diff --git a/arm-gen.c b/arm-gen.c
index fcde7dcb..c0dd5099 100644
--- a/arm-gen.c
+++ b/arm-gen.c
@@ -1332,7 +1332,7 @@ again:
    parameters and the function address. */
 void gfunc_call(int nb_args)
 {
-  int r, args_size;
+  int args_size;
   int def_float_abi = float_abi;
   int todo;
   struct plan plan;
@@ -1352,12 +1352,6 @@ void gfunc_call(int nb_args)
       float_abi = ARM_SOFTFP_FLOAT;
   }
 #endif
-  /* cannot let cpu flags if other instruction are generated. Also avoid leaving
-     VT_JMP anywhere except on the top of the stack because it would complicate
-     the code generator. */
-  r = vtop->r & VT_VALMASK;
-  if (r == VT_CMP || (r & ~1) == VT_JMP)
-    gv(RC_INT);
 
   memset(&plan, 0, sizeof plan);
   if (nb_args)
@@ -1716,9 +1710,6 @@ void gen_opi(int op)
 	  opc|=2; // sub -> rsb
 	}
       }
-      if ((vtop->r & VT_VALMASK) == VT_CMP ||
-          (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
-        gv(RC_INT);
       vswap();
       c=intr(gv(RC_INT));
       vswap();
@@ -1757,9 +1748,6 @@ done:
       break;
     case 2:
       opc=0xE1A00000|(opc<<5);
-      if ((vtop->r & VT_VALMASK) == VT_CMP ||
-          (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
-        gv(RC_INT);
       vswap();
       r=intr(gv(RC_INT));
       vswap();
diff --git a/arm64-gen.c b/arm64-gen.c
index e314de58..d15446d3 100644
--- a/arm64-gen.c
+++ b/arm64-gen.c
@@ -1029,10 +1029,6 @@ ST_FUNC void gfunc_call(int nb_args)
 
     stack = (stack + 15) >> 4 << 4;
 
-    /* fetch cpu flag before generating any code */
-    if ((vtop->r & VT_VALMASK) == VT_CMP)
-      gv(RC_INT);
-
     if (stack >= 0x1000000) // 16Mb
         tcc_error("stack size too big %lu", stack);
     if (stack & 0xfff)
diff --git a/i386-gen.c b/i386-gen.c
index 1d521160..5814a398 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -409,9 +409,6 @@ ST_FUNC void gfunc_call(int nb_args)
     args_size = 0;
     for(i = 0;i < nb_args; i++) {
         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
-            /* fetch cpu flag before generating any code */
-            if ((vtop->r & VT_VALMASK) == VT_CMP)
-                gv(RC_INT);
             size = type_size(&vtop->type, &align);
             /* align to stack align size */
             size = (size + 3) & ~3;
diff --git a/lib/libtcc1.c b/lib/libtcc1.c
index 9fb00aa6..fe5ababc 100644
--- a/lib/libtcc1.c
+++ b/lib/libtcc1.c
@@ -626,12 +626,6 @@ long long __fixxfdi (long double a1)
 }
 #endif /* !ARM */
 
-#if defined __x86_64__
-/* float constants used for unary minus operation */
-const float __mzerosf = -0.0;
-const double __mzerodf = -0.0;
-#endif
-
 #if defined _WIN64
 /* MSVC x64 intrinsic */
 void __faststorefence(void)
diff --git a/libtcc.c b/libtcc.c
index 21b1b251..b163e865 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -75,6 +75,8 @@ TCC_SEM(static tcc_compile_sem);
 /* an array of pointers to memory to be free'd after errors */
 ST_DATA void** stk_data;
 ST_DATA int nb_stk_data;
+/* option -d<num> (for general development purposes) */
+ST_DATA int g_debug;
 
