1 /* Subroutines used for code generation on IBM S/390 and zSeries
2 Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006,
3 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
4 Contributed by Hartmut Penner (hpenner@de.ibm.com) and
5 Ulrich Weigand (uweigand@de.ibm.com) and
6 Andreas Krebbel (Andreas.Krebbel@de.ibm.com).
8 This file is part of GCC.
10 GCC is free software; you can redistribute it and/or modify it under
11 the terms of the GNU General Public License as published by the Free
12 Software Foundation; either version 3, or (at your option) any later
15 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
16 WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING3. If not see
22 <http://www.gnu.org/licenses/>. */
26 #include "coretypes.h"
32 #include "hard-reg-set.h"
33 #include "insn-config.h"
34 #include "conditions.h"
36 #include "insn-attr.h"
43 #include "diagnostic-core.h"
45 #include "basic-block.h"
46 #include "integrate.h"
49 #include "target-def.h"
51 #include "langhooks.h"
59 /* Define the specific costs for a given cpu. */
61 struct processor_costs
64 const int m
; /* cost of an M instruction. */
65 const int mghi
; /* cost of an MGHI instruction. */
66 const int mh
; /* cost of an MH instruction. */
67 const int mhi
; /* cost of an MHI instruction. */
68 const int ml
; /* cost of an ML instruction. */
69 const int mr
; /* cost of an MR instruction. */
70 const int ms
; /* cost of an MS instruction. */
71 const int msg
; /* cost of an MSG instruction. */
72 const int msgf
; /* cost of an MSGF instruction. */
73 const int msgfr
; /* cost of an MSGFR instruction. */
74 const int msgr
; /* cost of an MSGR instruction. */
75 const int msr
; /* cost of an MSR instruction. */
76 const int mult_df
; /* cost of multiplication in DFmode. */
79 const int sqxbr
; /* cost of square root in TFmode. */
80 const int sqdbr
; /* cost of square root in DFmode. */
81 const int sqebr
; /* cost of square root in SFmode. */
82 /* multiply and add */
83 const int madbr
; /* cost of multiply and add in DFmode. */
84 const int maebr
; /* cost of multiply and add in SFmode. */
96 const struct processor_costs
*s390_cost
;
99 struct processor_costs z900_cost
=
101 COSTS_N_INSNS (5), /* M */
102 COSTS_N_INSNS (10), /* MGHI */
103 COSTS_N_INSNS (5), /* MH */
104 COSTS_N_INSNS (4), /* MHI */
105 COSTS_N_INSNS (5), /* ML */
106 COSTS_N_INSNS (5), /* MR */
107 COSTS_N_INSNS (4), /* MS */
108 COSTS_N_INSNS (15), /* MSG */
109 COSTS_N_INSNS (7), /* MSGF */
110 COSTS_N_INSNS (7), /* MSGFR */
111 COSTS_N_INSNS (10), /* MSGR */
112 COSTS_N_INSNS (4), /* MSR */
113 COSTS_N_INSNS (7), /* multiplication in DFmode */
114 COSTS_N_INSNS (13), /* MXBR */
115 COSTS_N_INSNS (136), /* SQXBR */
116 COSTS_N_INSNS (44), /* SQDBR */
117 COSTS_N_INSNS (35), /* SQEBR */
118 COSTS_N_INSNS (18), /* MADBR */
119 COSTS_N_INSNS (13), /* MAEBR */
120 COSTS_N_INSNS (134), /* DXBR */
121 COSTS_N_INSNS (30), /* DDBR */
122 COSTS_N_INSNS (27), /* DEBR */
123 COSTS_N_INSNS (220), /* DLGR */
124 COSTS_N_INSNS (34), /* DLR */
125 COSTS_N_INSNS (34), /* DR */
126 COSTS_N_INSNS (32), /* DSGFR */
127 COSTS_N_INSNS (32), /* DSGR */
131 struct processor_costs z990_cost
=
133 COSTS_N_INSNS (4), /* M */
134 COSTS_N_INSNS (2), /* MGHI */
135 COSTS_N_INSNS (2), /* MH */
136 COSTS_N_INSNS (2), /* MHI */
137 COSTS_N_INSNS (4), /* ML */
138 COSTS_N_INSNS (4), /* MR */
139 COSTS_N_INSNS (5), /* MS */
140 COSTS_N_INSNS (6), /* MSG */
141 COSTS_N_INSNS (4), /* MSGF */
142 COSTS_N_INSNS (4), /* MSGFR */
143 COSTS_N_INSNS (4), /* MSGR */
144 COSTS_N_INSNS (4), /* MSR */
145 COSTS_N_INSNS (1), /* multiplication in DFmode */
146 COSTS_N_INSNS (28), /* MXBR */
147 COSTS_N_INSNS (130), /* SQXBR */
148 COSTS_N_INSNS (66), /* SQDBR */
149 COSTS_N_INSNS (38), /* SQEBR */
150 COSTS_N_INSNS (1), /* MADBR */
151 COSTS_N_INSNS (1), /* MAEBR */
152 COSTS_N_INSNS (60), /* DXBR */
153 COSTS_N_INSNS (40), /* DDBR */
154 COSTS_N_INSNS (26), /* DEBR */
155 COSTS_N_INSNS (176), /* DLGR */
156 COSTS_N_INSNS (31), /* DLR */
157 COSTS_N_INSNS (31), /* DR */
158 COSTS_N_INSNS (31), /* DSGFR */
159 COSTS_N_INSNS (31), /* DSGR */
163 struct processor_costs z9_109_cost
=
165 COSTS_N_INSNS (4), /* M */
166 COSTS_N_INSNS (2), /* MGHI */
167 COSTS_N_INSNS (2), /* MH */
168 COSTS_N_INSNS (2), /* MHI */
169 COSTS_N_INSNS (4), /* ML */
170 COSTS_N_INSNS (4), /* MR */
171 COSTS_N_INSNS (5), /* MS */
172 COSTS_N_INSNS (6), /* MSG */
173 COSTS_N_INSNS (4), /* MSGF */
174 COSTS_N_INSNS (4), /* MSGFR */
175 COSTS_N_INSNS (4), /* MSGR */
176 COSTS_N_INSNS (4), /* MSR */
177 COSTS_N_INSNS (1), /* multiplication in DFmode */
178 COSTS_N_INSNS (28), /* MXBR */
179 COSTS_N_INSNS (130), /* SQXBR */
180 COSTS_N_INSNS (66), /* SQDBR */
181 COSTS_N_INSNS (38), /* SQEBR */
182 COSTS_N_INSNS (1), /* MADBR */
183 COSTS_N_INSNS (1), /* MAEBR */
184 COSTS_N_INSNS (60), /* DXBR */
185 COSTS_N_INSNS (40), /* DDBR */
186 COSTS_N_INSNS (26), /* DEBR */
187 COSTS_N_INSNS (30), /* DLGR */
188 COSTS_N_INSNS (23), /* DLR */
189 COSTS_N_INSNS (23), /* DR */
190 COSTS_N_INSNS (24), /* DSGFR */
191 COSTS_N_INSNS (24), /* DSGR */
195 struct processor_costs z10_cost
=
197 COSTS_N_INSNS (10), /* M */
198 COSTS_N_INSNS (10), /* MGHI */
199 COSTS_N_INSNS (10), /* MH */
200 COSTS_N_INSNS (10), /* MHI */
201 COSTS_N_INSNS (10), /* ML */
202 COSTS_N_INSNS (10), /* MR */
203 COSTS_N_INSNS (10), /* MS */
204 COSTS_N_INSNS (10), /* MSG */
205 COSTS_N_INSNS (10), /* MSGF */
206 COSTS_N_INSNS (10), /* MSGFR */
207 COSTS_N_INSNS (10), /* MSGR */
208 COSTS_N_INSNS (10), /* MSR */
209 COSTS_N_INSNS (1) , /* multiplication in DFmode */
210 COSTS_N_INSNS (50), /* MXBR */
211 COSTS_N_INSNS (120), /* SQXBR */
212 COSTS_N_INSNS (52), /* SQDBR */
213 COSTS_N_INSNS (38), /* SQEBR */
214 COSTS_N_INSNS (1), /* MADBR */
215 COSTS_N_INSNS (1), /* MAEBR */
216 COSTS_N_INSNS (111), /* DXBR */
217 COSTS_N_INSNS (39), /* DDBR */
218 COSTS_N_INSNS (32), /* DEBR */
219 COSTS_N_INSNS (160), /* DLGR */
220 COSTS_N_INSNS (71), /* DLR */
221 COSTS_N_INSNS (71), /* DR */
222 COSTS_N_INSNS (71), /* DSGFR */
223 COSTS_N_INSNS (71), /* DSGR */
227 struct processor_costs z196_cost
=
229 COSTS_N_INSNS (7), /* M */
230 COSTS_N_INSNS (5), /* MGHI */
231 COSTS_N_INSNS (5), /* MH */
232 COSTS_N_INSNS (5), /* MHI */
233 COSTS_N_INSNS (7), /* ML */
234 COSTS_N_INSNS (7), /* MR */
235 COSTS_N_INSNS (6), /* MS */
236 COSTS_N_INSNS (8), /* MSG */
237 COSTS_N_INSNS (6), /* MSGF */
238 COSTS_N_INSNS (6), /* MSGFR */
239 COSTS_N_INSNS (8), /* MSGR */
240 COSTS_N_INSNS (6), /* MSR */
241 COSTS_N_INSNS (1) , /* multiplication in DFmode */
242 COSTS_N_INSNS (40), /* MXBR B+40 */
243 COSTS_N_INSNS (100), /* SQXBR B+100 */
244 COSTS_N_INSNS (42), /* SQDBR B+42 */
245 COSTS_N_INSNS (28), /* SQEBR B+28 */
246 COSTS_N_INSNS (1), /* MADBR B */
247 COSTS_N_INSNS (1), /* MAEBR B */
248 COSTS_N_INSNS (101), /* DXBR B+101 */
249 COSTS_N_INSNS (29), /* DDBR */
250 COSTS_N_INSNS (22), /* DEBR */
251 COSTS_N_INSNS (160), /* DLGR cracked */
252 COSTS_N_INSNS (160), /* DLR cracked */
253 COSTS_N_INSNS (160), /* DR expanded */
254 COSTS_N_INSNS (160), /* DSGFR cracked */
255 COSTS_N_INSNS (160), /* DSGR cracked */
258 extern int reload_completed
;
260 /* Kept up to date using the SCHED_VARIABLE_ISSUE hook. */
261 static rtx last_scheduled_insn
;
263 /* Structure used to hold the components of a S/390 memory
264 address. A legitimate address on S/390 is of the general
266 base + index + displacement
267 where any of the components is optional.
269 base and index are registers of the class ADDR_REGS,
270 displacement is an unsigned 12-bit immediate constant. */
281 /* Which cpu are we tuning for. */
282 enum processor_type s390_tune
= PROCESSOR_max
;
284 /* Which instruction set architecture to use. */
285 enum processor_type s390_arch
;
288 HOST_WIDE_INT s390_warn_framesize
= 0;
289 HOST_WIDE_INT s390_stack_size
= 0;
290 HOST_WIDE_INT s390_stack_guard
= 0;
292 /* The following structure is embedded in the machine
293 specific part of struct function. */
295 struct GTY (()) s390_frame_layout
297 /* Offset within stack frame. */
298 HOST_WIDE_INT gprs_offset
;
299 HOST_WIDE_INT f0_offset
;
300 HOST_WIDE_INT f4_offset
;
301 HOST_WIDE_INT f8_offset
;
302 HOST_WIDE_INT backchain_offset
;
304 /* Number of first and last gpr where slots in the register
305 save area are reserved for. */
306 int first_save_gpr_slot
;
307 int last_save_gpr_slot
;
309 /* Number of first and last gpr to be saved, restored. */
311 int first_restore_gpr
;
313 int last_restore_gpr
;
315 /* Bits standing for floating point registers. Set, if the
316 respective register has to be saved. Starting with reg 16 (f0)
317 at the rightmost bit.
318 Bit 15 - 8 7 6 5 4 3 2 1 0
319 fpr 15 - 8 7 5 3 1 6 4 2 0
320 reg 31 - 24 23 22 21 20 19 18 17 16 */
321 unsigned int fpr_bitmap
;
323 /* Number of floating point registers f8-f15 which must be saved. */
326 /* Set if return address needs to be saved.
327 This flag is set by s390_return_addr_rtx if it could not use
328 the initial value of r14 and therefore depends on r14 saved
330 bool save_return_addr_p
;
332 /* Size of stack frame. */
333 HOST_WIDE_INT frame_size
;
336 /* Define the structure for the machine field in struct function. */
338 struct GTY(()) machine_function
340 struct s390_frame_layout frame_layout
;
342 /* Literal pool base register. */
345 /* True if we may need to perform branch splitting. */
346 bool split_branches_pending_p
;
348 /* Some local-dynamic TLS symbol name. */
349 const char *some_ld_name
;
351 bool has_landing_pad_p
;
354 /* Few accessor macros for struct cfun->machine->s390_frame_layout. */
356 #define cfun_frame_layout (cfun->machine->frame_layout)
357 #define cfun_save_high_fprs_p (!!cfun_frame_layout.high_fprs)
358 #define cfun_gprs_save_area_size ((cfun_frame_layout.last_save_gpr_slot - \
359 cfun_frame_layout.first_save_gpr_slot + 1) * UNITS_PER_LONG)
360 #define cfun_set_fpr_bit(BITNUM) (cfun->machine->frame_layout.fpr_bitmap |= \
362 #define cfun_fpr_bit_p(BITNUM) (!!(cfun->machine->frame_layout.fpr_bitmap & \
365 /* Number of GPRs and FPRs used for argument passing. */
366 #define GP_ARG_NUM_REG 5
367 #define FP_ARG_NUM_REG (TARGET_64BIT? 4 : 2)
369 /* A couple of shortcuts. */
370 #define CONST_OK_FOR_J(x) \
371 CONST_OK_FOR_CONSTRAINT_P((x), 'J', "J")
372 #define CONST_OK_FOR_K(x) \
373 CONST_OK_FOR_CONSTRAINT_P((x), 'K', "K")
374 #define CONST_OK_FOR_Os(x) \
375 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Os")
376 #define CONST_OK_FOR_Op(x) \
377 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Op")
378 #define CONST_OK_FOR_On(x) \
379 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "On")
381 #define REGNO_PAIR_OK(REGNO, MODE) \
382 (HARD_REGNO_NREGS ((REGNO), (MODE)) == 1 || !((REGNO) & 1))
384 /* That's the read ahead of the dynamic branch prediction unit in
385 bytes on a z10 (or higher) CPU. */
386 #define PREDICT_DISTANCE (TARGET_Z10 ? 384 : 2048)
388 static enum machine_mode
389 s390_libgcc_cmp_return_mode (void)
391 return TARGET_64BIT
? DImode
: SImode
;
394 static enum machine_mode
395 s390_libgcc_shift_count_mode (void)
397 return TARGET_64BIT
? DImode
: SImode
;
400 static enum machine_mode
401 s390_unwind_word_mode (void)
403 return TARGET_64BIT
? DImode
: SImode
;
406 /* Return true if the back end supports mode MODE. */
408 s390_scalar_mode_supported_p (enum machine_mode mode
)
410 /* In contrast to the default implementation reject TImode constants on 31bit
411 TARGET_ZARCH for ABI compliance. */
412 if (!TARGET_64BIT
&& TARGET_ZARCH
&& mode
== TImode
)
415 if (DECIMAL_FLOAT_MODE_P (mode
))
416 return default_decimal_float_supported_p ();
418 return default_scalar_mode_supported_p (mode
);
421 /* Set the has_landing_pad_p flag in struct machine_function to VALUE. */
424 s390_set_has_landing_pad_p (bool value
)
426 cfun
->machine
->has_landing_pad_p
= value
;
429 /* If two condition code modes are compatible, return a condition code
430 mode which is compatible with both. Otherwise, return
433 static enum machine_mode
434 s390_cc_modes_compatible (enum machine_mode m1
, enum machine_mode m2
)
442 if (m2
== CCUmode
|| m2
== CCTmode
|| m2
== CCZ1mode
443 || m2
== CCSmode
|| m2
== CCSRmode
|| m2
== CCURmode
)
464 /* Return true if SET either doesn't set the CC register, or else
465 the source and destination have matching CC modes and that
466 CC mode is at least as constrained as REQ_MODE. */
469 s390_match_ccmode_set (rtx set
, enum machine_mode req_mode
)
471 enum machine_mode set_mode
;
473 gcc_assert (GET_CODE (set
) == SET
);
475 if (GET_CODE (SET_DEST (set
)) != REG
|| !CC_REGNO_P (REGNO (SET_DEST (set
))))
478 set_mode
= GET_MODE (SET_DEST (set
));
492 if (req_mode
!= set_mode
)
497 if (req_mode
!= CCSmode
&& req_mode
!= CCUmode
&& req_mode
!= CCTmode
498 && req_mode
!= CCSRmode
&& req_mode
!= CCURmode
)
504 if (req_mode
!= CCAmode
)
512 return (GET_MODE (SET_SRC (set
)) == set_mode
);
515 /* Return true if every SET in INSN that sets the CC register
516 has source and destination with matching CC modes and that
517 CC mode is at least as constrained as REQ_MODE.
518 If REQ_MODE is VOIDmode, always return false. */
521 s390_match_ccmode (rtx insn
, enum machine_mode req_mode
)
525 /* s390_tm_ccmode returns VOIDmode to indicate failure. */
526 if (req_mode
== VOIDmode
)
529 if (GET_CODE (PATTERN (insn
)) == SET
)
530 return s390_match_ccmode_set (PATTERN (insn
), req_mode
);
532 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
533 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
535 rtx set
= XVECEXP (PATTERN (insn
), 0, i
);
536 if (GET_CODE (set
) == SET
)
537 if (!s390_match_ccmode_set (set
, req_mode
))
544 /* If a test-under-mask instruction can be used to implement
545 (compare (and ... OP1) OP2), return the CC mode required
546 to do that. Otherwise, return VOIDmode.
547 MIXED is true if the instruction can distinguish between
548 CC1 and CC2 for mixed selected bits (TMxx), it is false
549 if the instruction cannot (TM). */
552 s390_tm_ccmode (rtx op1
, rtx op2
, bool mixed
)
556 /* ??? Fixme: should work on CONST_DOUBLE as well. */
557 if (GET_CODE (op1
) != CONST_INT
|| GET_CODE (op2
) != CONST_INT
)
560 /* Selected bits all zero: CC0.
561 e.g.: int a; if ((a & (16 + 128)) == 0) */
562 if (INTVAL (op2
) == 0)
565 /* Selected bits all one: CC3.
566 e.g.: int a; if ((a & (16 + 128)) == 16 + 128) */
567 if (INTVAL (op2
) == INTVAL (op1
))
570 /* Exactly two bits selected, mixed zeroes and ones: CC1 or CC2. e.g.:
572 if ((a & (16 + 128)) == 16) -> CCT1
573 if ((a & (16 + 128)) == 128) -> CCT2 */
576 bit1
= exact_log2 (INTVAL (op2
));
577 bit0
= exact_log2 (INTVAL (op1
) ^ INTVAL (op2
));
578 if (bit0
!= -1 && bit1
!= -1)
579 return bit0
> bit1
? CCT1mode
: CCT2mode
;
585 /* Given a comparison code OP (EQ, NE, etc.) and the operands
586 OP0 and OP1 of a COMPARE, return the mode to be used for the
590 s390_select_ccmode (enum rtx_code code
, rtx op0
, rtx op1
)
596 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
597 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
599 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
600 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
602 if ((GET_CODE (op0
) == PLUS
|| GET_CODE (op0
) == MINUS
603 || GET_CODE (op1
) == NEG
)
604 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
607 if (GET_CODE (op0
) == AND
)
609 /* Check whether we can potentially do it via TM. */
610 enum machine_mode ccmode
;
611 ccmode
= s390_tm_ccmode (XEXP (op0
, 1), op1
, 1);
612 if (ccmode
!= VOIDmode
)
614 /* Relax CCTmode to CCZmode to allow fall-back to AND
615 if that turns out to be beneficial. */
616 return ccmode
== CCTmode
? CCZmode
: ccmode
;
620 if (register_operand (op0
, HImode
)
621 && GET_CODE (op1
) == CONST_INT
622 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 65535))
624 if (register_operand (op0
, QImode
)
625 && GET_CODE (op1
) == CONST_INT
626 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 255))
635 /* The only overflow condition of NEG and ABS happens when
636 -INT_MAX is used as parameter, which stays negative. So
637 we have an overflow from a positive value to a negative.
638 Using CCAP mode the resulting cc can be used for comparisons. */
639 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
640 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
643 /* If constants are involved in an add instruction it is possible to use
644 the resulting cc for comparisons with zero. Knowing the sign of the
645 constant the overflow behavior gets predictable. e.g.:
646 int a, b; if ((b = a + c) > 0)
647 with c as a constant value: c < 0 -> CCAN and c >= 0 -> CCAP */
648 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
649 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
651 if (INTVAL (XEXP((op0
), 1)) < 0)
665 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
666 && GET_CODE (op1
) != CONST_INT
)
672 if (GET_CODE (op0
) == PLUS
673 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
676 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
677 && GET_CODE (op1
) != CONST_INT
)
683 if (GET_CODE (op0
) == MINUS
684 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
687 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
688 && GET_CODE (op1
) != CONST_INT
)
697 /* Replace the comparison OP0 CODE OP1 by a semantically equivalent one
698 that we can implement more efficiently. */
701 s390_canonicalize_comparison (enum rtx_code
*code
, rtx
*op0
, rtx
*op1
)
703 /* Convert ZERO_EXTRACT back to AND to enable TM patterns. */
704 if ((*code
== EQ
|| *code
== NE
)
705 && *op1
== const0_rtx
706 && GET_CODE (*op0
) == ZERO_EXTRACT
707 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
708 && GET_CODE (XEXP (*op0
, 2)) == CONST_INT
709 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
711 rtx inner
= XEXP (*op0
, 0);
712 HOST_WIDE_INT modesize
= GET_MODE_BITSIZE (GET_MODE (inner
));
713 HOST_WIDE_INT len
= INTVAL (XEXP (*op0
, 1));
714 HOST_WIDE_INT pos
= INTVAL (XEXP (*op0
, 2));
716 if (len
> 0 && len
< modesize
717 && pos
>= 0 && pos
+ len
<= modesize
718 && modesize
<= HOST_BITS_PER_WIDE_INT
)
720 unsigned HOST_WIDE_INT block
;
721 block
= ((unsigned HOST_WIDE_INT
) 1 << len
) - 1;
722 block
<<= modesize
- pos
- len
;
724 *op0
= gen_rtx_AND (GET_MODE (inner
), inner
,
725 gen_int_mode (block
, GET_MODE (inner
)));
729 /* Narrow AND of memory against immediate to enable TM. */
730 if ((*code
== EQ
|| *code
== NE
)
731 && *op1
== const0_rtx
732 && GET_CODE (*op0
) == AND
733 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
734 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
736 rtx inner
= XEXP (*op0
, 0);
737 rtx mask
= XEXP (*op0
, 1);
739 /* Ignore paradoxical SUBREGs if all extra bits are masked out. */
740 if (GET_CODE (inner
) == SUBREG
741 && SCALAR_INT_MODE_P (GET_MODE (SUBREG_REG (inner
)))
742 && (GET_MODE_SIZE (GET_MODE (inner
))
743 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner
))))
745 & GET_MODE_MASK (GET_MODE (inner
))
746 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (inner
))))
748 inner
= SUBREG_REG (inner
);
750 /* Do not change volatile MEMs. */
751 if (MEM_P (inner
) && !MEM_VOLATILE_P (inner
))
753 int part
= s390_single_part (XEXP (*op0
, 1),
754 GET_MODE (inner
), QImode
, 0);
757 mask
= gen_int_mode (s390_extract_part (mask
, QImode
, 0), QImode
);
758 inner
= adjust_address_nv (inner
, QImode
, part
);
759 *op0
= gen_rtx_AND (QImode
, inner
, mask
);
764 /* Narrow comparisons against 0xffff to HImode if possible. */
765 if ((*code
== EQ
|| *code
== NE
)
766 && GET_CODE (*op1
) == CONST_INT
767 && INTVAL (*op1
) == 0xffff
768 && SCALAR_INT_MODE_P (GET_MODE (*op0
))
769 && (nonzero_bits (*op0
, GET_MODE (*op0
))
770 & ~(unsigned HOST_WIDE_INT
) 0xffff) == 0)
772 *op0
= gen_lowpart (HImode
, *op0
);
776 /* Remove redundant UNSPEC_CCU_TO_INT conversions if possible. */
777 if (GET_CODE (*op0
) == UNSPEC
778 && XINT (*op0
, 1) == UNSPEC_CCU_TO_INT
779 && XVECLEN (*op0
, 0) == 1
780 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCUmode
781 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
782 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
783 && *op1
== const0_rtx
)
785 enum rtx_code new_code
= UNKNOWN
;
788 case EQ
: new_code
= EQ
; break;
789 case NE
: new_code
= NE
; break;
790 case LT
: new_code
= GTU
; break;
791 case GT
: new_code
= LTU
; break;
792 case LE
: new_code
= GEU
; break;
793 case GE
: new_code
= LEU
; break;
797 if (new_code
!= UNKNOWN
)
799 *op0
= XVECEXP (*op0
, 0, 0);
804 /* Remove redundant UNSPEC_CCZ_TO_INT conversions if possible. */
805 if (GET_CODE (*op0
) == UNSPEC
806 && XINT (*op0
, 1) == UNSPEC_CCZ_TO_INT
807 && XVECLEN (*op0
, 0) == 1
808 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCZmode
809 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
810 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
811 && *op1
== const0_rtx
)
813 enum rtx_code new_code
= UNKNOWN
;
816 case EQ
: new_code
= EQ
; break;
817 case NE
: new_code
= NE
; break;
821 if (new_code
!= UNKNOWN
)
823 *op0
= XVECEXP (*op0
, 0, 0);
828 /* Simplify cascaded EQ, NE with const0_rtx. */
829 if ((*code
== NE
|| *code
== EQ
)
830 && (GET_CODE (*op0
) == EQ
|| GET_CODE (*op0
) == NE
)
831 && GET_MODE (*op0
) == SImode
832 && GET_MODE (XEXP (*op0
, 0)) == CCZ1mode
833 && REG_P (XEXP (*op0
, 0))
834 && XEXP (*op0
, 1) == const0_rtx
835 && *op1
== const0_rtx
)
837 if ((*code
== EQ
&& GET_CODE (*op0
) == NE
)
838 || (*code
== NE
&& GET_CODE (*op0
) == EQ
))
842 *op0
= XEXP (*op0
, 0);
845 /* Prefer register over memory as first operand. */
846 if (MEM_P (*op0
) && REG_P (*op1
))
848 rtx tem
= *op0
; *op0
= *op1
; *op1
= tem
;
849 *code
= swap_condition (*code
);
853 /* Emit a compare instruction suitable to implement the comparison
854 OP0 CODE OP1. Return the correct condition RTL to be placed in
855 the IF_THEN_ELSE of the conditional branch testing the result. */
858 s390_emit_compare (enum rtx_code code
, rtx op0
, rtx op1
)
860 enum machine_mode mode
= s390_select_ccmode (code
, op0
, op1
);
863 /* Do not output a redundant compare instruction if a compare_and_swap
864 pattern already computed the result and the machine modes are compatible. */
865 if (GET_MODE_CLASS (GET_MODE (op0
)) == MODE_CC
)
867 gcc_assert (s390_cc_modes_compatible (GET_MODE (op0
), mode
)
873 cc
= gen_rtx_REG (mode
, CC_REGNUM
);
874 emit_insn (gen_rtx_SET (VOIDmode
, cc
, gen_rtx_COMPARE (mode
, op0
, op1
)));
877 return gen_rtx_fmt_ee (code
, VOIDmode
, cc
, const0_rtx
);
880 /* Emit a SImode compare and swap instruction setting MEM to NEW_RTX if OLD
882 Return the correct condition RTL to be placed in the IF_THEN_ELSE of the
883 conditional branch testing the result. */
886 s390_emit_compare_and_swap (enum rtx_code code
, rtx old
, rtx mem
, rtx cmp
, rtx new_rtx
)
888 emit_insn (gen_sync_compare_and_swapsi (old
, mem
, cmp
, new_rtx
));
889 return s390_emit_compare (code
, gen_rtx_REG (CCZ1mode
, CC_REGNUM
), const0_rtx
);
892 /* Emit a jump instruction to TARGET. If COND is NULL_RTX, emit an
893 unconditional jump, else a conditional jump under condition COND. */
896 s390_emit_jump (rtx target
, rtx cond
)
900 target
= gen_rtx_LABEL_REF (VOIDmode
, target
);
902 target
= gen_rtx_IF_THEN_ELSE (VOIDmode
, cond
, target
, pc_rtx
);
904 insn
= gen_rtx_SET (VOIDmode
, pc_rtx
, target
);
905 emit_jump_insn (insn
);
908 /* Return branch condition mask to implement a branch
909 specified by CODE. Return -1 for invalid comparisons. */
912 s390_branch_condition_mask (rtx code
)
914 const int CC0
= 1 << 3;
915 const int CC1
= 1 << 2;
916 const int CC2
= 1 << 1;
917 const int CC3
= 1 << 0;
919 gcc_assert (GET_CODE (XEXP (code
, 0)) == REG
);
920 gcc_assert (REGNO (XEXP (code
, 0)) == CC_REGNUM
);
921 gcc_assert (XEXP (code
, 1) == const0_rtx
);
923 switch (GET_MODE (XEXP (code
, 0)))
927 switch (GET_CODE (code
))
930 case NE
: return CC1
| CC2
| CC3
;
936 switch (GET_CODE (code
))
939 case NE
: return CC0
| CC2
| CC3
;
945 switch (GET_CODE (code
))
948 case NE
: return CC0
| CC1
| CC3
;
954 switch (GET_CODE (code
))
957 case NE
: return CC0
| CC1
| CC2
;
963 switch (GET_CODE (code
))
965 case EQ
: return CC0
| CC2
;
966 case NE
: return CC1
| CC3
;
972 switch (GET_CODE (code
))
974 case LTU
: return CC2
| CC3
; /* carry */
975 case GEU
: return CC0
| CC1
; /* no carry */
981 switch (GET_CODE (code
))
983 case GTU
: return CC0
| CC1
; /* borrow */
984 case LEU
: return CC2
| CC3
; /* no borrow */
990 switch (GET_CODE (code
))
992 case EQ
: return CC0
| CC2
;
993 case NE
: return CC1
| CC3
;
994 case LTU
: return CC1
;
995 case GTU
: return CC3
;
996 case LEU
: return CC1
| CC2
;
997 case GEU
: return CC2
| CC3
;
1002 switch (GET_CODE (code
))
1004 case EQ
: return CC0
;
1005 case NE
: return CC1
| CC2
| CC3
;
1006 case LTU
: return CC1
;
1007 case GTU
: return CC2
;
1008 case LEU
: return CC0
| CC1
;
1009 case GEU
: return CC0
| CC2
;
1015 switch (GET_CODE (code
))
1017 case EQ
: return CC0
;
1018 case NE
: return CC2
| CC1
| CC3
;
1019 case LTU
: return CC2
;
1020 case GTU
: return CC1
;
1021 case LEU
: return CC0
| CC2
;
1022 case GEU
: return CC0
| CC1
;
1028 switch (GET_CODE (code
))
1030 case EQ
: return CC0
;
1031 case NE
: return CC1
| CC2
| CC3
;
1032 case LT
: return CC1
| CC3
;
1033 case GT
: return CC2
;
1034 case LE
: return CC0
| CC1
| CC3
;
1035 case GE
: return CC0
| CC2
;
1041 switch (GET_CODE (code
))
1043 case EQ
: return CC0
;
1044 case NE
: return CC1
| CC2
| CC3
;
1045 case LT
: return CC1
;
1046 case GT
: return CC2
| CC3
;
1047 case LE
: return CC0
| CC1
;
1048 case GE
: return CC0
| CC2
| CC3
;
1054 switch (GET_CODE (code
))
1056 case EQ
: return CC0
;
1057 case NE
: return CC1
| CC2
| CC3
;
1058 case LT
: return CC1
;
1059 case GT
: return CC2
;
1060 case LE
: return CC0
| CC1
;
1061 case GE
: return CC0
| CC2
;
1062 case UNORDERED
: return CC3
;
1063 case ORDERED
: return CC0
| CC1
| CC2
;
1064 case UNEQ
: return CC0
| CC3
;
1065 case UNLT
: return CC1
| CC3
;
1066 case UNGT
: return CC2
| CC3
;
1067 case UNLE
: return CC0
| CC1
| CC3
;
1068 case UNGE
: return CC0
| CC2
| CC3
;
1069 case LTGT
: return CC1
| CC2
;
1075 switch (GET_CODE (code
))
1077 case EQ
: return CC0
;
1078 case NE
: return CC2
| CC1
| CC3
;
1079 case LT
: return CC2
;
1080 case GT
: return CC1
;
1081 case LE
: return CC0
| CC2
;
1082 case GE
: return CC0
| CC1
;
1083 case UNORDERED
: return CC3
;
1084 case ORDERED
: return CC0
| CC2
| CC1
;
1085 case UNEQ
: return CC0
| CC3
;
1086 case UNLT
: return CC2
| CC3
;
1087 case UNGT
: return CC1
| CC3
;
1088 case UNLE
: return CC0
| CC2
| CC3
;
1089 case UNGE
: return CC0
| CC1
| CC3
;
1090 case LTGT
: return CC2
| CC1
;
1101 /* Return branch condition mask to implement a compare and branch
1102 specified by CODE. Return -1 for invalid comparisons. */
1105 s390_compare_and_branch_condition_mask (rtx code
)
1107 const int CC0
= 1 << 3;
1108 const int CC1
= 1 << 2;
1109 const int CC2
= 1 << 1;
1111 switch (GET_CODE (code
))
1135 /* If INV is false, return assembler mnemonic string to implement
1136 a branch specified by CODE. If INV is true, return mnemonic
1137 for the corresponding inverted branch. */
1140 s390_branch_condition_mnemonic (rtx code
, int inv
)
1144 static const char *const mnemonic
[16] =
1146 NULL
, "o", "h", "nle",
1147 "l", "nhe", "lh", "ne",
1148 "e", "nlh", "he", "nl",
1149 "le", "nh", "no", NULL
1152 if (GET_CODE (XEXP (code
, 0)) == REG
1153 && REGNO (XEXP (code
, 0)) == CC_REGNUM
1154 && XEXP (code
, 1) == const0_rtx
)
1155 mask
= s390_branch_condition_mask (code
);
1157 mask
= s390_compare_and_branch_condition_mask (code
);
1159 gcc_assert (mask
>= 0);
1164 gcc_assert (mask
>= 1 && mask
<= 14);
1166 return mnemonic
[mask
];
1169 /* Return the part of op which has a value different from def.
1170 The size of the part is determined by mode.
1171 Use this function only if you already know that op really
1172 contains such a part. */
1174 unsigned HOST_WIDE_INT
1175 s390_extract_part (rtx op
, enum machine_mode mode
, int def
)
1177 unsigned HOST_WIDE_INT value
= 0;
1178 int max_parts
= HOST_BITS_PER_WIDE_INT
/ GET_MODE_BITSIZE (mode
);
1179 int part_bits
= GET_MODE_BITSIZE (mode
);
1180 unsigned HOST_WIDE_INT part_mask
1181 = ((unsigned HOST_WIDE_INT
)1 << part_bits
) - 1;
1184 for (i
= 0; i
< max_parts
; i
++)
1187 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1189 value
>>= part_bits
;
1191 if ((value
& part_mask
) != (def
& part_mask
))
1192 return value
& part_mask
;
1198 /* If OP is an integer constant of mode MODE with exactly one
1199 part of mode PART_MODE unequal to DEF, return the number of that
1200 part. Otherwise, return -1. */
1203 s390_single_part (rtx op
,
1204 enum machine_mode mode
,
1205 enum machine_mode part_mode
,
1208 unsigned HOST_WIDE_INT value
= 0;
1209 int n_parts
= GET_MODE_SIZE (mode
) / GET_MODE_SIZE (part_mode
);
1210 unsigned HOST_WIDE_INT part_mask
1211 = ((unsigned HOST_WIDE_INT
)1 << GET_MODE_BITSIZE (part_mode
)) - 1;
1214 if (GET_CODE (op
) != CONST_INT
)
1217 for (i
= 0; i
< n_parts
; i
++)
1220 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1222 value
>>= GET_MODE_BITSIZE (part_mode
);
1224 if ((value
& part_mask
) != (def
& part_mask
))
1232 return part
== -1 ? -1 : n_parts
- 1 - part
;
1235 /* Return true if IN contains a contiguous bitfield in the lower SIZE
1236 bits and no other bits are set in IN. POS and LENGTH can be used
1237 to obtain the start position and the length of the bitfield.
1239 POS gives the position of the first bit of the bitfield counting
1240 from the lowest order bit starting with zero. In order to use this
1241 value for S/390 instructions this has to be converted to "bits big
1245 s390_contiguous_bitmask_p (unsigned HOST_WIDE_INT in
, int size
,
1246 int *pos
, int *length
)
1251 unsigned HOST_WIDE_INT mask
= 1ULL;
1252 bool contiguous
= false;
1254 for (i
= 0; i
< size
; mask
<<= 1, i
++)
1278 /* Calculate a mask for all bits beyond the contiguous bits. */
1279 mask
= (-1LL & ~(((1ULL << (tmp_length
+ tmp_pos
- 1)) << 1) - 1));
1284 if (tmp_length
+ tmp_pos
- 1 > size
)
1288 *length
= tmp_length
;
1296 /* Check whether we can (and want to) split a double-word
1297 move in mode MODE from SRC to DST into two single-word
1298 moves, moving the subword FIRST_SUBWORD first. */
1301 s390_split_ok_p (rtx dst
, rtx src
, enum machine_mode mode
, int first_subword
)
1303 /* Floating point registers cannot be split. */
1304 if (FP_REG_P (src
) || FP_REG_P (dst
))
1307 /* We don't need to split if operands are directly accessible. */
1308 if (s_operand (src
, mode
) || s_operand (dst
, mode
))
1311 /* Non-offsettable memory references cannot be split. */
1312 if ((GET_CODE (src
) == MEM
&& !offsettable_memref_p (src
))
1313 || (GET_CODE (dst
) == MEM
&& !offsettable_memref_p (dst
)))
1316 /* Moving the first subword must not clobber a register
1317 needed to move the second subword. */
1318 if (register_operand (dst
, mode
))
1320 rtx subreg
= operand_subword (dst
, first_subword
, 0, mode
);
1321 if (reg_overlap_mentioned_p (subreg
, src
))
1328 /* Return true if it can be proven that [MEM1, MEM1 + SIZE]
1329 and [MEM2, MEM2 + SIZE] do overlap and false
1333 s390_overlap_p (rtx mem1
, rtx mem2
, HOST_WIDE_INT size
)
1335 rtx addr1
, addr2
, addr_delta
;
1336 HOST_WIDE_INT delta
;
1338 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1344 addr1
= XEXP (mem1
, 0);
1345 addr2
= XEXP (mem2
, 0);
1347 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1349 /* This overlapping check is used by peepholes merging memory block operations.
