2 * x86-64 code generator for TCC
4 * Copyright (c) 2008 Shinichiro Hamaji
6 * Based on i386-gen.c by Fabrice Bellard
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2 of the License, or (at your option) any later version.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 #ifdef TARGET_DEFS_ONLY
25 /* number of available registers */
29 /* a register can belong to several classes. The classes must be
30 sorted from more general to more precise (see gv2() code which does
31 assumptions on it). */
32 #define RC_INT 0x0001 /* generic integer register */
33 #define RC_FLOAT 0x0002 /* generic float register */
37 #define RC_ST0 0x0080 /* only for long double */
42 #define RC_XMM0 0x1000
43 #define RC_XMM1 0x2000
44 #define RC_XMM2 0x4000
45 #define RC_XMM3 0x8000
46 #define RC_XMM4 0x10000
47 #define RC_XMM5 0x20000
48 #define RC_XMM6 0x40000
49 #define RC_XMM7 0x80000
50 #define RC_IRET RC_RAX /* function return: integer register */
51 #define RC_LRET RC_RDX /* function return: second integer register */
52 #define RC_FRET RC_XMM0 /* function return: float register */
53 #define RC_QRET RC_XMM1 /* function return: second float register */
55 /* pretty names for the registers */
83 #define REX_BASE(reg) (((reg) >> 3) & 1)
84 #define REG_VALUE(reg) ((reg) & 7)
86 /* return registers for function */
87 #define REG_IRET TREG_RAX /* single word int return register */
88 #define REG_LRET TREG_RDX /* second word return register (for long long) */
89 #define REG_FRET TREG_XMM0 /* float return register */
90 #define REG_QRET TREG_XMM1 /* second float return register */
92 /* defined if function parameters must be evaluated in reverse order */
93 #define INVERT_FUNC_PARAMS
95 /* pointer size, in bytes */
98 /* long double size and alignment, in bytes */
99 #define LDOUBLE_SIZE 16
100 #define LDOUBLE_ALIGN 16
101 /* maximum alignment (for aligned attribute support) */
104 /******************************************************/
107 #define EM_TCC_TARGET EM_X86_64
109 /* relocation type for 32 bit data relocation */
110 #define R_DATA_32 R_X86_64_32
111 #define R_DATA_PTR R_X86_64_64
112 #define R_JMP_SLOT R_X86_64_JUMP_SLOT
113 #define R_COPY R_X86_64_COPY
115 #define ELF_START_ADDR 0x08048000
116 #define ELF_PAGE_SIZE 0x1000
118 /******************************************************/
119 #else /* ! TARGET_DEFS_ONLY */
120 /******************************************************/
124 ST_DATA
const int reg_classes
[NB_REGS
] = {
125 /* eax */ RC_INT
| RC_RAX
,
126 /* ecx */ RC_INT
| RC_RCX
,
127 /* edx */ RC_INT
| RC_RDX
,
141 /* xmm0 */ RC_FLOAT
| RC_XMM0
,
142 /* xmm1 */ RC_FLOAT
| RC_XMM1
,
143 /* xmm2 */ RC_FLOAT
| RC_XMM2
,
144 /* xmm3 */ RC_FLOAT
| RC_XMM3
,
145 /* xmm4 */ RC_FLOAT
| RC_XMM4
,
146 /* xmm5 */ RC_FLOAT
| RC_XMM5
,
147 /* xmm6 an xmm7 are included so gv() can be used on them,
148 but they are not tagged with RC_FLOAT because they are
149 callee saved on Windows */
155 static unsigned long func_sub_sp_offset
;
156 static int func_ret_sub
;
158 /* XXX: make it faster ? */
163 if (ind1
> cur_text_section
->data_allocated
)
164 section_realloc(cur_text_section
, ind1
);
165 cur_text_section
->data
[ind
] = c
;
169 void o(unsigned int c
)
191 void gen_le64(int64_t c
)
203 void orex(int ll
, int r
, int r2
, int b
)
205 if ((r
& VT_VALMASK
) >= VT_CONST
)
207 if ((r2
& VT_VALMASK
) >= VT_CONST
)
209 if (ll
|| REX_BASE(r
) || REX_BASE(r2
))
210 o(0x40 | REX_BASE(r
) | (REX_BASE(r2
) << 2) | (ll
<< 3));
214 /* output a symbol and patch all calls to it */
215 void gsym_addr(int t
, int a
)
219 ptr
= (int *)(cur_text_section
->data
+ t
);
220 n
= *ptr
; /* next value */
231 /* psym is used to put an instruction with a data field which is a
232 reference to a symbol. It is in fact the same as oad ! */
235 static int is64_type(int t
)
237 return ((t
& VT_BTYPE
) == VT_PTR
||
238 (t
& VT_BTYPE
) == VT_FUNC
||
239 (t
& VT_BTYPE
) == VT_LLONG
);
242 /* instruction + 4 bytes data. Return the address of the data */
243 ST_FUNC
int oad(int c
, int s
)
249 if (ind1
> cur_text_section
->data_allocated
)
250 section_realloc(cur_text_section
, ind1
);
251 *(int *)(cur_text_section
->data
+ ind
) = s
;
257 ST_FUNC
void gen_addr32(int r
, Sym
*sym
, int c
)
260 greloc(cur_text_section
, sym
, ind
, R_X86_64_32
);
264 /* output constant with relocation if 'r & VT_SYM' is true */
265 ST_FUNC
void gen_addr64(int r
, Sym
*sym
, int64_t c
)
268 greloc(cur_text_section
, sym
