diff --git a/Makefile.target b/Makefile.target
index f31cc788cc..fe0cf372a2 100644
--- a/Makefile.target
+++ b/Makefile.target
@@ -658,7 +658,7 @@ endif
 ifeq ($(TARGET_BASE_ARCH), mips)
 helper.o: cpu.h exec-all.h
 op.o: op_template.c fop_template.c op_mem.c exec.h cpu.h
-op_helper.o: op_helper_mem.c exec.h softmmu_template.h cpu.h
+op_helper.o: exec.h softmmu_template.h cpu.h
 translate.o: translate_init.c exec-all.h disas.h
 endif
 
diff --git a/target-mips/exec.h b/target-mips/exec.h
index 8e909787e8..ea95da06e2 100644
--- a/target-mips/exec.h
+++ b/target-mips/exec.h
@@ -100,36 +100,6 @@ void fpu_dump_state(CPUState *env, FILE *f,
                     int (*fpu_fprintf)(FILE *f, const char *fmt, ...),
                     int flags);
 void dump_sc (void);
-void do_lwl_raw (uint32_t);
-void do_lwr_raw (uint32_t);
-uint32_t do_swl_raw (uint32_t);
-uint32_t do_swr_raw (uint32_t);
-#if defined(TARGET_MIPSN32) || defined(TARGET_MIPS64)
-void do_ldl_raw (uint64_t);
-void do_ldr_raw (uint64_t);
-uint64_t do_sdl_raw (uint64_t);
-uint64_t do_sdr_raw (uint64_t);
-#endif
-#if !defined(CONFIG_USER_ONLY)
-void do_lwl_user (uint32_t);
-void do_lwl_kernel (uint32_t);
-void do_lwr_user (uint32_t);
-void do_lwr_kernel (uint32_t);
-uint32_t do_swl_user (uint32_t);
-uint32_t do_swl_kernel (uint32_t);
-uint32_t do_swr_user (uint32_t);
-uint32_t do_swr_kernel (uint32_t);
-#if defined(TARGET_MIPSN32) || defined(TARGET_MIPS64)
-void do_ldl_user (uint64_t);
-void do_ldl_kernel (uint64_t);
-void do_ldr_user (uint64_t);
-void do_ldr_kernel (uint64_t);
-uint64_t do_sdl_user (uint64_t);
-uint64_t do_sdl_kernel (uint64_t);
-uint64_t do_sdr_user (uint64_t);
-uint64_t do_sdr_kernel (uint64_t);
-#endif
-#endif
 void do_pmon (int function);
 
 void dump_sc (void);
diff --git a/target-mips/op_helper.c b/target-mips/op_helper.c
index f4313280e8..3c23d8c362 100644
--- a/target-mips/op_helper.c
+++ b/target-mips/op_helper.c
@@ -62,18 +62,6 @@ void do_raise_exception_direct (uint32_t exception)
     do_raise_exception_direct_err (exception, 0);
 }
 
-#define MEMSUFFIX _raw
-#include "op_helper_mem.c"
-#undef MEMSUFFIX
-#if !defined(CONFIG_USER_ONLY)
-#define MEMSUFFIX _user
-#include "op_helper_mem.c"
-#undef MEMSUFFIX
-#define MEMSUFFIX _kernel
-#include "op_helper_mem.c"
-#undef MEMSUFFIX
-#endif
-
 #if defined(TARGET_MIPSN32) || defined(TARGET_MIPS64)
 #if TARGET_LONG_BITS > HOST_LONG_BITS
 /* Those might call libgcc functions.  */
diff --git a/target-mips/op_mem.c b/target-mips/op_mem.c
index a15ad5a961..d402d616b1 100644
--- a/target-mips/op_mem.c
+++ b/target-mips/op_mem.c
@@ -75,34 +75,92 @@ void glue(op_sw, MEMSUFFIX) (void)
 
