Multi-word fetch bundles and icache with realistic dram latency

1 files changed, 39 insertions, 16 deletions

diff --git a/uarch/core.cpp b/uarch/core.cpp
index 34f422a..8b2db9b 100644
--- a/uarch/core.cpp
+++ b/uarch/core.cpp
@@ -25,21 +25,38 @@ void fetch_stage::clock() {
         gen = r.new_gen;
         pc = r.new_pc;
         didrestart = true;
+        outstandingfill = false;
         c.fetch_restarto.reset();
     }
 
+    if (c.fetch_mem_responsep.can_read()) {
+        auto r = c.fetch_mem_responsep.read();
+        cache.handle_response(r);
+    }
+
     if (c.fetch_bundlep.can_write()) {
         fetch_bundle b;
-        b.tr = infra::pt::toplevel();
-        b.gen = gen;
-        b.pc = pc;
-        b.word = c.mem.fetch(pc);
-        if (didrestart)
-            infra::pt::event(b.tr, ">", now-1, "");
-        pte(b.tr, "F");
-        c.fetch_bundlep.write(std::move(b));
-        pc = (pc & 070000) | ((pc + 1) & 007777);
-        didrestart = false;
+        if (auto t = cache.fetchline(b.data, pc); t.has_value()) {
+            b.tr = infra::pt::toplevel();
+            b.gen = gen;
+            b.pc = pc;
+            if (didrestart)
+                infra::pt::event(b.tr, ">", now-1, "");
+            pte(b.tr, "F", fmt::format("pc={:05o}", b.pc));
+            c.fetch_bundlep.write(std::move(b));
+            pc = (pc & 070000) | (((pc & ~memory::LINE_BYTE_OFFSET_MASK) + memory::LINE_BYTES) & 007777);
+            didrestart = false;
+            outstandingfill = false;
+        }
+    }
+
+    if (!outstandingfill && c./*fetch_*/mem_commandp.can_write() && !cache.probe(pc)) {
+        memory::dram::command fr;
+        fr.transaction = infra::pt::toplevel();
+        fr.line_address = pc >> memory::LINE_BYTES_LOG2;
+        fr.responsep = &c.fetch_mem_responsep;
+        c./*fetch_*/mem_commandp.write(std::move(fr));
+        outstandingfill = true;
     }
 }
 
@@ -47,28 +64,31 @@ void decode_stage::clock() {
     bool progress = ctlregs[HALTED];
 
     if (!ctlregs[HALTED] && c.fetch_bundlep.can_read()) {
-        auto b = c.fetch_bundlep.read();
+        auto b = c.fetch_bundlep.peek();
 
         if (b.gen != gen)
             goto bail_out;
 
-        if (b.pc != pc) {
+        if ((b.pc >> memory::LINE_BYTES_LOG2) != (pc >> memory::LINE_BYTES_LOG2)) {
             pte(b.tr, "~");
             fetch_restart r;
             r.tr = b.tr;
             r.new_gen = ++gen;
             r.new_pc = pc;
             c.fetch_restarto = std::move(r);
+            c.fetch_bundlep.discard();
             goto bail_out;
         }
 
         progress = true;
 
-        pte(b.tr, "E");
+        auto tr = infra::pt::child(b.tr);
+
+        pte(tr, "E");
 
         inst = decode(ctlregs[FLAGS],
                       pc,
-                      c.mem.fetch(pc),
+                      b.data[pc & memory::LINE_BYTE_OFFSET_MASK],
                       interrupt);
         auto next_pc = inst.next_pc;
 
@@ -84,7 +104,7 @@ void decode_stage::clock() {
             assert(!inst.need_autoinc_store);
         }
 
-        pte(b.tr, "", inst.disasm());
+        pte(tr, "", inst.disasm());
 
         if (inst.need_exec_load)
             inst.data = c.mem.fetch(inst.final_address.value());
@@ -114,6 +134,9 @@ void decode_stage::clock() {
 
         assert(inst.next_pc == next_pc || inst.possibly_redirects);
         pc = inst.next_pc;
+
+        if ((b.pc >> memory::LINE_BYTES_LOG2) != (pc >> memory::LINE_BYTES_LOG2))
+            c.fetch_bundlep.discard();
     }
 bail_out:
 
@@ -125,7 +148,7 @@ bail_out:
 
         c.checker.execute();
         assert(c.checker.icount == icount);
-        std::cerr << fmt::format("icount={:} pc={:05o} checkerpc={:05o}\n", icount, pc, c.checker.pc);
+//        std::cerr << fmt::format("icount={:} pc={:05o} checkerpc={:05o}\n", icount, pc, c.checker.pc);
         assert(pc == c.checker.pc);
         assert(acc == c.checker.acc);
         assert(link == c.checker.link);