2 * X86 code generator for TCC
4 * Copyright (c) 2001 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 /* function call context */
72 typedef struct GFuncContext
{
74 int func_call
; /* func call type (FUNC_STDCALL or FUNC_CDECL) */
77 /******************************************************/
79 static int *func_sub_sp_ptr
;
80 static unsigned char *func_bound_ptr
;
81 static int func_ret_sub
;
104 /* patch relocation entry with value 'val' */
105 void greloc_patch1(Reloc
*p
, int val
)
109 *(int *)p
->addr
= val
;
112 *(int *)p
->addr
= val
- p
->addr
- 4;
117 /* output a symbol and patch all calls to it */
122 n
= *(int *)t
; /* next value */
123 *(int *)t
= a
- t
- 4;
133 /* psym is used to put an instruction with a data field which is a
134 reference to a symbol. It is in fact the same as oad ! */
137 /* instruction + 4 bytes data. Return the address of the data */
138 int oad(int c
, int s
)
147 /* output constant with relocation if 'r & VT_FORWARD' is true */
148 void gen_addr32(int r
, int c
)
150 if (!(r
& VT_FORWARD
)) {
153 greloc((Sym
*)c
, ind
, RELOC_ADDR32
);
158 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
160 void gen_modrm(int op_reg
, int r
, int c
)
162 op_reg
= op_reg
<< 3;
163 if ((r
& VT_VALMASK
) == VT_CONST
) {
164 /* constant memory reference */
167 } else if ((r
& VT_VALMASK
) == VT_LOCAL
) {
168 /* currently, we use only ebp as base */
170 /* short reference */
174 oad(0x85 | op_reg
, c
);
177 g(0x00 | op_reg
| (r
& VT_VALMASK
));
182 /* load 'r' from value 'sv' */
183 void load(int r
, SValue
*sv
)
185 int v
, t
, ft
, fc
, fr
;
194 if (v
== VT_LLOCAL
) {
196 v1
.r
= VT_LOCAL
| VT_LVAL
;
201 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
204 } else if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
207 } else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
210 } else if ((ft
& VT_TYPE
) == VT_BYTE
) {
211 o(0xbe0f); /* movsbl */
212 } else if ((ft
& VT_TYPE
) == (VT_BYTE
| VT_UNSIGNED
)) {
213 o(0xb60f); /* movzbl */
214 } else if ((ft
& VT_TYPE
) == VT_SHORT
) {
215 o(0xbf0f); /* movswl */
216 } else if ((ft
& VT_TYPE
) == (VT_SHORT
| VT_UNSIGNED
)) {
217 o(0xb70f); /* movzwl */
221 gen_modrm(r
, fr
, fc
);
224 o(0xb8 + r
); /* mov $xx, r */
226 } else if (v
== VT_LOCAL
) {
227 o(0x8d); /* lea xxx(%ebp), r */
228 gen_modrm(r
, VT_LOCAL
, fc
);
229 } else if (v
== VT_CMP
) {
230 oad(0xb8 + r
, 0); /* mov $0, r */
231 o(0x0f); /* setxx %br */
234 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
236 oad(0xb8 + r
, t
); /* mov $1, r */
237 oad(0xe9, 5); /* jmp after */
239 oad(0xb8 + r
, t
^ 1); /* mov $0, r */
242 o(0xc0 + r
+ v
* 8); /* mov v, r */
247 /* store register 'r' in lvalue 'v' */
248 void store(int r
, SValue
*v
)
254 fr
= v
->r
& VT_VALMASK
;
256 /* XXX: incorrect if float reg to reg */
257 if (bt
== VT_FLOAT
) {
260 } else if (bt
== VT_DOUBLE
) {
263 } else if (bt
== VT_LDOUBLE
) {
264 o(0xc0d9); /* fld %st(0) */
275 if (fr
== VT_CONST
||
278 gen_modrm(r
, v
->r
, fc
);
279 } else if (fr
!= r
) {
280 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
284 /* start function call and return function call context */
285 void gfunc_start(GFuncContext
*c
, int func_call
)
288 c
->func_call
= func_call
;
291 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
293 void gfunc_param(GFuncContext
*c
)
297 if ((vtop
->t
& VT_BTYPE
) == VT_STRUCT
) {
298 size
= type_size(vtop
->t
, &align
);