 /* "half" load and stores.  We must do the memory access inline,
    or fault handling won't work.  */
-/* XXX: This is broken, CP0_BADVADDR has the wrong (aligned) value. */
+
+#ifdef TARGET_WORDS_BIGENDIAN
+#define GET_LMASK(v) ((v) & 3)
+#define GET_OFFSET(addr, offset) (addr + (offset))
+#else
+#define GET_LMASK(v) (((v) & 3) ^ 3)
+#define GET_OFFSET(addr, offset) (addr - (offset))
+#endif
+
 void glue(op_lwl, MEMSUFFIX) (void)
 {
-    uint32_t tmp = glue(ldl, MEMSUFFIX)(T0 & ~3);
-    CALL_FROM_TB1(glue(do_lwl, MEMSUFFIX), tmp);
+    target_ulong tmp;
+
+    tmp = glue(ldub, MEMSUFFIX)(T0);
+    T1 = (int32_t)((T1 & 0x00FFFFFF) | (tmp << 24));
+
+    if (GET_LMASK(T0) <= 2) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 1));
+        T1 = (T1 & 0xFF00FFFF) | (tmp << 16);
+    }
+
+    if (GET_LMASK(T0) <= 1) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 2));
+        T1 = (T1 & 0xFFFF00FF) | (tmp << 8);
+    }
+
+    if (GET_LMASK(T0) == 0) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 3));
+        T1 = (T1 & 0xFFFFFF00) | tmp;
+    }
     RETURN();
 }
 
 void glue(op_lwr, MEMSUFFIX) (void)
 {
-    uint32_t tmp = glue(ldl, MEMSUFFIX)(T0 & ~3);
-    CALL_FROM_TB1(glue(do_lwr, MEMSUFFIX), tmp);
+    target_ulong tmp;
+
+    tmp = glue(ldub, MEMSUFFIX)(T0);
+    T1 = (T1 & 0xFFFFFF00) | tmp;
+
+    if (GET_LMASK(T0) >= 1) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -1));
+        T1 = (T1 & 0xFFFF00FF) | (tmp << 8);
+    }
+
+    if (GET_LMASK(T0) >= 2) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -2));
+        T1 = (T1 & 0xFF00FFFF) | (tmp << 16);
+    }
+
+    if (GET_LMASK(T0) == 3) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -3));
+        T1 = (T1 & 0x00FFFFFF) | (tmp << 24);
+    }
     RETURN();
 }
 
 void glue(op_swl, MEMSUFFIX) (void)
 {
-    uint32_t tmp = glue(ldl, MEMSUFFIX)(T0 & ~3);
-    tmp = CALL_FROM_TB1(glue(do_swl, MEMSUFFIX), tmp);
-    glue(stl, MEMSUFFIX)(T0 & ~3, tmp);
+    glue(stb, MEMSUFFIX)(T0, (uint8_t)(T1 >> 24));
+
+    if (GET_LMASK(T0) <= 2)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 1), (uint8_t)(T1 >> 16));
+
+    if (GET_LMASK(T0) <= 1)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 2), (uint8_t)(T1 >> 8));
+
+    if (GET_LMASK(T0) == 0)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 3), (uint8_t)T1);
+
     RETURN();
 }
 
 void glue(op_swr, MEMSUFFIX) (void)
 {
-    uint32_t tmp = glue(ldl, MEMSUFFIX)(T0 & ~3);
-    tmp = CALL_FROM_TB1(glue(do_swr, MEMSUFFIX), tmp);
-    glue(stl, MEMSUFFIX)(T0 & ~3, tmp);
+    glue(stb, MEMSUFFIX)(T0, (uint8_t)T1);
+
+    if (GET_LMASK(T0) >= 1)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -1), (uint8_t)(T1 >> 8));
+
+    if (GET_LMASK(T0) >= 2)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -2), (uint8_t)(T1 >> 16));
+
+    if (GET_LMASK(T0) == 3)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -3), (uint8_t)(T1 >> 24));
+
     RETURN();
 }
 
@@ -145,33 +203,156 @@ void glue(op_sd, MEMSUFFIX) (void)
 