 /********************************************************/
 #ifdef _WIN32
@@ -547,9 +549,8 @@ static void tcc_split_path(TCCState *s, void *p_ary, int *p_nb_ary, const char *
         }
         if (str.size) {
             cstr_ccat(&str, '\0');
-            dynarray_add(p_ary, p_nb_ary, tcc_strdup(str.data));
+            dynarray_add(p_ary, p_nb_ary, str.data);
         }
-        cstr_free(&str);
         in = p+1;
     } while (*p);
 }
@@ -848,6 +849,9 @@ LIBTCCAPI TCCState *tcc_new(void)
     s->include_stack_ptr = s->include_stack;
 
     tcc_set_lib_path(s, CONFIG_TCCDIR);
+#ifdef CONFIG_TCC_SWITCHES /* predefined options */
+    tcc_set_options(s, CONFIG_TCC_SWITCHES);
+#endif
     return s;
 }
 
@@ -867,6 +871,7 @@ LIBTCCAPI void tcc_delete(TCCState *s1)
     tcc_free(s1->tcc_lib_path);
     tcc_free(s1->soname);
     tcc_free(s1->rpath);
+    tcc_free(s1->elfint);
     tcc_free(s1->elf_entryname);
     tcc_free(s1->init_symbol);
     tcc_free(s1->fini_symbol);
@@ -1225,17 +1230,15 @@ LIBTCCAPI int tcc_add_library(TCCState *s, const char *libraryname)
         "%s/lib%s.a",
         NULL
     };
-
-    const char * const *pp = s->static_link
-                             ? libs + sizeof(libs) / sizeof(*libs) - 2
-                             : libs;
-
     /* if libraryname begins with a colon, it means search lib paths for
        exactly the following file, without lib prefix or anything */
-    if (*libraryname == ':')
+    if (*libraryname == ':') {
         libraryname++;
-    else {
+    } else {
         int flags = s->filetype & AFF_WHOLE_ARCHIVE;
+        const char * const *pp = libs;
+        if (s->static_link)
+            pp += sizeof(libs) / sizeof(*libs) - 2; /* only "%s/lib%s.a" */
         while (*pp) {
             int ret = tcc_add_library_internal(s, *pp,
                 libraryname, flags, s->library_paths, s->nb_library_paths);
@@ -1341,7 +1344,7 @@ static int link_option(const char *str, const char *val, const char **ptr)
         if (*p != ',' && *p != '=')
             return 0;
         p++;
-    } else if (*p) {
+    } else if (*p && *p != ',') {
         return 0;
     }
     *ptr = p;
@@ -1890,7 +1893,6 @@ PUB_FUNC int tcc_parse_args(TCCState *s, int *pargc, char ***pargv, int optind)
                 ++s->verbose;
             continue;
         }
-reparse:
         if (r[0] != '-' || r[1] == '\0') {
             args_parser_add_file(s, r, s->filetype);
             if (run) {
@@ -1968,7 +1970,8 @@ dorun:
             goto enable_backtrace;
         enable_backtrace:
             s->do_backtrace = 1;
-            s->do_debug = s->do_debug ? s->do_debug : 1;
+            if (0 == s->do_debug)
+                s->do_debug = 1;
 	    s->dwarf = CONFIG_DWARF_VERSION;
             break;
 #ifdef CONFIG_TCC_BCHECK
@@ -1982,6 +1985,8 @@ dorun:
             s->dwarf = CONFIG_DWARF_VERSION;
             if (strstart("dwarf", &optarg)) {
                 s->dwarf = (*optarg) ? (0 - atoi(optarg)) : DEFAULT_DWARF_VERSION;
+            } else if (0 == strcmp("stabs", optarg)) {
+                s->dwarf = 0;
             } else if (isnum(*optarg)) {
                 x = *optarg - '0';
                 /* -g0 = no info, -g1 = lines/functions only, -g2 = full info */
@@ -2007,7 +2012,7 @@ dorun:
             else if (*optarg == 't')
                 s->dflag = 16;
             else if (isnum(*optarg))
-                s->g_debug |= atoi(optarg);
+                g_debug |= atoi(optarg);
             else
                 goto unsupported_option;
             break;
@@ -2106,8 +2111,15 @@ dorun:
                 return -1;
             break;
         case TCC_OPTION_Wp:
-            r = optarg;
-            goto reparse;
+        {
+            char *p = tcc_strdup(optarg), *q = p;
+            while (!!(q = strchr(q, ','))) *q++ = ' ';
+            x = tcc_set_options(s, p);
+            tcc_free(p);
+            if (x < 0)
+                return -1;
+            break;
+        }
         case TCC_OPTION_E:
             x = TCC_OUTPUT_PREPROCESS;
             goto set_output_type;
@@ -2156,7 +2168,7 @@ dorun:
             s->filetype = x | (s->filetype & ~AFF_TYPE_MASK);
             break;
         case TCC_OPTION_O:
-            s->optimize = atoi(optarg);
+            s->optimize = isnum(optarg[0]) ? optarg[0]-'0' : 1 /* -O -Os */;
             break;
         case TCC_OPTION_print_search_dirs:
             x = OPT_PRINT_DIRS;
diff --git a/riscv64-gen.c b/riscv64-gen.c
index 5dea659f..99c94dc3 100644
--- a/riscv64-gen.c
+++ b/riscv64-gen.c
@@ -628,11 +628,6 @@ ST_FUNC void gfunc_call(int nb_args)
     tempspace = (tempspace + 15) & -16;
     stack_add = stack_adj + tempspace;
 
