Support PutPartial in ScratchpadSlavePort

src/main/scala/rocket/DCache.scala

@@ -121,7 +121,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   tlb.io.req.bits.sfence.valid := s1_sfence
   tlb.io.req.bits.sfence.bits.rs1 := s1_req.typ(0)
   tlb.io.req.bits.sfence.bits.rs2 := s1_req.typ(1)
-  tlb.io.req.bits.sfence.bits.asid := io.cpu.s1_data
+  tlb.io.req.bits.sfence.bits.asid := io.cpu.s1_data.data
   tlb.io.req.bits.passthrough := s1_req.phys
   tlb.io.req.bits.vaddr := s1_req.addr
   tlb.io.req.bits.instruction := false
@@ -155,6 +155,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
     }
   val s1_data_way = Mux(inWriteback, releaseWay, s1_hit_way)
   val s1_data = Mux1H(s1_data_way, data.io.resp) // retime into s2 if critical
+  val s1_mask = Mux(s1_req.cmd === M_PWR, io.cpu.s1_data.mask, new StoreGen(s1_req.typ, s1_req.addr, UInt(0), wordBytes).mask)
 
   val s2_valid = Reg(next=s1_valid_masked && !s1_sfence, init=Bool(false)) && !io.cpu.s2_xcpt.asUInt.orR
   val s2_probe = Reg(next=s1_probe, init=Bool(false))
@@ -229,10 +230,10 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   val pstore1_cmd = RegEnable(s1_req.cmd, s1_valid_not_nacked && s1_write)
   val pstore1_typ = RegEnable(s1_req.typ, s1_valid_not_nacked && s1_write)
   val pstore1_addr = RegEnable(s1_paddr, s1_valid_not_nacked && s1_write)
-  val pstore1_data = RegEnable(io.cpu.s1_data, s1_valid_not_nacked && s1_write)
+  val pstore1_data = RegEnable(io.cpu.s1_data.data, s1_valid_not_nacked && s1_write)
   val pstore1_way = RegEnable(s1_hit_way, s1_valid_not_nacked && s1_write)
-  val pstore1_storegen = new StoreGen(pstore1_typ, pstore1_addr, pstore1_data, wordBytes)
+  val pstore1_mask = RegEnable(s1_mask, s1_valid_not_nacked && s1_write)
-  val pstore1_storegen_data = Wire(init = pstore1_storegen.data)
+  val pstore1_storegen_data = Wire(init = pstore1_data)
   val pstore1_amo = Bool(usingAtomics) && isRead(pstore1_cmd)
   val pstore_drain_structural = pstore1_valid && pstore2_valid && ((s1_valid && s1_write) || pstore1_amo)
   val pstore_drain_opportunistic = !(io.cpu.req.valid && isRead(io.cpu.req.bits.cmd))
@@ -252,21 +253,20 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   val pstore2_addr = RegEnable(pstore1_addr, advance_pstore1)
   val pstore2_way = RegEnable(pstore1_way, advance_pstore1)
   val pstore2_storegen_data = RegEnable(pstore1_storegen_data, advance_pstore1)
-  val pstore2_storegen_mask = RegEnable(pstore1_storegen.mask, advance_pstore1)
+  val pstore2_storegen_mask = RegEnable(pstore1_mask, advance_pstore1)
   dataArb.io.in(0).valid := pstore_drain
   dataArb.io.in(0).bits.write := true
   dataArb.io.in(0).bits.addr := Mux(pstore2_valid, pstore2_addr, pstore1_addr)
   dataArb.io.in(0).bits.way_en := Mux(pstore2_valid, pstore2_way, pstore1_way)
   dataArb.io.in(0).bits.wdata := Fill(rowWords, Mux(pstore2_valid, pstore2_storegen_data, pstore1_storegen_data))
   val pstore_mask_shift = Mux(pstore2_valid, pstore2_addr, pstore1_addr).extract(rowOffBits-1,offsetlsb) << wordOffBits
-  dataArb.io.in(0).bits.wmask := Mux(pstore2_valid, pstore2_storegen_mask, pstore1_storegen.mask) << pstore_mask_shift
+  dataArb.io.in(0).bits.wmask := Mux(pstore2_valid, pstore2_storegen_mask, pstore1_mask) << pstore_mask_shift
 
