diff --git a/Makefile b/Makefile
index 066f30cbe5..a2498001a1 100644
--- a/Makefile
+++ b/Makefile
@@ -84,6 +84,7 @@ dis-asm.h    gen-i386.h  op-i386.h         syscall.c\
 dis-buf.c    i386-dis.c  opreg_template.h  syscall_defs.h\
 i386.ld ppc.ld\
 tests/test-i386.c tests/test-i386-shift.h tests/test-i386.h\
+tests/test-i386-muldiv.h\
 tests/test2.c tests/hello.c tests/sha1.c tests/test1.c
 
 FILE=gemu-$(VERSION)
diff --git a/cpu-i386.h b/cpu-i386.h
index e528f61981..c245466630 100644
--- a/cpu-i386.h
+++ b/cpu-i386.h
@@ -149,9 +149,8 @@ typedef struct CPUX86State {
     uint32_t segs[6];
 
     /* emulator internal variables */
-    
     CPU86_LDouble ft0;
-
+    
     /* exception handling */
     jmp_buf jmp_env;
     int exception_index;
diff --git a/ops_template.h b/ops_template.h
index e7317eae6a..ce92db097b 100644
--- a/ops_template.h
+++ b/ops_template.h
@@ -175,12 +175,13 @@ static int glue(compute_all_dec, SUFFIX)(void)
 static int glue(compute_all_shl, SUFFIX)(void)
 {
     int cf, pf, af, zf, sf, of;
-    cf = CC_SRC & 1;
+    cf = (CC_SRC >> (DATA_BITS - 1)) & CC_C;
     pf = parity_table[(uint8_t)CC_DST];
     af = 0; /* undefined */
     zf = ((DATA_TYPE)CC_DST == 0) << 6;
     sf = lshift(CC_DST, 8 - DATA_BITS) & 0x80;
-    of = lshift(CC_SRC, 12 - DATA_BITS) & CC_O; /* only meaniful for shr with count == 1 */
+    /* of is defined if shift count == 1 */
+    of = lshift(CC_SRC ^ CC_DST, 12 - DATA_BITS) & CC_O;
     return cf | pf | af | zf | sf | of;
 }
 
@@ -199,7 +200,8 @@ static int glue(compute_all_sar, SUFFIX)(void)
     af = 0; /* undefined */
     zf = ((DATA_TYPE)CC_DST == 0) << 6;
     sf = lshift(CC_DST, 8 - DATA_BITS) & 0x80;
-    of = 0; /* only meaniful for shr with count == 1 */
+    /* of is defined if shift count == 1 */
+    of = lshift(CC_SRC ^ CC_DST, 12 - DATA_BITS) & CC_O; 
     return cf | pf | af | zf | sf | of;
 }
 
@@ -415,13 +417,8 @@ void OPPROTO glue(glue(op_shl, SUFFIX), _T0_T1_cc)(void)
 {
     int count;
     count = T1 & 0x1f;
-    if (count == 1) {
-        CC_SRC = T0;
-        T0 = T0 << 1;
-        CC_DST = T0;
-        CC_OP = CC_OP_ADDB + SHIFT;
-    } else if (count) {
-        CC_SRC = (DATA_TYPE)T0 >> (DATA_BITS - count);
+    if (count) {
+        CC_SRC = (DATA_TYPE)T0 << (count - 1);
         T0 = T0 << count;
         CC_DST = T0;
         CC_OP = CC_OP_SHLB + SHIFT;
@@ -438,7 +435,7 @@ void OPPROTO glue(glue(op_shr, SUFFIX), _T0_T1_cc)(void)
         CC_SRC = T0 >> (count - 1);
         T0 = T0 >> count;
         CC_DST = T0;
-        CC_OP = CC_OP_SHLB + SHIFT;
+        CC_OP = CC_OP_SARB + SHIFT;
     }
     FORCE_RET();
 }
@@ -449,7 +446,7 @@ void OPPROTO glue(glue(op_sar, SUFFIX), _T0_T1_cc)(void)
     count = T1 & 0x1f;
     if (count) {
         src = (DATA_STYPE)T0;
-        CC_SRC =  src >> (count - 1);
+        CC_SRC = src >> (count - 1);
         T0 = src >> count;
         CC_DST = T0;
         CC_OP = CC_OP_SARB + SHIFT;
@@ -457,6 +454,129 @@ void OPPROTO glue(glue(op_sar, SUFFIX), _T0_T1_cc)(void)
     FORCE_RET();
 }
 
