2 * X86 code generator for TCC
4 * Copyright (c) 2001, 2002 Fabrice Bellard
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program 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
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, 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 REG_EAX /* single word int return register */
54 #define REG_LRET REG_EDX /* second word return register (for long long) */
55 #define REG_FRET REG_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
71 /* relocation type for 32 bit data relocation */
72 #define R_DATA_32 R_386_32
74 /* function call context */
75 typedef struct GFuncContext
{
77 int func_call
; /* func call type (FUNC_STDCALL or FUNC_CDECL) */
80 /******************************************************/
82 static unsigned long func_sub_sp_offset
;
83 static unsigned long func_bound_offset
;
84 static int func_ret_sub
;
86 /* XXX: make it faster ? */
91 if (ind1
> cur_text_section
->data_allocated
)
92 section_realloc(cur_text_section
, ind1
);
93 cur_text_section
->data
[ind
] = c
;
113 /* output a symbol and patch all calls to it */
114 void gsym_addr(int t
, int a
)
118 ptr
= (int *)(cur_text_section
->data
+ t
);
119 n
= *ptr
; /* next value */
130 /* psym is used to put an instruction with a data field which is a
131 reference to a symbol. It is in fact the same as oad ! */
134 /* instruction + 4 bytes data. Return the address of the data */
135 static int oad(int c
, int s
)
141 if (ind1
> cur_text_section
->data_allocated
)
142 section_realloc(cur_text_section
, ind1
);
143 *(int *)(cur_text_section
->data
+ ind
) = s
;
149 /* output constant with relocation if 'r & VT_SYM' is true */
150 static void gen_addr32(int r
, Sym
*sym
, int c
)
153 greloc(cur_text_section
, sym
, ind
, R_386_32
);
157 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
159 static void gen_modrm(int op_reg
, int r
, Sym
*sym
, int c
)
161 op_reg
= op_reg
<< 3;
162 if ((r
& VT_VALMASK
) == VT_CONST
) {
163 /* constant memory reference */
165 gen_addr32(r
, sym
, c
);
166 } else if ((r
& VT_VALMASK
) == VT_LOCAL
) {
167 /* currently, we use only ebp as base */
169 /* short reference */
173 oad(0x85 | op_reg
, c
);
176 g(0x00 | op_reg
| (r
& VT_VALMASK
));
181 /* load 'r' from value 'sv' */
182 void load(int r
, SValue
*sv
)
184 int v
, t
, ft
, fc
, fr
;
193 if (v
== VT_LLOCAL
) {
195 v1
.r
= VT_LOCAL
| VT_LVAL
;
200 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
203 } else if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
206 } else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
209 } else if ((ft
& VT_TYPE
) == VT_BYTE
) {
210 o(0xbe0f); /* movsbl */
211 } else if ((ft
& VT_TYPE
) == (VT_BYTE
| VT_UNSIGNED
)) {
212 o(0xb60f); /* movzbl */
213 } else if ((ft
& VT_TYPE
) == VT_SHORT
) {
214 o(0xbf0f); /* movswl */
215 } else if ((ft
& VT_TYPE
) == (VT_SHORT
| VT_UNSIGNED
)) {
216 o(0xb70f); /* movzwl */
220 gen_modrm(r
, fr
, sv
->sym
, fc
);
223 o(0xb8 + r
); /* mov $xx, r */
224 gen_addr32(fr
, sv
->sym
, fc
);
225 } else if (v
== VT_LOCAL
) {
226 o(0x8d); /* lea xxx(%ebp), r */
227 gen_modrm(r
, VT_LOCAL
, sv
->sym
, fc
);
228 } else if (v
== VT_CMP
) {
229 oad(0xb8 + r
, 0); /* mov $0, r */
230 o(0x0f); /* setxx %br */
233 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
235 oad(0xb8 + r
, t
); /* mov $1, r */
236 oad(0xe9, 5); /* jmp after */
238 oad(0xb8 + r
, t
^ 1); /* mov $0, r */
241 o(0xc0 + r
+ v
* 8); /* mov v, r */
246 /* store register 'r' in lvalue 'v' */
247 void store(int r
, SValue
*v
)
253 fr
= v
->r
& VT_VALMASK
;
255 /* XXX: incorrect if float reg to reg */
256 if (bt
== VT_FLOAT
) {
259 } else if (bt
== VT_DOUBLE
) {
262 } else if (bt
== VT_LDOUBLE
) {
263 o(0xc0d9); /* fld %st(0) */
274 if (fr
== VT_CONST
||
277 gen_modrm(r
, v
->r
, v
->sym
, fc
);
278 } else if (fr
