updated riscv-bmarks and riscv-tests to build with new toolchain

rocket/src/main/scala/cpu.scala

@@ -53,12 +53,11 @@ class rocketProc extends Component
   dpath.io.host             ^^ io.host;
   ctrl.io.host.start        := io.host.start;
   dpath.io.debug            ^^ io.debug;
-//   dpath.io.imem.resp_data   ^^ io.imem.resp_data;
 
-
+  // FIXME: try to make this more compact
-  // FIXME: make this less verbose
+  
   // connect ITLB to I$, ctrl, dpath
-  itlb.io.cpu.invalidate  := dpath.io.ptbr_wen || ctrl.io.flush_inst;
+  itlb.io.cpu.invalidate  := dpath.io.ptbr_wen;
   itlb.io.cpu.status      := dpath.io.ctrl.status;
   itlb.io.cpu.req_val     := ctrl.io.imem.req_val;  
   itlb.io.cpu.req_asid    := Bits(0,ASID_BITS); // FIXME: connect to PCR
@@ -76,7 +75,6 @@ class rocketProc extends Component
 
   
   // connect DTLB to D$ arbiter, ctrl+dpath
-//   dtlb.io.cpu.invalidate  := Bool(false); // FIXME
   dtlb.io.cpu.invalidate  := dpath.io.ptbr_wen;
   dtlb.io.cpu.status      := dpath.io.ctrl.status;
   dtlb.io.cpu.req_val     := ctrl.io.dmem.req_val;  
@@ -99,14 +97,10 @@ class rocketProc extends Component
   arb.io.mem              ^^ io.dmem
   
   // connect arbiter to ctrl+dpath+DTLB
-//   arb.io.cpu.req_val      := dtlb.io.cpu.resp_val;
   arb.io.cpu.req_val      := ctrl.io.dmem.req_val;
   arb.io.cpu.req_cmd      := ctrl.io.dmem.req_cmd;
   arb.io.cpu.req_type     := ctrl.io.dmem.req_type;
-//   arb.io.cpu.dtlb_busy    := dtlb.io.cpu.resp_busy;
   arb.io.cpu.dtlb_miss    := dtlb.io.cpu.resp_miss;
-
-//   arb.io.cpu.req_addr     := dtlb.io.cpu.resp_addr;
   arb.io.cpu.req_idx      := dpath.io.dmem.req_addr(PGIDX_BITS-1,0);
   arb.io.cpu.req_ppn      := dtlb.io.cpu.resp_ppn;
   arb.io.cpu.req_data     := dpath.io.dmem.req_data;

rocket/src/main/scala/dcache.scala

@@ -48,7 +48,7 @@ class rocketDCacheStoreGen extends Component {
     val req_type      = Bits(3, 'input);
     val req_addr_lsb  = Bits(3, 'input);
     val req_data      = Bits(64, 'input);
-    val store_wmask   = Bits(64, 'output);
+    val store_wmask   = Bits(8, 'output);
     val store_data    = Bits(64, 'output);
   }
   
@@ -86,16 +86,8 @@ class rocketDCacheStoreGen extends Component {
     Mux(io.req_type === MT_D, wmask_d,
         UFix(0, 8)))));
   
-  io.store_wmask := 
+  io.store_wmask := store_wmask_byte;
-    Cat(Fill(8, store_wmask_byte(7)),
+  
-  		Fill(8, store_wmask_byte(6)),
-  		Fill(8, store_wmask_byte(5)),
-  		Fill(8, store_wmask_byte(4)),
-  		Fill(8, store_wmask_byte(3)),
-  		Fill(8, store_wmask_byte(2)),
-  		Fill(8, store_wmask_byte(1)),
-  		Fill(8, store_wmask_byte(0)));
-    
   io.store_data :=
     Mux(io.req_type === MT_B, Fill(8, io.req_data( 7,0)),
     Mux(io.req_type === MT_H, Fill(4, io.req_data(15,0)),
@@ -183,7 +175,7 @@ class rocketDCacheDM(lines: Int) extends Component {
   val offsetmsb = indexlsb-1;
   val offsetlsb = 3;
   
