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
27 #ifndef TCC_ARM_VFP /* Avoid useless warning */
32 /* 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
146 /******************************************************/
147 #else /* ! TARGET_DEFS_ONLY */
148 /******************************************************/
151 ST_DATA
const int reg_classes
[NB_REGS
] = {
152 /* r0 */ RC_INT
| RC_R0
,
153 /* r1 */ RC_INT
| RC_R1
,
154 /* r2 */ RC_INT
| RC_R2
,
155 /* r3 */ RC_INT
| RC_R3
,
156 /* r12 */ RC_INT
| RC_R12
,
157 /* f0 */ RC_FLOAT
| RC_F0
,
158 /* f1 */ RC_FLOAT
| RC_F1
,
159 /* f2 */ RC_FLOAT
| RC_F2
,
160 /* f3 */ RC_FLOAT
| RC_F3
,
162 /* d4/s8 */ RC_FLOAT
| RC_F4
,
163 /* d5/s10 */ RC_FLOAT
| RC_F5
,
164 /* d6/s12 */ RC_FLOAT
| RC_F6
,
165 /* d7/s14 */ RC_FLOAT
| RC_F7
,
169 static int func_sub_sp_offset
, last_itod_magic
;
172 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
173 static CType float_type
, double_type
, func_float_type
, func_double_type
;
174 ST_FUNC
void arm_init_types(void)
176 float_type
.t
= VT_FLOAT
;
177 double_type
.t
= VT_DOUBLE
;
178 func_float_type
.t
= VT_FUNC
;
179 func_float_type
.ref
= sym_push(SYM_FIELD
, &float_type
, FUNC_CDECL
, FUNC_OLD
);
180 func_double_type
.t
= VT_FUNC
;
181 func_double_type
.ref
= sym_push(SYM_FIELD
, &double_type
, FUNC_CDECL
, FUNC_OLD
);
184 #define func_float_type func_old_type
185 #define func_double_type func_old_type
186 #define func_ldouble_type func_old_type
187 ST_FUNC
void arm_init_types(void) {}
190 static int two2mask(int a
,int b
) {
191 return (reg_classes
[a
]|reg_classes
[b
])&~(RC_INT
|RC_FLOAT
);
194 static int regmask(int r
) {
195 return reg_classes
[r
]&~(RC_INT
|RC_FLOAT
);
198 /******************************************************/
202 /* this is a good place to start adding big-endian support*/
206 if (!cur_text_section
)
207 tcc_error("compiler error! This happens f.ex. if the compiler\n"
208 "can't evaluate constant expressions outside of a function.");
209 if (ind1
> cur_text_section
->data_allocated
)
210 section_realloc(cur_text_section
, ind1
);
211 cur_text_section
->data
[ind
++] = i
&255;
213 cur_text_section
->data
[ind
++] = i
&255;
215 cur_text_section
->data
[ind
++] = i
&255;
217 cur_text_section
->data
[ind
++] = i
;
220 static uint32_t stuff_const(uint32_t op
, uint32_t c
)
223 uint32_t nc
= 0, negop
= 0;
233 case 0x1A00000: //mov
234 case 0x1E00000: //mvn
241 return (op
&0xF010F000)|((op
>>16)&0xF)|0x1E00000;
245 return (op
&0xF010F000)|((op
>>16)&0xF)|0x1A00000;
246 case 0x1C00000: //bic
251 case 0x1800000: //orr
253 return (op
&0xFFF0FFFF)|0x1E00000;
259 if(c
<256) /* catch undefined <<32 */
262 m
=(0xff>>i
)|(0xff<<(32-i
));
264 return op
|(i
<<7)|(c
<<i
)|(c
>>(32-i
));
274 void stuff_const_harder(uint32_t op
, uint32_t v
) {
280 uint32_t a
[16], nv
, no
, o2
, n2
;
283 o2
=(op
&0xfff0ffff)|((op
&0xf000)<<4);;
285 a
[i
]=(a
[i
-1]>>2)|(a
[i
-1]<<30);
287 for(j
=i
<4?i
+12:15;j
>=i
+4;j
--)
288 if((v
&(a
[i
]|a
[j
]))==v
) {
289 o(stuff_const(op
,v
&a
[i
]));
290 o(stuff_const(o2
,v
&a
[j
]));
297 for(j
=i
<4?i
+12:15;j
>=i
+4;j
--)
298 if((nv
&(a
[i
]|a
[j
]))==nv
) {
299 o(stuff_const(no
,nv
&a
[i
]));
300 o(stuff_const(n2
,nv
&a
[j
]));
305 for(k
=i
<4?i
+12:15;k
>=j
+4;k
--)
306 if((v
&(a
[i
]|a
[j
]|a
[k
]))==v
) {
307 o(stuff_const(op
,v
&a
[i
]));
308 o(stuff_const(o2
,v
&a
[j
]));
309 o(stuff_const(o2
,v
&a
[k
]));
316 for(k
=i
<4?i
+12:15;k
>=j
+4;k
--)
317 if((nv
&(a
[i
]|a
[j
]|a
[k
]))==nv
) {
318 o(stuff_const(no
,nv
&a
[i
]));
319 o(stuff_const(n2
,nv
&a
[j
]));
320 o(stuff_const(n2
,nv
&a
[k
]));
323 o(stuff_const(op
,v
&a
[0]));
324 o(stuff_const(o2
,v
&a
[4]));
325 o(stuff_const(o2
,v
&a
[8]));
326 o(stuff_const(o2
,v
&a
[12]));
330 ST_FUNC
uint32_t encbranch(int pos
, int addr
, int fail
)
334 if(addr
>=0x1000000 || addr
<-0x1000000) {
336 tcc_error("FIXME: function bigger than 32MB");
339 return 0x0A000000|(addr
&0xffffff);
342 int decbranch(int pos
)
345 x
=*(uint32_t *)(cur_text_section
->data
+ pos
);
352 /* output a symbol and patch all calls to it */
353 void gsym_addr(int t
, int a
)
358 x
=(uint32_t *)(cur_text_section
->data
+ t
);
361 *x
=0xE1A00000; // nop
364 *x
|= encbranch(lt
,a
,1);
375 static uint32_t vfpr(int r
)
377 if(r
<TREG_F0
|| r
>TREG_F7
)
378 tcc_error("compiler error! register %i is no vfp register",r
);
382 static uint32_t fpr(int r
)
384 if(r
<TREG_F0
|| r
>TREG_F3
)
385 tcc_error("compiler error! register %i is no fpa register",r
);
390 static uint32_t intr(int r
)