1350 Overlapping operations would otherwise be recognized by the S/390 hardware
1351 and would fall back to a slower implementation. Allowing overlapping
1352 operations would lead to slow code but not to wrong code. Therefore we are
1353 somewhat optimistic if we cannot prove that the memory blocks are
1355 That's why we return false here although this may accept operations on
1356 overlapping memory areas. */
1357 if (!addr_delta
|| GET_CODE (addr_delta
) != CONST_INT
)
1360 delta
= INTVAL (addr_delta
);
1363 || (delta
> 0 && delta
< size
)
1364 || (delta
< 0 && -delta
< size
))
1370 /* Check whether the address of memory reference MEM2 equals exactly
1371 the address of memory reference MEM1 plus DELTA. Return true if
1372 we can prove this to be the case, false otherwise. */
1375 s390_offset_p (rtx mem1
, rtx mem2
, rtx delta
)
1377 rtx addr1
, addr2
, addr_delta
;
1379 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1382 addr1
= XEXP (mem1
, 0);
1383 addr2
= XEXP (mem2
, 0);
1385 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1386 if (!addr_delta
|| !rtx_equal_p (addr_delta
, delta
))
1392 /* Expand logical operator CODE in mode MODE with operands OPERANDS. */
1395 s390_expand_logical_operator (enum rtx_code code
, enum machine_mode mode
,
1398 enum machine_mode wmode
= mode
;
1399 rtx dst
= operands
[0];
1400 rtx src1
= operands
[1];
1401 rtx src2
= operands
[2];
1404 /* If we cannot handle the operation directly, use a temp register. */
1405 if (!s390_logical_operator_ok_p (operands
))
1406 dst
= gen_reg_rtx (mode
);
1408 /* QImode and HImode patterns make sense only if we have a destination
1409 in memory. Otherwise perform the operation in SImode. */
1410 if ((mode
== QImode
|| mode
== HImode
) && GET_CODE (dst
) != MEM
)
1413 /* Widen operands if required. */
1416 if (GET_CODE (dst
) == SUBREG
1417 && (tem
= simplify_subreg (wmode
, dst
, mode
, 0)) != 0)
1419 else if (REG_P (dst
))
1420 dst
= gen_rtx_SUBREG (wmode
, dst
, 0);
1422 dst
= gen_reg_rtx (wmode
);
1424 if (GET_CODE (src1
) == SUBREG
1425 && (tem
= simplify_subreg (wmode
, src1
, mode
, 0)) != 0)
1427 else if (GET_MODE (src1
) != VOIDmode
)
1428 src1
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src1
), 0);
1430 if (GET_CODE (src2
) == SUBREG
1431 && (tem
= simplify_subreg (wmode
, src2
, mode
, 0)) != 0)
1433 else if (GET_MODE (src2
) != VOIDmode
)
1434 src2
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src2
), 0);
1437 /* Emit the instruction. */
1438 op
= gen_rtx_SET (VOIDmode
, dst
, gen_rtx_fmt_ee (code
, wmode
, src1
, src2
));
1439 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
1440 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clob
)));
1442 /* Fix up the destination if needed. */
1443 if (dst
!= operands
[0])
1444 emit_move_insn (operands
[0], gen_lowpart (mode
, dst
));
1447 /* Check whether OPERANDS are OK for a logical operation (AND, IOR, XOR). */
1450 s390_logical_operator_ok_p (rtx
*operands
)
1452 /* If the destination operand is in memory, it needs to coincide
1453 with one of the source operands. After reload, it has to be
1454 the first source operand. */
1455 if (GET_CODE (operands
[0]) == MEM
)
1456 return rtx_equal_p (operands
[0], operands
[1])
1457 || (!reload_completed
&& rtx_equal_p (operands
[0], operands
[2]));
1462 /* Narrow logical operation CODE of memory operand MEMOP with immediate
1463 operand IMMOP to switch from SS to SI type instructions. */
1466 s390_narrow_logical_operator (enum rtx_code code
, rtx
*memop
, rtx
*immop
)
1468 int def
= code
== AND
? -1 : 0;
1472 gcc_assert (GET_CODE (*memop
) == MEM
);
1473 gcc_assert (!MEM_VOLATILE_P (*memop
));
1475 mask
= s390_extract_part (*immop
, QImode
, def
);
1476 part
= s390_single_part (*immop
, GET_MODE (*memop
), QImode
, def
);
1477 gcc_assert (part
>= 0);
1479 *memop
= adjust_address (*memop
, QImode
, part
);
1480 *immop
= gen_int_mode (mask
, QImode
);
1484 /* How to allocate a 'struct machine_function'. */
1486 static struct machine_function
*
1487 s390_init_machine_status (void)
1489 return ggc_alloc_cleared_machine_function ();
1492 /* Change optimizations to be performed, depending on the
1493 optimization level. */
1495 static const struct default_options s390_option_optimization_table
[] =
1497 { OPT_LEVELS_1_PLUS
, OPT_fomit_frame_pointer
, NULL
, 1 },
1499 /* ??? There are apparently still problems with -fcaller-saves. */
1500 { OPT_LEVELS_ALL
, OPT_fcaller_saves
, NULL
, 0 },
1502 /* Use MVCLE instructions to decrease code size if requested. */
1503 { OPT_LEVELS_SIZE
, OPT_mmvcle
, NULL
, 1 },
1505 { OPT_LEVELS_NONE
, 0, NULL
, 0 }
1508 /* Implement TARGET_OPTION_INIT_STRUCT. */
1511 s390_option_init_struct (struct gcc_options
*opts
)
1513 /* By default, always emit DWARF-2 unwind info. This allows debugging
1514 without maintaining a stack frame back-chain. */
1515 opts
->x_flag_asynchronous_unwind_tables
= 1;
1518 /* Return true if ARG is the name of a processor. Set *TYPE and *FLAGS
1519 to the associated processor_type and processor_flags if so. */
1522 s390_handle_arch_option (const char *arg
,
1523 enum processor_type
*type
,
1528 const char *const name
; /* processor name or nickname. */
1529 const enum processor_type processor
;
1530 const int flags
; /* From enum processor_flags. */
1532 const processor_alias_table
[] =
1534 {"g5", PROCESSOR_9672_G5
, PF_IEEE_FLOAT
},
1535 {"g6", PROCESSOR_9672_G6
, PF_IEEE_FLOAT
},
1536 {"z900", PROCESSOR_2064_Z900
, PF_IEEE_FLOAT
| PF_ZARCH
},
1537 {"z990", PROCESSOR_2084_Z990
, PF_IEEE_FLOAT
| PF_ZARCH
1538 | PF_LONG_DISPLACEMENT
},
1539 {"z9-109", PROCESSOR_2094_Z9_109
, PF_IEEE_FLOAT
| PF_ZARCH
1540 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
},
1541 {"z9-ec", PROCESSOR_2094_Z9_109
, PF_IEEE_FLOAT
| PF_ZARCH
1542 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
| PF_DFP
},
1543 {"z10", PROCESSOR_2097_Z10
, PF_IEEE_FLOAT
| PF_ZARCH
1544 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
| PF_DFP
| PF_Z10
},
1545 {"z196", PROCESSOR_2817_Z196
, PF_IEEE_FLOAT
| PF_ZARCH
1546 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
| PF_DFP
| PF_Z10
| PF_Z196
},
1550 for (i
= 0; i
< ARRAY_SIZE (processor_alias_table
); i
++)
1551 if (strcmp (arg
, processor_alias_table
[i
].name
) == 0)
1553 *type
= processor_alias_table
[i
].processor
;
1554 *flags
= processor_alias_table
[i
].flags
;
1558 *type
= PROCESSOR_max
;
1563 /* Implement TARGET_HANDLE_OPTION. */
1566 s390_handle_option (size_t code
, const char *arg
, int value ATTRIBUTE_UNUSED
)
1571 return s390_handle_arch_option (arg
, &s390_arch
, &s390_arch_flags
);
1573 case OPT_mstack_guard_
:
1574 if (sscanf (arg
, HOST_WIDE_INT_PRINT_DEC
, &s390_stack_guard
) != 1)
1576 if (exact_log2 (s390_stack_guard
) == -1)
1577 error ("stack guard value must be an exact power of 2");
1580 case OPT_mstack_size_
:
1581 if (sscanf (arg
, HOST_WIDE_INT_PRINT_DEC
, &s390_stack_size
) != 1)
1583 if (exact_log2 (s390_stack_size
) == -1)
1584 error ("stack size must be an exact power of 2");
1588 return s390_handle_arch_option (arg
, &s390_tune
, &s390_tune_flags
);
1590 case OPT_mwarn_framesize_
:
1591 return sscanf (arg
, HOST_WIDE_INT_PRINT_DEC
, &s390_warn_framesize
) == 1;
1599 s390_option_override (void)
1601 /* Set up function hooks. */
1602 init_machine_status
= s390_init_machine_status
;
1604 /* Architecture mode defaults according to ABI. */
1605 if (!(target_flags_explicit
& MASK_ZARCH
))
1608 target_flags
|= MASK_ZARCH
;
1610 target_flags
&= ~MASK_ZARCH
;
1613 /* Determine processor architectural level. */
1614 if (!s390_arch_string
)
1616 s390_arch_string
= TARGET_ZARCH
? "z900" : "g5";
1617 s390_handle_arch_option (s390_arch_string
, &s390_arch
, &s390_arch_flags
);
1620 /* This check is triggered when the user specified a wrong -march=
1621 string and prevents subsequent error messages from being
1623 if (s390_arch
== PROCESSOR_max
)
1626 /* Determine processor to tune for. */
1627 if (s390_tune
== PROCESSOR_max
)
1629 s390_tune
= s390_arch
;
1630 s390_tune_flags
= s390_arch_flags
;
1633 /* Sanity checks. */
1634 if (TARGET_ZARCH
&& !TARGET_CPU_ZARCH
)
1635 error ("z/Architecture mode not supported on %s", s390_arch_string
);
1636 if (TARGET_64BIT
&& !TARGET_ZARCH
)
1637 error ("64-bit ABI not supported in ESA/390 mode");
1639 if (TARGET_HARD_DFP
&& !TARGET_DFP
)
1641 if (target_flags_explicit
& MASK_HARD_DFP
)
1643 if (!TARGET_CPU_DFP
)
1644 error ("hardware decimal floating point instructions"
1645 " not available on %s", s390_arch_string
);
1647 error ("hardware decimal floating point instructions"
1648 " not available in ESA/390 mode");
1651 target_flags
&= ~MASK_HARD_DFP
;
1654 if ((target_flags_explicit
& MASK_SOFT_FLOAT
) && TARGET_SOFT_FLOAT
)
1656 if ((target_flags_explicit
& MASK_HARD_DFP
) && TARGET_HARD_DFP
)
1657 error ("-mhard-dfp can%'t be used in conjunction with -msoft-float");
1659 target_flags
&= ~MASK_HARD_DFP
;
1662 /* Set processor cost function. */
1665 case PROCESSOR_2084_Z990
:
1666 s390_cost
= &z990_cost
;
1668 case PROCESSOR_2094_Z9_109
:
1669 s390_cost
= &z9_109_cost
;
1671 case PROCESSOR_2097_Z10
:
1672 s390_cost
= &z10_cost
;
1673 case PROCESSOR_2817_Z196
:
1674 s390_cost
= &z196_cost
;
1677 s390_cost
= &z900_cost
;
1680 if (TARGET_BACKCHAIN
&& TARGET_PACKED_STACK
&& TARGET_HARD_FLOAT
)
1681 error ("-mbackchain -mpacked-stack -mhard-float are not supported "
1684 if (s390_stack_size
)
1686 if (s390_stack_guard
>= s390_stack_size
)
1687 error ("stack size must be greater than the stack guard value");
1688 else if (s390_stack_size
> 1 << 16)
1689 error ("stack size must not be greater than 64k");
1691 else if (s390_stack_guard
)
1692 error ("-mstack-guard implies use of -mstack-size");
1694 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1695 if (!(target_flags_explicit
& MASK_LONG_DOUBLE_128
))
1696 target_flags
|= MASK_LONG_DOUBLE_128
;
1699 if (s390_tune
== PROCESSOR_2097_Z10
1700 || s390_tune
== PROCESSOR_2817_Z196
)
1702 maybe_set_param_value (PARAM_MAX_UNROLLED_INSNS
, 100,
1703 global_options
.x_param_values
,
1704 global_options_set
.x_param_values
);
1705 maybe_set_param_value (PARAM_MAX_UNROLL_TIMES
, 32,
1706 global_options
.x_param_values
,
1707 global_options_set
.x_param_values
);
1708 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEELED_INSNS
, 2000,
1709 global_options
.x_param_values
,
1710 global_options_set
.x_param_values
);
1711 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEEL_TIMES
, 64,
1712 global_options
.x_param_values
,
1713 global_options_set
.x_param_values
);
1716 maybe_set_param_value (PARAM_MAX_PENDING_LIST_LENGTH
, 256,
1717 global_options
.x_param_values
,
1718 global_options_set
.x_param_values
);
1719 /* values for loop prefetching */
1720 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE
, 256,
1721 global_options
.x_param_values
,
1722 global_options_set
.x_param_values
);
1723 maybe_set_param_value (PARAM_L1_CACHE_SIZE
, 128,
1724 global_options
.x_param_values
,
1725 global_options_set
.x_param_values
);
1726 /* s390 has more than 2 levels and the size is much larger. Since
1727 we are always running virtualized assume that we only get a small
1728 part of the caches above l1. */
1729 maybe_set_param_value (PARAM_L2_CACHE_SIZE
, 1500,
1730 global_options
.x_param_values
,
1731 global_options_set
.x_param_values
);
1732 maybe_set_param_value (PARAM_PREFETCH_MIN_INSN_TO_MEM_RATIO
, 2,
1733 global_options
.x_param_values
,
1734 global_options_set
.x_param_values
);
1735 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
, 6,
1736 global_options
.x_param_values
,
1737 global_options_set
.x_param_values
);
1739 /* This cannot reside in s390_option_optimization_table since HAVE_prefetch
1740 requires the arch flags to be evaluated already. Since prefetching
1741 is beneficial on s390, we enable it if available. */
1742 if (flag_prefetch_loop_arrays
< 0 && HAVE_prefetch
&& optimize
>= 3)
1743 flag_prefetch_loop_arrays
= 1;
1746 /* Map for smallest class containing reg regno. */
1748 const enum reg_class regclass_map
[FIRST_PSEUDO_REGISTER
] =
1749 { GENERAL_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1750 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1751 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1752 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1753 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1754 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1755 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1756 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1757 ADDR_REGS
, CC_REGS
, ADDR_REGS
, ADDR_REGS
,
1758 ACCESS_REGS
, ACCESS_REGS
1761 /* Return attribute type of insn. */
1763 static enum attr_type
1764 s390_safe_attr_type (rtx insn
)
1766 if (recog_memoized (insn
) >= 0)
1767 return get_attr_type (insn
);
1772 /* Return true if DISP is a valid short displacement. */
1775 s390_short_displacement (rtx disp
)
1777 /* No displacement is OK. */
1781 /* Without the long displacement facility we don't need to
1782 distingiush between long and short displacement. */
1783 if (!TARGET_LONG_DISPLACEMENT
)
1786 /* Integer displacement in range. */
1787 if (GET_CODE (disp
) == CONST_INT
)
1788 return INTVAL (disp
) >= 0 && INTVAL (disp
) < 4096;
1790 /* GOT offset is not OK, the GOT can be large. */
1791 if (GET_CODE (disp
) == CONST
1792 && GET_CODE (XEXP (disp
, 0)) == UNSPEC
1793 && (XINT (XEXP (disp
, 0), 1) == UNSPEC_GOT
1794 || XINT (XEXP (disp
, 0), 1) == UNSPEC_GOTNTPOFF
))
1797 /* All other symbolic constants are literal pool references,
1798 which are OK as the literal pool must be small. */
1799 if (GET_CODE (disp
) == CONST
)
1805 /* Decompose a RTL expression ADDR for a memory address into
1806 its components, returned in OUT.
1808 Returns false if ADDR is not a valid memory address, true
1809 otherwise. If OUT is NULL, don't return the components,
1810 but check for validity only.
1812 Note: Only addresses in canonical form are recognized.
1813 LEGITIMIZE_ADDRESS should convert non-canonical forms to the
1814 canonical form so that they will be recognized. */
1817 s390_decompose_address (rtx addr
, struct s390_address
*out
)
1819 HOST_WIDE_INT offset
= 0;
1820 rtx base
= NULL_RTX
;
1821 rtx indx
= NULL_RTX
;
1822 rtx disp
= NULL_RTX
;
1824 bool pointer
= false;
1825 bool base_ptr
= false;
1826 bool indx_ptr
= false;
1827 bool literal_pool
= false;
1829 /* We may need to substitute the literal pool base register into the address
1830 below. However, at this point we do not know which register is going to
1831 be used as base, so we substitute the arg pointer register. This is going
1832 to be treated as holding a pointer below -- it shouldn't be used for any
1834 rtx fake_pool_base
= gen_rtx_REG (Pmode
, ARG_POINTER_REGNUM
);
1836 /* Decompose address into base + index + displacement. */
1838 if (GET_CODE (addr
) == REG
|| GET_CODE (addr
) == UNSPEC
)
1841 else if (GET_CODE (addr
) == PLUS
)
1843 rtx op0
= XEXP (addr
, 0);
1844 rtx op1
= XEXP (addr
, 1);
1845 enum rtx_code code0
= GET_CODE (op0
);
1846 enum rtx_code code1
= GET_CODE (op1
);
1848 if (code0
== REG
|| code0
== UNSPEC
)
1850 if (code1
== REG
|| code1
== UNSPEC
)
1852 indx
= op0
; /* index + base */
1858 base
= op0
; /* base + displacement */
1863 else if (code0
== PLUS
)
1865 indx
= XEXP (op0
, 0); /* index + base + disp */
1866 base
= XEXP (op0
, 1);
1877 disp
= addr
; /* displacement */
1879 /* Extract integer part of displacement. */
1883 if (GET_CODE (disp
) == CONST_INT
)
1885 offset
= INTVAL (disp
);
1888 else if (GET_CODE (disp
) == CONST
1889 && GET_CODE (XEXP (disp
, 0)) == PLUS
1890 && GET_CODE (XEXP (XEXP (disp
, 0), 1)) == CONST_INT
)
1892 offset
= INTVAL (XEXP (XEXP (disp
, 0), 1));
1893 disp
= XEXP (XEXP (disp
, 0), 0);
1897 /* Strip off CONST here to avoid special case tests later. */
1898 if (disp
&& GET_CODE (disp
) == CONST
)
1899 disp
= XEXP (disp
, 0);
1901 /* We can convert literal pool addresses to
1902 displacements by basing them off the base register. */
1903 if (disp
&& GET_CODE (disp
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (disp
))
1905 /* Either base or index must be free to hold the base register. */
1907 base
= fake_pool_base
, literal_pool
= true;
1909 indx
= fake_pool_base
, literal_pool
= true;
1913 /* Mark up the displacement. */
1914 disp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, disp
),
1915 UNSPEC_LTREL_OFFSET
);
1918 /* Validate base register. */
1921 if (GET_CODE (base
) == UNSPEC
)
1922 switch (XINT (base
, 1))
1926 disp
= gen_rtx_UNSPEC (Pmode
,
1927 gen_rtvec (1, XVECEXP (base
, 0, 0)),
1928 UNSPEC_LTREL_OFFSET
);
1932 base
= XVECEXP (base
, 0, 1);
1935 case UNSPEC_LTREL_BASE
:
1936 if (XVECLEN (base
, 0) == 1)
1937 base
= fake_pool_base
, literal_pool
= true;
1939 base
= XVECEXP (base
, 0, 1);
1947 || (GET_MODE (base
) != SImode
1948 && GET_MODE (base
) != Pmode
))
1951 if (REGNO (base
) == STACK_POINTER_REGNUM
1952 || REGNO (base
) == FRAME_POINTER_REGNUM
1953 || ((reload_completed
|| reload_in_progress
)
1954 && frame_pointer_needed
1955 && REGNO (base
) == HARD_FRAME_POINTER_REGNUM
)
1956 || REGNO (base
) == ARG_POINTER_REGNUM
1958 && REGNO (base
) == PIC_OFFSET_TABLE_REGNUM
))
1959 pointer
= base_ptr
= true;
1961 if ((reload_completed
|| reload_in_progress
)
1962 && base
== cfun
->machine
->base_reg
)
1963 pointer
= base_ptr
= literal_pool
= true;
1966 /* Validate index register. */
1969 if (GET_CODE (indx
) == UNSPEC
)
1970 switch (XINT (indx
, 1))
1974 disp
= gen_rtx_UNSPEC (Pmode
,
1975 gen_rtvec (1, XVECEXP (indx
, 0, 0)),
1976 UNSPEC_LTREL_OFFSET
);
1980 indx
= XVECEXP (indx
, 0, 1);
1983 case UNSPEC_LTREL_BASE
:
1984 if (XVECLEN (indx
, 0) == 1)
1985 indx
= fake_pool_base
, literal_pool
= true;
1987 indx
= XVECEXP (indx
, 0, 1);
1995 || (GET_MODE (indx
) != SImode
1996 && GET_MODE (indx
) != Pmode
))
1999 if (REGNO (indx
) == STACK_POINTER_REGNUM
2000 || REGNO (indx
) == FRAME_POINTER_REGNUM
2001 || ((reload_completed
|| reload_in_progress
)
2002 && frame_pointer_needed
2003 && REGNO (indx
) == HARD_FRAME_POINTER_REGNUM
)
2004 || REGNO (indx
) == ARG_POINTER_REGNUM
2006 && REGNO (indx
) == PIC_OFFSET_TABLE_REGNUM
))
2007 pointer
= indx_ptr
= true;
2009 if ((reload_completed
|| reload_in_progress
)
2010 && indx
== cfun
->machine
->base_reg
)
2011 pointer
= indx_ptr
= literal_pool
= true;
2014 /* Prefer to use pointer as base, not index. */
2015 if (base
&& indx
&& !base_ptr
2016 && (indx_ptr
|| (!REG_POINTER (base
) && REG_POINTER (indx
))))
2023 /* Validate displacement. */
2026 /* If virtual registers are involved, the displacement will change later
2027 anyway as the virtual registers get eliminated. This could make a
2028 valid displacement invalid, but it is more likely to make an invalid
2029 displacement valid, because we sometimes access the register save area
2030 via negative offsets to one of those registers.
2031 Thus we don't check the displacement for validity here. If after
2032 elimination the displacement turns out to be invalid after all,
2033 this is fixed up by reload in any case. */
2034 if (base
!= arg_pointer_rtx
2035 && indx
!= arg_pointer_rtx
2036 && base
!= return_address_pointer_rtx
2037 && indx
!= return_address_pointer_rtx
2038 && base
!= frame_pointer_rtx
2039 && indx
!= frame_pointer_rtx
2040 && base
!= virtual_stack_vars_rtx
2041 && indx
!= virtual_stack_vars_rtx
)
2042 if (!DISP_IN_RANGE (offset
))
2047 /* All the special cases are pointers. */
2050 /* In the small-PIC case, the linker converts @GOT
2051 and @GOTNTPOFF offsets to possible displacements. */
2052 if (GET_CODE (disp
) == UNSPEC
2053 && (XINT (disp
, 1) == UNSPEC_GOT
2054 || XINT (disp
, 1) == UNSPEC_GOTNTPOFF
)
2060 /* Accept pool label offsets. */
2061 else if (GET_CODE (disp
) == UNSPEC
2062 && XINT (disp
, 1) == UNSPEC_POOL_OFFSET
)
2065 /* Accept literal pool references. */
2066 else if (GET_CODE (disp
) == UNSPEC
2067 && XINT (disp
, 1) == UNSPEC_LTREL_OFFSET
)
2069 orig_disp
= gen_rtx_CONST (Pmode
, disp
);
2072 /* If we have an offset, make sure it does not
2073 exceed the size of the constant pool entry. */
2074 rtx sym
= XVECEXP (disp
, 0, 0);
2075 if (offset
>= GET_MODE_SIZE (get_pool_mode (sym
)))
2078 orig_disp
= plus_constant (orig_disp
, offset
);
2093 out
->disp
= orig_disp
;
2094 out
->pointer
= pointer
;
2095 out
->literal_pool
= literal_pool
;
2101 /* Decompose a RTL expression OP for a shift count into its components,
2102 and return the base register in BASE and the offset in OFFSET.
2104 Return true if OP is a valid shift count, false if not. */
2107 s390_decompose_shift_count (rtx op
, rtx
*base
, HOST_WIDE_INT
*offset
)
2109 HOST_WIDE_INT off
= 0;
2111 /* We can have an integer constant, an address register,
2112 or a sum of the two. */
2113 if (GET_CODE (op
) == CONST_INT
)
2118 if (op
&& GET_CODE (op
) == PLUS
&& GET_CODE (XEXP (op
, 1)) == CONST_INT
)
2120 off
= INTVAL (XEXP (op
, 1));
2123 while (op
&& GET_CODE (op
) == SUBREG
)
2124 op
= SUBREG_REG (op
);
2126 if (op
&& GET_CODE (op
) != REG
)
2138 /* Return true if CODE is a valid address without index. */
2141 s390_legitimate_address_without_index_p (rtx op
)
2143 struct s390_address addr
;
2145 if (!s390_decompose_address (XEXP (op
, 0), &addr
))
2154 /* Return true if ADDR is of kind symbol_ref or symbol_ref + const_int
2155 and return these parts in SYMREF and ADDEND. You can pass NULL in
2156 SYMREF and/or ADDEND if you are not interested in these values.
2157 Literal pool references are *not* considered symbol references. */
2160 s390_symref_operand_p (rtx addr
, rtx
*symref
, HOST_WIDE_INT
*addend
)
2162 HOST_WIDE_INT tmpaddend
= 0;
2164 if (GET_CODE (addr
) == CONST
)
2165 addr
= XEXP (addr
, 0);
2167 if (GET_CODE (addr
) == PLUS
)
2169 if (GET_CODE (XEXP (addr
, 0)) == SYMBOL_REF
2170 && !CONSTANT_POOL_ADDRESS_P (XEXP (addr
, 0))
2171 && CONST_INT_P (XEXP (addr
, 1)))
2173 tmpaddend
= INTVAL (XEXP (addr
, 1));
2174 addr
= XEXP (addr
, 0);
2180 if (GET_CODE (addr
) != SYMBOL_REF
|| CONSTANT_POOL_ADDRESS_P (addr
))
2186 *addend
= tmpaddend
;
2192 /* Return true if the address in OP is valid for constraint letter C
2193 if wrapped in a MEM rtx. Set LIT_POOL_OK to true if it literal
2194 pool MEMs should be accepted. Only the Q, R, S, T constraint
2195 letters are allowed for C. */
2198 s390_check_qrst_address (char c
, rtx op
, bool lit_pool_ok
)
2200 struct s390_address addr
;
2201 bool decomposed
= false;
2203 /* This check makes sure that no symbolic address (except literal
2204 pool references) are accepted by the R or T constraints. */
2205 if (s390_symref_operand_p (op
, NULL
, NULL
))
2208 /* Ensure literal pool references are only accepted if LIT_POOL_OK. */
2211 if (!s390_decompose_address (op
, &addr
))
2213 if (addr
.literal_pool
)
2220 case 'Q': /* no index short displacement */
2221 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2225 if (!s390_short_displacement (addr
.disp
))
2229 case 'R': /* with index short displacement */
2230 if (TARGET_LONG_DISPLACEMENT
)
2232 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2234 if (!s390_short_displacement (addr
.disp
))
2237 /* Any invalid address here will be fixed up by reload,
2238 so accept it for the most generic constraint. */
2241 case 'S': /* no index long displacement */
2242 if (!TARGET_LONG_DISPLACEMENT
)
2244 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2248 if (s390_short_displacement (addr
.disp
))
2252 case 'T': /* with index long displacement */
2253 if (!TARGET_LONG_DISPLACEMENT
)
2255 /* Any invalid address here will be fixed up by reload,
2256 so accept it for the most generic constraint. */
2257 if ((decomposed
|| s390_decompose_address (op
, &addr
))
2258 && s390_short_displacement (addr
.disp
))
2268 /* Evaluates constraint strings described by the regular expression
2269 ([A|B|Z](Q|R|S|T))|U|W|Y and returns 1 if OP is a valid operand for
2270 the constraint given in STR, or 0 else. */
2273 s390_mem_constraint (const char *str
, rtx op
)
2280 /* Check for offsettable variants of memory constraints. */
2281 if (!MEM_P (op
) || MEM_VOLATILE_P (op
))
2283 if ((reload_completed
|| reload_in_progress
)
2284 ? !offsettable_memref_p (op
) : !offsettable_nonstrict_memref_p (op
))
2286 return s390_check_qrst_address (str
[1], XEXP (op
, 0), true);
2288 /* Check for non-literal-pool variants of memory constraints. */
2291 return s390_check_qrst_address (str
[1], XEXP (op
, 0), false);
2296 if (GET_CODE (op
) != MEM
)
2298 return s390_check_qrst_address (c
, XEXP (op
, 0), true);
2300 return (s390_check_qrst_address ('Q', op
, true)
2301 || s390_check_qrst_address ('R', op
, true));
2303 return (s390_check_qrst_address ('S', op
, true)
2304 || s390_check_qrst_address ('T', op
, true));
2306 /* Simply check for the basic form of a shift count. Reload will
2307 take care of making sure we have a proper base register. */
2308 if (!s390_decompose_shift_count (op
, NULL
, NULL
))
2312 return s390_check_qrst_address (str
[1], op
, true);
2320 /* Evaluates constraint strings starting with letter O. Input
2321 parameter C is the second letter following the "O" in the constraint
2322 string. Returns 1 if VALUE meets the respective constraint and 0
2326 s390_O_constraint_str (const char c
, HOST_WIDE_INT value
)
2334 return trunc_int_for_mode (value
, SImode
) == value
;
2338 || s390_single_part (GEN_INT (value
), DImode
, SImode
, 0) == 1;
2341 return s390_single_part (GEN_INT (value
- 1), DImode
, SImode
, -1) == 1;
2349 /* Evaluates constraint strings starting with letter N. Parameter STR
2350 contains the letters following letter "N" in the constraint string.
2351 Returns true if VALUE matches the constraint. */
2354 s390_N_constraint_str (const char *str
, HOST_WIDE_INT value
)
2356 enum machine_mode mode
, part_mode
;
2358 int part
, part_goal
;
2364 part_goal
= str
[0] - '0';
2408 if (GET_MODE_SIZE (mode
) <= GET_MODE_SIZE (part_mode
))
2411 part
= s390_single_part (GEN_INT (value
), mode
, part_mode
, def
);
2414 if (part_goal
!= -1 && part_goal
!= part
)
2421 /* Returns true if the input parameter VALUE is a float zero. */
2424 s390_float_const_zero_p (rtx value
)
2426 return (GET_MODE_CLASS (GET_MODE (value
)) == MODE_FLOAT
2427 && value
== CONST0_RTX (GET_MODE (value
)));
2431 /* Compute a (partial) cost for rtx X. Return true if the complete
2432 cost has been computed, and false if subexpressions should be
2433 scanned. In either case, *TOTAL contains the cost result.
2434 CODE contains GET_CODE (x), OUTER_CODE contains the code
2435 of the superexpression of x. */
2438 s390_rtx_costs (rtx x
, int code
, int outer_code
, int *total
,
2439 bool speed ATTRIBUTE_UNUSED
)
2462 *total
= COSTS_N_INSNS (1);
2467 *total
= COSTS_N_INSNS (1);
2471 switch (GET_MODE (x
))
2475 rtx left
= XEXP (x
, 0);
2476 rtx right
= XEXP (x
, 1);
2477 if (GET_CODE (right
) == CONST_INT
2478 && CONST_OK_FOR_K (INTVAL (right
)))
2479 *total
= s390_cost
->mhi
;
2480 else if (GET_CODE (left
) == SIGN_EXTEND
)
2481 *total
= s390_cost
->mh
;
2483 *total
= s390_cost
->ms
; /* msr, ms, msy */
2488 rtx left
= XEXP (x
, 0);
2489 rtx right
= XEXP (x
, 1);
2492 if (GET_CODE (right
) == CONST_INT
2493 && CONST_OK_FOR_K (INTVAL (right
)))
2494 *total
= s390_cost
->mghi
;
2495 else if (GET_CODE (left
) == SIGN_EXTEND
)
2496 *total
= s390_cost
->msgf
;
2498 *total
= s390_cost
->msg
; /* msgr, msg */
2500 else /* TARGET_31BIT */
2502 if (GET_CODE (left
) == SIGN_EXTEND
2503 && GET_CODE (right
) == SIGN_EXTEND
)
2504 /* mulsidi case: mr, m */
2505 *total
= s390_cost
->m
;
2506 else if (GET_CODE (left
) == ZERO_EXTEND
2507 && GET_CODE (right
) == ZERO_EXTEND
2508 && TARGET_CPU_ZARCH
)
2509 /* umulsidi case: ml, mlr */
2510 *total
= s390_cost
->ml
;
2512 /* Complex calculation is required. */
2513 *total
= COSTS_N_INSNS (40);
2519 *total
= s390_cost
->mult_df
;
2522 *total
= s390_cost
->mxbr
;
2530 switch (GET_MODE (x
))
2533 *total
= s390_cost
->madbr
;
2536 *total
= s390_cost
->maebr
;
2541 /* Negate in the third argument is free: FMSUB. */
2542 if (GET_CODE (XEXP (x
, 2)) == NEG
)
2544 *total
+= (rtx_cost (XEXP (x
, 0), FMA
, speed
)
2545 + rtx_cost (XEXP (x
, 1), FMA
, speed
)
2546 + rtx_cost (XEXP (XEXP (x
, 2), 0), FMA
, speed
));
2553 if (GET_MODE (x
) == TImode
) /* 128 bit division */
2554 *total
= s390_cost
->dlgr
;
2555 else if (GET_MODE (x
) == DImode
)
2557 rtx right
= XEXP (x
, 1);
2558 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2559 *total
= s390_cost
->dlr
;
2560 else /* 64 by 64 bit division */
2561 *total
= s390_cost
->dlgr
;
2563 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2564 *total
= s390_cost
->dlr
;
2569 if (GET_MODE (x
) == DImode
)
2571 rtx right
= XEXP (x
, 1);
2572 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2574 *total
= s390_cost
->dsgfr
;
2576 *total
= s390_cost
->dr
;
2577 else /* 64 by 64 bit division */
2578 *total
= s390_cost
->dsgr
;
2580 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2581 *total
= s390_cost
->dlr
;
2582 else if (GET_MODE (x
) == SFmode
)
2584 *total
= s390_cost
->debr
;
2586 else if (GET_MODE (x
) == DFmode
)
2588 *total
= s390_cost
->ddbr
;
2590 else if (GET_MODE (x
) == TFmode
)
2592 *total
= s390_cost
->dxbr
;
2597 if (GET_MODE (x
) == SFmode
)
2598 *total
= s390_cost
->sqebr
;
2599 else if (GET_MODE (x
) == DFmode
)
2600 *total
= s390_cost
->sqdbr
;
2602 *total
= s390_cost
->sqxbr
;
2607 if (outer_code
== MULT
|| outer_code
== DIV
|| outer_code
== MOD
2608 || outer_code
== PLUS
|| outer_code
== MINUS
2609 || outer_code
== COMPARE
)
2614 *total
= COSTS_N_INSNS (1);
2615 if (GET_CODE (XEXP (x
, 0)) == AND
2616 && GET_CODE (XEXP (x
, 1)) == CONST_INT
2617 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
)
2619 rtx op0
= XEXP (XEXP (x
, 0), 0);
2620 rtx op1
= XEXP (XEXP (x
, 0), 1);
2621 rtx op2
= XEXP (x
, 1);
2623 if (memory_operand (op0
, GET_MODE (op0
))
2624 && s390_tm_ccmode (op1
, op2
, 0) != VOIDmode
)
2626 if (register_operand (op0
, GET_MODE (op0
))
2627 && s390_tm_ccmode (op1
, op2
, 1) != VOIDmode
)
2637 /* Return the cost of an address rtx ADDR. */
2640 s390_address_cost (rtx addr
, bool speed ATTRIBUTE_UNUSED
)
2642 struct s390_address ad
;
2643 if (!s390_decompose_address (addr
, &ad
))
2646 return ad
.indx
? COSTS_N_INSNS (1) + 1 : COSTS_N_INSNS (1);
2649 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
2650 otherwise return 0. */
2653 tls_symbolic_operand (rtx op
)
2655 if (GET_CODE (op
) != SYMBOL_REF
)
2657 return SYMBOL_REF_TLS_MODEL (op
);
2660 /* Split DImode access register reference REG (on 64-bit) into its constituent
2661 low and high parts, and store them into LO and HI. Note that gen_lowpart/
2662 gen_highpart cannot be used as they assume all registers are word-sized,
2663 while our access registers have only half that size. */
2666 s390_split_access_reg (rtx reg
, rtx
*lo
, rtx
*hi
)
2668 gcc_assert (TARGET_64BIT
);
2669 gcc_assert (ACCESS_REG_P (reg
));
2670 gcc_assert (GET_MODE (reg
) == DImode
);
2671 gcc_assert (!(REGNO (reg
) & 1));
2673 *lo
= gen_rtx_REG (SImode
, REGNO (reg
) + 1);
2674 *hi
= gen_rtx_REG (SImode
, REGNO (reg
));
2677 /* Return true if OP contains a symbol reference */
2680 symbolic_reference_mentioned_p (rtx op
)
2685 if (GET_CODE (op
) == SYMBOL_REF
|| GET_CODE (op
) == LABEL_REF
)
2688 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2689 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2695 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2696 if (symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2700 else if (fmt
[i
] == 'e' && symbolic_reference_mentioned_p (XEXP (op
, i
)))
2707 /* Return true if OP contains a reference to a thread-local symbol. */
2710 tls_symbolic_reference_mentioned_p (rtx op
)
2715 if (GET_CODE (op
) == SYMBOL_REF
)
2716 return tls_symbolic_operand (op
);
2718 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2719 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2725 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2726 if (tls_symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2730 else if (fmt
[i
] == 'e' && tls_symbolic_reference_mentioned_p (XEXP (op
, i
)))
2738 /* Return true if OP is a legitimate general operand when
2739 generating PIC code. It is given that flag_pic is on
2740 and that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2743 legitimate_pic_operand_p (rtx op
)
2745 /* Accept all non-symbolic constants. */
2746 if (!SYMBOLIC_CONST (op
))
2749 /* Reject everything else; must be handled
2750 via emit_symbolic_move. */
2754 /* Returns true if the constant value OP is a legitimate general operand.