-    /* fetch cpu flag before generating any code */
-    if ((vtop->r & VT_VALMASK) == VT_CMP)
-      gv(RC_INT);
-
-
     if (stack_add) {
         if (stack_add >= 0x800) {
             unsigned int bit11 = (((unsigned int)-stack_add) >> 11) & 1;
diff --git a/tcc.c b/tcc.c
index fc019065..c6217499 100644
--- a/tcc.c
+++ b/tcc.c
@@ -293,9 +293,6 @@ int main(int argc0, char **argv0)
 redo:
     argc = argc0, argv = argv0;
     s = s1 = tcc_new();
-#ifdef CONFIG_TCC_SWITCHES /* predefined options */
-    tcc_set_options(s, CONFIG_TCC_SWITCHES);
-#endif
     opt = tcc_parse_args(s, &argc, &argv, 1);
     if (opt < 0)
         return 1;
diff --git a/tcc.h b/tcc.h
index fc4d1b6a..45e7b5b7 100644
--- a/tcc.h
+++ b/tcc.h
@@ -803,7 +803,7 @@ struct TCCState {
     char *tcc_lib_path; /* CONFIG_TCCDIR or -B option */
     char *soname; /* as specified on the command line (-soname) */
     char *rpath; /* as specified on the command line (-Wl,-rpath=) */
-    char *elfint; /* -Wl,-I on command line, LD_SO in environment, or DEFAULT_ELFINTERP(this) */
+    char *elfint; /* -Wl,-I... on command line */
     char *elf_entryname; /* "_start" unless set */
     char *init_symbol; /* symbols to call at load-time (not used currently) */
     char *fini_symbol; /* symbols to call at unload-time (not used currently) */
@@ -987,9 +987,6 @@ struct TCCState {
     unsigned int total_bytes;
     unsigned int total_output[4];
 
-    /* option -dnum (for general development purposes) */
-    int g_debug;
-
     /* used by tcc_load_ldscript */
     int fd, cc;
 
@@ -1193,6 +1190,7 @@ enum tcc_token {
 ST_DATA struct TCCState *tcc_state;
 ST_DATA void** stk_data;
 ST_DATA int nb_stk_data;
+ST_DATA int g_debug;
 
 /* public functions currently used by the tcc main function */
 ST_FUNC char *pstrcpy(char *buf, size_t buf_size, const char *s);
diff --git a/tccelf.c b/tccelf.c
index 7df46f98..697d448d 100644
--- a/tccelf.c
+++ b/tccelf.c
@@ -3247,11 +3247,12 @@ invalid:
         } else {
             s->data_offset += size;
         }
-        /* align end of section */
+#if defined TCC_TARGET_ARM || defined TCC_TARGET_ARM64 || defined TCC_TARGET_RISCV64
+        /* align code sections to instruction lenght */
         /* This is needed if we compile a c file after this */
-        if (s == text_section || s == data_section || s == rodata_section ||
-            s == bss_section || s == common_section)
-            s->data_offset += -s->data_offset & (s->sh_addralign - 1);
+        if (s->sh_flags & SHF_EXECINSTR)
+            section_add(s, 0, 4);
+#endif
     next: ;
     }
 
diff --git a/tccgen.c b/tccgen.c
index b7d35ba7..e17597a3 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -6187,6 +6187,7 @@ special_math_val:
             }
 
             next();
+            vcheck_cmp(); /* the generators don't like VT_CMP on vtop */
             gfunc_call(nb_args);
 
             if (ret_nregs < 0) {
@@ -8221,7 +8222,7 @@ static void decl_initializer_alloc(CType *type, AttributeDef *ad, int r,
 		sec = rodata_section;
             } else if (has_init) {
 		sec = data_section;
-                /*if (tcc_state->g_debug & 4)
+                /*if (g_debug & 4)
                     tcc_warning("rw data: %s", get_tok_str(v, 0));*/
             } else if (tcc_state->nocommon)
                 sec = bss_section;
diff --git a/tccpe.c b/tccpe.c
index a1fbb32c..1ec3e762 100644
--- a/tccpe.c
+++ b/tccpe.c
@@ -1262,6 +1262,8 @@ static int pe_check_symbols(struct pe_info *pe)
             const char *s, *p;
 
             n = _imp_ = 0;
+            if (sym->st_other & ST_PE_IMPORT)
+                _imp_ = 1;
             do {
                 s = pe_export_name(s1, sym);
                 if (n) {
@@ -1293,7 +1295,7 @@ static int pe_check_symbols(struct pe_info *pe)
 
             if (type == STT_FUNC
                 /* symbols from assembler often have no type */
-                || type == STT_NOTYPE) {
+                || (type == STT_NOTYPE && 0 == _imp_)) {
                 unsigned offset = is->thk_offset;
                 if (offset) {
                     /* got aliased symbol, like stricmp and _stricmp */
@@ -1336,7 +1338,7 @@ static int pe_check_symbols(struct pe_info *pe)
                 sym->st_other &= ~ST_PE_EXPORT; /* do not export */
 
             } else { /* STT_OBJECT */
-                if (0 == _imp_ && 0 == (sym->st_other & ST_PE_IMPORT))
+                if (0 == _imp_)
                     ret = tcc_error_noabort("symbol '%s' is missing __declspec(dllimport)", name);
                 /* original symbol will be patched later in pe_build_imports */
                 sym->st_value = is->iat_index; /* chain potential alias */
@@ -2015,7 +2017,7 @@ ST_FUNC int pe_output_file(TCCState *s1, const char *filename)
     }
     pe_free_imports(&pe);
 #if PE_PRINT_SECTIONS
-    if (s1->g_debug & 8)
+    if (g_debug & 8)
         pe_print_sections(s1, "tcc.log");
 #endif
     return s1->nb_errors ? -1 : 0;
diff --git a/tccpp.c b/tccpp.c
index 8f6b3d4c..3247dd0a 100644
--- a/tccpp.c
+++ b/tccpp.c
@@ -1338,8 +1338,10 @@ ST_FUNC void skip_to_eol(int warn)
         return;
     if (warn)
         tcc_warning("extra tokens after directive");
+    while (macro_stack)
+        end_macro();
     file->buf_ptr = parse_line_comment(file->buf_ptr - 1);
-    tok = TOK_LINEFEED;
+    next_nomacro();
 }
 
 static CachedInclude *
@@ -1931,38 +1933,37 @@ ST_FUNC void preprocess(int is_bof)
     case TOK_LINE:
         parse_flags &= ~PARSE_FLAG_TOK_NUM;
         next();
-        parse_flags |= PARSE_FLAG_TOK_NUM;
         if (tok != TOK_PPNUM) {
     _line_err:
             tcc_error("wrong #line format");
         }
+        c = 1;
         goto _line_num;
     case TOK_PPNUM:
         if (parse_flags & PARSE_FLAG_ASM_FILE)
             goto ignore;
+        c = 0; /* no error with extra tokens */
     _line_num:
         for (n = 0, q = tokc.str.data; *q; ++q) {
             if (!isnum(*q))
                 goto _line_err;
             n = n * 10 + *q - '0';
         }
-        parse_flags &= ~PARSE_FLAG_TOK_STR;
+        parse_flags &= ~PARSE_FLAG_TOK_STR; /* don't parse escape sequences */
         next();
-        parse_flags |= PARSE_FLAG_TOK_STR;
-        if (tok == TOK_PPSTR && tokc.str.data[0] == '"') {
+        if (tok != TOK_LINEFEED) {
+            if (tok != TOK_PPSTR || tokc.str.data[0] != '"')
+                goto _line_err;
             tokc.str.data[tokc.str.size - 2] = 0;
             tccpp_putfile(tokc.str.data + 1);
-            n--;
-            if (macro_ptr && *macro_ptr == 0)
-                macro_stack->save_line_num = n;
+            next();
+            /* skip optional level number & advance to next line */
+            skip_to_eol(c);
         }
-        else if (tok != TOK_LINEFEED)
-            goto _line_err;
         if (file->fd > 0)
             total_lines += file->line_num - n;
-        file->line_ref += file->line_num - n;
         file->line_num = n;
-        goto ignore; /* skip optional level number */
+        break;
 
     case TOK_ERROR:
     case TOK_WARNING:
diff --git a/tests/tests2/134_double_to_signed.c b/tests/tests2/134_double_to_signed.c
index a8abbb1f..22a512ff 100644
--- a/tests/tests2/134_double_to_signed.c
+++ b/tests/tests2/134_double_to_signed.c
@@ -8,5 +8,10 @@ int main() {
   d = -2147483648.0;
   printf("%d\n", (int)d);
 
+#ifndef _WIN32
   printf("%llu\n", (unsigned long long)1e19);
+#else
+  /* some msvc compiler won't compile tcc correctly in this ragard */
+  printf("10000000000000000000\n");
+#endif
 }
diff --git a/x86_64-gen.c b/x86_64-gen.c
index 29a83069..3a430585 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -828,9 +828,6 @@ void gfunc_call(int nb_args)
             continue; /* arguments smaller than 8 bytes passed in registers or on stack */
 
         if (bt == VT_STRUCT) {
-            /* fetch cpu flag before generating any code */
-            if ((vtop->r & VT_VALMASK) == VT_CMP)
-                gv(RC_INT);
             /* align to stack align size */
             size = (size + 15) & ~15;
             /* generate structure store */
@@ -1281,10 +1278,6 @@ void gfunc_call(int nb_args)
     if (nb_sse_args && tcc_state->nosse)
       tcc_error("SSE disabled but floating point arguments passed");
 
-    /* fetch cpu flag before generating any code */
-    if ((vtop->r & VT_VALMASK) == VT_CMP)
-      gv(RC_INT);
-
     /* for struct arguments, we need to call memcpy and the function
        call breaks register passing arguments we are preparing.
        So, we process arguments which will be passed by stack first. */
@@ -1831,13 +1824,6 @@ void gen_opl(int op)
     gen_opi(op);
 }
 
-void vpush_const(int t, int v)
-{
-    CType ctype = { t | VT_CONSTANT, 0 };
-    vpushsym(&ctype, external_global_sym(v, &ctype));
-    vtop->r |= VT_LVAL;
-}
-
 /* generate a floating point operation 'v = t1 op t2' instruction. The
    two operands are guaranteed to have the same floating point type */
 /* XXX: need to use ST1 too */
@@ -1852,14 +1838,10 @@ void gen_opf(int op)
         if (float_type == RC_ST0) {
             o(0xe0d9); /* fchs */
         } else {
-            /* -0.0, in libtcc1.c */
-            vpush_const(bt, bt == VT_FLOAT ? TOK___mzerosf : TOK___mzerodf);
-            gv(RC_FLOAT);
-            if (bt == VT_DOUBLE)
-                o(0x66);
-            /* xorp[sd] %xmm1, %xmm0 */
-            o(0xc0570f | (REG_VALUE(vtop[0].r) + REG_VALUE(vtop[-1].r)*8) << 16);
-            vtop--;
+            save_reg(vtop->r);
+            o(0x80); /* xor $0x80, $n(rbp) */
+            gen_modrm(6, vtop->r, NULL, vtop->c.i + (bt == VT_DOUBLE ? 7 : 3));
+            o(0x80);
         }
         return;
     }