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 /******************************************************/
76 #define EM_TCC_TARGET EM_386
78 /* relocation type for 32 bit data relocation */
79 #define R_DATA_32 R_386_32
80 #define R_JMP_SLOT R_386_JMP_SLOT
81 #define R_COPY R_386_COPY
83 #define ELF_START_ADDR 0x08048000
84 #define ELF_PAGE_SIZE 0x1000
86 /******************************************************/
88 static unsigned long func_sub_sp_offset
;
89 static unsigned long func_bound_offset
;
90 static int func_ret_sub
;
92 /* XXX: make it faster ? */
97 if (ind1
> cur_text_section
->data_allocated
)
98 section_realloc(cur_text_section
, ind1
);
99 cur_text_section
->data
[ind
] = c
;
103 void o(unsigned int c
)
119 /* output a symbol and patch all calls to it */
120 void gsym_addr(int t
, int a
)
124 ptr
= (int *)(cur_text_section
->data
+ t
);
125 n
= *ptr
; /* next value */
136 /* psym is used to put an instruction with a data field which is a
137 reference to a symbol. It is in fact the same as oad ! */
140 /* instruction + 4 bytes data. Return the address of the data */
141 static int oad(int c
, int s
)
147 if (ind1
> cur_text_section
->data_allocated
)
148 section_realloc(cur_text_section
, ind1
);
149 *(int *)(cur_text_section
->data
+ ind
) = s
;
155 /* output constant with relocation if 'r & VT_SYM' is true */
156 static void gen_addr32(int r
, Sym
*sym
, int c
)
159 greloc(cur_text_section
, sym
, ind
, R_386_32
);
163 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
165 static void gen_modrm(int op_reg
, int r
, Sym
*sym
, int c
)
167 op_reg
= op_reg
<< 3;
168 if ((r
& VT_VALMASK
) == VT_CONST
) {
169 /* constant memory reference */
171 gen_addr32(r
, sym
, c
);
172 } else if ((r
& VT_VALMASK
) == VT_LOCAL
) {
173 /* currently, we use only ebp as base */
175 /* short reference */
179 oad(0x85 | op_reg
, c
);
182 g(0x00 | op_reg
| (r
& VT_VALMASK
));
187 /* load 'r' from value 'sv' */
188 void load(int r
, SValue
*sv
)
190 int v
, t
, ft
, fc
, fr
;
199 if (v
== VT_LLOCAL
) {
201 v1
.r
= VT_LOCAL
| VT_LVAL
;
206 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
209 } else if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
212 } else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
215 } else if ((ft
& VT_TYPE
) == VT_BYTE
) {
216 o(0xbe0f); /* movsbl */
217 } else if ((ft
& VT_TYPE
) == (VT_BYTE
| VT_UNSIGNED
)) {
218 o(0xb60f); /* movzbl */
219 } else if ((ft
& VT_TYPE
) == VT_SHORT
) {
220 o(0xbf0f); /* movswl */
221 } else if ((ft
& VT_TYPE
) == (VT_SHORT
| VT_UNSIGNED
)) {
222 o(0xb70f); /* movzwl */
226 gen_modrm(r
, fr
, sv
->sym
, fc
);
229 o(0xb8 + r
); /* mov $xx, r */
230 gen_addr32(fr
, sv
->sym
, fc
);
231 } else if (v
== VT_LOCAL
) {
232 o(0x8d); /* lea xxx(%ebp), r */
233 gen_modrm(r
, VT_LOCAL
, sv
->sym
, fc
);
234 } else if (v
== VT_CMP
) {
235 oad(0xb8 + r
, 0); /* mov $0, r */
236 o(0x0f); /* setxx %br */
239 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
241 oad(0xb8 + r
, t
); /* mov $1, r */
242 o(0x05eb); /* jmp after */
244 oad(0xb8 + r
, t
^ 1); /* mov $0, r */
247 o(0xc0 + r
+ v
* 8); /* mov v, r */
252 /* store register 'r' in lvalue 'v' */
253 void store(int r
, SValue
*v
)
259 fr