, ind
, R_X86_64_64
);
272 /* output constant with relocation if 'r & VT_SYM' is true */
273 ST_FUNC
void gen_addrpc32(int r
, Sym
*sym
, int c
)
276 greloc(cur_text_section
, sym
, ind
, R_X86_64_PC32
);
280 /* output got address with relocation */
281 static void gen_gotpcrel(int r
, Sym
*sym
, int c
)
283 #ifndef TCC_TARGET_PE
286 greloc(cur_text_section
, sym
, ind
, R_X86_64_GOTPCREL
);
287 sr
= cur_text_section
->reloc
;
288 rel
= (ElfW(Rela
) *)(sr
->data
+ sr
->data_offset
- sizeof(ElfW(Rela
)));
291 printf("picpic: %s %x %x | %02x %02x %02x\n", get_tok_str(sym
->v
, NULL
), c
, r
,
292 cur_text_section
->data
[ind
-3],
293 cur_text_section
->data
[ind
-2],
294 cur_text_section
->data
[ind
-1]
296 greloc(cur_text_section
, sym
, ind
, R_X86_64_PC32
);
300 /* we use add c, %xxx for displacement */
302 o(0xc0 + REG_VALUE(r
));
307 static void gen_modrm_impl(int op_reg
, int r
, Sym
*sym
, int c
, int is_got
)
309 op_reg
= REG_VALUE(op_reg
) << 3;
310 if ((r
& VT_VALMASK
) == VT_CONST
) {
311 /* constant memory reference */
314 gen_gotpcrel(r
, sym
, c
);
316 gen_addrpc32(r
, sym
, c
);
318 } else if ((r
& VT_VALMASK
) == VT_LOCAL
) {
319 /* currently, we use only ebp as base */
321 /* short reference */
325 oad(0x85 | op_reg
, c
);
327 } else if ((r
& VT_VALMASK
) >= TREG_MEM
) {
329 g(0x80 | op_reg
| REG_VALUE(r
));
332 g(0x00 | op_reg
| REG_VALUE(r
));
335 g(0x00 | op_reg
| REG_VALUE(r
));
339 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
341 static void gen_modrm(int op_reg
, int r
, Sym
*sym
, int c
)
343 gen_modrm_impl(op_reg
, r
, sym
, c
, 0);
346 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
348 static void gen_modrm64(int opcode
, int op_reg
, int r
, Sym
*sym
, int c
)
351 is_got
= (op_reg
& TREG_MEM
) && !(sym
->type
.t
& VT_STATIC
);
352 orex(1, r
, op_reg
, opcode
);
353 gen_modrm_impl(op_reg
, r
, sym
, c
, is_got
);
357 /* load 'r' from value 'sv' */
358 void load(int r
, SValue
*sv
)
360 int v
, t
, ft
, fc
, fr
;
365 sv
= pe_getimport(sv
, &v2
);
372 #ifndef TCC_TARGET_PE
373 /* we use indirect access via got */
374 if ((fr
& VT_VALMASK
) == VT_CONST
&& (fr
& VT_SYM
) &&
375 (fr
& VT_LVAL
) && !(sv
->sym
->type
.t
& VT_STATIC
)) {
376 /* use the result register as a temporal register */
377 int tr
= r
| TREG_MEM
;
379 /* we cannot use float registers as a temporal register */
380 tr
= get_reg(RC_INT
) | TREG_MEM
;
382 gen_modrm64(0x8b, tr
, fr
, sv
->sym
, 0);
384 /* load from the temporal register */
392 if (v
== VT_LLOCAL
) {
394 v1
.r
= VT_LOCAL
| VT_LVAL
;
397 if (!(reg_classes
[fr
] & RC_INT
))
398 fr
= get_reg(RC_INT
);
402 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
404 r
= REG_VALUE(r
); /* movd */
405 } else if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
406 b
= 0x7e0ff3; /* movq */
408 } else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
) {
409 b
= 0xdb, r
= 5; /* fldt */
410 } else if ((ft
& VT_TYPE
) == VT_BYTE
|| (ft
& VT_TYPE
) == VT_BOOL
) {
411 b
= 0xbe0f; /* movsbl */
412 } else if ((ft
& VT_TYPE
) == (VT_BYTE
| VT_UNSIGNED
)) {
413 b
= 0xb60f; /* movzbl */
414 } else if ((ft
& VT_TYPE
) == VT_SHORT
) {
415 b
= 0xbf0f; /* movswl */
416 } else if ((ft
& VT_TYPE
) == (VT_SHORT
| VT_UNSIGNED
)) {
417 b
= 0xb70f; /* movzwl */
419 assert(((ft
& VT_BTYPE
) == VT_INT
) || ((ft
& VT_BTYPE
) == VT_LLONG
)
420 || ((ft
& VT_BTYPE
) == VT_PTR
) || ((ft
& VT_BTYPE
) == VT_ENUM
)
421 || ((ft
& VT_BTYPE
) == VT_FUNC
));
426 gen_modrm64(b
, r
, fr
, sv
->sym
, fc
);
429 gen_modrm(r
, fr
, sv
->sym
, fc
);
436 o(0x05 + REG_VALUE(r
) * 8); /* lea xx(%rip), r */
437 gen_addrpc32(fr
, sv
->sym
, fc
);
439 if (sv
->sym
->type
.t
& VT_STATIC
) {
441 o(0x05 + REG_VALUE(r
) * 8); /* lea xx(%rip), r */
442 gen_addrpc32(fr
, sv
->sym
, fc
);
445 o(0x05 + REG_VALUE(r
) * 8); /* mov xx(%rip), r */
446 gen_gotpcrel(r
, sv
->sym
, fc
);
449 } else if (is64_type(ft
)) {
450 orex(1,r
,0, 0xb8 + REG_VALUE(r
)); /* mov $xx, r */
453 orex(0,r
,0, 0xb8 + REG_VALUE(r
)); /* mov $xx, r */
456 } else if (v
== VT_LOCAL
) {
457 orex(1,0,r
,0x8d); /* lea xxx(%ebp), r */
458 gen_modrm(r
, VT_LOCAL
, sv
->sym
, fc
);
459 } else if (v
== VT_CMP
) {
461 if ((fc
& ~0x100) != TOK_NE
)
462 oad(0xb8 + REG_VALUE(r
), 0); /* mov $0, r */
464 oad(0xb8 + REG_VALUE(r
), 1); /* mov $1, r */
467 /* This was a float compare. If the parity bit is
468 set the result was unordered, meaning false for everything
469 except TOK_NE, and true for TOK_NE. */
