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 1 if this function returns via an sret pointer, 0 otherwise */
660 ST_FUNC
int gfunc_sret(CType
*vt
, CType
*ret
, int *ret_align
)
663 *ret_align
= 1; // Never have to re-align return values for x86-64
664 size
= type_size(vt
, &align
);
668 } else if (size
> 4) {
671 } else if (size
> 2) {
674 } else if (size
> 1) {
683 static int is_sse_float(int t
) {
686 return bt
== VT_DOUBLE
|| bt
== VT_FLOAT
;
689 int gfunc_arg_size(CType
*type
) {
691 if (type
->t
& (VT_ARRAY
|VT_BITFIELD
))
693 return type_size(type
, &align
);
696 void gfunc_call(int nb_args
)
698 int size
, r
, args_size
, i
, d
, bt
, struct_size
;
701 args_size
= (nb_args
< REGN
? REGN
: nb_args
) * PTR_SIZE
;
704 /* for struct arguments, we need to call memcpy and the function
705 call breaks register passing arguments we are preparing.
706 So, we process arguments which will be passed by stack first. */
707 struct_size
= args_size
;
708 for(i
= 0; i
< nb_args
; i
++) {
713 bt
= (sv
->type
.t
& VT_BTYPE
);
714 size
= gfunc_arg_size(&sv
->type
);
717 continue; /* arguments smaller than 8 bytes passed in registers or on stack */
719 if (bt
== VT_STRUCT
) {
720 /* align to stack align size */
721 size
= (size
+ 15) & ~15;
722 /* generate structure store */
724 gen_offs_sp(0x8d, r
, struct_size
);
727 /* generate memcpy call */
728 vset(&sv
->type
, r
| VT_LVAL
, 0);
732 } else if (bt
== VT_LDOUBLE
) {
734 gen_offs_sp(0xdb, 0x107, struct_size
);
739 if (func_scratch
< struct_size
)
740 func_scratch
= struct_size
;
743 struct_size
= args_size
;
745 for(i
= 0; i
< nb_args
; i
++) {
747 bt
= (vtop
->type
.t
& VT_BTYPE
);
749 size
= gfunc_arg_size(&vtop
->type
);
751 /* align to stack align size */
752 size
= (size
+ 15) & ~15;
755 gen_offs_sp(0x8d, d
, struct_size
);
756 gen_offs_sp(0x89, d
, arg
*8);
758 d
= arg_prepare_reg(arg
);
759 gen_offs_sp(0x8d, d
, struct_size
);
763 if (is_sse_float(vtop
->type
.t
)) {
764 gv(RC_XMM0
); /* only use one float register */
766 /* movq %xmm0, j*8(%rsp) */
767 gen_offs_sp(0xd60f66, 0x100, arg
*8);
769 /* movaps %xmm0, %xmmN */
771 o(0xc0 + (arg
<< 3));
772 d
= arg_prepare_reg(arg
);
773 /* mov %xmm0, %rxx */
776 o(0xc0 + REG_VALUE(d
));
779 if (bt
== VT_STRUCT
) {
780 vtop
->type
.ref
= NULL
;
781 vtop
->type
.t
= size
> 4 ? VT_LLONG
: size
> 2 ? VT_INT
782 : size
> 1 ? VT_SHORT
: VT_BYTE
;
787 gen_offs_sp(0x89, r
, arg
*8);
789 d
= arg_prepare_reg(arg
);
790 orex(1,d
,r
,0x89); /* mov */
791 o(0xc0 + REG_VALUE(r
) * 8 + REG_VALUE(d
));
799 /* Copy R10 and R11 into RCX and RDX, respectively */
801 o(0xd1894c); /* mov %r10, %rcx */
803 o(0xda894c); /* mov %r11, %rdx */
812 #define FUNC_PROLOG_SIZE 11
814 /* generate function prolog of type 't' */
815 void gfunc_prolog(CType
*func_type
)
817 int addr
, reg_param_index
, bt
, size
;
826 ind
+= FUNC_PROLOG_SIZE
;
827 func_sub_sp_offset
= ind
;
830 sym
= func_type
->ref
;
832 /* if the function returns a structure, then add an
833 implicit pointer parameter */
835 size
= gfunc_arg_size(&func_vt
);
837 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
843 /* define parameters */
844 while ((sym
= sym
->next
) != NULL
) {
846 bt
= type
->t
& VT_BTYPE
;
847 size
= gfunc_arg_size(type
);
849 if (reg_param_index
< REGN
) {
850 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
852 sym_push(sym
