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 */
37 #ifndef TCC_CPU_VERSION
38 # define TCC_CPU_VERSION 5
41 /* a register can belong to several classes. The classes must be
42 sorted from more general to more precise (see gv2() code which does
43 assumptions on it). */
44 #define RC_INT 0x0001 /* generic integer register */
45 #define RC_FLOAT 0x0002 /* generic float register */
61 #define RC_IRET RC_R0 /* function return: integer register */
62 #define RC_IRE2 RC_R1 /* function return: second integer register */
63 #define RC_FRET RC_F0 /* function return: float register */
65 /* 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_IRE2 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 /******************************************************/
133 #else /* ! TARGET_DEFS_ONLY */
134 /******************************************************/
135 #define USING_GLOBALS
138 enum float_abi float_abi
;
140 ST_DATA
const int reg_classes
[NB_REGS
] = {
141 /* r0 */ RC_INT
| RC_R0
,
142 /* r1 */ RC_INT
| RC_R1
,
143 /* r2 */ RC_INT
| RC_R2
,
144 /* r3 */ RC_INT
| RC_R3
,
145 /* r12 */ RC_INT
| RC_R12
,
146 /* f0 */ RC_FLOAT
| RC_F0
,
147 /* f1 */ RC_FLOAT
| RC_F1
,
148 /* f2 */ RC_FLOAT
| RC_F2
,
149 /* f3 */ RC_FLOAT
| RC_F3
,
151 /* d4/s8 */ RC_FLOAT
| RC_F4
,
152 /* d5/s10 */ RC_FLOAT
| RC_F5
,
153 /* d6/s12 */ RC_FLOAT
| RC_F6
,
154 /* d7/s14 */ RC_FLOAT
| RC_F7
,
158 static int func_sub_sp_offset
, last_itod_magic
;
161 #if defined(CONFIG_TCC_BCHECK)
162 static addr_t func_bound_offset
;
163 static unsigned long func_bound_ind
;
164 static int func_bound_add_epilog
;
167 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
168 static CType float_type
, double_type
, func_float_type
, func_double_type
;
169 ST_FUNC
void arm_init(struct TCCState
*s
)
171 float_type
.t
= VT_FLOAT
;
172 double_type
.t
= VT_DOUBLE
;
173 func_float_type
.t
= VT_FUNC
;
174 func_float_type
.ref
= sym_push(SYM_FIELD
, &float_type
, FUNC_CDECL
, FUNC_OLD
);
175 func_double_type
.t
= VT_FUNC
;
176 func_double_type
.ref
= sym_push(SYM_FIELD
, &double_type
, FUNC_CDECL
, FUNC_OLD
);
178 float_abi
= s
->float_abi
;
179 #ifndef TCC_ARM_HARDFLOAT
180 # warning "soft float ABI currently not supported: default to softfp"
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(struct TCCState
*s
)
190 #if !defined (TCC_ARM_VFP)
191 tcc_warning("Support for FPA is deprecated and will be removed in next"
194 #if !defined (TCC_ARM_EABI)
195 tcc_warning("Support for OABI is deprecated and will be removed in next"
202 #define CHECK_R(r) ((r) >= TREG_R0 && (r) <= TREG_LR)
204 static int two2mask(int a
,int b
) {
205 if (!CHECK_R(a
) || !CHECK_R(b
))
206 tcc_error("compiler error! registers %i,%i is not valid",a
,b
);
207 return (reg_classes
[a
]|reg_classes
[b
])&~(RC_INT
|RC_FLOAT
);
210 static int regmask(int r
) {
212 tcc_error("compiler error! register %i is not valid",r
);
213 return reg_classes
[r
]&~(RC_INT
|RC_FLOAT
);
216 /******************************************************/
218 #if defined(TCC_ARM_EABI) && !defined(CONFIG_TCC_ELFINTERP)
219 const char *default_elfinterp(struct TCCState
*s
)
221 if (s
->float_abi
== ARM_HARD_FLOAT
)
222 return "/lib/ld-linux-armhf.so.3";
224 return "/lib/ld-linux.so.3";
230 /* this is a good place to start adding big-endian support*/
235 if (!cur_text_section
)
236 tcc_error("compiler error! This happens f.ex. if the compiler\n"
237 "can't evaluate constant expressions outside of a function.");
238 if (ind1
> cur_text_section
->data_allocated
)
239 section_realloc(cur_text_section
, ind1
);
240 cur_text_section
->data
[ind
++] = i
&255;
242 cur_text_section
->data
[ind
++] = i
&255;
244 cur_text_section
->data
[ind
++] = i
&255;
246 cur_text_section
->data
[ind
++] = i
;
249 static uint32_t stuff_const(uint32_t op
, uint32_t c
)
252 uint32_t nc
= 0, negop
= 0;
262 case 0x1A00000: //mov
263 case 0x1E00000: //mvn
270 return (op
&0xF010F000)|((op
>>16)&0xF)|0x1E00000;
274 return (op
&0xF010F000)|((op
>>16)&0xF)|0x1A00000;
275 case 0x1C00000: //bic
280 case 0x1800000: //orr
282 return (op
&0xFFF0FFFF)|0x1E00000;
288 if(c
<256) /* catch undefined <<32 */
291 m
=(0xff>>i
)|(0xff<<(32-i
));
293 return op
|(i
<<7)|(c
<<i
)|(c
>>(32-i
));
303 void stuff_const_harder(uint32_t op
, uint32_t v
) {
309 uint32_t a
[16], nv
, no
, o2
, n2
;
312 o2
=(op
&0xfff0ffff)|((op
&0xf000)<<4);;
314 a
[i
]=(a
[i
-1]>>2)|(a
[i
-1]<<30);
316 for(j
=i
<4?i
+12:15;j
>=i
+4;j
--)
317 if((v
&(a
[i
]|a
[j
]))==v
) {
318 o(stuff_const(op
,v
&a
[i
]));
319 o(stuff_const(o2
,v
&a
[j
]));
326 for(j
=i
<4?i
+12:15;j
>=i
+4;j
--)
327 if((nv
&(a
[i
]|a
[j
]))==nv
) {
328 o(stuff_const(no
,nv
&a
[i
]));
329 o(stuff_const(n2
,nv
&a
[j
]));
334 for(k
=i
<4?i
+12:15;k
>=j
+4;k
--)
335 if((v
&(a
[i
]|a
[j
]|a
[k
]))==v
) {
336 o(stuff_const(op
,v
&a
[i
]));
337 o(stuff_const(o2
,v
&a
[j
]));
338 o(stuff_const(o2
,v
&a
[k
]));
345 for(k
=i
<4?i
+12:15;k
>=j
+4;k
--)
346 if((nv
&(a
[i
]|a
[j
]|a
[k
]))==nv
) {
347 o(stuff_const(no
,nv
&a
[i
]));
348 o(stuff_const(n2
,nv
&a
[j
]));
349 o(stuff_const(n2
,nv
&a
[k
]));
352 o(stuff_const(op
,v
&a
[0]));
353 o(stuff_const(o2
,v
&a
[4]));
354 o(stuff_const(o2
,v
&a
[8]));
355 o(stuff_const(o2