471 o(0x037a + (REX_BASE(r
) << 8));
473 orex(0,r
,0, 0x0f); /* setxx %br */
475 o(0xc0 + REG_VALUE(r
));
476 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
479 oad(0xb8 + REG_VALUE(r
), t
); /* mov $1, r */
480 o(0x05eb + (REX_BASE(r
) << 8)); /* jmp after */
483 oad(0xb8 + REG_VALUE(r
), t
^ 1); /* mov $0, r */
485 if ((r
>= TREG_XMM0
) && (r
<= TREG_XMM7
)) {
487 /* gen_cvt_ftof(VT_DOUBLE); */
488 o(0xf0245cdd); /* fstpl -0x10(%rsp) */
489 /* movsd -0x10(%rsp),%xmmN */
491 o(0x44 + REG_VALUE(r
)*8); /* %xmmN */
494 assert((v
>= TREG_XMM0
) && (v
<= TREG_XMM7
));
495 if ((ft
& VT_BTYPE
) == VT_FLOAT
) {
498 assert((ft
& VT_BTYPE
) == VT_DOUBLE
);
501 o(0xc0 + REG_VALUE(v
) + REG_VALUE(r
)*8);
503 } else if (r
== TREG_ST0
) {
504 assert((v
>= TREG_XMM0
) || (v
<= TREG_XMM7
));
505 /* gen_cvt_ftof(VT_LDOUBLE); */
506 /* movsd %xmmN,-0x10(%rsp) */
508 o(0x44 + REG_VALUE(r
)*8); /* %xmmN */
510 o(0xf02444dd); /* fldl -0x10(%rsp) */
513 o(0xc0 + REG_VALUE(r
) + REG_VALUE(v
) * 8); /* mov v, r */
519 /* store register 'r' in lvalue 'v' */
520 void store(int r
, SValue
*v
)
524 /* store the REX prefix in this variable when PIC is enabled */
529 v
= pe_getimport(v
, &v2
);
534 fr
= v
->r
& VT_VALMASK
;
537 #ifndef TCC_TARGET_PE
538 /* we need to access the variable via got */
539 if (fr
== VT_CONST
&& (v
->r
& VT_SYM
)) {
540 /* mov xx(%rip), %r11 */
542 gen_gotpcrel(TREG_R11
, v
->sym
, v
->c
.ul
);
543 pic
= is64_type(bt
) ? 0x49 : 0x41;
547 /* XXX: incorrect if float reg to reg */
548 if (bt
== VT_FLOAT
) {
551 o(0x7e0f); /* movd */
553 } else if (bt
== VT_DOUBLE
) {
556 o(0xd60f); /* movq */
558 } else if (bt
== VT_LDOUBLE
) {
559 o(0xc0d9); /* fld %st(0) */
567 if (bt
== VT_BYTE
|| bt
== VT_BOOL
)
569 else if (is64_type(bt
))
575 /* xxx r, (%r11) where xxx is mov, movq, fld, or etc */
580 if (fr
== VT_CONST
|| fr
== VT_LOCAL
|| (v
->r
& VT_LVAL
)) {
581 gen_modrm64(op64
, r
, v
->r
, v
->sym
, fc
);
582 } else if (fr
!= r
) {
583 /* XXX: don't we really come here? */
585 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
588 if (fr
== VT_CONST
|| fr
== VT_LOCAL
|| (v
->r
& VT_LVAL
)) {
589 gen_modrm(r
, v
->r
, v
->sym
, fc
);
590 } else if (fr
!= r
) {
591 /* XXX: don't we really come here? */
593 o(0xc0 + fr
+ r
* 8); /* mov r, fr */
598 /* 'is_jmp' is '1' if it is a jump */
599 static void gcall_or_jmp(int is_jmp
)
602 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
604 if (vtop
->r
& VT_SYM
) {
605 /* relocation case */
606 greloc(cur_text_section
, vtop
->sym
,
607 ind
+ 1, R_X86_64_PC32
);
609 /* put an empty PC32 relocation */
610 put_elf_reloc(symtab_section
, cur_text_section
,
611 ind
+ 1, R_X86_64_PC32
, 0);
613 oad(0xe8 + is_jmp
, vtop
->c
.ul
- 4); /* call/jmp im */
615 /* otherwise, indirect call */
619 o(0xff); /* call/jmp *r */
620 o(0xd0 + REG_VALUE(r
) + (is_jmp
<< 4));
627 static const uint8_t arg_regs
[REGN
] = {
628 TREG_RCX
, TREG_RDX
, TREG_R8
, TREG_R9
631 /* Prepare arguments in R10 and R11 rather than RCX and RDX
632 because gv() will not ever use these */
633 static int arg_prepare_reg(int idx
) {
634 if (idx
== 0 || idx
== 1)
635 /* idx=0: r10, idx=1: r11 */
638 return arg_regs
[idx
];
641 static int func_scratch
;
643 /* Generate function call. The function address is pushed first, then
644 all the parameters in call order. This functions pops all the
645 parameters and the function address. */
647 void gen_offs_sp(int b
, int r
, int d
)
649 orex(1,0,r
& 0x100 ? 0 : r
, b
);
651 o(0x2444 | (REG_VALUE(r
) << 3));
654 o(0x2484 | (REG_VALUE(r
) << 3));
659 /* Return the number of registers needed to return the struct, or 0 if
660 returning via struct pointer. */
661 ST_FUNC
int gfunc_sret(CType
*vt
, CType
*ret
, int *ret_align
)
664 *ret_align
= 1; // Never have to re-align return values for x86-64
665 size
= type_size(vt
, &align
);
669 } else if (size
> 4) {
672 } else if (size
> 2) {
675 } else if (size
> 1) {
684 static int is_sse_float(int t
) {
687 return bt
== VT_DOUBLE
|| bt
== VT_FLOAT
;
690 int gfunc_arg_size(CType
*type
) {
692 if (type
->t
& (VT_ARRAY
|VT_BITFIELD
))
694 return type_size(type
, &align
);
697 void gfunc_call(int nb_args
)
699 int size
, r
, args_size
, i
, d
, bt
, struct_size
;
702 args_size
= (nb_args
< REGN
? REGN
: nb_args
) * PTR_SIZE
;
705 /* for struct arguments, we need to call memcpy and the function
706 call breaks register passing arguments we are preparing.