!= r
) {
279 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
283 /* start function call and return function call context */
284 void gfunc_start(GFuncContext
*c
, int func_call
)
287 c
->func_call
= func_call
;
290 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
292 void gfunc_param(GFuncContext
*c
)
296 if ((vtop
->t
& VT_BTYPE
) == VT_STRUCT
) {
297 size
= type_size(vtop
->t
, &align
);
298 /* align to stack align size */
299 size
= (size
+ 3) & ~3;
300 /* allocate the necessary size on stack */
301 oad(0xec81, size
); /* sub $xxx, %esp */
302 /* generate structure store */
304 o(0x89); /* mov %esp, r */
306 vset(vtop
->t
, r
| VT_LVAL
, 0);
309 c
->args_size
+= size
;
310 } else if (is_float(vtop
->t
)) {
311 gv(RC_FLOAT
); /* only one float register */
312 if ((vtop
->t
& VT_BTYPE
) == VT_FLOAT
)
314 else if ((vtop
->t
& VT_BTYPE
) == VT_DOUBLE
)
318 oad(0xec81, size
); /* sub $xxx, %esp */
322 o(0x5cd9 + size
- 4); /* fstp[s|l] 0(%esp) */
325 c
->args_size
+= size
;
327 /* simple type (currently always same size) */
328 /* XXX: implicit cast ? */
330 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
332 o(0x50 + vtop
->r2
); /* push r */
336 o(0x50 + r
); /* push r */
337 c
->args_size
+= size
;
342 static void gadd_sp(int val
)
344 if (val
== (char)val
) {
348 oad(0xc481, val
); /* add $xxx, %esp */
352 /* generate function call with address in (vtop->t, vtop->c) and free function
353 context. Stack entry is popped */
354 void gfunc_call(GFuncContext
*c
)
357 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
359 if (vtop
->r
& VT_SYM
) {
360 /* relocation case */
361 greloc(cur_text_section
, vtop
->sym
,
362 ind
+ 1, R_386_PC32
);
363 oad(0xe8, vtop
->c
.ul
- 4);
365 oad(0xe8, vtop
->c
.ul
- ind
- 5);
368 /* otherwise, indirect call */
370 o(0xff); /* call *r */
373 if (c
->args_size
&& c
->func_call
== FUNC_CDECL
)
374 gadd_sp(c
->args_size
);
378 /* generate function prolog of type 't' */
379 void gfunc_prolog(int t
)
381 int addr
, align
, size
, u
, func_call
;
384 sym
= sym_find((unsigned)t
>> VT_STRUCT_SHIFT
);
387 /* if the function returns a structure, then add an
388 implicit pointer parameter */
390 if ((func_vt
& VT_BTYPE
) == VT_STRUCT
) {
394 /* define parameters */
395 while ((sym
= sym
->next
) != NULL
) {
397 sym_push(sym
->v
& ~SYM_FIELD
, u
,
398 VT_LOCAL
| VT_LVAL
, addr
);
399 size
= type_size(u
, &align
);
400 size
= (size
+ 3) & ~3;
401 #ifdef FUNC_STRUCT_PARAM_AS_PTR
402 /* structs are passed as pointer */
403 if ((u
& VT_BTYPE
) == VT_STRUCT
) {
410 /* pascal type call ? */
411 if (func_call
== FUNC_STDCALL
)
412 func_ret_sub
= addr
- 8;
413 o(0xe58955); /* push %ebp, mov %esp, %ebp */
414 func_sub_sp_offset
= oad(0xec81, 0); /* sub $xxx, %esp */
415 /* leave some room for bound checking code */
416 if (do_bounds_check
) {
417 oad(0xb8, 0); /* lbound section pointer */
418 oad(0xb8, 0); /* call to function */
419 func_bound_offset
= lbounds_section
->data_offset
;
423 /* generate function epilog */
424 void gfunc_epilog(void)
426 #ifdef CONFIG_TCC_BCHECK
427 if (do_bounds_check
&& func_bound_offset
!= lbounds_section
->data_offset
) {
431 /* add end of table info */
432 bounds_ptr
= section_ptr_add(lbounds_section
, sizeof(int));
434 /* generate bound local allocation */
436 ind
= func_sub_sp_offset
+ 4;
437 sym_data
= get_sym_ref(char_pointer_type
, lbounds_section
,
438 func_bound_offset
, lbounds_section
->data_offset
);
439 greloc(cur_text_section
, sym_data
,
441 oad(0xb8, 0); /* mov %eax, xxx */
442 sym
= external_global_sym(TOK___bound_local_new
, func_old_type
, 0);
443 greloc(cur_text_section
, sym
,
444 ind
+ 1, R_386_PC32
);
447 /* generate bound check local freeing */
448 o(0x5250); /* save returned value, if any */
449 greloc(cur_text_section
, sym_data
,
451 oad(0xb8, 0); /* mov %eax, xxx */
