cache_subsys: cut timing path of AMOs with a buffer (#3341)

Recent ASIC synthesis with the CVA6+HPDcache exhibit a critical path concerning the response path for AMOs (Atomic Memory Operations) from the cache to the CVA6.

This PR adds a pipeline buffer into this path to enhance the timing (clock frequency). However, it increases the latency of AMOs by one cycle. This extra cycle is anyway acceptable as AMOs are not part of the the most frequent operations, and the clock frequency speedup largely compensates this.

---------

Signed-off-by: Cesar Fuguet <15080006+cfuguet@users.noreply.github.com>
Co-authored-by: JeanRochCoulon <jean-roch.coulon@thalesgroup.com>

Author: cfuguet

.gitlab-ci/expected_synth.yml

@@ -1,2 +1,2 @@
 cv32a65x:
-  gates: 183130
+  gates: 183727

core/cache_subsystem/cva6_hpdcache_if_adapter.sv

@@ -326,6 +326,8 @@ module cva6_hpdcache_if_adapter
 
       //  Response forwarding
       //  {{{
+      ariane_pkg::amo_resp_t cva6_amo_resp;
+
       if (CVA6Cfg.XLEN == 64) begin : amo_resp_64_gen
         assign amo_resp_word = amo_is_word_hi
                              ? hpdcache_rsp_i.rdata[0][32 +: 32]
@@ -339,18 +341,25 @@ module cva6_hpdcache_if_adapter
       assign cva6_req_o.data_rid = hpdcache_rsp_i.tid;
       assign cva6_req_o.data_gnt = hpdcache_req_ready_i;
 
-      assign cva6_amo_resp_o.ack = hpdcache_rsp_valid_i && (hpdcache_rsp_i.tid == '1);
-      assign cva6_amo_resp_o.result = amo_is_word ? {{32{amo_resp_word[31]}}, amo_resp_word}
-                                                        : hpdcache_rsp_i.rdata[0];
+      assign cva6_amo_resp.ack = hpdcache_rsp_valid_i && (hpdcache_rsp_i.tid == '1);
+      assign cva6_amo_resp.result = amo_is_word ? {{32{amo_resp_word[31]}}, amo_resp_word}
+                                                : hpdcache_rsp_i.rdata[0];
       //  }}}
 
       always_ff @(posedge clk_i or negedge rst_ni) begin : amo_pending_ff
         if (!rst_ni) begin
-          amo_pending_q <= 1'b0;
+          amo_pending_q   <= 1'b0;
+          cva6_amo_resp_o <= '0;
         end else begin
           amo_pending_q <=
-              ( cva6_amo_req_i.req  & hpdcache_req_ready_i & ~amo_pending_q) |
-              (~cva6_amo_resp_o.ack & amo_pending_q);
+              (~amo_pending_q &  cva6_amo_req_i.req & hpdcache_req_ready_i) |
+              ( amo_pending_q & ~cva6_amo_resp_o.ack);
+
+          if (cva6_amo_resp_o.ack) begin
+            cva6_amo_resp_o <= '0;
+          end else if (cva6_amo_resp.ack) begin
+            cva6_amo_resp_o <= cva6_amo_resp;
+          end
         end
       end