diff --git a/arm64-gen.c b/arm64-gen.c
index 6e6e4784..268319af 100644
--- a/arm64-gen.c
+++ b/arm64-gen.c
@@ -296,22 +296,17 @@ static void arm64_spoff(int reg, uint64_t off)
     }
 }
 
-/* invert 0: return value to use for store/load */
-/* invert 1: return value to use for arm64_sym */
-static uint64_t arm64_check_offset(int invert, int sz_, uint64_t off)
+static uint64_t arm64_check_offset(int sz_, uint64_t off)
 {
     uint32_t sz = sz_;
     uint64_t scaled_mask = 0xffful << sz;
-
-    if (!(off & ~scaled_mask) ||
-        (off < 256 || -off <= 256))
-        return invert ? off : 0ul;
-    else if (off & scaled_mask)
-        return invert ? off & scaled_mask : off & ~scaled_mask;
-    else if (off & 0x1fful)
-        return invert ? off & 0x1fful : off & ~0x1fful;
-    else
-        return invert ? 0ul : off;
+    if (!(off & ~scaled_mask) || (off < 256 || -off <= 256))
+        return ~0ull;
+    if (off & scaled_mask)
+        return scaled_mask;
+    if (off & 0x1fful)
+        return 0x1fful;
+    return 0;
 }
 
 static void arm64_ldrx(int sg, int sz_, int dst, int bas, uint64_t off)
@@ -468,8 +463,41 @@ static void arm64_strv(int sz_, int dst, int bas, uint64_t off)
     }
 }
 
-static void arm64_add_offset(int r, unsigned long addend)
+static void arm64_sym(int r, Sym *sym, unsigned long addend)
 {
+    if (sym->type.t & VT_TLS) {
+#if TCC_TARGET_PE
+        Sym *s2 = external_global_sym(TOK___tls_index, &int_type);
+        int r2 = get_reg(RC_INT);
+        arm64_sym(30, s2, 0);
+        o(0xb94003de); // ldr w30, [x30]
+        o(0xf9402e40 | r2); // ldr r2, [x18, #88]
+        o(0x8b1e0c1e | r2 << 5); // add x30, r2, x30, lsl #3
+        o(0xf94003c0 | r); // ldr xr, [x30]
+#else
+        o(0xd53bd040 | r); /* mrs xr, tpidr_el0 */
+#endif
+        greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_HI12, 0);
+        /* add xr, xr, #0, lsl #12 */
+        o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) | ARM64_RN(r) | ARM64_RD(r));
+        greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_LO12, 0);
+        /* add xr, xr, #0 */
+        o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | ARM64_RD(r));
+        goto add_addend;
+    }
+#ifdef TCC_TARGET_PE
+    /* PE links symbol addresses directly; there is no ELF-style GOT here. */
+    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_PREL_PG_HI21, 0);
+    o(ARM64_ADRP | r);            // adrp xr, #sym
+    greloca(cur_text_section, sym, ind, R_AARCH64_ADD_ABS_LO12_NC, 0);
+    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | r); // add xr, xr, #sym
+#else
+    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_GOT_PAGE, 0);
+    o(ARM64_ADRP | r);            // adrp xr, #sym
+    greloca(cur_text_section, sym, ind, R_AARCH64_LD64_GOT_LO12_NC, 0);
+    o(ARM64_LDR_X | ARM64_RN(r) | r); // ld xr,[xr, #sym]
+#endif
+add_addend:
     if (addend) {
         // add xr, xr, #addend
 	if (addend & 0xffful)
@@ -484,7 +512,7 @@ static void arm64_add_offset(int r, unsigned long addend)
 		/* very unlikely */
 		int t = r ? 0 : 1;
 		o(ARM64_STR_X_PRE | 0x001F0FE0U | t); /* str xt, [sp, #-16]! */
-		arm64_movimm(t, addend & ~0xfffffful); // use xt for addent
+		arm64_movimm(t, addend & ~0xffffffull); // use xt for addent
 		o(ARM64_ADD_REG | ARM64_SF(1) | ARM64_RM(t) | ARM64_RN(r) | r); /* add xr, xr, xt */
 		o(ARM64_LDR_X_POST | 0x000107E0U | t); /* ldr xt, [sp], #16 */
 	    }
@@ -492,111 +520,60 @@ static void arm64_add_offset(int r, unsigned long addend)
     }
 }
 
-static void arm64_sym(int r, Sym *sym, unsigned long addend)
-{
-#ifdef TCC_TARGET_PE
-    /* PE links symbol addresses directly; there is no ELF-style GOT here. */
-    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_PREL_PG_HI21, 0);
-    o(ARM64_ADRP | r);            // adrp xr, #sym
-    greloca(cur_text_section, sym, ind, R_AARCH64_ADD_ABS_LO12_NC, 0);
-    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | r); // add xr, xr, #sym
-#else
-    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_GOT_PAGE, 0);
-    o(ARM64_ADRP | r);            // adrp xr, #sym
-    greloca(cur_text_section, sym, ind, R_AARCH64_LD64_GOT_LO12_NC, 0);
-    o(ARM64_LDR_X | ARM64_RN(r) | r); // ld xr,[xr, #sym]
-#endif
-    arm64_add_offset(r, addend);
-}
-
 static void arm64_load_cmp(int r, SValue *sv);
 
