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
{
76 /******************************************************/
78 static int *func_sub_sp_ptr
;
79 static unsigned char *func_bound_ptr
;
102 /* patch relocation entry with value 'val' */
103 void greloc_patch1(Reloc
*p
, int val
)
107 *(int *)p
->addr
= val
;
110 *(int *)p
->addr
= val
- p
->addr
- 4;
115 /* output a symbol and patch all calls to it */
120 n
= *(int *)t
; /* next value */
121 *(int *)t
= a
- t
- 4;
131 /* psym is used to put an instruction with a data field which is a
132 reference to a symbol. It is in fact the same as oad ! */
135 /* instruction + 4 bytes data. Return the address of the data */
136 int oad(int c
, int s
)
145 /* output constant with relocation if 'r & VT_FORWARD' is true */
146 void gen_addr32(int r
, int c
)
148 if (!(r
& VT_FORWARD
)) {
151 greloc((Sym
*)c
, ind
, RELOC_ADDR32
);
156 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
158 void gen_modrm(int op_reg
, int r
, int c
)
160 op_reg
= op_reg
<< 3;
161 if ((r
& VT_VALMASK
) == VT_CONST
) {
162 /* constant memory reference */
165 } else if ((r
& VT_VALMASK
) == VT_LOCAL
) {
166 /* currently, we use only ebp as base */
168 /* short reference */
172 oad(0x85 | op_reg
, c
);
175 g(0x00 | op_reg
| (r
& VT_VALMASK
));
180 /* load 'r' from value 'sv' */
181 void load(int r
, SValue
*sv
)
183 int v
, t
, ft
, fc
, fr
;
192 if (v
== VT_LLOCAL
) {
194 v1
.r
= VT_LOCAL
| VT_LVAL
;
199 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
202 } else if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
205 } else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
208 } else if ((ft
& VT_TYPE
) == VT_BYTE
)
209 o(0xbe0f); /* movsbl */
210 else if ((ft
& VT_TYPE
) == (VT_BYTE
| VT_UNSIGNED
))
211 o(0xb60f); /* movzbl */
212 else if ((ft
& VT_TYPE
) == VT_SHORT
)
213 o(0xbf0f); /* movswl */
214 else if ((ft
& VT_TYPE
) == (VT_SHORT
| VT_UNSIGNED
))
215 o(0xb70f); /* movzwl */
218 gen_modrm(r
, fr
, fc
);
221 o(0xb8 + r
); /* mov $xx, r */
223 } else if (v
== VT_LOCAL
) {
224 o(0x8d); /* lea xxx(%ebp), r */
225 gen_modrm(r
, VT_LOCAL
, fc
);
226 } else if (v
== VT_CMP
) {
227 oad(0xb8 + r
, 0); /* mov $0, r */
228 o(0x0f); /* setxx %br */
231 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
233 oad(0xb8 + r
, t
); /* mov $1, r */
234 oad(0xe9, 5); /* jmp after */
236 oad(0xb8 + r
, t
^ 1); /* mov $0, r */
239 o(0xc0 + r
+ v
* 8); /* mov v, r */
244 /* store register 'r' in lvalue 'v' */
245 void store(int r
, SValue
*v
)
251 fr
= v
->r
& VT_VALMASK
;
253 /* XXX: incorrect if float reg to reg */
254 if (bt
== VT_FLOAT
) {
257 } else if (bt
== VT_DOUBLE
) {
260 } else if (bt
== VT_LDOUBLE
) {
261 o(0xc0d9); /* fld %st(0) */
272 if (fr
== VT_CONST
||
275 gen_modrm(r
, v
->r
, fc
);
276 } else if (fr
!= r
) {
277 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
281 /* start function call and return function call context */
282 void gfunc_start(GFuncContext
*c
)
287 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
289 void gfunc_param(GFuncContext
*c
)
293 if ((vtop
->t
& VT_BTYPE
) == VT_STRUCT
) {
294 size
= type_size(vtop
->t
, &align
);
295 /* align to stack align size */
296 size
= (size
+ 3) & ~3;
297 /* allocate the necessary size on stack */
298 oad(0xec81, size
); /* sub $xxx, %esp */
299 /* generate structure store */
301 o(0x89); /* mov %esp, r */
306 c
->args_size
+= size
;
307 } else if (is_float(vtop
->t
)) {
308 gv(RC_FLOAT
); /* only one float register */
309 if ((vtop
->t
& VT_BTYPE
) == VT_FLOAT
)
311 else if ((vtop
->t
& VT_BTYPE
) == VT_DOUBLE
)
315 oad(0xec81, size
); /* sub $xxx, %esp */
319 o(0x5cd9 + size
- 4); /* fstp[s|l] 0(%esp) */
322 c
->args_size
+= size
;
324 /* simple type (currently always same size) */
325 /* XXX: implicit cast ? */
327 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
329 o(0x50 + vtop
->r2
); /* push r */
333 o(0x50 + r
); /* push r */
334 c
->args_size
+= size
;
339 /* generate function call with address in (vtop->t, vtop->c) and free function
340 context. Stack entry is popped */
341 void gfunc_call(GFuncContext
*c
)
344 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
