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, 2011, 2012 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"
44 #include "basic-block.h"
45 #include "integrate.h"
48 #include "target-def.h"
50 #include "langhooks.h"
58 /* Define the specific costs for a given cpu. */
60 struct processor_costs
63 const int m
; /* cost of an M instruction. */
64 const int mghi
; /* cost of an MGHI instruction. */
65 const int mh
; /* cost of an MH instruction. */
66 const int mhi
; /* cost of an MHI instruction. */
67 const int ml
; /* cost of an ML instruction. */
68 const int mr
; /* cost of an MR instruction. */
69 const int ms
; /* cost of an MS instruction. */
70 const int msg
; /* cost of an MSG instruction. */
71 const int msgf
; /* cost of an MSGF instruction. */
72 const int msgfr
; /* cost of an MSGFR instruction. */
73 const int msgr
; /* cost of an MSGR instruction. */
74 const int msr
; /* cost of an MSR instruction. */
75 const int mult_df
; /* cost of multiplication in DFmode. */
78 const int sqxbr
; /* cost of square root in TFmode. */
79 const int sqdbr
; /* cost of square root in DFmode. */
80 const int sqebr
; /* cost of square root in SFmode. */
81 /* multiply and add */
82 const int madbr
; /* cost of multiply and add in DFmode. */
83 const int maebr
; /* cost of multiply and add in SFmode. */
95 const struct processor_costs
*s390_cost
;
98 struct processor_costs z900_cost
=
100 COSTS_N_INSNS (5), /* M */
101 COSTS_N_INSNS (10), /* MGHI */
102 COSTS_N_INSNS (5), /* MH */
103 COSTS_N_INSNS (4), /* MHI */
104 COSTS_N_INSNS (5), /* ML */
105 COSTS_N_INSNS (5), /* MR */
106 COSTS_N_INSNS (4), /* MS */
107 COSTS_N_INSNS (15), /* MSG */
108 COSTS_N_INSNS (7), /* MSGF */
109 COSTS_N_INSNS (7), /* MSGFR */
110 COSTS_N_INSNS (10), /* MSGR */
111 COSTS_N_INSNS (4), /* MSR */
112 COSTS_N_INSNS (7), /* multiplication in DFmode */
113 COSTS_N_INSNS (13), /* MXBR */
114 COSTS_N_INSNS (136), /* SQXBR */
115 COSTS_N_INSNS (44), /* SQDBR */
116 COSTS_N_INSNS (35), /* SQEBR */
117 COSTS_N_INSNS (18), /* MADBR */
118 COSTS_N_INSNS (13), /* MAEBR */
119 COSTS_N_INSNS (134), /* DXBR */
120 COSTS_N_INSNS (30), /* DDBR */
121 COSTS_N_INSNS (27), /* DEBR */
122 COSTS_N_INSNS (220), /* DLGR */
123 COSTS_N_INSNS (34), /* DLR */
124 COSTS_N_INSNS (34), /* DR */
125 COSTS_N_INSNS (32), /* DSGFR */
126 COSTS_N_INSNS (32), /* DSGR */
130 struct processor_costs z990_cost
=
132 COSTS_N_INSNS (4), /* M */
133 COSTS_N_INSNS (2), /* MGHI */
134 COSTS_N_INSNS (2), /* MH */
135 COSTS_N_INSNS (2), /* MHI */
136 COSTS_N_INSNS (4), /* ML */
137 COSTS_N_INSNS (4), /* MR */
138 COSTS_N_INSNS (5), /* MS */
139 COSTS_N_INSNS (6), /* MSG */
140 COSTS_N_INSNS (4), /* MSGF */
141 COSTS_N_INSNS (4), /* MSGFR */
142 COSTS_N_INSNS (4), /* MSGR */
143 COSTS_N_INSNS (4), /* MSR */
144 COSTS_N_INSNS (1), /* multiplication in DFmode */
145 COSTS_N_INSNS (28), /* MXBR */
146 COSTS_N_INSNS (130), /* SQXBR */
147 COSTS_N_INSNS (66), /* SQDBR */
148 COSTS_N_INSNS (38), /* SQEBR */
149 COSTS_N_INSNS (1), /* MADBR */
150 COSTS_N_INSNS (1), /* MAEBR */
151 COSTS_N_INSNS (60), /* DXBR */
152 COSTS_N_INSNS (40), /* DDBR */
153 COSTS_N_INSNS (26), /* DEBR */
154 COSTS_N_INSNS (176), /* DLGR */
155 COSTS_N_INSNS (31), /* DLR */
156 COSTS_N_INSNS (31), /* DR */
157 COSTS_N_INSNS (31), /* DSGFR */
158 COSTS_N_INSNS (31), /* DSGR */
162 struct processor_costs z9_109_cost
=
164 COSTS_N_INSNS (4), /* M */
165 COSTS_N_INSNS (2), /* MGHI */
166 COSTS_N_INSNS (2), /* MH */
167 COSTS_N_INSNS (2), /* MHI */
168 COSTS_N_INSNS (4), /* ML */
169 COSTS_N_INSNS (4), /* MR */
170 COSTS_N_INSNS (5), /* MS */
171 COSTS_N_INSNS (6), /* MSG */
172 COSTS_N_INSNS (4), /* MSGF */
173 COSTS_N_INSNS (4), /* MSGFR */
174 COSTS_N_INSNS (4), /* MSGR */
175 COSTS_N_INSNS (4), /* MSR */
176 COSTS_N_INSNS (1), /* multiplication in DFmode */
177 COSTS_N_INSNS (28), /* MXBR */
178 COSTS_N_INSNS (130), /* SQXBR */
179 COSTS_N_INSNS (66), /* SQDBR */
180 COSTS_N_INSNS (38), /* SQEBR */
181 COSTS_N_INSNS (1), /* MADBR */
182 COSTS_N_INSNS (1), /* MAEBR */
183 COSTS_N_INSNS (60), /* DXBR */
184 COSTS_N_INSNS (40), /* DDBR */
185 COSTS_N_INSNS (26), /* DEBR */
186 COSTS_N_INSNS (30), /* DLGR */
187 COSTS_N_INSNS (23), /* DLR */
188 COSTS_N_INSNS (23), /* DR */
189 COSTS_N_INSNS (24), /* DSGFR */
190 COSTS_N_INSNS (24), /* DSGR */
194 struct processor_costs z10_cost
=
196 COSTS_N_INSNS (10), /* M */
197 COSTS_N_INSNS (10), /* MGHI */
198 COSTS_N_INSNS (10), /* MH */
199 COSTS_N_INSNS (10), /* MHI */
200 COSTS_N_INSNS (10), /* ML */
201 COSTS_N_INSNS (10), /* MR */
202 COSTS_N_INSNS (10), /* MS */
203 COSTS_N_INSNS (10), /* MSG */
204 COSTS_N_INSNS (10), /* MSGF */
205 COSTS_N_INSNS (10), /* MSGFR */
206 COSTS_N_INSNS (10), /* MSGR */
207 COSTS_N_INSNS (10), /* MSR */
208 COSTS_N_INSNS (1) , /* multiplication in DFmode */
209 COSTS_N_INSNS (50), /* MXBR */
210 COSTS_N_INSNS (120), /* SQXBR */
211 COSTS_N_INSNS (52), /* SQDBR */
212 COSTS_N_INSNS (38), /* SQEBR */
213 COSTS_N_INSNS (1), /* MADBR */
214 COSTS_N_INSNS (1), /* MAEBR */
215 COSTS_N_INSNS (111), /* DXBR */
216 COSTS_N_INSNS (39), /* DDBR */
217 COSTS_N_INSNS (32), /* DEBR */
218 COSTS_N_INSNS (160), /* DLGR */
219 COSTS_N_INSNS (71), /* DLR */
220 COSTS_N_INSNS (71), /* DR */
221 COSTS_N_INSNS (71), /* DSGFR */
222 COSTS_N_INSNS (71), /* DSGR */
226 struct processor_costs z196_cost
=
228 COSTS_N_INSNS (7), /* M */
229 COSTS_N_INSNS (5), /* MGHI */
230 COSTS_N_INSNS (5), /* MH */
231 COSTS_N_INSNS (5), /* MHI */
232 COSTS_N_INSNS (7), /* ML */
233 COSTS_N_INSNS (7), /* MR */
234 COSTS_N_INSNS (6), /* MS */
235 COSTS_N_INSNS (8), /* MSG */
236 COSTS_N_INSNS (6), /* MSGF */
237 COSTS_N_INSNS (6), /* MSGFR */
238 COSTS_N_INSNS (8), /* MSGR */
239 COSTS_N_INSNS (6), /* MSR */
240 COSTS_N_INSNS (1) , /* multiplication in DFmode */
241 COSTS_N_INSNS (40), /* MXBR B+40 */
242 COSTS_N_INSNS (100), /* SQXBR B+100 */
243 COSTS_N_INSNS (42), /* SQDBR B+42 */
244 COSTS_N_INSNS (28), /* SQEBR B+28 */
245 COSTS_N_INSNS (1), /* MADBR B */
246 COSTS_N_INSNS (1), /* MAEBR B */
247 COSTS_N_INSNS (101), /* DXBR B+101 */
248 COSTS_N_INSNS (29), /* DDBR */
249 COSTS_N_INSNS (22), /* DEBR */
250 COSTS_N_INSNS (160), /* DLGR cracked */
251 COSTS_N_INSNS (160), /* DLR cracked */
252 COSTS_N_INSNS (160), /* DR expanded */
253 COSTS_N_INSNS (160), /* DSGFR cracked */
254 COSTS_N_INSNS (160), /* DSGR cracked */
257 extern int reload_completed
;
259 /* Kept up to date using the SCHED_VARIABLE_ISSUE hook. */
260 static rtx last_scheduled_insn
;
262 /* Structure used to hold the components of a S/390 memory
263 address. A legitimate address on S/390 is of the general
265 base + index + displacement
266 where any of the components is optional.
268 base and index are registers of the class ADDR_REGS,
269 displacement is an unsigned 12-bit immediate constant. */
280 /* The following structure is embedded in the machine
281 specific part of struct function. */
283 struct GTY (()) s390_frame_layout
285 /* Offset within stack frame. */
286 HOST_WIDE_INT gprs_offset
;
287 HOST_WIDE_INT f0_offset
;
288 HOST_WIDE_INT f4_offset
;
289 HOST_WIDE_INT f8_offset
;
290 HOST_WIDE_INT backchain_offset
;
292 /* Number of first and last gpr where slots in the register
293 save area are reserved for. */
294 int first_save_gpr_slot
;
295 int last_save_gpr_slot
;
297 /* Number of first and last gpr to be saved, restored. */
299 int first_restore_gpr
;
301 int last_restore_gpr
;
303 /* Bits standing for floating point registers. Set, if the
304 respective register has to be saved. Starting with reg 16 (f0)
305 at the rightmost bit.
306 Bit 15 - 8 7 6 5 4 3 2 1 0
307 fpr 15 - 8 7 5 3 1 6 4 2 0
308 reg 31 - 24 23 22 21 20 19 18 17 16 */
309 unsigned int fpr_bitmap
;
311 /* Number of floating point registers f8-f15 which must be saved. */
314 /* Set if return address needs to be saved.
315 This flag is set by s390_return_addr_rtx if it could not use
316 the initial value of r14 and therefore depends on r14 saved
318 bool save_return_addr_p
;
320 /* Size of stack frame. */
321 HOST_WIDE_INT frame_size
;
324 /* Define the structure for the machine field in struct function. */
326 struct GTY(()) machine_function
328 struct s390_frame_layout frame_layout
;
330 /* Literal pool base register. */
333 /* True if we may need to perform branch splitting. */
334 bool split_branches_pending_p
;
336 /* Some local-dynamic TLS symbol name. */
337 const char *some_ld_name
;
339 bool has_landing_pad_p
;
342 /* Few accessor macros for struct cfun->machine->s390_frame_layout. */
344 #define cfun_frame_layout (cfun->machine->frame_layout)
345 #define cfun_save_high_fprs_p (!!cfun_frame_layout.high_fprs)
346 #define cfun_gprs_save_area_size ((cfun_frame_layout.last_save_gpr_slot - \
347 cfun_frame_layout.first_save_gpr_slot + 1) * UNITS_PER_LONG)
348 #define cfun_set_fpr_bit(BITNUM) (cfun->machine->frame_layout.fpr_bitmap |= \
350 #define cfun_fpr_bit_p(BITNUM) (!!(cfun->machine->frame_layout.fpr_bitmap & \
353 /* Number of GPRs and FPRs used for argument passing. */
354 #define GP_ARG_NUM_REG 5
355 #define FP_ARG_NUM_REG (TARGET_64BIT? 4 : 2)
357 /* A couple of shortcuts. */
358 #define CONST_OK_FOR_J(x) \
359 CONST_OK_FOR_CONSTRAINT_P((x), 'J', "J")
360 #define CONST_OK_FOR_K(x) \
361 CONST_OK_FOR_CONSTRAINT_P((x), 'K', "K")
362 #define CONST_OK_FOR_Os(x) \
363 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Os")
364 #define CONST_OK_FOR_Op(x) \
365 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Op")
366 #define CONST_OK_FOR_On(x) \
367 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "On")
369 #define REGNO_PAIR_OK(REGNO, MODE) \
370 (HARD_REGNO_NREGS ((REGNO), (MODE)) == 1 || !((REGNO) & 1))
372 /* That's the read ahead of the dynamic branch prediction unit in
373 bytes on a z10 (or higher) CPU. */
374 #define PREDICT_DISTANCE (TARGET_Z10 ? 384 : 2048)
376 /* Return the alignment for LABEL. We default to the -falign-labels
377 value except for the literal pool base label. */
379 s390_label_align (rtx label
)
381 rtx prev_insn
= prev_active_insn (label
);
383 if (prev_insn
== NULL_RTX
)
386 prev_insn
= single_set (prev_insn
);
388 if (prev_insn
== NULL_RTX
)
391 prev_insn
= SET_SRC (prev_insn
);
393 /* Don't align literal pool base labels. */
394 if (GET_CODE (prev_insn
) == UNSPEC
395 && XINT (prev_insn
, 1) == UNSPEC_MAIN_BASE
)
399 return align_labels_log
;
402 static enum machine_mode
403 s390_libgcc_cmp_return_mode (void)
405 return TARGET_64BIT
? DImode
: SImode
;
408 static enum machine_mode
409 s390_libgcc_shift_count_mode (void)
411 return TARGET_64BIT
? DImode
: SImode
;
414 static enum machine_mode
415 s390_unwind_word_mode (void)
417 return TARGET_64BIT
? DImode
: SImode
;
420 /* Return true if the back end supports mode MODE. */
422 s390_scalar_mode_supported_p (enum machine_mode mode
)
424 /* In contrast to the default implementation reject TImode constants on 31bit
425 TARGET_ZARCH for ABI compliance. */
426 if (!TARGET_64BIT
&& TARGET_ZARCH
&& mode
== TImode
)
429 if (DECIMAL_FLOAT_MODE_P (mode
))
430 return default_decimal_float_supported_p ();
432 return default_scalar_mode_supported_p (mode
);
435 /* Set the has_landing_pad_p flag in struct machine_function to VALUE. */
438 s390_set_has_landing_pad_p (bool value
)
440 cfun
->machine
->has_landing_pad_p
= value
;
443 /* If two condition code modes are compatible, return a condition code
444 mode which is compatible with both. Otherwise, return
447 static enum machine_mode
448 s390_cc_modes_compatible (enum machine_mode m1
, enum machine_mode m2
)
456 if (m2
== CCUmode
|| m2
== CCTmode
|| m2
== CCZ1mode
457 || m2
== CCSmode
|| m2
== CCSRmode
|| m2
== CCURmode
)
478 /* Return true if SET either doesn't set the CC register, or else
479 the source and destination have matching CC modes and that
480 CC mode is at least as constrained as REQ_MODE. */
483 s390_match_ccmode_set (rtx set
, enum machine_mode req_mode
)
485 enum machine_mode set_mode
;
487 gcc_assert (GET_CODE (set
) == SET
);
489 if (GET_CODE (SET_DEST (set
)) != REG
|| !CC_REGNO_P (REGNO (SET_DEST (set
))))
492 set_mode
= GET_MODE (SET_DEST (set
));
506 if (req_mode
!= set_mode
)
511 if (req_mode
!= CCSmode
&& req_mode
!= CCUmode
&& req_mode
!= CCTmode
512 && req_mode
!= CCSRmode
&& req_mode
!= CCURmode
)
518 if (req_mode
!= CCAmode
)
526 return (GET_MODE (SET_SRC (set
)) == set_mode
);
529 /* Return true if every SET in INSN that sets the CC register
530 has source and destination with matching CC modes and that
531 CC mode is at least as constrained as REQ_MODE.
532 If REQ_MODE is VOIDmode, always return false. */
535 s390_match_ccmode (rtx insn
, enum machine_mode req_mode
)
539 /* s390_tm_ccmode returns VOIDmode to indicate failure. */
540 if (req_mode
== VOIDmode
)
543 if (GET_CODE (PATTERN (insn
)) == SET
)
544 return s390_match_ccmode_set (PATTERN (insn
), req_mode
);
546 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
547 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
549 rtx set
= XVECEXP (PATTERN (insn
), 0, i
);
550 if (GET_CODE (set
) == SET
)
551 if (!s390_match_ccmode_set (set
, req_mode
))
558 /* If a test-under-mask instruction can be used to implement
559 (compare (and ... OP1) OP2), return the CC mode required
560 to do that. Otherwise, return VOIDmode.
561 MIXED is true if the instruction can distinguish between
562 CC1 and CC2 for mixed selected bits (TMxx), it is false
563 if the instruction cannot (TM). */
566 s390_tm_ccmode (rtx op1
, rtx op2
, bool mixed
)
570 /* ??? Fixme: should work on CONST_DOUBLE as well. */
571 if (GET_CODE (op1
) != CONST_INT
|| GET_CODE (op2
) != CONST_INT
)
574 /* Selected bits all zero: CC0.
575 e.g.: int a; if ((a & (16 + 128)) == 0) */
576 if (INTVAL (op2
) == 0)
579 /* Selected bits all one: CC3.
580 e.g.: int a; if ((a & (16 + 128)) == 16 + 128) */
581 if (INTVAL (op2
) == INTVAL (op1
))
584 /* Exactly two bits selected, mixed zeroes and ones: CC1 or CC2. e.g.:
586 if ((a & (16 + 128)) == 16) -> CCT1
587 if ((a & (16 + 128)) == 128) -> CCT2 */
590 bit1
= exact_log2 (INTVAL (op2
));
591 bit0
= exact_log2 (INTVAL (op1
) ^ INTVAL (op2
));
592 if (bit0
!= -1 && bit1
!= -1)
593 return bit0
> bit1
? CCT1mode
: CCT2mode
;
599 /* Given a comparison code OP (EQ, NE, etc.) and the operands
600 OP0 and OP1 of a COMPARE, return the mode to be used for the
604 s390_select_ccmode (enum rtx_code code
, rtx op0
, rtx op1
)
610 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
611 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
613 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
614 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
616 if ((GET_CODE (op0
) == PLUS
|| GET_CODE (op0
) == MINUS
617 || GET_CODE (op1
) == NEG
)
618 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
621 if (GET_CODE (op0
) == AND
)
623 /* Check whether we can potentially do it via TM. */
624 enum machine_mode ccmode
;
625 ccmode
= s390_tm_ccmode (XEXP (op0
, 1), op1
, 1);
626 if (ccmode
!= VOIDmode
)
628 /* Relax CCTmode to CCZmode to allow fall-back to AND
629 if that turns out to be beneficial. */
630 return ccmode
== CCTmode
? CCZmode
: ccmode
;
634 if (register_operand (op0
, HImode
)
635 && GET_CODE (op1
) == CONST_INT
636 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 65535))
638 if (register_operand (op0
, QImode
)
639 && GET_CODE (op1
) == CONST_INT
640 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 255))
649 /* The only overflow condition of NEG and ABS happens when
650 -INT_MAX is used as parameter, which stays negative. So
651 we have an overflow from a positive value to a negative.