+#if DATA_BITS == 16
+/* XXX: overflow flag might be incorrect in some cases in shldw */
+void OPPROTO glue(glue(op_shld, SUFFIX), _T0_T1_im_cc)(void)
+{
+    int count;
+    unsigned int res;
+    count = PARAM1;
+    T1 &= 0xffff;
+    res = T1 | (T0 << 16);
+    CC_SRC = res >> (32 - count);
+    res <<= count;
+    if (count > 16)
+        res |= T1 << (count - 16);
+    T0 = res >> 16;
+    CC_DST = T0;
+}
+
+void OPPROTO glue(glue(op_shld, SUFFIX), _T0_T1_ECX_cc)(void)
+{
+    int count;
+    unsigned int res;
+    count = ECX & 0x1f;
+    if (count) {
+        T1 &= 0xffff;
+        res = T1 | (T0 << 16);
+        CC_SRC = res >> (32 - count);
+        res <<= count;
+        if (count > 16)
+          res |= T1 << (count - 16);
+        T0 = res >> 16;
+        CC_DST = T0;
+        CC_OP = CC_OP_SARB + SHIFT;
+    }
+}
+
+void OPPROTO glue(glue(op_shrd, SUFFIX), _T0_T1_im_cc)(void)
+{
+    int count;
+    unsigned int res;
+
+    count = PARAM1;
+    res = (T0 & 0xffff) | (T1 << 16);
+    CC_SRC = res >> (count - 1);
+    res >>= count;
+    if (count > 16)
+        res |= T1 << (32 - count);
+    T0 = res;
+    CC_DST = T0;
+}
+
+
+void OPPROTO glue(glue(op_shrd, SUFFIX), _T0_T1_ECX_cc)(void)
+{
+    int count;
+    unsigned int res;
+
+    count = ECX & 0x1f;
+    if (count) {
+        res = (T0 & 0xffff) | (T1 << 16);
+        CC_SRC = res >> (count - 1);
+        res >>= count;
+        if (count > 16)
+            res |= T1 << (32 - count);
+        T0 = res;
+        CC_DST = T0;
+        CC_OP = CC_OP_SARB + SHIFT;
+    }
+}
+#endif
+
+#if DATA_BITS == 32
+void OPPROTO glue(glue(op_shld, SUFFIX), _T0_T1_im_cc)(void)
+{
+    int count;
+    count = PARAM1;
+    T0 &= DATA_MASK;
+    T1 &= DATA_MASK;
+    CC_SRC = T0 << (count - 1);
+    T0 = (T0 << count) | (T1 >> (DATA_BITS - count));
+    CC_DST = T0;
+}
+
+void OPPROTO glue(glue(op_shld, SUFFIX), _T0_T1_ECX_cc)(void)
+{
+    int count;
+    count = ECX & 0x1f;
+    if (count) {
+        T0 &= DATA_MASK;
+        T1 &= DATA_MASK;
+        CC_SRC = T0 << (count - 1);
+        T0 = (T0 << count) | (T1 >> (DATA_BITS - count));
+        CC_DST = T0;
+        CC_OP = CC_OP_SHLB + SHIFT;
+    }
+}
+
+void OPPROTO glue(glue(op_shrd, SUFFIX), _T0_T1_im_cc)(void)
+{
+    int count;
+    count = PARAM1;
+    T0 &= DATA_MASK;
+    T1 &= DATA_MASK;
+    CC_SRC = T0 >> (count - 1);
+    T0 = (T0 >> count) | (T1 << (DATA_BITS - count));
+    CC_DST = T0;
+}
+
+
+void OPPROTO glue(glue(op_shrd, SUFFIX), _T0_T1_ECX_cc)(void)
+{
+    int count;
+    count = ECX & 0x1f;
+    if (count) {
+        T0 &= DATA_MASK;
+        T1 &= DATA_MASK;
+        CC_SRC = T0 >> (count - 1);
+        T0 = (T0 >> count) | (T1 << (DATA_BITS - count));
+        CC_DST = T0;
+        CC_OP = CC_OP_SARB + SHIFT;
+    }
+}
+#endif
+
 /* carry add/sub (we only need to set CC_OP differently) */
 