2755 It is given that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2758 legitimate_constant_p (rtx op
)
2760 /* Accept all non-symbolic constants. */
2761 if (!SYMBOLIC_CONST (op
))
2764 /* Accept immediate LARL operands. */
2765 if (TARGET_CPU_ZARCH
&& larl_operand (op
, VOIDmode
))
2768 /* Thread-local symbols are never legal constants. This is
2769 so that emit_call knows that computing such addresses
2770 might require a function call. */
2771 if (TLS_SYMBOLIC_CONST (op
))
2774 /* In the PIC case, symbolic constants must *not* be
2775 forced into the literal pool. We accept them here,
2776 so that they will be handled by emit_symbolic_move. */
2780 /* All remaining non-PIC symbolic constants are
2781 forced into the literal pool. */
2785 /* Determine if it's legal to put X into the constant pool. This
2786 is not possible if X contains the address of a symbol that is
2787 not constant (TLS) or not known at final link time (PIC). */
2790 s390_cannot_force_const_mem (rtx x
)
2792 switch (GET_CODE (x
))
2796 /* Accept all non-symbolic constants. */
2800 /* Labels are OK iff we are non-PIC. */
2801 return flag_pic
!= 0;
2804 /* 'Naked' TLS symbol references are never OK,
2805 non-TLS symbols are OK iff we are non-PIC. */
2806 if (tls_symbolic_operand (x
))
2809 return flag_pic
!= 0;
2812 return s390_cannot_force_const_mem (XEXP (x
, 0));
2815 return s390_cannot_force_const_mem (XEXP (x
, 0))
2816 || s390_cannot_force_const_mem (XEXP (x
, 1));
2819 switch (XINT (x
, 1))
2821 /* Only lt-relative or GOT-relative UNSPECs are OK. */
2822 case UNSPEC_LTREL_OFFSET
:
2830 case UNSPEC_GOTNTPOFF
:
2831 case UNSPEC_INDNTPOFF
:
2834 /* If the literal pool shares the code section, be put
2835 execute template placeholders into the pool as well. */
2837 return TARGET_CPU_ZARCH
;
2849 /* Returns true if the constant value OP is a legitimate general
2850 operand during and after reload. The difference to
2851 legitimate_constant_p is that this function will not accept
2852 a constant that would need to be forced to the literal pool
2853 before it can be used as operand.
2854 This function accepts all constants which can be loaded directly
2858 legitimate_reload_constant_p (rtx op
)
2860 /* Accept la(y) operands. */
2861 if (GET_CODE (op
) == CONST_INT
2862 && DISP_IN_RANGE (INTVAL (op
)))
2865 /* Accept l(g)hi/l(g)fi operands. */
2866 if (GET_CODE (op
) == CONST_INT
2867 && (CONST_OK_FOR_K (INTVAL (op
)) || CONST_OK_FOR_Os (INTVAL (op
))))
2870 /* Accept lliXX operands. */
2872 && GET_CODE (op
) == CONST_INT
2873 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2874 && s390_single_part (op
, word_mode
, HImode
, 0) >= 0)
2878 && GET_CODE (op
) == CONST_INT
2879 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2880 && s390_single_part (op
, word_mode
, SImode
, 0) >= 0)
2883 /* Accept larl operands. */
2884 if (TARGET_CPU_ZARCH
2885 && larl_operand (op
, VOIDmode
))
2888 /* Accept floating-point zero operands that fit into a single GPR. */
2889 if (GET_CODE (op
) == CONST_DOUBLE
2890 && s390_float_const_zero_p (op
)
2891 && GET_MODE_SIZE (GET_MODE (op
)) <= UNITS_PER_WORD
)
2894 /* Accept double-word operands that can be split. */
2895 if (GET_CODE (op
) == CONST_INT
2896 && trunc_int_for_mode (INTVAL (op
), word_mode
) != INTVAL (op
))
2898 enum machine_mode dword_mode
= word_mode
== SImode
? DImode
: TImode
;
2899 rtx hi
= operand_subword (op
, 0, 0, dword_mode
);
2900 rtx lo
= operand_subword (op
, 1, 0, dword_mode
);
2901 return legitimate_reload_constant_p (hi
)
2902 && legitimate_reload_constant_p (lo
);
2905 /* Everything else cannot be handled without reload. */
2909 /* Returns true if the constant value OP is a legitimate fp operand
2910 during and after reload.
2911 This function accepts all constants which can be loaded directly
2915 legitimate_reload_fp_constant_p (rtx op
)
2917 /* Accept floating-point zero operands if the load zero instruction
2920 && GET_CODE (op
) == CONST_DOUBLE
2921 && s390_float_const_zero_p (op
))
2927 /* Given an rtx OP being reloaded into a reg required to be in class RCLASS,
2928 return the class of reg to actually use. */
2931 s390_preferred_reload_class (rtx op
, enum reg_class rclass
)
2933 switch (GET_CODE (op
))
2935 /* Constants we cannot reload into general registers
2936 must be forced into the literal pool. */
2939 if (reg_class_subset_p (GENERAL_REGS
, rclass
)
2940 && legitimate_reload_constant_p (op
))
2941 return GENERAL_REGS
;
2942 else if (reg_class_subset_p (ADDR_REGS
, rclass
)
2943 && legitimate_reload_constant_p (op
))
2945 else if (reg_class_subset_p (FP_REGS
, rclass
)
2946 && legitimate_reload_fp_constant_p (op
))
2950 /* If a symbolic constant or a PLUS is reloaded,
2951 it is most likely being used as an address, so
2952 prefer ADDR_REGS. If 'class' is not a superset
2953 of ADDR_REGS, e.g. FP_REGS, reject this reload. */
2958 if (reg_class_subset_p (ADDR_REGS
, rclass
))
2970 /* Return true if ADDR is SYMBOL_REF + addend with addend being a
2971 multiple of ALIGNMENT and the SYMBOL_REF being naturally
2975 s390_check_symref_alignment (rtx addr
, HOST_WIDE_INT alignment
)
2977 HOST_WIDE_INT addend
;
2980 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2983 return (!SYMBOL_REF_NOT_NATURALLY_ALIGNED_P (symref
)
2984 && !(addend
& (alignment
- 1)));
2987 /* ADDR is moved into REG using larl. If ADDR isn't a valid larl
2988 operand SCRATCH is used to reload the even part of the address and
2992 s390_reload_larl_operand (rtx reg
, rtx addr
, rtx scratch
)
2994 HOST_WIDE_INT addend
;
2997 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
3001 /* Easy case. The addend is even so larl will do fine. */
3002 emit_move_insn (reg
, addr
);
3005 /* We can leave the scratch register untouched if the target
3006 register is a valid base register. */
3007 if (REGNO (reg
) < FIRST_PSEUDO_REGISTER
3008 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
)
3011 gcc_assert (REGNO (scratch
) < FIRST_PSEUDO_REGISTER
);
3012 gcc_assert (REGNO_REG_CLASS (REGNO (scratch
)) == ADDR_REGS
);
3015 emit_move_insn (scratch
,
3016 gen_rtx_CONST (Pmode
,
3017 gen_rtx_PLUS (Pmode
, symref
,
3018 GEN_INT (addend
- 1))));
3020 emit_move_insn (scratch
, symref
);
3022 /* Increment the address using la in order to avoid clobbering cc. */
3023 emit_move_insn (reg
, gen_rtx_PLUS (Pmode
, scratch
, const1_rtx
));
3027 /* Generate what is necessary to move between REG and MEM using
3028 SCRATCH. The direction is given by TOMEM. */
3031 s390_reload_symref_address (rtx reg
, rtx mem
, rtx scratch
, bool tomem
)
3033 /* Reload might have pulled a constant out of the literal pool.
3034 Force it back in. */
3035 if (CONST_INT_P (mem
) || GET_CODE (mem
) == CONST_DOUBLE
3036 || GET_CODE (mem
) == CONST
)
3037 mem
= force_const_mem (GET_MODE (reg
), mem
);
3039 gcc_assert (MEM_P (mem
));
3041 /* For a load from memory we can leave the scratch register
3042 untouched if the target register is a valid base register. */
3044 && REGNO (reg
) < FIRST_PSEUDO_REGISTER
3045 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
3046 && GET_MODE (reg
) == GET_MODE (scratch
))
3049 /* Load address into scratch register. Since we can't have a
3050 secondary reload for a secondary reload we have to cover the case
3051 where larl would need a secondary reload here as well. */
3052 s390_reload_larl_operand (scratch
, XEXP (mem
, 0), scratch
);
3054 /* Now we can use a standard load/store to do the move. */
3056 emit_move_insn (replace_equiv_address (mem
, scratch
), reg
);
3058 emit_move_insn (reg
, replace_equiv_address (mem
, scratch
));
3061 /* Inform reload about cases where moving X with a mode MODE to a register in
3062 RCLASS requires an extra scratch or immediate register. Return the class
3063 needed for the immediate register. */
3066 s390_secondary_reload (bool in_p
, rtx x
, reg_class_t rclass_i
,
3067 enum machine_mode mode
, secondary_reload_info
*sri
)
3069 enum reg_class rclass
= (enum reg_class
) rclass_i
;
3071 /* Intermediate register needed. */
3072 if (reg_classes_intersect_p (CC_REGS
, rclass
))
3073 return GENERAL_REGS
;
3077 /* On z10 several optimizer steps may generate larl operands with
3080 && s390_symref_operand_p (x
, NULL
, NULL
)
3082 && !s390_check_symref_alignment (x
, 2))
3083 sri
->icode
= ((mode
== DImode
) ? CODE_FOR_reloaddi_larl_odd_addend_z10
3084 : CODE_FOR_reloadsi_larl_odd_addend_z10
);
3086 /* On z10 we need a scratch register when moving QI, TI or floating
3087 point mode values from or to a memory location with a SYMBOL_REF
3088 or if the symref addend of a SI or DI move is not aligned to the
3089 width of the access. */
3091 && s390_symref_operand_p (XEXP (x
, 0), NULL
, NULL
)
3092 && (mode
== QImode
|| mode
== TImode
|| FLOAT_MODE_P (mode
)
3093 || (!TARGET_ZARCH
&& mode
== DImode
)
3094 || ((mode
== HImode
|| mode
== SImode
|| mode
== DImode
)
3095 && (!s390_check_symref_alignment (XEXP (x
, 0),
3096 GET_MODE_SIZE (mode
))))))
3098 #define __SECONDARY_RELOAD_CASE(M,m) \
3101 sri->icode = in_p ? CODE_FOR_reload##m##di_toreg_z10 : \
3102 CODE_FOR_reload##m##di_tomem_z10; \
3104 sri->icode = in_p ? CODE_FOR_reload##m##si_toreg_z10 : \
3105 CODE_FOR_reload##m##si_tomem_z10; \
3108 switch (GET_MODE (x
))
3110 __SECONDARY_RELOAD_CASE (QI
, qi
);
3111 __SECONDARY_RELOAD_CASE (HI
, hi
);
3112 __SECONDARY_RELOAD_CASE (SI
, si
);
3113 __SECONDARY_RELOAD_CASE (DI
, di
);
3114 __SECONDARY_RELOAD_CASE (TI
, ti
);
3115 __SECONDARY_RELOAD_CASE (SF
, sf
);
3116 __SECONDARY_RELOAD_CASE (DF
, df
);
3117 __SECONDARY_RELOAD_CASE (TF
, tf
);
3118 __SECONDARY_RELOAD_CASE (SD
, sd
);
3119 __SECONDARY_RELOAD_CASE (DD
, dd
);
3120 __SECONDARY_RELOAD_CASE (TD
, td
);
3125 #undef __SECONDARY_RELOAD_CASE
3129 /* We need a scratch register when loading a PLUS expression which
3130 is not a legitimate operand of the LOAD ADDRESS instruction. */
3131 if (in_p
&& s390_plus_operand (x
, mode
))
3132 sri
->icode
= (TARGET_64BIT
?
3133 CODE_FOR_reloaddi_plus
: CODE_FOR_reloadsi_plus
);
3135 /* Performing a multiword move from or to memory we have to make sure the
3136 second chunk in memory is addressable without causing a displacement
3137 overflow. If that would be the case we calculate the address in
3138 a scratch register. */
3140 && GET_CODE (XEXP (x
, 0)) == PLUS
3141 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
3142 && !DISP_IN_RANGE (INTVAL (XEXP (XEXP (x
, 0), 1))
3143 + GET_MODE_SIZE (mode
) - 1))
3145 /* For GENERAL_REGS a displacement overflow is no problem if occurring
3146 in a s_operand address since we may fallback to lm/stm. So we only
3147 have to care about overflows in the b+i+d case. */
3148 if ((reg_classes_intersect_p (GENERAL_REGS
, rclass
)
3149 && s390_class_max_nregs (GENERAL_REGS
, mode
) > 1
3150 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == PLUS
)
3151 /* For FP_REGS no lm/stm is available so this check is triggered
3152 for displacement overflows in b+i+d and b+d like addresses. */
3153 || (reg_classes_intersect_p (FP_REGS
, rclass
)
3154 && s390_class_max_nregs (FP_REGS
, mode
) > 1))
3157 sri
->icode
= (TARGET_64BIT
?
3158 CODE_FOR_reloaddi_nonoffmem_in
:
3159 CODE_FOR_reloadsi_nonoffmem_in
);
3161 sri
->icode
= (TARGET_64BIT
?
3162 CODE_FOR_reloaddi_nonoffmem_out
:
3163 CODE_FOR_reloadsi_nonoffmem_out
);
3167 /* A scratch address register is needed when a symbolic constant is
3168 copied to r0 compiling with -fPIC. In other cases the target
3169 register might be used as temporary (see legitimize_pic_address). */
3170 if (in_p
&& SYMBOLIC_CONST (x
) && flag_pic
== 2 && rclass
!= ADDR_REGS
)
3171 sri
->icode
= (TARGET_64BIT
?
3172 CODE_FOR_reloaddi_PIC_addr
:
3173 CODE_FOR_reloadsi_PIC_addr
);
3175 /* Either scratch or no register needed. */
3179 /* Generate code to load SRC, which is PLUS that is not a
3180 legitimate operand for the LA instruction, into TARGET.
3181 SCRATCH may be used as scratch register. */
3184 s390_expand_plus_operand (rtx target
, rtx src
,
3188 struct s390_address ad
;
3190 /* src must be a PLUS; get its two operands. */
3191 gcc_assert (GET_CODE (src
) == PLUS
);
3192 gcc_assert (GET_MODE (src
) == Pmode
);
3194 /* Check if any of the two operands is already scheduled
3195 for replacement by reload. This can happen e.g. when
3196 float registers occur in an address. */
3197 sum1
= find_replacement (&XEXP (src
, 0));
3198 sum2
= find_replacement (&XEXP (src
, 1));
3199 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3201 /* If the address is already strictly valid, there's nothing to do. */
3202 if (!s390_decompose_address (src
, &ad
)
3203 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3204 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
3206 /* Otherwise, one of the operands cannot be an address register;
3207 we reload its value into the scratch register. */
3208 if (true_regnum (sum1
) < 1 || true_regnum (sum1
) > 15)
3210 emit_move_insn (scratch
, sum1
);
3213 if (true_regnum (sum2
) < 1 || true_regnum (sum2
) > 15)
3215 emit_move_insn (scratch
, sum2
);
3219 /* According to the way these invalid addresses are generated
3220 in reload.c, it should never happen (at least on s390) that
3221 *neither* of the PLUS components, after find_replacements
3222 was applied, is an address register. */
3223 if (sum1
== scratch
&& sum2
== scratch
)
3229 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3232 /* Emit the LOAD ADDRESS pattern. Note that reload of PLUS
3233 is only ever performed on addresses, so we can mark the
3234 sum as legitimate for LA in any case. */
3235 s390_load_address (target
, src
);
3239 /* Return true if ADDR is a valid memory address.
3240 STRICT specifies whether strict register checking applies. */
3243 s390_legitimate_address_p (enum machine_mode mode
, rtx addr
, bool strict
)
3245 struct s390_address ad
;
3248 && larl_operand (addr
, VOIDmode
)
3249 && (mode
== VOIDmode
3250 || s390_check_symref_alignment (addr
, GET_MODE_SIZE (mode
))))
3253 if (!s390_decompose_address (addr
, &ad
))
3258 if (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3261 if (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
)))
3267 && !(REGNO (ad
.base
) >= FIRST_PSEUDO_REGISTER
3268 || REGNO_REG_CLASS (REGNO (ad
.base
)) == ADDR_REGS
))
3272 && !(REGNO (ad
.indx
) >= FIRST_PSEUDO_REGISTER
3273 || REGNO_REG_CLASS (REGNO (ad
.indx
)) == ADDR_REGS
))
3279 /* Return true if OP is a valid operand for the LA instruction.
3280 In 31-bit, we need to prove that the result is used as an
3281 address, as LA performs only a 31-bit addition. */
3284 legitimate_la_operand_p (rtx op
)
3286 struct s390_address addr
;
3287 if (!s390_decompose_address (op
, &addr
))
3290 return (TARGET_64BIT
|| addr
.pointer
);
3293 /* Return true if it is valid *and* preferable to use LA to
3294 compute the sum of OP1 and OP2. */
3297 preferred_la_operand_p (rtx op1
, rtx op2
)
3299 struct s390_address addr
;
3301 if (op2
!= const0_rtx
)
3302 op1
= gen_rtx_PLUS (Pmode
, op1
, op2
);
3304 if (!s390_decompose_address (op1
, &addr
))
3306 if (addr
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (addr
.base
)))
3308 if (addr
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (addr
.indx
)))
3311 /* Avoid LA instructions with index register on z196; it is
3312 preferable to use regular add instructions when possible. */
3313 if (addr
.indx
&& s390_tune
== PROCESSOR_2817_Z196
)
3316 if (!TARGET_64BIT
&& !addr
.pointer
)
3322 if ((addr
.base
&& REG_P (addr
.base
) && REG_POINTER (addr
.base
))
3323 || (addr
.indx
&& REG_P (addr
.indx
) && REG_POINTER (addr
.indx
)))
3329 /* Emit a forced load-address operation to load SRC into DST.
3330 This will use the LOAD ADDRESS instruction even in situations
3331 where legitimate_la_operand_p (SRC) returns false. */
3334 s390_load_address (rtx dst
, rtx src
)
3337 emit_move_insn (dst
, src
);
3339 emit_insn (gen_force_la_31 (dst
, src
));
3342 /* Return a legitimate reference for ORIG (an address) using the
3343 register REG. If REG is 0, a new pseudo is generated.
3345 There are two types of references that must be handled:
3347 1. Global data references must load the address from the GOT, via
3348 the PIC reg. An insn is emitted to do this load, and the reg is
3351 2. Static data references, constant pool addresses, and code labels
3352 compute the address as an offset from the GOT, whose base is in
3353 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
3354 differentiate them from global data objects. The returned
3355 address is the PIC reg + an unspec constant.
3357 TARGET_LEGITIMIZE_ADDRESS_P rejects symbolic references unless the PIC
3358 reg also appears in the address. */
3361 legitimize_pic_address (rtx orig
, rtx reg
)
3367 gcc_assert (!TLS_SYMBOLIC_CONST (addr
));
3369 if (GET_CODE (addr
) == LABEL_REF
3370 || (GET_CODE (addr
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (addr
)))
3372 /* This is a local symbol. */
3373 if (TARGET_CPU_ZARCH
&& larl_operand (addr
, VOIDmode
))
3375 /* Access local symbols PC-relative via LARL.
3376 This is the same as in the non-PIC case, so it is
3377 handled automatically ... */
3381 /* Access local symbols relative to the GOT. */
3383 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3385 if (reload_in_progress
|| reload_completed
)
3386 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3388 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
3389 addr
= gen_rtx_CONST (Pmode
, addr
);
3390 addr
= force_const_mem (Pmode
, addr
);
3391 emit_move_insn (temp
, addr
);
3393 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3396 s390_load_address (reg
, new_rtx
);
3401 else if (GET_CODE (addr
) == SYMBOL_REF
)
3404 reg
= gen_reg_rtx (Pmode
);
3408 /* Assume GOT offset < 4k. This is handled the same way
3409 in both 31- and 64-bit code (@GOT). */
3411 if (reload_in_progress
|| reload_completed
)
3412 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3414 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3415 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3416 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3417 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3418 emit_move_insn (reg
, new_rtx
);
3421 else if (TARGET_CPU_ZARCH
)
3423 /* If the GOT offset might be >= 4k, we determine the position
3424 of the GOT entry via a PC-relative LARL (@GOTENT). */
3426 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3428 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3429 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3431 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTENT
);
3432 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3433 emit_move_insn (temp
, new_rtx
);
3435 new_rtx
= gen_const_mem (Pmode
, temp
);
3436 emit_move_insn (reg
, new_rtx
);
3441 /* If the GOT offset might be >= 4k, we have to load it
3442 from the literal pool (@GOT). */
3444 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3446 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3447 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3449 if (reload_in_progress
|| reload_completed
)
3450 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3452 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3453 addr
= gen_rtx_CONST (Pmode
, addr
);
3454 addr
= force_const_mem (Pmode
, addr
);
3455 emit_move_insn (temp
, addr
);
3457 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3458 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3459 emit_move_insn (reg
, new_rtx
);
3465 if (GET_CODE (addr
) == CONST
)
3467 addr
= XEXP (addr
, 0);
3468 if (GET_CODE (addr
) == UNSPEC
)
3470 gcc_assert (XVECLEN (addr
, 0) == 1);
3471 switch (XINT (addr
, 1))
3473 /* If someone moved a GOT-relative UNSPEC
3474 out of the literal pool, force them back in. */
3477 new_rtx
= force_const_mem (Pmode
, orig
);
3480 /* @GOT is OK as is if small. */
3483 new_rtx
= force_const_mem (Pmode
, orig
);
3486 /* @GOTENT is OK as is. */
3490 /* @PLT is OK as is on 64-bit, must be converted to
3491 GOT-relative @PLTOFF on 31-bit. */
3493 if (!TARGET_CPU_ZARCH
)
3495 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3497 if (reload_in_progress
|| reload_completed
)
3498 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3500 addr
= XVECEXP (addr
, 0, 0);
3501 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
),
3503 addr
= gen_rtx_CONST (Pmode
, addr
);
3504 addr
= force_const_mem (Pmode
, addr
);
3505 emit_move_insn (temp
, addr
);
3507 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3510 s390_load_address (reg
, new_rtx
);
3516 /* Everything else cannot happen. */
3522 gcc_assert (GET_CODE (addr
) == PLUS
);
3524 if (GET_CODE (addr
) == PLUS
)
3526 rtx op0
= XEXP (addr
, 0), op1
= XEXP (addr
, 1);
3528 gcc_assert (!TLS_SYMBOLIC_CONST (op0
));
3529 gcc_assert (!TLS_SYMBOLIC_CONST (op1
));
3531 /* Check first to see if this is a constant offset
3532 from a local symbol reference. */
3533 if ((GET_CODE (op0
) == LABEL_REF
3534 || (GET_CODE (op0
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (op0
)))
3535 && GET_CODE (op1
) == CONST_INT
)
3537 if (TARGET_CPU_ZARCH
3538 && larl_operand (op0
, VOIDmode
)
3539 && INTVAL (op1
) < (HOST_WIDE_INT
)1 << 31
3540 && INTVAL (op1
) >= -((HOST_WIDE_INT
)1 << 31))
3542 if (INTVAL (op1
) & 1)
3544 /* LARL can't handle odd offsets, so emit a
3545 pair of LARL and LA. */
3546 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3548 if (!DISP_IN_RANGE (INTVAL (op1
)))
3550 HOST_WIDE_INT even
= INTVAL (op1
) - 1;
3551 op0
= gen_rtx_PLUS (Pmode
, op0
, GEN_INT (even
));
3552 op0
= gen_rtx_CONST (Pmode
, op0
);
3556 emit_move_insn (temp
, op0
);
3557 new_rtx
= gen_rtx_PLUS (Pmode
, temp
, op1
);
3561 s390_load_address (reg
, new_rtx
);
3567 /* If the offset is even, we can just use LARL.
3568 This will happen automatically. */
3573 /* Access local symbols relative to the GOT. */
3575 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3577 if (reload_in_progress
|| reload_completed
)
3578 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3580 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op0
),
3582 addr
= gen_rtx_PLUS (Pmode
, addr
, op1
);
3583 addr
= gen_rtx_CONST (Pmode
, addr
);
3584 addr
= force_const_mem (Pmode
, addr
);
3585 emit_move_insn (temp
, addr
);
3587 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3590 s390_load_address (reg
, new_rtx
);
3596 /* Now, check whether it is a GOT relative symbol plus offset
3597 that was pulled out of the literal pool. Force it back in. */
3599 else if (GET_CODE (op0
) == UNSPEC
3600 && GET_CODE (op1
) == CONST_INT
3601 && XINT (op0
, 1) == UNSPEC_GOTOFF
)
3603 gcc_assert (XVECLEN (op0
, 0) == 1);
3605 new_rtx
= force_const_mem (Pmode
, orig
);
3608 /* Otherwise, compute the sum. */
3611 base
= legitimize_pic_address (XEXP (addr
, 0), reg
);
3612 new_rtx
= legitimize_pic_address (XEXP (addr
, 1),
3613 base
== reg
? NULL_RTX
: reg
);
3614 if (GET_CODE (new_rtx
) == CONST_INT
)
3615 new_rtx
= plus_constant (base
, INTVAL (new_rtx
));
3618 if (GET_CODE (new_rtx
) == PLUS
&& CONSTANT_P (XEXP (new_rtx
, 1)))
3620 base
= gen_rtx_PLUS (Pmode
, base
, XEXP (new_rtx
, 0));
3621 new_rtx
= XEXP (new_rtx
, 1);
3623 new_rtx
= gen_rtx_PLUS (Pmode
, base
, new_rtx
);
3626 if (GET_CODE (new_rtx
) == CONST
)
3627 new_rtx
= XEXP (new_rtx
, 0);
3628 new_rtx
= force_operand (new_rtx
, 0);
3635 /* Load the thread pointer into a register. */
3638 s390_get_thread_pointer (void)
3640 rtx tp
= gen_reg_rtx (Pmode
);
3642 emit_move_insn (tp
, gen_rtx_REG (Pmode
, TP_REGNUM
));
3643 mark_reg_pointer (tp
, BITS_PER_WORD
);
3648 /* Emit a tls call insn. The call target is the SYMBOL_REF stored
3649 in s390_tls_symbol which always refers to __tls_get_offset.
3650 The returned offset is written to RESULT_REG and an USE rtx is
3651 generated for TLS_CALL. */
3653 static GTY(()) rtx s390_tls_symbol
;
3656 s390_emit_tls_call_insn (rtx result_reg
, rtx tls_call
)
3660 gcc_assert (flag_pic
);
3662 if (!s390_tls_symbol
)
3663 s390_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tls_get_offset");
3665 insn
= s390_emit_call (s390_tls_symbol
, tls_call
, result_reg
,
3666 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
3668 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), result_reg
);
3669 RTL_CONST_CALL_P (insn
) = 1;
3672 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3673 this (thread-local) address. REG may be used as temporary. */
3676 legitimize_tls_address (rtx addr
, rtx reg
)
3678 rtx new_rtx
, tls_call
, temp
, base
, r2
, insn
;
3680 if (GET_CODE (addr
) == SYMBOL_REF
)
3681 switch (tls_symbolic_operand (addr
))
3683 case TLS_MODEL_GLOBAL_DYNAMIC
:
3685 r2
= gen_rtx_REG (Pmode
, 2);
3686 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_TLSGD
);
3687 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3688 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3689 emit_move_insn (r2
, new_rtx
);
3690 s390_emit_tls_call_insn (r2
, tls_call
);
3691 insn
= get_insns ();
3694 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3695 temp
= gen_reg_rtx (Pmode
);
3696 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3698 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3701 s390_load_address (reg
, new_rtx
);
3706 case TLS_MODEL_LOCAL_DYNAMIC
:
3708 r2
= gen_rtx_REG (Pmode
, 2);
3709 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM
);
3710 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3711 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3712 emit_move_insn (r2
, new_rtx
);
3713 s390_emit_tls_call_insn (r2
, tls_call
);
3714 insn
= get_insns ();
3717 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM_NTPOFF
);
3718 temp
= gen_reg_rtx (Pmode
);
3719 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3721 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3722 base
= gen_reg_rtx (Pmode
);
3723 s390_load_address (base
, new_rtx
);
3725 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_DTPOFF
);
3726 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3727 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3728 temp
= gen_reg_rtx (Pmode
);
3729 emit_move_insn (temp
, new_rtx
);
3731 new_rtx
= gen_rtx_PLUS (Pmode
, base
, temp
);
3734 s390_load_address (reg
, new_rtx
);
3739 case TLS_MODEL_INITIAL_EXEC
:
3742 /* Assume GOT offset < 4k. This is handled the same way
3743 in both 31- and 64-bit code. */
3745 if (reload_in_progress
|| reload_completed
)
3746 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3748 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3749 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3750 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3751 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3752 temp
= gen_reg_rtx (Pmode
);
3753 emit_move_insn (temp
, new_rtx
);
3755 else if (TARGET_CPU_ZARCH
)
3757 /* If the GOT offset might be >= 4k, we determine the position
3758 of the GOT entry via a PC-relative LARL. */
3760 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3761 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3762 temp
= gen_reg_rtx (Pmode
);
3763 emit_move_insn (temp
, new_rtx
);
3765 new_rtx
= gen_const_mem (Pmode
, temp
);
3766 temp
= gen_reg_rtx (Pmode
);
3767 emit_move_insn (temp
, new_rtx
);
3771 /* If the GOT offset might be >= 4k, we have to load it
3772 from the literal pool. */
3774 if (reload_in_progress
|| reload_completed
)
3775 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3777 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3778 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3779 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3780 temp
= gen_reg_rtx (Pmode
);
3781 emit_move_insn (temp
, new_rtx
);
3783 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3784 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3786 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3787 temp
= gen_reg_rtx (Pmode
);
3788 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3792 /* In position-dependent code, load the absolute address of
3793 the GOT entry from the literal pool. */
3795 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3796 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3797 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3798 temp
= gen_reg_rtx (Pmode
);
3799 emit_move_insn (temp
, new_rtx
);
3802 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3803 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3804 temp
= gen_reg_rtx (Pmode
);
3805 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3808 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3811 s390_load_address (reg
, new_rtx
);
3816 case TLS_MODEL_LOCAL_EXEC
:
3817 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3818 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3819 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3820 temp
= gen_reg_rtx (Pmode
);
3821 emit_move_insn (temp
, new_rtx
);
3823 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3826 s390_load_address (reg
, new_rtx
);
3835 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == UNSPEC
)
3837 switch (XINT (XEXP (addr
, 0), 1))
3839 case UNSPEC_INDNTPOFF
:
3840 gcc_assert (TARGET_CPU_ZARCH
);
3849 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
3850 && GET_CODE (XEXP (XEXP (addr
, 0), 1)) == CONST_INT
)
3852 new_rtx
= XEXP (XEXP (addr
, 0), 0);
3853 if (GET_CODE (new_rtx
) != SYMBOL_REF
)
3854 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3856 new_rtx
= legitimize_tls_address (new_rtx
, reg
);
3857 new_rtx
= plus_constant (new_rtx
, INTVAL (XEXP (XEXP (addr
, 0), 1)));
3858 new_rtx
= force_operand (new_rtx
, 0);
3862 gcc_unreachable (); /* for now ... */
3867 /* Emit insns making the address in operands[1] valid for a standard
3868 move to operands[0]. operands[1] is replaced by an address which
3869 should be used instead of the former RTX to emit the move
3873 emit_symbolic_move (rtx
*operands
)
3875 rtx temp
= !can_create_pseudo_p () ? operands
[0] : gen_reg_rtx (Pmode
);
3877 if (GET_CODE (operands
[0]) == MEM
)
3878 operands
[1] = force_reg (Pmode
, operands
[1]);
3879 else if (TLS_SYMBOLIC_CONST (operands
[1]))
3880 operands
[1] = legitimize_tls_address (operands
[1], temp
);
3882 operands
[1] = legitimize_pic_address (operands
[1], temp
);
3885 /* Try machine-dependent ways of modifying an illegitimate address X
3886 to be legitimate. If we find one, return the new, valid address.
3888 OLDX is the address as it was before break_out_memory_refs was called.
3889 In some cases it is useful to look at this to decide what needs to be done.
3891 MODE is the mode of the operand pointed to by X.
3893 When -fpic is used, special handling is needed for symbolic references.
3894 See comments by legitimize_pic_address for details. */
3897 s390_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
3898 enum machine_mode mode ATTRIBUTE_UNUSED
)
3900 rtx constant_term
= const0_rtx
;
3902 if (TLS_SYMBOLIC_CONST (x
))
3904 x
= legitimize_tls_address (x
, 0);
3906 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3909 else if (GET_CODE (x
) == PLUS
3910 && (TLS_SYMBOLIC_CONST (XEXP (x
, 0))
3911 || TLS_SYMBOLIC_CONST (XEXP (x
, 1))))
3917 if (SYMBOLIC_CONST (x
)
3918 || (GET_CODE (x
) == PLUS
3919 && (SYMBOLIC_CONST (XEXP (x
, 0))
3920 || SYMBOLIC_CONST (XEXP (x
, 1)))))
3921 x
= legitimize_pic_address (x
, 0);
3923 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3927 x
= eliminate_constant_term (x
, &constant_term
);
3929 /* Optimize loading of large displacements by splitting them
3930 into the multiple of 4K and the rest; this allows the
3931 former to be CSE'd if possible.
3933 Don't do this if the displacement is added to a register
3934 pointing into the stack frame, as the offsets will
3935 change later anyway. */
3937 if (GET_CODE (constant_term
) == CONST_INT
3938 && !TARGET_LONG_DISPLACEMENT
3939 && !DISP_IN_RANGE (INTVAL (constant_term
))
3940 && !(REG_P (x
) && REGNO_PTR_FRAME_P (REGNO (x
))))
3942 HOST_WIDE_INT lower
= INTVAL (constant_term
) & 0xfff;
3943 HOST_WIDE_INT upper
= INTVAL (constant_term
) ^ lower
;
3945 rtx temp
= gen_reg_rtx (Pmode
);
3946 rtx val
= force_operand (GEN_INT (upper
), temp
);
3948 emit_move_insn (temp
, val
);
3950 x
= gen_rtx_PLUS (Pmode
, x
, temp
);
3951 constant_term
= GEN_INT (lower
);
3954 if (GET_CODE (x
) == PLUS
)
3956 if (GET_CODE (XEXP (x
, 0)) == REG
)
3958 rtx temp
= gen_reg_rtx (Pmode
);
3959 rtx val
= force_operand (XEXP (x
, 1), temp
);
3961 emit_move_insn (temp
, val
);
3963 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0), temp
);
3966 else if (GET_CODE (XEXP (x
, 1)) == REG
)
3968 rtx temp
= gen_reg_rtx (Pmode
);
3969 rtx val
= force_operand (XEXP (x
, 0), temp
);
3971 emit_move_insn (temp
, val
);
3973 x
= gen_rtx_PLUS (Pmode
, temp
, XEXP (x
, 1));
3977 if (constant_term
!= const0_rtx
)
3978 x
= gen_rtx_PLUS (Pmode
, x
, constant_term
);
3983 /* Try a machine-dependent way of reloading an illegitimate address AD
3984 operand. If we find one, push the reload and and return the new address.
3986 MODE is the mode of the enclosing MEM. OPNUM is the operand number
3987 and TYPE is the reload type of the current reload. */
3990 legitimize_reload_address (rtx ad
, enum machine_mode mode ATTRIBUTE_UNUSED
,
3991 int opnum
, int type
)
3993 if (!optimize
|| TARGET_LONG_DISPLACEMENT
)
3996 if (GET_CODE (ad
) == PLUS
)
3998 rtx tem
= simplify_binary_operation (PLUS
, Pmode
,
3999 XEXP (ad
, 0), XEXP (ad
, 1));
4004 if (GET_CODE (ad
) == PLUS
4005 && GET_CODE (XEXP (ad
, 0)) == REG
4006 && GET_CODE (XEXP (ad
, 1)) == CONST_INT
4007 && !DISP_IN_RANGE (INTVAL (XEXP (ad
, 1))))
4009 HOST_WIDE_INT lower
= INTVAL (XEXP (ad
, 1)) & 0xfff;
4010 HOST_WIDE_INT upper
= INTVAL (XEXP (ad
, 1)) ^ lower
;
4011 rtx cst
, tem
, new_rtx
;
4013 cst
= GEN_INT (upper
);
4014 if (!legitimate_reload_constant_p (cst
))
4015 cst
= force_const_mem (Pmode
, cst
);
4017 tem
= gen_rtx_PLUS (Pmode
, XEXP (ad
, 0), cst
);
4018 new_rtx
= gen_rtx_PLUS (Pmode
, tem
, GEN_INT (lower
));
4020 push_reload (XEXP (tem
, 1), 0, &XEXP (tem
, 1), 0,
4021 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0,
4022 opnum
, (enum reload_type
) type
);
4029 /* Emit code to move LEN bytes from DST to SRC. */
4032 s390_expand_movmem (rtx dst
, rtx src
, rtx len
)
4034 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4036 if (INTVAL (len
) > 0)
4037 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (INTVAL (len
) - 1)));
4040 else if (TARGET_MVCLE
)
4042 emit_insn (gen_movmem_long (dst
, src
, convert_to_mode (Pmode
, len
, 1)));
4047 rtx dst_addr
, src_addr
, count
, blocks
, temp
;
4048 rtx loop_start_label
= gen_label_rtx ();
4049 rtx loop_end_label
= gen_label_rtx ();
4050 rtx end_label
= gen_label_rtx ();
4051 enum machine_mode mode
;
4053 mode
= GET_MODE (len
);
4054 if (mode
== VOIDmode
)
4057 dst_addr
= gen_reg_rtx (Pmode
);
4058 src_addr
= gen_reg_rtx (Pmode
);
4059 count
= gen_reg_rtx (mode
);
4060 blocks
= gen_reg_rtx (mode
);
4062 convert_move (count
, len
, 1);
4063 emit_cmp_and_jump_insns (count
, const0_rtx
,
4064 EQ
, NULL_RTX
, mode
, 1, end_label
);
4066 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4067 emit_move_insn (src_addr
, force_operand (XEXP (src
, 0), NULL_RTX
));
4068 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4069 src
= change_address (src
, VOIDmode
, src_addr
);
4071 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4074 emit_move_insn (count
, temp
);
4076 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4079 emit_move_insn (blocks
, temp
);
4081 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4082 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4084 emit_label (loop_start_label
);
4087 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 768))
4091 /* Issue a read prefetch for the +3 cache line. */
4092 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (768)),
4093 const0_rtx
, const0_rtx
);
4094 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4095 emit_insn (prefetch
);
4097 /* Issue a write prefetch for the +3 cache line. */
4098 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (768)),
4099 const1_rtx
, const0_rtx
);
4100 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4101 emit_insn (prefetch
);
4104 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (255)));
4105 s390_load_address (dst_addr
,
4106 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4107 s390_load_address (src_addr
,
4108 gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (256)));
4110 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4113 emit_move_insn (blocks
, temp
);
4115 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4116 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4118 emit_jump (loop_start_label
);
4119 emit_label (loop_end_label
);
4121 emit_insn (gen_movmem_short (dst
, src
,
4122 convert_to_mode (Pmode
, count
, 1)));
4123 emit_label (end_label
);
4127 /* Emit code to set LEN bytes at DST to VAL.