= v
->r
& VT_VALMASK
;
261 /* XXX: incorrect if float reg to reg */
262 if (bt
== VT_FLOAT
) {
265 } else if (bt
== VT_DOUBLE
) {
268 } else if (bt
== VT_LDOUBLE
) {
269 o(0xc0d9); /* fld %st(0) */
280 if (fr
== VT_CONST
||
283 gen_modrm(r
, v
->r
, v
->sym
, fc
);
284 } else if (fr
!= r
) {
285 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
289 static void gadd_sp(int val
)
291 if (val
== (char)val
) {
295 oad(0xc481, val
); /* add $xxx, %esp */
299 /* 'is_jmp' is '1' if it is a jump */
300 static void gcall_or_jmp(int is_jmp
)
303 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
305 if (vtop
->r
& VT_SYM
) {
306 /* relocation case */
307 greloc(cur_text_section
, vtop
->sym
,
308 ind
+ 1, R_386_PC32
);
310 /* put an empty PC32 relocation */
311 put_elf_reloc(symtab_section
, cur_text_section
,
312 ind
+ 1, R_386_PC32
, 0);
314 oad(0xe8 + is_jmp
, vtop
->c
.ul
- 4); /* call/jmp im */
316 /* otherwise, indirect call */
318 o(0xff); /* call/jmp *r */
319 o(0xd0 + r
+ (is_jmp
<< 4));
323 static uint8_t fastcall_regs
[3] = { TREG_EAX
, TREG_EDX
, TREG_ECX
};
325 /* Generate function call. The function address is pushed first, then
326 all the parameters in call order. This functions pops all the
327 parameters and the function address. */
328 void gfunc_call(int nb_args
)
330 int size
, align
, r
, args_size
, i
, func_call
;
334 for(i
= 0;i
< nb_args
; i
++) {
335 if ((vtop
->type
.t
& VT_BTYPE
) == VT_STRUCT
) {
336 size
= type_size(&vtop
->type
, &align
);
337 /* align to stack align size */
338 size
= (size
+ 3) & ~3;
339 /* allocate the necessary size on stack */
340 oad(0xec81, size
); /* sub $xxx, %esp */
341 /* generate structure store */
343 o(0x89); /* mov %esp, r */
345 vset(&vtop
->type
, r
| VT_LVAL
, 0);
349 } else if (is_float(vtop
->type
.t
)) {
350 gv(RC_FLOAT
); /* only one float register */
351 if ((vtop
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
353 else if ((vtop
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
357 oad(0xec81, size
); /* sub $xxx, %esp */
361 o(0x5cd9 + size
- 4); /* fstp[s|l] 0(%esp) */
366 /* simple type (currently always same size) */
367 /* XXX: implicit cast ? */
369 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
371 o(0x50 + vtop
->r2
); /* push r */
375 o(0x50 + r
); /* push r */
380 save_regs(0); /* save used temporary registers */
381 func_sym
= vtop
->type
.ref
;
382 func_call
= func_sym
->r
;
384 if (func_call
>= FUNC_FASTCALL1
&& func_call
<= FUNC_FASTCALL3
) {
385 int fastcall_nb_regs
;
386 fastcall_nb_regs
= func_call
- FUNC_FASTCALL1
+ 1;
387 for(i
= 0;i
< fastcall_nb_regs
; i
++) {
390 o(0x58 + fastcall_regs
[i
]); /* pop r */
391 /* XXX: incorrect for struct/floats */
396 if (args_size
&& func_sym
->r
!= FUNC_STDCALL
)
401 /* generate function prolog of type 't' */
402 void gfunc_prolog(CType
*func_type
)
404 int addr
, align
, size
, func_call
, fastcall_nb_regs
;
405 int param_index
, param_addr
;
409 sym
= func_type
->ref
;
413 if (func_call
>= FUNC_FASTCALL1
&& func_call
<= FUNC_FASTCALL3
) {
414 fastcall_nb_regs
= func_call
- FUNC_FASTCALL1
+ 1;
416 fastcall_nb_regs
= 0;
420 o(0xe58955); /* push %ebp, mov %esp, %ebp */
421 func_sub_sp_offset
= oad(0xec81, 0); /* sub $xxx, %esp */