->v
& ~SYM_FIELD
, type
, VT_LOCAL
| VT_LVAL
| VT_REF
, addr
);
854 if (reg_param_index
< REGN
) {
855 /* save arguments passed by register */
856 if ((bt
== VT_FLOAT
) || (bt
== VT_DOUBLE
)) {
857 o(0xd60f66); /* movq */
858 gen_modrm(reg_param_index
, VT_LOCAL
, NULL
, addr
);
860 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
863 sym_push(sym
->v
& ~SYM_FIELD
, type
, VT_LOCAL
| VT_LVAL
, addr
);
869 while (reg_param_index
< REGN
) {
870 if (func_type
->ref
->c
== FUNC_ELLIPSIS
) {
871 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, addr
);
878 /* generate function epilog */
879 void gfunc_epilog(void)
884 if (func_ret_sub
== 0) {
889 g(func_ret_sub
>> 8);
893 ind
= func_sub_sp_offset
- FUNC_PROLOG_SIZE
;
894 /* align local size to word & save local variables */
895 v
= (func_scratch
+ -loc
+ 15) & -16;
898 Sym
*sym
= external_global_sym(TOK___chkstk
, &func_old_type
, 0);
899 oad(0xb8, v
); /* mov stacksize, %eax */
900 oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
901 greloc(cur_text_section
, sym
, ind
-4, R_X86_64_PC32
);
902 o(0x90); /* fill for FUNC_PROLOG_SIZE = 11 bytes */
904 o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
905 o(0xec8148); /* sub rsp, stacksize */
909 cur_text_section
->data_offset
= saved_ind
;
910 pe_add_unwind_data(ind
, saved_ind
, v
);
911 ind
= cur_text_section
->data_offset
;
916 static void gadd_sp(int val
)
918 if (val
== (char)val
) {
922 oad(0xc48148, val
); /* add $xxx, %rsp */
926 typedef enum X86_64_Mode
{
934 static X86_64_Mode
classify_x86_64_merge(X86_64_Mode a
, X86_64_Mode b
) {
937 else if (a
== x86_64_mode_none
)
939 else if (b
== x86_64_mode_none
)
941 else if ((a
== x86_64_mode_memory
) || (b
== x86_64_mode_memory
))
942 return x86_64_mode_memory
;
943 else if ((a
== x86_64_mode_integer
) || (b
== x86_64_mode_integer
))
944 return x86_64_mode_integer
;
945 else if ((a
== x86_64_mode_x87
) || (b
== x86_64_mode_x87
))
946 return x86_64_mode_memory
;
948 return x86_64_mode_sse
;
951 static X86_64_Mode
classify_x86_64_inner(CType
*ty
) {
955 switch (ty
->t
& VT_BTYPE
) {
956 case VT_VOID
: return x86_64_mode_none
;
965 case VT_ENUM
: return x86_64_mode_integer
;
968 case VT_DOUBLE
: return x86_64_mode_sse
;
970 case VT_LDOUBLE
: return x86_64_mode_x87
;
976 if (f
->next
&& (f
->c
== f
->next
->c
))
977 return x86_64_mode_memory
;
979 mode
= x86_64_mode_none
;
980 for (; f
; f
= f
->next
)
981 mode
= classify_x86_64_merge(mode
, classify_x86_64_inner(&f
->type
));
989 static X86_64_Mode
classify_x86_64_arg(CType
*ty
, CType
*ret
, int *psize
, int *palign
, int *reg_count
) {
991 int size
, align
, ret_t
= 0;
993 if (ty
->t
& (VT_BITFIELD
|VT_ARRAY
)) {
997 mode
= x86_64_mode_integer
;
999 size
= type_size(ty
, &align
);
1000 *psize
= (size
+ 7) & ~7;
1001 *palign
= (align
+ 7) & ~7;
1004 mode
= x86_64_mode_memory
;
1006 mode
= classify_x86_64_inner(ty
);
1008 case x86_64_mode_integer
:
1014 ret_t
= (size
> 4) ? VT_LLONG
: VT_INT
;
1018 case x86_64_mode_x87
:
1023 case x86_64_mode_sse
:
1029 ret_t
= (size
> 4) ? VT_DOUBLE
: VT_FLOAT
;
1032 default: break; /* nothing to be done for x86_64_mode_memory and x86_64_mode_none*/
1045 ST_FUNC
int classify_x86_64_va_arg(CType
*ty
) {
1046 /* This definition must be synced with stdarg.h */
1047 enum __va_arg_type
{
1048 __va_gen_reg
, __va_float_reg
, __va_stack
1050 int size
