2 * X86 code generator for TCC
4 * Copyright (c) 2001, 2002, 2003 Fabrice Bellard
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 /* number of available registers */
24 /* a register can belong to several classes. The classes must be
25 sorted from more general to more precise (see gv2() code which does
26 assumptions on it). */
27 #define RC_INT 0x0001 /* generic integer register */
28 #define RC_FLOAT 0x0002 /* generic float register */
33 #define RC_IRET RC_EAX /* function return: integer register */
34 #define RC_LRET RC_EDX /* function return: second integer register */
35 #define RC_FRET RC_ST0 /* function return: float register */
37 /* pretty names for the registers */
45 int reg_classes
[NB_REGS
] = {
46 /* eax */ RC_INT
| RC_EAX
,
47 /* ecx */ RC_INT
| RC_ECX
,
48 /* edx */ RC_INT
| RC_EDX
,
49 /* st0 */ RC_FLOAT
| RC_ST0
,
52 /* return registers for function */
53 #define REG_IRET TREG_EAX /* single word int return register */
54 #define REG_LRET TREG_EDX /* second word return register (for long long) */
55 #define REG_FRET TREG_ST0 /* float return register */
57 /* defined if function parameters must be evaluated in reverse order */
58 #define INVERT_FUNC_PARAMS
60 /* defined if structures are passed as pointers. Otherwise structures
61 are directly pushed on stack. */
62 //#define FUNC_STRUCT_PARAM_AS_PTR
64 /* pointer size, in bytes */
67 /* long double size and alignment, in bytes */
68 #define LDOUBLE_SIZE 12
69 #define LDOUBLE_ALIGN 4
70 /* maximum alignment (for aligned attribute support) */
73 /* relocation type for 32 bit data relocation */
74 #define R_DATA_32 R_386_32
76 /******************************************************/
78 static unsigned long func_sub_sp_offset
;
79 static unsigned long func_bound_offset
;
80 static int func_ret_sub
;
82 /* XXX: make it faster ? */
87 if (ind1
> cur_text_section
->data_allocated
)
88 section_realloc(cur_text_section
, ind1
);
89 cur_text_section
->data
[ind
] = c
;
109 /* output a symbol and patch all calls to it */
110 void gsym_addr(int t
, int a
)
114 ptr
= (int *)(cur_text_section
->data
+ t
);
115 n
= *ptr
; /* next value */
126 /* psym is used to put an instruction with a data field which is a
127 reference to a symbol. It is in fact the same as oad ! */
130 /* instruction + 4 bytes data. Return the address of the data */
131 static int oad(int c
, int s
)
137 if (ind1
> cur_text_section
->data_allocated
)
138 section_realloc(cur_text_section
, ind1
);
139 *(int *)(cur_text_section
->data
+ ind
) = s
;
145 /* output constant with relocation if 'r & VT_SYM' is true */
146 static void gen_addr32(int r
, Sym
*sym
, int c
)
149 greloc(cur_text_section
, sym
, ind
, R_386_32
);
153 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
155 static void gen_modrm(int op_reg
, int r
, Sym
*sym
, int c
)
157 op_reg
= op_reg
<< 3;
158 if ((r
& VT_VALMASK
) == VT_CONST
) {
159 /* constant memory reference */
161 gen_addr32(r
, sym
, c
);
162 } else if ((r
& VT_VALMASK
) == VT_LOCAL
) {
163 /* currently, we use only ebp as base */
165 /* short reference */
169 oad(0x85 | op_reg
, c
);
172 g(0x00 | op_reg
| (r
& VT_VALMASK
));
177 /* load 'r' from value 'sv' */
178 void load(int r
, SValue
*sv
)
180 int v
, t
, ft
, fc
, fr
;
189 if (v
== VT_LLOCAL
) {
191 v1
.r
= VT_LOCAL
| VT_LVAL