 /* "half" load and stores.  We must do the memory access inline,
    or fault handling won't work.  */
+
+#ifdef TARGET_WORDS_BIGENDIAN
+#define GET_LMASK64(v) ((v) & 7)
+#else
+#define GET_LMASK64(v) (((v) & 7) ^ 7)
+#endif
+
 void glue(op_ldl, MEMSUFFIX) (void)
 {
-    target_long tmp = glue(ldq, MEMSUFFIX)(T0 & ~7);
-    CALL_FROM_TB1(glue(do_ldl, MEMSUFFIX), tmp);
+    uint64_t tmp;
+
+    tmp = glue(ldub, MEMSUFFIX)(T0);
+    T1 = (T1 & 0x00FFFFFFFFFFFFFFULL) | (tmp << 56);
+
+    if (GET_LMASK64(T0) <= 6) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 1));
+        T1 = (T1 & 0xFF00FFFFFFFFFFFFULL) | (tmp << 48);
+    }
+
+    if (GET_LMASK64(T0) <= 5) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 2));
+        T1 = (T1 & 0xFFFF00FFFFFFFFFFULL) | (tmp << 40);
+    }
+
+    if (GET_LMASK64(T0) <= 4) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 3));
+        T1 = (T1 & 0xFFFFFF00FFFFFFFFULL) | (tmp << 32);
+    }
+
+    if (GET_LMASK64(T0) <= 3) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 4));
+        T1 = (T1 & 0xFFFFFFFF00FFFFFFULL) | (tmp << 24);
+    }
+
+    if (GET_LMASK64(T0) <= 2) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 5));
+        T1 = (T1 & 0xFFFFFFFFFF00FFFFULL) | (tmp << 16);
+    }
+
+    if (GET_LMASK64(T0) <= 1) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 6));
+        T1 = (T1 & 0xFFFFFFFFFFFF00FFULL) | (tmp << 8);
+    }
+
+    if (GET_LMASK64(T0) == 0) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, 7));
+        T1 = (T1 & 0xFFFFFFFFFFFFFF00ULL) | tmp;
+    }
+
     RETURN();
 }
 
 void glue(op_ldr, MEMSUFFIX) (void)
 {
-    target_long tmp = glue(ldq, MEMSUFFIX)(T0 & ~7);
-    CALL_FROM_TB1(glue(do_ldr, MEMSUFFIX), tmp);
+    uint64_t tmp;
+
+    tmp = glue(ldub, MEMSUFFIX)(T0);
+    T1 = (T1 & 0xFFFFFFFFFFFFFF00ULL) | tmp;
+
+    if (GET_LMASK64(T0) >= 1) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -1));
+        T1 = (T1 & 0xFFFFFFFFFFFF00FFULL) | (tmp  << 8);
+    }
+
+    if (GET_LMASK64(T0) >= 2) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -2));
+        T1 = (T1 & 0xFFFFFFFFFF00FFFFULL) | (tmp << 16);
+    }
+
+    if (GET_LMASK64(T0) >= 3) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -3));
+        T1 = (T1 & 0xFFFFFFFF00FFFFFFULL) | (tmp << 24);
+    }
+
+    if (GET_LMASK64(T0) >= 4) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -4));
+        T1 = (T1 & 0xFFFFFF00FFFFFFFFULL) | (tmp << 32);
+    }
+
+    if (GET_LMASK64(T0) >= 5) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -5));
+        T1 = (T1 & 0xFFFF00FFFFFFFFFFULL) | (tmp << 40);
+    }
+
+    if (GET_LMASK64(T0) >= 6) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -6));
+        T1 = (T1 & 0xFF00FFFFFFFFFFFFULL) | (tmp << 48);
+    }
+
+    if (GET_LMASK64(T0) == 7) {
+        tmp = glue(ldub, MEMSUFFIX)(GET_OFFSET(T0, -7));
+        T1 = (T1 & 0x00FFFFFFFFFFFFFFULL) | (tmp << 56);
+    }
+
     RETURN();
 }
 
