Avoid saving ldr dest reg twice

Grarak · Jan 12, 2025 · b95c155 · b95c155
1 parent afedac8
commit b95c155
Show file tree

Hide file tree

Showing 2 changed files with 31 additions and 8 deletions.
diff --git a/src/jit/emitter/emit_transfer.rs b/src/jit/emitter/emit_transfer.rs
@@ -230,6 +230,7 @@ impl<const CPU: CpuType> JitAsm<'_, CPU> {
 
         let slow_read_patch_label = block_asm.new_label();
         let slow_read_label = block_asm.new_label();
+        let fast_mem_mark_dirty_label = block_asm.new_label();
         let continue_label = block_asm.new_label();
 
         block_asm.branch(slow_read_label, Cond::NV);
@@ -260,7 +261,11 @@ impl<const CPU: CpuType> JitAsm<'_, CPU> {
             block_asm.mov(op0, (fast_read_value_reg, ShiftType::Ror, fast_read_addr_masked_reg));
         }
 
-        block_asm.branch_fallthrough(continue_label, Cond::AL);
+        block_asm.mark_reg_dirty(op0, false);
+        if amount == MemoryAmount::Double {
+            block_asm.mark_reg_dirty(Reg::from(op0 as u8 + 1), false);
+        }
+        block_asm.branch_fallthrough(fast_mem_mark_dirty_label, Cond::AL);
         block_asm.branch(slow_read_patch_label, Cond::AL);
 
         block_asm.label_unlikely(slow_read_patch_label);
@@ -306,6 +311,13 @@ impl<const CPU: CpuType> JitAsm<'_, CPU> {
         block_asm.restore_reg(Reg::CPSR);
 
         block_asm.branch(continue_label, Cond::AL);
+
+        block_asm.label(fast_mem_mark_dirty_label);
+        block_asm.mark_reg_dirty(op0, true);
+        if amount == MemoryAmount::Double {
+            block_asm.mark_reg_dirty(Reg::from(op0 as u8 + 1), true);
+        }
+
         block_asm.label(continue_label);
 
         block_asm.free_reg(cpsr_backup_reg);

diff --git a/src/jit/inst_jit_handler.rs b/src/jit/inst_jit_handler.rs
@@ -49,17 +49,28 @@ pub unsafe extern "C" fn inst_slow_mem_patch(pc: u32) {
     debug_assert_ne!(fast_mem_end, 0);
 
     let mut fast_mem_start = 0;
-    let mut found_non_op = false;
+    let mut found_nop_op = 0;
     for pc_offset in (4..256).step_by(4) {
         let ptr = (fast_mem_end - pc_offset) as *const u32;
         let opcode = ptr.read();
-        if found_non_op {
-            if opcode == nop_opcode {
-                fast_mem_start = ptr as usize;
-                break;
+        match found_nop_op {
+            0 => {
+                if opcode == nop_opcode {
+                    found_nop_op = 1;
+                }
             }
-        } else if opcode != nop_opcode {
-            found_non_op = true;
+            1 => {
+                if opcode != nop_opcode {
+                    found_nop_op = 2;
+                }
+            }
+            2 => {
+                if opcode == nop_opcode {
+                    fast_mem_start = ptr as usize;
+                    break;
+                }
+            }
+            _ => {}
         }
     }
     debug_assert_ne!(fast_mem_start, 0);