-  val s_reset :: s_ready :: s_replay_load :: s_start_writeback :: s_writeback :: s_req_refill :: s_refill :: s_resolve_miss :: Nil = Enum(8) { UFix() };
+  val s_reset :: s_ready :: s_replay_load :: s_write_amo :: s_start_writeback :: s_writeback :: s_req_refill :: s_refill :: s_resolve_miss :: Nil = Enum(9) { UFix() };
   val state = Reg(resetVal = s_reset);
   
   // idx arrives one clock cycle prior to ppn b/c of DTLB
@@ -194,6 +186,7 @@ class rocketDCacheDM(lines: Int) extends Component {
   val r_cpu_req_type   = Reg(resetVal = Bits(0,3));
   val r_cpu_req_tag    = Reg(resetVal = Bits(0,5));
   val r_cpu_resp_val   = Reg(resetVal = Bool(false));
+  val r_amo_data       = Reg(resetVal = Bits(0,64));
 
   val p_store_data      = Reg(resetVal = Bits(0,64));
   val p_store_idx       = Reg(resetVal = Bits(0,PGIDX_BITS));
@@ -203,11 +196,13 @@ class rocketDCacheDM(lines: Int) extends Component {
   val req_store   = (io.cpu.req_cmd === M_XWR);
   val req_load    = (io.cpu.req_cmd === M_XRD) || (io.cpu.req_cmd === M_PRD);
   val req_flush   = (io.cpu.req_cmd === M_FLA);
+  val req_amo     = io.cpu.req_cmd(3).toBool;
   val r_req_load  = (r_cpu_req_cmd === M_XRD) || (r_cpu_req_cmd === M_PRD);
   val r_req_store = (r_cpu_req_cmd === M_XWR);
   val r_req_flush = (r_cpu_req_cmd === M_FLA);
   val r_req_ptw_load = (r_cpu_req_cmd === M_PRD);
-
+  val r_req_amo     = r_cpu_req_cmd(3).toBool;
+  
   when (io.cpu.req_val && io.cpu.req_rdy) {
     r_cpu_req_idx   <== io.cpu.req_idx;
     r_cpu_req_cmd   <== io.cpu.req_cmd;
@@ -273,7 +268,7 @@ class rocketDCacheDM(lines: Int) extends Component {
 
   // load/store addresses conflict if they are to any part of the same 64 bit word
   val addr_match    = (r_cpu_req_idx(PGIDX_BITS-1,offsetlsb) === p_store_idx(PGIDX_BITS-1,offsetlsb));
-  val ldst_conflict = tag_valid && tag_match && r_req_load && p_store_valid && addr_match;
+  val ldst_conflict = tag_valid && tag_match && (r_req_load || r_req_amo) && p_store_valid && addr_match;
   val store_hit = r_cpu_req_val && !io.cpu.dtlb_miss && tag_hit && r_req_store ;
 
   // write the pending store data when the cache is idle, when the next command isn't a load
@@ -297,6 +292,10 @@ class rocketDCacheDM(lines: Int) extends Component {
     p_store_type  <== io.cpu.req_type;
   }
     
+  when (io.cpu.req_val && io.cpu.req_rdy && req_amo) {
+    r_amo_data <== io.cpu.req_data;
+  }
+
   when (store_hit && !drain_store) {
     p_store_valid <== Bool(true);
   }
@@ -304,52 +303,85 @@ class rocketDCacheDM(lines: Int) extends Component {
     p_store_valid <== Bool(false);
     db_array <== db_array.bitSet(p_store_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(1,1));
   }
-  when (resolve_store) {
+  when (resolve_store || (state === s_write_amo)) {
     db_array <== db_array.bitSet(p_store_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(1,1));
   }
   when (tag_we) {
     db_array <== db_array.bitSet(r_cpu_req_idx(PGIDX_BITS-1,offsetbits).toUFix, UFix(0,1));
   }
   