, align
, reg_count
;
1051 X86_64_Mode mode
= classify_x86_64_arg(ty
, NULL
, &size
, &align
, ®_count
);
1053 default: return __va_stack
;
1054 case x86_64_mode_integer
: return __va_gen_reg
;
1055 case x86_64_mode_sse
: return __va_float_reg
;
1059 /* Return 1 if this function returns via an sret pointer, 0 otherwise */
1060 int gfunc_sret(CType
*vt
, CType
*ret
, int *ret_align
) {
1061 int size
, align
, reg_count
;
1062 *ret_align
= 1; // Never have to re-align return values for x86-64
1063 return (classify_x86_64_arg(vt
, ret
, &size
, &align
, ®_count
) == x86_64_mode_memory
);
1067 static const uint8_t arg_regs
[REGN
] = {
1068 TREG_RDI
, TREG_RSI
, TREG_RDX
, TREG_RCX
, TREG_R8
, TREG_R9
1071 static int arg_prepare_reg(int idx
) {
1072 if (idx
== 2 || idx
== 3)
1073 /* idx=2: r10, idx=3: r11 */
1076 return arg_regs
[idx
];
1079 /* Generate function call. The function address is pushed first, then
1080 all the parameters in call order. This functions pops all the
1081 parameters and the function address. */
1082 void gfunc_call(int nb_args
)
1086 int size
, align
, r
, args_size
, stack_adjust
, run_start
, run_end
, i
, reg_count
;
1087 int nb_reg_args
= 0;
1088 int nb_sse_args
= 0;
1089 int sse_reg
, gen_reg
;
1091 /* calculate the number of integer/float register arguments */
1092 for(i
= 0; i
< nb_args
; i
++) {
1093 mode
= classify_x86_64_arg(&vtop
[-i
].type
, NULL
, &size
, &align
, ®_count
);
1094 if (mode
== x86_64_mode_sse
)
1095 nb_sse_args
+= reg_count
;
1096 else if (mode
== x86_64_mode_integer
)
1097 nb_reg_args
+= reg_count
;
1100 /* arguments are collected in runs. Each run is a collection of 8-byte aligned arguments
1101 and ended by a 16-byte aligned argument. This is because, from the point of view of
1102 the callee, argument alignment is computed from the bottom up. */
1103 /* for struct arguments, we need to call memcpy and the function
1104 call breaks register passing arguments we are preparing.
1105 So, we process arguments which will be passed by stack first. */
1106 gen_reg
= nb_reg_args
;
1107 sse_reg
= nb_sse_args
;
1110 while (run_start
!= nb_args
) {
1111 int run_gen_reg
= gen_reg
, run_sse_reg
= sse_reg
;
1115 for(i
= run_start
; (i
< nb_args
) && (run_end
== nb_args
); i
++) {
1116 mode
= classify_x86_64_arg(&vtop
[-i
].type
, NULL
, &size
, &align
, ®_count
);
1118 case x86_64_mode_memory
:
1119 case x86_64_mode_x87
:
1124 stack_adjust
+= size
;
1127 case x86_64_mode_sse
:
1128 sse_reg
-= reg_count
;
1129 if (sse_reg
+ reg_count
> 8) goto stack_arg
;
1132 case x86_64_mode_integer
:
1133 gen_reg
-= reg_count
;
1134 if (gen_reg
+ reg_count
> REGN
) goto stack_arg
;
1136 default: break; /* nothing to be done for x86_64_mode_none */
1140 gen_reg
= run_gen_reg
;
1141 sse_reg
= run_sse_reg
;
1143 /* adjust stack to align SSE boundary */
1144 if (stack_adjust
&= 15) {
1145 /* fetch cpu flag before the following sub will change the value */
1146 if (vtop
>= vstack
&& (vtop
->r
& VT_VALMASK
) == VT_CMP
)
1149 stack_adjust
= 16 - stack_adjust
;
1151 oad(0xec81, stack_adjust
); /* sub $xxx, %rsp */
1152 args_size
+= stack_adjust
;
1155 for(i
= run_start
; i
< run_end
;) {
1156 /* Swap argument to top, it will possibly be changed here,
1157 and might use more temps. At the end of the loop we keep
1158 in on the stack and swap it back to its original position
1159 if it is a register. */
1160 SValue tmp
= vtop
[0];
1164 mode
= classify_x86_64_arg(&vtop
->type
, NULL
, &size
, &align
, ®_count
);
1167 switch (vtop
->type
.t
& VT_BTYPE
) {
1169 if (mode
== x86_64_mode_sse
) {
1171 sse_reg
-= reg_count
;
1174 } else if (mode
== x86_64_mode_integer
) {
1176 gen_reg
-= reg_count
;
1182 /* allocate the necessary size on stack */
1184 oad(0xec81, size
); /* sub $xxx, %rsp */
1185 /* generate structure store */
1186 r
= get_reg(RC_INT
);
1187 orex(1, r
, 0, 0x89); /* mov %rsp, r */
1188 o(0xe0 + REG_VALUE(r
));
1189 vset(&vtop
->type
, r
| VT_LVAL
, 0);
1202 assert(mode
== x86_64_mode_sse
);
1206 o(0x50); /* push $rax */
1207 /* movq %xmmN, (%rsp) */
1209 o(0x04 + REG_VALUE(r
)*8);
1218 assert(mode
== x86_64_mode_integer
);
1220 /* XXX: implicit cast ? */
1221 if (gen_reg
> REGN
) {
1224 orex(0,r
,0,0x50 + REG_VALUE(r
)); /* push r */
1232 /* And swap the argument back to it's original position. */
1239 assert((vtop
->type
.t
== tmp
.type
.t
) && (vtop
->r
== tmp
.r
));
1248 /* handle 16 byte aligned arguments at end of run */
1249 run_start
= i
= run_end
;
1250 while (i
< nb_args
) {
1251 /* Rotate argument to top since it will always be popped */
1252 mode
= classify_x86_64_arg(&vtop
[-i
].type
, NULL
, &size
, &align
, ®_count
);
1258 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
) {
1260 oad(0xec8148, size
); /* sub $xxx, %rsp */
1261 o(0x7cdb); /* fstpt 0(%rsp) */
1266 assert(mode
== x86_64_mode_memory
);
1268 /* allocate the necessary size on stack */
1270 oad(0xec81, size
); /* sub $xxx, %rsp */
1271 /* generate structure store */
1272 r
= get_reg(RC_INT
);
1273 orex(1, r
, 0, 0x89); /* mov %rsp, r */
1274 o(0xe0 + REG_VALUE(r
));
1275 vset(&vtop
->type
, r
| VT_LVAL
, 0);
1286 /* XXX This should be superfluous. */
1287 save_regs(0); /* save used temporary registers */
1289 /* then, we prepare register passing arguments.
1290 Note that we cannot set RDX and RCX in this loop because gv()
1291 may break these temporary registers. Let's use R10 and R11
1293 assert(gen_reg
<= REGN
);
1294 assert(sse_reg
<= 8);
1295 for(i
= 0; i
< nb_args
; i
++) {
1296 mode
= classify_x86_64_arg(&vtop
->type
, &type
, &size
, &align
, ®_count
);
1297 /* Alter stack entry type so that gv() knows how to treat it */
1299 if (mode
== x86_64_mode_sse
) {
1300 if (reg_count
== 2) {
1302 gv(RC_FRET
); /* Use pair load into xmm0 & xmm1 */
1303 if (sse_reg
) { /* avoid redundant movaps %xmm0, %xmm0 */
1304 /* movaps %xmm0, %xmmN */
1306 o(0xc0 + (sse_reg
<< 3));
1307 /* movaps %xmm1, %xmmN */
1309 o(0xc1 + ((sse_reg
+1) << 3));
1312 assert(reg_count
== 1);
1314 /* Load directly to register */
1315 gv(RC_XMM0
<< sse_reg
);
1317 } else if (mode
== x86_64_mode_integer
) {
1319 /* XXX: implicit cast ? */
1320 gen_reg
-= reg_count
;
1322 int d
= arg_prepare_reg(gen_reg
);
1323 orex(1,d
,r
,0x89); /* mov */
1324 o(0xc0 + REG_VALUE(r
) * 8 + REG_VALUE(d
));
1325 if (reg_count
== 2) {
1326 d
= arg_prepare_reg(gen_reg
+1);
1327 orex(1,d
,vtop
->r2
,0x89); /* mov */
1328 o(0xc0 + REG_VALUE(vtop
->r2
) * 8 + REG_VALUE(d
));
1333 assert(gen_reg
== 0);
1334 assert(sse_reg
== 0);
1336 /* We shouldn't have many operands on the stack anymore, but the
1337 call address itself is still there, and it might be in %eax
1338 (or edx/ecx) currently, which the below writes would clobber.
1339 So evict all remaining operands here. */
1342 /* Copy R10 and R11 into RDX and RCX, respectively */
1343 if (nb_reg_args
> 2) {
1344 o(0xd2894c); /* mov %r10, %rdx */
1345 if (nb_reg_args
> 3) {
1346 o(0xd9894c); /* mov %r11, %rcx */
1350 oad(0xb8, nb_sse_args
< 8 ? nb_sse_args
: 8); /* mov nb_sse_args, %eax */
1358 #define FUNC_PROLOG_SIZE 11
1360 static void push_arg_reg(int i
) {
1362 gen_modrm64(0x89, arg_regs
[i
], VT_LOCAL
, NULL
, loc
);
1365 /* generate function prolog of type 't' */
1366 void gfunc_prolog(CType
*func_type
)
1369 int i
, addr
, align
, size
, reg_count
;
1370 int param_addr
= 0, reg_param_index
, sse_param_index
;
1374 sym
= func_type
->ref
;
1375 addr
= PTR_SIZE
* 2;
1377 ind
+= FUNC_PROLOG_SIZE
;
1378 func_sub_sp_offset
= ind
;
1381 if (func_type
->ref
->c
== FUNC_ELLIPSIS
) {
1382 int seen_reg_num
, seen_sse_num
, seen_stack_size
;
1383 seen_reg_num
= seen_sse_num
= 0;
1384 /* frame pointer and return address */
1385 seen_stack_size
= PTR_SIZE
* 2;
1386 /* count the number of seen parameters */
1387 sym
= func_type
->ref
;
1388 while ((sym
= sym
->next
) != NULL
) {
1390 mode
= classify_x86_64_arg(type
, NULL
, &size
, &align
, ®_count
);
1394 seen_stack_size
= ((seen_stack_size
+ align
- 1) & -align
) + size
;
1397 case x86_64_mode_integer
:
1398 if (seen_reg_num
+ reg_count
<= 8) {
1399 seen_reg_num
+= reg_count
;
1406 case x86_64_mode_sse
:
1407 if (seen_sse_num
+ reg_count
<= 8) {
1408 seen_sse_num
+= reg_count
;
1418 /* movl $0x????????, -0x10(%rbp) */
1420 gen_le32(seen_reg_num
* 8);
1421 /* movl $0x????????, -0xc(%rbp) */
1423 gen_le32(seen_sse_num
* 16 + 48);
1424 /* movl $0x????????, -0x8(%rbp) */
1426 gen_le32(seen_stack_size
);
1428 /* save all register passing arguments */
1429 for (i
= 0; i
< 8; i
++) {
1431 o(0xd60f66); /* movq */
1432 gen_modrm(7 - i
, VT_LOCAL
, NULL
, loc
);
1433 /* movq $0, loc+8(%rbp) */
1438 for (i
= 0; i
< REGN
; i
++) {
1439 push_arg_reg(REGN
-1-i
);
1443 sym
= func_type
->ref
;
1444 reg_param_index
= 0;
1445 sse_param_index
= 0;
1447 /* if the function returns a structure, then add an
1448 implicit pointer parameter */
1449 func_vt
= sym
->type
;
1450 mode
= classify_x86_64_arg(&func_vt
, NULL
, &size
, &align
, ®_count
);
1451 if (mode
== x86_64_mode_memory
) {
1452 push_arg_reg(reg_param_index
);
1456 /* define parameters */
1457 while ((sym
= sym
->next
) != NULL
) {
1459 mode
= classify_x86_64_arg(type
, NULL
, &size
, &align
, ®_count
);
1461 case x86_64_mode_sse
:
1462 if (sse_param_index
+ reg_count
<= 8) {
1463 /* save arguments passed by register */
1464 loc
-= reg_count
* 8;
1466 for (i
= 0; i
< reg_count
; ++i
) {
1467 o(0xd60f66); /* movq */
1468 gen_modrm(sse_param_index
, VT_LOCAL
, NULL
, param_addr
+ i
*8);
1472 addr
= (addr
+ align
- 1) & -align
;
1475 sse_param_index
+= reg_count
;
1479 case x86_64_mode_memory
:
1480 case x86_64_mode_x87
:
1481 addr
= (addr
+ align
- 1) & -align
;
1486 case x86_64_mode_integer
: {
1487 if (reg_param_index
+ reg_count
<= REGN
) {
1488 /* save arguments passed by register */
1489 loc
-= reg_count
* 8;
1491 for (i
= 0; i
< reg_count
; ++i
) {
1492 gen_modrm64(0x89, arg_regs
[reg_param_index
], VT_LOCAL
, NULL
, param_addr
+ i
*8);
1496 addr
= (addr
+ align
- 1) & -align
;
1499 reg_param_index
+= reg_count
;
1503 default: break; /* nothing to be done for x86_64_mode_none */
1505 sym_push(sym
->v
& ~SYM_FIELD
, type
,
1506 VT_LOCAL
| VT_LVAL
, param_addr
);
1510 /* generate function epilog */
1511 void gfunc_epilog(void)
1515 o(0xc9); /* leave */
1516 if (func_ret_sub
== 0) {
1519 o(0xc2); /* ret n */
1521 g(func_ret_sub
>> 8);
1523 /* align local size to word & save local variables */
1524 v
= (-loc
+ 15) & -16;
1526 ind
= func_sub_sp_offset
- FUNC_PROLOG_SIZE
;
1527 o(0xe5894855); /* push %rbp, mov %rsp, %rbp */
1528 o(0xec8148); /* sub rsp, stacksize */
1535 /* generate a jump to a label */
1538 return psym(0xe9, t
);
1541 /* generate a jump to a fixed address */
1542 void gjmp_addr(int a
)
1550 oad(0xe9, a
- ind
- 5);
1554 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1555 int gtst(int inv
, int t
)
1559 v
= vtop
->r
& VT_VALMASK
;
1561 /* fast case : can jump directly since flags are set */
1562 if (vtop
->c
.i
& 0x100)
1564 /* This was a float compare. If the parity flag is set
1565 the result was unordered. For anything except != this
1566 means false and we don't jump (anding both conditions).