,v
&a
[12]));
359 uint32_t encbranch(int pos
, int addr
, int fail
)
363 if(addr
>=0x1000000 || addr
<-0x1000000) {
365 tcc_error("FIXME: function bigger than 32MB");
368 return 0x0A000000|(addr
&0xffffff);
371 int decbranch(int pos
)
374 x
=*(uint32_t *)(cur_text_section
->data
+ pos
);
381 /* output a symbol and patch all calls to it */
382 void gsym_addr(int t
, int a
)
387 x
=(uint32_t *)(cur_text_section
->data
+ t
);
390 *x
=0xE1A00000; // nop
393 *x
|= encbranch(lt
,a
,1);
399 static uint32_t vfpr(int r
)
401 if(r
<TREG_F0
|| r
>TREG_F7
)
402 tcc_error("compiler error! register %i is no vfp register",r
);
406 static uint32_t fpr(int r
)
408 if(r
<TREG_F0
|| r
>TREG_F3
)
409 tcc_error("compiler error! register %i is no fpa register",r
);
414 static uint32_t intr(int r
)
418 if(r
>= TREG_R0
&& r
<= TREG_R3
)
420 if (!(r
>= TREG_SP
&& r
<= TREG_LR
))
421 tcc_error("compiler error! register %i is no int register",r
);
422 return r
+ (13 - TREG_SP
);
425 static void calcaddr(uint32_t *base
, int *off
, int *sgn
, int maxoff
, unsigned shift
)
427 if(*off
>maxoff
|| *off
&((1<<shift
)-1)) {
434 y
=stuff_const(x
,*off
&~maxoff
);
440 y
=stuff_const(x
,(*off
+maxoff
)&~maxoff
);
444 *off
=((*off
+maxoff
)&~maxoff
)-*off
;
447 stuff_const_harder(x
,*off
&~maxoff
);
452 static uint32_t mapcc(int cc
)
457 return 0x30000000; /* CC/LO */
459 return 0x20000000; /* CS/HS */
461 return 0x00000000; /* EQ */
463 return 0x10000000; /* NE */
465 return 0x90000000; /* LS */
467 return 0x80000000; /* HI */
469 return 0x40000000; /* MI */
471 return 0x50000000; /* PL */
473 return 0xB0000000; /* LT */
475 return 0xA0000000; /* GE */
477 return 0xD0000000; /* LE */
479 return 0xC0000000; /* GT */
481 tcc_error("unexpected condition code");
482 return 0xE0000000; /* AL */
485 static int negcc(int cc
)
514 tcc_error("unexpected condition code");
518 /* load 'r' from value 'sv' */
519 void load(int r
, SValue
*sv
)
521 int v
, ft
, fc
, fr
, sign
;
538 uint32_t base
= 0xB; // fp
541 v1
.r
= VT_LOCAL
| VT_LVAL
;
547 } else if(v
== VT_CONST
) {
556 } else if(v
< VT_CONST
) {
563 calcaddr(&base
,&fc
,&sign
,1020,2);
565 op
=0xED100A00; /* flds */
568 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
569 op
|=0x100; /* flds -> fldd */
570 o(op
|(vfpr(r
)<<12)|(fc
>>2)|(base
<<16));
575 #if LDOUBLE_SIZE == 8
576 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
579 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
581 else if ((ft
& VT_BTYPE
) == VT_LDOUBLE
)
584 o(op
|(fpr(r
)<<12)|(fc
>>2)|(base
<<16));
586 } else if((ft
& (VT_BTYPE
|VT_UNSIGNED
)) == VT_BYTE
587 || (ft
& VT_BTYPE
) == VT_SHORT
) {
588 calcaddr(&base
,&fc
,&sign
,255,0);
590 if ((ft
& VT_BTYPE
) == VT_SHORT
)
592 if ((ft
& VT_UNSIGNED
) == 0)
596 o(op
|(intr(r
)<<12)|(base
<<16)|((fc
&0xf0)<<4)|(fc
&0xf));
598 calcaddr(&base
,&fc
,&sign
,4095,0);
602 if ((ft
& VT_BTYPE
) == VT_BYTE
|| (ft
& VT_BTYPE
) == VT_BOOL
)
604 o(op
|(intr(r
)<<12)|fc
|(base
<<16));
610 op
=stuff_const(0xE3A00000|(intr(r
)<<12),sv
->c
.i
);
611 if (fr
& VT_SYM
|| !op
) {
612 o(0xE59F0000|(intr(r
)<<12));
615 greloc(cur_text_section
, sv
->sym
, ind
, R_ARM_ABS32
);
620 } else if (v
== VT_LOCAL
) {
621 op
=stuff_const(0xE28B0000|(intr(r
)<<12),sv
->c
.i
);
622 if (fr
& VT_SYM
|| !op
) {
623 o(0xE59F0000|(intr(r
)<<12));
625 if(fr
& VT_SYM
) // needed ?
626 greloc(cur_text_section
, sv
->sym
, ind
, R_ARM_ABS32
);
628 o(0xE08B0000|(intr(r
)<<12)|intr(r
));
632 } else if(v
== VT_CMP
) {
633 o(mapcc(sv
->c
.i
)|0x3A00001|(intr(r
)<<12));
634 o(mapcc(negcc(sv
->c
.i
))|0x3A00000|(intr(r
)<<12));
636 } else if (v
== VT_JMP
|| v
== VT_JMPI
) {
639 o(0xE3A00000|(intr(r
)<<12)|t
);
642 o(0xE3A00000|(intr(r
)<<12)|(t
^1));
644 } else if (v
< VT_CONST
) {
647 o(0xEEB00A40|(vfpr(r
)<<12)|vfpr(v
)|T2CPR(ft
)); /* fcpyX */
649 o(0xEE008180|(fpr(r
)<<12)|fpr(v
));
652 o(0xE1A00000|(intr(r
)<<12)|intr(v
));
656 tcc_error("load unimplemented!");
659 /* store register 'r' in lvalue 'v' */
660 void store(int r
, SValue
*sv
)
663 int v
, ft
, fc
, fr
, sign
;
678 if (fr
& VT_LVAL
|| fr
== VT_LOCAL
) {
679 uint32_t base
= 0xb; /* fp */
684 } else if(v
== VT_CONST
) {
696 calcaddr(&base
,&fc
,&sign
,1020,2);
698 op
=0xED000A00; /* fsts */
701 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
702 op
|=0x100; /* fsts -> fstd */
703 o(op
|(vfpr(r
)<<12)|(fc
>>2)|(base
<<16));
708 #if LDOUBLE_SIZE == 8
709 if ((ft
& VT_BTYPE
) != VT_FLOAT
)
712 if ((ft
& VT_BTYPE
) == VT_DOUBLE
)
714 if ((ft
& VT_BTYPE
) == VT_LDOUBLE
)
717 o(op
|(fpr(r
)<<12)|(fc
>>2)|(base
<<16));
720 } else if((ft
& VT_BTYPE
) == VT_SHORT
) {
721 calcaddr(&base
,&fc
,&sign
,255,0);
725 o(op
|(intr(r
)<<12)|(base
<<16)|((fc
&0xf0)<<4)|(fc
&0xf));
727 calcaddr(&base
,&fc
,&sign
,4095,0);
731 if ((ft
& VT_BTYPE
) == VT_BYTE
|| (ft
& VT_BTYPE
) == VT_BOOL
)
733 o(op
|(intr(r
)<<12)|fc
|(base
<<16));
738 tcc_error("store unimplemented");
741 static void gadd_sp(int val
)
743 stuff_const_harder(0xE28DD000,val
);
746 /* 'is_jmp' is '1' if it is a jump */
747 static void gcall_or_jmp(int is_jmp
)
751 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
)) == VT_CONST
) {
753 if(vtop
->r
& VT_SYM
){
754 x
=encbranch(ind
,ind
+vtop
