diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index eeba8c984767..028eebcb445b 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -737,8 +737,7 @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
         address |= TLB_INVALID_MASK;
     }
     if (attrs.byte_swap) {
-        /* Force the access through the I/O slow path.  */
-        address |= TLB_MMIO;
+        address |= TLB_BSWAP;
     }
     if (!memory_region_is_ram(section->mr) &&
         !memory_region_is_romd(section->mr)) {
@@ -901,10 +900,6 @@ static uint64_t io_readx(CPUArchState *env, CPUIOTLBEntry *iotlbentry,
     bool locked = false;
     MemTxResult r;
 
-    if (iotlbentry->attrs.byte_swap) {
-        op ^= MO_BSWAP;
-    }
-
     section = iotlb_to_section(cpu, iotlbentry->addr, iotlbentry->attrs);
     mr = section->mr;
     mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr;
@@ -947,10 +942,6 @@ static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry,
     bool locked = false;
     MemTxResult r;
 
-    if (iotlbentry->attrs.byte_swap) {
-        op ^= MO_BSWAP;
-    }
-
     section = iotlb_to_section(cpu, iotlbentry->addr, iotlbentry->attrs);
     mr = section->mr;
     mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr;
@@ -1133,8 +1124,8 @@ void *probe_access(CPUArchState *env, target_ulong addr, int size,
                              wp_access, retaddr);
     }
 
-    if (tlb_addr & (TLB_NOTDIRTY | TLB_MMIO)) {
-        /* I/O access */
+    /* Reject I/O access, or other required slow-path.  */
+    if (tlb_addr & (TLB_NOTDIRTY | TLB_MMIO | TLB_BSWAP)) {
         return NULL;
     }
 
@@ -1344,6 +1335,7 @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi,
     /* Handle anything that isn't just a straight memory access.  */
     if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
         CPUIOTLBEntry *iotlbentry;
+        bool need_swap;
 
         /* For anything that is unaligned, recurse through full_load.  */
         if ((addr & (size - 1)) != 0) {
@@ -1357,17 +1349,27 @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi,
             /* On watchpoint hit, this will longjmp out.  */
             cpu_check_watchpoint(env_cpu(env), addr, size,
                                  iotlbentry->attrs, BP_MEM_READ, retaddr);
-
-            /* The backing page may or may not require I/O.  */
-            tlb_addr &= ~TLB_WATCHPOINT;
-            if ((tlb_addr & ~TARGET_PAGE_MASK) == 0) {
-                goto do_aligned_access;
-            }
         }
 
+        need_swap = size > 1 && (tlb_addr & TLB_BSWAP);
+
         /* Handle I/O access.  */
-        return io_readx(env, iotlbentry, mmu_idx, addr,
-                        retaddr, access_type, op);
+        if (likely(tlb_addr & TLB_MMIO)) {
+            return io_readx(env, iotlbentry, mmu_idx, addr, retaddr,
+                            access_type, op ^ (need_swap * MO_BSWAP));
+        }
+
+        haddr = (void *)((uintptr_t)addr + entry->addend);
+
+        /*
+         * Keep these two load_memop separate to ensure that the compiler
+         * is able to fold the entire function to a single instruction.
+         * There is a build-time assert inside to remind you of this.  ;-)
+         */
+        if (unlikely(need_swap)) {
+            return load_memop(haddr, op ^ MO_BSWAP);
+        }
+        return load_memop(haddr, op);
     }
 
     /* Handle slow unaligned access (it spans two pages or IO).  */
@@ -1394,7 +1396,6 @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi,
         return res & MAKE_64BIT_MASK(0, size * 8);
     }
 
- do_aligned_access:
     haddr = (void *)((uintptr_t)addr + entry->addend);
     return load_memop(haddr, op);
 }
@@ -1591,6 +1592,7 @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val,
     /* Handle anything that isn't just a straight memory access.  */
     if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
         CPUIOTLBEntry *iotlbentry;
+        bool need_swap;
 
         /* For anything that is unaligned, recurse through byte stores.  */
         if ((addr & (size - 1)) != 0) {
@@ -1604,16 +1606,29 @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val,
             /* On watchpoint hit, this will longjmp out.  */
             cpu_check_watchpoint(env_cpu(env), addr, size,
                                  iotlbentry->attrs, BP_MEM_WRITE, retaddr);
-
-            /* The backing page may or may not require I/O.  */
-            tlb_addr &= ~TLB_WATCHPOINT;
-            if ((tlb_addr & ~TARGET_PAGE_MASK) == 0) {
-                goto do_aligned_access;
-            }
         }
 
+        need_swap = size > 1 && (tlb_addr & TLB_BSWAP);
+
         /* Handle I/O access.  */
-        io_writex(env, iotlbentry, mmu_idx, val, addr, retaddr, op);
+        if (likely(tlb_addr & (TLB_MMIO | TLB_NOTDIRTY))) {
+            io_writex(env, iotlbentry, mmu_idx, val, addr, retaddr,
+                      op ^ (need_swap * MO_BSWAP));
+            return;
+        }
+
+        haddr = (void *)((uintptr_t)addr + entry->addend);
+
+        /*
+         * Keep these two store_memop separate to ensure that the compiler
+         * is able to fold the entire function to a single instruction.
+         * There is a build-time assert inside to remind you of this.  ;-)
+         */
+        if (unlikely(need_swap)) {
+            store_memop(haddr, val, op ^ MO_BSWAP);
+        } else {
+            store_memop(haddr, val, op);
+        }
         return;
     }
 
@@ -1682,7 +1697,6 @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val,
         return;
     }
 
- do_aligned_access:
     haddr = (void *)((uintptr_t)addr + entry->addend);
     store_memop(haddr, val, op);
 }
diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h
index e0c8dc540c7b..d148bded3599 100644
--- a/include/exec/cpu-all.h
+++ b/include/exec/cpu-all.h
@@ -335,12 +335,14 @@ CPUArchState *cpu_copy(CPUArchState *env);
 #define TLB_MMIO            (1 << (TARGET_PAGE_BITS_MIN - 3))
 /* Set if TLB entry contains a watchpoint.  */
 #define TLB_WATCHPOINT      (1 << (TARGET_PAGE_BITS_MIN - 4))
+/* Set if TLB entry requires byte swap.  */
+#define TLB_BSWAP           (1 << (TARGET_PAGE_BITS_MIN - 5))
 
 /* Use this mask to check interception with an alignment mask
  * in a TCG backend.
  */
 #define TLB_FLAGS_MASK \
-    (TLB_INVALID_MASK | TLB_NOTDIRTY | TLB_MMIO | TLB_WATCHPOINT)
+    (TLB_INVALID_MASK | TLB_NOTDIRTY | TLB_MMIO | TLB_WATCHPOINT | TLB_BSWAP)
 
 /**
  * tlb_hit_page: return true if page aligned @addr is a hit against the