ARM TCG conversion 13/16.

git-svn-id: svn://svn.savannah.nongnu.org/qemu/trunk@4150 c046a42c-6fe2-441c-8c8c-71466251a162

7 files changed, 204 insertions, 280 deletions

diff --git a/target-arm/exec.h b/target-arm/exec.h
index af18181..bd4910d 100644
--- a/target-arm/exec.h
+++ b/target-arm/exec.h
@@ -23,7 +23,6 @@
 register struct CPUARMState *env asm(AREG0);
 register uint32_t T0 asm(AREG1);
 register uint32_t T1 asm(AREG2);
-register uint32_t T2 asm(AREG3);
 
 #define M0   env->iwmmxt.val
 
@@ -59,15 +58,8 @@ static inline int cpu_halted(CPUState *env) {
 #include "softmmu_exec.h"
 #endif
 
-/* In op_helper.c */
-
-void helper_mark_exclusive(CPUARMState *, uint32_t addr);
-int helper_test_exclusive(CPUARMState *, uint32_t addr);
-void helper_clrex(CPUARMState *env);
-
 void cpu_loop_exit(void);
 
 void raise_exception(int);
 
-void helper_neon_tbl(int rn, int maxindex);
 uint32_t helper_neon_mul_p8(uint32_t op1, uint32_t op2);
diff --git a/target-arm/helper.c b/target-arm/helper.c
index 6438882..48cd6c8 100644
--- a/target-arm/helper.c
+++ b/target-arm/helper.c
@@ -432,7 +432,7 @@ static void flush_mmon(uint32_t addr)
 }
 
 /* Mark an address for exclusive access.  */
-void helper_mark_exclusive(CPUState *env, uint32_t addr)
+void HELPER(mark_exclusive)(CPUState *env, uint32_t addr)
 {
     if (!env->mmon_entry)
         allocate_mmon_state(env);
@@ -443,7 +443,7 @@ void helper_mark_exclusive(CPUState *env, uint32_t addr)
 
 /* Test if an exclusive address is still exclusive.  Returns zero
    if the address is still exclusive.   */
-int helper_test_exclusive(CPUState *env, uint32_t addr)
+uint32_t HELPER(test_exclusive)(CPUState *env, uint32_t addr)
 {
     int res;
 
@@ -457,7 +457,7 @@ int helper_test_exclusive(CPUState *env, uint32_t addr)
     return res;
 }
 
-void helper_clrex(CPUState *env)
+void HELPER(clrex)(CPUState *env)
 {
     if (!(env->mmon_entry && env->mmon_entry->addr))
         return;
@@ -1176,17 +1176,17 @@ target_phys_addr_t cpu_get_phys_page_debug(CPUState *env, target_ulong addr)
 /* Not really implemented.  Need to figure out a sane way of doing this.
    Maybe add generic watchpoint support and use that.  */
 
-void helper_mark_exclusive(CPUState *env, uint32_t addr)
+void HELPER(mark_exclusive)(CPUState *env, uint32_t addr)
 {
     env->mmon_addr = addr;
 }
 
-int helper_test_exclusive(CPUState *env, uint32_t addr)
+uint32_t HELPER(test_exclusive)(CPUState *env, uint32_t addr)
 {
     return (env->mmon_addr != addr);
 }
 
-void helper_clrex(CPUState *env)
+void HELPER(clrex)(CPUState *env)
 {
     env->mmon_addr = -1;
 }
@@ -2496,6 +2496,8 @@ float32 HELPER(rsqrts_f32)(float32 a, float32 b, CPUState *env)
     return float32_sub(three, float32_mul(a, b, s), s);
 }
 
+/* NEON helpers.  */
+
 /* TODO: The architecture specifies the value that the estimate functions
    should return.  We return the exact reciprocal/root instead.  */
 float32 HELPER(recpe_f32)(float32 a, CPUState *env)
diff --git a/target-arm/helpers.h b/target-arm/helpers.h
index 08eb590..8ebd25f 100644
--- a/target-arm/helpers.h
+++ b/target-arm/helpers.h
@@ -51,6 +51,13 @@ static inline void gen_helper_##name(TCGv ret, \
 { \
     tcg_gen_helper_1_3(helper_##name, ret, arg1, arg2, arg3); \
 }
+#define DEF_HELPER_1_4(name, ret, args) \
+DEF_HELPER(name, ret, args) \
+static inline void gen_helper_##name(TCGv ret, \
+    TCGv arg1, TCGv arg2, TCGv arg3, TCGv arg4) \
+{ \
+    tcg_gen_helper_1_4(helper_##name, ret, arg1, arg2, arg3, arg4); \
+}
 #else /* !GEN_HELPER */
 #define DEF_HELPER_0_0 DEF_HELPER
 #define DEF_HELPER_0_1 DEF_HELPER
@@ -60,6 +67,7 @@ static inline void gen_helper_##name(TCGv ret, \
 #define DEF_HELPER_1_1 DEF_HELPER
 #define DEF_HELPER_1_2 DEF_HELPER
 #define DEF_HELPER_1_3 DEF_HELPER
+#define DEF_HELPER_1_4 DEF_HELPER
 #define HELPER(x) glue(helper_,x)
 #endif
 
@@ -130,6 +138,10 @@ DEF_HELPER_1_2(get_cp, uint32_t, (CPUState *, uint32_t))
 DEF_HELPER_1_2(get_r13_banked, uint32_t, (CPUState *, uint32_t))
 DEF_HELPER_0_3(set_r13_banked, void, (CPUState *, uint32_t, uint32_t))
 
