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 int *func_sub_sp_ptr
;
83 static unsigned char *func_bound_ptr
;
84 static int func_ret_sub
;
107 void greloc_patch(unsigned char *ptr
,
108 unsigned long addr
, unsigned long val
, int type
)
115 *(int *)ptr
+= val
- addr
;
120 /* output a symbol and patch all calls to it */
121 void gsym_addr(int t
, int a
)
125 n
= *(int *)t
; /* next value */
126 *(int *)t
= a
- t
- 4;
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 int oad(int c
, int s
)
150 /* output constant with relocation if 'r & VT_SYM' is true */
151 void gen_addr32(int r
, int c
)
156 greloc(cur_text_section
,
157 (Sym
*)c
, ind
- (int)cur_text_section
->data
, R_386_32
);
162 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
164 void gen_modrm(int op_reg
, int r
, int c
)
166 op_reg
= op_reg
<< 3;
167 if ((r
& VT_VALMASK
) == VT_CONST
) {
168 /* constant memory reference */
171 } else if ((r
& VT_VALMASK
) == VT_LOCAL
) {
172 /* currently, we use only ebp as base */
174 /* short reference */
178 oad(0x85 | op_reg
, c
);
181 g(0x00 | op_reg
| (r
& VT_VALMASK
));
186 /* load 'r' from value 'sv' */
187 void load(int r
, SValue
*sv
)
189 int v
, t
, ft
, fc
, fr
;
198 if (v
== VT_LLOCAL
) {
200 v1
.r
= VT_LOCAL
| VT_LVAL
;
205 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
208 } else if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
211 } else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
214 } else if ((ft
& VT_TYPE
) == VT_BYTE
) {
215 o(0xbe0f); /* movsbl */
216 } else if ((ft
& VT_TYPE
) == (VT_BYTE
| VT_UNSIGNED
)) {
217 o(0xb60f); /* movzbl */
218 } else if ((ft
& VT_TYPE
) == VT_SHORT
) {
219 o(0xbf0f); /* movswl */
220 } else if ((ft
& VT_TYPE
) == (VT_SHORT
| VT_UNSIGNED
)) {
221 o(0xb70f); /* movzwl */
225 gen_modrm(r
, fr
, fc
);
228 o(0xb8 + r
); /* mov $xx, r */
230 } else if (v
== VT_LOCAL
) {
231 o(0x8d); /* lea xxx(%ebp), r */
232 gen_modrm(r
, VT_LOCAL
, fc
);
233 } else if (v
== VT_CMP
) {
234 oad(0xb8 + r
, 0); /* mov $0, r */
235 o(0x0f); /* setxx %br */
238 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
240 oad(0xb8 + r
, t
); /* mov $1, r */
241 oad(0xe9, 5); /* jmp after */
243 oad(0xb8 + r
, t
^ 1); /* mov $0, r */
246 o(0xc0 + r
+ v
* 8); /* mov v, r */
251 /* store register 'r' in lvalue 'v' */
252 void store(int r
, SValue
*v
)
258 fr
= v
->r
& VT_VALMASK
;
260 /* XXX: incorrect if float reg to reg */
261 if (bt
== VT_FLOAT
) {
264 } else if (bt
== VT_DOUBLE
) {
267 } else if (bt
== VT_LDOUBLE
) {
268 o(0xc0d9); /* fld %st(0) */
279 if (fr
== VT_CONST
||
282 gen_modrm(r
, v
->r
, fc
);
283 } else if (fr
!= r
) {
284 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
288 /* start function call and return function call context */
289 void gfunc_start(GFuncContext
*c
, int func_call
)
292 c
->func_call
= func_call
;
295 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
297 void gfunc_param(GFuncContext
*c
)
301 if ((vtop
->t
& VT_BTYPE
) == VT_STRUCT
) {
302 size
= type_size(vtop
->t
, &align
);
303 /* align to stack align size */
304 size
= (size
+ 3) & ~3;
305 /* allocate the necessary size on stack */
306 oad(0xec81, size
); /* sub $xxx, %esp */
307 /* generate structure store */
309 o(0x89); /* mov %esp, r */
311 vset(vtop
->t
, r
| VT_LVAL
, 0);
314 c
->args_size
+= size
;