-static void arm64_tls_sym(int r, Sym *sym, unsigned long addend)
-{
-    o(0xd53bd040 | r); /* mrs xr, tpidr_el0 */
-    greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_HI12, 0);
-    /* add xr, xr, #0, lsl #12 */
-    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) | ARM64_RN(r) | ARM64_RD(r));
-    greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_LO12, 0);
-    /* add xr, xr, #0 */
-    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | ARM64_RD(r));
-    arm64_add_offset(r, addend);
-}
-
 ST_FUNC void load(int r, SValue *sv)
 {
     int svtt = sv->type.t;
-    int svr = sv->r & ~(VT_BOUNDED | VT_NONCONST);
+    int svr = sv->r & (VT_VALMASK | VT_LVAL | VT_SYM);
     int svrv = svr & VT_VALMASK;
     uint64_t svcul = sv->c.i;
-    uint64_t svcoff = (uint64_t)(int64_t)(int32_t)sv->c.i;
+    uint64_t svcoff = (int32_t)sv->c.i;
+    int sb = !(svtt & VT_UNSIGNED);
+    int sz = arm64_type_size(svtt);
 
     if (svr == (VT_LOCAL | VT_LVAL)) {
         if (IS_FREG(r))
-            arm64_ldrv(arm64_type_size(svtt), fltr(r), 29, svcoff);
+            arm64_ldrv(sz, fltr(r), 29, svcoff);
         else
-            arm64_ldrx(!(svtt & VT_UNSIGNED), arm64_type_size(svtt),
-                       intr(r), 29, svcoff);
+            arm64_ldrx(sb, sz, intr(r), 29, svcoff);
         return;
     }
 
     if (svr == (VT_CONST | VT_LVAL)) {
-	uint64_t i = sv->c.i;
-
-	if (sv->sym && (sv->sym->type.t & VT_TLS))
-            arm64_tls_sym(30, sv->sym, 0);
-        else if (sv->sym)
-            arm64_sym(30, sv->sym, // use x30 for address
-	              arm64_check_offset(0, arm64_type_size(svtt), i));
-	else
-	    arm64_movimm (30, i), i = 0;
+	arm64_movimm (30, svcul);
         if (IS_FREG(r))
-            arm64_ldrv(arm64_type_size(svtt), fltr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), i));
+            arm64_ldrv(sz, fltr(r), 30, 0);
         else
-            arm64_ldrx(!(svtt&VT_UNSIGNED), arm64_type_size(svtt), intr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), i));
+            arm64_ldrx(sb, sz, intr(r), 30, 0);
         return;
     }
 
-    if ((svr & ~VT_VALMASK) == VT_LVAL && svrv < VT_CONST) {
-        if ((svtt & VT_BTYPE) != VT_VOID) {
-            if (IS_FREG(r))
-                arm64_ldrv(arm64_type_size(svtt), fltr(r), intr(svrv), 0);
-            else
-                arm64_ldrx(!(svtt & VT_UNSIGNED), arm64_type_size(svtt),
-                           intr(r), intr(svrv), 0);
-        }
+    if (svrv < VT_CONST && (svr & VT_LVAL)) {
+        if (IS_FREG(r))
+            arm64_ldrv(sz, fltr(r), intr(svrv), 0);
+        else
+            arm64_ldrx(sb, sz, intr(r), intr(svrv), 0);
         return;
     }
 
     if (svr == (VT_CONST | VT_LVAL | VT_SYM)) {
-        if (sv->sym->type.t & VT_TLS)
-            arm64_tls_sym(30, sv->sym, 0);
-        else
-        arm64_sym(30, sv->sym, // use x30 for address
-		  arm64_check_offset(0, arm64_type_size(svtt), svcoff));
+        uint64_t mask = arm64_check_offset(sz, svcoff);
+        arm64_sym(30, sv->sym, svcoff & ~mask);
         if (IS_FREG(r))
-            arm64_ldrv(arm64_type_size(svtt), fltr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), svcoff));
+            arm64_ldrv(sz, fltr(r), 30, svcoff & mask);
         else
-            arm64_ldrx(!(svtt&VT_UNSIGNED), arm64_type_size(svtt), intr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), svcoff));
+            arm64_ldrx(sb, sz, intr(r), 30, svcoff & mask);
         return;
     }
 
     if (svr == (VT_CONST | VT_SYM)) {
-        if (sv->sym->type.t & VT_TLS)
-            arm64_tls_sym(intr(r), sv->sym, svcul);
-        else
-            arm64_sym(intr(r), sv->sym, svcul);
+        arm64_sym(intr(r), sv->sym, svcul);
         return;
     }
 
     if (svr == VT_CONST) {
-        if ((svtt & VT_BTYPE) != VT_VOID)
-            arm64_movimm(intr(r), arm64_type_size(svtt) == 3 ?
-                         sv->c.i : (uint32_t)svcul);
+        arm64_movimm(intr(r), sz == 3 ? svcul : (uint32_t)svcul);
         return;
     }
 
@@ -636,9 +613,9 @@ ST_FUNC void load(int r, SValue *sv)
     if (svr == (VT_LLOCAL | VT_LVAL)) {
         arm64_ldrx(0, 3, 30, 29, svcoff); // use x30 for offset
         if (IS_FREG(r))
-            arm64_ldrv(arm64_type_size(svtt), fltr(r), 30, 0);
+            arm64_ldrv(sz, fltr(r), 30, 0);
         else
-            arm64_ldrx(!(svtt & VT_UNSIGNED), arm64_type_size(svtt),
+            arm64_ldrx(sb, sz,
                        intr(r), 30, 0);
         return;
     }
@@ -655,63 +632,43 @@ ST_FUNC void load(int r, SValue *sv)
 ST_FUNC void store(int r, SValue *sv)
 {
     int svtt = sv->type.t;
-    int svr = sv->r & ~VT_BOUNDED;
+    int svr = sv->r & (VT_VALMASK | VT_LVAL | VT_SYM);
     int svrv = svr & VT_VALMASK;
-    uint64_t svcoff = (uint64_t)(int64_t)(int32_t)sv->c.i;
+    uint64_t svcoff = (int32_t)sv->c.i;
+    int sz = arm64_type_size(svtt);
 
     if (svr == (VT_LOCAL | VT_LVAL)) {
         if (IS_FREG(r))
-            arm64_strv(arm64_type_size(svtt), fltr(r), 29, svcoff);
+            arm64_strv(sz, fltr(r), 29, svcoff);
         else
-            arm64_strx(arm64_type_size(svtt), intr(r), 29, svcoff);
+            arm64_strx(sz, intr(r), 29, svcoff);
         return;
     }
 
     if (svr == (VT_CONST | VT_LVAL)) {
-	uint64_t i = sv->c.i;
-	if (sv->sym && (sv->sym->type.t & VT_TLS))
-            arm64_tls_sym(30, sv->sym, 0);
-        else
-	if (sv->sym)
-            arm64_sym(30, sv->sym, // use x30 for address
-		      arm64_check_offset(0, arm64_type_size(svtt), i));
-	else
-	    arm64_movimm (30, i), i = 0;
-
+	arm64_movimm (30, sv->c.i);
         if (IS_FREG(r))
-            arm64_strv(arm64_type_size(svtt), fltr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), i));
+            arm64_strv(sz, fltr(r), 30, 0);
         else
-            arm64_strx(arm64_type_size(svtt), intr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), i));
+            arm64_strx(sz, intr(r), 30, 0);
         return;
     }
 