+DEF_HELPER_0_2(mark_exclusive, void, (CPUState *, uint32_t))
+DEF_HELPER_1_2(test_exclusive, uint32_t, (CPUState *, uint32_t))
+DEF_HELPER_0_1(clrex, void, (CPUState *))
+
 DEF_HELPER_1_1(get_user_reg, uint32_t, (uint32_t))
 DEF_HELPER_0_2(set_user_reg, void, (uint32_t, uint32_t))
 
@@ -195,6 +207,7 @@ DEF_HELPER_1_2(recpe_f32, float32, (float32, CPUState *))
 DEF_HELPER_1_2(rsqrte_f32, float32, (float32, CPUState *))
 DEF_HELPER_1_2(recpe_u32, uint32_t, (uint32_t, CPUState *))
 DEF_HELPER_1_2(rsqrte_u32, uint32_t, (uint32_t, CPUState *))
+DEF_HELPER_1_4(neon_tbl, uint32_t, (uint32_t, uint32_t, uint32_t, uint32_t))
 
 DEF_HELPER_1_2(add_cc, uint32_t, (uint32_t, uint32_t))
 DEF_HELPER_1_2(adc_cc, uint32_t, (uint32_t, uint32_t))
diff --git a/target-arm/op_helper.c b/target-arm/op_helper.c
index 79b3f1a..d1ce3a6 100644
--- a/target-arm/op_helper.c
+++ b/target-arm/op_helper.c
@@ -40,27 +40,26 @@ void cpu_unlock(void)
     spin_unlock(&global_cpu_lock);
 }
 
-void helper_neon_tbl(int rn, int maxindex)
+uint32_t HELPER(neon_tbl)(uint32_t ireg, uint32_t def,
+                          uint32_t rn, uint32_t maxindex)
 {
     uint32_t val;
-    uint32_t mask;
     uint32_t tmp;
     int index;
     int shift;
     uint64_t *table;
     table = (uint64_t *)&env->vfp.regs[rn];
     val = 0;
-    mask = 0;
     for (shift = 0; shift < 32; shift += 8) {
-        index = (T1 >> shift) & 0xff;
-        if (index <= maxindex) {
+        index = (ireg >> shift) & 0xff;
+        if (index < maxindex) {
             tmp = (table[index >> 3] >> (index & 7)) & 0xff;
             val |= tmp << shift;
         } else {
-            val |= T0 & (0xff << shift);
+            val |= def & (0xff << shift);
         }
     }
-    T0 = val;
+    return val;
 }
 
 #if !defined(CONFIG_USER_ONLY)
diff --git a/target-arm/op_mem.h b/target-arm/op_mem.h
index 519544d..daaab51 100644
--- a/target-arm/op_mem.h
+++ b/target-arm/op_mem.h
@@ -1,63 +1,5 @@
 /* ARM memory operations.  */
 
-/* Load-locked, store exclusive.  */
-#define EXCLUSIVE_OP(suffix, ldsuffix) \
-void OPPROTO glue(op_ld##suffix##ex,MEMSUFFIX)(void) \
-{ \
-    cpu_lock(); \
-    helper_mark_exclusive(env, T1); \
-    T0 = glue(ld##ldsuffix,MEMSUFFIX)(T1); \
-    cpu_unlock(); \
-    FORCE_RET(); \
-} \
- \
-void OPPROTO glue(op_st##suffix##ex,MEMSUFFIX)(void) \
-{ \
-    int failed; \
-    cpu_lock(); \
-    failed = helper_test_exclusive(env, T1); \
-    /* ??? Is it safe to hold the cpu lock over a store?  */ \
-    if (!failed) { \
-        glue(st##suffix,MEMSUFFIX)(T1, T0); \
-    } \
-    T0 = failed; \
-    cpu_unlock(); \
-    FORCE_RET(); \
-}
-
-EXCLUSIVE_OP(b, ub)
-EXCLUSIVE_OP(w, uw)
-EXCLUSIVE_OP(l, l)
-
-#undef EXCLUSIVE_OP
-
-/* Load exclusive T0:T1 from address T1.  */
-void OPPROTO glue(op_ldqex,MEMSUFFIX)(void)
-{
-    cpu_lock();
-    helper_mark_exclusive(env, T1);
-    T0 = glue(ldl,MEMSUFFIX)(T1);
-    T1 = glue(ldl,MEMSUFFIX)((T1 + 4));
-    cpu_unlock();
-    FORCE_RET();
-}
-
-/* Store exclusive T0:T2 to address T1.  */
-void OPPROTO glue(op_stqex,MEMSUFFIX)(void)
-{
-    int failed;
-    cpu_lock();
-    failed = helper_test_exclusive(env, T1);
-    /* ??? Is it safe to hold the cpu lock over a store?  */
-    if (!failed) {
-        glue(stl,MEMSUFFIX)(T1, T0);
-        glue(stl,MEMSUFFIX)((T1 + 4), T2);
-    }
-    T0 = failed;
-    cpu_unlock();
-    FORCE_RET();
-}
-
 /* iwMMXt load/store.  Address is in T1 */
 #define MMX_MEM_OP(name, ldname) \
 void OPPROTO glue(op_iwmmxt_ld##name,MEMSUFFIX)(void) \
diff --git a/target-arm/op_neon.h b/target-arm/op_neon.h
index 095165e..df3b7cb 100644
--- a/target-arm/op_neon.h
+++ b/target-arm/op_neon.h
@@ -47,11 +47,6 @@ NEON_OP(getreg_T1)
     T1 = *(uint32_t *)((char *) env + PARAM1);
 }
 
