diff --git a/src/main/scala/rocket/PMP.scala b/src/main/scala/rocket/PMP.scala
index b2c97537..5cae16ef 100644
--- a/src/main/scala/rocket/PMP.scala
+++ b/src/main/scala/rocket/PMP.scala
@@ -56,22 +56,23 @@ class PMP(implicit p: Parameters) extends PMPReg {
     }
   }
 
-  private def decomposedBoundMatch(xHi: UInt, xLo: UInt, lsbMask: UInt, lgMaxSize: Int) = {
-    val msbsLess = xHi < (comparand >> lgMaxSize)
-    val msbsEqual = (xHi ^ (comparand >> lgMaxSize)) === 0
-    val lsbsLess = xLo < (comparand(lgMaxSize-1, 0) & lsbMask(lgMaxSize-1, 0))
-    msbsLess || (msbsEqual && lsbsLess)
+  private def boundMatch(x: UInt, lsbMask: UInt, lgMaxSize: Int) = {
+    if (lgMaxSize <= lgAlign) {
+      x < comparand
+    } else {
+      // break up the circuit; the MSB part will be CSE'd
+      val msbsLess = (x >> lgMaxSize) < (comparand >> lgMaxSize)
+      val msbsEqual = ((x >> lgMaxSize) ^ (comparand >> lgMaxSize)) === 0
+      val lsbsLess =  (x(lgMaxSize-1, 0) | lsbMask) < comparand(lgMaxSize-1, 0)
+      msbsLess || (msbsEqual && lsbsLess)
+    }
   }
 
-  private def boundMatch(x: UInt, lsbMask: UInt, lgMaxSize: Int) =
-    if (lgMaxSize <= lgAlign) x < comparand
-    else decomposedBoundMatch(x >> lgMaxSize, x(lgMaxSize-1, 0), lsbMask, lgMaxSize)
-
   private def lowerBoundMatch(x: UInt, lgSize: UInt, lgMaxSize: Int) =
-    !boundMatch(x, ((BigInt(1) << lgMaxSize) - 1).U << lgSize, lgMaxSize)
+    !boundMatch(x, ~(((BigInt(1) << lgMaxSize) - 1).U << lgSize)(lgMaxSize-1, 0), lgMaxSize)
 
   private def upperBoundMatch(x: UInt, lgMaxSize: Int) =
-    boundMatch(x, ((BigInt(1) << lgMaxSize) - 1).U, lgMaxSize)
+    boundMatch(x, 0.U, lgMaxSize)
 
   private def rangeMatch(x: UInt, lgSize: UInt, lgMaxSize: Int, prev: PMP) =
     prev.lowerBoundMatch(x, lgSize, lgMaxSize) && upperBoundMatch(x, lgMaxSize)
@@ -85,26 +86,25 @@ class PMP(implicit p: Parameters) extends PMPReg {
     f(pgIdxBits + (pgLevels - 1 - i) * pgLevelBits)
   }
 
-  private def rangeHomogeneous(x: UInt, pgLevel: UInt, lgMaxSize: Int, prev: PMP) = {
-    val beginsAfterLower = !prev.boundMatch(x, ((BigInt(1) << lgMaxSize) - 1).U, lgMaxSize) // CSE with rangeMatch
-    val beginsAfterUpper = !boundMatch(x, ((BigInt(1) << lgMaxSize) - 1).U, lgMaxSize) // CSE with rangeMatch
+  private def rangeHomogeneous(x: UInt, pgLevel: UInt, prev: PMP) = {
+    val beginsAfterLower = !(x < prev.comparand)
+    val beginsAfterUpper = !(x < comparand)
     val endsBeforeLower = pgLevelMap { idxBits => (x >> idxBits) < (prev.comparand << idxBits) } (pgLevel)
-    val endsBeforeUpper = pgLevelMap { idxBits => (x >> idxBits) > (comparand << idxBits) } (pgLevel)
+    val endsBeforeUpper = pgLevelMap { idxBits => (x >> idxBits) < (comparand << idxBits) } (pgLevel)
     endsBeforeLower || beginsAfterUpper || (beginsAfterLower && endsBeforeUpper)
   }
 
   // returns whether this PMP completely contains, or contains none of, a page
-  def homogeneous(x: UInt, pgLevel: UInt, lgMaxSize: Int, prev: PMP): Bool =
-    !cfg.p(0) || Mux(cfg.a(1), rangeHomogeneous(x, pgLevel, lgMaxSize, prev), pow2Homogeneous(x, pgLevel))
+  def homogeneous(x: UInt, pgLevel: UInt, prev: PMP): Bool =
+    !cfg.p(0) || Mux(cfg.a(1), rangeHomogeneous(x, pgLevel, prev), pow2Homogeneous(x, pgLevel))
 