   // store->load RAW hazard detection
-  val s1_storegen = new StoreGen(s1_req.typ, s1_req.addr, UInt(0), wordBytes)
   val s1_idx = s1_req.addr(idxMSB, wordOffBits)
   val s1_raw_hazard = s1_read &&
-    ((pstore1_valid && pstore1_addr(idxMSB, wordOffBits) === s1_idx && (pstore1_storegen.mask & s1_storegen.mask).orR) ||
+    ((pstore1_valid && pstore1_addr(idxMSB, wordOffBits) === s1_idx && (pstore1_mask & s1_mask).orR) ||
-     (pstore2_valid && pstore2_addr(idxMSB, wordOffBits) === s1_idx && (pstore2_storegen_mask & s1_storegen.mask).orR))
+     (pstore2_valid && pstore2_addr(idxMSB, wordOffBits) === s1_idx && (pstore2_storegen_mask & s1_mask).orR))
   when (s1_valid && s1_raw_hazard) { s1_nack := true }
 
   metaWriteArb.io.in(0).valid := (s2_valid_hit && s2_update_meta) || (s2_victimize && !s2_victim_dirty)
@@ -279,8 +279,8 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   val a_source = PriorityEncoder(~uncachedInFlight.asUInt << mmioOffset) // skip the MSHR
   val acquire_address = s2_req_block_addr
   val access_address = s2_req.addr
-  val a_size = s2_req.typ(MT_SZ-2, 0)
+  val a_size = mtSize(s2_req.typ)
-  val a_data = Fill(beatWords, pstore1_storegen.data)
+  val a_data = Fill(beatWords, pstore1_data)
   val acquire = if (edge.manager.anySupportAcquireB) {
     edge.Acquire(UInt(0), acquire_address, lgCacheBlockBytes, s2_grow_param)._2 // Cacheability checked by tlb
   } else {
@@ -523,9 +523,8 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   // AMOs
   if (usingAtomics) {
     val amoalu = Module(new AMOALU(xLen))
-    amoalu.io.addr := pstore1_addr
+    amoalu.io.mask := pstore1_mask
     amoalu.io.cmd := pstore1_cmd
-    amoalu.io.typ := pstore1_typ
     amoalu.io.lhs := s2_data_word
     amoalu.io.rhs := pstore1_data
     pstore1_storegen_data := amoalu.io.out

src/main/scala/rocket/HellaCache.scala

@@ -118,11 +118,16 @@ class HellaCacheExceptions extends Bundle {
   val ae = new AlignmentExceptions
 }
 
+class HellaCacheWriteData(implicit p: Parameters) extends CoreBundle()(p) {
+  val data = UInt(width = coreDataBits)
+  val mask = UInt(width = coreDataBytes)
+}
+
 // interface between D$ and processor/DTLB
 class HellaCacheIO(implicit p: Parameters) extends CoreBundle()(p) {
   val req = Decoupled(new HellaCacheReq)
   val s1_kill = Bool(OUTPUT) // kill previous cycle's req
-  val s1_data = Bits(OUTPUT, coreDataBits) // data for previous cycle's req
+  val s1_data = new HellaCacheWriteData().asOutput // data for previous cycle's req
   val s2_nack = Bool(INPUT) // req from two cycles ago is rejected
 
   // performance events

src/main/scala/rocket/IDecode.scala

@@ -129,7 +129,7 @@ class IDecode(implicit val p: Parameters) extends DecodeConstants
 class SDecode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    SFENCE_VMA->List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_SFENCE,   MT_X, N,N,N,N,N,N,CSR.N,N,N,N,N),
+    SFENCE_VMA->List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_SFENCE,   MT_W, N,N,N,N,N,N,CSR.N,N,N,N,N),
     SRET->      List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.I,N,N,N,N))
 }
 

src/main/scala/rocket/NBDcache.scala

@@ -69,20 +69,18 @@ class IOMSHR(id: Int)(implicit edge: TLEdgeOut, p: Parameters) extends L1HellaCa
   }
 
   val req = Reg(new HellaCacheReq)
-  val req_cmd_sc = req.cmd === M_XSC
   val grant_word = Reg(UInt(width = wordBits))
 
   val s_idle :: s_mem_access :: s_mem_ack :: s_resp :: Nil = Enum(Bits(), 4)
   val state = Reg(init = s_idle)
   io.req.ready := (state === s_idle)
 
