Runtime lib functions

Yesterday I felt the urge to change a few things in TinyCC.
This is the first and biggest change of all of them.

- use __aeabi_*divmod functions in ARM EABI to make binaries depend
  solely on standardized library functions

- refactor ARM floating point <-> integer conversion a bit

- rename long long->float and shift library functions to correspond to
  the names used by GCC

- compile more tokens conditionally to reduce the size of TinyCC

The intention is primarily to allow users of the ARM target to use
libgcc (which is usually available as a shared library) instead of
libtcc1 (which can't be compiled for ARM due to lack of an inline
assembler).

Changing the EABI target to use the divmod functions in theory allows
to use it without libtcc1 on any (not necessarily GCC based) ARM EABI
system.

  Daniel

4 files changed, 140 insertions, 94 deletions

diff --git a/arm-gen.c b/arm-gen.c
index e3a5629..67a6f85 100644
--- a/arm-gen.c
+++ b/arm-gen.c
@@ -110,6 +110,13 @@ static int regmask(int r) {
 #define REG_LRET TREG_R1 /* second word return register (for long long) */
 #define REG_FRET TREG_F0 /* float return register */
 
+#ifdef TCC_ARM_EABI
+#define TOK___divdi3 TOK___aeabi_ldivmod
+#define TOK___moddi3 TOK___aeabi_ldivmod
+#define TOK___udivdi3 TOK___aeabi_uldivmod
+#define TOK___umoddi3 TOK___aeabi_uldivmod
+#endif
+
 /* defined if function parameters must be evaluated in reverse order */
 #define INVERT_FUNC_PARAMS
 
@@ -119,6 +126,11 @@ static int regmask(int r) {
 
 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
 static CType float_type, double_type, func_float_type, func_double_type;
+#define func_ldouble_type func_double_type
+#else
+#define func_float_type func_old_type
+#define func_double_type func_old_type
+#define func_ldouble_type func_old_type
 #endif
 
 /* pointer size, in bytes */
@@ -1083,6 +1095,7 @@ void gen_opi(int op)
 {
   int c, func = 0;
   unsigned long opc = 0,r,fr;
+  unsigned short retreg = REG_IRET;
 
   c=0;
   switch(op) {
@@ -1151,11 +1164,21 @@ void gen_opi(int op)
       c=3;
       break;
     case '%':
+#ifdef TCC_ARM_EABI
+      func=TOK___aeabi_idivmod;
+      retreg=REG_LRET;
+#else
       func=TOK___modsi3;
+#endif
       c=3;
       break;
     case TOK_UMOD:
+#ifdef TCC_ARM_EABI
+      func=TOK___aeabi_uidivmod;
+      retreg=REG_LRET;
+#else
       func=TOK___umodsi3;
+#endif
       c=3;
       break;
     case TOK_UMULL:
@@ -1230,7 +1253,7 @@ done:
       vrott(3);
       gfunc_call(2);
       vpushi(0);
-      vtop->r = REG_IRET;
+      vtop->r = retreg;
       break;
     default:
       error("gen_opi %i unimplemented!",op);
@@ -1547,11 +1570,14 @@ void gen_opf(int op)
 