 void glue(op_sdl, MEMSUFFIX) (void)
 {
-    target_long tmp = glue(ldq, MEMSUFFIX)(T0 & ~7);
-    tmp = CALL_FROM_TB1(glue(do_sdl, MEMSUFFIX), tmp);
-    glue(stq, MEMSUFFIX)(T0 & ~7, tmp);
+    glue(stb, MEMSUFFIX)(T0, (uint8_t)(T1 >> 56));
+
+    if (GET_LMASK64(T0) <= 6)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 1), (uint8_t)(T1 >> 48));
+
+    if (GET_LMASK64(T0) <= 5)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 2), (uint8_t)(T1 >> 40));
+
+    if (GET_LMASK64(T0) <= 4)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 3), (uint8_t)(T1 >> 32));
+
+    if (GET_LMASK64(T0) <= 3)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 4), (uint8_t)(T1 >> 24));
+
+    if (GET_LMASK64(T0) <= 2)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 5), (uint8_t)(T1 >> 16));
+
+    if (GET_LMASK64(T0) <= 1)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 6), (uint8_t)(T1 >> 8));
+
+    if (GET_LMASK64(T0) <= 0)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, 7), (uint8_t)T1);
+
     RETURN();
 }
 
 void glue(op_sdr, MEMSUFFIX) (void)
 {
-    target_long tmp = glue(ldq, MEMSUFFIX)(T0 & ~7);
-    tmp = CALL_FROM_TB1(glue(do_sdr, MEMSUFFIX), tmp);
-    glue(stq, MEMSUFFIX)(T0 & ~7, tmp);
+    glue(stb, MEMSUFFIX)(T0, (uint8_t)T1);
+
+    if (GET_LMASK64(T0) >= 1)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -1), (uint8_t)(T1 >> 8));
+
+    if (GET_LMASK64(T0) >= 2)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -2), (uint8_t)(T1 >> 16));
+
+    if (GET_LMASK64(T0) >= 3)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -3), (uint8_t)(T1 >> 24));
+
+    if (GET_LMASK64(T0) >= 4)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -4), (uint8_t)(T1 >> 32));
+
+    if (GET_LMASK64(T0) >= 5)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -5), (uint8_t)(T1 >> 40));
+
+    if (GET_LMASK64(T0) >= 6)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -6), (uint8_t)(T1 >> 48));
+
+    if (GET_LMASK64(T0) == 7)
+        glue(stb, MEMSUFFIX)(GET_OFFSET(T0, -7), (uint8_t)(T1 >> 56));
+
     RETURN();
 }
 
diff --git a/target-mips/translate.c b/target-mips/translate.c
index d3f80f2983..77499e83e5 100644
--- a/target-mips/translate.c
+++ b/target-mips/translate.c
@@ -889,7 +889,7 @@ static void gen_ldst (DisasContext *ctx, uint32_t opc, int rt,
     case OPC_LDL:
         GEN_LOAD_REG_TN(T1, rt);
         op_ldst(ldl);
-        GEN_STORE_TN_REG(rt, T0);
+        GEN_STORE_TN_REG(rt, T1);
         opn = "ldl";
         break;
     case OPC_SDL:
@@ -900,7 +900,7 @@ static void gen_ldst (DisasContext *ctx, uint32_t opc, int rt,
     case OPC_LDR:
         GEN_LOAD_REG_TN(T1, rt);
         op_ldst(ldr);
-        GEN_STORE_TN_REG(rt, T0);
+        GEN_STORE_TN_REG(rt, T1);
         opn = "ldr";
         break;
     case OPC_SDR:
@@ -952,7 +952,7 @@ static void gen_ldst (DisasContext *ctx, uint32_t opc, int rt,
     case OPC_LWL:
 	GEN_LOAD_REG_TN(T1, rt);
         op_ldst(lwl);
-        GEN_STORE_TN_REG(rt, T0);
+        GEN_STORE_TN_REG(rt, T1);
         opn = "lwl";
         break;
     case OPC_SWL:
@@ -963,7 +963,7 @@ static void gen_ldst (DisasContext *ctx, uint32_t opc, int rt,
     case OPC_LWR:
 	GEN_LOAD_REG_TN(T1, rt);
         op_ldst(lwr);
-        GEN_STORE_TN_REG(rt, T0);
+        GEN_STORE_TN_REG(rt, T1);
         opn = "lwr";
         break;
     case OPC_SWR: