1 /* Medium-level subroutines: convert bit-field store and extract
2 and shifts, multiplies and divides to rtl instructions.
3 Copyright (C) 1987-2014 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
26 #include "diagnostic-core.h"
29 #include "stor-layout.h"
32 #include "insn-config.h"
34 #include "insn-codes.h"
37 #include "langhooks.h"
39 #include "basic-block.h"
44 struct target_expmed default_target_expmed
;
46 struct target_expmed
*this_target_expmed
= &default_target_expmed
;
49 static void store_fixed_bit_field (rtx
, unsigned HOST_WIDE_INT
,
50 unsigned HOST_WIDE_INT
,
51 unsigned HOST_WIDE_INT
,
52 unsigned HOST_WIDE_INT
,
54 static void store_fixed_bit_field_1 (rtx
, unsigned HOST_WIDE_INT
,
55 unsigned HOST_WIDE_INT
,
57 static void store_split_bit_field (rtx
, unsigned HOST_WIDE_INT
,
58 unsigned HOST_WIDE_INT
,
59 unsigned HOST_WIDE_INT
,
60 unsigned HOST_WIDE_INT
,
62 static rtx
extract_fixed_bit_field (machine_mode
, rtx
,
63 unsigned HOST_WIDE_INT
,
64 unsigned HOST_WIDE_INT
, rtx
, int);
65 static rtx
extract_fixed_bit_field_1 (machine_mode
, rtx
,
66 unsigned HOST_WIDE_INT
,
67 unsigned HOST_WIDE_INT
, rtx
, int);
68 static rtx
lshift_value (machine_mode
, unsigned HOST_WIDE_INT
, int);
69 static rtx
extract_split_bit_field (rtx
, unsigned HOST_WIDE_INT
,
70 unsigned HOST_WIDE_INT
, int);
71 static void do_cmp_and_jump (rtx
, rtx
, enum rtx_code
, machine_mode
, rtx_code_label
*);
72 static rtx
expand_smod_pow2 (machine_mode
, rtx
, HOST_WIDE_INT
);
73 static rtx
expand_sdiv_pow2 (machine_mode
, rtx
, HOST_WIDE_INT
);
75 /* Return a constant integer mask value of mode MODE with BITSIZE ones
76 followed by BITPOS zeros, or the complement of that if COMPLEMENT.
77 The mask is truncated if necessary to the width of mode MODE. The
78 mask is zero-extended if BITSIZE+BITPOS is too small for MODE. */
81 mask_rtx (machine_mode mode
, int bitpos
, int bitsize
, bool complement
)
83 return immed_wide_int_const
84 (wi::shifted_mask (bitpos
, bitsize
, complement
,
85 GET_MODE_PRECISION (mode
)), mode
);
88 /* Test whether a value is zero of a power of two. */
89 #define EXACT_POWER_OF_2_OR_ZERO_P(x) \
90 (((x) & ((x) - (unsigned HOST_WIDE_INT) 1)) == 0)
92 struct init_expmed_rtl
113 rtx pow2
[MAX_BITS_PER_WORD
];
114 rtx cint
[MAX_BITS_PER_WORD
];
118 init_expmed_one_conv (struct init_expmed_rtl
*all
, machine_mode to_mode
,
119 machine_mode from_mode
, bool speed
)
121 int to_size
, from_size
;
124 to_size
= GET_MODE_PRECISION (to_mode
);
125 from_size
= GET_MODE_PRECISION (from_mode
);
127 /* Most partial integers have a precision less than the "full"
128 integer it requires for storage. In case one doesn't, for
129 comparison purposes here, reduce the bit size by one in that
131 if (GET_MODE_CLASS (to_mode
) == MODE_PARTIAL_INT
132 && exact_log2 (to_size
) != -1)
134 if (GET_MODE_CLASS (from_mode
) == MODE_PARTIAL_INT
135 && exact_log2 (from_size
) != -1)
138 /* Assume cost of zero-extend and sign-extend is the same. */
139 which
= (to_size
< from_size
? all
->trunc
: all
->zext
);
141 PUT_MODE (all
->reg
, from_mode
);
142 set_convert_cost (to_mode
, from_mode
, speed
, set_src_cost (which
, speed
));
146 init_expmed_one_mode (struct init_expmed_rtl
*all
,
147 machine_mode mode
, int speed
)
149 int m
, n
, mode_bitsize
;
150 machine_mode mode_from
;
152 mode_bitsize
= GET_MODE_UNIT_BITSIZE (mode
);
154 PUT_MODE (all
->reg
, mode
);
155 PUT_MODE (all
->plus
, mode
);
156 PUT_MODE (all
->neg
, mode
);
157 PUT_MODE (all
->mult
, mode
);
158 PUT_MODE (all
->sdiv
, mode
);
159 PUT_MODE (all
->udiv
, mode
);
160 PUT_MODE (all
->sdiv_32
, mode
);
161 PUT_MODE (all
->smod_32
, mode
);
162 PUT_MODE (all
->wide_trunc
, mode
);
163 PUT_MODE (all
->shift
, mode
);
164 PUT_MODE (all
->shift_mult
, mode
);
165 PUT_MODE (all
->shift_add
, mode
);
166 PUT_MODE (all
->shift_sub0
, mode
);
167 PUT_MODE (all
->shift_sub1
, mode
);
168 PUT_MODE (all
->zext
, mode
);
169 PUT_MODE (all
->trunc
, mode
);
171 set_add_cost (speed
, mode
, set_src_cost (all
->plus
, speed
));
172 set_neg_cost (speed
, mode
, set_src_cost (all
->neg
, speed
));
173 set_mul_cost (speed
, mode
, set_src_cost (all
->mult
, speed
));
174 set_sdiv_cost (speed
, mode
, set_src_cost (all
->sdiv
, speed
));
175 set_udiv_cost (speed
, mode
, set_src_cost (all
->udiv
, speed
));
177 set_sdiv_pow2_cheap (speed
, mode
, (set_src_cost (all
->sdiv_32
, speed
)
178 <= 2 * add_cost (speed
, mode
)));
179 set_smod_pow2_cheap (speed
, mode
, (set_src_cost (all
->smod_32
, speed
)
180 <= 4 * add_cost (speed
, mode
)));
182 set_shift_cost (speed
, mode
, 0, 0);
184 int cost
= add_cost (speed
, mode
);
185 set_shiftadd_cost (speed
, mode
, 0, cost
);
186 set_shiftsub0_cost (speed
, mode
, 0, cost
);
187 set_shiftsub1_cost (speed
, mode
, 0, cost
);
190 n
= MIN (MAX_BITS_PER_WORD
, mode_bitsize
);
191 for (m
= 1; m
< n
; m
++)
193 XEXP (all
->shift
, 1) = all
->cint
[m
];
194 XEXP (all
->shift_mult
, 1) = all
->pow2
[m
];
196 set_shift_cost (speed
, mode
, m
, set_src_cost (all
->shift
, speed
));
197 set_shiftadd_cost (speed
, mode
, m
, set_src_cost (all
->shift_add
, speed
));
198 set_shiftsub0_cost (speed
, mode
, m
, set_src_cost (all
->shift_sub0
, speed
));
199 set_shiftsub1_cost (speed
, mode
, m
, set_src_cost (all
->shift_sub1
, speed
));
202 if (SCALAR_INT_MODE_P (mode
))
204 for (mode_from
= MIN_MODE_INT
; mode_from
<= MAX_MODE_INT
;
205 mode_from
= (machine_mode
)(mode_from
+ 1))
206 init_expmed_one_conv (all
, mode
, mode_from
, speed
);
208 if (GET_MODE_CLASS (mode
) == MODE_INT
)
210 machine_mode wider_mode
= GET_MODE_WIDER_MODE (mode
);
211 if (wider_mode
!= VOIDmode
)
213 PUT_MODE (all
->zext
, wider_mode
);
214 PUT_MODE (all
->wide_mult
, wider_mode
);
215 PUT_MODE (all
->wide_lshr
, wider_mode
);
216 XEXP (all
->wide_lshr
, 1) = GEN_INT (mode_bitsize
);
218 set_mul_widen_cost (speed
, wider_mode
,
219 set_src_cost (all
->wide_mult
, speed
));
220 set_mul_highpart_cost (speed
, mode
,
221 set_src_cost (all
->wide_trunc
, speed
));
229 struct init_expmed_rtl all
;
230 machine_mode mode
= QImode
;
233 memset (&all
, 0, sizeof all
);
234 for (m
= 1; m
< MAX_BITS_PER_WORD
; m
++)
236 all
.pow2
[m
] = GEN_INT ((HOST_WIDE_INT
) 1 << m
);
237 all
.cint
[m
] = GEN_INT (m
);
240 /* Avoid using hard regs in ways which may be unsupported. */
241 all
.reg
= gen_rtx_raw_REG (mode
, LAST_VIRTUAL_REGISTER
+ 1);
242 all
.plus
= gen_rtx_PLUS (mode
, all
.reg
, all
.reg
);
243 all
.neg
= gen_rtx_NEG (mode
, all
.reg
);
244 all
.mult
= gen_rtx_MULT (mode
, all
.reg
, all
.reg
);
245 all
.sdiv
= gen_rtx_DIV (mode
, all
.reg
, all
.reg
);
246 all
.udiv
= gen_rtx_UDIV (mode
, all
.reg
, all
.reg
);
247 all
.sdiv_32
= gen_rtx_DIV (mode
, all
.reg
, all
.pow2
[5]);
248 all
.smod_32
= gen_rtx_MOD (mode
, all
.reg
, all
.pow2
[5]);
249 all
.zext
= gen_rtx_ZERO_EXTEND (mode
, all
.reg
);
250 all
.wide_mult
= gen_rtx_MULT (mode
, all
.zext
, all
.zext
);
251 all
.wide_lshr
= gen_rtx_LSHIFTRT (mode
, all
.wide_mult
, all
.reg
);
252 all
.wide_trunc
= gen_rtx_TRUNCATE (mode
, all
.wide_lshr
);
253 all
.shift
= gen_rtx_ASHIFT (mode
, all
.reg
, all
.reg
);
254 all
.shift_mult
= gen_rtx_MULT (mode
, all
.reg
, all
.reg
);
255 all
.shift_add
= gen_rtx_PLUS (mode
, all
.shift_mult
, all
.reg
);
256 all
.shift_sub0
= gen_rtx_MINUS (mode
, all
.shift_mult
, all
.reg
);
257 all
.shift_sub1
= gen_rtx_MINUS (mode
, all
.reg
, all
.shift_mult
);
258 all
.trunc
= gen_rtx_TRUNCATE (mode
, all
.reg
);
260 for (speed
= 0; speed
< 2; speed
++)
262 crtl
->maybe_hot_insn_p
= speed
;
263 set_zero_cost (speed
, set_src_cost (const0_rtx
, speed
));
265 for (mode
= MIN_MODE_INT
; mode
<= MAX_MODE_INT
;
266 mode
= (machine_mode
)(mode
+ 1))
267 init_expmed_one_mode (&all
, mode
, speed
);
269 if (MIN_MODE_PARTIAL_INT
!= VOIDmode
)
270 for (mode
= MIN_MODE_PARTIAL_INT
; mode
<= MAX_MODE_PARTIAL_INT
;
271 mode
= (machine_mode
)(mode
+ 1))
272 init_expmed_one_mode (&all
, mode
, speed
);
274 if (MIN_MODE_VECTOR_INT
!= VOIDmode
)
275 for (mode
= MIN_MODE_VECTOR_INT
; mode
<= MAX_MODE_VECTOR_INT
;
276 mode
= (machine_mode
)(mode
+ 1))
277 init_expmed_one_mode (&all
, mode
, speed
);
280 if (alg_hash_used_p ())
282 struct alg_hash_entry
*p
= alg_hash_entry_ptr (0);
283 memset (p
, 0, sizeof (*p
) * NUM_ALG_HASH_ENTRIES
);
286 set_alg_hash_used_p (true);
287 default_rtl_profile ();
289 ggc_free (all
.trunc
);
290 ggc_free (all
.shift_sub1
);
291 ggc_free (all
.shift_sub0
);
292 ggc_free (all
.shift_add
);
293 ggc_free (all
.shift_mult
);
294 ggc_free (all
.shift
);
295 ggc_free (all
.wide_trunc
);
296 ggc_free (all
.wide_lshr
);
297 ggc_free (all
.wide_mult
);
299 ggc_free (all
.smod_32
);
300 ggc_free (all
.sdiv_32
);
309 /* Return an rtx representing minus the value of X.
310 MODE is the intended mode of the result,
311 useful if X is a CONST_INT. */
314 negate_rtx (machine_mode mode
, rtx x
)
316 rtx result
= simplify_unary_operation (NEG
, mode
, x
, mode
);
319 result
= expand_unop (mode
, neg_optab
, x
, NULL_RTX
, 0);
324 /* Adjust bitfield memory MEM so that it points to the first unit of mode
325 MODE that contains a bitfield of size BITSIZE at bit position BITNUM.
326 If MODE is BLKmode, return a reference to every byte in the bitfield.
327 Set *NEW_BITNUM to the bit position of the field within the new memory. */
330 narrow_bit_field_mem (rtx mem
, machine_mode mode
,
331 unsigned HOST_WIDE_INT bitsize
,
332 unsigned HOST_WIDE_INT bitnum
,
333 unsigned HOST_WIDE_INT
*new_bitnum
)
337 *new_bitnum
= bitnum
% BITS_PER_UNIT
;
338 HOST_WIDE_INT offset
= bitnum
/ BITS_PER_UNIT
;
339 HOST_WIDE_INT size
= ((*new_bitnum
+ bitsize
+ BITS_PER_UNIT
- 1)
341 return adjust_bitfield_address_size (mem
, mode
, offset
, size
);
345 unsigned int unit
= GET_MODE_BITSIZE (mode
);
346 *new_bitnum
= bitnum
% unit
;
347 HOST_WIDE_INT offset
= (bitnum
- *new_bitnum
) / BITS_PER_UNIT
;
348 return adjust_bitfield_address (mem
, mode
, offset
);
352 /* The caller wants to perform insertion or extraction PATTERN on a
353 bitfield of size BITSIZE at BITNUM bits into memory operand OP0.
354 BITREGION_START and BITREGION_END are as for store_bit_field
355 and FIELDMODE is the natural mode of the field.
357 Search for a mode that is compatible with the memory access
358 restrictions and (where applicable) with a register insertion or
359 extraction. Return the new memory on success, storing the adjusted
360 bit position in *NEW_BITNUM. Return null otherwise. */
363 adjust_bit_field_mem_for_reg (enum extraction_pattern pattern
,
364 rtx op0
, HOST_WIDE_INT bitsize
,
365 HOST_WIDE_INT bitnum
,
366 unsigned HOST_WIDE_INT bitregion_start
,
367 unsigned HOST_WIDE_INT bitregion_end
,
368 machine_mode fieldmode
,
369 unsigned HOST_WIDE_INT
*new_bitnum
)
371 bit_field_mode_iterator
iter (bitsize
, bitnum
, bitregion_start
,
372 bitregion_end
, MEM_ALIGN (op0
),
373 MEM_VOLATILE_P (op0
));
374 machine_mode best_mode
;
375 if (iter
.next_mode (&best_mode
))
377 /* We can use a memory in BEST_MODE. See whether this is true for
378 any wider modes. All other things being equal, we prefer to
379 use the widest mode possible because it tends to expose more
380 CSE opportunities. */
381 if (!iter
.prefer_smaller_modes ())
383 /* Limit the search to the mode required by the corresponding
384 register insertion or extraction instruction, if any. */
385 machine_mode limit_mode
= word_mode
;
386 extraction_insn insn
;
387 if (get_best_reg_extraction_insn (&insn
, pattern
,
388 GET_MODE_BITSIZE (best_mode
),
390 limit_mode
= insn
.field_mode
;
392 machine_mode wider_mode
;
393 while (iter
.next_mode (&wider_mode
)
394 && GET_MODE_SIZE (wider_mode
) <= GET_MODE_SIZE (limit_mode
))
395 best_mode
= wider_mode
;
397 return narrow_bit_field_mem (op0
, best_mode
, bitsize
, bitnum
,
403 /* Return true if a bitfield of size BITSIZE at bit number BITNUM within
404 a structure of mode STRUCT_MODE represents a lowpart subreg. The subreg
405 offset is then BITNUM / BITS_PER_UNIT. */
408 lowpart_bit_field_p (unsigned HOST_WIDE_INT bitnum
,
409 unsigned HOST_WIDE_INT bitsize
,
410 machine_mode struct_mode
)
412 if (BYTES_BIG_ENDIAN
)
413 return (bitnum
% BITS_PER_UNIT
== 0
414 && (bitnum
+ bitsize
== GET_MODE_BITSIZE (struct_mode
)
415 || (bitnum
+ bitsize
) % BITS_PER_WORD
== 0));
417 return bitnum
% BITS_PER_WORD
== 0;
420 /* Return true if -fstrict-volatile-bitfields applies to an access of OP0
421 containing BITSIZE bits starting at BITNUM, with field mode FIELDMODE.
422 Return false if the access would touch memory outside the range
423 BITREGION_START to BITREGION_END for conformance to the C++ memory
427 strict_volatile_bitfield_p (rtx op0
, unsigned HOST_WIDE_INT bitsize
,
428 unsigned HOST_WIDE_INT bitnum
,
429 machine_mode fieldmode
,
430 unsigned HOST_WIDE_INT bitregion_start
,
431 unsigned HOST_WIDE_INT bitregion_end
)
433 unsigned HOST_WIDE_INT modesize
= GET_MODE_BITSIZE (fieldmode
);
435 /* -fstrict-volatile-bitfields must be enabled and we must have a
438 || !MEM_VOLATILE_P (op0
)
439 || flag_strict_volatile_bitfields
<= 0)
442 /* Non-integral modes likely only happen with packed structures.
444 if (!SCALAR_INT_MODE_P (fieldmode
))
447 /* The bit size must not be larger than the field mode, and
448 the field mode must not be larger than a word. */
449 if (bitsize
> modesize
|| modesize
> BITS_PER_WORD
)
452 /* Check for cases of unaligned fields that must be split. */
453 if (bitnum
% BITS_PER_UNIT
+ bitsize
> modesize
455 && bitnum
% GET_MODE_ALIGNMENT (fieldmode
) + bitsize
> modesize
))
458 /* Check for cases where the C++ memory model applies. */
459 if (bitregion_end
!= 0
460 && (bitnum
- bitnum
% modesize
< bitregion_start
461 || bitnum
- bitnum
% modesize
+ modesize
- 1 > bitregion_end
))
467 /* Return true if OP is a memory and if a bitfield of size BITSIZE at
468 bit number BITNUM can be treated as a simple value of mode MODE. */
471 simple_mem_bitfield_p (rtx op0
, unsigned HOST_WIDE_INT bitsize
,
472 unsigned HOST_WIDE_INT bitnum
, machine_mode mode
)
475 && bitnum
% BITS_PER_UNIT
== 0
476 && bitsize
== GET_MODE_BITSIZE (mode
)
477 && (!SLOW_UNALIGNED_ACCESS (mode
, MEM_ALIGN (op0
))
478 || (bitnum
% GET_MODE_ALIGNMENT (mode
) == 0
479 && MEM_ALIGN (op0
) >= GET_MODE_ALIGNMENT (mode
))));
482 /* Try to use instruction INSV to store VALUE into a field of OP0.
483 BITSIZE and BITNUM are as for store_bit_field. */
486 store_bit_field_using_insv (const extraction_insn
*insv
, rtx op0
,
487 unsigned HOST_WIDE_INT bitsize
,
488 unsigned HOST_WIDE_INT bitnum
,
491 struct expand_operand ops
[4];
494 rtx_insn
*last
= get_last_insn ();
495 bool copy_back
= false;
497 machine_mode op_mode
= insv
->field_mode
;
498 unsigned int unit
= GET_MODE_BITSIZE (op_mode
);
499 if (bitsize
== 0 || bitsize
> unit
)
503 /* Get a reference to the first byte of the field. */
504 xop0
= narrow_bit_field_mem (xop0
, insv
->struct_mode
, bitsize
, bitnum
,
508 /* Convert from counting within OP0 to counting in OP_MODE. */
509 if (BYTES_BIG_ENDIAN
)
510 bitnum
+= unit
- GET_MODE_BITSIZE (GET_MODE (op0
));
512 /* If xop0 is a register, we need it in OP_MODE
513 to make it acceptable to the format of insv. */
514 if (GET_CODE (xop0
) == SUBREG
)
515 /* We can't just change the mode, because this might clobber op0,
516 and we will need the original value of op0 if insv fails. */
517 xop0
= gen_rtx_SUBREG (op_mode
, SUBREG_REG (xop0
), SUBREG_BYTE (xop0
));
518 if (REG_P (xop0
) && GET_MODE (xop0
) != op_mode
)
519 xop0
= gen_lowpart_SUBREG (op_mode
, xop0
);
522 /* If the destination is a paradoxical subreg such that we need a
523 truncate to the inner mode, perform the insertion on a temporary and
524 truncate the result to the original destination. Note that we can't
525 just truncate the paradoxical subreg as (truncate:N (subreg:W (reg:N
526 X) 0)) is (reg:N X). */
527 if (GET_CODE (xop0
) == SUBREG
528 && REG_P (SUBREG_REG (xop0
))
529 && !TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (SUBREG_REG (xop0
)),
532 rtx tem
= gen_reg_rtx (op_mode
);
533 emit_move_insn (tem
, xop0
);
538 /* If BITS_BIG_ENDIAN is zero on a BYTES_BIG_ENDIAN machine, we count
539 "backwards" from the size of the unit we are inserting into.
540 Otherwise, we count bits from the most significant on a
541 BYTES/BITS_BIG_ENDIAN machine. */
543 if (BITS_BIG_ENDIAN
!= BYTES_BIG_ENDIAN
)
544 bitnum
= unit
- bitsize
- bitnum
;
546 /* Convert VALUE to op_mode (which insv insn wants) in VALUE1. */
548 if (GET_MODE (value
) != op_mode
)
550 if (GET_MODE_BITSIZE (GET_MODE (value
)) >= bitsize
)
552 /* Optimization: Don't bother really extending VALUE
553 if it has all the bits we will actually use. However,
554 if we must narrow it, be sure we do it correctly. */
556 if (GET_MODE_SIZE (GET_MODE (value
)) < GET_MODE_SIZE (op_mode
))
560 tmp
= simplify_subreg (op_mode
, value1
, GET_MODE (value
), 0);
562 tmp
= simplify_gen_subreg (op_mode
,
563 force_reg (GET_MODE (value
),
565 GET_MODE (value
), 0);
569 value1
= gen_lowpart (op_mode
, value1
);
571 else if (CONST_INT_P (value
))
572 value1
= gen_int_mode (INTVAL (value
), op_mode
);
574 /* Parse phase is supposed to make VALUE's data type
575 match that of the component reference, which is a type
576 at least as wide as the field; so VALUE should have
577 a mode that corresponds to that type. */
578 gcc_assert (CONSTANT_P (value
));
581 create_fixed_operand (&ops
[0], xop0
);
582 create_integer_operand (&ops
[1], bitsize
);
583 create_integer_operand (&ops
[2], bitnum
);
584 create_input_operand (&ops
[3], value1
, op_mode
);
585 if (maybe_expand_insn (insv
->icode
, 4, ops
))
588 convert_move (op0
, xop0
, true);
591 delete_insns_since (last
);
595 /* A subroutine of store_bit_field, with the same arguments. Return true
596 if the operation could be implemented.
598 If FALLBACK_P is true, fall back to store_fixed_bit_field if we have
599 no other way of implementing the operation. If FALLBACK_P is false,
600 return false instead. */
603 store_bit_field_1 (rtx str_rtx
, unsigned HOST_WIDE_INT bitsize
,
604 unsigned HOST_WIDE_INT bitnum
,
605 unsigned HOST_WIDE_INT bitregion_start
,
606 unsigned HOST_WIDE_INT bitregion_end
,
607 machine_mode fieldmode
,
608 rtx value
, bool fallback_p
)
613 while (GET_CODE (op0
) == SUBREG
)
615 /* The following line once was done only if WORDS_BIG_ENDIAN,
616 but I think that is a mistake. WORDS_BIG_ENDIAN is
617 meaningful at a much higher level; when structures are copied
618 between memory and regs, the higher-numbered regs
619 always get higher addresses. */
620 int inner_mode_size
= GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0
)));
621 int outer_mode_size
= GET_MODE_SIZE (GET_MODE (op0
));
624 /* Paradoxical subregs need special handling on big endian machines. */
625 if (SUBREG_BYTE (op0
) == 0 && inner_mode_size
< outer_mode_size
)
627 int difference
= inner_mode_size
- outer_mode_size
;
629 if (WORDS_BIG_ENDIAN
)
630 byte_offset
+= (difference
/ UNITS_PER_WORD
) * UNITS_PER_WORD
;
631 if (BYTES_BIG_ENDIAN
)
632 byte_offset
+= difference
% UNITS_PER_WORD
;
635 byte_offset
= SUBREG_BYTE (op0
);
637 bitnum
+= byte_offset
* BITS_PER_UNIT
;
638 op0
= SUBREG_REG (op0
);
641 /* No action is needed if the target is a register and if the field
642 lies completely outside that register. This can occur if the source
643 code contains an out-of-bounds access to a small array. */
644 if (REG_P (op0
) && bitnum
>= GET_MODE_BITSIZE (GET_MODE (op0
)))
647 /* Use vec_set patterns for inserting parts of vectors whenever
649 if (VECTOR_MODE_P (GET_MODE (op0
))
651 && optab_handler (vec_set_optab
, GET_MODE (op0
)) != CODE_FOR_nothing
652 && fieldmode
== GET_MODE_INNER (GET_MODE (op0
))
653 && bitsize
== GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0
)))
654 && !(bitnum
% GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0
)))))
656 struct expand_operand ops
[3];
657 machine_mode outermode
= GET_MODE (op0
);
658 machine_mode innermode
= GET_MODE_INNER (outermode
);
659 enum insn_code icode
= optab_handler (vec_set_optab
, outermode
);
660 int pos
= bitnum
/ GET_MODE_BITSIZE (innermode
);
662 create_fixed_operand (&ops
[0], op0
);
663 create_input_operand (&ops
[1], value
, innermode
);
664 create_integer_operand (&ops
[2], pos
);
665 if (maybe_expand_insn (icode
, 3, ops
))
669 /* If the target is a register, overwriting the entire object, or storing
670 a full-word or multi-word field can be done with just a SUBREG. */
672 && bitsize
== GET_MODE_BITSIZE (fieldmode
)
673 && ((bitsize
== GET_MODE_BITSIZE (GET_MODE (op0
)) && bitnum
== 0)
674 || (bitsize
% BITS_PER_WORD
== 0 && bitnum
% BITS_PER_WORD
== 0)))
676 /* Use the subreg machinery either to narrow OP0 to the required
677 words or to cope with mode punning between equal-sized modes.
678 In the latter case, use subreg on the rhs side, not lhs. */
681 if (bitsize
== GET_MODE_BITSIZE (GET_MODE (op0
)))
683 sub
= simplify_gen_subreg (GET_MODE (op0
), value
, fieldmode
, 0);
686 emit_move_insn (op0
, sub
);
692 sub
= simplify_gen_subreg (fieldmode
, op0
, GET_MODE (op0
),
693 bitnum
/ BITS_PER_UNIT
);
696 emit_move_insn (sub
, value
);
702 /* If the target is memory, storing any naturally aligned field can be
703 done with a simple store. For targets that support fast unaligned
704 memory, any naturally sized, unit aligned field can be done directly. */
705 if (simple_mem_bitfield_p (op0
, bitsize
, bitnum
, fieldmode
))
707 op0
= adjust_bitfield_address (op0
, fieldmode
, bitnum
/ BITS_PER_UNIT
);
708 emit_move_insn (op0
, value
);
712 /* Make sure we are playing with integral modes. Pun with subregs
713 if we aren't. This must come after the entire register case above,
714 since that case is valid for any mode. The following cases are only
715 valid for integral modes. */
717 machine_mode imode
= int_mode_for_mode (GET_MODE (op0
));
718 if (imode
!= GET_MODE (op0
))
721 op0
= adjust_bitfield_address_size (op0
, imode
, 0, MEM_SIZE (op0
));
724 gcc_assert (imode
!= BLKmode
);
725 op0
= gen_lowpart (imode
, op0
);
730 /* Storing an lsb-aligned field in a register
731 can be done with a movstrict instruction. */
734 && lowpart_bit_field_p (bitnum
, bitsize
, GET_MODE (op0
))
735 && bitsize
== GET_MODE_BITSIZE (fieldmode
)
736 && optab_handler (movstrict_optab
, fieldmode
) != CODE_FOR_nothing
)
738 struct expand_operand ops
[2];
739 enum insn_code icode
= optab_handler (movstrict_optab
, fieldmode
);
741 unsigned HOST_WIDE_INT subreg_off
;
743 if (GET_CODE (arg0
) == SUBREG
)
745 /* Else we've got some float mode source being extracted into
746 a different float mode destination -- this combination of
747 subregs results in Severe Tire Damage. */
748 gcc_assert (GET_MODE (SUBREG_REG (arg0
)) == fieldmode
749 || GET_MODE_CLASS (fieldmode
) == MODE_INT
750 || GET_MODE_CLASS (fieldmode
) == MODE_PARTIAL_INT
);
751 arg0
= SUBREG_REG (arg0
);
754 subreg_off
= bitnum
/ BITS_PER_UNIT
;
755 if (validate_subreg (fieldmode
, GET_MODE (arg0
), arg0
, subreg_off
))
757 arg0
= gen_rtx_SUBREG (fieldmode
, arg0
, subreg_off
);
759 create_fixed_operand (&ops
[0], arg0
);
760 /* Shrink the source operand to FIELDMODE. */
761 create_convert_operand_to (&ops
[1], value
, fieldmode
, false);
762 if (maybe_expand_insn (icode
, 2, ops
))
767 /* Handle fields bigger than a word. */
769 if (bitsize
> BITS_PER_WORD
)
771 /* Here we transfer the words of the field
772 in the order least significant first.
773 This is because the most significant word is the one which may
775 However, only do that if the value is not BLKmode. */
777 unsigned int backwards
= WORDS_BIG_ENDIAN
&& fieldmode
!= BLKmode
;
778 unsigned int nwords
= (bitsize
+ (BITS_PER_WORD
- 1)) / BITS_PER_WORD
;
782 /* This is the mode we must force value to, so that there will be enough
783 subwords to extract. Note that fieldmode will often (always?) be
784 VOIDmode, because that is what store_field uses to indicate that this
785 is a bit field, but passing VOIDmode to operand_subword_force
787 fieldmode
= GET_MODE (value
);
788 if (fieldmode
== VOIDmode
)
789 fieldmode
= smallest_mode_for_size (nwords
* BITS_PER_WORD
, MODE_INT
);
791 last
= get_last_insn ();
792 for (i
= 0; i
< nwords
; i
++)
794 /* If I is 0, use the low-order word in both field and target;
795 if I is 1, use the next to lowest word; and so on. */
796 unsigned int wordnum
= (backwards
797 ? GET_MODE_SIZE (fieldmode
) / UNITS_PER_WORD
800 unsigned int bit_offset
= (backwards
801 ? MAX ((int) bitsize
- ((int) i
+ 1)
804 : (int) i
* BITS_PER_WORD
);
805 rtx value_word
= operand_subword_force (value
, wordnum
, fieldmode
);
806 unsigned HOST_WIDE_INT new_bitsize
=
807 MIN (BITS_PER_WORD
, bitsize
- i
* BITS_PER_WORD
);
809 /* If the remaining chunk doesn't have full wordsize we have
810 to make sure that for big endian machines the higher order
812 if (new_bitsize
< BITS_PER_WORD
&& BYTES_BIG_ENDIAN
&& !backwards
)
813 value_word
= simplify_expand_binop (word_mode
, lshr_optab
,
815 GEN_INT (BITS_PER_WORD
820 if (!store_bit_field_1 (op0
, new_bitsize
,
822 bitregion_start
, bitregion_end
,
824 value_word
, fallback_p
))
826 delete_insns_since (last
);
833 /* If VALUE has a floating-point or complex mode, access it as an
834 integer of the corresponding size. This can occur on a machine
835 with 64 bit registers that uses SFmode for float. It can also
836 occur for unaligned float or complex fields. */
838 if (GET_MODE (value
) != VOIDmode
839 && GET_MODE_CLASS (GET_MODE (value
)) != MODE_INT
840 && GET_MODE_CLASS (GET_MODE (value
)) != MODE_PARTIAL_INT
)
842 value
= gen_reg_rtx (int_mode_for_mode (GET_MODE (value
)));
843 emit_move_insn (gen_lowpart (GET_MODE (orig_value
), value
), orig_value
);
846 /* If OP0 is a multi-word register, narrow it to the affected word.
847 If the region spans two words, defer to store_split_bit_field. */
848 if (!MEM_P (op0
) && GET_MODE_SIZE (GET_MODE (op0
)) > UNITS_PER_WORD
)
850 op0
= simplify_gen_subreg (word_mode
, op0
, GET_MODE (op0
),
851 bitnum
/ BITS_PER_WORD
* UNITS_PER_WORD
);
853 bitnum
%= BITS_PER_WORD
;
854 if (bitnum
+ bitsize
> BITS_PER_WORD
)
859 store_split_bit_field (op0
, bitsize
, bitnum
, bitregion_start
,
860 bitregion_end
, value
);
865 /* From here on we can assume that the field to be stored in fits
866 within a word. If the destination is a register, it too fits
869 extraction_insn insv
;
871 && get_best_reg_extraction_insn (&insv
, EP_insv
,
872 GET_MODE_BITSIZE (GET_MODE (op0
)),
874 && store_bit_field_using_insv (&insv
, op0
, bitsize
, bitnum
, value
))
877 /* If OP0 is a memory, try copying it to a register and seeing if a
878 cheap register alternative is available. */
881 if (get_best_mem_extraction_insn (&insv
, EP_insv
, bitsize
, bitnum
,
883 && store_bit_field_using_insv (&insv
, op0
, bitsize
, bitnum
, value
))
886 rtx_insn
*last
= get_last_insn ();
888 /* Try loading part of OP0 into a register, inserting the bitfield
889 into that, and then copying the result back to OP0. */
890 unsigned HOST_WIDE_INT bitpos
;
891 rtx xop0
= adjust_bit_field_mem_for_reg (EP_insv
, op0
, bitsize
, bitnum
,
892 bitregion_start
, bitregion_end
,
896 rtx tempreg
= copy_to_reg (xop0
);
897 if (store_bit_field_1 (tempreg
, bitsize
, bitpos
,
898 bitregion_start
, bitregion_end
,
899 fieldmode
, orig_value
, false))
901 emit_move_insn (xop0
, tempreg
);
904 delete_insns_since (last
);
911 store_fixed_bit_field (op0
, bitsize
, bitnum
, bitregion_start
,
912 bitregion_end
, value
);
916 /* Generate code to store value from rtx VALUE
917 into a bit-field within structure STR_RTX
918 containing BITSIZE bits starting at bit BITNUM.
920 BITREGION_START is bitpos of the first bitfield in this region.
921 BITREGION_END is the bitpos of the ending bitfield in this region.
922 These two fields are 0, if the C++ memory model does not apply,
923 or we are not interested in keeping track of bitfield regions.
925 FIELDMODE is the machine-mode of the FIELD_DECL node for this field. */
928 store_bit_field (rtx str_rtx
, unsigned HOST_WIDE_INT bitsize
,
929 unsigned HOST_WIDE_INT bitnum
,
930 unsigned HOST_WIDE_INT bitregion_start
,
931 unsigned HOST_WIDE_INT bitregion_end
,
932 machine_mode fieldmode
,
935 /* Handle -fstrict-volatile-bitfields in the cases where it applies. */
936 if (strict_volatile_bitfield_p (str_rtx
, bitsize
, bitnum
, fieldmode
,
937 bitregion_start
, bitregion_end
))
939 /* Storing any naturally aligned field can be done with a simple
940 store. For targets that support fast unaligned memory, any
941 naturally sized, unit aligned field can be done directly. */
942 if (simple_mem_bitfield_p (str_rtx
, bitsize
, bitnum
, fieldmode
))
944 str_rtx
= adjust_bitfield_address (str_rtx
, fieldmode
,
945 bitnum
/ BITS_PER_UNIT
);
946 emit_move_insn (str_rtx
, value
);
950 str_rtx
= narrow_bit_field_mem (str_rtx
, fieldmode
, bitsize
, bitnum
,
952 /* Explicitly override the C/C++ memory model; ignore the
953 bit range so that we can do the access in the mode mandated
954 by -fstrict-volatile-bitfields instead. */
955 store_fixed_bit_field_1 (str_rtx
, bitsize
, bitnum
, value
);
961 /* Under the C++0x memory model, we must not touch bits outside the
962 bit region. Adjust the address to start at the beginning of the
964 if (MEM_P (str_rtx
) && bitregion_start
> 0)
966 machine_mode bestmode
;
967 HOST_WIDE_INT offset
, size
;
969 gcc_assert ((bitregion_start
% BITS_PER_UNIT
) == 0);
971 offset
= bitregion_start
/ BITS_PER_UNIT
;
972 bitnum
-= bitregion_start
;
973 size
= (bitnum
+ bitsize
+ BITS_PER_UNIT
- 1) / BITS_PER_UNIT
;
974 bitregion_end
-= bitregion_start
;
976 bestmode
= get_best_mode (bitsize
, bitnum
,
977 bitregion_start
, bitregion_end
,
978 MEM_ALIGN (str_rtx
), VOIDmode
,
979 MEM_VOLATILE_P (str_rtx
));
980 str_rtx
= adjust_bitfield_address_size (str_rtx
, bestmode
, offset
, size
);
983 if (!store_bit_field_1 (str_rtx
, bitsize
, bitnum
,
984 bitregion_start
, bitregion_end
,
985 fieldmode
, value
, true))
989 /* Use shifts and boolean operations to store VALUE into a bit field of
990 width BITSIZE in OP0, starting at bit BITNUM. */
993 store_fixed_bit_field (rtx op0
, unsigned HOST_WIDE_INT bitsize
,
994 unsigned HOST_WIDE_INT bitnum
,
995 unsigned HOST_WIDE_INT bitregion_start
,
996 unsigned HOST_WIDE_INT bitregion_end
,
999 /* There is a case not handled here:
1000 a structure with a known alignment of just a halfword
1001 and a field split across two aligned halfwords within the structure.
1002 Or likewise a structure with a known alignment of just a byte
1003 and a field split across two bytes.
1004 Such cases are not supposed to be able to occur. */
1008 machine_mode mode
= GET_MODE (op0
);
1009 if (GET_MODE_BITSIZE (mode
) == 0
1010 || GET_MODE_BITSIZE (mode
) > GET_MODE_BITSIZE (word_mode
))
1012 mode
= get_best_mode (bitsize
, bitnum
, bitregion_start
, bitregion_end
,
1013 MEM_ALIGN (op0
), mode
, MEM_VOLATILE_P (op0
));
1015 if (mode
== VOIDmode
)
1017 /* The only way this should occur is if the field spans word
1019 store_split_bit_field (op0
, bitsize
, bitnum
, bitregion_start
,
1020 bitregion_end
, value
);
1024 op0
= narrow_bit_field_mem (op0
, mode
, bitsize
, bitnum
, &bitnum
);
1027 store_fixed_bit_field_1 (op0
, bitsize
, bitnum
, value
);
1030 /* Helper function for store_fixed_bit_field, stores
1031 the bit field always using the MODE of OP0. */
1034 store_fixed_bit_field_1 (rtx op0
, unsigned HOST_WIDE_INT bitsize
,
1035 unsigned HOST_WIDE_INT bitnum
,
1043 mode
= GET_MODE (op0
);
1044 gcc_assert (SCALAR_INT_MODE_P (mode
));
1046 /* Note that bitsize + bitnum can be greater than GET_MODE_BITSIZE (mode)
1047 for invalid input, such as f5 from gcc.dg/pr48335-2.c. */
1049 if (BYTES_BIG_ENDIAN
)
1050 /* BITNUM is the distance between our msb
1051 and that of the containing datum.
1052 Convert it to the distance from the lsb. */
1053 bitnum
= GET_MODE_BITSIZE (mode
) - bitsize
- bitnum
;
1055 /* Now BITNUM is always the distance between our lsb
1058 /* Shift VALUE left by BITNUM bits. If VALUE is not constant,
1059 we must first convert its mode to MODE. */
1061 if (CONST_INT_P (value
))
1063 unsigned HOST_WIDE_INT v
= UINTVAL (value
);
1065 if (bitsize
< HOST_BITS_PER_WIDE_INT
)
1066 v
&= ((unsigned HOST_WIDE_INT
) 1 << bitsize
) - 1;
1070 else if ((bitsize
< HOST_BITS_PER_WIDE_INT
1071 && v
== ((unsigned HOST_WIDE_INT
) 1 << bitsize
) - 1)
1072 || (bitsize
== HOST_BITS_PER_WIDE_INT
1073 && v
== (unsigned HOST_WIDE_INT
) -1))
1076 value
= lshift_value (mode
, v
, bitnum
);
1080 int must_and
= (GET_MODE_BITSIZE (GET_MODE (value
)) != bitsize
1081 && bitnum
+ bitsize
!= GET_MODE_BITSIZE (mode
));
1083 if (GET_MODE (value
) != mode
)
1084 value
= convert_to_mode (mode
, value
, 1);
1087 value
= expand_binop (mode
, and_optab
, value
,
1088 mask_rtx (mode
, 0, bitsize
, 0),
1089 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
1091 value
= expand_shift (LSHIFT_EXPR
, mode
, value
,
1092 bitnum
, NULL_RTX
, 1);
1095 /* Now clear the chosen bits in OP0,
1096 except that if VALUE is -1 we need not bother. */
1097 /* We keep the intermediates in registers to allow CSE to combine
1098 consecutive bitfield assignments. */
1100 temp
= force_reg (mode
, op0
);
1104 temp
= expand_binop (mode
, and_optab
, temp
,
1105 mask_rtx (mode
, bitnum
, bitsize
, 1),
1106 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
1107 temp
= force_reg (mode
, temp
);
1110 /* Now logical-or VALUE into OP0, unless it is zero. */
1114 temp
= expand_binop (mode
, ior_optab
, temp
, value
,
1115 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
1116 temp
= force_reg (mode
, temp
);
1121 op0
= copy_rtx (op0
);
1122 emit_move_insn (op0
, temp
);
1126 /* Store a bit field that is split across multiple accessible memory objects.
1128 OP0 is the REG, SUBREG or MEM rtx for the first of the objects.
1129 BITSIZE is the field width; BITPOS the position of its first bit
1131 VALUE is the value to store.
1133 This does not yet handle fields wider than BITS_PER_WORD. */
1136 store_split_bit_field (rtx op0
, unsigned HOST_WIDE_INT bitsize
,
1137 unsigned HOST_WIDE_INT bitpos
,
1138 unsigned HOST_WIDE_INT bitregion_start
,
1139 unsigned HOST_WIDE_INT bitregion_end
,
1143 unsigned int bitsdone
= 0;
1145 /* Make sure UNIT isn't larger than BITS_PER_WORD, we can only handle that
1147 if (REG_P (op0
) || GET_CODE (op0
) == SUBREG
)
1148 unit
= BITS_PER_WORD
;
1150 unit
= MIN (MEM_ALIGN (op0
), BITS_PER_WORD
);
1152 /* If OP0 is a memory with a mode, then UNIT must not be larger than
1153 OP0's mode as well. Otherwise, store_fixed_bit_field will call us
1154 again, and we will mutually recurse forever. */
1155 if (MEM_P (op0
) && GET_MODE_BITSIZE (GET_MODE (op0
)) > 0)
1156 unit
= MIN (unit
, GET_MODE_BITSIZE (GET_MODE (op0
)));
1158 /* If VALUE is a constant other than a CONST_INT, get it into a register in
1159 WORD_MODE. If we can do this using gen_lowpart_common, do so. Note
1160 that VALUE might be a floating-point constant. */
1161 if (CONSTANT_P (value
) && !CONST_INT_P (value
))
1163 rtx word
= gen_lowpart_common (word_mode
, value
);
1165 if (word
&& (value
!= word
))
1168 value
= gen_lowpart_common (word_mode
,
1169 force_reg (GET_MODE (value
) != VOIDmode
1171 : word_mode
, value
));
1174 while (bitsdone
< bitsize
)
1176 unsigned HOST_WIDE_INT thissize
;
1178 unsigned HOST_WIDE_INT thispos
;
1179 unsigned HOST_WIDE_INT offset
;
1181 offset
= (bitpos
+ bitsdone
) / unit
;
1182 thispos
= (bitpos
+ bitsdone
) % unit
;
1184 /* When region of bytes we can touch is restricted, decrease
1185 UNIT close to the end of the region as needed. If op0 is a REG
1186 or SUBREG of REG, don't do this, as there can't be data races
1187 on a register and we can expand shorter code in some cases. */
1189 && unit
> BITS_PER_UNIT
1190 && bitpos
+ bitsdone
- thispos
+ unit
> bitregion_end
+ 1
1192 && (GET_CODE (op0
) != SUBREG
|| !REG_P (SUBREG_REG (op0
))))
1198 /* THISSIZE must not overrun a word boundary. Otherwise,
1199 store_fixed_bit_field will call us again, and we will mutually
1201 thissize
= MIN (bitsize
- bitsdone
, BITS_PER_WORD
);
1202 thissize
= MIN (thissize
, unit
- thispos
);
1204 if (BYTES_BIG_ENDIAN
)
1206 /* Fetch successively less significant portions. */
1207 if (CONST_INT_P (value
))
1208 part
= GEN_INT (((unsigned HOST_WIDE_INT
) (INTVAL (value
))
1209 >> (bitsize
- bitsdone
- thissize
))
1210 & (((HOST_WIDE_INT
) 1 << thissize
) - 1));
1213 int total_bits
= GET_MODE_BITSIZE (GET_MODE (value
));
1214 /* The args are chosen so that the last part includes the
1215 lsb. Give extract_bit_field the value it needs (with
1216 endianness compensation) to fetch the piece we want. */
1217 part
= extract_fixed_bit_field (word_mode
, value
, thissize
,
1218 total_bits
- bitsize
+ bitsdone
,
1224 /* Fetch successively more significant portions. */
1225 if (CONST_INT_P (value
))
1226 part
= GEN_INT (((unsigned HOST_WIDE_INT
) (INTVAL (value
))
1228 & (((HOST_WIDE_INT
) 1 << thissize
) - 1));
1230 part
= extract_fixed_bit_field (word_mode
, value
, thissize
,
1231 bitsdone
, NULL_RTX
, 1);
1234 /* If OP0 is a register, then handle OFFSET here.
1236 When handling multiword bitfields, extract_bit_field may pass
1237 down a word_mode SUBREG of a larger REG for a bitfield that actually
1238 crosses a word boundary. Thus, for a SUBREG, we must find
1239 the current word starting from the base register. */
1240 if (GET_CODE (op0
) == SUBREG
)
1242 int word_offset
= (SUBREG_BYTE (op0
) / UNITS_PER_WORD
)
1243 + (offset
* unit
/ BITS_PER_WORD
);
1244 machine_mode sub_mode
= GET_MODE (SUBREG_REG (op0
));
1245 if (sub_mode
!= BLKmode
&& GET_MODE_SIZE (sub_mode
) < UNITS_PER_WORD
)
1246 word
= word_offset
? const0_rtx
: op0
;
1248 word
= operand_subword_force (SUBREG_REG (op0
), word_offset
,
1249 GET_MODE (SUBREG_REG (op0
)));
1250 offset
&= BITS_PER_WORD
/ unit
- 1;
1252 else if (REG_P (op0
))
1254 machine_mode op0_mode
= GET_MODE (op0
);
1255 if (op0_mode
!= BLKmode
&& GET_MODE_SIZE (op0_mode
) < UNITS_PER_WORD
)
1256 word
= offset
? const0_rtx
: op0
;
1258 word
= operand_subword_force (op0
, offset
* unit
/ BITS_PER_WORD
,
1260 offset
&= BITS_PER_WORD
/ unit
- 1;
1265 /* OFFSET is in UNITs, and UNIT is in bits. If WORD is const0_rtx,
1266 it is just an out-of-bounds access. Ignore it. */
1267 if (word
!= const0_rtx
)
1268 store_fixed_bit_field (word
, thissize
, offset
* unit
+ thispos
,
1269 bitregion_start
, bitregion_end
, part
);
1270 bitsdone
+= thissize
;
1274 /* A subroutine of extract_bit_field_1 that converts return value X
1275 to either MODE or TMODE. MODE, TMODE and UNSIGNEDP are arguments
1276 to extract_bit_field. */
1279 convert_extracted_bit_field (rtx x
, machine_mode mode
,
1280 machine_mode tmode
, bool unsignedp
)
1282 if (GET_MODE (x
) == tmode
|| GET_MODE (x
) == mode
)
1285 /* If the x mode is not a scalar integral, first convert to the
1286 integer mode of that size and then access it as a floating-point
1287 value via a SUBREG. */
1288 if (!SCALAR_INT_MODE_P (tmode
))
1292 smode
= mode_for_size (GET_MODE_BITSIZE (tmode
), MODE_INT
, 0);
1293 x
= convert_to_mode (smode
, x
, unsignedp
);
1294 x
= force_reg (smode
, x
);
1295 return gen_lowpart (tmode
, x
);
1298 return convert_to_mode (tmode
, x
, unsignedp
);
1301 /* Try to use an ext(z)v pattern to extract a field from OP0.
1302 Return the extracted value on success, otherwise return null.
1303 EXT_MODE is the mode of the extraction and the other arguments
1304 are as for extract_bit_field. */
1307 extract_bit_field_using_extv (const extraction_insn
*extv
, rtx op0
,
1308 unsigned HOST_WIDE_INT bitsize
,
1309 unsigned HOST_WIDE_INT bitnum
,
1310 int unsignedp
, rtx target
,
1311 machine_mode mode
, machine_mode tmode
)
1313 struct expand_operand ops
[4];
1314 rtx spec_target
= target
;
1315 rtx spec_target_subreg
= 0;
1316 machine_mode ext_mode
= extv
->field_mode
;
1317 unsigned unit
= GET_MODE_BITSIZE (ext_mode
);
1319 if (bitsize
== 0 || unit
< bitsize
)
1323 /* Get a reference to the first byte of the field. */
1324 op0
= narrow_bit_field_mem (op0
, extv
->struct_mode
, bitsize
, bitnum
,
1328 /* Convert from counting within OP0 to counting in EXT_MODE. */
1329 if (BYTES_BIG_ENDIAN
)
1330 bitnum
+= unit
- GET_MODE_BITSIZE (GET_MODE (op0
));
1332 /* If op0 is a register, we need it in EXT_MODE to make it
1333 acceptable to the format of ext(z)v. */
1334 if (GET_CODE (op0
) == SUBREG
&& GET_MODE (op0
) != ext_mode
)
1336 if (REG_P (op0
) && GET_MODE (op0
) != ext_mode
)
1337 op0
= gen_lowpart_SUBREG (ext_mode
, op0
);
1340 /* If BITS_BIG_ENDIAN is zero on a BYTES_BIG_ENDIAN machine, we count
1341 "backwards" from the size of the unit we are extracting from.
1342 Otherwise, we count bits from the most significant on a
1343 BYTES/BITS_BIG_ENDIAN machine. */
1345 if (BITS_BIG_ENDIAN
!= BYTES_BIG_ENDIAN
)
1346 bitnum
= unit
- bitsize
- bitnum
;
1349 target
= spec_target
= gen_reg_rtx (tmode
);
1351 if (GET_MODE (target
) != ext_mode
)
1353 /* Don't use LHS paradoxical subreg if explicit truncation is needed
1354 between the mode of the extraction (word_mode) and the target
1355 mode. Instead, create a temporary and use convert_move to set
1358 && TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (target
), ext_mode
))
1360 target
= gen_lowpart (ext_mode
, target
);
1361 if (GET_MODE_PRECISION (ext_mode
)
1362 > GET_MODE_PRECISION (GET_MODE (spec_target
)))
1363 spec_target_subreg
= target
;
1366 target
= gen_reg_rtx (ext_mode
);
1369 create_output_operand (&ops
[0], target
, ext_mode
);
1370 create_fixed_operand (&ops
[1], op0
);
1371 create_integer_operand (&ops
[2], bitsize
);
1372 create_integer_operand (&ops
[3], bitnum
);
1373 if (maybe_expand_insn (extv
->icode
, 4, ops
))
1375 target
= ops
[0].value
;
1376 if (target
== spec_target
)
1378 if (target
== spec_target_subreg
)
1380 return convert_extracted_bit_field (target
, mode
, tmode
, unsignedp
);
1385 /* A subroutine of extract_bit_field, with the same arguments.
1386 If FALLBACK_P is true, fall back to extract_fixed_bit_field
1387 if we can find no other means of implementing the operation.