 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
    and 'long long' cases. */
-void gen_cvt_itof(int t)
+void gen_cvt_itof1(int t)
 {
   int r,r2,bt;
   bt=vtop->type.t & VT_BTYPE;
   if(bt == VT_INT || bt == VT_SHORT || bt == VT_BYTE) {
+#ifndef TCC_ARM_VFP
+    unsigned int dsize=0;
+#endif
     r=intr(gv(RC_INT));
 #ifdef TCC_ARM_VFP
     r2=vfpr(vtop->r=get_reg(RC_FLOAT));
@@ -1562,10 +1588,12 @@ void gen_cvt_itof(int t)
     o(0xEEB80A40|r2|T2CPR(t)); /* fYitoX*/
 #else
     r2=fpr(vtop->r=get_reg(RC_FLOAT));
-    o(0xEE000190|(r2<<16)|(r<<12));
+    if((t & VT_BTYPE) != VT_FLOAT)
+      dsize=0x80;    /* flts -> fltd */
+    o(0xEE000110|dsize|(r2<<16)|(r<<12)); /* flts */
     if((vtop->type.t & (VT_UNSIGNED|VT_BTYPE)) == (VT_UNSIGNED|VT_INT)) {
       unsigned int off=0;
-      o(0xE3500000|(r<<12));
+      o(0xE3500000|(r<<12));        /* cmp */
       r=fpr(get_reg(RC_FLOAT));
       if(last_itod_magic) {
 	off=ind+8-last_itod_magic;
@@ -1573,44 +1601,50 @@ void gen_cvt_itof(int t)
 	if(off>255)
 	  off=0;
       }
-      o(0xBD1F8100|(r<<12)|off);
+      o(0xBD1F0100|(r<<12)|off);    /* ldflts */
       if(!off) {
-        o(0xEA000001);
+        o(0xEA000000);              /* b */
         last_itod_magic=ind;
-        o(0x41F00000);
-        o(0);
+        o(0x4F800000);              /* 4294967296.0f */
       }
-      o(0xBE000180|(r2<<16)|(r2<<12)|r);
+      o(0xBE000100|dsize|(r2<<16)|(r2<<12)|r); /* adflt */
     }
 #endif
     return;
   } else if(bt == VT_LLONG) {
     int func;
-    CType *func_type = &func_old_type;
-#ifdef TCC_ARM_VFP
-#ifdef TCC_ARM_EABI
-    func_type = &func_double_type;
-#endif
+    CType *func_type = 0;
     if((t & VT_BTYPE) == VT_FLOAT) {
-#ifdef TCC_ARM_EABI
       func_type = &func_float_type;
-#endif
       if(vtop->type.t & VT_UNSIGNED)
-        func=TOK___ulltof;
+        func=TOK___floatundisf;
       else
-        func=TOK___slltof;
-    } else
+        func=TOK___floatdisf;
+#if LDOUBLE_SIZE != 8
+    } else if((t & VT_BTYPE) == VT_LDOUBLE) {
+      func_type = &func_ldouble_type;
+      if(vtop->type.t & VT_UNSIGNED)
+        func=TOK___floatundixf;
+      else
+        func=TOK___floatdixf;
+    } else if((t & VT_BTYPE) == VT_DOUBLE) {
+#else
+    } else if((t & VT_BTYPE) == VT_DOUBLE || (t & VT_BTYPE) == VT_LDOUBLE) {
 #endif
-    if(vtop->type.t & VT_UNSIGNED)
-      func=TOK___ulltold;
-    else
-      func=TOK___slltold;
-    vpush_global_sym(func_type, func);
-    vswap();
-    gfunc_call(1);
-    vpushi(0);
-    vtop->r=TREG_F0;
-    return;
+      func_type = &func_double_type;
+      if(vtop->type.t & VT_UNSIGNED)
+        func=TOK___floatundidf;
+      else
+        func=TOK___floatdidf;
+    }
+    if(func_type) {
+      vpush_global_sym(func_type, func);
+      vswap();
+      gfunc_call(1);
+      vpushi(0);
+      vtop->r=TREG_F0;
+      return;
+    }
   }
   error("unimplemented gen_cvt_itof %x!",vtop->type.t);
 }
@@ -1634,33 +1668,33 @@ void gen_cvt_ftoi(int t)
     if(u) {
       if(r2 == VT_FLOAT)
         func=TOK___fixunssfsi;
-      else if(r2 == VT_DOUBLE)
-	func=TOK___fixunsdfsi;
+#if LDOUBLE_SIZE != 8
       else if(r2 == VT_LDOUBLE)
-#if LDOUBLE_SIZE == 8
-	func=TOK___fixunsdfsi;
-#else
 	func=TOK___fixunsxfsi;
+      else if(r2 == VT_DOUBLE)
+#else
+      else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
 #endif
+	func=TOK___fixunsdfsi;
     } else {
       r=fpr(gv(RC_FLOAT));
       r2=intr(vtop->r=get_reg(RC_INT));
       o(0xEE100170|(r2<<12)|r);
-    return;
+      return;
     }
 #endif
   } else if(t == VT_LLONG) { // unsigned handled in gen_cvt_ftoi1
     if(r2 == VT_FLOAT)
       func=TOK___fixsfdi;
-    else if(r2 == VT_DOUBLE)
-      func=TOK___fixdfdi;
+#if LDOUBLE_SIZE != 8
     else if(r2 == VT_LDOUBLE)
-#if LDOUBLE_SIZE == 8
-      func=TOK___fixdfdi;
-#else
       func=TOK___fixxfdi;
+    else if(r2 == VT_DOUBLE)
+#else
+    else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
 #endif
-    }
+      func=TOK___fixdfdi;
+  }
   if(func) {
     vpush_global_sym(&func_old_type, func);
     vswap();
diff --git a/libtcc1.c b/libtcc1.c
index 11393e3..892baba 100644
--- a/libtcc1.c
+++ b/libtcc1.c
@@ -419,7 +419,7 @@ unsigned long long __umoddi3(unsigned long long u, unsigned long long v)
 }
 