-  val storegen = new StoreGen(req.typ, req.addr, req.data, wordBytes)
+  val loadgen = new LoadGen(req.typ, mtSigned(req.typ), req.addr, grant_word, false.B, wordBytes)
-  val loadgen = new LoadGen(req.typ, mtSigned(req.typ), req.addr, grant_word, req_cmd_sc, wordBytes)
  
   val a_source = UInt(id)
   val a_address = req.addr
-  val a_size = storegen.size
+  val a_size = mtSize(req.typ)
-  val a_data = Fill(beatWords, storegen.data)
+  val a_data = Fill(beatWords, req.data)
 
   val get     = edge.Get(a_source, a_address, a_size)._2
   val put     = edge.Put(a_source, a_address, a_size, a_data)._2
@@ -99,9 +97,10 @@ class IOMSHR(id: Int)(implicit edge: TLEdgeOut, p: Parameters) extends L1HellaCa
       M_XA_MAXU -> edge.Arithmetic(a_source, a_address, a_size, a_data, TLAtomics.MAXU)._2))
   } else {
     // If no managers support atomics, assert fail if processor asks for them
-    assert (!isAMO(req.cmd))
+    assert(state === s_idle || !isAMO(req.cmd))
     Wire(new TLBundleA(edge.bundle))
   }
+  assert(state === s_idle || req.cmd =/= M_XSC)
 
   io.mem_access.valid := (state === s_mem_access)
   io.mem_access.bits := Mux(isAMO(req.cmd), atomics, Mux(isRead(req.cmd), get, put))
@@ -110,7 +109,7 @@ class IOMSHR(id: Int)(implicit edge: TLEdgeOut, p: Parameters) extends L1HellaCa
   io.resp.valid := (state === s_resp)
   io.resp.bits := req
   io.resp.bits.has_data := isRead(req.cmd)
-  io.resp.bits.data := loadgen.data | req_cmd_sc
+  io.resp.bits.data := loadgen.data
   io.resp.bits.store_data := req.data
   io.resp.bits.replay := Bool(true)
 
@@ -696,6 +695,8 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   val s1_read  = isRead(s1_req.cmd)
   val s1_write = isWrite(s1_req.cmd)
   val s1_readwrite = s1_read || s1_write || isPrefetch(s1_req.cmd)
+  // check for unsupported operations
+  assert(!s1_valid || !s1_req.cmd.isOneOf(M_PWR))
 
   val dtlb = Module(new TLB(log2Ceil(coreDataBytes), nTLBEntries))
   io.ptw <> dtlb.io.ptw
@@ -703,7 +704,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   dtlb.io.req.bits.sfence.valid := s1_sfence
   dtlb.io.req.bits.sfence.bits.rs1 := s1_req.typ(0)
   dtlb.io.req.bits.sfence.bits.rs2 := s1_req.typ(1)
-  dtlb.io.req.bits.sfence.bits.asid := io.cpu.s1_data
+  dtlb.io.req.bits.sfence.bits.asid := io.cpu.s1_data.data
   dtlb.io.req.bits.passthrough := s1_req.phys
   dtlb.io.req.bits.vaddr := s1_req.addr
   dtlb.io.req.bits.instruction := Bool(false)
@@ -736,7 +737,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
     s2_req.phys := s1_req.phys
     s2_req.addr := s1_addr
     when (s1_write) {
-      s2_req.data := Mux(s1_replay, mshrs.io.replay.bits.data, io.cpu.s1_data)
+      s2_req.data := Mux(s1_replay, mshrs.io.replay.bits.data, io.cpu.s1_data.data)
     }
     when (s1_recycled) { s2_req.data := s1_req.data }
     s2_req.tag := s1_req.tag
@@ -927,10 +928,9 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   val s2_data_word_prebypass = s2_data_uncorrected >> Cat(s2_word_idx, Bits(0,log2Up(coreDataBits)))
   val s2_data_word = Mux(s2_store_bypass, s2_store_bypass_data, s2_data_word_prebypass)
   val loadgen = new LoadGen(s2_req.typ, mtSigned(s2_req.typ), s2_req.addr, s2_data_word, s2_sc, wordBytes)
-  
+
-  amoalu.io.addr := s2_req.addr
+  amoalu.io.mask := new StoreGen(s2_req.typ, s2_req.addr, 0.U, xLen/8).mask
   amoalu.io.cmd := s2_req.cmd
-  amoalu.io.typ := s2_req.typ
   amoalu.io.lhs := s2_data_word
   amoalu.io.rhs := s2_req.data
 