299 /* align to stack align size */
300 size
= (size
+ 3) & ~3;
301 /* allocate the necessary size on stack */
302 oad(0xec81, size
); /* sub $xxx, %esp */
303 /* generate structure store */
305 o(0x89); /* mov %esp, r */
307 vset(vtop
->t
, r
| VT_LVAL
, 0);
310 c
->args_size
+= size
;
311 } else if (is_float(vtop
->t
)) {
312 gv(RC_FLOAT
); /* only one float register */
313 if ((vtop
->t
& VT_BTYPE
) == VT_FLOAT
)
315 else if ((vtop
->t
& VT_BTYPE
) == VT_DOUBLE
)
319 oad(0xec81, size
); /* sub $xxx, %esp */
323 o(0x5cd9 + size
- 4); /* fstp[s|l] 0(%esp) */
326 c
->args_size
+= size
;
328 /* simple type (currently always same size) */
329 /* XXX: implicit cast ? */
331 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
333 o(0x50 + vtop
->r2
); /* push r */
337 o(0x50 + r
); /* push r */
338 c
->args_size
+= size
;
343 /* generate function call with address in (vtop->t, vtop->c) and free function
344 context. Stack entry is popped */
345 void gfunc_call(GFuncContext
*c
)
348 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
350 /* forward reference */
351 if (vtop
->r
& VT_FORWARD
) {
352 greloc(vtop
->c
.sym
, ind
+ 1, RELOC_REL32
);
355 oad(0xe8, vtop
->c
.ul
- ind
- 5);
358 /* otherwise, indirect call */
360 o(0xff); /* call *r */
363 if (c
->args_size
&& c
->func_call
== FUNC_CDECL
)
364 oad(0xc481, c
->args_size
); /* add $xxx, %esp */
368 /* generate function prolog of type 't' */
369 void gfunc_prolog(int t
)
371 int addr
, align
, size
, u
, func_call
;
374 sym
= sym_find((unsigned)t
>> VT_STRUCT_SHIFT
);
377 /* if the function returns a structure, then add an
378 implicit pointer parameter */
380 if ((func_vt
& VT_BTYPE
) == VT_STRUCT
) {
384 /* define parameters */
385 while ((sym
= sym
->next
) != NULL
) {
387 sym_push(sym
->v
& ~SYM_FIELD
, u
,
388 VT_LOCAL
| VT_LVAL
, addr
);
389 size
= type_size(u
, &align
);
390 size
= (size
+ 3) & ~3;
391 #ifdef FUNC_STRUCT_PARAM_AS_PTR
392 /* structs are passed as pointer */
393 if ((u
& VT_BTYPE
) == VT_STRUCT
) {
400 /* pascal type call ? */
401 if (func_call
== FUNC_STDCALL
)
402 func_ret_sub
= addr
- 8;
403 o(0xe58955); /* push %ebp, mov %esp, %ebp */
404 func_sub_sp_ptr
= (int *)oad(0xec81, 0); /* sub $xxx, %esp */
405 /* leave some room for bound checking code */
406 if (do_bounds_check
) {
407 oad(0xb8, 0); /* lbound section pointer */
408 oad(0xb8, 0); /* call to function */
409 func_bound_ptr
= lbounds_section
->data_ptr
;
413 /* generate function epilog */
414 void gfunc_epilog(void)
416 #ifdef CONFIG_TCC_BCHECK
417 if (do_bounds_check
&& func_bound_ptr
!= lbounds_section
->data_ptr
) {
420 /* add end of table info */
421 bounds_ptr
= (int *)lbounds_section
->data_ptr
;
423 lbounds_section
->data_ptr
= (unsigned char *)bounds_ptr
;
424 /* generate bound local allocation */
426 ind
= (int)func_sub_sp_ptr
+ 4;
427 oad(0xb8, (int)func_bound_ptr
); /* mov %eax, xxx */
428 oad(0xe8, (int)__bound_local_new
- ind
- 5);
430 /* generate bound check local freeing */
431 o(0x5250); /* save returned value, if any */
432 oad(0xb8, (int)func_bound_ptr
); /* mov %eax, xxx */
433 oad(0xe8, (int)__bound_local_delete
- ind
- 5);
434 o(0x585a); /* restore returned value, if any */
438 if (func_ret_sub
== 0) {
443 g(func_ret_sub
>> 8);
445 /* align local size to word & save local variables */
446 *func_sub_sp_ptr
= (-loc
+ 3) & -4;
449 /* generate a jump to a label */
452 return psym(0xe9, t
);
455 /* generate a jump to a fixed address */