652 Using CCAP mode the resulting cc can be used for comparisons. */
653 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
654 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
657 /* If constants are involved in an add instruction it is possible to use
658 the resulting cc for comparisons with zero. Knowing the sign of the
659 constant the overflow behavior gets predictable. e.g.:
660 int a, b; if ((b = a + c) > 0)
661 with c as a constant value: c < 0 -> CCAN and c >= 0 -> CCAP */
662 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
663 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
665 if (INTVAL (XEXP((op0
), 1)) < 0)
679 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
680 && GET_CODE (op1
) != CONST_INT
)
686 if (GET_CODE (op0
) == PLUS
687 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
690 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
691 && GET_CODE (op1
) != CONST_INT
)
697 if (GET_CODE (op0
) == MINUS
698 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
701 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
702 && GET_CODE (op1
) != CONST_INT
)
711 /* Replace the comparison OP0 CODE OP1 by a semantically equivalent one
712 that we can implement more efficiently. */
715 s390_canonicalize_comparison (enum rtx_code
*code
, rtx
*op0
, rtx
*op1
)
717 /* Convert ZERO_EXTRACT back to AND to enable TM patterns. */
718 if ((*code
== EQ
|| *code
== NE
)
719 && *op1
== const0_rtx
720 && GET_CODE (*op0
) == ZERO_EXTRACT
721 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
722 && GET_CODE (XEXP (*op0
, 2)) == CONST_INT
723 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
725 rtx inner
= XEXP (*op0
, 0);
726 HOST_WIDE_INT modesize
= GET_MODE_BITSIZE (GET_MODE (inner
));
727 HOST_WIDE_INT len
= INTVAL (XEXP (*op0
, 1));
728 HOST_WIDE_INT pos
= INTVAL (XEXP (*op0
, 2));
730 if (len
> 0 && len
< modesize
731 && pos
>= 0 && pos
+ len
<= modesize
732 && modesize
<= HOST_BITS_PER_WIDE_INT
)
734 unsigned HOST_WIDE_INT block
;
735 block
= ((unsigned HOST_WIDE_INT
) 1 << len
) - 1;
736 block
<<= modesize
- pos
- len
;
738 *op0
= gen_rtx_AND (GET_MODE (inner
), inner
,
739 gen_int_mode (block
, GET_MODE (inner
)));
743 /* Narrow AND of memory against immediate to enable TM. */
744 if ((*code
== EQ
|| *code
== NE
)
745 && *op1
== const0_rtx
746 && GET_CODE (*op0
) == AND
747 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
748 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
750 rtx inner
= XEXP (*op0
, 0);
751 rtx mask
= XEXP (*op0
, 1);
753 /* Ignore paradoxical SUBREGs if all extra bits are masked out. */
754 if (GET_CODE (inner
) == SUBREG
755 && SCALAR_INT_MODE_P (GET_MODE (SUBREG_REG (inner
)))
756 && (GET_MODE_SIZE (GET_MODE (inner
))
757 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner
))))
759 & GET_MODE_MASK (GET_MODE (inner
))
760 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (inner
))))
762 inner
= SUBREG_REG (inner
);
764 /* Do not change volatile MEMs. */
765 if (MEM_P (inner
) && !MEM_VOLATILE_P (inner
))
767 int part
= s390_single_part (XEXP (*op0
, 1),
768 GET_MODE (inner
), QImode
, 0);
771 mask
= gen_int_mode (s390_extract_part (mask
, QImode
, 0), QImode
);
772 inner
= adjust_address_nv (inner
, QImode
, part
);
773 *op0
= gen_rtx_AND (QImode
, inner
, mask
);
778 /* Narrow comparisons against 0xffff to HImode if possible. */
779 if ((*code
== EQ
|| *code
== NE
)
780 && GET_CODE (*op1
) == CONST_INT
781 && INTVAL (*op1
) == 0xffff
782 && SCALAR_INT_MODE_P (GET_MODE (*op0
))
783 && (nonzero_bits (*op0
, GET_MODE (*op0
))
784 & ~(unsigned HOST_WIDE_INT
) 0xffff) == 0)
786 *op0
= gen_lowpart (HImode
, *op0
);
790 /* Remove redundant UNSPEC_CCU_TO_INT conversions if possible. */
791 if (GET_CODE (*op0
) == UNSPEC
792 && XINT (*op0
, 1) == UNSPEC_CCU_TO_INT
793 && XVECLEN (*op0
, 0) == 1
794 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCUmode
795 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
796 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
797 && *op1
== const0_rtx
)
799 enum rtx_code new_code
= UNKNOWN
;
802 case EQ
: new_code
= EQ
; break;
803 case NE
: new_code
= NE
; break;
804 case LT
: new_code
= GTU
; break;
805 case GT
: new_code
= LTU
; break;
806 case LE
: new_code
= GEU
; break;
807 case GE
: new_code
= LEU
; break;
811 if (new_code
!= UNKNOWN
)
813 *op0
= XVECEXP (*op0
, 0, 0);
818 /* Remove redundant UNSPEC_CCZ_TO_INT conversions if possible. */
819 if (GET_CODE (*op0
) == UNSPEC
820 && XINT (*op0
, 1) == UNSPEC_CCZ_TO_INT
821 && XVECLEN (*op0
, 0) == 1
822 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCZmode
823 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
824 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
825 && *op1
== const0_rtx
)
827 enum rtx_code new_code
= UNKNOWN
;
830 case EQ
: new_code
= EQ
; break;
831 case NE
: new_code
= NE
; break;
835 if (new_code
!= UNKNOWN
)
837 *op0
= XVECEXP (*op0
, 0, 0);
842 /* Simplify cascaded EQ, NE with const0_rtx. */
843 if ((*code
== NE
|| *code
== EQ
)
844 && (GET_CODE (*op0
) == EQ
|| GET_CODE (*op0
) == NE
)
845 && GET_MODE (*op0
) == SImode
846 && GET_MODE (XEXP (*op0
, 0)) == CCZ1mode
847 && REG_P (XEXP (*op0
, 0))
848 && XEXP (*op0
, 1) == const0_rtx
849 && *op1
== const0_rtx
)
851 if ((*code
== EQ
&& GET_CODE (*op0
) == NE
)
852 || (*code
== NE
&& GET_CODE (*op0
) == EQ
))
856 *op0
= XEXP (*op0
, 0);
859 /* Prefer register over memory as first operand. */
860 if (MEM_P (*op0
) && REG_P (*op1
))
862 rtx tem
= *op0
; *op0
= *op1
; *op1
= tem
;
863 *code
= swap_condition (*code
);
867 /* Emit a compare instruction suitable to implement the comparison
868 OP0 CODE OP1. Return the correct condition RTL to be placed in
869 the IF_THEN_ELSE of the conditional branch testing the result. */
872 s390_emit_compare (enum rtx_code code
, rtx op0
, rtx op1
)
874 enum machine_mode mode
= s390_select_ccmode (code
, op0
, op1
);
877 /* Do not output a redundant compare instruction if a compare_and_swap
878 pattern already computed the result and the machine modes are compatible. */
879 if (GET_MODE_CLASS (GET_MODE (op0
)) == MODE_CC
)
881 gcc_assert (s390_cc_modes_compatible (GET_MODE (op0
), mode
)
887 cc
= gen_rtx_REG (mode
, CC_REGNUM
);
888 emit_insn (gen_rtx_SET (VOIDmode
, cc
, gen_rtx_COMPARE (mode
, op0
, op1
)));
891 return gen_rtx_fmt_ee (code
, VOIDmode
, cc
, const0_rtx
);
894 /* Emit a SImode compare and swap instruction setting MEM to NEW_RTX if OLD
896 Return the correct condition RTL to be placed in the IF_THEN_ELSE of the
897 conditional branch testing the result. */
900 s390_emit_compare_and_swap (enum rtx_code code
, rtx old
, rtx mem
, rtx cmp
, rtx new_rtx
)
902 emit_insn (gen_sync_compare_and_swapsi (old
, mem
, cmp
, new_rtx
));
903 return s390_emit_compare (code
, gen_rtx_REG (CCZ1mode
, CC_REGNUM
), const0_rtx
);
906 /* Emit a jump instruction to TARGET. If COND is NULL_RTX, emit an
907 unconditional jump, else a conditional jump under condition COND. */
910 s390_emit_jump (rtx target
, rtx cond
)
914 target
= gen_rtx_LABEL_REF (VOIDmode
, target
);
916 target
= gen_rtx_IF_THEN_ELSE (VOIDmode
, cond
, target
, pc_rtx
);
918 insn
= gen_rtx_SET (VOIDmode
, pc_rtx
, target
);
919 emit_jump_insn (insn
);
922 /* Return branch condition mask to implement a branch
923 specified by CODE. Return -1 for invalid comparisons. */
926 s390_branch_condition_mask (rtx code
)
928 const int CC0
= 1 << 3;
929 const int CC1
= 1 << 2;
930 const int CC2
= 1 << 1;
931 const int CC3
= 1 << 0;
933 gcc_assert (GET_CODE (XEXP (code
, 0)) == REG
);
934 gcc_assert (REGNO (XEXP (code
, 0)) == CC_REGNUM
);
935 gcc_assert (XEXP (code
, 1) == const0_rtx
);
937 switch (GET_MODE (XEXP (code
, 0)))
941 switch (GET_CODE (code
))
944 case NE
: return CC1
| CC2
| CC3
;
950 switch (GET_CODE (code
))
953 case NE
: return CC0
| CC2
| CC3
;
959 switch (GET_CODE (code
))
962 case NE
: return CC0
| CC1
| CC3
;
968 switch (GET_CODE (code
))
971 case NE
: return CC0
| CC1
| CC2
;
977 switch (GET_CODE (code
))
979 case EQ
: return CC0
| CC2
;
980 case NE
: return CC1
| CC3
;
986 switch (GET_CODE (code
))
988 case LTU
: return CC2
| CC3
; /* carry */
989 case GEU
: return CC0
| CC1
; /* no carry */
995 switch (GET_CODE (code
))
997 case GTU
: return CC0
| CC1
; /* borrow */
998 case LEU
: return CC2
| CC3
; /* no borrow */
1004 switch (GET_CODE (code
))
1006 case EQ
: return CC0
| CC2
;
1007 case NE
: return CC1
| CC3
;
1008 case LTU
: return CC1
;
1009 case GTU
: return CC3
;
1010 case LEU
: return CC1
| CC2
;
1011 case GEU
: return CC2
| CC3
;
1016 switch (GET_CODE (code
))
1018 case EQ
: return CC0
;
1019 case NE
: return CC1
| CC2
| CC3
;
1020 case LTU
: return CC1
;
1021 case GTU
: return CC2
;
1022 case LEU
: return CC0
| CC1
;
1023 case GEU
: return CC0
| CC2
;
1029 switch (GET_CODE (code
))
1031 case EQ
: return CC0
;
1032 case NE
: return CC2
| CC1
| CC3
;
1033 case LTU
: return CC2
;
1034 case GTU
: return CC1
;
1035 case LEU
: return CC0
| CC2
;
1036 case GEU
: return CC0
| CC1
;
1042 switch (GET_CODE (code
))
1044 case EQ
: return CC0
;
1045 case NE
: return CC1
| CC2
| CC3
;
1046 case LT
: return CC1
| CC3
;
1047 case GT
: return CC2
;
1048 case LE
: return CC0
| CC1
| CC3
;
1049 case GE
: return CC0
| CC2
;
1055 switch (GET_CODE (code
))
1057 case EQ
: return CC0
;
1058 case NE
: return CC1
| CC2
| CC3
;
1059 case LT
: return CC1
;
1060 case GT
: return CC2
| CC3
;
1061 case LE
: return CC0
| CC1
;
1062 case GE
: return CC0
| CC2
| CC3
;
1068 switch (GET_CODE (code
))
1070 case EQ
: return CC0
;
1071 case NE
: return CC1
| CC2
| CC3
;
1072 case LT
: return CC1
;
1073 case GT
: return CC2
;
1074 case LE
: return CC0
| CC1
;
1075 case GE
: return CC0
| CC2
;
1076 case UNORDERED
: return CC3
;
1077 case ORDERED
: return CC0
| CC1
| CC2
;
1078 case UNEQ
: return CC0
| CC3
;
1079 case UNLT
: return CC1
| CC3
;
1080 case UNGT
: return CC2
| CC3
;
1081 case UNLE
: return CC0
| CC1
| CC3
;
1082 case UNGE
: return CC0
| CC2
| CC3
;
1083 case LTGT
: return CC1
| CC2
;
1089 switch (GET_CODE (code
))
1091 case EQ
: return CC0
;
1092 case NE
: return CC2
| CC1
| CC3
;
1093 case LT
: return CC2
;
1094 case GT
: return CC1
;
1095 case LE
: return CC0
| CC2
;
1096 case GE
: return CC0
| CC1
;
1097 case UNORDERED
: return CC3
;
1098 case ORDERED
: return CC0
| CC2
| CC1
;
1099 case UNEQ
: return CC0
| CC3
;
1100 case UNLT
: return CC2
| CC3
;
1101 case UNGT
: return CC1
| CC3
;
1102 case UNLE
: return CC0
| CC2
| CC3
;
1103 case UNGE
: return CC0
| CC1
| CC3
;
1104 case LTGT
: return CC2
| CC1
;
1115 /* Return branch condition mask to implement a compare and branch
1116 specified by CODE. Return -1 for invalid comparisons. */
1119 s390_compare_and_branch_condition_mask (rtx code
)
1121 const int CC0
= 1 << 3;
1122 const int CC1
= 1 << 2;
1123 const int CC2
= 1 << 1;
1125 switch (GET_CODE (code
))
1149 /* If INV is false, return assembler mnemonic string to implement
1150 a branch specified by CODE. If INV is true, return mnemonic
1151 for the corresponding inverted branch. */
1154 s390_branch_condition_mnemonic (rtx code
, int inv
)
1158 static const char *const mnemonic
[16] =
1160 NULL
, "o", "h", "nle",
1161 "l", "nhe", "lh", "ne",
1162 "e", "nlh", "he", "nl",
1163 "le", "nh", "no", NULL
1166 if (GET_CODE (XEXP (code
, 0)) == REG
1167 && REGNO (XEXP (code
, 0)) == CC_REGNUM
1168 && XEXP (code
, 1) == const0_rtx
)
1169 mask
= s390_branch_condition_mask (code
);
1171 mask
= s390_compare_and_branch_condition_mask (code
);
1173 gcc_assert (mask
>= 0);
1178 gcc_assert (mask
>= 1 && mask
<= 14);
1180 return mnemonic
[mask
];
1183 /* Return the part of op which has a value different from def.
1184 The size of the part is determined by mode.
1185 Use this function only if you already know that op really
1186 contains such a part. */
1188 unsigned HOST_WIDE_INT
1189 s390_extract_part (rtx op
, enum machine_mode mode
, int def
)
1191 unsigned HOST_WIDE_INT value
= 0;
1192 int max_parts
= HOST_BITS_PER_WIDE_INT
/ GET_MODE_BITSIZE (mode
);
1193 int part_bits
= GET_MODE_BITSIZE (mode
);
1194 unsigned HOST_WIDE_INT part_mask
1195 = ((unsigned HOST_WIDE_INT
)1 << part_bits
) - 1;
1198 for (i
= 0; i
< max_parts
; i
++)
1201 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1203 value
>>= part_bits
;
1205 if ((value
& part_mask
) != (def
& part_mask
))
1206 return value
& part_mask
;
1212 /* If OP is an integer constant of mode MODE with exactly one
1213 part of mode PART_MODE unequal to DEF, return the number of that
1214 part. Otherwise, return -1. */
1217 s390_single_part (rtx op
,
1218 enum machine_mode mode
,
1219 enum machine_mode part_mode
,
1222 unsigned HOST_WIDE_INT value
= 0;
1223 int n_parts
= GET_MODE_SIZE (mode
) / GET_MODE_SIZE (part_mode
);
1224 unsigned HOST_WIDE_INT part_mask
1225 = ((unsigned HOST_WIDE_INT
)1 << GET_MODE_BITSIZE (part_mode
)) - 1;
1228 if (GET_CODE (op
) != CONST_INT
)
1231 for (i
= 0; i
< n_parts
; i
++)
1234 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1236 value
>>= GET_MODE_BITSIZE (part_mode
);
1238 if ((value
& part_mask
) != (def
& part_mask
))
1246 return part
== -1 ? -1 : n_parts
- 1 - part
;
1249 /* Return true if IN contains a contiguous bitfield in the lower SIZE
1250 bits and no other bits are set in IN. POS and LENGTH can be used
1251 to obtain the start position and the length of the bitfield.
1253 POS gives the position of the first bit of the bitfield counting
1254 from the lowest order bit starting with zero. In order to use this
1255 value for S/390 instructions this has to be converted to "bits big
1259 s390_contiguous_bitmask_p (unsigned HOST_WIDE_INT in
, int size
,
1260 int *pos
, int *length
)
1265 unsigned HOST_WIDE_INT mask
= 1ULL;
1266 bool contiguous
= false;
1268 for (i
= 0; i
< size
; mask
<<= 1, i
++)
1292 /* Calculate a mask for all bits beyond the contiguous bits. */
1293 mask
= (-1LL & ~(((1ULL << (tmp_length
+ tmp_pos
- 1)) << 1) - 1));
1298 if (tmp_length
+ tmp_pos
- 1 > size
)
1302 *length
= tmp_length
;
1310 /* Check whether we can (and want to) split a double-word
1311 move in mode MODE from SRC to DST into two single-word
1312 moves, moving the subword FIRST_SUBWORD first. */
1315 s390_split_ok_p (rtx dst
, rtx src
, enum machine_mode mode
, int first_subword
)
1317 /* Floating point registers cannot be split. */
1318 if (FP_REG_P (src
) || FP_REG_P (dst
))
1321 /* We don't need to split if operands are directly accessible. */
1322 if (s_operand (src
, mode
) || s_operand (dst
, mode
))
1325 /* Non-offsettable memory references cannot be split. */
1326 if ((GET_CODE (src
) == MEM
&& !offsettable_memref_p (src
))
1327 || (GET_CODE (dst
) == MEM
&& !offsettable_memref_p (dst
)))
1330 /* Moving the first subword must not clobber a register
1331 needed to move the second subword. */
1332 if (register_operand (dst
, mode
))
1334 rtx subreg
= operand_subword (dst
, first_subword
, 0, mode
);
1335 if (reg_overlap_mentioned_p (subreg
, src
))
1342 /* Return true if it can be proven that [MEM1, MEM1 + SIZE]
1343 and [MEM2, MEM2 + SIZE] do overlap and false
1347 s390_overlap_p (rtx mem1
, rtx mem2
, HOST_WIDE_INT size
)
1349 rtx addr1
, addr2
, addr_delta
;
1350 HOST_WIDE_INT delta
;
1352 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1358 addr1
= XEXP (mem1
, 0);
1359 addr2
= XEXP (mem2
, 0);
1361 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1363 /* This overlapping check is used by peepholes merging memory block operations.
1364 Overlapping operations would otherwise be recognized by the S/390 hardware
1365 and would fall back to a slower implementation. Allowing overlapping
1366 operations would lead to slow code but not to wrong code. Therefore we are
1367 somewhat optimistic if we cannot prove that the memory blocks are
1369 That's why we return false here although this may accept operations on
1370 overlapping memory areas. */
1371 if (!addr_delta
|| GET_CODE (addr_delta
) != CONST_INT
)
1374 delta
= INTVAL (addr_delta
);
1377 || (delta
> 0 && delta
< size
)
1378 || (delta
< 0 && -delta
< size
))
1384 /* Check whether the address of memory reference MEM2 equals exactly
1385 the address of memory reference MEM1 plus DELTA. Return true if
1386 we can prove this to be the case, false otherwise. */
1389 s390_offset_p (rtx mem1
, rtx mem2
, rtx delta
)
1391 rtx addr1
, addr2
, addr_delta
;
1393 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1396 addr1
= XEXP (mem1
, 0);
1397 addr2
= XEXP (mem2
, 0);
1399 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1400 if (!addr_delta
|| !rtx_equal_p (addr_delta
, delta
))
1406 /* Expand logical operator CODE in mode MODE with operands OPERANDS. */
1409 s390_expand_logical_operator (enum rtx_code code
, enum machine_mode mode
,
1412 enum machine_mode wmode
= mode
;
1413 rtx dst
= operands
[0];
1414 rtx src1
= operands
[1];
1415 rtx src2
= operands
[2];
1418 /* If we cannot handle the operation directly, use a temp register. */
1419 if (!s390_logical_operator_ok_p (operands
))
1420 dst
= gen_reg_rtx (mode
);
1422 /* QImode and HImode patterns make sense only if we have a destination
1423 in memory. Otherwise perform the operation in SImode. */
1424 if ((mode
== QImode
|| mode
== HImode
) && GET_CODE (dst
) != MEM
)
1427 /* Widen operands if required. */
1430 if (GET_CODE (dst
) == SUBREG
1431 && (tem
= simplify_subreg (wmode
, dst
, mode
, 0)) != 0)
1433 else if (REG_P (dst
))
1434 dst
= gen_rtx_SUBREG (wmode
, dst
, 0);
1436 dst
= gen_reg_rtx (wmode
);
1438 if (GET_CODE (src1
) == SUBREG
1439 && (tem
= simplify_subreg (wmode
, src1
, mode
, 0)) != 0)
1441 else if (GET_MODE (src1
) != VOIDmode
)
1442 src1
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src1
), 0);
1444 if (GET_CODE (src2
) == SUBREG
1445 && (tem
= simplify_subreg (wmode
, src2
, mode
, 0)) != 0)
1447 else if (GET_MODE (src2
) != VOIDmode
)
1448 src2
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src2
), 0);
1451 /* Emit the instruction. */
1452 op
= gen_rtx_SET (VOIDmode
, dst
, gen_rtx_fmt_ee (code
, wmode
, src1
, src2
));
1453 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
1454 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clob
)));
1456 /* Fix up the destination if needed. */
1457 if (dst
!= operands
[0])
1458 emit_move_insn (operands
[0], gen_lowpart (mode
, dst
));
1461 /* Check whether OPERANDS are OK for a logical operation (AND, IOR, XOR). */
1464 s390_logical_operator_ok_p (rtx
*operands
)
1466 /* If the destination operand is in memory, it needs to coincide
1467 with one of the source operands. After reload, it has to be
1468 the first source operand. */
1469 if (GET_CODE (operands
[0]) == MEM
)
1470 return rtx_equal_p (operands
[0], operands
[1])
1471 || (!reload_completed
&& rtx_equal_p (operands
[0], operands
[2]));
1476 /* Narrow logical operation CODE of memory operand MEMOP with immediate
1477 operand IMMOP to switch from SS to SI type instructions. */
1480 s390_narrow_logical_operator (enum rtx_code code
, rtx
*memop
, rtx
*immop
)
1482 int def
= code
== AND
? -1 : 0;
1486 gcc_assert (GET_CODE (*memop
) == MEM
);
1487 gcc_assert (!MEM_VOLATILE_P (*memop
));
1489 mask
= s390_extract_part (*immop
, QImode
, def
);
1490 part
= s390_single_part (*immop
, GET_MODE (*memop
), QImode
, def
);
1491 gcc_assert (part
>= 0);
1493 *memop
= adjust_address (*memop
, QImode
, part
);
1494 *immop
= gen_int_mode (mask
, QImode
);
1498 /* How to allocate a 'struct machine_function'. */
1500 static struct machine_function
*
1501 s390_init_machine_status (void)
1503 return ggc_alloc_cleared_machine_function ();
1507 s390_option_override (void)
1509 /* Set up function hooks. */
1510 init_machine_status
= s390_init_machine_status
;
1512 /* Architecture mode defaults according to ABI. */
1513 if (!(target_flags_explicit
& MASK_ZARCH
))
1516 target_flags
|= MASK_ZARCH
;
1518 target_flags
&= ~MASK_ZARCH
;
1521 /* Set the march default in case it hasn't been specified on
1523 if (s390_arch
== PROCESSOR_max
)
1525 s390_arch_string
= TARGET_ZARCH
? "z900" : "g5";
1526 s390_arch
= TARGET_ZARCH
? PROCESSOR_2064_Z900
: PROCESSOR_9672_G5
;
1527 s390_arch_flags
= processor_flags_table
[(int)s390_arch
];
1530 /* Determine processor to tune for. */
1531 if (s390_tune
== PROCESSOR_max
)
1533 s390_tune
= s390_arch
;
1534 s390_tune_flags
= s390_arch_flags
;
1537 /* Sanity checks. */
1538 if (TARGET_ZARCH
&& !TARGET_CPU_ZARCH
)
1539 error ("z/Architecture mode not supported on %s", s390_arch_string
);
1540 if (TARGET_64BIT
&& !TARGET_ZARCH
)
1541 error ("64-bit ABI not supported in ESA/390 mode");
1543 /* Use hardware DFP if available and not explicitly disabled by
1544 user. E.g. with -m31 -march=z10 -mzarch */
1545 if (!(target_flags_explicit
& MASK_HARD_DFP
) && TARGET_DFP
)
1546 target_flags
|= MASK_HARD_DFP
;
1548 if (TARGET_HARD_DFP
&& !TARGET_DFP
)
1550 if (target_flags_explicit
& MASK_HARD_DFP
)
1552 if (!TARGET_CPU_DFP
)
1553 error ("hardware decimal floating point instructions"
1554 " not available on %s", s390_arch_string
);
1556 error ("hardware decimal floating point instructions"
1557 " not available in ESA/390 mode");
1560 target_flags
&= ~MASK_HARD_DFP
;
1563 if ((target_flags_explicit
& MASK_SOFT_FLOAT
) && TARGET_SOFT_FLOAT
)
1565 if ((target_flags_explicit
& MASK_HARD_DFP
) && TARGET_HARD_DFP
)
1566 error ("-mhard-dfp can%'t be used in conjunction with -msoft-float");
1568 target_flags
&= ~MASK_HARD_DFP
;
1571 /* Set processor cost function. */
1574 case PROCESSOR_2084_Z990
:
1575 s390_cost
= &z990_cost
;
1577 case PROCESSOR_2094_Z9_109
:
1578 s390_cost
= &z9_109_cost
;
1580 case PROCESSOR_2097_Z10
:
1581 s390_cost
= &z10_cost
;
1582 case PROCESSOR_2817_Z196
:
1583 s390_cost
= &z196_cost
;
1586 s390_cost
= &z900_cost
;
1589 if (TARGET_BACKCHAIN
&& TARGET_PACKED_STACK
&& TARGET_HARD_FLOAT
)
1590 error ("-mbackchain -mpacked-stack -mhard-float are not supported "
1593 if (s390_stack_size
)
1595 if (s390_stack_guard
>= s390_stack_size
)
1596 error ("stack size must be greater than the stack guard value");
1597 else if (s390_stack_size
> 1 << 16)
1598 error ("stack size must not be greater than 64k");
1600 else if (s390_stack_guard
)
1601 error ("-mstack-guard implies use of -mstack-size");
1603 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1604 if (!(target_flags_explicit
& MASK_LONG_DOUBLE_128
))
1605 target_flags
|= MASK_LONG_DOUBLE_128
;
1608 if (s390_tune
== PROCESSOR_2097_Z10
1609 || s390_tune
== PROCESSOR_2817_Z196
)
1611 maybe_set_param_value (PARAM_MAX_UNROLLED_INSNS
, 100,
1612 global_options
.x_param_values
,
1613 global_options_set
.x_param_values
);
1614 maybe_set_param_value (PARAM_MAX_UNROLL_TIMES
, 32,
1615 global_options
.x_param_values
,
1616 global_options_set
.x_param_values
);
1617 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEELED_INSNS
, 2000,
1618 global_options
.x_param_values
,
1619 global_options_set
.x_param_values
);
1620 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEEL_TIMES
, 64,
1621 global_options
.x_param_values
,
1622 global_options_set
.x_param_values
);
1625 maybe_set_param_value (PARAM_MAX_PENDING_LIST_LENGTH
, 256,
1626 global_options
.x_param_values
,
1627 global_options_set
.x_param_values
);
1628 /* values for loop prefetching */
1629 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE
, 256,
1630 global_options
.x_param_values
,
1631 global_options_set
.x_param_values
);
1632 maybe_set_param_value (PARAM_L1_CACHE_SIZE
, 128,
1633 global_options
.x_param_values
,
1634 global_options_set
.x_param_values
);
1635 /* s390 has more than 2 levels and the size is much larger. Since
1636 we are always running virtualized assume that we only get a small
1637 part of the caches above l1. */
1638 maybe_set_param_value (PARAM_L2_CACHE_SIZE
, 1500,
1639 global_options
.x_param_values
,
1640 global_options_set
.x_param_values
);
1641 maybe_set_param_value (PARAM_PREFETCH_MIN_INSN_TO_MEM_RATIO
, 2,
1642 global_options
.x_param_values
,
1643 global_options_set
.x_param_values
);
1644 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
, 6,
1645 global_options
.x_param_values
,
1646 global_options_set
.x_param_values
);
1648 /* This cannot reside in s390_option_optimization_table since HAVE_prefetch
1649 requires the arch flags to be evaluated already. Since prefetching
1650 is beneficial on s390, we enable it if available. */
1651 if (flag_prefetch_loop_arrays
< 0 && HAVE_prefetch
&& optimize
>= 3)
1652 flag_prefetch_loop_arrays
= 1;
1655 /* Map for smallest class containing reg regno. */
1657 const enum reg_class regclass_map
[FIRST_PSEUDO_REGISTER
] =
1658 { GENERAL_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1659 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1660 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1661 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1662 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1663 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1664 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1665 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1666 ADDR_REGS
, CC_REGS
, ADDR_REGS
, ADDR_REGS
,
1667 ACCESS_REGS
, ACCESS_REGS
1670 /* Return attribute type of insn. */
1672 static enum attr_type
1673 s390_safe_attr_type (rtx insn
)
1675 if (recog_memoized (insn
) >= 0)
1676 return get_attr_type (insn
);
1681 /* Return true if DISP is a valid short displacement. */
1684 s390_short_displacement (rtx disp
)
1686 /* No displacement is OK. */
1690 /* Without the long displacement facility we don't need to
1691 distingiush between long and short displacement. */
1692 if (!TARGET_LONG_DISPLACEMENT
)
1695 /* Integer displacement in range. */
1696 if (GET_CODE (disp
) == CONST_INT
)
1697 return INTVAL (disp
) >= 0 && INTVAL (disp
) < 4096;
1699 /* GOT offset is not OK, the GOT can be large. */
1700 if (GET_CODE (disp
) == CONST
1701 && GET_CODE (XEXP (disp
, 0)) == UNSPEC
1702 && (XINT (XEXP (disp
, 0), 1) == UNSPEC_GOT
1703 || XINT (XEXP (disp
, 0), 1) == UNSPEC_GOTNTPOFF
))
1706 /* All other symbolic constants are literal pool references,
1707 which are OK as the literal pool must be small. */
1708 if (GET_CODE (disp
) == CONST
)
1714 /* Decompose a RTL expression ADDR for a memory address into
1715 its components, returned in OUT.