346 /* forward reference */
347 if (vtop
->r
& VT_FORWARD
) {
348 greloc(vtop
->c
.sym
, ind
+ 1, RELOC_REL32
);
351 oad(0xe8, vtop
->c
.ul
- ind
- 5);
354 /* otherwise, indirect call */
356 o(0xff); /* call *r */
360 oad(0xc481, c
->args_size
); /* add $xxx, %esp */
364 /* generate function prolog of type 't' */
365 void gfunc_prolog(int t
)
367 int addr
, align
, size
, u
;
370 sym
= sym_find((unsigned)t
>> VT_STRUCT_SHIFT
);
372 /* if the function returns a structure, then add an
373 implicit pointer parameter */
375 if ((func_vt
& VT_BTYPE
) == VT_STRUCT
) {
379 /* define parameters */
380 while ((sym
= sym
->next
) != NULL
) {
382 sym_push(sym
->v
& ~SYM_FIELD
, u
,
383 VT_LOCAL
| VT_LVAL
, addr
);
384 size
= type_size(u
, &align
);
385 size
= (size
+ 3) & ~3;
386 #ifdef FUNC_STRUCT_PARAM_AS_PTR
387 /* structs are passed as pointer */
388 if ((u
& VT_BTYPE
) == VT_STRUCT
) {
394 o(0xe58955); /* push %ebp, mov %esp, %ebp */
395 func_sub_sp_ptr
= (int *)oad(0xec81, 0); /* sub $xxx, %esp */
396 /* leave some room for bound checking code */
397 if (do_bounds_check
) {
398 oad(0xb8, 0); /* lbound section pointer */
399 oad(0xb8, 0); /* call to function */
400 func_bound_ptr
= lbounds_section
->data_ptr
;
404 /* generate function epilog */
405 void gfunc_epilog(void)
407 if (do_bounds_check
&& func_bound_ptr
!= lbounds_section
->data_ptr
) {
410 /* add end of table info */
411 bounds_ptr
= (int *)lbounds_section
->data_ptr
;
413 lbounds_section
->data_ptr
= (unsigned char *)bounds_ptr
;
414 /* generate bound local allocation */
416 ind
= (int)func_sub_sp_ptr
+ 4;
417 oad(0xb8, (int)func_bound_ptr
); /* mov %eax, xxx */
418 oad(0xe8, (int)__bound_local_new
- ind
- 5);
420 /* generate bound check local freeing */
421 o(0x5250); /* save returned value, if any */
422 oad(0xb8, (int)func_bound_ptr
); /* mov %eax, xxx */
423 oad(0xe8, (int)__bound_local_delete
- ind
- 5);
424 o(0x585a); /* restore returned value, if any */
426 o(0xc3c9); /* leave, ret */
427 /* align local size to word & save local variables */
428 *func_sub_sp_ptr
= (-loc
+ 3) & -4;
433 return psym(0xe9, t
);
436 /* generate a test. set 'inv' to invert test. Stack entry is popped */
437 int gtst(int inv
, int t
)
440 v
= vtop
->r
& VT_VALMASK
;
442 /* fast case : can jump directly since flags are set */
444 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
445 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
446 /* && or || optimization */
447 if ((v
& 1) == inv
) {
448 /* insert vtop->c jump list in t */
459 if (is_float(vtop
->t
)) {
463 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_FORWARD
)) == VT_CONST
) {
464 /* constant jmp optimization */
465 if ((vtop
->c
.i
!= 0) != inv
)
472 t
= psym(0x85 ^ inv
, t
);
479 /* generate an integer binary operation */
486 case TOK_ADDC1
: /* add with carry generation */
489 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_FORWARD
)) == VT_CONST
) {
496 /* XXX: generate inc and dec for smaller code ? */
498 o(0xc0 | (opc
<< 3) | r
);
502 oad(0xc0 | (opc
<< 3) | r
, c
);
508 o((opc
<< 3) | 0x01);
509 o(0xc0 + r
+ fr
* 8);
512 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
514 vset(VT_INT
, VT_CMP
, op
);
518 case TOK_SUBC1
: /* sub with carry generation */
521 case TOK_ADDC2
: /* add with carry use */
524 case TOK_SUBC2
: /* sub with carry use */
541 o(0xaf0f); /* imul fr, r */
542 o(0xc0 + fr
+ r
* 8);
553 opc
= 0xc0 | (opc
<< 3);
554 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_FORWARD
)) == VT_CONST
) {
559 c
= vtop
->c
.i
& 0x1f;
560 o(0xc1); /* shl/shr/sar $xxx, r */
564 /* we generate the shift in ecx */
567 o(0xd3); /* shl/shr/sar %cl, r */
578 /* first operand must be in eax */
579 /* XXX: need better constraint for second operand */
585 if (op
== TOK_UMULL
) {
586 o(0xf7); /* mul fr */
591 if (op
== TOK_UDIV
|| op
== TOK_UMOD
) {
592 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
595 o(0xf799); /* cltd, idiv fr, %eax */
598 if (op
== '%' || op
== TOK_UMOD
)
611 /* generate a floating point operation 'v = t1 op t2' instruction. The
612 two operands are guaranted to have the same floating point type */