->c
.i
,0);
756 /* relocation case */
757 greloc(cur_text_section
, vtop
->sym
, ind
, R_ARM_PC24
);
758 o(x
|(is_jmp
?0xE0000000:0xE1000000));
761 o(0xE28FE004); // add lr,pc,#4
762 o(0xE51FF004); // ldr pc,[pc,#-4]
763 greloc(cur_text_section
, vtop
->sym
, ind
, R_ARM_ABS32
);
766 #ifdef CONFIG_TCC_BCHECK
767 if (tcc_state
->do_bounds_check
&&
768 (vtop
->sym
->v
== TOK_setjmp
||
769 vtop
->sym
->v
== TOK__setjmp
||
770 vtop
->sym
->v
== TOK_sigsetjmp
||
771 vtop
->sym
->v
== TOK___sigsetjmp
))
772 func_bound_add_epilog
= 1;
776 o(0xE28FE004); // add lr,pc,#4
777 o(0xE51FF004); // ldr pc,[pc,#-4]
781 /* otherwise, indirect call */
782 #ifdef CONFIG_TCC_BCHECK
783 vtop
->r
&= ~VT_MUSTBOUND
;
787 o(0xE1A0E00F); // mov lr,pc
788 o(0xE1A0F000|intr(r
)); // mov pc,r
792 #if defined(CONFIG_TCC_BCHECK)
794 static void gen_bounds_call(int v
)
796 Sym
*sym
= external_global_sym(v
, &func_old_type
);
798 greloc(cur_text_section
, sym
, ind
, R_ARM_PC24
);
802 /* generate a bounded pointer addition */
803 ST_FUNC
void gen_bounded_ptr_add(void)
805 vpush_global_sym(&func_old_type
, TOK___bound_ptr_add
);
809 /* returned pointer is in REG_IRET */
810 vtop
->r
= REG_IRET
| VT_BOUNDED
;
813 /* relocation offset of the bounding function call point */
814 vtop
->c
.i
= (cur_text_section
->reloc
->data_offset
- sizeof(Elf32_Rel
));
817 /* patch pointer addition in vtop so that pointer dereferencing is
819 ST_FUNC
void gen_bounded_ptr_deref(void)
829 size
= type_size(&vtop
->type
, &align
);
831 case 1: func
= TOK___bound_ptr_indir1
; break;
832 case 2: func
= TOK___bound_ptr_indir2
; break;
833 case 4: func
= TOK___bound_ptr_indir4
; break;
834 case 8: func
= TOK___bound_ptr_indir8
; break;
835 case 12: func
= TOK___bound_ptr_indir12
; break;
836 case 16: func
= TOK___bound_ptr_indir16
; break;
838 /* may happen with struct member access */
840 //tcc_error("unhandled size when dereferencing bounded pointer");
844 sym
= external_global_sym(func
, &func_old_type
);
846 put_extern_sym(sym
, NULL
, 0, 0);
847 /* patch relocation */
848 /* XXX: find a better solution ? */
849 rel
= (Elf32_Rel
*)(cur_text_section
->reloc
->data
+ vtop
->c
.i
);
850 rel
->r_info
= ELF32_R_INFO(sym
->c
, ELF32_R_TYPE(rel
->r_info
));
853 static void gen_bounds_prolog(void)
855 /* leave some room for bound checking code */
856 func_bound_offset
= lbounds_section
->data_offset
;
857 func_bound_ind
= ind
;
858 func_bound_add_epilog
= 0;
859 o(0xe1a00000); /* ld r0,lbounds_section->data_offset */
862 o(0xe1a00000); /* call __bound_local_new */
865 static void gen_bounds_epilog(void)
870 int offset_modified
= func_bound_offset
!= lbounds_section
->data_offset
;
872 if (!offset_modified
&& !func_bound_add_epilog
)
875 /* add end of table info */
876 bounds_ptr
= section_ptr_add(lbounds_section
, sizeof(addr_t
));
879 sym_data
= get_sym_ref(&char_pointer_type
, lbounds_section
,
880 func_bound_offset
, lbounds_section
->data_offset
);
882 /* generate bound local allocation */
883 if (offset_modified
) {
885 ind
= func_bound_ind
;
886 o(0xe59f0000); /* ldr r0, [pc] */
887 o(0xea000000); /* b $+4 */
888 greloc(cur_text_section
, sym_data
, ind
, R_ARM_ABS32
);
889 o(0x00000000); /* lbounds_section->data_offset */
890 gen_bounds_call(TOK___bound_local_new
);
894 /* generate bound check local freeing */
895 o(0xe92d0003); /* push {r0,r1} */
896 o(0xed2d0b02); /* vpush {d0} */
897 o(0xe59f0000); /* ldr r0, [pc] */
898 o(0xea000000); /* b $+4 */
899 greloc(cur_text_section
, sym_data
, ind
, R_ARM_ABS32
);
900 o(0x00000000); /* lbounds_section->data_offset */
901 gen_bounds_call(TOK___bound_local_delete
);
902 o(0xecbd0b02); /* vpop {d0} */
903 o(0xe8bd0003); /* pop {r0,r1} */
907 static int unalias_ldbl(int btype
)
909 #if LDOUBLE_SIZE == 8
910 if (btype
== VT_LDOUBLE
)
916 /* Return whether a structure is an homogeneous float aggregate or not.
917 The answer is true if all the elements of the structure are of the same
918 primitive float type and there is less than 4 elements.
920 type: the type corresponding to the structure to be tested */
921 static int is_hgen_float_aggr(CType
*type
)
923 if ((type
->t
& VT_BTYPE
) == VT_STRUCT
) {
925 int btype
, nb_fields
= 0;
927 ref
= type
->ref
->next
;
928 btype
= unalias_ldbl(ref
->type
.t
& VT_BTYPE
);
929 if (btype
== VT_FLOAT
|| btype
== VT_DOUBLE
) {
930 for(; ref
&& btype
== unalias_ldbl(ref
->type
.t
& VT_BTYPE
); ref
= ref
->next
, nb_fields
++);
931 return !ref
&& nb_fields
<= 4;
938 signed char avail
[3]; /* 3 holes max with only float and double alignments */
939 int first_hole
; /* first available hole */
940 int last_hole
; /* last available hole (none if equal to first_hole) */
941 int first_free_reg
; /* next free register in the sequence, hole excluded */
944 #define AVAIL_REGS_INITIALIZER (struct avail_regs) { { 0, 0, 0}, 0, 0, 0 }
946 /* Find suitable registers for a VFP Co-Processor Register Candidate (VFP CPRC
947 param) according to the rules described in the procedure call standard for
948 the ARM architecture (AAPCS). If found, the registers are assigned to this
949 VFP CPRC parameter. Registers are allocated in sequence unless a hole exists
950 and the parameter is a single float.