1567 For != this means true (oring both).
1568 Take care about inverting the test. We need to jump
1569 to our target if the result was unordered and test wasn't NE,
1570 otherwise if unordered we don't want to jump. */
1571 vtop
->c
.i
&= ~0x100;
1572 if (!inv
== (vtop
->c
.i
!= TOK_NE
))
1573 o(0x067a); /* jp +6 */
1577 t
= psym(0x8a, t
); /* jp t */
1581 t
= psym((vtop
->c
.i
- 16) ^ inv
, t
);
1582 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
1583 /* && or || optimization */
1584 if ((v
& 1) == inv
) {
1585 /* insert vtop->c jump list in t */
1588 p
= (int *)(cur_text_section
->data
+ *p
);
1596 if (is_float(vtop
->type
.t
) ||
1597 (vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1601 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1602 /* constant jmp optimization */
1603 if ((vtop
->c
.i
!= 0) != inv
)
1608 o(0xc0 + REG_VALUE(v
) * 9);
1610 t
= psym(0x85 ^ inv
, t
);
1617 /* generate an integer binary operation */
1618 void gen_opi(int op
)
1623 ll
= is64_type(vtop
[-1].type
.t
);
1624 uu
= (vtop
[-1].type
.t
& VT_UNSIGNED
) != 0;
1625 cc
= (vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
;
1629 case TOK_ADDC1
: /* add with carry generation */
1632 if (cc
&& (!ll
|| (int)vtop
->c
.ll
== vtop
->c
.ll
)) {
1639 /* XXX: generate inc and dec for smaller code ? */
1640 orex(ll
, r
, 0, 0x83);
1641 o(0xc0 | (opc
<< 3) | REG_VALUE(r
));
1644 orex(ll
, r
, 0, 0x81);
1645 oad(0xc0 | (opc
<< 3) | REG_VALUE(r
), c
);
1648 gv2(RC_INT
, RC_INT
);
1651 orex(ll
, r
, fr
, (opc
<< 3) | 0x01);
1652 o(0xc0 + REG_VALUE(r
) + REG_VALUE(fr
) * 8);
1655 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1661 case TOK_SUBC1
: /* sub with carry generation */
1664 case TOK_ADDC2
: /* add with carry use */
1667 case TOK_SUBC2
: /* sub with carry use */
1680 gv2(RC_INT
, RC_INT
);
1683 orex(ll
, fr
, r
, 0xaf0f); /* imul fr, r */
1684 o(0xc0 + REG_VALUE(fr
) + REG_VALUE(r
) * 8);
1696 opc
= 0xc0 | (opc
<< 3);
1702 orex(ll
, r
, 0, 0xc1); /* shl/shr/sar $xxx, r */
1703 o(opc
| REG_VALUE(r
));
1704 g(vtop
->c
.i
& (ll
? 63 : 31));
1706 /* we generate the shift in ecx */
1707 gv2(RC_INT
, RC_RCX
);
1709 orex(ll
, r
, 0, 0xd3); /* shl/shr/sar %cl, r */
1710 o(opc
| REG_VALUE(r
));
1723 /* first operand must be in eax */
1724 /* XXX: need better constraint for second operand */
1725 gv2(RC_RAX
, RC_RCX
);
1730 orex(ll
, 0, 0, uu
? 0xd231 : 0x99); /* xor %edx,%edx : cqto */
1731 orex(ll
, fr
, 0, 0xf7); /* div fr, %eax */
1732 o((uu
? 0xf0 : 0xf8) + REG_VALUE(fr
));
1733 if (op
== '%' || op
== TOK_UMOD
)
1745 void gen_opl(int op
)
1750 /* generate a floating point operation 'v = t1 op t2' instruction. The
1751 two operands are guaranted to have the same floating point type */
1752 /* XXX: need to use ST1 too */
1753 void gen_opf(int op
)
1755 int a
, ft
, fc
, swapped
, r
;
1757 (vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
? RC_ST0
: RC_FLOAT
;
1759 /* convert constants to memory references */
1760 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
1765 if ((vtop
[0].r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
)
1768 /* must put at least one value in the floating point register */
