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 o(0x05eb); /* 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
->type
.t
& VT_BTYPE
) == VT_STRUCT
) {
297 size
= type_size(&vtop
->type
, &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
->type
, r
| VT_LVAL
, 0);
309 c
->args_size
+= size
;
310 } else if (is_float(vtop
->type
.t
)) {
311 gv(RC_FLOAT
); /* only one float register */
312 if ((vtop
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
314 else if ((vtop
->type
.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
->type
.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 /* 'is_jmp' is '1' if it is a jump */
353 static void gcall_or_jmp(int is_jmp
)
356 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
358 if (vtop
->r
& VT_SYM
) {
359 /* relocation case */
360 greloc(cur_text_section
, vtop
->sym
,
361 ind
+ 1, R_386_PC32
);
363 /* put an empty PC32 relocation */
364 put_elf_reloc(symtab_section
, cur_text_section
,
365 ind
+ 1, R_386_PC32
, 0);
367 oad(0xe8 + is_jmp
, vtop
->c
.ul
- 4); /* call/jmp im */
369 /* otherwise, indirect call */
371 o(0xff); /* call/jmp *r */
372 o(0xd0 + r
+ (is_jmp
<< 4));
376 /* generate function call with address in (vtop->t, vtop->c) and free function
377 context. Stack entry is popped */
378 void gfunc_call(GFuncContext
*c
)
381 if (c
->args_size
&& c
->func_call
== FUNC_CDECL
)
382 gadd_sp(c
->args_size
);
386 /* generate function prolog of type 't' */
387 void gfunc_prolog(CType
*func_type
)
389 int addr
, align
, size
, func_call
;
393 sym
= func_type
->ref
;
396 /* if the function returns a structure, then add an
397 implicit pointer parameter */
399 if ((func_vt
.t
& VT_BTYPE
) == VT_STRUCT
) {
403 /* define parameters */
404 while ((sym
= sym
->next
) != NULL
) {
406 sym_push(sym
->v
& ~SYM_FIELD
, type
,
407 VT_LOCAL
| VT_LVAL
, addr
);
408 size
= type_size(type
, &align
);
409 size
= (size
+ 3) & ~3;
410 #ifdef FUNC_STRUCT_PARAM_AS_PTR
411 /* structs are passed as pointer */
412 if ((type
->t
& VT_BTYPE
) == VT_STRUCT
) {
419 /* pascal type call ? */
420 if (func_call
== FUNC_STDCALL
)
421 func_ret_sub
= addr
- 8;
422 o(0xe58955); /* push %ebp, mov %esp, %ebp */
423 func_sub_sp_offset
= oad(0xec81, 0); /* sub $xxx, %esp */
424 /* leave some room for bound checking code */
425 if (do_bounds_check
) {
426 oad(0xb8, 0); /* lbound section pointer */
427 oad(0xb8, 0); /* call to function */
428 func_bound_offset
= lbounds_section
->data_offset
;
432 /* generate function epilog */
433 void gfunc_epilog(void)
435 #ifdef CONFIG_TCC_BCHECK
436 if (do_bounds_check
&& func_bound_offset
!= lbounds_section
->data_offset
) {
440 /* add end of table info */
441 bounds_ptr
= section_ptr_add(lbounds_section
, sizeof(int));
443 /* generate bound local allocation */
445 ind
= func_sub_sp_offset
+ 4;
446 sym_data
= get_sym_ref(&char_pointer_type
, lbounds_section
,
447 func_bound_offset
, lbounds_section
->data_offset
);
448 greloc(cur_text_section
, sym_data
,
450 oad(0xb8, 0); /* mov %eax, xxx */
451 sym
= external_global_sym(TOK___bound_local_new
, &func_old_type
, 0);
452 greloc(cur_text_section
, sym
,
453 ind
+ 1, R_386_PC32
);
456 /* generate bound check local freeing */
457 o(0x5250); /* save returned value, if any */
458 greloc(cur_text_section
, sym_data
,
460 oad(0xb8, 0); /* mov %eax, xxx */
461 sym
= external_global_sym(TOK___bound_local_delete
, &func_old_type
, 0);
462 greloc(cur_text_section
, sym
,
463 ind
+ 1, R_386_PC32
);
465 o(0x585a); /* restore returned value, if any */
469 if (func_ret_sub
== 0) {
474 g(func_ret_sub
>> 8);
476 /* align local size to word & save local variables */
477 *(int *)(cur_text_section
->data
+ func_sub_sp_offset