 /* XXX: fix tcc's code generator to do this instead */
-long long __sardi3(long long a, int b)
+long long __ashrdi3(long long a, int b)
 {
 #ifdef __TINYC__
     DWunion u;
@@ -438,7 +438,7 @@ long long __sardi3(long long a, int b)
 }
 
 /* XXX: fix tcc's code generator to do this instead */
-unsigned long long __shrdi3(unsigned long long a, int b)
+unsigned long long __lshrdi3(unsigned long long a, int b)
 {
 #ifdef __TINYC__
     DWunion u;
@@ -457,7 +457,7 @@ unsigned long long __shrdi3(unsigned long long a, int b)
 }
 
 /* XXX: fix tcc's code generator to do this instead */
-long long __shldi3(long long a, int b)
+long long __ashldi3(long long a, int b)
 {
 #ifdef __TINYC__
     DWunion u;
@@ -483,7 +483,7 @@ unsigned short __tcc_int_fpu_control = 0x137f | 0x0c00;
 #endif
 
 /* XXX: fix tcc's code generator to do this instead */
-float __ulltof(unsigned long long a)
+float __floatundisf(unsigned long long a)
 {
     DWunion uu; 
     XFtype r;
@@ -498,7 +498,7 @@ float __ulltof(unsigned long long a)
     }
 }
 
-double __ulltod(unsigned long long a)
+double __floatundidf(unsigned long long a)
 {
     DWunion uu; 
     XFtype r;
@@ -513,7 +513,7 @@ double __ulltod(unsigned long long a)
     }
 }
 
-long double __ulltold(unsigned long long a)
+long double __floatundixf(unsigned long long a)
 {
     DWunion uu; 
     XFtype r;
diff --git a/tcc.c b/tcc.c
index 9855a13..aa34658 100644
--- a/tcc.c
+++ b/tcc.c
@@ -5261,6 +5261,8 @@ void gen_opl(int op)
 {
     int t, a, b, op1, c, i;
     int func;
+    unsigned short reg_iret = REG_IRET;
+    unsigned short reg_lret = REG_LRET;
     SValue tmp;
 
     switch(op) {
@@ -5273,17 +5275,22 @@ void gen_opl(int op)
         goto gen_func;
     case '%':
         func = TOK___moddi3;
-        goto gen_func;
+        goto gen_mod_func;
     case TOK_UMOD:
         func = TOK___umoddi3;
+    gen_mod_func:
+#ifdef TCC_ARM_EABI
+	reg_iret = TREG_R2;
+	reg_lret = TREG_R3;
+#endif
     gen_func:
         /* call generic long long function */
         vpush_global_sym(&func_old_type, func);
         vrott(3);
         gfunc_call(2);
         vpushi(0);
-        vtop->r = REG_IRET;
-        vtop->r2 = REG_LRET;
+        vtop->r = reg_iret;
+        vtop->r2 = reg_lret;
         break;
     case '^':
     case '&':
@@ -5407,13 +5414,13 @@ void gen_opl(int op)
             /* XXX: should provide a faster fallback on x86 ? */
             switch(op) {
             case TOK_SAR:
-                func = TOK___sardi3;
+                func = TOK___ashrdi3;
                 goto gen_func;
             case TOK_SHR:
-                func = TOK___shrdi3;
+                func = TOK___lshrdi3;
                 goto gen_func;
             case TOK_SHL:
-                func = TOK___shldi3;
+                func = TOK___ashldi3;
                 goto gen_func;
             }
         }
@@ -5878,6 +5885,7 @@ void gen_op(int op)
     }
 }
 
