diff --git a/src/codegen/codegen_ops_x86-64.h b/src/codegen/codegen_ops_x86-64.h
index e46f55a05..01fb292cf 100644
--- a/src/codegen/codegen_ops_x86-64.h
+++ b/src/codegen/codegen_ops_x86-64.h
@@ -3571,6 +3571,8 @@ FP_FLD(int reg)
     addbyte(0x89); /*MOV [TOP], EBX*/
     addbyte(0x5d);
     addbyte((uint8_t) cpu_state_offset(TOP));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3688,6 +3690,8 @@ FP_LOAD_S(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+    
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_D(void)
@@ -3717,6 +3721,8 @@ FP_LOAD_D(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3754,6 +3760,8 @@ FP_LOAD_IW(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IL(void)
@@ -3787,6 +3795,8 @@ FP_LOAD_IL(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IQ(void)
@@ -3831,6 +3841,8 @@ FP_LOAD_IQ(void)
     addbyte(0x44);
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -3863,6 +3875,8 @@ FP_LOAD_IMM_Q(uint64_t v)
     addbyte(0x1d);
     addbyte((uint8_t) cpu_state_offset(tag));
     addbyte(v ? 0 : 1);
+
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
diff --git a/src/codegen/codegen_ops_x86.h b/src/codegen/codegen_ops_x86.h
index c48324c2a..3b47d81a9 100644
--- a/src/codegen/codegen_ops_x86.h
+++ b/src/codegen/codegen_ops_x86.h
@@ -1794,6 +1794,7 @@ FP_FLD(int reg)
         addbyte(0x5d);
         addbyte((uint8_t) cpu_state_offset(TOP));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -2037,6 +2038,7 @@ FP_LOAD_S(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_D(void)
@@ -2096,6 +2098,7 @@ FP_LOAD_D(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IW(void)
@@ -2154,6 +2157,7 @@ FP_LOAD_IW(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IL(void)
@@ -2210,6 +2214,7 @@ FP_LOAD_IL(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 static __inline void
 FP_LOAD_IQ(void)
@@ -2285,6 +2290,7 @@ FP_LOAD_IQ(void)
         addbyte(0x1d);
         addbyte((uint8_t) cpu_state_offset(tag[0]));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline void
@@ -2336,6 +2342,7 @@ FP_LOAD_IMM_Q(uint64_t v)
         addbyte(0x5d);
         addbyte((uint8_t) cpu_state_offset(TOP));
     }
+    CALL_FUNC((uintptr_t) x87_to_mmxreg);
 }
 
 static __inline int
diff --git a/src/codegen_new/codegen_ops_helpers.h b/src/codegen_new/codegen_ops_helpers.h
index 92b721099..2304aa542 100644
--- a/src/codegen_new/codegen_ops_helpers.h
+++ b/src/codegen_new/codegen_ops_helpers.h
@@ -64,6 +64,9 @@ fpu_POP2(codeblock_t *block, ir_data_t *ir)
 static inline void
 fpu_PUSH(codeblock_t *block, ir_data_t *ir)
 {
+    uop_LOAD_FUNC_ARG_IMM(ir, 0, ((uint16_t)cpu_state.TOP - 1));
+    uop_CALL_FUNC(ir, x87_to_mmxreg);
+
     if (block->flags & CODEBLOCK_STATIC_TOP)
         uop_MOV_IMM(ir, IREG_FPU_TOP, cpu_state.TOP - 1);
     else
diff --git a/src/cpu/386_dynarec.c b/src/cpu/386_dynarec.c
index 5f41c416a..5cb385122 100644
--- a/src/cpu/386_dynarec.c
+++ b/src/cpu/386_dynarec.c
@@ -74,6 +74,51 @@ x386_dynarec_log(const char *fmt, ...)
 #    define x386_dynarec_log(fmt, ...)
 #endif
 
+/* Deliberately stashed here; this function is only relevant for dynamic recompilers. */
+#if defined(_MSC_VER) && !defined(__clang__)
+#    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86
+#        define X87_INLINE_ASM
+#    endif
+#else
+#    if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86 || defined _M_X64 || defined __amd64__
+#        define X87_INLINE_ASM
+#    endif
+#endif
+
+#ifdef USE_NEW_DYNAREC
+void
+x87_to_mmxreg(uint16_t reg)
+#else
+void
+x87_to_mmxreg(void)
+#endif
+{
+#ifndef USE_NEW_DYNAREC
+    uint32_t reg = cpu_state.TOP & 7;
+#endif
+    double val = cpu_state.ST[reg & 7];
+#ifdef X87_INLINE_ASM
+    unsigned char buffer[10];
+#else
+    x87_conv_t test;
+#endif
+
+#ifdef X87_INLINE_ASM
+    __asm volatile(""
+        :
+        :
+        : "memory");
+
+    __asm volatile("fldl %1\n"
+                    "fstpt %0\n" : "=m"(buffer) : "m"(val));
+
+    cpu_state.MM[reg & 7].q = (*(uint64_t*)buffer);
+#else
+    x87_to80(val, &test);
+    cpu_state.MM[reg & 7].q = test.eind.ll;
+#endif
+}
+
 static __inline void
 fetch_ea_32_long(uint32_t rmdat)
 {
diff --git a/src/cpu/cpu.h b/src/cpu/cpu.h
index 1d003ddc9..98c6f4938 100644
--- a/src/cpu/cpu.h
+++ b/src/cpu/cpu.h
@@ -829,6 +829,12 @@ extern uint16_t prefetch_queue_get_ip(void);
 extern int      prefetch_queue_get_prefetching(void);
 extern int      prefetch_queue_get_size(void);
 
+#ifdef USE_NEW_DYNAREC
+extern void     x87_to_mmxreg(uint16_t reg);
+#else
+extern void     x87_to_mmxreg(void);
+#endif
+
 #define prefetch_queue_set_suspended(s) prefetch_queue_set_prefetching(!s)
 #define prefetch_queue_get_suspended !prefetch_queue_get_prefetching