-    if ((svr & ~VT_VALMASK) == VT_LVAL && svrv < VT_CONST) {
+    if (svrv < VT_CONST && (svr & VT_LVAL)) {
         if (IS_FREG(r))
-            arm64_strv(arm64_type_size(svtt), fltr(r), intr(svrv), 0);
+            arm64_strv(sz, fltr(r), intr(svrv), 0);
         else
-            arm64_strx(arm64_type_size(svtt), intr(r), intr(svrv), 0);
+            arm64_strx(sz, intr(r), intr(svrv), 0);
         return;
     }
 
     if (svr == (VT_CONST | VT_LVAL | VT_SYM)) {
-        if (sv->sym->type.t & VT_TLS) {
-            arm64_tls_sym(30, sv->sym, 0);
-            if (IS_FREG(r))
-                arm64_strv(arm64_type_size(svtt), fltr(r), 30, svcoff);
-            else
-                arm64_strx(arm64_type_size(svtt), intr(r), 30, svcoff);
-            return;
-        }
-        arm64_sym(30, sv->sym, // use x30 for address
-		  arm64_check_offset(0, arm64_type_size(svtt), svcoff));
+        uint64_t mask = arm64_check_offset(sz, svcoff);
+        arm64_sym(30, sv->sym, svcoff & ~mask);
         if (IS_FREG(r))
-            arm64_strv(arm64_type_size(svtt), fltr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), svcoff));
+            arm64_strv(sz, fltr(r), 30, svcoff & mask);
         else
-            arm64_strx(arm64_type_size(svtt), intr(r), 30,
-		       arm64_check_offset(1, arm64_type_size(svtt), svcoff));
+            arm64_strx(sz, intr(r), 30, svcoff & mask);
         return;
     }
 
@@ -1789,9 +1746,9 @@ static int arm64_gen_opic(int op, uint32_t l, int rev, uint64_t val,
         uint32_t s = l ? val >> 63 : val >> 31;
         val = s ? -val : val;
         val = l ? val : (uint32_t)val;
-        if (!(val & ~0xffful))
+        if (!(val & ~0xfffull))
             o(0x11000000 | l << 31 | s << 30 | x | a << 5 | val << 10);
-        else if (!(val & ~0xfff000ul))
+        else if (!(val & ~0xfff000ull))
             o(0x11400000 | l << 31 | s << 30 | x | a << 5 | val >> 12 << 10);
         else {
             arm64_movimm(30, val); // use x30
diff --git a/arm64-link.c b/arm64-link.c
index dd02479a..d3d925dd 100644
--- a/arm64-link.c
+++ b/arm64-link.c
@@ -371,8 +371,12 @@ ST_FUNC void relocate(TCCState *s1, ElfW_Rel *rel, int type, unsigned char *ptr,
             return;
         case R_AARCH64_TLSLE_ADD_TPREL_HI12:
         case R_AARCH64_TLSLE_ADD_TPREL_LO12: {
+#if TCC_TARGET_PE
+            int64_t tp_offset = val - s1->tls_start;
+#else
             /* glibc arm64: tp points to tcbhead_t (DTV), TLS data starts after it */
             int64_t tp_offset = val - s1->tls_start + 16;
+#endif
             int64_t imm;
             if (type == R_AARCH64_TLSLE_ADD_TPREL_HI12)
                 imm = (tp_offset >> 12) & 0xfff;
diff --git a/configure b/configure
index 97999e1c..c63c7403 100755
--- a/configure
+++ b/configure
@@ -363,8 +363,10 @@ case $targetos in
     default_conf "codesign"
     DLLSUF=".dylib"
     if test -z "$build_cross"; then
-      cc=`command -v cc`
-      cc=`readlink $cc || echo clang`
+      if test "$cc" = "gcc"; then
+        cc=`command -v cc`
+        cc=`readlink $cc || echo clang`
+      fi
       tcc_usrinclude="`xcrun --show-sdk-path`/usr/include"
       default tcc_sysincludepaths "{B}/include:$tcc_usrinclude"
       if test "${confvars%new_macho*}" = "${confvars}"; then
diff --git a/i386-gen.c b/i386-gen.c
index 54d24806..36a246ed 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -242,12 +242,20 @@ static void gen_modrm(int opc, int op_r2, int r, Sym *sym, int c)
     int op_reg = REG_VALUE(op_r2) << 3;
 