4128 Make use of clrmem if VAL is zero. */
4131 s390_expand_setmem (rtx dst
, rtx len
, rtx val
)
4133 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) == 0)
4136 gcc_assert (GET_CODE (val
) == CONST_INT
|| GET_MODE (val
) == QImode
);
4138 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) > 0 && INTVAL (len
) <= 257)
4140 if (val
== const0_rtx
&& INTVAL (len
) <= 256)
4141 emit_insn (gen_clrmem_short (dst
, GEN_INT (INTVAL (len
) - 1)));
4144 /* Initialize memory by storing the first byte. */
4145 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4147 if (INTVAL (len
) > 1)
4149 /* Initiate 1 byte overlap move.
4150 The first byte of DST is propagated through DSTP1.
4151 Prepare a movmem for: DST+1 = DST (length = LEN - 1).
4152 DST is set to size 1 so the rest of the memory location
4153 does not count as source operand. */
4154 rtx dstp1
= adjust_address (dst
, VOIDmode
, 1);
4155 set_mem_size (dst
, const1_rtx
);
4157 emit_insn (gen_movmem_short (dstp1
, dst
,
4158 GEN_INT (INTVAL (len
) - 2)));
4163 else if (TARGET_MVCLE
)
4165 val
= force_not_mem (convert_modes (Pmode
, QImode
, val
, 1));
4166 emit_insn (gen_setmem_long (dst
, convert_to_mode (Pmode
, len
, 1), val
));
4171 rtx dst_addr
, count
, blocks
, temp
, dstp1
= NULL_RTX
;
4172 rtx loop_start_label
= gen_label_rtx ();
4173 rtx loop_end_label
= gen_label_rtx ();
4174 rtx end_label
= gen_label_rtx ();
4175 enum machine_mode mode
;
4177 mode
= GET_MODE (len
);
4178 if (mode
== VOIDmode
)
4181 dst_addr
= gen_reg_rtx (Pmode
);
4182 count
= gen_reg_rtx (mode
);
4183 blocks
= gen_reg_rtx (mode
);
4185 convert_move (count
, len
, 1);
4186 emit_cmp_and_jump_insns (count
, const0_rtx
,
4187 EQ
, NULL_RTX
, mode
, 1, end_label
);
4189 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4190 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4192 if (val
== const0_rtx
)
4193 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4197 dstp1
= adjust_address (dst
, VOIDmode
, 1);
4198 set_mem_size (dst
, const1_rtx
);
4200 /* Initialize memory by storing the first byte. */
4201 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4203 /* If count is 1 we are done. */
4204 emit_cmp_and_jump_insns (count
, const1_rtx
,
4205 EQ
, NULL_RTX
, mode
, 1, end_label
);
4207 temp
= expand_binop (mode
, add_optab
, count
, GEN_INT (-2), count
, 1,
4211 emit_move_insn (count
, temp
);
4213 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4216 emit_move_insn (blocks
, temp
);
4218 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4219 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4221 emit_label (loop_start_label
);
4224 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 1024))
4226 /* Issue a write prefetch for the +4 cache line. */
4227 rtx prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
,
4229 const1_rtx
, const0_rtx
);
4230 emit_insn (prefetch
);
4231 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4234 if (val
== const0_rtx
)
4235 emit_insn (gen_clrmem_short (dst
, GEN_INT (255)));
4237 emit_insn (gen_movmem_short (dstp1
, dst
, GEN_INT (255)));
4238 s390_load_address (dst_addr
,
4239 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4241 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4244 emit_move_insn (blocks
, temp
);
4246 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4247 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4249 emit_jump (loop_start_label
);
4250 emit_label (loop_end_label
);
4252 if (val
== const0_rtx
)
4253 emit_insn (gen_clrmem_short (dst
, convert_to_mode (Pmode
, count
, 1)));
4255 emit_insn (gen_movmem_short (dstp1
, dst
, convert_to_mode (Pmode
, count
, 1)));
4256 emit_label (end_label
);
4260 /* Emit code to compare LEN bytes at OP0 with those at OP1,
4261 and return the result in TARGET. */
4264 s390_expand_cmpmem (rtx target
, rtx op0
, rtx op1
, rtx len
)
4266 rtx ccreg
= gen_rtx_REG (CCUmode
, CC_REGNUM
);
4269 /* As the result of CMPINT is inverted compared to what we need,
4270 we have to swap the operands. */
4271 tmp
= op0
; op0
= op1
; op1
= tmp
;
4273 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4275 if (INTVAL (len
) > 0)
4277 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (INTVAL (len
) - 1)));
4278 emit_insn (gen_cmpint (target
, ccreg
));
4281 emit_move_insn (target
, const0_rtx
);
4283 else if (TARGET_MVCLE
)
4285 emit_insn (gen_cmpmem_long (op0
, op1
, convert_to_mode (Pmode
, len
, 1)));
4286 emit_insn (gen_cmpint (target
, ccreg
));
4290 rtx addr0
, addr1
, count
, blocks
, temp
;
4291 rtx loop_start_label
= gen_label_rtx ();
4292 rtx loop_end_label
= gen_label_rtx ();
4293 rtx end_label
= gen_label_rtx ();
4294 enum machine_mode mode
;
4296 mode
= GET_MODE (len
);
4297 if (mode
== VOIDmode
)
4300 addr0
= gen_reg_rtx (Pmode
);
4301 addr1
= gen_reg_rtx (Pmode
);
4302 count
= gen_reg_rtx (mode
);
4303 blocks
= gen_reg_rtx (mode
);
4305 convert_move (count
, len
, 1);
4306 emit_cmp_and_jump_insns (count
, const0_rtx
,
4307 EQ
, NULL_RTX
, mode
, 1, end_label
);
4309 emit_move_insn (addr0
, force_operand (XEXP (op0
, 0), NULL_RTX
));
4310 emit_move_insn (addr1
, force_operand (XEXP (op1
, 0), NULL_RTX
));
4311 op0
= change_address (op0
, VOIDmode
, addr0
);
4312 op1
= change_address (op1
, VOIDmode
, addr1
);
4314 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4317 emit_move_insn (count
, temp
);
4319 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4322 emit_move_insn (blocks
, temp
);
4324 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4325 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4327 emit_label (loop_start_label
);
4330 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 512))
4334 /* Issue a read prefetch for the +2 cache line of operand 1. */
4335 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (512)),
4336 const0_rtx
, const0_rtx
);
4337 emit_insn (prefetch
);
4338 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4340 /* Issue a read prefetch for the +2 cache line of operand 2. */
4341 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (512)),
4342 const0_rtx
, const0_rtx
);
4343 emit_insn (prefetch
);
4344 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4347 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (255)));
4348 temp
= gen_rtx_NE (VOIDmode
, ccreg
, const0_rtx
);
4349 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
4350 gen_rtx_LABEL_REF (VOIDmode
, end_label
), pc_rtx
);
4351 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
4352 emit_jump_insn (temp
);
4354 s390_load_address (addr0
,
4355 gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (256)));
4356 s390_load_address (addr1
,
4357 gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (256)));
4359 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4362 emit_move_insn (blocks
, temp
);
4364 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4365 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4367 emit_jump (loop_start_label
);
4368 emit_label (loop_end_label
);
4370 emit_insn (gen_cmpmem_short (op0
, op1
,
4371 convert_to_mode (Pmode
, count
, 1)));
4372 emit_label (end_label
);
4374 emit_insn (gen_cmpint (target
, ccreg
));
4379 /* Expand conditional increment or decrement using alc/slb instructions.
4380 Should generate code setting DST to either SRC or SRC + INCREMENT,
4381 depending on the result of the comparison CMP_OP0 CMP_CODE CMP_OP1.
4382 Returns true if successful, false otherwise.
4384 That makes it possible to implement some if-constructs without jumps e.g.:
4385 (borrow = CC0 | CC1 and carry = CC2 | CC3)
4386 unsigned int a, b, c;
4387 if (a < b) c++; -> CCU b > a -> CC2; c += carry;
4388 if (a < b) c--; -> CCL3 a - b -> borrow; c -= borrow;
4389 if (a <= b) c++; -> CCL3 b - a -> borrow; c += carry;
4390 if (a <= b) c--; -> CCU a <= b -> borrow; c -= borrow;
4392 Checks for EQ and NE with a nonzero value need an additional xor e.g.:
4393 if (a == b) c++; -> CCL3 a ^= b; 0 - a -> borrow; c += carry;
4394 if (a == b) c--; -> CCU a ^= b; a <= 0 -> CC0 | CC1; c -= borrow;
4395 if (a != b) c++; -> CCU a ^= b; a > 0 -> CC2; c += carry;
4396 if (a != b) c--; -> CCL3 a ^= b; 0 - a -> borrow; c -= borrow; */
4399 s390_expand_addcc (enum rtx_code cmp_code
, rtx cmp_op0
, rtx cmp_op1
,
4400 rtx dst
, rtx src
, rtx increment
)
4402 enum machine_mode cmp_mode
;
4403 enum machine_mode cc_mode
;
4409 if ((GET_MODE (cmp_op0
) == SImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4410 && (GET_MODE (cmp_op1
) == SImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4412 else if ((GET_MODE (cmp_op0
) == DImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4413 && (GET_MODE (cmp_op1
) == DImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4418 /* Try ADD LOGICAL WITH CARRY. */
4419 if (increment
== const1_rtx
)
4421 /* Determine CC mode to use. */
4422 if (cmp_code
== EQ
|| cmp_code
== NE
)
4424 if (cmp_op1
!= const0_rtx
)
4426 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4427 NULL_RTX
, 0, OPTAB_WIDEN
);
4428 cmp_op1
= const0_rtx
;
4431 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4434 if (cmp_code
== LTU
|| cmp_code
== LEU
)
4439 cmp_code
= swap_condition (cmp_code
);
4456 /* Emit comparison instruction pattern. */
4457 if (!register_operand (cmp_op0
, cmp_mode
))
4458 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4460 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4461 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4462 /* We use insn_invalid_p here to add clobbers if required. */
4463 ret
= insn_invalid_p (emit_insn (insn
));
4466 /* Emit ALC instruction pattern. */
4467 op_res
= gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4468 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4471 if (src
!= const0_rtx
)
4473 if (!register_operand (src
, GET_MODE (dst
)))
4474 src
= force_reg (GET_MODE (dst
), src
);
4476 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, src
);
4477 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, const0_rtx
);
4480 p
= rtvec_alloc (2);
4482 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4484 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4485 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4490 /* Try SUBTRACT LOGICAL WITH BORROW. */
4491 if (increment
== constm1_rtx
)
4493 /* Determine CC mode to use. */
4494 if (cmp_code
== EQ
|| cmp_code
== NE
)
4496 if (cmp_op1
!= const0_rtx
)
4498 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4499 NULL_RTX
, 0, OPTAB_WIDEN
);
4500 cmp_op1
= const0_rtx
;
4503 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4506 if (cmp_code
== GTU
|| cmp_code
== GEU
)
4511 cmp_code
= swap_condition (cmp_code
);
4528 /* Emit comparison instruction pattern. */
4529 if (!register_operand (cmp_op0
, cmp_mode
))
4530 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4532 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4533 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4534 /* We use insn_invalid_p here to add clobbers if required. */
4535 ret
= insn_invalid_p (emit_insn (insn
));
4538 /* Emit SLB instruction pattern. */
4539 if (!register_operand (src
, GET_MODE (dst
)))
4540 src
= force_reg (GET_MODE (dst
), src
);
4542 op_res
= gen_rtx_MINUS (GET_MODE (dst
),
4543 gen_rtx_MINUS (GET_MODE (dst
), src
, const0_rtx
),
4544 gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4545 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4547 p
= rtvec_alloc (2);
4549 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4551 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4552 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4560 /* Expand code for the insv template. Return true if successful. */
4563 s390_expand_insv (rtx dest
, rtx op1
, rtx op2
, rtx src
)
4565 int bitsize
= INTVAL (op1
);
4566 int bitpos
= INTVAL (op2
);
4568 /* On z10 we can use the risbg instruction to implement insv. */
4570 && ((GET_MODE (dest
) == DImode
&& GET_MODE (src
) == DImode
)
4571 || (GET_MODE (dest
) == SImode
&& GET_MODE (src
) == SImode
)))
4576 op
= gen_rtx_SET (GET_MODE(src
),
4577 gen_rtx_ZERO_EXTRACT (GET_MODE (dest
), dest
, op1
, op2
),
4579 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4580 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4585 /* We need byte alignment. */
4586 if (bitsize
% BITS_PER_UNIT
)
4590 && memory_operand (dest
, VOIDmode
)
4591 && (register_operand (src
, word_mode
)
4592 || const_int_operand (src
, VOIDmode
)))
4594 /* Emit standard pattern if possible. */
4595 enum machine_mode mode
= smallest_mode_for_size (bitsize
, MODE_INT
);
4596 if (GET_MODE_BITSIZE (mode
) == bitsize
)
4597 emit_move_insn (adjust_address (dest
, mode
, 0), gen_lowpart (mode
, src
));
4599 /* (set (ze (mem)) (const_int)). */
4600 else if (const_int_operand (src
, VOIDmode
))
4602 int size
= bitsize
/ BITS_PER_UNIT
;
4603 rtx src_mem
= adjust_address (force_const_mem (word_mode
, src
), BLKmode
,
4604 GET_MODE_SIZE (word_mode
) - size
);
4606 dest
= adjust_address (dest
, BLKmode
, 0);
4607 set_mem_size (dest
, GEN_INT (size
));
4608 s390_expand_movmem (dest
, src_mem
, GEN_INT (size
));
4611 /* (set (ze (mem)) (reg)). */
4612 else if (register_operand (src
, word_mode
))
4614 if (bitsize
<= GET_MODE_BITSIZE (SImode
))
4615 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, op1
,
4619 /* Emit st,stcmh sequence. */
4620 int stcmh_width
= bitsize
- GET_MODE_BITSIZE (SImode
);
4621 int size
= stcmh_width
/ BITS_PER_UNIT
;
4623 emit_move_insn (adjust_address (dest
, SImode
, size
),
4624 gen_lowpart (SImode
, src
));
4625 set_mem_size (dest
, GEN_INT (size
));
4626 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, GEN_INT
4627 (stcmh_width
), const0_rtx
),
4628 gen_rtx_LSHIFTRT (word_mode
, src
, GEN_INT
4629 (GET_MODE_BITSIZE (SImode
))));
4638 /* (set (ze (reg)) (const_int)). */
4640 && register_operand (dest
, word_mode
)
4641 && (bitpos
% 16) == 0
4642 && (bitsize
% 16) == 0
4643 && const_int_operand (src
, VOIDmode
))
4645 HOST_WIDE_INT val
= INTVAL (src
);
4646 int regpos
= bitpos
+ bitsize
;
4648 while (regpos
> bitpos
)
4650 enum machine_mode putmode
;
4653 if (TARGET_EXTIMM
&& (regpos
% 32 == 0) && (regpos
>= bitpos
+ 32))
4658 putsize
= GET_MODE_BITSIZE (putmode
);
4660 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4663 gen_int_mode (val
, putmode
));
4666 gcc_assert (regpos
== bitpos
);
4673 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic which returns a
4674 register that holds VAL of mode MODE shifted by COUNT bits. */
4677 s390_expand_mask_and_shift (rtx val
, enum machine_mode mode
, rtx count
)
4679 val
= expand_simple_binop (SImode
, AND
, val
, GEN_INT (GET_MODE_MASK (mode
)),
4680 NULL_RTX
, 1, OPTAB_DIRECT
);
4681 return expand_simple_binop (SImode
, ASHIFT
, val
, count
,
4682 NULL_RTX
, 1, OPTAB_DIRECT
);
4685 /* Structure to hold the initial parameters for a compare_and_swap operation
4686 in HImode and QImode. */
4688 struct alignment_context
4690 rtx memsi
; /* SI aligned memory location. */
4691 rtx shift
; /* Bit offset with regard to lsb. */
4692 rtx modemask
; /* Mask of the HQImode shifted by SHIFT bits. */
4693 rtx modemaski
; /* ~modemask */
4694 bool aligned
; /* True if memory is aligned, false else. */
4697 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic to initialize
4698 structure AC for transparent simplifying, if the memory alignment is known
4699 to be at least 32bit. MEM is the memory location for the actual operation
4700 and MODE its mode. */
4703 init_alignment_context (struct alignment_context
*ac
, rtx mem
,
4704 enum machine_mode mode
)
4706 ac
->shift
= GEN_INT (GET_MODE_SIZE (SImode
) - GET_MODE_SIZE (mode
));
4707 ac
->aligned
= (MEM_ALIGN (mem
) >= GET_MODE_BITSIZE (SImode
));
4710 ac
->memsi
= adjust_address (mem
, SImode
, 0); /* Memory is aligned. */
4713 /* Alignment is unknown. */
4714 rtx byteoffset
, addr
, align
;
4716 /* Force the address into a register. */
4717 addr
= force_reg (Pmode
, XEXP (mem
, 0));
4719 /* Align it to SImode. */
4720 align
= expand_simple_binop (Pmode
, AND
, addr
,
4721 GEN_INT (-GET_MODE_SIZE (SImode
)),
4722 NULL_RTX
, 1, OPTAB_DIRECT
);
4724 ac
->memsi
= gen_rtx_MEM (SImode
, align
);
4725 MEM_VOLATILE_P (ac
->memsi
) = MEM_VOLATILE_P (mem
);
4726 set_mem_alias_set (ac
->memsi
, ALIAS_SET_MEMORY_BARRIER
);
4727 set_mem_align (ac
->memsi
, GET_MODE_BITSIZE (SImode
));
4729 /* Calculate shiftcount. */
4730 byteoffset
= expand_simple_binop (Pmode
, AND
, addr
,
4731 GEN_INT (GET_MODE_SIZE (SImode
) - 1),
4732 NULL_RTX
, 1, OPTAB_DIRECT
);
4733 /* As we already have some offset, evaluate the remaining distance. */
4734 ac
->shift
= expand_simple_binop (SImode
, MINUS
, ac
->shift
, byteoffset
,
4735 NULL_RTX
, 1, OPTAB_DIRECT
);
4738 /* Shift is the byte count, but we need the bitcount. */
4739 ac
->shift
= expand_simple_binop (SImode
, MULT
, ac
->shift
, GEN_INT (BITS_PER_UNIT
),
4740 NULL_RTX
, 1, OPTAB_DIRECT
);
4741 /* Calculate masks. */
4742 ac
->modemask
= expand_simple_binop (SImode
, ASHIFT
,
4743 GEN_INT (GET_MODE_MASK (mode
)), ac
->shift
,
4744 NULL_RTX
, 1, OPTAB_DIRECT
);
4745 ac
->modemaski
= expand_simple_unop (SImode
, NOT
, ac
->modemask
, NULL_RTX
, 1);
4748 /* Expand an atomic compare and swap operation for HImode and QImode. MEM is
4749 the memory location, CMP the old value to compare MEM with and NEW_RTX the value
4750 to set if CMP == MEM.
4751 CMP is never in memory for compare_and_swap_cc because
4752 expand_bool_compare_and_swap puts it into a register for later compare. */
4755 s390_expand_cs_hqi (enum machine_mode mode
, rtx target
, rtx mem
, rtx cmp
, rtx new_rtx
)
4757 struct alignment_context ac
;
4758 rtx cmpv
, newv
, val
, resv
, cc
;
4759 rtx res
= gen_reg_rtx (SImode
);
4760 rtx csloop
= gen_label_rtx ();
4761 rtx csend
= gen_label_rtx ();
4763 gcc_assert (register_operand (target
, VOIDmode
));
4764 gcc_assert (MEM_P (mem
));
4766 init_alignment_context (&ac
, mem
, mode
);
4768 /* Shift the values to the correct bit positions. */
4769 if (!(ac
.aligned
&& MEM_P (cmp
)))
4770 cmp
= s390_expand_mask_and_shift (cmp
, mode
, ac
.shift
);
4771 if (!(ac
.aligned
&& MEM_P (new_rtx
)))
4772 new_rtx
= s390_expand_mask_and_shift (new_rtx
, mode
, ac
.shift
);
4774 /* Load full word. Subsequent loads are performed by CS. */
4775 val
= expand_simple_binop (SImode
, AND
, ac
.memsi
, ac
.modemaski
,
4776 NULL_RTX
, 1, OPTAB_DIRECT
);
4778 /* Start CS loop. */
4779 emit_label (csloop
);
4780 /* val = "<mem>00..0<mem>"
4781 * cmp = "00..0<cmp>00..0"
4782 * new = "00..0<new>00..0"
4785 /* Patch cmp and new with val at correct position. */
4786 if (ac
.aligned
&& MEM_P (cmp
))
4788 cmpv
= force_reg (SImode
, val
);
4789 store_bit_field (cmpv
, GET_MODE_BITSIZE (mode
), 0, SImode
, cmp
);
4792 cmpv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, cmp
, val
,
4793 NULL_RTX
, 1, OPTAB_DIRECT
));
4794 if (ac
.aligned
&& MEM_P (new_rtx
))
4796 newv
= force_reg (SImode
, val
);
4797 store_bit_field (newv
, GET_MODE_BITSIZE (mode
), 0, SImode
, new_rtx
);
4800 newv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4801 NULL_RTX
, 1, OPTAB_DIRECT
));
4803 /* Jump to end if we're done (likely?). */
4804 s390_emit_jump (csend
, s390_emit_compare_and_swap (EQ
, res
, ac
.memsi
,
4807 /* Check for changes outside mode. */
4808 resv
= expand_simple_binop (SImode
, AND
, res
, ac
.modemaski
,
4809 NULL_RTX
, 1, OPTAB_DIRECT
);
4810 cc
= s390_emit_compare (NE
, resv
, val
);
4811 emit_move_insn (val
, resv
);
4812 /* Loop internal if so. */
4813 s390_emit_jump (csloop
, cc
);
4817 /* Return the correct part of the bitfield. */
4818 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
, res
, ac
.shift
,
4819 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4822 /* Expand an atomic operation CODE of mode MODE. MEM is the memory location
4823 and VAL the value to play with. If AFTER is true then store the value
4824 MEM holds after the operation, if AFTER is false then store the value MEM
4825 holds before the operation. If TARGET is zero then discard that value, else
4826 store it to TARGET. */
4829 s390_expand_atomic (enum machine_mode mode
, enum rtx_code code
,
4830 rtx target
, rtx mem
, rtx val
, bool after
)
4832 struct alignment_context ac
;
4834 rtx new_rtx
= gen_reg_rtx (SImode
);
4835 rtx orig
= gen_reg_rtx (SImode
);
4836 rtx csloop
= gen_label_rtx ();
4838 gcc_assert (!target
|| register_operand (target
, VOIDmode
));
4839 gcc_assert (MEM_P (mem
));
4841 init_alignment_context (&ac
, mem
, mode
);
4843 /* Shift val to the correct bit positions.
4844 Preserve "icm", but prevent "ex icm". */
4845 if (!(ac
.aligned
&& code
== SET
&& MEM_P (val
)))
4846 val
= s390_expand_mask_and_shift (val
, mode
, ac
.shift
);
4848 /* Further preparation insns. */
4849 if (code
== PLUS
|| code
== MINUS
)
4850 emit_move_insn (orig
, val
);
4851 else if (code
== MULT
|| code
== AND
) /* val = "11..1<val>11..1" */
4852 val
= expand_simple_binop (SImode
, XOR
, val
, ac
.modemaski
,
4853 NULL_RTX
, 1, OPTAB_DIRECT
);
4855 /* Load full word. Subsequent loads are performed by CS. */
4856 cmp
= force_reg (SImode
, ac
.memsi
);
4858 /* Start CS loop. */
4859 emit_label (csloop
);
4860 emit_move_insn (new_rtx
, cmp
);
4862 /* Patch new with val at correct position. */
4867 val
= expand_simple_binop (SImode
, code
, new_rtx
, orig
,
4868 NULL_RTX
, 1, OPTAB_DIRECT
);
4869 val
= expand_simple_binop (SImode
, AND
, val
, ac
.modemask
,
4870 NULL_RTX
, 1, OPTAB_DIRECT
);
4873 if (ac
.aligned
&& MEM_P (val
))
4874 store_bit_field (new_rtx
, GET_MODE_BITSIZE (mode
), 0, SImode
, val
);
4877 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, ac
.modemaski
,
4878 NULL_RTX
, 1, OPTAB_DIRECT
);
4879 new_rtx
= expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4880 NULL_RTX
, 1, OPTAB_DIRECT
);
4886 new_rtx
= expand_simple_binop (SImode
, code
, new_rtx
, val
,
4887 NULL_RTX
, 1, OPTAB_DIRECT
);
4889 case MULT
: /* NAND */
4890 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, val
,
4891 NULL_RTX
, 1, OPTAB_DIRECT
);
4892 new_rtx
= expand_simple_binop (SImode
, XOR
, new_rtx
, ac
.modemask
,
4893 NULL_RTX
, 1, OPTAB_DIRECT
);
4899 s390_emit_jump (csloop
, s390_emit_compare_and_swap (NE
, cmp
,
4900 ac
.memsi
, cmp
, new_rtx
));
4902 /* Return the correct part of the bitfield. */
4904 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
,
4905 after
? new_rtx
: cmp
, ac
.shift
,
4906 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4909 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
4910 We need to emit DTP-relative relocations. */
4912 static void s390_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
4915 s390_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
4920 fputs ("\t.long\t", file
);
4923 fputs ("\t.quad\t", file
);
4928 output_addr_const (file
, x
);
4929 fputs ("@DTPOFF", file
);
4932 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
4933 /* Implement TARGET_MANGLE_TYPE. */
4936 s390_mangle_type (const_tree type
)
4938 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
4939 && TARGET_LONG_DOUBLE_128
)
4942 /* For all other types, use normal C++ mangling. */
4947 /* In the name of slightly smaller debug output, and to cater to
4948 general assembler lossage, recognize various UNSPEC sequences
4949 and turn them back into a direct symbol reference. */
4952 s390_delegitimize_address (rtx orig_x
)
4956 orig_x
= delegitimize_mem_from_attrs (orig_x
);
4958 if (GET_CODE (x
) != MEM
)
4962 if (GET_CODE (x
) == PLUS
4963 && GET_CODE (XEXP (x
, 1)) == CONST
4964 && GET_CODE (XEXP (x
, 0)) == REG
4965 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
4967 y
= XEXP (XEXP (x
, 1), 0);
4968 if (GET_CODE (y
) == UNSPEC
4969 && XINT (y
, 1) == UNSPEC_GOT
)
4970 return XVECEXP (y
, 0, 0);
4974 if (GET_CODE (x
) == CONST
)
4977 if (GET_CODE (y
) == UNSPEC
4978 && XINT (y
, 1) == UNSPEC_GOTENT
)
4979 return XVECEXP (y
, 0, 0);
4986 /* Output operand OP to stdio stream FILE.
4987 OP is an address (register + offset) which is not used to address data;
4988 instead the rightmost bits are interpreted as the value. */
4991 print_shift_count_operand (FILE *file
, rtx op
)
4993 HOST_WIDE_INT offset
;
4996 /* Extract base register and offset. */
4997 if (!s390_decompose_shift_count (op
, &base
, &offset
))
5003 gcc_assert (GET_CODE (base
) == REG
);
5004 gcc_assert (REGNO (base
) < FIRST_PSEUDO_REGISTER
);
5005 gcc_assert (REGNO_REG_CLASS (REGNO (base
)) == ADDR_REGS
);
5008 /* Offsets are constricted to twelve bits. */
5009 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, offset
& ((1 << 12) - 1));
5011 fprintf (file
, "(%s)", reg_names
[REGNO (base
)]);
5014 /* See 'get_some_local_dynamic_name'. */
5017 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
5021 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
5023 x
= get_pool_constant (x
);
5024 return for_each_rtx (&x
, get_some_local_dynamic_name_1
, 0);
5027 if (GET_CODE (x
) == SYMBOL_REF
5028 && tls_symbolic_operand (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
5030 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
5037 /* Locate some local-dynamic symbol still in use by this function
5038 so that we can print its name in local-dynamic base patterns. */
5041 get_some_local_dynamic_name (void)
5045 if (cfun
->machine
->some_ld_name
)
5046 return cfun
->machine
->some_ld_name
;
5048 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5050 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
5051 return cfun
->machine
->some_ld_name
;
5056 /* Output machine-dependent UNSPECs occurring in address constant X
5057 in assembler syntax to stdio stream FILE. Returns true if the
5058 constant X could be recognized, false otherwise. */
5061 s390_output_addr_const_extra (FILE *file
, rtx x
)
5063 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 1)
5064 switch (XINT (x
, 1))
5067 output_addr_const (file
, XVECEXP (x
, 0, 0));
5068 fprintf (file
, "@GOTENT");
5071 output_addr_const (file
, XVECEXP (x
, 0, 0));
5072 fprintf (file
, "@GOT");
5075 output_addr_const (file
, XVECEXP (x
, 0, 0));
5076 fprintf (file
, "@GOTOFF");
5079 output_addr_const (file
, XVECEXP (x
, 0, 0));
5080 fprintf (file
, "@PLT");
5083 output_addr_const (file
, XVECEXP (x
, 0, 0));
5084 fprintf (file
, "@PLTOFF");
5087 output_addr_const (file
, XVECEXP (x
, 0, 0));
5088 fprintf (file
, "@TLSGD");
5091 assemble_name (file
, get_some_local_dynamic_name ());
5092 fprintf (file
, "@TLSLDM");
5095 output_addr_const (file
, XVECEXP (x
, 0, 0));
5096 fprintf (file
, "@DTPOFF");
5099 output_addr_const (file
, XVECEXP (x
, 0, 0));
5100 fprintf (file
, "@NTPOFF");
5102 case UNSPEC_GOTNTPOFF
:
5103 output_addr_const (file
, XVECEXP (x
, 0, 0));
5104 fprintf (file
, "@GOTNTPOFF");
5106 case UNSPEC_INDNTPOFF
:
5107 output_addr_const (file
, XVECEXP (x
, 0, 0));
5108 fprintf (file
, "@INDNTPOFF");
5112 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 2)
5113 switch (XINT (x
, 1))
5115 case UNSPEC_POOL_OFFSET
:
5116 x
= gen_rtx_MINUS (GET_MODE (x
), XVECEXP (x
, 0, 0), XVECEXP (x
, 0, 1));
5117 output_addr_const (file
, x
);
5123 /* Output address operand ADDR in assembler syntax to
5124 stdio stream FILE. */
5127 print_operand_address (FILE *file
, rtx addr
)
5129 struct s390_address ad
;
5131 if (s390_symref_operand_p (addr
, NULL
, NULL
))
5135 error ("symbolic memory references are only supported on z10 or later");
5138 output_addr_const (file
, addr
);
5142 if (!s390_decompose_address (addr
, &ad
)
5143 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5144 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
5145 output_operand_lossage ("cannot decompose address");
5148 output_addr_const (file
, ad
.disp
);
5150 fprintf (file
, "0");
5152 if (ad
.base
&& ad
.indx
)
5153 fprintf (file
, "(%s,%s)", reg_names
[REGNO (ad
.indx
)],
5154 reg_names
[REGNO (ad
.base
)]);
5156 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5159 /* Output operand X in assembler syntax to stdio stream FILE.
5160 CODE specified the format flag. The following format flags
5163 'C': print opcode suffix for branch condition.
5164 'D': print opcode suffix for inverse branch condition.
5165 'E': print opcode suffix for branch on index instruction.
5166 'J': print tls_load/tls_gdcall/tls_ldcall suffix
5167 'G': print the size of the operand in bytes.
5168 'O': print only the displacement of a memory reference.
5169 'R': print only the base register of a memory reference.
5170 'S': print S-type memory reference (base+displacement).
5171 'N': print the second word of a DImode operand.
5172 'M': print the second word of a TImode operand.
5173 'Y': print shift count operand.
5175 'b': print integer X as if it's an unsigned byte.
5176 'c': print integer X as if it's an signed byte.
5177 'x': print integer X as if it's an unsigned halfword.
5178 'h': print integer X as if it's a signed halfword.
5179 'i': print the first nonzero HImode part of X.
5180 'j': print the first HImode part unequal to -1 of X.
5181 'k': print the first nonzero SImode part of X.
5182 'm': print the first SImode part unequal to -1 of X.
5183 'o': print integer X as if it's an unsigned 32bit word. */
5186 print_operand (FILE *file
, rtx x
, int code
)
5191 fprintf (file
, s390_branch_condition_mnemonic (x
, FALSE
));
5195 fprintf (file
, s390_branch_condition_mnemonic (x
, TRUE
));
5199 if (GET_CODE (x
) == LE
)
5200 fprintf (file
, "l");
5201 else if (GET_CODE (x
) == GT
)
5202 fprintf (file
, "h");
5204 error ("invalid comparison operator for 'E' output modifier");
5208 if (GET_CODE (x
) == SYMBOL_REF
)
5210 fprintf (file
, "%s", ":tls_load:");
5211 output_addr_const (file
, x
);
5213 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD
)
5215 fprintf (file
, "%s", ":tls_gdcall:");
5216 output_addr_const (file
, XVECEXP (x
, 0, 0));
5218 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM
)
5220 fprintf (file
, "%s", ":tls_ldcall:");
5221 assemble_name (file
, get_some_local_dynamic_name ());
5224 error ("invalid reference for 'J' output modifier");
5228 fprintf (file
, "%u", GET_MODE_SIZE (GET_MODE (x
)));
5233 struct s390_address ad
;
5238 error ("memory reference expected for 'O' output modifier");
5242 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5245 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5248 error ("invalid address for 'O' output modifier");
5253 output_addr_const (file
, ad
.disp
);
5255 fprintf (file
, "0");
5261 struct s390_address ad
;
5266 error ("memory reference expected for 'R' output modifier");
5270 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5273 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5276 error ("invalid address for 'R' output modifier");
5281 fprintf (file
, "%s", reg_names
[REGNO (ad
.base
)]);
5283 fprintf (file
, "0");
5289 struct s390_address ad
;
5294 error ("memory reference expected for 'S' output modifier");
5297 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5300 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5303 error ("invalid address for 'S' output modifier");
5308 output_addr_const (file
, ad
.disp
);
5310 fprintf (file
, "0");
5313 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5318 if (GET_CODE (x
) == REG
)
5319 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5320 else if (GET_CODE (x
) == MEM
)
5321 x
= change_address (x
, VOIDmode
, plus_constant (XEXP (x
, 0), 4));
5323 error ("register or memory expression expected for 'N' output modifier");
5327 if (GET_CODE (x
) == REG
)
5328 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5329 else if (GET_CODE (x
) == MEM
)
5330 x
= change_address (x
, VOIDmode
, plus_constant (XEXP (x
, 0), 8));
5332 error ("register or memory expression expected for 'M' output modifier");
5336 print_shift_count_operand (file
, x
);
5340 switch (GET_CODE (x
))
5343 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5347 output_address (XEXP (x
, 0));
5354 output_addr_const (file
, x
);
5359 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xff);
5360 else if (code
== 'c')
5361 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xff) ^ 0x80) - 0x80);
5362 else if (code
== 'x')
5363 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffff);
5364 else if (code
== 'h')
5365 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xffff) ^ 0x8000) - 0x8000);
5366 else if (code
== 'i')
5367 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5368 s390_extract_part (x
, HImode
, 0));
5369 else if (code
== 'j')
5370 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5371 s390_extract_part (x
, HImode
, -1));
5372 else if (code
== 'k')
5373 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5374 s390_extract_part (x
, SImode
, 0));
5375 else if (code
== 'm')
5376 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5377 s390_extract_part (x
, SImode
, -1));
5378 else if (code
== 'o')
5379 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffffffff);
5381 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5385 gcc_assert (GET_MODE (x
) == VOIDmode
);
5387 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xff);
5388 else if (code
== 'x')
5389 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xffff);
5390 else if (code
== 'h')
5391 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((CONST_DOUBLE_LOW (x
) & 0xffff) ^ 0x8000) - 0x8000);
5395 error ("invalid constant - try using an output modifier");
5397 error ("invalid constant for output modifier '%c'", code
);
5403 error ("invalid expression - try using an output modifier");
5405 error ("invalid expression for output modifier '%c'", code
);
5410 /* Target hook for assembling integer objects. We need to define it
5411 here to work a round a bug in some versions of GAS, which couldn't
5412 handle values smaller than INT_MIN when printed in decimal. */
5415 s390_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
5417 if (size
== 8 && aligned_p
5418 && GET_CODE (x
) == CONST_INT
&& INTVAL (x
) < INT_MIN
)
5420 fprintf (asm_out_file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n",
5424 return default_assemble_integer (x
, size
, aligned_p
);
5427 /* Returns true if register REGNO is used for forming
5428 a memory address in expression X. */
5431 reg_used_in_mem_p (int regno
, rtx x
)
5433 enum rtx_code code
= GET_CODE (x
);
5439 if (refers_to_regno_p (regno
, regno
+1,
5443 else if (code
== SET
5444 && GET_CODE (SET_DEST (x
)) == PC
)
5446 if (refers_to_regno_p (regno
, regno
+1,
5451 fmt
= GET_RTX_FORMAT (code
);
5452 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
5455 && reg_used_in_mem_p (regno
, XEXP (x
, i
)))
5458 else if (fmt
[i
] == 'E')
5459 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5460 if (reg_used_in_mem_p (regno
, XVECEXP (x
, i
, j
)))
5466 /* Returns true if expression DEP_RTX sets an address register
5467 used by instruction INSN to address memory. */
5470 addr_generation_dependency_p (rtx dep_rtx
, rtx insn
)
5474 if (GET_CODE (dep_rtx
) == INSN
)
5475 dep_rtx
= PATTERN (dep_rtx
);
5477 if (GET_CODE (dep_rtx
) == SET
)
5479 target
= SET_DEST (dep_rtx
);
5480 if (GET_CODE (target
) == STRICT_LOW_PART
)
5481 target
= XEXP (target
, 0);
5482 while (GET_CODE (target
) == SUBREG
)
5483 target
= SUBREG_REG (target
);
5485 if (GET_CODE (target
) == REG
)
5487 int regno
= REGNO (target
);
5489 if (s390_safe_attr_type (insn
) == TYPE_LA
)
5491 pat
= PATTERN (insn
);
5492 if (GET_CODE (pat
) == PARALLEL
)
5494 gcc_assert (XVECLEN (pat
, 0) == 2);
5495 pat
= XVECEXP (pat
, 0, 0);
5497 gcc_assert (GET_CODE (pat
) == SET
);
5498 return refers_to_regno_p (regno
, regno
+1, SET_SRC (pat
), 0);
5500 else if (get_attr_atype (insn
) == ATYPE_AGEN
)
5501 return reg_used_in_mem_p (regno
, PATTERN (insn
));
5507 /* Return 1, if dep_insn sets register used in insn in the agen unit. */
5510 s390_agen_dep_p (rtx dep_insn
, rtx insn
)
5512 rtx dep_rtx
= PATTERN (dep_insn
);
5515 if (GET_CODE (dep_rtx
) == SET
5516 && addr_generation_dependency_p (dep_rtx
, insn
))
5518 else if (GET_CODE (dep_rtx
) == PARALLEL
)
5520 for (i
= 0; i
< XVECLEN (dep_rtx
, 0); i
++)
5522 if (addr_generation_dependency_p (XVECEXP (dep_rtx
, 0, i
), insn
))
5530 /* A C statement (sans semicolon) to update the integer scheduling priority
5531 INSN_PRIORITY (INSN). Increase the priority to execute the INSN earlier,
5532 reduce the priority to execute INSN later. Do not define this macro if
5533 you do not need to adjust the scheduling priorities of insns.
