Reduce stalling due to stores by using a global counter.

2 files changed, 16 insertions, 15 deletions

diff --git a/uarch/core.cpp b/uarch/core.cpp
index 7304442..76af173 100644
--- a/uarch/core.cpp
+++ b/uarch/core.cpp
@@ -77,13 +77,7 @@ void decode_stage::clock() {
         speculative_stores_sent = r.stores_sent;
     }
 
-    if (c.decode_store_completep.can_read()) {
-        ++stores_done;
-        assert((int)(speculative_stores_sent - stores_done) >= 0);
-        c.decode_store_completep.discard();
-    }
-
-    if (speculative_stores_sent == stores_done && c.fetch_bundlep.can_read() && c.decode_mem_commandp.can_write() && c.indir_instp.can_write() && c.decode_to_exec_instp.can_write()) {
+    if (c.fetch_bundlep.can_read() && c.decode_mem_commandp.can_write() && c.indir_instp.can_write() && c.decode_to_exec_instp.can_write()) {
         auto b = c.fetch_bundlep.peek();
 
         if (b.gen != c.gen) {
@@ -106,6 +100,11 @@ void decode_stage::clock() {
             return;
         }
 
+        if (speculative_stores_sent != c.stores_done) {
+            pte(b.tr, "z");
+            return;
+        }
+
         inst_bundle i;
 
         i.tr = infra::pt::child(b.tr);
@@ -190,7 +189,6 @@ void indir_stage::clock() {
                 sr.mask.fill(false);
                 sr.mask[*i.inst.init_address & memory::LINE_BYTE_OFFSET_MASK] = true;
                 sr.write = true;
-                sr.responsep = &c.decode_store_completep;
                 c.indir_mem_store_commandp.write(std::move(sr));
             } else {
                 pte(i.tr, "I");
@@ -238,9 +236,9 @@ void exec_stage::clock() {
                     sr.mask.fill(false);
                     sr.mask[*i.inst.init_address & memory::LINE_BYTE_OFFSET_MASK] = true;
                     sr.write = true;
-                    sr.responsep = &c.decode_store_completep;
                     c.exec_mem_commandp.write(std::move(sr));
                     stores_sent += 2; // Original store sent by Indir stage plus unstore here
+                    c.stores_done += 2;
                 }
                 c.decode_to_exec_instp.discard();
             } else if (i.icount == c.icount) {
@@ -261,9 +259,9 @@ void exec_stage::clock() {
                     sr.mask.fill(false);
                     sr.mask[*i.inst.init_address & memory::LINE_BYTE_OFFSET_MASK] = true;
                     sr.write = true;
-                    sr.responsep = &c.decode_store_completep;
                     c.exec_mem_commandp.write(std::move(sr));
                     stores_sent += 2; // Original store sent by Indir stage plus unstore here
+                    c.stores_done += 2;
                 }
                 c.indir_to_exec_instp.discard();
             } else if (i.icount == c.icount) {
@@ -305,8 +303,10 @@ void exec_stage::clock() {
 
         auto next_pc = inst.next_pc;
 
-        if (inst.need_autoinc_store)
+        if (inst.need_autoinc_store) {
             ++stores_sent; // It was sent by Indir stage
+            ++c.stores_done;
+        }
         if (inst.need_read_acc)
             inst.acc = acc;
         if (inst.need_read_link)
@@ -338,9 +338,9 @@ void exec_stage::clock() {
             sr.mask.fill(false);
             sr.mask[*inst.final_address & memory::LINE_BYTE_OFFSET_MASK] = true;
             sr.write = true;
-            sr.responsep = &c.decode_store_completep;
             c.exec_mem_commandp.write(std::move(sr));
             ++stores_sent;
+            ++c.stores_done;
         }
 
         assert(inst.next_pc == next_pc || inst.possibly_redirects);
diff --git a/uarch/core.h b/uarch/core.h
index 21725b3..b8473e6 100644
--- a/uarch/core.h
+++ b/uarch/core.h
@@ -59,7 +59,6 @@ struct decode_stage : public infra::sim {
     std::uint64_t icount;
 
     unsigned int speculative_stores_sent = 0;
-    unsigned int stores_done = 0;
 
     decode_stage(core &c);
 
@@ -114,7 +113,6 @@ struct core {
 
     infra::port<memory::dram::command> decode_mem_commandp;
     infra::port<memory::dram::response> decode_mem_responsep;
-    infra::port<memory::dram::response> decode_store_completep;
     infra::port<inst_bundle> decode_to_exec_instp;
 
     infra::port<inst_bundle> indir_instp;
@@ -126,7 +124,10 @@ struct core {
     infra::port<memory::dram::command> exec_mem_commandp;
     infra::port<memory::dram::response> exec_mem_responsep;
 
-    // Construction order is execution order within a cycle, so this list should be back-to-front (for zero-cycle restarts)
+    // Global counters (should be Gray code in FPGA implementation, only do == comparisons)
+    unsigned int stores_done = 0;
+
+    // Construction order is execution order within a cycle, so this list should be back-to-front (for zero-cycle restarts and store count propagation)
     exec_stage exec{*this};
     indir_stage indir{*this};
     decode_stage decode{*this};