;
196 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
199 } else if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
202 } else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
205 } else if ((ft
& VT_TYPE
) == VT_BYTE
) {
206 o(0xbe0f); /* movsbl */
207 } else if ((ft
& VT_TYPE
) == (VT_BYTE
| VT_UNSIGNED
)) {
208 o(0xb60f); /* movzbl */
209 } else if ((ft
& VT_TYPE
) == VT_SHORT
) {
210 o(0xbf0f); /* movswl */
211 } else if ((ft
& VT_TYPE
) == (VT_SHORT
| VT_UNSIGNED
)) {
212 o(0xb70f); /* movzwl */
216 gen_modrm(r
, fr
, sv
->sym
, fc
);
219 o(0xb8 + r
); /* mov $xx, r */
220 gen_addr32(fr
, sv
->sym
, fc
);
221 } else if (v
== VT_LOCAL
) {
222 o(0x8d); /* lea xxx(%ebp), r */
223 gen_modrm(r
, VT_LOCAL
, sv
->sym
, fc
);
224 } else if (v
== VT_CMP
) {
225 oad(0xb8 + r
, 0); /* mov $0, r */
226 o(0x0f); /* setxx %br */
229 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
231 oad(0xb8 + r
, t
); /* mov $1, r */
232 o(0x05eb); /* jmp after */
234 oad(0xb8 + r
, t
^ 1); /* mov $0, r */
237 o(0xc0 + r
+ v
* 8); /* mov v, r */
242 /* store register 'r' in lvalue 'v' */
243 void store(int r
, SValue
*v
)
249 fr
= v
->r
& VT_VALMASK
;
251 /* XXX: incorrect if float reg to reg */
252 if (bt
== VT_FLOAT
) {
255 } else if (bt
== VT_DOUBLE
) {
258 } else if (bt
== VT_LDOUBLE
) {
259 o(0xc0d9); /* fld %st(0) */
270 if (fr
== VT_CONST
||
273 gen_modrm(r
, v
->r
, v
->sym
, fc
);
274 } else if (fr
!= r
) {
275 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
279 static void gadd_sp(int val
)
281 if (val
== (char)val
) {
285 oad(0xc481, val
); /* add $xxx, %esp */
289 /* 'is_jmp' is '1' if it is a jump */
290 static void gcall_or_jmp(int is_jmp
)
293 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
295 if (vtop
->r
& VT_SYM
) {
296 /* relocation case */
297 greloc(cur_text_section
, vtop
->sym
,
298 ind
+ 1, R_386_PC32
);
300 /* put an empty PC32 relocation */
301 put_elf_reloc(symtab_section
, cur_text_section
,
302 ind
+ 1, R_386_PC32
, 0);
304 oad(0xe8 + is_jmp
, vtop
->c
.ul
- 4); /* call/jmp im */
306 /* otherwise, indirect call */
308 o(0xff); /* call/jmp *r */
309 o(0xd0 + r
+ (is_jmp
<< 4));
313 /* Generate function call. The function address is pushed first, then
314 all the parameters in call order. This functions pops all the
315 parameters and the function address. */
316 void gfunc_call(int nb_args
)
318 int size
, align
, r
, args_size
, i
;
322 for(i
= 0;i
< nb_args
; i
++) {
323 if ((vtop
->type
.t
& VT_BTYPE
) == VT_STRUCT
) {
324 size
= type_size(&vtop
->type
, &align
);
325 /* align to stack align size */
326 size
= (size
+ 3) & ~3;
327 /* allocate the necessary size on stack */
328 oad(0xec81, size
); /* sub $xxx, %esp */
329 /* generate structure store */
331 o(0x89); /* mov %esp, r */
333 vset(&vtop
->type
, r
| VT_LVAL
, 0);
337 } else if (is_float(vtop
->type
.t
)) {
338 gv(RC_FLOAT
); /* only one float register */
339 if ((vtop
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
341 else if ((vtop
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
345 oad(0xec81, size
); /* sub $xxx, %esp */
349 o(0x5cd9 + size
- 4); /* fstp[s|l] 0(%esp) */
354 /* simple type (currently always same size) */
355 /* XXX: implicit cast ? */
357 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
359 o(0x50 + vtop