     if ((r & VT_SYM) && (sym->type.t & VT_TLS)) {
-	if (opc == 0xdbc0d9) // VT_LDOUBLE
-	    o(0xc0d9), opc = 0xdb;
-        o(0x65); /* gs segment prefix */
-        o(opc);
-        oad(0x05 | op_reg, c);
+#ifdef TCC_TARGET_PE
+        Sym *s2 = external_global_sym(TOK___tls_index, &int_type);
+        r = get_reg(RC_INT);
+        gen_modrm(0x8b, r, VT_SYM|VT_CONST, s2, 0); /* mov __tls_index, r */
+        o(0x02e0c1 | r << 8); /* shl 2,r */
+        oad(0x050364 | r << 19, 11*PTR_SIZE); /* add fs:0x2c,r */
+        gen_modrm(0x8b, r, r | VT_LVAL, 0, 0); /* mov (r),r */
+        o(opc), oad(0x80 | op_reg | r, c); /* mov #c(r),op_reg */
         greloc(cur_text_section, sym, ind - 4, R_386_TLS_LE);
+#else
+        o(0x65); /* gs segment prefix */
+        o(opc), oad(0x05 | op_reg, c);
+        greloc(cur_text_section, sym, ind - 4, R_386_TLS_LE);
+#endif
 #if defined CONFIG_TCC_PIC
     } else if ((r & (VT_VALMASK|VT_SYM)) == (VT_CONST|VT_SYM)) {
         int is_got = (op_r2 & TREG_MEM) && !(sym->type.t & VT_STATIC);
@@ -353,9 +361,13 @@ ST_FUNC void load(int r, SValue *sv)
         gen_modrm(opc, r, fr, sv->sym, fc);
     } else {
         if ((fr & VT_SYM) && (sv->sym->type.t & VT_TLS)) {
+#ifdef TCC_TARGET_PE
+            gen_modrm(0x8d, r, fr, sv->sym, fc);
+#else
             oad(0x058b65 | REG_VALUE(r) << 19, 0); /* mov gs:0,r */
             oad(0xC081 | REG_VALUE(r) << 8, fc); /* add tpoffs,r */
             greloc(cur_text_section, sv->sym, ind - 4, R_386_TLS_LE);
+#endif
 #if defined CONFIG_TCC_PIC
         } else if ((fr & (VT_VALMASK|VT_SYM)) == (VT_CONST|VT_SYM)) {
             if (sv->sym->type.t & VT_STATIC) {
@@ -412,7 +424,8 @@ ST_FUNC void store(int r, SValue *v)
         opc = 0xdd; /* fstpl */
         r = 2;
     } else if (bt == VT_LDOUBLE) {
-        opc = 0xdbc0d9; /* fld %st(0), fstpt */
+        o(0xc0d9); /* fld %st(0), fstpt */
+        opc = 0xdb;
         r = 7;
     } else if (bt == VT_SHORT) {
         opc = 0x8966;
diff --git a/libtcc.c b/libtcc.c
index fcc035a9..44ea25fa 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -1689,6 +1689,7 @@ static const TCCOption tcc_options[] = {
     { "-param", 0, TCC_OPTION_HAS_ARG },
     { "pedantic", 0, 0 },
     { "pie", 0, 0 },
+    { "no-pie", 0, 0 },
     { "pipe", 0, 0 },
     { "s", 0, 0 },
     { "traditional", 0, 0 },
diff --git a/tcc.h b/tcc.h
index 15333b3d..abffcbfe 100644
--- a/tcc.h
+++ b/tcc.h
@@ -71,6 +71,9 @@ extern long double strtold (const char *__nptr, char **__endptr);
 #  define strtoll _strtoi64
 #  define strtoull _strtoui64
 # endif
+  /* some compilers don't know ldexpl and windows doesn't have long doubles anyway */
+# undef ldexpl
+# define ldexpl ldexp
 # ifdef LIBTCC_AS_DLL
 #  define LIBTCCAPI __declspec(dllexport)
 #  define PUB_FUNC LIBTCCAPI
diff --git a/tccgen.c b/tccgen.c
index 218fb063..227ff618 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -842,7 +842,7 @@ ST_FUNC void label_pop(Sym **ptop, Sym *slast, int keep)
             if (s->c) {
                 /* define corresponding symbol. A size of
                    1 is put. */
-                put_extern_sym(s, cur_text_section, s->jnext, 1);
+                put_extern_sym(s, cur_text_section, s->jind, 1);
             }
         }
         /* remove label */
diff --git a/tccpe.c b/tccpe.c
index ab30268e..90eae16b 100644
--- a/tccpe.c
+++ b/tccpe.c
@@ -241,6 +241,15 @@ typedef struct _IMAGE_EXPORT_DIRECTORY {
     DWORD AddressOfNameOrdinals;
 } IMAGE_EXPORT_DIRECTORY,*PIMAGE_EXPORT_DIRECTORY;
 
+typedef struct _IMAGE_TLS_DIRECTORY {
+    ADDR3264 StartAddressOfRawData;
+    ADDR3264 EndAddressOfRawData;
+    ADDR3264 AddressOfIndex;
+    ADDR3264 AddressOfCallBacks;
+    DWORD SizeOfZeroFill;
+    DWORD Characteristics;
+} IMAGE_TLS_DIRECTORY;
+
 typedef struct _IMAGE_IMPORT_DESCRIPTOR {
     union {
         DWORD Characteristics;
@@ -342,6 +351,7 @@ enum {
     sec_bss ,
     sec_idata ,
     sec_pdata ,
+    sec_tls ,
     sec_other ,
     sec_rsrc ,
     sec_debug ,
@@ -404,6 +414,9 @@ struct pe_info {
     DWORD iat_size;
     DWORD exp_offs;
     DWORD exp_size;
+    DWORD tls_dir;
+    DWORD tls_data;
+    DWORD tls_size;
     int subsystem;
     DWORD section_align;
     DWORD file_align;
@@ -758,6 +771,10 @@ static int pe_write(struct pe_info *pe)
             pe_set_datadir(&pe_header, IMAGE_DIRECTORY_ENTRY_EXPORT,
                 pe->exp_offs, pe->exp_size);
         }
+        if (pe->tls_size) {
+            pe_set_datadir(&pe_header, IMAGE_DIRECTORY_ENTRY_TLS,
+                pe->tls_dir + (pe->thunk->sh_addr - pe->imagebase), pe->tls_size);
+        }
 
         memcpy(psh->Name, sh_name, umin(strlen(sh_name), sizeof psh->Name));
         if (pe->coffstr && strlen(sh_name) > 8) {
@@ -1172,6 +1189,39 @@ static void pe_build_reloc (struct pe_info *pe)
     }
 }
 