452 sym
= external_global_sym(TOK___bound_local_delete
, func_old_type
, 0);
453 greloc(cur_text_section
, sym
,
454 ind
+ 1, R_386_PC32
);
456 o(0x585a); /* restore returned value, if any */
460 if (func_ret_sub
== 0) {
465 g(func_ret_sub
>> 8);
467 /* align local size to word & save local variables */
468 *(int *)(cur_text_section
->data
+ func_sub_sp_offset
) = (-loc
+ 3) & -4;
471 /* generate a jump to a label */
474 return psym(0xe9, t
);
477 /* generate a jump to a fixed address */
478 void gjmp_addr(int a
)
480 oad(0xe9, a
- ind
- 5);
483 /* generate a test. set 'inv' to invert test. Stack entry is popped */
484 int gtst(int inv
, int t
)
487 v
= vtop
->r
& VT_VALMASK
;
489 /* fast case : can jump directly since flags are set */
491 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
492 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
493 /* && or || optimization */
494 if ((v
& 1) == inv
) {
495 /* insert vtop->c jump list in t */
498 p
= (int *)(cur_text_section
->data
+ *p
);
506 if (is_float(vtop
->t
)) {
510 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
511 /* constant jmp optimization */
512 if ((vtop
->c
.i
!= 0) != inv
)
519 t
= psym(0x85 ^ inv
, t
);
526 /* generate an integer binary operation */
533 case TOK_ADDC1
: /* add with carry generation */
536 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
543 /* XXX: generate inc and dec for smaller code ? */
545 o(0xc0 | (opc
<< 3) | r
);
549 oad(0xc0 | (opc
<< 3) | r
, c
);
555 o((opc
<< 3) | 0x01);
556 o(0xc0 + r
+ fr
* 8);
559 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
561 vset(VT_INT
, VT_CMP
, op
);
565 case TOK_SUBC1
: /* sub with carry generation */
568 case TOK_ADDC2
: /* add with carry use */
571 case TOK_SUBC2
: /* sub with carry use */
588 o(0xaf0f); /* imul fr, r */
589 o(0xc0 + fr
+ r
* 8);
600 opc
= 0xc0 | (opc
<< 3);
601 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
606 c
= vtop
->c
.i
& 0x1f;
607 o(0xc1); /* shl/shr/sar $xxx, r */
611 /* we generate the shift in ecx */
614 o(0xd3); /* shl/shr/sar %cl, r */
625 /* first operand must be in eax */
626 /* XXX: need better constraint for second operand */
632 if (op
== TOK_UMULL
) {
633 o(0xf7); /* mul fr */
638 if (op
== TOK_UDIV
|| op
== TOK_UMOD
) {
639 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
642 o(0xf799); /* cltd, idiv fr, %eax */
645 if (op
== '%' || op
== TOK_UMOD
)
658 /* generate a floating point operation 'v = t1 op t2' instruction. The
659 two operands are guaranted to have the same floating point type */
660 /* XXX: need to use ST1 too */
663 int a
, ft
, fc
, swapped
, r
;
665 /* convert constants to memory references */
666 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
671 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
674 /* must put at least one value in the floating point register */
675 if ((vtop
[-1].r
& VT_LVAL
) &&
676 (vtop
[0].r
& VT_LVAL
)) {
682 /* swap the stack if needed so that t1 is the register and t2 is
683 the memory reference */
684 if (vtop
[-1].r
& VT_LVAL
) {
688 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
689 /* load on stack second operand */
691 save_reg(REG_EAX
); /* eax is used by FP comparison code */
692 if (op
== TOK_GE
|| op
== TOK_GT
)
694 else if (op
== TOK_EQ
|| op
== TOK_NE
)
697 o(0xc9d9); /* fxch %st(1) */
698 o(0xe9da); /* fucompp */
699 o(0xe0df); /* fnstsw %ax */
701 o(0x45e480); /* and $0x45, %ah */
702 o(0x40fC80); /* cmp $0x40, %ah */
703 } else if (op
== TOK_NE
) {
704 o(0x45e480); /* and $0x45, %ah */
705 o(0x40f480); /* xor $0x40, %ah */
707 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
708 o(0x05c4f6); /* test $0x05, %ah */
711 o(0x45c4f6); /* test $0x45, %ah */
718 /* no memory reference possible for long double operations */
719 if ((vtop
->t
& VT_BTYPE
) == VT_LDOUBLE
) {