315 } else if (is_float(vtop
->t
)) {
316 gv(RC_FLOAT
); /* only one float register */
317 if ((vtop
->t
& VT_BTYPE
) == VT_FLOAT
)
319 else if ((vtop
->t
& VT_BTYPE
) == VT_DOUBLE
)
323 oad(0xec81, size
); /* sub $xxx, %esp */
327 o(0x5cd9 + size
- 4); /* fstp[s|l] 0(%esp) */
330 c
->args_size
+= size
;
332 /* simple type (currently always same size) */
333 /* XXX: implicit cast ? */
335 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
337 o(0x50 + vtop
->r2
); /* push r */
341 o(0x50 + r
); /* push r */
342 c
->args_size
+= size
;
347 /* generate function call with address in (vtop->t, vtop->c) and free function
348 context. Stack entry is popped */
349 void gfunc_call(GFuncContext
*c
)
352 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
354 if (vtop
->r
& VT_SYM
) {
355 /* relocation case */
356 greloc(cur_text_section
, vtop
->c
.sym
,
357 ind
+ 1 - (int)cur_text_section
->data
, R_386_PC32
);
360 oad(0xe8, vtop
->c
.ul
- ind
- 5);
363 /* otherwise, indirect call */
365 o(0xff); /* call *r */
368 if (c
->args_size
&& c
->func_call
== FUNC_CDECL
)
369 oad(0xc481, c
->args_size
); /* add $xxx, %esp */
373 /* generate function prolog of type 't' */
374 void gfunc_prolog(int t
)
376 int addr
, align
, size
, u
, func_call
;
379 sym
= sym_find((unsigned)t
>> VT_STRUCT_SHIFT
);
382 /* if the function returns a structure, then add an
383 implicit pointer parameter */
385 if ((func_vt
& VT_BTYPE
) == VT_STRUCT
) {
389 /* define parameters */
390 while ((sym
= sym
->next
) != NULL
) {
392 sym_push(sym
->v
& ~SYM_FIELD
, u
,
393 VT_LOCAL
| VT_LVAL
, addr
);
394 size
= type_size(u
, &align
);
395 size
= (size
+ 3) & ~3;
396 #ifdef FUNC_STRUCT_PARAM_AS_PTR
397 /* structs are passed as pointer */
398 if ((u
& VT_BTYPE
) == VT_STRUCT
) {
405 /* pascal type call ? */
406 if (func_call
== FUNC_STDCALL
)
407 func_ret_sub
= addr
- 8;
408 o(0xe58955); /* push %ebp, mov %esp, %ebp */
409 func_sub_sp_ptr
= (int *)oad(0xec81, 0); /* sub $xxx, %esp */
410 /* leave some room for bound checking code */
411 if (do_bounds_check
) {
412 oad(0xb8, 0); /* lbound section pointer */
413 oad(0xb8, 0); /* call to function */
414 func_bound_ptr
= lbounds_section
->data_ptr
;
418 /* generate function epilog */
419 void gfunc_epilog(void)
421 #ifdef CONFIG_TCC_BCHECK
422 if (do_bounds_check
&& func_bound_ptr
!= lbounds_section
->data_ptr
) {
425 /* add end of table info */
426 bounds_ptr
= (int *)lbounds_section
->data_ptr
;
428 lbounds_section
->data_ptr
= (unsigned char *)bounds_ptr
;
429 /* generate bound local allocation */
431 ind
= (int)func_sub_sp_ptr
+ 4;
432 oad(0xb8, (int)func_bound_ptr
); /* mov %eax, xxx */
433 oad(0xe8, (int)__bound_local_new
- ind
- 5);
435 /* generate bound check local freeing */
436 o(0x5250); /* save returned value, if any */
437 oad(0xb8, (int)func_bound_ptr
); /* mov %eax, xxx */
438 oad(0xe8, (int)__bound_local_delete
- ind
- 5);
439 o(0x585a); /* restore returned value, if any */
443 if (func_ret_sub
== 0) {
448 g(func_ret_sub
>> 8);
450 /* align local size to word & save local variables */
451 *func_sub_sp_ptr
= (-loc
+ 3) & -4;
454 /* generate a jump to a label */
457 return psym(0xe9, t
);
460 /* generate a jump to a fixed address */
461 void gjmp_addr(int a
)
463 oad(0xe9, a
- ind
- 5);
466 /* generate a test. set 'inv' to invert test. Stack entry is popped */