-NEON_OP(getreg_T2)
-{
-    T2 = *(uint32_t *)((char *) env + PARAM1);
-}
-
 NEON_OP(setreg_T0)
 {
     *(uint32_t *)((char *) env + PARAM1) = T0;
@@ -62,11 +57,6 @@ NEON_OP(setreg_T1)
     *(uint32_t *)((char *) env + PARAM1) = T1;
 }
 
-NEON_OP(setreg_T2)
-{
-    *(uint32_t *)((char *) env + PARAM1) = T2;
-}
-
 #define NEON_TYPE1(name, type) \
 typedef struct \
 { \
@@ -293,28 +283,6 @@ NEON_OP(hsub_u32)
     FORCE_RET();
 }
 
-/* ??? bsl, bif and bit are all the same op, just with the oparands in a
-   differnet order.  It's currently easier to have 3 differnt ops than
-   rearange the operands.  */
-
-/* Bitwise Select.  */
-NEON_OP(bsl)
-{
-    T0 = (T0 & T2) | (T1 & ~T2);
-}
-
-/* Bitwise Insert If True.  */
-NEON_OP(bit)
-{
-    T0 = (T0 & T1) | (T2 & ~T1);
-}
-
-/* Bitwise Insert If False.  */
-NEON_OP(bif)
-{
-    T0 = (T2 & T1) | (T0 & ~T1);
-}
-
 #define NEON_USAT(dest, src1, src2, type) do { \
     uint32_t tmp = (uint32_t)src1 + (uint32_t)src2; \
     if (tmp != (type)tmp) { \
@@ -423,7 +391,7 @@ NEON_VOP(shl_u32, neon_u32, 1)
 
 NEON_OP(shl_u64)
 {
-    int8_t shift = T2;
+    int8_t shift = env->vfp.scratch[0];
     uint64_t val = T0 | ((uint64_t)T1 << 32);
     if (shift < 0) {
         val >>= -shift;
@@ -437,7 +405,7 @@ NEON_OP(shl_u64)
 
 NEON_OP(shl_s64)
 {
-    int8_t shift = T2;
+    int8_t shift = env->vfp.scratch[0];
     int64_t val = T0 | ((uint64_t)T1 << 32);
     if (shift < 0) {
         val >>= -shift;
@@ -468,7 +436,7 @@ NEON_VOP(rshl_u32, neon_u32, 1)
 
 NEON_OP(rshl_u64)
 {
-    int8_t shift = T2;
+    int8_t shift = env->vfp.scratch[0];
     uint64_t val = T0 | ((uint64_t)T1 << 32);
     if (shift < 0) {
         val = (val + ((uint64_t)1 << (-1 - shift))) >> -shift;
@@ -483,7 +451,7 @@ NEON_OP(rshl_u64)
 
 NEON_OP(rshl_s64)
 {
-    int8_t shift = T2;
+    int8_t shift = env->vfp.scratch[0];
     int64_t val = T0 | ((uint64_t)T1 << 32);
     if (shift < 0) {
         val = (val + ((int64_t)1 << (-1 - shift))) >> -shift;
@@ -514,7 +482,7 @@ NEON_VOP(qshl_s32, neon_s32, 1)
 
 NEON_OP(qshl_s64)
 {
-    int8_t shift = T2;
+    int8_t shift = env->vfp.scratch[0];
     int64_t val = T0 | ((uint64_t)T1 << 32);
     if (shift < 0) {
         val >>= -shift;
@@ -550,7 +518,7 @@ NEON_VOP(qshl_u32, neon_u32, 1)
 
 NEON_OP(qshl_u64)
 {
-    int8_t shift = T2;
+    int8_t shift = env->vfp.scratch[0];
     uint64_t val = T0 | ((uint64_t)T1 << 32);
     if (shift < 0) {
         val >>= -shift;
@@ -1713,12 +1681,6 @@ NEON_OP(zip_u16)
     FORCE_RET();
 }
 
-/* Table lookup.  This accessed the register file directly.  */
-NEON_OP(tbl)
-{
-    helper_neon_tbl(PARAM1, PARAM2);
-}
-
 NEON_OP(dup_u8)
 {
     T0 = (T0 >> PARAM1) & 0xff;
@@ -1726,20 +1688,3 @@ NEON_OP(dup_u8)
     T0 |= T0 << 16;
     FORCE_RET();
 }
-
-/* Helpers for element load/store.  */
-NEON_OP(insert_elt)
-{
-    int shift = PARAM1;
-    uint32_t mask = PARAM2;
-    T2 = (T2 & mask) | (T0 << shift);
-    FORCE_RET();
-}
-
-NEON_OP(extract_elt)
-{
-    int shift = PARAM1;
-    uint32_t mask = PARAM2;
-    T0 = (T2 & mask) >> shift;
-    FORCE_RET();
-}
diff --git a/target-arm/translate.c b/target-arm/translate.c
index 0220d6a..a4b1df5 100644
--- a/target-arm/translate.c
+++ b/target-arm/translate.c
@@ -78,7 +78,7 @@ extern int loglevel;
 
 static TCGv cpu_env;
 /* FIXME:  These should be removed.  */
-static TCGv cpu_T[3];
+static TCGv cpu_T[2];
 static TCGv cpu_F0s, cpu_F1s, cpu_F0d, cpu_F1d;
 
 /* initialize TCG globals.  */
@@ -88,7 +88,6 @@ void arm_translate_init(void)
 
     cpu_T[0] = tcg_global_reg_new(TCG_TYPE_I32, TCG_AREG1, "T0");
     cpu_T[1] = tcg_global_reg_new(TCG_TYPE_I32, TCG_AREG2, "T1");
-    cpu_T[2] = tcg_global_reg_new(TCG_TYPE_I32, TCG_AREG3, "T2");
 }
 