) = (-loc
+ 3) & -4;
480 /* generate a jump to a label */
483 return psym(0xe9, t
);
486 /* generate a jump to a fixed address */
487 void gjmp_addr(int a
)
495 oad(0xe9, a
- ind
- 5);
499 /* generate a test. set 'inv' to invert test. Stack entry is popped */
500 int gtst(int inv
, int t
)
503 v
= vtop
->r
& VT_VALMASK
;
505 /* fast case : can jump directly since flags are set */
507 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
508 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
509 /* && or || optimization */
510 if ((v
& 1) == inv
) {
511 /* insert vtop->c jump list in t */
514 p
= (int *)(cur_text_section
->data
+ *p
);
522 if (is_float(vtop
->type
.t
)) {
526 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
527 /* constant jmp optimization */
528 if ((vtop
->c
.i
!= 0) != inv
)
535 t
= psym(0x85 ^ inv
, t
);
542 /* generate an integer binary operation */
549 case TOK_ADDC1
: /* add with carry generation */
552 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
559 /* XXX: generate inc and dec for smaller code ? */
561 o(0xc0 | (opc
<< 3) | r
);
565 oad(0xc0 | (opc
<< 3) | r
, c
);
571 o((opc
<< 3) | 0x01);
572 o(0xc0 + r
+ fr
* 8);
575 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
581 case TOK_SUBC1
: /* sub with carry generation */
584 case TOK_ADDC2
: /* add with carry use */
587 case TOK_SUBC2
: /* sub with carry use */
604 o(0xaf0f); /* imul fr, r */
605 o(0xc0 + fr
+ r
* 8);
616 opc
= 0xc0 | (opc
<< 3);
617 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
622 c
= vtop
->c
.i
& 0x1f;
623 o(0xc1); /* shl/shr/sar $xxx, r */
627 /* we generate the shift in ecx */
630 o(0xd3); /* shl/shr/sar %cl, r */
641 /* first operand must be in eax */
642 /* XXX: need better constraint for second operand */
648 if (op
== TOK_UMULL
) {
649 o(0xf7); /* mul fr */
654 if (op
== TOK_UDIV
|| op
== TOK_UMOD
) {
655 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
658 o(0xf799); /* cltd, idiv fr, %eax */
661 if (op
== '%' || op
== TOK_UMOD
)
674 /* generate a floating point operation 'v = t1 op t2' instruction. The
675 two operands are guaranted to have the same floating point type */
676 /* XXX: need to use ST1 too */
679 int a
, ft
, fc
, swapped
, r
;
681 /* convert constants to memory references */
682 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
687 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
690 /* must put at least one value in the floating point register */
691 if ((vtop
[-1].r
& VT_LVAL
) &&
692 (vtop
[0].r
& VT_LVAL
)) {
698 /* swap the stack if needed so that t1 is the register and t2 is
699 the memory reference */
700 if (vtop
[-1].r
& VT_LVAL
) {
704 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
705 /* load on stack second operand */
707 save_reg(REG_EAX
); /* eax is used by FP comparison code */
708 if (op
== TOK_GE
|| op
== TOK_GT
)
710 else if (op
== TOK_EQ
|| op
== TOK_NE
)
713 o(0xc9d9); /* fxch %st(1) */
714 o(0xe9da); /* fucompp */
715 o(0xe0df); /* fnstsw %ax */
717 o(0x45e480); /* and $0x45, %ah */
718 o(0x40fC80); /* cmp $0x40, %ah */
719 } else if (op
== TOK_NE
) {
720 o(0x45e480); /* and $0x45, %ah */
721 o(0x40f480); /* xor $0x40, %ah */
723 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
724 o(0x05c4f6); /* test $0x05, %ah */
727 o(0x45c4f6); /* test $0x45, %ah */
734 /* no memory reference possible for long double operations */
735 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
) {
761 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
762 o(0xde); /* fxxxp %st, %st(1) */
765 /* if saved lvalue, then we must reload it */
767 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
771 v1
.r
= VT_LOCAL
| VT_LVAL
;