1388 if FALLBACK_P is false, return NULL instead. */
1391 extract_bit_field_1 (rtx str_rtx
, unsigned HOST_WIDE_INT bitsize
,
1392 unsigned HOST_WIDE_INT bitnum
, int unsignedp
, rtx target
,
1393 machine_mode mode
, machine_mode tmode
,
1397 machine_mode int_mode
;
1400 if (tmode
== VOIDmode
)
1403 while (GET_CODE (op0
) == SUBREG
)
1405 bitnum
+= SUBREG_BYTE (op0
) * BITS_PER_UNIT
;
1406 op0
= SUBREG_REG (op0
);
1409 /* If we have an out-of-bounds access to a register, just return an
1410 uninitialized register of the required mode. This can occur if the
1411 source code contains an out-of-bounds access to a small array. */
1412 if (REG_P (op0
) && bitnum
>= GET_MODE_BITSIZE (GET_MODE (op0
)))
1413 return gen_reg_rtx (tmode
);
1416 && mode
== GET_MODE (op0
)
1418 && bitsize
== GET_MODE_BITSIZE (GET_MODE (op0
)))
1420 /* We're trying to extract a full register from itself. */
1424 /* See if we can get a better vector mode before extracting. */
1425 if (VECTOR_MODE_P (GET_MODE (op0
))
1427 && GET_MODE_INNER (GET_MODE (op0
)) != tmode
)
1429 machine_mode new_mode
;
1431 if (GET_MODE_CLASS (tmode
) == MODE_FLOAT
)
1432 new_mode
= MIN_MODE_VECTOR_FLOAT
;
1433 else if (GET_MODE_CLASS (tmode
) == MODE_FRACT
)
1434 new_mode
= MIN_MODE_VECTOR_FRACT
;
1435 else if (GET_MODE_CLASS (tmode
) == MODE_UFRACT
)
1436 new_mode
= MIN_MODE_VECTOR_UFRACT
;
1437 else if (GET_MODE_CLASS (tmode
) == MODE_ACCUM
)
1438 new_mode
= MIN_MODE_VECTOR_ACCUM
;
1439 else if (GET_MODE_CLASS (tmode
) == MODE_UACCUM
)
1440 new_mode
= MIN_MODE_VECTOR_UACCUM
;
1442 new_mode
= MIN_MODE_VECTOR_INT
;
1444 for (; new_mode
!= VOIDmode
; new_mode
= GET_MODE_WIDER_MODE (new_mode
))
1445 if (GET_MODE_SIZE (new_mode
) == GET_MODE_SIZE (GET_MODE (op0
))
1446 && targetm
.vector_mode_supported_p (new_mode
))
1448 if (new_mode
!= VOIDmode
)
1449 op0
= gen_lowpart (new_mode
, op0
);
1452 /* Use vec_extract patterns for extracting parts of vectors whenever
1454 if (VECTOR_MODE_P (GET_MODE (op0
))
1456 && optab_handler (vec_extract_optab
, GET_MODE (op0
)) != CODE_FOR_nothing
1457 && ((bitnum
+ bitsize
- 1) / GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0
)))
1458 == bitnum
/ GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0
)))))
1460 struct expand_operand ops
[3];
1461 machine_mode outermode
= GET_MODE (op0
);
1462 machine_mode innermode
= GET_MODE_INNER (outermode
);
1463 enum insn_code icode
= optab_handler (vec_extract_optab
, outermode
);
1464 unsigned HOST_WIDE_INT pos
= bitnum
/ GET_MODE_BITSIZE (innermode
);
1466 create_output_operand (&ops
[0], target
, innermode
);
1467 create_input_operand (&ops
[1], op0
, outermode
);
1468 create_integer_operand (&ops
[2], pos
);
1469 if (maybe_expand_insn (icode
, 3, ops
))
1471 target
= ops
[0].value
;
1472 if (GET_MODE (target
) != mode
)
1473 return gen_lowpart (tmode
, target
);
1478 /* Make sure we are playing with integral modes. Pun with subregs
1481 machine_mode imode
= int_mode_for_mode (GET_MODE (op0
));
1482 if (imode
!= GET_MODE (op0
))
1485 op0
= adjust_bitfield_address_size (op0
, imode
, 0, MEM_SIZE (op0
));
1486 else if (imode
!= BLKmode
)
1488 op0
= gen_lowpart (imode
, op0
);
1490 /* If we got a SUBREG, force it into a register since we
1491 aren't going to be able to do another SUBREG on it. */
1492 if (GET_CODE (op0
) == SUBREG
)
1493 op0
= force_reg (imode
, op0
);
1495 else if (REG_P (op0
))
1498 imode
= smallest_mode_for_size (GET_MODE_BITSIZE (GET_MODE (op0
)),
1500 reg
= gen_reg_rtx (imode
);
1501 subreg
= gen_lowpart_SUBREG (GET_MODE (op0
), reg
);
1502 emit_move_insn (subreg
, op0
);
1504 bitnum
+= SUBREG_BYTE (subreg
) * BITS_PER_UNIT
;
1508 HOST_WIDE_INT size
= GET_MODE_SIZE (GET_MODE (op0
));
1509 rtx mem
= assign_stack_temp (GET_MODE (op0
), size
);
1510 emit_move_insn (mem
, op0
);
1511 op0
= adjust_bitfield_address_size (mem
, BLKmode
, 0, size
);
1516 /* ??? We currently assume TARGET is at least as big as BITSIZE.
1517 If that's wrong, the solution is to test for it and set TARGET to 0
1520 /* Get the mode of the field to use for atomic access or subreg
1523 if (SCALAR_INT_MODE_P (tmode
))
1525 machine_mode try_mode
= mode_for_size (bitsize
,
1526 GET_MODE_CLASS (tmode
), 0);
1527 if (try_mode
!= BLKmode
)
1530 gcc_assert (mode1
!= BLKmode
);
1532 /* Extraction of a full MODE1 value can be done with a subreg as long
1533 as the least significant bit of the value is the least significant
1534 bit of either OP0 or a word of OP0. */
1536 && lowpart_bit_field_p (bitnum
, bitsize
, GET_MODE (op0
))
1537 && bitsize
== GET_MODE_BITSIZE (mode1
)
1538 && TRULY_NOOP_TRUNCATION_MODES_P (mode1
, GET_MODE (op0
)))
1540 rtx sub
= simplify_gen_subreg (mode1
, op0
, GET_MODE (op0
),
1541 bitnum
/ BITS_PER_UNIT
);
1543 return convert_extracted_bit_field (sub
, mode
, tmode
, unsignedp
);
1546 /* Extraction of a full MODE1 value can be done with a load as long as
1547 the field is on a byte boundary and is sufficiently aligned. */
1548 if (simple_mem_bitfield_p (op0
, bitsize
, bitnum
, mode1
))
1550 op0
= adjust_bitfield_address (op0
, mode1
, bitnum
/ BITS_PER_UNIT
);
1551 return convert_extracted_bit_field (op0
, mode
, tmode
, unsignedp
);
1554 /* Handle fields bigger than a word. */
1556 if (bitsize
> BITS_PER_WORD
)
1558 /* Here we transfer the words of the field
1559 in the order least significant first.
1560 This is because the most significant word is the one which may
1561 be less than full. */
1563 unsigned int backwards
= WORDS_BIG_ENDIAN
;
1564 unsigned int nwords
= (bitsize
+ (BITS_PER_WORD
- 1)) / BITS_PER_WORD
;
1568 if (target
== 0 || !REG_P (target
) || !valid_multiword_target_p (target
))
1569 target
= gen_reg_rtx (mode
);
1571 /* Indicate for flow that the entire target reg is being set. */
1572 emit_clobber (target
);
1574 last
= get_last_insn ();
1575 for (i
= 0; i
< nwords
; i
++)
1577 /* If I is 0, use the low-order word in both field and target;
1578 if I is 1, use the next to lowest word; and so on. */
1579 /* Word number in TARGET to use. */
1580 unsigned int wordnum
1582 ? GET_MODE_SIZE (GET_MODE (target
)) / UNITS_PER_WORD
- i
- 1
1584 /* Offset from start of field in OP0. */
1585 unsigned int bit_offset
= (backwards
1586 ? MAX ((int) bitsize
- ((int) i
+ 1)
1589 : (int) i
* BITS_PER_WORD
);
1590 rtx target_part
= operand_subword (target
, wordnum
, 1, VOIDmode
);
1592 = extract_bit_field_1 (op0
, MIN (BITS_PER_WORD
,
1593 bitsize
- i
* BITS_PER_WORD
),
1594 bitnum
+ bit_offset
, 1, target_part
,
1595 mode
, word_mode
, fallback_p
);
1597 gcc_assert (target_part
);
1600 delete_insns_since (last
);
1604 if (result_part
!= target_part
)
1605 emit_move_insn (target_part
, result_part
);
1610 /* Unless we've filled TARGET, the upper regs in a multi-reg value
1611 need to be zero'd out. */
1612 if (GET_MODE_SIZE (GET_MODE (target
)) > nwords
* UNITS_PER_WORD
)
1614 unsigned int i
, total_words
;
1616 total_words
= GET_MODE_SIZE (GET_MODE (target
)) / UNITS_PER_WORD
;
1617 for (i
= nwords
; i
< total_words
; i
++)
1619 (operand_subword (target
,
1620 backwards
? total_words
- i
- 1 : i
,
1627 /* Signed bit field: sign-extend with two arithmetic shifts. */
1628 target
= expand_shift (LSHIFT_EXPR
, mode
, target
,
1629 GET_MODE_BITSIZE (mode
) - bitsize
, NULL_RTX
, 0);
1630 return expand_shift (RSHIFT_EXPR
, mode
, target
,
1631 GET_MODE_BITSIZE (mode
) - bitsize
, NULL_RTX
, 0);
1634 /* If OP0 is a multi-word register, narrow it to the affected word.
1635 If the region spans two words, defer to extract_split_bit_field. */
1636 if (!MEM_P (op0
) && GET_MODE_SIZE (GET_MODE (op0
)) > UNITS_PER_WORD
)
1638 op0
= simplify_gen_subreg (word_mode
, op0
, GET_MODE (op0
),
1639 bitnum
/ BITS_PER_WORD
* UNITS_PER_WORD
);
1640 bitnum
%= BITS_PER_WORD
;
1641 if (bitnum
+ bitsize
> BITS_PER_WORD
)
1645 target
= extract_split_bit_field (op0
, bitsize
, bitnum
, unsignedp
);
1646 return convert_extracted_bit_field (target
, mode
, tmode
, unsignedp
);
1650 /* From here on we know the desired field is smaller than a word.
1651 If OP0 is a register, it too fits within a word. */
1652 enum extraction_pattern pattern
= unsignedp
? EP_extzv
: EP_extv
;
1653 extraction_insn extv
;
1655 /* ??? We could limit the structure size to the part of OP0 that
1656 contains the field, with appropriate checks for endianness
1657 and TRULY_NOOP_TRUNCATION. */
1658 && get_best_reg_extraction_insn (&extv
, pattern
,
1659 GET_MODE_BITSIZE (GET_MODE (op0
)),
1662 rtx result
= extract_bit_field_using_extv (&extv
, op0
, bitsize
, bitnum
,
1663 unsignedp
, target
, mode
,
1669 /* If OP0 is a memory, try copying it to a register and seeing if a
1670 cheap register alternative is available. */
1673 if (get_best_mem_extraction_insn (&extv
, pattern
, bitsize
, bitnum
,
1676 rtx result
= extract_bit_field_using_extv (&extv
, op0
, bitsize
,
1684 rtx_insn
*last
= get_last_insn ();
1686 /* Try loading part of OP0 into a register and extracting the
1687 bitfield from that. */
1688 unsigned HOST_WIDE_INT bitpos
;
1689 rtx xop0
= adjust_bit_field_mem_for_reg (pattern
, op0
, bitsize
, bitnum
,
1690 0, 0, tmode
, &bitpos
);
1693 xop0
= copy_to_reg (xop0
);
1694 rtx result
= extract_bit_field_1 (xop0
, bitsize
, bitpos
,
1696 mode
, tmode
, false);
1699 delete_insns_since (last
);
1706 /* Find a correspondingly-sized integer field, so we can apply
1707 shifts and masks to it. */
1708 int_mode
= int_mode_for_mode (tmode
);
1709 if (int_mode
== BLKmode
)
1710 int_mode
= int_mode_for_mode (mode
);
1711 /* Should probably push op0 out to memory and then do a load. */
1712 gcc_assert (int_mode
!= BLKmode
);
1714 target
= extract_fixed_bit_field (int_mode
, op0
, bitsize
, bitnum
,
1716 return convert_extracted_bit_field (target
, mode
, tmode
, unsignedp
);
1719 /* Generate code to extract a byte-field from STR_RTX
1720 containing BITSIZE bits, starting at BITNUM,
1721 and put it in TARGET if possible (if TARGET is nonzero).
1722 Regardless of TARGET, we return the rtx for where the value is placed.
1724 STR_RTX is the structure containing the byte (a REG or MEM).
1725 UNSIGNEDP is nonzero if this is an unsigned bit field.
1726 MODE is the natural mode of the field value once extracted.
1727 TMODE is the mode the caller would like the value to have;
1728 but the value may be returned with type MODE instead.
1730 If a TARGET is specified and we can store in it at no extra cost,
1731 we do so, and return TARGET.
1732 Otherwise, we return a REG of mode TMODE or MODE, with TMODE preferred
1733 if they are equally easy. */
1736 extract_bit_field (rtx str_rtx
, unsigned HOST_WIDE_INT bitsize
,
1737 unsigned HOST_WIDE_INT bitnum
, int unsignedp
, rtx target
,
1738 machine_mode mode
, machine_mode tmode
)
1742 /* Handle -fstrict-volatile-bitfields in the cases where it applies. */
1743 if (GET_MODE_BITSIZE (GET_MODE (str_rtx
)) > 0)
1744 mode1
= GET_MODE (str_rtx
);
1745 else if (target
&& GET_MODE_BITSIZE (GET_MODE (target
)) > 0)
1746 mode1
= GET_MODE (target
);
1750 if (strict_volatile_bitfield_p (str_rtx
, bitsize
, bitnum
, mode1
, 0, 0))
1754 /* Extraction of a full MODE1 value can be done with a load as long as
1755 the field is on a byte boundary and is sufficiently aligned. */
1756 if (simple_mem_bitfield_p (str_rtx
, bitsize
, bitnum
, mode1
))
1757 result
= adjust_bitfield_address (str_rtx
, mode1
,
1758 bitnum
/ BITS_PER_UNIT
);
1761 str_rtx
= narrow_bit_field_mem (str_rtx
, mode1
, bitsize
, bitnum
,
1763 result
= extract_fixed_bit_field_1 (mode
, str_rtx
, bitsize
, bitnum
,
1767 return convert_extracted_bit_field (result
, mode
, tmode
, unsignedp
);
1770 return extract_bit_field_1 (str_rtx
, bitsize
, bitnum
, unsignedp
,
1771 target
, mode
, tmode
, true);
1774 /* Use shifts and boolean operations to extract a field of BITSIZE bits
1775 from bit BITNUM of OP0.
1777 UNSIGNEDP is nonzero for an unsigned bit field (don't sign-extend value).
1778 If TARGET is nonzero, attempts to store the value there
1779 and return TARGET, but this is not guaranteed.
1780 If TARGET is not used, create a pseudo-reg of mode TMODE for the value. */
1783 extract_fixed_bit_field (machine_mode tmode
, rtx op0
,
1784 unsigned HOST_WIDE_INT bitsize
,
1785 unsigned HOST_WIDE_INT bitnum
, rtx target
,
1791 = get_best_mode (bitsize
, bitnum
, 0, 0, MEM_ALIGN (op0
), word_mode
,
1792 MEM_VOLATILE_P (op0
));
1794 if (mode
== VOIDmode
)
1795 /* The only way this should occur is if the field spans word
1797 return extract_split_bit_field (op0
, bitsize
, bitnum
, unsignedp
);
1799 op0
= narrow_bit_field_mem (op0
, mode
, bitsize
, bitnum
, &bitnum
);
1802 return extract_fixed_bit_field_1 (tmode
, op0
, bitsize
, bitnum
,
1806 /* Helper function for extract_fixed_bit_field, extracts
1807 the bit field always using the MODE of OP0. */
1810 extract_fixed_bit_field_1 (machine_mode tmode
, rtx op0
,
1811 unsigned HOST_WIDE_INT bitsize
,
1812 unsigned HOST_WIDE_INT bitnum
, rtx target
,
1815 machine_mode mode
= GET_MODE (op0
);
1816 gcc_assert (SCALAR_INT_MODE_P (mode
));
1818 /* Note that bitsize + bitnum can be greater than GET_MODE_BITSIZE (mode)
1819 for invalid input, such as extract equivalent of f5 from
1820 gcc.dg/pr48335-2.c. */
1822 if (BYTES_BIG_ENDIAN
)
1823 /* BITNUM is the distance between our msb and that of OP0.
1824 Convert it to the distance from the lsb. */
1825 bitnum
= GET_MODE_BITSIZE (mode
) - bitsize
- bitnum
;
1827 /* Now BITNUM is always the distance between the field's lsb and that of OP0.
1828 We have reduced the big-endian case to the little-endian case. */
1834 /* If the field does not already start at the lsb,
1835 shift it so it does. */
1836 /* Maybe propagate the target for the shift. */
1837 rtx subtarget
= (target
!= 0 && REG_P (target
) ? target
: 0);
1840 op0
= expand_shift (RSHIFT_EXPR
, mode
, op0
, bitnum
, subtarget
, 1);
1842 /* Convert the value to the desired mode. */
1844 op0
= convert_to_mode (tmode
, op0
, 1);
1846 /* Unless the msb of the field used to be the msb when we shifted,
1847 mask out the upper bits. */
1849 if (GET_MODE_BITSIZE (mode
) != bitnum
+ bitsize
)
1850 return expand_binop (GET_MODE (op0
), and_optab
, op0
,
1851 mask_rtx (GET_MODE (op0
), 0, bitsize
, 0),
1852 target
, 1, OPTAB_LIB_WIDEN
);
1856 /* To extract a signed bit-field, first shift its msb to the msb of the word,
1857 then arithmetic-shift its lsb to the lsb of the word. */
1858 op0
= force_reg (mode
, op0
);
1860 /* Find the narrowest integer mode that contains the field. */
1862 for (mode
= GET_CLASS_NARROWEST_MODE (MODE_INT
); mode
!= VOIDmode
;
1863 mode
= GET_MODE_WIDER_MODE (mode
))
1864 if (GET_MODE_BITSIZE (mode
) >= bitsize
+ bitnum
)
1866 op0
= convert_to_mode (mode
, op0
, 0);
1873 if (GET_MODE_BITSIZE (mode
) != (bitsize
+ bitnum
))
1875 int amount
= GET_MODE_BITSIZE (mode
) - (bitsize
+ bitnum
);
1876 /* Maybe propagate the target for the shift. */
1877 rtx subtarget
= (target
!= 0 && REG_P (target
) ? target
: 0);
1878 op0
= expand_shift (LSHIFT_EXPR
, mode
, op0
, amount
, subtarget
, 1);
1881 return expand_shift (RSHIFT_EXPR
, mode
, op0
,
1882 GET_MODE_BITSIZE (mode
) - bitsize
, target
, 0);
1885 /* Return a constant integer (CONST_INT or CONST_DOUBLE) rtx with the value
1889 lshift_value (machine_mode mode
, unsigned HOST_WIDE_INT value
,
1892 return immed_wide_int_const (wi::lshift (value
, bitpos
), mode
);
1895 /* Extract a bit field that is split across two words
1896 and return an RTX for the result.
1898 OP0 is the REG, SUBREG or MEM rtx for the first of the two words.
1899 BITSIZE is the field width; BITPOS, position of its first bit, in the word.
1900 UNSIGNEDP is 1 if should zero-extend the contents; else sign-extend. */
1903 extract_split_bit_field (rtx op0
, unsigned HOST_WIDE_INT bitsize
,
1904 unsigned HOST_WIDE_INT bitpos
, int unsignedp
)
1907 unsigned int bitsdone
= 0;
1908 rtx result
= NULL_RTX
;
1911 /* Make sure UNIT isn't larger than BITS_PER_WORD, we can only handle that
1913 if (REG_P (op0
) || GET_CODE (op0
) == SUBREG
)
1914 unit
= BITS_PER_WORD
;
1916 unit
= MIN (MEM_ALIGN (op0
), BITS_PER_WORD
);
1918 while (bitsdone
< bitsize
)
1920 unsigned HOST_WIDE_INT thissize
;
1922 unsigned HOST_WIDE_INT thispos
;
1923 unsigned HOST_WIDE_INT offset
;
1925 offset
= (bitpos
+ bitsdone
) / unit
;
1926 thispos
= (bitpos
+ bitsdone
) % unit
;
1928 /* THISSIZE must not overrun a word boundary. Otherwise,
1929 extract_fixed_bit_field will call us again, and we will mutually
1931 thissize
= MIN (bitsize
- bitsdone
, BITS_PER_WORD
);
1932 thissize
= MIN (thissize
, unit
- thispos
);
1934 /* If OP0 is a register, then handle OFFSET here.
1936 When handling multiword bitfields, extract_bit_field may pass
1937 down a word_mode SUBREG of a larger REG for a bitfield that actually
1938 crosses a word boundary. Thus, for a SUBREG, we must find
1939 the current word starting from the base register. */
1940 if (GET_CODE (op0
) == SUBREG
)
1942 int word_offset
= (SUBREG_BYTE (op0
) / UNITS_PER_WORD
) + offset
;
1943 word
= operand_subword_force (SUBREG_REG (op0
), word_offset
,
1944 GET_MODE (SUBREG_REG (op0
)));
1947 else if (REG_P (op0
))
1949 word
= operand_subword_force (op0
, offset
, GET_MODE (op0
));
1955 /* Extract the parts in bit-counting order,
1956 whose meaning is determined by BYTES_PER_UNIT.
1957 OFFSET is in UNITs, and UNIT is in bits. */
1958 part
= extract_fixed_bit_field (word_mode
, word
, thissize
,
1959 offset
* unit
+ thispos
, 0, 1);
1960 bitsdone
+= thissize
;
1962 /* Shift this part into place for the result. */
1963 if (BYTES_BIG_ENDIAN
)
1965 if (bitsize
!= bitsdone
)
1966 part
= expand_shift (LSHIFT_EXPR
, word_mode
, part
,
1967 bitsize
- bitsdone
, 0, 1);
1971 if (bitsdone
!= thissize
)
1972 part
= expand_shift (LSHIFT_EXPR
, word_mode
, part
,
1973 bitsdone
- thissize
, 0, 1);
1979 /* Combine the parts with bitwise or. This works
1980 because we extracted each part as an unsigned bit field. */
1981 result
= expand_binop (word_mode
, ior_optab
, part
, result
, NULL_RTX
, 1,
1987 /* Unsigned bit field: we are done. */
1990 /* Signed bit field: sign-extend with two arithmetic shifts. */
1991 result
= expand_shift (LSHIFT_EXPR
, word_mode
, result
,
1992 BITS_PER_WORD
- bitsize
, NULL_RTX
, 0);
1993 return expand_shift (RSHIFT_EXPR
, word_mode
, result
,
1994 BITS_PER_WORD
- bitsize
, NULL_RTX
, 0);
1997 /* Try to read the low bits of SRC as an rvalue of mode MODE, preserving
1998 the bit pattern. SRC_MODE is the mode of SRC; if this is smaller than
1999 MODE, fill the upper bits with zeros. Fail if the layout of either
2000 mode is unknown (as for CC modes) or if the extraction would involve
2001 unprofitable mode punning. Return the value on success, otherwise
2004 This is different from gen_lowpart* in these respects:
2006 - the returned value must always be considered an rvalue
2008 - when MODE is wider than SRC_MODE, the extraction involves
2011 - when MODE is smaller than SRC_MODE, the extraction involves
2012 a truncation (and is thus subject to TRULY_NOOP_TRUNCATION).
2014 In other words, this routine performs a computation, whereas the
2015 gen_lowpart* routines are conceptually lvalue or rvalue subreg
2019 extract_low_bits (machine_mode mode
, machine_mode src_mode
, rtx src
)
2021 machine_mode int_mode
, src_int_mode
;
2023 if (mode
== src_mode
)
2026 if (CONSTANT_P (src
))
2028 /* simplify_gen_subreg can't be used here, as if simplify_subreg
2029 fails, it will happily create (subreg (symbol_ref)) or similar
2031 unsigned int byte
= subreg_lowpart_offset (mode
, src_mode
);
2032 rtx ret
= simplify_subreg (mode
, src
, src_mode
, byte
);
2036 if (GET_MODE (src
) == VOIDmode
2037 || !validate_subreg (mode
, src_mode
, src
, byte
))
2040 src
= force_reg (GET_MODE (src
), src
);
2041 return gen_rtx_SUBREG (mode
, src
, byte
);
2044 if (GET_MODE_CLASS (mode
) == MODE_CC
|| GET_MODE_CLASS (src_mode
) == MODE_CC
)
2047 if (GET_MODE_BITSIZE (mode
) == GET_MODE_BITSIZE (src_mode
)
2048 && MODES_TIEABLE_P (mode
, src_mode
))
2050 rtx x
= gen_lowpart_common (mode
, src
);
2055 src_int_mode
= int_mode_for_mode (src_mode
);
2056 int_mode
= int_mode_for_mode (mode
);
2057 if (src_int_mode
== BLKmode
|| int_mode
== BLKmode
)
2060 if (!MODES_TIEABLE_P (src_int_mode
, src_mode
))
2062 if (!MODES_TIEABLE_P (int_mode
, mode
))
2065 src
= gen_lowpart (src_int_mode
, src
);
2066 src
= convert_modes (int_mode
, src_int_mode
, src
, true);
2067 src
= gen_lowpart (mode
, src
);
2071 /* Add INC into TARGET. */
2074 expand_inc (rtx target
, rtx inc
)
2076 rtx value
= expand_binop (GET_MODE (target
), add_optab
,
2078 target
, 0, OPTAB_LIB_WIDEN
);
2079 if (value
!= target
)
2080 emit_move_insn (target
, value
);
2083 /* Subtract DEC from TARGET. */
2086 expand_dec (rtx target
, rtx dec
)
2088 rtx value
= expand_binop (GET_MODE (target
), sub_optab
,
2090 target
, 0, OPTAB_LIB_WIDEN
);
2091 if (value
!= target
)
2092 emit_move_insn (target
, value
);
2095 /* Output a shift instruction for expression code CODE,
2096 with SHIFTED being the rtx for the value to shift,
2097 and AMOUNT the rtx for the amount to shift by.
