notdirty_write: fix store-related performance problems

Every store would always cause the tb_invalidate_phys_page_fast path to be invoked,
amounting to a 40x slowdown of stores compared to loads.

Change this code to only worry about TB invalidation for regions marked as
executable (i.e. emulated executable).

Even without uc_set_native_thunks, this change fixes most of the performance
issues seen with thunking to native calls.

Signed-off-by: Andrei Warkentin <andrei.warkentin@intel.com>

qemu/accel/tcg/translate-all.c

@@ -1843,6 +1843,11 @@ TranslationBlock *tb_gen_code(CPUState *cpu,
     if ((pc & TARGET_PAGE_MASK) != virt_page2) {
         phys_page2 = get_page_addr_code(env, virt_page2);
     }
+
+    /* Undoes tlb_set_dirty in notdirty_write. */
+    tlb_reset_dirty_by_vaddr(cpu, pc & TARGET_PAGE_MASK,
+                             (pc & ~TARGET_PAGE_MASK) + tb->size);
+
     /*
      * No explicit memory barrier is required -- tb_link_page() makes the
      * TB visible in a consistent state.