1
0
rocket-chip/src/main/scala/rocket/RocketCore.scala

812 lines
35 KiB
Scala
Raw Normal View History

// See LICENSE.Berkeley for license details.
// See LICENSE.SiFive for license details.
package freechips.rocketchip.rocket
import Chisel._
import Chisel.ImplicitConversions._
import chisel3.core.withReset
import freechips.rocketchip.config.Parameters
import freechips.rocketchip.tile._
import freechips.rocketchip.util._
import freechips.rocketchip.util.property._
import scala.collection.immutable.ListMap
import scala.collection.mutable.ArrayBuffer
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
case class RocketCoreParams(
2017-03-02 01:47:10 +01:00
bootFreqHz: BigInt = 0,
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
useVM: Boolean = true,
useUser: Boolean = false,
useDebug: Boolean = true,
useAtomics: Boolean = true,
useAtomicsOnlyForIO: Boolean = false,
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
useCompressed: Boolean = true,
nLocalInterrupts: Int = 0,
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
nBreakpoints: Int = 1,
2017-03-24 23:55:51 +01:00
nPMPs: Int = 8,
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
nPerfCounters: Int = 0,
haveBasicCounters: Boolean = true,
2017-09-21 04:16:34 +02:00
misaWritable: Boolean = true,
2017-07-06 08:53:52 +02:00
nL2TLBEntries: Int = 0,
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
mtvecInit: Option[BigInt] = Some(BigInt(0)),
mtvecWritable: Boolean = true,
fastLoadWord: Boolean = true,
fastLoadByte: Boolean = false,
2017-09-16 03:49:40 +02:00
tileControlAddr: Option[BigInt] = None,
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
mulDiv: Option[MulDivParams] = Some(MulDivParams()),
fpu: Option[FPUParams] = Some(FPUParams())
) extends CoreParams {
val fetchWidth: Int = if (useCompressed) 2 else 1
// fetchWidth doubled, but coreInstBytes halved, for RVC:
val decodeWidth: Int = fetchWidth / (if (useCompressed) 2 else 1)
val retireWidth: Int = 1
val instBits: Int = if (useCompressed) 16 else 32
2015-11-25 03:27:07 +01:00
}
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
trait HasRocketCoreParameters extends HasCoreParameters {
val rocketParams: RocketCoreParams = tileParams.core.asInstanceOf[RocketCoreParams]
2015-11-25 03:27:07 +01:00
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
val fastLoadWord = rocketParams.fastLoadWord
val fastLoadByte = rocketParams.fastLoadByte
val mulDivParams = rocketParams.mulDiv.getOrElse(MulDivParams()) // TODO ask andrew about this
require(!fastLoadByte || fastLoadWord)
2015-11-25 03:27:07 +01:00
}
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
class Rocket(implicit p: Parameters) extends CoreModule()(p)
with HasRocketCoreParameters
with HasCoreIO {
2017-03-09 09:28:19 +01:00
// performance counters
def pipelineIDToWB[T <: Data](x: T): T =
RegEnable(RegEnable(RegEnable(x, !ctrl_killd), ex_pc_valid), mem_pc_valid)
val perfEvents = new EventSets(Seq(
new EventSet((mask, hits) => Mux(mask(0), wb_xcpt, wb_valid && pipelineIDToWB((mask & hits).orR)), Seq(
("exception", () => false.B),
("load", () => id_ctrl.mem && id_ctrl.mem_cmd === M_XRD && !id_ctrl.fp),
("store", () => id_ctrl.mem && id_ctrl.mem_cmd === M_XWR && !id_ctrl.fp),
("amo", () => Bool(usingAtomics) && id_ctrl.mem && (isAMO(id_ctrl.mem_cmd) || id_ctrl.mem_cmd.isOneOf(M_XLR, M_XSC))),
("system", () => id_ctrl.csr =/= CSR.N),
("arith", () => id_ctrl.wxd && !(id_ctrl.jal || id_ctrl.jalr || id_ctrl.mem || id_ctrl.fp || id_ctrl.div || id_ctrl.csr =/= CSR.N)),
("branch", () => id_ctrl.branch),
("jal", () => id_ctrl.jal),
("jalr", () => id_ctrl.jalr))
++ (if (!usingMulDiv) Seq() else Seq(
("mul", () => id_ctrl.div && (id_ctrl.alu_fn & ALU.FN_DIV) =/= ALU.FN_DIV),
("div", () => id_ctrl.div && (id_ctrl.alu_fn & ALU.FN_DIV) === ALU.FN_DIV)))
++ (if (!usingFPU) Seq() else Seq(
("fp load", () => id_ctrl.fp && io.fpu.dec.ldst && io.fpu.dec.wen),
("fp store", () => id_ctrl.fp && io.fpu.dec.ldst && !io.fpu.dec.wen),
("fp add", () => id_ctrl.fp && io.fpu.dec.fma && io.fpu.dec.swap23),
("fp mul", () => id_ctrl.fp && io.fpu.dec.fma && !io.fpu.dec.swap23 && !io.fpu.dec.ren3),
("fp mul-add", () => id_ctrl.fp && io.fpu.dec.fma && io.fpu.dec.ren3),
("fp div/sqrt", () => id_ctrl.fp && (io.fpu.dec.div || io.fpu.dec.sqrt)),
("fp other", () => id_ctrl.fp && !(io.fpu.dec.ldst || io.fpu.dec.fma || io.fpu.dec.div || io.fpu.dec.sqrt))))),
new EventSet((mask, hits) => (mask & hits).orR, Seq(
("load-use interlock", () => id_ex_hazard && ex_ctrl.mem || id_mem_hazard && mem_ctrl.mem || id_wb_hazard && wb_ctrl.mem),
("long-latency interlock", () => id_sboard_hazard),
