2 * ARMv4 code generator for TCC
4 * Copyright (c) 2003 Daniel Glöckner
5 * Copyright (c) 2012 Thomas Preud'homme
7 * Based on i386-gen.c by Fabrice Bellard
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2 of the License, or (at your option) any later version.
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
24 #ifdef TARGET_DEFS_ONLY
26 #if defined(TCC_ARM_EABI) && !defined(TCC_ARM_VFP)
27 #error "Currently TinyCC only supports float computation with VFP instructions"
30 /* number of available registers */
39 #ifndef TCC_ARM_VERSION
40 #define TCC_ARM_VERSION 5
43 /* a register can belong to several classes. The classes must be
44 sorted from more general to more precise (see gv2() code which does
45 assumptions on it). */
46 #define RC_INT 0x0001 /* generic integer register */
47 #define RC_FLOAT 0x0002 /* generic float register */
63 #define RC_IRET RC_R0 /* function return: integer register */
64 #define RC_LRET RC_R1 /* function return: second integer register */
65 #define RC_FRET RC_F0 /* function return: float register */
67 /* pretty names for the registers */
87 #define T2CPR(t) (((t)&VT_BTYPE) != VT_FLOAT ? 0x100 : 0)
90 /* return registers for function */
91 #define REG_IRET TREG_R0 /* single word int return register */
92 #define REG_LRET TREG_R1 /* second word return register (for long long) */
93 #define REG_FRET TREG_F0 /* float return register */
96 #define TOK___divdi3 TOK___aeabi_ldivmod
97 #define TOK___moddi3 TOK___aeabi_ldivmod
98 #define TOK___udivdi3 TOK___aeabi_uldivmod
99 #define TOK___umoddi3 TOK___aeabi_uldivmod
102 /* defined if function parameters must be evaluated in reverse order */
103 #define INVERT_FUNC_PARAMS
105 /* defined if structures are passed as pointers. Otherwise structures
106 are directly pushed on stack. */
107 /* #define FUNC_STRUCT_PARAM_AS_PTR */
109 /* pointer size, in bytes */
112 /* long double size and alignment, in bytes */
114 #define LDOUBLE_SIZE 8
118 #define LDOUBLE_SIZE 8
122 #define LDOUBLE_ALIGN 8
124 #define LDOUBLE_ALIGN 4
127 /* maximum alignment (for aligned attribute support) */
130 #define CHAR_IS_UNSIGNED
132 /******************************************************/
135 #define EM_TCC_TARGET EM_ARM
137 /* relocation type for 32 bit data relocation */
138 #define R_DATA_32 R_ARM_ABS32
139 #define R_DATA_PTR R_ARM_ABS32
140 #define R_JMP_SLOT R_ARM_JUMP_SLOT
141 #define R_COPY R_ARM_COPY
143 #define ELF_START_ADDR 0x00008000
144 #define ELF_PAGE_SIZE 0x1000
151 /******************************************************/
152 #else /* ! TARGET_DEFS_ONLY */
153 /******************************************************/
156 enum float_abi float_abi
;
158 ST_DATA
const int reg_classes
[NB_REGS
] = {
159 /* r0 */ RC_INT
| RC_R0
,
160 /* r1 */ RC_INT
| RC_R1
,
161 /* r2 */ RC_INT
| RC_R2
,
162 /* r3 */ RC_INT
| RC_R3
,
163 /* r12 */ RC_INT
| RC_R12
,
164 /* f0 */ RC_FLOAT
| RC_F0
,
165 /* f1 */ RC_FLOAT
| RC_F1
,
166 /* f2 */ RC_FLOAT
| RC_F2
,
167 /* f3 */ RC_FLOAT
| RC_F3
,
169 /* d4/s8 */ RC_FLOAT
| RC_F4
,
170 /* d5/s10 */ RC_FLOAT
| RC_F5
,
171 /* d6/s12 */ RC_FLOAT
| RC_F6
,
172 /* d7/s14 */ RC_FLOAT
| RC_F7
,
176 static int func_sub_sp_offset
, last_itod_magic
;
179 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
180 static CType float_type
, double_type
, func_float_type
, func_double_type
;
181 ST_FUNC
void arm_init(struct TCCState
* s
)
183 float_type
.t
= VT_FLOAT
;
184 double_type
.t
= VT_DOUBLE
;
185 func_float_type
.t
= VT_FUNC
;
186 func_float_type
.ref
=
187 sym_push(SYM_FIELD
, &float_type
, FUNC_CDECL
, FUNC_OLD
);
188 func_double_type
.t
= VT_FUNC
;
189 func_double_type
.ref
=
190 sym_push(SYM_FIELD
, &double_type
, FUNC_CDECL
, FUNC_OLD
);
192 float_abi
= s
->float_abi
;
193 #ifndef TCC_ARM_HARDFLOAT
194 tcc_warning("soft float ABI currently not supported: default to softfp");
198 #define func_float_type func_old_type
199 #define func_double_type func_old_type
200 #define func_ldouble_type func_old_type
201 ST_FUNC
void arm_init(struct TCCState
* s
)
203 #if !defined(TCC_ARM_VFP)
204 tcc_warning("Support for FPA is deprecated and will be removed in next"
207 #if !defined(TCC_ARM_EABI)
208 tcc_warning("Support for OABI is deprecated and will be removed in next"
214 static int two2mask(int a
, int b
)
216 return (reg_classes
[a
] | reg_classes
[b
]) & ~(RC_INT
| RC_FLOAT
);
219 static int regmask(int r
)
221 return reg_classes
[r
] & ~(RC_INT
| RC_FLOAT
);
224 /******************************************************/
226 #if defined(TCC_ARM_EABI) && !defined(CONFIG_TCC_ELFINTERP)
227 char* default_elfinterp(struct TCCState
* s
)
229 if (s
->float_abi
== ARM_HARD_FLOAT
)
230 return "/lib/ld-linux-armhf.so.3";
232 return "/lib/ld-linux.so.3";
238 /* this is a good place to start adding big-endian support*/
242 if (!cur_text_section
)
243 tcc_error("compiler error! This happens f.ex. if the compiler\n"
244 "can't evaluate constant expressions outside of a function.");
245 if (ind1
> cur_text_section
->data_allocated
)
246 section_realloc(cur_text_section
, ind1
);
247 cur_text_section
->data
[ind
++] = i
& 255;
249 cur_text_section
->data
[ind
++] = i
& 255;
251 cur_text_section
->data
[ind
++] = i
& 255;
253 cur_text_section
->data
[ind
++] = i
;
256 static uint32_t stuff_const(uint32_t op
, uint32_t c
)
259 uint32_t nc
= 0, negop
= 0;
261 switch (op
& 0x1F00000) {
262 case 0x800000: // add
263 case 0x400000: // sub
265 negop
= op
^ 0xC00000;
268 case 0x1A00000: // mov
269 case 0x1E00000: // mvn
271 negop
= op
^ 0x400000;
274 case 0x200000: // xor
276 return (op
& 0xF010F000) | ((op
>> 16) & 0xF) | 0x1E00000;
280 return (op
& 0xF010F000) | ((op
>> 16) & 0xF) | 0x1A00000;
281 case 0x1C00000: // bic
283 negop
= op
^ 0x1C00000;
286 case 0x1800000: // orr
288 return (op
& 0xFFF0FFFF) | 0x1E00000;
294 if (c
< 256) /* catch undefined <<32 */
296 for (i
= 2; i
< 32; i
+= 2) {
297 m
= (0xff >> i
) | (0xff << (32 - i
));
299 return op
| (i
<< 7) | (c
<< i
) | (c
>> (32 - i
));
308 void stuff_const_harder(uint32_t op
, uint32_t v
)
311 x
= stuff_const(op
, v
);
315 uint32_t a
[16], nv
, no
, o2
, n2
;
318 o2
= (op
& 0xfff0ffff) | ((op
& 0xf000) << 4);
320 for (i
= 1; i
< 16; i
++)
321 a
[i
] = (a
[i
- 1] >> 2) | (a
[i
- 1] << 30);
322 for (i
= 0; i
< 12; i
++)
323 for (j
= i
< 4 ? i
+ 12 : 15; j
>= i
+ 4; j
--)
324 if ((v
& (a
[i
] | a
[j
])) == v
) {
325 o(stuff_const(op
, v
& a
[i
]));
326 o(stuff_const(o2
, v
& a
[j
]));
332 for (i
= 0; i
< 12; i
++)
333 for (j
= i
< 4 ? i
+ 12 : 15; j
>= i
+ 4; j
--)
334 if ((nv
& (a
[i
] | a
[j
])) == nv
) {
335 o(stuff_const(no
, nv
& a
[i
]));
336 o(stuff_const(n2
, nv
& a
[j
]));
339 for (i
= 0; i
< 8; i
++)
340 for (j
= i
+ 4; j
< 12; j
++)
341 for (k
= i
< 4 ? i
+ 12 : 15; k
>= j
+ 4; k
--)
342 if ((v
& (a
[i
] | a
[j
] | a
[k
])) == v
) {
343 o(stuff_const(op
, v
& a
[i
]));
344 o(stuff_const(o2
, v
& a
[j
]));
345 o(stuff_const(o2
, v
& a
[k
]));
350 for (i
= 0; i
< 8; i
++)
351 for (j
= i
+ 4; j
< 12; j
++)
352 for (k
= i
< 4 ? i
+ 12 : 15; k
>= j
+ 4; k
--)
353 if ((nv
& (a
[i
] | a
[j
] | a
[k
])) == nv
) {
354 o(stuff_const(no
, nv
& a
[i
]));
355 o(stuff_const(n2
, nv
& a
[j
]));
356 o(stuff_const(n2
, nv
& a
[k
]));
359 o(stuff_const(op
, v
& a
[0]));
360 o(stuff_const(o2
, v
& a
[4]));
361 o(stuff_const(o2
, v
& a
[8]));
362 o(stuff_const(o2
, v
& a
[12]));
366 ST_FUNC
uint32_t encbranch(int pos
, int addr
, int fail
)
370 if (addr
>= 0x1000000 || addr
< -0x1000000) {
372 tcc_error("FIXME: function bigger than 32MB");
375 return 0x0A000000 | (addr
& 0xffffff);
378 int decbranch(int pos
)
381 x
= *(uint32_t*)(cur_text_section
->data
+ pos
);
385 return x
* 4 + pos
+ 8;
388 /* output a symbol and patch all calls to it */
389 void gsym_addr(int t
, int a
)
394 x
= (uint32_t*)(cur_text_section
->data
+ t
);
395 t
= decbranch(lt
= t
);
397 *x
= 0xE1A00000; // nop
400 *x
|= encbranch(lt
, a
, 1);
411 static uint32_t vfpr(int r
)
413 if (r
< TREG_F0
|| r
> TREG_F7
)
414 tcc_error("compiler error! register %i is no vfp register", r
);
418 static uint32_t fpr(int r
)
420 if (r
< TREG_F0
|| r
> TREG_F3
)
421 tcc_error("compiler error! register %i is no fpa register", r
);
426 static uint32_t intr(int r
)
430 if ((r
< 0 || r
> 4) && r
!= 14)
431 tcc_error("compiler error! register %i is no int register", r
);
435 static void calcaddr(uint32_t* base
, int* off
, int* sgn
, int maxoff
,
438 if (*off
> maxoff
|| *off
& ((1 << shift
) - 1)) {
445 y
= stuff_const(x
, *off
& ~maxoff
);
451 y
= stuff_const(x
, (*off
+ maxoff
) & ~maxoff
);
455 *off
= ((*off
+ maxoff
) & ~maxoff
) - *off
;
458 stuff_const_harder(x
, *off
& ~maxoff
);
463 static uint32_t mapcc(int cc
)
467 return 0x30000000; /* CC/LO */
469 return 0x20000000; /* CS/HS */
471 return 0x00000000; /* EQ */
473 return 0x10000000; /* NE */
475 return 0x90000000; /* LS */
477 return 0x80000000; /* HI */
479 return 0x40000000; /* MI */
481 return 0x50000000; /* PL */
483 return 0xB0000000; /* LT */
485 return 0xA0000000; /* GE */
487 return 0xD0000000; /* LE */
489 return 0xC0000000; /* GT */
491 tcc_error("unexpected condition code");
492 return 0xE0000000; /* AL */
495 static int negcc(int cc
)
523 tcc_error("unexpected condition code");
527 /* load 'r' from value 'sv' */
528 void load(int r
, SValue
* sv
)
530 int v
, ft
, fc
, fr
, sign
;
547 uint32_t base
= 0xB; // fp
548 if (v
== VT_LLOCAL
) {
550 v1
.r
= VT_LOCAL
| VT_LVAL
;
552 load(base
= 14 /* lr */, &v1
);
555 } else if (v
== VT_CONST
) {
557 v1
.r
= fr
& ~VT_LVAL
;
560 load(base
= 14, &v1
);
563 } else if (v
< VT_CONST
) {
570 calcaddr(&base
, &fc
, &sign
, 1020, 2);
572 op
= 0xED100A00; /* flds */
575 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
576 op
|= 0x100; /* flds -> fldd */
577 o(op
| (vfpr(r
) << 12) | (fc
>> 2) | (base
<< 16));
582 #if LDOUBLE_SIZE == 8
583 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
586 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
588 else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
)
591 o(op
| (fpr(r
) << 12) | (fc
>> 2) | (base
<< 16));
593 } else if ((ft
& (VT_BTYPE
| VT_UNSIGNED
)) == VT_BYTE
||
594 (ft
& VT_BTYPE
) == VT_SHORT
) {
595 calcaddr(&base
, &fc
, &sign
, 255, 0);
597 if ((ft
& VT_BTYPE
) == VT_SHORT
)
599 if ((ft
& VT_UNSIGNED
) == 0)
603 o(op
| (intr(r
) << 12) | (base
<< 16) | ((fc
& 0xf0) << 4) |
606 calcaddr(&base
, &fc
, &sign
, 4095, 0);
610 if ((ft
& VT_BTYPE
) == VT_BYTE
|| (ft
& VT_BTYPE
) == VT_BOOL
)
612 o(op
| (intr(r
) << 12) | fc
| (base
<< 16));
618 op
= stuff_const(0xE3A00000 | (intr(r
) << 12), sv
->c
.ul
);
619 if (fr
& VT_SYM
|| !op
) {
620 o(0xE59F0000 | (intr(r
) << 12));
623 greloc(cur_text_section
, sv
->sym
, ind
, R_ARM_ABS32
);
628 } else if (v
== VT_LOCAL
) {
629 op
= stuff_const(0xE28B0000 | (intr(r
) << 12), sv
->c
.ul
);
630 if (fr
& VT_SYM
|| !op
) {
631 o(0xE59F0000 | (intr(r
) << 12));
633 if (fr
& VT_SYM
) // needed ?