1717 Returns false if ADDR is not a valid memory address, true
1718 otherwise. If OUT is NULL, don't return the components,
1719 but check for validity only.
1721 Note: Only addresses in canonical form are recognized.
1722 LEGITIMIZE_ADDRESS should convert non-canonical forms to the
1723 canonical form so that they will be recognized. */
1726 s390_decompose_address (rtx addr
, struct s390_address
*out
)
1728 HOST_WIDE_INT offset
= 0;
1729 rtx base
= NULL_RTX
;
1730 rtx indx
= NULL_RTX
;
1731 rtx disp
= NULL_RTX
;
1733 bool pointer
= false;
1734 bool base_ptr
= false;
1735 bool indx_ptr
= false;
1736 bool literal_pool
= false;
1738 /* We may need to substitute the literal pool base register into the address
1739 below. However, at this point we do not know which register is going to
1740 be used as base, so we substitute the arg pointer register. This is going
1741 to be treated as holding a pointer below -- it shouldn't be used for any
1743 rtx fake_pool_base
= gen_rtx_REG (Pmode
, ARG_POINTER_REGNUM
);
1745 /* Decompose address into base + index + displacement. */
1747 if (GET_CODE (addr
) == REG
|| GET_CODE (addr
) == UNSPEC
)
1750 else if (GET_CODE (addr
) == PLUS
)
1752 rtx op0
= XEXP (addr
, 0);
1753 rtx op1
= XEXP (addr
, 1);
1754 enum rtx_code code0
= GET_CODE (op0
);
1755 enum rtx_code code1
= GET_CODE (op1
);
1757 if (code0
== REG
|| code0
== UNSPEC
)
1759 if (code1
== REG
|| code1
== UNSPEC
)
1761 indx
= op0
; /* index + base */
1767 base
= op0
; /* base + displacement */
1772 else if (code0
== PLUS
)
1774 indx
= XEXP (op0
, 0); /* index + base + disp */
1775 base
= XEXP (op0
, 1);
1786 disp
= addr
; /* displacement */
1788 /* Extract integer part of displacement. */
1792 if (GET_CODE (disp
) == CONST_INT
)
1794 offset
= INTVAL (disp
);
1797 else if (GET_CODE (disp
) == CONST
1798 && GET_CODE (XEXP (disp
, 0)) == PLUS
1799 && GET_CODE (XEXP (XEXP (disp
, 0), 1)) == CONST_INT
)
1801 offset
= INTVAL (XEXP (XEXP (disp
, 0), 1));
1802 disp
= XEXP (XEXP (disp
, 0), 0);
1806 /* Strip off CONST here to avoid special case tests later. */
1807 if (disp
&& GET_CODE (disp
) == CONST
)
1808 disp
= XEXP (disp
, 0);
1810 /* We can convert literal pool addresses to
1811 displacements by basing them off the base register. */
1812 if (disp
&& GET_CODE (disp
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (disp
))
1814 /* Either base or index must be free to hold the base register. */
1816 base
= fake_pool_base
, literal_pool
= true;
1818 indx
= fake_pool_base
, literal_pool
= true;
1822 /* Mark up the displacement. */
1823 disp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, disp
),
1824 UNSPEC_LTREL_OFFSET
);
1827 /* Validate base register. */
1830 if (GET_CODE (base
) == UNSPEC
)
1831 switch (XINT (base
, 1))
1835 disp
= gen_rtx_UNSPEC (Pmode
,
1836 gen_rtvec (1, XVECEXP (base
, 0, 0)),
1837 UNSPEC_LTREL_OFFSET
);
1841 base
= XVECEXP (base
, 0, 1);
1844 case UNSPEC_LTREL_BASE
:
1845 if (XVECLEN (base
, 0) == 1)
1846 base
= fake_pool_base
, literal_pool
= true;
1848 base
= XVECEXP (base
, 0, 1);
1856 || (GET_MODE (base
) != SImode
1857 && GET_MODE (base
) != Pmode
))
1860 if (REGNO (base
) == STACK_POINTER_REGNUM
1861 || REGNO (base
) == FRAME_POINTER_REGNUM
1862 || ((reload_completed
|| reload_in_progress
)
1863 && frame_pointer_needed
1864 && REGNO (base
) == HARD_FRAME_POINTER_REGNUM
)
1865 || REGNO (base
) == ARG_POINTER_REGNUM
1867 && REGNO (base
) == PIC_OFFSET_TABLE_REGNUM
))
1868 pointer
= base_ptr
= true;
1870 if ((reload_completed
|| reload_in_progress
)
1871 && base
== cfun
->machine
->base_reg
)
1872 pointer
= base_ptr
= literal_pool
= true;
1875 /* Validate index register. */
1878 if (GET_CODE (indx
) == UNSPEC
)
1879 switch (XINT (indx
, 1))
1883 disp
= gen_rtx_UNSPEC (Pmode
,
1884 gen_rtvec (1, XVECEXP (indx
, 0, 0)),
1885 UNSPEC_LTREL_OFFSET
);
1889 indx
= XVECEXP (indx
, 0, 1);
1892 case UNSPEC_LTREL_BASE
:
1893 if (XVECLEN (indx
, 0) == 1)
1894 indx
= fake_pool_base
, literal_pool
= true;
1896 indx
= XVECEXP (indx
, 0, 1);
1904 || (GET_MODE (indx
) != SImode
1905 && GET_MODE (indx
) != Pmode
))
1908 if (REGNO (indx
) == STACK_POINTER_REGNUM
1909 || REGNO (indx
) == FRAME_POINTER_REGNUM
1910 || ((reload_completed
|| reload_in_progress
)
1911 && frame_pointer_needed
1912 && REGNO (indx
) == HARD_FRAME_POINTER_REGNUM
)
1913 || REGNO (indx
) == ARG_POINTER_REGNUM
1915 && REGNO (indx
) == PIC_OFFSET_TABLE_REGNUM
))
1916 pointer
= indx_ptr
= true;
1918 if ((reload_completed
|| reload_in_progress
)
1919 && indx
== cfun
->machine
->base_reg
)
1920 pointer
= indx_ptr
= literal_pool
= true;
1923 /* Prefer to use pointer as base, not index. */
1924 if (base
&& indx
&& !base_ptr
1925 && (indx_ptr
|| (!REG_POINTER (base
) && REG_POINTER (indx
))))
1932 /* Validate displacement. */
1935 /* If virtual registers are involved, the displacement will change later
1936 anyway as the virtual registers get eliminated. This could make a
1937 valid displacement invalid, but it is more likely to make an invalid
1938 displacement valid, because we sometimes access the register save area
1939 via negative offsets to one of those registers.
1940 Thus we don't check the displacement for validity here. If after
1941 elimination the displacement turns out to be invalid after all,
1942 this is fixed up by reload in any case. */
1943 if (base
!= arg_pointer_rtx
1944 && indx
!= arg_pointer_rtx
1945 && base
!= return_address_pointer_rtx
1946 && indx
!= return_address_pointer_rtx
1947 && base
!= frame_pointer_rtx
1948 && indx
!= frame_pointer_rtx
1949 && base
!= virtual_stack_vars_rtx
1950 && indx
!= virtual_stack_vars_rtx
)
1951 if (!DISP_IN_RANGE (offset
))
1956 /* All the special cases are pointers. */
1959 /* In the small-PIC case, the linker converts @GOT
1960 and @GOTNTPOFF offsets to possible displacements. */
1961 if (GET_CODE (disp
) == UNSPEC
1962 && (XINT (disp
, 1) == UNSPEC_GOT
1963 || XINT (disp
, 1) == UNSPEC_GOTNTPOFF
)
1969 /* Accept pool label offsets. */
1970 else if (GET_CODE (disp
) == UNSPEC
1971 && XINT (disp
, 1) == UNSPEC_POOL_OFFSET
)
1974 /* Accept literal pool references. */
1975 else if (GET_CODE (disp
) == UNSPEC
1976 && XINT (disp
, 1) == UNSPEC_LTREL_OFFSET
)
1978 /* In case CSE pulled a non literal pool reference out of
1979 the pool we have to reject the address. This is
1980 especially important when loading the GOT pointer on non
1981 zarch CPUs. In this case the literal pool contains an lt
1982 relative offset to the _GLOBAL_OFFSET_TABLE_ label which
1983 will most likely exceed the displacement. */
1984 if (GET_CODE (XVECEXP (disp
, 0, 0)) != SYMBOL_REF
1985 || !CONSTANT_POOL_ADDRESS_P (XVECEXP (disp
, 0, 0)))
1988 orig_disp
= gen_rtx_CONST (Pmode
, disp
);
1991 /* If we have an offset, make sure it does not
1992 exceed the size of the constant pool entry. */
1993 rtx sym
= XVECEXP (disp
, 0, 0);
1994 if (offset
>= GET_MODE_SIZE (get_pool_mode (sym
)))
1997 orig_disp
= plus_constant (orig_disp
, offset
);
2012 out
->disp
= orig_disp
;
2013 out
->pointer
= pointer
;
2014 out
->literal_pool
= literal_pool
;
2020 /* Decompose a RTL expression OP for a shift count into its components,
2021 and return the base register in BASE and the offset in OFFSET.
2023 Return true if OP is a valid shift count, false if not. */
2026 s390_decompose_shift_count (rtx op
, rtx
*base
, HOST_WIDE_INT
*offset
)
2028 HOST_WIDE_INT off
= 0;
2030 /* We can have an integer constant, an address register,
2031 or a sum of the two. */
2032 if (GET_CODE (op
) == CONST_INT
)
2037 if (op
&& GET_CODE (op
) == PLUS
&& GET_CODE (XEXP (op
, 1)) == CONST_INT
)
2039 off
= INTVAL (XEXP (op
, 1));
2042 while (op
&& GET_CODE (op
) == SUBREG
)
2043 op
= SUBREG_REG (op
);
2045 if (op
&& GET_CODE (op
) != REG
)
2057 /* Return true if CODE is a valid address without index. */
2060 s390_legitimate_address_without_index_p (rtx op
)
2062 struct s390_address addr
;
2064 if (!s390_decompose_address (XEXP (op
, 0), &addr
))
2073 /* Return true if ADDR is of kind symbol_ref or symbol_ref + const_int
2074 and return these parts in SYMREF and ADDEND. You can pass NULL in
2075 SYMREF and/or ADDEND if you are not interested in these values.
2076 Literal pool references are *not* considered symbol references. */
2079 s390_symref_operand_p (rtx addr
, rtx
*symref
, HOST_WIDE_INT
*addend
)
2081 HOST_WIDE_INT tmpaddend
= 0;
2083 if (GET_CODE (addr
) == CONST
)
2084 addr
= XEXP (addr
, 0);
2086 if (GET_CODE (addr
) == PLUS
)
2088 if (GET_CODE (XEXP (addr
, 0)) == SYMBOL_REF
2089 && !CONSTANT_POOL_ADDRESS_P (XEXP (addr
, 0))
2090 && CONST_INT_P (XEXP (addr
, 1)))
2092 tmpaddend
= INTVAL (XEXP (addr
, 1));
2093 addr
= XEXP (addr
, 0);
2099 if (GET_CODE (addr
) != SYMBOL_REF
|| CONSTANT_POOL_ADDRESS_P (addr
))
2105 *addend
= tmpaddend
;
2111 /* Return true if the address in OP is valid for constraint letter C
2112 if wrapped in a MEM rtx. Set LIT_POOL_OK to true if it literal
2113 pool MEMs should be accepted. Only the Q, R, S, T constraint
2114 letters are allowed for C. */
2117 s390_check_qrst_address (char c
, rtx op
, bool lit_pool_ok
)
2119 struct s390_address addr
;
2120 bool decomposed
= false;
2122 /* This check makes sure that no symbolic address (except literal
2123 pool references) are accepted by the R or T constraints. */
2124 if (s390_symref_operand_p (op
, NULL
, NULL
))
2127 /* Ensure literal pool references are only accepted if LIT_POOL_OK. */
2130 if (!s390_decompose_address (op
, &addr
))
2132 if (addr
.literal_pool
)
2139 case 'Q': /* no index short displacement */
2140 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2144 if (!s390_short_displacement (addr
.disp
))
2148 case 'R': /* with index short displacement */
2149 if (TARGET_LONG_DISPLACEMENT
)
2151 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2153 if (!s390_short_displacement (addr
.disp
))
2156 /* Any invalid address here will be fixed up by reload,
2157 so accept it for the most generic constraint. */
2160 case 'S': /* no index long displacement */
2161 if (!TARGET_LONG_DISPLACEMENT
)
2163 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2167 if (s390_short_displacement (addr
.disp
))
2171 case 'T': /* with index long displacement */
2172 if (!TARGET_LONG_DISPLACEMENT
)
2174 /* Any invalid address here will be fixed up by reload,
2175 so accept it for the most generic constraint. */
2176 if ((decomposed
|| s390_decompose_address (op
, &addr
))
2177 && s390_short_displacement (addr
.disp
))
2187 /* Evaluates constraint strings described by the regular expression
2188 ([A|B|Z](Q|R|S|T))|U|W|Y and returns 1 if OP is a valid operand for
2189 the constraint given in STR, or 0 else. */
2192 s390_mem_constraint (const char *str
, rtx op
)
2199 /* Check for offsettable variants of memory constraints. */
2200 if (!MEM_P (op
) || MEM_VOLATILE_P (op
))
2202 if ((reload_completed
|| reload_in_progress
)
2203 ? !offsettable_memref_p (op
) : !offsettable_nonstrict_memref_p (op
))
2205 return s390_check_qrst_address (str
[1], XEXP (op
, 0), true);
2207 /* Check for non-literal-pool variants of memory constraints. */
2210 return s390_check_qrst_address (str
[1], XEXP (op
, 0), false);
2215 if (GET_CODE (op
) != MEM
)
2217 return s390_check_qrst_address (c
, XEXP (op
, 0), true);
2219 return (s390_check_qrst_address ('Q', op
, true)
2220 || s390_check_qrst_address ('R', op
, true));
2222 return (s390_check_qrst_address ('S', op
, true)
2223 || s390_check_qrst_address ('T', op
, true));
2225 /* Simply check for the basic form of a shift count. Reload will
2226 take care of making sure we have a proper base register. */
2227 if (!s390_decompose_shift_count (op
, NULL
, NULL
))
2231 return s390_check_qrst_address (str
[1], op
, true);
2239 /* Evaluates constraint strings starting with letter O. Input
2240 parameter C is the second letter following the "O" in the constraint
2241 string. Returns 1 if VALUE meets the respective constraint and 0
2245 s390_O_constraint_str (const char c
, HOST_WIDE_INT value
)
2253 return trunc_int_for_mode (value
, SImode
) == value
;
2257 || s390_single_part (GEN_INT (value
), DImode
, SImode
, 0) == 1;
2260 return s390_single_part (GEN_INT (value
- 1), DImode
, SImode
, -1) == 1;
2268 /* Evaluates constraint strings starting with letter N. Parameter STR
2269 contains the letters following letter "N" in the constraint string.
2270 Returns true if VALUE matches the constraint. */
2273 s390_N_constraint_str (const char *str
, HOST_WIDE_INT value
)
2275 enum machine_mode mode
, part_mode
;
2277 int part
, part_goal
;
2283 part_goal
= str
[0] - '0';
2327 if (GET_MODE_SIZE (mode
) <= GET_MODE_SIZE (part_mode
))
2330 part
= s390_single_part (GEN_INT (value
), mode
, part_mode
, def
);
2333 if (part_goal
!= -1 && part_goal
!= part
)
2340 /* Returns true if the input parameter VALUE is a float zero. */
2343 s390_float_const_zero_p (rtx value
)
2345 return (GET_MODE_CLASS (GET_MODE (value
)) == MODE_FLOAT
2346 && value
== CONST0_RTX (GET_MODE (value
)));
2349 /* Implement TARGET_REGISTER_MOVE_COST. */
2352 s390_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2353 reg_class_t from
, reg_class_t to
)
2355 /* On s390, copy between fprs and gprs is expensive. */
2356 if ((reg_classes_intersect_p (from
, GENERAL_REGS
)
2357 && reg_classes_intersect_p (to
, FP_REGS
))
2358 || (reg_classes_intersect_p (from
, FP_REGS
)
2359 && reg_classes_intersect_p (to
, GENERAL_REGS
)))
2365 /* Implement TARGET_MEMORY_MOVE_COST. */
2368 s390_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2369 reg_class_t rclass ATTRIBUTE_UNUSED
,
2370 bool in ATTRIBUTE_UNUSED
)
2375 /* Compute a (partial) cost for rtx X. Return true if the complete
2376 cost has been computed, and false if subexpressions should be
2377 scanned. In either case, *TOTAL contains the cost result.
2378 CODE contains GET_CODE (x), OUTER_CODE contains the code
2379 of the superexpression of x. */
2382 s390_rtx_costs (rtx x
, int code
, int outer_code
, int opno ATTRIBUTE_UNUSED
,
2383 int *total
, bool speed ATTRIBUTE_UNUSED
)
2406 *total
= COSTS_N_INSNS (1);
2411 *total
= COSTS_N_INSNS (1);
2415 switch (GET_MODE (x
))
2419 rtx left
= XEXP (x
, 0);
2420 rtx right
= XEXP (x
, 1);
2421 if (GET_CODE (right
) == CONST_INT
2422 && CONST_OK_FOR_K (INTVAL (right
)))
2423 *total
= s390_cost
->mhi
;
2424 else if (GET_CODE (left
) == SIGN_EXTEND
)
2425 *total
= s390_cost
->mh
;
2427 *total
= s390_cost
->ms
; /* msr, ms, msy */
2432 rtx left
= XEXP (x
, 0);
2433 rtx right
= XEXP (x
, 1);
2436 if (GET_CODE (right
) == CONST_INT
2437 && CONST_OK_FOR_K (INTVAL (right
)))
2438 *total
= s390_cost
->mghi
;
2439 else if (GET_CODE (left
) == SIGN_EXTEND
)
2440 *total
= s390_cost
->msgf
;
2442 *total
= s390_cost
->msg
; /* msgr, msg */
2444 else /* TARGET_31BIT */
2446 if (GET_CODE (left
) == SIGN_EXTEND
2447 && GET_CODE (right
) == SIGN_EXTEND
)
2448 /* mulsidi case: mr, m */
2449 *total
= s390_cost
->m
;
2450 else if (GET_CODE (left
) == ZERO_EXTEND
2451 && GET_CODE (right
) == ZERO_EXTEND
2452 && TARGET_CPU_ZARCH
)
2453 /* umulsidi case: ml, mlr */
2454 *total
= s390_cost
->ml
;
2456 /* Complex calculation is required. */
2457 *total
= COSTS_N_INSNS (40);
2463 *total
= s390_cost
->mult_df
;
2466 *total
= s390_cost
->mxbr
;
2474 switch (GET_MODE (x
))
2477 *total
= s390_cost
->madbr
;
2480 *total
= s390_cost
->maebr
;
2485 /* Negate in the third argument is free: FMSUB. */
2486 if (GET_CODE (XEXP (x
, 2)) == NEG
)
2488 *total
+= (rtx_cost (XEXP (x
, 0), FMA
, 0, speed
)
2489 + rtx_cost (XEXP (x
, 1), FMA
, 1, speed
)
2490 + rtx_cost (XEXP (XEXP (x
, 2), 0), FMA
, 2, speed
));
2497 if (GET_MODE (x
) == TImode
) /* 128 bit division */
2498 *total
= s390_cost
->dlgr
;
2499 else if (GET_MODE (x
) == DImode
)
2501 rtx right
= XEXP (x
, 1);
2502 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2503 *total
= s390_cost
->dlr
;
2504 else /* 64 by 64 bit division */
2505 *total
= s390_cost
->dlgr
;
2507 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2508 *total
= s390_cost
->dlr
;
2513 if (GET_MODE (x
) == DImode
)
2515 rtx right
= XEXP (x
, 1);
2516 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2518 *total
= s390_cost
->dsgfr
;
2520 *total
= s390_cost
->dr
;
2521 else /* 64 by 64 bit division */
2522 *total
= s390_cost
->dsgr
;
2524 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2525 *total
= s390_cost
->dlr
;
2526 else if (GET_MODE (x
) == SFmode
)
2528 *total
= s390_cost
->debr
;
2530 else if (GET_MODE (x
) == DFmode
)
2532 *total
= s390_cost
->ddbr
;
2534 else if (GET_MODE (x
) == TFmode
)
2536 *total
= s390_cost
->dxbr
;
2541 if (GET_MODE (x
) == SFmode
)
2542 *total
= s390_cost
->sqebr
;
2543 else if (GET_MODE (x
) == DFmode
)
2544 *total
= s390_cost
->sqdbr
;
2546 *total
= s390_cost
->sqxbr
;
2551 if (outer_code
== MULT
|| outer_code
== DIV
|| outer_code
== MOD
2552 || outer_code
== PLUS
|| outer_code
== MINUS
2553 || outer_code
== COMPARE
)
2558 *total
= COSTS_N_INSNS (1);
2559 if (GET_CODE (XEXP (x
, 0)) == AND
2560 && GET_CODE (XEXP (x
, 1)) == CONST_INT
2561 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
)
2563 rtx op0
= XEXP (XEXP (x
, 0), 0);
2564 rtx op1
= XEXP (XEXP (x
, 0), 1);
2565 rtx op2
= XEXP (x
, 1);
2567 if (memory_operand (op0
, GET_MODE (op0
))
2568 && s390_tm_ccmode (op1
, op2
, 0) != VOIDmode
)
2570 if (register_operand (op0
, GET_MODE (op0
))
2571 && s390_tm_ccmode (op1
, op2
, 1) != VOIDmode
)
2581 /* Return the cost of an address rtx ADDR. */
2584 s390_address_cost (rtx addr
, bool speed ATTRIBUTE_UNUSED
)
2586 struct s390_address ad
;
2587 if (!s390_decompose_address (addr
, &ad
))
2590 return ad
.indx
? COSTS_N_INSNS (1) + 1 : COSTS_N_INSNS (1);
2593 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
2594 otherwise return 0. */
2597 tls_symbolic_operand (rtx op
)
2599 if (GET_CODE (op
) != SYMBOL_REF
)
2601 return SYMBOL_REF_TLS_MODEL (op
);
2604 /* Split DImode access register reference REG (on 64-bit) into its constituent
2605 low and high parts, and store them into LO and HI. Note that gen_lowpart/
2606 gen_highpart cannot be used as they assume all registers are word-sized,
2607 while our access registers have only half that size. */
2610 s390_split_access_reg (rtx reg
, rtx
*lo
, rtx
*hi
)
2612 gcc_assert (TARGET_64BIT
);
2613 gcc_assert (ACCESS_REG_P (reg
));
2614 gcc_assert (GET_MODE (reg
) == DImode
);
2615 gcc_assert (!(REGNO (reg
) & 1));
2617 *lo
= gen_rtx_REG (SImode
, REGNO (reg
) + 1);
2618 *hi
= gen_rtx_REG (SImode
, REGNO (reg
));
2621 /* Return true if OP contains a symbol reference */
2624 symbolic_reference_mentioned_p (rtx op
)
2629 if (GET_CODE (op
) == SYMBOL_REF
|| GET_CODE (op
) == LABEL_REF
)
2632 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2633 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2639 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2640 if (symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2644 else if (fmt
[i
] == 'e' && symbolic_reference_mentioned_p (XEXP (op
, i
)))
2651 /* Return true if OP contains a reference to a thread-local symbol. */
2654 tls_symbolic_reference_mentioned_p (rtx op
)
2659 if (GET_CODE (op
) == SYMBOL_REF
)
2660 return tls_symbolic_operand (op
);
2662 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2663 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2669 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2670 if (tls_symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2674 else if (fmt
[i
] == 'e' && tls_symbolic_reference_mentioned_p (XEXP (op
, i
)))
2682 /* Return true if OP is a legitimate general operand when
2683 generating PIC code. It is given that flag_pic is on
2684 and that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2687 legitimate_pic_operand_p (rtx op
)
2689 /* Accept all non-symbolic constants. */
2690 if (!SYMBOLIC_CONST (op
))
2693 /* Reject everything else; must be handled
2694 via emit_symbolic_move. */
2698 /* Returns true if the constant value OP is a legitimate general operand.