("csr interlock", () => id_ex_hazard && ex_ctrl.csr =/= CSR.N || id_mem_hazard && mem_ctrl.csr =/= CSR.N || id_wb_hazard && wb_ctrl.csr =/= CSR.N),
2017-07-28 22:14:04 +02:00
("I$ blocked", () => icache_blocked),
2017-03-09 09:28:19 +01:00
("D$ blocked", () => id_ctrl.mem && dcache_blocked),
("branch misprediction", () => take_pc_mem && mem_direction_misprediction),
2017-07-28 22:14:04 +02:00
("control-flow target misprediction", () => take_pc_mem && mem_misprediction && mem_cfi && !mem_direction_misprediction && !icache_blocked),
("flush", () => wb_reg_flush_pipe),
2017-03-09 09:28:19 +01:00
("replay", () => replay_wb))
++ (if (!usingMulDiv) Seq() else Seq(
("mul/div interlock", () => id_ex_hazard && ex_ctrl.div || id_mem_hazard && mem_ctrl.div || id_wb_hazard && wb_ctrl.div)))
++ (if (!usingFPU) Seq() else Seq(
("fp interlock", () => id_ex_hazard && ex_ctrl.fp || id_mem_hazard && mem_ctrl.fp || id_wb_hazard && wb_ctrl.fp || id_ctrl.fp && id_stall_fpu)))),
new EventSet((mask, hits) => (mask & hits).orR, Seq(
("I$ miss", () => io.imem.perf.acquire),
("D$ miss", () => io.dmem.perf.acquire),
("D$ release", () => io.dmem.perf.release),
("ITLB miss", () => io.imem.perf.tlbMiss),
2017-07-25 20:59:53 +02:00
("DTLB miss", () => io.dmem.perf.tlbMiss),
("L2 TLB miss", () => io.ptw.perf.l2miss)))))
2017-03-09 09:28:19 +01:00
val decode_table = {
2016-08-30 00:56:28 +02:00
(if (usingMulDiv) new MDecode +: (xLen > 32).option(new M64Decode).toSeq else Nil) ++:
(if (usingAtomics) new ADecode +: (xLen > 32).option(new A64Decode).toSeq else Nil) ++:
(if (usingFPU) new FDecode +: (xLen > 32).option(new F64Decode).toSeq else Nil) ++:
2016-09-07 08:53:12 +02:00
(if (usingFPU && xLen > 32) Seq(new DDecode, new D64Decode) else Nil) ++:
2016-08-30 00:56:28 +02:00
(usingRoCC.option(new RoCCDecode)) ++:
((xLen > 32).option(new I64Decode)) ++:
(usingVM.option(new SDecode)) ++:
(usingDebug.option(new DebugDecode)) ++:
Seq(new IDecode)
} flatMap(_.table)
val ex_ctrl = Reg(new IntCtrlSigs)
val mem_ctrl = Reg(new IntCtrlSigs)
val wb_ctrl = Reg(new IntCtrlSigs)
val ex_reg_xcpt_interrupt = Reg(Bool())
val ex_reg_valid = Reg(Bool())
2016-07-30 01:36:07 +02:00
val ex_reg_rvc = Reg(Bool())
val ex_reg_btb_resp = Reg(new BTBResp)
val ex_reg_xcpt = Reg(Bool())
val ex_reg_flush_pipe = Reg(Bool())
val ex_reg_load_use = Reg(Bool())
val ex_reg_cause = Reg(UInt())
val ex_reg_replay = Reg(Bool())
val ex_reg_pc = Reg(UInt())
val ex_reg_inst = Reg(Bits())
2017-09-20 07:59:28 +02:00
val ex_reg_raw_inst = Reg(UInt())
val mem_reg_xcpt_interrupt = Reg(Bool())
val mem_reg_valid = Reg(Bool())
2016-07-30 01:36:07 +02:00
val mem_reg_rvc = Reg(Bool())
val mem_reg_btb_resp = Reg(new BTBResp)
val mem_reg_xcpt = Reg(Bool())
val mem_reg_replay = Reg(Bool())
val mem_reg_flush_pipe = Reg(Bool())
val mem_reg_cause = Reg(UInt())
val mem_reg_slow_bypass = Reg(Bool())
2016-06-09 05:19:52 +02:00
val mem_reg_load = Reg(Bool())
val mem_reg_store = Reg(Bool())
val mem_reg_sfence = Reg(Bool())
val mem_reg_pc = Reg(UInt())
val mem_reg_inst = Reg(Bits())
2017-09-20 07:59:28 +02:00
val mem_reg_raw_inst = Reg(UInt())
val mem_reg_wdata = Reg(Bits())
val mem_reg_rs2 = Reg(Bits())
val mem_br_taken = Reg(Bool())
val take_pc_mem = Wire(Bool())
val wb_reg_valid = Reg(Bool())
val wb_reg_xcpt = Reg(Bool())
val wb_reg_replay = Reg(Bool())
val wb_reg_flush_pipe = Reg(Bool())
val wb_reg_cause = Reg(UInt())
val wb_reg_sfence = Reg(Bool())
val wb_reg_pc = Reg(UInt())
val wb_reg_inst = Reg(Bits())
2017-09-20 07:59:28 +02:00
val wb_reg_raw_inst = Reg(UInt())
val wb_reg_wdata = Reg(Bits())
val wb_reg_rs2 = Reg(Bits())
val take_pc_wb = Wire(Bool())
val take_pc_mem_wb = take_pc_wb || take_pc_mem
val take_pc = take_pc_mem_wb
// decode stage
2016-07-30 01:36:07 +02:00
val ibuf = Module(new IBuf)
val id_expanded_inst = ibuf.io.inst.map(_.bits.inst)
2017-09-20 07:59:28 +02:00
val id_raw_inst = ibuf.io.inst.map(_.bits.raw)
2016-07-30 01:36:07 +02:00
val id_inst = id_expanded_inst.map(_.bits)
ibuf.io.imem <> io.imem.resp
2016-07-30 01:36:07 +02:00
ibuf.io.kill := take_pc
require(decodeWidth == 1 /* TODO */ && retireWidth == decodeWidth)
2016-07-30 01:36:07 +02:00
val id_ctrl = Wire(new IntCtrlSigs()).decode(id_inst(0), decode_table)
val id_raddr3 = id_expanded_inst(0).rs3
val id_raddr2 = id_expanded_inst(0).rs2
val id_raddr1 = id_expanded_inst(0).rs1
val id_waddr = id_expanded_inst(0).rd
val id_load_use = Wire(Bool())
val id_reg_fence = Reg(init=Bool(false))
val id_ren = IndexedSeq(id_ctrl.rxs1, id_ctrl.rxs2)
val id_raddr = IndexedSeq(id_raddr1, id_raddr2)
2015-11-25 03:27:07 +01:00
val rf = new RegFile(31, xLen)
val id_rs = id_raddr.map(rf.read _)
