i386-gen: preserve fp control word in gen_cvt_ftoi

- Use runtime function for conversion
- Also initialize fp with tcc -run on windows

This fixes a bug where
  double x = 1.0;
  double y = 1.0000000000000001;
  double z = x < y ? 0 : sqrt (x*x - y*y);
caused a bad sqrt because rounding precision for the x < y comparison
was different to the one used within the sqrt function.

This also fixes a bug where
  printf("%d, %d", (int)pow(10, 2), (int)pow(10, 2));
would print
  100, 99

Unrelated:
  win32: document relative include & lib lookup
  win32: normalize_slashes: do not mirror silly gcc behavior
  This reverts part of commit 8a81f9e1036637e21a47e14fb56bf64133546890
  winapi: add missing WINAPI decl. for some functions

11 files changed, 77 insertions, 107 deletions

diff --git a/i386-gen.c b/i386-gen.c
index 844a482..0a6d4d3 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -338,6 +338,15 @@ static void gadd_sp(int val)
     }
 }
 
+static void gen_static_call(int v)
+{
+    Sym *sym;
+
+    sym = external_global_sym(v, &func_old_type, 0);
+    oad(0xe8, -4);
+    greloc(cur_text_section, sym, ind-4, R_386_PC32);
+}
+
 /* 'is_jmp' is '1' if it is a jump */
 static void gcall_or_jmp(int is_jmp)
 {
@@ -570,6 +579,12 @@ ST_FUNC void gfunc_prolog(CType *func_type)
         func_bound_offset = lbounds_section->data_offset;
     }
 #endif
+
+#ifndef TCC_TARGET_PE
+    if (0 == strcmp(funcname, "main"))
+        gen_static_call(TOK___tcc_fpinit);
+#endif
+
 }
 
 /* generate function epilog */
@@ -582,7 +597,7 @@ ST_FUNC void gfunc_epilog(void)
      && func_bound_offset != lbounds_section->data_offset) {
         int saved_ind;
         int *bounds_ptr;
-        Sym *sym, *sym_data;
+        Sym *sym_data;
         /* add end of table info */
         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
         *bounds_ptr = 0;
@@ -594,20 +609,16 @@ ST_FUNC void gfunc_epilog(void)
         greloc(cur_text_section, sym_data,
                ind + 1, R_386_32);
         oad(0xb8, 0); /* mov %eax, xxx */
-        sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
-        greloc(cur_text_section, sym, 
-               ind + 1, R_386_PC32);
-        oad(0xe8, -4);
+        gen_static_call(TOK___bound_local_new);
+
         ind = saved_ind;
         /* generate bound check local freeing */
         o(0x5250); /* save returned value, if any */
         greloc(cur_text_section, sym_data,
                ind + 1, R_386_32);
         oad(0xb8, 0); /* mov %eax, xxx */
-        sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
-        greloc(cur_text_section, sym, 
-               ind + 1, R_386_PC32);
-        oad(0xe8, -4);
+        gen_static_call(TOK___bound_local_delete);
+
         o(0x585a); /* restore returned value, if any */
     }
 #endif
@@ -626,10 +637,8 @@ ST_FUNC void gfunc_epilog(void)
     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 #ifdef TCC_TARGET_PE
     if (v >= 4096) {
-        Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
         oad(0xb8, v); /* mov stacksize, %eax */
-        oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
-        greloc(cur_text_section, sym, ind-4, R_386_PC32);
+        gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
     } else
 #endif
     {
@@ -992,52 +1001,13 @@ ST_FUNC void gen_cvt_itof(int t)
 /* XXX: handle long long case */
 ST_FUNC void gen_cvt_ftoi(int t)
 {
-    int r, r2, size;
-    Sym *sym;
-    CType ushort_type;
-
-    ushort_type.t = VT_SHORT | VT_UNSIGNED;
-    ushort_type.ref = 0;
-
     gv(RC_FLOAT);
-    if (t != VT_INT)
-        size = 8;
-    else 
-        size = 4;
-    
-    o(0x2dd9); /* ldcw xxx */
-    sym = external_global_sym(TOK___tcc_int_fpu_control, 
-                              &ushort_type, VT_LVAL);
-    greloc(cur_text_section, sym, 
-           ind, R_386_32);
-    gen_le32(0);
-    
-    oad(0xec81, size); /* sub $xxx, %esp */
-    if (size == 4)
-        o(0x1cdb); /* fistpl */
-    else
-        o(0x3cdf); /* fistpll */
-    o(0x24);
-    o(0x2dd9); /* ldcw xxx */
-    sym = external_global_sym(TOK___tcc_fpu_control, 
-                              &ushort_type, VT_LVAL);
-    greloc(cur_text_section, sym, 
-           ind, R_386_32);
-    gen_le32(0);
-
-    r = get_reg(RC_INT);
-    o(0x58 + r); /* pop r */
-    if (size == 8) {
-        if (t == VT_LLONG) {
-            vtop->r = r; /* mark reg as used */
-            r2 = get_reg(RC_INT);
-            o(0x58 + r2); /* pop r2 */
-            vtop->r2 = r2;
-        } else {
-            o(0x04c483); /* add $4, %esp */
-        }
-    }
-    vtop->r = r;
+    save_reg(TREG_EAX);
+    save_reg(TREG_EDX);
+    gen_static_call(TOK___tcc_cvt_ftol);
+    vtop->r = TREG_EAX; /* mark reg as used */
+    if (t == VT_LLONG)
+        vtop->r2 = TREG_EDX;
 }
 