2098 Store the result in the rtx TARGET, if that is convenient.
2099 If UNSIGNEDP is nonzero, do a logical shift; otherwise, arithmetic.
2100 Return the rtx for where the value is. */
2103 expand_shift_1 (enum tree_code code
, machine_mode mode
, rtx shifted
,
2104 rtx amount
, rtx target
, int unsignedp
)
2107 int left
= (code
== LSHIFT_EXPR
|| code
== LROTATE_EXPR
);
2108 int rotate
= (code
== LROTATE_EXPR
|| code
== RROTATE_EXPR
);
2109 optab lshift_optab
= ashl_optab
;
2110 optab rshift_arith_optab
= ashr_optab
;
2111 optab rshift_uns_optab
= lshr_optab
;
2112 optab lrotate_optab
= rotl_optab
;
2113 optab rrotate_optab
= rotr_optab
;
2114 machine_mode op1_mode
;
2115 machine_mode scalar_mode
= mode
;
2117 bool speed
= optimize_insn_for_speed_p ();
2119 if (VECTOR_MODE_P (mode
))
2120 scalar_mode
= GET_MODE_INNER (mode
);
2122 op1_mode
= GET_MODE (op1
);
2124 /* Determine whether the shift/rotate amount is a vector, or scalar. If the
2125 shift amount is a vector, use the vector/vector shift patterns. */
2126 if (VECTOR_MODE_P (mode
) && VECTOR_MODE_P (op1_mode
))
2128 lshift_optab
= vashl_optab
;
2129 rshift_arith_optab
= vashr_optab
;
2130 rshift_uns_optab
= vlshr_optab
;
2131 lrotate_optab
= vrotl_optab
;
2132 rrotate_optab
= vrotr_optab
;
2135 /* Previously detected shift-counts computed by NEGATE_EXPR
2136 and shifted in the other direction; but that does not work
2139 if (SHIFT_COUNT_TRUNCATED
)
2141 if (CONST_INT_P (op1
)
2142 && ((unsigned HOST_WIDE_INT
) INTVAL (op1
) >=
2143 (unsigned HOST_WIDE_INT
) GET_MODE_BITSIZE (scalar_mode
)))
2144 op1
= GEN_INT ((unsigned HOST_WIDE_INT
) INTVAL (op1
)
2145 % GET_MODE_BITSIZE (scalar_mode
));
2146 else if (GET_CODE (op1
) == SUBREG
2147 && subreg_lowpart_p (op1
)
2148 && SCALAR_INT_MODE_P (GET_MODE (SUBREG_REG (op1
)))
2149 && SCALAR_INT_MODE_P (GET_MODE (op1
)))
2150 op1
= SUBREG_REG (op1
);
2153 /* Canonicalize rotates by constant amount. If op1 is bitsize / 2,
2154 prefer left rotation, if op1 is from bitsize / 2 + 1 to
2155 bitsize - 1, use other direction of rotate with 1 .. bitsize / 2 - 1
2158 && CONST_INT_P (op1
)
2159 && IN_RANGE (INTVAL (op1
), GET_MODE_BITSIZE (scalar_mode
) / 2 + left
,
2160 GET_MODE_BITSIZE (scalar_mode
) - 1))
2162 op1
= GEN_INT (GET_MODE_BITSIZE (scalar_mode
) - INTVAL (op1
));
2164 code
= left
? LROTATE_EXPR
: RROTATE_EXPR
;
2167 if (op1
== const0_rtx
)
2170 /* Check whether its cheaper to implement a left shift by a constant
2171 bit count by a sequence of additions. */
2172 if (code
== LSHIFT_EXPR
2173 && CONST_INT_P (op1
)
2175 && INTVAL (op1
) < GET_MODE_PRECISION (scalar_mode
)
2176 && INTVAL (op1
) < MAX_BITS_PER_WORD
2177 && (shift_cost (speed
, mode
, INTVAL (op1
))
2178 > INTVAL (op1
) * add_cost (speed
, mode
))
2179 && shift_cost (speed
, mode
, INTVAL (op1
)) != MAX_COST
)
2182 for (i
= 0; i
< INTVAL (op1
); i
++)
2184 temp
= force_reg (mode
, shifted
);
2185 shifted
= expand_binop (mode
, add_optab
, temp
, temp
, NULL_RTX
,
2186 unsignedp
, OPTAB_LIB_WIDEN
);
2191 for (attempt
= 0; temp
== 0 && attempt
< 3; attempt
++)
2193 enum optab_methods methods
;
2196 methods
= OPTAB_DIRECT
;
2197 else if (attempt
== 1)
2198 methods
= OPTAB_WIDEN
;
2200 methods
= OPTAB_LIB_WIDEN
;
2204 /* Widening does not work for rotation. */
2205 if (methods
== OPTAB_WIDEN
)
2207 else if (methods
== OPTAB_LIB_WIDEN
)
2209 /* If we have been unable to open-code this by a rotation,
2210 do it as the IOR of two shifts. I.e., to rotate A
2212 (A << N) | ((unsigned) A >> ((-N) & (C - 1)))
2213 where C is the bitsize of A.
2215 It is theoretically possible that the target machine might
2216 not be able to perform either shift and hence we would
2217 be making two libcalls rather than just the one for the
2218 shift (similarly if IOR could not be done). We will allow
2219 this extremely unlikely lossage to avoid complicating the
2222 rtx subtarget
= target
== shifted
? 0 : target
;
2223 rtx new_amount
, other_amount
;
2227 if (op1
== const0_rtx
)
2229 else if (CONST_INT_P (op1
))
2230 other_amount
= GEN_INT (GET_MODE_BITSIZE (scalar_mode
)
2235 = simplify_gen_unary (NEG
, GET_MODE (op1
),
2236 op1
, GET_MODE (op1
));
2237 HOST_WIDE_INT mask
= GET_MODE_PRECISION (scalar_mode
) - 1;
2239 = simplify_gen_binary (AND
, GET_MODE (op1
), other_amount
,
2240 gen_int_mode (mask
, GET_MODE (op1
)));
2243 shifted
= force_reg (mode
, shifted
);
2245 temp
= expand_shift_1 (left
? LSHIFT_EXPR
: RSHIFT_EXPR
,
2246 mode
, shifted
, new_amount
, 0, 1);
2247 temp1
= expand_shift_1 (left
? RSHIFT_EXPR
: LSHIFT_EXPR
,
2248 mode
, shifted
, other_amount
,
2250 return expand_binop (mode
, ior_optab
, temp
, temp1
, target
,
2251 unsignedp
, methods
);
2254 temp
= expand_binop (mode
,
2255 left
? lrotate_optab
: rrotate_optab
,
2256 shifted
, op1
, target
, unsignedp
, methods
);
2259 temp
= expand_binop (mode
,
2260 left
? lshift_optab
: rshift_uns_optab
,
2261 shifted
, op1
, target
, unsignedp
, methods
);
2263 /* Do arithmetic shifts.
2264 Also, if we are going to widen the operand, we can just as well
2265 use an arithmetic right-shift instead of a logical one. */
2266 if (temp
== 0 && ! rotate
2267 && (! unsignedp
|| (! left
&& methods
== OPTAB_WIDEN
)))
2269 enum optab_methods methods1
= methods
;
2271 /* If trying to widen a log shift to an arithmetic shift,
2272 don't accept an arithmetic shift of the same size. */
2274 methods1
= OPTAB_MUST_WIDEN
;
2276 /* Arithmetic shift */
2278 temp
= expand_binop (mode
,
2279 left
? lshift_optab
: rshift_arith_optab
,
2280 shifted
, op1
, target
, unsignedp
, methods1
);
2283 /* We used to try extzv here for logical right shifts, but that was
2284 only useful for one machine, the VAX, and caused poor code
2285 generation there for lshrdi3, so the code was deleted and a
2286 define_expand for lshrsi3 was added to vax.md. */
2293 /* Output a shift instruction for expression code CODE,
2294 with SHIFTED being the rtx for the value to shift,
2295 and AMOUNT the amount to shift by.
2296 Store the result in the rtx TARGET, if that is convenient.
2297 If UNSIGNEDP is nonzero, do a logical shift; otherwise, arithmetic.
2298 Return the rtx for where the value is. */
2301 expand_shift (enum tree_code code
, machine_mode mode
, rtx shifted
,
2302 int amount
, rtx target
, int unsignedp
)
2304 return expand_shift_1 (code
, mode
,
2305 shifted
, GEN_INT (amount
), target
, unsignedp
);
2308 /* Output a shift instruction for expression code CODE,
2309 with SHIFTED being the rtx for the value to shift,
2310 and AMOUNT the tree for the amount to shift by.
2311 Store the result in the rtx TARGET, if that is convenient.
2312 If UNSIGNEDP is nonzero, do a logical shift; otherwise, arithmetic.
2313 Return the rtx for where the value is. */
2316 expand_variable_shift (enum tree_code code
, machine_mode mode
, rtx shifted
,
2317 tree amount
, rtx target
, int unsignedp
)
2319 return expand_shift_1 (code
, mode
,
2320 shifted
, expand_normal (amount
), target
, unsignedp
);
2324 /* Indicates the type of fixup needed after a constant multiplication.
2325 BASIC_VARIANT means no fixup is needed, NEGATE_VARIANT means that
2326 the result should be negated, and ADD_VARIANT means that the
2327 multiplicand should be added to the result. */
2328 enum mult_variant
{basic_variant
, negate_variant
, add_variant
};
2330 static void synth_mult (struct algorithm
*, unsigned HOST_WIDE_INT
,
2331 const struct mult_cost
*, machine_mode mode
);
2332 static bool choose_mult_variant (machine_mode
, HOST_WIDE_INT
,
2333 struct algorithm
*, enum mult_variant
*, int);
2334 static rtx
expand_mult_const (machine_mode
, rtx
, HOST_WIDE_INT
, rtx
,
2335 const struct algorithm
*, enum mult_variant
);
2336 static unsigned HOST_WIDE_INT
invert_mod2n (unsigned HOST_WIDE_INT
, int);
2337 static rtx
extract_high_half (machine_mode
, rtx
);
2338 static rtx
expmed_mult_highpart (machine_mode
, rtx
, rtx
, rtx
, int, int);
2339 static rtx
expmed_mult_highpart_optab (machine_mode
, rtx
, rtx
, rtx
,
2341 /* Compute and return the best algorithm for multiplying by T.
2342 The algorithm must cost less than cost_limit
2343 If retval.cost >= COST_LIMIT, no algorithm was found and all
2344 other field of the returned struct are undefined.
2345 MODE is the machine mode of the multiplication. */
2348 synth_mult (struct algorithm
*alg_out
, unsigned HOST_WIDE_INT t
,
2349 const struct mult_cost
*cost_limit
, machine_mode mode
)
2352 struct algorithm
*alg_in
, *best_alg
;
2353 struct mult_cost best_cost
;
2354 struct mult_cost new_limit
;
2355 int op_cost
, op_latency
;
2356 unsigned HOST_WIDE_INT orig_t
= t
;
2357 unsigned HOST_WIDE_INT q
;
2358 int maxm
, hash_index
;
2359 bool cache_hit
= false;
2360 enum alg_code cache_alg
= alg_zero
;
2361 bool speed
= optimize_insn_for_speed_p ();
2363 struct alg_hash_entry
*entry_ptr
;
2365 /* Indicate that no algorithm is yet found. If no algorithm
2366 is found, this value will be returned and indicate failure. */
2367 alg_out
->cost
.cost
= cost_limit
->cost
+ 1;
2368 alg_out
->cost
.latency
= cost_limit
->latency
+ 1;
2370 if (cost_limit
->cost
< 0
2371 || (cost_limit
->cost
== 0 && cost_limit
->latency
<= 0))
2374 /* Be prepared for vector modes. */
2375 imode
= GET_MODE_INNER (mode
);
2376 if (imode
== VOIDmode
)
2379 maxm
= MIN (BITS_PER_WORD
, GET_MODE_BITSIZE (imode
));
2381 /* Restrict the bits of "t" to the multiplication's mode. */
2382 t
&= GET_MODE_MASK (imode
);
2384 /* t == 1 can be done in zero cost. */
2388 alg_out
->cost
.cost
= 0;
2389 alg_out
->cost
.latency
= 0;
2390 alg_out
->op
[0] = alg_m
;
2394 /* t == 0 sometimes has a cost. If it does and it exceeds our limit,
2398 if (MULT_COST_LESS (cost_limit
, zero_cost (speed
)))
2403 alg_out
->cost
.cost
= zero_cost (speed
);
2404 alg_out
->cost
.latency
= zero_cost (speed
);
2405 alg_out
->op
[0] = alg_zero
;
2410 /* We'll be needing a couple extra algorithm structures now. */
2412 alg_in
= XALLOCA (struct algorithm
);
2413 best_alg
= XALLOCA (struct algorithm
);
2414 best_cost
= *cost_limit
;
2416 /* Compute the hash index. */
2417 hash_index
= (t
^ (unsigned int) mode
^ (speed
* 256)) % NUM_ALG_HASH_ENTRIES
;
2419 /* See if we already know what to do for T. */
2420 entry_ptr
= alg_hash_entry_ptr (hash_index
);
2421 if (entry_ptr
->t
== t
2422 && entry_ptr
->mode
== mode
2423 && entry_ptr
->mode
== mode
2424 && entry_ptr
->speed
== speed
2425 && entry_ptr
->alg
!= alg_unknown
)
2427 cache_alg
= entry_ptr
->alg
;
2429 if (cache_alg
== alg_impossible
)
2431 /* The cache tells us that it's impossible to synthesize
2432 multiplication by T within entry_ptr->cost. */
2433 if (!CHEAPER_MULT_COST (&entry_ptr
->cost
, cost_limit
))
2434 /* COST_LIMIT is at least as restrictive as the one
2435 recorded in the hash table, in which case we have no
2436 hope of synthesizing a multiplication. Just
2440 /* If we get here, COST_LIMIT is less restrictive than the
2441 one recorded in the hash table, so we may be able to
2442 synthesize a multiplication. Proceed as if we didn't
2443 have the cache entry. */
2447 if (CHEAPER_MULT_COST (cost_limit
, &entry_ptr
->cost
))
2448 /* The cached algorithm shows that this multiplication
2449 requires more cost than COST_LIMIT. Just return. This
2450 way, we don't clobber this cache entry with
2451 alg_impossible but retain useful information. */
2463 goto do_alg_addsub_t_m2
;
2465 case alg_add_factor
:
2466 case alg_sub_factor
:
2467 goto do_alg_addsub_factor
;
2470 goto do_alg_add_t2_m
;
2473 goto do_alg_sub_t2_m
;
2481 /* If we have a group of zero bits at the low-order part of T, try
2482 multiplying by the remaining bits and then doing a shift. */
2487 m
= floor_log2 (t
& -t
); /* m = number of low zero bits */
2491 /* The function expand_shift will choose between a shift and
2492 a sequence of additions, so the observed cost is given as
2493 MIN (m * add_cost(speed, mode), shift_cost(speed, mode, m)). */
2494 op_cost
= m
* add_cost (speed
, mode
);
2495 if (shift_cost (speed
, mode
, m
) < op_cost
)
2496 op_cost
= shift_cost (speed
, mode
, m
);
2497 new_limit
.cost
= best_cost
.cost
- op_cost
;
2498 new_limit
.latency
= best_cost
.latency
- op_cost
;
2499 synth_mult (alg_in
, q
, &new_limit
, mode
);
2501 alg_in
->cost
.cost
+= op_cost
;
2502 alg_in
->cost
.latency
+= op_cost
;
2503 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2505 struct algorithm
*x
;
2506 best_cost
= alg_in
->cost
;
2507 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2508 best_alg
->log
[best_alg
->ops
] = m
;
2509 best_alg
->op
[best_alg
->ops
] = alg_shift
;
2512 /* See if treating ORIG_T as a signed number yields a better
2513 sequence. Try this sequence only for a negative ORIG_T
2514 as it would be useless for a non-negative ORIG_T. */
2515 if ((HOST_WIDE_INT
) orig_t
< 0)
2517 /* Shift ORIG_T as follows because a right shift of a
2518 negative-valued signed type is implementation
2520 q
= ~(~orig_t
>> m
);
2521 /* The function expand_shift will choose between a shift
2522 and a sequence of additions, so the observed cost is
2523 given as MIN (m * add_cost(speed, mode),
2524 shift_cost(speed, mode, m)). */
2525 op_cost
= m
* add_cost (speed
, mode
);
2526 if (shift_cost (speed
, mode
, m
) < op_cost
)
2527 op_cost
= shift_cost (speed
, mode
, m
);
2528 new_limit
.cost
= best_cost
.cost
- op_cost
;
2529 new_limit
.latency
= best_cost
.latency
- op_cost
;
2530 synth_mult (alg_in
, q
, &new_limit
, mode
);
2532 alg_in
->cost
.cost
+= op_cost
;
2533 alg_in
->cost
.latency
+= op_cost
;
2534 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2536 struct algorithm
*x
;
2537 best_cost
= alg_in
->cost
;
2538 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2539 best_alg
->log
[best_alg
->ops
] = m
;
2540 best_alg
->op
[best_alg
->ops
] = alg_shift
;
2548 /* If we have an odd number, add or subtract one. */
2551 unsigned HOST_WIDE_INT w
;
2554 for (w
= 1; (w
& t
) != 0; w
<<= 1)
2556 /* If T was -1, then W will be zero after the loop. This is another
2557 case where T ends with ...111. Handling this with (T + 1) and
2558 subtract 1 produces slightly better code and results in algorithm
2559 selection much faster than treating it like the ...0111 case
2563 /* Reject the case where t is 3.
2564 Thus we prefer addition in that case. */
2567 /* T ends with ...111. Multiply by (T + 1) and subtract 1. */
2569 op_cost
= add_cost (speed
, mode
);
2570 new_limit
.cost
= best_cost
.cost
- op_cost
;
2571 new_limit
.latency
= best_cost
.latency
- op_cost
;
2572 synth_mult (alg_in
, t
+ 1, &new_limit
, mode
);
2574 alg_in
->cost
.cost
+= op_cost
;
2575 alg_in
->cost
.latency
+= op_cost
;
2576 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2578 struct algorithm
*x
;
2579 best_cost
= alg_in
->cost
;
2580 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2581 best_alg
->log
[best_alg
->ops
] = 0;
2582 best_alg
->op
[best_alg
->ops
] = alg_sub_t_m2
;
2587 /* T ends with ...01 or ...011. Multiply by (T - 1) and add 1. */
2589 op_cost
= add_cost (speed
, mode
);
2590 new_limit
.cost
= best_cost
.cost
- op_cost
;
2591 new_limit
.latency
= best_cost
.latency
- op_cost
;
2592 synth_mult (alg_in
, t
- 1, &new_limit
, mode
);
2594 alg_in
->cost
.cost
+= op_cost
;
2595 alg_in
->cost
.latency
+= op_cost
;
2596 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2598 struct algorithm
*x
;
2599 best_cost
= alg_in
->cost
;
2600 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2601 best_alg
->log
[best_alg
->ops
] = 0;
2602 best_alg
->op
[best_alg
->ops
] = alg_add_t_m2
;
2606 /* We may be able to calculate a * -7, a * -15, a * -31, etc
2607 quickly with a - a * n for some appropriate constant n. */
2608 m
= exact_log2 (-orig_t
+ 1);
2609 if (m
>= 0 && m
< maxm
)
2611 op_cost
= shiftsub1_cost (speed
, mode
, m
);
2612 new_limit
.cost
= best_cost
.cost
- op_cost
;
2613 new_limit
.latency
= best_cost
.latency
- op_cost
;
2614 synth_mult (alg_in
, (unsigned HOST_WIDE_INT
) (-orig_t
+ 1) >> m
,
2617 alg_in
->cost
.cost
+= op_cost
;
2618 alg_in
->cost
.latency
+= op_cost
;
2619 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2621 struct algorithm
*x
;
2622 best_cost
= alg_in
->cost
;
2623 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2624 best_alg
->log
[best_alg
->ops
] = m
;
2625 best_alg
->op
[best_alg
->ops
] = alg_sub_t_m2
;
2633 /* Look for factors of t of the form
2634 t = q(2**m +- 1), 2 <= m <= floor(log2(t - 1)).
2635 If we find such a factor, we can multiply by t using an algorithm that
2636 multiplies by q, shift the result by m and add/subtract it to itself.
2638 We search for large factors first and loop down, even if large factors
2639 are less probable than small; if we find a large factor we will find a
2640 good sequence quickly, and therefore be able to prune (by decreasing
2641 COST_LIMIT) the search. */
2643 do_alg_addsub_factor
:
2644 for (m
= floor_log2 (t
- 1); m
>= 2; m
--)
2646 unsigned HOST_WIDE_INT d
;
2648 d
= ((unsigned HOST_WIDE_INT
) 1 << m
) + 1;
2649 if (t
% d
== 0 && t
> d
&& m
< maxm
2650 && (!cache_hit
|| cache_alg
== alg_add_factor
))
2652 /* If the target has a cheap shift-and-add instruction use
2653 that in preference to a shift insn followed by an add insn.
2654 Assume that the shift-and-add is "atomic" with a latency
2655 equal to its cost, otherwise assume that on superscalar
2656 hardware the shift may be executed concurrently with the
2657 earlier steps in the algorithm. */
2658 op_cost
= add_cost (speed
, mode
) + shift_cost (speed
, mode
, m
);
2659 if (shiftadd_cost (speed
, mode
, m
) < op_cost
)
2661 op_cost
= shiftadd_cost (speed
, mode
, m
);
2662 op_latency
= op_cost
;
2665 op_latency
= add_cost (speed
, mode
);
2667 new_limit
.cost
= best_cost
.cost
- op_cost
;
2668 new_limit
.latency
= best_cost
.latency
- op_latency
;
2669 synth_mult (alg_in
, t
/ d
, &new_limit
, mode
);
2671 alg_in
->cost
.cost
+= op_cost
;
2672 alg_in
->cost
.latency
+= op_latency
;
2673 if (alg_in
->cost
.latency
< op_cost
)
2674 alg_in
->cost
.latency
= op_cost
;
2675 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2677 struct algorithm
*x
;
2678 best_cost
= alg_in
->cost
;
2679 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2680 best_alg
->log
[best_alg
->ops
] = m
;
2681 best_alg
->op
[best_alg
->ops
] = alg_add_factor
;
2683 /* Other factors will have been taken care of in the recursion. */
2687 d
= ((unsigned HOST_WIDE_INT
) 1 << m
) - 1;
2688 if (t
% d
== 0 && t
> d
&& m
< maxm
2689 && (!cache_hit
|| cache_alg
== alg_sub_factor
))
2691 /* If the target has a cheap shift-and-subtract insn use
2692 that in preference to a shift insn followed by a sub insn.
2693 Assume that the shift-and-sub is "atomic" with a latency
2694 equal to it's cost, otherwise assume that on superscalar
2695 hardware the shift may be executed concurrently with the
2696 earlier steps in the algorithm. */
2697 op_cost
= add_cost (speed
, mode
) + shift_cost (speed
, mode
, m
);
2698 if (shiftsub0_cost (speed
, mode
, m
) < op_cost
)
2700 op_cost
= shiftsub0_cost (speed
, mode
, m
);
2701 op_latency
= op_cost
;
2704 op_latency
= add_cost (speed
, mode
);
2706 new_limit
.cost
= best_cost
.cost
- op_cost
;
2707 new_limit
.latency
= best_cost
.latency
- op_latency
;
2708 synth_mult (alg_in
, t
/ d
, &new_limit
, mode
);
2710 alg_in
->cost
.cost
+= op_cost
;
2711 alg_in
->cost
.latency
+= op_latency
;
2712 if (alg_in
->cost
.latency
< op_cost
)
2713 alg_in
->cost
.latency
= op_cost
;
2714 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2716 struct algorithm
*x
;
2717 best_cost
= alg_in
->cost
;
2718 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2719 best_alg
->log
[best_alg
->ops
] = m
;
2720 best_alg
->op
[best_alg
->ops
] = alg_sub_factor
;
2728 /* Try shift-and-add (load effective address) instructions,
2729 i.e. do a*3, a*5, a*9. */
2736 if (m
>= 0 && m
< maxm
)
2738 op_cost
= shiftadd_cost (speed
, mode
, m
);
2739 new_limit
.cost
= best_cost
.cost
- op_cost
;
2740 new_limit
.latency
= best_cost
.latency
- op_cost
;
2741 synth_mult (alg_in
, (t
- 1) >> m
, &new_limit
, mode
);
2743 alg_in
->cost
.cost
+= op_cost
;
2744 alg_in
->cost
.latency
+= op_cost
;
2745 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2747 struct algorithm
*x
;
2748 best_cost
= alg_in
->cost
;
2749 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2750 best_alg
->log
[best_alg
->ops
] = m
;
2751 best_alg
->op
[best_alg
->ops
] = alg_add_t2_m
;
2761 if (m
>= 0 && m
< maxm
)
2763 op_cost
= shiftsub0_cost (speed
, mode
, m
);
2764 new_limit
.cost
= best_cost
.cost
- op_cost
;
2765 new_limit
.latency
= best_cost
.latency
- op_cost
;
2766 synth_mult (alg_in
, (t
+ 1) >> m
, &new_limit
, mode
);
2768 alg_in
->cost
.cost
+= op_cost
;
2769 alg_in
->cost
.latency
+= op_cost
;
2770 if (CHEAPER_MULT_COST (&alg_in
->cost
, &best_cost
))
2772 struct algorithm
*x
;
2773 best_cost
= alg_in
->cost
;
2774 x
= alg_in
, alg_in
= best_alg
, best_alg
= x
;
2775 best_alg
->log
[best_alg
->ops
] = m
;
2776 best_alg
->op
[best_alg
->ops
] = alg_sub_t2_m
;
2784 /* If best_cost has not decreased, we have not found any algorithm. */
2785 if (!CHEAPER_MULT_COST (&best_cost
, cost_limit
))
2787 /* We failed to find an algorithm. Record alg_impossible for
2788 this case (that is, <T, MODE, COST_LIMIT>) so that next time
2789 we are asked to find an algorithm for T within the same or
2790 lower COST_LIMIT, we can immediately return to the
2793 entry_ptr
->mode
= mode
;
2794 entry_ptr
->speed
= speed
;
2795 entry_ptr
->alg
= alg_impossible
;
2796 entry_ptr
->cost
= *cost_limit
;
2800 /* Cache the result. */
2804 entry_ptr
->mode
= mode
;
2805 entry_ptr
->speed
= speed
;
2806 entry_ptr
->alg
= best_alg
->op
[best_alg
->ops
];
2807 entry_ptr
->cost
.cost
= best_cost
.cost
;
2808 entry_ptr
->cost
.latency
= best_cost
.latency
;
2811 /* If we are getting a too long sequence for `struct algorithm'
2812 to record, make this search fail. */
2813 if (best_alg
->ops
== MAX_BITS_PER_WORD
)
2816 /* Copy the algorithm from temporary space to the space at alg_out.
2817 We avoid using structure assignment because the majority of
2818 best_alg is normally undefined, and this is a critical function. */
2819 alg_out
->ops
= best_alg
->ops
+ 1;
2820 alg_out
->cost
= best_cost
;
2821 memcpy (alg_out
->op
, best_alg
->op
,
2822 alg_out
->ops
* sizeof *alg_out
->op
);
2823 memcpy (alg_out
->log
, best_alg
->log
,
2824 alg_out
->ops
* sizeof *alg_out
->log
);
2827 /* Find the cheapest way of multiplying a value of mode MODE by VAL.
2828 Try three variations:
2830 - a shift/add sequence based on VAL itself
2831 - a shift/add sequence based on -VAL, followed by a negation
2832 - a shift/add sequence based on VAL - 1, followed by an addition.
2834 Return true if the cheapest of these cost less than MULT_COST,
2835 describing the algorithm in *ALG and final fixup in *VARIANT. */
2838 choose_mult_variant (machine_mode mode
, HOST_WIDE_INT val
,
2839 struct algorithm
*alg
, enum mult_variant
*variant
,
2842 struct algorithm alg2
;
2843 struct mult_cost limit
;
2845 bool speed
= optimize_insn_for_speed_p ();
2847 /* Fail quickly for impossible bounds. */
2851 /* Ensure that mult_cost provides a reasonable upper bound.