456 void gjmp_addr(int a
)
458 oad(0xe9, a
- ind
- 5);
461 /* generate a test. set 'inv' to invert test. Stack entry is popped */
462 int gtst(int inv
, int t
)
465 v
= vtop
->r
& VT_VALMASK
;
467 /* fast case : can jump directly since flags are set */
469 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
470 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
471 /* && or || optimization */
472 if ((v
& 1) == inv
) {
473 /* insert vtop->c jump list in t */
484 if (is_float(vtop
->t
)) {
488 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_FORWARD
)) == VT_CONST
) {
489 /* constant jmp optimization */
490 if ((vtop
->c
.i
!= 0) != inv
)
497 t
= psym(0x85 ^ inv
, t
);
504 /* generate an integer binary operation */
511 case TOK_ADDC1
: /* add with carry generation */
514 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_FORWARD
)) == VT_CONST
) {
521 /* XXX: generate inc and dec for smaller code ? */
523 o(0xc0 | (opc
<< 3) | r
);
527 oad(0xc0 | (opc
<< 3) | r
, c
);
533 o((opc
<< 3) | 0x01);
534 o(0xc0 + r
+ fr
* 8);
537 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
539 vset(VT_INT
, VT_CMP
, op
);
543 case TOK_SUBC1
: /* sub with carry generation */
546 case TOK_ADDC2
: /* add with carry use */
549 case TOK_SUBC2
: /* sub with carry use */
566 o(0xaf0f); /* imul fr, r */
567 o(0xc0 + fr
+ r
* 8);
578 opc
= 0xc0 | (opc
<< 3);
579 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_FORWARD
)) == VT_CONST
) {
584 c
= vtop
->c
.i
& 0x1f;
585 o(0xc1); /* shl/shr/sar $xxx, r */
589 /* we generate the shift in ecx */
592 o(0xd3); /* shl/shr/sar %cl, r */
603 /* first operand must be in eax */
604 /* XXX: need better constraint for second operand */
610 if (op
== TOK_UMULL
) {
611 o(0xf7); /* mul fr */
616 if (op
== TOK_UDIV
|| op
== TOK_UMOD
) {
617 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
620 o(0xf799); /* cltd, idiv fr, %eax */
623 if (op
== '%' || op
== TOK_UMOD
)
636 /* generate a floating point operation 'v = t1 op t2' instruction. The
637 two operands are guaranted to have the same floating point type */
638 /* XXX: need to use ST1 too */
641 int a
, ft
, fc
, swapped
, r
;
643 /* convert constants to memory references */
644 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
649 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
652 /* must put at least one value in the floating point register */
653 if ((vtop
[-1].r
& VT_LVAL
) &&
654 (vtop
[0].r
& VT_LVAL
)) {
660 /* swap the stack if needed so that t1 is the register and t2 is
661 the memory reference */
662 if (vtop
[-1].r
& VT_LVAL
) {
666 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
667 /* load on stack second operand */
669 save_reg(REG_EAX
); /* eax is used by FP comparison code */
670 if (op
== TOK_GE
|| op
== TOK_GT
)
672 else if (op
== TOK_EQ
|| op
== TOK_NE
)
675 o(0xc9d9); /* fxch %st(1) */
676 o(0xe9da); /* fucompp */
677 o(0xe0df); /* fnstsw %ax */
679 o(0x45e480); /* and $0x45, %ah */
680 o(0x40fC80); /* cmp $0x40, %ah */
681 } else if (op
== TOK_NE
) {
682 o(0x45e480); /* and $0x45, %ah */
683 o(0x40f480); /* xor $0x40, %ah */
685 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
686 o(0x05c4f6); /* test $0x05, %ah */
689 o(0x45c4f6); /* test $0x45, %ah */
696 /* no memory reference possible for long double operations */
697 if ((vtop
->t
& VT_BTYPE
) == VT_LDOUBLE
) {
723 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