   // returns whether this matching PMP fully contains the access
   def aligned(x: UInt, lgSize: UInt, lgMaxSize: Int, prev: PMP): Bool = if (lgMaxSize <= lgAlign) true.B else {
-    val lsbMask = ((BigInt(1) << lgMaxSize) - 1).U << lgSize
-    val alignMask = ~lsbMask(lgMaxSize-1, 0)
-    val lowerBoundOK = !prev.upperBoundMatch(x, lgMaxSize)
-    val upperBoundOK = decomposedBoundMatch(x >> lgMaxSize, x(lgMaxSize-1, 0) | alignMask, lsbMask, lgMaxSize)
-    val rangeAligned = lowerBoundOK && upperBoundOK
-    val pow2Aligned = (alignMask & ~mask(lgMaxSize-1, 0)) === 0
+    val lsbMask = ~(((BigInt(1) << lgMaxSize) - 1).U << lgSize)(lgMaxSize-1, 0)
+    val straddlesLowerBound = ((x >> lgMaxSize) ^ (prev.comparand >> lgMaxSize)) === 0 && (prev.comparand(lgMaxSize-1, 0) & ~x(lgMaxSize-1, 0)) =/= 0
+    val straddlesUpperBound = ((x >> lgMaxSize) ^ (comparand >> lgMaxSize)) === 0 && (comparand(lgMaxSize-1, 0) & (x(lgMaxSize-1, 0) | lsbMask)) =/= 0
+    val rangeAligned = !(straddlesLowerBound || straddlesUpperBound)
+    val pow2Aligned = (lsbMask & ~mask(lgMaxSize-1, 0)) === 0
     Mux(cfg.a(1), rangeAligned, pow2Aligned)
   }
 
@@ -113,6 +113,14 @@ class PMP(implicit p: Parameters) extends PMPReg {
     cfg.p(0) && Mux(cfg.a(1), rangeMatch(x, lgSize, lgMaxSize, prev), pow2Match(x, lgSize, lgMaxSize))
 }
 
+class PMPHomogeneityChecker(pmps: Seq[PMP])(implicit p: Parameters) {
+  def apply(addr: UInt, pgLevel: UInt): Bool = {
+    ((true.B, 0.U.asTypeOf(new PMP)) /: pmps) { case ((h, prev), pmp) =>
+      (h && pmp.homogeneous(addr, pgLevel, prev), pmp)
+    }._1
+  }
+}
+
 class PMPChecker(lgMaxSize: Int)(implicit p: Parameters) extends CoreModule()(p)
     with HasRocketCoreParameters {
   val io = new Bundle {
@@ -123,8 +131,6 @@ class PMPChecker(lgMaxSize: Int)(implicit p: Parameters) extends CoreModule()(p)
     val r = Bool(OUTPUT)
     val w = Bool(OUTPUT)
     val x = Bool(OUTPUT)
-    val pgLevel = UInt(INPUT, log2Ceil(pgLevels))
-    val homogeneous = Bool(OUTPUT)
   }
 
   val default = io.prv > PRV.S
@@ -133,10 +139,6 @@ class PMPChecker(lgMaxSize: Int)(implicit p: Parameters) extends CoreModule()(p)
   pmp0.cfg.w := default
   pmp0.cfg.x := default
 
-  io.homogeneous := ((true.B, pmp0) /: io.pmp) { case ((h, prev), pmp) =>
-    (h && pmp.homogeneous(io.addr, io.pgLevel, lgMaxSize, prev), pmp)
-  }._1
-
   val res = (pmp0 /: (io.pmp zip (pmp0 +: io.pmp)).reverse) { case (prev, (pmp, prevPMP)) =>
     val hit = pmp.hit(io.addr, io.size, lgMaxSize, prevPMP)
     val ignore = default && !pmp.cfg.m
diff --git a/src/main/scala/rocket/PTW.scala b/src/main/scala/rocket/PTW.scala
index b50b61b9..6e150cc7 100644
--- a/src/main/scala/rocket/PTW.scala
+++ b/src/main/scala/rocket/PTW.scala
@@ -7,7 +7,9 @@ import Chisel._
 import Chisel.ImplicitConversions._
 import config._
 import tile._
+import coreplex.CacheBlockBytes
 import uncore.constants._
+import uncore.tilelink2._
 import util._
 
 import scala.collection.mutable.ListBuffer
@@ -19,6 +21,7 @@ class PTWReq(implicit p: Parameters) extends CoreBundle()(p) {
 class PTWResp(implicit p: Parameters) extends CoreBundle()(p) {
   val pte = new PTE
   val level = UInt(width = log2Ceil(pgLevels))
+  val homogeneous = Bool()
 }
 
 class TLBPTWIO(implicit p: Parameters) extends CoreBundle()(p)
@@ -60,7 +63,7 @@ class PTE(implicit p: Parameters) extends CoreBundle()(p) {
   def sx(dummy: Int = 0) = leaf() && x
 }
 
-class PTW(n: Int)(implicit p: Parameters) extends CoreModule()(p) {
+class PTW(n: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(p) {
   val io = new Bundle {
     val requestor = Vec(n, new TLBPTWIO).flip
     val mem = new HellaCacheIO
@@ -131,11 +134,18 @@ class PTW(n: Int)(implicit p: Parameters) extends CoreModule()(p) {
   io.mem.s1_kill := s1_kill
   io.mem.invalidate_lr := Bool(false)
   