2852 Any constant multiplication can be performed with less
2853 than 2 * bits additions. */
2854 op_cost
= 2 * GET_MODE_UNIT_BITSIZE (mode
) * add_cost (speed
, mode
);
2855 if (mult_cost
> op_cost
)
2856 mult_cost
= op_cost
;
2858 *variant
= basic_variant
;
2859 limit
.cost
= mult_cost
;
2860 limit
.latency
= mult_cost
;
2861 synth_mult (alg
, val
, &limit
, mode
);
2863 /* This works only if the inverted value actually fits in an
2865 if (HOST_BITS_PER_INT
>= GET_MODE_UNIT_BITSIZE (mode
))
2867 op_cost
= neg_cost (speed
, mode
);
2868 if (MULT_COST_LESS (&alg
->cost
, mult_cost
))
2870 limit
.cost
= alg
->cost
.cost
- op_cost
;
2871 limit
.latency
= alg
->cost
.latency
- op_cost
;
2875 limit
.cost
= mult_cost
- op_cost
;
2876 limit
.latency
= mult_cost
- op_cost
;
2879 synth_mult (&alg2
, -val
, &limit
, mode
);
2880 alg2
.cost
.cost
+= op_cost
;
2881 alg2
.cost
.latency
+= op_cost
;
2882 if (CHEAPER_MULT_COST (&alg2
.cost
, &alg
->cost
))
2883 *alg
= alg2
, *variant
= negate_variant
;
2886 /* This proves very useful for division-by-constant. */
2887 op_cost
= add_cost (speed
, mode
);
2888 if (MULT_COST_LESS (&alg
->cost
, mult_cost
))
2890 limit
.cost
= alg
->cost
.cost
- op_cost
;
2891 limit
.latency
= alg
->cost
.latency
- op_cost
;
2895 limit
.cost
= mult_cost
- op_cost
;
2896 limit
.latency
= mult_cost
- op_cost
;
2899 synth_mult (&alg2
, val
- 1, &limit
, mode
);
2900 alg2
.cost
.cost
+= op_cost
;
2901 alg2
.cost
.latency
+= op_cost
;
2902 if (CHEAPER_MULT_COST (&alg2
.cost
, &alg
->cost
))
2903 *alg
= alg2
, *variant
= add_variant
;
2905 return MULT_COST_LESS (&alg
->cost
, mult_cost
);
2908 /* A subroutine of expand_mult, used for constant multiplications.
2909 Multiply OP0 by VAL in mode MODE, storing the result in TARGET if
2910 convenient. Use the shift/add sequence described by ALG and apply
2911 the final fixup specified by VARIANT. */
2914 expand_mult_const (machine_mode mode
, rtx op0
, HOST_WIDE_INT val
,
2915 rtx target
, const struct algorithm
*alg
,
2916 enum mult_variant variant
)
2918 HOST_WIDE_INT val_so_far
;
2924 /* Avoid referencing memory over and over and invalid sharing
2926 op0
= force_reg (mode
, op0
);
2928 /* ACCUM starts out either as OP0 or as a zero, depending on
2929 the first operation. */
2931 if (alg
->op
[0] == alg_zero
)
2933 accum
= copy_to_mode_reg (mode
, CONST0_RTX (mode
));
2936 else if (alg
->op
[0] == alg_m
)
2938 accum
= copy_to_mode_reg (mode
, op0
);
2944 for (opno
= 1; opno
< alg
->ops
; opno
++)
2946 int log
= alg
->log
[opno
];
2947 rtx shift_subtarget
= optimize
? 0 : accum
;
2949 = (opno
== alg
->ops
- 1 && target
!= 0 && variant
!= add_variant
2952 rtx accum_target
= optimize
? 0 : accum
;
2955 switch (alg
->op
[opno
])
2958 tem
= expand_shift (LSHIFT_EXPR
, mode
, accum
, log
, NULL_RTX
, 0);
2959 /* REG_EQUAL note will be attached to the following insn. */
2960 emit_move_insn (accum
, tem
);
2965 tem
= expand_shift (LSHIFT_EXPR
, mode
, op0
, log
, NULL_RTX
, 0);
2966 accum
= force_operand (gen_rtx_PLUS (mode
, accum
, tem
),
2967 add_target
? add_target
: accum_target
);
2968 val_so_far
+= (HOST_WIDE_INT
) 1 << log
;
2972 tem
= expand_shift (LSHIFT_EXPR
, mode
, op0
, log
, NULL_RTX
, 0);
2973 accum
= force_operand (gen_rtx_MINUS (mode
, accum
, tem
),
2974 add_target
? add_target
: accum_target
);
2975 val_so_far
-= (HOST_WIDE_INT
) 1 << log
;
2979 accum
= expand_shift (LSHIFT_EXPR
, mode
, accum
,
2980 log
, shift_subtarget
, 0);
2981 accum
= force_operand (gen_rtx_PLUS (mode
, accum
, op0
),
2982 add_target
? add_target
: accum_target
);
2983 val_so_far
= (val_so_far
<< log
) + 1;
2987 accum
= expand_shift (LSHIFT_EXPR
, mode
, accum
,
2988 log
, shift_subtarget
, 0);
2989 accum
= force_operand (gen_rtx_MINUS (mode
, accum
, op0
),
2990 add_target
? add_target
: accum_target
);
2991 val_so_far
= (val_so_far
<< log
) - 1;
2994 case alg_add_factor
:
2995 tem
= expand_shift (LSHIFT_EXPR
, mode
, accum
, log
, NULL_RTX
, 0);
2996 accum
= force_operand (gen_rtx_PLUS (mode
, accum
, tem
),
2997 add_target
? add_target
: accum_target
);
2998 val_so_far
+= val_so_far
<< log
;
3001 case alg_sub_factor
:
3002 tem
= expand_shift (LSHIFT_EXPR
, mode
, accum
, log
, NULL_RTX
, 0);
3003 accum
= force_operand (gen_rtx_MINUS (mode
, tem
, accum
),
3005 ? add_target
: (optimize
? 0 : tem
)));
3006 val_so_far
= (val_so_far
<< log
) - val_so_far
;
3013 if (SCALAR_INT_MODE_P (mode
))
3015 /* Write a REG_EQUAL note on the last insn so that we can cse
3016 multiplication sequences. Note that if ACCUM is a SUBREG,
3017 we've set the inner register and must properly indicate that. */
3018 tem
= op0
, nmode
= mode
;
3019 accum_inner
= accum
;
3020 if (GET_CODE (accum
) == SUBREG
)
3022 accum_inner
= SUBREG_REG (accum
);
3023 nmode
= GET_MODE (accum_inner
);
3024 tem
= gen_lowpart (nmode
, op0
);
3027 insn
= get_last_insn ();
3028 set_dst_reg_note (insn
, REG_EQUAL
,
3029 gen_rtx_MULT (nmode
, tem
,
3030 gen_int_mode (val_so_far
, nmode
)),
3035 if (variant
== negate_variant
)
3037 val_so_far
= -val_so_far
;
3038 accum
= expand_unop (mode
, neg_optab
, accum
, target
, 0);
3040 else if (variant
== add_variant
)
3042 val_so_far
= val_so_far
+ 1;
3043 accum
= force_operand (gen_rtx_PLUS (mode
, accum
, op0
), target
);
3046 /* Compare only the bits of val and val_so_far that are significant
3047 in the result mode, to avoid sign-/zero-extension confusion. */
3048 nmode
= GET_MODE_INNER (mode
);
3049 if (nmode
== VOIDmode
)
3051 val
&= GET_MODE_MASK (nmode
);
3052 val_so_far
&= GET_MODE_MASK (nmode
);
3053 gcc_assert (val
== val_so_far
);
3058 /* Perform a multiplication and return an rtx for the result.
3059 MODE is mode of value; OP0 and OP1 are what to multiply (rtx's);
3060 TARGET is a suggestion for where to store the result (an rtx).
3062 We check specially for a constant integer as OP1.
3063 If you want this check for OP0 as well, then before calling
3064 you should swap the two operands if OP0 would be constant. */
3067 expand_mult (machine_mode mode
, rtx op0
, rtx op1
, rtx target
,
3070 enum mult_variant variant
;
3071 struct algorithm algorithm
;
3074 bool speed
= optimize_insn_for_speed_p ();
3075 bool do_trapv
= flag_trapv
&& SCALAR_INT_MODE_P (mode
) && !unsignedp
;
3077 if (CONSTANT_P (op0
))
3084 /* For vectors, there are several simplifications that can be made if
3085 all elements of the vector constant are identical. */
3087 if (GET_CODE (op1
) == CONST_VECTOR
)
3089 int i
, n
= CONST_VECTOR_NUNITS (op1
);
3090 scalar_op1
= CONST_VECTOR_ELT (op1
, 0);
3091 for (i
= 1; i
< n
; ++i
)
3092 if (!rtx_equal_p (scalar_op1
, CONST_VECTOR_ELT (op1
, i
)))
3096 if (INTEGRAL_MODE_P (mode
))
3099 HOST_WIDE_INT coeff
;
3103 if (op1
== CONST0_RTX (mode
))
3105 if (op1
== CONST1_RTX (mode
))
3107 if (op1
== CONSTM1_RTX (mode
))
3108 return expand_unop (mode
, do_trapv
? negv_optab
: neg_optab
,
3114 /* If mode is integer vector mode, check if the backend supports
3115 vector lshift (by scalar or vector) at all. If not, we can't use
3116 synthetized multiply. */
3117 if (GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
3118 && optab_handler (vashl_optab
, mode
) == CODE_FOR_nothing
3119 && optab_handler (ashl_optab
, mode
) == CODE_FOR_nothing
)
3122 /* These are the operations that are potentially turned into
3123 a sequence of shifts and additions. */
3124 mode_bitsize
= GET_MODE_UNIT_BITSIZE (mode
);
3126 /* synth_mult does an `unsigned int' multiply. As long as the mode is
3127 less than or equal in size to `unsigned int' this doesn't matter.
3128 If the mode is larger than `unsigned int', then synth_mult works
3129 only if the constant value exactly fits in an `unsigned int' without
3130 any truncation. This means that multiplying by negative values does
3131 not work; results are off by 2^32 on a 32 bit machine. */
3132 if (CONST_INT_P (scalar_op1
))
3134 coeff
= INTVAL (scalar_op1
);
3137 #if TARGET_SUPPORTS_WIDE_INT
3138 else if (CONST_WIDE_INT_P (scalar_op1
))
3140 else if (CONST_DOUBLE_AS_INT_P (scalar_op1
))
3143 int shift
= wi::exact_log2 (std::make_pair (scalar_op1
, mode
));
3144 /* Perfect power of 2 (other than 1, which is handled above). */
3146 return expand_shift (LSHIFT_EXPR
, mode
, op0
,
3147 shift
, target
, unsignedp
);
3154 /* We used to test optimize here, on the grounds that it's better to
3155 produce a smaller program when -O is not used. But this causes
3156 such a terrible slowdown sometimes that it seems better to always
3159 /* Special case powers of two. */
3160 if (EXACT_POWER_OF_2_OR_ZERO_P (coeff
)
3161 && !(is_neg
&& mode_bitsize
> HOST_BITS_PER_WIDE_INT
))
3162 return expand_shift (LSHIFT_EXPR
, mode
, op0
,
3163 floor_log2 (coeff
), target
, unsignedp
);
3165 fake_reg
= gen_raw_REG (mode
, LAST_VIRTUAL_REGISTER
+ 1);
3167 /* Attempt to handle multiplication of DImode values by negative
3168 coefficients, by performing the multiplication by a positive
3169 multiplier and then inverting the result. */
3170 if (is_neg
&& mode_bitsize
> HOST_BITS_PER_WIDE_INT
)
3172 /* Its safe to use -coeff even for INT_MIN, as the
3173 result is interpreted as an unsigned coefficient.
3174 Exclude cost of op0 from max_cost to match the cost
3175 calculation of the synth_mult. */
3176 coeff
= -(unsigned HOST_WIDE_INT
) coeff
;
3177 max_cost
= (set_src_cost (gen_rtx_MULT (mode
, fake_reg
, op1
), speed
)
3178 - neg_cost (speed
, mode
));
3182 /* Special case powers of two. */
3183 if (EXACT_POWER_OF_2_OR_ZERO_P (coeff
))
3185 rtx temp
= expand_shift (LSHIFT_EXPR
, mode
, op0
,
3186 floor_log2 (coeff
), target
, unsignedp
);
3187 return expand_unop (mode
, neg_optab
, temp
, target
, 0);
3190 if (choose_mult_variant (mode
, coeff
, &algorithm
, &variant
,
3193 rtx temp
= expand_mult_const (mode
, op0
, coeff
, NULL_RTX
,
3194 &algorithm
, variant
);
3195 return expand_unop (mode
, neg_optab
, temp
, target
, 0);
3200 /* Exclude cost of op0 from max_cost to match the cost
3201 calculation of the synth_mult. */
3202 max_cost
= set_src_cost (gen_rtx_MULT (mode
, fake_reg
, op1
), speed
);
3203 if (choose_mult_variant (mode
, coeff
, &algorithm
, &variant
, max_cost
))
3204 return expand_mult_const (mode
, op0
, coeff
, target
,
3205 &algorithm
, variant
);
3209 /* Expand x*2.0 as x+x. */
3210 if (CONST_DOUBLE_AS_FLOAT_P (scalar_op1
))
3213 REAL_VALUE_FROM_CONST_DOUBLE (d
, scalar_op1
);
3215 if (REAL_VALUES_EQUAL (d
, dconst2
))
3217 op0
= force_reg (GET_MODE (op0
), op0
);
3218 return expand_binop (mode
, add_optab
, op0
, op0
,
3219 target
, unsignedp
, OPTAB_LIB_WIDEN
);
3224 /* This used to use umul_optab if unsigned, but for non-widening multiply
3225 there is no difference between signed and unsigned. */
3226 op0
= expand_binop (mode
, do_trapv
? smulv_optab
: smul_optab
,
3227 op0
, op1
, target
, unsignedp
, OPTAB_LIB_WIDEN
);
3232 /* Return a cost estimate for multiplying a register by the given
3233 COEFFicient in the given MODE and SPEED. */
3236 mult_by_coeff_cost (HOST_WIDE_INT coeff
, machine_mode mode
, bool speed
)
3239 struct algorithm algorithm
;
3240 enum mult_variant variant
;
3242 rtx fake_reg
= gen_raw_REG (mode
, LAST_VIRTUAL_REGISTER
+ 1);
3243 max_cost
= set_src_cost (gen_rtx_MULT (mode
, fake_reg
, fake_reg
), speed
);
3244 if (choose_mult_variant (mode
, coeff
, &algorithm
, &variant
, max_cost
))
3245 return algorithm
.cost
.cost
;
3250 /* Perform a widening multiplication and return an rtx for the result.
3251 MODE is mode of value; OP0 and OP1 are what to multiply (rtx's);
3252 TARGET is a suggestion for where to store the result (an rtx).
3253 THIS_OPTAB is the optab we should use, it must be either umul_widen_optab
3254 or smul_widen_optab.
3256 We check specially for a constant integer as OP1, comparing the
3257 cost of a widening multiply against the cost of a sequence of shifts
3261 expand_widening_mult (machine_mode mode
, rtx op0
, rtx op1
, rtx target
,
3262 int unsignedp
, optab this_optab
)
3264 bool speed
= optimize_insn_for_speed_p ();
3267 if (CONST_INT_P (op1
)
3268 && GET_MODE (op0
) != VOIDmode
3269 && (cop1
= convert_modes (mode
, GET_MODE (op0
), op1
,
3270 this_optab
== umul_widen_optab
))
3271 && CONST_INT_P (cop1
)
3272 && (INTVAL (cop1
) >= 0
3273 || HWI_COMPUTABLE_MODE_P (mode
)))
3275 HOST_WIDE_INT coeff
= INTVAL (cop1
);
3277 enum mult_variant variant
;
3278 struct algorithm algorithm
;
3280 /* Special case powers of two. */
3281 if (EXACT_POWER_OF_2_OR_ZERO_P (coeff
))
3283 op0
= convert_to_mode (mode
, op0
, this_optab
== umul_widen_optab
);
3284 return expand_shift (LSHIFT_EXPR
, mode
, op0
,
3285 floor_log2 (coeff
), target
, unsignedp
);
3288 /* Exclude cost of op0 from max_cost to match the cost
3289 calculation of the synth_mult. */
3290 max_cost
= mul_widen_cost (speed
, mode
);
3291 if (choose_mult_variant (mode
, coeff
, &algorithm
, &variant
,
3294 op0
= convert_to_mode (mode
, op0
, this_optab
== umul_widen_optab
);
3295 return expand_mult_const (mode
, op0
, coeff
, target
,
3296 &algorithm
, variant
);
3299 return expand_binop (mode
, this_optab
, op0
, op1
, target
,
3300 unsignedp
, OPTAB_LIB_WIDEN
);
3303 /* Choose a minimal N + 1 bit approximation to 1/D that can be used to
3304 replace division by D, and put the least significant N bits of the result
3305 in *MULTIPLIER_PTR and return the most significant bit.
3307 The width of operations is N (should be <= HOST_BITS_PER_WIDE_INT), the
3308 needed precision is in PRECISION (should be <= N).
3310 PRECISION should be as small as possible so this function can choose
3311 multiplier more freely.
3313 The rounded-up logarithm of D is placed in *lgup_ptr. A shift count that
3314 is to be used for a final right shift is placed in *POST_SHIFT_PTR.
3316 Using this function, x/D will be equal to (x * m) >> (*POST_SHIFT_PTR),
3317 where m is the full HOST_BITS_PER_WIDE_INT + 1 bit multiplier. */
3319 unsigned HOST_WIDE_INT
3320 choose_multiplier (unsigned HOST_WIDE_INT d
, int n
, int precision
,
3321 unsigned HOST_WIDE_INT
*multiplier_ptr
,
3322 int *post_shift_ptr
, int *lgup_ptr
)
3324 int lgup
, post_shift
;
3327 /* lgup = ceil(log2(divisor)); */
3328 lgup
= ceil_log2 (d
);
3330 gcc_assert (lgup
<= n
);
3333 pow2
= n
+ lgup
- precision
;
3335 /* mlow = 2^(N + lgup)/d */
3336 wide_int val
= wi::set_bit_in_zero (pow
, HOST_BITS_PER_DOUBLE_INT
);
3337 wide_int mlow
= wi::udiv_trunc (val
, d
);
3339 /* mhigh = (2^(N + lgup) + 2^(N + lgup - precision))/d */
3340 val
|= wi::set_bit_in_zero (pow2
, HOST_BITS_PER_DOUBLE_INT
);
3341 wide_int mhigh
= wi::udiv_trunc (val
, d
);
3343 /* If precision == N, then mlow, mhigh exceed 2^N
3344 (but they do not exceed 2^(N+1)). */
3346 /* Reduce to lowest terms. */
3347 for (post_shift
= lgup
; post_shift
> 0; post_shift
--)
3349 unsigned HOST_WIDE_INT ml_lo
= wi::extract_uhwi (mlow
, 1,
3350 HOST_BITS_PER_WIDE_INT
);
3351 unsigned HOST_WIDE_INT mh_lo
= wi::extract_uhwi (mhigh
, 1,
3352 HOST_BITS_PER_WIDE_INT
);
3356 mlow
= wi::uhwi (ml_lo
, HOST_BITS_PER_DOUBLE_INT
);
3357 mhigh
= wi::uhwi (mh_lo
, HOST_BITS_PER_DOUBLE_INT
);
3360 *post_shift_ptr
= post_shift
;
3362 if (n
< HOST_BITS_PER_WIDE_INT
)
3364 unsigned HOST_WIDE_INT mask
= ((unsigned HOST_WIDE_INT
) 1 << n
) - 1;
3365 *multiplier_ptr
= mhigh
.to_uhwi () & mask
;
3366 return mhigh
.to_uhwi () >= mask
;
3370 *multiplier_ptr
= mhigh
.to_uhwi ();
3371 return wi::extract_uhwi (mhigh
, HOST_BITS_PER_WIDE_INT
, 1);
3375 /* Compute the inverse of X mod 2**n, i.e., find Y such that X * Y is
3376 congruent to 1 (mod 2**N). */
3378 static unsigned HOST_WIDE_INT
3379 invert_mod2n (unsigned HOST_WIDE_INT x
, int n
)
3381 /* Solve x*y == 1 (mod 2^n), where x is odd. Return y. */
3383 /* The algorithm notes that the choice y = x satisfies
3384 x*y == 1 mod 2^3, since x is assumed odd.
3385 Each iteration doubles the number of bits of significance in y. */
3387 unsigned HOST_WIDE_INT mask
;
3388 unsigned HOST_WIDE_INT y
= x
;
3391 mask
= (n
== HOST_BITS_PER_WIDE_INT
3392 ? ~(unsigned HOST_WIDE_INT
) 0
3393 : ((unsigned HOST_WIDE_INT
) 1 << n
) - 1);
3397 y
= y
* (2 - x
*y
) & mask
; /* Modulo 2^N */
3403 /* Emit code to adjust ADJ_OPERAND after multiplication of wrong signedness
3404 flavor of OP0 and OP1. ADJ_OPERAND is already the high half of the
3405 product OP0 x OP1. If UNSIGNEDP is nonzero, adjust the signed product
3406 to become unsigned, if UNSIGNEDP is zero, adjust the unsigned product to
3409 The result is put in TARGET if that is convenient.
3411 MODE is the mode of operation. */
3414 expand_mult_highpart_adjust (machine_mode mode
, rtx adj_operand
, rtx op0
,
3415 rtx op1
, rtx target
, int unsignedp
)
3418 enum rtx_code adj_code
= unsignedp
? PLUS
: MINUS
;
3420 tem
= expand_shift (RSHIFT_EXPR
, mode
, op0
,
3421 GET_MODE_BITSIZE (mode
) - 1, NULL_RTX
, 0);
3422 tem
= expand_and (mode
, tem
, op1
, NULL_RTX
);
3424 = force_operand (gen_rtx_fmt_ee (adj_code
, mode
, adj_operand
, tem
),
3427 tem
= expand_shift (RSHIFT_EXPR
, mode
, op1
,
3428 GET_MODE_BITSIZE (mode
) - 1, NULL_RTX
, 0);
3429 tem
= expand_and (mode
, tem
, op0
, NULL_RTX
);
3430 target
= force_operand (gen_rtx_fmt_ee (adj_code
, mode
, adj_operand
, tem
),
3436 /* Subroutine of expmed_mult_highpart. Return the MODE high part of OP. */
3439 extract_high_half (machine_mode mode
, rtx op
)
3441 machine_mode wider_mode
;
3443 if (mode
== word_mode
)
3444 return gen_highpart (mode
, op
);
3446 gcc_assert (!SCALAR_FLOAT_MODE_P (mode
));
3448 wider_mode
= GET_MODE_WIDER_MODE (mode
);
3449 op
= expand_shift (RSHIFT_EXPR
, wider_mode
, op
,
3450 GET_MODE_BITSIZE (mode
), 0, 1);
3451 return convert_modes (mode
, wider_mode
, op
, 0);
3454 /* Like expmed_mult_highpart, but only consider using a multiplication
3455 optab. OP1 is an rtx for the constant operand. */
3458 expmed_mult_highpart_optab (machine_mode mode
, rtx op0
, rtx op1
,
3459 rtx target
, int unsignedp
, int max_cost
)
3461 rtx narrow_op1
= gen_int_mode (INTVAL (op1
), mode
);
3462 machine_mode wider_mode
;
3466 bool speed
= optimize_insn_for_speed_p ();
3468 gcc_assert (!SCALAR_FLOAT_MODE_P (mode
));
3470 wider_mode
= GET_MODE_WIDER_MODE (mode
);
3471 size
= GET_MODE_BITSIZE (mode
);
3473 /* Firstly, try using a multiplication insn that only generates the needed
3474 high part of the product, and in the sign flavor of unsignedp. */
3475 if (mul_highpart_cost (speed
, mode
) < max_cost
)
3477 moptab
= unsignedp
? umul_highpart_optab
: smul_highpart_optab
;
3478 tem
= expand_binop (mode
, moptab
, op0
, narrow_op1
, target
,
3479 unsignedp
, OPTAB_DIRECT
);
3484 /* Secondly, same as above, but use sign flavor opposite of unsignedp.
3485 Need to adjust the result after the multiplication. */
3486 if (size
- 1 < BITS_PER_WORD
3487 && (mul_highpart_cost (speed
, mode
)
3488 + 2 * shift_cost (speed
, mode
, size
-1)
3489 + 4 * add_cost (speed
, mode
) < max_cost
))
3491 moptab
= unsignedp
? smul_highpart_optab
: umul_highpart_optab
;
3492 tem
= expand_binop (mode
, moptab
, op0
, narrow_op1
, target
,
3493 unsignedp
, OPTAB_DIRECT
);
3495 /* We used the wrong signedness. Adjust the result. */
3496 return expand_mult_highpart_adjust (mode
, tem
, op0
, narrow_op1
,
3500 /* Try widening multiplication. */
3501 moptab
= unsignedp
? umul_widen_optab
: smul_widen_optab
;
3502 if (widening_optab_handler (moptab
, wider_mode
, mode
) != CODE_FOR_nothing
3503 && mul_widen_cost (speed
, wider_mode
) < max_cost
)
3505 tem
= expand_binop (wider_mode
, moptab
, op0
, narrow_op1
, 0,
3506 unsignedp
, OPTAB_WIDEN
);
3508 return extract_high_half (mode
, tem
);
3511 /* Try widening the mode and perform a non-widening multiplication. */
3512 if (optab_handler (smul_optab
, wider_mode
) != CODE_FOR_nothing
3513 && size
- 1 < BITS_PER_WORD
3514 && (mul_cost (speed
, wider_mode
) + shift_cost (speed
, mode
, size
-1)
3520 /* We need to widen the operands, for example to ensure the
3521 constant multiplier is correctly sign or zero extended.
3522 Use a sequence to clean-up any instructions emitted by
3523 the conversions if things don't work out. */
3525 wop0
= convert_modes (wider_mode
, mode
, op0
, unsignedp
);
3526 wop1
= convert_modes (wider_mode
, mode
, op1
, unsignedp
);
3527 tem
= expand_binop (wider_mode
, smul_optab
, wop0
, wop1
, 0,
3528 unsignedp
, OPTAB_WIDEN
);
3529 insns
= get_insns ();
3535 return extract_high_half (mode
, tem
);
3539 /* Try widening multiplication of opposite signedness, and adjust. */
3540 moptab
= unsignedp
? smul_widen_optab
: umul_widen_optab
;
3541 if (widening_optab_handler (moptab
, wider_mode
, mode
) != CODE_FOR_nothing
3542 && size
- 1 < BITS_PER_WORD
3543 && (mul_widen_cost (speed
, wider_mode
)
3544 + 2 * shift_cost (speed
, mode
, size
-1)
3545 + 4 * add_cost (speed
, mode
) < max_cost
))
3547 tem
= expand_binop (wider_mode
, moptab
, op0
, narrow_op1
,
3548 NULL_RTX
, ! unsignedp
, OPTAB_WIDEN
);
3551 tem
= extract_high_half (mode
, tem
);
3552 /* We used the wrong signedness. Adjust the result. */
3553 return expand_mult_highpart_adjust (mode
, tem
, op0
, narrow_op1
,
3561 /* Emit code to multiply OP0 and OP1 (where OP1 is an integer constant),
3562 putting the high half of the result in TARGET if that is convenient,
3563 and return where the result is. If the operation can not be performed,
3566 MODE is the mode of operation and result.
3568 UNSIGNEDP nonzero means unsigned multiply.
3570 MAX_COST is the total allowed cost for the expanded RTL. */
3573 expmed_mult_highpart (machine_mode mode
, rtx op0
, rtx op1
,
3574 rtx target
, int unsignedp
, int max_cost
)
3576 machine_mode wider_mode
= GET_MODE_WIDER_MODE (mode
);
3577 unsigned HOST_WIDE_INT cnst1
;
3579 bool sign_adjust
= false;
3580 enum mult_variant variant
;
3581 struct algorithm alg
;
3583 bool speed
= optimize_insn_for_speed_p ();
3585 gcc_assert (!SCALAR_FLOAT_MODE_P (mode
));
3586 /* We can't support modes wider than HOST_BITS_PER_INT. */
3587 gcc_assert (HWI_COMPUTABLE_MODE_P (mode
));
3589 cnst1
= INTVAL (op1
) & GET_MODE_MASK (mode
);
3591 /* We can't optimize modes wider than BITS_PER_WORD.
