x86_64 fixes

git-svn-id: svn://svn.savannah.nongnu.org/qemu/trunk@1324 c046a42c-6fe2-441c-8c8c-71466251a162

4 files changed, 33 insertions, 18 deletions

diff --git a/target-i386/helper.c b/target-i386/helper.c
index b727e93d66..53b85d3aaa 100644
--- a/target-i386/helper.c
+++ b/target-i386/helper.c
@@ -1012,7 +1012,8 @@ void helper_sysret(int dflag)
                                DESC_G_MASK | DESC_B_MASK | DESC_P_MASK |
                                DESC_S_MASK | (3 << DESC_DPL_SHIFT) |
                                DESC_W_MASK | DESC_A_MASK);
-        load_eflags((uint32_t)(env->regs[11]), 0xffffffff);
+        load_eflags((uint32_t)(env->regs[11]), TF_MASK | AC_MASK | ID_MASK | 
+                    IF_MASK | IOPL_MASK | VM_MASK | RF_MASK | NT_MASK);
         cpu_x86_set_cpl(env, 3);
     } else {
         cpu_x86_load_seg_cache(env, R_CS, selector | 3, 
@@ -1209,7 +1210,7 @@ void helper_divl_EAX_T0(void)
     unsigned int den, q, r;
     uint64_t num;
     
-    num = EAX | ((uint64_t)EDX << 32);
+    num = ((uint32_t)EAX) | ((uint64_t)((uint32_t)EDX) << 32);
     den = T0;
     if (den == 0) {
         raise_exception(EXCP00_DIVZ);
@@ -1229,7 +1230,7 @@ void helper_idivl_EAX_T0(void)
     int den, q, r;
     int64_t num;
     
-    num = EAX | ((uint64_t)EDX << 32);
+    num = ((uint32_t)EAX) | ((uint64_t)((uint32_t)EDX) << 32);
     den = T0;
     if (den == 0) {
         raise_exception(EXCP00_DIVZ);
@@ -3003,7 +3004,7 @@ void helper_fxrstor(target_ulong ptr, int data64)
     }
 
     if (env->cr[4] & CR4_OSFXSR_MASK) {
-        /* XXX: finish it, endianness */
+        /* XXX: finish it */
         env->mxcsr = ldl(ptr + 0x18);
         //ldl(ptr + 0x1c);
         nb_xmm_regs = 8 << data64;
@@ -3170,7 +3171,7 @@ static void div64(uint64_t *plow, uint64_t *phigh, uint64_t b)
     }
 }
 
-static void idiv64(uint64_t *plow, uint64_t *phigh, uint64_t b)
+static void idiv64(uint64_t *plow, uint64_t *phigh, int64_t b)
 {
     int sa, sb;
     sa = ((int64_t)*phigh < 0);
@@ -3182,7 +3183,7 @@ static void idiv64(uint64_t *plow, uint64_t *phigh, uint64_t b)
     div64(plow, phigh, b);
     if (sa ^ sb)
         *plow = - *plow;
-    if (sb)
+    if (sa)
         *phigh = - *phigh;
 }
 