2699 It is given that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2702 s390_legitimate_constant_p (enum machine_mode mode
, rtx op
)
2704 /* Accept all non-symbolic constants. */
2705 if (!SYMBOLIC_CONST (op
))
2708 /* Accept immediate LARL operands. */
2709 if (TARGET_CPU_ZARCH
&& larl_operand (op
, mode
))
2712 /* Thread-local symbols are never legal constants. This is
2713 so that emit_call knows that computing such addresses
2714 might require a function call. */
2715 if (TLS_SYMBOLIC_CONST (op
))
2718 /* In the PIC case, symbolic constants must *not* be
2719 forced into the literal pool. We accept them here,
2720 so that they will be handled by emit_symbolic_move. */
2724 /* All remaining non-PIC symbolic constants are
2725 forced into the literal pool. */
2729 /* Determine if it's legal to put X into the constant pool. This
2730 is not possible if X contains the address of a symbol that is
2731 not constant (TLS) or not known at final link time (PIC). */
2734 s390_cannot_force_const_mem (enum machine_mode mode
, rtx x
)
2736 switch (GET_CODE (x
))
2740 /* Accept all non-symbolic constants. */
2744 /* Labels are OK iff we are non-PIC. */
2745 return flag_pic
!= 0;
2748 /* 'Naked' TLS symbol references are never OK,
2749 non-TLS symbols are OK iff we are non-PIC. */
2750 if (tls_symbolic_operand (x
))
2753 return flag_pic
!= 0;
2756 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0));
2759 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0))
2760 || s390_cannot_force_const_mem (mode
, XEXP (x
, 1));
2763 switch (XINT (x
, 1))
2765 /* Only lt-relative or GOT-relative UNSPECs are OK. */
2766 case UNSPEC_LTREL_OFFSET
:
2774 case UNSPEC_GOTNTPOFF
:
2775 case UNSPEC_INDNTPOFF
:
2778 /* If the literal pool shares the code section, be put
2779 execute template placeholders into the pool as well. */
2781 return TARGET_CPU_ZARCH
;
2793 /* Returns true if the constant value OP is a legitimate general
2794 operand during and after reload. The difference to
2795 legitimate_constant_p is that this function will not accept
2796 a constant that would need to be forced to the literal pool
2797 before it can be used as operand.
2798 This function accepts all constants which can be loaded directly
2802 legitimate_reload_constant_p (rtx op
)
2804 /* Accept la(y) operands. */
2805 if (GET_CODE (op
) == CONST_INT
2806 && DISP_IN_RANGE (INTVAL (op
)))
2809 /* Accept l(g)hi/l(g)fi operands. */
2810 if (GET_CODE (op
) == CONST_INT
2811 && (CONST_OK_FOR_K (INTVAL (op
)) || CONST_OK_FOR_Os (INTVAL (op
))))
2814 /* Accept lliXX operands. */
2816 && GET_CODE (op
) == CONST_INT
2817 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2818 && s390_single_part (op
, word_mode
, HImode
, 0) >= 0)
2822 && GET_CODE (op
) == CONST_INT
2823 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2824 && s390_single_part (op
, word_mode
, SImode
, 0) >= 0)
2827 /* Accept larl operands. */
2828 if (TARGET_CPU_ZARCH
2829 && larl_operand (op
, VOIDmode
))
2832 /* Accept floating-point zero operands that fit into a single GPR. */
2833 if (GET_CODE (op
) == CONST_DOUBLE
2834 && s390_float_const_zero_p (op
)
2835 && GET_MODE_SIZE (GET_MODE (op
)) <= UNITS_PER_WORD
)
2838 /* Accept double-word operands that can be split. */
2839 if (GET_CODE (op
) == CONST_INT
2840 && trunc_int_for_mode (INTVAL (op
), word_mode
) != INTVAL (op
))
2842 enum machine_mode dword_mode
= word_mode
== SImode
? DImode
: TImode
;
2843 rtx hi
= operand_subword (op
, 0, 0, dword_mode
);
2844 rtx lo
= operand_subword (op
, 1, 0, dword_mode
);
2845 return legitimate_reload_constant_p (hi
)
2846 && legitimate_reload_constant_p (lo
);
2849 /* Everything else cannot be handled without reload. */
2853 /* Returns true if the constant value OP is a legitimate fp operand
2854 during and after reload.
2855 This function accepts all constants which can be loaded directly
2859 legitimate_reload_fp_constant_p (rtx op
)
2861 /* Accept floating-point zero operands if the load zero instruction
2864 && GET_CODE (op
) == CONST_DOUBLE
2865 && s390_float_const_zero_p (op
))
2871 /* Given an rtx OP being reloaded into a reg required to be in class RCLASS,
2872 return the class of reg to actually use. */
2875 s390_preferred_reload_class (rtx op
, reg_class_t rclass
)
2877 switch (GET_CODE (op
))
2879 /* Constants we cannot reload into general registers
2880 must be forced into the literal pool. */
2883 if (reg_class_subset_p (GENERAL_REGS
, rclass
)
2884 && legitimate_reload_constant_p (op
))
2885 return GENERAL_REGS
;
2886 else if (reg_class_subset_p (ADDR_REGS
, rclass
)
2887 && legitimate_reload_constant_p (op
))
2889 else if (reg_class_subset_p (FP_REGS
, rclass
)
2890 && legitimate_reload_fp_constant_p (op
))
2894 /* If a symbolic constant or a PLUS is reloaded,
2895 it is most likely being used as an address, so
2896 prefer ADDR_REGS. If 'class' is not a superset
2897 of ADDR_REGS, e.g. FP_REGS, reject this reload. */
2901 if (!legitimate_reload_constant_p (op
))
2905 /* load address will be used. */
2906 if (reg_class_subset_p (ADDR_REGS
, rclass
))
2918 /* Return true if ADDR is SYMBOL_REF + addend with addend being a
2919 multiple of ALIGNMENT and the SYMBOL_REF being naturally
2923 s390_check_symref_alignment (rtx addr
, HOST_WIDE_INT alignment
)
2925 HOST_WIDE_INT addend
;
2928 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2931 return (!SYMBOL_REF_NOT_NATURALLY_ALIGNED_P (symref
)
2932 && !(addend
& (alignment
- 1)));
2935 /* ADDR is moved into REG using larl. If ADDR isn't a valid larl
2936 operand SCRATCH is used to reload the even part of the address and
2940 s390_reload_larl_operand (rtx reg
, rtx addr
, rtx scratch
)
2942 HOST_WIDE_INT addend
;
2945 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2949 /* Easy case. The addend is even so larl will do fine. */
2950 emit_move_insn (reg
, addr
);
2953 /* We can leave the scratch register untouched if the target
2954 register is a valid base register. */
2955 if (REGNO (reg
) < FIRST_PSEUDO_REGISTER
2956 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
)
2959 gcc_assert (REGNO (scratch
) < FIRST_PSEUDO_REGISTER
);
2960 gcc_assert (REGNO_REG_CLASS (REGNO (scratch
)) == ADDR_REGS
);
2963 emit_move_insn (scratch
,
2964 gen_rtx_CONST (Pmode
,
2965 gen_rtx_PLUS (Pmode
, symref
,
2966 GEN_INT (addend
- 1))));
2968 emit_move_insn (scratch
, symref
);
2970 /* Increment the address using la in order to avoid clobbering cc. */
2971 emit_move_insn (reg
, gen_rtx_PLUS (Pmode
, scratch
, const1_rtx
));
2975 /* Generate what is necessary to move between REG and MEM using
2976 SCRATCH. The direction is given by TOMEM. */
2979 s390_reload_symref_address (rtx reg
, rtx mem
, rtx scratch
, bool tomem
)
2981 /* Reload might have pulled a constant out of the literal pool.
2982 Force it back in. */
2983 if (CONST_INT_P (mem
) || GET_CODE (mem
) == CONST_DOUBLE
2984 || GET_CODE (mem
) == CONST
)
2985 mem
= force_const_mem (GET_MODE (reg
), mem
);
2987 gcc_assert (MEM_P (mem
));
2989 /* For a load from memory we can leave the scratch register
2990 untouched if the target register is a valid base register. */
2992 && REGNO (reg
) < FIRST_PSEUDO_REGISTER
2993 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
2994 && GET_MODE (reg
) == GET_MODE (scratch
))
2997 /* Load address into scratch register. Since we can't have a
2998 secondary reload for a secondary reload we have to cover the case
2999 where larl would need a secondary reload here as well. */
3000 s390_reload_larl_operand (scratch
, XEXP (mem
, 0), scratch
);
3002 /* Now we can use a standard load/store to do the move. */
3004 emit_move_insn (replace_equiv_address (mem
, scratch
), reg
);
3006 emit_move_insn (reg
, replace_equiv_address (mem
, scratch
));
3009 /* Inform reload about cases where moving X with a mode MODE to a register in
3010 RCLASS requires an extra scratch or immediate register. Return the class
3011 needed for the immediate register. */
3014 s390_secondary_reload (bool in_p
, rtx x
, reg_class_t rclass_i
,
3015 enum machine_mode mode
, secondary_reload_info
*sri
)
3017 enum reg_class rclass
= (enum reg_class
) rclass_i
;
3019 /* Intermediate register needed. */
3020 if (reg_classes_intersect_p (CC_REGS
, rclass
))
3021 return GENERAL_REGS
;
3025 HOST_WIDE_INT offset
;
3028 /* On z10 several optimizer steps may generate larl operands with
3031 && s390_symref_operand_p (x
, &symref
, &offset
)
3033 && !SYMBOL_REF_ALIGN1_P (symref
)
3034 && (offset
& 1) == 1)
3035 sri
->icode
= ((mode
== DImode
) ? CODE_FOR_reloaddi_larl_odd_addend_z10
3036 : CODE_FOR_reloadsi_larl_odd_addend_z10
);
3038 /* On z10 we need a scratch register when moving QI, TI or floating
3039 point mode values from or to a memory location with a SYMBOL_REF
3040 or if the symref addend of a SI or DI move is not aligned to the
3041 width of the access. */
3043 && s390_symref_operand_p (XEXP (x
, 0), NULL
, NULL
)
3044 && (mode
== QImode
|| mode
== TImode
|| FLOAT_MODE_P (mode
)
3045 || (!TARGET_ZARCH
&& mode
== DImode
)
3046 || ((mode
== HImode
|| mode
== SImode
|| mode
== DImode
)
3047 && (!s390_check_symref_alignment (XEXP (x
, 0),
3048 GET_MODE_SIZE (mode
))))))
3050 #define __SECONDARY_RELOAD_CASE(M,m) \
3053 sri->icode = in_p ? CODE_FOR_reload##m##di_toreg_z10 : \
3054 CODE_FOR_reload##m##di_tomem_z10; \
3056 sri->icode = in_p ? CODE_FOR_reload##m##si_toreg_z10 : \
3057 CODE_FOR_reload##m##si_tomem_z10; \
3060 switch (GET_MODE (x
))
3062 __SECONDARY_RELOAD_CASE (QI
, qi
);
3063 __SECONDARY_RELOAD_CASE (HI
, hi
);
3064 __SECONDARY_RELOAD_CASE (SI
, si
);
3065 __SECONDARY_RELOAD_CASE (DI
, di
);
3066 __SECONDARY_RELOAD_CASE (TI
, ti
);
3067 __SECONDARY_RELOAD_CASE (SF
, sf
);
3068 __SECONDARY_RELOAD_CASE (DF
, df
);
3069 __SECONDARY_RELOAD_CASE (TF
, tf
);
3070 __SECONDARY_RELOAD_CASE (SD
, sd
);
3071 __SECONDARY_RELOAD_CASE (DD
, dd
);
3072 __SECONDARY_RELOAD_CASE (TD
, td
);
3077 #undef __SECONDARY_RELOAD_CASE
3081 /* We need a scratch register when loading a PLUS expression which
3082 is not a legitimate operand of the LOAD ADDRESS instruction. */
3083 if (in_p
&& s390_plus_operand (x
, mode
))
3084 sri
->icode
= (TARGET_64BIT
?
3085 CODE_FOR_reloaddi_plus
: CODE_FOR_reloadsi_plus
);
3087 /* Performing a multiword move from or to memory we have to make sure the
3088 second chunk in memory is addressable without causing a displacement
3089 overflow. If that would be the case we calculate the address in
3090 a scratch register. */
3092 && GET_CODE (XEXP (x
, 0)) == PLUS
3093 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
3094 && !DISP_IN_RANGE (INTVAL (XEXP (XEXP (x
, 0), 1))
3095 + GET_MODE_SIZE (mode
) - 1))
3097 /* For GENERAL_REGS a displacement overflow is no problem if occurring
3098 in a s_operand address since we may fallback to lm/stm. So we only
3099 have to care about overflows in the b+i+d case. */
3100 if ((reg_classes_intersect_p (GENERAL_REGS
, rclass
)
3101 && s390_class_max_nregs (GENERAL_REGS
, mode
) > 1
3102 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == PLUS
)
3103 /* For FP_REGS no lm/stm is available so this check is triggered
3104 for displacement overflows in b+i+d and b+d like addresses. */
3105 || (reg_classes_intersect_p (FP_REGS
, rclass
)
3106 && s390_class_max_nregs (FP_REGS
, mode
) > 1))
3109 sri
->icode
= (TARGET_64BIT
?
3110 CODE_FOR_reloaddi_nonoffmem_in
:
3111 CODE_FOR_reloadsi_nonoffmem_in
);
3113 sri
->icode
= (TARGET_64BIT
?
3114 CODE_FOR_reloaddi_nonoffmem_out
:
3115 CODE_FOR_reloadsi_nonoffmem_out
);
3119 /* A scratch address register is needed when a symbolic constant is
3120 copied to r0 compiling with -fPIC. In other cases the target
3121 register might be used as temporary (see legitimize_pic_address). */
3122 if (in_p
&& SYMBOLIC_CONST (x
) && flag_pic
== 2 && rclass
!= ADDR_REGS
)
3123 sri
->icode
= (TARGET_64BIT
?
3124 CODE_FOR_reloaddi_PIC_addr
:
3125 CODE_FOR_reloadsi_PIC_addr
);
3127 /* Either scratch or no register needed. */
3131 /* Generate code to load SRC, which is PLUS that is not a
3132 legitimate operand for the LA instruction, into TARGET.
3133 SCRATCH may be used as scratch register. */
3136 s390_expand_plus_operand (rtx target
, rtx src
,
3140 struct s390_address ad
;
3142 /* src must be a PLUS; get its two operands. */
3143 gcc_assert (GET_CODE (src
) == PLUS
);
3144 gcc_assert (GET_MODE (src
) == Pmode
);
3146 /* Check if any of the two operands is already scheduled
3147 for replacement by reload. This can happen e.g. when
3148 float registers occur in an address. */
3149 sum1
= find_replacement (&XEXP (src
, 0));
3150 sum2
= find_replacement (&XEXP (src
, 1));
3151 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3153 /* If the address is already strictly valid, there's nothing to do. */
3154 if (!s390_decompose_address (src
, &ad
)
3155 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3156 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
3158 /* Otherwise, one of the operands cannot be an address register;
3159 we reload its value into the scratch register. */
3160 if (true_regnum (sum1
) < 1 || true_regnum (sum1
) > 15)
3162 emit_move_insn (scratch
, sum1
);
3165 if (true_regnum (sum2
) < 1 || true_regnum (sum2
) > 15)
3167 emit_move_insn (scratch
, sum2
);
3171 /* According to the way these invalid addresses are generated
3172 in reload.c, it should never happen (at least on s390) that
3173 *neither* of the PLUS components, after find_replacements
3174 was applied, is an address register. */
3175 if (sum1
== scratch
&& sum2
== scratch
)
3181 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3184 /* Emit the LOAD ADDRESS pattern. Note that reload of PLUS
3185 is only ever performed on addresses, so we can mark the
3186 sum as legitimate for LA in any case. */
3187 s390_load_address (target
, src
);
3191 /* Return true if ADDR is a valid memory address.
3192 STRICT specifies whether strict register checking applies. */
3195 s390_legitimate_address_p (enum machine_mode mode
, rtx addr
, bool strict
)
3197 struct s390_address ad
;
3200 && larl_operand (addr
, VOIDmode
)
3201 && (mode
== VOIDmode
3202 || s390_check_symref_alignment (addr
, GET_MODE_SIZE (mode
))))
3205 if (!s390_decompose_address (addr
, &ad
))
3210 if (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3213 if (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
)))
3219 && !(REGNO (ad
.base
) >= FIRST_PSEUDO_REGISTER
3220 || REGNO_REG_CLASS (REGNO (ad
.base
)) == ADDR_REGS
))
3224 && !(REGNO (ad
.indx
) >= FIRST_PSEUDO_REGISTER
3225 || REGNO_REG_CLASS (REGNO (ad
.indx
)) == ADDR_REGS
))
3231 /* Return true if OP is a valid operand for the LA instruction.
3232 In 31-bit, we need to prove that the result is used as an
3233 address, as LA performs only a 31-bit addition. */
3236 legitimate_la_operand_p (rtx op
)
3238 struct s390_address addr
;
3239 if (!s390_decompose_address (op
, &addr
))
3242 return (TARGET_64BIT
|| addr
.pointer
);
3245 /* Return true if it is valid *and* preferable to use LA to
3246 compute the sum of OP1 and OP2. */
3249 preferred_la_operand_p (rtx op1
, rtx op2
)
3251 struct s390_address addr
;
3253 if (op2
!= const0_rtx
)
3254 op1
= gen_rtx_PLUS (Pmode
, op1
, op2
);
3256 if (!s390_decompose_address (op1
, &addr
))
3258 if (addr
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (addr
.base
)))
3260 if (addr
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (addr
.indx
)))
3263 /* Avoid LA instructions with index register on z196; it is
3264 preferable to use regular add instructions when possible. */
3265 if (addr
.indx
&& s390_tune
== PROCESSOR_2817_Z196
)
3268 if (!TARGET_64BIT
&& !addr
.pointer
)
3274 if ((addr
.base
&& REG_P (addr
.base
) && REG_POINTER (addr
.base
))
3275 || (addr
.indx
&& REG_P (addr
.indx
) && REG_POINTER (addr
.indx
)))
3281 /* Emit a forced load-address operation to load SRC into DST.
3282 This will use the LOAD ADDRESS instruction even in situations
3283 where legitimate_la_operand_p (SRC) returns false. */
3286 s390_load_address (rtx dst
, rtx src
)
3289 emit_move_insn (dst
, src
);
3291 emit_insn (gen_force_la_31 (dst
, src
));
3294 /* Return a legitimate reference for ORIG (an address) using the
3295 register REG. If REG is 0, a new pseudo is generated.
3297 There are two types of references that must be handled:
3299 1. Global data references must load the address from the GOT, via
3300 the PIC reg. An insn is emitted to do this load, and the reg is
3303 2. Static data references, constant pool addresses, and code labels
3304 compute the address as an offset from the GOT, whose base is in
3305 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
3306 differentiate them from global data objects. The returned
3307 address is the PIC reg + an unspec constant.
3309 TARGET_LEGITIMIZE_ADDRESS_P rejects symbolic references unless the PIC
3310 reg also appears in the address. */
3313 legitimize_pic_address (rtx orig
, rtx reg
)
3319 gcc_assert (!TLS_SYMBOLIC_CONST (addr
));
3321 if (GET_CODE (addr
) == LABEL_REF
3322 || (GET_CODE (addr
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (addr
)))
3324 /* This is a local symbol. */
3325 if (TARGET_CPU_ZARCH
&& larl_operand (addr
, VOIDmode
))
3327 /* Access local symbols PC-relative via LARL.