613 /* XXX: need to use ST1 too */
616 int a
, ft
, fc
, swapped
;
618 /* convert constants to memory references */
619 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
624 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
627 /* must put at least one value in the floating point register */
628 if ((vtop
[-1].r
& VT_LVAL
) &&
629 (vtop
[0].r
& VT_LVAL
)) {
634 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
635 /* load on stack second operand */
637 if (op
== TOK_GE
|| op
== TOK_GT
)
638 o(0xc9d9); /* fxch %st(1) */
639 o(0xe9da); /* fucompp */
640 o(0xe0df); /* fnstsw %ax */
642 o(0x45e480); /* and $0x45, %ah */
643 o(0x40fC80); /* cmp $0x40, %ah */
644 } else if (op
== TOK_NE
) {
645 o(0x45e480); /* and $0x45, %ah */
646 o(0x40f480); /* xor $0x40, %ah */
648 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
649 o(0x05c4f6); /* test $0x05, %ah */
652 o(0x45c4f6); /* test $0x45, %ah */
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 /* no memory reference possible for long double operations */
667 if ((vtop
->t
& VT_BTYPE
) == VT_LDOUBLE
) {
693 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
694 o(0xde); /* fxxxp %st, %st(1) */
697 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
701 gen_modrm(a
, vtop
->r
, fc
);
707 /* FPU control word for rounding to nearest mode */
708 /* XXX: should move that into tcc lib support code ! */
709 static unsigned short __tcc_fpu_control
= 0x137f;
710 /* FPU control word for round to zero mode for int convertion */
711 static unsigned short __tcc_int_fpu_control
= 0x137f | 0x0c00;
713 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
714 and 'long long' cases. */
715 void gen_cvt_itof(int t
)
718 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
719 /* signed long long to float/double/long double (unsigned case
720 is handled generically) */
721 o(0x50 + vtop
->r2
); /* push r2 */
722 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
723 o(0x242cdf); /* fildll (%esp) */
724 o(0x08c483); /* add $8, %esp */
725 } else if ((vtop
->t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
726 (VT_INT
| VT_UNSIGNED
)) {
727 /* unsigned int to float/double/long double */
728 o(0x6a); /* push $0 */
730 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
731 o(0x242cdf); /* fildll (%esp) */
732 o(0x08c483); /* add $8, %esp */
734 /* int to float/double/long double */
735 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
736 o(0x2404db); /* fildl (%esp) */
737 o(0x04c483); /* add $4, %esp */
742 /* convert fp to int 't' type */
743 /* XXX: handle long long case */
744 void gen_cvt_ftoi(int t
)
754 oad(0x2dd9, (int)&__tcc_int_fpu_control
); /* ldcw xxx */
755 oad(0xec81, size
); /* sub $xxx, %esp */
757 o(0x1cdb); /* fistpl */
759 o(0x3cdf); /* fistpll */
761 oad(0x2dd9, (int)&__tcc_fpu_control
); /* ldcw xxx */
763 o(0x58 + r
); /* pop r */
766 vtop
->r
= r
; /* mark reg as used */
767 r2
= get_reg(RC_INT
);
768 o(0x58 + r2
); /* pop r2 */
771 o(0x04c483); /* add $4, %esp */
777 /* convert from one floating point type to another */
778 void gen_cvt_ftof(int t
)
780 /* all we have to do on i386 is to put the float in a register */
784 /* bound check support functions */
786 /* generate first part of bounded pointer addition */
787 void gen_bounded_ptr_add1(void)
789 /* prepare fast i386 function call (args in eax and edx) */
791 /* save all temporary registers */
797 /* if deref is true, then also test dereferencing */
798 void gen_bounded_ptr_add2(int deref
)
804 size
= type_size(vtop
->t
, &align
);
806 case 1: func
= __bound_ptr_indir1
; break;
807 case 2: func
= __bound_ptr_indir2
; break;
808 case 4: func
= __bound_ptr_indir4
; break;
809 case 8: func
= __bound_ptr_indir8
; break;
810 case 12: func
= __bound_ptr_indir12
; break;
811 case 16: func
= __bound_ptr_indir16
; break;
813 error("unhandled size when derefencing bounded pointer");
818 func
= __bound_ptr_add
;
821 /* do a fast function call */
822 oad(0xe8, (int)func
- ind
- 5);
823 /* return pointer is there */
827 /* end of X86 code generator */
828 /*************************************************************/