+  val pmaPgLevelHomogeneous = (0 until pgLevels) map { i =>
+    TLBPageLookup(edge.manager.managers, xLen, p(CacheBlockBytes), BigInt(1) << (pgIdxBits + ((pgLevels - 1 - i) * pgLevelBits)))(pte_addr >> pgIdxBits << pgIdxBits).homogeneous
+  }
+  val pmaHomogeneous = pmaPgLevelHomogeneous(count)
+  val pmpHomogeneous = new PMPHomogeneityChecker(io.dpath.pmp).apply(pte_addr >> pgIdxBits << pgIdxBits, count)
+
   for (i <- 0 until io.requestor.size) {
     io.requestor(i).resp.valid := resp_valid(i)
     io.requestor(i).resp.bits.pte := r_pte
     io.requestor(i).resp.bits.level := count
     io.requestor(i).resp.bits.pte.ppn := pte_addr >> pgIdxBits
+    io.requestor(i).resp.bits.homogeneous := pmpHomogeneous && pmaHomogeneous
     io.requestor(i).ptbr := io.dpath.ptbr
     io.requestor(i).status := io.dpath.status
     io.requestor(i).pmp := io.dpath.pmp
@@ -195,6 +205,6 @@ trait CanHavePTW extends HasHellaCache {
 trait CanHavePTWModule extends HasHellaCacheModule {
   val outer: CanHavePTW
   val ptwPorts = ListBuffer(outer.dcache.module.io.ptw)
-  val ptwOpt = if (outer.usingPTW) { Some(Module(new PTW(outer.nPTWPorts)(outer.p))) } else None
+  val ptwOpt = if (outer.usingPTW) { Some(Module(new PTW(outer.nPTWPorts)(outer.dcache.node.edgesOut(0), outer.p))) } else None
   ptwOpt foreach { ptw => dcachePorts += ptw.io.mem }
 }
diff --git a/src/main/scala/rocket/TLB.scala b/src/main/scala/rocket/TLB.scala
index be1fece5..58cbdee1 100644
--- a/src/main/scala/rocket/TLB.scala
+++ b/src/main/scala/rocket/TLB.scala
@@ -73,14 +73,14 @@ class TLB(lgMaxSize: Int, entries: Int)(implicit edge: TLEdgeOut, p: Parameters)
   val refill_ppn = io.ptw.resp.bits.pte.ppn(ppnBits-1, 0)
   val do_refill = Bool(usingVM) && io.ptw.resp.valid
   val invalidate_refill = state.isOneOf(s_request /* don't care */, s_wait_invalidate)
-  val mpu_physaddr = Mux(do_refill, refill_ppn << pgIdxBits,
-                     Cat(Mux(vm_enabled, ppns.last, vpn(ppnBits-1, 0)), io.req.bits.vaddr(pgIdxBits-1, 0)))
+  val mpu_ppn = Mux(do_refill, refill_ppn,
+                Mux(vm_enabled, ppns.last, vpn(ppnBits-1, 0)))
+  val mpu_physaddr = Cat(mpu_ppn, io.req.bits.vaddr(pgIdxBits-1, 0))
   val pmp = Module(new PMPChecker(lgMaxSize))
   pmp.io.addr := mpu_physaddr
   pmp.io.size := io.req.bits.size
   pmp.io.pmp := io.ptw.pmp
   pmp.io.prv := Mux(do_refill || io.req.bits.passthrough /* PTW */, PRV.S, priv)
-  pmp.io.pgLevel := io.ptw.resp.bits.level
   val legal_address = edge.manager.findSafe(mpu_physaddr).reduce(_||_)
   def fastCheck(member: TLManagerParameters => Boolean) =
     legal_address && Mux1H(edge.manager.findFast(mpu_physaddr), edge.manager.managers.map(m => Bool(member(m))))
@@ -88,15 +88,7 @@ class TLB(lgMaxSize: Int, entries: Int)(implicit edge: TLEdgeOut, p: Parameters)
   val prot_w = fastCheck(_.supportsPutFull) && pmp.io.w
   val prot_x = fastCheck(_.executable) && pmp.io.x
   val cacheable = fastCheck(_.supportsAcquireB)
-  val isSpecial = !pmp.io.homogeneous || {
-    val homogeneous = Wire(init = false.B)
-    for (i <- 0 until pgLevels) {
-      when (io.ptw.resp.bits.level >= i) {
-        homogeneous := TLBPageLookup(edge.manager.managers, xLen, p(CacheBlockBytes), BigInt(1) << (pgIdxBits + ((pgLevels - 1 - i) * pgLevelBits)))(mpu_physaddr).homogeneous
-      }
-    }
-    !homogeneous
-  }
+  val isSpecial = !io.ptw.resp.bits.homogeneous
 
   val lookup_tag = Cat(io.ptw.ptbr.asid, vpn(vpnBits-1,0))
   val hitsVec = (0 until totalEntries).map { i => vm_enabled && {