+/* ------------------------------------------------------------- */
+static void pe_build_tls(struct pe_info *pe, Section *s)
+{
+    TCCState *s1 = pe->s1;
+    IMAGE_TLS_DIRECTORY *d;
+    int c, n;
+
+    if (0 == s) {
+        pe->tls_dir = section_add(pe->thunk, pe->tls_size, 16);
+        pe->tls_data = section_add(data_section, PTR_SIZE * (1+3), 16);
+        /* put relocations on entries */
+        c = put_elf_sym(symtab_section, 0, 0, 0, 0, data_section->sh_num, 0);
+        for (n = 0; n < 4; ++n)
+            put_elf_reloc(symtab_section, pe->thunk, pe->tls_dir + PTR_SIZE*n, REL_TYPE_DIRECT, c);
+        /* for generators */
+        set_elf_sym(symtab_section, pe->tls_data, PTR_SIZE * 4,
+                ELFW(ST_INFO)(STB_GLOBAL, STT_OBJECT),
+                0, data_section->sh_num, "__tls_index");
+        return;
+    }
+    if (0 == s1->tls_start)
+        s1->tls_start = s->sh_addr;
+    d = (void*)(pe->thunk->data + pe->tls_dir);
+    d->StartAddressOfRawData = s1->tls_start - data_section->sh_addr;
+    d->EndAddressOfRawData = s->sh_addr + s->data_offset - data_section->sh_addr;
+    d->AddressOfIndex = pe->tls_data;
+    d->AddressOfCallBacks = pe->tls_data + PTR_SIZE;
+    d->SizeOfZeroFill = 0;
+    d->Characteristics = 0;
+    /* to reuse logic from linux in xxx-link.c */
+    s1->tls_end = s1->tls_start;
+}
+
 /* ------------------------------------------------------------- */
 static int pe_section_class(Section *s)
 {
@@ -1184,6 +1234,8 @@ static int pe_section_class(Section *s)
     if (0 == memcmp(name, ".stab", 5) || 0 == memcmp(name, ".debug_", 7)) {
         return sec_debug;
     } else if (flags & SHF_ALLOC) {
+        if (flags & SHF_TLS)
+            return sec_tls;
         if (type == SHT_PROGBITS
          || type == SHT_INIT_ARRAY
          || type == SHT_FINI_ARRAY) {
@@ -1232,6 +1284,8 @@ static int pe_assign_addresses (struct pe_info *pe)
         for (n = i; n > 1 && k < (c = sec_cls[n - 1]); --n)
             sec_cls[n] = c, sec_order[n] = sec_order[n - 1];
         sec_cls[n] = k, sec_order[n] = i;
+        if (k == sec_tls)
+            pe->tls_size = sizeof (IMAGE_TLS_DIRECTORY);
     }
     si = NULL;
     addr = pe->imagebase + 1;
@@ -1257,7 +1311,10 @@ static int pe_assign_addresses (struct pe_info *pe)
         if (s == pe->thunk) {
             pe_build_imports(pe);
             pe_build_exports(pe);
+            if (pe->tls_size)
+                pe_build_tls(pe, NULL);
         }
+
         if (s == pe->reloc)
             pe_build_reloc (pe);
 
@@ -1277,7 +1334,7 @@ static int pe_assign_addresses (struct pe_info *pe)
         si->pe_flags = IMAGE_SCN_MEM_READ;
         if (s->sh_flags & SHF_EXECINSTR)
             si->pe_flags |= IMAGE_SCN_MEM_EXECUTE | IMAGE_SCN_CNT_CODE;
-        else if (s->sh_type == SHT_NOBITS)
+        else if (s->sh_type == SHT_NOBITS && !(s->sh_flags & SHF_TLS))
             si->pe_flags |= IMAGE_SCN_CNT_UNINITIALIZED_DATA;
         else
             si->pe_flags |= IMAGE_SCN_CNT_INITIALIZED_DATA;
@@ -1297,6 +1354,12 @@ add_section:
             *ps = s, s->prev = NULL;
             si->data_size = si->sh_size;
         }
+
+        if (s->sh_flags & SHF_TLS) {
+            strcpy(si->name, ".tls");
+            pe_build_tls(pe, s);
+        }
+
         //printf("%08x %05x %08x %s\n", si->sh_addr, si->sh_size, si->pe_flags, s->name);
     }
 #if 0
@@ -1766,35 +1829,28 @@ static char *trimback(char *a, char *e)
 
 static char *get_token(char **s, char *f)
 {
-    char *p = *s, *e;
-    p = e = trimfront(p);
-    while ((unsigned char)*e > ' ')
-        ++e;
-    *s = trimfront(e);
-    *f = **s; *e = 0;
-    return p;
-}
+    char *p, *e;
+    int q;
 
-/* Read a LIBRARY argument whose quotes may protect whitespace. */
-static char *get_libname(char **s, char *f)
-{
-    char *p = trimfront(*s), *e;
-    if (*p != '"')
-        return get_token(s, f);
-    for (e = ++p; *e && *e != '"' && *e != '\r' && *e != '\n'; ++e)
+    p = trimfront(*s);
+    q = *p;
+    if (q == '"') /* support quoted LIBRARY "xyz.dll" */
+        ++p;
+    else
+        q = ' ';
+    for (e = p; (unsigned char)*e >= ' ' && *e != q; ++e)
         ;
-    if (*e != '"' || (unsigned char)e[1] > ' ')
-        return NULL;
-    *s = trimfront(e + 1);
-    *f = **s;
-    *e = 0;
+    if (*e == '"')
+        *e++ = 0;
+    *s = trimfront(e);
+    *f = **s, *e = 0;
     return p;
 }
 
 static int pe_load_def(TCCState *s1, int fd)
 {
     int state = 0, ret = -1, dllindex = 0, ord;
-    char *dllname = NULL, *buf, *line, *p, *x, next;
+    char dllname[80], *buf, *line, *p, *x, next;
 
     buf = tcc_load_text(fd);
     if (!buf)
@@ -1808,13 +1864,9 @@ static int pe_load_def(TCCState *s1, int fd)
         case 0:
             if (0 != stricmp(p, "LIBRARY") || next == '\n')
                 goto quit;
-            p = get_libname(&line, &next);
-            if (!p || !*p)
-                goto quit;
-            dllname = tcc_malloc(strlen(p) + 5);
-            strcpy(dllname, p);
+            pstrcpy(dllname, sizeof dllname, get_token(&line, &next));
             if (!*tcc_fileextension(dllname))
-                strcat(dllname, ".dll");
+                pstrcat(dllname, sizeof dllname, ".dll");
             ++state;
             break;
         case 1:
@@ -1845,7 +1897,6 @@ skip:
     }
     ret = 0;
 quit:
-    tcc_free(dllname);
     tcc_free(buf);
     return ret;
 }
diff --git a/tccpp.c b/tccpp.c
index c6bf1860..214e9cd4 100644
--- a/tccpp.c
+++ b/tccpp.c
@@ -21,21 +21,6 @@
 #define USING_GLOBALS
 #include "tcc.h"
 
