diff --git a/src/main/scala/rocket/DCache.scala b/src/main/scala/rocket/DCache.scala
index e83796ea..2feae0e0 100644
--- a/src/main/scala/rocket/DCache.scala
+++ b/src/main/scala/rocket/DCache.scala
@@ -144,7 +144,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   when (!metaArb.io.in(7).ready) { io.cpu.req.ready := false }
 
   // address translation
-  val tlb = Module(new TLB(log2Ceil(coreDataBytes), nTLBEntries))
+  val tlb = Module(new TLB(false, log2Ceil(coreDataBytes), nTLBEntries))
   io.ptw <> tlb.io.ptw
   tlb.io.req.valid := s1_valid && !io.cpu.s1_kill && (s1_readwrite || s1_sfence)
   tlb.io.req.bits.sfence.valid := s1_sfence
@@ -154,7 +154,6 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   tlb.io.req.bits.sfence.bits.addr := s1_req.addr
   tlb.io.req.bits.passthrough := s1_req.phys
   tlb.io.req.bits.vaddr := s1_req.addr
-  tlb.io.req.bits.instruction := false
   tlb.io.req.bits.size := s1_req.typ
   tlb.io.req.bits.cmd := s1_req.cmd
   when (!tlb.io.req.ready && !tlb.io.ptw.resp.valid && !io.cpu.req.bits.phys) { io.cpu.req.ready := false }
@@ -201,7 +200,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   when (s1_valid_not_nacked || s1_flush_valid) {
     s2_req := s1_req
     s2_req.addr := s1_paddr
-    s2_uncached := !tlb.io.resp.cacheable || Bool(usingDataScratchpad)
+    s2_uncached := !tlb.io.resp.cacheable
   }
   val s2_read = isRead(s2_req.cmd)
   val s2_write = isWrite(s2_req.cmd)
diff --git a/src/main/scala/rocket/Frontend.scala b/src/main/scala/rocket/Frontend.scala
index b4088bfe..465e521d 100644
--- a/src/main/scala/rocket/Frontend.scala
+++ b/src/main/scala/rocket/Frontend.scala
@@ -82,7 +82,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   require(fetchWidth*coreInstBytes == outer.icacheParams.fetchBytes)
 
   val fetchBytes = coreInstBytes * fetchWidth
-  val tlb = Module(new TLB(log2Ceil(fetchBytes), nTLBEntries))
+  val tlb = Module(new TLB(true, log2Ceil(fetchBytes), nTLBEntries))
   val fq = withReset(reset || io.cpu.req.valid) { Module(new ShiftQueue(new FrontendResp, 5, flow = true)) }
 
   val s0_valid = io.cpu.req.valid || !fq.io.mask(fq.io.mask.getWidth-3)
@@ -130,7 +130,6 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   tlb.io.req.valid := !s2_replay
   tlb.io.req.bits.vaddr := s1_pc
   tlb.io.req.bits.passthrough := Bool(false)
-  tlb.io.req.bits.instruction := Bool(true)
   tlb.io.req.bits.sfence := io.cpu.sfence
   tlb.io.req.bits.size := log2Ceil(coreInstBytes*fetchWidth)
 
diff --git a/src/main/scala/rocket/NBDcache.scala b/src/main/scala/rocket/NBDcache.scala
index fd96ec7f..8d9899d4 100644
--- a/src/main/scala/rocket/NBDcache.scala
+++ b/src/main/scala/rocket/NBDcache.scala
@@ -700,7 +700,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   // check for unsupported operations
   assert(!s1_valid || !s1_req.cmd.isOneOf(M_PWR))
 
-  val dtlb = Module(new TLB(log2Ceil(coreDataBytes), nTLBEntries))
+  val dtlb = Module(new TLB(false, log2Ceil(coreDataBytes), nTLBEntries))
   io.ptw <> dtlb.io.ptw
   dtlb.io.req.valid := s1_valid && !io.cpu.s1_kill && (s1_readwrite || s1_sfence)
   dtlb.io.req.bits.sfence.valid := s1_sfence
@@ -710,7 +710,6 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   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)
   dtlb.io.req.bits.size := s1_req.typ
   dtlb.io.req.bits.cmd := s1_req.cmd
   when (!dtlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := Bool(false) }
diff --git a/src/main/scala/rocket/TLB.scala b/src/main/scala/rocket/TLB.scala
index 14b65732..f01260b2 100644
--- a/src/main/scala/rocket/TLB.scala
+++ b/src/main/scala/rocket/TLB.scala
@@ -52,7 +52,7 @@ class TLBResp(implicit p: Parameters) extends CoreBundle()(p) {
   val prefetchable = Bool()
 }
 
-class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(p) {
+class TLB(instruction: Boolean, lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(p) {
   val io = new Bundle {
     val req = Decoupled(new TLBReq(lgMaxSize)).flip
     val resp = new TLBResp().asOutput
@@ -92,7 +92,7 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
   val r_refill_waddr = Reg(UInt(width = log2Ceil(normalEntries)))
   val r_req = Reg(new TLBReq(lgMaxSize))
 
-  val priv = Mux(io.req.bits.instruction, io.ptw.status.prv, io.ptw.status.dprv)
+  val priv = if (instruction) io.ptw.status.prv else io.ptw.status.dprv
   val priv_s = priv(0)
   val priv_uses_vm = priv <= PRV.S
   val vm_enabled = Bool(usingVM) && io.ptw.ptbr.mode(io.ptw.ptbr.mode.getWidth-1) && priv_uses_vm && !io.req.bits.passthrough
@@ -113,7 +113,7 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
   val legal_address = edge.manager.findSafe(mpu_physaddr).reduce(_||_)
   def fastCheck(member: TLManagerParameters => Boolean) =
     legal_address && edge.manager.fastProperty(mpu_physaddr, member, (b:Boolean) => Bool(b))
-  val cacheable = fastCheck(_.supportsAcquireB)
+  val cacheable = fastCheck(_.supportsAcquireB) && (instruction || !usingDataScratchpad)
   val homogeneous = TLBPageLookup(edge.manager.managers, xLen, p(CacheBlockBytes), BigInt(1) << pgIdxBits)(mpu_physaddr).homogeneous
   val prot_r = fastCheck(_.supportsGet) && pmp.io.r
   val prot_w = fastCheck(_.supportsPutFull) && pmp.io.w