3328 This is the same as in the non-PIC case, so it is
3329 handled automatically ... */
3333 /* Access local symbols relative to the GOT. */
3335 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3337 if (reload_in_progress
|| reload_completed
)
3338 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3340 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
3341 addr
= gen_rtx_CONST (Pmode
, addr
);
3342 addr
= force_const_mem (Pmode
, addr
);
3343 emit_move_insn (temp
, addr
);
3345 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3348 s390_load_address (reg
, new_rtx
);
3353 else if (GET_CODE (addr
) == SYMBOL_REF
)
3356 reg
= gen_reg_rtx (Pmode
);
3360 /* Assume GOT offset < 4k. This is handled the same way
3361 in both 31- and 64-bit code (@GOT). */
3363 if (reload_in_progress
|| reload_completed
)
3364 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3366 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3367 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3368 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3369 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3370 emit_move_insn (reg
, new_rtx
);
3373 else if (TARGET_CPU_ZARCH
)
3375 /* If the GOT offset might be >= 4k, we determine the position
3376 of the GOT entry via a PC-relative LARL (@GOTENT). */
3378 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3380 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3381 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3383 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTENT
);
3384 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3385 emit_move_insn (temp
, new_rtx
);
3387 new_rtx
= gen_const_mem (Pmode
, temp
);
3388 emit_move_insn (reg
, new_rtx
);
3393 /* If the GOT offset might be >= 4k, we have to load it
3394 from the literal pool (@GOT). */
3396 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3398 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3399 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3401 if (reload_in_progress
|| reload_completed
)
3402 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3404 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3405 addr
= gen_rtx_CONST (Pmode
, addr
);
3406 addr
= force_const_mem (Pmode
, addr
);
3407 emit_move_insn (temp
, addr
);
3409 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3410 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3411 emit_move_insn (reg
, new_rtx
);
3417 if (GET_CODE (addr
) == CONST
)
3419 addr
= XEXP (addr
, 0);
3420 if (GET_CODE (addr
) == UNSPEC
)
3422 gcc_assert (XVECLEN (addr
, 0) == 1);
3423 switch (XINT (addr
, 1))
3425 /* If someone moved a GOT-relative UNSPEC
3426 out of the literal pool, force them back in. */
3429 new_rtx
= force_const_mem (Pmode
, orig
);
3432 /* @GOT is OK as is if small. */
3435 new_rtx
= force_const_mem (Pmode
, orig
);
3438 /* @GOTENT is OK as is. */
3442 /* @PLT is OK as is on 64-bit, must be converted to
3443 GOT-relative @PLTOFF on 31-bit. */
3445 if (!TARGET_CPU_ZARCH
)
3447 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3449 if (reload_in_progress
|| reload_completed
)
3450 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3452 addr
= XVECEXP (addr
, 0, 0);
3453 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
),
3455 addr
= gen_rtx_CONST (Pmode
, addr
);
3456 addr
= force_const_mem (Pmode
, addr
);
3457 emit_move_insn (temp
, addr
);
3459 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3462 s390_load_address (reg
, new_rtx
);
3468 /* Everything else cannot happen. */
3474 gcc_assert (GET_CODE (addr
) == PLUS
);
3476 if (GET_CODE (addr
) == PLUS
)
3478 rtx op0
= XEXP (addr
, 0), op1
= XEXP (addr
, 1);
3480 gcc_assert (!TLS_SYMBOLIC_CONST (op0
));
3481 gcc_assert (!TLS_SYMBOLIC_CONST (op1
));
3483 /* Check first to see if this is a constant offset
3484 from a local symbol reference. */
3485 if ((GET_CODE (op0
) == LABEL_REF
3486 || (GET_CODE (op0
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (op0
)))
3487 && GET_CODE (op1
) == CONST_INT
)
3489 if (TARGET_CPU_ZARCH
3490 && larl_operand (op0
, VOIDmode
)
3491 && INTVAL (op1
) < (HOST_WIDE_INT
)1 << 31
3492 && INTVAL (op1
) >= -((HOST_WIDE_INT
)1 << 31))
3494 if (INTVAL (op1
) & 1)
3496 /* LARL can't handle odd offsets, so emit a
3497 pair of LARL and LA. */
3498 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3500 if (!DISP_IN_RANGE (INTVAL (op1
)))
3502 HOST_WIDE_INT even
= INTVAL (op1
) - 1;
3503 op0
= gen_rtx_PLUS (Pmode
, op0
, GEN_INT (even
));
3504 op0
= gen_rtx_CONST (Pmode
, op0
);
3508 emit_move_insn (temp
, op0
);
3509 new_rtx
= gen_rtx_PLUS (Pmode
, temp
, op1
);
3513 s390_load_address (reg
, new_rtx
);
3519 /* If the offset is even, we can just use LARL.
3520 This will happen automatically. */
3525 /* Access local symbols relative to the GOT. */
3527 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3529 if (reload_in_progress
|| reload_completed
)
3530 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3532 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op0
),
3534 addr
= gen_rtx_PLUS (Pmode
, addr
, op1
);
3535 addr
= gen_rtx_CONST (Pmode
, addr
);
3536 addr
= force_const_mem (Pmode
, addr
);
3537 emit_move_insn (temp
, addr
);
3539 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3542 s390_load_address (reg
, new_rtx
);
3548 /* Now, check whether it is a GOT relative symbol plus offset
3549 that was pulled out of the literal pool. Force it back in. */
3551 else if (GET_CODE (op0
) == UNSPEC
3552 && GET_CODE (op1
) == CONST_INT
3553 && XINT (op0
, 1) == UNSPEC_GOTOFF
)
3555 gcc_assert (XVECLEN (op0
, 0) == 1);
3557 new_rtx
= force_const_mem (Pmode
, orig
);
3560 /* Otherwise, compute the sum. */
3563 base
= legitimize_pic_address (XEXP (addr
, 0), reg
);
3564 new_rtx
= legitimize_pic_address (XEXP (addr
, 1),
3565 base
== reg
? NULL_RTX
: reg
);
3566 if (GET_CODE (new_rtx
) == CONST_INT
)
3567 new_rtx
= plus_constant (base
, INTVAL (new_rtx
));
3570 if (GET_CODE (new_rtx
) == PLUS
&& CONSTANT_P (XEXP (new_rtx
, 1)))
3572 base
= gen_rtx_PLUS (Pmode
, base
, XEXP (new_rtx
, 0));
3573 new_rtx
= XEXP (new_rtx
, 1);
3575 new_rtx
= gen_rtx_PLUS (Pmode
, base
, new_rtx
);
3578 if (GET_CODE (new_rtx
) == CONST
)
3579 new_rtx
= XEXP (new_rtx
, 0);
3580 new_rtx
= force_operand (new_rtx
, 0);
3587 /* Load the thread pointer into a register. */
3590 s390_get_thread_pointer (void)
3592 rtx tp
= gen_reg_rtx (Pmode
);
3594 emit_move_insn (tp
, gen_rtx_REG (Pmode
, TP_REGNUM
));
3595 mark_reg_pointer (tp
, BITS_PER_WORD
);
3600 /* Emit a tls call insn. The call target is the SYMBOL_REF stored
3601 in s390_tls_symbol which always refers to __tls_get_offset.
3602 The returned offset is written to RESULT_REG and an USE rtx is
3603 generated for TLS_CALL. */
3605 static GTY(()) rtx s390_tls_symbol
;
3608 s390_emit_tls_call_insn (rtx result_reg
, rtx tls_call
)
3613 emit_insn (s390_load_got ());
3615 if (!s390_tls_symbol
)
3616 s390_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tls_get_offset");
3618 insn
= s390_emit_call (s390_tls_symbol
, tls_call
, result_reg
,
3619 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
3621 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), result_reg
);
3622 RTL_CONST_CALL_P (insn
) = 1;
3625 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3626 this (thread-local) address. REG may be used as temporary. */
3629 legitimize_tls_address (rtx addr
, rtx reg
)
3631 rtx new_rtx
, tls_call
, temp
, base
, r2
, insn
;
3633 if (GET_CODE (addr
) == SYMBOL_REF
)
3634 switch (tls_symbolic_operand (addr
))
3636 case TLS_MODEL_GLOBAL_DYNAMIC
:
3638 r2
= gen_rtx_REG (Pmode
, 2);
3639 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_TLSGD
);
3640 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3641 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3642 emit_move_insn (r2
, new_rtx
);
3643 s390_emit_tls_call_insn (r2
, tls_call
);
3644 insn
= get_insns ();
3647 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3648 temp
= gen_reg_rtx (Pmode
);
3649 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3651 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3654 s390_load_address (reg
, new_rtx
);
3659 case TLS_MODEL_LOCAL_DYNAMIC
:
3661 r2
= gen_rtx_REG (Pmode
, 2);
3662 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM
);
3663 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3664 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3665 emit_move_insn (r2
, new_rtx
);
3666 s390_emit_tls_call_insn (r2
, tls_call
);
3667 insn
= get_insns ();
3670 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM_NTPOFF
);
3671 temp
= gen_reg_rtx (Pmode
);
3672 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3674 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3675 base
= gen_reg_rtx (Pmode
);
3676 s390_load_address (base
, new_rtx
);
3678 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_DTPOFF
);
3679 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3680 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3681 temp
= gen_reg_rtx (Pmode
);
3682 emit_move_insn (temp
, new_rtx
);
3684 new_rtx
= gen_rtx_PLUS (Pmode
, base
, temp
);
3687 s390_load_address (reg
, new_rtx
);
3692 case TLS_MODEL_INITIAL_EXEC
:
3695 /* Assume GOT offset < 4k. This is handled the same way
3696 in both 31- and 64-bit code. */
3698 if (reload_in_progress
|| reload_completed
)
3699 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3701 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3702 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3703 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3704 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3705 temp
= gen_reg_rtx (Pmode
);
3706 emit_move_insn (temp
, new_rtx
);
3708 else if (TARGET_CPU_ZARCH
)
3710 /* If the GOT offset might be >= 4k, we determine the position
3711 of the GOT entry via a PC-relative LARL. */
3713 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3714 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3715 temp
= gen_reg_rtx (Pmode
);
3716 emit_move_insn (temp
, new_rtx
);
3718 new_rtx
= gen_const_mem (Pmode
, temp
);
3719 temp
= gen_reg_rtx (Pmode
);
3720 emit_move_insn (temp
, new_rtx
);
3724 /* If the GOT offset might be >= 4k, we have to load it
3725 from the literal pool. */
3727 if (reload_in_progress
|| reload_completed
)
3728 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3730 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3731 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3732 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3733 temp
= gen_reg_rtx (Pmode
);
3734 emit_move_insn (temp
, new_rtx
);
3736 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3737 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3739 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3740 temp
= gen_reg_rtx (Pmode
);
3741 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3745 /* In position-dependent code, load the absolute address of
3746 the GOT entry from the literal pool. */
3748 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3749 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3750 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3751 temp
= gen_reg_rtx (Pmode
);
3752 emit_move_insn (temp
, new_rtx
);
3755 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3756 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3757 temp
= gen_reg_rtx (Pmode
);
3758 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3761 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3764 s390_load_address (reg
, new_rtx
);
3769 case TLS_MODEL_LOCAL_EXEC
:
3770 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3771 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3772 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3773 temp
= gen_reg_rtx (Pmode
);
3774 emit_move_insn (temp
, new_rtx
);
3776 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3779 s390_load_address (reg
, new_rtx
);
3788 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == UNSPEC
)
3790 switch (XINT (XEXP (addr
, 0), 1))
3792 case UNSPEC_INDNTPOFF
:
3793 gcc_assert (TARGET_CPU_ZARCH
);
3802 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
3803 && GET_CODE (XEXP (XEXP (addr
, 0), 1)) == CONST_INT
)
3805 new_rtx
= XEXP (XEXP (addr
, 0), 0);
3806 if (GET_CODE (new_rtx
) != SYMBOL_REF
)
3807 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3809 new_rtx
= legitimize_tls_address (new_rtx
, reg
);
3810 new_rtx
= plus_constant (new_rtx
, INTVAL (XEXP (XEXP (addr
, 0), 1)));
3811 new_rtx
= force_operand (new_rtx
, 0);
3815 gcc_unreachable (); /* for now ... */
3820 /* Emit insns making the address in operands[1] valid for a standard
3821 move to operands[0]. operands[1] is replaced by an address which
3822 should be used instead of the former RTX to emit the move
3826 emit_symbolic_move (rtx
*operands
)
3828 rtx temp
= !can_create_pseudo_p () ? operands
[0] : gen_reg_rtx (Pmode
);
3830 if (GET_CODE (operands
[0]) == MEM
)
3831 operands
[1] = force_reg (Pmode
, operands
[1]);
3832 else if (TLS_SYMBOLIC_CONST (operands
[1]))
3833 operands
[1] = legitimize_tls_address (operands
[1], temp
);
3835 operands
[1] = legitimize_pic_address (operands
[1], temp
);
3838 /* Try machine-dependent ways of modifying an illegitimate address X
3839 to be legitimate. If we find one, return the new, valid address.
3841 OLDX is the address as it was before break_out_memory_refs was called.
3842 In some cases it is useful to look at this to decide what needs to be done.
3844 MODE is the mode of the operand pointed to by X.
3846 When -fpic is used, special handling is needed for symbolic references.
3847 See comments by legitimize_pic_address for details. */
3850 s390_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
3851 enum machine_mode mode ATTRIBUTE_UNUSED
)
3853 rtx constant_term
= const0_rtx
;
3855 if (TLS_SYMBOLIC_CONST (x
))
3857 x
= legitimize_tls_address (x
, 0);
3859 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3862 else if (GET_CODE (x
) == PLUS
3863 && (TLS_SYMBOLIC_CONST (XEXP (x
, 0))
3864 || TLS_SYMBOLIC_CONST (XEXP (x
, 1))))
3870 if (SYMBOLIC_CONST (x
)
3871 || (GET_CODE (x
) == PLUS
3872 && (SYMBOLIC_CONST (XEXP (x
, 0))
3873 || SYMBOLIC_CONST (XEXP (x
, 1)))))
3874 x
= legitimize_pic_address (x
, 0);
3876 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3880 x
= eliminate_constant_term (x
, &constant_term
);
3882 /* Optimize loading of large displacements by splitting them
3883 into the multiple of 4K and the rest; this allows the
3884 former to be CSE'd if possible.
3886 Don't do this if the displacement is added to a register
3887 pointing into the stack frame, as the offsets will
3888 change later anyway. */
3890 if (GET_CODE (constant_term
) == CONST_INT
3891 && !TARGET_LONG_DISPLACEMENT
3892 && !DISP_IN_RANGE (INTVAL (constant_term
))
3893 && !(REG_P (x
) && REGNO_PTR_FRAME_P (REGNO (x
))))
3895 HOST_WIDE_INT lower
= INTVAL (constant_term
) & 0xfff;
3896 HOST_WIDE_INT upper
= INTVAL (constant_term
) ^ lower
;
3898 rtx temp
= gen_reg_rtx (Pmode
);
3899 rtx val
= force_operand (GEN_INT (upper
), temp
);
3901 emit_move_insn (temp
, val
);
3903 x
= gen_rtx_PLUS (Pmode
, x
, temp
);
3904 constant_term
= GEN_INT (lower
);
3907 if (GET_CODE (x
) == PLUS
)
3909 if (GET_CODE (XEXP (x
, 0)) == REG
)
3911 rtx temp
= gen_reg_rtx (Pmode
);
3912 rtx val
= force_operand (XEXP (x
, 1), temp
);
3914 emit_move_insn (temp
, val
);
3916 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0), temp
);
3919 else if (GET_CODE (XEXP (x
, 1)) == REG
)
3921 rtx temp
= gen_reg_rtx (Pmode
);
3922 rtx val
= force_operand (XEXP (x
, 0), temp
);
3924 emit_move_insn (temp
, val
);
3926 x
= gen_rtx_PLUS (Pmode
, temp
, XEXP (x
, 1));
3930 if (constant_term
!= const0_rtx
)
3931 x
= gen_rtx_PLUS (Pmode
, x
, constant_term
);
3936 /* Try a machine-dependent way of reloading an illegitimate address AD
3937 operand. If we find one, push the reload and return the new address.
3939 MODE is the mode of the enclosing MEM. OPNUM is the operand number
3940 and TYPE is the reload type of the current reload. */
3943 legitimize_reload_address (rtx ad
, enum machine_mode mode ATTRIBUTE_UNUSED
,
3944 int opnum
, int type
)
3946 if (!optimize
|| TARGET_LONG_DISPLACEMENT
)
3949 if (GET_CODE (ad
) == PLUS
)
3951 rtx tem
= simplify_binary_operation (PLUS
, Pmode
,
3952 XEXP (ad
, 0), XEXP (ad
, 1));
3957 if (GET_CODE (ad
) == PLUS
3958 && GET_CODE (XEXP (ad
, 0)) == REG
3959 && GET_CODE (XEXP (ad
, 1)) == CONST_INT
3960 && !DISP_IN_RANGE (INTVAL (XEXP (ad
, 1))))
3962 HOST_WIDE_INT lower
= INTVAL (XEXP (ad
, 1)) & 0xfff;
3963 HOST_WIDE_INT upper
= INTVAL (XEXP (ad
, 1)) ^ lower
;
3964 rtx cst
, tem
, new_rtx
;
3966 cst
= GEN_INT (upper
);
3967 if (!legitimate_reload_constant_p (cst
))
3968 cst
= force_const_mem (Pmode
, cst
);
3970 tem
= gen_rtx_PLUS (Pmode
, XEXP (ad
, 0), cst
);
3971 new_rtx
= gen_rtx_PLUS (Pmode
, tem
, GEN_INT (lower
));
3973 push_reload (XEXP (tem
, 1), 0, &XEXP (tem
, 1), 0,
3974 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0,
3975 opnum
, (enum reload_type
) type
);
3982 /* Emit code to move LEN bytes from DST to SRC. */
3985 s390_expand_movmem (rtx dst
, rtx src
, rtx len
)
3987 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
3989 if (INTVAL (len
) > 0)
3990 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (INTVAL (len
) - 1)));
3993 else if (TARGET_MVCLE
)
3995 emit_insn (gen_movmem_long (dst
, src
, convert_to_mode (Pmode
, len
, 1)));
4000 rtx dst_addr
, src_addr
, count
, blocks
, temp
;
4001 rtx loop_start_label
= gen_label_rtx ();
4002 rtx loop_end_label
= gen_label_rtx ();
4003 rtx end_label
= gen_label_rtx ();
4004 enum machine_mode mode
;
4006 mode
= GET_MODE (len
);
4007 if (mode
== VOIDmode
)
4010 dst_addr
= gen_reg_rtx (Pmode
);
4011 src_addr
= gen_reg_rtx (Pmode
);
4012 count
= gen_reg_rtx (mode
);
4013 blocks
= gen_reg_rtx (mode
);
4015 convert_move (count
, len
, 1);
4016 emit_cmp_and_jump_insns (count
, const0_rtx
,
4017 EQ
, NULL_RTX
, mode
, 1, end_label
);
4019 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4020 emit_move_insn (src_addr
, force_operand (XEXP (src
, 0), NULL_RTX
));
4021 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4022 src
= change_address (src
, VOIDmode
, src_addr
);
4024 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4027 emit_move_insn (count
, temp
);
4029 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4032 emit_move_insn (blocks
, temp
);
4034 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4035 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4037 emit_label (loop_start_label
);
4040 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 768))
4044 /* Issue a read prefetch for the +3 cache line. */
4045 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (768)),
4046 const0_rtx
, const0_rtx
);
4047 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4048 emit_insn (prefetch
);
4050 /* Issue a write prefetch for the +3 cache line. */
4051 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (768)),
4052 const1_rtx
, const0_rtx
);
4053 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4054 emit_insn (prefetch
);
4057 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (255)));
4058 s390_load_address (dst_addr
,
4059 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4060 s390_load_address (src_addr
,
4061 gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (256)));
4063 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4066 emit_move_insn (blocks
, temp
);
4068 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4069 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4071 emit_jump (loop_start_label
);
4072 emit_label (loop_end_label
);
4074 emit_insn (gen_movmem_short (dst
, src
,
4075 convert_to_mode (Pmode
, count
, 1)));
4076 emit_label (end_label
);
4080 /* Emit code to set LEN bytes at DST to VAL.
4081 Make use of clrmem if VAL is zero. */
4084 s390_expand_setmem (rtx dst
, rtx len
, rtx val
)
4086 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) == 0)
4089 gcc_assert (GET_CODE (val
) == CONST_INT
|| GET_MODE (val
) == QImode
);
4091 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) > 0 && INTVAL (len
) <= 257)
4093 if (val
== const0_rtx
&& INTVAL (len
) <= 256)
4094 emit_insn (gen_clrmem_short (dst
, GEN_INT (INTVAL (len
) - 1)));
4097 /* Initialize memory by storing the first byte. */
4098 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4100 if (INTVAL (len
) > 1)
4102 /* Initiate 1 byte overlap move.
4103 The first byte of DST is propagated through DSTP1.
4104 Prepare a movmem for: DST+1 = DST (length = LEN - 1).
4105 DST is set to size 1 so the rest of the memory location
4106 does not count as source operand. */
4107 rtx dstp1
= adjust_address (dst
, VOIDmode
, 1);
4108 set_mem_size (dst
, 1);
4110 emit_insn (gen_movmem_short (dstp1
, dst
,
4111 GEN_INT (INTVAL (len
) - 2)));
4116 else if (TARGET_MVCLE
)
4118 val
= force_not_mem (convert_modes (Pmode
, QImode
, val
, 1));
4119 emit_insn (gen_setmem_long (dst
, convert_to_mode (Pmode
, len
, 1), val
));
4124 rtx dst_addr
, count
, blocks
, temp
, dstp1
= NULL_RTX
;
4125 rtx loop_start_label
= gen_label_rtx ();
4126 rtx loop_end_label
= gen_label_rtx ();
4127 rtx end_label
= gen_label_rtx ();
4128 enum machine_mode mode
;
4130 mode
= GET_MODE (len
);
4131 if (mode
== VOIDmode
)
4134 dst_addr
= gen_reg_rtx (Pmode
);
4135 count
= gen_reg_rtx (mode
);
4136 blocks
= gen_reg_rtx (mode
);
4138 convert_move (count
, len
, 1);
4139 emit_cmp_and_jump_insns (count
, const0_rtx
,
4140 EQ
, NULL_RTX
, mode
, 1, end_label
);
4142 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4143 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4145 if (val
== const0_rtx
)
4146 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4150 dstp1
= adjust_address (dst
, VOIDmode
, 1);
4151 set_mem_size (dst
, 1);
4153 /* Initialize memory by storing the first byte. */
4154 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4156 /* If count is 1 we are done. */
4157 emit_cmp_and_jump_insns (count
, const1_rtx
,
4158 EQ
, NULL_RTX
, mode
, 1, end_label
);
4160 temp
= expand_binop (mode
, add_optab
, count
, GEN_INT (-2), count
, 1,
4164 emit_move_insn (count
, temp
);
4166 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4169 emit_move_insn (blocks
, temp
);
4171 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4172 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4174 emit_label (loop_start_label
);
4177 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 1024))
4179 /* Issue a write prefetch for the +4 cache line. */
4180 rtx prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
,
4182 const1_rtx
, const0_rtx
);
4183 emit_insn (prefetch
);
4184 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4187 if (val
== const0_rtx
)
4188 emit_insn (gen_clrmem_short (dst
, GEN_INT (255)));
4190 emit_insn (gen_movmem_short (dstp1
, dst
, GEN_INT (255)));
4191 s390_load_address (dst_addr
,
4192 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4194 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4197 emit_move_insn (blocks
, temp
);
4199 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4200 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4202 emit_jump (loop_start_label
);
4203 emit_label (loop_end_label
);
4205 if (val
== const0_rtx
)
4206 emit_insn (gen_clrmem_short (dst
, convert_to_mode (Pmode
, count
, 1)));
4208 emit_insn (gen_movmem_short (dstp1
, dst
, convert_to_mode (Pmode
, count
, 1)));
4209 emit_label (end_label
);
4213 /* Emit code to compare LEN bytes at OP0 with those at OP1,
4214 and return the result in TARGET. */
4217 s390_expand_cmpmem (rtx target
, rtx op0
, rtx op1
, rtx len
)
4219 rtx ccreg
= gen_rtx_REG (CCUmode
, CC_REGNUM
);
4222 /* As the result of CMPINT is inverted compared to what we need,
4223 we have to swap the operands. */
4224 tmp
= op0
; op0
= op1
; op1
= tmp
;
4226 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4228 if (INTVAL (len
) > 0)
4230 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (INTVAL (len
) - 1)));
4231 emit_insn (gen_cmpint (target
, ccreg
));
4234 emit_move_insn (target
, const0_rtx
);
4236 else if (TARGET_MVCLE
)
4238 emit_insn (gen_cmpmem_long (op0
, op1
, convert_to_mode (Pmode
, len
, 1)));
4239 emit_insn (gen_cmpint (target
, ccreg
));
4243 rtx addr0
, addr1
, count
, blocks
, temp
;
4244 rtx loop_start_label
= gen_label_rtx ();
4245 rtx loop_end_label
= gen_label_rtx ();
4246 rtx end_label
= gen_label_rtx ();
4247 enum machine_mode mode
;
4249 mode
= GET_MODE (len
);
4250 if (mode
== VOIDmode
)
4253 addr0
= gen_reg_rtx (Pmode
);
4254 addr1
= gen_reg_rtx (Pmode
);
4255 count
= gen_reg_rtx (mode
);
4256 blocks
= gen_reg_rtx (mode
);
4258 convert_move (count
, len
, 1);
4259 emit_cmp_and_jump_insns (count
, const0_rtx
,
4260 EQ
, NULL_RTX
, mode
, 1, end_label
);
4262 emit_move_insn (addr0
, force_operand (XEXP (op0
, 0), NULL_RTX
));
4263 emit_move_insn (addr1
, force_operand (XEXP (op1
, 0), NULL_RTX
));
4264 op0
= change_address (op0
, VOIDmode
, addr0
);
4265 op1
= change_address (op1
, VOIDmode
, addr1
);
4267 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4270 emit_move_insn (count
, temp
);
4272 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4275 emit_move_insn (blocks
, temp
);
4277 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4278 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4280 emit_label (loop_start_label
);
4283 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 512))
4287 /* Issue a read prefetch for the +2 cache line of operand 1. */
4288 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (512)),
4289 const0_rtx
, const0_rtx
);
4290 emit_insn (prefetch
);
4291 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4293 /* Issue a read prefetch for the +2 cache line of operand 2. */
4294 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (512)),
4295 const0_rtx
, const0_rtx
);
4296 emit_insn (prefetch
);
4297 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4300 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (255)));
4301 temp
= gen_rtx_NE (VOIDmode
, ccreg
, const0_rtx
);
4302 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
4303 gen_rtx_LABEL_REF (VOIDmode
, end_label
), pc_rtx
);
4304 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
4305 emit_jump_insn (temp
);
4307 s390_load_address (addr0
,
4308 gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (256)));
4309 s390_load_address (addr1
,
4310 gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (256)));
4312 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4315 emit_move_insn (blocks
, temp
);
4317 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4318 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4320 emit_jump (loop_start_label
);
4321 emit_label (loop_end_label
);
4323 emit_insn (gen_cmpmem_short (op0
, op1
,
4324 convert_to_mode (Pmode
, count
, 1)));
4325 emit_label (end_label
);
4327 emit_insn (gen_cmpint (target
, ccreg
));
4332 /* Expand conditional increment or decrement using alc/slb instructions.