-#if defined(_WIN32) && defined(__TINYC__)
-  /* allow self-host build with tcc 0.9.27 - doesn't have ldexpl in tcc_libm.h .
-   *
-   * in tcc 0.9.27 both ldexp and ldexpl are declared in win32/include/math.h,
-   * but only ldexp can be linked - via win32/lib/msvcrt.def. we can't test
-   * whether we have ldexpl or not, so map ldexpl to ldexp unconditionally.
-   *
-   * note that ldexpl takes long double while ldexp takes double, however, on
-   * windows these types are identical, and current ldexpl in tcc_libm.h also
-   * uses the "normal" double scalbn - just like ldexp, so do the same here.
-   */
-  #undef ldexpl
-  #define ldexpl ldexp
-#endif
-
 /* #define to 1 to enable (see parse_pp_string()) */
 #define ACCEPT_LF_IN_STRINGS 0
 
diff --git a/tcctok.h b/tcctok.h
index 458e106d..d8e7b445 100644
--- a/tcctok.h
+++ b/tcctok.h
@@ -306,6 +306,7 @@
 
 #if defined TCC_TARGET_PE
      DEF(TOK___chkstk, "__chkstk")
+     DEF(TOK___tls_index, "__tls_index")
 #endif
 #ifdef TCC_TARGET_ARM64
      DEF(TOK___arm64_clear_cache, "__arm64_clear_cache")
diff --git a/tests/test-win32.bat b/tests/test-win32.bat
index bf593cbf..ef3b4afe 100644
--- a/tests/test-win32.bat
+++ b/tests/test-win32.bat
@@ -70,5 +70,5 @@ echo>>%CFG_MAK% TOPSRC = $(TOP)
 echo>>%CFG_MAK% SHELL = sh
 echo>>%CFG_MAK% test.ref: CFLAGS += %REF_LINK%
 
-set GMAKE=make
+if "%GMAKE%"=="" set GMAKE=make
 %GMAKE% TCC_LOCAL=tcc.exe LIBTCC=win32/%LIBTCC% %TESTS%
diff --git a/tests/tests2/124_atomic_counter.c b/tests/tests2/124_atomic_counter.c
index 42de43d4..5017ad50 100644
--- a/tests/tests2/124_atomic_counter.c
+++ b/tests/tests2/124_atomic_counter.c
@@ -2,7 +2,7 @@
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
-#include <pthread.h>
+#include "threads.h"
 #include <stdatomic.h>
 
 #define NR_THREADS 16
diff --git a/tests/tests2/144_tls.c b/tests/tests2/144_tls.c
index 82c844ce..c7134ede 100644
--- a/tests/tests2/144_tls.c
+++ b/tests/tests2/144_tls.c
@@ -1,5 +1,11 @@
 #include <stdio.h>
-#include <pthread.h>
+#include "threads.h"
+
+#if __SIZEOF_POINTER__ == 8
+# define pFMT "0x%llx"
+#else
+# define pFMT "0x%x"
+#endif
 
 #define	CHECK(var,fmt,val)	printf(fmt "\n",var); if (var != val) errors = 1
 
@@ -70,7 +76,7 @@ static int check(void)
     CHECK(tls_float, "%g", 46.0);
     CHECK(tls_double, "%g", 47.0);
     CHECK(tls_long_double, "%Lg", 48.0);
-    CHECK(tls_ptr, "%p", (int *)49);
+    CHECK(tls_ptr, pFMT, (int *)49);
     CHECK(tls_enum, "%d", tls_b);
     CHECK(tls_struct.tls_d, "%d", 50);
     CHECK(tls_struct.tls_e, "%d", 51);
@@ -121,7 +127,7 @@ static void *thread_func(void *arg)
     CHECK(tls_double, "%g", 70.0);
     CHECK(tls_long_double, "%Lg", 80.0);
     CHECK(tls_enum, "%d", tls_c);
-    CHECK(tls_ptr, "%p", (int *)90);
+    CHECK(tls_ptr, pFMT, (int *)90);
     CHECK(tls_struct.tls_d, "%d", 100);
     CHECK(tls_struct.tls_e, "%d", 110);
     CHECK(tls_struct.tls_f, "%d", 120);
@@ -130,7 +136,7 @@ static void *thread_func(void *arg)
         || tls_array[index] != -3333 || check_extra(1))
         errors = 1;
 