diff --git a/target-i386/op.c b/target-i386/op.c
index fed0fca93c..dd9d7bad1f 100644
--- a/target-i386/op.c
+++ b/target-i386/op.c
@@ -286,8 +286,8 @@ void OPPROTO op_imull_EAX_T0(void)
 {
     int64_t res;
     res = (int64_t)((int32_t)EAX) * (int64_t)((int32_t)T0);
-    EAX = res;
-    EDX = res >> 32;
+    EAX = (uint32_t)(res);
+    EDX = (uint32_t)(res >> 32);
     CC_DST = res;
     CC_SRC = (res != (int32_t)res);
 }
diff --git a/target-i386/ops_template_mem.h b/target-i386/ops_template_mem.h
index 00454e6d05..a999e2d440 100644
--- a/target-i386/ops_template_mem.h
+++ b/target-i386/ops_template_mem.h
@@ -160,7 +160,7 @@ void OPPROTO glue(glue(op_rcl, MEM_SUFFIX), _T0_T1_cc)(void)
         eflags = cc_table[CC_OP].compute_all();
         T0 &= DATA_MASK;
         src = T0;
-        res = (T0 << count) | ((eflags & CC_C) << (count - 1));
+        res = (T0 << count) | ((target_ulong)(eflags & CC_C) << (count - 1));
         if (count > 1)
             res |= T0 >> (DATA_BITS + 1 - count);
         T0 = res;
@@ -191,7 +191,7 @@ void OPPROTO glue(glue(op_rcr, MEM_SUFFIX), _T0_T1_cc)(void)
         eflags = cc_table[CC_OP].compute_all();
         T0 &= DATA_MASK;
         src = T0;
-        res = (T0 >> count) | ((eflags & CC_C) << (DATA_BITS - count));
+        res = (T0 >> count) | ((target_ulong)(eflags & CC_C) << (DATA_BITS - count));
         if (count > 1)
             res |= T0 << (DATA_BITS + 1 - count);
         T0 = res;
diff --git a/target-i386/translate.c b/target-i386/translate.c
index 093723944c..ad5acc9c48 100644
--- a/target-i386/translate.c
+++ b/target-i386/translate.c
@@ -600,12 +600,14 @@ static GenOpFunc *gen_op_shift_mem_T0_T1_cc[3 * 4][8] = {
     {\
         gen_op_shldw ## SUFFIX ## _T0_T1_ ## op ## _cc,\
         gen_op_shrdw ## SUFFIX ## _T0_T1_ ## op ## _cc,\
-    },\
+     },\
     {\
         gen_op_shldl ## SUFFIX ## _T0_T1_ ## op ## _cc,\
         gen_op_shrdl ## SUFFIX ## _T0_T1_ ## op ## _cc,\
     },\
     {\
+X86_64_DEF(gen_op_shldq ## SUFFIX ## _T0_T1_ ## op ## _cc,\
+           gen_op_shrdq ## SUFFIX ## _T0_T1_ ## op ## _cc,)\
     },
 
 static GenOpFunc1 *gen_op_shiftd_T0_T1_im_cc[4][2] = {
@@ -2723,14 +2725,14 @@ static void gen_sse(DisasContext *s, int b, target_ulong pc_start, int rex_r)
             sse_op2(op2_offset, op1_offset);
             break;
         case 0x050: /* movmskps */
-            gen_op_movmskps(offsetof(CPUX86State,xmm_regs[reg]));
             rm = (modrm & 7) | REX_B(s);
-            gen_op_mov_reg_T0[OT_LONG][rm]();
+            gen_op_movmskps(offsetof(CPUX86State,xmm_regs[rm]));
+            gen_op_mov_reg_T0[OT_LONG][reg]();
             break;
         case 0x150: /* movmskpd */
-            gen_op_movmskpd(offsetof(CPUX86State,xmm_regs[reg]));
             rm = (modrm & 7) | REX_B(s);
-            gen_op_mov_reg_T0[OT_LONG][rm]();
+            gen_op_movmskpd(offsetof(CPUX86State,xmm_regs[rm]));
+            gen_op_mov_reg_T0[OT_LONG][reg]();
             break;
         case 0x02a: /* cvtpi2ps */
         case 0x12a: /* cvtpi2pd */
@@ -2795,10 +2797,22 @@ static void gen_sse(DisasContext *s, int b, target_ulong pc_start, int rex_r)
         case 0x22d: /* cvtss2si */
         case 0x32d: /* cvtsd2si */
             ot = (s->dflag == 2) ? OT_QUAD : OT_LONG;
-            op1_offset = offsetof(CPUX86State,xmm_regs[reg]);
+            if (mod != 3) {
+                gen_lea_modrm(s, modrm, &reg_addr, &offset_addr);
+                if ((b >> 8) & 1) {
+                    gen_ldq_env_A0[s->mem_index >> 2](offsetof(CPUX86State,xmm_t0.XMM_Q(0)));
+                } else {
+                    gen_op_ld_T0_A0[OT_LONG + s->mem_index]();
+                    gen_op_movl_env_T0(offsetof(CPUX86State,xmm_t0.XMM_L(0)));
+                }
+                op2_offset = offsetof(CPUX86State,xmm_t0);
+            } else {
+                rm = (modrm & 7) | REX_B(s);
+                op2_offset = offsetof(CPUX86State,xmm_regs[rm]);
+            }
             sse_op_table3[(s->dflag == 2) * 2 + ((b >> 8) - 2) + 4 + 
-                          (b & 1) * 4](op1_offset);
-            gen_ldst_modrm(s, modrm, ot, OR_TMP0, 1);
+                          (b & 1) * 4](op2_offset);
+            gen_op_mov_reg_T0[ot][reg]();
             break;
         case 0xc4: /* pinsrw */
         case 0x1c4: