Issue 1766 fix: lw returns wrong value when access coincides with ITLB miss

sim/questa/coverage-exclusions-rv64gc.do

@@ -328,6 +328,9 @@ set line [GetLineNum ${SRC}/mmu/pmpchecker.sv "EnforcePMP & ExecuteAccessF"]
 coverage exclude -scope /dut/core/lsu/dmmu/dmmu/pmp/pmpchecker -linerange $line-$line -item e 1 -fecexprrow 1,2,4,5,6
 set line [GetLineNum ${SRC}/mmu/pmpchecker.sv "EnforcePMP & ExecuteAccessF"]
 coverage exclude -scope /dut/core/ifu/immu/immu/pmp/pmpchecker -linerange $line-$line -item e 1 -fecexprrow 3
+## A single data access is never simultaneously read and write, so WriteAccessM=1 can never occur
+## while the ReadAccessM-qualified load-access-fault term is evaluated (PMPLoadAccessFaultM, row 2).
+set line [GetLineNum ${SRC}/mmu/pmpchecker.sv "EnforcePMP & ReadAccessM"]
 
 
 ## The IFU has ReadAccess = WriteAccess = 0 and ExecuteAccess = 1 hardwired, so exclude alternatives
@@ -412,10 +415,6 @@ coverage exclude -scope /dut/core/lsu/hptw/hptw -linerange $line-$line -item e 1
 set line [GetLineNum ${SRC}/mmu/hptw.sv "assign HPTWUpdateDA"]
 coverage exclude -scope /dut/core/lsu/hptw/hptw -linerange $line-$line -item e 1 -fecexprrow 3
 
-# NOTE: the hptw instance is at /dut/core/lsu/hptw/hptw (single lsu).  The four exclusions above use a
-# stale /dut/core/lsu/hptw/hptw path that no longer resolves (silent no-ops) -- left untouched here;
-# the two exclusions below use the correct path.
-
 # UPDATE_PTE never self-loops on a cache-bus stall: the PTE being A/D-updated was just read during the
 # same uninterrupted walk (LSU stalled, line cannot be evicted), so it is resident/writable in the D$ and
 # the UPDATE_PTE store hits.  DCacheBusStallM therefore cannot assert in UPDATE_PTE, so the

src/mmu/hptw.sv

@@ -88,6 +88,8 @@ module hptw import cvw::*;  #(parameter cvw_t P) (
   logic                     ValidPTE, LeafPTE, ValidLeafPTE, ValidNonLeafPTE;
   logic                     StartWalk;
   logic                     TLBMissOrUpdateDA;
+  logic                     ITLBMissReady;
+  logic                     MStageMemPending;
   logic                     PRegEn;
   logic [2:0]               NextPageType;
   logic [P.SVMODE_BITS-1:0] SvMode;
@@ -143,7 +145,25 @@ module hptw import cvw::*;  #(parameter cvw_t P) (
   // Extract bits from CSRs and inputs
   assign SvMode = SATP_REGW[P.XLEN-1:P.XLEN-P.SVMODE_BITS];
   assign BasePageTablePPN = SATP_REGW[P.PPN_BITS-1:0];
-  assign TLBMissOrUpdateDA = DTLBMissOrUpdateDAM | ITLBMissOrUpdateAF;
+  // Defer servicing an instruction-side (ITLB) miss while a committed M-stage data access
+  // (load/store/AMO/CMO) is still present.  At IDLE, SelHPTW=0 so DCacheBusStallM reflects the LSU
+  // data op (not an HPTW access).  Two hazards motivate the deferral:
+  //   (#1538) Starting the walk while the data op's cache/bus access is mid-flight would seize the
+  //           data cache and drop a committed store.  Covered by ~DCacheBusStallM.
+  //   (#1766) Even after the data op has *hit* in the cache (DCacheBusStallM low), the load's read
+  //           data is only valid combinationally while the access still owns the cache.  If the
+  //           pipeline is stalled for another reason (here HPTWStall itself, from the pending ITLB
+  //           miss) the load is held in M; starting the walk now re-points the cache SRAMs at the
+  //           PTE sets and the load captures stale data when it finally retires.  Covered by
+  //           ~MStageMemPending: hold the walk until the data op leaves the M stage (MemRWM/CMOpM
+  //           deassert), then start the walk against an idle cache.
+  // This cannot deadlock: the data op's retirement never depends on the (younger) instruction-fetch
+  // walk, so it drains through the normal LSU stall and the walk starts once it leaves M.  Because
+  // TLBMissOrUpdateDA is low while deferred, HPTWStall is also low for the ITLB miss, so the HPTW
+  // does not freeze the pipeline against the data op.  DTLB misses are never deferred.
+  assign MStageMemPending  = (|MemRWM) | (|CMOpM);
+  assign ITLBMissReady     = ITLBMissOrUpdateAF & ~DCacheBusStallM & ~MStageMemPending;
+  assign TLBMissOrUpdateDA = DTLBMissOrUpdateDAM | ITLBMissReady;
 
   // Determine which address to translate
   mux2 #(P.XLEN) vadrmux(PCSpillF, IEUAdrExtM[P.XLEN-1:0], DTLBWalk, TranslationVAdr);