4333 Should generate code setting DST to either SRC or SRC + INCREMENT,
4334 depending on the result of the comparison CMP_OP0 CMP_CODE CMP_OP1.
4335 Returns true if successful, false otherwise.
4337 That makes it possible to implement some if-constructs without jumps e.g.:
4338 (borrow = CC0 | CC1 and carry = CC2 | CC3)
4339 unsigned int a, b, c;
4340 if (a < b) c++; -> CCU b > a -> CC2; c += carry;
4341 if (a < b) c--; -> CCL3 a - b -> borrow; c -= borrow;
4342 if (a <= b) c++; -> CCL3 b - a -> borrow; c += carry;
4343 if (a <= b) c--; -> CCU a <= b -> borrow; c -= borrow;
4345 Checks for EQ and NE with a nonzero value need an additional xor e.g.:
4346 if (a == b) c++; -> CCL3 a ^= b; 0 - a -> borrow; c += carry;
4347 if (a == b) c--; -> CCU a ^= b; a <= 0 -> CC0 | CC1; c -= borrow;
4348 if (a != b) c++; -> CCU a ^= b; a > 0 -> CC2; c += carry;
4349 if (a != b) c--; -> CCL3 a ^= b; 0 - a -> borrow; c -= borrow; */
4352 s390_expand_addcc (enum rtx_code cmp_code
, rtx cmp_op0
, rtx cmp_op1
,
4353 rtx dst
, rtx src
, rtx increment
)
4355 enum machine_mode cmp_mode
;
4356 enum machine_mode cc_mode
;
4362 if ((GET_MODE (cmp_op0
) == SImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4363 && (GET_MODE (cmp_op1
) == SImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4365 else if ((GET_MODE (cmp_op0
) == DImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4366 && (GET_MODE (cmp_op1
) == DImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4371 /* Try ADD LOGICAL WITH CARRY. */
4372 if (increment
== const1_rtx
)
4374 /* Determine CC mode to use. */
4375 if (cmp_code
== EQ
|| cmp_code
== NE
)
4377 if (cmp_op1
!= const0_rtx
)
4379 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4380 NULL_RTX
, 0, OPTAB_WIDEN
);
4381 cmp_op1
= const0_rtx
;
4384 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4387 if (cmp_code
== LTU
|| cmp_code
== LEU
)
4392 cmp_code
= swap_condition (cmp_code
);
4409 /* Emit comparison instruction pattern. */
4410 if (!register_operand (cmp_op0
, cmp_mode
))
4411 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4413 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4414 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4415 /* We use insn_invalid_p here to add clobbers if required. */
4416 ret
= insn_invalid_p (emit_insn (insn
), false);
4419 /* Emit ALC instruction pattern. */
4420 op_res
= gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4421 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4424 if (src
!= const0_rtx
)
4426 if (!register_operand (src
, GET_MODE (dst
)))
4427 src
= force_reg (GET_MODE (dst
), src
);
4429 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, src
);
4430 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, const0_rtx
);
4433 p
= rtvec_alloc (2);
4435 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4437 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4438 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4443 /* Try SUBTRACT LOGICAL WITH BORROW. */
4444 if (increment
== constm1_rtx
)
4446 /* Determine CC mode to use. */
4447 if (cmp_code
== EQ
|| cmp_code
== NE
)
4449 if (cmp_op1
!= const0_rtx
)
4451 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4452 NULL_RTX
, 0, OPTAB_WIDEN
);
4453 cmp_op1
= const0_rtx
;
4456 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4459 if (cmp_code
== GTU
|| cmp_code
== GEU
)
4464 cmp_code
= swap_condition (cmp_code
);
4481 /* Emit comparison instruction pattern. */
4482 if (!register_operand (cmp_op0
, cmp_mode
))
4483 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4485 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4486 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4487 /* We use insn_invalid_p here to add clobbers if required. */
4488 ret
= insn_invalid_p (emit_insn (insn
), false);
4491 /* Emit SLB instruction pattern. */
4492 if (!register_operand (src
, GET_MODE (dst
)))
4493 src
= force_reg (GET_MODE (dst
), src
);
4495 op_res
= gen_rtx_MINUS (GET_MODE (dst
),
4496 gen_rtx_MINUS (GET_MODE (dst
), src
, const0_rtx
),
4497 gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4498 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4500 p
= rtvec_alloc (2);
4502 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4504 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4505 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4513 /* Expand code for the insv template. Return true if successful. */
4516 s390_expand_insv (rtx dest
, rtx op1
, rtx op2
, rtx src
)
4518 int bitsize
= INTVAL (op1
);
4519 int bitpos
= INTVAL (op2
);
4521 /* On z10 we can use the risbg instruction to implement insv. */
4523 && ((GET_MODE (dest
) == DImode
&& GET_MODE (src
) == DImode
)
4524 || (GET_MODE (dest
) == SImode
&& GET_MODE (src
) == SImode
)))
4529 op
= gen_rtx_SET (GET_MODE(src
),
4530 gen_rtx_ZERO_EXTRACT (GET_MODE (dest
), dest
, op1
, op2
),
4532 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4533 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4538 /* We need byte alignment. */
4539 if (bitsize
% BITS_PER_UNIT
)
4543 && memory_operand (dest
, VOIDmode
)
4544 && (register_operand (src
, word_mode
)
4545 || const_int_operand (src
, VOIDmode
)))
4547 /* Emit standard pattern if possible. */
4548 enum machine_mode mode
= smallest_mode_for_size (bitsize
, MODE_INT
);
4549 if (GET_MODE_BITSIZE (mode
) == bitsize
)
4550 emit_move_insn (adjust_address (dest
, mode
, 0), gen_lowpart (mode
, src
));
4552 /* (set (ze (mem)) (const_int)). */
4553 else if (const_int_operand (src
, VOIDmode
))
4555 int size
= bitsize
/ BITS_PER_UNIT
;
4556 rtx src_mem
= adjust_address (force_const_mem (word_mode
, src
), BLKmode
,
4557 GET_MODE_SIZE (word_mode
) - size
);
4559 dest
= adjust_address (dest
, BLKmode
, 0);
4560 set_mem_size (dest
, size
);
4561 s390_expand_movmem (dest
, src_mem
, GEN_INT (size
));
4564 /* (set (ze (mem)) (reg)). */
4565 else if (register_operand (src
, word_mode
))
4567 if (bitsize
<= GET_MODE_BITSIZE (SImode
))
4568 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, op1
,
4572 /* Emit st,stcmh sequence. */
4573 int stcmh_width
= bitsize
- GET_MODE_BITSIZE (SImode
);
4574 int size
= stcmh_width
/ BITS_PER_UNIT
;
4576 emit_move_insn (adjust_address (dest
, SImode
, size
),
4577 gen_lowpart (SImode
, src
));
4578 set_mem_size (dest
, size
);
4579 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, GEN_INT
4580 (stcmh_width
), const0_rtx
),
4581 gen_rtx_LSHIFTRT (word_mode
, src
, GEN_INT
4582 (GET_MODE_BITSIZE (SImode
))));
4591 /* (set (ze (reg)) (const_int)). */
4593 && register_operand (dest
, word_mode
)
4594 && (bitpos
% 16) == 0
4595 && (bitsize
% 16) == 0
4596 && const_int_operand (src
, VOIDmode
))
4598 HOST_WIDE_INT val
= INTVAL (src
);
4599 int regpos
= bitpos
+ bitsize
;
4601 while (regpos
> bitpos
)
4603 enum machine_mode putmode
;
4606 if (TARGET_EXTIMM
&& (regpos
% 32 == 0) && (regpos
>= bitpos
+ 32))
4611 putsize
= GET_MODE_BITSIZE (putmode
);
4613 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4616 gen_int_mode (val
, putmode
));
4619 gcc_assert (regpos
== bitpos
);
4626 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic which returns a
4627 register that holds VAL of mode MODE shifted by COUNT bits. */
4630 s390_expand_mask_and_shift (rtx val
, enum machine_mode mode
, rtx count
)
4632 val
= expand_simple_binop (SImode
, AND
, val
, GEN_INT (GET_MODE_MASK (mode
)),
4633 NULL_RTX
, 1, OPTAB_DIRECT
);
4634 return expand_simple_binop (SImode
, ASHIFT
, val
, count
,
4635 NULL_RTX
, 1, OPTAB_DIRECT
);
4638 /* Structure to hold the initial parameters for a compare_and_swap operation
4639 in HImode and QImode. */
4641 struct alignment_context
4643 rtx memsi
; /* SI aligned memory location. */
4644 rtx shift
; /* Bit offset with regard to lsb. */
4645 rtx modemask
; /* Mask of the HQImode shifted by SHIFT bits. */
4646 rtx modemaski
; /* ~modemask */
4647 bool aligned
; /* True if memory is aligned, false else. */
4650 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic to initialize
4651 structure AC for transparent simplifying, if the memory alignment is known
4652 to be at least 32bit. MEM is the memory location for the actual operation
4653 and MODE its mode. */
4656 init_alignment_context (struct alignment_context
*ac
, rtx mem
,
4657 enum machine_mode mode
)
4659 ac
->shift
= GEN_INT (GET_MODE_SIZE (SImode
) - GET_MODE_SIZE (mode
));
4660 ac
->aligned
= (MEM_ALIGN (mem
) >= GET_MODE_BITSIZE (SImode
));
4663 ac
->memsi
= adjust_address (mem
, SImode
, 0); /* Memory is aligned. */
4666 /* Alignment is unknown. */
4667 rtx byteoffset
, addr
, align
;
4669 /* Force the address into a register. */
4670 addr
= force_reg (Pmode
, XEXP (mem
, 0));
4672 /* Align it to SImode. */
4673 align
= expand_simple_binop (Pmode
, AND
, addr
,
4674 GEN_INT (-GET_MODE_SIZE (SImode
)),
4675 NULL_RTX
, 1, OPTAB_DIRECT
);
4677 ac
->memsi
= gen_rtx_MEM (SImode
, align
);
4678 MEM_VOLATILE_P (ac
->memsi
) = MEM_VOLATILE_P (mem
);
4679 set_mem_alias_set (ac
->memsi
, ALIAS_SET_MEMORY_BARRIER
);
4680 set_mem_align (ac
->memsi
, GET_MODE_BITSIZE (SImode
));
4682 /* Calculate shiftcount. */
4683 byteoffset
= expand_simple_binop (Pmode
, AND
, addr
,
4684 GEN_INT (GET_MODE_SIZE (SImode
) - 1),
4685 NULL_RTX
, 1, OPTAB_DIRECT
);
4686 /* As we already have some offset, evaluate the remaining distance. */
4687 ac
->shift
= expand_simple_binop (SImode
, MINUS
, ac
->shift
, byteoffset
,
4688 NULL_RTX
, 1, OPTAB_DIRECT
);
4691 /* Shift is the byte count, but we need the bitcount. */
4692 ac
->shift
= expand_simple_binop (SImode
, MULT
, ac
->shift
, GEN_INT (BITS_PER_UNIT
),
4693 NULL_RTX
, 1, OPTAB_DIRECT
);
4694 /* Calculate masks. */
4695 ac
->modemask
= expand_simple_binop (SImode
, ASHIFT
,
4696 GEN_INT (GET_MODE_MASK (mode
)), ac
->shift
,
4697 NULL_RTX
, 1, OPTAB_DIRECT
);
4698 ac
->modemaski
= expand_simple_unop (SImode
, NOT
, ac
->modemask
, NULL_RTX
, 1);
4701 /* Expand an atomic compare and swap operation for HImode and QImode. MEM is
4702 the memory location, CMP the old value to compare MEM with and NEW_RTX the value
4703 to set if CMP == MEM.
4704 CMP is never in memory for compare_and_swap_cc because
4705 expand_bool_compare_and_swap puts it into a register for later compare. */
4708 s390_expand_cs_hqi (enum machine_mode mode
, rtx target
, rtx mem
, rtx cmp
, rtx new_rtx
)
4710 struct alignment_context ac
;
4711 rtx cmpv
, newv
, val
, resv
, cc
;
4712 rtx res
= gen_reg_rtx (SImode
);
4713 rtx csloop
= gen_label_rtx ();
4714 rtx csend
= gen_label_rtx ();
4716 gcc_assert (register_operand (target
, VOIDmode
));
4717 gcc_assert (MEM_P (mem
));
4719 init_alignment_context (&ac
, mem
, mode
);
4721 /* Shift the values to the correct bit positions. */
4722 if (!(ac
.aligned
&& MEM_P (cmp
)))
4723 cmp
= s390_expand_mask_and_shift (cmp
, mode
, ac
.shift
);
4724 if (!(ac
.aligned
&& MEM_P (new_rtx
)))
4725 new_rtx
= s390_expand_mask_and_shift (new_rtx
, mode
, ac
.shift
);
4727 /* Load full word. Subsequent loads are performed by CS. */
4728 val
= expand_simple_binop (SImode
, AND
, ac
.memsi
, ac
.modemaski
,
4729 NULL_RTX
, 1, OPTAB_DIRECT
);
4731 /* Start CS loop. */
4732 emit_label (csloop
);
4733 /* val = "<mem>00..0<mem>"
4734 * cmp = "00..0<cmp>00..0"
4735 * new = "00..0<new>00..0"
4738 /* Patch cmp and new with val at correct position. */
4739 if (ac
.aligned
&& MEM_P (cmp
))
4741 cmpv
= force_reg (SImode
, val
);
4742 store_bit_field (cmpv
, GET_MODE_BITSIZE (mode
), 0,
4746 cmpv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, cmp
, val
,
4747 NULL_RTX
, 1, OPTAB_DIRECT
));
4748 if (ac
.aligned
&& MEM_P (new_rtx
))
4750 newv
= force_reg (SImode
, val
);
4751 store_bit_field (newv
, GET_MODE_BITSIZE (mode
), 0,
4752 0, 0, SImode
, new_rtx
);
4755 newv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4756 NULL_RTX
, 1, OPTAB_DIRECT
));
4758 /* Jump to end if we're done (likely?). */
4759 s390_emit_jump (csend
, s390_emit_compare_and_swap (EQ
, res
, ac
.memsi
,
4762 /* Check for changes outside mode. */
4763 resv
= expand_simple_binop (SImode
, AND
, res
, ac
.modemaski
,
4764 NULL_RTX
, 1, OPTAB_DIRECT
);
4765 cc
= s390_emit_compare (NE
, resv
, val
);
4766 emit_move_insn (val
, resv
);
4767 /* Loop internal if so. */
4768 s390_emit_jump (csloop
, cc
);
4772 /* Return the correct part of the bitfield. */
4773 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
, res
, ac
.shift
,
4774 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4777 /* Expand an atomic operation CODE of mode MODE. MEM is the memory location
4778 and VAL the value to play with. If AFTER is true then store the value
4779 MEM holds after the operation, if AFTER is false then store the value MEM
4780 holds before the operation. If TARGET is zero then discard that value, else
4781 store it to TARGET. */
4784 s390_expand_atomic (enum machine_mode mode
, enum rtx_code code
,
4785 rtx target
, rtx mem
, rtx val
, bool after
)
4787 struct alignment_context ac
;
4789 rtx new_rtx
= gen_reg_rtx (SImode
);
4790 rtx orig
= gen_reg_rtx (SImode
);
4791 rtx csloop
= gen_label_rtx ();
4793 gcc_assert (!target
|| register_operand (target
, VOIDmode
));
4794 gcc_assert (MEM_P (mem
));
4796 init_alignment_context (&ac
, mem
, mode
);
4798 /* Shift val to the correct bit positions.
4799 Preserve "icm", but prevent "ex icm". */
4800 if (!(ac
.aligned
&& code
== SET
&& MEM_P (val
)))
4801 val
= s390_expand_mask_and_shift (val
, mode
, ac
.shift
);
4803 /* Further preparation insns. */
4804 if (code
== PLUS
|| code
== MINUS
)
4805 emit_move_insn (orig
, val
);
4806 else if (code
== MULT
|| code
== AND
) /* val = "11..1<val>11..1" */
4807 val
= expand_simple_binop (SImode
, XOR
, val
, ac
.modemaski
,
4808 NULL_RTX
, 1, OPTAB_DIRECT
);
4810 /* Load full word. Subsequent loads are performed by CS. */
4811 cmp
= force_reg (SImode
, ac
.memsi
);
4813 /* Start CS loop. */
4814 emit_label (csloop
);
4815 emit_move_insn (new_rtx
, cmp
);
4817 /* Patch new with val at correct position. */
4822 val
= expand_simple_binop (SImode
, code
, new_rtx
, orig
,
4823 NULL_RTX
, 1, OPTAB_DIRECT
);
4824 val
= expand_simple_binop (SImode
, AND
, val
, ac
.modemask
,
4825 NULL_RTX
, 1, OPTAB_DIRECT
);
4828 if (ac
.aligned
&& MEM_P (val
))
4829 store_bit_field (new_rtx
, GET_MODE_BITSIZE (mode
), 0,
4833 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, ac
.modemaski
,
4834 NULL_RTX
, 1, OPTAB_DIRECT
);
4835 new_rtx
= expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4836 NULL_RTX
, 1, OPTAB_DIRECT
);
4842 new_rtx
= expand_simple_binop (SImode
, code
, new_rtx
, val
,
4843 NULL_RTX
, 1, OPTAB_DIRECT
);
4845 case MULT
: /* NAND */
4846 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, val
,
4847 NULL_RTX
, 1, OPTAB_DIRECT
);
4848 new_rtx
= expand_simple_binop (SImode
, XOR
, new_rtx
, ac
.modemask
,
4849 NULL_RTX
, 1, OPTAB_DIRECT
);
4855 s390_emit_jump (csloop
, s390_emit_compare_and_swap (NE
, cmp
,
4856 ac
.memsi
, cmp
, new_rtx
));
4858 /* Return the correct part of the bitfield. */
4860 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
,
4861 after
? new_rtx
: cmp
, ac
.shift
,
4862 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4865 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
4866 We need to emit DTP-relative relocations. */
4868 static void s390_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
4871 s390_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
4876 fputs ("\t.long\t", file
);
4879 fputs ("\t.quad\t", file
);
4884 output_addr_const (file
, x
);
4885 fputs ("@DTPOFF", file
);
4888 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
4889 /* Implement TARGET_MANGLE_TYPE. */
4892 s390_mangle_type (const_tree type
)
4894 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
4895 && TARGET_LONG_DOUBLE_128
)
4898 /* For all other types, use normal C++ mangling. */
4903 /* In the name of slightly smaller debug output, and to cater to
4904 general assembler lossage, recognize various UNSPEC sequences
4905 and turn them back into a direct symbol reference. */
4908 s390_delegitimize_address (rtx orig_x
)
4912 orig_x
= delegitimize_mem_from_attrs (orig_x
);
4915 /* Extract the symbol ref from:
4916 (plus:SI (reg:SI 12 %r12)
4917 (const:SI (unspec:SI [(symbol_ref/f:SI ("*.LC0"))]
4918 UNSPEC_GOTOFF/PLTOFF)))
4920 (plus:SI (reg:SI 12 %r12)
4921 (const:SI (plus:SI (unspec:SI [(symbol_ref:SI ("L"))]
4922 UNSPEC_GOTOFF/PLTOFF)
4923 (const_int 4 [0x4])))) */
4924 if (GET_CODE (x
) == PLUS
4925 && REG_P (XEXP (x
, 0))
4926 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
4927 && GET_CODE (XEXP (x
, 1)) == CONST
)
4929 HOST_WIDE_INT offset
= 0;
4931 /* The const operand. */
4932 y
= XEXP (XEXP (x
, 1), 0);
4934 if (GET_CODE (y
) == PLUS
4935 && GET_CODE (XEXP (y
, 1)) == CONST_INT
)
4937 offset
= INTVAL (XEXP (y
, 1));
4941 if (GET_CODE (y
) == UNSPEC
4942 && (XINT (y
, 1) == UNSPEC_GOTOFF
4943 || XINT (y
, 1) == UNSPEC_PLTOFF
))
4944 return plus_constant (XVECEXP (y
, 0, 0), offset
);
4947 if (GET_CODE (x
) != MEM
)
4951 if (GET_CODE (x
) == PLUS
4952 && GET_CODE (XEXP (x
, 1)) == CONST
4953 && GET_CODE (XEXP (x
, 0)) == REG
4954 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
4956 y
= XEXP (XEXP (x
, 1), 0);
4957 if (GET_CODE (y
) == UNSPEC
4958 && XINT (y
, 1) == UNSPEC_GOT
)
4959 y
= XVECEXP (y
, 0, 0);
4963 else if (GET_CODE (x
) == CONST
)
4965 /* Extract the symbol ref from:
4966 (mem:QI (const:DI (unspec:DI [(symbol_ref:DI ("foo"))]
4967 UNSPEC_PLT/GOTENT))) */
4970 if (GET_CODE (y
) == UNSPEC
4971 && (XINT (y
, 1) == UNSPEC_GOTENT
4972 || XINT (y
, 1) == UNSPEC_PLT
))
4973 y
= XVECEXP (y
, 0, 0);
4980 if (GET_MODE (orig_x
) != Pmode
)
4982 if (GET_MODE (orig_x
) == BLKmode
)
4984 y
= lowpart_subreg (GET_MODE (orig_x
), y
, Pmode
);
4991 /* Output operand OP to stdio stream FILE.
4992 OP is an address (register + offset) which is not used to address data;
4993 instead the rightmost bits are interpreted as the value. */
4996 print_shift_count_operand (FILE *file
, rtx op
)
4998 HOST_WIDE_INT offset
;
5001 /* Extract base register and offset. */
5002 if (!s390_decompose_shift_count (op
, &base
, &offset
))
5008 gcc_assert (GET_CODE (base
) == REG
);
5009 gcc_assert (REGNO (base
) < FIRST_PSEUDO_REGISTER
);
5010 gcc_assert (REGNO_REG_CLASS (REGNO (base
)) == ADDR_REGS
);
5013 /* Offsets are constricted to twelve bits. */
5014 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, offset
& ((1 << 12) - 1));
5016 fprintf (file
, "(%s)", reg_names
[REGNO (base
)]);
5019 /* See 'get_some_local_dynamic_name'. */
5022 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
5026 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
5028 x
= get_pool_constant (x
);
5029 return for_each_rtx (&x
, get_some_local_dynamic_name_1
, 0);
5032 if (GET_CODE (x
) == SYMBOL_REF
5033 && tls_symbolic_operand (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
5035 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
5042 /* Locate some local-dynamic symbol still in use by this function
5043 so that we can print its name in local-dynamic base patterns. */
5046 get_some_local_dynamic_name (void)
5050 if (cfun
->machine
->some_ld_name
)
5051 return cfun
->machine
->some_ld_name
;
5053 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5055 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
5056 return cfun
->machine
->some_ld_name
;
5061 /* Output machine-dependent UNSPECs occurring in address constant X
5062 in assembler syntax to stdio stream FILE. Returns true if the
5063 constant X could be recognized, false otherwise. */
5066 s390_output_addr_const_extra (FILE *file
, rtx x
)
5068 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 1)
5069 switch (XINT (x
, 1))
5072 output_addr_const (file
, XVECEXP (x
, 0, 0));
5073 fprintf (file
, "@GOTENT");
5076 output_addr_const (file
, XVECEXP (x
, 0, 0));
5077 fprintf (file
, "@GOT");
5080 output_addr_const (file
, XVECEXP (x
, 0, 0));
5081 fprintf (file
, "@GOTOFF");
5084 output_addr_const (file
, XVECEXP (x
, 0, 0));
5085 fprintf (file
, "@PLT");
5088 output_addr_const (file
, XVECEXP (x
, 0, 0));
5089 fprintf (file
, "@PLTOFF");
5092 output_addr_const (file
, XVECEXP (x
, 0, 0));
5093 fprintf (file
, "@TLSGD");
5096 assemble_name (file
, get_some_local_dynamic_name ());
5097 fprintf (file
, "@TLSLDM");
5100 output_addr_const (file
, XVECEXP (x
, 0, 0));
5101 fprintf (file
, "@DTPOFF");
5104 output_addr_const (file
, XVECEXP (x
, 0, 0));
5105 fprintf (file
, "@NTPOFF");
5107 case UNSPEC_GOTNTPOFF
:
5108 output_addr_const (file
, XVECEXP (x
, 0, 0));
5109 fprintf (file
, "@GOTNTPOFF");
5111 case UNSPEC_INDNTPOFF
:
5112 output_addr_const (file
, XVECEXP (x
, 0, 0));
5113 fprintf (file
, "@INDNTPOFF");
5117 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 2)
5118 switch (XINT (x
, 1))
5120 case UNSPEC_POOL_OFFSET
:
5121 x
= gen_rtx_MINUS (GET_MODE (x
), XVECEXP (x
, 0, 0), XVECEXP (x
, 0, 1));
5122 output_addr_const (file
, x
);
5128 /* Output address operand ADDR in assembler syntax to
5129 stdio stream FILE. */
5132 print_operand_address (FILE *file
, rtx addr
)
5134 struct s390_address ad
;
5136 if (s390_symref_operand_p (addr
, NULL
, NULL
))
5140 output_operand_lossage ("symbolic memory references are "
5141 "only supported on z10 or later");
5144 output_addr_const (file
, addr
);
5148 if (!s390_decompose_address (addr
, &ad
)
5149 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5150 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
5151 output_operand_lossage ("cannot decompose address");
5154 output_addr_const (file
, ad
.disp
);
5156 fprintf (file
, "0");
5158 if (ad
.base
&& ad
.indx
)
5159 fprintf (file
, "(%s,%s)", reg_names
[REGNO (ad
.indx
)],
5160 reg_names
[REGNO (ad
.base
)]);
5162 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5165 /* Output operand X in assembler syntax to stdio stream FILE.
5166 CODE specified the format flag. The following format flags
5169 'C': print opcode suffix for branch condition.
5170 'D': print opcode suffix for inverse branch condition.
5171 'E': print opcode suffix for branch on index instruction.
5172 'J': print tls_load/tls_gdcall/tls_ldcall suffix
5173 'G': print the size of the operand in bytes.
5174 'O': print only the displacement of a memory reference.
5175 'R': print only the base register of a memory reference.
5176 'S': print S-type memory reference (base+displacement).
5177 'N': print the second word of a DImode operand.
5178 'M': print the second word of a TImode operand.
5179 'Y': print shift count operand.
5181 'b': print integer X as if it's an unsigned byte.
5182 'c': print integer X as if it's an signed byte.
5183 'x': print integer X as if it's an unsigned halfword.
5184 'h': print integer X as if it's a signed halfword.
5185 'i': print the first nonzero HImode part of X.
5186 'j': print the first HImode part unequal to -1 of X.
5187 'k': print the first nonzero SImode part of X.
5188 'm': print the first SImode part unequal to -1 of X.
5189 'o': print integer X as if it's an unsigned 32bit word. */
5192 print_operand (FILE *file
, rtx x
, int code
)
5197 fprintf (file
, s390_branch_condition_mnemonic (x
, FALSE
));
5201 fprintf (file
, s390_branch_condition_mnemonic (x
, TRUE
));
5205 if (GET_CODE (x
) == LE
)
5206 fprintf (file
, "l");
5207 else if (GET_CODE (x
) == GT
)
5208 fprintf (file
, "h");
5210 output_operand_lossage ("invalid comparison operator "
5211 "for 'E' output modifier");
5215 if (GET_CODE (x
) == SYMBOL_REF
)
5217 fprintf (file
, "%s", ":tls_load:");
5218 output_addr_const (file
, x
);
5220 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD
)
5222 fprintf (file
, "%s", ":tls_gdcall:");
5223 output_addr_const (file
, XVECEXP (x
, 0, 0));
5225 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM
)
5227 fprintf (file
, "%s", ":tls_ldcall:");
5228 assemble_name (file
, get_some_local_dynamic_name ());
5231 output_operand_lossage ("invalid reference for 'J' output modifier");
5235 fprintf (file
, "%u", GET_MODE_SIZE (GET_MODE (x
)));
5240 struct s390_address ad
;
5245 output_operand_lossage ("memory reference expected for "
5246 "'O' output modifier");
5250 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5253 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5256 output_operand_lossage ("invalid address for 'O' output modifier");
5261 output_addr_const (file
, ad
.disp
);
5263 fprintf (file
, "0");
5269 struct s390_address ad
;
5274 output_operand_lossage ("memory reference expected for "
5275 "'R' output modifier");
5279 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5282 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5285 output_operand_lossage ("invalid address for 'R' output modifier");
5290 fprintf (file
, "%s", reg_names
[REGNO (ad
.base
)]);
5292 fprintf (file
, "0");
5298 struct s390_address ad
;
5303 output_operand_lossage ("memory reference expected for "
5304 "'S' output modifier");
5307 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5310 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5313 output_operand_lossage ("invalid address for 'S' output modifier");
5318 output_addr_const (file
, ad
.disp
);
5320 fprintf (file
, "0");
5323 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5328 if (GET_CODE (x
) == REG
)
5329 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5330 else if (GET_CODE (x
) == MEM
)
5331 x
= change_address (x
, VOIDmode
, plus_constant (XEXP (x
, 0), 4));
5333 output_operand_lossage ("register or memory expression expected "
5334 "for 'N' output modifier");
5338 if (GET_CODE (x
) == REG
)
5339 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5340 else if (GET_CODE (x
) == MEM
)
5341 x
= change_address (x
, VOIDmode
, plus_constant (XEXP (x
, 0), 8));
5343 output_operand_lossage ("register or memory expression expected "
5344 "for 'M' output modifier");
5348 print_shift_count_operand (file
, x
);
5352 switch (GET_CODE (x
))
5355 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5359 output_address (XEXP (x
, 0));
5366 output_addr_const (file
, x
);
5371 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xff);
5372 else if (code
== 'c')
5373 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xff) ^ 0x80) - 0x80);
5374 else if (code
== 'x')
5375 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffff);
5376 else if (code
== 'h')
5377 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xffff) ^ 0x8000) - 0x8000);
5378 else if (code
== 'i')
5379 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5380 s390_extract_part (x
, HImode
, 0));
5381 else if (code
== 'j')
5382 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5383 s390_extract_part (x
, HImode
, -1));
5384 else if (code
== 'k')
5385 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5386 s390_extract_part (x
, SImode
, 0));
5387 else if (code
== 'm')
5388 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5389 s390_extract_part (x
, SImode
, -1));
5390 else if (code
== 'o')
5391 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffffffff);
5393 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5397 gcc_assert (GET_MODE (x
) == VOIDmode
);
5399 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xff);
5400 else if (code
== 'x')
5401 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xffff);
5402 else if (code
== 'h')
5403 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5404 ((CONST_DOUBLE_LOW (x
) & 0xffff) ^ 0x8000) - 0x8000);
5408 output_operand_lossage ("invalid constant - try using "
5409 "an output modifier");
5411 output_operand_lossage ("invalid constant for output modifier '%c'",
5418 output_operand_lossage ("invalid expression - try using "
5419 "an output modifier");
5421 output_operand_lossage ("invalid expression for output "
5422 "modifier '%c'", code
);
5427 /* Target hook for assembling integer objects. We need to define it
5428 here to work a round a bug in some versions of GAS, which couldn't
5429 handle values smaller than INT_MIN when printed in decimal. */
5432 s390_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
5434 if (size
== 8 && aligned_p
5435 && GET_CODE (x
) == CONST_INT
&& INTVAL (x
) < INT_MIN
)
5437 fprintf (asm_out_file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n",
5441 return default_assemble_integer (x
, size
, aligned_p
);
5444 /* Returns true if register REGNO is used for forming
5445 a memory address in expression X. */
5448 reg_used_in_mem_p (int regno
, rtx x
)
5450 enum rtx_code code
= GET_CODE (x
);
5456 if (refers_to_regno_p (regno
, regno
+1,
5460 else if (code
== SET
5461 && GET_CODE (SET_DEST (x
)) == PC
)
5463 if (refers_to_regno_p (regno
, regno
+1,
5468 fmt
= GET_RTX_FORMAT (code
);
5469 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
5472 && reg_used_in_mem_p (regno
, XEXP (x
, i
)))
5475 else if (fmt
[i
] == 'E')
5476 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5477 if (reg_used_in_mem_p (regno
, XVECEXP (x
, i
, j
)))
5483 /* Returns true if expression DEP_RTX sets an address register
5484 used by instruction INSN to address memory. */
5487 addr_generation_dependency_p (rtx dep_rtx
, rtx insn
)
5491 if (GET_CODE (dep_rtx
) == INSN
)
5492 dep_rtx
= PATTERN (dep_rtx
);
5494 if (GET_CODE (dep_rtx
) == SET
)
5496 target
= SET_DEST (dep_rtx
);
5497 if (GET_CODE (target
) == STRICT_LOW_PART
)
5498 target
= XEXP (target
, 0);
5499 while (GET_CODE (target
) == SUBREG
)
5500 target
= SUBREG_REG (target
);
5502 if (GET_CODE (target
) == REG
)
5504 int regno
= REGNO (target
);
5506 if (s390_safe_attr_type (insn
) == TYPE_LA
)
5508 pat
= PATTERN (insn
);
5509 if (GET_CODE (pat
) == PARALLEL
)
5511 gcc_assert (XVECLEN (pat
, 0) == 2);
5512 pat
= XVECEXP (pat
, 0, 0);
5514 gcc_assert (GET_CODE (pat
) == SET
);
5515 return refers_to_regno_p (regno
, regno
+1, SET_SRC (pat
), 0);
5517 else if (get_attr_atype (insn
) == ATYPE_AGEN
)
5518 return reg_used_in_mem_p (regno
, PATTERN (insn
));
5524 /* Return 1, if dep_insn sets register used in insn in the agen unit. */
5527 s390_agen_dep_p (rtx dep_insn
, rtx insn
)
5529 rtx dep_rtx
= PATTERN (dep_insn
);
5532 if (GET_CODE (dep_rtx
) == SET
5533 && addr_generation_dependency_p (dep_rtx
, insn
))
5535 else if (GET_CODE (dep_rtx
) == PARALLEL
)
5537 for (i
= 0; i
< XVECLEN (dep_rtx
, 0); i
++)
5539 if (addr_generation_dependency_p (XVECEXP (dep_rtx
, 0, i
), insn
))
5547 /* A C statement (sans semicolon) to update the integer scheduling priority
5548 INSN_PRIORITY (INSN). Increase the priority to execute the INSN earlier,
5549 reduce the priority to execute INSN later. Do not define this macro if
5550 you do not need to adjust the scheduling priorities of insns.
5552 A STD instruction should be scheduled earlier,
5553 in order to use the bypass. */
5555 s390_adjust_priority (rtx insn ATTRIBUTE_UNUSED
, int priority
)
5557 if (! INSN_P (insn
))
5560 if (s390_tune
!= PROCESSOR_2084_Z990
5561 && s390_tune
!= PROCESSOR_2094_Z9_109
5562 && s390_tune
!= PROCESSOR_2097_Z10
5563 && s390_tune
!= PROCESSOR_2817_Z196
)
5566 switch (s390_safe_attr_type (insn
))
5570 priority
= priority
<< 3;
5574 priority
= priority
<< 1;
5583 /* The number of instructions that can be issued per cycle. */
5586 s390_issue_rate (void)
5590 case PROCESSOR_2084_Z990
:
5591 case PROCESSOR_2094_Z9_109
:
5592 case PROCESSOR_2817_Z196
:
5594 case PROCESSOR_2097_Z10
:
5602 s390_first_cycle_multipass_dfa_lookahead (void)
5607 /* Annotate every literal pool reference in X by an UNSPEC_LTREF expression.
5608 Fix up MEMs as required. */
5611 annotate_constant_pool_refs (rtx
*x
)
5616 gcc_assert (GET_CODE (*x
) != SYMBOL_REF
5617 || !CONSTANT_POOL_ADDRESS_P (*x
));
5619 /* Literal pool references can only occur inside a MEM ... */
5620 if (GET_CODE (*x
) == MEM
)
5622 rtx memref
= XEXP (*x
, 0);
5624 if (GET_CODE (memref
) == SYMBOL_REF
5625 && CONSTANT_POOL_ADDRESS_P (memref
))
5627 rtx base
= cfun
->machine
->base_reg
;
5628 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, memref
, base
),
5631 *x
= replace_equiv_address (*x
, addr
);
5635 if (GET_CODE (memref
) == CONST
5636 && GET_CODE (XEXP (memref
, 0)) == PLUS
5637 && GET_CODE (XEXP (XEXP (memref
, 0), 1)) == CONST_INT
5638 && GET_CODE (XEXP (XEXP (memref
, 0), 0)) == SYMBOL_REF
5639 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (memref
, 0), 0)))
5641 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (memref
, 0), 1));
5642 rtx sym
= XEXP (XEXP (memref
, 0), 0);
5643 rtx base
= cfun
->machine
->base_reg
;
5644 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5647 *x
= replace_equiv_address (*x
, plus_constant (addr
, off
));
5652 /* ... or a load-address type pattern. */
5653 if (GET_CODE (*x
) == SET
)
5655 rtx addrref
= SET_SRC (*x
);
5657 if (GET_CODE (addrref
) == SYMBOL_REF
5658 && CONSTANT_POOL_ADDRESS_P (addrref
))
5660 rtx base
= cfun
->machine
->base_reg
;
5661 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, addrref
, base
),
5664 SET_SRC (*x
) = addr
;
5668 if (GET_CODE (addrref
) == CONST
5669 && GET_CODE (XEXP (addrref
, 0)) == PLUS
5670 && GET_CODE (XEXP (XEXP (addrref
, 0), 1)) == CONST_INT
5671 && GET_CODE (XEXP (XEXP (addrref
, 0), 0)) == SYMBOL_REF
5672 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (addrref
, 0), 0)))
5674 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (addrref
, 0), 1));
5675 rtx sym
= XEXP (XEXP (addrref
, 0), 0);
5676 rtx base
= cfun
->machine
->base_reg
;
5677 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5680 SET_SRC (*x
) = plus_constant (addr
, off
);
5685 /* Annotate LTREL_BASE as well. */
5686 if (GET_CODE (*x
) == UNSPEC
5687 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5689 rtx base
= cfun
->machine
->base_reg
;
5690 *x
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XVECEXP (*x
, 0, 0), base
),
5695 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5696 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5700 annotate_constant_pool_refs (&XEXP (*x
, i
));
5702 else if (fmt
[i
] == 'E')
5704 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5705 annotate_constant_pool_refs (&XVECEXP (*x
, i
, j
));
5710 /* Split all branches that exceed the maximum distance.
5711 Returns true if this created a new literal pool entry. */
5714 s390_split_branches (void)
5716 rtx temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
5717 int new_literal
= 0, ret
;
5718 rtx insn
, pat
, tmp
, target
;
5721 /* We need correct insn addresses. */
5723 shorten_branches (get_insns ());
5725 /* Find all branches that exceed 64KB, and split them. */
5727 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5729 if (GET_CODE (insn
) != JUMP_INSN
)
5732 pat
= PATTERN (insn
);
5733 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
5734 pat
= XVECEXP (pat
, 0, 0);
5735 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
5738 if (GET_CODE (SET_SRC (pat
)) == LABEL_REF
)
5740 label
= &SET_SRC (pat
);
5742 else if (GET_CODE (SET_SRC (pat
)) == IF_THEN_ELSE
)
5744 if (GET_CODE (XEXP (SET_SRC (pat
), 1)) == LABEL_REF
)
5745 label
= &XEXP (SET_SRC (pat
), 1);
5746 else if (GET_CODE (XEXP (SET_SRC (pat
), 2)) == LABEL_REF
)
5747 label
= &XEXP (SET_SRC (pat
), 2);
5754 if (get_attr_length (insn
) <= 4)
5757 /* We are going to use the return register as scratch register,
5758 make sure it will be saved/restored by the prologue/epilogue. */
5759 cfun_frame_layout
.save_return_addr_p
= 1;
5764 tmp
= force_const_mem (Pmode
, *label
);
5765 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, tmp
), insn
);
5766 INSN_ADDRESSES_NEW (tmp
, -1);
5767 annotate_constant_pool_refs (&PATTERN (tmp
));
5774 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, *label
),
5775 UNSPEC_LTREL_OFFSET
);
5776 target
= gen_rtx_CONST (Pmode
, target
);
5777 target
= force_const_mem (Pmode
, target
);
5778 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, target
), insn
);
5779 INSN_ADDRESSES_NEW (tmp
, -1);
5780 annotate_constant_pool_refs (&PATTERN (tmp
));
5782 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XEXP (target
, 0),
5783 cfun
->machine
->base_reg
),
5785 target
= gen_rtx_PLUS (Pmode
, temp_reg
, target
);
5788 ret
= validate_change (insn
, label
, target
, 0);
5796 /* Find an annotated literal pool symbol referenced in RTX X,
5797 and store it at REF. Will abort if X contains references to
5798 more than one such pool symbol; multiple references to the same
5799 symbol are allowed, however.
5801 The rtx pointed to by REF must be initialized to NULL_RTX
5802 by the caller before calling this routine. */
5805 find_constant_pool_ref (rtx x
, rtx
*ref
)
5810 /* Ignore LTREL_BASE references. */
5811 if (GET_CODE (x
) == UNSPEC
5812 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5814 /* Likewise POOL_ENTRY insns. */
5815 if (GET_CODE (x
) == UNSPEC_VOLATILE
5816 && XINT (x
, 1) == UNSPECV_POOL_ENTRY
)
5819 gcc_assert (GET_CODE (x
) != SYMBOL_REF
5820 || !CONSTANT_POOL_ADDRESS_P (x
));
5822 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_LTREF
)
5824 rtx sym
= XVECEXP (x
, 0, 0);
5825 gcc_assert (GET_CODE (sym
) == SYMBOL_REF
5826 && CONSTANT_POOL_ADDRESS_P (sym
));
5828 if (*ref
== NULL_RTX
)
5831 gcc_assert (*ref
== sym
);
5836 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5837 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5841 find_constant_pool_ref (XEXP (x
, i
), ref
);
5843 else if (fmt
[i
] == 'E')
5845 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5846 find_constant_pool_ref (XVECEXP (x
, i
, j
), ref
);
5851 /* Replace every reference to the annotated literal pool
5852 symbol REF in X by its base plus OFFSET. */
5855 replace_constant_pool_ref (rtx
*x
, rtx ref
, rtx offset
)
5860 gcc_assert (*x
!= ref
);
5862 if (GET_CODE (*x
) == UNSPEC
5863 && XINT (*x
, 1) == UNSPEC_LTREF
5864 && XVECEXP (*x
, 0, 0) == ref
)
5866 *x
= gen_rtx_PLUS (Pmode
, XVECEXP (*x
, 0, 1), offset
);
5870 if (GET_CODE (*x
) == PLUS
5871 && GET_CODE (XEXP (*x
, 1)) == CONST_INT
5872 && GET_CODE (XEXP (*x
, 0)) == UNSPEC
5873 && XINT (XEXP (*x
, 0), 1) == UNSPEC_LTREF
5874 && XVECEXP (XEXP (*x
, 0), 0, 0) == ref
)
5876 rtx addr
= gen_rtx_PLUS (Pmode
, XVECEXP (XEXP (*x
, 0), 0, 1), offset
);
5877 *x
= plus_constant (addr
, INTVAL (XEXP (*x
, 1)));
5881 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5882 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5886 replace_constant_pool_ref (&XEXP (*x
, i
), ref
, offset
);
5888 else if (fmt
[i
] == 'E')
5890 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5891 replace_constant_pool_ref (&XVECEXP (*x
, i
, j
), ref
, offset
);
5896 /* Check whether X contains an UNSPEC_LTREL_BASE.