 /* The code generator doesn't like lots of temporaries, so maintain our own
@@ -188,13 +187,9 @@ static void store_reg(DisasContext *s, int reg, TCGv var)
 
 /* Basic operations.  */
 #define gen_op_movl_T0_T1() tcg_gen_mov_i32(cpu_T[0], cpu_T[1])
-#define gen_op_movl_T0_T2() tcg_gen_mov_i32(cpu_T[0], cpu_T[2])
 #define gen_op_movl_T1_T0() tcg_gen_mov_i32(cpu_T[1], cpu_T[0])
-#define gen_op_movl_T1_T2() tcg_gen_mov_i32(cpu_T[1], cpu_T[2])
-#define gen_op_movl_T2_T0() tcg_gen_mov_i32(cpu_T[2], cpu_T[0])
 #define gen_op_movl_T0_im(im) tcg_gen_movi_i32(cpu_T[0], im)
 #define gen_op_movl_T1_im(im) tcg_gen_movi_i32(cpu_T[1], im)
-#define gen_op_movl_T2_im(im) tcg_gen_movi_i32(cpu_T[2], im)
 
 #define gen_op_addl_T1_im(im) tcg_gen_addi_i32(cpu_T[1], cpu_T[1], im)
 #define gen_op_addl_T0_T1() tcg_gen_add_i32(cpu_T[0], cpu_T[0], cpu_T[1])
@@ -310,9 +305,9 @@ static void gen_sbfx(TCGv var, int shift, int width)
 /* Bitfield insertion.  Insert val into base.  Clobbers base and val.  */
 static void gen_bfi(TCGv dest, TCGv base, TCGv val, int shift, uint32_t mask)
 {
-    tcg_gen_shli_i32(val, val, shift);
     tcg_gen_andi_i32(val, val, mask);
-    tcg_gen_andi_i32(base, base, ~mask);
+    tcg_gen_shli_i32(val, val, shift);
+    tcg_gen_andi_i32(base, base, ~(mask << shift));
     tcg_gen_or_i32(dest, base, val);
 }
 
@@ -460,6 +455,13 @@ static inline void tcg_gen_not_i32(TCGv t0, TCGv t1)
 
 /* T0 &= ~T1.  Clobbers T1.  */
 /* FIXME: Implement bic natively.  */
+static inline void tcg_gen_bic_i32(TCGv dest, TCGv t0, TCGv t1)
+{
+    TCGv tmp = new_tmp();
+    tcg_gen_not_i32(tmp, t1);
+    tcg_gen_and_i32(dest, t0, tmp);
+    dead_tmp(tmp);
+}
 static inline void gen_op_bicl_T0_T1(void)
 {
     gen_op_notl_T1();
@@ -1167,6 +1169,19 @@ neon_reg_offset (int reg, int n)
 #define NEON_GET_REG(T, reg, n) gen_op_neon_getreg_##T(neon_reg_offset(reg, n))
 #define NEON_SET_REG(T, reg, n) gen_op_neon_setreg_##T(neon_reg_offset(reg, n))
 
+static TCGv neon_load_reg(int reg, int pass)
+{
+    TCGv tmp = new_tmp();
+    tcg_gen_ld_i32(tmp, cpu_env, neon_reg_offset(reg, pass));
+    return tmp;
+}
+
+static void neon_store_reg(int reg, int pass, TCGv var)
+{
+    tcg_gen_st_i32(var, cpu_env, neon_reg_offset(reg, pass));
+    dead_tmp(var);
+}
+
 #define tcg_gen_ld_f32 tcg_gen_ld_i32
 #define tcg_gen_ld_f64 tcg_gen_ld_i64
 #define tcg_gen_st_f32 tcg_gen_st_i32
@@ -2500,19 +2515,14 @@ static int disas_vfp_insn(CPUState * env, DisasContext *s, uint32_t insn)
                         /* VMOV */
                         switch (size) {
                         case 0:
-                            NEON_GET_REG(T2, rn, pass);
-                            gen_op_movl_T1_im(0xff);
-                            gen_op_andl_T0_T1();
-                            gen_op_neon_insert_elt(offset, ~(0xff << offset));
-                            NEON_SET_REG(T2, rn, pass);
+                            tmp = neon_load_reg(rn, pass);
+                            gen_bfi(tmp, tmp, cpu_T[0], offset, 0xff);
+                            neon_store_reg(rn, pass, tmp);
                             break;
                         case 1:
-                            NEON_GET_REG(T2, rn, pass);
-                            gen_op_movl_T1_im(0xffff);
-                            gen_op_andl_T0_T1();
-                            bank_mask = offset ? 0xffff : 0xffff0000;
-                            gen_op_neon_insert_elt(offset, bank_mask);
-                            NEON_SET_REG(T2, rn, pass);
+                            tmp = neon_load_reg(rn, pass);
+                            gen_bfi(tmp, tmp, cpu_T[0], offset, 0xffff);
+                            neon_store_reg(rn, pass, tmp);
                             break;
                         case 2:
                             NEON_SET_REG(T0, rn, pass);
@@ -3480,9 +3490,9 @@ static int disas_neon_ls_insn(CPUState * env, DisasContext *s, uint32_t insn)
     int pass;
     int load;
     int shift;
-    uint32_t mask;
     int n;
     TCGv tmp;
+    TCGv tmp2;
 