val ctrl_killd = Wire(Bool())
val id_npc = (ibuf.io.pc.asSInt + ImmGen(IMM_UJ, id_inst(0))).asUInt
2017-03-09 09:28:19 +01:00
val csr = Module(new CSRFile(perfEvents))
val id_csr_en = id_ctrl.csr.isOneOf(CSR.S, CSR.C, CSR.W)
val id_system_insn = id_ctrl.csr >= CSR.I
val id_csr_ren = id_ctrl.csr.isOneOf(CSR.S, CSR.C) && id_raddr1 === UInt(0)
val id_csr = Mux(id_csr_ren, CSR.R, id_ctrl.csr)
val id_sfence = id_ctrl.mem && id_ctrl.mem_cmd === M_SFENCE
val id_csr_flush = id_sfence || id_system_insn || (id_csr_en && !id_csr_ren && csr.io.decode(0).write_flush)
val id_illegal_insn = !id_ctrl.legal ||
id_ctrl.div && !csr.io.status.isa('m'-'a') ||
id_ctrl.amo && !csr.io.status.isa('a'-'a') ||
id_ctrl.fp && (csr.io.decode(0).fp_illegal || io.fpu.illegal_rm) ||
id_ctrl.dp && !csr.io.status.isa('d'-'a') ||
ibuf.io.inst(0).bits.rvc && !csr.io.status.isa('c'-'a') ||
id_ctrl.rocc && csr.io.decode(0).rocc_illegal ||
id_csr_en && (csr.io.decode(0).read_illegal || !id_csr_ren && csr.io.decode(0).write_illegal) ||
!ibuf.io.inst(0).bits.rvc && ((id_sfence || id_system_insn) && csr.io.decode(0).system_illegal)
// stall decode for fences (now, for AMO.aq; later, for AMO.rl and FENCE)
2016-07-30 01:36:07 +02:00
val id_amo_aq = id_inst(0)(26)
val id_amo_rl = id_inst(0)(25)
val id_fence_next = id_ctrl.fence || id_ctrl.amo && id_amo_rl
val id_mem_busy = !io.dmem.ordered || io.dmem.req.valid
when (!id_mem_busy) { id_reg_fence := false }
2015-10-06 06:48:05 +02:00
val id_rocc_busy = Bool(usingRoCC) &&
(io.rocc.busy || ex_reg_valid && ex_ctrl.rocc ||
mem_reg_valid && mem_ctrl.rocc || wb_reg_valid && wb_ctrl.rocc)
2017-04-05 17:29:45 +02:00
val id_do_fence = Wire(init = id_rocc_busy && id_ctrl.fence ||
id_mem_busy && (id_ctrl.amo && id_amo_aq || id_ctrl.fence_i || id_reg_fence && (id_ctrl.mem || id_ctrl.rocc)))
val bpu = Module(new BreakpointUnit(nBreakpoints))
2016-06-09 21:41:52 +02:00
bpu.io.status := csr.io.status
bpu.io.bp := csr.io.bp
2016-07-30 01:36:07 +02:00
bpu.io.pc := ibuf.io.pc
2016-06-09 05:19:52 +02:00
bpu.io.ea := mem_reg_wdata
val id_xcpt0 = ibuf.io.inst(0).bits.xcpt0
val id_xcpt1 = ibuf.io.inst(0).bits.xcpt1
val (id_xcpt, id_cause) = checkExceptions(List(
2016-07-30 01:36:07 +02:00
(csr.io.interrupt, csr.io.interrupt_cause),
(bpu.io.debug_if, UInt(CSR.debugTriggerCause)),
2016-07-30 01:36:07 +02:00
(bpu.io.xcpt_if, UInt(Causes.breakpoint)),
(id_xcpt0.pf.inst, UInt(Causes.fetch_page_fault)),
(id_xcpt0.ae.inst, UInt(Causes.fetch_access)),
(id_xcpt1.pf.inst, UInt(Causes.fetch_page_fault)),
(id_xcpt1.ae.inst, UInt(Causes.fetch_access)),
2016-07-30 01:36:07 +02:00
(id_illegal_insn, UInt(Causes.illegal_instruction))))
val idCoverCauses = List(
(CSR.debugTriggerCause, "DEBUG_TRIGGER"),
(Causes.breakpoint, "BREAKPOINT"),
(Causes.fetch_page_fault, "FETCH_PAGE_FAULT"),
(Causes.fetch_access, "FETCH_ACCESS"),
(Causes.illegal_instruction, "ILLEGAL_INSTRUCTION")
)
coverExceptions(id_xcpt, id_cause, "DECODE", idCoverCauses)
val dcache_bypass_data =
2015-10-06 06:48:05 +02:00
if (fastLoadByte) io.dmem.resp.bits.data
else if (fastLoadWord) io.dmem.resp.bits.data_word_bypass
else wb_reg_wdata
// detect bypass opportunities
val ex_waddr = ex_reg_inst(11,7)
val mem_waddr = mem_reg_inst(11,7)
val wb_waddr = wb_reg_inst(11,7)
val bypass_sources = IndexedSeq(
(Bool(true), UInt(0), UInt(0)), // treat reading x0 as a bypass
(ex_reg_valid && ex_ctrl.wxd, ex_waddr, mem_reg_wdata),
(mem_reg_valid && mem_ctrl.wxd && !mem_ctrl.mem, mem_waddr, wb_reg_wdata),
(mem_reg_valid && mem_ctrl.wxd, mem_waddr, dcache_bypass_data))
val id_bypass_src = id_raddr.map(raddr => bypass_sources.map(s => s._1 && s._2 === raddr))
// execute stage
2017-04-14 00:54:24 +02:00
val bypass_mux = bypass_sources.map(_._3)
val ex_reg_rs_bypass = Reg(Vec(id_raddr.size, Bool()))
val ex_reg_rs_lsb = Reg(Vec(id_raddr.size, UInt(width = log2Ceil(bypass_sources.size))))
val ex_reg_rs_msb = Reg(Vec(id_raddr.size, UInt()))
val ex_rs = for (i <- 0 until id_raddr.size)
yield Mux(ex_reg_rs_bypass(i), bypass_mux(ex_reg_rs_lsb(i)), Cat(ex_reg_rs_msb(i), ex_reg_rs_lsb(i)))
2015-11-25 03:27:07 +01:00
val ex_imm = ImmGen(ex_ctrl.sel_imm, ex_reg_inst)
val ex_op1 = MuxLookup(ex_ctrl.sel_alu1, SInt(0), Seq(
A1_RS1 -> ex_rs(0).asSInt,
A1_PC -> ex_reg_pc.asSInt))
val ex_op2 = MuxLookup(ex_ctrl.sel_alu2, SInt(0), Seq(
A2_RS2 -> ex_rs(1).asSInt,
A2_IMM -> ex_imm,
2016-07-30 01:36:07 +02:00
A2_SIZE -> Mux(ex_reg_rvc, SInt(2), SInt(4))))
2015-12-01 02:35:33 +01:00
val alu = Module(new ALU)
alu.io.dw := ex_ctrl.alu_dw
alu.io.fn := ex_ctrl.alu_fn
alu.io.in2 := ex_op2.asUInt
alu.io.in1 := ex_op1.asUInt
// multiplier and divider