394 if((r
<0 || r
>4) && r
!=14)
395 tcc_error("compiler error! register %i is no int register",r
);
399 static void calcaddr(uint32_t *base
, int *off
, int *sgn
, int maxoff
, unsigned shift
)
401 if(*off
>maxoff
|| *off
&((1<<shift
)-1)) {
408 y
=stuff_const(x
,*off
&~maxoff
);
414 y
=stuff_const(x
,(*off
+maxoff
)&~maxoff
);
418 *off
=((*off
+maxoff
)&~maxoff
)-*off
;
421 stuff_const_harder(x
,*off
&~maxoff
);
426 static uint32_t mapcc(int cc
)
431 return 0x30000000; /* CC/LO */
433 return 0x20000000; /* CS/HS */
435 return 0x00000000; /* EQ */
437 return 0x10000000; /* NE */
439 return 0x90000000; /* LS */
441 return 0x80000000; /* HI */
443 return 0x40000000; /* MI */
445 return 0x50000000; /* PL */
447 return 0xB0000000; /* LT */
449 return 0xA0000000; /* GE */
451 return 0xD0000000; /* LE */
453 return 0xC0000000; /* GT */
455 tcc_error("unexpected condition code");
456 return 0xE0000000; /* AL */
459 static int negcc(int cc
)
488 tcc_error("unexpected condition code");
492 /* load 'r' from value 'sv' */
493 void load(int r
, SValue
*sv
)
495 int v
, ft
, fc
, fr
, sign
;
512 uint32_t base
= 0xB; // fp
515 v1
.r
= VT_LOCAL
| VT_LVAL
;
517 load(base
=14 /* lr */, &v1
);
520 } else if(v
== VT_CONST
) {
528 } else if(v
< VT_CONST
) {
535 calcaddr(&base
,&fc
,&sign
,1020,2);
537 op
=0xED100A00; /* flds */
540 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
541 op
|=0x100; /* flds -> fldd */
542 o(op
|(vfpr(r
)<<12)|(fc
>>2)|(base
<<16));
547 #if LDOUBLE_SIZE == 8
548 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
551 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
553 else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
)
556 o(op
|(fpr(r
)<<12)|(fc
>>2)|(base
<<16));
558 } else if((ft
& (VT_BTYPE
|VT_UNSIGNED
)) == VT_BYTE
559 || (ft
& VT_BTYPE
) == VT_SHORT
) {
560 calcaddr(&base
,&fc
,&sign
,255,0);
562 if ((ft
& VT_BTYPE
) == VT_SHORT
)
564 if ((ft
& VT_UNSIGNED
) == 0)
568 o(op
|(intr(r
)<<12)|(base
<<16)|((fc
&0xf0)<<4)|(fc
&0xf));
570 calcaddr(&base
,&fc
,&sign
,4095,0);
574 if ((ft
& VT_BTYPE
) == VT_BYTE
|| (ft
& VT_BTYPE
) == VT_BOOL
)
576 o(op
|(intr(r
)<<12)|fc
|(base
<<16));
582 op
=stuff_const(0xE3A00000|(intr(r
)<<12),sv
->c
.ul
);
583 if (fr
& VT_SYM
|| !op
) {
584 o(0xE59F0000|(intr(r
)<<12));
587 greloc(cur_text_section
, sv
->sym
, ind
, R_ARM_ABS32
);
592 } else if (v
== VT_LOCAL
) {
593 op
=stuff_const(0xE28B0000|(intr(r
)<<12),sv
->c
.ul
);
594 if (fr
& VT_SYM
|| !op
) {
595 o(0xE59F0000|(intr(r
)<<12));
597 if(fr
& VT_SYM
) // needed ?
598 greloc(cur_text_section
, sv
->sym
, ind
, R_ARM_ABS32
);
600 o(0xE08B0000|(intr(r
)<<12)|intr(r
));
604 } else if(v
== VT_CMP
) {
605 o(mapcc(sv
->c
.ul
)|0x3A00001|(intr(r
)<<12));
606 o(mapcc(negcc(sv
->c
.ul
))|0x3A00000|(intr(r
)<<12));
608 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
611 o(0xE3A00000|(intr(r
)<<12)|t
);
614 o(0xE3A00000|(intr(r
)<<12)|(t
^1));
616 } else if (v
< VT_CONST
) {
619 o(0xEEB00A40|(vfpr(r
)<<12)|vfpr(v
)|T2CPR(ft
)); /* fcpyX */
621 o(0xEE008180|(fpr(r
)<<12)|fpr(v
));
624 o(0xE1A00000|(intr(r
)<<12)|intr(v
));
628 tcc_error("load unimplemented!");
631 /* store register 'r' in lvalue 'v' */
632 void store(int r
, SValue
*sv
)
635 int v
, ft
, fc
, fr
, sign
;
650 if (fr
& VT_LVAL
|| fr
== VT_LOCAL
) {
656 } else if(v
== VT_CONST
) {
667 calcaddr(&base
,&fc
,&sign
,1020,2);
669 op
=0xED000A00; /* fsts */
672 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
673 op
|=0x100; /* fsts -> fstd */
674 o(op
|(vfpr(r
)<<12)|(fc
>>2)|(base
<<16));
679 #if LDOUBLE_SIZE == 8
680 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
683 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
685 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
)
688 o(op
|(fpr(r
)<<12)|(fc
>>2)|(base
<<16));
691 } else if((ft
& VT_BTYPE
) == VT_SHORT
) {
692 calcaddr(&base
,&fc
,&sign
,255,0);
696 o(op
|(intr(r
)<<12)|(base
<<16)|((fc
&0xf0)<<4)|(fc
&0xf));
698 calcaddr(&base
,&fc
,&sign
,4095,0);
702 if ((ft
& VT_BTYPE
) == VT_BYTE
|| (ft
& VT_BTYPE
) == VT_BOOL
)
704 o(op
|(intr(r
)<<12)|fc
|(base
<<16));
709 tcc_error("store unimplemented");
712 static void gadd_sp(int val
)
714 stuff_const_harder(0xE28DD000,val
);
717 /* 'is_jmp' is '1' if it is a jump */
718 static void gcall_or_jmp(int is_jmp
)
721 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
724 x
=encbranch(ind
,ind
+vtop
->c
.ul
,0);
726 if (vtop
->r
& VT_SYM
) {
727 /* relocation case */
728 greloc(cur_text_section
, vtop
->sym
, ind
, R_ARM_PC24
);