952 avregs: opaque structure to keep track of available VFP co-processor regs
953 align: alignment constraints for the param, as returned by type_size()
954 size: size of the parameter, as returned by type_size() */
955 int assign_vfpreg(struct avail_regs
*avregs
, int align
, int size
)
959 if (avregs
->first_free_reg
== -1)
961 if (align
>> 3) { /* double alignment */
962 first_reg
= avregs
->first_free_reg
;
963 /* alignment constraint not respected so use next reg and record hole */
965 avregs
->avail
[avregs
->last_hole
++] = first_reg
++;
966 } else { /* no special alignment (float or array of float) */
967 /* if single float and a hole is available, assign the param to it */
968 if (size
== 4 && avregs
->first_hole
!= avregs
->last_hole
)
969 return avregs
->avail
[avregs
->first_hole
++];
971 first_reg
= avregs
->first_free_reg
;
973 if (first_reg
+ size
/ 4 <= 16) {
974 avregs
->first_free_reg
= first_reg
+ size
/ 4;
977 avregs
->first_free_reg
= -1;
981 /* Returns whether all params need to be passed in core registers or not.
982 This is the case for function part of the runtime ABI. */
983 int floats_in_core_regs(SValue
*sval
)
988 switch (sval
->sym
->v
) {
989 case TOK___floatundisf
:
990 case TOK___floatundidf
:
991 case TOK___fixunssfdi
:
992 case TOK___fixunsdfdi
:
994 case TOK___fixunsxfdi
:
996 case TOK___floatdisf
:
997 case TOK___floatdidf
:
1007 /* Return the number of registers needed to return the struct, or 0 if
1008 returning via struct pointer. */
1009 ST_FUNC
int gfunc_sret(CType
*vt
, int variadic
, CType
*ret
, int *ret_align
, int *regsize
) {
1012 size
= type_size(vt
, &align
);
1013 if (float_abi
== ARM_HARD_FLOAT
&& !variadic
&&
1014 (is_float(vt
->t
) || is_hgen_float_aggr(vt
))) {
1019 return (size
+ 7) >> 3;
1020 } else if (size
<= 4) {
1033 /* Parameters are classified according to how they are copied to their final
1034 destination for the function call. Because the copying is performed class
1035 after class according to the order in the union below, it is important that
1036 some constraints about the order of the members of this union are respected:
1037 - CORE_STRUCT_CLASS must come after STACK_CLASS;
1038 - CORE_CLASS must come after STACK_CLASS, CORE_STRUCT_CLASS and
1040 - VFP_STRUCT_CLASS must come after VFP_CLASS.
1041 See the comment for the main loop in copy_params() for the reason. */
1052 int start
; /* first reg or addr used depending on the class */
1053 int end
; /* last reg used or next free addr depending on the class */
1054 SValue
*sval
; /* pointer to SValue on the value stack */
1055 struct param_plan
*prev
; /* previous element in this class */
1059 struct param_plan
*pplans
; /* array of all the param plans */
1060 struct param_plan
*clsplans
[NB_CLASSES
]; /* per class lists of param plans */
1063 #define add_param_plan(plan,pplan,class) \
1065 pplan.prev = plan->clsplans[class]; \
1066 plan->pplans[plan ## _nb] = pplan; \
1067 plan->clsplans[class] = &plan->pplans[plan ## _nb++]; \
1070 /* Assign parameters to registers and stack with alignment according to the
1071 rules in the procedure call standard for the ARM architecture (AAPCS).
1072 The overall assignment is recorded in an array of per parameter structures
1073 called parameter plans. The parameter plans are also further organized in a
1074 number of linked lists, one per class of parameter (see the comment for the
1075 definition of union reg_class).
1077 nb_args: number of parameters of the function for which a call is generated
1078 float_abi: float ABI in use for this function call
1079 plan: the structure where the overall assignment is recorded
1080 todo: a bitmap that record which core registers hold a parameter
1082 Returns the amount of stack space needed for parameter passing
1084 Note: this function allocated an array in plan->pplans with tcc_malloc. It
1085 is the responsibility of the caller to free this array once used (ie not
1086 before copy_params). */
1087 static int assign_regs(int nb_args
, int float_abi
, struct plan
*plan
, int *todo
)
1090 int ncrn
/* next core register number */, nsaa
/* next stacked argument address*/;
1092 struct param_plan pplan
;
1093 struct avail_regs avregs
= AVAIL_REGS_INITIALIZER
;
1097 plan
->pplans
= nb_args
? tcc_malloc(nb_args
* sizeof(*plan
->pplans
)) : NULL
;
1098 memset(plan
->clsplans
, 0, sizeof(plan
->clsplans
));
1099 for(i
= nb_args
; i
-- ;) {
1100 int j
, start_vfpreg
= 0;
1101 CType type
= vtop
[-i
].type
;
1102 type
.t
&= ~VT_ARRAY
;
1103 size
= type_size(&type
, &align
);
1104 size
= (size
+ 3) & ~3;
1105 align
= (align
+ 3) & ~3;
1106 switch(vtop
[-i
].type
.t
& VT_BTYPE
) {
1111 if (float_abi
== ARM_HARD_FLOAT
) {
1112 int is_hfa
= 0; /* Homogeneous float aggregate */
1114 if (is_float(vtop
[-i
].type
.t
)
1115 || (is_hfa
= is_hgen_float_aggr(&vtop
[-i
].type
))) {
1118 start_vfpreg
= assign_vfpreg(&avregs
, align
, size
);
1119 end_vfpreg
= start_vfpreg
+ ((size
- 1) >> 2);
1120 if (start_vfpreg
>= 0) {
1121 pplan
= (struct param_plan
) {start_vfpreg
, end_vfpreg
, &vtop
[-i
]};
1123 add_param_plan(plan
, pplan
, VFP_STRUCT_CLASS
);
1125 add_param_plan(plan
, pplan
, VFP_CLASS
);
1131 ncrn
= (ncrn
+ (align
-1)/4) & ~((align
/4) - 1);
1132 if (ncrn
+ size
/4 <= 4 || (ncrn
< 4 && start_vfpreg
!= -1)) {
1133 /* The parameter is allocated both in core register and on stack. As
1134 * such, it can be of either class: it would either be the last of
1135 * CORE_STRUCT_CLASS or the first of STACK_CLASS. */
1136 for (j
= ncrn
; j
< 4 && j
< ncrn
+ size
/ 4; j
++)
1138 pplan
= (struct param_plan
) {ncrn
, j
, &vtop
[-i
]};
1139 add_param_plan(plan
, pplan
, CORE_STRUCT_CLASS
);
1142 nsaa
= (ncrn
- 4) * 4;
1150 int is_long
= (vtop
[-i
].type
.t
& VT_BTYPE
) == VT_LLONG
;
1153 ncrn
= (ncrn
+ 1) & -2;
1157 pplan
= (struct param_plan
) {ncrn
, ncrn
, &vtop
[-i
]};
1161 add_param_plan(plan
, pplan
, CORE_CLASS
);
1165 nsaa
= (nsaa
+ (align
- 1)) & ~(align
- 1);
1166 pplan
= (struct param_plan
) {nsaa
, nsaa
+ size
, &vtop
[-i
]};
1167 add_param_plan(plan
, pplan
, STACK_CLASS
);
1168 nsaa
+= size
; /* size already rounded up before */
1173 #undef add_param_plan
1175 /* Copy parameters to their final destination (core reg, VFP reg or stack) for
1178 nb_args: number of parameters the function take
1179 plan: the overall assignment plan for parameters
1180 todo: a bitmap indicating what core reg will hold a parameter
1182 Returns the number of SValue added by this function on the value stack */
1183 static int copy_params(int nb_args
, struct plan
*plan
, int todo
)
1185 int size
, align
, r
, i
, nb_extra_sval
= 0;
1186 struct param_plan
*pplan
;
1189 /* Several constraints require parameters to be copied in a specific order:
1190 - structures are copied to the stack before being loaded in a reg;
1191 - floats loaded to an odd numbered VFP reg are first copied to the
1192 preceding even numbered VFP reg and then moved to the next VFP reg.
