Merge pull request #104 from mihaip/upstream-mmu-perf

ppcopcodes: avoid TLB flushes for noop SR changes
dingusdev · Jul 28, 2024 · 6f62f9d · 6f62f9d
2 parents f192d11 + 31d7b05
commit 6f62f9d
Show file tree

Hide file tree

Showing 2 changed files with 24 additions and 7 deletions.
diff --git a/cpu/ppc/ppcopcodes.cpp b/cpu/ppc/ppcopcodes.cpp
@@ -728,8 +728,10 @@ void dppc_interpreter::ppc_mtsr() {
     }
     int reg_s             = (ppc_cur_instruction >> 21) & 0x1F;
     uint32_t grab_sr      = (ppc_cur_instruction >> 16) & 0x0F;
-    ppc_state.sr[grab_sr] = ppc_state.gpr[reg_s];
-    mmu_pat_ctx_changed();
+   if (ppc_state.sr[grab_sr] != ppc_state.gpr[reg_s]) {
+        ppc_state.sr[grab_sr] = ppc_state.gpr[reg_s];
+        mmu_pat_ctx_changed();
+   }
 }
 
 void dppc_interpreter::ppc_mtsrin() {
@@ -741,8 +743,10 @@ void dppc_interpreter::ppc_mtsrin() {
     }
     ppc_grab_regssb(ppc_cur_instruction);
     uint32_t grab_sr      = ppc_result_b >> 28;
-    ppc_state.sr[grab_sr] = ppc_result_d;
-    mmu_pat_ctx_changed();
+    if (ppc_state.sr[grab_sr] != ppc_result_d) {
+        ppc_state.sr[grab_sr] = ppc_result_d;
+        mmu_pat_ctx_changed();
+    }
 }
 
 void dppc_interpreter::ppc_mfsr() {
@@ -969,8 +973,10 @@ void dppc_interpreter::ppc_mtspr() {
         ppc_state.spr[ref_spr] = val & 0xe000ff7f;
         break;
     case SPR::SDR1:
-        ppc_state.spr[ref_spr] = val;
-        mmu_pat_ctx_changed(); // adapt to SDR1 changes
+        if (ppc_state.spr[ref_spr] != val) {
+            ppc_state.spr[ref_spr] = val;
+            mmu_pat_ctx_changed(); // adapt to SDR1 changes
+        }
         break;
     case SPR::RTCL_S:
         calc_rtcl_value();

diff --git a/devices/memctrl/memctrlbase.cpp b/devices/memctrl/memctrlbase.cpp
@@ -168,7 +168,18 @@ bool MemCtrlBase::add_mem_region(uint32_t start_addr, uint32_t size,
     entry->devobj  = nullptr;
     entry->mem_ptr = reg_content;
 
-    this->address_map.push_back(entry);
+    // Keep address_map sorted, that way the RAM region (which starts at 0 and
+    // is most often requested) will be found by find_range on the first
+    // iteration.
+    this->address_map.insert(
+        std::upper_bound(
+            this->address_map.begin(),
+            this->address_map.end(),
+            entry,
+            [](const auto& lhs, const auto& rhs) {
+                return lhs->start < rhs->start;
+            }),
+            entry);
 
     LOG_F(INFO, "Added mem region 0x%X..0x%X (%s%s%s%s) -> 0x%X", start_addr, end,
         entry->type & RT_ROM ? "ROM," : "",