423 /* if the function returns a structure, then add an
424 implicit pointer parameter */
426 if ((func_vt
.t
& VT_BTYPE
) == VT_STRUCT
) {
427 /* XXX: fastcall case ? */
432 /* define parameters */
433 while ((sym
= sym
->next
) != NULL
) {
435 size
= type_size(type
, &align
);
436 size
= (size
+ 3) & ~3;
437 #ifdef FUNC_STRUCT_PARAM_AS_PTR
438 /* structs are passed as pointer */
439 if ((type
->t
& VT_BTYPE
) == VT_STRUCT
) {
443 if (param_index
< fastcall_nb_regs
) {
444 /* save FASTCALL register */
447 gen_modrm(fastcall_regs
[param_index
], VT_LOCAL
, NULL
, loc
);
453 sym_push(sym
->v
& ~SYM_FIELD
, type
,
454 VT_LOCAL
| VT_LVAL
, param_addr
);
458 /* pascal type call ? */
459 if (func_call
== FUNC_STDCALL
)
460 func_ret_sub
= addr
- 8;
462 /* leave some room for bound checking code */
463 if (do_bounds_check
) {
464 oad(0xb8, 0); /* lbound section pointer */
465 oad(0xb8, 0); /* call to function */
466 func_bound_offset
= lbounds_section
->data_offset
;
470 /* generate function epilog */
471 void gfunc_epilog(void)
473 #ifdef CONFIG_TCC_BCHECK
474 if (do_bounds_check
&& func_bound_offset
!= lbounds_section
->data_offset
) {
478 /* add end of table info */
479 bounds_ptr
= section_ptr_add(lbounds_section
, sizeof(int));
481 /* generate bound local allocation */
483 ind
= func_sub_sp_offset
+ 4;
484 sym_data
= get_sym_ref(&char_pointer_type
, lbounds_section
,
485 func_bound_offset
, lbounds_section
->data_offset
);
486 greloc(cur_text_section
, sym_data
,
488 oad(0xb8, 0); /* mov %eax, xxx */
489 sym
= external_global_sym(TOK___bound_local_new
, &func_old_type
, 0);
490 greloc(cur_text_section
, sym
,
491 ind
+ 1, R_386_PC32
);
494 /* generate bound check local freeing */
495 o(0x5250); /* save returned value, if any */
496 greloc(cur_text_section
, sym_data
,
498 oad(0xb8, 0); /* mov %eax, xxx */
499 sym
= external_global_sym(TOK___bound_local_delete
, &func_old_type
, 0);
500 greloc(cur_text_section
, sym
,
501 ind
+ 1, R_386_PC32
);
503 o(0x585a); /* restore returned value, if any */
507 if (func_ret_sub
== 0) {
512 g(func_ret_sub
>> 8);
514 /* align local size to word & save local variables */
515 *(int *)(cur_text_section
->data
+ func_sub_sp_offset
) = (-loc
+ 3) & -4;
518 /* generate a jump to a label */
521 return psym(0xe9, t
);
524 /* generate a jump to a fixed address */
525 void gjmp_addr(int a
)
533 oad(0xe9, a
- ind
- 5);
537 /* generate a test. set 'inv' to invert test. Stack entry is popped */
538 int gtst(int inv
, int t
)
542 v
= vtop
->r
& VT_VALMASK
;
544 /* fast case : can jump directly since flags are set */
546 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
547 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
548 /* && or || optimization */
549 if ((v
& 1) == inv
) {
550 /* insert vtop->c jump list in t */
553 p
= (int *)(cur_text_section
->data
+ *p
);
561 if (is_float(vtop
->type
.t
)) {
565 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
566 /* constant jmp optimization */
567 if ((vtop
->c
.i
!= 0) != inv
)
574 t
= psym(0x85 ^ inv
, t
);
581 /* generate an integer binary operation */
588 case TOK_ADDC1
: /* add with carry generation */
591 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