1194 It is thus important that:
1195 - structures assigned to core regs must be copied after parameters
1196 assigned to the stack but before structures assigned to VFP regs because
1197 a structure can lie partly in core registers and partly on the stack;
1198 - parameters assigned to the stack and all structures be copied before
1199 parameters assigned to a core reg since copying a parameter to the stack
1200 require using a core reg;
1201 - parameters assigned to VFP regs be copied before structures assigned to
1202 VFP regs as the copy might use an even numbered VFP reg that already
1203 holds part of a structure. */
1205 for(i
= 0; i
< NB_CLASSES
; i
++) {
1206 for(pplan
= plan
->clsplans
[i
]; pplan
; pplan
= pplan
->prev
) {
1209 && (i
!= CORE_CLASS
|| pplan
->sval
->r
< VT_CONST
))
1212 vpushv(pplan
->sval
);
1213 pplan
->sval
->r
= pplan
->sval
->r2
= VT_CONST
; /* disable entry */
1216 case CORE_STRUCT_CLASS
:
1217 case VFP_STRUCT_CLASS
:
1218 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_STRUCT
) {
1220 size
= type_size(&pplan
->sval
->type
, &align
);
1221 /* align to stack align size */
1222 size
= (size
+ 3) & ~3;
1223 if (i
== STACK_CLASS
&& pplan
->prev
)
1224 padding
= pplan
->start
- pplan
->prev
->end
;
1225 size
+= padding
; /* Add padding if any */
1226 /* allocate the necessary size on stack */
1228 /* generate structure store */
1229 r
= get_reg(RC_INT
);
1230 o(0xE28D0000|(intr(r
)<<12)|padding
); /* add r, sp, padding */
1231 vset(&vtop
->type
, r
| VT_LVAL
, 0);
1233 vstore(); /* memcpy to current sp + potential padding */
1235 /* Homogeneous float aggregate are loaded to VFP registers
1236 immediately since there is no way of loading data in multiple
1237 non consecutive VFP registers as what is done for other
1238 structures (see the use of todo). */
1239 if (i
== VFP_STRUCT_CLASS
) {
1240 int first
= pplan
->start
, nb
= pplan
->end
- first
+ 1;
1241 /* vpop.32 {pplan->start, ..., pplan->end} */
1242 o(0xECBD0A00|(first
&1)<<22|(first
>>1)<<12|nb
);
1243 /* No need to write the register used to a SValue since VFP regs
1244 cannot be used for gcall_or_jmp */
1247 if (is_float(pplan
->sval
->type
.t
)) {
1249 r
= vfpr(gv(RC_FLOAT
)) << 12;
1250 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
1254 r
|= 0x101; /* vpush.32 -> vpush.64 */
1256 o(0xED2D0A01 + r
); /* vpush */
1258 r
= fpr(gv(RC_FLOAT
)) << 12;
1259 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_FLOAT
)
1261 else if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1264 size
= LDOUBLE_SIZE
;
1271 o(0xED2D0100|r
|(size
>>2)); /* some kind of vpush for FPA */
1274 /* simple type (currently always same size) */
1275 /* XXX: implicit cast ? */
1277 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1281 o(0xE52D0004|(intr(r
)<<12)); /* push r */
1285 o(0xE52D0004|(intr(r
)<<12)); /* push r */
1287 if (i
== STACK_CLASS
&& pplan
->prev
)
1288 gadd_sp(pplan
->prev
->end
- pplan
->start
); /* Add padding if any */
1293 gv(regmask(TREG_F0
+ (pplan
->start
>> 1)));
1294 if (pplan
->start
& 1) { /* Must be in upper part of double register */
1295 o(0xEEF00A40|((pplan
->start
>>1)<<12)|(pplan
->start
>>1)); /* vmov.f32 s(n+1), sn */
1296 vtop
->r
= VT_CONST
; /* avoid being saved on stack by gv for next float */
1301 if ((pplan
->sval
->type
.t
& VT_BTYPE
) == VT_LLONG
) {
1303 gv(regmask(pplan
->end
));
1304 pplan
->sval
->r2
= vtop
->r
;
1307 gv(regmask(pplan
->start
));
1308 /* Mark register as used so that gcall_or_jmp use another one
1309 (regs >=4 are free as never used to pass parameters) */
1310 pplan
->sval
->r
= vtop
->r
;
1317 /* second pass to restore registers that were saved on stack by accident.