->r2
); /* push r */
363 o(0x50 + r
); /* push r */
368 save_regs(0); /* save used temporary registers */
369 func_sym
= vtop
->type
.ref
;
371 if (args_size
&& func_sym
->r
== FUNC_CDECL
)
376 /* generate function prolog of type 't' */
377 void gfunc_prolog(CType
*func_type
)
379 int addr
, align
, size
, func_call
;
383 sym
= func_type
->ref
;
386 /* if the function returns a structure, then add an
387 implicit pointer parameter */
389 if ((func_vt
.t
& VT_BTYPE
) == VT_STRUCT
) {
393 /* define parameters */
394 while ((sym
= sym
->next
) != NULL
) {
396 sym_push(sym
->v
& ~SYM_FIELD
, type
,
397 VT_LOCAL
| VT_LVAL
, addr
);
398 size
= type_size(type
, &align
);
399 size
= (size
+ 3) & ~3;
400 #ifdef FUNC_STRUCT_PARAM_AS_PTR
401 /* structs are passed as pointer */
402 if ((type
->t
& VT_BTYPE
) == VT_STRUCT
) {
409 /* pascal type call ? */
410 if (func_call
== FUNC_STDCALL
)
411 func_ret_sub
= addr
- 8;
412 o(0xe58955); /* push %ebp, mov %esp, %ebp */
413 func_sub_sp_offset
= oad(0xec81, 0); /* sub $xxx, %esp */
414 /* leave some room for bound checking code */
415 if (do_bounds_check
) {
416 oad(0xb8, 0); /* lbound section pointer */
417 oad(0xb8, 0); /* call to function */
418 func_bound_offset
= lbounds_section
->data_offset
;
422 /* generate function epilog */
423 void gfunc_epilog(void)
425 #ifdef CONFIG_TCC_BCHECK
426 if (do_bounds_check
&& func_bound_offset
!= lbounds_section
->data_offset
) {
430 /* add end of table info */
431 bounds_ptr
= section_ptr_add(lbounds_section
, sizeof(int));
433 /* generate bound local allocation */
435 ind
= func_sub_sp_offset
+ 4;
436 sym_data
= get_sym_ref(&char_pointer_type
, lbounds_section
,
437 func_bound_offset
, lbounds_section
->data_offset
);
438 greloc(cur_text_section
, sym_data
,
440 oad(0xb8, 0); /* mov %eax, xxx */
441 sym
= external_global_sym(TOK___bound_local_new
, &func_old_type
, 0);
442 greloc(cur_text_section
, sym
,
443 ind
+ 1, R_386_PC32
);
446 /* generate bound check local freeing */
447 o(0x5250); /* save returned value, if any */
448 greloc(cur_text_section
, sym_data
,
450 oad(0xb8, 0); /* mov %eax, xxx */
451 sym
= external_global_sym(TOK___bound_local_delete
, &func_old_type
, 0);
452 greloc(cur_text_section
, sym
,
453 ind
+ 1, R_386_PC32
);
455 o(0x585a); /* restore returned value, if any */
459 if (func_ret_sub
== 0) {
464 g(func_ret_sub
>> 8);
466 /* align local size to word & save local variables */
467 *(int *)(cur_text_section
->data
+ func_sub_sp_offset
) = (-loc
+ 3) & -4;
470 /* generate a jump to a label */
473 return psym(0xe9, t
);
476 /* generate a jump to a fixed address */
477 void gjmp_addr(int a
)
485 oad(0xe9, a
- ind
- 5);
489 /* generate a test. set 'inv' to invert test. Stack entry is popped */
490 int gtst(int inv
, int t
)
494 v
= vtop
->r
& VT_VALMASK
;
496 /* fast case : can jump directly since flags are set */
498 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
499 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
500 /* && or || optimization */
501 if ((v
& 1) == inv
) {
502 /* insert vtop->c jump list in t */
505 p
= (int *)(cur_text_section
->data
+ *p
);
513 if (is_float(vtop
->type
.t
)) {
517 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
518 /* constant jmp optimization */
519 if ((vtop
->c
.i
!= 0) != inv
)
526 t
= psym(0x85 ^ inv
, t
);
533 /* generate an integer binary operation */
540 case TOK_ADDC1
: /* add with carry generation */
543 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
550 /* XXX: generate inc and dec for smaller code ? */
552 o(0xc0 | (opc
<< 3) | r
);
556 oad(0xc0 | (opc
<< 3) | r
, c
);
562 o((opc
<< 3) | 0x01);
563 o(0xc0 + r
+ fr
* 8);
566 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
572 case TOK_SUBC1
: /* sub with carry generation */
575 case TOK_ADDC2
: /* add with carry use */
578 case TOK_SUBC2
: /* sub with carry use */
595 o(0xaf0f); /* imul fr, r */
596 o(0xc0 + fr
+ r
* 8);
607 opc
= 0xc0 | (opc
<< 3);
608 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
613 c
= vtop
->c
.i
& 0x1f;
614 o(0xc1); /* shl/shr/sar $xxx, r */
618 /* we generate the shift in ecx */
621 o(0xd3); /* shl/shr/sar %cl, r */
632 /* first operand must be in eax */
633 /* XXX: need better constraint for second operand */
639 if (op
== TOK_UMULL
) {
640 o(0xf7); /* mul fr */
645 if (op
== TOK_UDIV
|| op
== TOK_UMOD
) {
646 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
649 o(0xf799); /* cltd, idiv fr, %eax */
652 if (op
== '%' || op
== TOK_UMOD
)
665 /* generate a floating point operation 'v = t1 op t2' instruction. The
666 two operands are guaranted to have the same floating point type */
667 /* XXX: need to use ST1 too */
670 int a
, ft
, fc
, swapped
, r
;
672 /* convert constants to memory references */
673 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
678 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
681 /* must put at least one value in the floating point register */
682 if ((vtop
[-1].r
& VT_LVAL
) &&
683 (vtop
[0].r
& VT_LVAL
)) {
689 /* swap the stack if needed so that t1 is the register and t2 is
690 the memory reference */
691 if (vtop
[-1].r
& VT_LVAL
) {
695 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
696 /* load on stack second operand */
697 load(TREG_ST0
, vtop
);
698 save_reg(TREG_EAX
); /* eax is used by FP comparison code */
699 if (op
== TOK_GE
|| op
== TOK_GT
)
701 else if (op
== TOK_EQ
|| op
== TOK_NE
)
704 o(0xc9d9); /* fxch %st(1) */
705 o(0xe9da); /* fucompp */
706 o(0xe0df); /* fnstsw %ax */
708 o(0x45e480); /* and $0x45, %ah */
709 o(0x40fC80); /* cmp $0x40, %ah */
710 } else if (op
== TOK_NE
) {
711 o(0x45e480); /* and $0x45, %ah */
712 o(0x40f480); /* xor $0x40, %ah */
714 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
715 o(0x05c4f6); /* test $0x05, %ah */
718 o(0x45c4f6); /* test $0x45, %ah */
725 /* no memory reference possible for long double operations */
726 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
) {
727 load(TREG_ST0
, vtop
);
752 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
753 o(0xde); /* fxxxp %st, %st(1) */
756 /* if saved lvalue, then we must reload it */
758 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
762 v1
.r
= VT_LOCAL
| VT_LVAL
;
768 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
772 gen_modrm(a
, r
, vtop
->sym
, fc
);
778 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
779 and 'long long' cases. */
780 void gen_cvt_itof(int t
)
784 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
785 /* signed long long to float/double/long double (unsigned case
786 is handled generically) */
787 o(0x50 + vtop
