of issuing on multiple ports.
+def HWPort23  : ProcResGroup<[HWPort2, HWPort3]>;
+def HWPort237 : ProcResGroup<[HWPort2, HWPort3, HWPort7]>;
+def HWPort05  : ProcResGroup<[HWPort0, HWPort5]>;
+def HWPort056 : ProcResGroup<[HWPort0, HWPort5, HWPort6]>;
+def HWPort15  : ProcResGroup<[HWPort1, HWPort5]>;
+def HWPort015 : ProcResGroup<[HWPort0, HWPort1, HWPort5]>;
+def HWPort0156: ProcResGroup<[HWPort0, HWPort1, HWPort5, HWPort6]>;
+
+// Integer division issued on port 0, but uses the non-pipelined divider.
+def HWDivider : ProcResource<1> { let Buffered = 0; }
+
+// Loads are 4 cycles, so ReadAfterLd registers needn't be available until 4
+// cycles after the memory operand.
+def : ReadAdvance<ReadAfterLd, 4>;
+
+// Many SchedWrites are defined in pairs with and without a folded load.
+// Instructions with folded loads are usually micro-fused, so they only appear
+// as two micro-ops when queued in the reservation station.
+// This multiclass defines the resource usage for variants with and without
+// folded loads.
+multiclass HWWriteResPair<X86FoldableSchedWrite SchedRW,
+                          ProcResourceKind ExePort,
+                          int Lat> {
+  // Register variant is using a single cycle on ExePort.
+  def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }
+
+  // Memory variant also uses a cycle on port 2/3 and adds 4 cycles to the
+  // latency.
+  def : WriteRes<SchedRW.Folded, [HWPort23, ExePort]> {
+     let Latency = !add(Lat, 4);
+  }
+}
+
+// A folded store needs a cycle on port 4 for the store data, but it does not
+// need an extra port 2/3 cycle to recompute the address.
+def : WriteRes<WriteRMW, [HWPort4]>;
+
+def : WriteRes<WriteStore, [HWPort237, HWPort4]>;
+def : WriteRes<WriteLoad,  [HWPort23]> { let Latency = 4; }
+def : WriteRes<WriteMove,  [HWPort0156]>;
+def : WriteRes<WriteZero,  []>;
+
+defm : HWWriteResPair<WriteALU,   HWPort0156, 1>;
+defm : HWWriteResPair<WriteIMul,  HWPort1,   3>;
+defm : HWWriteResPair<WriteShift, HWPort056,  1>;
+defm : HWWriteResPair<WriteJump,  HWPort5,   1>;
+
+// This is for simple LEAs with one or two input operands.
+// The complex ones can only execute on port 1, and they require two cycles on
+// the port to read all inputs. We don't model that.
+def : WriteRes<WriteLEA, [HWPort15]>;
+
+// This is quite rough, latency depends on the dividend.
+def : WriteRes<WriteIDiv, [HWPort0, HWDivider]> {
+  let Latency = 25;
+  let ResourceCycles = [1, 10];
+}
+def : WriteRes<WriteIDivLd, [HWPort23, HWPort0, HWDivider]> {
+  let Latency = 29;
+  let ResourceCycles = [1, 1, 10];
+}
+
+// Scalar and vector floating point.
+defm : HWWriteResPair<WriteFAdd,   HWPort1, 3>;
+defm : HWWriteResPair<WriteFMul,   HWPort0, 5>;
+defm : HWWriteResPair<WriteFDiv,   HWPort0, 12>; // 10-14 cycles.
+defm : HWWriteResPair<WriteFRcp,   HWPort0, 5>;
+defm : HWWriteResPair<WriteFSqrt,  HWPort0, 15>;
+defm : HWWriteResPair<WriteCvtF2I, HWPort1, 3>;
+defm : HWWriteResPair<WriteCvtI2F, HWPort1, 4>;
+defm : HWWriteResPair<WriteCvtF2F, HWPort1, 3>;
+
+// Vector integer operations.
+defm : HWWriteResPair<WriteVecShift, HWPort05,  1>;
+defm : HWWriteResPair<WriteVecLogic, HWPort015, 1>;
+defm : HWWriteResPair<WriteVecALU,   HWPort15,  1>;
+defm : HWWriteResPair<WriteVecIMul,  HWPort0,   5>;
+defm : HWWriteResPair<WriteShuffle,  HWPort15,  1>;
+
+def : WriteRes<WriteSystem,     [HWPort0156]> { let Latency = 100; }
+def : WriteRes<WriteMicrocoded, [HWPort0156]> { let Latency = 100; }
+} // SchedModel
diff --git a/lib/Target/X86/X86Schedule.td b/lib/Target/X86/X86Schedule.td
index 782e771fd0..0159500ac0 100644
--- a/lib/Target/X86/X86Schedule.td
+++ b/lib/Target/X86/X86Schedule.td
@@ -566,3 +566,4 @@ def GenericModel : SchedMachineModel {
 
 include "X86ScheduleAtom.td"
 include "X86SchedSandyBridge.td"
+include "X86SchedHaswell.td"