730 put_elf_reloc(symtab_section
, cur_text_section
, ind
, R_ARM_PC24
, 0);
731 o(x
|(is_jmp
?0xE0000000:0xE1000000));
734 o(0xE28FE004); // add lr,pc,#4
735 o(0xE51FF004); // ldr pc,[pc,#-4]
736 if (vtop
->r
& VT_SYM
)
737 greloc(cur_text_section
, vtop
->sym
, ind
, R_ARM_ABS32
);
741 /* otherwise, indirect call */
744 o(0xE1A0E00F); // mov lr,pc
745 o(0xE1A0F000|intr(r
)); // mov pc,r
749 /* Return 1 if this function returns via an sret pointer, 0 otherwise */
750 ST_FUNC
int gfunc_sret(CType
*vt
, CType
*ret
, int *ret_align
) {
753 size
= type_size(vt
, &align
);
767 #ifdef TCC_ARM_HARDFLOAT
768 /* Return whether a structure is an homogeneous float aggregate or not.
769 The answer is true if all the elements of the structure are of the same
770 primitive float type and there is less than 4 elements.
772 type: the type corresponding to the structure to be tested */
773 static int is_hgen_float_aggr(CType
*type
)
775 if ((type
->t
& VT_BTYPE
) == VT_STRUCT
) {
777 int btype
, nb_fields
= 0;
780 btype
= ref
->type
.t
& VT_BTYPE
;
781 if (btype
== VT_FLOAT
|| btype
== VT_DOUBLE
) {
782 for(; ref
&& btype
== (ref
->type
.t
& VT_BTYPE
); ref
= ref
->next
, nb_fields
++);
783 return !ref
&& nb_fields
<= 4;
790 signed char avail
[3]; /* 3 holes max with only float and double alignments */
791 int first_hole
; /* first available hole */
792 int last_hole
; /* last available hole (none if equal to first_hole) */
793 int first_free_reg
; /* next free register in the sequence, hole excluded */
796 #define AVAIL_REGS_INITIALIZER (struct avail_regs) { { 0, 0, 0}, 0, 0, 0 }
798 /* Find suitable registers for a VFP Co-Processor Register Candidate (VFP CPRC
799 param) according to the rules described in the procedure call standard for
800 the ARM architecture (AAPCS). If found, the registers are assigned to this
801 VFP CPRC parameter. Registers are allocated in sequence unless a hole exists
802 and the parameter is a single float.
804 avregs: opaque structure to keep track of available VFP co-processor regs
805 align: alignment contraints for the param, as returned by type_size()
806 size: size of the parameter, as returned by type_size() */
807 int assign_vfpreg(struct avail_regs
*avregs
, int align
, int size
)
811 if (avregs
->first_free_reg
== -1)
813 if (align
>> 3) { /* double alignment */
814 first_reg
= avregs
->first_free_reg
;
815 /* alignment contraint not respected so use next reg and record hole */
817 avregs
->avail
[avregs
->last_hole
++] = first_reg
++;
818 } else { /* no special alignment (float or array of float) */
819 /* if single float and a hole is available, assign the param to it */
820 if (size
== 4 && avregs
->first_hole
!= avregs
->last_hole
)
821 return avregs
->avail
[avregs
->first_hole
++];
823 first_reg
= avregs
->first_free_reg
;
825 if (first_reg
+ size
/ 4 <= 16) {
826 avregs
->first_free_reg
= first_reg
+ size
/ 4;
829 avregs
->first_free_reg
= -1;
834 /* Parameters are classified according to how they are copied to their final
835 destination for the function call. Because the copying is performed class
836 after class according to the order in the union below, it is important that
837 some constraints about the order of the members of this union are respected:
838 - CORE_STRUCT_CLASS must come after STACK_CLASS;
839 - CORE_CLASS must come after STACK_CLASS, CORE_STRUCT_CLASS and
841 - VFP_STRUCT_CLASS must come after VFP_CLASS.
842 See the comment for the main loop in copy_params() for the reason. */
853 int start
; /* first reg or addr used depending on the class */
854 int end
; /* last reg used or next free addr depending on the class */
855 SValue
*sval
; /* pointer to SValue on the value stack */
856 struct param_plan
*prev
; /* previous element in this class */
860 struct param_plan
*pplans
; /* array of all the param plans */
861 struct param_plan
*clsplans
[NB_CLASSES
]; /* per class lists of param plans */
864 #define add_param_plan(plan,pplan,class) \
866 pplan.prev = plan->clsplans[class]; \
867 plan->pplans[plan ## _nb] = pplan; \
868 plan->clsplans[class] = &plan->pplans[plan ## _nb++]; \
871 /* Assign parameters to registers and stack with alignment according to the
872 rules in the procedure call standard for the ARM architecture (AAPCS).
873 The overall assignment is recorded in an array of per parameter structures
874 called parameter plans. The parameter plans are also further organized in a
875 number of linked lists, one per class of parameter (see the comment for the
876 definition of union reg_class).