707 So, we process arguments which will be passed by stack first. */
708 struct_size
= args_size
;
709 for(i
= 0; i
< nb_args
; i
++) {
714 bt
= (sv
->type
.t
& VT_BTYPE
);
715 size
= gfunc_arg_size(&sv
->type
);
718 continue; /* arguments smaller than 8 bytes passed in registers or on stack */
720 if (bt
== VT_STRUCT
) {
721 /* align to stack align size */
722 size
= (size
+ 15) & ~15;
723 /* generate structure store */
725 gen_offs_sp(0x8d, r
, struct_size
);
728 /* generate memcpy call */
729 vset(&sv
->type
, r
| VT_LVAL
, 0);
733 } else if (bt
== VT_LDOUBLE
) {
735 gen_offs_sp(0xdb, 0x107, struct_size
);
740 if (func_scratch
< struct_size
)
741 func_scratch
= struct_size
;
744 struct_size
= args_size
;
746 for(i
= 0; i
< nb_args
; i
++) {
748 bt
= (vtop
->type
.t
& VT_BTYPE
);
750 size
= gfunc_arg_size(&vtop
->type
);
752 /* align to stack align size */
753 size
= (size
+ 15) & ~15;
756 gen_offs_sp(0x8d, d
, struct_size
);
757 gen_offs_sp(0x89, d
, arg
*8);
759 d
= arg_prepare_reg(arg
);
760 gen_offs_sp(0x8d, d
, struct_size
);
764 if (is_sse_float(vtop
->type
.t
)) {
765 gv(RC_XMM0
); /* only use one float register */
767 /* movq %xmm0, j*8(%rsp) */
768 gen_offs_sp(0xd60f66, 0x100, arg
*8);
770 /* movaps %xmm0, %xmmN */
772 o(0xc0 + (arg
<< 3));
773 d
= arg_prepare_reg(arg
);
774 /* mov %xmm0, %rxx */
777 o(0xc0 + REG_VALUE(d
));
780 if (bt
== VT_STRUCT
) {
781 vtop
->type
.ref
= NULL
;
782 vtop
->type
.t
= size
> 4 ? VT_LLONG
: size
> 2 ? VT_INT
783 : size
> 1 ? VT_SHORT
: VT_BYTE
;
788 gen_offs_sp(0x89, r
, arg
*8);
790 d
= arg_prepare_reg(arg
);
791 orex(1,d
,r
,0x89); /* mov */
792 o(0xc0 + REG_VALUE(r
) * 8 + REG_VALUE(d
));
800 /* Copy R10 and R11 into RCX and RDX, respectively */
802 o(0xd1894c); /* mov %r10, %rcx */
804 o(0xda894c); /* mov %r11, %rdx */
813 #define FUNC_PROLOG_SIZE 11
815 /* generate function prolog of type 't' */
816 void gfunc_prolog(CType
*func_type
)
818 int addr
, reg_param_index
, bt
, size
;
827 ind
+= FUNC_PROLOG_SIZE
;
828 func_sub_sp_offset
= ind
;
831 sym
= func_type
->ref
;
833 /* if the function returns a structure, then add an
834 implicit pointer parameter */
836 size
= gfunc_arg_size(&func_vt
);
838 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
844 /* define parameters */
845 while ((sym
= sym
->next
) != NULL
) {
847 bt
= type
->t
& VT_BTYPE
;
848 size
= gfunc_arg_size(type
);
850 if (reg_param_index
< REGN
) {
851 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
853 sym_push(sym
->v
& ~SYM_FIELD
, type
, VT_LOCAL
| VT_LVAL
| VT_REF
, addr
);
855 if (reg_param_index
< REGN
) {
856 /* save arguments passed by register */
857 if ((bt
== VT_FLOAT
) || (bt
== VT_DOUBLE
)) {
858 o(0xd60f66); /* movq */
859 gen_modrm(reg_param_index
, VT_LOCAL
, NULL
, addr
);
861 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
864 sym_push(sym
->v
& ~SYM_FIELD
, type
, VT_LOCAL
| VT_LVAL
, addr
);
870 while (reg_param_index
< REGN
) {
871 if (func_type
->ref
->c
== FUNC_ELLIPSIS
) {
872 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
879 /* generate function epilog */
880 void gfunc_epilog(void)
885 if (func_ret_sub
== 0) {
890 g(func_ret_sub
>> 8);
894 ind
= func_sub_sp_offset
- FUNC_PROLOG_SIZE
;
895 /* align local size to word & save local variables */
896 v
= (func_scratch
+ -loc
+ 15) & -16;
899 Sym
*sym
= external_global_sym(TOK___chkstk
, &func_old_type
, 0);
900 oad(0xb8, v
); /* mov stacksize, %eax */
901 oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
902 greloc(cur_text_section
, sym
, ind
-4, R_X86_64_PC32
);
903 o(0x90); /* fill for FUNC_PROLOG_SIZE = 11 bytes */
905 o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
906 o(0xec8148); /* sub rsp, stacksize */
910 cur_text_section
->data_offset
= saved_ind
;
911 pe_add_unwind_data(ind
, saved_ind
, v
);
912 ind
= cur_text_section
->data_offset
;
917 static void gadd_sp(int val
)
919 if (val
== (char)val
) {
923 oad(0xc48148, val
); /* add $xxx, %rsp */
927 typedef enum X86_64_Mode
{
935 static X86_64_Mode
classify_x86_64_merge(X86_64_Mode a
, X86_64_Mode b
) {
938 else if (a
== x86_64_mode_none
)
940 else if (b
== x86_64_mode_none
)
942 else if ((a
== x86_64_mode_memory
) || (b
== x86_64_mode_memory
))
943 return x86_64_mode_memory
;
944 else if ((a
== x86_64_mode_integer
) || (b
== x86_64_mode_integer
))
945 return x86_64_mode_integer
;
946 else if ((a
== x86_64_mode_x87
) || (b
== x86_64_mode_x87
))
947 return x86_64_mode_memory
;
949 return x86_64_mode_sse
;
952 static X86_64_Mode
classify_x86_64_inner(CType
*ty
) {
956 switch (ty
->t
& VT_BTYPE
) {
957 case VT_VOID
: return x86_64_mode_none
;
966 case VT_ENUM
: return x86_64_mode_integer
;
969 case VT_DOUBLE
: return x86_64_mode_sse
;
971 case VT_LDOUBLE
: return x86_64_mode_x87
;
977 if (f
->next
&& (f
->c
== f
->next
->c
))
978 return x86_64_mode_memory
;
980 mode
= x86_64_mode_none
;
981 for (; f
; f
= f
->next
)
982 mode
= classify_x86_64_merge(mode
, classify_x86_64_inner(&f
->type
));
990 static X86_64_Mode
classify_x86_64_arg(CType
*ty
, CType
*ret
, int *psize
, int *palign
, int *reg_count
) {
992 int size
, align
, ret_t
= 0;
994 if (ty
->t
& (VT_BITFIELD
|VT_ARRAY
)) {
998 mode
= x86_64_mode_integer
;
1000 size
= type_size(ty
, &align
);
1001 *psize
= (size
+ 7) & ~7;
1002 *palign
= (align
+ 7) & ~7;
1005 mode
= x86_64_mode_memory
;
1007 mode
= classify_x86_64_inner(ty
);
1009 case x86_64_mode_integer
:
1015 ret_t
= (size
> 4) ? VT_LLONG
: VT_INT
;
1019 case x86_64_mode_x87
:
1024 case x86_64_mode_sse
:
1030 ret_t
= (size
> 4) ? VT_DOUBLE
: VT_FLOAT
;
