1
0

Improve TLB simulation performance

This commit is contained in:
Andrew Waterman 2016-07-02 14:26:05 -07:00
parent a9e0a5e2df
commit 663002ec0c

View File

@ -18,44 +18,6 @@ trait HasTLBParameters extends HasCoreParameters {
val camTagBits = asIdBits + vpnBits val camTagBits = asIdBits + vpnBits
} }
abstract class TLBModule(implicit val p: Parameters) extends Module
with HasTLBParameters
abstract class TLBBundle(implicit val p: Parameters) extends ParameterizedBundle()(p)
with HasTLBParameters
class CAMIO(implicit p: Parameters) extends TLBBundle()(p) {
val clear = Bool(INPUT)
val clear_mask = Bits(INPUT, entries)
val tag = Bits(INPUT, camTagBits)
val hit = Bool(OUTPUT)
val hits = UInt(OUTPUT, entries)
val valid_bits = Bits(OUTPUT, entries)
val write = Bool(INPUT)
val write_tag = Bits(INPUT, camTagBits)
val write_addr = UInt(INPUT, camAddrBits)
}
class RocketCAM(implicit p: Parameters) extends TLBModule()(p) {
val io = new CAMIO
val cam_tags = Mem(entries, Bits(width = camTagBits))
val vb_array = Reg(init=Bits(0, entries))
when (io.write) {
vb_array := vb_array.bitSet(io.write_addr, Bool(true))
cam_tags(io.write_addr) := io.write_tag
}
when (io.clear) {
vb_array := vb_array & ~io.clear_mask
}
val hits = (0 until entries).map(i => vb_array(i) && cam_tags(i) === io.tag)
io.valid_bits := vb_array
io.hits := Vec(hits).toBits
io.hit := io.hits.orR
}
class TLBReq(implicit p: Parameters) extends CoreBundle()(p) { class TLBReq(implicit p: Parameters) extends CoreBundle()(p) {
val vpn = UInt(width = vpnBitsExtended) val vpn = UInt(width = vpnBitsExtended)
val passthrough = Bool() val passthrough = Bool()
@ -63,7 +25,7 @@ class TLBReq(implicit p: Parameters) extends CoreBundle()(p) {
val store = Bool() val store = Bool()
} }
class TLBRespNoHitIndex(implicit p: Parameters) extends CoreBundle()(p) { class TLBResp(implicit p: Parameters) extends CoreBundle()(p) {
// lookup responses // lookup responses
val miss = Bool(OUTPUT) val miss = Bool(OUTPUT)
val ppn = UInt(OUTPUT, ppnBits) val ppn = UInt(OUTPUT, ppnBits)
@ -72,100 +34,88 @@ class TLBRespNoHitIndex(implicit p: Parameters) extends CoreBundle()(p) {
val xcpt_if = Bool(OUTPUT) val xcpt_if = Bool(OUTPUT)
} }
class TLBResp(implicit p: Parameters) extends TLBRespNoHitIndex()(p) with HasTLBParameters { class TLB(implicit val p: Parameters) extends Module with HasTLBParameters {
val hit_idx = UInt(OUTPUT, entries)
}
class TLB(implicit p: Parameters) extends TLBModule()(p) {
val io = new Bundle { val io = new Bundle {
val req = Decoupled(new TLBReq).flip val req = Decoupled(new TLBReq).flip
val resp = new TLBResp val resp = new TLBResp
val ptw = new TLBPTWIO val ptw = new TLBPTWIO
} }
val tag_cam = Module(new RocketCAM) val valid = Reg(init = UInt(0, entries))
val tag_ram = Mem(entries, io.ptw.resp.bits.pte.ppn) val ppns = Reg(Vec(entries, io.ptw.resp.bits.pte.ppn))
val tags = Reg(Vec(entries, UInt(width = asIdBits + vpnBits)))
val s_ready :: s_request :: s_wait :: s_wait_invalidate :: Nil = Enum(UInt(), 4) val s_ready :: s_request :: s_wait :: s_wait_invalidate :: Nil = Enum(UInt(), 4)
val state = Reg(init=s_ready) val state = Reg(init=s_ready)
val r_refill_tag = Reg(tag_cam.io.write_tag) val r_refill_tag = Reg(UInt(width = asIdBits + vpnBits))
val r_refill_waddr = Reg(tag_cam.io.write_addr) val r_refill_waddr = Reg(UInt(width = log2Ceil(entries)))