878 nb_args: number of parameters of the function for which a call is generated
879 variadic: whether the function is a variadic function or not
880 plan: the structure where the overall assignment is recorded
881 todo: a bitmap that record which core registers hold a parameter
883 Returns the amount of stack space needed for parameter passing
885 Note: this function allocated an array in plan->pplans with tcc_malloc. It
886 is the responsability of the caller to free this array once used (ie not
887 before copy_params). */
888 static int assign_regs(int nb_args
, int variadic
, struct plan
*plan
, int *todo
)
891 int ncrn
/* next core register number */, nsaa
/* next stacked argument address*/;
893 struct param_plan pplan
;
894 #ifdef TCC_ARM_HARDFLOAT
895 struct avail_regs avregs
= AVAIL_REGS_INITIALIZER
;
900 plan
->pplans
= tcc_malloc(nb_args
* sizeof(*plan
->pplans
));
901 memset(plan
->clsplans
, 0, sizeof(plan
->clsplans
));
902 for(i
= nb_args
; i
-- ;) {
903 int j
, start_vfpreg
= 0;
904 size
= type_size(&vtop
[-i
].type
, &align
);
905 switch(vtop
[-i
].type
.t
& VT_BTYPE
) {
910 #ifdef TCC_ARM_HARDFLOAT
912 int is_hfa
= 0; /* Homogeneous float aggregate */
914 if (is_float(vtop
[-i
].type
.t
)
915 || (is_hfa
= is_hgen_float_aggr(&vtop
[-i
].type
))) {
918 start_vfpreg
= assign_vfpreg(&avregs
, align
, size
);
919 end_vfpreg
= start_vfpreg
+ ((size
- 1) >> 2);
920 if (start_vfpreg
>= 0) {
921 pplan
= (struct param_plan
) {start_vfpreg
, end_vfpreg
, &vtop
[-i
]};
923 add_param_plan(plan
, pplan
, VFP_STRUCT_CLASS
);
925 add_param_plan(plan
, pplan
, VFP_CLASS
);
932 ncrn
= (ncrn
+ (align
-1)/4) & -(align
/4);
933 size
= (size
+ 3) & -4;
934 if (ncrn
+ size
/4 <= 4 || (ncrn
< 4 && start_vfpreg
!= -1)) {
935 /* The parameter is allocated both in core register and on stack. As
936 * such, it can be of either class: it would either be the last of
937 * CORE_STRUCT_CLASS or the first of STACK_CLASS. */
938 for (j
= ncrn
; j
< 4 && j
< ncrn
+ size
/ 4; j
++)
940 pplan
= (struct param_plan
) {ncrn
, j
, &vtop
[-i
]};
941 add_param_plan(plan
, pplan
, CORE_STRUCT_CLASS
);
944 nsaa
= (ncrn
- 4) * 4;
952 int is_long
= (vtop
[-i
].type
.t
& VT_BTYPE
) == VT_LLONG
;
955 ncrn
= (ncrn
+ 1) & -2;
959 pplan
= (struct param_plan
) {ncrn
, ncrn
, &vtop
[-i
]};
963 add_param_plan(plan
, pplan
, CORE_CLASS
);
967 nsaa
= (nsaa
+ (align
- 1)) & ~(align
- 1);
968 pplan
= (struct param_plan
) {nsaa
, nsaa
+ size
, &vtop
[-i
]};
969 add_param_plan(plan
, pplan
, STACK_CLASS
);
970 nsaa
+= size
; /* size already rounded up before */
975 #undef add_param_plan
977 /* Copy parameters to their final destination (core reg, VFP reg or stack) for
980 nb_args: number of parameters the function take
981 plan: the overall assignment plan for parameters
982 todo: a bitmap indicating what core reg will hold a parameter
984 Returns the number of SValue added by this function on the value stack */
985 static int copy_params(int nb_args
, struct plan
*plan
, int todo
)
987 int size
, align
, r
, i
, nb_extra_sval
= 0;
988 struct param_plan
*pplan
;
990 /* Several constraints require parameters to be copied in a specific order:
991 - structures are copied to the stack before being loaded in a reg;
992 - floats loaded to an odd numbered VFP reg are first copied to the
993 preceding even numbered VFP reg and then moved to the next VFP reg.
995 It is thus important that:
996 - structures assigned to core regs must be copied after parameters
997 assigned to the stack but before structures assigned to VFP regs because
998 a structure can lie partly in core registers and partly on the stack;
999 - parameters assigned to the stack and all structures be copied before
1000 parameters assigned to a core reg since copying a parameter to the stack
1001 require using a core reg;
1002 - parameters assigned to VFP regs be copied before structures assigned to
1003 VFP regs as the copy might use an even numbered VFP reg that already
1004 holds part of a structure. */
1005 for(i
= 0; i
< NB_CLASSES
; i
++) {
1006 for(pplan
= plan
->clsplans
[i
]; pplan
; pplan
= pplan
->prev
) {
1007 vpushv(pplan
->sval
);
1008 pplan
->sval
->r
= pplan
->sval
->r2
= VT_CONST
; /* disable entry */
1011 case CORE_STRUCT_CLASS
:
1012 case VFP_STRUCT_CLASS
:
1013 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_STRUCT
) {
1014 size
= type_size(&pplan
->sval
->type
, &align
);
1015 /* align to stack align size */
1016 size
= (size
+ 3) & ~3;
1017 if (i
== STACK_CLASS
&& pplan
->prev
)
1018 size
+= pplan
->start
- pplan
->prev
->end
; /* Add padding if any */
1019 /* allocate the necessary size on stack */
1021 /* generate structure store */
1022 r
= get_reg(RC_INT
);
1023 o(0xE1A0000D|(intr(r
)<<12)); /* mov r, sp */
1024 vset(&vtop
->type
, r
| VT_LVAL
, 0);
1026 vstore(); /* memcpy to current sp */
1027 /* Homogeneous float aggregate are loaded to VFP registers
1028 immediately since there is no way of loading data in multiple
1029 non consecutive VFP registers as what is done for other
1030 structures (see the use of todo). */
1031 if (i
== VFP_STRUCT_CLASS
) {
1032 int first
= pplan
->start
, nb
= pplan
->end
- first
+ 1;
1033 /* vpop.32 {pplan->start, ..., pplan->end} */
1034 o(0xECBD0A00|(first
&1)<<22|(first
>>1)<<12|nb
);
1035 /* No need to write the register used to a SValue since VFP regs
1036 cannot be used for gcall_or_jmp */
1039 if (is_float(pplan
->sval
->type
.t
)) {
1041 r
= vfpr(gv(RC_FLOAT
)) << 12;
1042 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
1046 r
|= 0x101; /* vpush.32 -> vpush.64 */
1048 o(0xED2D0A01 + r
); /* vpush */
1050 r
= fpr(gv(RC_FLOAT
)) << 12;
1051 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
1053 else if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1056 size
= LDOUBLE_SIZE
;
1063 o(0xED2D0100|r
|(size
>>2)); /* some kind of vpush for FPA */
1066 /* simple type (currently always same size) */
1067 /* XXX: implicit cast ? */
1069 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1073 o(0xE52D0004|(intr(r
)<<12)); /* push r */
1077 o(0xE52D0004|(intr(r
)<<12)); /* push r */
1079 if (i
== STACK_CLASS
&& pplan
->prev
)
1080 gadd_sp(pplan
->prev
->end
- pplan
->start
); /* Add padding if any */
1085 gv(regmask(TREG_F0
+ (pplan
->start
>> 1)));
1086 if (pplan
->start
& 1) { /* Must be in upper part of double register */
1087 o(0xEEF00A40|((pplan
->start
>>1)<<12)|(pplan
->start
>>1)); /* vmov.f32 s(n+1), sn */
1088 vtop
->r
= VT_CONST
; /* avoid being saved on stack by gv for next float */
1093 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1095 gv(regmask(pplan
->end
));
1096 pplan
->sval
->r2
= vtop
->r
;
1099 gv(regmask(pplan
->start
));
1100 /* Mark register as used so that gcall_or_jmp use another one
1101 (regs >=4 are free as never used to pass parameters) */
1102 pplan
->sval
->r
= vtop
->r
;
1109 /* Manually free remaining registers since next parameters are loaded
1110 * manually, without the help of gv(int). */
1114 o(0xE8BD0000|todo
); /* pop {todo} */
1115 for(pplan
= plan
->clsplans
[CORE_STRUCT_CLASS
]; pplan
; pplan
= pplan
->prev
) {
1117 pplan
->sval
->r
= pplan
->start
;
1118 /* TODO: why adding fake param */
1119 for (r
= pplan
->start
+ 1; r
<= pplan
->end
; r
++) {
1120 if (todo
& (1 << r
)) {
1128 return nb_extra_sval
;
1131 /* Generate function call. The function address is pushed first, then
1132 all the parameters in call order. This functions pops all the
1133 parameters and the function address. */
1134 void gfunc_call(int nb_args
)
1136 int align
, r
, args_size
;
1141 variadic
= (vtop
[-nb_args
].type
.ref
->c
== FUNC_ELLIPSIS
);
1142 /* cannot let cpu flags if other instruction are generated. Also avoid leaving
1143 VT_JMP anywhere except on the top of the stack because it would complicate
1144 the code generator. */
1145 r
= vtop
->r
& VT_VALMASK
;
1146 if (r
== VT_CMP
|| (r
& ~1) == VT_JMP
)
1149 /* return type is a struct so caller of gfunc_call (unary(void) in tccgen.c)
1150 assumed it had to be passed by a pointer. Since it's less than 4 bytes, we
1151 can actually pass it directly in a register. */
1152 if((vtop
[-nb_args
].type
.ref
->type
.t
& VT_BTYPE
) == VT_STRUCT
1153 && type_size(&vtop
[-nb_args
].type
.ref
->type
, &align
) <= 4) {
1156 vtop
[-nb_args
]=vtop
[-nb_args
+1];
1157 vtop
[-nb_args
+1]=tmp
;
1162 args_size
= assign_regs(nb_args
, variadic
, &plan
, &todo
);
1165 if (args_size
& 7) { /* Stack must be 8 byte aligned at fct call for EABI */
1166 args_size
= (args_size
+ 7) & ~7;
1167 o(0xE24DD004); /* sub sp, sp, #4 */
1171 nb_args
+= copy_params(nb_args
, &plan
, todo
);
1172 tcc_free(plan
.pplans
);
1174 /* Move fct SValue on top as required by gcall_or_jmp */
1178 gadd_sp(args_size
); /* pop all parameters passed on the stack */
1180 if((vtop
->type
.ref
->type
.t
& VT_BTYPE
) == VT_STRUCT
1181 && type_size(&vtop
->type
.ref
->type
, &align
) <= 4) {
1182 store(REG_IRET
,vtop
-nb_args
-1);
1186 #ifdef TCC_ARM_HARDFLOAT
1187 else if(variadic
&& is_float(vtop
->type
.ref
->type
.t
)) {
1189 else if(is_float(vtop
->type
.ref
->type
.t
)) {
1191 if((vtop
->type
.ref
->type
.t
& VT_BTYPE
) == VT_FLOAT
) {
1192 o(0xEE000A10); /*vmov s0, r0 */
1194 o(0xEE000B10); /* vmov.32 d0[0], r0 */
1195 o(0xEE201B10); /* vmov.32 d0[1], r1 */
1200 vtop
-= nb_args
+ 1; /* Pop all params and fct address from value stack */
1201 leaffunc
= 0; /* we are calling a function, so we aren't in a leaf function */
1204 /* generate function prolog of type 't' */
1205 void gfunc_prolog(CType
*func_type
)
1208 int n
,nf
,size
,align
, variadic
, struct_ret
= 0;
1209 #ifdef TCC_ARM_HARDFLOAT
1210 struct avail_regs avregs
= AVAIL_REGS_INITIALIZER
;
1213 sym
= func_type
->ref
;
1214 func_vt
= sym
->type
;
1217 variadic
= (func_type
->ref
->c
== FUNC_ELLIPSIS
);
1218 if((func_vt
.t
& VT_BTYPE
) == VT_STRUCT
1219 && type_size(&func_vt
,&align
) > 4)
1223 func_vc
= 12; /* Offset from fp of the place to store the result */
1225 for(sym2
=sym
->next
;sym2
&& (n
<4 || nf
<16);sym2
=sym2
->next
) {
1226 size
= type_size(&sym2
->type
, &align
);
1227 #ifdef TCC_ARM_HARDFLOAT
1228 if (!variadic
&& (is_float(sym2
->type
.t
)
1229 || is_hgen_float_aggr(&sym2
->type
))) {
1230 int tmpnf
= assign_vfpreg(&avregs
, align
, size
) + 1;
1231 nf
= (tmpnf
> nf
) ? tmpnf
: nf
;
1235 n
+= (size
+ 3) / 4;
1237 o(0xE1A0C00D); /* mov ip,sp */
1246 o(0xE92D0000|((1<<n
)-1)); /* save r0-r4 on stack if needed */
1251 nf
=(nf
+1)&-2; /* nf => HARDFLOAT => EABI */
1252 o(0xED2D0A00|nf
); /* save s0-s15 on stack if needed */
1254 o(0xE92D5800); /* save fp, ip, lr */
1255 o(0xE1A0B00D); /* mov fp, sp */
1256 func_sub_sp_offset
= ind
;
1257 o(0xE1A00000); /* nop, leave space for stack adjustment in epilogue */
1259 int addr
, pn
= struct_ret
, sn
= 0; /* pn=core, sn=stack */
1261 #ifdef TCC_ARM_HARDFLOAT
1263 avregs
= AVAIL_REGS_INITIALIZER
;
1265 while ((sym
= sym
->next
)) {
1268 size
= type_size(type
, &align
);
1269 size
= (size
+ 3) >> 2;
1270 align
= (align
+ 3) & ~3;
1271 #ifdef TCC_ARM_HARDFLOAT
1272 if (!variadic
&& (is_float(sym
->type
.t
)
1273 || is_hgen_float_aggr(&sym
->type
))) {
1274 int fpn
= assign_vfpreg(&avregs
, align
, size
<< 2);
1283 pn
= (pn
+ (align
-1)/4) & -(align
/4);
1285 addr
= (nf
+ pn
) * 4;
1290 #ifdef TCC_ARM_HARDFLOAT
1294 sn
= (sn
+ (align
-1)/4) & -(align
/4);
1296 addr
= (n
+ nf
+ sn
) * 4;
1299 sym_push(sym
->v
& ~SYM_FIELD
, type
, VT_LOCAL
| lvalue_type(type
->t
), addr
+12);
1307 /* generate function epilog */
1308 void gfunc_epilog(void)
1313 /* Useless but harmless copy of the float result into main register(s) in case
1314 of variadic function in the hardfloat variant */
1315 if(is_float(func_vt
.t
)) {
1316 if((func_vt
.t
& VT_BTYPE
) == VT_FLOAT
)
1317 o(0xEE100A10); /* fmrs r0, s0 */
1319 o(0xEE100B10); /* fmrdl r0, d0 */
1320 o(0xEE301B10); /* fmrdh r1, d0 */
1324 o(0xE89BA800); /* restore fp, sp, pc */
1325 diff
= (-loc
+ 3) & -4;
1328 diff
= ((diff
+ 11) & -8) - 4;
1331 x
=stuff_const(0xE24BD000, diff
); /* sub sp,fp,# */
1333 *(uint32_t *)(cur_text_section
->data
+ func_sub_sp_offset
) = x
;
1337 o(0xE59FC004); /* ldr ip,[pc+4] */
1338 o(0xE04BD00C); /* sub sp,fp,ip */
1339 o(0xE1A0F00E); /* mov pc,lr */
1341 *(uint32_t *)(cur_text_section
->data
+ func_sub_sp_offset
) = 0xE1000000|encbranch(func_sub_sp_offset
,addr
,1);
1346 /* generate a jump to a label */
1351 o(0xE0000000|encbranch(r
,t
,1));
1355 /* generate a jump to a fixed address */
1356 void gjmp_addr(int a
)
1361 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1362 int gtst(int inv
, int t
)
1366 v
= vtop
->r
& VT_VALMASK
;
1369 op
=mapcc(inv
?negcc(vtop
->c
.i
):vtop
->c
.i
);
1370 op
|=encbranch(r
,t
,1);
1373 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
1374 if ((v
& 1) == inv
) {
1383 p
= decbranch(lp
=p
);
1385 x
= (uint32_t *)(cur_text_section
->data
+ lp
);
1387 *x
|= encbranch(lp
,t
,1);
1396 if (is_float(vtop
->type
.t
)) {
1399 o(0xEEB50A40|(vfpr(r
)<<12)|T2CPR(vtop
->type
.t
)); /* fcmpzX */
1400 o(0xEEF1FA10); /* fmstat */
1402 o(0xEE90F118|(fpr(r
)<<16));
1406 return gtst(inv
, t
);
1407 } else if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1408 /* constant jmp optimization */
1409 if ((vtop
->c
.i
!= 0) != inv
)
1413 o(0xE3300000|(intr(v
)<<16));
1416 return gtst(inv
, t
);