1033 default: break; /* nothing to be done for x86_64_mode_memory and x86_64_mode_none*/
1046 ST_FUNC
int classify_x86_64_va_arg(CType
*ty
) {
1047 /* This definition must be synced with stdarg.h */
1048 enum __va_arg_type
{
1049 __va_gen_reg
, __va_float_reg
, __va_stack
1051 int size
, align
, reg_count
;
1052 X86_64_Mode mode
= classify_x86_64_arg(ty
, NULL
, &size
, &align
, ®_count
);
1054 default: return __va_stack
;
1055 case x86_64_mode_integer
: return __va_gen_reg
;
1056 case x86_64_mode_sse
: return __va_float_reg
;
1060 /* Return the number of registers needed to return the struct, or 0 if
1061 returning via struct pointer. */
1062 int gfunc_sret(CType
*vt
, CType
*ret
, int *ret_align
) {
1063 int size
, align
, reg_count
;
1064 *ret_align
= 1; // Never have to re-align return values for x86-64
1065 return (classify_x86_64_arg(vt
, ret
, &size
, &align
, ®_count
) != x86_64_mode_memory
);
1069 static const uint8_t arg_regs
[REGN
] = {
1070 TREG_RDI
, TREG_RSI
, TREG_RDX
, TREG_RCX
, TREG_R8
, TREG_R9
1073 static int arg_prepare_reg(int idx
) {
1074 if (idx
== 2 || idx
== 3)
1075 /* idx=2: r10, idx=3: r11 */
1078 return arg_regs
[idx
];
1081 /* Generate function call. The function address is pushed first, then
1082 all the parameters in call order. This functions pops all the
1083 parameters and the function address. */
1084 void gfunc_call(int nb_args
)
1088 int size
, align
, r
, args_size
, stack_adjust
, run_start
, run_end
, i
, reg_count
;
1089 int nb_reg_args
= 0;
1090 int nb_sse_args
= 0;
1091 int sse_reg
, gen_reg
;
1093 /* calculate the number of integer/float register arguments */
1094 for(i
= 0; i
< nb_args
; i
++) {
1095 mode
= classify_x86_64_arg(&vtop
[-i
].type
, NULL
, &size
, &align
, ®_count
);
1096 if (mode
== x86_64_mode_sse
)
1097 nb_sse_args
+= reg_count
;
1098 else if (mode
== x86_64_mode_integer
)
1099 nb_reg_args
+= reg_count
;
1102 /* arguments are collected in runs. Each run is a collection of 8-byte aligned arguments
1103 and ended by a 16-byte aligned argument. This is because, from the point of view of
1104 the callee, argument alignment is computed from the bottom up. */
1105 /* for struct arguments, we need to call memcpy and the function
1106 call breaks register passing arguments we are preparing.
1107 So, we process arguments which will be passed by stack first. */
1108 gen_reg
= nb_reg_args
;
1109 sse_reg
= nb_sse_args
;
1112 while (run_start
!= nb_args
) {
1113 int run_gen_reg
= gen_reg
, run_sse_reg
= sse_reg
;
1117 for(i
= run_start
; (i
< nb_args
) && (run_end
== nb_args
); i
++) {
1118 mode
= classify_x86_64_arg(&vtop
[-i
].type
, NULL
, &size
, &align
, ®_count
);
1120 case x86_64_mode_memory
:
1121 case x86_64_mode_x87
:
1126 stack_adjust
+= size
;
1129 case x86_64_mode_sse
:
1130 sse_reg
-= reg_count
;
1131 if (sse_reg
+ reg_count
> 8) goto stack_arg
;
1134 case x86_64_mode_integer
:
1135 gen_reg
-= reg_count
;
1136 if (gen_reg
+ reg_count
> REGN
) goto stack_arg
;
1138 default: break; /* nothing to be done for x86_64_mode_none */
1142 gen_reg
= run_gen_reg
;
1143 sse_reg
= run_sse_reg
;
1145 /* adjust stack to align SSE boundary */
1146 if (stack_adjust
&= 15) {
1147 /* fetch cpu flag before the following sub will change the value */
1148 if (vtop
>= vstack
&& (vtop
->r
& VT_VALMASK
) == VT_CMP
)
1151 stack_adjust
= 16 - stack_adjust
;
1153 oad(0xec81, stack_adjust
); /* sub $xxx, %rsp */
1154 args_size
+= stack_adjust
;
1157 for(i
= run_start
; i
< run_end
;) {
1158 /* Swap argument to top, it will possibly be changed here,
1159 and might use more temps. At the end of the loop we keep
1160 in on the stack and swap it back to its original position
1161 if it is a register. */
1162 SValue tmp
= vtop
[0];
1166 mode
= classify_x86_64_arg(&vtop
->type
, NULL
, &size
, &align
, ®_count
);
1169 switch (vtop
->type
.t
& VT_BTYPE
) {
1171 if (mode
== x86_64_mode_sse
) {
1173 sse_reg
-= reg_count
;
1176 } else if (mode
== x86_64_mode_integer
) {
1178 gen_reg
-= reg_count
;
1184 /* allocate the necessary size on stack */
1186 oad(0xec81, size
); /* sub $xxx, %rsp */
1187 /* generate structure store */
1188 r
= get_reg(RC_INT
);
1189 orex(1, r
, 0, 0x89); /* mov %rsp, r */
1190 o(0xe0 + REG_VALUE(r
));
1191 vset(&vtop
->type
, r
| VT_LVAL
, 0);
1204 assert(mode
== x86_64_mode_sse
);
1208 o(0x50); /* push $rax */
1209 /* movq %xmmN, (%rsp) */
1211 o(0x04 + REG_VALUE(r
)*8);
1220 assert(mode
== x86_64_mode_integer
);
1222 /* XXX: implicit cast ? */
1223 if (gen_reg
> REGN
) {
1226 orex(0,r
,0,0x50 + REG_VALUE(r
)); /* push r */
1234 /* And swap the argument back to it's original position. */
1241 assert((vtop
->type
.t
== tmp
.type
.t
) && (vtop
->r
== tmp
.r
));
1250 /* handle 16 byte aligned arguments at end of run */
1251 run_start
= i
= run_end
;
1252 while (i
< nb_args
) {
1253 /* Rotate argument to top since it will always be popped */
1254 mode
= classify_x86_64_arg(&vtop
[-i
].type
, NULL
, &size
, &align
, ®_count
);
1260 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
) {
1262 oad(0xec8148, size
); /* sub $xxx, %rsp */
1263 o(0x7cdb); /* fstpt 0(%rsp) */
1268 assert(mode
== x86_64_mode_memory
);
1270 /* allocate the necessary size on stack */
1272 oad(0xec81, size
); /* sub $xxx, %rsp */
1273 /* generate structure store */
1274 r
= get_reg(RC_INT
);
1275 orex(1, r
, 0, 0x89); /* mov %rsp, r */
1276 o(0xe0 + REG_VALUE(r
));
1277 vset(&vtop
->type
, r
| VT_LVAL
, 0);
1288 /* XXX This should be superfluous. */
1289 save_regs(0); /* save used temporary registers */
1291 /* then, we prepare register passing arguments.