1769 if ((vtop
[-1].r
& VT_LVAL
) &&
1770 (vtop
[0].r
& VT_LVAL
)) {
1776 /* swap the stack if needed so that t1 is the register and t2 is
1777 the memory reference */
1778 if (vtop
[-1].r
& VT_LVAL
) {
1782 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
) {
1783 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1784 /* load on stack second operand */
1785 load(TREG_ST0
, vtop
);
1786 save_reg(TREG_RAX
); /* eax is used by FP comparison code */
1787 if (op
== TOK_GE
|| op
== TOK_GT
)
1789 else if (op
== TOK_EQ
|| op
== TOK_NE
)
1792 o(0xc9d9); /* fxch %st(1) */
1793 o(0xe9da); /* fucompp */
1794 o(0xe0df); /* fnstsw %ax */
1796 o(0x45e480); /* and $0x45, %ah */
1797 o(0x40fC80); /* cmp $0x40, %ah */
1798 } else if (op
== TOK_NE
) {
1799 o(0x45e480); /* and $0x45, %ah */
1800 o(0x40f480); /* xor $0x40, %ah */
1802 } else if (op
== TOK_GE
|| op
== TOK_LE
) {
1803 o(0x05c4f6); /* test $0x05, %ah */
1806 o(0x45c4f6); /* test $0x45, %ah */
1813 /* no memory reference possible for long double operations */
1814 load(TREG_ST0
, vtop
);
1838 o(0xde); /* fxxxp %st, %st(1) */
1843 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1844 /* if saved lvalue, then we must reload it */
1847 if ((r
& VT_VALMASK
) == VT_LLOCAL
) {
1849 r
= get_reg(RC_INT
);
1851 v1
.r
= VT_LOCAL
| VT_LVAL
;
1857 if (op
== TOK_EQ
|| op
== TOK_NE
) {
1860 if (op
== TOK_LE
|| op
== TOK_LT
)
1862 if (op
== TOK_LE
|| op
== TOK_GE
) {
1863 op
= 0x93; /* setae */
1865 op
= 0x97; /* seta */
1873 assert(!(vtop
[-1].r
& VT_LVAL
));
1875 if ((vtop
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1877 o(0x2e0f); /* ucomisd */
1879 if (vtop
->r
& VT_LVAL
) {
1880 gen_modrm(vtop
[-1].r
, r
, vtop
->sym
, fc
);
1882 o(0xc0 + REG_VALUE(vtop
[0].r
) + REG_VALUE(vtop
[-1].r
)*8);
1887 vtop
->c
.i
= op
| 0x100;
1889 assert((vtop
->type
.t
& VT_BTYPE
) != VT_LDOUBLE
);
1907 assert((ft
& VT_BTYPE
) != VT_LDOUBLE
);
1910 /* if saved lvalue, then we must reload it */
1911 if ((vtop
->r
& VT_VALMASK
) == VT_LLOCAL
) {
1913 r
= get_reg(RC_INT
);
1915 v1
.r
= VT_LOCAL
| VT_LVAL
;
1921 assert(!(vtop
[-1].r
& VT_LVAL
));
1923 assert(vtop
->r
& VT_LVAL
);
1928 if ((ft
& VT_BTYPE
) == VT_DOUBLE
) {
1936 if (vtop
->r
& VT_LVAL
) {
1937 gen_modrm(vtop
[-1].r
, r
, vtop
->sym
, fc
);
1939 o(0xc0 + REG_VALUE(vtop
[0].r
) + REG_VALUE(vtop
[-1].r
)*8);
1947 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1948 and 'long long' cases. */
1949 void gen_cvt_itof(int t
)
1951 if ((t
& VT_BTYPE
) == VT_LDOUBLE
) {
1954 if ((vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1955 /* signed long long to float/double/long double (unsigned case
1956 is handled generically) */
1957 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
1958 o(0x242cdf); /* fildll (%rsp) */
1959 o(0x08c48348); /* add $8, %rsp */
1960 } else if ((vtop
->type
.t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
1961 (VT_INT
| VT_UNSIGNED
)) {
1962 /* unsigned int to float/double/long double */
1963 o(0x6a); /* push $0 */
1965 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
1966 o(0x242cdf); /* fildll (%rsp) */
1967 o(0x10c48348); /* add $16, %rsp */