5535 A STD instruction should be scheduled earlier,
5536 in order to use the bypass. */
5538 s390_adjust_priority (rtx insn ATTRIBUTE_UNUSED
, int priority
)
5540 if (! INSN_P (insn
))
5543 if (s390_tune
!= PROCESSOR_2084_Z990
5544 && s390_tune
!= PROCESSOR_2094_Z9_109
5545 && s390_tune
!= PROCESSOR_2097_Z10
5546 && s390_tune
!= PROCESSOR_2817_Z196
)
5549 switch (s390_safe_attr_type (insn
))
5553 priority
= priority
<< 3;
5557 priority
= priority
<< 1;
5566 /* The number of instructions that can be issued per cycle. */
5569 s390_issue_rate (void)
5573 case PROCESSOR_2084_Z990
:
5574 case PROCESSOR_2094_Z9_109
:
5575 case PROCESSOR_2817_Z196
:
5577 case PROCESSOR_2097_Z10
:
5585 s390_first_cycle_multipass_dfa_lookahead (void)
5590 /* Annotate every literal pool reference in X by an UNSPEC_LTREF expression.
5591 Fix up MEMs as required. */
5594 annotate_constant_pool_refs (rtx
*x
)
5599 gcc_assert (GET_CODE (*x
) != SYMBOL_REF
5600 || !CONSTANT_POOL_ADDRESS_P (*x
));
5602 /* Literal pool references can only occur inside a MEM ... */
5603 if (GET_CODE (*x
) == MEM
)
5605 rtx memref
= XEXP (*x
, 0);
5607 if (GET_CODE (memref
) == SYMBOL_REF
5608 && CONSTANT_POOL_ADDRESS_P (memref
))
5610 rtx base
= cfun
->machine
->base_reg
;
5611 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, memref
, base
),
5614 *x
= replace_equiv_address (*x
, addr
);
5618 if (GET_CODE (memref
) == CONST
5619 && GET_CODE (XEXP (memref
, 0)) == PLUS
5620 && GET_CODE (XEXP (XEXP (memref
, 0), 1)) == CONST_INT
5621 && GET_CODE (XEXP (XEXP (memref
, 0), 0)) == SYMBOL_REF
5622 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (memref
, 0), 0)))
5624 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (memref
, 0), 1));
5625 rtx sym
= XEXP (XEXP (memref
, 0), 0);
5626 rtx base
= cfun
->machine
->base_reg
;
5627 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5630 *x
= replace_equiv_address (*x
, plus_constant (addr
, off
));
5635 /* ... or a load-address type pattern. */
5636 if (GET_CODE (*x
) == SET
)
5638 rtx addrref
= SET_SRC (*x
);
5640 if (GET_CODE (addrref
) == SYMBOL_REF
5641 && CONSTANT_POOL_ADDRESS_P (addrref
))
5643 rtx base
= cfun
->machine
->base_reg
;
5644 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, addrref
, base
),
5647 SET_SRC (*x
) = addr
;
5651 if (GET_CODE (addrref
) == CONST
5652 && GET_CODE (XEXP (addrref
, 0)) == PLUS
5653 && GET_CODE (XEXP (XEXP (addrref
, 0), 1)) == CONST_INT
5654 && GET_CODE (XEXP (XEXP (addrref
, 0), 0)) == SYMBOL_REF
5655 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (addrref
, 0), 0)))
5657 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (addrref
, 0), 1));
5658 rtx sym
= XEXP (XEXP (addrref
, 0), 0);
5659 rtx base
= cfun
->machine
->base_reg
;
5660 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5663 SET_SRC (*x
) = plus_constant (addr
, off
);
5668 /* Annotate LTREL_BASE as well. */
5669 if (GET_CODE (*x
) == UNSPEC
5670 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5672 rtx base
= cfun
->machine
->base_reg
;
5673 *x
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XVECEXP (*x
, 0, 0), base
),
5678 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5679 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5683 annotate_constant_pool_refs (&XEXP (*x
, i
));
5685 else if (fmt
[i
] == 'E')
5687 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5688 annotate_constant_pool_refs (&XVECEXP (*x
, i
, j
));
5693 /* Split all branches that exceed the maximum distance.
5694 Returns true if this created a new literal pool entry. */
5697 s390_split_branches (void)
5699 rtx temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
5700 int new_literal
= 0, ret
;
5701 rtx insn
, pat
, tmp
, target
;
5704 /* We need correct insn addresses. */
5706 shorten_branches (get_insns ());
5708 /* Find all branches that exceed 64KB, and split them. */
5710 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5712 if (GET_CODE (insn
) != JUMP_INSN
)
5715 pat
= PATTERN (insn
);
5716 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
5717 pat
= XVECEXP (pat
, 0, 0);
5718 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
5721 if (GET_CODE (SET_SRC (pat
)) == LABEL_REF
)
5723 label
= &SET_SRC (pat
);
5725 else if (GET_CODE (SET_SRC (pat
)) == IF_THEN_ELSE
)
5727 if (GET_CODE (XEXP (SET_SRC (pat
), 1)) == LABEL_REF
)
5728 label
= &XEXP (SET_SRC (pat
), 1);
5729 else if (GET_CODE (XEXP (SET_SRC (pat
), 2)) == LABEL_REF
)
5730 label
= &XEXP (SET_SRC (pat
), 2);
5737 if (get_attr_length (insn
) <= 4)
5740 /* We are going to use the return register as scratch register,
5741 make sure it will be saved/restored by the prologue/epilogue. */
5742 cfun_frame_layout
.save_return_addr_p
= 1;
5747 tmp
= force_const_mem (Pmode
, *label
);
5748 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, tmp
), insn
);
5749 INSN_ADDRESSES_NEW (tmp
, -1);
5750 annotate_constant_pool_refs (&PATTERN (tmp
));
5757 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, *label
),
5758 UNSPEC_LTREL_OFFSET
);
5759 target
= gen_rtx_CONST (Pmode
, target
);
5760 target
= force_const_mem (Pmode
, target
);
5761 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, target
), insn
);
5762 INSN_ADDRESSES_NEW (tmp
, -1);
5763 annotate_constant_pool_refs (&PATTERN (tmp
));
5765 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XEXP (target
, 0),
5766 cfun
->machine
->base_reg
),
5768 target
= gen_rtx_PLUS (Pmode
, temp_reg
, target
);
5771 ret
= validate_change (insn
, label
, target
, 0);
5779 /* Find an annotated literal pool symbol referenced in RTX X,
5780 and store it at REF. Will abort if X contains references to
5781 more than one such pool symbol; multiple references to the same
5782 symbol are allowed, however.
5784 The rtx pointed to by REF must be initialized to NULL_RTX
5785 by the caller before calling this routine. */
5788 find_constant_pool_ref (rtx x
, rtx
*ref
)
5793 /* Ignore LTREL_BASE references. */
5794 if (GET_CODE (x
) == UNSPEC
5795 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5797 /* Likewise POOL_ENTRY insns. */
5798 if (GET_CODE (x
) == UNSPEC_VOLATILE
5799 && XINT (x
, 1) == UNSPECV_POOL_ENTRY
)
5802 gcc_assert (GET_CODE (x
) != SYMBOL_REF
5803 || !CONSTANT_POOL_ADDRESS_P (x
));
5805 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_LTREF
)
5807 rtx sym
= XVECEXP (x
, 0, 0);
5808 gcc_assert (GET_CODE (sym
) == SYMBOL_REF
5809 && CONSTANT_POOL_ADDRESS_P (sym
));
5811 if (*ref
== NULL_RTX
)
5814 gcc_assert (*ref
== sym
);
5819 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5820 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5824 find_constant_pool_ref (XEXP (x
, i
), ref
);
5826 else if (fmt
[i
] == 'E')
5828 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5829 find_constant_pool_ref (XVECEXP (x
, i
, j
), ref
);
5834 /* Replace every reference to the annotated literal pool
5835 symbol REF in X by its base plus OFFSET. */
5838 replace_constant_pool_ref (rtx
*x
, rtx ref
, rtx offset
)
5843 gcc_assert (*x
!= ref
);
5845 if (GET_CODE (*x
) == UNSPEC
5846 && XINT (*x
, 1) == UNSPEC_LTREF
5847 && XVECEXP (*x
, 0, 0) == ref
)
5849 *x
= gen_rtx_PLUS (Pmode
, XVECEXP (*x
, 0, 1), offset
);
5853 if (GET_CODE (*x
) == PLUS
5854 && GET_CODE (XEXP (*x
, 1)) == CONST_INT
5855 && GET_CODE (XEXP (*x
, 0)) == UNSPEC
5856 && XINT (XEXP (*x
, 0), 1) == UNSPEC_LTREF
5857 && XVECEXP (XEXP (*x
, 0), 0, 0) == ref
)
5859 rtx addr
= gen_rtx_PLUS (Pmode
, XVECEXP (XEXP (*x
, 0), 0, 1), offset
);
5860 *x
= plus_constant (addr
, INTVAL (XEXP (*x
, 1)));
5864 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5865 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5869 replace_constant_pool_ref (&XEXP (*x
, i
), ref
, offset
);
5871 else if (fmt
[i
] == 'E')
5873 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5874 replace_constant_pool_ref (&XVECEXP (*x
, i
, j
), ref
, offset
);
5879 /* Check whether X contains an UNSPEC_LTREL_BASE.
5880 Return its constant pool symbol if found, NULL_RTX otherwise. */
5883 find_ltrel_base (rtx x
)
5888 if (GET_CODE (x
) == UNSPEC
5889 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5890 return XVECEXP (x
, 0, 0);
5892 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5893 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5897 rtx fnd
= find_ltrel_base (XEXP (x
, i
));
5901 else if (fmt
[i
] == 'E')
5903 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5905 rtx fnd
= find_ltrel_base (XVECEXP (x
, i
, j
));
5915 /* Replace any occurrence of UNSPEC_LTREL_BASE in X with its base. */
5918 replace_ltrel_base (rtx
*x
)
5923 if (GET_CODE (*x
) == UNSPEC
5924 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5926 *x
= XVECEXP (*x
, 0, 1);
5930 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5931 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5935 replace_ltrel_base (&XEXP (*x
, i
));
5937 else if (fmt
[i
] == 'E')
5939 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5940 replace_ltrel_base (&XVECEXP (*x
, i
, j
));
5946 /* We keep a list of constants which we have to add to internal
5947 constant tables in the middle of large functions. */
5949 #define NR_C_MODES 11
5950 enum machine_mode constant_modes
[NR_C_MODES
] =
5952 TFmode
, TImode
, TDmode
,
5953 DFmode
, DImode
, DDmode
,
5954 SFmode
, SImode
, SDmode
,
5961 struct constant
*next
;
5966 struct constant_pool
5968 struct constant_pool
*next
;
5972 rtx emit_pool_after
;
5974 struct constant
*constants
[NR_C_MODES
];
5975 struct constant
*execute
;
5980 /* Allocate new constant_pool structure. */
5982 static struct constant_pool
*
5983 s390_alloc_pool (void)
5985 struct constant_pool
*pool
;
5988 pool
= (struct constant_pool
*) xmalloc (sizeof *pool
);
5990 for (i
= 0; i
< NR_C_MODES
; i
++)
5991 pool
->constants
[i
] = NULL
;
5993 pool
->execute
= NULL
;
5994 pool
->label
= gen_label_rtx ();
5995 pool
->first_insn
= NULL_RTX
;
5996 pool
->pool_insn
= NULL_RTX
;
5997 pool
->insns
= BITMAP_ALLOC (NULL
);
5999 pool
->emit_pool_after
= NULL_RTX
;
6004 /* Create new constant pool covering instructions starting at INSN
6005 and chain it to the end of POOL_LIST. */
6007 static struct constant_pool
*
6008 s390_start_pool (struct constant_pool
**pool_list
, rtx insn
)
6010 struct constant_pool
*pool
, **prev
;
6012 pool
= s390_alloc_pool ();
6013 pool
->first_insn
= insn
;
6015 for (prev
= pool_list
; *prev
; prev
= &(*prev
)->next
)
6022 /* End range of instructions covered by POOL at INSN and emit
6023 placeholder insn representing the pool. */
6026 s390_end_pool (struct constant_pool
*pool
, rtx insn
)
6028 rtx pool_size
= GEN_INT (pool
->size
+ 8 /* alignment slop */);
6031 insn
= get_last_insn ();
6033 pool
->pool_insn
= emit_insn_after (gen_pool (pool_size
), insn
);
6034 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6037 /* Add INSN to the list of insns covered by POOL. */
6040 s390_add_pool_insn (struct constant_pool
*pool
, rtx insn
)
6042 bitmap_set_bit (pool
->insns
, INSN_UID (insn
));
6045 /* Return pool out of POOL_LIST that covers INSN. */
6047 static struct constant_pool
*
6048 s390_find_pool (struct constant_pool
*pool_list
, rtx insn
)
6050 struct constant_pool
*pool
;
6052 for (pool
= pool_list
; pool
; pool
= pool
->next
)
6053 if (bitmap_bit_p (pool
->insns
, INSN_UID (insn
)))
6059 /* Add constant VAL of mode MODE to the constant pool POOL. */
6062 s390_add_constant (struct constant_pool
*pool
, rtx val
, enum machine_mode mode
)
6067 for (i
= 0; i
< NR_C_MODES
; i
++)
6068 if (constant_modes
[i
] == mode
)
6070 gcc_assert (i
!= NR_C_MODES
);
6072 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6073 if (rtx_equal_p (val
, c
->value
))
6078 c
= (struct constant
*) xmalloc (sizeof *c
);
6080 c
->label
= gen_label_rtx ();
6081 c
->next
= pool
->constants
[i
];
6082 pool
->constants
[i
] = c
;
6083 pool
->size
+= GET_MODE_SIZE (mode
);
6087 /* Return an rtx that represents the offset of X from the start of
6091 s390_pool_offset (struct constant_pool
*pool
, rtx x
)
6095 label
= gen_rtx_LABEL_REF (GET_MODE (x
), pool
->label
);
6096 x
= gen_rtx_UNSPEC (GET_MODE (x
), gen_rtvec (2, x
, label
),
6097 UNSPEC_POOL_OFFSET
);
6098 return gen_rtx_CONST (GET_MODE (x
), x
);
6101 /* Find constant VAL of mode MODE in the constant pool POOL.
6102 Return an RTX describing the distance from the start of
6103 the pool to the location of the new constant. */
6106 s390_find_constant (struct constant_pool
*pool
, rtx val
,
6107 enum machine_mode mode
)
6112 for (i
= 0; i
< NR_C_MODES
; i
++)
6113 if (constant_modes
[i
] == mode
)
6115 gcc_assert (i
!= NR_C_MODES
);
6117 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6118 if (rtx_equal_p (val
, c
->value
))
6123 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6126 /* Check whether INSN is an execute. Return the label_ref to its
6127 execute target template if so, NULL_RTX otherwise. */
6130 s390_execute_label (rtx insn
)
6132 if (GET_CODE (insn
) == INSN
6133 && GET_CODE (PATTERN (insn
)) == PARALLEL
6134 && GET_CODE (XVECEXP (PATTERN (insn
), 0, 0)) == UNSPEC
6135 && XINT (XVECEXP (PATTERN (insn
), 0, 0), 1) == UNSPEC_EXECUTE
)
6136 return XVECEXP (XVECEXP (PATTERN (insn
), 0, 0), 0, 2);
6141 /* Add execute target for INSN to the constant pool POOL. */
6144 s390_add_execute (struct constant_pool
*pool
, rtx insn
)
6148 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6149 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6154 c
= (struct constant
*) xmalloc (sizeof *c
);
6156 c
->label
= gen_label_rtx ();
6157 c
->next
= pool
->execute
;
6163 /* Find execute target for INSN in the constant pool POOL.
6164 Return an RTX describing the distance from the start of
6165 the pool to the location of the execute target. */
6168 s390_find_execute (struct constant_pool
*pool
, rtx insn
)
6172 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6173 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6178 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6181 /* For an execute INSN, extract the execute target template. */
6184 s390_execute_target (rtx insn
)
6186 rtx pattern
= PATTERN (insn
);
6187 gcc_assert (s390_execute_label (insn
));
6189 if (XVECLEN (pattern
, 0) == 2)
6191 pattern
= copy_rtx (XVECEXP (pattern
, 0, 1));
6195 rtvec vec
= rtvec_alloc (XVECLEN (pattern
, 0) - 1);
6198 for (i
= 0; i
< XVECLEN (pattern
, 0) - 1; i
++)
6199 RTVEC_ELT (vec
, i
) = copy_rtx (XVECEXP (pattern
, 0, i
+ 1));
6201 pattern
= gen_rtx_PARALLEL (VOIDmode
, vec
);
6207 /* Indicate that INSN cannot be duplicated. This is the case for
6208 execute insns that carry a unique label. */
6211 s390_cannot_copy_insn_p (rtx insn
)
6213 rtx label
= s390_execute_label (insn
);
6214 return label
&& label
!= const0_rtx
;
6217 /* Dump out the constants in POOL. If REMOTE_LABEL is true,
6218 do not emit the pool base label. */
6221 s390_dump_pool (struct constant_pool
*pool
, bool remote_label
)
6224 rtx insn
= pool
->pool_insn
;
6227 /* Switch to rodata section. */
6228 if (TARGET_CPU_ZARCH
)
6230 insn
= emit_insn_after (gen_pool_section_start (), insn
);
6231 INSN_ADDRESSES_NEW (insn
, -1);
6234 /* Ensure minimum pool alignment. */
6235 if (TARGET_CPU_ZARCH
)
6236 insn
= emit_insn_after (gen_pool_align (GEN_INT (8)), insn
);
6238 insn
= emit_insn_after (gen_pool_align (GEN_INT (4)), insn
);
6239 INSN_ADDRESSES_NEW (insn
, -1);
6241 /* Emit pool base label. */
6244 insn
= emit_label_after (pool
->label
, insn
);
6245 INSN_ADDRESSES_NEW (insn
, -1);
6248 /* Dump constants in descending alignment requirement order,
6249 ensuring proper alignment for every constant. */
6250 for (i
= 0; i
< NR_C_MODES
; i
++)
6251 for (c
= pool
->constants
[i
]; c
; c
= c
->next
)
6253 /* Convert UNSPEC_LTREL_OFFSET unspecs to pool-relative references. */
6254 rtx value
= copy_rtx (c
->value
);
6255 if (GET_CODE (value
) == CONST
6256 && GET_CODE (XEXP (value
, 0)) == UNSPEC
6257 && XINT (XEXP (value
, 0), 1) == UNSPEC_LTREL_OFFSET
6258 && XVECLEN (XEXP (value
, 0), 0) == 1)
6259 value
= s390_pool_offset (pool
, XVECEXP (XEXP (value
, 0), 0, 0));
6261 insn
= emit_label_after (c
->label
, insn
);
6262 INSN_ADDRESSES_NEW (insn
, -1);
6264 value
= gen_rtx_UNSPEC_VOLATILE (constant_modes
[i
],
6265 gen_rtvec (1, value
),
6266 UNSPECV_POOL_ENTRY
);
6267 insn
= emit_insn_after (value
, insn
);
6268 INSN_ADDRESSES_NEW (insn
, -1);
6271 /* Ensure minimum alignment for instructions. */
6272 insn
= emit_insn_after (gen_pool_align (GEN_INT (2)), insn
);
6273 INSN_ADDRESSES_NEW (insn
, -1);
6275 /* Output in-pool execute template insns. */
6276 for (c
= pool
->execute
; c
; c
= c
->next
)
6278 insn
= emit_label_after (c
->label
, insn
);
6279 INSN_ADDRESSES_NEW (insn
, -1);
6281 insn
= emit_insn_after (s390_execute_target (c
->value
), insn
);
6282 INSN_ADDRESSES_NEW (insn
, -1);
6285 /* Switch back to previous section. */
6286 if (TARGET_CPU_ZARCH
)
6288 insn
= emit_insn_after (gen_pool_section_end (), insn
);
6289 INSN_ADDRESSES_NEW (insn
, -1);
6292 insn
= emit_barrier_after (insn
);
6293 INSN_ADDRESSES_NEW (insn
, -1);
6295 /* Remove placeholder insn. */
6296 remove_insn (pool
->pool_insn
);
6299 /* Free all memory used by POOL. */
6302 s390_free_pool (struct constant_pool
*pool
)
6304 struct constant
*c
, *next
;
6307 for (i
= 0; i
< NR_C_MODES
; i
++)
6308 for (c
= pool
->constants
[i
]; c
; c
= next
)
6314 for (c
= pool
->execute
; c
; c
= next
)
6320 BITMAP_FREE (pool
->insns
);
6325 /* Collect main literal pool. Return NULL on overflow. */
6327 static struct constant_pool
*
6328 s390_mainpool_start (void)
6330 struct constant_pool
*pool
;
6333 pool
= s390_alloc_pool ();
6335 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6337 if (GET_CODE (insn
) == INSN
6338 && GET_CODE (PATTERN (insn
)) == SET
6339 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC_VOLATILE
6340 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPECV_MAIN_POOL
)
6342 gcc_assert (!pool
->pool_insn
);
6343 pool
->pool_insn
= insn
;
6346 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6348 s390_add_execute (pool
, insn
);
6350 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6352 rtx pool_ref
= NULL_RTX
;
6353 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6356 rtx constant
= get_pool_constant (pool_ref
);
6357 enum machine_mode mode
= get_pool_mode (pool_ref
);
6358 s390_add_constant (pool
, constant
, mode
);
6362 /* If hot/cold partitioning is enabled we have to make sure that
6363 the literal pool is emitted in the same section where the
6364 initialization of the literal pool base pointer takes place.
6365 emit_pool_after is only used in the non-overflow case on non
6366 Z cpus where we can emit the literal pool at the end of the
6367 function body within the text section. */
6369 && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
6370 && !pool
->emit_pool_after
)
6371 pool
->emit_pool_after
= PREV_INSN (insn
);
6374 gcc_assert (pool
->pool_insn
|| pool
->size
== 0);
6376 if (pool
->size
>= 4096)
6378 /* We're going to chunkify the pool, so remove the main
6379 pool placeholder insn. */
6380 remove_insn (pool
->pool_insn
);
6382 s390_free_pool (pool
);
6386 /* If the functions ends with the section where the literal pool
6387 should be emitted set the marker to its end. */
6388 if (pool
&& !pool
->emit_pool_after
)
6389 pool
->emit_pool_after
= get_last_insn ();
6394 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6395 Modify the current function to output the pool constants as well as
6396 the pool register setup instruction. */
6399 s390_mainpool_finish (struct constant_pool
*pool
)
6401 rtx base_reg
= cfun
->machine
->base_reg
;
6404 /* If the pool is empty, we're done. */
6405 if (pool
->size
== 0)
6407 /* We don't actually need a base register after all. */
6408 cfun
->machine
->base_reg
= NULL_RTX
;
6410 if (pool
->pool_insn
)
6411 remove_insn (pool
->pool_insn
);
6412 s390_free_pool (pool
);
6416 /* We need correct insn addresses. */
6417 shorten_branches (get_insns ());
6419 /* On zSeries, we use a LARL to load the pool register. The pool is
6420 located in the .rodata section, so we emit it after the function. */
6421 if (TARGET_CPU_ZARCH
)
6423 insn
= gen_main_base_64 (base_reg
, pool
->label
);
6424 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6425 INSN_ADDRESSES_NEW (insn
, -1);
6426 remove_insn (pool
->pool_insn
);
6428 insn
= get_last_insn ();
6429 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6430 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6432 s390_dump_pool (pool
, 0);
6435 /* On S/390, if the total size of the function's code plus literal pool
6436 does not exceed 4096 bytes, we use BASR to set up a function base
6437 pointer, and emit the literal pool at the end of the function. */
6438 else if (INSN_ADDRESSES (INSN_UID (pool
->emit_pool_after
))
6439 + pool
->size
+ 8 /* alignment slop */ < 4096)
6441 insn
= gen_main_base_31_small (base_reg
, pool
->label
);
6442 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6443 INSN_ADDRESSES_NEW (insn
, -1);
6444 remove_insn (pool
->pool_insn
);
6446 insn
= emit_label_after (pool
->label
, insn
);
6447 INSN_ADDRESSES_NEW (insn
, -1);
6449 /* emit_pool_after will be set by s390_mainpool_start to the
6450 last insn of the section where the literal pool should be
6452 insn
= pool
->emit_pool_after
;
6454 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6455 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6457 s390_dump_pool (pool
, 1);
6460 /* Otherwise, we emit an inline literal pool and use BASR to branch
6461 over it, setting up the pool register at the same time. */
6464 rtx pool_end
= gen_label_rtx ();
6466 insn
= gen_main_base_31_large (base_reg
, pool
->label
, pool_end
);
6467 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6468 INSN_ADDRESSES_NEW (insn
, -1);
6469 remove_insn (pool
->pool_insn
);
6471 insn
= emit_label_after (pool
->label
, insn
);
6472 INSN_ADDRESSES_NEW (insn
, -1);
6474 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6475 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6477 insn
= emit_label_after (pool_end
, pool
->pool_insn
);
6478 INSN_ADDRESSES_NEW (insn
, -1);
6480 s390_dump_pool (pool
, 1);
6484 /* Replace all literal pool references. */
6486 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6489 replace_ltrel_base (&PATTERN (insn
));
6491 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6493 rtx addr
, pool_ref
= NULL_RTX
;
6494 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6497 if (s390_execute_label (insn
))
6498 addr
= s390_find_execute (pool
, insn
);
6500 addr
= s390_find_constant (pool
, get_pool_constant (pool_ref
),
6501 get_pool_mode (pool_ref
));
6503 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6504 INSN_CODE (insn
) = -1;
6510 /* Free the pool. */
6511 s390_free_pool (pool
);
6514 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6515 We have decided we cannot use this pool, so revert all changes
6516 to the current function that were done by s390_mainpool_start. */
6518 s390_mainpool_cancel (struct constant_pool
*pool
)
6520 /* We didn't actually change the instruction stream, so simply
6521 free the pool memory. */
6522 s390_free_pool (pool
);
6526 /* Chunkify the literal pool. */
6528 #define S390_POOL_CHUNK_MIN 0xc00
6529 #define S390_POOL_CHUNK_MAX 0xe00
6531 static struct constant_pool
*
6532 s390_chunkify_start (void)
6534 struct constant_pool
*curr_pool
= NULL
, *pool_list
= NULL
;
6537 rtx pending_ltrel
= NULL_RTX
;
6540 rtx (*gen_reload_base
) (rtx
, rtx
) =
6541 TARGET_CPU_ZARCH
? gen_reload_base_64
: gen_reload_base_31
;
6544 /* We need correct insn addresses. */
6546 shorten_branches (get_insns ());
6548 /* Scan all insns and move literals to pool chunks. */
6550 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6552 bool section_switch_p
= false;
6554 /* Check for pending LTREL_BASE. */
6557 rtx ltrel_base
= find_ltrel_base (PATTERN (insn
));
6560 gcc_assert (ltrel_base
== pending_ltrel
);
6561 pending_ltrel
= NULL_RTX
;
6565 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6568 curr_pool
= s390_start_pool (&pool_list
, insn
);
6570 s390_add_execute (curr_pool
, insn
);
6571 s390_add_pool_insn (curr_pool
, insn
);
6573 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6575 rtx pool_ref
= NULL_RTX
;
6576 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6579 rtx constant
= get_pool_constant (pool_ref
);
6580 enum machine_mode mode
= get_pool_mode (pool_ref
);
6583 curr_pool
= s390_start_pool (&pool_list
, insn
);
6585 s390_add_constant (curr_pool
, constant
, mode
);
6586 s390_add_pool_insn (curr_pool
, insn
);
6588 /* Don't split the pool chunk between a LTREL_OFFSET load
6589 and the corresponding LTREL_BASE. */
6590 if (GET_CODE (constant
) == CONST
6591 && GET_CODE (XEXP (constant
, 0)) == UNSPEC
6592 && XINT (XEXP (constant
, 0), 1) == UNSPEC_LTREL_OFFSET
)
6594 gcc_assert (!pending_ltrel
);
6595 pending_ltrel
= pool_ref
;
6600 if (GET_CODE (insn
) == JUMP_INSN
|| GET_CODE (insn
) == CODE_LABEL
)
6603 s390_add_pool_insn (curr_pool
, insn
);
6604 /* An LTREL_BASE must follow within the same basic block. */
6605 gcc_assert (!pending_ltrel
);
6608 if (NOTE_P (insn
) && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
)
6609 section_switch_p
= true;
6612 || INSN_ADDRESSES_SIZE () <= (size_t) INSN_UID (insn
)
6613 || INSN_ADDRESSES (INSN_UID (insn
)) == -1)
6616 if (TARGET_CPU_ZARCH
)
6618 if (curr_pool
->size
< S390_POOL_CHUNK_MAX
)
6621 s390_end_pool (curr_pool
, NULL_RTX
);
6626 int chunk_size
= INSN_ADDRESSES (INSN_UID (insn
))
6627 - INSN_ADDRESSES (INSN_UID (curr_pool
->first_insn
))
6630 /* We will later have to insert base register reload insns.
6631 Those will have an effect on code size, which we need to
6632 consider here. This calculation makes rather pessimistic
6633 worst-case assumptions. */
6634 if (GET_CODE (insn
) == CODE_LABEL
)
6637 if (chunk_size
< S390_POOL_CHUNK_MIN
6638 && curr_pool
->size
< S390_POOL_CHUNK_MIN
6639 && !section_switch_p
)
6642 /* Pool chunks can only be inserted after BARRIERs ... */
6643 if (GET_CODE (insn
) == BARRIER
)
6645 s390_end_pool (curr_pool
, insn
);
6650 /* ... so if we don't find one in time, create one. */
6651 else if (chunk_size
> S390_POOL_CHUNK_MAX
6652 || curr_pool
->size
> S390_POOL_CHUNK_MAX
6653 || section_switch_p
)
6655 rtx label
, jump
, barrier
;
6657 if (!section_switch_p
)
6659 /* We can insert the barrier only after a 'real' insn. */
6660 if (GET_CODE (insn
) != INSN
&& GET_CODE (insn
) != CALL_INSN
)
6662 if (get_attr_length (insn
) == 0)
6664 /* Don't separate LTREL_BASE from the corresponding
6665 LTREL_OFFSET load. */
6671 gcc_assert (!pending_ltrel
);
6673 /* The old pool has to end before the section switch
6674 note in order to make it part of the current
6676 insn
= PREV_INSN (insn
);
6679 label
= gen_label_rtx ();
6680 jump
= emit_jump_insn_after (gen_jump (label
), insn
);
6681 barrier
= emit_barrier_after (jump
);
6682 insn
= emit_label_after (label
, barrier
);
6683 JUMP_LABEL (jump
) = label
;
6684 LABEL_NUSES (label
) = 1;
6686 INSN_ADDRESSES_NEW (jump
, -1);
6687 INSN_ADDRESSES_NEW (barrier
, -1);
6688 INSN_ADDRESSES_NEW (insn
, -1);
6690 s390_end_pool (curr_pool
, barrier
);
6698 s390_end_pool (curr_pool
, NULL_RTX
);
6699 gcc_assert (!pending_ltrel
);
6701 /* Find all labels that are branched into
6702 from an insn belonging to a different chunk. */
6704 far_labels
= BITMAP_ALLOC (NULL
);
6706 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6708 /* Labels marked with LABEL_PRESERVE_P can be target
6709 of non-local jumps, so we have to mark them.
6710 The same holds for named labels.
6712 Don't do that, however, if it is the label before
6715 if (GET_CODE (insn
) == CODE_LABEL
6716 && (LABEL_PRESERVE_P (insn
) || LABEL_NAME (insn
)))
6718 rtx vec_insn
= next_real_insn (insn
);
6719 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6720 PATTERN (vec_insn
) : NULL_RTX
;
6722 || !(GET_CODE (vec_pat
) == ADDR_VEC
6723 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6724 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (insn
));
6727 /* If we have a direct jump (conditional or unconditional)
6728 or a casesi jump, check all potential targets. */
6729 else if (GET_CODE (insn
) == JUMP_INSN
)
6731 rtx pat
= PATTERN (insn
);
6732 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
6733 pat
= XVECEXP (pat
, 0, 0);
6735 if (GET_CODE (pat
) == SET
)
6737 rtx label
= JUMP_LABEL (insn
);
6740 if (s390_find_pool (pool_list
, label
)
6741 != s390_find_pool (pool_list
, insn
))
6742 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6745 else if (GET_CODE (pat
) == PARALLEL
6746 && XVECLEN (pat
, 0) == 2
6747 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
6748 && GET_CODE (XVECEXP (pat
, 0, 1)) == USE
6749 && GET_CODE (XEXP (XVECEXP (pat
, 0, 1), 0)) == LABEL_REF
)
6751 /* Find the jump table used by this casesi jump. */
6752 rtx vec_label
= XEXP (XEXP (XVECEXP (pat
, 0, 1), 0), 0);
6753 rtx vec_insn
= next_real_insn (vec_label
);
6754 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6755 PATTERN (vec_insn
) : NULL_RTX
;
6757 && (GET_CODE (vec_pat
) == ADDR_VEC
6758 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6760 int i
, diff_p
= GET_CODE (vec_pat
) == ADDR_DIFF_VEC
;
6762 for (i
= 0; i
< XVECLEN (vec_pat
, diff_p
); i
++)
6764 rtx label
= XEXP (XVECEXP (vec_pat
, diff_p
, i
), 0);
6766 if (s390_find_pool (pool_list
, label
)
6767 != s390_find_pool (pool_list
, insn
))
6768 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6775 /* Insert base register reload insns before every pool. */
6777 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6779 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6781 rtx insn
= curr_pool
->first_insn
;
6782 INSN_ADDRESSES_NEW (emit_insn_before (new_insn
, insn
), -1);
6785 /* Insert base register reload insns at every far label. */
6787 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6788 if (GET_CODE (insn
) == CODE_LABEL
6789 && bitmap_bit_p (far_labels
, CODE_LABEL_NUMBER (insn
)))
6791 struct constant_pool
*pool
= s390_find_pool (pool_list
, insn
);
6794 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6796 INSN_ADDRESSES_NEW (emit_insn_after (new_insn
, insn
), -1);
6801 BITMAP_FREE (far_labels
);
6804 /* Recompute insn addresses. */
6806 init_insn_lengths ();
6807 shorten_branches (get_insns ());
6812 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6813 After we have decided to use this list, finish implementing
6814 all changes to the current function as required. */
6817 s390_chunkify_finish (struct constant_pool
*pool_list
)
6819 struct constant_pool
*curr_pool
= NULL
;
6823 /* Replace all literal pool references. */
6825 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6828 replace_ltrel_base (&PATTERN (insn
));
6830 curr_pool
= s390_find_pool (pool_list
, insn
);
6834 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6836 rtx addr
, pool_ref
= NULL_RTX
;
6837 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6840 if (s390_execute_label (insn
))
6841 addr
= s390_find_execute (curr_pool
, insn
);
6843 addr
= s390_find_constant (curr_pool
,
6844 get_pool_constant (pool_ref
),
6845 get_pool_mode (pool_ref
));
6847 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6848 INSN_CODE (insn
) = -1;
6853 /* Dump out all literal pools. */
6855 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6856 s390_dump_pool (curr_pool
, 0);
6858 /* Free pool list. */
6862 struct constant_pool
*next
= pool_list
->next
;
6863 s390_free_pool (pool_list
);
6868 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6869 We have decided we cannot use this list, so revert all changes
6870 to the current function that were done by s390_chunkify_start. */
6873 s390_chunkify_cancel (struct constant_pool
*pool_list
)
6875 struct constant_pool
*curr_pool
= NULL
;
6878 /* Remove all pool placeholder insns. */
6880 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6882 /* Did we insert an extra barrier? Remove it. */
6883 rtx barrier
= PREV_INSN (curr_pool
->pool_insn
);
6884 rtx jump
= barrier
? PREV_INSN (barrier
) : NULL_RTX
;
6885 rtx label
= NEXT_INSN (curr_pool
->pool_insn
);
6887 if (jump
&& GET_CODE (jump
) == JUMP_INSN
6888 && barrier
&& GET_CODE (barrier
) == BARRIER
6889 && label
&& GET_CODE (label
) == CODE_LABEL
6890 && GET_CODE (PATTERN (jump
)) == SET
6891 && SET_DEST (PATTERN (jump
)) == pc_rtx
6892 && GET_CODE (SET_SRC (PATTERN (jump
))) == LABEL_REF
6893 && XEXP (SET_SRC (PATTERN (jump
)), 0) == label
)
6896 remove_insn (barrier
);
6897 remove_insn (label
);
6900 remove_insn (curr_pool
->pool_insn
);
6903 /* Remove all base register reload insns. */
6905 for (insn
= get_insns (); insn
; )
6907 rtx next_insn
= NEXT_INSN (insn
);
6909 if (GET_CODE (insn
) == INSN
6910 && GET_CODE (PATTERN (insn
)) == SET
6911 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC
6912 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPEC_RELOAD_BASE
)
6918 /* Free pool list. */
6922 struct constant_pool
*next
= pool_list
->next
;
6923 s390_free_pool (pool_list
);
6928 /* Output the constant pool entry EXP in mode MODE with alignment ALIGN. */
6931 s390_output_pool_entry (rtx exp
, enum machine_mode mode
, unsigned int align
)
6935 switch (GET_MODE_CLASS (mode
))
6938 case MODE_DECIMAL_FLOAT
:
6939 gcc_assert (GET_CODE (exp
) == CONST_DOUBLE
);
6941 REAL_VALUE_FROM_CONST_DOUBLE (r
, exp
);
6942 assemble_real (r
, mode
, align
);
6946 assemble_integer (exp
, GET_MODE_SIZE (mode
), align
, 1);
6947 mark_symbol_refs_as_used (exp
);
6956 /* Return an RTL expression representing the value of the return address
6957 for the frame COUNT steps up from the current frame. FRAME is the
6958 frame pointer of that frame. */
6961 s390_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
6966 /* Without backchain, we fail for all but the current frame. */
6968 if (!TARGET_BACKCHAIN
&& count
> 0)
6971 /* For the current frame, we need to make sure the initial
6972 value of RETURN_REGNUM is actually saved. */
6976 /* On non-z architectures branch splitting could overwrite r14. */
6977 if (TARGET_CPU_ZARCH
)
6978 return get_hard_reg_initial_val (Pmode
, RETURN_REGNUM
);
6981 cfun_frame_layout
.save_return_addr_p
= true;
6982 return gen_rtx_MEM (Pmode
, return_address_pointer_rtx
);
6986 if (TARGET_PACKED_STACK
)
6987 offset
= -2 * UNITS_PER_LONG
;
6989 offset
= RETURN_REGNUM
* UNITS_PER_LONG
;
6991 addr
= plus_constant (frame
, offset
);
6992 addr
= memory_address (Pmode
, addr
);
6993 return gen_rtx_MEM (Pmode
, addr
);
6996 /* Return an RTL expression representing the back chain stored in
6997 the current stack frame. */
7000 s390_back_chain_rtx (void)
7004 gcc_assert (TARGET_BACKCHAIN
);
7006 if (TARGET_PACKED_STACK
)
7007 chain
= plus_constant (stack_pointer_rtx
,
7008 STACK_POINTER_OFFSET
- UNITS_PER_LONG
);
7010 chain
= stack_pointer_rtx
;
7012 chain
= gen_rtx_MEM (Pmode
, chain
);
7016 /* Find first call clobbered register unused in a function.