val r_req = Reg(new TLBReq) val r_req = Reg(new TLBReq)
val lookup_tag = Cat(io.ptw.ptbr.asid, io.req.bits.vpn(vpnBits-1,0)).toUInt val lookup_tag = Cat(io.ptw.ptbr.asid, io.req.bits.vpn(vpnBits-1,0)).toUInt
tag_cam.io.tag := lookup_tag val hitsVec = (0 until entries).map(i => valid(i) && tags(i) === lookup_tag)
tag_cam.io.write := state === s_wait && io.ptw.resp.valid val hits = hitsVec.toBits
tag_cam.io.write_tag := r_refill_tag
tag_cam.io.write_addr := r_refill_waddr
val tag_hit_addr = OHToUInt(tag_cam.io.hits)
// permission bit arrays // permission bit arrays
val ur_array = Reg(Vec(entries, Bool())) // user read permission val ur_array = Reg(UInt(width = entries)) // user read permission
val uw_array = Reg(Vec(entries, Bool())) // user write permission val uw_array = Reg(UInt(width = entries)) // user write permission
val ux_array = Reg(Vec(entries, Bool())) // user execute permission val ux_array = Reg(UInt(width = entries)) // user execute permission
val sr_array = Reg(Vec(entries, Bool())) // supervisor read permission val sr_array = Reg(UInt(width = entries)) // supervisor read permission
val sw_array = Reg(Vec(entries, Bool())) // supervisor write permission val sw_array = Reg(UInt(width = entries)) // supervisor write permission
val sx_array = Reg(Vec(entries, Bool())) // supervisor execute permission val sx_array = Reg(UInt(width = entries)) // supervisor execute permission
val dirty_array = Reg(Vec(entries, Bool())) // PTE dirty bit val dirty_array = Reg(UInt(width = entries)) // PTE dirty bit
when (io.ptw.resp.valid) { when (io.ptw.resp.valid) {
val pte = io.ptw.resp.bits.pte val pte = io.ptw.resp.bits.pte
tag_ram(r_refill_waddr) := pte.ppn ppns(r_refill_waddr) := pte.ppn
ur_array(r_refill_waddr) := pte.ur() tags(r_refill_waddr) := r_refill_tag
uw_array(r_refill_waddr) := pte.uw()
ux_array(r_refill_waddr) := pte.ux() val mask = UIntToOH(r_refill_waddr)
sr_array(r_refill_waddr) := pte.sr() valid := valid | mask
sw_array(r_refill_waddr) := pte.sw() ur_array := Mux(pte.ur(), ur_array | mask, ur_array & ~mask)
sx_array(r_refill_waddr) := pte.sx() uw_array := Mux(pte.uw(), uw_array | mask, uw_array & ~mask)
dirty_array(r_refill_waddr) := pte.d ux_array := Mux(pte.ux(), ux_array | mask, ux_array & ~mask)
sr_array := Mux(pte.sr(), sr_array | mask, sr_array & ~mask)
sw_array := Mux(pte.sw(), sw_array | mask, sw_array & ~mask)
sx_array := Mux(pte.sx(), sx_array | mask, sx_array & ~mask)
dirty_array := Mux(pte.d, dirty_array | mask, dirty_array & ~mask)
} }
// high if there are any unused (invalid) entries in the TLB // high if there are any unused (invalid) entries in the TLB
val has_invalid_entry = !tag_cam.io.valid_bits.andR
val invalid_entry = PriorityEncoder(~tag_cam.io.valid_bits)
val plru = new PseudoLRU(entries) val plru = new PseudoLRU(entries)
val repl_waddr = Mux(has_invalid_entry, invalid_entry, plru.replace) val repl_waddr = Mux(!valid.andR, PriorityEncoder(~valid), plru.replace)
val do_mprv = io.ptw.status.mprv && !io.req.bits.instruction val do_mprv = io.ptw.status.mprv && !io.req.bits.instruction
val priv = Mux(do_mprv, io.ptw.status.mpp, io.ptw.status.prv) val priv = Mux(do_mprv, io.ptw.status.mpp, io.ptw.status.prv)
val priv_s = priv === PRV.S val priv_s = priv === PRV.S
val priv_uses_vm = priv <= PRV.S && !io.ptw.status.debug val priv_uses_vm = priv <= PRV.S && !io.ptw.status.debug
val req_xwr = Cat(!r_req.store, r_req.store, !(r_req.instruction || r_req.store))
val ur_bits = ur_array.toBits val pum_ok = ~Mux(io.ptw.status.pum, ur_array, UInt(0))
val pum_ok = ~Mux(io.ptw.status.pum, ur_bits, UInt(0)) val r_array = Mux(priv_s, sr_array & pum_ok, ur_array)
val r_array = Mux(priv_s, sr_array.toBits & pum_ok, ur_bits) val w_array = Mux(priv_s, sw_array & pum_ok, uw_array)
val w_array = Mux(priv_s, sw_array.toBits & pum_ok, uw_array.toBits) val x_array = Mux(priv_s, sx_array, ux_array)
val x_array = Mux(priv_s, sx_array.toBits, ux_array.toBits)
val vm_enabled = Bool(usingVM) && io.ptw.status.vm(3) && priv_uses_vm && !io.req.bits.passthrough val vm_enabled = Bool(usingVM) && io.ptw.status.vm(3) && priv_uses_vm && !io.req.bits.passthrough
val bad_va = val bad_va =
if (vpnBits == vpnBitsExtended) Bool(false) if (vpnBits == vpnBitsExtended) Bool(false)
else io.req.bits.vpn(vpnBits) =/= io.req.bits.vpn(vpnBits-1) else io.req.bits.vpn(vpnBits) =/= io.req.bits.vpn(vpnBits-1)
// it's only a store hit if the dirty bit is set // it's only a store hit if the dirty bit is set
val tag_hits = tag_cam.io.hits & (dirty_array.toBits | ~Mux(io.req.bits.store, w_array, UInt(0))) val tag_hits = hits & (dirty_array | ~Mux(io.req.bits.store, w_array, UInt(0)))
val tag_hit = tag_hits.orR val tag_hit = tag_hits.orR
val tlb_hit = vm_enabled && tag_hit val tlb_hit = vm_enabled && tag_hit
val tlb_miss = vm_enabled && !tag_hit && !bad_va val tlb_miss = vm_enabled && !tag_hit && !bad_va
when (io.req.valid && tlb_hit) { when (io.req.valid && tlb_hit) {
plru.access(OHToUInt(tag_cam.io.hits)) plru.access(OHToUInt(hits))
} }
val paddr = Cat(io.resp.ppn, UInt(0, pgIdxBits)) val paddr = Cat(io.resp.ppn, UInt(0, pgIdxBits))
val addr_prot = addrMap.getProt(paddr) val addr_prot = addrMap.getProt(paddr)
io.req.ready := state === s_ready io.req.ready := state === s_ready
io.resp.xcpt_ld := bad_va || (!tlb_miss && !addr_prot.r) || (tlb_hit && !(r_array & tag_cam.io.hits).orR) io.resp.xcpt_ld := bad_va || (!tlb_miss && !addr_prot.r) || (tlb_hit && !(r_array & hits).orR)
io.resp.xcpt_st := bad_va || (!tlb_miss && !addr_prot.w) || (tlb_hit && !(w_array & tag_cam.io.hits).orR) io.resp.xcpt_st := bad_va || (!tlb_miss && !addr_prot.w) || (tlb_hit && !(w_array & hits).orR)
io.resp.xcpt_if := bad_va || (!tlb_miss && !addr_prot.x) || (tlb_hit && !(x_array & tag_cam.io.hits).orR) io.resp.xcpt_if := bad_va || (!tlb_miss && !addr_prot.x) || (tlb_hit && !(x_array & hits).orR)
io.resp.miss := tlb_miss io.resp.miss := tlb_miss
io.resp.ppn := Mux(vm_enabled, Mux1H(tag_cam.io.hits, tag_ram), io.req.bits.vpn(ppnBits-1,0)) io.resp.ppn := Mux(vm_enabled, Mux1H(hitsVec, ppns), io.req.bits.vpn(ppnBits-1,0))
io.resp.hit_idx := tag_cam.io.hits
// clear entries on a TLB flush.
// TODO: selective flushing. careful with superpage mappings (flush it all)
tag_cam.io.clear := io.ptw.invalidate
tag_cam.io.clear_mask := ~UInt(0, entries)
io.ptw.req.valid := state === s_request io.ptw.req.valid := state === s_request
io.ptw.req.bits.addr := r_refill_tag io.ptw.req.bits.addr := r_refill_tag
@ -195,6 +145,10 @@ class TLB(implicit p: Parameters) extends TLBModule()(p) {
when (io.ptw.resp.valid) { when (io.ptw.resp.valid) {
state := s_ready state := s_ready
} }
when (io.ptw.invalidate) {
valid := 0
}
} }
} }