634 greloc(cur_text_section
, sv
->sym
, ind
, R_ARM_ABS32
);
636 o(0xE08B0000 | (intr(r
) << 12) | intr(r
));
640 } else if (v
== VT_CMP
) {
641 o(mapcc(sv
->c
.ul
) | 0x3A00001 | (intr(r
) << 12));
642 o(mapcc(negcc(sv
->c
.ul
)) | 0x3A00000 | (intr(r
) << 12));
644 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
647 o(0xE3A00000 | (intr(r
) << 12) | t
);
650 o(0xE3A00000 | (intr(r
) << 12) | (t
^ 1));
652 } else if (v
< VT_CONST
) {
655 o(0xEEB00A40 | (vfpr(r
) << 12) | vfpr(v
) |
656 T2CPR(ft
)); /* fcpyX */
658 o(0xEE008180 | (fpr(r
) << 12) | fpr(v
));
661 o(0xE1A00000 | (intr(r
) << 12) | intr(v
));
665 tcc_error("load unimplemented!");
668 /* store register 'r' in lvalue 'v' */
669 void store(int r
, SValue
* sv
)
672 int v
, ft
, fc
, fr
, sign
;
687 if (fr
& VT_LVAL
|| fr
== VT_LOCAL
) {
693 } else if (v
== VT_CONST
) {
695 v1
.r
= fr
& ~VT_LVAL
;
698 load(base
= 14, &v1
);
704 calcaddr(&base
, &fc
, &sign
, 1020, 2);
706 op
= 0xED000A00; /* fsts */
709 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
710 op
|= 0x100; /* fsts -> fstd */
711 o(op
| (vfpr(r
) << 12) | (fc
>> 2) | (base
<< 16));
716 #if LDOUBLE_SIZE == 8
717 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
720 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
722 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
)
725 o(op
| (fpr(r
) << 12) | (fc
>> 2) | (base
<< 16));
728 } else if ((ft
& VT_BTYPE
) == VT_SHORT
) {
729 calcaddr(&base
, &fc
, &sign
, 255, 0);
733 o(op
| (intr(r
) << 12) | (base
<< 16) | ((fc
& 0xf0) << 4) |
736 calcaddr(&base
, &fc
, &sign
, 4095, 0);
740 if ((ft
& VT_BTYPE
) == VT_BYTE
|| (ft
& VT_BTYPE
) == VT_BOOL
)
742 o(op
| (intr(r
) << 12) | fc
| (base
<< 16));
747 tcc_error("store unimplemented");
750 static void gadd_sp(int val
)
752 stuff_const_harder(0xE28DD000, val
);
755 /* 'is_jmp' is '1' if it is a jump */
756 static void gcall_or_jmp(int is_jmp
)
759 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
762 x
= encbranch(ind
, ind
+ vtop
->c
.ul
, 0);
764 if (vtop
->r
& VT_SYM
) {
765 /* relocation case */
766 greloc(cur_text_section
, vtop
->sym
, ind
, R_ARM_PC24
);
768 put_elf_reloc(symtab_section
, cur_text_section
, ind
, R_ARM_PC24
,
770 o(x
| (is_jmp
? 0xE0000000 : 0xE1000000));
773 o(0xE28FE004); // add lr,pc,#4
774 o(0xE51FF004); // ldr pc,[pc,#-4]
775 if (vtop
->r
& VT_SYM
)
776 greloc(cur_text_section
, vtop
->sym
, ind
, R_ARM_ABS32
);
780 /* otherwise, indirect call */
783 o(0xE1A0E00F); // mov lr,pc
784 o(0xE1A0F000 | intr(r
)); // mov pc,r
788 /* Return whether a structure is an homogeneous float aggregate or not.
789 The answer is true if all the elements of the structure are of the same
790 primitive float type and there is less than 4 elements.
792 type: the type corresponding to the structure to be tested */
793 static int is_hgen_float_aggr(CType
* type
)
795 if ((type
->t
& VT_BTYPE
) == VT_STRUCT
) {
797 int btype
, nb_fields
= 0;
799 ref
= type
->ref
->next
;
800 btype
= ref
->type
.t
& VT_BTYPE
;
801 if (btype
== VT_FLOAT
|| btype
== VT_DOUBLE
) {
802 for (; ref
&& btype
== (ref
->type
.t
& VT_BTYPE
);
803 ref
= ref
->next
, nb_fields
++)
805 return !ref
&& nb_fields
<= 4;
813 avail
[3]; /* 3 holes max with only float and double alignments */
814 int first_hole
; /* first available hole */
815 int last_hole
; /* last available hole (none if equal to first_hole) */
816 int first_free_reg
; /* next free register in the sequence, hole excluded */
819 #define AVAIL_REGS_INITIALIZER (struct avail_regs) { { 0, 0, 0}, 0, 0, 0 }
821 /* Find suitable registers for a VFP Co-Processor Register Candidate (VFP CPRC
822 param) according to the rules described in the procedure call standard for
823 the ARM architecture (AAPCS). If found, the registers are assigned to this
824 VFP CPRC parameter. Registers are allocated in sequence unless a hole exists
825 and the parameter is a single float.
827 avregs: opaque structure to keep track of available VFP co-processor regs
828 align: alignment contraints for the param, as returned by type_size()
829 size: size of the parameter, as returned by type_size() */
830 int assign_vfpreg(struct avail_regs
* avregs
, int align
, int size
)
834 if (avregs
->first_free_reg
== -1)
836 if (align
>> 3) { /* double alignment */
837 first_reg
= avregs
->first_free_reg
;
838 /* alignment contraint not respected so use next reg and record hole */
840 avregs
->avail
[avregs
->last_hole
++] = first_reg
++;
841 } else { /* no special alignment (float or array of float) */
842 /* if single float and a hole is available, assign the param to it */
843 if (size
== 4 && avregs
->first_hole
!= avregs
->last_hole
)
844 return avregs
->avail
[avregs
->first_hole
++];
846 first_reg
= avregs
->first_free_reg
;
848 if (first_reg
+ size
/ 4 <= 16) {
849 avregs
->first_free_reg
= first_reg
+ size
/ 4;
852 avregs
->first_free_reg
= -1;
856 /* Returns whether all params need to be passed in core registers or not.