745 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
746 o(0xde); /* fxxxp %st, %st(1) */
749 /* if saved lvalue, then we must reload it */
751 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
755 v1
.r
= VT_LOCAL
| VT_LVAL
;
761 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
765 gen_modrm(a
, r
, vtop
->sym
, fc
);
771 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
772 and 'long long' cases. */
773 void gen_cvt_itof(int t
)
777 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
778 /* signed long long to float/double/long double (unsigned case
779 is handled generically) */
780 o(0x50 + vtop
->r2
); /* push r2 */
781 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
782 o(0x242cdf); /* fildll (%esp) */
783 o(0x08c483); /* add $8, %esp */
784 } else if ((vtop
->t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
785 (VT_INT
| VT_UNSIGNED
)) {
786 /* unsigned int to float/double/long double */
787 o(0x6a); /* push $0 */
789 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
790 o(0x242cdf); /* fildll (%esp) */
791 o(0x08c483); /* add $8, %esp */
793 /* int to float/double/long double */
794 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
795 o(0x2404db); /* fildl (%esp) */
796 o(0x04c483); /* add $4, %esp */
801 /* convert fp to int 't' type */
802 /* XXX: handle long long case */
803 void gen_cvt_ftoi(int t
)
814 o(0x2dd9); /* ldcw xxx */
815 sym
= external_global_sym(TOK___tcc_int_fpu_control
,
816 VT_SHORT
| VT_UNSIGNED
, VT_LVAL
);
817 greloc(cur_text_section
, sym
,
821 oad(0xec81, size
); /* sub $xxx, %esp */
823 o(0x1cdb); /* fistpl */
825 o(0x3cdf); /* fistpll */
827 o(0x2dd9); /* ldcw xxx */
828 sym
= external_global_sym(TOK___tcc_fpu_control
,
829 VT_SHORT
| VT_UNSIGNED
, VT_LVAL
);
830 greloc(cur_text_section
, sym
,
835 o(0x58 + r
); /* pop r */
838 vtop
->r
= r
; /* mark reg as used */
839 r2
= get_reg(RC_INT
);
840 o(0x58 + r2
); /* pop r2 */
843 o(0x04c483); /* add $4, %esp */
849 /* convert from one floating point type to another */
850 void gen_cvt_ftof(int t
)
852 /* all we have to do on i386 is to put the float in a register */
856 /* bound check support functions */
857 #ifdef CONFIG_TCC_BCHECK
859 /* generate a bounded pointer addition */
860 void gen_bounded_ptr_add(void)
864 /* prepare fast i386 function call (args in eax and edx) */
866 /* save all temporary registers */
869 /* do a fast function call */
870 sym
= external_global_sym(TOK___bound_ptr_add
, func_old_type
, 0);
871 greloc(cur_text_section
, sym
,
872 ind
+ 1, R_386_PC32
);
874 /* returned pointer is in eax */
876 vtop
->r
= REG_EAX
| VT_BOUNDED
;
877 /* address of bounding function call point */
878 vtop
->c
.ul
= (cur_text_section
->reloc
->data_offset
- sizeof(Elf32_Rel
));
881 /* patch pointer addition in vtop so that pointer dereferencing is
883 void gen_bounded_ptr_deref(void)
891 /* XXX: put that code in generic part of tcc */
892 if (!is_float(vtop
->t
)) {
893 if (vtop
->r
& VT_LVAL_BYTE
)
895 else if (vtop
->r
& VT_LVAL_SHORT
)
899 size
= type_size(vtop
->t
, &align
);
901 case 1: func
= TOK___bound_ptr_indir1
; break;
902 case 2: func
= TOK___bound_ptr_indir2
; break;
903 case 4: func
= TOK___bound_ptr_indir4
; break;
904 case 8: func
= TOK___bound_ptr_indir8
; break;
905 case 12: func
= TOK___bound_ptr_indir12
; break;
906 case 16: func
= TOK___bound_ptr_indir16
; break;
908 error("unhandled size when derefencing bounded pointer");
913 /* patch relocation */
914 /* XXX: find a better solution ? */
915 rel
= (Elf32_Rel
*)(cur_text_section
->reloc
->data
+ vtop
->c
.ul
);
916 sym
= external_global_sym(func
, func_old_type
, 0);
918 put_extern_sym(sym
, NULL
, 0, 0);
919 rel
->r_info
= ELF32_R_INFO(sym
->c
, ELF32_R_TYPE(rel
->r_info
));
923 /* end of X86 code generator */
924 /*************************************************************/