7017 This could be used as base register in a leaf function
7018 or for holding the return address before epilogue. */
7021 find_unused_clobbered_reg (void)
7024 for (i
= 0; i
< 6; i
++)
7025 if (!df_regs_ever_live_p (i
))
7031 /* Helper function for s390_regs_ever_clobbered. Sets the fields in DATA for all
7032 clobbered hard regs in SETREG. */
7035 s390_reg_clobbered_rtx (rtx setreg
, const_rtx set_insn ATTRIBUTE_UNUSED
, void *data
)
7037 int *regs_ever_clobbered
= (int *)data
;
7038 unsigned int i
, regno
;
7039 enum machine_mode mode
= GET_MODE (setreg
);
7041 if (GET_CODE (setreg
) == SUBREG
)
7043 rtx inner
= SUBREG_REG (setreg
);
7044 if (!GENERAL_REG_P (inner
))
7046 regno
= subreg_regno (setreg
);
7048 else if (GENERAL_REG_P (setreg
))
7049 regno
= REGNO (setreg
);
7054 i
< regno
+ HARD_REGNO_NREGS (regno
, mode
);
7056 regs_ever_clobbered
[i
] = 1;
7059 /* Walks through all basic blocks of the current function looking
7060 for clobbered hard regs using s390_reg_clobbered_rtx. The fields
7061 of the passed integer array REGS_EVER_CLOBBERED are set to one for
7062 each of those regs. */
7065 s390_regs_ever_clobbered (int *regs_ever_clobbered
)
7071 memset (regs_ever_clobbered
, 0, 16 * sizeof (int));
7073 /* For non-leaf functions we have to consider all call clobbered regs to be
7075 if (!current_function_is_leaf
)
7077 for (i
= 0; i
< 16; i
++)
7078 regs_ever_clobbered
[i
] = call_really_used_regs
[i
];
7081 /* Make the "magic" eh_return registers live if necessary. For regs_ever_live
7082 this work is done by liveness analysis (mark_regs_live_at_end).
7083 Special care is needed for functions containing landing pads. Landing pads
7084 may use the eh registers, but the code which sets these registers is not
7085 contained in that function. Hence s390_regs_ever_clobbered is not able to
7086 deal with this automatically. */
7087 if (crtl
->calls_eh_return
|| cfun
->machine
->has_landing_pad_p
)
7088 for (i
= 0; EH_RETURN_DATA_REGNO (i
) != INVALID_REGNUM
; i
++)
7089 if (crtl
->calls_eh_return
7090 || (cfun
->machine
->has_landing_pad_p
7091 && df_regs_ever_live_p (EH_RETURN_DATA_REGNO (i
))))
7092 regs_ever_clobbered
[EH_RETURN_DATA_REGNO (i
)] = 1;
7094 /* For nonlocal gotos all call-saved registers have to be saved.
7095 This flag is also set for the unwinding code in libgcc.
7096 See expand_builtin_unwind_init. For regs_ever_live this is done by
7098 if (cfun
->has_nonlocal_label
)
7099 for (i
= 0; i
< 16; i
++)
7100 if (!call_really_used_regs
[i
])
7101 regs_ever_clobbered
[i
] = 1;
7103 FOR_EACH_BB (cur_bb
)
7105 FOR_BB_INSNS (cur_bb
, cur_insn
)
7107 if (INSN_P (cur_insn
))
7108 note_stores (PATTERN (cur_insn
),
7109 s390_reg_clobbered_rtx
,
7110 regs_ever_clobbered
);
7115 /* Determine the frame area which actually has to be accessed
7116 in the function epilogue. The values are stored at the
7117 given pointers AREA_BOTTOM (address of the lowest used stack
7118 address) and AREA_TOP (address of the first item which does
7119 not belong to the stack frame). */
7122 s390_frame_area (int *area_bottom
, int *area_top
)
7130 if (cfun_frame_layout
.first_restore_gpr
!= -1)
7132 b
= (cfun_frame_layout
.gprs_offset
7133 + cfun_frame_layout
.first_restore_gpr
* UNITS_PER_LONG
);
7134 t
= b
+ (cfun_frame_layout
.last_restore_gpr
7135 - cfun_frame_layout
.first_restore_gpr
+ 1) * UNITS_PER_LONG
;
7138 if (TARGET_64BIT
&& cfun_save_high_fprs_p
)
7140 b
= MIN (b
, cfun_frame_layout
.f8_offset
);
7141 t
= MAX (t
, (cfun_frame_layout
.f8_offset
7142 + cfun_frame_layout
.high_fprs
* 8));
7146 for (i
= 2; i
< 4; i
++)
7147 if (cfun_fpr_bit_p (i
))
7149 b
= MIN (b
, cfun_frame_layout
.f4_offset
+ (i
- 2) * 8);
7150 t
= MAX (t
, cfun_frame_layout
.f4_offset
+ (i
- 1) * 8);
7157 /* Fill cfun->machine with info about register usage of current function.
7158 Return in CLOBBERED_REGS which GPRs are currently considered set. */
7161 s390_register_info (int clobbered_regs
[])
7165 /* fprs 8 - 15 are call saved for 64 Bit ABI. */
7166 cfun_frame_layout
.fpr_bitmap
= 0;
7167 cfun_frame_layout
.high_fprs
= 0;
7169 for (i
= 24; i
< 32; i
++)
7170 if (df_regs_ever_live_p (i
) && !global_regs
[i
])
7172 cfun_set_fpr_bit (i
- 16);
7173 cfun_frame_layout
.high_fprs
++;
7176 /* Find first and last gpr to be saved. We trust regs_ever_live
7177 data, except that we don't save and restore global registers.
7179 Also, all registers with special meaning to the compiler need
7180 to be handled extra. */
7182 s390_regs_ever_clobbered (clobbered_regs
);
7184 for (i
= 0; i
< 16; i
++)
7185 clobbered_regs
[i
] = clobbered_regs
[i
] && !global_regs
[i
] && !fixed_regs
[i
];
7187 if (frame_pointer_needed
)
7188 clobbered_regs
[HARD_FRAME_POINTER_REGNUM
] = 1;
7191 clobbered_regs
[PIC_OFFSET_TABLE_REGNUM
]
7192 |= df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
);
7194 clobbered_regs
[BASE_REGNUM
]
7195 |= (cfun
->machine
->base_reg
7196 && REGNO (cfun
->machine
->base_reg
) == BASE_REGNUM
);
7198 clobbered_regs
[RETURN_REGNUM
]
7199 |= (!current_function_is_leaf
7200 || TARGET_TPF_PROFILING
7201 || cfun
->machine
->split_branches_pending_p
7202 || cfun_frame_layout
.save_return_addr_p
7203 || crtl
->calls_eh_return
7206 clobbered_regs
[STACK_POINTER_REGNUM
]
7207 |= (!current_function_is_leaf
7208 || TARGET_TPF_PROFILING
7209 || cfun_save_high_fprs_p
7210 || get_frame_size () > 0
7211 || cfun
->calls_alloca
7214 for (i
= 6; i
< 16; i
++)
7215 if (df_regs_ever_live_p (i
) || clobbered_regs
[i
])
7217 for (j
= 15; j
> i
; j
--)
7218 if (df_regs_ever_live_p (j
) || clobbered_regs
[j
])
7223 /* Nothing to save/restore. */
7224 cfun_frame_layout
.first_save_gpr_slot
= -1;
7225 cfun_frame_layout
.last_save_gpr_slot
= -1;
7226 cfun_frame_layout
.first_save_gpr
= -1;
7227 cfun_frame_layout
.first_restore_gpr
= -1;
7228 cfun_frame_layout
.last_save_gpr
= -1;
7229 cfun_frame_layout
.last_restore_gpr
= -1;
7233 /* Save slots for gprs from i to j. */
7234 cfun_frame_layout
.first_save_gpr_slot
= i
;
7235 cfun_frame_layout
.last_save_gpr_slot
= j
;
7237 for (i
= cfun_frame_layout
.first_save_gpr_slot
;
7238 i
< cfun_frame_layout
.last_save_gpr_slot
+ 1;
7240 if (clobbered_regs
[i
])
7243 for (j
= cfun_frame_layout
.last_save_gpr_slot
; j
> i
; j
--)
7244 if (clobbered_regs
[j
])
7247 if (i
== cfun_frame_layout
.last_save_gpr_slot
+ 1)
7249 /* Nothing to save/restore. */
7250 cfun_frame_layout
.first_save_gpr
= -1;
7251 cfun_frame_layout
.first_restore_gpr
= -1;
7252 cfun_frame_layout
.last_save_gpr
= -1;
7253 cfun_frame_layout
.last_restore_gpr
= -1;
7257 /* Save / Restore from gpr i to j. */
7258 cfun_frame_layout
.first_save_gpr
= i
;
7259 cfun_frame_layout
.first_restore_gpr
= i
;
7260 cfun_frame_layout
.last_save_gpr
= j
;
7261 cfun_frame_layout
.last_restore_gpr
= j
;
7267 /* Varargs functions need to save gprs 2 to 6. */
7268 if (cfun
->va_list_gpr_size
7269 && crtl
->args
.info
.gprs
< GP_ARG_NUM_REG
)
7271 int min_gpr
= crtl
->args
.info
.gprs
;
7272 int max_gpr
= min_gpr
+ cfun
->va_list_gpr_size
;
7273 if (max_gpr
> GP_ARG_NUM_REG
)
7274 max_gpr
= GP_ARG_NUM_REG
;
7276 if (cfun_frame_layout
.first_save_gpr
== -1
7277 || cfun_frame_layout
.first_save_gpr
> 2 + min_gpr
)
7279 cfun_frame_layout
.first_save_gpr
= 2 + min_gpr
;
7280 cfun_frame_layout
.first_save_gpr_slot
= 2 + min_gpr
;
7283 if (cfun_frame_layout
.last_save_gpr
== -1
7284 || cfun_frame_layout
.last_save_gpr
< 2 + max_gpr
- 1)
7286 cfun_frame_layout
.last_save_gpr
= 2 + max_gpr
- 1;
7287 cfun_frame_layout
.last_save_gpr_slot
= 2 + max_gpr
- 1;
7291 /* Mark f0, f2 for 31 bit and f0-f4 for 64 bit to be saved. */
7292 if (TARGET_HARD_FLOAT
&& cfun
->va_list_fpr_size
7293 && crtl
->args
.info
.fprs
< FP_ARG_NUM_REG
)
7295 int min_fpr
= crtl
->args
.info
.fprs
;
7296 int max_fpr
= min_fpr
+ cfun
->va_list_fpr_size
;
7297 if (max_fpr
> FP_ARG_NUM_REG
)
7298 max_fpr
= FP_ARG_NUM_REG
;
7300 /* ??? This is currently required to ensure proper location
7301 of the fpr save slots within the va_list save area. */
7302 if (TARGET_PACKED_STACK
)
7305 for (i
= min_fpr
; i
< max_fpr
; i
++)
7306 cfun_set_fpr_bit (i
);
7311 for (i
= 2; i
< 4; i
++)
7312 if (df_regs_ever_live_p (i
+ 16) && !global_regs
[i
+ 16])
7313 cfun_set_fpr_bit (i
);
7316 /* Fill cfun->machine with info about frame of current function. */
7319 s390_frame_info (void)
7323 cfun_frame_layout
.frame_size
= get_frame_size ();
7324 if (!TARGET_64BIT
&& cfun_frame_layout
.frame_size
> 0x7fff0000)
7325 fatal_error ("total size of local variables exceeds architecture limit");
7327 if (!TARGET_PACKED_STACK
)
7329 cfun_frame_layout
.backchain_offset
= 0;
7330 cfun_frame_layout
.f0_offset
= 16 * UNITS_PER_LONG
;
7331 cfun_frame_layout
.f4_offset
= cfun_frame_layout
.f0_offset
+ 2 * 8;
7332 cfun_frame_layout
.f8_offset
= -cfun_frame_layout
.high_fprs
* 8;
7333 cfun_frame_layout
.gprs_offset
= (cfun_frame_layout
.first_save_gpr_slot
7336 else if (TARGET_BACKCHAIN
) /* kernel stack layout */
7338 cfun_frame_layout
.backchain_offset
= (STACK_POINTER_OFFSET
7340 cfun_frame_layout
.gprs_offset
7341 = (cfun_frame_layout
.backchain_offset
7342 - (STACK_POINTER_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
+ 1)
7347 cfun_frame_layout
.f4_offset
7348 = (cfun_frame_layout
.gprs_offset
7349 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7351 cfun_frame_layout
.f0_offset
7352 = (cfun_frame_layout
.f4_offset
7353 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7357 /* On 31 bit we have to care about alignment of the
7358 floating point regs to provide fastest access. */
7359 cfun_frame_layout
.f0_offset
7360 = ((cfun_frame_layout
.gprs_offset
7361 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1))
7362 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7364 cfun_frame_layout
.f4_offset
7365 = (cfun_frame_layout
.f0_offset
7366 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7369 else /* no backchain */
7371 cfun_frame_layout
.f4_offset
7372 = (STACK_POINTER_OFFSET
7373 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7375 cfun_frame_layout
.f0_offset
7376 = (cfun_frame_layout
.f4_offset
7377 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7379 cfun_frame_layout
.gprs_offset
7380 = cfun_frame_layout
.f0_offset
- cfun_gprs_save_area_size
;
7383 if (current_function_is_leaf
7384 && !TARGET_TPF_PROFILING
7385 && cfun_frame_layout
.frame_size
== 0
7386 && !cfun_save_high_fprs_p
7387 && !cfun
->calls_alloca
7391 if (!TARGET_PACKED_STACK
)
7392 cfun_frame_layout
.frame_size
+= (STACK_POINTER_OFFSET
7393 + crtl
->outgoing_args_size
7394 + cfun_frame_layout
.high_fprs
* 8);
7397 if (TARGET_BACKCHAIN
)
7398 cfun_frame_layout
.frame_size
+= UNITS_PER_LONG
;
7400 /* No alignment trouble here because f8-f15 are only saved under
7402 cfun_frame_layout
.f8_offset
= (MIN (MIN (cfun_frame_layout
.f0_offset
,
7403 cfun_frame_layout
.f4_offset
),
7404 cfun_frame_layout
.gprs_offset
)
7405 - cfun_frame_layout
.high_fprs
* 8);
7407 cfun_frame_layout
.frame_size
+= cfun_frame_layout
.high_fprs
* 8;
7409 for (i
= 0; i
< 8; i
++)
7410 if (cfun_fpr_bit_p (i
))
7411 cfun_frame_layout
.frame_size
+= 8;
7413 cfun_frame_layout
.frame_size
+= cfun_gprs_save_area_size
;
7415 /* If under 31 bit an odd number of gprs has to be saved we have to adjust
7416 the frame size to sustain 8 byte alignment of stack frames. */
7417 cfun_frame_layout
.frame_size
= ((cfun_frame_layout
.frame_size
+
7418 STACK_BOUNDARY
/ BITS_PER_UNIT
- 1)
7419 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1));
7421 cfun_frame_layout
.frame_size
+= crtl
->outgoing_args_size
;
7425 /* Generate frame layout. Fills in register and frame data for the current
7426 function in cfun->machine. This routine can be called multiple times;
7427 it will re-do the complete frame layout every time. */
7430 s390_init_frame_layout (void)
7432 HOST_WIDE_INT frame_size
;
7434 int clobbered_regs
[16];
7436 /* On S/390 machines, we may need to perform branch splitting, which
7437 will require both base and return address register. We have no
7438 choice but to assume we're going to need them until right at the
7439 end of the machine dependent reorg phase. */
7440 if (!TARGET_CPU_ZARCH
)
7441 cfun
->machine
->split_branches_pending_p
= true;
7445 frame_size
= cfun_frame_layout
.frame_size
;
7447 /* Try to predict whether we'll need the base register. */
7448 base_used
= cfun
->machine
->split_branches_pending_p
7449 || crtl
->uses_const_pool
7450 || (!DISP_IN_RANGE (frame_size
)
7451 && !CONST_OK_FOR_K (frame_size
));
7453 /* Decide which register to use as literal pool base. In small
7454 leaf functions, try to use an unused call-clobbered register
7455 as base register to avoid save/restore overhead. */
7457 cfun
->machine
->base_reg
= NULL_RTX
;
7458 else if (current_function_is_leaf
&& !df_regs_ever_live_p (5))
7459 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, 5);
7461 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, BASE_REGNUM
);
7463 s390_register_info (clobbered_regs
);
7466 while (frame_size
!= cfun_frame_layout
.frame_size
);
7469 /* Update frame layout. Recompute actual register save data based on
7470 current info and update regs_ever_live for the special registers.
7471 May be called multiple times, but may never cause *more* registers
7472 to be saved than s390_init_frame_layout allocated room for. */
7475 s390_update_frame_layout (void)
7477 int clobbered_regs
[16];
7479 s390_register_info (clobbered_regs
);
7481 df_set_regs_ever_live (BASE_REGNUM
,
7482 clobbered_regs
[BASE_REGNUM
] ? true : false);
7483 df_set_regs_ever_live (RETURN_REGNUM
,
7484 clobbered_regs
[RETURN_REGNUM
] ? true : false);
7485 df_set_regs_ever_live (STACK_POINTER_REGNUM
,
7486 clobbered_regs
[STACK_POINTER_REGNUM
] ? true : false);
7488 if (cfun
->machine
->base_reg
)
7489 df_set_regs_ever_live (REGNO (cfun
->machine
->base_reg
), true);
7492 /* Return true if it is legal to put a value with MODE into REGNO. */
7495 s390_hard_regno_mode_ok (unsigned int regno
, enum machine_mode mode
)
7497 switch (REGNO_REG_CLASS (regno
))
7500 if (REGNO_PAIR_OK (regno
, mode
))
7502 if (mode
== SImode
|| mode
== DImode
)
7505 if (FLOAT_MODE_P (mode
) && GET_MODE_CLASS (mode
) != MODE_VECTOR_FLOAT
)
7510 if (FRAME_REGNO_P (regno
) && mode
== Pmode
)
7515 if (REGNO_PAIR_OK (regno
, mode
))
7518 || (mode
!= TFmode
&& mode
!= TCmode
&& mode
!= TDmode
))
7523 if (GET_MODE_CLASS (mode
) == MODE_CC
)
7527 if (REGNO_PAIR_OK (regno
, mode
))
7529 if (mode
== SImode
|| mode
== Pmode
)
7540 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
7543 s390_hard_regno_rename_ok (unsigned int old_reg
, unsigned int new_reg
)
7545 /* Once we've decided upon a register to use as base register, it must
7546 no longer be used for any other purpose. */
7547 if (cfun
->machine
->base_reg
)
7548 if (REGNO (cfun
->machine
->base_reg
) == old_reg
7549 || REGNO (cfun
->machine
->base_reg
) == new_reg
)
7555 /* Maximum number of registers to represent a value of mode MODE
7556 in a register of class RCLASS. */
7559 s390_class_max_nregs (enum reg_class rclass
, enum machine_mode mode
)
7564 if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7565 return 2 * ((GET_MODE_SIZE (mode
) / 2 + 8 - 1) / 8);
7567 return (GET_MODE_SIZE (mode
) + 8 - 1) / 8;
7569 return (GET_MODE_SIZE (mode
) + 4 - 1) / 4;
7573 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
7576 /* Return true if register FROM can be eliminated via register TO. */
7579 s390_can_eliminate (const int from
, const int to
)
7581 /* On zSeries machines, we have not marked the base register as fixed.
7582 Instead, we have an elimination rule BASE_REGNUM -> BASE_REGNUM.
7583 If a function requires the base register, we say here that this
7584 elimination cannot be performed. This will cause reload to free
7585 up the base register (as if it were fixed). On the other hand,
7586 if the current function does *not* require the base register, we
7587 say here the elimination succeeds, which in turn allows reload
7588 to allocate the base register for any other purpose. */
7589 if (from
== BASE_REGNUM
&& to
== BASE_REGNUM
)
7591 if (TARGET_CPU_ZARCH
)
7593 s390_init_frame_layout ();
7594 return cfun
->machine
->base_reg
== NULL_RTX
;
7600 /* Everything else must point into the stack frame. */
7601 gcc_assert (to
== STACK_POINTER_REGNUM
7602 || to
== HARD_FRAME_POINTER_REGNUM
);
7604 gcc_assert (from
== FRAME_POINTER_REGNUM
7605 || from
== ARG_POINTER_REGNUM
7606 || from
== RETURN_ADDRESS_POINTER_REGNUM
);
7608 /* Make sure we actually saved the return address. */
7609 if (from
== RETURN_ADDRESS_POINTER_REGNUM
)
7610 if (!crtl
->calls_eh_return
7612 && !cfun_frame_layout
.save_return_addr_p
)
7618 /* Return offset between register FROM and TO initially after prolog. */
7621 s390_initial_elimination_offset (int from
, int to
)
7623 HOST_WIDE_INT offset
;
7626 /* ??? Why are we called for non-eliminable pairs? */
7627 if (!s390_can_eliminate (from
, to
))
7632 case FRAME_POINTER_REGNUM
:
7633 offset
= (get_frame_size()
7634 + STACK_POINTER_OFFSET
7635 + crtl
->outgoing_args_size
);
7638 case ARG_POINTER_REGNUM
:
7639 s390_init_frame_layout ();
7640 offset
= cfun_frame_layout
.frame_size
+ STACK_POINTER_OFFSET
;
7643 case RETURN_ADDRESS_POINTER_REGNUM
:
7644 s390_init_frame_layout ();
7645 index
= RETURN_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
;
7646 gcc_assert (index
>= 0);
7647 offset
= cfun_frame_layout
.frame_size
+ cfun_frame_layout
.gprs_offset
;
7648 offset
+= index
* UNITS_PER_LONG
;
7662 /* Emit insn to save fpr REGNUM at offset OFFSET relative
7663 to register BASE. Return generated insn. */
7666 save_fpr (rtx base
, int offset
, int regnum
)
7669 addr
= gen_rtx_MEM (DFmode
, plus_constant (base
, offset
));
7671 if (regnum
>= 16 && regnum
<= (16 + FP_ARG_NUM_REG
))
7672 set_mem_alias_set (addr
, get_varargs_alias_set ());
7674 set_mem_alias_set (addr
, get_frame_alias_set ());
7676 return emit_move_insn (addr
, gen_rtx_REG (DFmode
, regnum
));
7679 /* Emit insn to restore fpr REGNUM from offset OFFSET relative
7680 to register BASE. Return generated insn. */
7683 restore_fpr (rtx base
, int offset
, int regnum
)
7686 addr
= gen_rtx_MEM (DFmode
, plus_constant (base
, offset
));
7687 set_mem_alias_set (addr
, get_frame_alias_set ());
7689 return emit_move_insn (gen_rtx_REG (DFmode
, regnum
), addr
);
7692 /* Return true if REGNO is a global register, but not one
7693 of the special ones that need to be saved/restored in anyway. */
7696 global_not_special_regno_p (int regno
)
7698 return (global_regs
[regno
]
7699 /* These registers are special and need to be
7700 restored in any case. */
7701 && !(regno
== STACK_POINTER_REGNUM
7702 || regno
== RETURN_REGNUM
7703 || regno
== BASE_REGNUM
7704 || (flag_pic
&& regno
== (int)PIC_OFFSET_TABLE_REGNUM
)));
7707 /* Generate insn to save registers FIRST to LAST into
7708 the register save area located at offset OFFSET
7709 relative to register BASE. */
7712 save_gprs (rtx base
, int offset
, int first
, int last
)
7714 rtx addr
, insn
, note
;
7717 addr
= plus_constant (base
, offset
);
7718 addr
= gen_rtx_MEM (Pmode
, addr
);
7720 set_mem_alias_set (addr
, get_frame_alias_set ());
7722 /* Special-case single register. */
7726 insn
= gen_movdi (addr
, gen_rtx_REG (Pmode
, first
));
7728 insn
= gen_movsi (addr
, gen_rtx_REG (Pmode
, first
));
7730 if (!global_not_special_regno_p (first
))
7731 RTX_FRAME_RELATED_P (insn
) = 1;
7736 insn
= gen_store_multiple (addr
,
7737 gen_rtx_REG (Pmode
, first
),
7738 GEN_INT (last
- first
+ 1));
7740 if (first
<= 6 && cfun
->stdarg
)
7741 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
7743 rtx mem
= XEXP (XVECEXP (PATTERN (insn
), 0, i
), 0);
7746 set_mem_alias_set (mem
, get_varargs_alias_set ());
7749 /* We need to set the FRAME_RELATED flag on all SETs
7750 inside the store-multiple pattern.
7752 However, we must not emit DWARF records for registers 2..5
7753 if they are stored for use by variable arguments ...
7755 ??? Unfortunately, it is not enough to simply not the
7756 FRAME_RELATED flags for those SETs, because the first SET
7757 of the PARALLEL is always treated as if it had the flag
7758 set, even if it does not. Therefore we emit a new pattern
7759 without those registers as REG_FRAME_RELATED_EXPR note. */
7761 if (first
>= 6 && !global_not_special_regno_p (first
))
7763 rtx pat
= PATTERN (insn
);
7765 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
7766 if (GET_CODE (XVECEXP (pat
, 0, i
)) == SET
7767 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (pat
,
7769 RTX_FRAME_RELATED_P (XVECEXP (pat
, 0, i
)) = 1;
7771 RTX_FRAME_RELATED_P (insn
) = 1;
7777 for (start
= first
>= 6 ? first
: 6; start
<= last
; start
++)
7778 if (!global_not_special_regno_p (start
))
7784 addr
= plus_constant (base
, offset
+ (start
- first
) * UNITS_PER_LONG
);
7785 note
= gen_store_multiple (gen_rtx_MEM (Pmode
, addr
),
7786 gen_rtx_REG (Pmode
, start
),
7787 GEN_INT (last
- start
+ 1));
7788 note
= PATTERN (note
);
7790 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, note
);
7792 for (i
= 0; i
< XVECLEN (note
, 0); i
++)
7793 if (GET_CODE (XVECEXP (note
, 0, i
)) == SET
7794 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (note
,
7796 RTX_FRAME_RELATED_P (XVECEXP (note
, 0, i
)) = 1;
7798 RTX_FRAME_RELATED_P (insn
) = 1;
7804 /* Generate insn to restore registers FIRST to LAST from
7805 the register save area located at offset OFFSET
7806 relative to register BASE. */
7809 restore_gprs (rtx base
, int offset
, int first
, int last
)
7813 addr
= plus_constant (base
, offset
);
7814 addr
= gen_rtx_MEM (Pmode
, addr
);
7815 set_mem_alias_set (addr
, get_frame_alias_set ());
7817 /* Special-case single register. */
7821 insn
= gen_movdi (gen_rtx_REG (Pmode
, first
), addr
);
7823 insn
= gen_movsi (gen_rtx_REG (Pmode
, first
), addr
);
7828 insn
= gen_load_multiple (gen_rtx_REG (Pmode
, first
),
7830 GEN_INT (last
- first
+ 1));
7834 /* Return insn sequence to load the GOT register. */
7836 static GTY(()) rtx got_symbol
;
7838 s390_load_got (void)
7844 got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
7845 SYMBOL_REF_FLAGS (got_symbol
) = SYMBOL_FLAG_LOCAL
;
7850 if (TARGET_CPU_ZARCH
)
7852 emit_move_insn (pic_offset_table_rtx
, got_symbol
);
7858 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, got_symbol
),
7859 UNSPEC_LTREL_OFFSET
);
7860 offset
= gen_rtx_CONST (Pmode
, offset
);
7861 offset
= force_const_mem (Pmode
, offset
);
7863 emit_move_insn (pic_offset_table_rtx
, offset
);
7865 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (offset
, 0)),
7867 offset
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, offset
);
7869 emit_move_insn (pic_offset_table_rtx
, offset
);
7872 insns
= get_insns ();
7877 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
7878 and the change to the stack pointer. */
7881 s390_emit_stack_tie (void)
7883 rtx mem
= gen_frame_mem (BLKmode
,
7884 gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
));
7886 emit_insn (gen_stack_tie (mem
));
7889 /* Expand the prologue into a bunch of separate insns. */
7892 s390_emit_prologue (void)
7900 /* Complete frame layout. */
7902 s390_update_frame_layout ();
7904 /* Annotate all constant pool references to let the scheduler know
7905 they implicitly use the base register. */
7907 push_topmost_sequence ();
7909 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
7912 annotate_constant_pool_refs (&PATTERN (insn
));
7913 df_insn_rescan (insn
);
7916 pop_topmost_sequence ();
7918 /* Choose best register to use for temp use within prologue.
7919 See below for why TPF must use the register 1. */
7921 if (!has_hard_reg_initial_val (Pmode
, RETURN_REGNUM
)
7922 && !current_function_is_leaf
7923 && !TARGET_TPF_PROFILING
)
7924 temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
7926 temp_reg
= gen_rtx_REG (Pmode
, 1);
7928 /* Save call saved gprs. */
7929 if (cfun_frame_layout
.first_save_gpr
!= -1)
7931 insn
= save_gprs (stack_pointer_rtx
,
7932 cfun_frame_layout
.gprs_offset
+
7933 UNITS_PER_LONG
* (cfun_frame_layout
.first_save_gpr
7934 - cfun_frame_layout
.first_save_gpr_slot
),
7935 cfun_frame_layout
.first_save_gpr
,
7936 cfun_frame_layout
.last_save_gpr
);
7940 /* Dummy insn to mark literal pool slot. */
7942 if (cfun
->machine
->base_reg
)
7943 emit_insn (gen_main_pool (cfun
->machine
->base_reg
));
7945 offset
= cfun_frame_layout
.f0_offset
;
7947 /* Save f0 and f2. */
7948 for (i
= 0; i
< 2; i
++)
7950 if (cfun_fpr_bit_p (i
))
7952 save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
7955 else if (!TARGET_PACKED_STACK
)
7959 /* Save f4 and f6. */
7960 offset
= cfun_frame_layout
.f4_offset
;
7961 for (i
= 2; i
< 4; i
++)
7963 if (cfun_fpr_bit_p (i
))
7965 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
7968 /* If f4 and f6 are call clobbered they are saved due to stdargs and
7969 therefore are not frame related. */
7970 if (!call_really_used_regs
[i
+ 16])
7971 RTX_FRAME_RELATED_P (insn
) = 1;
7973 else if (!TARGET_PACKED_STACK
)
7977 if (TARGET_PACKED_STACK
7978 && cfun_save_high_fprs_p
7979 && cfun_frame_layout
.f8_offset
+ cfun_frame_layout
.high_fprs
* 8 > 0)
7981 offset
= (cfun_frame_layout
.f8_offset
7982 + (cfun_frame_layout
.high_fprs
- 1) * 8);
7984 for (i
= 15; i
> 7 && offset
>= 0; i
--)
7985 if (cfun_fpr_bit_p (i
))
7987 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
7989 RTX_FRAME_RELATED_P (insn
) = 1;
7992 if (offset
>= cfun_frame_layout
.f8_offset
)
7996 if (!TARGET_PACKED_STACK
)
7997 next_fpr
= cfun_save_high_fprs_p
? 31 : 0;
7999 if (flag_stack_usage
)
8000 current_function_static_stack_size
= cfun_frame_layout
.frame_size
;
8002 /* Decrement stack pointer. */
8004 if (cfun_frame_layout
.frame_size
> 0)
8006 rtx frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8009 if (s390_stack_size
)
8011 HOST_WIDE_INT stack_guard
;
8013 if (s390_stack_guard
)
8014 stack_guard
= s390_stack_guard
;
8017 /* If no value for stack guard is provided the smallest power of 2
8018 larger than the current frame size is chosen. */
8020 while (stack_guard
< cfun_frame_layout
.frame_size
)
8024 if (cfun_frame_layout
.frame_size
>= s390_stack_size
)
8026 warning (0, "frame size of function %qs is "
8027 HOST_WIDE_INT_PRINT_DEC
8028 " bytes exceeding user provided stack limit of "
8029 HOST_WIDE_INT_PRINT_DEC
" bytes. "
8030 "An unconditional trap is added.",
8031 current_function_name(), cfun_frame_layout
.frame_size
,
8033 emit_insn (gen_trap ());
8037 /* stack_guard has to be smaller than s390_stack_size.
8038 Otherwise we would emit an AND with zero which would
8039 not match the test under mask pattern. */
8040 if (stack_guard
>= s390_stack_size
)
8042 warning (0, "frame size of function %qs is "
8043 HOST_WIDE_INT_PRINT_DEC
8044 " bytes which is more than half the stack size. "
8045 "The dynamic check would not be reliable. "
8046 "No check emitted for this function.",
8047 current_function_name(),
8048 cfun_frame_layout
.frame_size
);
8052 HOST_WIDE_INT stack_check_mask
= ((s390_stack_size
- 1)
8053 & ~(stack_guard
- 1));
8055 rtx t
= gen_rtx_AND (Pmode
, stack_pointer_rtx
,
8056 GEN_INT (stack_check_mask
));
8058 emit_insn (gen_ctrapdi4 (gen_rtx_EQ (VOIDmode
,
8060 t
, const0_rtx
, const0_rtx
));
8062 emit_insn (gen_ctrapsi4 (gen_rtx_EQ (VOIDmode
,
8064 t
, const0_rtx
, const0_rtx
));
8069 if (s390_warn_framesize
> 0
8070 && cfun_frame_layout
.frame_size
>= s390_warn_framesize
)
8071 warning (0, "frame size of %qs is " HOST_WIDE_INT_PRINT_DEC
" bytes",
8072 current_function_name (), cfun_frame_layout
.frame_size
);
8074 if (s390_warn_dynamicstack_p
&& cfun
->calls_alloca
)
8075 warning (0, "%qs uses dynamic stack allocation", current_function_name ());
8077 /* Save incoming stack pointer into temp reg. */
8078 if (TARGET_BACKCHAIN
|| next_fpr
)
8079 insn
= emit_insn (gen_move_insn (temp_reg
, stack_pointer_rtx
));
8081 /* Subtract frame size from stack pointer. */
8083 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8085 insn
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8086 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8088 insn
= emit_insn (insn
);
8092 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8093 frame_off
= force_const_mem (Pmode
, frame_off
);
8095 insn
= emit_insn (gen_add2_insn (stack_pointer_rtx
, frame_off
));
8096 annotate_constant_pool_refs (&PATTERN (insn
));
8099 RTX_FRAME_RELATED_P (insn
) = 1;
8100 real_frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8101 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8102 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8103 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8106 /* Set backchain. */
8108 if (TARGET_BACKCHAIN
)
8110 if (cfun_frame_layout
.backchain_offset
)
8111 addr
= gen_rtx_MEM (Pmode
,
8112 plus_constant (stack_pointer_rtx
,
8113 cfun_frame_layout
.backchain_offset
));
8115 addr
= gen_rtx_MEM (Pmode
, stack_pointer_rtx
);
8116 set_mem_alias_set (addr
, get_frame_alias_set ());
8117 insn
= emit_insn (gen_move_insn (addr
, temp_reg
));
8120 /* If we support non-call exceptions (e.g. for Java),
8121 we need to make sure the backchain pointer is set up
8122 before any possibly trapping memory access. */
8123 if (TARGET_BACKCHAIN
&& cfun
->can_throw_non_call_exceptions
)
8125 addr
= gen_rtx_MEM (BLKmode
, gen_rtx_SCRATCH (VOIDmode
));
8126 emit_clobber (addr
);
8130 /* Save fprs 8 - 15 (64 bit ABI). */
8132 if (cfun_save_high_fprs_p
&& next_fpr
)
8134 /* If the stack might be accessed through a different register
8135 we have to make sure that the stack pointer decrement is not
8136 moved below the use of the stack slots. */
8137 s390_emit_stack_tie ();
8139 insn
= emit_insn (gen_add2_insn (temp_reg
,
8140 GEN_INT (cfun_frame_layout
.f8_offset
)));
8144 for (i
= 24; i
<= next_fpr
; i
++)
8145 if (cfun_fpr_bit_p (i
- 16))
8147 rtx addr
= plus_constant (stack_pointer_rtx
,
8148 cfun_frame_layout
.frame_size
8149 + cfun_frame_layout
.f8_offset
8152 insn
= save_fpr (temp_reg
, offset
, i
);
8154 RTX_FRAME_RELATED_P (insn
) = 1;
8155 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8156 gen_rtx_SET (VOIDmode
,
8157 gen_rtx_MEM (DFmode
, addr
),
8158 gen_rtx_REG (DFmode
, i
)));
8162 /* Set frame pointer, if needed. */
8164 if (frame_pointer_needed
)
8166 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
8167 RTX_FRAME_RELATED_P (insn
) = 1;
8170 /* Set up got pointer, if needed. */
8172 if (flag_pic
&& df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
8174 rtx insns
= s390_load_got ();
8176 for (insn
= insns
; insn
; insn
= NEXT_INSN (insn
))
8177 annotate_constant_pool_refs (&PATTERN (insn
));
8182 if (TARGET_TPF_PROFILING
)
8184 /* Generate a BAS instruction to serve as a function
8185 entry intercept to facilitate the use of tracing
8186 algorithms located at the branch target. */
8187 emit_insn (gen_prologue_tpf ());
8189 /* Emit a blockage here so that all code
8190 lies between the profiling mechanisms. */
8191 emit_insn (gen_blockage ());
8195 /* Expand the epilogue into a bunch of separate insns. */
8198 s390_emit_epilogue (bool sibcall
)
8200 rtx frame_pointer
, return_reg
, cfa_restores
= NULL_RTX
;
8201 int area_bottom
, area_top
, offset
= 0;
8206 if (TARGET_TPF_PROFILING
)
8209 /* Generate a BAS instruction to serve as a function
8210 entry intercept to facilitate the use of tracing
8211 algorithms located at the branch target. */
8213 /* Emit a blockage here so that all code
8214 lies between the profiling mechanisms. */
8215 emit_insn (gen_blockage ());
8217 emit_insn (gen_epilogue_tpf ());
8220 /* Check whether to use frame or stack pointer for restore. */
8222 frame_pointer
= (frame_pointer_needed
8223 ? hard_frame_pointer_rtx
: stack_pointer_rtx
);
8225 s390_frame_area (&area_bottom
, &area_top
);
8227 /* Check whether we can access the register save area.
8228 If not, increment the frame pointer as required. */
8230 if (area_top
<= area_bottom
)
8232 /* Nothing to restore. */
8234 else if (DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_bottom
)
8235 && DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_top
- 1))
8237 /* Area is in range. */
8238 offset
= cfun_frame_layout
.frame_size
;
8242 rtx insn
, frame_off
, cfa
;
8244 offset
= area_bottom
< 0 ? -area_bottom
: 0;
8245 frame_off
= GEN_INT (cfun_frame_layout
.frame_size
- offset
);
8247 cfa
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8248 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8249 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8251 insn
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8252 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8253 insn
= emit_insn (insn
);
8257 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8258 frame_off
= force_const_mem (Pmode
, frame_off
);
8260 insn
= emit_insn (gen_add2_insn (frame_pointer
, frame_off
));
8261 annotate_constant_pool_refs (&PATTERN (insn
));
8263 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, cfa
);
8264 RTX_FRAME_RELATED_P (insn
) = 1;
8267 /* Restore call saved fprs. */
8271 if (cfun_save_high_fprs_p
)
8273 next_offset
= cfun_frame_layout
.f8_offset
;
8274 for (i
= 24; i
< 32; i
++)
8276 if (cfun_fpr_bit_p (i
- 16))
8278 restore_fpr (frame_pointer
,
8279 offset
+ next_offset
, i
);
8281 = alloc_reg_note (REG_CFA_RESTORE
,
8282 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8291 next_offset
= cfun_frame_layout
.f4_offset
;
8292 for (i
= 18; i
< 20; i
++)
8294 if (cfun_fpr_bit_p (i
- 16))
8296 restore_fpr (frame_pointer
,
8297 offset
+ next_offset
, i
);
8299 = alloc_reg_note (REG_CFA_RESTORE
,
8300 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8303 else if (!TARGET_PACKED_STACK
)
8309 /* Return register. */
8311 return_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8313 /* Restore call saved gprs. */
8315 if (cfun_frame_layout
.first_restore_gpr
!= -1)
8320 /* Check for global register and save them
8321 to stack location from where they get restored. */
8323 for (i
= cfun_frame_layout
.first_restore_gpr
;
8324 i
<= cfun_frame_layout
.last_restore_gpr
;
8327 if (global_not_special_regno_p (i
))
8329 addr
= plus_constant (frame_pointer
,
8330 offset
+ cfun_frame_layout
.gprs_offset
8331 + (i
- cfun_frame_layout
.first_save_gpr_slot
)
8333 addr
= gen_rtx_MEM (Pmode
, addr
);
8334 set_mem_alias_set (addr
, get_frame_alias_set ());
8335 emit_move_insn (addr
, gen_rtx_REG (Pmode
, i
));
8339 = alloc_reg_note (REG_CFA_RESTORE
,
8340 gen_rtx_REG (Pmode
, i
), cfa_restores
);
8345 /* Fetch return address from stack before load multiple,
8346 this will do good for scheduling. */
8348 if (cfun_frame_layout
.save_return_addr_p
8349 || (cfun_frame_layout
.first_restore_gpr
< BASE_REGNUM
8350 && cfun_frame_layout
.last_restore_gpr
> RETURN_REGNUM
))
8352 int return_regnum
= find_unused_clobbered_reg();
8355 return_reg
= gen_rtx_REG (Pmode
, return_regnum
);
8357 addr
= plus_constant (frame_pointer
,
8358 offset
+ cfun_frame_layout
.gprs_offset
8360 - cfun_frame_layout
.first_save_gpr_slot
)
8362 addr
= gen_rtx_MEM (Pmode
, addr
);
8363 set_mem_alias_set (addr
, get_frame_alias_set ());
8364 emit_move_insn (return_reg
, addr
);
8368 insn
= restore_gprs (frame_pointer
,
8369 offset
+ cfun_frame_layout
.gprs_offset
8370 + (cfun_frame_layout
.first_restore_gpr
8371 - cfun_frame_layout
.first_save_gpr_slot
)
8373 cfun_frame_layout
.first_restore_gpr
,
8374 cfun_frame_layout
.last_restore_gpr
);
8375 insn
= emit_insn (insn
);
8376 REG_NOTES (insn
) = cfa_restores
;
8377 add_reg_note (insn
, REG_CFA_DEF_CFA
,
8378 plus_constant (stack_pointer_rtx
, STACK_POINTER_OFFSET
));
8379 RTX_FRAME_RELATED_P (insn
) = 1;
8385 /* Return to caller. */
8387 p
= rtvec_alloc (2);
8389 RTVEC_ELT (p
, 0) = gen_rtx_RETURN (VOIDmode
);
8390 RTVEC_ELT (p
, 1) = gen_rtx_USE (VOIDmode
, return_reg
);
8391 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
8396 /* Return the size in bytes of a function argument of
8397 type TYPE and/or mode MODE. At least one of TYPE or
8398 MODE must be specified. */
8401 s390_function_arg_size (enum machine_mode mode
, const_tree type
)
8404 return int_size_in_bytes (type
);
8406 /* No type info available for some library calls ... */
8407 if (mode
!= BLKmode
)
8408 return GET_MODE_SIZE (mode
);
8410 /* If we have neither type nor mode, abort */
8414 /* Return true if a function argument of type TYPE and mode MODE
8415 is to be passed in a floating-point register, if available. */
8418 s390_function_arg_float (enum machine_mode mode
, const_tree type
)
8420 int size
= s390_function_arg_size (mode
, type
);
8424 /* Soft-float changes the ABI: no floating-point registers are used. */
8425 if (TARGET_SOFT_FLOAT
)
8428 /* No type info available for some library calls ... */
8430 return mode
== SFmode
|| mode
== DFmode
|| mode
== SDmode
|| mode
== DDmode
;
8432 /* The ABI says that record types with a single member are treated
8433 just like that member would be. */
8434 while (TREE_CODE (type
) == RECORD_TYPE
)
8436 tree field
, single
= NULL_TREE
;
8438 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
8440 if (TREE_CODE (field
) != FIELD_DECL
)
8443 if (single
== NULL_TREE
)
8444 single
= TREE_TYPE (field
);
8449 if (single
== NULL_TREE
)
8455 return TREE_CODE (type
) == REAL_TYPE
;
8458 /* Return true if a function argument of type TYPE and mode MODE
8459 is to be passed in an integer register, or a pair of integer
8460 registers, if available. */
8463 s390_function_arg_integer (enum machine_mode mode
, const_tree type
)
8465 int size
= s390_function_arg_size (mode
, type
);
8469 /* No type info available for some library calls ... */
8471 return GET_MODE_CLASS (mode
) == MODE_INT
8472 || (TARGET_SOFT_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
));
8474 /* We accept small integral (and similar) types. */
8475 if (INTEGRAL_TYPE_P (type
)
8476 || POINTER_TYPE_P (type
)
8477 || TREE_CODE (type
) == NULLPTR_TYPE
8478 || TREE_CODE (type
) == OFFSET_TYPE
8479 || (TARGET_SOFT_FLOAT
&& TREE_CODE (type
) == REAL_TYPE
))
8482 /* We also accept structs of size 1, 2, 4, 8 that are not
8483 passed in floating-point registers. */
8484 if (AGGREGATE_TYPE_P (type
)
8485 && exact_log2 (size
) >= 0
8486 && !s390_function_arg_float (mode
, type
))
8492 /* Return 1 if a function argument of type TYPE and mode MODE
8493 is to be passed by reference. The ABI specifies that only
8494 structures of size 1, 2, 4, or 8 bytes are passed by value,
8495 all other structures (and complex numbers) are passed by
8499 s390_pass_by_reference (CUMULATIVE_ARGS
*ca ATTRIBUTE_UNUSED
,
8500 enum machine_mode mode
, const_tree type
,
8501 bool named ATTRIBUTE_UNUSED
)
8503 int size
= s390_function_arg_size (mode
, type
);
8509 if (AGGREGATE_TYPE_P (type
) && exact_log2 (size
) < 0)
8512 if (TREE_CODE (type
) == COMPLEX_TYPE
8513 || TREE_CODE (type
) == VECTOR_TYPE
)
8520 /* Update the data in CUM to advance over an argument of mode MODE and
8521 data type TYPE. (TYPE is null for libcalls where that information
8522 may not be available.). The boolean NAMED specifies whether the
8523 argument is a named argument (as opposed to an unnamed argument
8524 matching an ellipsis). */
8527 s390_function_arg_advance (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
8528 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8530 if (s390_function_arg_float (mode
, type
))
8534 else if (s390_function_arg_integer (mode
, type
))
8536 int size
= s390_function_arg_size (mode
, type
);
8537 cum
->gprs
+= ((size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
);
8543 /* Define where to put the arguments to a function.
8544 Value is zero to push the argument on the stack,
8545 or a hard register in which to store the argument.
8547 MODE is the argument's machine mode.
8548 TYPE is the data type of the argument (as a tree).
8549 This is null for libcalls where that information may
8551 CUM is a variable of type CUMULATIVE_ARGS which gives info about
8552 the preceding args and about the function being called.
8553 NAMED is nonzero if this argument is a named parameter
8554 (otherwise it is an extra parameter matching an ellipsis).
8556 On S/390, we use general purpose registers 2 through 6 to
8557 pass integer, pointer, and certain structure arguments, and
8558 floating point registers 0 and 2 (0, 2, 4, and 6 on 64-bit)
8559 to pass floating point arguments. All remaining arguments
8560 are pushed to the stack. */
8563 s390_function_arg (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
8564 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8566 if (s390_function_arg_float (mode
, type
))
8568 if (cum
->fprs
+ 1 > FP_ARG_NUM_REG
)
8571 return gen_rtx_REG (mode
, cum
->fprs
+ 16);
8573 else if (s390_function_arg_integer (mode
, type
))
8575 int size
= s390_function_arg_size (mode
, type
);
8576 int n_gprs
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8578 if (cum
->gprs
+ n_gprs
> GP_ARG_NUM_REG
)
8580 else if (n_gprs
== 1 || UNITS_PER_WORD
== UNITS_PER_LONG
)
8581 return gen_rtx_REG (mode
, cum
->gprs
+ 2);
8582 else if (n_gprs
== 2)
8584 rtvec p
= rtvec_alloc (2);
8587 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 2),
8590 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 3),
8593 return gen_rtx_PARALLEL (mode
, p
);
8597 /* After the real arguments, expand_call calls us once again
8598 with a void_type_node type. Whatever we return here is
8599 passed as operand 2 to the call expanders.
8601 We don't need this feature ... */
8602 else if (type
== void_type_node
)
8608 /* Return true if return values of type TYPE should be returned
8609 in a memory buffer whose address is passed by the caller as
8610 hidden first argument. */
8613 s390_return_in_memory (const_tree type
, const_tree fundecl ATTRIBUTE_UNUSED
)
8615 /* We accept small integral (and similar) types. */
8616 if (INTEGRAL_TYPE_P (type
)
8617 || POINTER_TYPE_P (type
)
8618 || TREE_CODE (type
) == OFFSET_TYPE
8619 || TREE_CODE (type
) == REAL_TYPE
)
8620 return int_size_in_bytes (type
) > 8;
8622 /* Aggregates and similar constructs are always returned
8624 if (AGGREGATE_TYPE_P (type
)
8625 || TREE_CODE (type
) == COMPLEX_TYPE
8626 || TREE_CODE (type
) == VECTOR_TYPE
)
8629 /* ??? We get called on all sorts of random stuff from
8630 aggregate_value_p. We can't abort, but it's not clear
8631 what's safe to return. Pretend it's a struct I guess. */
8635 /* Function arguments and return values are promoted to word size. */
8637 static enum machine_mode
8638 s390_promote_function_mode (const_tree type
, enum machine_mode mode
,
8640 const_tree fntype ATTRIBUTE_UNUSED
,
8641 int for_return ATTRIBUTE_UNUSED
)
8643 if (INTEGRAL_MODE_P (mode
)
8644 && GET_MODE_SIZE (mode
) < UNITS_PER_LONG
)
8646 if (POINTER_TYPE_P (type
))
8647 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
8654 /* Define where to return a (scalar) value of type TYPE.
8655 If TYPE is null, define where to return a (scalar)
8656 value of mode MODE from a libcall. */
8659 s390_function_value (const_tree type
, const_tree fn
, enum machine_mode mode
)
8663 int unsignedp
= TYPE_UNSIGNED (type
);
8664 mode
= promote_function_mode (type
, TYPE_MODE (type
), &unsignedp
, fn
, 1);
8667 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
|| SCALAR_FLOAT_MODE_P (mode
));
8668 gcc_assert (GET_MODE_SIZE (mode
) <= 8);
8670 if (TARGET_HARD_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
))
8671 return gen_rtx_REG (mode
, 16);
8672 else if (GET_MODE_SIZE (mode
) <= UNITS_PER_LONG
8673 || UNITS_PER_LONG
== UNITS_PER_WORD
)
8674 return gen_rtx_REG (mode
, 2);
8675 else if (GET_MODE_SIZE (mode
) == 2 * UNITS_PER_LONG
)
8677 rtvec p
= rtvec_alloc (2);
8680 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 2), const0_rtx
);
8682 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 3), GEN_INT (4));
8684 return gen_rtx_PARALLEL (mode
, p
);
8691 /* Create and return the va_list datatype.
8693 On S/390, va_list is an array type equivalent to
8695 typedef struct __va_list_tag
8699 void *__overflow_arg_area;
8700 void *__reg_save_area;
8703 where __gpr and __fpr hold the number of general purpose
8704 or floating point arguments used up to now, respectively,
8705 __overflow_arg_area points to the stack location of the
8706 next argument passed on the stack, and __reg_save_area
8707 always points to the start of the register area in the
8708 call frame of the current function. The function prologue
8709 saves all registers used for argument passing into this
8710 area if the function uses variable arguments. */
8713 s390_build_builtin_va_list (void)
8715 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
8717 record
= lang_hooks
.types
.make_type (RECORD_TYPE
);
8720 build_decl (BUILTINS_LOCATION
,
8721 TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
8723 f_gpr
= build_decl (BUILTINS_LOCATION
,
8724 FIELD_DECL
, get_identifier ("__gpr"),
8725 long_integer_type_node
);
8726 f_fpr
= build_decl (BUILTINS_LOCATION
,
8727 FIELD_DECL
, get_identifier ("__fpr"),
8728 long_integer_type_node
);
8729 f_ovf
= build_decl (BUILTINS_LOCATION
,
8730 FIELD_DECL
, get_identifier ("__overflow_arg_area"),
8732 f_sav
= build_decl (BUILTINS_LOCATION
,
8733 FIELD_DECL
, get_identifier ("__reg_save_area"),
8736 va_list_gpr_counter_field
= f_gpr
;
8737 va_list_fpr_counter_field
= f_fpr
;
8739 DECL_FIELD_CONTEXT (f_gpr
) = record
;
8740 DECL_FIELD_CONTEXT (f_fpr
) = record
;
8741 DECL_FIELD_CONTEXT (f_ovf
) = record
;
8742 DECL_FIELD_CONTEXT (f_sav
) = record
;
8744 TYPE_STUB_DECL (record
) = type_decl
;
8745 TYPE_NAME (record
) = type_decl
;
8746 TYPE_FIELDS (record
) = f_gpr
;
8747 DECL_CHAIN (f_gpr
) = f_fpr
;
8748 DECL_CHAIN (f_fpr
) = f_ovf
;
8749 DECL_CHAIN (f_ovf
) = f_sav
;
8751 layout_type (record
);
8753 /* The correct type is an array type of one element. */
8754 return build_array_type (record
, build_index_type (size_zero_node
));
8757 /* Implement va_start by filling the va_list structure VALIST.
8758 STDARG_P is always true, and ignored.
8759 NEXTARG points to the first anonymous stack argument.
8761 The following global variables are used to initialize
8762 the va_list structure:
8765 holds number of gprs and fprs used for named arguments.
8766 crtl->args.arg_offset_rtx:
8767 holds the offset of the first anonymous stack argument
8768 (relative to the virtual arg pointer). */
8771 s390_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
8773 HOST_WIDE_INT n_gpr
, n_fpr
;
8775 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8776 tree gpr
, fpr
, ovf
, sav
, t
;
8778 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8779 f_fpr
= DECL_CHAIN (f_gpr
);
8780 f_ovf
= DECL_CHAIN (f_fpr
);
8781 f_sav
= DECL_CHAIN (f_ovf
);
8783 valist
= build_simple_mem_ref (valist
);
8784 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8785 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8786 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8787 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8789 /* Count number of gp and fp argument registers used. */
8791 n_gpr
= crtl
->args
.info
.gprs
;
8792 n_fpr
= crtl
->args
.info
.fprs
;
8794 if (cfun
->va_list_gpr_size
)
8796 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gpr
), gpr
,
8797 build_int_cst (NULL_TREE
, n_gpr
));
8798 TREE_SIDE_EFFECTS (t
) = 1;
8799 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8802 if (cfun
->va_list_fpr_size
)
8804 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fpr
), fpr
,
8805 build_int_cst (NULL_TREE
, n_fpr
));
8806 TREE_SIDE_EFFECTS (t
) = 1;
8807 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8810 /* Find the overflow area. */
8811 if (n_gpr
+ cfun
->va_list_gpr_size
> GP_ARG_NUM_REG
8812 || n_fpr
+ cfun
->va_list_fpr_size
> FP_ARG_NUM_REG
)
8814 t
= make_tree (TREE_TYPE (ovf
), virtual_incoming_args_rtx
);
8816 off
= INTVAL (crtl
->args
.arg_offset_rtx
);
8817 off
= off
< 0 ? 0 : off
;
8818 if (TARGET_DEBUG_ARG
)
8819 fprintf (stderr
, "va_start: n_gpr = %d, n_fpr = %d off %d\n",
8820 (int)n_gpr
, (int)n_fpr
, off
);
8822 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (ovf
), t
, size_int (off
));
8824 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovf
), ovf
, t
);
8825 TREE_SIDE_EFFECTS (t
) = 1;
8826 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8829 /* Find the register save area. */
8830 if ((cfun
->va_list_gpr_size
&& n_gpr
< GP_ARG_NUM_REG
)
8831 || (cfun
->va_list_fpr_size
&& n_fpr
< FP_ARG_NUM_REG
))
8833 t
= make_tree (TREE_TYPE (sav
), return_address_pointer_rtx
);
8834 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (sav
), t
,
8835 size_int (-RETURN_REGNUM
* UNITS_PER_LONG
));
8837 t
= build2 (MODIFY_EXPR
, TREE_TYPE (sav
), sav
, t
);
8838 TREE_SIDE_EFFECTS (t
) = 1;
8839 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8843 /* Implement va_arg by updating the va_list structure
8844 VALIST as required to retrieve an argument of type
8845 TYPE, and returning that argument.
8847 Generates code equivalent to:
8849 if (integral value) {
8850 if (size <= 4 && args.gpr < 5 ||
8851 size > 4 && args.gpr < 4 )
8852 ret = args.reg_save_area[args.gpr+8]
8854 ret = *args.overflow_arg_area++;
8855 } else if (float value) {
8857 ret = args.reg_save_area[args.fpr+64]
8859 ret = *args.overflow_arg_area++;
8860 } else if (aggregate value) {
8862 ret = *args.reg_save_area[args.gpr]
8864 ret = **args.overflow_arg_area++;
8868 s390_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
8869 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
8871 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8872 tree gpr
, fpr
, ovf
, sav
, reg
, t
, u
;
8873 int indirect_p
, size
, n_reg
, sav_ofs
, sav_scale
, max_reg
;
8874 tree lab_false
, lab_over
, addr
;
8876 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8877 f_fpr
= DECL_CHAIN (f_gpr
);
8878 f_ovf
= DECL_CHAIN (f_fpr
);
8879 f_sav
= DECL_CHAIN (f_ovf
);
8881 valist
= build_va_arg_indirect_ref (valist
);
8882 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8883 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8884 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8886 /* The tree for args* cannot be shared between gpr/fpr and ovf since
8887 both appear on a lhs. */
8888 valist
= unshare_expr (valist
);
8889 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8891 size
= int_size_in_bytes (type
);
8893 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
8895 if (TARGET_DEBUG_ARG
)
8897 fprintf (stderr
, "va_arg: aggregate type");
8901 /* Aggregates are passed by reference. */
8906 /* kernel stack layout on 31 bit: It is assumed here that no padding
8907 will be added by s390_frame_info because for va_args always an even
8908 number of gprs has to be saved r15-r2 = 14 regs. */
8909 sav_ofs
= 2 * UNITS_PER_LONG
;
8910 sav_scale
= UNITS_PER_LONG
;
8911 size
= UNITS_PER_LONG
;
8912 max_reg
= GP_ARG_NUM_REG
- n_reg
;
8914 else if (s390_function_arg_float (TYPE_MODE (type
), type
))
8916 if (TARGET_DEBUG_ARG
)
8918 fprintf (stderr
, "va_arg: float type");
8922 /* FP args go in FP registers, if present. */
8926 sav_ofs
= 16 * UNITS_PER_LONG
;
8928 max_reg
= FP_ARG_NUM_REG
- n_reg
;
8932 if (TARGET_DEBUG_ARG
)
8934 fprintf (stderr
, "va_arg: other type");
8938 /* Otherwise into GP registers. */
8941 n_reg
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8943 /* kernel stack layout on 31 bit: It is assumed here that no padding
8944 will be added by s390_frame_info because for va_args always an even
8945 number of gprs has to be saved r15-r2 = 14 regs. */
8946 sav_ofs
= 2 * UNITS_PER_LONG
;
8948 if (size
< UNITS_PER_LONG
)
8949 sav_ofs
+= UNITS_PER_LONG
- size
;
8951 sav_scale
= UNITS_PER_LONG
;
8952 max_reg
= GP_ARG_NUM_REG
- n_reg
;
8955 /* Pull the value out of the saved registers ... */
8957 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
8958 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
8959 addr
= create_tmp_var (ptr_type_node
, "addr");
8961 t
= fold_convert (TREE_TYPE (reg
), size_int (max_reg
));
8962 t
= build2 (GT_EXPR
, boolean_type_node
, reg
, t
);
8963 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
8964 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
8965 gimplify_and_add (t
, pre_p
);
8967 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, sav
,
8968 size_int (sav_ofs
));
8969 u
= build2 (MULT_EXPR
, TREE_TYPE (reg
), reg
,
8970 fold_convert (TREE_TYPE (reg
), size_int (sav_scale
)));
8971 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, t
, fold_convert (sizetype
, u
));
8973 gimplify_assign (addr
, t
, pre_p
);
8975 gimple_seq_add_stmt (pre_p
, gimple_build_goto (lab_over
));
8977 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_false
));
8980 /* ... Otherwise out of the overflow area. */
8983 if (size
< UNITS_PER_LONG
)
8984 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, t
,
8985 size_int (UNITS_PER_LONG
- size
));
8987 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
8989 gimplify_assign (addr
, t
, pre_p
);
8991 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, t
,
8993 gimplify_assign (ovf
, t
, pre_p
);
8995 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_over
));
8998 /* Increment register save count. */
9000 u
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (reg
), reg
,
9001 fold_convert (TREE_TYPE (reg
), size_int (n_reg
)));
9002 gimplify_and_add (u
, pre_p
);
9006 t
= build_pointer_type_for_mode (build_pointer_type (type
),
9008 addr
= fold_convert (t
, addr
);
9009 addr
= build_va_arg_indirect_ref (addr
);
9013 t
= build_pointer_type_for_mode (type
, ptr_mode
, true);
9014 addr
= fold_convert (t
, addr
);
9017 return build_va_arg_indirect_ref (addr
);
9025 S390_BUILTIN_THREAD_POINTER
,
9026 S390_BUILTIN_SET_THREAD_POINTER
,
9031 static enum insn_code
const code_for_builtin_64
[S390_BUILTIN_max
] = {
9036 static enum insn_code
const code_for_builtin_31
[S390_BUILTIN_max
] = {
9042 s390_init_builtins (void)
9046 ftype
= build_function_type (ptr_type_node
, void_list_node
);
9047 add_builtin_function ("__builtin_thread_pointer", ftype
,
9048 S390_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
9051 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
9052 add_builtin_function ("__builtin_set_thread_pointer", ftype
,
9053 S390_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
9057 /* Expand an expression EXP that calls a built-in function,
9058 with result going to TARGET if that's convenient
9059 (and in mode MODE if that's convenient).
9060 SUBTARGET may be used as the target for computing one of EXP's operands.
9061 IGNORE is nonzero if the value is to be ignored. */
9064 s390_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
9065 enum machine_mode mode ATTRIBUTE_UNUSED
,
9066 int ignore ATTRIBUTE_UNUSED
)
9070 enum insn_code
const *code_for_builtin
=
9071 TARGET_64BIT
? code_for_builtin_64
: code_for_builtin_31
;
9073 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
9074 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
9075 enum insn_code icode
;
9076 rtx op
[MAX_ARGS
], pat
;
9080 call_expr_arg_iterator iter
;
9082 if (fcode
>= S390_BUILTIN_max
)
9083 internal_error ("bad builtin fcode");
9084 icode
= code_for_builtin
[fcode
];
9086 internal_error ("bad builtin fcode");
9088 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
9091 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
9093 const struct insn_operand_data
*insn_op
;
9095 if (arg
== error_mark_node
)
9097 if (arity
> MAX_ARGS
)
9100 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
9102 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
9104 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
9105 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
9111 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
9113 || GET_MODE (target
) != tmode
9114 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
9115 target
= gen_reg_rtx (tmode
);
9121 pat
= GEN_FCN (icode
) (target
);
9125 pat
= GEN_FCN (icode
) (target
, op
[0]);
9127 pat
= GEN_FCN (icode
) (op
[0]);
9130 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
9146 /* Output assembly code for the trampoline template to
9149 On S/390, we use gpr 1 internally in the trampoline code;
9150 gpr 0 is used to hold the static chain. */
9153 s390_asm_trampoline_template (FILE *file
)
9156 op
[0] = gen_rtx_REG (Pmode
, 0);
9157 op
[1] = gen_rtx_REG (Pmode
, 1);
9161 output_asm_insn ("basr\t%1,0", op
);
9162 output_asm_insn ("lmg\t%0,%1,14(%1)", op
);
9163 output_asm_insn ("br\t%1", op
);
9164 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 10));
9168 output_asm_insn ("basr\t%1,0", op
);
9169 output_asm_insn ("lm\t%0,%1,6(%1)", op
);
9170 output_asm_insn ("br\t%1", op
);
9171 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 8));
9175 /* Emit RTL insns to initialize the variable parts of a trampoline.
9176 FNADDR is an RTX for the address of the function's pure code.
9177 CXT is an RTX for the static chain value for the function. */
9180 s390_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9182 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
9185 emit_block_move (m_tramp
, assemble_trampoline_template (),
9186 GEN_INT (2*UNITS_PER_WORD
), BLOCK_OP_NORMAL
);
9188 mem
= adjust_address (m_tramp
, Pmode
, 2*UNITS_PER_WORD
);
9189 emit_move_insn (mem
, cxt
);
9190 mem
= adjust_address (m_tramp
, Pmode
, 3*UNITS_PER_WORD
);
9191 emit_move_insn (mem
, fnaddr
);
9194 /* Output assembler code to FILE to increment profiler label # LABELNO
9195 for profiling a function entry. */
9198 s390_function_profiler (FILE *file
, int labelno
)
9203 ASM_GENERATE_INTERNAL_LABEL (label
, "LP", labelno
);
9205 fprintf (file
, "# function profiler \n");
9207 op
[0] = gen_rtx_REG (Pmode
, RETURN_REGNUM
);
9208 op
[1] = gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
);
9209 op
[1] = gen_rtx_MEM (Pmode
, plus_constant (op
[1], UNITS_PER_LONG
));
9211 op
[2] = gen_rtx_REG (Pmode
, 1);
9212 op
[3] = gen_rtx_SYMBOL_REF (Pmode
, label
);
9213 SYMBOL_REF_FLAGS (op
[3]) = SYMBOL_FLAG_LOCAL
;
9215 op
[4] = gen_rtx_SYMBOL_REF (Pmode
, "_mcount");
9218 op
[4] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[4]), UNSPEC_PLT
);
9219 op
[4] = gen_rtx_CONST (Pmode
, op
[4]);
9224 output_asm_insn ("stg\t%0,%1", op
);
9225 output_asm_insn ("larl\t%2,%3", op
);
9226 output_asm_insn ("brasl\t%0,%4", op
);
9227 output_asm_insn ("lg\t%0,%1", op
);
9231 op
[6] = gen_label_rtx ();
9233 output_asm_insn ("st\t%0,%1", op
);
9234 output_asm_insn ("bras\t%2,%l6", op
);
9235 output_asm_insn (".long\t%4", op
);
9236 output_asm_insn (".long\t%3", op
);
9237 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9238 output_asm_insn ("l\t%0,0(%2)", op
);
9239 output_asm_insn ("l\t%2,4(%2)", op
);
9240 output_asm_insn ("basr\t%0,%0", op
);
9241 output_asm_insn ("l\t%0,%1", op
);
9245 op
[5] = gen_label_rtx ();
9246 op
[6] = gen_label_rtx ();
9248 output_asm_insn ("st\t%0,%1", op
);
9249 output_asm_insn ("bras\t%2,%l6", op
);
9250 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[5]));
9251 output_asm_insn (".long\t%4-%l5", op
);
9252 output_asm_insn (".long\t%3-%l5", op
);
9253 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9254 output_asm_insn ("lr\t%0,%2", op
);
9255 output_asm_insn ("a\t%0,0(%2)", op
);
9256 output_asm_insn ("a\t%2,4(%2)", op
);
9257 output_asm_insn ("basr\t%0,%0", op
);
9258 output_asm_insn ("l\t%0,%1", op
);
9262 /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF
9263 into its SYMBOL_REF_FLAGS. */
9266 s390_encode_section_info (tree decl
, rtx rtl
, int first
)
9268 default_encode_section_info (decl
, rtl
, first
);
9270 if (TREE_CODE (decl
) == VAR_DECL
)
9272 /* If a variable has a forced alignment to < 2 bytes, mark it
9273 with SYMBOL_FLAG_ALIGN1 to prevent it from being used as LARL
9275 if (DECL_USER_ALIGN (decl
) && DECL_ALIGN (decl
) < 16)
9276 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_ALIGN1
;
9277 if (!DECL_SIZE (decl
)
9278 || !DECL_ALIGN (decl
)
9279 || !host_integerp (DECL_SIZE (decl
), 0)
9280 || (DECL_ALIGN (decl
) <= 64
9281 && DECL_ALIGN (decl
) != tree_low_cst (DECL_SIZE (decl
), 0)))
9282 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9285 /* Literal pool references don't have a decl so they are handled
9286 differently here. We rely on the information in the MEM_ALIGN
9287 entry to decide upon natural alignment. */
9289 && GET_CODE (XEXP (rtl
, 0)) == SYMBOL_REF
9290 && TREE_CONSTANT_POOL_ADDRESS_P (XEXP (rtl
, 0))
9291 && (MEM_ALIGN (rtl
) == 0
9292 || GET_MODE_BITSIZE (GET_MODE (rtl
)) == 0
9293 || MEM_ALIGN (rtl
) < GET_MODE_BITSIZE (GET_MODE (rtl
))))
9294 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9297 /* Output thunk to FILE that implements a C++ virtual function call (with
9298 multiple inheritance) to FUNCTION. The thunk adjusts the this pointer
9299 by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment
9300 stored at VCALL_OFFSET in the vtable whose address is located at offset 0
9301 relative to the resulting this pointer. */
9304 s390_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
9305 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
9311 /* Make sure unwind info is emitted for the thunk if needed. */
9312 final_start_function (emit_barrier (), file
, 1);
9314 /* Operand 0 is the target function. */
9315 op
[0] = XEXP (DECL_RTL (function
), 0);
9316 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (op
[0]))
9319 op
[0] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[0]),
9320 TARGET_64BIT
? UNSPEC_PLT
: UNSPEC_GOT
);
9321 op
[0] = gen_rtx_CONST (Pmode
, op
[0]);
9324 /* Operand 1 is the 'this' pointer. */
9325 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
9326 op
[1] = gen_rtx_REG (Pmode
, 3);
9328 op
[1] = gen_rtx_REG (Pmode
, 2);
9330 /* Operand 2 is the delta. */
9331 op
[2] = GEN_INT (delta
);
9333 /* Operand 3 is the vcall_offset. */
9334 op
[3] = GEN_INT (vcall_offset
);
9336 /* Operand 4 is the temporary register. */
9337 op
[4] = gen_rtx_REG (Pmode
, 1);
9339 /* Operands 5 to 8 can be used as labels. */
9345 /* Operand 9 can be used for temporary register. */
9348 /* Generate code. */
9351 /* Setup literal pool pointer if required. */
9352 if ((!DISP_IN_RANGE (delta
)
9353 && !CONST_OK_FOR_K (delta
)
9354 && !CONST_OK_FOR_Os (delta
))
9355 || (!DISP_IN_RANGE (vcall_offset
)
9356 && !CONST_OK_FOR_K (vcall_offset
)
9357 && !CONST_OK_FOR_Os (vcall_offset
)))
9359 op
[5] = gen_label_rtx ();
9360 output_asm_insn ("larl\t%4,%5", op
);
9363 /* Add DELTA to this pointer. */
9366 if (CONST_OK_FOR_J (delta
))
9367 output_asm_insn ("la\t%1,%2(%1)", op
);
9368 else if (DISP_IN_RANGE (delta
))
9369 output_asm_insn ("lay\t%1,%2(%1)", op
);
9370 else if (CONST_OK_FOR_K (delta
))
9371 output_asm_insn ("aghi\t%1,%2", op
);
9372 else if (CONST_OK_FOR_Os (delta
))
9373 output_asm_insn ("agfi\t%1,%2", op
);
9376 op
[6] = gen_label_rtx ();
9377 output_asm_insn ("agf\t%1,%6-%5(%4)", op
);
9381 /* Perform vcall adjustment. */
9384 if (DISP_IN_RANGE (vcall_offset
))
9386 output_asm_insn ("lg\t%4,0(%1)", op
);
9387 output_asm_insn ("ag\t%1,%3(%4)", op
);
9389 else if (CONST_OK_FOR_K (vcall_offset
))
9391 output_asm_insn ("lghi\t%4,%3", op
);
9392 output_asm_insn ("ag\t%4,0(%1)", op
);
9393 output_asm_insn ("ag\t%1,0(%4)", op
);
9395 else if (CONST_OK_FOR_Os (vcall_offset
))
9397 output_asm_insn ("lgfi\t%4,%3", op
);
9398 output_asm_insn ("ag\t%4,0(%1)", op
);
9399 output_asm_insn ("ag\t%1,0(%4)", op
);
9403 op
[7] = gen_label_rtx ();
9404 output_asm_insn ("llgf\t%4,%7-%5(%4)", op
);
9405 output_asm_insn ("ag\t%4,0(%1)", op
);
9406 output_asm_insn ("ag\t%1,0(%4)", op
);
9410 /* Jump to target. */
9411 output_asm_insn ("jg\t%0", op
);
9413 /* Output literal pool if required. */
9416 output_asm_insn (".align\t4", op
);
9417 targetm
.asm_out
.internal_label (file
, "L",
9418 CODE_LABEL_NUMBER (op
[5]));
9422 targetm
.asm_out
.internal_label (file
, "L",
9423 CODE_LABEL_NUMBER (op
[6]));
9424 output_asm_insn (".long\t%2", op
);
9428 targetm
.asm_out
.internal_label (file
, "L",
9429 CODE_LABEL_NUMBER (op
[7]));
9430 output_asm_insn (".long\t%3", op
);
9435 /* Setup base pointer if required. */
9437 || (!DISP_IN_RANGE (delta
)
9438 && !CONST_OK_FOR_K (delta
)
9439 && !CONST_OK_FOR_Os (delta
))
9440 || (!DISP_IN_RANGE (delta
)
9441 && !CONST_OK_FOR_K (vcall_offset
)
9442 && !CONST_OK_FOR_Os (vcall_offset
)))
9444 op
[5] = gen_label_rtx ();
9445 output_asm_insn ("basr\t%4,0", op
);
9446 targetm
.asm_out
.internal_label (file
, "L",
9447 CODE_LABEL_NUMBER (op
[5]));
9450 /* Add DELTA to this pointer. */
9453 if (CONST_OK_FOR_J (delta
))
9454 output_asm_insn ("la\t%1,%2(%1)", op
);
9455 else if (DISP_IN_RANGE (delta
))
9456 output_asm_insn ("lay\t%1,%2(%1)", op
);
9457 else if (CONST_OK_FOR_K (delta
))
9458 output_asm_insn ("ahi\t%1,%2", op
);
9459 else if (CONST_OK_FOR_Os (delta
))
9460 output_asm_insn ("afi\t%1,%2", op
);
9463 op
[6] = gen_label_rtx ();
9464 output_asm_insn ("a\t%1,%6-%5(%4)", op
);
9468 /* Perform vcall adjustment. */
9471 if (CONST_OK_FOR_J (vcall_offset
))
9473 output_asm_insn ("l\t%4,0(%1)", op
);
9474 output_asm_insn ("a\t%1,%3(%4)", op
);
9476 else if (DISP_IN_RANGE (vcall_offset
))
9478 output_asm_insn ("l\t%4,0(%1)", op
);
9479 output_asm_insn ("ay\t%1,%3(%4)", op
);
9481 else if (CONST_OK_FOR_K (vcall_offset
))
9483 output_asm_insn ("lhi\t%4,%3", op
);
9484 output_asm_insn ("a\t%4,0(%1)", op
);
9485 output_asm_insn ("a\t%1,0(%4)", op
);
9487 else if (CONST_OK_FOR_Os (vcall_offset
))
9489 output_asm_insn ("iilf\t%4,%3", op
);
9490 output_asm_insn ("a\t%4,0(%1)", op
);
9491 output_asm_insn ("a\t%1,0(%4)", op
);
9495 op
[7] = gen_label_rtx ();
9496 output_asm_insn ("l\t%4,%7-%5(%4)", op
);
9497 output_asm_insn ("a\t%4,0(%1)", op
);
9498 output_asm_insn ("a\t%1,0(%4)", op
);
9501 /* We had to clobber the base pointer register.