857 This is the case for function part of the runtime ABI. */
858 int floats_in_core_regs(SValue
* sval
)
863 switch (sval
->sym
->v
) {
864 case TOK___floatundisf
:
865 case TOK___floatundidf
:
866 case TOK___fixunssfdi
:
867 case TOK___fixunsdfdi
:
869 case TOK___fixunsxfdi
:
871 case TOK___floatdisf
:
872 case TOK___floatdidf
:
882 ST_FUNC
int regargs_nregs(RegArgs
* args
)
887 /* Return the number of registers needed to return the struct, or 0 if
888 returning via struct pointer. */
889 ST_FUNC
int gfunc_sret(CType
* vt
, int variadic
, CType
* ret
, int* ret_align
,
890 int* regsize
, RegArgs
* args
)
894 size
= type_size(vt
, &align
);
895 if (float_abi
== ARM_HARD_FLOAT
&& !variadic
&&
896 (is_float(vt
->t
) || is_hgen_float_aggr(vt
))) {
901 *args
= (size
+ 7) >> 3;
902 } else if (size
<= 4) {
917 /* Parameters are classified according to how they are copied to their final
918 destination for the function call. Because the copying is performed class
919 after class according to the order in the union below, it is important that
920 some constraints about the order of the members of this union are respected:
921 - CORE_STRUCT_CLASS must come after STACK_CLASS;
922 - CORE_CLASS must come after STACK_CLASS, CORE_STRUCT_CLASS and
924 - VFP_STRUCT_CLASS must come after VFP_CLASS.
925 See the comment for the main loop in copy_params() for the reason. */
936 int start
; /* first reg or addr used depending on the class */
937 int end
; /* last reg used or next free addr depending on the class */
938 SValue
* sval
; /* pointer to SValue on the value stack */
939 struct param_plan
* prev
; /* previous element in this class */
943 struct param_plan
* pplans
; /* array of all the param plans */
945 clsplans
[NB_CLASSES
]; /* per class lists of param plans */
948 #define add_param_plan(plan, pplan, class) \
950 pplan.prev = plan->clsplans[class]; \
951 plan->pplans[plan##_nb] = pplan; \
952 plan->clsplans[class] = &plan->pplans[plan##_nb++]; \
955 /* Assign parameters to registers and stack with alignment according to the
956 rules in the procedure call standard for the ARM architecture (AAPCS).
957 The overall assignment is recorded in an array of per parameter structures
958 called parameter plans. The parameter plans are also further organized in a
959 number of linked lists, one per class of parameter (see the comment for the
960 definition of union reg_class).
962 nb_args: number of parameters of the function for which a call is generated
963 float_abi: float ABI in use for this function call
964 plan: the structure where the overall assignment is recorded
965 todo: a bitmap that record which core registers hold a parameter
967 Returns the amount of stack space needed for parameter passing
969 Note: this function allocated an array in plan->pplans with tcc_malloc. It
970 is the responsibility of the caller to free this array once used (ie not
971 before copy_params). */
972 static int assign_regs(int nb_args
, int float_abi
, struct plan
* plan
, int* todo
)
975 int ncrn
/* next core register number */,
976 nsaa
/* next stacked argument address*/;
978 struct param_plan pplan
;
979 struct avail_regs avregs
= AVAIL_REGS_INITIALIZER
;
983 plan
->pplans
= tcc_malloc(nb_args
* sizeof(*plan
->pplans
));
984 memset(plan
->clsplans
, 0, sizeof(plan
->clsplans
));
985 for (i
= nb_args
; i
--;) {
986 int j
, start_vfpreg
= 0;
987 CType type
= vtop
[-i
].type
;
989 size
= type_size(&type
, &align
);
990 size
= (size
+ 3) & ~3;
991 align
= (align
+ 3) & ~3;
992 switch (vtop
[-i
].type
.t
& VT_BTYPE
) {
997 if (float_abi
== ARM_HARD_FLOAT
) {
998 int is_hfa
= 0; /* Homogeneous float aggregate */
1000 if (is_float(vtop
[-i
].type
.t
) ||
1001 (is_hfa
= is_hgen_float_aggr(&vtop
[-i
].type
))) {
1004 start_vfpreg
= assign_vfpreg(&avregs
, align
, size
);
1005 end_vfpreg
= start_vfpreg
+ ((size
- 1) >> 2);
1006 if (start_vfpreg
>= 0) {
1007 pplan
= (struct param_plan
){start_vfpreg
, end_vfpreg
,
1010 add_param_plan(plan
, pplan
, VFP_STRUCT_CLASS
);
1012 add_param_plan(plan
, pplan
, VFP_CLASS
);
1018 ncrn
= (ncrn
+ (align
- 1) / 4) & ~((align
/ 4) - 1);
1019 if (ncrn
+ size
/ 4 <= 4 || (ncrn
< 4 && start_vfpreg
!= -1)) {
1020 /* The parameter is allocated both in core register and on
1022 * such, it can be of either class: it would either be the last of
1023 * CORE_STRUCT_CLASS or the first of STACK_CLASS. */
1024 for (j
= ncrn
; j
< 4 && j
< ncrn
+ size
/ 4; j
++)
1026 pplan
= (struct param_plan
){ncrn
, j
, &vtop
[-i
]};
1027 add_param_plan(plan
, pplan
, CORE_STRUCT_CLASS
);
1030 nsaa
= (ncrn
- 4) * 4;
1038 int is_long
= (vtop
[-i
].type
.t
& VT_BTYPE
) == VT_LLONG
;
1041 ncrn
= (ncrn
+ 1) & -2;
1045 pplan
= (struct param_plan
){ncrn
, ncrn
, &vtop
[-i
]};
1049 add_param_plan(plan
, pplan
, CORE_CLASS
);
1053 nsaa
= (nsaa
+ (align
- 1)) & ~(align
- 1);
1054 pplan
= (struct param_plan
){nsaa
, nsaa
+ size
, &vtop
[-i
]};
1055 add_param_plan(plan
, pplan
, STACK_CLASS
);
1056 nsaa
+= size
; /* size already rounded up before */
1061 #undef add_param_plan
1063 /* Copy parameters to their final destination (core reg, VFP reg or stack) for
1066 nb_args: number of parameters the function take
1067 plan: the overall assignment plan for parameters
1068 todo: a bitmap indicating what core reg will hold a parameter
1070 Returns the number of SValue added by this function on the value stack */
1071 static int copy_params(int nb_args
, struct plan
* plan
, int todo
)
1073 int size
, align
, r
, i
, nb_extra_sval
= 0;
1074 struct param_plan
* pplan
;
1076 /* Several constraints require parameters to be copied in a specific order:
1077 - structures are copied to the stack before being loaded in a reg;
1078 - floats loaded to an odd numbered VFP reg are first copied to the
1079 preceding even numbered VFP reg and then moved to the next VFP reg.