->r2
); /* push r2 */
788 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
789 o(0x242cdf); /* fildll (%esp) */
790 o(0x08c483); /* add $8, %esp */
791 } else if ((vtop
->type
.t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
792 (VT_INT
| VT_UNSIGNED
)) {
793 /* unsigned int to float/double/long double */
794 o(0x6a); /* push $0 */
796 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
797 o(0x242cdf); /* fildll (%esp) */
798 o(0x08c483); /* add $8, %esp */
800 /* int to float/double/long double */
801 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
802 o(0x2404db); /* fildl (%esp) */
803 o(0x04c483); /* add $4, %esp */
808 /* convert fp to int 't' type */
809 /* XXX: handle long long case */
810 void gen_cvt_ftoi(int t
)
816 ushort_type
.t
= VT_SHORT
| VT_UNSIGNED
;
824 o(0x2dd9); /* ldcw xxx */
825 sym
= external_global_sym(TOK___tcc_int_fpu_control
,
826 &ushort_type
, VT_LVAL
);
827 greloc(cur_text_section
, sym
,
831 oad(0xec81, size
); /* sub $xxx, %esp */
833 o(0x1cdb); /* fistpl */
835 o(0x3cdf); /* fistpll */
837 o(0x2dd9); /* ldcw xxx */
838 sym
= external_global_sym(TOK___tcc_fpu_control
,
839 &ushort_type
, VT_LVAL
);
840 greloc(cur_text_section
, sym
,
845 o(0x58 + r
); /* pop r */
848 vtop
->r
= r
; /* mark reg as used */
849 r2
= get_reg(RC_INT
);
850 o(0x58 + r2
); /* pop r2 */
853 o(0x04c483); /* add $4, %esp */
859 /* convert from one floating point type to another */
860 void gen_cvt_ftof(int t
)
862 /* all we have to do on i386 is to put the float in a register */
866 /* computed goto support */
873 /* bound check support functions */
874 #ifdef CONFIG_TCC_BCHECK
876 /* generate a bounded pointer addition */
877 void gen_bounded_ptr_add(void)
881 /* prepare fast i386 function call (args in eax and edx) */
883 /* save all temporary registers */
886 /* do a fast function call */
887 sym
= external_global_sym(TOK___bound_ptr_add
, &func_old_type
, 0);
888 greloc(cur_text_section
, sym
,
889 ind
+ 1, R_386_PC32
);
891 /* returned pointer is in eax */
893 vtop
->r
= TREG_EAX
| VT_BOUNDED
;
894 /* address of bounding function call point */
895 vtop
->c
.ul
= (cur_text_section
->reloc
->data_offset
- sizeof(Elf32_Rel
));
898 /* patch pointer addition in vtop so that pointer dereferencing is
900 void gen_bounded_ptr_deref(void)
908 /* XXX: put that code in generic part of tcc */
909 if (!is_float(vtop
->type
.t
)) {
910 if (vtop
->r
& VT_LVAL_BYTE
)
912 else if (vtop
->r
& VT_LVAL_SHORT
)
916 size
= type_size(&vtop
->type
, &align
);
918 case 1: func
= TOK___bound_ptr_indir1
; break;
919 case 2: func
= TOK___bound_ptr_indir2
; break;
920 case 4: func
= TOK___bound_ptr_indir4
; break;
921 case 8: func
= TOK___bound_ptr_indir8
; break;
922 case 12: func
= TOK___bound_ptr_indir12
; break;
923 case 16: func
= TOK___bound_ptr_indir16
; break;
925 error("unhandled size when derefencing bounded pointer");
930 /* patch relocation */
931 /* XXX: find a better solution ? */
932 rel
= (Elf32_Rel
*)(cur_text_section
->reloc
->data
+ vtop
->c
.ul
);
933 sym
= external_global_sym(func
, &func_old_type
, 0);
935 put_extern_sym(sym
, NULL
, 0, 0);
936 rel
->r_info
= ELF32_R_INFO(sym
->c
, ELF32_R_TYPE(rel
->r_info
));
940 /* end of X86 code generator */
941 /*************************************************************/