+#ifndef TCC_TARGET_ARM
 /* generic itof for unsigned long long case */
 void gen_cvt_itof1(int t)
 {
@@ -5885,11 +5893,13 @@ void gen_cvt_itof1(int t)
         (VT_LLONG | VT_UNSIGNED)) {
 
         if (t == VT_FLOAT)
-            vpush_global_sym(&func_old_type, TOK___ulltof);
-        else if (t == VT_DOUBLE)
-            vpush_global_sym(&func_old_type, TOK___ulltod);
+            vpush_global_sym(&func_old_type, TOK___floatundisf);
+#if LDOUBLE_SIZE != 8
+        else if (t == VT_LDOUBLE)
+            vpush_global_sym(&func_old_type, TOK___floatundixf);
+#endif
         else
-            vpush_global_sym(&func_old_type, TOK___ulltold);
+            vpush_global_sym(&func_old_type, TOK___floatundidf);
         vrott(2);
         gfunc_call(1);
         vpushi(0);
@@ -5898,6 +5908,7 @@ void gen_cvt_itof1(int t)
         gen_cvt_itof(t);
     }
 }
+#endif
 
 /* generic ftoi for unsigned long long case */
 void gen_cvt_ftoi1(int t)
@@ -5909,10 +5920,12 @@ void gen_cvt_ftoi1(int t)
         st = vtop->type.t & VT_BTYPE;
         if (st == VT_FLOAT)
             vpush_global_sym(&func_old_type, TOK___fixunssfdi);
-        else if (st == VT_DOUBLE)
-            vpush_global_sym(&func_old_type, TOK___fixunsdfdi);
-        else
+#if LDOUBLE_SIZE != 8
+        else if (st == VT_LDOUBLE)
             vpush_global_sym(&func_old_type, TOK___fixunsxfdi);
+#endif
+        else
+            vpush_global_sym(&func_old_type, TOK___fixunsdfdi);
         vrott(2);
         gfunc_call(1);
         vpushi(0);
@@ -6020,11 +6033,7 @@ static void gen_cast(CType *type)
                 }
             } else {
             do_itof:
-#if !defined(TCC_TARGET_ARM)
                 gen_cvt_itof1(dbt);
-#else
-                gen_cvt_itof(dbt);
-#endif
             }
         } else if (sf) {
             /* convert fp to int */
diff --git a/tcctok.h b/tcctok.h
index e019476..d717024 100644
--- a/tcctok.h
+++ b/tcctok.h
@@ -126,41 +126,42 @@
      DEF(TOK_memcpy4, "__aeabi_memcpy4")
      DEF(TOK_memcpy8, "__aeabi_memcpy8")
      DEF(TOK_memset, "__aeabi_memset")
+     DEF(TOK___aeabi_ldivmod, "__aeabi_ldivmod")
+     DEF(TOK___aeabi_uldivmod, "__aeabi_uldivmod")
 #else
      DEF(TOK_memcpy, "memcpy")
      DEF(TOK_memset, "memset")
-#endif
      DEF(TOK___divdi3, "__divdi3")
      DEF(TOK___moddi3, "__moddi3")
      DEF(TOK___udivdi3, "__udivdi3")
      DEF(TOK___umoddi3, "__umoddi3")
+#endif
 #if defined(TCC_TARGET_ARM)
-     DEF(TOK___modsi3, "__modsi3")
-     DEF(TOK___umodsi3, "__umodsi3")
 #ifdef TCC_ARM_EABI
+     DEF(TOK___aeabi_idivmod, "__aeabi_idivmod")
+     DEF(TOK___aeabi_uidivmod, "__aeabi_uidivmod")
      DEF(TOK___divsi3, "__aeabi_idiv")
      DEF(TOK___udivsi3, "__aeabi_uidiv")
-     DEF(TOK___sardi3, "__aeabi_lasr")
-     DEF(TOK___shrdi3, "__aeabi_llsr")
-     DEF(TOK___shldi3, "__aeabi_llsl")