 void OPPROTO glue(glue(op_adc, SUFFIX), _T0_T1_cc)(void)
diff --git a/tests/Makefile b/tests/Makefile
index 489e6b557f..30022b19aa 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -19,7 +19,7 @@ test2: test2.c
 	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $<
 
 # i386 emulation test (dump various opcodes) */
-test-i386: test-i386.c test-i386.h test-i386-shift.h
+test-i386: test-i386.c test-i386.h test-i386-shift.h test-i386-muldiv.h
 	$(CC) $(CFLAGS) $(LDFLAGS) -static -o $@ $<
 
 test: test-i386
diff --git a/tests/test-i386-muldiv.h b/tests/test-i386-muldiv.h
new file mode 100644
index 0000000000..5dba315c24
--- /dev/null
+++ b/tests/test-i386-muldiv.h
@@ -0,0 +1,56 @@
+
+void glue(glue(test_, OP), b)(int op0, int op1) 
+{
+    int res, s1, s0, flags;
+    s0 = op0;
+    s1 = op1;
+    res = s0;
+    flags = 0;
+    asm ("push %4\n\t"
+         "popf\n\t"
+         stringify(OP)"b %b2\n\t" 
+         "pushf\n\t"
+         "popl %1\n\t"
+         : "=a" (res), "=g" (flags)
+         : "q" (s1), "0" (res), "1" (flags));
+    printf("%-10s A=%08x B=%08x R=%08x CC=%04x\n",
+           stringify(OP) "b", s0, s1, res, flags & CC_MASK);
+}
+
+void glue(glue(test_, OP), w)(int op0h, int op0, int op1) 
+{
+    int res, s1, flags, resh;
+    s1 = op1;
+    resh = op0h;
+    res = op0;
+    flags = 0;
+    asm ("push %5\n\t"
+         "popf\n\t"
+         stringify(OP) "w %w3\n\t" 
+         "pushf\n\t"
+         "popl %1\n\t"
+         : "=a" (res), "=g" (flags), "=d" (resh)
+         : "q" (s1), "0" (res), "1" (flags), "2" (resh));
+    printf("%-10s AH=%08x AL=%08x B=%08x RH=%08x RL=%08x CC=%04x\n",
+           stringify(OP) "w", op0h, op0, s1, resh, res, flags & CC_MASK);
+}
+
+void glue(glue(test_, OP), l)(int op0h, int op0, int op1) 
+{
+    int res, s1, flags, resh;
+    s1 = op1;
+    resh = op0h;
+    res = op0;
+    flags = 0;
+    asm ("push %5\n\t"
+         "popf\n\t"
+         stringify(OP) "l %3\n\t" 
+         "pushf\n\t"
+         "popl %1\n\t"
+         : "=a" (res), "=g" (flags), "=d" (resh)
+         : "q" (s1), "0" (res), "1" (flags), "2" (resh));
+    printf("%-10s AH=%08x AL=%08x B=%08x RH=%08x RL=%08x CC=%04x\n",
+           stringify(OP) "l", op0h, op0, s1, resh, res, flags & CC_MASK);
+}
+
+#undef OP
diff --git a/tests/test-i386-shift.h b/tests/test-i386-shift.h
index af892f6c7d..f3795d9f36 100644
--- a/tests/test-i386-shift.h
+++ b/tests/test-i386-shift.h
@@ -4,7 +4,19 @@
 #define exec_opw glue(glue(exec_, OP), w)
 #define exec_opb glue(glue(exec_, OP), b)
 
-#define EXECSHIFT(size, res, s1, flags) \
+#ifndef OP_SHIFTD
+
+#ifdef OP_NOBYTE
+#define EXECSHIFT(size, res, s1, s2, flags) \
+    asm ("push %4\n\t"\
+         "popf\n\t"\
+         stringify(OP) size " %" size "2, %" size "0\n\t" \
+         "pushf\n\t"\
+         "popl %1\n\t"\
+         : "=g" (res), "=g" (flags)\
+         : "r" (s1), "0" (res), "1" (flags));
+#else
+#define EXECSHIFT(size, res, s1, s2, flags) \
     asm ("push %4\n\t"\
          "popf\n\t"\
          stringify(OP) size " %%cl, %" size "0\n\t" \
@@ -12,13 +24,14 @@
          "popl %1\n\t"\
          : "=q" (res), "=g" (flags)\
          : "c" (s1), "0" (res), "1" (flags));
+#endif
 
-void exec_opl(int s0, int s1, int iflags)
+void exec_opl(int s2, int s0, int s1, int iflags)
 {
     int res, flags;
     res = s0;
     flags = iflags;
-    EXECSHIFT("", res, s1, flags);
+    EXECSHIFT("", res, s1, s2, flags);
     /* overflow is undefined if count != 1 */
     if (s1 != 1)
       flags &= ~CC_O;
@@ -26,12 +39,12 @@ void exec_opl(int s0, int s1, int iflags)
            stringify(OP) "l", s0, s1, res, iflags, flags & CC_MASK);
 }
 
-void exec_opw(int s0, int s1, int iflags)
+void exec_opw(int s2, int s0, int s1, int iflags)
 {
     int res, flags;
     res = s0;
     flags = iflags;
-    EXECSHIFT("w", res, s1, flags);
+    EXECSHIFT("w", res, s1, s2, flags);
     /* overflow is undefined if count != 1 */
     if (s1 != 1)
       flags &= ~CC_O;
@@ -39,27 +52,69 @@ void exec_opw(int s0, int s1, int iflags)
            stringify(OP) "w", s0, s1, res, iflags, flags & CC_MASK);
 }
 
+#else
+#define EXECSHIFT(size, res, s1, s2, flags) \
+    asm ("push %4\n\t"\
+         "popf\n\t"\
+         stringify(OP) size " %%cl, %" size "5, %" size "0\n\t" \
+         "pushf\n\t"\
+         "popl %1\n\t"\
+         : "=g" (res), "=g" (flags)\
+         : "c" (s1), "0" (res), "1" (flags), "r" (s2));
+
+void exec_opl(int s2, int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECSHIFT("", res, s1, s2, flags);
+    /* overflow is undefined if count != 1 */
+    if (s1 != 1)
+      flags &= ~CC_O;
+    printf("%-10s A=%08x B=%08x C=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "l", s0, s2, s1, res, iflags, flags & CC_MASK);
+}
+
+void exec_opw(int s2, int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECSHIFT("w", res, s1, s2, flags);
+    /* overflow is undefined if count != 1 */
+    if (s1 != 1)
+      flags &= ~CC_O;
+    printf("%-10s A=%08x B=%08x C=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "w", s0, s2, s1, res, iflags, flags & CC_MASK);
+}
+
+#endif
+
+#ifndef OP_NOBYTE
 void exec_opb(int s0, int s1, int iflags)
 {
     int res, flags;
     res = s0;
     flags = iflags;
-    EXECSHIFT("b", res, s1, flags);
+    EXECSHIFT("b", res, s1, 0, flags);
     /* overflow is undefined if count != 1 */
     if (s1 != 1)
       flags &= ~CC_O;
     printf("%-10s A=%08x B=%08x R=%08x CCIN=%04x CC=%04x\n",
            stringify(OP) "b", s0, s1, res, iflags, flags & CC_MASK);
 }
+#endif
 
-void exec_op(int s0, int s1)
+void exec_op(int s2, int s0, int s1)
 {
-    exec_opl(s0, s1, 0);
-    exec_opw(s0, s1, 0);
+    exec_opl(s2, s0, s1, 0);
+    exec_opw(s2, s0, s1, 0);
+#ifndef OP_NOBYTE
     exec_opb(s0, s1, 0);
+#endif
 #ifdef OP_CC
-    exec_opl(s0, s1, CC_C);
-    exec_opw(s0, s1, CC_C);
+    exec_opl(s2, s0, s1, CC_C);
+    exec_opw(s2, s0, s1, CC_C);
     exec_opb(s0, s1, CC_C);
 #endif
 }
@@ -68,12 +123,16 @@ void glue(test_, OP)(void)
 {
     int i;
     for(i = 0; i < 32; i++)
-        exec_op(0x12345678, i);
+        exec_op(0x21ad3d34, 0x12345678, i);
     for(i = 0; i < 32; i++)
-        exec_op(0x82345678, i);
+        exec_op(0x813f3421, 0x82345678, i);
 }
 
 void *glue(_test_, OP) __init_call = glue(test_, OP);
 
 #undef OP
 #undef OP_CC
+#undef OP_SHIFTD
+#undef OP_NOBYTE
+#undef EXECSHIFT
+
diff --git a/tests/test-i386.c b/tests/test-i386.c
index 55dd9eb2c7..b3438ebbf9 100644
--- a/tests/test-i386.c
+++ b/tests/test-i386.c
@@ -92,6 +92,35 @@ static void *call_start __init_call = NULL;
 #define OP_CC
 #include "test-i386-shift.h"
 
+#define OP shld
+#define OP_SHIFTD
+#define OP_NOBYTE
+#include "test-i386-shift.h"
+
+#define OP shrd
+#define OP_SHIFTD
+#define OP_NOBYTE
+#include "test-i386-shift.h"
+
+/* XXX: should be more precise ? */
+#undef CC_MASK
+#define CC_MASK (CC_C)
+
+#define OP bt
+#define OP_NOBYTE
+#include "test-i386-shift.h"
+
+#define OP bts
+#define OP_NOBYTE
+#include "test-i386-shift.h"
+
+#define OP btr
+#define OP_NOBYTE
+#include "test-i386-shift.h"
+
+#define OP btc
+#define OP_NOBYTE
+#include "test-i386-shift.h"
 
 /* lea test (modrm support) */
 #define TEST_LEA(STR)\
@@ -403,15 +432,13 @@ int main(int argc, char **argv)
     void **ptr;
     void (*func)(void);
 