1318 Maybe redundant after the "lvalue_save" patch in tccgen.c:gv() */
1322 /* Manually free remaining registers since next parameters are loaded
1323 * manually, without the help of gv(int). */
1327 o(0xE8BD0000|todo
); /* pop {todo} */
1328 for(pplan
= plan
->clsplans
[CORE_STRUCT_CLASS
]; pplan
; pplan
= pplan
->prev
) {
1330 pplan
->sval
->r
= pplan
->start
;
1331 /* An SValue can only pin 2 registers at best (r and r2) but a structure
1332 can occupy more than 2 registers. Thus, we need to push on the value
1333 stack some fake parameter to have on SValue for each registers used
1334 by a structure (r2 is not used). */
1335 for (r
= pplan
->start
+ 1; r
<= pplan
->end
; r
++) {
1336 if (todo
& (1 << r
)) {
1344 return nb_extra_sval
;
1347 /* Generate function call. The function address is pushed first, then
1348 all the parameters in call order. This functions pops all the
1349 parameters and the function address. */
1350 void gfunc_call(int nb_args
)
1353 int def_float_abi
= float_abi
;
1360 #ifdef CONFIG_TCC_BCHECK
1361 if (tcc_state
->do_bounds_check
)
1362 gbound_args(nb_args
);
1366 if (float_abi
== ARM_HARD_FLOAT
) {
1367 variadic
= (vtop
[-nb_args
].type
.ref
->f
.func_type
== FUNC_ELLIPSIS
);
1368 if (variadic
|| floats_in_core_regs(&vtop
[-nb_args
]))
1369 float_abi
= ARM_SOFTFP_FLOAT
;
1372 /* cannot let cpu flags if other instruction are generated. Also avoid leaving
1373 VT_JMP anywhere except on the top of the stack because it would complicate
1374 the code generator. */
1375 r
= vtop
->r
& VT_VALMASK
;
1376 if (r
== VT_CMP
|| (r
& ~1) == VT_JMP
)
1379 args_size
= assign_regs(nb_args
, float_abi
, &plan
, &todo
);
1382 if (args_size
& 7) { /* Stack must be 8 byte aligned at fct call for EABI */
1383 args_size
= (args_size
+ 7) & ~7;
1384 o(0xE24DD004); /* sub sp, sp, #4 */
1388 nb_args
+= copy_params(nb_args
, &plan
, todo
);
1389 tcc_free(plan
.pplans
);
1391 /* Move fct SValue on top as required by gcall_or_jmp */
1395 gadd_sp(args_size
); /* pop all parameters passed on the stack */
1396 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
1397 if(float_abi
== ARM_SOFTFP_FLOAT
&& is_float(vtop
->type
.ref
->type
.t
)) {
1398 if((vtop
->type
.ref
->type
.t
& VT_BTYPE
) == VT_FLOAT
) {
1399 o(0xEE000A10); /*vmov s0, r0 */
1401 o(0xEE000B10); /* vmov.32 d0[0], r0 */
1402 o(0xEE201B10); /* vmov.32 d0[1], r1 */
1406 vtop
-= nb_args
+ 1; /* Pop all params and fct address from value stack */
1407 leaffunc
= 0; /* we are calling a function, so we aren't in a leaf function */
1408 float_abi
= def_float_abi
;
1411 /* generate function prolog of type 't' */
1412 void gfunc_prolog(Sym
*func_sym
)
1414 CType
*func_type
= &func_sym
->type
;
1416 int n
, nf
, size
, align
, rs
, struct_ret
= 0;
1417 int addr
, pn
, sn
; /* pn=core, sn=stack */
1421 struct avail_regs avregs
= AVAIL_REGS_INITIALIZER
;
1424 sym
= func_type
->ref
;
1427 if ((func_vt
.t
& VT_BTYPE
) == VT_STRUCT
&&
1428 !gfunc_sret(&func_vt
, func_var
, &ret_type
, &align
, &rs
))
1432 func_vc
= 12; /* Offset from fp of the place to store the result */
1434 for(sym2
= sym
->next
; sym2
&& (n
< 4 || nf
< 16); sym2
= sym2
->next
) {
1435 size
= type_size(&sym2
->type
, &align
);
1437 if (float_abi
== ARM_HARD_FLOAT
&& !func_var
&&
1438 (is_float(sym2
->type
.t
) || is_hgen_float_aggr(&sym2
->type
))) {
1439 int tmpnf
= assign_vfpreg(&avregs
, align
, size
);
1440 tmpnf
+= (size
+ 3) / 4;
1441 nf
= (tmpnf
> nf
) ? tmpnf
: nf
;
1445 n
+= (size
+ 3) / 4;
1447 o(0xE1A0C00D); /* mov ip,sp */
1456 o(0xE92D0000|((1<<n
)-1)); /* save r0-r4 on stack if needed */
1461 nf
=(nf
+1)&-2; /* nf => HARDFLOAT => EABI */
1462 o(0xED2D0A00|nf
); /* save s0-s15 on stack if needed */
1464 o(0xE92D5800); /* save fp, ip, lr */
1465 o(0xE1A0B00D); /* mov fp, sp */
1466 func_sub_sp_offset
= ind
;
1467 o(0xE1A00000); /* nop, leave space for stack adjustment in epilog */
1470 if (float_abi
== ARM_HARD_FLOAT
) {
1472 avregs
= AVAIL_REGS_INITIALIZER
;
1475 pn
= struct_ret
, sn
= 0;
1476 while ((sym
= sym
->next
)) {
1479 size
= type_size(type
, &align
);
1480 size
= (size
+ 3) >> 2;
1481 align
= (align
+ 3) & ~3;
1483 if (float_abi
== ARM_HARD_FLOAT
&& !func_var
&& (is_float(sym
->type
.t
)
1484 || is_hgen_float_aggr(&sym
->type
))) {
1485 int fpn
= assign_vfpreg(&avregs
, align
, size
<< 2);
1494 pn
= (pn
+ (align
-1)/4) & -(align
/4);
1496 addr
= (nf
+ pn
) * 4;
1503 sn
= (sn
+ (align
-1)/4) & -(align
/4);
1505 addr
= (n
+ nf
+ sn
) * 4;
1508 sym_push(sym
->v
& ~SYM_FIELD
, type
, VT_LOCAL
| VT_LVAL
,
1514 #ifdef CONFIG_TCC_BCHECK
1515 if (tcc_state
->do_bounds_check
)
1516 gen_bounds_prolog();
1520 /* generate function epilog */
1521 void gfunc_epilog(void)
1526 #ifdef CONFIG_TCC_BCHECK
1527 if (tcc_state
->do_bounds_check
)
1528 gen_bounds_epilog();
1530 /* Copy float return value to core register if base standard is used and
1531 float computation is made with VFP */
1532 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
1533 if ((float_abi
== ARM_SOFTFP_FLOAT
|| func_var
) && is_float(func_vt
.t
)) {
1534 if((func_vt
.t
& VT_BTYPE
) == VT_FLOAT
)
1535 o(0xEE100A10); /* fmrs r0, s0 */
1537 o(0xEE100B10); /* fmrdl r0, d0 */
1538 o(0xEE301B10); /* fmrdh r1, d0 */
1542 o(0xE89BA800); /* restore fp, sp, pc */
1543 diff
= (-loc
+ 3) & -4;
1546 diff
= ((diff
+ 11) & -8) - 4;
1549 x
=stuff_const(0xE24BD000, diff