-    return (void *)errors;
+    return (void *)(size_t)errors;
 }
 
 int main()
diff --git a/tests/tests2/147_def_library.c b/tests/tests2/147_def_library.c
deleted file mode 100644
index 8e1d9b3f..00000000
--- a/tests/tests2/147_def_library.c
+++ /dev/null
@@ -1,21 +0,0 @@
-#ifdef DEF_LIBRARY_DLL
-
-__declspec(dllexport) int def_library_value(void)
-{
-    return 42;
-}
-
-#else
-
-#include <stdio.h>
-
-int def_library_value(void);
-
-int main(void)
-{
-    int value = def_library_value();
-    printf("value: %d\n", value);
-    return value != 42;
-}
-
-#endif
diff --git a/tests/tests2/147_def_library.def.in b/tests/tests2/147_def_library.def.in
deleted file mode 100644
index aedf95d0..00000000
--- a/tests/tests2/147_def_library.def.in
+++ /dev/null
@@ -1,3 +0,0 @@
-LIBRARY "def library test"
-EXPORTS
-    def_library_value
diff --git a/tests/tests2/147_def_library.expect b/tests/tests2/147_def_library.expect
deleted file mode 100644
index 84a64c37..00000000
--- a/tests/tests2/147_def_library.expect
+++ /dev/null
@@ -1 +0,0 @@
-value: 42
diff --git a/tests/tests2/Makefile b/tests/tests2/Makefile
index d8c4ebb8..91a043cf 100644
--- a/tests/tests2/Makefile
+++ b/tests/tests2/Makefile
@@ -15,9 +15,6 @@ endif
 ifeq (,$(filter i386,$(ARCH)))
  SKIP += 98_al_ax_extend.test 99_fastcall.test
 endif
-ifeq (,$(filter x86_64 riscv64 arm64 i386,$(ARCH)))
- SKIP += 146_tls_extern.test # TLS only implemented on these architectures
-endif
 ifeq (,$(filter i386 x86_64,$(ARCH)))
  SKIP += 85_asm-outside-function.test # x86 asm
  SKIP += 127_asm_goto.test    # hardcodes x86 asm
@@ -37,15 +34,12 @@ ifeq ($(CONFIG_bcheck),no)
  SKIP += 132_bound_test.test
 endif
 ifeq ($(CONFIG_dll),no)
- SKIP += 113_btdll.test 147_def_library.test # no shared lib support yet
+ SKIP += 113_btdll.test # no shared lib support yet
  SKIP += 149_end_copy_reloc.test 151_dso_linker_symbol.test
 endif
 ifeq (-$(findstring gcc,$(CC))-,--)
  SKIP += $(patsubst %.expect,%.test,$(GEN-ALWAYS))
 endif
-ifeq ($(CC_NAME),tcc)
- SKIP += 146_tls_extern.test 150_linker_boundaries.test # needs external tools
-endif
 ifeq (-$(CONFIG_WIN32)-$(CONFIG_i386)$(CONFIG_arm)-,--yes-)
  SKIP += 95_bitfields%.test # type_align is different on 32bit-non-windows
 endif
@@ -53,20 +47,15 @@ ifeq (-$(CONFIG_WIN32)-,-yes-)
  SKIP += 106_versym.test # No pthread support
  SKIP += 114_bound_signal.test # No pthread support
  SKIP += 117_builtins.test # win32 port doesn't define __builtins
- SKIP += 124_atomic_counter.test # No pthread support
- SKIP += 144_tls.test 146_tls_extern.test # TLS runtime not supported
  SKIP += 148_linker_symbols.test 149_end_copy_reloc.test # ELF linker tests
  SKIP += 150_linker_boundaries.test 151_dso_linker_symbol.test
 endif
-ifneq (,$(filter OpenBSD FreeBSD NetBSD,$(TARGETOS)))
+ifneq (,$(filter OpenBSD FreeBSD NetBSD DragonFly,$(TARGETOS)))
  SKIP += 106_versym.test # no pthread_condattr_setpshared
  SKIP += 114_bound_signal.test # libc problem signal/fork
  SKIP += 116_bound_setjmp2.test # No TLS_FUNC/TLS_VAR in bcheck.c
  SKIP += 144_tls.test 146_tls_extern.test # No tls support
 endif
-ifeq ($(TARGETOS),DragonFly)
- SKIP += 144_tls.test 146_tls_extern.test # TLS runtime not supported
-endif
 ifeq ($(CONFIG_OSX),yes)
  SKIP += 144_tls.test 146_tls_extern.test # TLS runtime not supported on Mach-O
  SKIP += 148_linker_symbols.test 149_end_copy_reloc.test # ELF linker tests
@@ -89,9 +78,6 @@ endif
 ifneq (-$(ARCH)-$(CONFIG_WIN32)-,-arm64-yes-)
  SKIP += 145_winarm64_interlocked.test
 endif
-ifneq ($(CONFIG_WIN32),yes)
- SKIP += 147_def_library.test
-endif
 
 # Some tests might need arguments
 ARGS =
@@ -141,7 +127,6 @@ GEN-ALWAYS =
 # TLS needs executable and pthread, not -run
 144_tls.test: FLAGS += -pthread
 144_tls.test: NORUN = true
-
 # Verify that undefined TLS symbols in TCC objects work with the system linker.
 146_tls_extern.test: NORUN = true
 146_tls_extern.test: T1 = \
@@ -150,6 +135,9 @@ GEN-ALWAYS =
 	$(CC) -no-pie -Wl,-z,noexecstack 146_tls_extern-main.o \
 		146_tls_extern-def.o -o $(basename $@).exe && \
 	./$(basename $@).exe
+ifeq (-$(CONFIG_WIN32)-,-yes-)
+146_tls_extern.test : CC=$(TCC)
+endif
 
 # Check optional ELF linker symbols, including references from an archive.
 148_linker_symbols.test: NORUN = true
@@ -204,13 +192,6 @@ GEN-ALWAYS =
     $(TCC) -bt $1 a1$(DLLSUF) a2$(DLLSUF) -Wl,-rpath=. -o $(basename $@).exe && \
     ./$(basename $@).exe
 
-# Verify quoted LIBRARY names and the default .dll extension in DEF files.
-147_def_library.test: T1 = \
-    cp $(SRC)/147_def_library.def.in 147_def_library.def && \
-    $(TCC) -D DEF_LIBRARY_DLL -shared $1 -o "def library test.dll" && \
-    $(TCC) $1 147_def_library.def -o $(basename $@).exe && \
-    grep -a -q "def library test.dll" $(basename $@).exe && \
-    ./$(basename $@).exe
 
 114_bound_signal.test: FLAGS += -b
 114_bound_signal.test: NORUN = true # tcc -run does not support fork and -b and SELINUX
diff --git a/tests/tests2/threads.h b/tests/tests2/threads.h
new file mode 100644
index 00000000..33b7aa2b
--- /dev/null
+++ b/tests/tests2/threads.h
@@ -0,0 +1,22 @@
+#ifdef _WIN32
+#include <windows.h>
+typedef HANDLE pthread_t;
+int pthread_create(pthread_t *p, void *x, void *fn, void *args)
+{
+    DWORD tid;
+    *p = CreateThread(NULL, 0, fn, args, 0, &tid);
+    return *p ? 0 : -1;
+}
+int pthread_join(pthread_t h, void **ret)
+{
+    DWORD r;
+    WaitForSingleObject(h, INFINITE);
+    GetExitCodeThread(h, &r);
+    CloseHandle(h);
+    if (ret)
+        *ret = (void*)(size_t)r;
+    return 0;
+}
+#else
+#include <pthread.h>
+#endif
diff --git a/x86_64-gen.c b/x86_64-gen.c
index 8509a3ae..ffa4cbc8 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -221,8 +221,11 @@ static void orex(int ll, int r, int r2, int b)
         r = 0;
     if ((r2 & VT_VALMASK) >= VT_CONST)
         r2 = 0;
-    if (ll || REX_BASE(r) || REX_BASE(r2))
+    if (ll || REX_BASE(r) || REX_BASE(r2)) {
+        if ((b & 0xff) == 0x66) /* output prefix before rex byte */
+            o(0x66), b >>= 8;
         o(0x40 | REX_BASE(r) | (REX_BASE(r2) << 2) | (ll << 3));
+    }
     o(b);
 }
 