598 /* XXX: generate inc and dec for smaller code ? */
600 o(0xc0 | (opc
<< 3) | r
);
604 oad(0xc0 | (opc
<< 3) | r
, c
);
610 o((opc
<< 3) | 0x01);
611 o(0xc0 + r
+ fr
* 8);
614 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
620 case TOK_SUBC1
: /* sub with carry generation */
623 case TOK_ADDC2
: /* add with carry use */
626 case TOK_SUBC2
: /* sub with carry use */
643 o(0xaf0f); /* imul fr, r */
644 o(0xc0 + fr
+ r
* 8);
655 opc
= 0xc0 | (opc
<< 3);
656 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
661 c
= vtop
->c
.i
& 0x1f;
662 o(0xc1); /* shl/shr/sar $xxx, r */
666 /* we generate the shift in ecx */
669 o(0xd3); /* shl/shr/sar %cl, r */
680 /* first operand must be in eax */
681 /* XXX: need better constraint for second operand */
687 if (op
== TOK_UMULL
) {
688 o(0xf7); /* mul fr */
693 if (op
== TOK_UDIV
|| op
== TOK_UMOD
) {
694 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
697 o(0xf799); /* cltd, idiv fr, %eax */
700 if (op
== '%' || op
== TOK_UMOD
)
713 /* generate a floating point operation 'v = t1 op t2' instruction. The
714 two operands are guaranted to have the same floating point type */
715 /* XXX: need to use ST1 too */
718 int a
, ft
, fc
, swapped
, r
;
720 /* convert constants to memory references */
721 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
726 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
729 /* must put at least one value in the floating point register */
730 if ((vtop
[-1].r
& VT_LVAL
) &&
731 (vtop
[0].r
& VT_LVAL
)) {
737 /* swap the stack if needed so that t1 is the register and t2 is
738 the memory reference */
739 if (vtop
[-1].r
& VT_LVAL
) {
743 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
744 /* load on stack second operand */
745 load(TREG_ST0
, vtop
);
746 save_reg(TREG_EAX
); /* eax is used by FP comparison code */
747 if (op
== TOK_GE
|| op
== TOK_GT
)
749 else if (op
== TOK_EQ
|| op
== TOK_NE
)
752 o(0xc9d9); /* fxch %st(1) */
753 o(0xe9da); /* fucompp */
754 o(0xe0df); /* fnstsw %ax */
756 o(0x45e480); /* and $0x45, %ah */
757 o(0x40fC80); /* cmp $0x40, %ah */
758 } else if (op
== TOK_NE
) {
759 o(0x45e480); /* and $0x45, %ah */
760 o(0x40f480); /* xor $0x40, %ah */
762 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
763 o(0x05c4f6); /* test $0x05, %ah */
766 o(0x45c4f6); /* test $0x45, %ah */
773 /* no memory reference possible for long double operations */
774 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
) {
775 load(TREG_ST0
, vtop
);
800 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
801 o(0xde); /* fxxxp %st, %st(1) */
804 /* if saved lvalue, then we must reload it */
806 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
810 v1
.r
= VT_LOCAL
| VT_LVAL
;
816 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
820 gen_modrm(a
, r
, vtop
->sym
, fc
);
826 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
827 and 'long long' cases. */
828 void gen_cvt_itof(int t
)
832 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
833 /* signed long long to float/double/long double (unsigned case
834 is handled generically) */
835 o(0x50 + vtop
->r2
); /* push r2 */
836 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