     if (!vfp_enabled(env))
       return 1;
@@ -3525,60 +3535,47 @@ static int disas_neon_ls_insn(CPUState * env, DisasContext *s, uint32_t insn)
                 } else if (size == 1) {
                     if (load) {
                         tmp = gen_ld16u(cpu_T[1], IS_USER(s));
-                        tcg_gen_mov_i32(cpu_T[0], tmp);
-                        dead_tmp(tmp);
                         gen_op_addl_T1_im(stride);
-                        gen_op_movl_T2_T0();
-                        tmp = gen_ld16u(cpu_T[1], IS_USER(s));
-                        tcg_gen_mov_i32(cpu_T[0], tmp);
-                        dead_tmp(tmp);
+                        tmp2 = gen_ld16u(cpu_T[1], IS_USER(s));
                         gen_op_addl_T1_im(stride);
-                        gen_op_neon_insert_elt(16, 0xffff);
-                        NEON_SET_REG(T2, rd, pass);
+                        gen_bfi(tmp, tmp, tmp2, 16, 0xffff);
+                        dead_tmp(tmp2);
+                        neon_store_reg(rd, pass, tmp);
                     } else {
-                        NEON_GET_REG(T2, rd, pass);
-                        gen_op_movl_T0_T2();
-                        tmp = new_tmp();
-                        tcg_gen_mov_i32(tmp, cpu_T[0]);
+                        tmp = neon_load_reg(rd, pass);
+                        tmp2 = new_tmp();
+                        tcg_gen_shri_i32(tmp2, tmp, 16);
                         gen_st16(tmp, cpu_T[1], IS_USER(s));
                         gen_op_addl_T1_im(stride);
-                        gen_op_neon_extract_elt(16, 0xffff0000);
-                        tmp = new_tmp();
-                        tcg_gen_mov_i32(tmp, cpu_T[0]);
-                        gen_st16(tmp, cpu_T[1], IS_USER(s));
+                        gen_st16(tmp2, cpu_T[1], IS_USER(s));
                         gen_op_addl_T1_im(stride);
                     }
                 } else /* size == 0 */ {
                     if (load) {
-                        mask = 0xff;
                         for (n = 0; n < 4; n++) {
                             tmp = gen_ld8u(cpu_T[1], IS_USER(s));
-                            tcg_gen_mov_i32(cpu_T[0], tmp);
-                            dead_tmp(tmp);
                             gen_op_addl_T1_im(stride);
                             if (n == 0) {
-                                gen_op_movl_T2_T0();
+                                tmp2 = tmp;
                             } else {
-                                gen_op_neon_insert_elt(n * 8, ~mask);
+                                gen_bfi(tmp2, tmp2, tmp, n * 8, 0xff);
+                                dead_tmp(tmp);
                             }
-                            mask <<= 8;
                         }
-                        NEON_SET_REG(T2, rd, pass);
+                        neon_store_reg(rd, pass, tmp2);
                     } else {
-                        NEON_GET_REG(T2, rd, pass);
-                        mask = 0xff;
+                        tmp2 = neon_load_reg(rd, pass);
                         for (n = 0; n < 4; n++) {
+                            tmp = new_tmp();
                             if (n == 0) {
-                                gen_op_movl_T0_T2();
+                                tcg_gen_mov_i32(tmp, tmp2);
                             } else {
-                                gen_op_neon_extract_elt(n * 8, mask);
+                                tcg_gen_shri_i32(tmp, tmp2, n * 8);
                             }
-                            tmp = new_tmp();
-                            tcg_gen_mov_i32(tmp, cpu_T[0]);
                             gen_st8(tmp, cpu_T[1], IS_USER(s));
                             gen_op_addl_T1_im(stride);
-                            mask <<= 8;
                         }
+                        dead_tmp(tmp2);
                     }
                 }
             }
@@ -3629,17 +3626,14 @@ static int disas_neon_ls_insn(CPUState * env, DisasContext *s, uint32_t insn)
             switch (size) {
             case 0:
                 shift = ((insn >> 5) & 3) * 8;
-                mask = 0xff << shift;
                 stride = 1;
                 break;
             case 1:
                 shift = ((insn >> 6) & 1) * 16;
-                mask = shift ? 0xffff0000 : 0xffff;
                 stride = (insn & (1 << 5)) ? 2 : 1;
                 break;
             case 2:
                 shift = 0;
-                mask = 0xffffffff;
                 stride = (insn & (1 << 6)) ? 2 : 1;
                 break;
             default:
@@ -3649,9 +3643,6 @@ static int disas_neon_ls_insn(CPUState * env, DisasContext *s, uint32_t insn)
             gen_movl_T1_reg(s, rn);
             for (reg = 0; reg < nregs; reg++) {
                 if (load) {
-                    if (size != 2) {
-                        NEON_GET_REG(T2, rd, pass);
-                    }
                     switch (size) {
                     case 0:
                         tmp = gen_ld8u(cpu_T[1], IS_USER(s));
@@ -3663,23 +3654,16 @@ static int disas_neon_ls_insn(CPUState * env, DisasContext *s, uint32_t insn)
                         tmp = gen_ld32(cpu_T[1], IS_USER(s));
                         break;
                     }
-                    tcg_gen_mov_i32(cpu_T[0], tmp);
-                    dead_tmp(tmp);
                     if (size != 2) {
-                        gen_op_neon_insert_elt(shift, ~mask);
-                        NEON_SET_REG(T0, rd, pass);
-                    } else {
-                        NEON_SET_REG(T0, rd, pass);
+                        tmp2 = neon_load_reg(rd, pass);
+                        gen_bfi(tmp, tmp2, tmp, shift, size ? 0xffff : 0xff);
+                        dead_tmp(tmp2);
                     }
+                    neon_store_reg(rd, pass, tmp);
                 } else { /* Store */
-                    if (size == 2) {
-                        NEON_GET_REG(T0, rd, pass);
-                    } else {
-                        NEON_GET_REG(T2, rd, pass);
-                        gen_op_neon_extract_elt(shift, mask);
-                    }
-                    tmp = new_tmp();
-                    tcg_gen_mov_i32(tmp, cpu_T[0]);
+                    tmp = neon_load_reg(rd, pass);
+                    if (shift)
+                        tcg_gen_shri_i32(tmp, tmp, shift);
                     switch (size) {
                     case 0:
                         gen_st8(tmp, cpu_T[1], IS_USER(s));
@@ -3715,6 +3699,14 @@ static int disas_neon_ls_insn(CPUState * env, DisasContext *s, uint32_t insn)
     return 0;
 }
 
+/* Bitwise select.  dest = c ? t : f.  Clobbers T and F.  */
+static void gen_neon_bsl(TCGv dest, TCGv t, TCGv f, TCGv c)
+{
+    tcg_gen_and_i32(t, t, c);
+    tcg_gen_bic_i32(f, f, c);
+    tcg_gen_or_i32(dest, t, f);
+}
+
 /* Translate a NEON data processing instruction.  Return nonzero if the
    instruction is invalid.
    In general we process vectors in 32-bit chunks.  This means we can reuse
@@ -3735,6 +3727,9 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
     int u;
     int n;
     uint32_t imm;
+    TCGv tmp;
+    TCGv tmp2;
+    TCGv tmp3;
 
     if (!vfp_enabled(env))
       return 1;
@@ -3875,16 +3870,19 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                 gen_op_xorl_T0_T1();
                 break;
             case 5: /* VBSL */
-                NEON_GET_REG(T2, rd, pass);
-                gen_op_neon_bsl();
+                tmp = neon_load_reg(rd, pass);
+                gen_neon_bsl(cpu_T[0], cpu_T[0], cpu_T[1], tmp);
+                dead_tmp(tmp);
                 break;
             case 6: /* VBIT */
-                NEON_GET_REG(T2, rd, pass);
-                gen_op_neon_bit();
+                tmp = neon_load_reg(rd, pass);
+                gen_neon_bsl(cpu_T[0], cpu_T[0], tmp, cpu_T[1]);
+                dead_tmp(tmp);
                 break;
             case 7: /* VBIF */
-                NEON_GET_REG(T2, rd, pass);
-                gen_op_neon_bif();
+                tmp = neon_load_reg(rd, pass);
+                gen_neon_bsl(cpu_T[0], tmp, cpu_T[0], cpu_T[1]);
+                dead_tmp(tmp);
                 break;
             }
             break;
@@ -4190,8 +4188,6 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                    element size in bits.  */
                 if (op <= 4)
                     shift = shift - (1 << (size + 3));
-                else
-                    shift++;
                 if (size == 3) {
                     count = q + 1;
                 } else {
@@ -4276,9 +4272,10 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                         default:
                             abort();
                         }
-                        NEON_GET_REG(T1, rd, pass);
-                        gen_op_movl_T2_im(imm);
-                        gen_op_neon_bsl();
+                        tmp = neon_load_reg(rd, pass);
+                        tcg_gen_andi_i32(cpu_T[0], cpu_T[0], imm);
+                        tcg_gen_andi_i32(tmp, tmp, ~imm);
+                        tcg_gen_or_i32(cpu_T[0], cpu_T[0], tmp);
                     }
                     if (size == 3) {
                         NEON_SET_REG(T0, rd, pass * 2);
@@ -4519,24 +4516,26 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                 /* Avoid overlapping operands.  Wide source operands are
                    always aligned so will never overlap with wide
                    destinations in problematic ways.  */
-                if (rd == rm) {
-                    NEON_GET_REG(T2, rm, 1);
-                } else if (rd == rn) {
-                    NEON_GET_REG(T2, rn, 1);
+                if (rd == rm && !src2_wide) {
+                    NEON_GET_REG(T0, rm, 1);
+                    gen_neon_movl_scratch_T0(2);
+                } else if (rd == rn && !src1_wide) {
+                    NEON_GET_REG(T0, rn, 1);
+                    gen_neon_movl_scratch_T0(2);
                 }
                 for (pass = 0; pass < 2; pass++) {
                     /* Load the second operand into env->vfp.scratch.
                        Also widen narrow operands.  */
-                    if (pass == 1 && rd == rm) {
-                        if (prewiden) {
-                            gen_op_movl_T0_T2();
-                        } else {
-                            gen_op_movl_T1_T2();
-                        }
+                    if (src2_wide) {
+                        NEON_GET_REG(T0, rm, pass * 2);
+                        NEON_GET_REG(T1, rm, pass * 2 + 1);
                     } else {
-                        if (src2_wide) {
-                            NEON_GET_REG(T0, rm, pass * 2);
-                            NEON_GET_REG(T1, rm, pass * 2 + 1);
+                        if (pass == 1 && rd == rm) {
+                            if (prewiden) {
+                                gen_neon_movl_T0_scratch(2);
+                            } else {
+                                gen_neon_movl_T1_scratch(2);
+                            }
                         } else {
                             if (prewiden) {
                                 NEON_GET_REG(T0, rm, pass);
@@ -4554,12 +4553,12 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                     }
 
                     /* Load the first operand.  */
-                    if (pass == 1 && rd == rn) {
-                        gen_op_movl_T0_T2();
+                    if (src1_wide) {
+                        NEON_GET_REG(T0, rn, pass * 2);
+                        NEON_GET_REG(T1, rn, pass * 2 + 1);
                     } else {
-                        if (src1_wide) {
-                            NEON_GET_REG(T0, rn, pass * 2);
-                            NEON_GET_REG(T1, rn, pass * 2 + 1);
+                        if (pass == 1 && rd == rn) {
+                            gen_neon_movl_T0_scratch(2);
                         } else {
                             NEON_GET_REG(T0, rn, pass);
                         }
@@ -4696,10 +4695,10 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                 case 12: /* VQDMULH scalar */
                 case 13: /* VQRDMULH scalar */
                     gen_neon_get_scalar(size, rm);
-                    gen_op_movl_T2_T0();
+                    gen_neon_movl_scratch_T0(0);
                     for (pass = 0; pass < (u ? 4 : 2); pass++) {
                         if (pass != 0)
-                            gen_op_movl_T0_T2();
+                            gen_neon_movl_T0_scratch(0);
                         NEON_GET_REG(T1, rn, pass);
                         if (op == 12) {
                             if (size == 1) {
@@ -4764,10 +4763,10 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                         gen_neon_movl_scratch_T0(2);
                     }
                     gen_neon_get_scalar(size, rm);
-                    gen_op_movl_T2_T0();
+                    gen_neon_movl_scratch_T0(3);
                     for (pass = 0; pass < 2; pass++) {
                         if (pass != 0) {
-                            gen_op_movl_T0_T2();
+                            gen_neon_movl_T0_scratch(3);
                         }
                         if (pass != 0 && rd == rn) {
                             gen_neon_movl_T1_scratch(2);
@@ -5025,11 +5024,12 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
                     if (q)
                         return 1;
                     if (rm == rd) {
-                        NEON_GET_REG(T2, rm, 1);
+                        NEON_GET_REG(T0, rm, 1);
+                        gen_neon_movl_scratch_T0(0);
                     }
                     for (pass = 0; pass < 2; pass++) {
                         if (pass == 1 && rm == rd) {
-                            gen_op_movl_T0_T2();
+                            gen_neon_movl_T0_scratch(0);
                         } else {
                             NEON_GET_REG(T0, rm, pass);
                         }
@@ -5253,23 +5253,26 @@ static int disas_neon_data_insn(CPUState * env, DisasContext *s, uint32_t insn)
             } else if ((insn & (1 << 10)) == 0) {
                 /* VTBL, VTBX.  */
                 n = (insn >> 5) & 0x18;
-                NEON_GET_REG(T1, rm, 0);
                 if (insn & (1 << 6)) {
-                    NEON_GET_REG(T0, rd, 0);
+                    tmp = neon_load_reg(rd, 0);
                 } else {
-                    gen_op_movl_T0_im(0);
+                    tmp = new_tmp();
+                    tcg_gen_movi_i32(tmp, 0);
                 }
-                gen_op_neon_tbl(rn, n);
-                gen_op_movl_T2_T0();
-                NEON_GET_REG(T1, rm, 1);
+                tmp2 = neon_load_reg(rm, 0);
+                gen_helper_neon_tbl(tmp2, tmp2, tmp, tcg_const_i32(rn),
+                                    tcg_const_i32(n));
                 if (insn & (1 << 6)) {
-                    NEON_GET_REG(T0, rd, 0);
+                    tmp = neon_load_reg(rd, 1);
                 } else {
-                    gen_op_movl_T0_im(0);
+                    tmp = new_tmp();
+                    tcg_gen_movi_i32(tmp, 0);
                 }
-                gen_op_neon_tbl(rn, n);
-                NEON_SET_REG(T2, rd, 0);
-                NEON_SET_REG(T0, rd, 1);
+                tmp3 = neon_load_reg(rm, 1);
+                gen_helper_neon_tbl(tmp3, tmp3, tmp, tcg_const_i32(rn),
+                                    tcg_const_i32(n));
+                neon_store_reg(rd, 0, tmp2);
+                neon_store_reg(rd, 1, tmp2);
             } else if ((insn & 0x380) == 0) {
                 /* VDUP */
                 if (insn & (1 << 19)) {
@@ -5430,7 +5433,7 @@ static void disas_arm_insn(CPUState * env, DisasContext *s)
             switch ((insn >> 4) & 0xf) {
             case 1: /* clrex */
                 ARCH(6K);
-                gen_op_clrex();
+                gen_helper_clrex(cpu_env);
                 return;
             case 4: /* dsb */
             case 5: /* dmb */
@@ -5977,13 +5980,19 @@ static void disas_arm_insn(CPUState * env, DisasContext *s)
                         /* load/store exclusive */
                         gen_movl_T1_reg(s, rn);
                         if (insn & (1 << 20)) {
-                            gen_ldst(ldlex, s);
+                            gen_helper_mark_exclusive(cpu_env, cpu_T[1]);
+                            tmp = gen_ld32(addr, IS_USER(s));
+                            store_reg(s, rd, tmp);
                         } else {
+                            int label = gen_new_label();
                             rm = insn & 0xf;
-                            gen_movl_T0_reg(s, rm);
-                            gen_ldst(stlex, s);
+                            gen_helper_test_exclusive(cpu_T[0], cpu_env, addr);
+                            tcg_gen_brcond_i32(TCG_COND_NE, cpu_T[0],
+                                               tcg_const_i32(0), label);
+                            tmp = load_reg(s,rm);
+                            gen_st32(tmp, cpu_T[1], IS_USER(s));
+                            gen_movl_reg_T0(s, rd);
                         }
-                        gen_movl_reg_T0(s, rd);
                     } else {
                         /* SWP instruction */
                         rm = (insn) & 0xf;
@@ -6287,8 +6296,7 @@ static void disas_arm_insn(CPUState * env, DisasContext *s)
                         }
                         if (i != 32) {
                             tmp2 = load_reg(s, rd);
-                            gen_bfi(tmp, tmp2, tmp,
-                                    shift, ((1u << i) - 1) << shift);
+                            gen_bfi(tmp, tmp2, tmp, shift, (1u << i) - 1);
                             dead_tmp(tmp2);
                         }
                         store_reg(s, rd, tmp);
@@ -6720,14 +6728,21 @@ static int disas_thumb2_insn(CPUState *env, DisasContext *s, uint16_t insn_hw1)
                 }
             } else if ((insn & (1 << 23)) == 0) {
                 /* Load/store exclusive word.  */
-                gen_movl_T0_reg(s, rd);
                 gen_movl_T1_reg(s, rn);
                 if (insn & (1 << 20)) {
-                    gen_ldst(ldlex, s);
+                    gen_helper_mark_exclusive(cpu_env, cpu_T[1]);
+                    tmp = gen_ld32(addr, IS_USER(s));
+                    store_reg(s, rd, tmp);
                 } else {
-                    gen_ldst(stlex, s);
+                    int label = gen_new_label();
+                    gen_helper_test_exclusive(cpu_T[0], cpu_env, addr);
+                    tcg_gen_brcond_i32(TCG_COND_NE, cpu_T[0],
+                                       tcg_const_i32(0), label);
+                    tmp = load_reg(s, rs);
+                    gen_st32(tmp, cpu_T[1], IS_USER(s));
+                    gen_set_label(label);
+                    gen_movl_reg_T0(s, rd);
                 }
-                gen_movl_reg_T0(s, rd);
             } else if ((insn & (1 << 6)) == 0) {
                 /* Table Branch.  */
                 if (rn == 15) {
@@ -6753,40 +6768,57 @@ static int disas_thumb2_insn(CPUState *env, DisasContext *s, uint16_t insn_hw1)
                 store_reg(s, 15, tmp);
             } else {
                 /* Load/store exclusive byte/halfword/doubleword.  */
+                /* ??? These are not really atomic.  However we know
+                   we never have multiple CPUs running in parallel,
+                   so it is good enough.  */
                 op = (insn >> 4) & 0x3;
+                /* Must use a global reg for the address because we have
+                   a conditional branch in the store instruction.  */
                 gen_movl_T1_reg(s, rn);
+                addr = cpu_T[1];
                 if (insn & (1 << 20)) {
+                    gen_helper_mark_exclusive(cpu_env, addr);
                     switch (op) {
                     case 0:
-                        gen_ldst(ldbex, s);
+                        tmp = gen_ld8u(addr, IS_USER(s));
                         break;
                     case 1:
-                        gen_ldst(ldwex, s);
+                        tmp = gen_ld16u(addr, IS_USER(s));
                         break;
                     case 3:
-                        gen_ldst(ldqex, s);
-                        gen_movl_reg_T1(s, rd);
+                        tmp = gen_ld32(addr, IS_USER(s));
+                        tcg_gen_addi_i32(addr, addr, 4);
+                        tmp2 = gen_ld32(addr, IS_USER(s));
+                        store_reg(s, rd, tmp2);
                         break;
                     default:
                         goto illegal_op;
                     }
-                    gen_movl_reg_T0(s, rs);
+                    store_reg(s, rs, tmp);
                 } else {
-                    gen_movl_T0_reg(s, rs);
+                    int label = gen_new_label();
+                    /* Must use a global that is not killed by the branch.  */
+                    gen_helper_test_exclusive(cpu_T[0], cpu_env, addr);
+                    tcg_gen_brcond_i32(TCG_COND_NE, cpu_T[0], tcg_const_i32(0),
+                                       label);
+                    tmp = load_reg(s, rs);
                     switch (op) {
                     case 0:
-                        gen_ldst(stbex, s);
+                        gen_st8(tmp, addr, IS_USER(s));
                         break;
                     case 1:
-                        gen_ldst(stwex, s);
+                        gen_st16(tmp, addr, IS_USER(s));
                         break;
                     case 3:
-                        gen_movl_T2_reg(s, rd);
-                        gen_ldst(stqex, s);
+                        gen_st32(tmp, addr, IS_USER(s));
+                        tcg_gen_addi_i32(addr, addr, 4);
+                        tmp = load_reg(s, rd);
+                        gen_st32(tmp, addr, IS_USER(s));
                         break;
                     default:
                         goto illegal_op;
                     }
+                    gen_set_label(label);
                     gen_movl_reg_T0(s, rm);
                 }
             }
@@ -7271,7 +7303,7 @@ static int disas_thumb2_insn(CPUState *env, DisasContext *s, uint16_t insn_hw1)
                         op = (insn >> 4) & 0xf;
                         switch (op) {
                         case 2: /* clrex */
-                            gen_op_clrex();
+                            gen_helper_clrex(cpu_env);
                             break;
                         case 4: /* dsb */
                         case 5: /* dmb */
@@ -7369,8 +7401,7 @@ static int disas_thumb2_insn(CPUState *env, DisasContext *s, uint16_t insn_hw1)
                         imm = imm + 1 - shift;
                         if (imm != 32) {
                             tmp2 = load_reg(s, rd);
-                            gen_bfi(tmp, tmp2, tmp,
-                                    shift, ((1u << imm) - 1) << shift);
+                            gen_bfi(tmp, tmp2, tmp, shift, (1u << imm) - 1);
                             dead_tmp(tmp2);
                         }
                         break;