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
val div = Module(new MulDiv(mulDivParams, width = xLen))
div.io.req.valid := ex_reg_valid && ex_ctrl.div
div.io.req.bits.dw := ex_ctrl.alu_dw
div.io.req.bits.fn := ex_ctrl.alu_fn
div.io.req.bits.in1 := ex_rs(0)
div.io.req.bits.in2 := ex_rs(1)
div.io.req.bits.tag := ex_waddr
ex_reg_valid := !ctrl_killd
2016-07-30 01:36:07 +02:00
ex_reg_replay := !take_pc && ibuf.io.inst(0).valid && ibuf.io.inst(0).bits.replay
ex_reg_xcpt := !ctrl_killd && id_xcpt
2016-07-30 01:36:07 +02:00
ex_reg_xcpt_interrupt := !take_pc && ibuf.io.inst(0).valid && csr.io.interrupt
when (!ctrl_killd) {
ex_ctrl := id_ctrl
2016-07-30 01:36:07 +02:00
ex_reg_rvc := ibuf.io.inst(0).bits.rvc
ex_ctrl.csr := id_csr
when (id_fence_next) { id_reg_fence := true }
2016-07-30 01:36:07 +02:00
when (id_xcpt) { // pass PC down ALU writeback pipeline for badaddr
ex_ctrl.alu_fn := ALU.FN_ADD
ex_ctrl.alu_dw := DW_XPR
ex_ctrl.sel_alu1 := A1_RS1 // badaddr := instruction
2016-07-30 01:36:07 +02:00
ex_ctrl.sel_alu2 := A2_ZERO
when (id_xcpt1.asUInt.orR) { // badaddr := PC+2
ex_ctrl.sel_alu1 := A1_PC
2016-07-30 01:36:07 +02:00
ex_ctrl.sel_alu2 := A2_SIZE
ex_reg_rvc := true
}
when (bpu.io.xcpt_if || id_xcpt0.asUInt.orR) { // badaddr := PC
ex_ctrl.sel_alu1 := A1_PC
ex_ctrl.sel_alu2 := A2_ZERO
}
2016-07-30 01:36:07 +02:00
}
ex_reg_flush_pipe := id_ctrl.fence_i || id_csr_flush
ex_reg_load_use := id_load_use
when (id_sfence) {
ex_ctrl.mem_type := Cat(id_raddr2 =/= UInt(0), id_raddr1 =/= UInt(0))
}
for (i <- 0 until id_raddr.size) {
val do_bypass = id_bypass_src(i).reduce(_||_)
val bypass_src = PriorityEncoder(id_bypass_src(i))
ex_reg_rs_bypass(i) := do_bypass
ex_reg_rs_lsb(i) := bypass_src
when (id_ren(i) && !do_bypass) {
ex_reg_rs_lsb(i) := id_rs(i)(log2Ceil(bypass_sources.size)-1, 0)
ex_reg_rs_msb(i) := id_rs(i) >> log2Ceil(bypass_sources.size)
}
}
when (id_illegal_insn) {
2017-09-20 07:59:28 +02:00
val inst = Mux(ibuf.io.inst(0).bits.rvc, id_raw_inst(0)(15, 0), id_raw_inst(0))
ex_reg_rs_bypass(0) := false
ex_reg_rs_lsb(0) := inst(log2Ceil(bypass_sources.size)-1, 0)
ex_reg_rs_msb(0) := inst >> log2Ceil(bypass_sources.size)
}
}
2016-07-30 01:36:07 +02:00
when (!ctrl_killd || csr.io.interrupt || ibuf.io.inst(0).bits.replay) {
ex_reg_cause := id_cause
2016-07-30 01:36:07 +02:00
ex_reg_inst := id_inst(0)
2017-09-20 07:59:28 +02:00
ex_reg_raw_inst := id_raw_inst(0)
2016-07-30 01:36:07 +02:00
ex_reg_pc := ibuf.io.pc
ex_reg_btb_resp := ibuf.io.btb_resp
}
// replay inst in ex stage?
val ex_pc_valid = ex_reg_valid || ex_reg_replay || ex_reg_xcpt_interrupt
val wb_dcache_miss = wb_ctrl.mem && !io.dmem.resp.valid
val replay_ex_structural = ex_ctrl.mem && !io.dmem.req.ready ||
ex_ctrl.div && !div.io.req.ready
val replay_ex_load_use = wb_dcache_miss && ex_reg_load_use
val replay_ex = ex_reg_replay || (ex_reg_valid && (replay_ex_structural || replay_ex_load_use))
val ctrl_killx = take_pc_mem_wb || replay_ex || !ex_reg_valid
// detect 2-cycle load-use delay for LB/LH/SC
val ex_slow_bypass = ex_ctrl.mem_cmd === M_XSC || Vec(MT_B, MT_BU, MT_H, MT_HU).contains(ex_ctrl.mem_type)
val ex_sfence = Bool(usingVM) && ex_ctrl.mem && ex_ctrl.mem_cmd === M_SFENCE
val (ex_xcpt, ex_cause) = checkExceptions(List(
(ex_reg_xcpt_interrupt || ex_reg_xcpt, ex_reg_cause)))
val exCoverCauses = idCoverCauses
coverExceptions(ex_xcpt, ex_cause, "EXECUTE", exCoverCauses)
// memory stage
2017-03-09 09:28:19 +01:00
val mem_pc_valid = mem_reg_valid || mem_reg_replay || mem_reg_xcpt_interrupt
val mem_br_target = mem_reg_pc.asSInt +
2015-11-25 03:27:07 +01:00
Mux(mem_ctrl.branch && mem_br_taken, ImmGen(IMM_SB, mem_reg_inst),
Mux(mem_ctrl.jal, ImmGen(IMM_UJ, mem_reg_inst),
2016-07-30 01:36:07 +02:00
Mux(mem_reg_rvc, SInt(2), SInt(4))))
val mem_npc = (Mux(mem_ctrl.jalr || mem_reg_sfence, encodeVirtualAddress(mem_reg_wdata, mem_reg_wdata).asSInt, mem_br_target) & SInt(-2)).asUInt
val mem_wrong_npc =
Mux(ex_pc_valid, mem_npc =/= ex_reg_pc,
Mux(ibuf.io.inst(0).valid || ibuf.io.imem.valid, mem_npc =/= ibuf.io.pc, Bool(true)))
val mem_npc_misaligned = !csr.io.status.isa('c'-'a') && mem_npc(1) && !mem_reg_sfence
val mem_int_wdata = Mux(!mem_reg_xcpt && (mem_ctrl.jalr ^ mem_npc_misaligned), mem_br_target, mem_reg_wdata.asSInt).asUInt
val mem_cfi = mem_ctrl.branch || mem_ctrl.jalr || mem_ctrl.jal
val mem_cfi_taken = (mem_ctrl.branch && mem_br_taken) || mem_ctrl.jalr || mem_ctrl.jal
val mem_direction_misprediction = mem_ctrl.branch && mem_br_taken =/= (usingBTB && mem_reg_btb_resp.taken)
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
val mem_misprediction = if (usingBTB) mem_wrong_npc else mem_cfi_taken
take_pc_mem := mem_reg_valid && (mem_misprediction || mem_reg_sfence)
mem_reg_valid := !ctrl_killx
mem_reg_replay := !take_pc_mem_wb && replay_ex
mem_reg_xcpt := !ctrl_killx && ex_xcpt
mem_reg_xcpt_interrupt := !take_pc_mem_wb && ex_reg_xcpt_interrupt
// on pipeline flushes, cause mem_npc to hold the sequential npc, which
// will drive the W-stage npc mux
when (mem_reg_valid && mem_reg_flush_pipe) {
mem_reg_sfence := false
}.elsewhen (ex_pc_valid) {
mem_ctrl := ex_ctrl
2016-07-30 01:36:07 +02:00
mem_reg_rvc := ex_reg_rvc
2016-06-09 05:19:52 +02:00
mem_reg_load := ex_ctrl.mem && isRead(ex_ctrl.mem_cmd)
mem_reg_store := ex_ctrl.mem && isWrite(ex_ctrl.mem_cmd)
mem_reg_sfence := ex_sfence
mem_reg_btb_resp := ex_reg_btb_resp
mem_reg_flush_pipe := ex_reg_flush_pipe
mem_reg_slow_bypass := ex_slow_bypass
mem_reg_cause := ex_cause
mem_reg_inst := ex_reg_inst
2017-09-20 07:59:28 +02:00
mem_reg_raw_inst := ex_reg_raw_inst
mem_reg_pc := ex_reg_pc
mem_reg_wdata := alu.io.out
mem_br_taken := alu.io.cmp_out
when (ex_ctrl.rxs2 && (ex_ctrl.mem || ex_ctrl.rocc || ex_sfence)) {
val typ = Mux(ex_ctrl.rocc, log2Ceil(xLen/8).U, ex_ctrl.mem_type)
mem_reg_rs2 := new StoreGen(typ, 0.U, ex_rs(1), coreDataBytes).data
}
when (ex_ctrl.jalr && csr.io.status.debug) {
// flush I$ on D-mode JALR to effect uncached fetch without D$ flush
mem_ctrl.fence_i := true
mem_reg_flush_pipe := true
}
}
val mem_breakpoint = (mem_reg_load && bpu.io.xcpt_ld) || (mem_reg_store && bpu.io.xcpt_st)
val mem_debug_breakpoint = (mem_reg_load && bpu.io.debug_ld) || (mem_reg_store && bpu.io.debug_st)
2016-06-09 21:33:43 +02:00
val (mem_new_xcpt, mem_new_cause) = checkExceptions(List(
(mem_debug_breakpoint, UInt(CSR.debugTriggerCause)),
(mem_breakpoint, UInt(Causes.breakpoint)),
2017-04-15 08:57:32 +02:00
(mem_npc_misaligned, UInt(Causes.misaligned_fetch))))
2016-06-09 21:33:43 +02:00
val (mem_xcpt, mem_cause) = checkExceptions(List(
2016-06-09 21:33:43 +02:00
(mem_reg_xcpt_interrupt || mem_reg_xcpt, mem_reg_cause),
(mem_reg_valid && mem_new_xcpt, mem_new_cause)))
val memCoverCauses = (exCoverCauses ++ List(
(CSR.debugTriggerCause, "DEBUG_TRIGGER"),
2017-12-04 23:04:24 +01:00
(Causes.breakpoint, "BREAKPOINT"),
(Causes.misaligned_fetch, "MISALIGNED_FETCH")
)).distinct
coverExceptions(mem_xcpt, mem_cause, "MEMORY", memCoverCauses)
2016-05-14 02:54:23 +02:00
val dcache_kill_mem = mem_reg_valid && mem_ctrl.wxd && io.dmem.replay_next // structural hazard on writeback port
val fpu_kill_mem = mem_reg_valid && mem_ctrl.fp && io.fpu.nack_mem
val replay_mem = dcache_kill_mem || mem_reg_replay || fpu_kill_mem
val killm_common = dcache_kill_mem || take_pc_wb || mem_reg_xcpt || !mem_reg_valid
div.io.kill := killm_common && Reg(next = div.io.req.fire())
val ctrl_killm = killm_common || mem_xcpt || fpu_kill_mem
2015-07-23 02:32:44 +02:00
// writeback stage
wb_reg_valid := !ctrl_killm
wb_reg_replay := replay_mem && !take_pc_wb
wb_reg_xcpt := mem_xcpt && !take_pc_wb
wb_reg_flush_pipe := !ctrl_killm && mem_reg_flush_pipe
2017-03-09 09:28:19 +01:00
when (mem_pc_valid) {
wb_ctrl := mem_ctrl
wb_reg_sfence := mem_reg_sfence
2016-07-30 01:36:07 +02:00
wb_reg_wdata := Mux(!mem_reg_xcpt && mem_ctrl.fp && mem_ctrl.wxd, io.fpu.toint_data, mem_int_wdata)
when (mem_ctrl.rocc || mem_reg_sfence) {
wb_reg_rs2 := mem_reg_rs2
}
wb_reg_cause := mem_cause
wb_reg_inst := mem_reg_inst
2017-09-20 07:59:28 +02:00
wb_reg_raw_inst := mem_reg_raw_inst
wb_reg_pc := mem_reg_pc
}
2017-03-16 02:00:32 +01:00
val (wb_xcpt, wb_cause) = checkExceptions(List(
(wb_reg_xcpt, wb_reg_cause),
2017-04-15 08:57:32 +02:00
(wb_reg_valid && wb_ctrl.mem && io.dmem.s2_xcpt.ma.st, UInt(Causes.misaligned_store)),
(wb_reg_valid && wb_ctrl.mem && io.dmem.s2_xcpt.ma.ld, UInt(Causes.misaligned_load)),
(wb_reg_valid && wb_ctrl.mem && io.dmem.s2_xcpt.pf.st, UInt(Causes.store_page_fault)),
(wb_reg_valid && wb_ctrl.mem && io.dmem.s2_xcpt.pf.ld, UInt(Causes.load_page_fault)),
(wb_reg_valid && wb_ctrl.mem && io.dmem.s2_xcpt.ae.st, UInt(Causes.store_access)),
(wb_reg_valid && wb_ctrl.mem && io.dmem.s2_xcpt.ae.ld, UInt(Causes.load_access))
2017-03-16 02:00:32 +01:00
))
val wbCoverCauses = List(
(Causes.misaligned_store, "MISALIGNED_STORE"),
(Causes.misaligned_load, "MISALIGNED_LOAD"),
(Causes.store_page_fault, "STORE_PAGE_FAULT"),
(Causes.load_page_fault, "LOAD_PAGE_FAULT"),
(Causes.store_access, "STORE_ACCESS"),
(Causes.load_access, "LOAD_ACCESS")
)
coverExceptions(wb_xcpt, wb_cause, "WRITEBACK", wbCoverCauses)
val wb_wxd = wb_reg_valid && wb_ctrl.wxd
val wb_set_sboard = wb_ctrl.div || wb_dcache_miss || wb_ctrl.rocc
val replay_wb_common = io.dmem.s2_nack || wb_reg_replay
2016-07-14 21:09:34 +02:00
val replay_wb_rocc = wb_reg_valid && wb_ctrl.rocc && !io.rocc.cmd.ready
val replay_wb = replay_wb_common || replay_wb_rocc
take_pc_wb := replay_wb || wb_xcpt || csr.io.eret || wb_reg_flush_pipe
// writeback arbitration
val dmem_resp_xpu = !io.dmem.resp.bits.tag(0).toBool
val dmem_resp_fpu = io.dmem.resp.bits.tag(0).toBool
2016-07-30 01:36:07 +02:00
val dmem_resp_waddr = io.dmem.resp.bits.tag(5, 1)
val dmem_resp_valid = io.dmem.resp.valid && io.dmem.resp.bits.has_data
val dmem_resp_replay = dmem_resp_valid && io.dmem.resp.bits.replay
div.io.resp.ready := !wb_wxd
val ll_wdata = Wire(init = div.io.resp.bits.data)
val ll_waddr = Wire(init = div.io.resp.bits.tag)
val ll_wen = Wire(init = div.io.resp.fire())
2015-10-06 06:48:05 +02:00
if (usingRoCC) {
io.rocc.resp.ready := !wb_wxd
when (io.rocc.resp.fire()) {
div.io.resp.ready := Bool(false)
ll_wdata := io.rocc.resp.bits.data
ll_waddr := io.rocc.resp.bits.rd
ll_wen := Bool(true)
}
}
when (dmem_resp_replay && dmem_resp_xpu) {
div.io.resp.ready := Bool(false)
2015-10-06 06:48:05 +02:00
if (usingRoCC)
io.rocc.resp.ready := Bool(false)
ll_waddr := dmem_resp_waddr
ll_wen := Bool(true)
}
2016-07-14 21:09:34 +02:00
val wb_valid = wb_reg_valid && !replay_wb && !wb_xcpt
val wb_wen = wb_valid && wb_ctrl.wxd
val rf_wen = wb_wen || ll_wen
val rf_waddr = Mux(ll_wen, ll_waddr, wb_waddr)
val rf_wdata = Mux(dmem_resp_valid && dmem_resp_xpu, io.dmem.resp.bits.data,
Mux(ll_wen, ll_wdata,
Mux(wb_ctrl.csr =/= CSR.N, csr.io.rw.rdata,
wb_reg_wdata)))
when (rf_wen) { rf.write(rf_waddr, rf_wdata) }
// hook up control/status regfile
csr.io.decode(0).csr := id_raw_inst(0)(31,20)
2017-03-16 02:00:32 +01:00
csr.io.exception := wb_xcpt
csr.io.cause := wb_cause
csr.io.retire := wb_valid
csr.io.inst(0) := (if (usingCompressed) Cat(Mux(wb_reg_raw_inst(1, 0).andR, wb_reg_inst >> 16, 0.U), wb_reg_raw_inst(15, 0)) else wb_reg_inst)
csr.io.interrupts := io.interrupts
csr.io.hartid := io.hartid
io.fpu.fcsr_rm := csr.io.fcsr_rm
csr.io.fcsr_flags := io.fpu.fcsr_flags
csr.io.rocc_interrupt := io.rocc.interrupt
csr.io.pc := wb_reg_pc
2016-07-30 01:36:07 +02:00
csr.io.badaddr := encodeVirtualAddress(wb_reg_wdata, wb_reg_wdata)
io.ptw.ptbr := csr.io.ptbr
io.ptw.status := csr.io.status
2017-03-15 09:18:39 +01:00
io.ptw.pmp := csr.io.pmp
csr.io.rw.addr := wb_reg_inst(31,20)
csr.io.rw.cmd := Mux(wb_reg_valid, wb_ctrl.csr, CSR.N)
csr.io.rw.wdata := wb_reg_wdata
2017-09-20 07:59:28 +02:00
io.trace := csr.io.trace
val hazard_targets = Seq((id_ctrl.rxs1 && id_raddr1 =/= UInt(0), id_raddr1),
(id_ctrl.rxs2 && id_raddr2 =/= UInt(0), id_raddr2),
(id_ctrl.wxd && id_waddr =/= UInt(0), id_waddr))
val fp_hazard_targets = Seq((io.fpu.dec.ren1, id_raddr1),
(io.fpu.dec.ren2, id_raddr2),
(io.fpu.dec.ren3, id_raddr3),
(io.fpu.dec.wen, id_waddr))
2016-07-30 01:36:07 +02:00
val sboard = new Scoreboard(32, true)
2015-07-23 02:32:44 +02:00
sboard.clear(ll_wen, ll_waddr)
2017-08-02 20:49:43 +02:00
def id_sboard_clear_bypass(r: UInt) = {
// ll_waddr arrives late when D$ has ECC, so reshuffle the hazard check
if (tileParams.dcache.get.dataECC.isInstanceOf[IdentityCode]) ll_wen && ll_waddr === r
else div.io.resp.fire() && div.io.resp.bits.tag === r || dmem_resp_replay && dmem_resp_xpu && dmem_resp_waddr === r
}
val id_sboard_hazard = checkHazards(hazard_targets, rd => sboard.read(rd) && !id_sboard_clear_bypass(rd))
2015-07-23 02:32:44 +02:00
sboard.set(wb_set_sboard && wb_wen, wb_waddr)
// stall for RAW/WAW hazards on CSRs, loads, AMOs, and mul/div in execute stage.
val ex_cannot_bypass = ex_ctrl.csr =/= CSR.N || ex_ctrl.jalr || ex_ctrl.mem || ex_ctrl.div || ex_ctrl.fp || ex_ctrl.rocc
val data_hazard_ex = ex_ctrl.wxd && checkHazards(hazard_targets, _ === ex_waddr)
val fp_data_hazard_ex = ex_ctrl.wfd && checkHazards(fp_hazard_targets, _ === ex_waddr)
val id_ex_hazard = ex_reg_valid && (data_hazard_ex && ex_cannot_bypass || fp_data_hazard_ex)
// stall for RAW/WAW hazards on CSRs, LB/LH, and mul/div in memory stage.
val mem_mem_cmd_bh =
2015-10-06 06:48:05 +02:00
if (fastLoadWord) Bool(!fastLoadByte) && mem_reg_slow_bypass
else Bool(true)
val mem_cannot_bypass = mem_ctrl.csr =/= CSR.N || mem_ctrl.mem && mem_mem_cmd_bh || mem_ctrl.div || mem_ctrl.fp || mem_ctrl.rocc
val data_hazard_mem = mem_ctrl.wxd && checkHazards(hazard_targets, _ === mem_waddr)
val fp_data_hazard_mem = mem_ctrl.wfd && checkHazards(fp_hazard_targets, _ === mem_waddr)
val id_mem_hazard = mem_reg_valid && (data_hazard_mem && mem_cannot_bypass || fp_data_hazard_mem)
id_load_use := mem_reg_valid && data_hazard_mem && mem_ctrl.mem
// stall for RAW/WAW hazards on load/AMO misses and mul/div in writeback.
val data_hazard_wb = wb_ctrl.wxd && checkHazards(hazard_targets, _ === wb_waddr)
val fp_data_hazard_wb = wb_ctrl.wfd && checkHazards(fp_hazard_targets, _ === wb_waddr)
val id_wb_hazard = wb_reg_valid && (data_hazard_wb && wb_set_sboard || fp_data_hazard_wb)
2015-10-06 06:48:05 +02:00
val id_stall_fpu = if (usingFPU) {
val fp_sboard = new Scoreboard(32)
fp_sboard.set((wb_dcache_miss && wb_ctrl.wfd || io.fpu.sboard_set) && wb_valid, wb_waddr)
fp_sboard.clear(dmem_resp_replay && dmem_resp_fpu, dmem_resp_waddr)
fp_sboard.clear(io.fpu.sboard_clr, io.fpu.sboard_clra)
id_csr_en && !io.fpu.fcsr_rdy || checkHazards(fp_hazard_targets, fp_sboard.read _)
} else Bool(false)
val dcache_blocked = Reg(Bool())
dcache_blocked := !io.dmem.req.ready && (io.dmem.req.valid || dcache_blocked)
val rocc_blocked = Reg(Bool())
2017-03-16 02:00:32 +01:00
rocc_blocked := !wb_xcpt && !io.rocc.cmd.ready && (io.rocc.cmd.valid || rocc_blocked)
val ctrl_stalld =
id_ex_hazard || id_mem_hazard || id_wb_hazard || id_sboard_hazard ||
csr.io.singleStep && (ex_reg_valid || mem_reg_valid || wb_reg_valid) ||
id_ctrl.fp && id_stall_fpu ||
id_ctrl.mem && dcache_blocked || // reduce activity during D$ misses
id_ctrl.rocc && rocc_blocked || // reduce activity while RoCC is busy
id_ctrl.div && (!(div.io.req.ready || (div.io.resp.valid && !wb_wxd)) || div.io.req.valid) || // reduce odds of replay
id_do_fence ||
csr.io.csr_stall
ctrl_killd := !ibuf.io.inst(0).valid || ibuf.io.inst(0).bits.replay || take_pc_mem_wb || ctrl_stalld || csr.io.interrupt
2015-07-23 02:32:44 +02:00
io.imem.req.valid := take_pc
io.imem.req.bits.speculative := !take_pc_wb
2015-07-23 02:32:44 +02:00
io.imem.req.bits.pc :=
Mux(wb_xcpt || csr.io.eret, csr.io.evec, // exception or [m|s]ret
Mux(replay_wb, wb_reg_pc, // replay
mem_npc)) // flush or branch misprediction
2016-06-01 04:27:28 +02:00
io.imem.flush_icache := wb_reg_valid && wb_ctrl.fence_i && !io.dmem.s2_nack
io.imem.sfence.valid := wb_reg_valid && wb_reg_sfence
io.imem.sfence.bits.rs1 := wb_ctrl.mem_type(0)
io.imem.sfence.bits.rs2 := wb_ctrl.mem_type(1)
2017-07-06 08:53:52 +02:00
io.imem.sfence.bits.addr := wb_reg_wdata
io.imem.sfence.bits.asid := wb_reg_rs2
2017-07-06 08:53:52 +02:00
io.ptw.sfence := io.imem.sfence
2015-07-23 02:32:44 +02:00
ibuf.io.inst(0).ready := !ctrl_stalld
2016-07-30 01:36:07 +02:00
io.imem.btb_update.valid := mem_reg_valid && !take_pc_wb && mem_wrong_npc && (!mem_cfi || mem_cfi_taken)
io.imem.btb_update.bits.isValid := mem_cfi
io.imem.btb_update.bits.cfiType :=
Mux((mem_ctrl.jal || mem_ctrl.jalr) && mem_waddr(0), CFIType.call,
Mux(mem_ctrl.jalr && mem_reg_inst(19,15) === BitPat("b00?01"), CFIType.ret,
Mux(mem_ctrl.jal || mem_ctrl.jalr, CFIType.jump,
CFIType.branch)))
2015-07-23 02:32:44 +02:00
io.imem.btb_update.bits.target := io.imem.req.bits.pc
2016-07-30 01:36:07 +02:00
io.imem.btb_update.bits.br_pc := (if (usingCompressed) mem_reg_pc + Mux(mem_reg_rvc, UInt(0), UInt(2)) else mem_reg_pc)
io.imem.btb_update.bits.pc := ~(~io.imem.btb_update.bits.br_pc | (coreInstBytes*fetchWidth-1))
io.imem.btb_update.bits.prediction := mem_reg_btb_resp
2015-07-23 02:32:44 +02:00
io.imem.bht_update.valid := mem_reg_valid && !take_pc_wb
2016-07-30 01:36:07 +02:00
io.imem.bht_update.bits.pc := io.imem.btb_update.bits.pc
2015-07-23 02:32:44 +02:00
io.imem.bht_update.bits.taken := mem_br_taken
io.imem.bht_update.bits.mispredict := mem_wrong_npc
io.imem.bht_update.bits.branch := mem_ctrl.branch
io.imem.bht_update.bits.prediction := mem_reg_btb_resp.bht
2015-07-23 02:32:44 +02:00
io.fpu.valid := !ctrl_killd && id_ctrl.fp
io.fpu.killx := ctrl_killx
io.fpu.killm := killm_common
2016-07-30 01:36:07 +02:00
io.fpu.inst := id_inst(0)
io.fpu.fromint_data := ex_rs(0)
io.fpu.dmem_resp_val := dmem_resp_valid && dmem_resp_fpu
io.fpu.dmem_resp_data := io.dmem.resp.bits.data_word_bypass
io.fpu.dmem_resp_type := io.dmem.resp.bits.typ
io.fpu.dmem_resp_tag := dmem_resp_waddr
io.dmem.req.valid := ex_reg_valid && ex_ctrl.mem
val ex_dcache_tag = Cat(ex_waddr, ex_ctrl.fp)
require(coreDCacheReqTagBits >= ex_dcache_tag.getWidth)
io.dmem.req.bits.tag := ex_dcache_tag
io.dmem.req.bits.cmd := ex_ctrl.mem_cmd
io.dmem.req.bits.typ := ex_ctrl.mem_type
io.dmem.req.bits.phys := Bool(false)
2016-03-11 02:32:00 +01:00
io.dmem.req.bits.addr := encodeVirtualAddress(ex_rs(0), alu.io.adder_out)
io.dmem.invalidate_lr := wb_xcpt
io.dmem.s1_data.data := Mux(mem_ctrl.fp, io.fpu.store_data, mem_reg_rs2)
io.dmem.s1_kill := killm_common || mem_breakpoint
io.rocc.cmd.valid := wb_reg_valid && wb_ctrl.rocc && !replay_wb_common
io.rocc.exception := wb_xcpt && csr.io.status.xs.orR
io.rocc.cmd.bits.status := csr.io.status
io.rocc.cmd.bits.inst := new RoCCInstruction().fromBits(wb_reg_inst)
io.rocc.cmd.bits.rs1 := wb_reg_wdata
io.rocc.cmd.bits.rs2 := wb_reg_rs2
2017-03-09 09:28:19 +01:00
// evaluate performance counters
2017-07-28 22:14:04 +02:00
val icache_blocked = !(io.imem.resp.valid || RegNext(io.imem.resp.valid))
2017-03-09 09:28:19 +01:00
csr.io.counters foreach { c => c.inc := RegNext(perfEvents.evaluate(c.eventSel)) }
2015-10-06 06:48:05 +02:00
if (enableCommitLog) {
2017-09-20 07:59:28 +02:00
val t = csr.io.trace(0)
val rd = wb_waddr
val wfd = wb_ctrl.wfd
val wxd = wb_ctrl.wxd