1081 It is thus important that:
1082 - structures assigned to core regs must be copied after parameters
1083 assigned to the stack but before structures assigned to VFP regs
1084 because a structure can lie partly in core registers and partly on
1086 - parameters assigned to the stack and all structures be copied before
1087 parameters assigned to a core reg since copying a parameter to the
1088 stack require using a core reg;
1089 - parameters assigned to VFP regs be copied before structures assigned to
1090 VFP regs as the copy might use an even numbered VFP reg that already
1091 holds part of a structure. */
1092 for (i
= 0; i
< NB_CLASSES
; i
++) {
1093 for (pplan
= plan
->clsplans
[i
]; pplan
; pplan
= pplan
->prev
) {
1094 vpushv(pplan
->sval
);
1095 pplan
->sval
->r
= pplan
->sval
->r2
= VT_CONST
; /* disable entry */
1098 case CORE_STRUCT_CLASS
:
1099 case VFP_STRUCT_CLASS
:
1100 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_STRUCT
) {
1102 size
= type_size(&pplan
->sval
->type
, &align
);
1103 /* align to stack align size */
1104 size
= (size
+ 3) & ~3;
1105 if (i
== STACK_CLASS
&& pplan
->prev
)
1106 padding
= pplan
->start
- pplan
->prev
->end
;
1107 size
+= padding
; /* Add padding if any */
1108 /* allocate the necessary size on stack */
1110 /* generate structure store */
1111 r
= get_reg(RC_INT
);
1112 o(0xE28D0000 | (intr(r
) << 12) |
1113 padding
); /* add r, sp, padding */
1114 vset(&vtop
->type
, r
| VT_LVAL
, 0);
1116 vstore(); /* memcpy to current sp + potential padding */
1118 /* Homogeneous float aggregate are loaded to VFP registers
1119 immediately since there is no way of loading data in
1120 multiple non consecutive VFP registers as what is done
1121 for other structures (see the use of todo). */
1122 if (i
== VFP_STRUCT_CLASS
) {
1123 int first
= pplan
->start
, nb
= pplan
->end
- first
+ 1;
1124 /* vpop.32 {pplan->start, ..., pplan->end} */
1125 o(0xECBD0A00 | (first
& 1) << 22 | (first
>> 1) << 12 |
1127 /* No need to write the register used to a SValue since
1128 VFP regs cannot be used for gcall_or_jmp */
1131 if (is_float(pplan
->sval
->type
.t
)) {
1133 r
= vfpr(gv(RC_FLOAT
)) << 12;
1134 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
1138 r
|= 0x101; /* vpush.32 -> vpush.64 */
1140 o(0xED2D0A01 + r
); /* vpush */
1142 r
= fpr(gv(RC_FLOAT
)) << 12;
1143 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
1145 else if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1148 size
= LDOUBLE_SIZE
;
1155 o(0xED2D0100 | r
| (size
>> 2)); /* some kind of vpush for FPA */
1158 /* simple type (currently always same size) */
1159 /* XXX: implicit cast ? */
1161 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1165 o(0xE52D0004 | (intr(r
) << 12)); /* push r */
1169 o(0xE52D0004 | (intr(r
) << 12)); /* push r */
1171 if (i
== STACK_CLASS
&& pplan
->prev
)
1172 gadd_sp(pplan
->prev
->end
-
1173 pplan
->start
); /* Add padding if any */
1178 gv(regmask(TREG_F0
+ (pplan
->start
>> 1)));
1180 1) { /* Must be in upper part of double register */
1181 o(0xEEF00A40 | ((pplan
->start
>> 1) << 12) |
1182 (pplan
->start
>> 1)); /* vmov.f32 s(n+1), sn */
1184 VT_CONST
; /* avoid being saved on stack by gv for next
1190 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1192 gv(regmask(pplan
->end
));
1193 pplan
->sval
->r2
= vtop
->r
;
1196 gv(regmask(pplan
->start
));
1197 /* Mark register as used so that gcall_or_jmp use another one
1198 (regs >=4 are free as never used to pass parameters) */
1199 pplan
->sval
->r
= vtop
->r
;
1206 /* Manually free remaining registers since next parameters are loaded
1207 * manually, without the help of gv(int). */
1211 o(0xE8BD0000 | todo
); /* pop {todo} */
1212 for (pplan
= plan
->clsplans
[CORE_STRUCT_CLASS
]; pplan
;
1213 pplan
= pplan
->prev
) {
1215 pplan
->sval
->r
= pplan
->start
;
1216 /* An SValue can only pin 2 registers at best (r and r2) but a
1217 structure can occupy more than 2 registers. Thus, we need to
1218 push on the value stack some fake parameter to have on SValue
1219 for each registers used by a structure (r2 is not used). */
1220 for (r
= pplan
->start
+ 1; r
<= pplan
->end
; r
++) {
1221 if (todo
& (1 << r
)) {
1229 return nb_extra_sval
;
1232 /* Generate function call. The function address is pushed first, then
1233 all the parameters in call order. This functions pops all the
1234 parameters and the function address. */
1235 void gfunc_call(int nb_args
)
1238 int def_float_abi
= float_abi
;
1245 if (float_abi
== ARM_HARD_FLOAT
) {
1246 variadic
= (vtop
[-nb_args
].type
.ref
->c
== FUNC_ELLIPSIS
);
1247 if (variadic
|| floats_in_core_regs(&vtop
[-nb_args
]))
1248 float_abi
= ARM_SOFTFP_FLOAT
;
1251 /* cannot let cpu flags if other instruction are generated. Also avoid
1252 leaving VT_JMP anywhere except on the top of the stack because it