467 int gtst(int inv
, int t
)
470 v
= vtop
->r
& VT_VALMASK
;
472 /* fast case : can jump directly since flags are set */
474 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
475 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
476 /* && or || optimization */
477 if ((v
& 1) == inv
) {
478 /* insert vtop->c jump list in t */
489 if (is_float(vtop
->t
)) {
493 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
494 /* constant jmp optimization */
495 if ((vtop
->c
.i
!= 0) != inv
)
502 t
= psym(0x85 ^ inv
, t
);
509 /* generate an integer binary operation */
516 case TOK_ADDC1
: /* add with carry generation */
519 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
526 /* XXX: generate inc and dec for smaller code ? */
528 o(0xc0 | (opc
<< 3) | r
);
532 oad(0xc0 | (opc
<< 3) | r
, c
);
538 o((opc
<< 3) | 0x01);
539 o(0xc0 + r
+ fr
* 8);
542 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
544 vset(VT_INT
, VT_CMP
, op
);
548 case TOK_SUBC1
: /* sub with carry generation */
551 case TOK_ADDC2
: /* add with carry use */
554 case TOK_SUBC2
: /* sub with carry use */
571 o(0xaf0f); /* imul fr, r */
572 o(0xc0 + fr
+ r
* 8);
583 opc
= 0xc0 | (opc
<< 3);
584 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
589 c
= vtop
->c
.i
& 0x1f;
590 o(0xc1); /* shl/shr/sar $xxx, r */
594 /* we generate the shift in ecx */
597 o(0xd3); /* shl/shr/sar %cl, r */
608 /* first operand must be in eax */
609 /* XXX: need better constraint for second operand */
615 if (op
== TOK_UMULL
) {
616 o(0xf7); /* mul fr */
621 if (op
== TOK_UDIV
|| op
== TOK_UMOD
) {
622 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
625 o(0xf799); /* cltd, idiv fr, %eax */
628 if (op
== '%' || op
== TOK_UMOD
)
641 /* generate a floating point operation 'v = t1 op t2' instruction. The
642 two operands are guaranted to have the same floating point type */
643 /* XXX: need to use ST1 too */
646 int a
, ft
, fc
, swapped
, r
;
648 /* convert constants to memory references */
649 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
654 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
657 /* must put at least one value in the floating point register */
658 if ((vtop
[-1].r
& VT_LVAL
) &&
659 (vtop
[0].r
& VT_LVAL
)) {
665 /* swap the stack if needed so that t1 is the register and t2 is
666 the memory reference */
667 if (vtop
[-1].r
& VT_LVAL
) {
671 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
672 /* load on stack second operand */
674 save_reg(REG_EAX
); /* eax is used by FP comparison code */
675 if (op
== TOK_GE
|| op
== TOK_GT
)
677 else if (op
== TOK_EQ
|| op
== TOK_NE
)
680 o(0xc9d9); /* fxch %st(1) */
681 o(0xe9da); /* fucompp */
682 o(0xe0df); /* fnstsw %ax */
684 o(0x45e480); /* and $0x45, %ah */
685 o(0x40fC80); /* cmp $0x40, %ah */
686 } else if (op
== TOK_NE
) {
687 o(0x45e480); /* and $0x45, %ah */
688 o(0x40f480); /* xor $0x40, %ah */
690 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
691 o(0x05c4f6); /* test $0x05, %ah */
694 o(0x45c4f6); /* test $0x45, %ah */
701 /* no memory reference possible for long double operations */
702 if ((vtop
->t
& VT_BTYPE
) == VT_LDOUBLE
) {
728 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
729 o(0xde); /* fxxxp %st, %st(1) */
732 /* if saved lvalue, then we must reload it */