1423 /* generate an integer binary operation */
1424 void gen_opi(int op
)
1427 uint32_t opc
= 0, r
, fr
;
1428 unsigned short retreg
= REG_IRET
;
1436 case TOK_ADDC1
: /* add with carry generation */
1444 case TOK_SUBC1
: /* sub with carry generation */
1448 case TOK_ADDC2
: /* add with carry use */
1452 case TOK_SUBC2
: /* sub with carry use */
1469 gv2(RC_INT
, RC_INT
);
1473 o(0xE0000090|(intr(r
)<<16)|(intr(r
)<<8)|intr(fr
));
1498 func
=TOK___aeabi_idivmod
;
1507 func
=TOK___aeabi_uidivmod
;
1515 gv2(RC_INT
, RC_INT
);
1516 r
=intr(vtop
[-1].r2
=get_reg(RC_INT
));
1518 vtop
[-1].r
=get_reg_ex(RC_INT
,regmask(c
));
1520 o(0xE0800090|(r
<<16)|(intr(vtop
->r
)<<12)|(intr(c
)<<8)|intr(vtop
[1].r
));
1529 if((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1530 if(opc
== 4 || opc
== 5 || opc
== 0xc) {
1532 opc
|=2; // sub -> rsb
1535 if ((vtop
->r
& VT_VALMASK
) == VT_CMP
||
1536 (vtop
->r
& (VT_VALMASK
& ~1)) == VT_JMP
)
1541 opc
=0xE0000000|(opc
<<20)|(c
<<16);
1542 if((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1544 x
=stuff_const(opc
|0x2000000,vtop
->c
.i
);
1546 r
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,regmask(vtop
[-1].r
)));
1551 fr
=intr(gv(RC_INT
));
1552 r
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,two2mask(vtop
->r
,vtop
[-1].r
)));
1556 if (op
>= TOK_ULT
&& op
<= TOK_GT
) {
1562 opc
=0xE1A00000|(opc
<<5);
1563 if ((vtop
->r
& VT_VALMASK
) == VT_CMP
||
1564 (vtop
->r
& (VT_VALMASK
& ~1)) == VT_JMP
)
1570 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1571 fr
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,regmask(vtop
[-1].r
)));
1572 c
= vtop
->c
.i
& 0x1f;
1573 o(opc
|(c
<<7)|(fr
<<12));
1575 fr
=intr(gv(RC_INT
));
1576 c
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,two2mask(vtop
->r
,vtop
[-1].r
)));
1577 o(opc
|(c
<<12)|(fr
<<8)|0x10);
1582 vpush_global_sym(&func_old_type
, func
);
1589 tcc_error("gen_opi %i unimplemented!",op
);
1594 static int is_zero(int i
)
1596 if((vtop
[i
].r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) != VT_CONST
)
1598 if (vtop
[i
].type
.t
== VT_FLOAT
)
1599 return (vtop
[i
].c
.f
== 0.f
);
1600 else if (vtop
[i
].type
.t
== VT_DOUBLE
)
1601 return (vtop
[i
].c
.d
== 0.0);
1602 return (vtop
[i
].c
.ld
== 0.l
);
1605 /* generate a floating point operation 'v = t1 op t2' instruction. The
1606 * two operands are guaranted to have the same floating point type */
1607 void gen_opf(int op
)
1611 x
=0xEE000A00|T2CPR(vtop
->type
.t
);
1629 x
|=0x810000; /* fsubX -> fnegX */
1642 if(op
< TOK_ULT
|| op
> TOK_GT
) {
1643 tcc_error("unknown fp op %x!",op
);
1649 case TOK_LT
: op
=TOK_GT
; break;
1650 case TOK_GE
: op
=TOK_ULE
; break;
1651 case TOK_LE
: op
=TOK_GE
; break;
1652 case TOK_GT
: op
=TOK_ULT
; break;
1655 x
|=0xB40040; /* fcmpX */
1656 if(op
!=TOK_EQ
&& op
!=TOK_NE
)
1657 x
|=0x80; /* fcmpX -> fcmpeX */
1660 o(x
|0x10000|(vfpr(gv(RC_FLOAT
))<<12)); /* fcmp(e)X -> fcmp(e)zX */
1662 x
|=vfpr(gv(RC_FLOAT
));
1664 o(x
|(vfpr(gv(RC_FLOAT
))<<12));
1667 o(0xEEF1FA10); /* fmstat */
1670 case TOK_LE
: op
=TOK_ULE
; break;
1671 case TOK_LT
: op
=TOK_ULT
; break;
1672 case TOK_UGE
: op
=TOK_GE
; break;
1673 case TOK_UGT
: op
=TOK_GT
; break;
1690 vtop
->r
=get_reg_ex(RC_FLOAT
,r
);
1693 o(x
|(vfpr(vtop
->r
)<<12));
1697 static uint32_t is_fconst()
1701 if((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) != VT_CONST
)
1703 if (vtop
->type
.t
== VT_FLOAT
)
1705 else if (vtop
->type
.t
== VT_DOUBLE
)
1735 /* generate a floating point operation 'v = t1 op t2' instruction. The
1736 two operands are guaranted to have the same floating point type */
1737 void gen_opf(int op
)
1739 uint32_t x
, r
, r2
, c1
, c2
;
1740 //fputs("gen_opf\n",stderr);
1746 #if LDOUBLE_SIZE == 8
1747 if ((vtop
->type
.t
& VT_BTYPE
) != VT_FLOAT
)
1750 if ((vtop
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1752 else if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
)
1763 r
=fpr(gv(RC_FLOAT
));
1770 r2
=fpr(gv(RC_FLOAT
));
1779 r
=fpr(gv(RC_FLOAT
));
1781 } else if(c1
&& c1
<=0xf) {
1784 r
=fpr(gv(RC_FLOAT
));
1789 r
=fpr(gv(RC_FLOAT
));
1791 r2
=fpr(gv(RC_FLOAT
));
1800 r
=fpr(gv(RC_FLOAT
));
1805 r2
=fpr(gv(RC_FLOAT
));
1813 r
=fpr(gv(RC_FLOAT
));
1815 } else if(c1
&& c1
<=0xf) {
1818 r
=fpr(gv(RC_FLOAT
));
1823 r
=fpr(gv(RC_FLOAT
));
1825 r2
=fpr(gv(RC_FLOAT
));
1829 if(op
>= TOK_ULT
&& op
<= TOK_GT
) {
1830 x
|=0xd0f110; // cmfe
1831 /* bug (intention?) in Linux FPU emulator