 /* convert from one floating point type to another */
@@ -1060,18 +1030,13 @@ ST_FUNC void ggoto(void)
 /* generate a bounded pointer addition */
 ST_FUNC void gen_bounded_ptr_add(void)
 {
-    Sym *sym;
-
     /* prepare fast i386 function call (args in eax and edx) */
     gv2(RC_EAX, RC_EDX);
     /* save all temporary registers */
     vtop -= 2;
     save_regs(0);
     /* do a fast function call */
-    sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
-    greloc(cur_text_section, sym, 
-           ind + 1, R_386_PC32);
-    oad(0xe8, -4);
+    gen_static_call(TOK___bound_ptr_add);
     /* returned pointer is in eax */
     vtop++;
     vtop->r = TREG_EAX | VT_BOUNDED;
diff --git a/lib/libtcc1.c b/lib/libtcc1.c
index 53dbec4..a94a82d 100644
--- a/lib/libtcc1.c
+++ b/lib/libtcc1.c
@@ -478,13 +478,24 @@ long long __ashldi3(long long a, int b)
 #endif
 }
 
-#if defined(__i386__)
-/* FPU control word for rounding to nearest mode */
-unsigned short __tcc_fpu_control = 0x137f;
-/* FPU control word for round to zero mode for int conversion */
-unsigned short __tcc_int_fpu_control = 0x137f | 0x0c00;
+#ifndef _WIN32
+void __tcc_fpinit(void)
+{
+    unsigned c = 0x137F;
+    __asm__ __volatile__ ("fldcw %0" : "=m" (c));
+}
 #endif
-
+long long __tcc_cvt_ftol(long double x)
+{
+    unsigned c0, c1;
+    long long ret;
+    __asm__ __volatile__ ("fnstcw %0" : "=m" (c0));
+    c1 = c0 | 0x0C00;
+    __asm__ __volatile__ ("fldcw %0" : "=m" (c1));
+    __asm__ __volatile__ ("fistpll %0"  : "=m" (ret));
+    __asm__ __volatile__ ("fldcw %0" : "=m" (c0));
+    return ret;
+}
 #endif /* !__x86_64__ */
 
 /* XXX: fix tcc's code generator to do this instead */
diff --git a/libtcc.c b/libtcc.c
index b10a2a7..3b018ae 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -78,21 +78,13 @@ ST_FUNC void asm_global_instr(void)
 #endif
 
 /********************************************************/
-
 #ifdef _WIN32
-// GCC appears to use '/' for relative paths and '\\' for absolute paths on Windows
 static char *normalize_slashes(char *path)
 {
     char *p;
-    if (path[1] == ':') {
-        for (p = path+2; *p; ++p)
-            if (*p == '/')
-                *p = '\\';
-    } else {
-        for (p = path; *p; ++p)
-            if (*p == '\\')
-                *p = '/';
-    }
+    for (p = path; *p; ++p)
+        if (*p == '\\')
+            *p = '/';
     return path;
 }
 