); /* sub sp,fp,# */
1551 *(uint32_t *)(cur_text_section
->data
+ func_sub_sp_offset
) = x
;
1555 o(0xE59FC004); /* ldr ip,[pc+4] */
1556 o(0xE04BD00C); /* sub sp,fp,ip */
1557 o(0xE1A0F00E); /* mov pc,lr */
1559 *(uint32_t *)(cur_text_section
->data
+ func_sub_sp_offset
) = 0xE1000000|encbranch(func_sub_sp_offset
,addr
,1);
1564 ST_FUNC
void gen_fill_nops(int bytes
)
1567 tcc_error("alignment of code section not multiple of 4");
1574 /* generate a jump to a label */
1575 ST_FUNC
int gjmp(int t
)
1581 o(0xE0000000|encbranch(r
,t
,1));
1585 /* generate a jump to a fixed address */
1586 ST_FUNC
void gjmp_addr(int a
)
1591 ST_FUNC
int gjmp_cond(int op
, int t
)
1598 op
|=encbranch(r
,t
,1);
1603 ST_FUNC
int gjmp_append(int n
, int t
)
1610 p
= decbranch(lp
=p
);
1612 x
= (uint32_t *)(cur_text_section
->data
+ lp
);
1614 *x
|= encbranch(lp
,t
,1);
1620 /* generate an integer binary operation */
1621 void gen_opi(int op
)
1624 uint32_t opc
= 0, r
, fr
;
1625 unsigned short retreg
= REG_IRET
;
1633 case TOK_ADDC1
: /* add with carry generation */
1641 case TOK_SUBC1
: /* sub with carry generation */
1645 case TOK_ADDC2
: /* add with carry use */
1649 case TOK_SUBC2
: /* sub with carry use */
1666 gv2(RC_INT
, RC_INT
);
1670 o(0xE0000090|(intr(r
)<<16)|(intr(r
)<<8)|intr(fr
));
1695 func
=TOK___aeabi_idivmod
;
1704 func
=TOK___aeabi_uidivmod
;
1712 gv2(RC_INT
, RC_INT
);
1713 r
=intr(vtop
[-1].r2
=get_reg(RC_INT
));
1715 vtop
[-1].r
=get_reg_ex(RC_INT
,regmask(c
));
1717 o(0xE0800090|(r
<<16)|(intr(vtop
->r
)<<12)|(intr(c
)<<8)|intr(vtop
[1].r
));
1726 if((vtop
[-1].r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1727 if(opc
== 4 || opc
== 5 || opc
== 0xc) {
1729 opc
|=2; // sub -> rsb
1732 if ((vtop
->r
& VT_VALMASK
) == VT_CMP
||
1733 (vtop
->r
& (VT_VALMASK
& ~1)) == VT_JMP
)
1738 opc
=0xE0000000|(opc
<<20);
1739 if((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1741 x
=stuff_const(opc
|0x2000000|(c
<<16),vtop
->c
.i
);
1743 r
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,regmask(vtop
[-1].r
)));
1748 fr
=intr(gv(RC_INT
));
1749 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
1754 r
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,two2mask(vtop
->r
,vtop
[-1].r
)));
1755 o(opc
|(c
<<16)|(r
<<12)|fr
);
1758 if (op
>= TOK_ULT
&& op
<= TOK_GT
)
1762 opc
=0xE1A00000|(opc
<<5);
1763 if ((vtop
->r
& VT_VALMASK
) == VT_CMP
||
1764 (vtop
->r
& (VT_VALMASK
& ~1)) == VT_JMP
)
1769 if ((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) == VT_CONST
) {
1770 fr
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,regmask(vtop
[-1].r
)));
1771 c
= vtop
->c
.i
& 0x1f;
1772 o(opc
|r
|(c
<<7)|(fr
<<12));
1774 fr
=intr(gv(RC_INT
));
1775 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
1780 c
=intr(vtop
[-1].r
=get_reg_ex(RC_INT
,two2mask(vtop
->r
,vtop
[-1].r
)));
1781 o(opc
|r
|(c
<<12)|(fr
<<8)|0x10);
1786 vpush_global_sym(&func_old_type
, func
);
1793 tcc_error("gen_opi %i unimplemented!",op
);
1798 static int is_zero(int i
)
1800 if((vtop
[i
].r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) != VT_CONST
)
1802 if (vtop
[i
].type
.t
== VT_FLOAT
)
1803 return (vtop
[i
].c
.f
== 0.f
);
1804 else if (vtop
[i
].type
.t
== VT_DOUBLE
)
1805 return (vtop
[i
].c
.d
== 0.0);
1806 return (vtop
[i
].c
.ld
== 0.l
);
1809 /* generate a floating point operation 'v = t1 op t2' instruction. The
1810 * two operands are guaranteed to have the same floating point type */
1811 void gen_opf(int op
)
1815 x
=0xEE000A00|T2CPR(vtop
->type
.t
);
1833 x
|=0x810000; /* fsubX -> fnegX */
1846 if(op
< TOK_ULT
|| op
> TOK_GT
) {
1847 tcc_error("unknown fp op %x!",op
);
1853 case TOK_LT
: op
=TOK_GT
; break;
1854 case TOK_GE
: op
=TOK_ULE
; break;
1855 case TOK_LE
: op
=TOK_GE
; break;
1856 case TOK_GT
: op
=TOK_ULT
; break;
1859 x
|=0xB40040; /* fcmpX */
1860 if(op
!=TOK_EQ
&& op
!=TOK_NE
)
1861 x
|=0x80; /* fcmpX -> fcmpeX */
1864 o(x
|0x10000|(vfpr(gv(RC_FLOAT
))<<12)); /* fcmp(e)X -> fcmp(e)zX */
1866 gv2(RC_FLOAT
,RC_FLOAT
);
1868 o(x
|(vfpr(vtop
[-1].r
) << 12));
1871 o(0xEEF1FA10); /* fmstat */
1874 case TOK_LE
: op
=TOK_ULE
; break;
1875 case TOK_LT
: op
=TOK_ULT
; break;
1876 case TOK_UGE
: op
=TOK_GE
; break;
1877 case TOK_UGT
: op
=TOK_GT
; break;
1891 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
1898 vtop
->r
=get_reg_ex(RC_FLOAT
,r
);
1901 o(x
|(vfpr(vtop
->r
)<<12));
1905 static uint32_t is_fconst()
1909 if((vtop
->r
& (VT_VALMASK
| VT_LVAL
| VT_SYM
)) != VT_CONST
)
1911 if (vtop
->type
.t
== VT_FLOAT
)
1913 else if (vtop
->type
.t
== VT_DOUBLE
)
1943 /* generate a floating point operation 'v = t1 op t2' instruction. The
1944 two operands are guaranteed to have the same floating point type */
1945 void gen_opf(int op
)
1947 uint32_t x
, r
, r2
, c1
, c2
;
1948 //fputs("gen_opf\n",stderr);
1954 #if LDOUBLE_SIZE == 8
1955 if ((vtop
->type
.t
& VT_BTYPE
) != VT_FLOAT
)
1958 if ((vtop
->type
.t
& VT_BTYPE
) == VT_DOUBLE
)
1960 else if ((vtop
->type
.t
& VT_BTYPE
) == VT_LDOUBLE
)
1971 r
=fpr(gv(RC_FLOAT
));
1978 r2
=fpr(gv(RC_FLOAT
));
1979 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
1981 r
=fpr(gv(RC_FLOAT
));
1992 r
=fpr(gv(RC_FLOAT
));
1994 } else if(c1
&& c1
<=0xf) {
1997 r
=fpr(gv(RC_FLOAT
));
2002 r
=fpr(gv(RC_FLOAT
));
2004 r2
=fpr(gv(RC_FLOAT
));
2005 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
2007 r
=fpr(gv(RC_FLOAT