src/main/scala/rocket/Rocket.scala

@@ -394,7 +394,8 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
     mem_reg_pc := ex_reg_pc
     mem_reg_wdata := alu.io.out
     when (ex_ctrl.rxs2 && (ex_ctrl.mem || ex_ctrl.rocc)) {
-      mem_reg_rs2 := ex_rs(1)
+      val typ = Mux(ex_ctrl.rocc, log2Ceil(xLen/8).U, ex_ctrl.mem_type)
+      mem_reg_rs2 := new uncore.util.StoreGen(typ, 0.U, ex_rs(1), coreDataBytes).data
     }
   }
 
@@ -625,7 +626,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   io.dmem.req.bits.phys := Bool(false)
   io.dmem.req.bits.addr := encodeVirtualAddress(ex_rs(0), alu.io.adder_out)
   io.dmem.invalidate_lr := wb_xcpt
-  io.dmem.s1_data := Mux(mem_ctrl.fp, io.fpu.store_data, mem_reg_rs2)
+  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

src/main/scala/rocket/ScratchpadSlavePort.scala

@@ -24,7 +24,7 @@ class ScratchpadSlavePort(address: AddressSet)(implicit p: Parameters) extends L
       executable         = true,
       supportsArithmetic = if (usingAtomics) TransferSizes(4, coreDataBytes) else TransferSizes.none,
       supportsLogical    = if (usingAtomics) TransferSizes(4, coreDataBytes) else TransferSizes.none,
-      supportsPutPartial = TransferSizes.none, // Can't support PutPartial
+      supportsPutPartial = TransferSizes(1, coreDataBytes),
       supportsPutFull    = TransferSizes(1, coreDataBytes),
       supportsGet        = TransferSizes(1, coreDataBytes),
       fifoId             = Some(0))), // requests handled in FIFO order
@@ -55,6 +55,7 @@ class ScratchpadSlavePort(address: AddressSet)(implicit p: Parameters) extends L
       val req = Wire(new HellaCacheReq)
       req.cmd := MuxLookup(a.opcode, Wire(M_XRD), Array(
         TLMessages.PutFullData    -> M_XWR,
+        TLMessages.PutPartialData -> M_PWR,
         TLMessages.ArithmeticData -> MuxLookup(a.param, Wire(M_XRD), Array(
           TLAtomics.MIN           -> M_XA_MIN,
           TLAtomics.MAX           -> M_XA_MAX,
@@ -67,9 +68,8 @@ class ScratchpadSlavePort(address: AddressSet)(implicit p: Parameters) extends L
           TLAtomics.AND           -> M_XA_AND,
           TLAtomics.SWAP          -> M_XA_SWAP)),
         TLMessages.Get            -> M_XRD))
-      // treat all loads as full words, so bytes appear in correct lane
+      req.typ := a.size
-      req.typ := Mux(edge.hasData(a), a.size, log2Ceil(coreDataBytes))
+      req.addr := a.address
-      req.addr := Mux(edge.hasData(a), a.address, ~(~a.address | (coreDataBytes-1)))
       req.tag := UInt(0)
       req.phys := true
       req
@@ -79,14 +79,13 @@ class ScratchpadSlavePort(address: AddressSet)(implicit p: Parameters) extends L
     io.dmem.req.valid := (tl_in.a.valid && ready) || state === s_replay
     tl_in.a.ready := io.dmem.req.ready && ready
     io.dmem.req.bits := formCacheReq(Mux(state === s_replay, acq, tl_in.a.bits))
-    // the TL data is already in the correct byte lane, but the D$
+    io.dmem.s1_data.data := acq.data
-    // expects right-justified store data, so that it can steer the bytes.
+    io.dmem.s1_data.mask := acq.mask
-    io.dmem.s1_data := new LoadGen(acq.size, Bool(false), acq.address(log2Ceil(coreDataBytes)-1,0), acq.data, Bool(false), coreDataBytes).data
     io.dmem.s1_kill := false
     io.dmem.invalidate_lr := false
 