1292 Note that we cannot set RDX and RCX in this loop because gv()
1293 may break these temporary registers. Let's use R10 and R11
1295 assert(gen_reg
<= REGN
);
1296 assert(sse_reg
<= 8);
1297 for(i
= 0; i
< nb_args
; i
++) {
1298 mode
= classify_x86_64_arg(&vtop
->type
, &type
, &size
, &align
, ®_count
);
1299 /* Alter stack entry type so that gv() knows how to treat it */
1301 if (mode
== x86_64_mode_sse
) {
1302 if (reg_count
== 2) {
1304 gv(RC_FRET
); /* Use pair load into xmm0 & xmm1 */
1305 if (sse_reg
) { /* avoid redundant movaps %xmm0, %xmm0 */
1306 /* movaps %xmm0, %xmmN */
1308 o(0xc0 + (sse_reg
<< 3));
1309 /* movaps %xmm1, %xmmN */
1311 o(0xc1 + ((sse_reg
+1) << 3));
1314 assert(reg_count
== 1);
1316 /* Load directly to register */
1317 gv(RC_XMM0
<< sse_reg
);
1319 } else if (mode
== x86_64_mode_integer
) {
1321 /* XXX: implicit cast ? */
1322 gen_reg
-= reg_count
;
1324 int d
= arg_prepare_reg(gen_reg
);
1325 orex(1,d
,r
,0x89); /* mov */
1326 o(0xc0 + REG_VALUE(r
) * 8 + REG_VALUE(d
));
1327 if (reg_count
== 2) {
1328 d
= arg_prepare_reg(gen_reg
+1);
1329 orex(1,d
,vtop
->r2
,0x89); /* mov */
1330 o(0xc0 + REG_VALUE(vtop
->r2
) * 8 + REG_VALUE(d
));
1335 assert(gen_reg
== 0);
1336 assert(sse_reg
== 0);
1338 /* We shouldn't have many operands on the stack anymore, but the
1339 call address itself is still there, and it might be in %eax
1340 (or edx/ecx) currently, which the below writes would clobber.
1341 So evict all remaining operands here. */
1344 /* Copy R10 and R11 into RDX and RCX, respectively */
1345 if (nb_reg_args
> 2) {
1346 o(0xd2894c); /* mov %r10, %rdx */
1347 if (nb_reg_args
> 3) {
1348 o(0xd9894c); /* mov %r11, %rcx */
1352 oad(0xb8, nb_sse_args
< 8 ? nb_sse_args
: 8); /* mov nb_sse_args, %eax */
1360 #define FUNC_PROLOG_SIZE 11
1362 static void push_arg_reg(int i
) {
1364 gen_modrm64(0x89, arg_regs
[i
], VT_LOCAL
, NULL
, loc
);
1367 /* generate function prolog of type 't' */
1368 void gfunc_prolog(CType
*func_type
)
1371 int i
, addr
, align
, size
, reg_count
;
1372 int param_addr
= 0, reg_param_index
, sse_param_index
;
1376 sym
= func_type
->ref
;
1377 addr
= PTR_SIZE
* 2;
1379 ind
+= FUNC_PROLOG_SIZE
;
1380 func_sub_sp_offset
= ind
;
1383 if (func_type
->ref
->c
== FUNC_ELLIPSIS
) {
1384 int seen_reg_num
, seen_sse_num
, seen_stack_size
;
1385 seen_reg_num
= seen_sse_num
= 0;
1386 /* frame pointer and return address */
1387 seen_stack_size
= PTR_SIZE
* 2;
1388 /* count the number of seen parameters */
1389 sym
= func_type
->ref
;
1390 while ((sym
= sym
->next
) != NULL
) {
1392 mode
= classify_x86_64_arg(type
, NULL
, &size
, &align
, ®_count
);
1396 seen_stack_size
= ((seen_stack_size
+ align
- 1) & -align
) + size
;
1399 case x86_64_mode_integer
:
1400 if (seen_reg_num
+ reg_count
<= 8) {
1401 seen_reg_num
+= reg_count
;
1408 case x86_64_mode_sse
:
1409 if (seen_sse_num
+ reg_count
<= 8) {
1410 seen_sse_num
+= reg_count
;
1420 /* movl $0x????????, -0x10(%rbp) */
1422 gen_le32(seen_reg_num
* 8);
1423 /* movl $0x????????, -0xc(%rbp) */
1425 gen_le32(seen_sse_num
* 16 + 48);
1426 /* movl $0x????????, -0x8(%rbp) */
1428 gen_le32(seen_stack_size
);
1430 /* save all register passing arguments */
1431 for (i
= 0; i
< 8; i
++) {
1433 o(0xd60f66); /* movq */
1434 gen_modrm(7 - i
, VT_LOCAL
, NULL
, loc
);
1435 /* movq $0, loc+8(%rbp) */
1440 for (i
= 0; i
< REGN
; i
++) {
1441 push_arg_reg(REGN
-1-i
);
1445 sym
= func_type
->ref
;
1446 reg_param_index
= 0;
1447 sse_param_index
= 0;
1449 /* if the function returns a structure, then add an
1450 implicit pointer parameter */
1451 func_vt
= sym
->type
;
1452 mode
= classify_x86_64_arg(&func_vt
, NULL
, &size
, &align
, ®_count
);
1453 if (mode
== x86_64_mode_memory
) {
1454 push_arg_reg(reg_param_index
);
1458 /* define parameters */
1459 while ((sym
= sym
->next
) != NULL
) {
1461 mode
= classify_x86_64_arg(type
, NULL
, &size
, &align
, ®_count
);
1463 case x86_64_mode_sse
:
1464 if (sse_param_index
+ reg_count
<= 8) {
1465 /* save arguments passed by register */
1466 loc
-= reg_count
* 8;
1468 for (i
= 0; i
< reg_count
; ++i
) {
1469 o(0xd60f66); /* movq */
1470 gen_modrm(sse_param_index
, VT_LOCAL
, NULL
, param_addr
+ i
*8);
1474 addr
= (addr
+ align
- 1) & -align
;
1477 sse_param_index
+= reg_count
;
1481 case x86_64_mode_memory
:
1482 case x86_64_mode_x87
:
1483 addr
= (addr
+ align
- 1) & -align
;
1488 case x86_64_mode_integer
: {
1489 if (reg_param_index
+ reg_count
<= REGN
) {
1490 /* save arguments passed by register */
1491 loc
-= reg_count
* 8;
1493 for (i
= 0; i
< reg_count
; ++i
) {
1494 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, param_addr
+ i
*8);
1498 addr
= (addr
+ align
- 1) & -align
;
1501 reg_param_index
+= reg_count
;
1505 default: break; /* nothing to be done for x86_64_mode_none */
1507 sym_push(sym
->v
& ~SYM_FIELD
, type
,
1508 VT_LOCAL
| VT_LVAL
, param_addr
);
1512 /* generate function epilog */
1513 void gfunc_epilog(void)
1517 o(0xc9); /* leave */
1518 if (func_ret_sub
== 0) {
1521 o(0xc2); /* ret n */
1523 g(func_ret_sub
>> 8);
1525 /* align local size to word & save local variables */
1526 v
= (-loc
+ 15) & -16;
1528 ind
= func_sub_sp_offset
- FUNC_PROLOG_SIZE
;
1529 o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
1530 o(0xec8148); /* sub rsp, stacksize */
1537 /* generate a jump to a label */
1540 return psym(0xe9, t
);
1543 /* generate a jump to a fixed address */
1544 void gjmp_addr(int a
)
1552 oad(0xe9, a
- ind
- 5);
1556 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1557 int gtst(int inv
, int t
)
1561 v
= vtop
->r
& VT_VALMASK
;
1563 /* fast case : can jump directly since flags are set */
1564 if (vtop
->c
.i
& 0x100)
1566 /* This was a float compare. If the parity flag is set
1567 the result was unordered. For anything except != this
1568 means false and we don't jump (anding both conditions).