-     DEF(TOK___slltof, "__aeabi_l2f")
-     DEF(TOK___slltold, "__aeabi_l2d")
+     DEF(TOK___floatdisf, "__aeabi_l2f")
+     DEF(TOK___floatdidf, "__aeabi_l2d")
      DEF(TOK___fixsfdi, "__aeabi_f2lz")
      DEF(TOK___fixdfdi, "__aeabi_d2lz")
-     DEF(TOK___fixxfdi, "__aeabi_d2lz")
 #else
+     DEF(TOK___modsi3, "__modsi3")
+     DEF(TOK___umodsi3, "__umodsi3")
      DEF(TOK___divsi3, "__divsi3")
      DEF(TOK___udivsi3, "__udivsi3")
-     DEF(TOK___sardi3, "__ashrdi3")
-     DEF(TOK___shrdi3, "__lshrdi3")
-     DEF(TOK___shldi3, "__ashldi3")
-     DEF(TOK___slltold, "__slltold")
+     DEF(TOK___floatdisf, "__floatdisf")
+     DEF(TOK___floatdidf, "__floatdidf")
+#ifndef TCC_ARM_VFP
+     DEF(TOK___floatdixf, "__floatdixf")
      DEF(TOK___fixunssfsi, "__fixunssfsi")
      DEF(TOK___fixunsdfsi, "__fixunsdfsi")
      DEF(TOK___fixunsxfsi, "__fixunsxfsi")
+     DEF(TOK___fixxfdi, "__fixxfdi")
+#endif
      DEF(TOK___fixsfdi, "__fixsfdi")
      DEF(TOK___fixdfdi, "__fixdfdi")
-     DEF(TOK___fixxfdi, "__fixxfdi")
 #endif
 #elif defined(TCC_TARGET_C67)
      DEF(TOK__divi, "_divi")
@@ -169,33 +170,35 @@
      DEF(TOK__divd, "_divd")
      DEF(TOK__remi, "_remi")
      DEF(TOK__remu, "_remu")
-     DEF(TOK___sardi3, "__sardi3")
-     DEF(TOK___shrdi3, "__shrdi3")
-     DEF(TOK___shldi3, "__shldi3")
-#else
-     /* XXX: same names on i386 ? */
-     DEF(TOK___sardi3, "__sardi3")
-     DEF(TOK___shrdi3, "__shrdi3")
-     DEF(TOK___shldi3, "__shldi3")
 #endif
+#ifdef TCC_TARGET_I386
      DEF(TOK___tcc_int_fpu_control, "__tcc_int_fpu_control")
      DEF(TOK___tcc_fpu_control, "__tcc_fpu_control")
+#endif
 #ifdef TCC_ARM_EABI
-     DEF(TOK___ulltof, "__aeabi_ul2f")
-     DEF(TOK___ulltod, "__aeabi_ul2d")
-     DEF(TOK___ulltold, "__aeabi_ul2d")
+     DEF(TOK___ashrdi3, "__aeabi_lasr")
+     DEF(TOK___lshrdi3, "__aeabi_llsr")
+     DEF(TOK___ashldi3, "__aeabi_llsl")
+     DEF(TOK___floatundisf, "__aeabi_ul2f")
+     DEF(TOK___floatundidf, "__aeabi_ul2d")
      DEF(TOK___fixunssfdi, "__aeabi_f2ulz")
      DEF(TOK___fixunsdfdi, "__aeabi_d2ulz")
-     DEF(TOK___fixunsxfdi, "__aeabi_d2ulz")
 #else
-     DEF(TOK___ulltof, "__ulltof")
-     DEF(TOK___ulltod, "__ulltod")
-     DEF(TOK___ulltold, "__ulltold")
+     DEF(TOK___ashrdi3, "__ashrdi3")
+     DEF(TOK___lshrdi3, "__lshrdi3")
+     DEF(TOK___ashldi3, "__ashldi3")
+     DEF(TOK___floatundisf, "__floatundisf")
+     DEF(TOK___floatundidf, "__floatundidf")
+#ifndef TCC_ARM_VFP
+     DEF(TOK___floatundixf, "__floatundixf")
+     DEF(TOK___fixunsxfdi, "__fixunsxfdi")
+#endif
      DEF(TOK___fixunssfdi, "__fixunssfdi")
      DEF(TOK___fixunsdfdi, "__fixunsdfdi")
-     DEF(TOK___fixunsxfdi, "__fixunsxfdi")
 #endif
+#ifdef TCC_TARGET_PE
      DEF(TOK___chkstk, "__chkstk")
+#endif
 
 /* bound checking symbols */
 #ifdef CONFIG_TCC_BCHECK