837 o(0x242cdf); /* fildll (%esp) */
838 o(0x08c483); /* add $8, %esp */
839 } else if ((vtop
->type
.t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
840 (VT_INT
| VT_UNSIGNED
)) {
841 /* unsigned int to float/double/long double */
842 o(0x6a); /* push $0 */
844 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
845 o(0x242cdf); /* fildll (%esp) */
846 o(0x08c483); /* add $8, %esp */
848 /* int to float/double/long double */
849 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
850 o(0x2404db); /* fildl (%esp) */
851 o(0x04c483); /* add $4, %esp */
856 /* convert fp to int 't' type */
857 /* XXX: handle long long case */
858 void gen_cvt_ftoi(int t
)
864 ushort_type
.t
= VT_SHORT
| VT_UNSIGNED
;
872 o(0x2dd9); /* ldcw xxx */
873 sym
= external_global_sym(TOK___tcc_int_fpu_control
,
874 &ushort_type
, VT_LVAL
);
875 greloc(cur_text_section
, sym
,
879 oad(0xec81, size
); /* sub $xxx, %esp */
881 o(0x1cdb); /* fistpl */
883 o(0x3cdf); /* fistpll */
885 o(0x2dd9); /* ldcw xxx */
886 sym
= external_global_sym(TOK___tcc_fpu_control
,
887 &ushort_type
, VT_LVAL
);
888 greloc(cur_text_section
, sym
,
893 o(0x58 + r
); /* pop r */
896 vtop
->r
= r
; /* mark reg as used */
897 r2
= get_reg(RC_INT
);
898 o(0x58 + r2
); /* pop r2 */
901 o(0x04c483); /* add $4, %esp */
907 /* convert from one floating point type to another */
908 void gen_cvt_ftof(int t
)
910 /* all we have to do on i386 is to put the float in a register */
914 /* computed goto support */
921 /* bound check support functions */
922 #ifdef CONFIG_TCC_BCHECK
924 /* generate a bounded pointer addition */
925 void gen_bounded_ptr_add(void)
929 /* prepare fast i386 function call (args in eax and edx) */
931 /* save all temporary registers */
934 /* do a fast function call */
935 sym
= external_global_sym(TOK___bound_ptr_add
, &func_old_type
, 0);
936 greloc(cur_text_section
, sym
,
937 ind
+ 1, R_386_PC32
);
939 /* returned pointer is in eax */
941 vtop
->r
= TREG_EAX
| VT_BOUNDED
;
942 /* address of bounding function call point */
943 vtop
->c
.ul
= (cur_text_section
->reloc
->data_offset
- sizeof(Elf32_Rel
));
946 /* patch pointer addition in vtop so that pointer dereferencing is
948 void gen_bounded_ptr_deref(void)
956 /* XXX: put that code in generic part of tcc */
957 if (!is_float(vtop
->type
.t
)) {
958 if (vtop
->r
& VT_LVAL_BYTE
)
960 else if (vtop
->r
& VT_LVAL_SHORT
)
964 size
= type_size(&vtop
->type
, &align
);
966 case 1: func
= TOK___bound_ptr_indir1
; break;
967 case 2: func
= TOK___bound_ptr_indir2
; break;
968 case 4: func
= TOK___bound_ptr_indir4
; break;
969 case 8: func
= TOK___bound_ptr_indir8
; break;
970 case 12: func
= TOK___bound_ptr_indir12
; break;
971 case 16: func
= TOK___bound_ptr_indir16
; break;
973 error("unhandled size when derefencing bounded pointer");
978 /* patch relocation */
979 /* XXX: find a better solution ? */
980 rel
= (Elf32_Rel
*)(cur_text_section
->reloc
->data
+ vtop
->c
.ul
);
981 sym
= external_global_sym(func
, &func_old_type
, 0);
983 put_extern_sym(sym
, NULL
, 0, 0);
984 rel
->r_info
= ELF32_R_INFO(sym
->c
, ELF32_R_TYPE(rel
->r_info
));
988 /* end of X86 code generator */
989 /*************************************************************/