@@ -1036,6 +1028,7 @@ LIBTCCAPI TCCState *tcc_new(void)
 #ifdef TCC_TARGET_I386
     s->seg_size = 32;
 #endif
+    s->runtime_main = "main";
     return s;
 }
 
diff --git a/tcc.h b/tcc.h
index 859d4fd..c988e4f 100644
--- a/tcc.h
+++ b/tcc.h
@@ -665,6 +665,7 @@ struct TCCState {
 #endif
 
 #ifdef TCC_IS_NATIVE
+    const char *runtime_main;
     /* for tcc_relocate */
     void *runtime_mem;
 # ifdef HAVE_SELINUX
diff --git a/tccpe.c b/tccpe.c
index 72c1572..05fed09 100644
--- a/tccpe.c
+++ b/tccpe.c
@@ -1726,7 +1726,6 @@ ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
 static void pe_add_runtime(TCCState *s1, struct pe_info *pe)
 {
     const char *start_symbol;
-    ADDR3264 addr = 0;
     int pe_type = 0;
 
     if (find_elf_sym(symtab_section, PE_STDSYM("WinMain","@16")))
@@ -1742,16 +1741,13 @@ static void pe_add_runtime(TCCState *s1, struct pe_info *pe)
 
     start_symbol =
         TCC_OUTPUT_MEMORY == s1->output_type
-        ? PE_GUI == pe_type ? "__runwinmain" : "_main"
+        ? PE_GUI == pe_type ? "__runwinmain" : "__runmain"
         : PE_DLL == pe_type ? PE_STDSYM("__dllstart","@12")
         : PE_GUI == pe_type ? "__winstart" : "__start"
         ;
 
-    if (!s1->leading_underscore || strchr(start_symbol, '@')) {
+    if (!s1->leading_underscore || strchr(start_symbol, '@'))
         ++start_symbol;
-        if (start_symbol[0] != '_')
-            start_symbol = NULL;
-    }
 
     /* grab the startup code from libtcc1 */
     if (start_symbol)
@@ -1776,21 +1772,13 @@ static void pe_add_runtime(TCCState *s1, struct pe_info *pe)
         }
     }
 
-    if (TCC_OUTPUT_MEMORY == s1->output_type)
+    if (TCC_OUTPUT_MEMORY == s1->output_type) {
         pe_type = PE_RUN;
-
-    if (start_symbol) {
-        addr = get_elf_sym_addr(s1, start_symbol, 1);
-        if (PE_RUN == pe_type && addr)
-            /* for -run GUI's, put '_runwinmain' instead of 'main' */
-            add_elf_sym(symtab_section,
-                    addr, 0,
-                    ELFW(ST_INFO)(STB_GLOBAL, STT_NOTYPE), 0,
-                    text_section->sh_num, "main");
+        s1->runtime_main = start_symbol;
     }
 
     pe->type = pe_type;
-    pe->start_addr = addr;
+    pe->start_addr = (DWORD)tcc_get_symbol_err(s1, start_symbol);
 }
 
 ST_FUNC int pe_output_file(TCCState * s1, const char *filename)
diff --git a/tccrun.c b/tccrun.c
index d858ae6..b07ab0f 100644
--- a/tccrun.c
+++ b/tccrun.c
@@ -97,7 +97,7 @@ LIBTCCAPI int tcc_run(TCCState *s1, int argc, char **argv)
     if (tcc_relocate(s1, TCC_RELOCATE_AUTO) < 0)
         return -1;
 
-    prog_main = tcc_get_symbol_err(s1, "main");
+    prog_main = tcc_get_symbol_err(s1, s1->runtime_main);
 