1569 For != this means true (oring both).
1570 Take care about inverting the test. We need to jump
1571 to our target if the result was unordered and test wasn't NE,
1572 otherwise if unordered we don't want to jump. */
1573 vtop
->c
.i
&= ~0x100;
1574 if (!inv
== (vtop
->c
.i
!= TOK_NE
))
1575 o(0x067a); /* jp +6 */
1579 t
= psym(0x8a, t
); /* jp t */
1583 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
1584 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
1585 /* && or || optimization */
1586 if ((v
& 1) == inv
) {
1587 /* insert vtop->c jump list in t */
1590 p
= (int *)(cur_text_section
->data
+ *p
);
1598 if (is_float(vtop
->type
.t
) ||
1599 (vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1603 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1604 /* constant jmp optimization */
1605 if ((vtop
->c
.i
!= 0) != inv
)
1610 o(0xc0 + REG_VALUE(v
) * 9);
1612 t
= psym(0x85 ^ inv
, t
);
1619 /* generate an integer binary operation */
1620 void gen_opi(int op
)
1625 ll
= is64_type(vtop
[-1].type
.t
);
1626 uu
= (vtop
[-1].type
.t
& VT_UNSIGNED
) != 0;
1627 cc
= (vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
;
1631 case TOK_ADDC1
: /* add with carry generation */
1634 if (cc
&& (!ll
|| (int)vtop
->c
.ll
== vtop
->c
.ll
)) {
1641 /* XXX: generate inc and dec for smaller code ? */
1642 orex(ll
, r
, 0, 0x83);
1643 o(0xc0 | (opc
<< 3) | REG_VALUE(r
));
1646 orex(ll
, r
, 0, 0x81);
1647 oad(0xc0 | (opc
<< 3) | REG_VALUE(r
), c
);
1650 gv2(RC_INT
, RC_INT
);
1653 orex(ll
, r
, fr
, (opc
<< 3) | 0x01);
1654 o(0xc0 + REG_VALUE(r
) + REG_VALUE(fr
) * 8);
1657 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1663 case TOK_SUBC1
: /* sub with carry generation */
1666 case TOK_ADDC2
: /* add with carry use */
1669 case TOK_SUBC2
: /* sub with carry use */
1682 gv2(RC_INT
, RC_INT
);
1685 orex(ll
, fr
, r
, 0xaf0f); /* imul fr, r */
1686 o(0xc0 + REG_VALUE(fr
) + REG_VALUE(r
) * 8);
1698 opc
= 0xc0 | (opc
<< 3);
1704 orex(ll
, r
, 0, 0xc1); /* shl/shr/sar $xxx, r */
1705 o(opc
| REG_VALUE(r
));
1706 g(vtop
->c
.i
& (ll
? 63 : 31));
1708 /* we generate the shift in ecx */
1709 gv2(RC_INT
, RC_RCX
);
1711 orex(ll
, r
, 0, 0xd3); /* shl/shr/sar %cl, r */
1712 o(opc
| REG_VALUE(r
));
1725 /* first operand must be in eax */
1726 /* XXX: need better constraint for second operand */
1727 gv2(RC_RAX
, RC_RCX
);
1732 orex(ll
, 0, 0, uu
? 0xd231 : 0x99); /* xor %edx,%edx : cqto */
1733 orex(ll
, fr
, 0, 0xf7); /* div fr, %eax */
1734 o((uu
? 0xf0 : 0xf8) + REG_VALUE(fr
));
1735 if (op
== '%' || op
== TOK_UMOD
)
1747 void gen_opl(int op
)
1752 /* generate a floating point operation 'v = t1 op t2' instruction. The
1753 two operands are guaranted to have the same floating point type */
1754 /* XXX: need to use ST1 too */
1755 void gen_opf(int op
)
1757 int a
, ft
, fc
, swapped
, r
;
1759 (vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
? RC_ST0
: RC_FLOAT
;
1761 /* convert constants to memory references */
1762 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
1767 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
1770 /* must put at least one value in the floating point register */
1771 if ((vtop
[-1].r
& VT_LVAL
) &&
1772 (vtop
[0].r
& VT_LVAL
)) {
1778 /* swap the stack if needed so that t1 is the register and t2 is
1779 the memory reference */
1780 if (vtop
[-1].r
& VT_LVAL
) {
1784 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
) {
1785 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1786 /* load on stack second operand */
1787 load(TREG_ST0
, vtop
);
1788 save_reg(TREG_RAX
); /* eax is used by FP comparison code */
1789 if (op
== TOK_GE
|| op
== TOK_GT
)
1791 else if (op
== TOK_EQ
|| op
== TOK_NE
)
1794 o(0xc9d9); /* fxch %st(1) */
1795 o(0xe9da); /* fucompp */
1796 o(0xe0df); /* fnstsw %ax */
1798 o(0x45e480); /* and $0x45, %ah */
1799 o(0x40fC80); /* cmp $0x40, %ah */
1800 } else if (op
== TOK_NE
) {
1801 o(0x45e480); /* and $0x45, %ah */
1802 o(0x40f480); /* xor $0x40, %ah */
1804 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
1805 o(0x05c4f6); /* test $0x05, %ah */
1808 o(0x45c4f6); /* test $0x45, %ah */
1815 /* no memory reference possible for long double operations */
1816 load(TREG_ST0
, vtop
);
1840 o(0xde); /* fxxxp %st, %st(1) */
1845 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1846 /* if saved lvalue, then we must reload it */
1849 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
1851 r
= get_reg(RC_INT
);
1853 v1
.r
= VT_LOCAL
| VT_LVAL
;
1859 if (op
== TOK_EQ
|| op
== TOK_NE
) {
1862 if (op
== TOK_LE
|| op
== TOK_LT
)
1864 if (op
== TOK_LE
|| op
== TOK_GE
) {
1865 op
= 0x93; /* setae */
1867 op
= 0x97; /* seta */
1875 assert(!(vtop
[-1].r
& VT_LVAL
));
1877 if ((vtop
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1879 o(0x2e0f); /* ucomisd */