-    test_mul();
-#if 0
     ptr = &call_start + 1;
     while (*ptr != NULL) {
         func = *ptr++;
         func();
     }
+    test_mul();
     test_jcc();
     test_lea();
-#endif
     return 0;
 }
diff --git a/tests/test-i386.h b/tests/test-i386.h
new file mode 100644
index 0000000000..7d1812c883
--- /dev/null
+++ b/tests/test-i386.h
@@ -0,0 +1,131 @@
+
+#define exec_op glue(exec_, OP)
+#define exec_opl glue(glue(exec_, OP), l)
+#define exec_opw glue(glue(exec_, OP), w)
+#define exec_opb glue(glue(exec_, OP), b)
+
+#define EXECOP2(size, res, s1, flags) \
+    asm ("push %4\n\t"\
+         "popf\n\t"\
+         stringify(OP) size " %" size "2, %" size "0\n\t" \
+         "pushf\n\t"\
+         "popl %1\n\t"\
+         : "=q" (res), "=g" (flags)\
+         : "q" (s1), "0" (res), "1" (flags));
+
+#define EXECOP1(size, res, flags) \
+    asm ("push %3\n\t"\
+         "popf\n\t"\
+         stringify(OP) size " %" size "0\n\t" \
+         "pushf\n\t"\
+         "popl %1\n\t"\
+         : "=q" (res), "=g" (flags)\
+         : "0" (res), "1" (flags));
+
+#ifdef OP1
+void exec_opl(int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECOP1("", res, flags);
+    printf("%-10s A=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "l", s0, res, iflags, flags & CC_MASK);
+}
+
+void exec_opw(int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECOP1("w", res, flags);
+    printf("%-10s A=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "w", s0, res, iflags, flags & CC_MASK);
+}
+
+void exec_opb(int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECOP1("b", res, flags);
+    printf("%-10s A=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "b", s0, res, iflags, flags & CC_MASK);
+}
+#else
+void exec_opl(int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECOP2("", res, s1, flags);
+    printf("%-10s A=%08x B=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "l", s0, s1, res, iflags, flags & CC_MASK);
+}
+
+void exec_opw(int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECOP2("w", res, s1, flags);
+    printf("%-10s A=%08x B=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "w", s0, s1, res, iflags, flags & CC_MASK);
+}
+
+void exec_opb(int s0, int s1, int iflags)
+{
+    int res, flags;
+    res = s0;
+    flags = iflags;
+    EXECOP2("b", res, s1, flags);
+    printf("%-10s A=%08x B=%08x R=%08x CCIN=%04x CC=%04x\n",
+           stringify(OP) "b", s0, s1, res, iflags, flags & CC_MASK);
+}
+#endif
+
+void exec_op(int s0, int s1)
+{
+    exec_opl(s0, s1, 0);
+    exec_opw(s0, s1, 0);
+    exec_opb(s0, s1, 0);
+#ifdef OP_CC
+    exec_opl(s0, s1, CC_C);
+    exec_opw(s0, s1, CC_C);
+    exec_opb(s0, s1, CC_C);
+#endif
+}
+
+void glue(test_, OP)(void)
+{
+    exec_op(0x12345678, 0x812FADA);
+    exec_op(0x12341, 0x12341);
+    exec_op(0x12341, -0x12341);
+    exec_op(0xffffffff, 0);
+    exec_op(0xffffffff, -1);
+    exec_op(0xffffffff, 1);
+    exec_op(0xffffffff, 2);
+    exec_op(0x7fffffff, 0);
+    exec_op(0x7fffffff, 1);
+    exec_op(0x7fffffff, -1);
+    exec_op(0x80000000, -1);
+    exec_op(0x80000000, 1);
+    exec_op(0x80000000, -2);
+    exec_op(0x12347fff, 0);
+    exec_op(0x12347fff, 1);
+    exec_op(0x12347fff, -1);
+    exec_op(0x12348000, -1);
+    exec_op(0x12348000, 1);
+    exec_op(0x12348000, -2);
+    exec_op(0x12347f7f, 0);
+    exec_op(0x12347f7f, 1);
+    exec_op(0x12347f7f, -1);
+    exec_op(0x12348080, -1);
+    exec_op(0x12348080, 1);
+    exec_op(0x12348080, -2);
+}
+
+void *glue(_test_, OP) __init_call = glue(test_, OP);
+
+#undef OP
+#undef OP_CC
diff --git a/translate-i386.c b/translate-i386.c
index f145a54068..69c769c198 100644
--- a/translate-i386.c
+++ b/translate-i386.c
@@ -394,6 +394,28 @@ static GenOpFunc *gen_op_shift_T0_T1_cc[3][8] = {
     },
 };
 
+static GenOpFunc1 *gen_op_shiftd_T0_T1_im_cc[2][2] = {
+    [0] = {
+        gen_op_shldw_T0_T1_im_cc,
+        gen_op_shrdw_T0_T1_im_cc,
+    },
+    [1] = {
+        gen_op_shldl_T0_T1_im_cc,
+        gen_op_shrdl_T0_T1_im_cc,
+    },
+};
+
+static GenOpFunc *gen_op_shiftd_T0_T1_ECX_cc[2][2] = {
+    [0] = {
+        gen_op_shldw_T0_T1_ECX_cc,
+        gen_op_shrdw_T0_T1_ECX_cc,
+    },
+    [1] = {
+        gen_op_shldl_T0_T1_ECX_cc,
+        gen_op_shrdl_T0_T1_ECX_cc,
+    },
+};
+
 static GenOpFunc *gen_op_btx_T0_T1_cc[2][4] = {
     [0] = {
         gen_op_btw_T0_T1_cc,
@@ -1689,6 +1711,59 @@ long disas_insn(DisasContext *s, uint8_t *pc_start, int *is_jmp_ptr)
         shift = 0;
         goto grp2;
 