-  // generate write mask and data signals for stores
+  // generate write mask and data signals for stores and amos
   val storegen = new rocketDCacheStoreGen();
   storegen.io.req_addr_lsb := p_store_idx(2,0);
   storegen.io.req_data := p_store_data;
-  storegen.io.req_type := p_store_type
+  storegen.io.req_type := p_store_type;
   val store_data = Fill(2, storegen.io.store_data);
-  val store_wmask_d = storegen.io.store_wmask;
+  val store_wmask_b = storegen.io.store_wmask;
+  val store_wmask_d = Cat(Fill(8, store_wmask_b(7)),
+  		Fill(8, store_wmask_b(6)),
+  		Fill(8, store_wmask_b(5)),
+  		Fill(8, store_wmask_b(4)),
+  		Fill(8, store_wmask_b(3)),
+  		Fill(8, store_wmask_b(2)),
+  		Fill(8, store_wmask_b(1)),
+  		Fill(8, store_wmask_b(0)));
   val store_idx_sel = p_store_idx(offsetlsb).toBool;
   val store_wmask = Mux(store_idx_sel, Cat(store_wmask_d, Bits(0,64)), Cat(Bits(0,64), store_wmask_d)); 
 
   // data array
   val data_array = new rocketSRAMsp(lines*4, 128);
+  val data_array_rdata = data_array.io.q; 
+  val resp_data = Mux(r_cpu_req_idx(offsetlsb).toBool, data_array_rdata(127, 64), data_array_rdata(63,0));
+  val r_resp_data = Reg(resp_data);
+  
+  // ALU for AMOs
+  val amo_wmask = 
+    Mux(r_cpu_req_type === MT_D, ~Bits(0,8),
+    Mux(r_cpu_req_idx(2).toBool, Cat(~Bits(0,4), Bits(0,4)),
+      Cat(Bits(0,4), ~Bits(0,4))));
+        
+  val amo_alu = new rocketDCacheAmoALU();
+  amo_alu.io.cmd := r_cpu_req_cmd;
+  amo_alu.io.wmask := amo_wmask;
+  amo_alu.io.lhs := r_resp_data.toUFix;
+  amo_alu.io.rhs := r_amo_data.toUFix;
+  val amo_alu_out = amo_alu.io.result;
+  
   data_array.io.a := 
     Mux(drain_store || resolve_store, p_store_idx(PGIDX_BITS-1, offsetmsb-1),
     Mux((state === s_writeback) && io.mem.req_rdy, Cat(r_cpu_req_idx(PGIDX_BITS-1, offsetbits), rr_count_next),
     Mux((state === s_start_writeback) || (state === s_writeback) || (state === s_refill), Cat(r_cpu_req_idx(PGIDX_BITS-1, offsetbits), rr_count),
-    Mux((state === s_resolve_miss) || (state === s_replay_load),  r_cpu_req_idx(PGIDX_BITS-1, offsetmsb-1),
+    Mux((state === s_resolve_miss) || (state === s_replay_load) || (state === s_write_amo),  r_cpu_req_idx(PGIDX_BITS-1, offsetmsb-1),
       io.cpu.req_idx(PGIDX_BITS-1, offsetmsb-1))))).toUFix;
       
