diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c
index 2c34adccce..99ca6f36b9 100644
--- a/accel/tcg/translate-all.c
+++ b/accel/tcg/translate-all.c
@@ -60,6 +60,7 @@
 #include "sysemu/cpu-timers.h"
 #include "sysemu/tcg.h"
 #include "qapi/error.h"
+#include "hw/core/tcg-cpu-ops.h"
 #include "internal.h"
 
 /* #define DEBUG_TB_INVALIDATE */
@@ -2421,6 +2422,7 @@ void cpu_io_recompile(CPUState *cpu, uintptr_t retaddr)
     CPUArchState *env = cpu->env_ptr;
 #endif
     TranslationBlock *tb;
+    CPUClass *cc;
     uint32_t n;
 
     tb = tcg_tb_lookup(retaddr);
@@ -2430,11 +2432,18 @@ void cpu_io_recompile(CPUState *cpu, uintptr_t retaddr)
     }
     cpu_restore_state_from_tb(cpu, tb, retaddr, true);
 
-    /* On MIPS and SH, delay slot instructions can only be restarted if
-       they were already the first instruction in the TB.  If this is not
-       the first instruction in a TB then re-execute the preceding
-       branch.  */
+    /*
+     * Some guests must re-execute the branch when re-executing a delay
+     * slot instruction.  When this is the case, adjust icount and N
+     * to account for the re-execution of the branch.
+     */
     n = 1;
+    cc = CPU_GET_CLASS(cpu);
+    if (cc->tcg_ops->io_recompile_replay_branch &&
+        cc->tcg_ops->io_recompile_replay_branch(cpu, tb)) {
+        cpu_neg(cpu)->icount_decr.u16.low++;
+        n = 2;
+    }
 #if defined(TARGET_MIPS)
     if ((env->hflags & MIPS_HFLAG_BMASK) != 0
         && env->active_tc.PC != tb->pc) {
diff --git a/include/hw/core/tcg-cpu-ops.h b/include/hw/core/tcg-cpu-ops.h
index ac3bb051f2..72d791438c 100644
--- a/include/hw/core/tcg-cpu-ops.h
+++ b/include/hw/core/tcg-cpu-ops.h
@@ -88,6 +88,16 @@ struct TCGCPUOps {
      */
     bool (*debug_check_watchpoint)(CPUState *cpu, CPUWatchpoint *wp);
 
+    /**
+     * @io_recompile_replay_branch: Callback for cpu_io_recompile.
+     *
+     * The cpu has been stopped, and cpu_restore_state_from_tb has been
+     * called.  If the faulting instruction is in a delay slot, and the
+     * target architecture requires re-execution of the branch, then
+     * adjust the cpu state as required and return true.
+     */
+    bool (*io_recompile_replay_branch)(CPUState *cpu,
+                                       const TranslationBlock *tb);
 #endif /* CONFIG_SOFTMMU */
 #endif /* NEED_CPU_H */