724 o(0xde); /* fxxxp %st, %st(1) */
727 /* if saved lvalue, then we must reload it */
729 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
733 v1
.r
= VT_LOCAL
| VT_LVAL
;
739 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
749 /* FPU control word for rounding to nearest mode */
750 /* XXX: should move that into tcc lib support code ! */
751 static unsigned short __tcc_fpu_control
= 0x137f;
752 /* FPU control word for round to zero mode for int convertion */
753 static unsigned short __tcc_int_fpu_control
= 0x137f | 0x0c00;
755 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
756 and 'long long' cases. */
757 void gen_cvt_itof(int t
)
761 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
762 /* signed long long to float/double/long double (unsigned case
763 is handled generically) */
764 o(0x50 + vtop
->r2
); /* push r2 */
765 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
766 o(0x242cdf); /* fildll (%esp) */
767 o(0x08c483); /* add $8, %esp */
768 } else if ((vtop
->t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
769 (VT_INT
| VT_UNSIGNED
)) {
770 /* unsigned int to float/double/long double */
771 o(0x6a); /* push $0 */
773 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
774 o(0x242cdf); /* fildll (%esp) */
775 o(0x08c483); /* add $8, %esp */
777 /* int to float/double/long double */
778 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
779 o(0x2404db); /* fildl (%esp) */
780 o(0x04c483); /* add $4, %esp */
785 /* convert fp to int 't' type */
786 /* XXX: handle long long case */
787 void gen_cvt_ftoi(int t
)
797 oad(0x2dd9, (int)&__tcc_int_fpu_control
); /* ldcw xxx */
798 oad(0xec81, size
); /* sub $xxx, %esp */
800 o(0x1cdb); /* fistpl */
802 o(0x3cdf); /* fistpll */
804 oad(0x2dd9, (int)&__tcc_fpu_control
); /* ldcw xxx */
806 o(0x58 + r
); /* pop r */
809 vtop
->r
= r
; /* mark reg as used */
810 r2
= get_reg(RC_INT
);
811 o(0x58 + r2
); /* pop r2 */
814 o(0x04c483); /* add $4, %esp */
820 /* convert from one floating point type to another */
821 void gen_cvt_ftof(int t
)
823 /* all we have to do on i386 is to put the float in a register */
827 /* bound check support functions */
828 #ifdef CONFIG_TCC_BCHECK
830 /* generate a bounded pointer addition */
831 void gen_bounded_ptr_add(void)
834 /* prepare fast i386 function call (args in eax and edx) */
836 /* save all temporary registers */
839 /* do a fast function call */
841 oad(0xe8, (int)__bound_ptr_add
- ind
- 5);
842 /* returned pointer is in eax */
844 vtop
->r
= REG_EAX
| VT_BOUNDED
;
845 vtop
->c
.ul
= addr
; /* address of bounding function call point */
848 /* patch pointer addition in vtop so that pointer dereferencing is
850 void gen_bounded_ptr_deref(void)
853 int size
, align
, addr
;
856 /* XXX: put that code in generic part of tcc */
857 if (!is_float(vtop
->t
)) {
858 if (vtop
->r
& VT_LVAL_BYTE
)
860 else if (vtop
->r
& VT_LVAL_SHORT
)
864 size
= type_size(vtop
->t
, &align
);
866 case 1: func
= __bound_ptr_indir1
; break;
867 case 2: func
= __bound_ptr_indir2
; break;
868 case 4: func
= __bound_ptr_indir4
; break;
869 case 8: func
= __bound_ptr_indir8
; break;
870 case 12: func
= __bound_ptr_indir12
; break;
871 case 16: func
= __bound_ptr_indir16
; break;
873 error("unhandled size when derefencing bounded pointer");
879 *(int *)(addr
+ 1) = (int)func
- addr
- 5;
883 /* end of X86 code generator */
884 /*************************************************************/