 #ifdef CONFIG_TCC_BACKTRACE
     if (s1->do_debug) {
diff --git a/tcctok.h b/tcctok.h
index fde13dd..9b47a60 100644
--- a/tcctok.h
+++ b/tcctok.h
@@ -194,8 +194,8 @@
      DEF(TOK__remu, "_remu")
 #endif
 #ifdef TCC_TARGET_I386
-     DEF(TOK___tcc_int_fpu_control, "__tcc_int_fpu_control")
-     DEF(TOK___tcc_fpu_control, "__tcc_fpu_control")
+     DEF(TOK___tcc_fpinit, "__tcc_fpinit")
+     DEF(TOK___tcc_cvt_ftol, "__tcc_cvt_ftol")
 #endif
 #ifdef TCC_ARM_EABI
      DEF(TOK___ashrdi3, "__aeabi_lasr")
diff --git a/win32/include/winapi/winbase.h b/win32/include/winapi/winbase.h
index a5d9b17..4a38006 100644
--- a/win32/include/winapi/winbase.h
+++ b/win32/include/winapi/winbase.h
@@ -968,15 +968,15 @@ extern "C" {
   LONG64 InterlockedExchangeAdd64(LONG64 volatile *Addend,LONG64 Value);
   LONG64 InterlockedCompareExchange64(LONG64 volatile *Destination,LONG64 ExChange,LONG64 Comperand);
 #else
-  LONG InterlockedIncrement(LONG volatile *lpAddend);
-  LONG InterlockedDecrement(LONG volatile *lpAddend);
-  LONG InterlockedExchange(LONG volatile *Target,LONG Value);
+  LONG WINAPI InterlockedIncrement(LONG volatile *lpAddend);
+  LONG WINAPI InterlockedDecrement(LONG volatile *lpAddend);
+  LONG WINAPI InterlockedExchange(LONG volatile *Target,LONG Value);
 
 #define InterlockedExchangePointer(Target,Value) (PVOID)InterlockedExchange((PLONG)(Target),(LONG)(Value))
 
-  LONG InterlockedExchangeAdd(LONG volatile *Addend,LONG Value);
-  LONG InterlockedCompareExchange(LONG volatile *Destination,LONG Exchange,LONG Comperand);
-  LONGLONG InterlockedCompareExchange64(LONGLONG volatile *Destination,LONGLONG Exchange,LONGLONG Comperand);
+  LONG WINAPI InterlockedExchangeAdd(LONG volatile *Addend,LONG Value);
+  LONG WINAPI InterlockedCompareExchange(LONG volatile *Destination,LONG Exchange,LONG Comperand);
+  LONGLONG WINAPI InterlockedCompareExchange64(LONGLONG volatile *Destination,LONGLONG Exchange,LONGLONG Comperand);
 
   __CRT_INLINE LONGLONG InterlockedAnd64 (LONGLONG volatile *Destination,LONGLONG Value) {
     LONGLONG Old;
diff --git a/win32/lib/crt1.c b/win32/lib/crt1.c
index 3e1d17f..cde3910 100644
--- a/win32/lib/crt1.c
+++ b/win32/lib/crt1.c
@@ -31,4 +31,10 @@ int _start(void)
     exit(ret);
 }
 
+int _runmain(int argc, char **argv)
+{
+    _controlfp(0x10000, 0x30000);
+    return main(argc, argv, NULL);
+}
+
 // =============================================
diff --git a/win32/lib/wincrt1.c b/win32/lib/wincrt1.c
index 77e74b8..663fd33 100644
--- a/win32/lib/wincrt1.c
+++ b/win32/lib/wincrt1.c
@@ -59,6 +59,6 @@ int _runwinmain(int argc, char **argv)
         szCmd = "";
     else if (szCmd > p && szCmd[-1] == '\"')
         --szCmd;
+    _controlfp(0x10000, 0x30000);
     return WinMain(GetModuleHandle(NULL), NULL, szCmd, SW_SHOWDEFAULT);
 }
-
diff --git a/win32/tcc-win32.txt b/win32/tcc-win32.txt
index dc06b8f..1cb35c5 100644
--- a/win32/tcc-win32.txt
+++ b/win32/tcc-win32.txt
@@ -18,6 +18,12 @@
     system PATH.
 
 
+    Include and library search paths
+    --------------------------------
+    On windows,  the standard "include" and "lib" directories are searched
+    relatively from the location of the executables (tcc.exe, libtcc.dll).
+
+
     Examples:
     ---------
     Open a console window (DOS box) and 'cd' to the examples directory.