+    case 0x1a4: /* shld imm */
+        op = 0;
+        shift = 1;
+        goto do_shiftd;
+    case 0x1a5: /* shld cl */
+        op = 0;
+        shift = 0;
+        goto do_shiftd;
+    case 0x1ac: /* shrd imm */
+        op = 1;
+        shift = 1;
+        goto do_shiftd;
+    case 0x1ad: /* shrd cl */
+        op = 1;
+        shift = 0;
+    do_shiftd:
+        ot = dflag ? OT_LONG : OT_WORD;
+        modrm = ldub(s->pc++);
+        mod = (modrm >> 6) & 3;
+        rm = modrm & 7;
+        reg = (modrm >> 3) & 7;
+        
+        if (mod != 3) {
+            gen_lea_modrm(s, modrm, &reg_addr, &offset_addr);
+            gen_op_ld_T0_A0[ot]();
+        } else {
+            gen_op_mov_TN_reg[ot][0][rm]();
+        }
+        gen_op_mov_TN_reg[ot][1][reg]();
+        
+        if (shift) {
+            val = ldub(s->pc++);
+            val &= 0x1f;
+            if (val) {
+                gen_op_shiftd_T0_T1_im_cc[ot - OT_WORD][op](val);
+                if (op == 0 && ot != OT_WORD)
+                    s->cc_op = CC_OP_SHLB + ot;
+                else
+                    s->cc_op = CC_OP_SARB + ot;
+            }
+        } else {
+            if (s->cc_op != CC_OP_DYNAMIC)
+                gen_op_set_cc_op(s->cc_op);
+            gen_op_shiftd_T0_T1_ECX_cc[ot - OT_WORD][op]();
+            s->cc_op = CC_OP_DYNAMIC; /* cannot predict flags after */
+        }
+        if (mod != 3) {
+            gen_op_st_T0_A0[ot]();
+        } else {
+            gen_op_mov_reg_T0[ot][rm]();
+        }
+        break;
+
         /************************/
         /* floats */
     case 0xd8 ... 0xdf: 
@@ -2002,6 +2077,7 @@ long disas_insn(DisasContext *s, uint8_t *pc_start, int *is_jmp_ptr)
                     break;
 #endif
                 default:
+                    error("unhandled FP df/4\n");
                     return -1;
                 }
                 break;
@@ -2291,7 +2367,7 @@ long disas_insn(DisasContext *s, uint8_t *pc_start, int *is_jmp_ptr)
             return -1;
         op -= 4;
         gen_op_btx_T0_T1_cc[ot - OT_WORD][op]();
-        s->cc_op = CC_OP_SHLB + ot;
+        s->cc_op = CC_OP_SARB + ot;
         if (op != 0) {
             if (mod != 3)
                 gen_op_st_T0_A0[ot]();
@@ -2329,7 +2405,7 @@ long disas_insn(DisasContext *s, uint8_t *pc_start, int *is_jmp_ptr)
             gen_op_mov_TN_reg[ot][0][rm]();
         }
         gen_op_btx_T0_T1_cc[ot - OT_WORD][op]();
-        s->cc_op = CC_OP_SHLB + ot;
+        s->cc_op = CC_OP_SARB + ot;
         if (op != 0) {
             if (mod != 3)
                 gen_op_st_T0_A0[ot]();
@@ -2417,7 +2493,8 @@ int cpu_x86_gen_code(uint8_t *gen_code_buf, int *gen_code_size_ptr,
     is_jmp = 0;
     ret = disas_insn(dc, pc_start, &is_jmp);
     if (ret == -1) 
-        error("unknown instruction at PC=0x%x", pc_start);
+        error("unknown instruction at PC=0x%x B=%02x %02x", 
+              pc_start, pc_start[0], pc_start[1]);
     /* we must store the eflags state if it is not already done */
     if (dc->cc_op != CC_OP_DYNAMIC)
         gen_op_set_cc_op(dc->cc_op);