1253 would complicate the code generator. */
1254 r
= vtop
->r
& VT_VALMASK
;
1255 if (r
== VT_CMP
|| (r
& ~1) == VT_JMP
)
1258 args_size
= assign_regs(nb_args
, float_abi
, &plan
, &todo
);
1261 if (args_size
& 7) { /* Stack must be 8 byte aligned at fct call for EABI */
1262 args_size
= (args_size
+ 7) & ~7;
1263 o(0xE24DD004); /* sub sp, sp, #4 */
1267 nb_args
+= copy_params(nb_args
, &plan
, todo
);
1268 tcc_free(plan
.pplans
);
1270 /* Move fct SValue on top as required by gcall_or_jmp */
1274 gadd_sp(args_size
); /* pop all parameters passed on the stack */
1275 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
1276 if (float_abi
== ARM_SOFTFP_FLOAT
&& is_float(vtop
->type
.ref
->type
.t
)) {
1277 if ((vtop
->type
.ref
->type
.t
& VT_BTYPE
) == VT_FLOAT
) {
1278 o(0xEE000A10); /*vmov s0, r0 */
1280 o(0xEE000B10); /* vmov.32 d0[0], r0 */
1281 o(0xEE201B10); /* vmov.32 d0[1], r1 */
1285 vtop
-= nb_args
+ 1; /* Pop all params and fct address from value stack */
1287 0; /* we are calling a function, so we aren't in a leaf function */
1288 float_abi
= def_float_abi
;
1291 /* generate function prolog of type 't' */
1292 void gfunc_prolog(CType
* func_type
)
1295 int n
, nf
, size
, align
, rs
, struct_ret
= 0;
1296 int addr
, pn
, sn
; /* pn=core, sn=stack */
1301 struct avail_regs avregs
= AVAIL_REGS_INITIALIZER
;
1304 sym
= func_type
->ref
;
1305 func_vt
= sym
->type
;
1306 func_var
= (func_type
->ref
->c
== FUNC_ELLIPSIS
);
1309 if ((func_vt
.t
& VT_BTYPE
) == VT_STRUCT
&&
1310 !gfunc_sret(&func_vt
, func_var
, &ret_type
, &align
, &rs
, &dummy
)) {
1313 func_vc
= 12; /* Offset from fp of the place to store the result */
1315 for (sym2
= sym
->next
; sym2
&& (n
< 4 || nf
< 16); sym2
= sym2
->next
) {
1316 size
= type_size(&sym2
->type
, &align
);
1318 if (float_abi
== ARM_HARD_FLOAT
&& !func_var
&&
1319 (is_float(sym2
->type
.t
) || is_hgen_float_aggr(&sym2
->type
))) {
1320 int tmpnf
= assign_vfpreg(&avregs
, align
, size
);
1321 tmpnf
+= (size
+ 3) / 4;
1322 nf
= (tmpnf
> nf
) ? tmpnf
: nf
;
1326 n
+= (size
+ 3) / 4;
1328 o(0xE1A0C00D); /* mov ip,sp */
1337 o(0xE92D0000 | ((1 << n
) - 1)); /* save r0-r4 on stack if needed */
1342 nf
= (nf
+ 1) & -2; /* nf => HARDFLOAT => EABI */
1343 o(0xED2D0A00 | nf
); /* save s0-s15 on stack if needed */
1345 o(0xE92D5800); /* save fp, ip, lr */
1346 o(0xE1A0B00D); /* mov fp, sp */
1347 func_sub_sp_offset
= ind
;
1348 o(0xE1A00000); /* nop, leave space for stack adjustment in epilog */
1351 if (float_abi
== ARM_HARD_FLOAT
) {
1353 avregs
= AVAIL_REGS_INITIALIZER
;
1356 pn
= struct_ret
, sn
= 0;
1357 while ((sym
= sym
->next
)) {
1360 size
= type_size(type
, &align
);
1361 size
= (size
+ 3) >> 2;
1362 align
= (align
+ 3) & ~3;
1364 if (float_abi
== ARM_HARD_FLOAT
&& !func_var
&&
1365 (is_float(sym
->type
.t
) || is_hgen_float_aggr(&sym
->type
))) {
1366 int fpn
= assign_vfpreg(&avregs
, align
, size
<< 2);
1375 pn
= (pn
+ (align
- 1) / 4) & -(align
/ 4);
1377 addr
= (nf
+ pn
) * 4;
1384 sn
= (sn
+ (align
- 1) / 4) & -(align
/ 4);
1386 addr
= (n
+ nf
+ sn
) * 4;
1389 sym_push(sym
->v
& ~SYM_FIELD
, type
, VT_LOCAL
| lvalue_type(type
->t
),
1392 last_itod_magic
= 0;
1397 /* generate function epilog */
1398 void gfunc_epilog(void)
1402 /* Copy float return value to core register if base standard is used and
1403 float computation is made with VFP */
1404 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
1405 if ((float_abi
== ARM_SOFTFP_FLOAT
|| func_var
) && is_float(func_vt
.t
)) {
1406 if ((func_vt
.t
& VT_BTYPE
) == VT_FLOAT
)
1407 o(0xEE100A10); /* fmrs r0, s0 */
1409 o(0xEE100B10); /* fmrdl r0, d0 */
1410 o(0xEE301B10); /* fmrdh r1, d0 */
1414 o(0xE89BA800); /* restore fp, sp, pc */
1415 diff
= (-loc
+ 3) & -4;
1418 diff
= ((diff
+ 11) & -8) - 4;
1421 x
= stuff_const(0xE24BD000, diff
); /* sub sp,fp,# */
1423 *(uint32_t*)(cur_text_section
->data
+ func_sub_sp_offset
) = x
;
1427 o(0xE59FC004); /* ldr ip,[pc+4] */
1428 o(0xE04BD00C); /* sub sp,fp,ip */
1429 o(0xE1A0F00E); /* mov pc,lr */
1431 *(uint32_t*)(cur_text_section
->data
+ func_sub_sp_offset
) =
1432 0xE1000000 | encbranch(func_sub_sp_offset
, addr
, 1);
1437 /* generate a jump to a label */
1442 o(0xE0000000 | encbranch(r
, t
, 1));
1446 /* generate a jump to a fixed address */
1447 void gjmp_addr(int a
)
1452 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1453 int gtst(int inv
, int t
)
1457 v
= vtop
->r
& VT_VALMASK
;
1460 op
= mapcc(inv
? negcc(vtop
->c
.i
) : vtop
->c
.i
);
1461 op
|= encbranch(r
, t
, 1);
1464 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
1465 if ((v
& 1) == inv
) {
1474 p
= decbranch(lp
= p
);
1476 x
= (uint32_t*)(cur_text_section
->data
+ lp
);
1478 *x
|= encbranch(lp
, t
, 1);
1491 /* generate an integer binary operation */
1492 void gen_opi(int op
)
1495 uint32_t opc
= 0, r
, fr
;
1496 unsigned short retreg