3592 ??? We might be able to perform double-word arithmetic if
3593 mode == word_mode, however all the cost calculations in
3594 synth_mult etc. assume single-word operations. */
3595 if (GET_MODE_BITSIZE (wider_mode
) > BITS_PER_WORD
)
3596 return expmed_mult_highpart_optab (mode
, op0
, op1
, target
,
3597 unsignedp
, max_cost
);
3599 extra_cost
= shift_cost (speed
, mode
, GET_MODE_BITSIZE (mode
) - 1);
3601 /* Check whether we try to multiply by a negative constant. */
3602 if (!unsignedp
&& ((cnst1
>> (GET_MODE_BITSIZE (mode
) - 1)) & 1))
3605 extra_cost
+= add_cost (speed
, mode
);
3608 /* See whether shift/add multiplication is cheap enough. */
3609 if (choose_mult_variant (wider_mode
, cnst1
, &alg
, &variant
,
3610 max_cost
- extra_cost
))
3612 /* See whether the specialized multiplication optabs are
3613 cheaper than the shift/add version. */
3614 tem
= expmed_mult_highpart_optab (mode
, op0
, op1
, target
, unsignedp
,
3615 alg
.cost
.cost
+ extra_cost
);
3619 tem
= convert_to_mode (wider_mode
, op0
, unsignedp
);
3620 tem
= expand_mult_const (wider_mode
, tem
, cnst1
, 0, &alg
, variant
);
3621 tem
= extract_high_half (mode
, tem
);
3623 /* Adjust result for signedness. */
3625 tem
= force_operand (gen_rtx_MINUS (mode
, tem
, op0
), tem
);
3629 return expmed_mult_highpart_optab (mode
, op0
, op1
, target
,
3630 unsignedp
, max_cost
);
3634 /* Expand signed modulus of OP0 by a power of two D in mode MODE. */
3637 expand_smod_pow2 (machine_mode mode
, rtx op0
, HOST_WIDE_INT d
)
3639 rtx result
, temp
, shift
;
3640 rtx_code_label
*label
;
3642 int prec
= GET_MODE_PRECISION (mode
);
3644 logd
= floor_log2 (d
);
3645 result
= gen_reg_rtx (mode
);
3647 /* Avoid conditional branches when they're expensive. */
3648 if (BRANCH_COST (optimize_insn_for_speed_p (), false) >= 2
3649 && optimize_insn_for_speed_p ())
3651 rtx signmask
= emit_store_flag (result
, LT
, op0
, const0_rtx
,
3655 HOST_WIDE_INT masklow
= ((HOST_WIDE_INT
) 1 << logd
) - 1;
3656 signmask
= force_reg (mode
, signmask
);
3657 shift
= GEN_INT (GET_MODE_BITSIZE (mode
) - logd
);
3659 /* Use the rtx_cost of a LSHIFTRT instruction to determine
3660 which instruction sequence to use. If logical right shifts
3661 are expensive the use 2 XORs, 2 SUBs and an AND, otherwise
3662 use a LSHIFTRT, 1 ADD, 1 SUB and an AND. */
3664 temp
= gen_rtx_LSHIFTRT (mode
, result
, shift
);
3665 if (optab_handler (lshr_optab
, mode
) == CODE_FOR_nothing
3666 || (set_src_cost (temp
, optimize_insn_for_speed_p ())
3667 > COSTS_N_INSNS (2)))
3669 temp
= expand_binop (mode
, xor_optab
, op0
, signmask
,
3670 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3671 temp
= expand_binop (mode
, sub_optab
, temp
, signmask
,
3672 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3673 temp
= expand_binop (mode
, and_optab
, temp
,
3674 gen_int_mode (masklow
, mode
),
3675 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3676 temp
= expand_binop (mode
, xor_optab
, temp
, signmask
,
3677 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3678 temp
= expand_binop (mode
, sub_optab
, temp
, signmask
,
3679 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3683 signmask
= expand_binop (mode
, lshr_optab
, signmask
, shift
,
3684 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3685 signmask
= force_reg (mode
, signmask
);
3687 temp
= expand_binop (mode
, add_optab
, op0
, signmask
,
3688 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3689 temp
= expand_binop (mode
, and_optab
, temp
,
3690 gen_int_mode (masklow
, mode
),
3691 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3692 temp
= expand_binop (mode
, sub_optab
, temp
, signmask
,
3693 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
3699 /* Mask contains the mode's signbit and the significant bits of the
3700 modulus. By including the signbit in the operation, many targets
3701 can avoid an explicit compare operation in the following comparison
3703 wide_int mask
= wi::mask (logd
, false, prec
);
3704 mask
= wi::set_bit (mask
, prec
- 1);
3706 temp
= expand_binop (mode
, and_optab
, op0
,
3707 immed_wide_int_const (mask
, mode
),
3708 result
, 1, OPTAB_LIB_WIDEN
);
3710 emit_move_insn (result
, temp
);
3712 label
= gen_label_rtx ();
3713 do_cmp_and_jump (result
, const0_rtx
, GE
, mode
, label
);
3715 temp
= expand_binop (mode
, sub_optab
, result
, const1_rtx
, result
,
3716 0, OPTAB_LIB_WIDEN
);
3718 mask
= wi::mask (logd
, true, prec
);
3719 temp
= expand_binop (mode
, ior_optab
, temp
,
3720 immed_wide_int_const (mask
, mode
),
3721 result
, 1, OPTAB_LIB_WIDEN
);
3722 temp
= expand_binop (mode
, add_optab
, temp
, const1_rtx
, result
,
3723 0, OPTAB_LIB_WIDEN
);
3725 emit_move_insn (result
, temp
);
3730 /* Expand signed division of OP0 by a power of two D in mode MODE.
3731 This routine is only called for positive values of D. */
3734 expand_sdiv_pow2 (machine_mode mode
, rtx op0
, HOST_WIDE_INT d
)
3737 rtx_code_label
*label
;
3740 logd
= floor_log2 (d
);
3743 && BRANCH_COST (optimize_insn_for_speed_p (),
3746 temp
= gen_reg_rtx (mode
);
3747 temp
= emit_store_flag (temp
, LT
, op0
, const0_rtx
, mode
, 0, 1);
3748 temp
= expand_binop (mode
, add_optab
, temp
, op0
, NULL_RTX
,
3749 0, OPTAB_LIB_WIDEN
);
3750 return expand_shift (RSHIFT_EXPR
, mode
, temp
, logd
, NULL_RTX
, 0);
3753 #ifdef HAVE_conditional_move
3754 if (BRANCH_COST (optimize_insn_for_speed_p (), false)
3760 temp2
= copy_to_mode_reg (mode
, op0
);
3761 temp
= expand_binop (mode
, add_optab
, temp2
, gen_int_mode (d
- 1, mode
),
3762 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
3763 temp
= force_reg (mode
, temp
);
3765 /* Construct "temp2 = (temp2 < 0) ? temp : temp2". */
3766 temp2
= emit_conditional_move (temp2
, LT
, temp2
, const0_rtx
,
3767 mode
, temp
, temp2
, mode
, 0);
3770 rtx_insn
*seq
= get_insns ();
3773 return expand_shift (RSHIFT_EXPR
, mode
, temp2
, logd
, NULL_RTX
, 0);
3779 if (BRANCH_COST (optimize_insn_for_speed_p (),
3782 int ushift
= GET_MODE_BITSIZE (mode
) - logd
;
3784 temp
= gen_reg_rtx (mode
);
3785 temp
= emit_store_flag (temp
, LT
, op0
, const0_rtx
, mode
, 0, -1);
3786 if (GET_MODE_BITSIZE (mode
) >= BITS_PER_WORD
3787 || shift_cost (optimize_insn_for_speed_p (), mode
, ushift
)
3788 > COSTS_N_INSNS (1))
3789 temp
= expand_binop (mode
, and_optab
, temp
, gen_int_mode (d
- 1, mode
),
3790 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
3792 temp
= expand_shift (RSHIFT_EXPR
, mode
, temp
,
3793 ushift
, NULL_RTX
, 1);
3794 temp
= expand_binop (mode
, add_optab
, temp
, op0
, NULL_RTX
,
3795 0, OPTAB_LIB_WIDEN
);
3796 return expand_shift (RSHIFT_EXPR
, mode
, temp
, logd
, NULL_RTX
, 0);
3799 label
= gen_label_rtx ();
3800 temp
= copy_to_mode_reg (mode
, op0
);
3801 do_cmp_and_jump (temp
, const0_rtx
, GE
, mode
, label
);
3802 expand_inc (temp
, gen_int_mode (d
- 1, mode
));
3804 return expand_shift (RSHIFT_EXPR
, mode
, temp
, logd
, NULL_RTX
, 0);
3807 /* Emit the code to divide OP0 by OP1, putting the result in TARGET
3808 if that is convenient, and returning where the result is.
3809 You may request either the quotient or the remainder as the result;
3810 specify REM_FLAG nonzero to get the remainder.
3812 CODE is the expression code for which kind of division this is;
3813 it controls how rounding is done. MODE is the machine mode to use.
3814 UNSIGNEDP nonzero means do unsigned division. */
3816 /* ??? For CEIL_MOD_EXPR, can compute incorrect remainder with ANDI
3817 and then correct it by or'ing in missing high bits
3818 if result of ANDI is nonzero.
3819 For ROUND_MOD_EXPR, can use ANDI and then sign-extend the result.
3820 This could optimize to a bfexts instruction.
3821 But C doesn't use these operations, so their optimizations are
3823 /* ??? For modulo, we don't actually need the highpart of the first product,
3824 the low part will do nicely. And for small divisors, the second multiply
3825 can also be a low-part only multiply or even be completely left out.
3826 E.g. to calculate the remainder of a division by 3 with a 32 bit
3827 multiply, multiply with 0x55555556 and extract the upper two bits;
3828 the result is exact for inputs up to 0x1fffffff.
3829 The input range can be reduced by using cross-sum rules.
3830 For odd divisors >= 3, the following table gives right shift counts
3831 so that if a number is shifted by an integer multiple of the given
3832 amount, the remainder stays the same:
3833 2, 4, 3, 6, 10, 12, 4, 8, 18, 6, 11, 20, 18, 0, 5, 10, 12, 0, 12, 20,
3834 14, 12, 23, 21, 8, 0, 20, 18, 0, 0, 6, 12, 0, 22, 0, 18, 20, 30, 0, 0,
3835 0, 8, 0, 11, 12, 10, 36, 0, 30, 0, 0, 12, 0, 0, 0, 0, 44, 12, 24, 0,
3836 20, 0, 7, 14, 0, 18, 36, 0, 0, 46, 60, 0, 42, 0, 15, 24, 20, 0, 0, 33,
3837 0, 20, 0, 0, 18, 0, 60, 0, 0, 0, 0, 0, 40, 18, 0, 0, 12
3839 Cross-sum rules for even numbers can be derived by leaving as many bits
3840 to the right alone as the divisor has zeros to the right.
3841 E.g. if x is an unsigned 32 bit number:
3842 (x mod 12) == (((x & 1023) + ((x >> 8) & ~3)) * 0x15555558 >> 2 * 3) >> 28
3846 expand_divmod (int rem_flag
, enum tree_code code
, machine_mode mode
,
3847 rtx op0
, rtx op1
, rtx target
, int unsignedp
)
3849 machine_mode compute_mode
;
3851 rtx quotient
= 0, remainder
= 0;
3855 optab optab1
, optab2
;
3856 int op1_is_constant
, op1_is_pow2
= 0;
3857 int max_cost
, extra_cost
;
3858 static HOST_WIDE_INT last_div_const
= 0;
3859 bool speed
= optimize_insn_for_speed_p ();
3861 op1_is_constant
= CONST_INT_P (op1
);
3862 if (op1_is_constant
)
3864 unsigned HOST_WIDE_INT ext_op1
= UINTVAL (op1
);
3866 ext_op1
&= GET_MODE_MASK (mode
);
3867 op1_is_pow2
= ((EXACT_POWER_OF_2_OR_ZERO_P (ext_op1
)
3868 || (! unsignedp
&& EXACT_POWER_OF_2_OR_ZERO_P (-ext_op1
))));
3872 This is the structure of expand_divmod:
3874 First comes code to fix up the operands so we can perform the operations
3875 correctly and efficiently.
3877 Second comes a switch statement with code specific for each rounding mode.
3878 For some special operands this code emits all RTL for the desired
3879 operation, for other cases, it generates only a quotient and stores it in
3880 QUOTIENT. The case for trunc division/remainder might leave quotient = 0,
3881 to indicate that it has not done anything.
3883 Last comes code that finishes the operation. If QUOTIENT is set and
3884 REM_FLAG is set, the remainder is computed as OP0 - QUOTIENT * OP1. If
3885 QUOTIENT is not set, it is computed using trunc rounding.
3887 We try to generate special code for division and remainder when OP1 is a
3888 constant. If |OP1| = 2**n we can use shifts and some other fast
3889 operations. For other values of OP1, we compute a carefully selected
3890 fixed-point approximation m = 1/OP1, and generate code that multiplies OP0
3893 In all cases but EXACT_DIV_EXPR, this multiplication requires the upper
3894 half of the product. Different strategies for generating the product are
3895 implemented in expmed_mult_highpart.
3897 If what we actually want is the remainder, we generate that by another
3898 by-constant multiplication and a subtraction. */
3900 /* We shouldn't be called with OP1 == const1_rtx, but some of the
3901 code below will malfunction if we are, so check here and handle
3902 the special case if so. */
3903 if (op1
== const1_rtx
)
3904 return rem_flag
? const0_rtx
: op0
;
3906 /* When dividing by -1, we could get an overflow.
3907 negv_optab can handle overflows. */
3908 if (! unsignedp
&& op1
== constm1_rtx
)
3912 return expand_unop (mode
, flag_trapv
&& GET_MODE_CLASS (mode
) == MODE_INT
3913 ? negv_optab
: neg_optab
, op0
, target
, 0);
3917 /* Don't use the function value register as a target
3918 since we have to read it as well as write it,
3919 and function-inlining gets confused by this. */
3920 && ((REG_P (target
) && REG_FUNCTION_VALUE_P (target
))
3921 /* Don't clobber an operand while doing a multi-step calculation. */
3922 || ((rem_flag
|| op1_is_constant
)
3923 && (reg_mentioned_p (target
, op0
)
3924 || (MEM_P (op0
) && MEM_P (target
))))
3925 || reg_mentioned_p (target
, op1
)
3926 || (MEM_P (op1
) && MEM_P (target
))))
3929 /* Get the mode in which to perform this computation. Normally it will
3930 be MODE, but sometimes we can't do the desired operation in MODE.
3931 If so, pick a wider mode in which we can do the operation. Convert
3932 to that mode at the start to avoid repeated conversions.
3934 First see what operations we need. These depend on the expression
3935 we are evaluating. (We assume that divxx3 insns exist under the
3936 same conditions that modxx3 insns and that these insns don't normally
3937 fail. If these assumptions are not correct, we may generate less
3938 efficient code in some cases.)
3940 Then see if we find a mode in which we can open-code that operation
3941 (either a division, modulus, or shift). Finally, check for the smallest
3942 mode for which we can do the operation with a library call. */
3944 /* We might want to refine this now that we have division-by-constant
3945 optimization. Since expmed_mult_highpart tries so many variants, it is
3946 not straightforward to generalize this. Maybe we should make an array
3947 of possible modes in init_expmed? Save this for GCC 2.7. */
3949 optab1
= ((op1_is_pow2
&& op1
!= const0_rtx
)
3950 ? (unsignedp
? lshr_optab
: ashr_optab
)
3951 : (unsignedp
? udiv_optab
: sdiv_optab
));
3952 optab2
= ((op1_is_pow2
&& op1
!= const0_rtx
)
3954 : (unsignedp
? udivmod_optab
: sdivmod_optab
));
3956 for (compute_mode
= mode
; compute_mode
!= VOIDmode
;
3957 compute_mode
= GET_MODE_WIDER_MODE (compute_mode
))
3958 if (optab_handler (optab1
, compute_mode
) != CODE_FOR_nothing
3959 || optab_handler (optab2
, compute_mode
) != CODE_FOR_nothing
)
3962 if (compute_mode
== VOIDmode
)
3963 for (compute_mode
= mode
; compute_mode
!= VOIDmode
;
3964 compute_mode
= GET_MODE_WIDER_MODE (compute_mode
))
3965 if (optab_libfunc (optab1
, compute_mode
)
3966 || optab_libfunc (optab2
, compute_mode
))
3969 /* If we still couldn't find a mode, use MODE, but expand_binop will
3971 if (compute_mode
== VOIDmode
)
3972 compute_mode
= mode
;
3974 if (target
&& GET_MODE (target
) == compute_mode
)
3977 tquotient
= gen_reg_rtx (compute_mode
);
3979 size
= GET_MODE_BITSIZE (compute_mode
);
3981 /* It should be possible to restrict the precision to GET_MODE_BITSIZE
3982 (mode), and thereby get better code when OP1 is a constant. Do that
3983 later. It will require going over all usages of SIZE below. */
3984 size
= GET_MODE_BITSIZE (mode
);
3987 /* Only deduct something for a REM if the last divide done was
3988 for a different constant. Then set the constant of the last
3990 max_cost
= (unsignedp
3991 ? udiv_cost (speed
, compute_mode
)
3992 : sdiv_cost (speed
, compute_mode
));
3993 if (rem_flag
&& ! (last_div_const
!= 0 && op1_is_constant
3994 && INTVAL (op1
) == last_div_const
))
3995 max_cost
-= (mul_cost (speed
, compute_mode
)
3996 + add_cost (speed
, compute_mode
));
3998 last_div_const
= ! rem_flag
&& op1_is_constant
? INTVAL (op1
) : 0;
4000 /* Now convert to the best mode to use. */
4001 if (compute_mode
!= mode
)
4003 op0
= convert_modes (compute_mode
, mode
, op0
, unsignedp
);
4004 op1
= convert_modes (compute_mode
, mode
, op1
, unsignedp
);
4006 /* convert_modes may have placed op1 into a register, so we
4007 must recompute the following. */
4008 op1_is_constant
= CONST_INT_P (op1
);
4009 op1_is_pow2
= (op1_is_constant
4010 && ((EXACT_POWER_OF_2_OR_ZERO_P (INTVAL (op1
))
4012 && EXACT_POWER_OF_2_OR_ZERO_P (-UINTVAL (op1
))))));
4015 /* If one of the operands is a volatile MEM, copy it into a register. */
4017 if (MEM_P (op0
) && MEM_VOLATILE_P (op0
))
4018 op0
= force_reg (compute_mode
, op0
);
4019 if (MEM_P (op1
) && MEM_VOLATILE_P (op1
))
4020 op1
= force_reg (compute_mode
, op1
);
4022 /* If we need the remainder or if OP1 is constant, we need to
4023 put OP0 in a register in case it has any queued subexpressions. */
4024 if (rem_flag
|| op1_is_constant
)
4025 op0
= force_reg (compute_mode
, op0
);
4027 last
= get_last_insn ();
4029 /* Promote floor rounding to trunc rounding for unsigned operations. */
4032 if (code
== FLOOR_DIV_EXPR
)
4033 code
= TRUNC_DIV_EXPR
;
4034 if (code
== FLOOR_MOD_EXPR
)
4035 code
= TRUNC_MOD_EXPR
;
4036 if (code
== EXACT_DIV_EXPR
&& op1_is_pow2
)
4037 code
= TRUNC_DIV_EXPR
;
4040 if (op1
!= const0_rtx
)
4043 case TRUNC_MOD_EXPR
:
4044 case TRUNC_DIV_EXPR
:
4045 if (op1_is_constant
)
4049 unsigned HOST_WIDE_INT mh
, ml
;
4050 int pre_shift
, post_shift
;
4052 unsigned HOST_WIDE_INT d
= (INTVAL (op1
)
4053 & GET_MODE_MASK (compute_mode
));
4055 if (EXACT_POWER_OF_2_OR_ZERO_P (d
))
4057 pre_shift
= floor_log2 (d
);
4060 unsigned HOST_WIDE_INT mask
4061 = ((unsigned HOST_WIDE_INT
) 1 << pre_shift
) - 1;
4063 = expand_binop (compute_mode
, and_optab
, op0
,
4064 gen_int_mode (mask
, compute_mode
),
4068 return gen_lowpart (mode
, remainder
);
4070 quotient
= expand_shift (RSHIFT_EXPR
, compute_mode
, op0
,
4071 pre_shift
, tquotient
, 1);
4073 else if (size
<= HOST_BITS_PER_WIDE_INT
)
4075 if (d
>= ((unsigned HOST_WIDE_INT
) 1 << (size
- 1)))
4077 /* Most significant bit of divisor is set; emit an scc
4079 quotient
= emit_store_flag_force (tquotient
, GEU
, op0
, op1
,
4080 compute_mode
, 1, 1);
4084 /* Find a suitable multiplier and right shift count
4085 instead of multiplying with D. */
4087 mh
= choose_multiplier (d
, size
, size
,
4088 &ml
, &post_shift
, &dummy
);
4090 /* If the suggested multiplier is more than SIZE bits,
4091 we can do better for even divisors, using an
4092 initial right shift. */
4093 if (mh
!= 0 && (d
& 1) == 0)
4095 pre_shift
= floor_log2 (d
& -d
);
4096 mh
= choose_multiplier (d
>> pre_shift
, size
,
4098 &ml
, &post_shift
, &dummy
);
4108 if (post_shift
- 1 >= BITS_PER_WORD
)
4112 = (shift_cost (speed
, compute_mode
, post_shift
- 1)
4113 + shift_cost (speed
, compute_mode
, 1)
4114 + 2 * add_cost (speed
, compute_mode
));
4115 t1
= expmed_mult_highpart
4117 gen_int_mode (ml
, compute_mode
),
4118 NULL_RTX
, 1, max_cost
- extra_cost
);
4121 t2
= force_operand (gen_rtx_MINUS (compute_mode
,
4124 t3
= expand_shift (RSHIFT_EXPR
, compute_mode
,
4125 t2
, 1, NULL_RTX
, 1);
4126 t4
= force_operand (gen_rtx_PLUS (compute_mode
,
4129 quotient
= expand_shift
4130 (RSHIFT_EXPR
, compute_mode
, t4
,
4131 post_shift
- 1, tquotient
, 1);
4137 if (pre_shift
>= BITS_PER_WORD
4138 || post_shift
>= BITS_PER_WORD
)
4142 (RSHIFT_EXPR
, compute_mode
, op0
,
4143 pre_shift
, NULL_RTX
, 1);
4145 = (shift_cost (speed
, compute_mode
, pre_shift
)
4146 + shift_cost (speed
, compute_mode
, post_shift
));
4147 t2
= expmed_mult_highpart
4149 gen_int_mode (ml
, compute_mode
),
4150 NULL_RTX
, 1, max_cost
- extra_cost
);
4153 quotient
= expand_shift
4154 (RSHIFT_EXPR
, compute_mode
, t2
,
4155 post_shift
, tquotient
, 1);
4159 else /* Too wide mode to use tricky code */
4162 insn
= get_last_insn ();
4164 set_dst_reg_note (insn
, REG_EQUAL
,
4165 gen_rtx_UDIV (compute_mode
, op0
, op1
),
4168 else /* TRUNC_DIV, signed */
4170 unsigned HOST_WIDE_INT ml
;
4171 int lgup
, post_shift
;
4173 HOST_WIDE_INT d
= INTVAL (op1
);
4174 unsigned HOST_WIDE_INT abs_d
;
4176 /* Since d might be INT_MIN, we have to cast to
4177 unsigned HOST_WIDE_INT before negating to avoid
4178 undefined signed overflow. */
4180 ? (unsigned HOST_WIDE_INT
) d
4181 : - (unsigned HOST_WIDE_INT
) d
);
4183 /* n rem d = n rem -d */
4184 if (rem_flag
&& d
< 0)
4187 op1
= gen_int_mode (abs_d
, compute_mode
);
4193 quotient
= expand_unop (compute_mode
, neg_optab
, op0
,
4195 else if (HOST_BITS_PER_WIDE_INT
>= size
4196 && abs_d
== (unsigned HOST_WIDE_INT
) 1 << (size
- 1))
4198 /* This case is not handled correctly below. */
4199 quotient
= emit_store_flag (tquotient
, EQ
, op0
, op1
,
4200 compute_mode
, 1, 1);
4204 else if (EXACT_POWER_OF_2_OR_ZERO_P (d
)
4206 ? smod_pow2_cheap (speed
, compute_mode
)
4207 : sdiv_pow2_cheap (speed
, compute_mode
))
4208 /* We assume that cheap metric is true if the
4209 optab has an expander for this mode. */
4210 && ((optab_handler ((rem_flag
? smod_optab
4213 != CODE_FOR_nothing
)
4214 || (optab_handler (sdivmod_optab
,
4216 != CODE_FOR_nothing
)))
4218 else if (EXACT_POWER_OF_2_OR_ZERO_P (abs_d
))
4222 remainder
= expand_smod_pow2 (compute_mode
, op0
, d
);
4224 return gen_lowpart (mode
, remainder
);
4227 if (sdiv_pow2_cheap (speed
, compute_mode
)
4228 && ((optab_handler (sdiv_optab
, compute_mode
)
4229 != CODE_FOR_nothing
)
4230 || (optab_handler (sdivmod_optab
, compute_mode
)
4231 != CODE_FOR_nothing
)))
4232 quotient
= expand_divmod (0, TRUNC_DIV_EXPR
,
4234 gen_int_mode (abs_d
,
4238 quotient
= expand_sdiv_pow2 (compute_mode
, op0
, abs_d
);
4240 /* We have computed OP0 / abs(OP1). If OP1 is negative,
4241 negate the quotient. */
4244 insn
= get_last_insn ();
4246 && abs_d
< ((unsigned HOST_WIDE_INT
) 1
4247 << (HOST_BITS_PER_WIDE_INT
- 1)))
4248 set_dst_reg_note (insn
, REG_EQUAL
,
4249 gen_rtx_DIV (compute_mode
, op0
,
4255 quotient
= expand_unop (compute_mode
, neg_optab
,
4256 quotient
, quotient
, 0);
4259 else if (size
<= HOST_BITS_PER_WIDE_INT
)
4261 choose_multiplier (abs_d
, size
, size
- 1,
4262 &ml
, &post_shift
, &lgup
);
4263 if (ml
< (unsigned HOST_WIDE_INT
) 1 << (size
- 1))
4267 if (post_shift
>= BITS_PER_WORD
4268 || size
- 1 >= BITS_PER_WORD
)
4271 extra_cost
= (shift_cost (speed
, compute_mode
, post_shift
)
4272 + shift_cost (speed
, compute_mode
, size
- 1)
4273 + add_cost (speed
, compute_mode
));
4274 t1
= expmed_mult_highpart
4275 (compute_mode
, op0
, gen_int_mode (ml
, compute_mode
),
4276 NULL_RTX
, 0, max_cost
- extra_cost
);
4280 (RSHIFT_EXPR
, compute_mode
, t1
,
4281 post_shift
, NULL_RTX
, 0);
4283 (RSHIFT_EXPR
, compute_mode
, op0
,
4284 size
- 1, NULL_RTX
, 0);
4287 = force_operand (gen_rtx_MINUS (compute_mode
,
4292 = force_operand (gen_rtx_MINUS (compute_mode
,
4300 if (post_shift
>= BITS_PER_WORD
4301 || size
- 1 >= BITS_PER_WORD
)
4304 ml
|= (~(unsigned HOST_WIDE_INT
) 0) << (size
- 1);
4305 mlr
= gen_int_mode (ml
, compute_mode
);
4306 extra_cost
= (shift_cost (speed
, compute_mode
, post_shift
)
4307 + shift_cost (speed
, compute_mode
, size
- 1)
4308 + 2 * add_cost (speed
, compute_mode
));
4309 t1
= expmed_mult_highpart (compute_mode
, op0
, mlr
,
4311 max_cost
- extra_cost
);
4314 t2
= force_operand (gen_rtx_PLUS (compute_mode
,
4318 (RSHIFT_EXPR
, compute_mode
, t2
,
4319 post_shift
, NULL_RTX
, 0);
4321 (RSHIFT_EXPR
, compute_mode
, op0
,
4322 size
- 1, NULL_RTX
, 0);
4325 = force_operand (gen_rtx_MINUS (compute_mode
,
4330 = force_operand (gen_rtx_MINUS (compute_mode
,
4335 else /* Too wide mode to use tricky code */
4338 insn
= get_last_insn ();
4340 set_dst_reg_note (insn
, REG_EQUAL
,
4341 gen_rtx_DIV (compute_mode
, op0
, op1
),
4347 delete_insns_since (last
);
4350 case FLOOR_DIV_EXPR
:
4351 case FLOOR_MOD_EXPR
:
4352 /* We will come here only for signed operations. */
4353 if (op1_is_constant
&& HOST_BITS_PER_WIDE_INT
>= size
)
4355 unsigned HOST_WIDE_INT mh
, ml
;
4356 int pre_shift
, lgup
, post_shift
;
4357 HOST_WIDE_INT d
= INTVAL (op1
);
4361 /* We could just as easily deal with negative constants here,
4362 but it does not seem worth the trouble for GCC 2.6. */
4363 if (EXACT_POWER_OF_2_OR_ZERO_P (d
))
4365 pre_shift
= floor_log2 (d
);
4368 unsigned HOST_WIDE_INT mask
4369 = ((unsigned HOST_WIDE_INT
) 1 << pre_shift
) - 1;
4370 remainder
= expand_binop
4371 (compute_mode
, and_optab
, op0
,
4372 gen_int_mode (mask
, compute_mode
),
4373 remainder
, 0, OPTAB_LIB_WIDEN
);
4375 return gen_lowpart (mode
, remainder
);
4377 quotient
= expand_shift
4378 (RSHIFT_EXPR
, compute_mode
, op0
,
4379 pre_shift
, tquotient
, 0);
4385 mh
= choose_multiplier (d
, size
, size
- 1,
4386 &ml
, &post_shift
, &lgup
);
4389 if (post_shift
< BITS_PER_WORD
4390 && size
- 1 < BITS_PER_WORD
)
4393 (RSHIFT_EXPR
, compute_mode
, op0
,
4394 size
- 1, NULL_RTX
, 0);
4395 t2
= expand_binop (compute_mode
, xor_optab
, op0
, t1
,
4396 NULL_RTX
, 0, OPTAB_WIDEN
);
4397 extra_cost
= (shift_cost (speed
, compute_mode
, post_shift
)
4398 + shift_cost (speed
, compute_mode
, size
- 1)
4399 + 2 * add_cost (speed
, compute_mode
));
4400 t3
= expmed_mult_highpart
4401 (compute_mode
, t2
, gen_int_mode (ml
, compute_mode
),
4402 NULL_RTX
, 1, max_cost
- extra_cost
);
4406 (RSHIFT_EXPR
, compute_mode
, t3
,
4407 post_shift
, NULL_RTX
, 1);
4408 quotient
= expand_binop (compute_mode
, xor_optab
,
4409 t4
, t1
, tquotient
, 0,
4417 rtx nsign
, t1
, t2
, t3
, t4
;
4418 t1
= force_operand (gen_rtx_PLUS (compute_mode
,
4419 op0
, constm1_rtx
), NULL_RTX
);
4420 t2
= expand_binop (compute_mode
, ior_optab
, op0
, t1
, NULL_RTX
,
4422 nsign
= expand_shift
4423 (RSHIFT_EXPR
, compute_mode
, t2
,
4424 size
- 1, NULL_RTX
, 0);
4425 t3
= force_operand (gen_rtx_MINUS (compute_mode
, t1
, nsign
),
4427 t4
= expand_divmod (0, TRUNC_DIV_EXPR
, compute_mode
, t3
, op1
,
4432 t5
= expand_unop (compute_mode
, one_cmpl_optab
, nsign
,
4434 quotient
= force_operand (gen_rtx_PLUS (compute_mode
,
4443 delete_insns_since (last
);
4445 /* Try using an instruction that produces both the quotient and
4446 remainder, using truncation. We can easily compensate the quotient
4447 or remainder to get floor rounding, once we have the remainder.