     tl_in.d.valid := io.dmem.resp.valid || state === s_grant
-    tl_in.d.bits := Mux(acq.opcode === TLMessages.PutFullData,
+    tl_in.d.bits := Mux(acq.opcode.isOneOf(TLMessages.PutFullData, TLMessages.PutPartialData),
       edge.AccessAck(acq, UInt(0)),
       edge.AccessAck(acq, UInt(0), UInt(0)))
     tl_in.d.bits.data := Mux(io.dmem.resp.valid, io.dmem.resp.bits.data_raw, acq.data)

src/main/scala/rocket/SimpleHellaCacheIF.scala

@@ -123,7 +123,7 @@ class SimpleHellaCacheIF(implicit p: Parameters) extends Module
   io.cache.invalidate_lr := io.requestor.invalidate_lr
   io.cache.req <> req_arb.io.out
   io.cache.s1_kill := io.cache.s2_nack
-  io.cache.s1_data := RegEnable(req_arb.io.out.bits.data, s0_req_fire)
+  io.cache.s1_data.data := RegEnable(req_arb.io.out.bits.data, s0_req_fire)
 
   replayq.io.nack.valid := (io.cache.s2_nack || s2_kill) && s2_req_fire
   replayq.io.nack.bits := s2_req_tag

src/main/scala/tile/FPU.scala

@@ -392,7 +392,7 @@ class FPToInt(implicit p: Parameters) extends FPUModule()(p) {
 
   val store = ieee(in.in1)
   val toint = Mux(in.rm(0), classify_out, store)
-  io.out.bits.store := store
+  io.out.bits.store := Mux(in.singleOut, Fill(xLen/32, store(31, 0)), store)
   io.out.bits.toint := Mux(in.singleOut, toint(31, 0).sextTo(xLen), toint)
   io.out.bits.exc := Bits(0)
 

src/main/scala/uncore/Consts.scala

@@ -28,7 +28,8 @@ trait MemoryOpConstants {
   def M_XA_MINU = UInt("b01110");
   def M_XA_MAXU = UInt("b01111");
   def M_FLUSH   = UInt("b10000") // write back dirty data and cede R/W permissions
-  def M_PRODUCE = UInt("b10001") // write back dirty data and cede W permissions
+  def M_PWR     = UInt("b10001") // partial (masked) store
+  def M_PRODUCE = UInt("b10010") // write back dirty data and cede W permissions
   def M_CLEAN   = UInt("b10011") // write back dirty data and retain R/W permissions
   def M_SFENCE  = UInt("b10100") // flush TLB
 
@@ -37,7 +38,7 @@ trait MemoryOpConstants {
   def isAMO(cmd: UInt) = isAMOLogical(cmd) || isAMOArithmetic(cmd)
   def isPrefetch(cmd: UInt) = cmd === M_PFR || cmd === M_PFW
   def isRead(cmd: UInt) = cmd === M_XRD || cmd === M_XLR || cmd === M_XSC || isAMO(cmd)
-  def isWrite(cmd: UInt) = cmd === M_XWR || cmd === M_XSC || isAMO(cmd)
+  def isWrite(cmd: UInt) = cmd === M_XWR || cmd === M_PWR || cmd === M_XSC || isAMO(cmd)
   def isWriteIntent(cmd: UInt) = isWrite(cmd) || cmd === M_PFW || cmd === M_XLR
 }
 

src/main/scala/uncore/util/AmoAlu.scala

@@ -31,10 +31,6 @@ class StoreGen(typ: UInt, addr: UInt, dat: UInt, maxSize: Int) {
   def wordData = genData(2)
 }
 
-class StoreGenAligned(typ: UInt, addr: UInt, dat: UInt, maxSize: Int) extends StoreGen(typ, addr, dat, maxSize) {
-  override def genData(i: Int) = dat
-}
-
 class LoadGen(typ: UInt, signed: Bool, addr: UInt, dat: UInt, zero: Bool, maxSize: Int) {
   private val size = new StoreGen(typ, addr, dat, maxSize).size
 
@@ -54,22 +50,16 @@ class LoadGen(typ: UInt, signed: Bool, addr: UInt, dat: UInt, zero: Bool, maxSiz
   def data = genData(0)
 }
 