= REG_IRET
;
1504 case TOK_ADDC1
: /* add with carry generation */
1512 case TOK_SUBC1
: /* sub with carry generation */
1516 case TOK_ADDC2
: /* add with carry use */
1520 case TOK_SUBC2
: /* sub with carry use */
1537 gv2(RC_INT
, RC_INT
);
1541 o(0xE0000090 | (intr(r
) << 16) | (intr(r
) << 8) | intr(fr
));
1557 func
= TOK___divsi3
;
1561 func
= TOK___udivsi3
;
1566 func
= TOK___aeabi_idivmod
;
1569 func
= TOK___modsi3
;
1575 func
= TOK___aeabi_uidivmod
;
1578 func
= TOK___umodsi3
;
1583 gv2(RC_INT
, RC_INT
);
1584 r
= intr(vtop
[-1].r2
= get_reg(RC_INT
));
1586 vtop
[-1].r
= get_reg_ex(RC_INT
, regmask(c
));
1588 o(0xE0800090 | (r
<< 16) | (intr(vtop
->r
) << 12) | (intr(c
) << 8) |
1598 if ((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1599 if (opc
== 4 || opc
== 5 || opc
== 0xc) {
1601 opc
|= 2; // sub -> rsb
1604 if ((vtop
->r
& VT_VALMASK
) == VT_CMP
||
1605 (vtop
->r
& (VT_VALMASK
& ~1)) == VT_JMP
)
1608 c
= intr(gv(RC_INT
));
1610 opc
= 0xE0000000 | (opc
<< 20) | (c
<< 16);
1611 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1613 x
= stuff_const(opc
| 0x2000000, vtop
->c
.i
);
1615 r
= intr(vtop
[-1].r
= get_reg_ex(RC_INT
, regmask(vtop
[-1].r
)));
1620 fr
= intr(gv(RC_INT
));
1621 r
= intr(vtop
[-1].r
=
1622 get_reg_ex(RC_INT
, two2mask(vtop
->r
, vtop
[-1].r
)));
1623 o(opc
| (r
<< 12) | fr
);
1626 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1632 opc
= 0xE1A00000 | (opc
<< 5);
1633 if ((vtop
->r
& VT_VALMASK
) == VT_CMP
||
1634 (vtop
->r
& (VT_VALMASK
& ~1)) == VT_JMP
)
1637 r
= intr(gv(RC_INT
));
1640 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1641 fr
= intr(vtop
[-1].r
= get_reg_ex(RC_INT
, regmask(vtop
[-1].r
)));
1642 c
= vtop
->c
.i
& 0x1f;
1643 o(opc
| (c
<< 7) | (fr
<< 12));
1645 fr
= intr(gv(RC_INT
));
1646 c
= intr(vtop
[-1].r
=
1647 get_reg_ex(RC_INT
, two2mask(vtop
->r
, vtop
[-1].r
)));
1648 o(opc
| (c
<< 12) | (fr
<< 8) | 0x10);
1653 vpush_global_sym(&func_old_type
, func
);
1660 tcc_error("gen_opi %i unimplemented!", op
);
1665 static int is_zero(int i
)
1667 if ((vtop
[i
].r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) != VT_CONST
)
1669 if (vtop
[i
].type
.t
== VT_FLOAT
)
1670 return (vtop
[i
].c
.f
== 0.f
);
1671 else if (vtop
[i
].type
.t
== VT_DOUBLE
)
1672 return (vtop
[i
].c
.d
== 0.0);
1673 return (vtop
[i
].c
.ld
== 0.l
);
1676 /* generate a floating point operation 'v = t1 op t2' instruction. The
1677 * two operands are guaranted to have the same floating point type */
1678 void gen_opf(int op
)
1682 x
= 0xEE000A00 | T2CPR(vtop
->type
.t
);
1700 x
|= 0x810000; /* fsubX -> fnegX */
1713 if (op
< TOK_ULT
|| op
> TOK_GT
) {
1714 tcc_error("unknown fp op %x!", op
);
1720 case TOK_LT
: op
= TOK_GT
; break;
1721 case TOK_GE
: op
= TOK_ULE
; break;
1722 case TOK_LE
: op
= TOK_GE
; break;
1723 case TOK_GT
: op
= TOK_ULT
; break;
1726 x
|= 0xB40040; /* fcmpX */
1727 if (op
!= TOK_EQ
&& op
!= TOK_NE
)
1728 x
|= 0x80; /* fcmpX -> fcmpeX */
1732 (vfpr(gv(RC_FLOAT
)) << 12)); /* fcmp(e)X -> fcmp(e)zX */
1734 x
|= vfpr(gv(RC_FLOAT
));
1736 o(x
| (vfpr(gv(RC_FLOAT
)) << 12));
1739 o(0xEEF1FA10); /* fmstat */
1742 case TOK_LE
: op
= TOK_ULE
; break;
1743 case TOK_LT
: op
= TOK_ULT
; break;
1744 case TOK_UGE
: op
= TOK_GE
; break;
1745 case TOK_UGT
: op
= TOK_GT
; break;
1759 x
|= vfpr(r2
) << 16;
1762 vtop
->r
= get_reg_ex(RC_FLOAT
, r
);
1765 o(x
| (vfpr(vtop
->r
) << 12));
1769 static uint32_t is_fconst()
1773 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) != VT_CONST
)
1775 if (vtop
->type
.t
== VT_FLOAT
)
1777 else if (vtop
->type
.t
== VT_DOUBLE
)
1781 if (!ieee_finite(f
))
1807 /* generate a floating point operation 'v = t1 op t2' instruction. The
1808 two operands are guaranted to have the same floating point type */
1809 void gen_opf(int op
)
1811 uint32_t x
, r
, r2
, c1
, c2
;
1812 // fputs("gen_opf\n",stderr);
1818 #if LDOUBLE_SIZE == 8
1819 if ((vtop
->type
.t
& VT_BTYPE
) != VT_FLOAT
)
1822 if ((vtop
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1824 else if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
)
1834 r
= fpr(gv(RC_FLOAT
));
1838 x
|= 0x200000; // suf
1841 r2
= fpr(gv(RC_FLOAT
));
1847 x
|= 0x200000; // suf
1850 r
= fpr(gv(RC_FLOAT
));
1852 } else if (c1
&& c1
<= 0xf) {
1853 x
|= 0x300000; // rsf
1855 r
= fpr(gv(RC_FLOAT
));
1858 x
|= 0x200000; // suf
1860 r
= fpr(gv(RC_FLOAT
));
1862 r2
= fpr(gv(RC_FLOAT
));
1866 if (!c2
|| c2
> 0xf) {
1871 r
= fpr(gv(RC_FLOAT
));
1873 if (c2
&& c2
<= 0xf)
1876 r2
= fpr(gv(RC_FLOAT
));
1877 x
|= 0x100000; // muf
1880 if (c2
&& c2
<= 0xf) {
1881 x
|= 0x400000; // dvf
1884 r
= fpr(gv(RC_FLOAT
));
1886 } else if (c1
&& c1
<= 0xf) {
1887 x
|= 0x500000; // rdf
1889 r
= fpr(gv(RC_FLOAT
));
1892 x
|= 0x400000; // dvf
1894 r
= fpr(gv(RC_FLOAT
));
1896 r2
= fpr(gv(RC_FLOAT
));
1900 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1901 x