@@ -301,6 +304,26 @@ static void gen_modrm_impl(int opcode, int ll, int op_reg_0, int r, Sym *sym, in
 {
     int op_reg = REG_VALUE(op_reg_0) << 3;
 
+    if ((r & VT_SYM) && (sym->type.t & VT_TLS)) {
+#ifdef TCC_TARGET_PE
+        Sym *s2 = external_global_sym(TOK___tls_index, &int_type);
+        r = get_reg(RC_INT);
+        gen_modrm_impl(0x8B, 0,  r, VT_SYM|VT_CONST, s2, 0);
+        o(0x03e0c148 | r << 16); /* shl 2,r */
+        o(0x4865), oad(0x250403 | r << 11, 11*PTR_SIZE); /* add gs:0x58,r */
+        gen_modrm_impl(0x8B, 1, r, r | VT_LVAL, 0, 0); /* mov (r),r */
+        orex(ll, r, op_reg_0, opcode), oad(0x80 | op_reg | r, 0);
+        greloca(cur_text_section, sym, ind - 4, R_X86_64_TPOFF32, c);
+#else
+        o(0x64); /* fs segment prefix */
+        orex(ll, r, op_reg_0, opcode);
+	o(0x04 | op_reg); /* [sib] | destreg */
+	oad(0x25, 0);     /* disp32 (relocated) */
+	greloca(cur_text_section, sym, ind - 4, R_X86_64_TPOFF32, c);
+#endif
+        return;
+    }
+
     orex(ll, r, op_reg_0, opcode);
 
     if ((r & VT_VALMASK) == VT_CONST) {
@@ -309,10 +332,6 @@ static void gen_modrm_impl(int opcode, int ll, int op_reg_0, int r, Sym *sym, in
 	    /* Absolute memory reference */
 	    o(0x04 | op_reg); /* [sib] | destreg */
 	    oad(0x25, c);     /* disp32 */
-	} else if (sym->type.t & VT_TLS) {
-	    o(0x04 | op_reg); /* [sib] | destreg */
-	    oad(0x25, 0);     /* disp32 (relocated) */
-	    greloca(cur_text_section, sym, ind - 4, R_X86_64_TPOFF32, c);
 	} else {
 	    o(0x05 | op_reg); /* (%rip)+disp32 | destreg */
 	    if (op_reg_0 & TREG_MEM) {
@@ -454,16 +473,12 @@ void load(int r, SValue *sv)
             ll = is64_type(ft);
             b = 0x8b;
         }
-        if ((sv->r & VT_SYM) && (sv->sym->type.t & VT_TLS))
-            o(0x64); /* fs segment prefix */
         gen_modrm_impl(b, ll, r, fr, sv->sym, fc);
     } else {
         if (v == VT_CONST) {
             if (fr & VT_SYM) {
 #ifdef TCC_TARGET_PE
-                orex(1,0,r,0x8d);
-                o(0x05 + REG_VALUE(r) * 8); /* lea xx(%rip), r */
-                gen_addrpc32(fr, sv->sym, fc);
+                gen_modrm64(0x8d, r, fr, sv->sym, fc);
 #else
                 if (sv->sym->type.t & VT_TLS) {
                     /* mov fs:0, r */
@@ -563,7 +578,7 @@ void store(int r, SValue *v)
 {
     int fr, bt, fc;
     /* store the REX prefix in this variable when PIC is enabled */
-    int prefix = 0, opc = 0, seg = 0, ll = 0;
+    int opc = 0, ll = 0;
 
     fr = v->r;
     fc = v->c.i;
@@ -584,12 +599,10 @@ void store(int r, SValue *v)
 
     /* XXX: incorrect if float reg to reg */
     if (bt == VT_FLOAT) {
-        prefix = 0x66;
-        opc = 0x7e0f;
+        opc = 0x7e0f66;
         r = REG_VALUE(r);
     } else if (bt == VT_DOUBLE) {
-        prefix = 0x66;
-        opc = 0xd60f;
+        opc = 0xd60f66;
         r = REG_VALUE(r);
     } else if (bt == VT_LDOUBLE) {
         o(0xc0d9); /* fld %st(0) */
@@ -600,15 +613,11 @@ void store(int r, SValue *v)
     } else {
         opc = 0x89;
         if (bt == VT_SHORT)
-            prefix = 0x66;
+            opc = 0x8966;
         else if (is64_type(bt))
             ll = 1;
     }
-
-    if ((fr & VT_SYM) && (v->sym->type.t & VT_TLS))
-        seg = 0x64; /* fs segment prefix */
-
-    o(seg), o(prefix), gen_modrm_impl(opc, ll, r, fr, v->sym, fc);
+    gen_modrm_impl(opc, ll, r, fr, v->sym, fc);
 }
 
 /* 'is_jmp' is '1' if it is a jump */
@@ -1932,15 +1941,12 @@ void gen_opf(int op)
             }
             assert(!(vtop[-1].r & VT_LVAL));
             
-	    if ((vtop->r & VT_SYM) && (vtop->sym->type.t & VT_TLS))
-                o(0x64); /* fs segment prefix */
-
-            if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
-                o(0x66);
             if (op == TOK_EQ || op == TOK_NE)
                 opc = 0x2e0f; /* ucomisd */
             else
                 opc = 0x2f0f; /* comisd */
+            if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
+                opc = opc << 8 | 0x66;
 
             gen_modrm32(opc, vtop[-1].r, vtop->r, vtop->sym, fc);
             vtop--;