-class AMOALU(operandBits: Int, rhsIsAligned: Boolean = false)(implicit p: Parameters) extends Module {
+class AMOALU(operandBits: Int)(implicit p: Parameters) extends Module {
   require(operandBits == 32 || operandBits == 64)
   val io = new Bundle {
-    val addr = Bits(INPUT, log2Ceil(operandBits/8))
+    val mask = UInt(INPUT, operandBits/8)
     val cmd = Bits(INPUT, M_SZ)
-    val typ = Bits(INPUT, log2Ceil(log2Ceil(operandBits/8) + 1))
     val lhs = Bits(INPUT, operandBits)
     val rhs = Bits(INPUT, operandBits)
     val out = Bits(OUTPUT, operandBits)
   }
 
-  val storegen =
-    if(rhsIsAligned) new StoreGenAligned(io.typ, io.addr, io.rhs, operandBits/8)
-    else new StoreGen(io.typ, io.addr, io.rhs, operandBits/8)
-  val rhs = storegen.wordData
-  
   val max = io.cmd === M_XA_MAX || io.cmd === M_XA_MAXU
   val min = io.cmd === M_XA_MIN || io.cmd === M_XA_MINU
   val add = io.cmd === M_XA_ADD
@@ -77,10 +67,10 @@ class AMOALU(operandBits: Int, rhsIsAligned: Boolean = false)(implicit p: Parame
   val logic_xor = io.cmd === M_XA_XOR || io.cmd === M_XA_OR
 
   val adder_out =
-    if (operandBits == 32) io.lhs + rhs
+    if (operandBits == 32) io.lhs + io.rhs
     else {
-      val mask = ~UInt(0,64) ^ (io.addr(2) << 31)
+      val mask = ~UInt(0,64) ^ (!io.mask(3) << 31)
-      (io.lhs & mask) + (rhs & mask)
+      (io.lhs & mask) + (io.rhs & mask)
     }
 
   val less = {
@@ -90,28 +80,29 @@ class AMOALU(operandBits: Int, rhsIsAligned: Boolean = false)(implicit p: Parame
     }
 
     if (operandBits == 32) {
-      Mux(io.lhs(31) === rhs(31), io.lhs < rhs, Mux(sgned, io.lhs(31), io.rhs(31)))
+      Mux(io.lhs(31) === io.rhs(31), io.lhs < io.rhs, Mux(sgned, io.lhs(31), io.rhs(31)))
     } else {
-      val word = !io.typ(0)
+      val cmp_lhs = Mux(!io.mask(4), io.lhs(31), io.lhs(63))
-      val cmp_lhs = Mux(word && !io.addr(2), io.lhs(31), io.lhs(63))
+      val cmp_rhs = Mux(!io.mask(4), io.rhs(31), io.rhs(63))
-      val cmp_rhs = Mux(word && !io.addr(2), rhs(31), rhs(63))
+      val lt_lo = io.lhs(31,0) < io.rhs(31,0)
-      val lt_lo = io.lhs(31,0) < rhs(31,0)
+      val lt_hi = io.lhs(63,32) < io.rhs(63,32)
-      val lt_hi = io.lhs(63,32) < rhs(63,32)
+      val eq_hi = io.lhs(63,32) === io.rhs(63,32)
-      val eq_hi = io.lhs(63,32) === rhs(63,32)
+      val lt =
-      val lt = Mux(word, Mux(io.addr(2), lt_hi, lt_lo), lt_hi || eq_hi && lt_lo)
+        Mux(io.mask(4) && io.mask(3), lt_hi || eq_hi && lt_lo,
+        Mux(io.mask(4), lt_hi, lt_lo))
       Mux(cmp_lhs === cmp_rhs, lt, Mux(sgned, cmp_lhs, cmp_rhs))
     }
   }
 
-  val minmax = Mux(Mux(less, min, max), io.lhs, storegen.data)
+  val minmax = Mux(Mux(less, min, max), io.lhs, io.rhs)
   val logic =
-    Mux(logic_and, io.lhs & rhs, 0.U) |
+    Mux(logic_and, io.lhs & io.rhs, 0.U) |
-    Mux(logic_xor, io.lhs ^ rhs, 0.U)
+    Mux(logic_xor, io.lhs ^ io.rhs, 0.U)
   val out =
     Mux(add,                    adder_out,
     Mux(logic_and || logic_xor, logic,
                                 minmax))
 
-  val wmask = FillInterleaved(8, storegen.mask)
+  val wmask = FillInterleaved(8, io.mask)
   io.out := wmask & out | ~wmask & io.lhs
 }