777 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
781 gen_modrm(a
, r
, vtop
->sym
, fc
);
787 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
788 and 'long long' cases. */
789 void gen_cvt_itof(int t
)
793 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
794 /* signed long long to float/double/long double (unsigned case
795 is handled generically) */
796 o(0x50 + vtop
->r2
); /* push r2 */
797 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
798 o(0x242cdf); /* fildll (%esp) */
799 o(0x08c483); /* add $8, %esp */
800 } else if ((vtop
->type
.t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
801 (VT_INT
| VT_UNSIGNED
)) {
802 /* unsigned int to float/double/long double */
803 o(0x6a); /* push $0 */
805 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
806 o(0x242cdf); /* fildll (%esp) */
807 o(0x08c483); /* add $8, %esp */
809 /* int to float/double/long double */
810 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
811 o(0x2404db); /* fildl (%esp) */
812 o(0x04c483); /* add $4, %esp */
817 /* convert fp to int 't' type */
818 /* XXX: handle long long case */
819 void gen_cvt_ftoi(int t
)
825 ushort_type
.t
= VT_SHORT
| VT_UNSIGNED
;
833 o(0x2dd9); /* ldcw xxx */
834 sym
= external_global_sym(TOK___tcc_int_fpu_control
,
835 &ushort_type
, VT_LVAL
);
836 greloc(cur_text_section
, sym
,
840 oad(0xec81, size
); /* sub $xxx, %esp */
842 o(0x1cdb); /* fistpl */
844 o(0x3cdf); /* fistpll */
846 o(0x2dd9); /* ldcw xxx */
847 sym
= external_global_sym(TOK___tcc_fpu_control
,
848 &ushort_type
, VT_LVAL
);
849 greloc(cur_text_section
, sym
,
854 o(0x58 + r
); /* pop r */
857 vtop
->r
= r
; /* mark reg as used */
858 r2
= get_reg(RC_INT
);
859 o(0x58 + r2
); /* pop r2 */
862 o(0x04c483); /* add $4, %esp */
868 /* convert from one floating point type to another */
869 void gen_cvt_ftof(int t
)
871 /* all we have to do on i386 is to put the float in a register */
875 /* computed goto support */
882 /* bound check support functions */
883 #ifdef CONFIG_TCC_BCHECK
885 /* generate a bounded pointer addition */
886 void gen_bounded_ptr_add(void)
890 /* prepare fast i386 function call (args in eax and edx) */
892 /* save all temporary registers */
895 /* do a fast function call */
896 sym
= external_global_sym(TOK___bound_ptr_add
, &func_old_type
, 0);
897 greloc(cur_text_section
, sym
,
898 ind
+ 1, R_386_PC32
);
900 /* returned pointer is in eax */
902 vtop
->r
= REG_EAX
| VT_BOUNDED
;
903 /* address of bounding function call point */
904 vtop
->c
.ul
= (cur_text_section
->reloc
->data_offset
- sizeof(Elf32_Rel
));
907 /* patch pointer addition in vtop so that pointer dereferencing is
909 void gen_bounded_ptr_deref(void)
917 /* XXX: put that code in generic part of tcc */
918 if (!is_float(vtop
->type
.t
)) {
919 if (vtop
->r
& VT_LVAL_BYTE
)
921 else if (vtop
->r
& VT_LVAL_SHORT
)
925 size
= type_size(&vtop
->type
, &align
);
927 case 1: func
= TOK___bound_ptr_indir1
; break;
928 case 2: func
= TOK___bound_ptr_indir2
; break;
929 case 4: func
= TOK___bound_ptr_indir4
; break;
930 case 8: func
= TOK___bound_ptr_indir8
; break;
931 case 12: func
= TOK___bound_ptr_indir12
; break;
932 case 16: func
= TOK___bound_ptr_indir16
; break;
934 error("unhandled size when derefencing bounded pointer");
939 /* patch relocation */
940 /* XXX: find a better solution ? */
941 rel
= (Elf32_Rel
*)(cur_text_section
->reloc
->data
+ vtop
->c
.ul
);
942 sym
= external_global_sym(func
, &func_old_type
, 0);
944 put_extern_sym(sym
, NULL
, 0, 0);
945 rel
->r_info
= ELF32_R_INFO(sym
->c
, ELF32_R_TYPE(rel
->r_info
));
949 /* end of X86 code generator */
950 /*************************************************************/