-  data_array.io.d :=  Mux((state === s_refill), io.mem.resp_data, store_data);
+  data_array.io.d := 
-  data_array.io.we := ((state === s_refill) && io.mem.resp_val) || drain_store || resolve_store;
+    Mux((state === s_refill), io.mem.resp_data, 
+    Mux((state === s_write_amo), amo_alu_out,
+      store_data));
+  data_array.io.we := 
+    ((state === s_refill) && io.mem.resp_val) ||
+    (state === s_write_amo) ||
+     drain_store || resolve_store;
+     
   data_array.io.bweb := Mux((state === s_refill), ~Bits(0,128), store_wmask);
   data_array.io.ce := 
-    (io.cpu.req_val && io.cpu.req_rdy && req_load) ||
+    (io.cpu.req_val && io.cpu.req_rdy && (req_load || req_amo)) ||
     (state === s_start_writeback) ||
     (state === s_writeback) ||
-    ((state === s_resolve_miss) && r_req_load) ||
+    ((state === s_resolve_miss) && (r_req_load || r_req_amo)) ||
     (state === s_replay_load);
-  
-  val data_array_rdata = data_array.io.q;
 
   // signal a load miss when the data isn't present in the cache and when it's in the pending store data register
-  // (causes the cache to block for 2 cycles and the load instruction is replayed)
+  // (causes the cache to block for 2 cycles and the load or amo instruction is replayed)
-  val load_miss = !io.cpu.dtlb_miss && (state === s_ready) && r_cpu_req_val && r_req_load && (!tag_hit || (p_store_valid && addr_match));
+  val load_miss = 
+    !io.cpu.dtlb_miss && 
+    (state === s_ready) && r_cpu_req_val && (r_req_load || r_req_amo) && (!tag_hit || (p_store_valid && addr_match));
 
   // output signals
   // busy when there's a load to the same address as a pending store, or on a cache miss, or when executing a flush
-  io.cpu.req_rdy   := (state === s_ready) && !io.cpu.dtlb_miss && !ldst_conflict && (!r_cpu_req_val || (tag_hit && !r_req_flush));
+  io.cpu.req_rdy   := (state === s_ready) && !io.cpu.dtlb_miss && !ldst_conflict && (!r_cpu_req_val || (tag_hit && !(r_req_flush || r_req_amo)));
-  io.cpu.resp_val  := !io.cpu.dtlb_miss && ((state === s_ready) &&  tag_hit && r_req_load && !(p_store_valid && addr_match)) || 
+  io.cpu.resp_val  := !io.cpu.dtlb_miss && 
+                      ((state === s_ready) &&  tag_hit && (r_req_load || r_req_amo) && !(p_store_valid && addr_match)) || 
                       ((state === s_resolve_miss) && r_req_flush) ||
                       r_cpu_resp_val;
                       
@@ -358,15 +390,13 @@ class rocketDCacheDM(lines: Int) extends Component {
     (((r_cpu_req_type === MT_W) || (r_cpu_req_type === MT_WU)) && (r_cpu_req_idx(1,0) != Bits(0,2))) ||
     ((r_cpu_req_type === MT_D) && (r_cpu_req_idx(2,0) != Bits(0,3)));
     
-  io.cpu.xcpt_ma_ld := r_cpu_req_val && r_req_load && misaligned;
+  io.cpu.xcpt_ma_ld := r_cpu_req_val && (r_req_load  || r_req_amo) && misaligned;
-  io.cpu.xcpt_ma_st := r_cpu_req_val && r_req_store && misaligned;
+  io.cpu.xcpt_ma_st := r_cpu_req_val && (r_req_store || r_req_amo) && misaligned;
     
   io.cpu.resp_miss := load_miss;
   // tag MSB distinguishes between loads destined for the PTW and CPU
   io.cpu.resp_tag  := Cat(r_req_ptw_load, r_cpu_req_type, r_cpu_req_idx(2,0), r_cpu_req_tag);
-  io.cpu.resp_data := 
+  io.cpu.resp_data := resp_data;
-    Mux(r_cpu_req_idx(offsetlsb).toBool, data_array_rdata(127, 64), 
-      data_array_rdata(63,0));
 