));
2018 r
=fpr(gv(RC_FLOAT
));
2023 r2
=fpr(gv(RC_FLOAT
));
2024 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
2026 r
=fpr(gv(RC_FLOAT
));
2037 r
=fpr(gv(RC_FLOAT
));
2039 } else if(c1
&& c1
<=0xf) {
2042 r
=fpr(gv(RC_FLOAT
));
2047 r
=fpr(gv(RC_FLOAT
));
2049 r2
=fpr(gv(RC_FLOAT
));
2050 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
2052 r
=fpr(gv(RC_FLOAT
));
2058 if(op
>= TOK_ULT
&& op
<= TOK_GT
) {
2059 x
|=0xd0f110; // cmfe
2060 /* bug (intention?) in Linux FPU emulator
2061 doesn't set carry if equal */
2067 tcc_error("unsigned comparison on floats?");
2073 op
=TOK_ULE
; /* correct in unordered case only if AC bit in FPSR set */
2077 x
&=~0x400000; // cmfe -> cmf
2099 r
=fpr(gv(RC_FLOAT
));
2106 r2
=fpr(gv(RC_FLOAT
));
2107 if ((vtop
[-1].r
& VT_VALMASK
) >= VT_CONST
) {
2109 r
=fpr(gv(RC_FLOAT
));
2117 tcc_error("unknown fp op %x!",op
);
2121 if(vtop
[-1].r
== VT_CMP
)
2127 vtop
[-1].r
=get_reg_ex(RC_FLOAT
,two2mask(vtop
[-1].r
,c1
));
2131 o(x
|(r
<<16)|(c1
<<12)|r2
);
2135 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
2136 and 'long long' cases. */
2137 ST_FUNC
void gen_cvt_itof(int t
)
2141 bt
=vtop
->type
.t
& VT_BTYPE
;
2142 if(bt
== VT_INT
|| bt
== VT_SHORT
|| bt
== VT_BYTE
) {
2148 r2
=vfpr(vtop
->r
=get_reg(RC_FLOAT
));
2149 o(0xEE000A10|(r
<<12)|(r2
<<16)); /* fmsr */
2151 if(!(vtop
->type
.t
& VT_UNSIGNED
))
2152 r2
|=0x80; /* fuitoX -> fsituX */
2153 o(0xEEB80A40|r2
|T2CPR(t
)); /* fYitoX*/
2155 r2
=fpr(vtop
->r
=get_reg(RC_FLOAT
));
2156 if((t
& VT_BTYPE
) != VT_FLOAT
)
2157 dsize
=0x80; /* flts -> fltd */
2158 o(0xEE000110|dsize
|(r2
<<16)|(r
<<12)); /* flts */
2159 if((vtop
->type
.t
& (VT_UNSIGNED
|VT_BTYPE
)) == (VT_UNSIGNED
|VT_INT
)) {
2161 o(0xE3500000|(r
<<12)); /* cmp */
2162 r
=fpr(get_reg(RC_FLOAT
));
2163 if(last_itod_magic
) {
2164 off
=ind
+8-last_itod_magic
;
2169 o(0xBD1F0100|(r
<<12)|off
); /* ldflts */
2171 o(0xEA000000); /* b */
2172 last_itod_magic
=ind
;
2173 o(0x4F800000); /* 4294967296.0f */
2175 o(0xBE000100|dsize
|(r2
<<16)|(r2
<<12)|r
); /* adflt */
2179 } else if(bt
== VT_LLONG
) {
2181 CType
*func_type
= 0;
2182 if((t
& VT_BTYPE
) == VT_FLOAT
) {
2183 func_type
= &func_float_type
;
2184 if(vtop
->type
.t
& VT_UNSIGNED
)
2185 func
=TOK___floatundisf
;
2187 func
=TOK___floatdisf
;
2188 #if LDOUBLE_SIZE != 8
2189 } else if((t
& VT_BTYPE
) == VT_LDOUBLE
) {
2190 func_type
= &func_ldouble_type
;
2191 if(vtop
->type
.t
& VT_UNSIGNED
)
2192 func
=TOK___floatundixf
;
2194 func
=TOK___floatdixf
;
2195 } else if((t
& VT_BTYPE
) == VT_DOUBLE
) {
2197 } else if((t
& VT_BTYPE
) == VT_DOUBLE
|| (t
& VT_BTYPE
) == VT_LDOUBLE
) {
2199 func_type
= &func_double_type
;
2200 if(vtop
->type
.t
& VT_UNSIGNED
)
2201 func
=TOK___floatundidf
;
2203 func
=TOK___floatdidf
;
2206 vpush_global_sym(func_type
, func
);
2214 tcc_error("unimplemented gen_cvt_itof %x!",vtop
->type
.t
);
2217 /* convert fp to int 't' type */
2218 void gen_cvt_ftoi(int t
)
2224 r2
=vtop
->type
.t
& VT_BTYPE
;
2227 r
=vfpr(gv(RC_FLOAT
));
2229 o(0xEEBC0AC0|(r
<<12)|r
|T2CPR(r2
)|u
); /* ftoXizY */
2230 r2
=intr(vtop
->r
=get_reg(RC_INT
));
2231 o(0xEE100A10|(r
<<16)|(r2
<<12));
2236 func
=TOK___fixunssfsi
;
2237 #if LDOUBLE_SIZE != 8
2238 else if(r2
== VT_LDOUBLE
)
2239 func
=TOK___fixunsxfsi
;
2240 else if(r2
== VT_DOUBLE
)
2242 else if(r2
== VT_LDOUBLE
|| r2
== VT_DOUBLE
)
2244 func
=TOK___fixunsdfsi
;
2246 r
=fpr(gv(RC_FLOAT
));
2247 r2
=intr(vtop
->r
=get_reg(RC_INT
));
2248 o(0xEE100170|(r2
<<12)|r
);
2252 } else if(t
== VT_LLONG
) { // unsigned handled in gen_cvt_ftoi1
2255 #if LDOUBLE_SIZE != 8
2256 else if(r2
== VT_LDOUBLE
)
2258 else if(r2
== VT_DOUBLE
)
2260 else if(r2
== VT_LDOUBLE
|| r2
== VT_DOUBLE
)
2265 vpush_global_sym(&func_old_type
, func
);
2270 vtop
->r2
= REG_IRE2
;
2274 tcc_error("unimplemented gen_cvt_ftoi!");
2277 /* convert from one floating point type to another */
2278 void gen_cvt_ftof(int t
)
2281 if(((vtop
->type
.t
& VT_BTYPE
) == VT_FLOAT
) != ((t
& VT_BTYPE
) == VT_FLOAT
)) {
2282 uint32_t r
= vfpr(gv(RC_FLOAT
));
2283 o(0xEEB70AC0|(r
<<12)|r
|T2CPR(vtop
->type
.t
));
2286 /* all we have to do on i386 and FPA ARM is to put the float in a register */
2291 /* computed goto support */
2298 /* Save the stack pointer onto the stack and return the location of its address */
2299 ST_FUNC
void gen_vla_sp_save(int addr
) {
2302 v
.r
= VT_LOCAL
| VT_LVAL
;
2307 /* Restore the SP from a location on the stack */
2308 ST_FUNC
void gen_vla_sp_restore(int addr
) {
2311 v
.r
= VT_LOCAL
| VT_LVAL
;
2316 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2317 ST_FUNC
void gen_vla_alloc(CType
*type
, int align
) {
2319 #if defined(CONFIG_TCC_BCHECK)
2320 if (tcc_state
->do_bounds_check
)
2323 r
= intr(gv(RC_INT
));
2324 o(0xE04D0000|(r
<<12)|r
); /* sub r, sp, r */
2332 if (align
& (align
- 1))
2333 tcc_error("alignment is not a power of 2: %i", align
);
2334 o(stuff_const(0xE3C0D000|(r
<<16), align
- 1)); /* bic sp, r, #align-1 */
2336 #if defined(CONFIG_TCC_BCHECK)
2337 if (tcc_state
->do_bounds_check
) {
2340 o(0xe1a0000d | (vtop
->r
<< 12)); // mov r0,sp
2342 vpush_global_sym(&func_old_type
, TOK___bound_new_region
);
2345 func_bound_add_epilog
= 1;
2350 /* end of ARM code generator */
2351 /*************************************************************/
2353 /*************************************************************/