1832 doesn't set carry if equal */
1838 tcc_error("unsigned comparision on floats?");
1844 op
=TOK_ULE
; /* correct in unordered case only if AC bit in FPSR set */
1848 x
&=~0x400000; // cmfe -> cmf
1870 r
=fpr(gv(RC_FLOAT
));
1877 r2
=fpr(gv(RC_FLOAT
));
1879 vtop
[-1].r
= VT_CMP
;
1882 tcc_error("unknown fp op %x!",op
);
1886 if(vtop
[-1].r
== VT_CMP
)
1892 vtop
[-1].r
=get_reg_ex(RC_FLOAT
,two2mask(vtop
[-1].r
,c1
));
1896 o(x
|(r
<<16)|(c1
<<12)|r2
);
1900 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1901 and 'long long' cases. */
1902 ST_FUNC
void gen_cvt_itof1(int t
)
1906 bt
=vtop
->type
.t
& VT_BTYPE
;
1907 if(bt
== VT_INT
|| bt
== VT_SHORT
|| bt
== VT_BYTE
) {
1913 r2
=vfpr(vtop
->r
=get_reg(RC_FLOAT
));
1914 o(0xEE000A10|(r
<<12)|(r2
<<16)); /* fmsr */
1916 if(!(vtop
->type
.t
& VT_UNSIGNED
))
1917 r2
|=0x80; /* fuitoX -> fsituX */
1918 o(0xEEB80A40|r2
|T2CPR(t
)); /* fYitoX*/
1920 r2
=fpr(vtop
->r
=get_reg(RC_FLOAT
));
1921 if((t
& VT_BTYPE
) != VT_FLOAT
)
1922 dsize
=0x80; /* flts -> fltd */
1923 o(0xEE000110|dsize
|(r2
<<16)|(r
<<12)); /* flts */
1924 if((vtop
->type
.t
& (VT_UNSIGNED
|VT_BTYPE
)) == (VT_UNSIGNED
|VT_INT
)) {
1926 o(0xE3500000|(r
<<12)); /* cmp */
1927 r
=fpr(get_reg(RC_FLOAT
));
1928 if(last_itod_magic
) {
1929 off
=ind
+8-last_itod_magic
;
1934 o(0xBD1F0100|(r
<<12)|off
); /* ldflts */
1936 o(0xEA000000); /* b */
1937 last_itod_magic
=ind
;
1938 o(0x4F800000); /* 4294967296.0f */
1940 o(0xBE000100|dsize
|(r2
<<16)|(r2
<<12)|r
); /* adflt */
1944 } else if(bt
== VT_LLONG
) {
1946 CType
*func_type
= 0;
1947 if((t
& VT_BTYPE
) == VT_FLOAT
) {
1948 func_type
= &func_float_type
;
1949 if(vtop
->type
.t
& VT_UNSIGNED
)
1950 func
=TOK___floatundisf
;
1952 func
=TOK___floatdisf
;
1953 #if LDOUBLE_SIZE != 8
1954 } else if((t
& VT_BTYPE
) == VT_LDOUBLE
) {
1955 func_type
= &func_ldouble_type
;
1956 if(vtop
->type
.t
& VT_UNSIGNED
)
1957 func
=TOK___floatundixf
;
1959 func
=TOK___floatdixf
;
1960 } else if((t
& VT_BTYPE
) == VT_DOUBLE
) {
1962 } else if((t
& VT_BTYPE
) == VT_DOUBLE
|| (t
& VT_BTYPE
) == VT_LDOUBLE
) {
1964 func_type
= &func_double_type
;
1965 if(vtop
->type
.t
& VT_UNSIGNED
)
1966 func
=TOK___floatundidf
;
1968 func
=TOK___floatdidf
;
1971 vpush_global_sym(func_type
, func
);
1979 tcc_error("unimplemented gen_cvt_itof %x!",vtop
->type
.t
);
1982 /* convert fp to int 't' type */
1983 void gen_cvt_ftoi(int t
)
1989 r2
=vtop
->type
.t
& VT_BTYPE
;
1992 r
=vfpr(gv(RC_FLOAT
));
1994 o(0xEEBC0AC0|(r
<<12)|r
|T2CPR(r2
)|u
); /* ftoXizY */
1995 r2
=intr(vtop
->r
=get_reg(RC_INT
));
1996 o(0xEE100A10|(r
<<16)|(r2
<<12));
2001 func
=TOK___fixunssfsi
;
2002 #if LDOUBLE_SIZE != 8
2003 else if(r2
== VT_LDOUBLE
)
2004 func
=TOK___fixunsxfsi
;
2005 else if(r2
== VT_DOUBLE
)
2007 else if(r2
== VT_LDOUBLE
|| r2
== VT_DOUBLE
)
2009 func
=TOK___fixunsdfsi
;
2011 r
=fpr(gv(RC_FLOAT
));
2012 r2
=intr(vtop
->r
=get_reg(RC_INT
));
2013 o(0xEE100170|(r2
<<12)|r
);
2017 } else if(t
== VT_LLONG
) { // unsigned handled in gen_cvt_ftoi1
2020 #if LDOUBLE_SIZE != 8
2021 else if(r2
== VT_LDOUBLE
)
2023 else if(r2
== VT_DOUBLE
)
2025 else if(r2
== VT_LDOUBLE
|| r2
== VT_DOUBLE
)
2030 vpush_global_sym(&func_old_type
, func
);
2035 vtop
->r2
= REG_LRET
;
2039 tcc_error("unimplemented gen_cvt_ftoi!");
2042 /* convert from one floating point type to another */
2043 void gen_cvt_ftof(int t
)
2046 if(((vtop
->type
.t
& VT_BTYPE
) == VT_FLOAT
) != ((t
& VT_BTYPE
) == VT_FLOAT
)) {
2047 uint32_t r
= vfpr(gv(RC_FLOAT
));
2048 o(0xEEB70AC0|(r
<<12)|r
|T2CPR(vtop
->type
.t
));
2051 /* all we have to do on i386 and FPA ARM is to put the float in a register */
2056 /* computed goto support */
2063 /* Save the stack pointer onto the stack and return the location of its address */
2064 ST_FUNC
void gen_vla_sp_save(int addr
) {
2065 tcc_error("variable length arrays unsupported for this target");
2068 /* Restore the SP from a location on the stack */
2069 ST_FUNC
void gen_vla_sp_restore(int addr
) {
2070 tcc_error("variable length arrays unsupported for this target");
2073 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2074 ST_FUNC
void gen_vla_alloc(CType
*type
, int align
) {
2075 tcc_error("variable length arrays unsupported for this target");
2078 /* end of ARM code generator */
2079 /*************************************************************/
2081 /*************************************************************/