4448 Notice that we compute also the final remainder value here,
4449 and return the result right away. */
4450 if (target
== 0 || GET_MODE (target
) != compute_mode
)
4451 target
= gen_reg_rtx (compute_mode
);
4456 = REG_P (target
) ? target
: gen_reg_rtx (compute_mode
);
4457 quotient
= gen_reg_rtx (compute_mode
);
4462 = REG_P (target
) ? target
: gen_reg_rtx (compute_mode
);
4463 remainder
= gen_reg_rtx (compute_mode
);
4466 if (expand_twoval_binop (sdivmod_optab
, op0
, op1
,
4467 quotient
, remainder
, 0))
4469 /* This could be computed with a branch-less sequence.
4470 Save that for later. */
4472 rtx_code_label
*label
= gen_label_rtx ();
4473 do_cmp_and_jump (remainder
, const0_rtx
, EQ
, compute_mode
, label
);
4474 tem
= expand_binop (compute_mode
, xor_optab
, op0
, op1
,
4475 NULL_RTX
, 0, OPTAB_WIDEN
);
4476 do_cmp_and_jump (tem
, const0_rtx
, GE
, compute_mode
, label
);
4477 expand_dec (quotient
, const1_rtx
);
4478 expand_inc (remainder
, op1
);
4480 return gen_lowpart (mode
, rem_flag
? remainder
: quotient
);
4483 /* No luck with division elimination or divmod. Have to do it
4484 by conditionally adjusting op0 *and* the result. */
4486 rtx_code_label
*label1
, *label2
, *label3
, *label4
, *label5
;
4490 quotient
= gen_reg_rtx (compute_mode
);
4491 adjusted_op0
= copy_to_mode_reg (compute_mode
, op0
);
4492 label1
= gen_label_rtx ();
4493 label2
= gen_label_rtx ();
4494 label3
= gen_label_rtx ();
4495 label4
= gen_label_rtx ();
4496 label5
= gen_label_rtx ();
4497 do_cmp_and_jump (op1
, const0_rtx
, LT
, compute_mode
, label2
);
4498 do_cmp_and_jump (adjusted_op0
, const0_rtx
, LT
, compute_mode
, label1
);
4499 tem
= expand_binop (compute_mode
, sdiv_optab
, adjusted_op0
, op1
,
4500 quotient
, 0, OPTAB_LIB_WIDEN
);
4501 if (tem
!= quotient
)
4502 emit_move_insn (quotient
, tem
);
4503 emit_jump_insn (gen_jump (label5
));
4505 emit_label (label1
);
4506 expand_inc (adjusted_op0
, const1_rtx
);
4507 emit_jump_insn (gen_jump (label4
));
4509 emit_label (label2
);
4510 do_cmp_and_jump (adjusted_op0
, const0_rtx
, GT
, compute_mode
, label3
);
4511 tem
= expand_binop (compute_mode
, sdiv_optab
, adjusted_op0
, op1
,
4512 quotient
, 0, OPTAB_LIB_WIDEN
);
4513 if (tem
!= quotient
)
4514 emit_move_insn (quotient
, tem
);
4515 emit_jump_insn (gen_jump (label5
));
4517 emit_label (label3
);
4518 expand_dec (adjusted_op0
, const1_rtx
);
4519 emit_label (label4
);
4520 tem
= expand_binop (compute_mode
, sdiv_optab
, adjusted_op0
, op1
,
4521 quotient
, 0, OPTAB_LIB_WIDEN
);
4522 if (tem
!= quotient
)
4523 emit_move_insn (quotient
, tem
);
4524 expand_dec (quotient
, const1_rtx
);
4525 emit_label (label5
);
4533 if (op1_is_constant
&& EXACT_POWER_OF_2_OR_ZERO_P (INTVAL (op1
)))
4536 unsigned HOST_WIDE_INT d
= INTVAL (op1
);
4537 t1
= expand_shift (RSHIFT_EXPR
, compute_mode
, op0
,
4538 floor_log2 (d
), tquotient
, 1);
4539 t2
= expand_binop (compute_mode
, and_optab
, op0
,
4540 gen_int_mode (d
- 1, compute_mode
),
4541 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
4542 t3
= gen_reg_rtx (compute_mode
);
4543 t3
= emit_store_flag (t3
, NE
, t2
, const0_rtx
,
4544 compute_mode
, 1, 1);
4547 rtx_code_label
*lab
;
4548 lab
= gen_label_rtx ();
4549 do_cmp_and_jump (t2
, const0_rtx
, EQ
, compute_mode
, lab
);
4550 expand_inc (t1
, const1_rtx
);
4555 quotient
= force_operand (gen_rtx_PLUS (compute_mode
,
4561 /* Try using an instruction that produces both the quotient and
4562 remainder, using truncation. We can easily compensate the
4563 quotient or remainder to get ceiling rounding, once we have the
4564 remainder. Notice that we compute also the final remainder
4565 value here, and return the result right away. */
4566 if (target
== 0 || GET_MODE (target
) != compute_mode
)
4567 target
= gen_reg_rtx (compute_mode
);
4571 remainder
= (REG_P (target
)
4572 ? target
: gen_reg_rtx (compute_mode
));
4573 quotient
= gen_reg_rtx (compute_mode
);
4577 quotient
= (REG_P (target
)
4578 ? target
: gen_reg_rtx (compute_mode
));
4579 remainder
= gen_reg_rtx (compute_mode
);
4582 if (expand_twoval_binop (udivmod_optab
, op0
, op1
, quotient
,
4585 /* This could be computed with a branch-less sequence.
4586 Save that for later. */
4587 rtx_code_label
*label
= gen_label_rtx ();
4588 do_cmp_and_jump (remainder
, const0_rtx
, EQ
,
4589 compute_mode
, label
);
4590 expand_inc (quotient
, const1_rtx
);
4591 expand_dec (remainder
, op1
);
4593 return gen_lowpart (mode
, rem_flag
? remainder
: quotient
);
4596 /* No luck with division elimination or divmod. Have to do it
4597 by conditionally adjusting op0 *and* the result. */
4599 rtx_code_label
*label1
, *label2
;
4600 rtx adjusted_op0
, tem
;
4602 quotient
= gen_reg_rtx (compute_mode
);
4603 adjusted_op0
= copy_to_mode_reg (compute_mode
, op0
);
4604 label1
= gen_label_rtx ();
4605 label2
= gen_label_rtx ();
4606 do_cmp_and_jump (adjusted_op0
, const0_rtx
, NE
,
4607 compute_mode
, label1
);
4608 emit_move_insn (quotient
, const0_rtx
);
4609 emit_jump_insn (gen_jump (label2
));
4611 emit_label (label1
);
4612 expand_dec (adjusted_op0
, const1_rtx
);
4613 tem
= expand_binop (compute_mode
, udiv_optab
, adjusted_op0
, op1
,
4614 quotient
, 1, OPTAB_LIB_WIDEN
);
4615 if (tem
!= quotient
)
4616 emit_move_insn (quotient
, tem
);
4617 expand_inc (quotient
, const1_rtx
);
4618 emit_label (label2
);
4623 if (op1_is_constant
&& EXACT_POWER_OF_2_OR_ZERO_P (INTVAL (op1
))
4624 && INTVAL (op1
) >= 0)
4626 /* This is extremely similar to the code for the unsigned case
4627 above. For 2.7 we should merge these variants, but for
4628 2.6.1 I don't want to touch the code for unsigned since that
4629 get used in C. The signed case will only be used by other
4633 unsigned HOST_WIDE_INT d
= INTVAL (op1
);
4634 t1
= expand_shift (RSHIFT_EXPR
, compute_mode
, op0
,
4635 floor_log2 (d
), tquotient
, 0);
4636 t2
= expand_binop (compute_mode
, and_optab
, op0
,
4637 gen_int_mode (d
- 1, compute_mode
),
4638 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
4639 t3
= gen_reg_rtx (compute_mode
);
4640 t3
= emit_store_flag (t3
, NE
, t2
, const0_rtx
,
4641 compute_mode
, 1, 1);
4644 rtx_code_label
*lab
;
4645 lab
= gen_label_rtx ();
4646 do_cmp_and_jump (t2
, const0_rtx
, EQ
, compute_mode
, lab
);
4647 expand_inc (t1
, const1_rtx
);
4652 quotient
= force_operand (gen_rtx_PLUS (compute_mode
,
4658 /* Try using an instruction that produces both the quotient and
4659 remainder, using truncation. We can easily compensate the
4660 quotient or remainder to get ceiling rounding, once we have the
4661 remainder. Notice that we compute also the final remainder
4662 value here, and return the result right away. */
4663 if (target
== 0 || GET_MODE (target
) != compute_mode
)
4664 target
= gen_reg_rtx (compute_mode
);
4667 remainder
= (REG_P (target
)
4668 ? target
: gen_reg_rtx (compute_mode
));
4669 quotient
= gen_reg_rtx (compute_mode
);
4673 quotient
= (REG_P (target
)
4674 ? target
: gen_reg_rtx (compute_mode
));
4675 remainder
= gen_reg_rtx (compute_mode
);
4678 if (expand_twoval_binop (sdivmod_optab
, op0
, op1
, quotient
,
4681 /* This could be computed with a branch-less sequence.
4682 Save that for later. */
4684 rtx_code_label
*label
= gen_label_rtx ();
4685 do_cmp_and_jump (remainder
, const0_rtx
, EQ
,
4686 compute_mode
, label
);
4687 tem
= expand_binop (compute_mode
, xor_optab
, op0
, op1
,
4688 NULL_RTX
, 0, OPTAB_WIDEN
);
4689 do_cmp_and_jump (tem
, const0_rtx
, LT
, compute_mode
, label
);
4690 expand_inc (quotient
, const1_rtx
);
4691 expand_dec (remainder
, op1
);
4693 return gen_lowpart (mode
, rem_flag
? remainder
: quotient
);
4696 /* No luck with division elimination or divmod. Have to do it
4697 by conditionally adjusting op0 *and* the result. */
4699 rtx_code_label
*label1
, *label2
, *label3
, *label4
, *label5
;
4703 quotient
= gen_reg_rtx (compute_mode
);
4704 adjusted_op0
= copy_to_mode_reg (compute_mode
, op0
);
4705 label1
= gen_label_rtx ();
4706 label2
= gen_label_rtx ();
4707 label3
= gen_label_rtx ();
4708 label4
= gen_label_rtx ();
4709 label5
= gen_label_rtx ();
4710 do_cmp_and_jump (op1
, const0_rtx
, LT
, compute_mode
, label2
);
4711 do_cmp_and_jump (adjusted_op0
, const0_rtx
, GT
,
4712 compute_mode
, label1
);
4713 tem
= expand_binop (compute_mode
, sdiv_optab
, adjusted_op0
, op1
,
4714 quotient
, 0, OPTAB_LIB_WIDEN
);
4715 if (tem
!= quotient
)
4716 emit_move_insn (quotient
, tem
);
4717 emit_jump_insn (gen_jump (label5
));
4719 emit_label (label1
);
4720 expand_dec (adjusted_op0
, const1_rtx
);
4721 emit_jump_insn (gen_jump (label4
));
4723 emit_label (label2
);
4724 do_cmp_and_jump (adjusted_op0
, const0_rtx
, LT
,
4725 compute_mode
, label3
);
4726 tem
= expand_binop (compute_mode
, sdiv_optab
, adjusted_op0
, op1
,
4727 quotient
, 0, OPTAB_LIB_WIDEN
);
4728 if (tem
!= quotient
)
4729 emit_move_insn (quotient
, tem
);
4730 emit_jump_insn (gen_jump (label5
));
4732 emit_label (label3
);
4733 expand_inc (adjusted_op0
, const1_rtx
);
4734 emit_label (label4
);
4735 tem
= expand_binop (compute_mode
, sdiv_optab
, adjusted_op0
, op1
,
4736 quotient
, 0, OPTAB_LIB_WIDEN
);
4737 if (tem
!= quotient
)
4738 emit_move_insn (quotient
, tem
);
4739 expand_inc (quotient
, const1_rtx
);
4740 emit_label (label5
);
4745 case EXACT_DIV_EXPR
:
4746 if (op1_is_constant
&& HOST_BITS_PER_WIDE_INT
>= size
)
4748 HOST_WIDE_INT d
= INTVAL (op1
);
4749 unsigned HOST_WIDE_INT ml
;
4753 pre_shift
= floor_log2 (d
& -d
);
4754 ml
= invert_mod2n (d
>> pre_shift
, size
);
4755 t1
= expand_shift (RSHIFT_EXPR
, compute_mode
, op0
,
4756 pre_shift
, NULL_RTX
, unsignedp
);
4757 quotient
= expand_mult (compute_mode
, t1
,
4758 gen_int_mode (ml
, compute_mode
),
4761 insn
= get_last_insn ();
4762 set_dst_reg_note (insn
, REG_EQUAL
,
4763 gen_rtx_fmt_ee (unsignedp
? UDIV
: DIV
,
4764 compute_mode
, op0
, op1
),
4769 case ROUND_DIV_EXPR
:
4770 case ROUND_MOD_EXPR
:
4774 rtx_code_label
*label
;
4775 label
= gen_label_rtx ();
4776 quotient
= gen_reg_rtx (compute_mode
);
4777 remainder
= gen_reg_rtx (compute_mode
);
4778 if (expand_twoval_binop (udivmod_optab
, op0
, op1
, quotient
, remainder
, 1) == 0)
4781 quotient
= expand_binop (compute_mode
, udiv_optab
, op0
, op1
,
4782 quotient
, 1, OPTAB_LIB_WIDEN
);
4783 tem
= expand_mult (compute_mode
, quotient
, op1
, NULL_RTX
, 1);
4784 remainder
= expand_binop (compute_mode
, sub_optab
, op0
, tem
,
4785 remainder
, 1, OPTAB_LIB_WIDEN
);
4787 tem
= plus_constant (compute_mode
, op1
, -1);
4788 tem
= expand_shift (RSHIFT_EXPR
, compute_mode
, tem
, 1, NULL_RTX
, 1);
4789 do_cmp_and_jump (remainder
, tem
, LEU
, compute_mode
, label
);
4790 expand_inc (quotient
, const1_rtx
);
4791 expand_dec (remainder
, op1
);
4796 rtx abs_rem
, abs_op1
, tem
, mask
;
4797 rtx_code_label
*label
;
4798 label
= gen_label_rtx ();
4799 quotient
= gen_reg_rtx (compute_mode
);
4800 remainder
= gen_reg_rtx (compute_mode
);
4801 if (expand_twoval_binop (sdivmod_optab
, op0
, op1
, quotient
, remainder
, 0) == 0)
4804 quotient
= expand_binop (compute_mode
, sdiv_optab
, op0
, op1
,
4805 quotient
, 0, OPTAB_LIB_WIDEN
);
4806 tem
= expand_mult (compute_mode
, quotient
, op1
, NULL_RTX
, 0);
4807 remainder
= expand_binop (compute_mode
, sub_optab
, op0
, tem
,
4808 remainder
, 0, OPTAB_LIB_WIDEN
);
4810 abs_rem
= expand_abs (compute_mode
, remainder
, NULL_RTX
, 1, 0);
4811 abs_op1
= expand_abs (compute_mode
, op1
, NULL_RTX
, 1, 0);
4812 tem
= expand_shift (LSHIFT_EXPR
, compute_mode
, abs_rem
,
4814 do_cmp_and_jump (tem
, abs_op1
, LTU
, compute_mode
, label
);
4815 tem
= expand_binop (compute_mode
, xor_optab
, op0
, op1
,
4816 NULL_RTX
, 0, OPTAB_WIDEN
);
4817 mask
= expand_shift (RSHIFT_EXPR
, compute_mode
, tem
,
4818 size
- 1, NULL_RTX
, 0);
4819 tem
= expand_binop (compute_mode
, xor_optab
, mask
, const1_rtx
,
4820 NULL_RTX
, 0, OPTAB_WIDEN
);
4821 tem
= expand_binop (compute_mode
, sub_optab
, tem
, mask
,
4822 NULL_RTX
, 0, OPTAB_WIDEN
);
4823 expand_inc (quotient
, tem
);
4824 tem
= expand_binop (compute_mode
, xor_optab
, mask
, op1
,
4825 NULL_RTX
, 0, OPTAB_WIDEN
);
4826 tem
= expand_binop (compute_mode
, sub_optab
, tem
, mask
,
4827 NULL_RTX
, 0, OPTAB_WIDEN
);
4828 expand_dec (remainder
, tem
);
4831 return gen_lowpart (mode
, rem_flag
? remainder
: quotient
);
4839 if (target
&& GET_MODE (target
) != compute_mode
)
4844 /* Try to produce the remainder without producing the quotient.
4845 If we seem to have a divmod pattern that does not require widening,
4846 don't try widening here. We should really have a WIDEN argument
4847 to expand_twoval_binop, since what we'd really like to do here is
4848 1) try a mod insn in compute_mode
4849 2) try a divmod insn in compute_mode
4850 3) try a div insn in compute_mode and multiply-subtract to get
4852 4) try the same things with widening allowed. */
4854 = sign_expand_binop (compute_mode
, umod_optab
, smod_optab
,
4857 ((optab_handler (optab2
, compute_mode
)
4858 != CODE_FOR_nothing
)
4859 ? OPTAB_DIRECT
: OPTAB_WIDEN
));
4862 /* No luck there. Can we do remainder and divide at once
4863 without a library call? */
4864 remainder
= gen_reg_rtx (compute_mode
);
4865 if (! expand_twoval_binop ((unsignedp
4869 NULL_RTX
, remainder
, unsignedp
))
4874 return gen_lowpart (mode
, remainder
);
4877 /* Produce the quotient. Try a quotient insn, but not a library call.
4878 If we have a divmod in this mode, use it in preference to widening
4879 the div (for this test we assume it will not fail). Note that optab2
4880 is set to the one of the two optabs that the call below will use. */
4882 = sign_expand_binop (compute_mode
, udiv_optab
, sdiv_optab
,
4883 op0
, op1
, rem_flag
? NULL_RTX
: target
,
4885 ((optab_handler (optab2
, compute_mode
)
4886 != CODE_FOR_nothing
)
4887 ? OPTAB_DIRECT
: OPTAB_WIDEN
));
4891 /* No luck there. Try a quotient-and-remainder insn,
4892 keeping the quotient alone. */
4893 quotient
= gen_reg_rtx (compute_mode
);
4894 if (! expand_twoval_binop (unsignedp
? udivmod_optab
: sdivmod_optab
,
4896 quotient
, NULL_RTX
, unsignedp
))
4900 /* Still no luck. If we are not computing the remainder,
4901 use a library call for the quotient. */
4902 quotient
= sign_expand_binop (compute_mode
,
4903 udiv_optab
, sdiv_optab
,
4905 unsignedp
, OPTAB_LIB_WIDEN
);
4912 if (target
&& GET_MODE (target
) != compute_mode
)
4917 /* No divide instruction either. Use library for remainder. */
4918 remainder
= sign_expand_binop (compute_mode
, umod_optab
, smod_optab
,
4920 unsignedp
, OPTAB_LIB_WIDEN
);
4921 /* No remainder function. Try a quotient-and-remainder
4922 function, keeping the remainder. */
4925 remainder
= gen_reg_rtx (compute_mode
);
4926 if (!expand_twoval_binop_libfunc
4927 (unsignedp
? udivmod_optab
: sdivmod_optab
,
4929 NULL_RTX
, remainder
,
4930 unsignedp
? UMOD
: MOD
))
4931 remainder
= NULL_RTX
;
4936 /* We divided. Now finish doing X - Y * (X / Y). */
4937 remainder
= expand_mult (compute_mode
, quotient
, op1
,
4938 NULL_RTX
, unsignedp
);
4939 remainder
= expand_binop (compute_mode
, sub_optab
, op0
,
4940 remainder
, target
, unsignedp
,
4945 return gen_lowpart (mode
, rem_flag
? remainder
: quotient
);
4948 /* Return a tree node with data type TYPE, describing the value of X.
4949 Usually this is an VAR_DECL, if there is no obvious better choice.
4950 X may be an expression, however we only support those expressions
4951 generated by loop.c. */
4954 make_tree (tree type
, rtx x
)
4958 switch (GET_CODE (x
))
4961 case CONST_WIDE_INT
:
4962 t
= wide_int_to_tree (type
, std::make_pair (x
, TYPE_MODE (type
)));
4966 STATIC_ASSERT (HOST_BITS_PER_WIDE_INT
* 2 <= MAX_BITSIZE_MODE_ANY_INT
);
4967 if (TARGET_SUPPORTS_WIDE_INT
== 0 && GET_MODE (x
) == VOIDmode
)
4968 t
= wide_int_to_tree (type
,
4969 wide_int::from_array (&CONST_DOUBLE_LOW (x
), 2,
4970 HOST_BITS_PER_WIDE_INT
* 2));
4975 REAL_VALUE_FROM_CONST_DOUBLE (d
, x
);
4976 t
= build_real (type
, d
);
4983 int units
= CONST_VECTOR_NUNITS (x
);
4984 tree itype
= TREE_TYPE (type
);
4988 /* Build a tree with vector elements. */
4989 elts
= XALLOCAVEC (tree
, units
);
4990 for (i
= units
- 1; i
>= 0; --i
)
4992 rtx elt
= CONST_VECTOR_ELT (x
, i
);
4993 elts
[i
] = make_tree (itype
, elt
);
4996 return build_vector (type
, elts
);
5000 return fold_build2 (PLUS_EXPR
, type
, make_tree (type
, XEXP (x
, 0)),
5001 make_tree (type
, XEXP (x
, 1)));
5004 return fold_build2 (MINUS_EXPR
, type
, make_tree (type
, XEXP (x
, 0)),
5005 make_tree (type
, XEXP (x
, 1)));
5008 return fold_build1 (NEGATE_EXPR
, type
, make_tree (type
, XEXP (x
, 0)));
5011 return fold_build2 (MULT_EXPR
, type
, make_tree (type
, XEXP (x
, 0)),
5012 make_tree (type
, XEXP (x
, 1)));
5015 return fold_build2 (LSHIFT_EXPR
, type
, make_tree (type
, XEXP (x
, 0)),
5016 make_tree (type
, XEXP (x
, 1)));
5019 t
= unsigned_type_for (type
);
5020 return fold_convert (type
, build2 (RSHIFT_EXPR
, t
,
5021 make_tree (t
, XEXP (x
, 0)),
5022 make_tree (type
, XEXP (x
, 1))));
5025 t
= signed_type_for (type
);
5026 return fold_convert (type
, build2 (RSHIFT_EXPR
, t
,
5027 make_tree (t
, XEXP (x
, 0)),
5028 make_tree (type
, XEXP (x
, 1))));
5031 if (TREE_CODE (type
) != REAL_TYPE
)
5032 t
= signed_type_for (type
);
5036 return fold_convert (type
, build2 (TRUNC_DIV_EXPR
, t
,
5037 make_tree (t
, XEXP (x
, 0)),
5038 make_tree (t
, XEXP (x
, 1))));
5040 t
= unsigned_type_for (type
);
5041 return fold_convert (type
, build2 (TRUNC_DIV_EXPR
, t
,
5042 make_tree (t
, XEXP (x
, 0)),
5043 make_tree (t
, XEXP (x
, 1))));
5047 t
= lang_hooks
.types
.type_for_mode (GET_MODE (XEXP (x
, 0)),
5048 GET_CODE (x
) == ZERO_EXTEND
);
5049 return fold_convert (type
, make_tree (t
, XEXP (x
, 0)));
5052 return make_tree (type
, XEXP (x
, 0));
5055 t
= SYMBOL_REF_DECL (x
);
5057 return fold_convert (type
, build_fold_addr_expr (t
));
5058 /* else fall through. */
5061 t
= build_decl (RTL_LOCATION (x
), VAR_DECL
, NULL_TREE
, type
);
5063 /* If TYPE is a POINTER_TYPE, we might need to convert X from
5064 address mode to pointer mode. */
5065 if (POINTER_TYPE_P (type
))
5066 x
= convert_memory_address_addr_space
5067 (TYPE_MODE (type
), x
, TYPE_ADDR_SPACE (TREE_TYPE (type
)));
5069 /* Note that we do *not* use SET_DECL_RTL here, because we do not
5070 want set_decl_rtl to go adjusting REG_ATTRS for this temporary. */
5071 t
->decl_with_rtl
.rtl
= x
;
5077 /* Compute the logical-and of OP0 and OP1, storing it in TARGET
5078 and returning TARGET.