|= 0xd0f110; // cmfe
1902 /* bug (intention?) in Linux FPU emulator
1903 doesn't set carry if equal */
1909 tcc_error("unsigned comparison on floats?");
1915 op
= TOK_ULE
; /* correct in unordered case only if AC bit in
1920 x
&= ~0x400000; // cmfe -> cmf
1942 r
= fpr(gv(RC_FLOAT
));
1949 r2
= fpr(gv(RC_FLOAT
));
1951 vtop
[-1].r
= VT_CMP
;
1954 tcc_error("unknown fp op %x!", op
);
1958 if (vtop
[-1].r
== VT_CMP
)
1964 vtop
[-1].r
= get_reg_ex(RC_FLOAT
, two2mask(vtop
[-1].r
, c1
));
1965 c1
= fpr(vtop
[-1].r
);
1968 o(x
| (r
<< 16) | (c1
<< 12) | r2
);
1972 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1973 and 'long long' cases. */
1974 ST_FUNC
void gen_cvt_itof1(int t
)
1978 bt
= vtop
->type
.t
& VT_BTYPE
;
1979 if (bt
== VT_INT
|| bt
== VT_SHORT
|| bt
== VT_BYTE
) {
1983 r
= intr(gv(RC_INT
));
1985 r2
= vfpr(vtop
->r
= get_reg(RC_FLOAT
));
1986 o(0xEE000A10 | (r
<< 12) | (r2
<< 16)); /* fmsr */
1988 if (!(vtop
->type
.t
& VT_UNSIGNED
))
1989 r2
|= 0x80; /* fuitoX -> fsituX */
1990 o(0xEEB80A40 | r2
| T2CPR(t
)); /* fYitoX*/
1992 r2
= fpr(vtop
->r
= get_reg(RC_FLOAT
));
1993 if ((t
& VT_BTYPE
) != VT_FLOAT
)
1994 dsize
= 0x80; /* flts -> fltd */
1995 o(0xEE000110 | dsize
| (r2
<< 16) | (r
<< 12)); /* flts */
1996 if ((vtop
->type
.t
& (VT_UNSIGNED
| VT_BTYPE
)) ==
1997 (VT_UNSIGNED
| VT_INT
)) {
1999 o(0xE3500000 | (r
<< 12)); /* cmp */
2000 r
= fpr(get_reg(RC_FLOAT
));
2001 if (last_itod_magic
) {
2002 off
= ind
+ 8 - last_itod_magic
;
2007 o(0xBD1F0100 | (r
<< 12) | off
); /* ldflts */
2009 o(0xEA000000); /* b */
2010 last_itod_magic
= ind
;
2011 o(0x4F800000); /* 4294967296.0f */
2013 o(0xBE000100 | dsize
| (r2
<< 16) | (r2
<< 12) | r
); /* adflt */
2017 } else if (bt
== VT_LLONG
) {
2019 CType
* func_type
= 0;
2020 if ((t
& VT_BTYPE
) == VT_FLOAT
) {
2021 func_type
= &func_float_type
;
2022 if (vtop
->type
.t
& VT_UNSIGNED
)
2023 func
= TOK___floatundisf
;
2025 func
= TOK___floatdisf
;
2026 #if LDOUBLE_SIZE != 8
2027 } else if ((t
& VT_BTYPE
) == VT_LDOUBLE
) {
2028 func_type
= &func_ldouble_type
;
2029 if (vtop
->type
.t
& VT_UNSIGNED
)
2030 func
= TOK___floatundixf
;
2032 func
= TOK___floatdixf
;
2033 } else if ((t
& VT_BTYPE
) == VT_DOUBLE
) {
2035 } else if ((t
& VT_BTYPE
) == VT_DOUBLE
||
2036 (t
& VT_BTYPE
) == VT_LDOUBLE
) {
2038 func_type
= &func_double_type
;
2039 if (vtop
->type
.t
& VT_UNSIGNED
)
2040 func
= TOK___floatundidf
;
2042 func
= TOK___floatdidf
;
2045 vpush_global_sym(func_type
, func
);
2053 tcc_error("unimplemented gen_cvt_itof %x!", vtop
->type
.t
);
2056 /* convert fp to int 't' type */
2057 void gen_cvt_ftoi(int t
)
2061 u
= t
& VT_UNSIGNED
;
2063 r2
= vtop
->type
.t
& VT_BTYPE
;
2066 r
= vfpr(gv(RC_FLOAT
));
2067 u
= u
? 0 : 0x10000;
2068 o(0xEEBC0AC0 | (r
<< 12) | r
| T2CPR(r2
) | u
); /* ftoXizY */
2069 r2
= intr(vtop
->r
= get_reg(RC_INT
));
2070 o(0xEE100A10 | (r
<< 16) | (r2
<< 12));
2075 func
= TOK___fixunssfsi
;
2076 #if LDOUBLE_SIZE != 8
2077 else if (r2
== VT_LDOUBLE
)
2078 func
= TOK___fixunsxfsi
;
2079 else if (r2
== VT_DOUBLE
)
2081 else if (r2
== VT_LDOUBLE
|| r2
== VT_DOUBLE
)
2083 func
= TOK___fixunsdfsi
;
2085 r
= fpr(gv(RC_FLOAT
));
2086 r2
= intr(vtop
->r
= get_reg(RC_INT
));
2087 o(0xEE100170 | (r2
<< 12) | r
);
2091 } else if (t
== VT_LLONG
) { // unsigned handled in gen_cvt_ftoi1
2093 func
= TOK___fixsfdi
;
2094 #if LDOUBLE_SIZE != 8
2095 else if (r2
== VT_LDOUBLE
)
2096 func
= TOK___fixxfdi
;
2097 else if (r2
== VT_DOUBLE
)
2099 else if (r2
== VT_LDOUBLE
|| r2
== VT_DOUBLE
)
2101 func
= TOK___fixdfdi
;
2104 vpush_global_sym(&func_old_type
, func
);
2109 vtop
->r2
= REG_LRET
;
2113 tcc_error("unimplemented gen_cvt_ftoi!");
2116 /* convert from one floating point type to another */
2117 void gen_cvt_ftof(int t
)
2120 if (((vtop
->type
.t
& VT_BTYPE
) == VT_FLOAT
) !=
2121 ((t
& VT_BTYPE
) == VT_FLOAT
)) {
2122 uint32_t r
= vfpr(gv(RC_FLOAT
));
2123 o(0xEEB70AC0 | (r
<< 12) | r
| T2CPR(vtop
->type
.t
));
2126 /* all we have to do on i386 and FPA ARM is to put the float in a register
2132 /* computed goto support */
2139 /* Save the stack pointer onto the stack and return the location of its address
2141 ST_FUNC
void gen_vla_sp_save(int addr
)
2143 tcc_error("variable length arrays unsupported for this target");
2146 /* Restore the SP from a location on the stack */
2147 ST_FUNC
void gen_vla_sp_restore(int addr
)
2149 tcc_error("variable length arrays unsupported for this target");
2152 /* Subtract from the stack pointer, and push the resulting value onto the stack
2154 ST_FUNC
void gen_vla_alloc(CType
* type
, int align
)
2156 tcc_error("variable length arrays unsupported for this target");
2159 /* end of ARM code generator */
2160 /*************************************************************/
2162 /*************************************************************/