   io.mem.req_val   := (state === s_req_refill) || (state === s_writeback);
   io.mem.req_rw    := (state === s_writeback);
@@ -388,9 +418,12 @@ class rocketDCacheDM(lines: Int) extends Component {
       when (ldst_conflict) {
         state <== s_replay_load;
       }
-      when (!r_cpu_req_val || (tag_hit && !r_req_flush)) {
+      when (!r_cpu_req_val || (tag_hit && !(r_req_flush || r_req_amo))) {
         state <== s_ready;
       }
+      when (tag_hit && r_req_amo) {
+        state <== s_write_amo;
+      }
       when (tag_valid & tag_dirty) {
         state <== s_start_writeback; 
       }
@@ -404,6 +437,9 @@ class rocketDCacheDM(lines: Int) extends Component {
     is (s_replay_load) {
       state <== s_ready;
     }
+    is (s_write_amo) {
+      state <== s_ready;
+    }
     is (s_start_writeback) {
       state <== s_writeback;
     }
@@ -425,9 +461,42 @@ class rocketDCacheDM(lines: Int) extends Component {
       when (io.mem.resp_val && (rr_count === UFix(3,2))) { state <== s_resolve_miss; }
     }
     is (s_resolve_miss) {
+      when (r_req_amo) {
+        state <== s_write_amo;
+      }
       state <== s_ready;
     }
   }  
 }
 
+class rocketDCacheAmoALU extends Component {
+  val io = new Bundle {
+    val cmd  = Bits(4, 'input);
+    val wmask = Bits(8, 'input);
+    val lhs   = UFix(64, 'input);
+    val rhs   = UFix(64, 'input);
+    val result    = UFix(64, 'output);
+  }
+  
+//   val signed_cmp = (op === M_XA_MIN) || (op === M_XA_MAX);
+//   val sub = (op === M_XA_MIN) || (op === M_XA_MINU) ||
+//             (op === M_XA_MAX) || (op === M_XA_MAXU);
+
+  val adder_lhs = Cat(io.lhs(63,32),io.wmask(3) & io.lhs(31), io.lhs(30,0)).toUFix;
+  val adder_rhs = Cat(io.rhs(63,32),io.wmask(3) & io.rhs(31), io.rhs(30,0)).toUFix;
+//   val adder_rhs = Cat(Mux(sub, ~io.rhs, io.rhs), sub).toUFix;
+//   val sum = adder_lhs + adder_rhs;
+//   val adder_out = sum(64,1);
+  val adder_out = adder_lhs + adder_rhs;
+  val alu_out = Wire() { UFix() };
+  switch (io.cmd) {
+//     is (M_XA_ADD)  { alu_out <== adder_out; }
+    is (M_XA_SWAP) { alu_out <== io.rhs; }
+    is (M_XA_AND)  { alu_out <== io.lhs & io.rhs; }
+    is (M_XA_OR)   { alu_out <== io.lhs | io.rhs; }
+  }
+  alu_out <== adder_out;
+  io.result := alu_out;
 }
+  
+}

rocket/src/main/scala/dpath.scala

@@ -45,7 +45,7 @@ class rocketDpath extends Component
 
   val alu = new rocketDpathALU(); 
   val ex_alu_out = alu.io.out; 
-  val ex_jr_target = ex_alu_out(31,0);
+  val ex_jr_target = ex_alu_out(VADDR_BITS,0);
 
   val div = new rocketDivider(64);
   val div_result = div.io.div_result_bits;
@@ -342,7 +342,7 @@ class rocketDpath extends Component
 
   // D$ request interface (registered inside D$ module)
   // other signals (req_val, req_rdy) connect to control module  
-  io.dmem.req_addr  := ex_alu_out(PADDR_BITS-1,0);
+  io.dmem.req_addr  := ex_alu_out(VADDR_BITS-1,0);
   io.dmem.req_data  := ex_reg_rs2;
   io.dmem.req_tag   := ex_reg_waddr;