734 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
738 v1
.r
= VT_LOCAL
| VT_LVAL
;
744 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
754 /* FPU control word for rounding to nearest mode */
755 /* XXX: should move that into tcc lib support code ! */
756 static unsigned short __tcc_fpu_control
= 0x137f;
757 /* FPU control word for round to zero mode for int convertion */
758 static unsigned short __tcc_int_fpu_control
= 0x137f | 0x0c00;
760 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
761 and 'long long' cases. */
762 void gen_cvt_itof(int t
)
766 if ((vtop
->t
& VT_BTYPE
) == VT_LLONG
) {
767 /* signed long long to float/double/long double (unsigned case
768 is handled generically) */
769 o(0x50 + vtop
->r2
); /* push r2 */
770 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
771 o(0x242cdf); /* fildll (%esp) */
772 o(0x08c483); /* add $8, %esp */
773 } else if ((vtop
->t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
774 (VT_INT
| VT_UNSIGNED
)) {
775 /* unsigned int to float/double/long double */
776 o(0x6a); /* push $0 */
778 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
779 o(0x242cdf); /* fildll (%esp) */
780 o(0x08c483); /* add $8, %esp */
782 /* int to float/double/long double */
783 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
784 o(0x2404db); /* fildl (%esp) */
785 o(0x04c483); /* add $4, %esp */
790 /* convert fp to int 't' type */
791 /* XXX: handle long long case */
792 void gen_cvt_ftoi(int t
)
802 oad(0x2dd9, (int)&__tcc_int_fpu_control
); /* ldcw xxx */
803 oad(0xec81, size
); /* sub $xxx, %esp */
805 o(0x1cdb); /* fistpl */
807 o(0x3cdf); /* fistpll */
809 oad(0x2dd9, (int)&__tcc_fpu_control
); /* ldcw xxx */
811 o(0x58 + r
); /* pop r */
814 vtop
->r
= r
; /* mark reg as used */
815 r2
= get_reg(RC_INT
);
816 o(0x58 + r2
); /* pop r2 */
819 o(0x04c483); /* add $4, %esp */
825 /* convert from one floating point type to another */
826 void gen_cvt_ftof(int t
)
828 /* all we have to do on i386 is to put the float in a register */
832 /* bound check support functions */
833 #ifdef CONFIG_TCC_BCHECK
835 /* generate a bounded pointer addition */
836 void gen_bounded_ptr_add(void)
839 /* prepare fast i386 function call (args in eax and edx) */
841 /* save all temporary registers */
844 /* do a fast function call */
846 oad(0xe8, (int)__bound_ptr_add
- ind
- 5);
847 /* returned pointer is in eax */
849 vtop
->r
= REG_EAX
| VT_BOUNDED
;
850 vtop
->c
.ul
= addr
; /* address of bounding function call point */
853 /* patch pointer addition in vtop so that pointer dereferencing is
855 void gen_bounded_ptr_deref(void)
858 int size
, align
, addr
;
861 /* XXX: put that code in generic part of tcc */
862 if (!is_float(vtop
->t
)) {
863 if (vtop
->r
& VT_LVAL_BYTE
)
865 else if (vtop
->r
& VT_LVAL_SHORT
)
869 size
= type_size(vtop
->t
, &align
);
871 case 1: func
= __bound_ptr_indir1
; break;
872 case 2: func
= __bound_ptr_indir2
; break;
873 case 4: func
= __bound_ptr_indir4
; break;
874 case 8: func
= __bound_ptr_indir8
; break;
875 case 12: func
= __bound_ptr_indir12
; break;
876 case 16: func
= __bound_ptr_indir16
; break;
878 error("unhandled size when derefencing bounded pointer");
884 *(int *)(addr
+ 1) = (int)func
- addr
- 5;
888 /* end of X86 code generator */
889 /*************************************************************/