val has_data = wb_wen && !wb_set_sboard
when (t.valid && !t.exception) {
when (wfd) {
2017-09-20 23:32:41 +02:00
printf ("%d 0x%x (0x%x) f%d p%d 0xXXXXXXXXXXXXXXXX\n", t.priv, t.iaddr, t.insn, rd, rd+UInt(32))
}
.elsewhen (wxd && rd =/= UInt(0) && has_data) {
2017-09-20 23:32:41 +02:00
printf ("%d 0x%x (0x%x) x%d 0x%x\n", t.priv, t.iaddr, t.insn, rd, rf_wdata)
}
.elsewhen (wxd && rd =/= UInt(0) && !has_data) {
2017-09-20 23:32:41 +02:00
printf ("%d 0x%x (0x%x) x%d p%d 0xXXXXXXXXXXXXXXXX\n", t.priv, t.iaddr, t.insn, rd, rd)
}
.otherwise {
2017-09-20 23:32:41 +02:00
printf ("%d 0x%x (0x%x)\n", t.priv, t.iaddr, t.insn)
}
}
when (ll_wen && rf_waddr =/= UInt(0)) {
printf ("x%d p%d 0x%x\n", rf_waddr, rf_waddr, rf_wdata)
}
}
else {
printf("C%d: %d [%d] pc=[%x] W[r%d=%x][%d] R[r%d=%x] R[r%d=%x] inst=[%x] DASM(%x)\n",
io.hartid, csr.io.time(31,0), csr.io.trace(0).valid && !csr.io.trace(0).exception,
csr.io.trace(0).iaddr(vaddrBitsExtended-1, 0),
Mux(rf_wen && !(wb_set_sboard && wb_wen), rf_waddr, UInt(0)), rf_wdata, rf_wen,
wb_reg_inst(19,15), Reg(next=Reg(next=ex_rs(0))),
wb_reg_inst(24,20), Reg(next=Reg(next=ex_rs(1))),
2017-09-20 07:59:28 +02:00
csr.io.trace(0).insn, csr.io.trace(0).insn)
}
val max_core_cycles = PlusArg("max-core-cycles",
default = 0,
docstring = "Maximum Core Clock cycles simulation may run before timeout. Ignored if 0 (Default).")
when (max_core_cycles > UInt(0)) {
assert (csr.io.time < max_core_cycles, "Maximum Core Cycles reached.")
}
def checkExceptions(x: Seq[(Bool, UInt)]) =
(x.map(_._1).reduce(_||_), PriorityMux(x))
def coverExceptions(exceptionValid: Bool, cause: UInt, labelPrefix: String, coverCausesLabels: Seq[(Int, String)]): Unit = {
for ((coverCause, label) <- coverCausesLabels) {
cover(exceptionValid && (cause === UInt(coverCause)), s"${labelPrefix}_${label}")
}
}
def checkHazards(targets: Seq[(Bool, UInt)], cond: UInt => Bool) =
targets.map(h => h._1 && cond(h._2)).reduce(_||_)
2016-06-09 21:33:43 +02:00
def encodeVirtualAddress(a0: UInt, ea: UInt) = if (vaddrBitsExtended == vaddrBits) ea else {
// efficient means to compress 64-bit VA into vaddrBits+1 bits
// (VA is bad if VA(vaddrBits) != VA(vaddrBits-1))
val a = a0.asSInt >> vaddrBits
val msb = Mux(a === 0.S || a === -1.S, ea(vaddrBits), !ea(vaddrBits-1))
2016-03-11 02:32:00 +01:00
Cat(msb, ea(vaddrBits-1,0))
}
2016-07-30 01:36:07 +02:00
class Scoreboard(n: Int, zero: Boolean = false)
{
def set(en: Bool, addr: UInt): Unit = update(en, _next | mask(en, addr))
def clear(en: Bool, addr: UInt): Unit = update(en, _next & ~mask(en, addr))
def read(addr: UInt): Bool = r(addr)
def readBypassed(addr: UInt): Bool = _next(addr)
2016-07-30 01:36:07 +02:00
private val _r = Reg(init=Bits(0, n))
private val r = if (zero) (_r >> 1 << 1) else _r
private var _next = r
private var ens = Bool(false)
private def mask(en: Bool, addr: UInt) = Mux(en, UInt(1) << addr, UInt(0))
private def update(en: Bool, update: UInt) = {
_next = update
ens = ens || en
2016-07-30 01:36:07 +02:00
when (ens) { _r := _next }
}
}
}
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
class RegFile(n: Int, w: Int, zero: Boolean = false) {
private val rf = Mem(n, UInt(width = w))
private def access(addr: UInt) = rf(~addr(log2Up(n)-1,0))
private val reads = ArrayBuffer[(UInt,UInt)]()
Heterogeneous Tiles (#550) Fundamental new features: * Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces. * Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile. * Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile. * Defined RocketCoreParams: All the parameters that can be varied per-core. * Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes. * Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created. * Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little. * Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support. Additional changes that got rolled in along the way: * rocket: Fix critical path through BTB for I$ index bits > pgIdxBits * coreplex: tiles connected via :=* * groundtest: updated to use TileParams * tilelink: cache cork requirements are relaxed to allow more cacheless masters
2017-02-09 22:59:09 +01:00
private var canRead = true
def read(addr: UInt) = {
require(canRead)
reads += addr -> Wire(UInt())
reads.last._2 := Mux(Bool(zero) && addr === UInt(0), UInt(0), access(addr))
reads.last._2
}
def write(addr: UInt, data: UInt) = {
canRead = false
when (addr =/= UInt(0)) {
access(addr) := data
for ((raddr, rdata) <- reads)
when (addr === raddr) { rdata := data }
}
}
}
object ImmGen {
def apply(sel: UInt, inst: UInt) = {
val sign = Mux(sel === IMM_Z, SInt(0), inst(31).asSInt)
val b30_20 = Mux(sel === IMM_U, inst(30,20).asSInt, sign)
val b19_12 = Mux(sel =/= IMM_U && sel =/= IMM_UJ, sign, inst(19,12).asSInt)
val b11 = Mux(sel === IMM_U || sel === IMM_Z, SInt(0),
Mux(sel === IMM_UJ, inst(20).asSInt,
Mux(sel === IMM_SB, inst(7).asSInt, sign)))
val b10_5 = Mux(sel === IMM_U || sel === IMM_Z, Bits(0), inst(30,25))
val b4_1 = Mux(sel === IMM_U, Bits(0),
Mux(sel === IMM_S || sel === IMM_SB, inst(11,8),
Mux(sel === IMM_Z, inst(19,16), inst(24,21))))
val b0 = Mux(sel === IMM_S, inst(7),
Mux(sel === IMM_I, inst(20),
Mux(sel === IMM_Z, inst(15), Bits(0))))
Cat(sign, b30_20, b19_12, b11, b10_5, b4_1, b0).asSInt
}
}