9502 Re-setup the base pointer (with a different base). */
9503 op
[5] = gen_label_rtx ();
9504 output_asm_insn ("basr\t%4,0", op
);
9505 targetm
.asm_out
.internal_label (file
, "L",
9506 CODE_LABEL_NUMBER (op
[5]));
9509 /* Jump to target. */
9510 op
[8] = gen_label_rtx ();
9513 output_asm_insn ("l\t%4,%8-%5(%4)", op
);
9515 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9516 /* We cannot call through .plt, since .plt requires %r12 loaded. */
9517 else if (flag_pic
== 1)
9519 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9520 output_asm_insn ("l\t%4,%0(%4)", op
);
9522 else if (flag_pic
== 2)
9524 op
[9] = gen_rtx_REG (Pmode
, 0);
9525 output_asm_insn ("l\t%9,%8-4-%5(%4)", op
);
9526 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9527 output_asm_insn ("ar\t%4,%9", op
);
9528 output_asm_insn ("l\t%4,0(%4)", op
);
9531 output_asm_insn ("br\t%4", op
);
9533 /* Output literal pool. */
9534 output_asm_insn (".align\t4", op
);
9536 if (nonlocal
&& flag_pic
== 2)
9537 output_asm_insn (".long\t%0", op
);
9540 op
[0] = gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
9541 SYMBOL_REF_FLAGS (op
[0]) = SYMBOL_FLAG_LOCAL
;
9544 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[8]));
9546 output_asm_insn (".long\t%0", op
);
9548 output_asm_insn (".long\t%0-%5", op
);
9552 targetm
.asm_out
.internal_label (file
, "L",
9553 CODE_LABEL_NUMBER (op
[6]));
9554 output_asm_insn (".long\t%2", op
);
9558 targetm
.asm_out
.internal_label (file
, "L",
9559 CODE_LABEL_NUMBER (op
[7]));
9560 output_asm_insn (".long\t%3", op
);
9563 final_end_function ();
9567 s390_valid_pointer_mode (enum machine_mode mode
)
9569 return (mode
== SImode
|| (TARGET_64BIT
&& mode
== DImode
));
9572 /* Checks whether the given CALL_EXPR would use a caller
9573 saved register. This is used to decide whether sibling call
9574 optimization could be performed on the respective function
9578 s390_call_saved_register_used (tree call_expr
)
9580 CUMULATIVE_ARGS cum
;
9582 enum machine_mode mode
;
9587 INIT_CUMULATIVE_ARGS (cum
, NULL
, NULL
, 0, 0);
9589 for (i
= 0; i
< call_expr_nargs (call_expr
); i
++)
9591 parameter
= CALL_EXPR_ARG (call_expr
, i
);
9592 gcc_assert (parameter
);
9594 /* For an undeclared variable passed as parameter we will get
9595 an ERROR_MARK node here. */
9596 if (TREE_CODE (parameter
) == ERROR_MARK
)
9599 type
= TREE_TYPE (parameter
);
9602 mode
= TYPE_MODE (type
);
9605 if (pass_by_reference (&cum
, mode
, type
, true))
9608 type
= build_pointer_type (type
);
9611 parm_rtx
= s390_function_arg (&cum
, mode
, type
, 0);
9613 s390_function_arg_advance (&cum
, mode
, type
, 0);
9618 if (REG_P (parm_rtx
))
9621 reg
< HARD_REGNO_NREGS (REGNO (parm_rtx
), GET_MODE (parm_rtx
));
9623 if (!call_used_regs
[reg
+ REGNO (parm_rtx
)])
9627 if (GET_CODE (parm_rtx
) == PARALLEL
)
9631 for (i
= 0; i
< XVECLEN (parm_rtx
, 0); i
++)
9633 rtx r
= XEXP (XVECEXP (parm_rtx
, 0, i
), 0);
9635 gcc_assert (REG_P (r
));
9638 reg
< HARD_REGNO_NREGS (REGNO (r
), GET_MODE (r
));
9640 if (!call_used_regs
[reg
+ REGNO (r
)])
9649 /* Return true if the given call expression can be
9650 turned into a sibling call.
9651 DECL holds the declaration of the function to be called whereas
9652 EXP is the call expression itself. */
9655 s390_function_ok_for_sibcall (tree decl
, tree exp
)
9657 /* The TPF epilogue uses register 1. */
9658 if (TARGET_TPF_PROFILING
)
9661 /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
9662 which would have to be restored before the sibcall. */
9663 if (!TARGET_64BIT
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
9666 /* Register 6 on s390 is available as an argument register but unfortunately
9667 "caller saved". This makes functions needing this register for arguments
9668 not suitable for sibcalls. */
9669 return !s390_call_saved_register_used (exp
);
9672 /* Return the fixed registers used for condition codes. */
9675 s390_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
9678 *p2
= INVALID_REGNUM
;
9683 /* This function is used by the call expanders of the machine description.
9684 It emits the call insn itself together with the necessary operations
9685 to adjust the target address and returns the emitted insn.
9686 ADDR_LOCATION is the target address rtx
9687 TLS_CALL the location of the thread-local symbol
9688 RESULT_REG the register where the result of the call should be stored
9689 RETADDR_REG the register where the return address should be stored
9690 If this parameter is NULL_RTX the call is considered
9691 to be a sibling call. */
9694 s390_emit_call (rtx addr_location
, rtx tls_call
, rtx result_reg
,
9697 bool plt_call
= false;
9703 /* Direct function calls need special treatment. */
9704 if (GET_CODE (addr_location
) == SYMBOL_REF
)
9706 /* When calling a global routine in PIC mode, we must
9707 replace the symbol itself with the PLT stub. */
9708 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (addr_location
))
9710 if (retaddr_reg
!= NULL_RTX
)
9712 addr_location
= gen_rtx_UNSPEC (Pmode
,
9713 gen_rtvec (1, addr_location
),
9715 addr_location
= gen_rtx_CONST (Pmode
, addr_location
);
9719 /* For -fpic code the PLT entries might use r12 which is
9720 call-saved. Therefore we cannot do a sibcall when
9721 calling directly using a symbol ref. When reaching
9722 this point we decided (in s390_function_ok_for_sibcall)
9723 to do a sibcall for a function pointer but one of the
9724 optimizers was able to get rid of the function pointer
9725 by propagating the symbol ref into the call. This
9726 optimization is illegal for S/390 so we turn the direct
9727 call into a indirect call again. */
9728 addr_location
= force_reg (Pmode
, addr_location
);
9731 /* Unless we can use the bras(l) insn, force the
9732 routine address into a register. */
9733 if (!TARGET_SMALL_EXEC
&& !TARGET_CPU_ZARCH
)
9736 addr_location
= legitimize_pic_address (addr_location
, 0);
9738 addr_location
= force_reg (Pmode
, addr_location
);
9742 /* If it is already an indirect call or the code above moved the
9743 SYMBOL_REF to somewhere else make sure the address can be found in
9745 if (retaddr_reg
== NULL_RTX
9746 && GET_CODE (addr_location
) != SYMBOL_REF
9749 emit_move_insn (gen_rtx_REG (Pmode
, SIBCALL_REGNUM
), addr_location
);
9750 addr_location
= gen_rtx_REG (Pmode
, SIBCALL_REGNUM
);
9753 addr_location
= gen_rtx_MEM (QImode
, addr_location
);
9754 call
= gen_rtx_CALL (VOIDmode
, addr_location
, const0_rtx
);
9756 if (result_reg
!= NULL_RTX
)
9757 call
= gen_rtx_SET (VOIDmode
, result_reg
, call
);
9759 if (retaddr_reg
!= NULL_RTX
)
9761 clobber
= gen_rtx_CLOBBER (VOIDmode
, retaddr_reg
);
9763 if (tls_call
!= NULL_RTX
)
9764 vec
= gen_rtvec (3, call
, clobber
,
9765 gen_rtx_USE (VOIDmode
, tls_call
));
9767 vec
= gen_rtvec (2, call
, clobber
);
9769 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
9772 insn
= emit_call_insn (call
);
9774 /* 31-bit PLT stubs and tls calls use the GOT register implicitly. */
9775 if ((!TARGET_64BIT
&& plt_call
) || tls_call
!= NULL_RTX
)
9777 /* s390_function_ok_for_sibcall should
9778 have denied sibcalls in this case. */
9779 gcc_assert (retaddr_reg
!= NULL_RTX
);
9781 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), pic_offset_table_rtx
);
9786 /* Implement TARGET_CONDITIONAL_REGISTER_USAGE. */
9789 s390_conditional_register_usage (void)
9795 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9796 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9798 if (TARGET_CPU_ZARCH
)
9800 fixed_regs
[BASE_REGNUM
] = 0;
9801 call_used_regs
[BASE_REGNUM
] = 0;
9802 fixed_regs
[RETURN_REGNUM
] = 0;
9803 call_used_regs
[RETURN_REGNUM
] = 0;
9807 for (i
= 24; i
< 32; i
++)
9808 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9812 for (i
= 18; i
< 20; i
++)
9813 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9816 if (TARGET_SOFT_FLOAT
)
9818 for (i
= 16; i
< 32; i
++)
9819 call_used_regs
[i
] = fixed_regs
[i
] = 1;
9823 /* Corresponding function to eh_return expander. */
9825 static GTY(()) rtx s390_tpf_eh_return_symbol
;
9827 s390_emit_tpf_eh_return (rtx target
)
9831 if (!s390_tpf_eh_return_symbol
)
9832 s390_tpf_eh_return_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tpf_eh_return");
9834 reg
= gen_rtx_REG (Pmode
, 2);
9836 emit_move_insn (reg
, target
);
9837 insn
= s390_emit_call (s390_tpf_eh_return_symbol
, NULL_RTX
, reg
,
9838 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
9839 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), reg
);
9841 emit_move_insn (EH_RETURN_HANDLER_RTX
, reg
);
9844 /* Rework the prologue/epilogue to avoid saving/restoring
9845 registers unnecessarily. */
9848 s390_optimize_prologue (void)
9850 rtx insn
, new_insn
, next_insn
;
9852 /* Do a final recompute of the frame-related data. */
9854 s390_update_frame_layout ();
9856 /* If all special registers are in fact used, there's nothing we
9857 can do, so no point in walking the insn list. */
9859 if (cfun_frame_layout
.first_save_gpr
<= BASE_REGNUM
9860 && cfun_frame_layout
.last_save_gpr
>= BASE_REGNUM
9861 && (TARGET_CPU_ZARCH
9862 || (cfun_frame_layout
.first_save_gpr
<= RETURN_REGNUM
9863 && cfun_frame_layout
.last_save_gpr
>= RETURN_REGNUM
)))
9866 /* Search for prologue/epilogue insns and replace them. */
9868 for (insn
= get_insns (); insn
; insn
= next_insn
)
9870 int first
, last
, off
;
9871 rtx set
, base
, offset
;
9873 next_insn
= NEXT_INSN (insn
);
9875 if (GET_CODE (insn
) != INSN
)
9878 if (GET_CODE (PATTERN (insn
)) == PARALLEL
9879 && store_multiple_operation (PATTERN (insn
), VOIDmode
))
9881 set
= XVECEXP (PATTERN (insn
), 0, 0);
9882 first
= REGNO (SET_SRC (set
));
9883 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
9884 offset
= const0_rtx
;
9885 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
9886 off
= INTVAL (offset
);
9888 if (GET_CODE (base
) != REG
|| off
< 0)
9890 if (cfun_frame_layout
.first_save_gpr
!= -1
9891 && (cfun_frame_layout
.first_save_gpr
< first
9892 || cfun_frame_layout
.last_save_gpr
> last
))
9894 if (REGNO (base
) != STACK_POINTER_REGNUM
9895 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
9897 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
9900 if (cfun_frame_layout
.first_save_gpr
!= -1)
9902 new_insn
= save_gprs (base
,
9903 off
+ (cfun_frame_layout
.first_save_gpr
9904 - first
) * UNITS_PER_LONG
,
9905 cfun_frame_layout
.first_save_gpr
,
9906 cfun_frame_layout
.last_save_gpr
);
9907 new_insn
= emit_insn_before (new_insn
, insn
);
9908 INSN_ADDRESSES_NEW (new_insn
, -1);
9915 if (cfun_frame_layout
.first_save_gpr
== -1
9916 && GET_CODE (PATTERN (insn
)) == SET
9917 && GET_CODE (SET_SRC (PATTERN (insn
))) == REG
9918 && (REGNO (SET_SRC (PATTERN (insn
))) == BASE_REGNUM
9919 || (!TARGET_CPU_ZARCH
9920 && REGNO (SET_SRC (PATTERN (insn
))) == RETURN_REGNUM
))
9921 && GET_CODE (SET_DEST (PATTERN (insn
))) == MEM
)
9923 set
= PATTERN (insn
);
9924 first
= REGNO (SET_SRC (set
));
9925 offset
= const0_rtx
;
9926 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
9927 off
= INTVAL (offset
);
9929 if (GET_CODE (base
) != REG
|| off
< 0)
9931 if (REGNO (base
) != STACK_POINTER_REGNUM
9932 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
9939 if (GET_CODE (PATTERN (insn
)) == PARALLEL
9940 && load_multiple_operation (PATTERN (insn
), VOIDmode
))
9942 set
= XVECEXP (PATTERN (insn
), 0, 0);
9943 first
= REGNO (SET_DEST (set
));
9944 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
9945 offset
= const0_rtx
;
9946 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
9947 off
= INTVAL (offset
);
9949 if (GET_CODE (base
) != REG
|| off
< 0)
9951 if (cfun_frame_layout
.first_restore_gpr
!= -1
9952 && (cfun_frame_layout
.first_restore_gpr
< first
9953 || cfun_frame_layout
.last_restore_gpr
> last
))
9955 if (REGNO (base
) != STACK_POINTER_REGNUM
9956 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
9958 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
9961 if (cfun_frame_layout
.first_restore_gpr
!= -1)
9963 new_insn
= restore_gprs (base
,
9964 off
+ (cfun_frame_layout
.first_restore_gpr
9965 - first
) * UNITS_PER_LONG
,
9966 cfun_frame_layout
.first_restore_gpr
,
9967 cfun_frame_layout
.last_restore_gpr
);
9968 new_insn
= emit_insn_before (new_insn
, insn
);
9969 INSN_ADDRESSES_NEW (new_insn
, -1);
9976 if (cfun_frame_layout
.first_restore_gpr
== -1
9977 && GET_CODE (PATTERN (insn
)) == SET
9978 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
9979 && (REGNO (SET_DEST (PATTERN (insn
))) == BASE_REGNUM
9980 || (!TARGET_CPU_ZARCH
9981 && REGNO (SET_DEST (PATTERN (insn
))) == RETURN_REGNUM
))
9982 && GET_CODE (SET_SRC (PATTERN (insn
))) == MEM
)
9984 set
= PATTERN (insn
);
9985 first
= REGNO (SET_DEST (set
));
9986 offset
= const0_rtx
;
9987 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
9988 off
= INTVAL (offset
);
9990 if (GET_CODE (base
) != REG
|| off
< 0)
9992 if (REGNO (base
) != STACK_POINTER_REGNUM
9993 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10002 /* On z10 and later the dynamic branch prediction must see the
10003 backward jump within a certain windows. If not it falls back to
10004 the static prediction. This function rearranges the loop backward
10005 branch in a way which makes the static prediction always correct.
10006 The function returns true if it added an instruction. */
10008 s390_fix_long_loop_prediction (rtx insn
)
10010 rtx set
= single_set (insn
);
10011 rtx code_label
, label_ref
, new_label
;
10017 /* This will exclude branch on count and branch on index patterns
10018 since these are correctly statically predicted. */
10020 || SET_DEST (set
) != pc_rtx
10021 || GET_CODE (SET_SRC(set
)) != IF_THEN_ELSE
)
10024 label_ref
= (GET_CODE (XEXP (SET_SRC (set
), 1)) == LABEL_REF
?
10025 XEXP (SET_SRC (set
), 1) : XEXP (SET_SRC (set
), 2));
10027 gcc_assert (GET_CODE (label_ref
) == LABEL_REF
);
10029 code_label
= XEXP (label_ref
, 0);
10031 if (INSN_ADDRESSES (INSN_UID (code_label
)) == -1
10032 || INSN_ADDRESSES (INSN_UID (insn
)) == -1
10033 || (INSN_ADDRESSES (INSN_UID (insn
))
10034 - INSN_ADDRESSES (INSN_UID (code_label
)) < PREDICT_DISTANCE
))
10037 for (distance
= 0, cur_insn
= PREV_INSN (insn
);
10038 distance
< PREDICT_DISTANCE
- 6;
10039 distance
+= get_attr_length (cur_insn
), cur_insn
= PREV_INSN (cur_insn
))
10040 if (!cur_insn
|| JUMP_P (cur_insn
) || LABEL_P (cur_insn
))
10043 new_label
= gen_label_rtx ();
10044 uncond_jump
= emit_jump_insn_after (
10045 gen_rtx_SET (VOIDmode
, pc_rtx
,
10046 gen_rtx_LABEL_REF (VOIDmode
, code_label
)),
10048 emit_label_after (new_label
, uncond_jump
);
10050 tmp
= XEXP (SET_SRC (set
), 1);
10051 XEXP (SET_SRC (set
), 1) = XEXP (SET_SRC (set
), 2);
10052 XEXP (SET_SRC (set
), 2) = tmp
;
10053 INSN_CODE (insn
) = -1;
10055 XEXP (label_ref
, 0) = new_label
;
10056 JUMP_LABEL (insn
) = new_label
;
10057 JUMP_LABEL (uncond_jump
) = code_label
;
10062 /* Returns 1 if INSN reads the value of REG for purposes not related
10063 to addressing of memory, and 0 otherwise. */
10065 s390_non_addr_reg_read_p (rtx reg
, rtx insn
)
10067 return reg_referenced_p (reg
, PATTERN (insn
))
10068 && !reg_used_in_mem_p (REGNO (reg
), PATTERN (insn
));
10071 /* Starting from INSN find_cond_jump looks downwards in the insn
10072 stream for a single jump insn which is the last user of the
10073 condition code set in INSN. */
10075 find_cond_jump (rtx insn
)
10077 for (; insn
; insn
= NEXT_INSN (insn
))
10081 if (LABEL_P (insn
))
10084 if (!JUMP_P (insn
))
10086 if (reg_mentioned_p (gen_rtx_REG (CCmode
, CC_REGNUM
), insn
))
10091 /* This will be triggered by a return. */
10092 if (GET_CODE (PATTERN (insn
)) != SET
)
10095 gcc_assert (SET_DEST (PATTERN (insn
)) == pc_rtx
);
10096 ite
= SET_SRC (PATTERN (insn
));
10098 if (GET_CODE (ite
) != IF_THEN_ELSE
)
10101 cc
= XEXP (XEXP (ite
, 0), 0);
10102 if (!REG_P (cc
) || !CC_REGNO_P (REGNO (cc
)))
10105 if (find_reg_note (insn
, REG_DEAD
, cc
))
10113 /* Swap the condition in COND and the operands in OP0 and OP1 so that
10114 the semantics does not change. If NULL_RTX is passed as COND the
10115 function tries to find the conditional jump starting with INSN. */
10117 s390_swap_cmp (rtx cond
, rtx
*op0
, rtx
*op1
, rtx insn
)
10121 if (cond
== NULL_RTX
)
10123 rtx jump
= find_cond_jump (NEXT_INSN (insn
));
10124 jump
= jump
? single_set (jump
) : NULL_RTX
;
10126 if (jump
== NULL_RTX
)
10129 cond
= XEXP (XEXP (jump
, 1), 0);
10134 PUT_CODE (cond
, swap_condition (GET_CODE (cond
)));
10137 /* On z10, instructions of the compare-and-branch family have the
10138 property to access the register occurring as second operand with
10139 its bits complemented. If such a compare is grouped with a second
10140 instruction that accesses the same register non-complemented, and
10141 if that register's value is delivered via a bypass, then the
10142 pipeline recycles, thereby causing significant performance decline.
10143 This function locates such situations and exchanges the two
10144 operands of the compare. The function return true whenever it
10147 s390_z10_optimize_cmp (rtx insn
)
10149 rtx prev_insn
, next_insn
;
10150 bool insn_added_p
= false;
10151 rtx cond
, *op0
, *op1
;
10153 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
10155 /* Handle compare and branch and branch on count
10157 rtx pattern
= single_set (insn
);
10160 || SET_DEST (pattern
) != pc_rtx
10161 || GET_CODE (SET_SRC (pattern
)) != IF_THEN_ELSE
)
10164 cond
= XEXP (SET_SRC (pattern
), 0);
10165 op0
= &XEXP (cond
, 0);
10166 op1
= &XEXP (cond
, 1);
10168 else if (GET_CODE (PATTERN (insn
)) == SET
)
10172 /* Handle normal compare instructions. */
10173 src
= SET_SRC (PATTERN (insn
));
10174 dest
= SET_DEST (PATTERN (insn
));
10177 || !CC_REGNO_P (REGNO (dest
))
10178 || GET_CODE (src
) != COMPARE
)
10181 /* s390_swap_cmp will try to find the conditional
10182 jump when passing NULL_RTX as condition. */
10184 op0
= &XEXP (src
, 0);
10185 op1
= &XEXP (src
, 1);
10190 if (!REG_P (*op0
) || !REG_P (*op1
))
10193 if (GET_MODE_CLASS (GET_MODE (*op0
)) != MODE_INT
)
10196 /* Swap the COMPARE arguments and its mask if there is a
10197 conflicting access in the previous insn. */
10198 prev_insn
= prev_active_insn (insn
);
10199 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10200 && reg_referenced_p (*op1
, PATTERN (prev_insn
)))
10201 s390_swap_cmp (cond
, op0
, op1
, insn
);
10203 /* Check if there is a conflict with the next insn. If there
10204 was no conflict with the previous insn, then swap the
10205 COMPARE arguments and its mask. If we already swapped
10206 the operands, or if swapping them would cause a conflict
10207 with the previous insn, issue a NOP after the COMPARE in
10208 order to separate the two instuctions. */
10209 next_insn
= next_active_insn (insn
);
10210 if (next_insn
!= NULL_RTX
&& INSN_P (next_insn
)
10211 && s390_non_addr_reg_read_p (*op1
, next_insn
))
10213 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10214 && s390_non_addr_reg_read_p (*op0
, prev_insn
))
10216 if (REGNO (*op1
) == 0)
10217 emit_insn_after (gen_nop1 (), insn
);
10219 emit_insn_after (gen_nop (), insn
);
10220 insn_added_p
= true;
10223 s390_swap_cmp (cond
, op0
, op1
, insn
);
10225 return insn_added_p
;
10228 /* Perform machine-dependent processing. */
10233 bool pool_overflow
= false;
10235 /* Make sure all splits have been performed; splits after
10236 machine_dependent_reorg might confuse insn length counts. */
10237 split_all_insns_noflow ();
10239 /* Install the main literal pool and the associated base
10240 register load insns.
10242 In addition, there are two problematic situations we need
10245 - the literal pool might be > 4096 bytes in size, so that
10246 some of its elements cannot be directly accessed
10248 - a branch target might be > 64K away from the branch, so that
10249 it is not possible to use a PC-relative instruction.
10251 To fix those, we split the single literal pool into multiple
10252 pool chunks, reloading the pool base register at various
10253 points throughout the function to ensure it always points to
10254 the pool chunk the following code expects, and / or replace
10255 PC-relative branches by absolute branches.
10257 However, the two problems are interdependent: splitting the
10258 literal pool can move a branch further away from its target,
10259 causing the 64K limit to overflow, and on the other hand,
10260 replacing a PC-relative branch by an absolute branch means
10261 we need to put the branch target address into the literal
10262 pool, possibly causing it to overflow.
10264 So, we loop trying to fix up both problems until we manage
10265 to satisfy both conditions at the same time. Note that the
10266 loop is guaranteed to terminate as every pass of the loop
10267 strictly decreases the total number of PC-relative branches
10268 in the function. (This is not completely true as there
10269 might be branch-over-pool insns introduced by chunkify_start.
10270 Those never need to be split however.) */
10274 struct constant_pool
*pool
= NULL
;
10276 /* Collect the literal pool. */
10277 if (!pool_overflow
)
10279 pool
= s390_mainpool_start ();
10281 pool_overflow
= true;
10284 /* If literal pool overflowed, start to chunkify it. */
10286 pool
= s390_chunkify_start ();
10288 /* Split out-of-range branches. If this has created new
10289 literal pool entries, cancel current chunk list and
10290 recompute it. zSeries machines have large branch
10291 instructions, so we never need to split a branch. */
10292 if (!TARGET_CPU_ZARCH
&& s390_split_branches ())
10295 s390_chunkify_cancel (pool
);
10297 s390_mainpool_cancel (pool
);
10302 /* If we made it up to here, both conditions are satisfied.
10303 Finish up literal pool related changes. */
10305 s390_chunkify_finish (pool
);
10307 s390_mainpool_finish (pool
);
10309 /* We're done splitting branches. */
10310 cfun
->machine
->split_branches_pending_p
= false;
10314 /* Generate out-of-pool execute target insns. */
10315 if (TARGET_CPU_ZARCH
)
10317 rtx insn
, label
, target
;
10319 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10321 label
= s390_execute_label (insn
);
10325 gcc_assert (label
!= const0_rtx
);
10327 target
= emit_label (XEXP (label
, 0));
10328 INSN_ADDRESSES_NEW (target
, -1);
10330 target
= emit_insn (s390_execute_target (insn
));
10331 INSN_ADDRESSES_NEW (target
, -1);
10335 /* Try to optimize prologue and epilogue further. */
10336 s390_optimize_prologue ();
10338 /* Walk over the insns and do some >=z10 specific changes. */
10339 if (s390_tune
== PROCESSOR_2097_Z10
10340 || s390_tune
== PROCESSOR_2817_Z196
)
10343 bool insn_added_p
= false;
10345 /* The insn lengths and addresses have to be up to date for the
10346 following manipulations. */
10347 shorten_branches (get_insns ());
10349 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10351 if (!INSN_P (insn
) || INSN_CODE (insn
) <= 0)
10355 insn_added_p
|= s390_fix_long_loop_prediction (insn
);
10357 if ((GET_CODE (PATTERN (insn
)) == PARALLEL
10358 || GET_CODE (PATTERN (insn
)) == SET
)
10359 && s390_tune
== PROCESSOR_2097_Z10
)
10360 insn_added_p
|= s390_z10_optimize_cmp (insn
);
10363 /* Adjust branches if we added new instructions. */
10365 shorten_branches (get_insns ());
10369 /* Return true if INSN is a fp load insn writing register REGNO. */
10371 s390_fpload_toreg (rtx insn
, unsigned int regno
)
10374 enum attr_type flag
= s390_safe_attr_type (insn
);
10376 if (flag
!= TYPE_FLOADSF
&& flag
!= TYPE_FLOADDF
)
10379 set
= single_set (insn
);
10381 if (set
== NULL_RTX
)
10384 if (!REG_P (SET_DEST (set
)) || !MEM_P (SET_SRC (set
)))
10387 if (REGNO (SET_DEST (set
)) != regno
)
10393 /* This value describes the distance to be avoided between an
10394 aritmetic fp instruction and an fp load writing the same register.
10395 Z10_EARLYLOAD_DISTANCE - 1 as well as Z10_EARLYLOAD_DISTANCE + 1 is
10396 fine but the exact value has to be avoided. Otherwise the FP
10397 pipeline will throw an exception causing a major penalty. */
10398 #define Z10_EARLYLOAD_DISTANCE 7
10400 /* Rearrange the ready list in order to avoid the situation described
10401 for Z10_EARLYLOAD_DISTANCE. A problematic load instruction is
10402 moved to the very end of the ready list. */
10404 s390_z10_prevent_earlyload_conflicts (rtx
*ready
, int *nready_p
)
10406 unsigned int regno
;
10407 int nready
= *nready_p
;
10412 enum attr_type flag
;
10415 /* Skip DISTANCE - 1 active insns. */
10416 for (insn
= last_scheduled_insn
, distance
= Z10_EARLYLOAD_DISTANCE
- 1;
10417 distance
> 0 && insn
!= NULL_RTX
;
10418 distance
--, insn
= prev_active_insn (insn
))
10419 if (CALL_P (insn
) || JUMP_P (insn
))
10422 if (insn
== NULL_RTX
)
10425 set
= single_set (insn
);
10427 if (set
== NULL_RTX
|| !REG_P (SET_DEST (set
))
10428 || GET_MODE_CLASS (GET_MODE (SET_DEST (set
))) != MODE_FLOAT
)
10431 flag
= s390_safe_attr_type (insn
);
10433 if (flag
== TYPE_FLOADSF
|| flag
== TYPE_FLOADDF
)
10436 regno
= REGNO (SET_DEST (set
));
10439 while (!s390_fpload_toreg (ready
[i
], regno
) && i
> 0)
10446 memmove (&ready
[1], &ready
[0], sizeof (rtx
) * i
);
10450 /* This function is called via hook TARGET_SCHED_REORDER before
10451 issueing one insn from list READY which contains *NREADYP entries.
10452 For target z10 it reorders load instructions to avoid early load
10453 conflicts in the floating point pipeline */
10455 s390_sched_reorder (FILE *file ATTRIBUTE_UNUSED
, int verbose ATTRIBUTE_UNUSED
,
10456 rtx
*ready
, int *nreadyp
, int clock ATTRIBUTE_UNUSED
)
10458 if (s390_tune
== PROCESSOR_2097_Z10
)
10459 if (reload_completed
&& *nreadyp
> 1)
10460 s390_z10_prevent_earlyload_conflicts (ready
, nreadyp
);
10462 return s390_issue_rate ();
10465 /* This function is called via hook TARGET_SCHED_VARIABLE_ISSUE after
10466 the scheduler has issued INSN. It stores the last issued insn into
10467 last_scheduled_insn in order to make it available for
10468 s390_sched_reorder. */
10470 s390_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED
,
10471 int verbose ATTRIBUTE_UNUSED
,
10472 rtx insn
, int more
)
10474 last_scheduled_insn
= insn
;
10476 if (GET_CODE (PATTERN (insn
)) != USE
10477 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
10484 s390_sched_init (FILE *file ATTRIBUTE_UNUSED
,
10485 int verbose ATTRIBUTE_UNUSED
,
10486 int max_ready ATTRIBUTE_UNUSED
)
10488 last_scheduled_insn
= NULL_RTX
;
10491 /* This function checks the whole of insn X for memory references. The
10492 function always returns zero because the framework it is called
10493 from would stop recursively analyzing the insn upon a return value
10494 other than zero. The real result of this function is updating
10495 counter variable MEM_COUNT. */
10497 check_dpu (rtx
*x
, unsigned *mem_count
)
10499 if (*x
!= NULL_RTX
&& MEM_P (*x
))
10504 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates
10505 a new number struct loop *loop should be unrolled if tuned for cpus with
10506 a built-in stride prefetcher.
10507 The loop is analyzed for memory accesses by calling check_dpu for
10508 each rtx of the loop. Depending on the loop_depth and the amount of
10509 memory accesses a new number <=nunroll is returned to improve the
10510 behaviour of the hardware prefetch unit. */
10512 s390_loop_unroll_adjust (unsigned nunroll
, struct loop
*loop
)
10517 unsigned mem_count
= 0;
10519 if (s390_tune
!= PROCESSOR_2097_Z10
&& s390_tune
!= PROCESSOR_2817_Z196
)
10522 /* Count the number of memory references within the loop body. */
10523 bbs
= get_loop_body (loop
);
10524 for (i
= 0; i
< loop
->num_nodes
; i
++)
10526 for (insn
= BB_HEAD (bbs
[i
]); insn
!= BB_END (bbs
[i
]); insn
= NEXT_INSN (insn
))
10527 if (INSN_P (insn
) && INSN_CODE (insn
) != -1)
10528 for_each_rtx (&insn
, (rtx_function
) check_dpu
, &mem_count
);
10532 /* Prevent division by zero, and we do not need to adjust nunroll in this case. */
10533 if (mem_count
== 0)
10536 switch (loop_depth(loop
))
10539 return MIN (nunroll
, 28 / mem_count
);
10541 return MIN (nunroll
, 22 / mem_count
);
10543 return MIN (nunroll
, 16 / mem_count
);
10547 /* Initialize GCC target structure. */
10549 #undef TARGET_ASM_ALIGNED_HI_OP
10550 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10551 #undef TARGET_ASM_ALIGNED_DI_OP
10552 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10553 #undef TARGET_ASM_INTEGER
10554 #define TARGET_ASM_INTEGER s390_assemble_integer
10556 #undef TARGET_ASM_OPEN_PAREN
10557 #define TARGET_ASM_OPEN_PAREN ""
10559 #undef TARGET_ASM_CLOSE_PAREN
10560 #define TARGET_ASM_CLOSE_PAREN ""
10562 #undef TARGET_DEFAULT_TARGET_FLAGS
10563 #define TARGET_DEFAULT_TARGET_FLAGS (TARGET_DEFAULT)
10565 #undef TARGET_HANDLE_OPTION
10566 #define TARGET_HANDLE_OPTION s390_handle_option
10568 #undef TARGET_OPTION_OVERRIDE
10569 #define TARGET_OPTION_OVERRIDE s390_option_override
10571 #undef TARGET_OPTION_OPTIMIZATION_TABLE
10572 #define TARGET_OPTION_OPTIMIZATION_TABLE s390_option_optimization_table
10574 #undef TARGET_OPTION_INIT_STRUCT
10575 #define TARGET_OPTION_INIT_STRUCT s390_option_init_struct
10577 #undef TARGET_ENCODE_SECTION_INFO
10578 #define TARGET_ENCODE_SECTION_INFO s390_encode_section_info
10580 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10581 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10584 #undef TARGET_HAVE_TLS
10585 #define TARGET_HAVE_TLS true
10587 #undef TARGET_CANNOT_FORCE_CONST_MEM
10588 #define TARGET_CANNOT_FORCE_CONST_MEM s390_cannot_force_const_mem
10590 #undef TARGET_DELEGITIMIZE_ADDRESS
10591 #define TARGET_DELEGITIMIZE_ADDRESS s390_delegitimize_address
10593 #undef TARGET_LEGITIMIZE_ADDRESS
10594 #define TARGET_LEGITIMIZE_ADDRESS s390_legitimize_address
10596 #undef TARGET_RETURN_IN_MEMORY
10597 #define TARGET_RETURN_IN_MEMORY s390_return_in_memory
10599 #undef TARGET_INIT_BUILTINS
10600 #define TARGET_INIT_BUILTINS s390_init_builtins
10601 #undef TARGET_EXPAND_BUILTIN
10602 #define TARGET_EXPAND_BUILTIN s390_expand_builtin
10604 #undef TARGET_ASM_OUTPUT_MI_THUNK
10605 #define TARGET_ASM_OUTPUT_MI_THUNK s390_output_mi_thunk
10606 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10607 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10609 #undef TARGET_SCHED_ADJUST_PRIORITY
10610 #define TARGET_SCHED_ADJUST_PRIORITY s390_adjust_priority
10611 #undef TARGET_SCHED_ISSUE_RATE
10612 #define TARGET_SCHED_ISSUE_RATE s390_issue_rate
10613 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10614 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD s390_first_cycle_multipass_dfa_lookahead
10616 #undef TARGET_SCHED_VARIABLE_ISSUE
10617 #define TARGET_SCHED_VARIABLE_ISSUE s390_sched_variable_issue
10618 #undef TARGET_SCHED_REORDER
10619 #define TARGET_SCHED_REORDER s390_sched_reorder
10620 #undef TARGET_SCHED_INIT
10621 #define TARGET_SCHED_INIT s390_sched_init
10623 #undef TARGET_CANNOT_COPY_INSN_P
10624 #define TARGET_CANNOT_COPY_INSN_P s390_cannot_copy_insn_p
10625 #undef TARGET_RTX_COSTS
10626 #define TARGET_RTX_COSTS s390_rtx_costs
10627 #undef TARGET_ADDRESS_COST
10628 #define TARGET_ADDRESS_COST s390_address_cost
10630 #undef TARGET_MACHINE_DEPENDENT_REORG
10631 #define TARGET_MACHINE_DEPENDENT_REORG s390_reorg
10633 #undef TARGET_VALID_POINTER_MODE
10634 #define TARGET_VALID_POINTER_MODE s390_valid_pointer_mode
10636 #undef TARGET_BUILD_BUILTIN_VA_LIST
10637 #define TARGET_BUILD_BUILTIN_VA_LIST s390_build_builtin_va_list
10638 #undef TARGET_EXPAND_BUILTIN_VA_START
10639 #define TARGET_EXPAND_BUILTIN_VA_START s390_va_start
10640 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10641 #define TARGET_GIMPLIFY_VA_ARG_EXPR s390_gimplify_va_arg
10643 #undef TARGET_PROMOTE_FUNCTION_MODE
10644 #define TARGET_PROMOTE_FUNCTION_MODE s390_promote_function_mode
10645 #undef TARGET_PASS_BY_REFERENCE
10646 #define TARGET_PASS_BY_REFERENCE s390_pass_by_reference
10648 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10649 #define TARGET_FUNCTION_OK_FOR_SIBCALL s390_function_ok_for_sibcall
10650 #undef TARGET_FUNCTION_ARG
10651 #define TARGET_FUNCTION_ARG s390_function_arg
10652 #undef TARGET_FUNCTION_ARG_ADVANCE
10653 #define TARGET_FUNCTION_ARG_ADVANCE s390_function_arg_advance
10655 #undef TARGET_FIXED_CONDITION_CODE_REGS
10656 #define TARGET_FIXED_CONDITION_CODE_REGS s390_fixed_condition_code_regs
10658 #undef TARGET_CC_MODES_COMPATIBLE
10659 #define TARGET_CC_MODES_COMPATIBLE s390_cc_modes_compatible
10661 #undef TARGET_INVALID_WITHIN_DOLOOP
10662 #define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null
10665 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
10666 #define TARGET_ASM_OUTPUT_DWARF_DTPREL s390_output_dwarf_dtprel
10669 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10670 #undef TARGET_MANGLE_TYPE
10671 #define TARGET_MANGLE_TYPE s390_mangle_type
10674 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10675 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10677 #undef TARGET_SECONDARY_RELOAD
10678 #define TARGET_SECONDARY_RELOAD s390_secondary_reload
10680 #undef TARGET_LIBGCC_CMP_RETURN_MODE
10681 #define TARGET_LIBGCC_CMP_RETURN_MODE s390_libgcc_cmp_return_mode
10683 #undef TARGET_LIBGCC_SHIFT_COUNT_MODE
10684 #define TARGET_LIBGCC_SHIFT_COUNT_MODE s390_libgcc_shift_count_mode
10686 #undef TARGET_LEGITIMATE_ADDRESS_P
10687 #define TARGET_LEGITIMATE_ADDRESS_P s390_legitimate_address_p
10689 #undef TARGET_CAN_ELIMINATE
10690 #define TARGET_CAN_ELIMINATE s390_can_eliminate
10692 #undef TARGET_CONDITIONAL_REGISTER_USAGE
10693 #define TARGET_CONDITIONAL_REGISTER_USAGE s390_conditional_register_usage
10695 #undef TARGET_LOOP_UNROLL_ADJUST
10696 #define TARGET_LOOP_UNROLL_ADJUST s390_loop_unroll_adjust
10698 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
10699 #define TARGET_ASM_TRAMPOLINE_TEMPLATE s390_asm_trampoline_template
10700 #undef TARGET_TRAMPOLINE_INIT
10701 #define TARGET_TRAMPOLINE_INIT s390_trampoline_init
10703 #undef TARGET_UNWIND_WORD_MODE
10704 #define TARGET_UNWIND_WORD_MODE s390_unwind_word_mode
10706 struct gcc_target targetm
= TARGET_INITIALIZER
;
10708 #include "gt-s390.h"