1969 /* int to float/double/long double */
1970 o(0x50 + (vtop
->r
& VT_VALMASK
)); /* push r */
1971 o(0x2404db); /* fildl (%rsp) */
1972 o(0x08c48348); /* add $8, %rsp */
1976 int r
= get_reg(RC_FLOAT
);
1978 o(0xf2 + ((t
& VT_BTYPE
) == VT_FLOAT
?1:0));
1979 if ((vtop
->type
.t
& (VT_BTYPE
| VT_UNSIGNED
)) ==
1980 (VT_INT
| VT_UNSIGNED
) ||
1981 (vtop
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1985 o(0xc0 + (vtop
->r
& VT_VALMASK
) + REG_VALUE(r
)*8); /* cvtsi2sd */
1990 /* convert from one floating point type to another */
1991 void gen_cvt_ftof(int t
)
1999 if (bt
== VT_FLOAT
) {
2001 if (tbt
== VT_DOUBLE
) {
2002 o(0x140f); /* unpcklps */
2003 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2004 o(0x5a0f); /* cvtps2pd */
2005 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2006 } else if (tbt
== VT_LDOUBLE
) {
2008 /* movss %xmm0,-0x10(%rsp) */
2010 o(0x44 + REG_VALUE(vtop
->r
)*8);
2012 o(0xf02444d9); /* flds -0x10(%rsp) */
2015 } else if (bt
== VT_DOUBLE
) {
2017 if (tbt
== VT_FLOAT
) {
2018 o(0x140f66); /* unpcklpd */
2019 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2020 o(0x5a0f66); /* cvtpd2ps */
2021 o(0xc0 + REG_VALUE(vtop
->r
)*9);
2022 } else if (tbt
== VT_LDOUBLE
) {
2024 /* movsd %xmm0,-0x10(%rsp) */
2026 o(0x44 + REG_VALUE(vtop
->r
)*8);
2028 o(0xf02444dd); /* fldl -0x10(%rsp) */
2034 r
= get_reg(RC_FLOAT
);
2035 if (tbt
== VT_DOUBLE
) {
2036 o(0xf0245cdd); /* fstpl -0x10(%rsp) */
2037 /* movsd -0x10(%rsp),%xmm0 */
2039 o(0x44 + REG_VALUE(r
)*8);
2042 } else if (tbt
== VT_FLOAT
) {
2043 o(0xf0245cd9); /* fstps -0x10(%rsp) */
2044 /* movss -0x10(%rsp),%xmm0 */
2046 o(0x44 + REG_VALUE(r
)*8);
2053 /* convert fp to int 't' type */
2054 void gen_cvt_ftoi(int t
)
2056 int ft
, bt
, size
, r
;
2059 if (bt
== VT_LDOUBLE
) {
2060 gen_cvt_ftof(VT_DOUBLE
);
2070 r
= get_reg(RC_INT
);
2071 if (bt
== VT_FLOAT
) {
2073 } else if (bt
== VT_DOUBLE
) {
2078 orex(size
== 8, r
, 0, 0x2c0f); /* cvttss2si or cvttsd2si */
2079 o(0xc0 + REG_VALUE(vtop
->r
) + REG_VALUE(r
)*8);
2083 /* computed goto support */
2090 /* Save the stack pointer onto the stack and return the location of its address */
2091 ST_FUNC
void gen_vla_sp_save(int addr
) {
2092 /* mov %rsp,addr(%rbp)*/
2093 gen_modrm64(0x89, TREG_RSP
, VT_LOCAL
, NULL
, addr
);
2096 /* Restore the SP from a location on the stack */
2097 ST_FUNC
void gen_vla_sp_restore(int addr
) {
2098 gen_modrm64(0x8b, TREG_RSP
, VT_LOCAL
, NULL
, addr
);
2101 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2102 ST_FUNC
void gen_vla_alloc(CType
*type
, int align
) {
2103 #ifdef TCC_TARGET_PE
2104 /* alloca does more than just adjust %rsp on Windows */
2105 vpush_global_sym(&func_old_type
, TOK_alloca
);
2106 vswap(); /* Move alloca ref past allocation size */
2108 vset(type
, REG_IRET
, 0);
2111 r
= gv(RC_INT
); /* allocation size */
2114 o(0xe0 | REG_VALUE(r
));
2115 /* We align to 16 bytes rather than align */
2120 o(0xe0 | REG_VALUE(r
));
2127 /* end of x86-64 code generator */
2128 /*************************************************************/
2129 #endif /* ! TARGET_DEFS_ONLY */
2130 /******************************************************/