1881 if (vtop
->r
& VT_LVAL
) {
1882 gen_modrm(vtop
[-1].r
, r
, vtop
->sym
, fc
);
1884 o(0xc0 + REG_VALUE(vtop
[0].r
) + REG_VALUE(vtop
[-1].r
)*8);
1889 vtop
->c
.i
= op
| 0x100;
1891 assert((vtop
->type
.t
& VT_BTYPE
) != VT_LDOUBLE
);
1909 assert((ft
& VT_BTYPE
) != VT_LDOUBLE
);
1912 /* if saved lvalue, then we must reload it */
1913 if ((vtop
->r
& VT_VALMASK
) == VT_LLOCAL
) {
1915 r
= get_reg(RC_INT
);
1917 v1
.r
= VT_LOCAL
| VT_LVAL
;
1923 assert(!(vtop
[-1].r
& VT_LVAL
));
1925 assert(vtop
->r
& VT_LVAL
);
1930 if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
1938 if (vtop
->r
& VT_LVAL
) {
1939 gen_modrm(vtop
[-1].r
, r
, vtop
->sym
, fc
);
1941 o(0xc0 + REG_VALUE(vtop
[0].r
) + REG_VALUE(vtop
[-1].r
)*8);
1949 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1950 and 'long long' cases. */
1951 void gen_cvt_itof(int t
)
1953 if ((t
& VT_BTYPE
) == VT_LDOUBLE
) {
1956 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1957 /* signed long long to float/double/long double (unsigned case
1958 is handled generically) */
1959 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
1960 o(0x242cdf); /* fildll (%rsp) */
1961 o(0x08c48348); /* add $8, %rsp */
1962 } else if ((vtop
->type
.t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
1963 (VT_INT
| VT_UNSIGNED
)) {
1964 /* unsigned int to float/double/long double */
1965 o(0x6a); /* push $0 */
1967 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
1968 o(0x242cdf); /* fildll (%rsp) */
1969 o(0x10c48348); /* add $16, %rsp */
1971 /* int to float/double/long double */
1972 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
1973 o(0x2404db); /* fildl (%rsp) */
1974 o(0x08c48348); /* add $8, %rsp */
1978 int r
= get_reg(RC_FLOAT
);
1980 o(0xf2 + ((t
& VT_BTYPE
) == VT_FLOAT
?1:0));
1981 if ((vtop
->type
.t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
1982 (VT_INT
| VT_UNSIGNED
) ||
1983 (vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1987 o(0xc0 + (vtop
->r
& VT_VALMASK
) + REG_VALUE(r
)*8); /* cvtsi2sd */
1992 /* convert from one floating point type to another */
1993 void gen_cvt_ftof(int t
)
2001 if (bt
== VT_FLOAT
) {
2003 if (tbt
== VT_DOUBLE
) {
2004 o(0x140f); /* unpcklps */
2005 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2006 o(0x5a0f); /* cvtps2pd */
2007 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2008 } else if (tbt
== VT_LDOUBLE
) {
2010 /* movss %xmm0,-0x10(%rsp) */
2012 o(0x44 + REG_VALUE(vtop
->r
)*8);
2014 o(0xf02444d9); /* flds -0x10(%rsp) */
2017 } else if (bt
== VT_DOUBLE
) {
2019 if (tbt
== VT_FLOAT
) {
2020 o(0x140f66); /* unpcklpd */
2021 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2022 o(0x5a0f66); /* cvtpd2ps */
2023 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2024 } else if (tbt
== VT_LDOUBLE
) {
2026 /* movsd %xmm0,-0x10(%rsp) */
2028 o(0x44 + REG_VALUE(vtop
->r
)*8);
2030 o(0xf02444dd); /* fldl -0x10(%rsp) */
2036 r
= get_reg(RC_FLOAT
);
2037 if (tbt
== VT_DOUBLE
) {
2038 o(0xf0245cdd); /* fstpl -0x10(%rsp) */
2039 /* movsd -0x10(%rsp),%xmm0 */
2041 o(0x44 + REG_VALUE(r
)*8);
2044 } else if (tbt
== VT_FLOAT
) {
2045 o(0xf0245cd9); /* fstps -0x10(%rsp) */
2046 /* movss -0x10(%rsp),%xmm0 */
2048 o(0x44 + REG_VALUE(r
)*8);
2055 /* convert fp to int 't' type */
2056 void gen_cvt_ftoi(int t
)
2058 int ft
, bt
, size
, r
;
2061 if (bt
== VT_LDOUBLE
) {
2062 gen_cvt_ftof(VT_DOUBLE
);
2072 r
= get_reg(RC_INT
);
2073 if (bt
== VT_FLOAT
) {
2075 } else if (bt
== VT_DOUBLE
) {
2080 orex(size
== 8, r
, 0, 0x2c0f); /* cvttss2si or cvttsd2si */
2081 o(0xc0 + REG_VALUE(vtop
->r
) + REG_VALUE(r
)*8);
2085 /* computed goto support */
2092 /* Save the stack pointer onto the stack and return the location of its address */
2093 ST_FUNC
void gen_vla_sp_save(int addr
) {
2094 /* mov %rsp,addr(%rbp)*/
2095 gen_modrm64(0x89, TREG_RSP
, VT_LOCAL
, NULL
, addr
);
2098 /* Restore the SP from a location on the stack */
2099 ST_FUNC
void gen_vla_sp_restore(int addr
) {
2100 gen_modrm64(0x8b, TREG_RSP
, VT_LOCAL
, NULL
, addr
);
2103 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2104 ST_FUNC
void gen_vla_alloc(CType
*type
, int align
) {
2105 #ifdef TCC_TARGET_PE
2106 /* alloca does more than just adjust %rsp on Windows */
2107 vpush_global_sym(&func_old_type
, TOK_alloca
);
2108 vswap(); /* Move alloca ref past allocation size */
2110 vset(type
, REG_IRET
, 0);
2113 r
= gv(RC_INT
); /* allocation size */
2116 o(0xe0 | REG_VALUE(r
));
2117 /* We align to 16 bytes rather than align */
2122 o(0xe0 | REG_VALUE(r
));
2129 /* end of x86-64 code generator */
2130 /*************************************************************/
2131 #endif /* ! TARGET_DEFS_ONLY */
2132 /******************************************************/