5080 If TARGET is 0, a pseudo-register or constant is returned. */
5083 expand_and (machine_mode mode
, rtx op0
, rtx op1
, rtx target
)
5087 if (GET_MODE (op0
) == VOIDmode
&& GET_MODE (op1
) == VOIDmode
)
5088 tem
= simplify_binary_operation (AND
, mode
, op0
, op1
);
5090 tem
= expand_binop (mode
, and_optab
, op0
, op1
, target
, 0, OPTAB_LIB_WIDEN
);
5094 else if (tem
!= target
)
5095 emit_move_insn (target
, tem
);
5099 /* Helper function for emit_store_flag. */
5101 emit_cstore (rtx target
, enum insn_code icode
, enum rtx_code code
,
5102 machine_mode mode
, machine_mode compare_mode
,
5103 int unsignedp
, rtx x
, rtx y
, int normalizep
,
5104 machine_mode target_mode
)
5106 struct expand_operand ops
[4];
5107 rtx op0
, comparison
, subtarget
;
5109 machine_mode result_mode
= targetm
.cstore_mode (icode
);
5111 last
= get_last_insn ();
5112 x
= prepare_operand (icode
, x
, 2, mode
, compare_mode
, unsignedp
);
5113 y
= prepare_operand (icode
, y
, 3, mode
, compare_mode
, unsignedp
);
5116 delete_insns_since (last
);
5120 if (target_mode
== VOIDmode
)
5121 target_mode
= result_mode
;
5123 target
= gen_reg_rtx (target_mode
);
5125 comparison
= gen_rtx_fmt_ee (code
, result_mode
, x
, y
);
5127 create_output_operand (&ops
[0], optimize
? NULL_RTX
: target
, result_mode
);
5128 create_fixed_operand (&ops
[1], comparison
);
5129 create_fixed_operand (&ops
[2], x
);
5130 create_fixed_operand (&ops
[3], y
);
5131 if (!maybe_expand_insn (icode
, 4, ops
))
5133 delete_insns_since (last
);
5136 subtarget
= ops
[0].value
;
5138 /* If we are converting to a wider mode, first convert to
5139 TARGET_MODE, then normalize. This produces better combining
5140 opportunities on machines that have a SIGN_EXTRACT when we are
5141 testing a single bit. This mostly benefits the 68k.
5143 If STORE_FLAG_VALUE does not have the sign bit set when
5144 interpreted in MODE, we can do this conversion as unsigned, which
5145 is usually more efficient. */
5146 if (GET_MODE_SIZE (target_mode
) > GET_MODE_SIZE (result_mode
))
5148 convert_move (target
, subtarget
,
5149 val_signbit_known_clear_p (result_mode
,
5152 result_mode
= target_mode
;
5157 /* If we want to keep subexpressions around, don't reuse our last
5162 /* Now normalize to the proper value in MODE. Sometimes we don't
5163 have to do anything. */
5164 if (normalizep
== 0 || normalizep
== STORE_FLAG_VALUE
)
5166 /* STORE_FLAG_VALUE might be the most negative number, so write
5167 the comparison this way to avoid a compiler-time warning. */
5168 else if (- normalizep
== STORE_FLAG_VALUE
)
5169 op0
= expand_unop (result_mode
, neg_optab
, op0
, subtarget
, 0);
5171 /* We don't want to use STORE_FLAG_VALUE < 0 below since this makes
5172 it hard to use a value of just the sign bit due to ANSI integer
5173 constant typing rules. */
5174 else if (val_signbit_known_set_p (result_mode
, STORE_FLAG_VALUE
))
5175 op0
= expand_shift (RSHIFT_EXPR
, result_mode
, op0
,
5176 GET_MODE_BITSIZE (result_mode
) - 1, subtarget
,
5180 gcc_assert (STORE_FLAG_VALUE
& 1);
5182 op0
= expand_and (result_mode
, op0
, const1_rtx
, subtarget
);
5183 if (normalizep
== -1)
5184 op0
= expand_unop (result_mode
, neg_optab
, op0
, op0
, 0);
5187 /* If we were converting to a smaller mode, do the conversion now. */
5188 if (target_mode
!= result_mode
)
5190 convert_move (target
, op0
, 0);
5198 /* A subroutine of emit_store_flag only including "tricks" that do not
5199 need a recursive call. These are kept separate to avoid infinite
5203 emit_store_flag_1 (rtx target
, enum rtx_code code
, rtx op0
, rtx op1
,
5204 machine_mode mode
, int unsignedp
, int normalizep
,
5205 machine_mode target_mode
)
5208 enum insn_code icode
;
5209 machine_mode compare_mode
;
5210 enum mode_class mclass
;
5211 enum rtx_code scode
;
5215 code
= unsigned_condition (code
);
5216 scode
= swap_condition (code
);
5218 /* If one operand is constant, make it the second one. Only do this
5219 if the other operand is not constant as well. */
5221 if (swap_commutative_operands_p (op0
, op1
))
5226 code
= swap_condition (code
);
5229 if (mode
== VOIDmode
)
5230 mode
= GET_MODE (op0
);
5232 /* For some comparisons with 1 and -1, we can convert this to
5233 comparisons with zero. This will often produce more opportunities for
5234 store-flag insns. */
5239 if (op1
== const1_rtx
)
5240 op1
= const0_rtx
, code
= LE
;
5243 if (op1
== constm1_rtx
)
5244 op1
= const0_rtx
, code
= LT
;
5247 if (op1
== const1_rtx
)
5248 op1
= const0_rtx
, code
= GT
;
5251 if (op1
== constm1_rtx
)
5252 op1
= const0_rtx
, code
= GE
;
5255 if (op1
== const1_rtx
)
5256 op1
= const0_rtx
, code
= NE
;
5259 if (op1
== const1_rtx
)
5260 op1
= const0_rtx
, code
= EQ
;
5266 /* If we are comparing a double-word integer with zero or -1, we can
5267 convert the comparison into one involving a single word. */
5268 if (GET_MODE_BITSIZE (mode
) == BITS_PER_WORD
* 2
5269 && GET_MODE_CLASS (mode
) == MODE_INT
5270 && (!MEM_P (op0
) || ! MEM_VOLATILE_P (op0
)))
5272 if ((code
== EQ
|| code
== NE
)
5273 && (op1
== const0_rtx
|| op1
== constm1_rtx
))
5277 /* Do a logical OR or AND of the two words and compare the
5279 op00
= simplify_gen_subreg (word_mode
, op0
, mode
, 0);
5280 op01
= simplify_gen_subreg (word_mode
, op0
, mode
, UNITS_PER_WORD
);
5281 tem
= expand_binop (word_mode
,
5282 op1
== const0_rtx
? ior_optab
: and_optab
,
5283 op00
, op01
, NULL_RTX
, unsignedp
,
5287 tem
= emit_store_flag (NULL_RTX
, code
, tem
, op1
, word_mode
,
5288 unsignedp
, normalizep
);
5290 else if ((code
== LT
|| code
== GE
) && op1
== const0_rtx
)
5294 /* If testing the sign bit, can just test on high word. */
5295 op0h
= simplify_gen_subreg (word_mode
, op0
, mode
,
5296 subreg_highpart_offset (word_mode
,
5298 tem
= emit_store_flag (NULL_RTX
, code
, op0h
, op1
, word_mode
,
5299 unsignedp
, normalizep
);
5306 if (target_mode
== VOIDmode
|| GET_MODE (tem
) == target_mode
)
5309 target
= gen_reg_rtx (target_mode
);
5311 convert_move (target
, tem
,
5312 !val_signbit_known_set_p (word_mode
,
5313 (normalizep
? normalizep
5314 : STORE_FLAG_VALUE
)));
5319 /* If this is A < 0 or A >= 0, we can do this by taking the ones
5320 complement of A (for GE) and shifting the sign bit to the low bit. */
5321 if (op1
== const0_rtx
&& (code
== LT
|| code
== GE
)
5322 && GET_MODE_CLASS (mode
) == MODE_INT
5323 && (normalizep
|| STORE_FLAG_VALUE
== 1
5324 || val_signbit_p (mode
, STORE_FLAG_VALUE
)))
5331 /* If the result is to be wider than OP0, it is best to convert it
5332 first. If it is to be narrower, it is *incorrect* to convert it
5334 else if (GET_MODE_SIZE (target_mode
) > GET_MODE_SIZE (mode
))
5336 op0
= convert_modes (target_mode
, mode
, op0
, 0);
5340 if (target_mode
!= mode
)
5344 op0
= expand_unop (mode
, one_cmpl_optab
, op0
,
5345 ((STORE_FLAG_VALUE
== 1 || normalizep
)
5346 ? 0 : subtarget
), 0);
5348 if (STORE_FLAG_VALUE
== 1 || normalizep
)
5349 /* If we are supposed to produce a 0/1 value, we want to do
5350 a logical shift from the sign bit to the low-order bit; for
5351 a -1/0 value, we do an arithmetic shift. */
5352 op0
= expand_shift (RSHIFT_EXPR
, mode
, op0
,
5353 GET_MODE_BITSIZE (mode
) - 1,
5354 subtarget
, normalizep
!= -1);
5356 if (mode
!= target_mode
)
5357 op0
= convert_modes (target_mode
, mode
, op0
, 0);
5362 mclass
= GET_MODE_CLASS (mode
);
5363 for (compare_mode
= mode
; compare_mode
!= VOIDmode
;
5364 compare_mode
= GET_MODE_WIDER_MODE (compare_mode
))
5366 machine_mode optab_mode
= mclass
== MODE_CC
? CCmode
: compare_mode
;
5367 icode
= optab_handler (cstore_optab
, optab_mode
);
5368 if (icode
!= CODE_FOR_nothing
)
5370 do_pending_stack_adjust ();
5371 tem
= emit_cstore (target
, icode
, code
, mode
, compare_mode
,
5372 unsignedp
, op0
, op1
, normalizep
, target_mode
);
5376 if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
5378 tem
= emit_cstore (target
, icode
, scode
, mode
, compare_mode
,
5379 unsignedp
, op1
, op0
, normalizep
, target_mode
);
5390 /* Emit a store-flags instruction for comparison CODE on OP0 and OP1
5391 and storing in TARGET. Normally return TARGET.
5392 Return 0 if that cannot be done.
5394 MODE is the mode to use for OP0 and OP1 should they be CONST_INTs. If
5395 it is VOIDmode, they cannot both be CONST_INT.
5397 UNSIGNEDP is for the case where we have to widen the operands
5398 to perform the operation. It says to use zero-extension.
5400 NORMALIZEP is 1 if we should convert the result to be either zero
5401 or one. Normalize is -1 if we should convert the result to be
5402 either zero or -1. If NORMALIZEP is zero, the result will be left
5403 "raw" out of the scc insn. */
5406 emit_store_flag (rtx target
, enum rtx_code code
, rtx op0
, rtx op1
,
5407 machine_mode mode
, int unsignedp
, int normalizep
)
5409 machine_mode target_mode
= target
? GET_MODE (target
) : VOIDmode
;
5410 enum rtx_code rcode
;
5415 /* If we compare constants, we shouldn't use a store-flag operation,
5416 but a constant load. We can get there via the vanilla route that
5417 usually generates a compare-branch sequence, but will in this case
5418 fold the comparison to a constant, and thus elide the branch. */
5419 if (CONSTANT_P (op0
) && CONSTANT_P (op1
))
5422 tem
= emit_store_flag_1 (target
, code
, op0
, op1
, mode
, unsignedp
, normalizep
,
5427 /* If we reached here, we can't do this with a scc insn, however there
5428 are some comparisons that can be done in other ways. Don't do any
5429 of these cases if branches are very cheap. */
5430 if (BRANCH_COST (optimize_insn_for_speed_p (), false) == 0)
5433 /* See what we need to return. We can only return a 1, -1, or the
5436 if (normalizep
== 0)
5438 if (STORE_FLAG_VALUE
== 1 || STORE_FLAG_VALUE
== -1)
5439 normalizep
= STORE_FLAG_VALUE
;
5441 else if (val_signbit_p (mode
, STORE_FLAG_VALUE
))
5447 last
= get_last_insn ();
5449 /* If optimizing, use different pseudo registers for each insn, instead
5450 of reusing the same pseudo. This leads to better CSE, but slows
5451 down the compiler, since there are more pseudos */
5452 subtarget
= (!optimize
5453 && (target_mode
== mode
)) ? target
: NULL_RTX
;
5454 trueval
= GEN_INT (normalizep
? normalizep
: STORE_FLAG_VALUE
);
5456 /* For floating-point comparisons, try the reverse comparison or try
5457 changing the "orderedness" of the comparison. */
5458 if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
5460 enum rtx_code first_code
;
5463 rcode
= reverse_condition_maybe_unordered (code
);
5464 if (can_compare_p (rcode
, mode
, ccp_store_flag
)
5465 && (code
== ORDERED
|| code
== UNORDERED
5466 || (! HONOR_NANS (mode
) && (code
== LTGT
|| code
== UNEQ
))
5467 || (! HONOR_SNANS (mode
) && (code
== EQ
|| code
== NE
))))
5469 int want_add
= ((STORE_FLAG_VALUE
== 1 && normalizep
== -1)
5470 || (STORE_FLAG_VALUE
== -1 && normalizep
== 1));
5472 /* For the reverse comparison, use either an addition or a XOR. */
5474 && rtx_cost (GEN_INT (normalizep
), PLUS
, 1,
5475 optimize_insn_for_speed_p ()) == 0)
5477 tem
= emit_store_flag_1 (subtarget
, rcode
, op0
, op1
, mode
, 0,
5478 STORE_FLAG_VALUE
, target_mode
);
5480 return expand_binop (target_mode
, add_optab
, tem
,
5481 gen_int_mode (normalizep
, target_mode
),
5482 target
, 0, OPTAB_WIDEN
);
5485 && rtx_cost (trueval
, XOR
, 1,
5486 optimize_insn_for_speed_p ()) == 0)
5488 tem
= emit_store_flag_1 (subtarget
, rcode
, op0
, op1
, mode
, 0,
5489 normalizep
, target_mode
);
5491 return expand_binop (target_mode
, xor_optab
, tem
, trueval
,
5492 target
, INTVAL (trueval
) >= 0, OPTAB_WIDEN
);
5496 delete_insns_since (last
);
5498 /* Cannot split ORDERED and UNORDERED, only try the above trick. */
5499 if (code
== ORDERED
|| code
== UNORDERED
)
5502 and_them
= split_comparison (code
, mode
, &first_code
, &code
);
5504 /* If there are no NaNs, the first comparison should always fall through.
5505 Effectively change the comparison to the other one. */
5506 if (!HONOR_NANS (mode
))
5508 gcc_assert (first_code
== (and_them
? ORDERED
: UNORDERED
));
5509 return emit_store_flag_1 (target
, code
, op0
, op1
, mode
, 0, normalizep
,
5513 #ifdef HAVE_conditional_move
5514 /* Try using a setcc instruction for ORDERED/UNORDERED, followed by a
5515 conditional move. */
5516 tem
= emit_store_flag_1 (subtarget
, first_code
, op0
, op1
, mode
, 0,
5517 normalizep
, target_mode
);
5522 tem
= emit_conditional_move (target
, code
, op0
, op1
, mode
,
5523 tem
, const0_rtx
, GET_MODE (tem
), 0);
5525 tem
= emit_conditional_move (target
, code
, op0
, op1
, mode
,
5526 trueval
, tem
, GET_MODE (tem
), 0);
5529 delete_insns_since (last
);
5536 /* The remaining tricks only apply to integer comparisons. */
5538 if (GET_MODE_CLASS (mode
) != MODE_INT
)
5541 /* If this is an equality comparison of integers, we can try to exclusive-or
5542 (or subtract) the two operands and use a recursive call to try the
5543 comparison with zero. Don't do any of these cases if branches are
5546 if ((code
== EQ
|| code
== NE
) && op1
!= const0_rtx
)
5548 tem
= expand_binop (mode
, xor_optab
, op0
, op1
, subtarget
, 1,
5552 tem
= expand_binop (mode
, sub_optab
, op0
, op1
, subtarget
, 1,
5555 tem
= emit_store_flag (target
, code
, tem
, const0_rtx
,
5556 mode
, unsignedp
, normalizep
);
5560 delete_insns_since (last
);
5563 /* For integer comparisons, try the reverse comparison. However, for
5564 small X and if we'd have anyway to extend, implementing "X != 0"
5565 as "-(int)X >> 31" is still cheaper than inverting "(int)X == 0". */
5566 rcode
= reverse_condition (code
);
5567 if (can_compare_p (rcode
, mode
, ccp_store_flag
)
5568 && ! (optab_handler (cstore_optab
, mode
) == CODE_FOR_nothing
5570 && GET_MODE_SIZE (mode
) < UNITS_PER_WORD
5571 && op1
== const0_rtx
))
5573 int want_add
= ((STORE_FLAG_VALUE
== 1 && normalizep
== -1)
5574 || (STORE_FLAG_VALUE
== -1 && normalizep
== 1));
5576 /* Again, for the reverse comparison, use either an addition or a XOR. */
5578 && rtx_cost (GEN_INT (normalizep
), PLUS
, 1,
5579 optimize_insn_for_speed_p ()) == 0)
5581 tem
= emit_store_flag_1 (subtarget
, rcode
, op0
, op1
, mode
, 0,
5582 STORE_FLAG_VALUE
, target_mode
);
5584 tem
= expand_binop (target_mode
, add_optab
, tem
,
5585 gen_int_mode (normalizep
, target_mode
),
5586 target
, 0, OPTAB_WIDEN
);
5589 && rtx_cost (trueval
, XOR
, 1,
5590 optimize_insn_for_speed_p ()) == 0)
5592 tem
= emit_store_flag_1 (subtarget
, rcode
, op0
, op1
, mode
, 0,
5593 normalizep
, target_mode
);
5595 tem
= expand_binop (target_mode
, xor_optab
, tem
, trueval
, target
,
5596 INTVAL (trueval
) >= 0, OPTAB_WIDEN
);
5601 delete_insns_since (last
);
5604 /* Some other cases we can do are EQ, NE, LE, and GT comparisons with
5605 the constant zero. Reject all other comparisons at this point. Only
5606 do LE and GT if branches are expensive since they are expensive on
5607 2-operand machines. */
5609 if (op1
!= const0_rtx
5610 || (code
!= EQ
&& code
!= NE
5611 && (BRANCH_COST (optimize_insn_for_speed_p (),
5612 false) <= 1 || (code
!= LE
&& code
!= GT
))))
5615 /* Try to put the result of the comparison in the sign bit. Assume we can't
5616 do the necessary operation below. */
5620 /* To see if A <= 0, compute (A | (A - 1)). A <= 0 iff that result has
5621 the sign bit set. */
5625 /* This is destructive, so SUBTARGET can't be OP0. */
5626 if (rtx_equal_p (subtarget
, op0
))
5629 tem
= expand_binop (mode
, sub_optab
, op0
, const1_rtx
, subtarget
, 0,
5632 tem
= expand_binop (mode
, ior_optab
, op0
, tem
, subtarget
, 0,
5636 /* To see if A > 0, compute (((signed) A) << BITS) - A, where BITS is the
5637 number of bits in the mode of OP0, minus one. */
5641 if (rtx_equal_p (subtarget
, op0
))
5644 tem
= expand_shift (RSHIFT_EXPR
, mode
, op0
,
5645 GET_MODE_BITSIZE (mode
) - 1,
5647 tem
= expand_binop (mode
, sub_optab
, tem
, op0
, subtarget
, 0,
5651 if (code
== EQ
|| code
== NE
)
5653 /* For EQ or NE, one way to do the comparison is to apply an operation
5654 that converts the operand into a positive number if it is nonzero
5655 or zero if it was originally zero. Then, for EQ, we subtract 1 and
5656 for NE we negate. This puts the result in the sign bit. Then we
5657 normalize with a shift, if needed.
5659 Two operations that can do the above actions are ABS and FFS, so try
5660 them. If that doesn't work, and MODE is smaller than a full word,
5661 we can use zero-extension to the wider mode (an unsigned conversion)
5662 as the operation. */
5664 /* Note that ABS doesn't yield a positive number for INT_MIN, but
5665 that is compensated by the subsequent overflow when subtracting
5668 if (optab_handler (abs_optab
, mode
) != CODE_FOR_nothing
)
5669 tem
= expand_unop (mode
, abs_optab
, op0
, subtarget
, 1);
5670 else if (optab_handler (ffs_optab
, mode
) != CODE_FOR_nothing
)
5671 tem
= expand_unop (mode
, ffs_optab
, op0
, subtarget
, 1);
5672 else if (GET_MODE_SIZE (mode
) < UNITS_PER_WORD
)
5674 tem
= convert_modes (word_mode
, mode
, op0
, 1);
5681 tem
= expand_binop (mode
, sub_optab
, tem
, const1_rtx
, subtarget
,
5684 tem
= expand_unop (mode
, neg_optab
, tem
, subtarget
, 0);
5687 /* If we couldn't do it that way, for NE we can "or" the two's complement
5688 of the value with itself. For EQ, we take the one's complement of
5689 that "or", which is an extra insn, so we only handle EQ if branches
5694 || BRANCH_COST (optimize_insn_for_speed_p (),
5697 if (rtx_equal_p (subtarget
, op0
))
5700 tem
= expand_unop (mode
, neg_optab
, op0
, subtarget
, 0);
5701 tem
= expand_binop (mode
, ior_optab
, tem
, op0
, subtarget
, 0,
5704 if (tem
&& code
== EQ
)
5705 tem
= expand_unop (mode
, one_cmpl_optab
, tem
, subtarget
, 0);
5709 if (tem
&& normalizep
)
5710 tem
= expand_shift (RSHIFT_EXPR
, mode
, tem
,
5711 GET_MODE_BITSIZE (mode
) - 1,
5712 subtarget
, normalizep
== 1);
5718 else if (GET_MODE (tem
) != target_mode
)
5720 convert_move (target
, tem
, 0);
5723 else if (!subtarget
)
5725 emit_move_insn (target
, tem
);
5730 delete_insns_since (last
);
5735 /* Like emit_store_flag, but always succeeds. */
5738 emit_store_flag_force (rtx target
, enum rtx_code code
, rtx op0
, rtx op1
,
5739 machine_mode mode
, int unsignedp
, int normalizep
)
5742 rtx_code_label
*label
;
5743 rtx trueval
, falseval
;
5745 /* First see if emit_store_flag can do the job. */
5746 tem
= emit_store_flag (target
, code
, op0
, op1
, mode
, unsignedp
, normalizep
);
5751 target
= gen_reg_rtx (word_mode
);
5753 /* If this failed, we have to do this with set/compare/jump/set code.
5754 For foo != 0, if foo is in OP0, just replace it with 1 if nonzero. */
5755 trueval
= normalizep
? GEN_INT (normalizep
) : const1_rtx
;
5757 && GET_MODE_CLASS (mode
) == MODE_INT
5760 && op1
== const0_rtx
)
5762 label
= gen_label_rtx ();
5763 do_compare_rtx_and_jump (target
, const0_rtx
, EQ
, unsignedp
,
5764 mode
, NULL_RTX
, NULL_RTX
, label
, -1);
5765 emit_move_insn (target
, trueval
);
5771 || reg_mentioned_p (target
, op0
) || reg_mentioned_p (target
, op1
))
5772 target
= gen_reg_rtx (GET_MODE (target
));
5774 /* Jump in the right direction if the target cannot implement CODE
5775 but can jump on its reverse condition. */
5776 falseval
= const0_rtx
;
5777 if (! can_compare_p (code
, mode
, ccp_jump
)
5778 && (! FLOAT_MODE_P (mode
)
5779 || code
== ORDERED
|| code
== UNORDERED
5780 || (! HONOR_NANS (mode
) && (code
== LTGT
|| code
== UNEQ
))
5781 || (! HONOR_SNANS (mode
) && (code
== EQ
|| code
== NE
))))
5783 enum rtx_code rcode
;
5784 if (FLOAT_MODE_P (mode
))
5785 rcode
= reverse_condition_maybe_unordered (code
);
5787 rcode
= reverse_condition (code
);
5789 /* Canonicalize to UNORDERED for the libcall. */
5790 if (can_compare_p (rcode
, mode
, ccp_jump
)
5791 || (code
== ORDERED
&& ! can_compare_p (ORDERED
, mode
, ccp_jump
)))
5794 trueval
= const0_rtx
;
5799 emit_move_insn (target
, trueval
);
5800 label
= gen_label_rtx ();
5801 do_compare_rtx_and_jump (op0
, op1
, code
, unsignedp
, mode
, NULL_RTX
,
5802 NULL_RTX
, label
, -1);
5804 emit_move_insn (target
, falseval
);
5810 /* Perform possibly multi-word comparison and conditional jump to LABEL
5811 if ARG1 OP ARG2 true where ARG1 and ARG2 are of mode MODE. This is
5812 now a thin wrapper around do_compare_rtx_and_jump. */
5815 do_cmp_and_jump (rtx arg1
, rtx arg2
, enum rtx_code op
, machine_mode mode
,
5816 rtx_code_label
*label
)
5818 int unsignedp
= (op
== LTU
|| op
== LEU
|| op
== GTU
|| op
== GEU
);
5819 do_compare_rtx_and_jump (arg1
, arg2
, op
, unsignedp
, mode
,
5820 NULL_RTX
, NULL_RTX
, label
, -1);