5897 Return its constant pool symbol if found, NULL_RTX otherwise. */
5900 find_ltrel_base (rtx x
)
5905 if (GET_CODE (x
) == UNSPEC
5906 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5907 return XVECEXP (x
, 0, 0);
5909 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5910 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5914 rtx fnd
= find_ltrel_base (XEXP (x
, i
));
5918 else if (fmt
[i
] == 'E')
5920 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5922 rtx fnd
= find_ltrel_base (XVECEXP (x
, i
, j
));
5932 /* Replace any occurrence of UNSPEC_LTREL_BASE in X with its base. */
5935 replace_ltrel_base (rtx
*x
)
5940 if (GET_CODE (*x
) == UNSPEC
5941 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5943 *x
= XVECEXP (*x
, 0, 1);
5947 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5948 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5952 replace_ltrel_base (&XEXP (*x
, i
));
5954 else if (fmt
[i
] == 'E')
5956 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5957 replace_ltrel_base (&XVECEXP (*x
, i
, j
));
5963 /* We keep a list of constants which we have to add to internal
5964 constant tables in the middle of large functions. */
5966 #define NR_C_MODES 11
5967 enum machine_mode constant_modes
[NR_C_MODES
] =
5969 TFmode
, TImode
, TDmode
,
5970 DFmode
, DImode
, DDmode
,
5971 SFmode
, SImode
, SDmode
,
5978 struct constant
*next
;
5983 struct constant_pool
5985 struct constant_pool
*next
;
5989 rtx emit_pool_after
;
5991 struct constant
*constants
[NR_C_MODES
];
5992 struct constant
*execute
;
5997 /* Allocate new constant_pool structure. */
5999 static struct constant_pool
*
6000 s390_alloc_pool (void)
6002 struct constant_pool
*pool
;
6005 pool
= (struct constant_pool
*) xmalloc (sizeof *pool
);
6007 for (i
= 0; i
< NR_C_MODES
; i
++)
6008 pool
->constants
[i
] = NULL
;
6010 pool
->execute
= NULL
;
6011 pool
->label
= gen_label_rtx ();
6012 pool
->first_insn
= NULL_RTX
;
6013 pool
->pool_insn
= NULL_RTX
;
6014 pool
->insns
= BITMAP_ALLOC (NULL
);
6016 pool
->emit_pool_after
= NULL_RTX
;
6021 /* Create new constant pool covering instructions starting at INSN
6022 and chain it to the end of POOL_LIST. */
6024 static struct constant_pool
*
6025 s390_start_pool (struct constant_pool
**pool_list
, rtx insn
)
6027 struct constant_pool
*pool
, **prev
;
6029 pool
= s390_alloc_pool ();
6030 pool
->first_insn
= insn
;
6032 for (prev
= pool_list
; *prev
; prev
= &(*prev
)->next
)
6039 /* End range of instructions covered by POOL at INSN and emit
6040 placeholder insn representing the pool. */
6043 s390_end_pool (struct constant_pool
*pool
, rtx insn
)
6045 rtx pool_size
= GEN_INT (pool
->size
+ 8 /* alignment slop */);
6048 insn
= get_last_insn ();
6050 pool
->pool_insn
= emit_insn_after (gen_pool (pool_size
), insn
);
6051 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6054 /* Add INSN to the list of insns covered by POOL. */
6057 s390_add_pool_insn (struct constant_pool
*pool
, rtx insn
)
6059 bitmap_set_bit (pool
->insns
, INSN_UID (insn
));
6062 /* Return pool out of POOL_LIST that covers INSN. */
6064 static struct constant_pool
*
6065 s390_find_pool (struct constant_pool
*pool_list
, rtx insn
)
6067 struct constant_pool
*pool
;
6069 for (pool
= pool_list
; pool
; pool
= pool
->next
)
6070 if (bitmap_bit_p (pool
->insns
, INSN_UID (insn
)))
6076 /* Add constant VAL of mode MODE to the constant pool POOL. */
6079 s390_add_constant (struct constant_pool
*pool
, rtx val
, enum machine_mode mode
)
6084 for (i
= 0; i
< NR_C_MODES
; i
++)
6085 if (constant_modes
[i
] == mode
)
6087 gcc_assert (i
!= NR_C_MODES
);
6089 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6090 if (rtx_equal_p (val
, c
->value
))
6095 c
= (struct constant
*) xmalloc (sizeof *c
);
6097 c
->label
= gen_label_rtx ();
6098 c
->next
= pool
->constants
[i
];
6099 pool
->constants
[i
] = c
;
6100 pool
->size
+= GET_MODE_SIZE (mode
);
6104 /* Return an rtx that represents the offset of X from the start of
6108 s390_pool_offset (struct constant_pool
*pool
, rtx x
)
6112 label
= gen_rtx_LABEL_REF (GET_MODE (x
), pool
->label
);
6113 x
= gen_rtx_UNSPEC (GET_MODE (x
), gen_rtvec (2, x
, label
),
6114 UNSPEC_POOL_OFFSET
);
6115 return gen_rtx_CONST (GET_MODE (x
), x
);
6118 /* Find constant VAL of mode MODE in the constant pool POOL.
6119 Return an RTX describing the distance from the start of
6120 the pool to the location of the new constant. */
6123 s390_find_constant (struct constant_pool
*pool
, rtx val
,
6124 enum machine_mode mode
)
6129 for (i
= 0; i
< NR_C_MODES
; i
++)
6130 if (constant_modes
[i
] == mode
)
6132 gcc_assert (i
!= NR_C_MODES
);
6134 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6135 if (rtx_equal_p (val
, c
->value
))
6140 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6143 /* Check whether INSN is an execute. Return the label_ref to its
6144 execute target template if so, NULL_RTX otherwise. */
6147 s390_execute_label (rtx insn
)
6149 if (GET_CODE (insn
) == INSN
6150 && GET_CODE (PATTERN (insn
)) == PARALLEL
6151 && GET_CODE (XVECEXP (PATTERN (insn
), 0, 0)) == UNSPEC
6152 && XINT (XVECEXP (PATTERN (insn
), 0, 0), 1) == UNSPEC_EXECUTE
)
6153 return XVECEXP (XVECEXP (PATTERN (insn
), 0, 0), 0, 2);
6158 /* Add execute target for INSN to the constant pool POOL. */
6161 s390_add_execute (struct constant_pool
*pool
, rtx insn
)
6165 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6166 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6171 c
= (struct constant
*) xmalloc (sizeof *c
);
6173 c
->label
= gen_label_rtx ();
6174 c
->next
= pool
->execute
;
6180 /* Find execute target for INSN in the constant pool POOL.
6181 Return an RTX describing the distance from the start of
6182 the pool to the location of the execute target. */
6185 s390_find_execute (struct constant_pool
*pool
, rtx insn
)
6189 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6190 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6195 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6198 /* For an execute INSN, extract the execute target template. */
6201 s390_execute_target (rtx insn
)
6203 rtx pattern
= PATTERN (insn
);
6204 gcc_assert (s390_execute_label (insn
));
6206 if (XVECLEN (pattern
, 0) == 2)
6208 pattern
= copy_rtx (XVECEXP (pattern
, 0, 1));
6212 rtvec vec
= rtvec_alloc (XVECLEN (pattern
, 0) - 1);
6215 for (i
= 0; i
< XVECLEN (pattern
, 0) - 1; i
++)
6216 RTVEC_ELT (vec
, i
) = copy_rtx (XVECEXP (pattern
, 0, i
+ 1));
6218 pattern
= gen_rtx_PARALLEL (VOIDmode
, vec
);
6224 /* Indicate that INSN cannot be duplicated. This is the case for
6225 execute insns that carry a unique label. */
6228 s390_cannot_copy_insn_p (rtx insn
)
6230 rtx label
= s390_execute_label (insn
);
6231 return label
&& label
!= const0_rtx
;
6234 /* Dump out the constants in POOL. If REMOTE_LABEL is true,
6235 do not emit the pool base label. */
6238 s390_dump_pool (struct constant_pool
*pool
, bool remote_label
)
6241 rtx insn
= pool
->pool_insn
;
6244 /* Switch to rodata section. */
6245 if (TARGET_CPU_ZARCH
)
6247 insn
= emit_insn_after (gen_pool_section_start (), insn
);
6248 INSN_ADDRESSES_NEW (insn
, -1);
6251 /* Ensure minimum pool alignment. */
6252 if (TARGET_CPU_ZARCH
)
6253 insn
= emit_insn_after (gen_pool_align (GEN_INT (8)), insn
);
6255 insn
= emit_insn_after (gen_pool_align (GEN_INT (4)), insn
);
6256 INSN_ADDRESSES_NEW (insn
, -1);
6258 /* Emit pool base label. */
6261 insn
= emit_label_after (pool
->label
, insn
);
6262 INSN_ADDRESSES_NEW (insn
, -1);
6265 /* Dump constants in descending alignment requirement order,
6266 ensuring proper alignment for every constant. */
6267 for (i
= 0; i
< NR_C_MODES
; i
++)
6268 for (c
= pool
->constants
[i
]; c
; c
= c
->next
)
6270 /* Convert UNSPEC_LTREL_OFFSET unspecs to pool-relative references. */
6271 rtx value
= copy_rtx (c
->value
);
6272 if (GET_CODE (value
) == CONST
6273 && GET_CODE (XEXP (value
, 0)) == UNSPEC
6274 && XINT (XEXP (value
, 0), 1) == UNSPEC_LTREL_OFFSET
6275 && XVECLEN (XEXP (value
, 0), 0) == 1)
6276 value
= s390_pool_offset (pool
, XVECEXP (XEXP (value
, 0), 0, 0));
6278 insn
= emit_label_after (c
->label
, insn
);
6279 INSN_ADDRESSES_NEW (insn
, -1);
6281 value
= gen_rtx_UNSPEC_VOLATILE (constant_modes
[i
],
6282 gen_rtvec (1, value
),
6283 UNSPECV_POOL_ENTRY
);
6284 insn
= emit_insn_after (value
, insn
);
6285 INSN_ADDRESSES_NEW (insn
, -1);
6288 /* Ensure minimum alignment for instructions. */
6289 insn
= emit_insn_after (gen_pool_align (GEN_INT (2)), insn
);
6290 INSN_ADDRESSES_NEW (insn
, -1);
6292 /* Output in-pool execute template insns. */
6293 for (c
= pool
->execute
; c
; c
= c
->next
)
6295 insn
= emit_label_after (c
->label
, insn
);
6296 INSN_ADDRESSES_NEW (insn
, -1);
6298 insn
= emit_insn_after (s390_execute_target (c
->value
), insn
);
6299 INSN_ADDRESSES_NEW (insn
, -1);
6302 /* Switch back to previous section. */
6303 if (TARGET_CPU_ZARCH
)
6305 insn
= emit_insn_after (gen_pool_section_end (), insn
);
6306 INSN_ADDRESSES_NEW (insn
, -1);
6309 insn
= emit_barrier_after (insn
);
6310 INSN_ADDRESSES_NEW (insn
, -1);
6312 /* Remove placeholder insn. */
6313 remove_insn (pool
->pool_insn
);
6316 /* Free all memory used by POOL. */
6319 s390_free_pool (struct constant_pool
*pool
)
6321 struct constant
*c
, *next
;
6324 for (i
= 0; i
< NR_C_MODES
; i
++)
6325 for (c
= pool
->constants
[i
]; c
; c
= next
)
6331 for (c
= pool
->execute
; c
; c
= next
)
6337 BITMAP_FREE (pool
->insns
);
6342 /* Collect main literal pool. Return NULL on overflow. */
6344 static struct constant_pool
*
6345 s390_mainpool_start (void)
6347 struct constant_pool
*pool
;
6350 pool
= s390_alloc_pool ();
6352 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6354 if (GET_CODE (insn
) == INSN
6355 && GET_CODE (PATTERN (insn
)) == SET
6356 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC_VOLATILE
6357 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPECV_MAIN_POOL
)
6359 gcc_assert (!pool
->pool_insn
);
6360 pool
->pool_insn
= insn
;
6363 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6365 s390_add_execute (pool
, insn
);
6367 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6369 rtx pool_ref
= NULL_RTX
;
6370 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6373 rtx constant
= get_pool_constant (pool_ref
);
6374 enum machine_mode mode
= get_pool_mode (pool_ref
);
6375 s390_add_constant (pool
, constant
, mode
);
6379 /* If hot/cold partitioning is enabled we have to make sure that
6380 the literal pool is emitted in the same section where the
6381 initialization of the literal pool base pointer takes place.
6382 emit_pool_after is only used in the non-overflow case on non
6383 Z cpus where we can emit the literal pool at the end of the
6384 function body within the text section. */
6386 && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
6387 && !pool
->emit_pool_after
)
6388 pool
->emit_pool_after
= PREV_INSN (insn
);
6391 gcc_assert (pool
->pool_insn
|| pool
->size
== 0);
6393 if (pool
->size
>= 4096)
6395 /* We're going to chunkify the pool, so remove the main
6396 pool placeholder insn. */
6397 remove_insn (pool
->pool_insn
);
6399 s390_free_pool (pool
);
6403 /* If the functions ends with the section where the literal pool
6404 should be emitted set the marker to its end. */
6405 if (pool
&& !pool
->emit_pool_after
)
6406 pool
->emit_pool_after
= get_last_insn ();
6411 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6412 Modify the current function to output the pool constants as well as
6413 the pool register setup instruction. */
6416 s390_mainpool_finish (struct constant_pool
*pool
)
6418 rtx base_reg
= cfun
->machine
->base_reg
;
6421 /* If the pool is empty, we're done. */
6422 if (pool
->size
== 0)
6424 /* We don't actually need a base register after all. */
6425 cfun
->machine
->base_reg
= NULL_RTX
;
6427 if (pool
->pool_insn
)
6428 remove_insn (pool
->pool_insn
);
6429 s390_free_pool (pool
);
6433 /* We need correct insn addresses. */
6434 shorten_branches (get_insns ());
6436 /* On zSeries, we use a LARL to load the pool register. The pool is
6437 located in the .rodata section, so we emit it after the function. */
6438 if (TARGET_CPU_ZARCH
)
6440 insn
= gen_main_base_64 (base_reg
, pool
->label
);
6441 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6442 INSN_ADDRESSES_NEW (insn
, -1);
6443 remove_insn (pool
->pool_insn
);
6445 insn
= get_last_insn ();
6446 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6447 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6449 s390_dump_pool (pool
, 0);
6452 /* On S/390, if the total size of the function's code plus literal pool
6453 does not exceed 4096 bytes, we use BASR to set up a function base
6454 pointer, and emit the literal pool at the end of the function. */
6455 else if (INSN_ADDRESSES (INSN_UID (pool
->emit_pool_after
))
6456 + pool
->size
+ 8 /* alignment slop */ < 4096)
6458 insn
= gen_main_base_31_small (base_reg
, pool
->label
);
6459 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6460 INSN_ADDRESSES_NEW (insn
, -1);
6461 remove_insn (pool
->pool_insn
);
6463 insn
= emit_label_after (pool
->label
, insn
);
6464 INSN_ADDRESSES_NEW (insn
, -1);
6466 /* emit_pool_after will be set by s390_mainpool_start to the
6467 last insn of the section where the literal pool should be
6469 insn
= pool
->emit_pool_after
;
6471 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6472 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6474 s390_dump_pool (pool
, 1);
6477 /* Otherwise, we emit an inline literal pool and use BASR to branch
6478 over it, setting up the pool register at the same time. */
6481 rtx pool_end
= gen_label_rtx ();
6483 insn
= gen_main_base_31_large (base_reg
, pool
->label
, pool_end
);
6484 insn
= emit_jump_insn_after (insn
, pool
->pool_insn
);
6485 JUMP_LABEL (insn
) = pool_end
;
6486 INSN_ADDRESSES_NEW (insn
, -1);
6487 remove_insn (pool
->pool_insn
);
6489 insn
= emit_label_after (pool
->label
, insn
);
6490 INSN_ADDRESSES_NEW (insn
, -1);
6492 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6493 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6495 insn
= emit_label_after (pool_end
, pool
->pool_insn
);
6496 INSN_ADDRESSES_NEW (insn
, -1);
6498 s390_dump_pool (pool
, 1);
6502 /* Replace all literal pool references. */
6504 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6507 replace_ltrel_base (&PATTERN (insn
));
6509 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6511 rtx addr
, pool_ref
= NULL_RTX
;
6512 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6515 if (s390_execute_label (insn
))
6516 addr
= s390_find_execute (pool
, insn
);
6518 addr
= s390_find_constant (pool
, get_pool_constant (pool_ref
),
6519 get_pool_mode (pool_ref
));
6521 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6522 INSN_CODE (insn
) = -1;
6528 /* Free the pool. */
6529 s390_free_pool (pool
);
6532 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6533 We have decided we cannot use this pool, so revert all changes
6534 to the current function that were done by s390_mainpool_start. */
6536 s390_mainpool_cancel (struct constant_pool
*pool
)
6538 /* We didn't actually change the instruction stream, so simply
6539 free the pool memory. */
6540 s390_free_pool (pool
);
6544 /* Chunkify the literal pool. */
6546 #define S390_POOL_CHUNK_MIN 0xc00
6547 #define S390_POOL_CHUNK_MAX 0xe00
6549 static struct constant_pool
*
6550 s390_chunkify_start (void)
6552 struct constant_pool
*curr_pool
= NULL
, *pool_list
= NULL
;
6555 rtx pending_ltrel
= NULL_RTX
;
6558 rtx (*gen_reload_base
) (rtx
, rtx
) =
6559 TARGET_CPU_ZARCH
? gen_reload_base_64
: gen_reload_base_31
;
6562 /* We need correct insn addresses. */
6564 shorten_branches (get_insns ());
6566 /* Scan all insns and move literals to pool chunks. */
6568 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6570 bool section_switch_p
= false;
6572 /* Check for pending LTREL_BASE. */
6575 rtx ltrel_base
= find_ltrel_base (PATTERN (insn
));
6578 gcc_assert (ltrel_base
== pending_ltrel
);
6579 pending_ltrel
= NULL_RTX
;
6583 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6586 curr_pool
= s390_start_pool (&pool_list
, insn
);
6588 s390_add_execute (curr_pool
, insn
);
6589 s390_add_pool_insn (curr_pool
, insn
);
6591 else if (GET_CODE (insn
) == INSN
|| CALL_P (insn
))
6593 rtx pool_ref
= NULL_RTX
;
6594 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6597 rtx constant
= get_pool_constant (pool_ref
);
6598 enum machine_mode mode
= get_pool_mode (pool_ref
);
6601 curr_pool
= s390_start_pool (&pool_list
, insn
);
6603 s390_add_constant (curr_pool
, constant
, mode
);
6604 s390_add_pool_insn (curr_pool
, insn
);
6606 /* Don't split the pool chunk between a LTREL_OFFSET load
6607 and the corresponding LTREL_BASE. */
6608 if (GET_CODE (constant
) == CONST
6609 && GET_CODE (XEXP (constant
, 0)) == UNSPEC
6610 && XINT (XEXP (constant
, 0), 1) == UNSPEC_LTREL_OFFSET
)
6612 gcc_assert (!pending_ltrel
);
6613 pending_ltrel
= pool_ref
;
6618 if (GET_CODE (insn
) == JUMP_INSN
|| GET_CODE (insn
) == CODE_LABEL
)
6621 s390_add_pool_insn (curr_pool
, insn
);
6622 /* An LTREL_BASE must follow within the same basic block. */
6623 gcc_assert (!pending_ltrel
);
6627 switch (NOTE_KIND (insn
))
6629 case NOTE_INSN_SWITCH_TEXT_SECTIONS
:
6630 section_switch_p
= true;
6632 case NOTE_INSN_VAR_LOCATION
:
6633 case NOTE_INSN_CALL_ARG_LOCATION
:
6640 || INSN_ADDRESSES_SIZE () <= (size_t) INSN_UID (insn
)
6641 || INSN_ADDRESSES (INSN_UID (insn
)) == -1)
6644 if (TARGET_CPU_ZARCH
)
6646 if (curr_pool
->size
< S390_POOL_CHUNK_MAX
)
6649 s390_end_pool (curr_pool
, NULL_RTX
);
6654 int chunk_size
= INSN_ADDRESSES (INSN_UID (insn
))
6655 - INSN_ADDRESSES (INSN_UID (curr_pool
->first_insn
))
6658 /* We will later have to insert base register reload insns.
6659 Those will have an effect on code size, which we need to
6660 consider here. This calculation makes rather pessimistic
6661 worst-case assumptions. */
6662 if (GET_CODE (insn
) == CODE_LABEL
)
6665 if (chunk_size
< S390_POOL_CHUNK_MIN
6666 && curr_pool
->size
< S390_POOL_CHUNK_MIN
6667 && !section_switch_p
)
6670 /* Pool chunks can only be inserted after BARRIERs ... */
6671 if (GET_CODE (insn
) == BARRIER
)
6673 s390_end_pool (curr_pool
, insn
);
6678 /* ... so if we don't find one in time, create one. */
6679 else if (chunk_size
> S390_POOL_CHUNK_MAX
6680 || curr_pool
->size
> S390_POOL_CHUNK_MAX
6681 || section_switch_p
)
6683 rtx label
, jump
, barrier
, next
, prev
;
6685 if (!section_switch_p
)
6687 /* We can insert the barrier only after a 'real' insn. */
6688 if (GET_CODE (insn
) != INSN
&& GET_CODE (insn
) != CALL_INSN
)
6690 if (get_attr_length (insn
) == 0)
6692 /* Don't separate LTREL_BASE from the corresponding
6693 LTREL_OFFSET load. */
6700 next
= NEXT_INSN (insn
);
6704 && (NOTE_KIND (next
) == NOTE_INSN_VAR_LOCATION
6705 || NOTE_KIND (next
) == NOTE_INSN_CALL_ARG_LOCATION
));
6709 gcc_assert (!pending_ltrel
);
6711 /* The old pool has to end before the section switch
6712 note in order to make it part of the current
6714 insn
= PREV_INSN (insn
);
6717 label
= gen_label_rtx ();
6719 if (prev
&& NOTE_P (prev
))
6720 prev
= prev_nonnote_insn (prev
);
6722 jump
= emit_jump_insn_after_setloc (gen_jump (label
), insn
,
6723 INSN_LOCATOR (prev
));
6725 jump
= emit_jump_insn_after_noloc (gen_jump (label
), insn
);
6726 barrier
= emit_barrier_after (jump
);
6727 insn
= emit_label_after (label
, barrier
);
6728 JUMP_LABEL (jump
) = label
;
6729 LABEL_NUSES (label
) = 1;
6731 INSN_ADDRESSES_NEW (jump
, -1);
6732 INSN_ADDRESSES_NEW (barrier
, -1);
6733 INSN_ADDRESSES_NEW (insn
, -1);
6735 s390_end_pool (curr_pool
, barrier
);
6743 s390_end_pool (curr_pool
, NULL_RTX
);
6744 gcc_assert (!pending_ltrel
);
6746 /* Find all labels that are branched into
6747 from an insn belonging to a different chunk. */
6749 far_labels
= BITMAP_ALLOC (NULL
);
6751 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6753 /* Labels marked with LABEL_PRESERVE_P can be target
6754 of non-local jumps, so we have to mark them.
6755 The same holds for named labels.
6757 Don't do that, however, if it is the label before
6760 if (GET_CODE (insn
) == CODE_LABEL
6761 && (LABEL_PRESERVE_P (insn
) || LABEL_NAME (insn
)))
6763 rtx vec_insn
= next_real_insn (insn
);
6764 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6765 PATTERN (vec_insn
) : NULL_RTX
;
6767 || !(GET_CODE (vec_pat
) == ADDR_VEC
6768 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6769 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (insn
));
6772 /* If we have a direct jump (conditional or unconditional)
6773 or a casesi jump, check all potential targets. */
6774 else if (GET_CODE (insn
) == JUMP_INSN
)
6776 rtx pat
= PATTERN (insn
);
6777 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
6778 pat
= XVECEXP (pat
, 0, 0);
6780 if (GET_CODE (pat
) == SET
)
6782 rtx label
= JUMP_LABEL (insn
);
6785 if (s390_find_pool (pool_list
, label
)
6786 != s390_find_pool (pool_list
, insn
))
6787 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6790 else if (GET_CODE (pat
) == PARALLEL
6791 && XVECLEN (pat
, 0) == 2
6792 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
6793 && GET_CODE (XVECEXP (pat
, 0, 1)) == USE
6794 && GET_CODE (XEXP (XVECEXP (pat
, 0, 1), 0)) == LABEL_REF
)
6796 /* Find the jump table used by this casesi jump. */
6797 rtx vec_label
= XEXP (XEXP (XVECEXP (pat
, 0, 1), 0), 0);
6798 rtx vec_insn
= next_real_insn (vec_label
);
6799 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6800 PATTERN (vec_insn
) : NULL_RTX
;
6802 && (GET_CODE (vec_pat
) == ADDR_VEC
6803 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6805 int i
, diff_p
= GET_CODE (vec_pat
) == ADDR_DIFF_VEC
;
6807 for (i
= 0; i
< XVECLEN (vec_pat
, diff_p
); i
++)
6809 rtx label
= XEXP (XVECEXP (vec_pat
, diff_p
, i
), 0);
6811 if (s390_find_pool (pool_list
, label
)
6812 != s390_find_pool (pool_list
, insn
))
6813 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6820 /* Insert base register reload insns before every pool. */
6822 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6824 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6826 rtx insn
= curr_pool
->first_insn
;
6827 INSN_ADDRESSES_NEW (emit_insn_before (new_insn
, insn
), -1);
6830 /* Insert base register reload insns at every far label. */
6832 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6833 if (GET_CODE (insn
) == CODE_LABEL
6834 && bitmap_bit_p (far_labels
, CODE_LABEL_NUMBER (insn
)))
6836 struct constant_pool
*pool
= s390_find_pool (pool_list
, insn
);
6839 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6841 INSN_ADDRESSES_NEW (emit_insn_after (new_insn
, insn
), -1);
6846 BITMAP_FREE (far_labels
);
6849 /* Recompute insn addresses. */
6851 init_insn_lengths ();
6852 shorten_branches (get_insns ());
6857 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6858 After we have decided to use this list, finish implementing
6859 all changes to the current function as required. */
6862 s390_chunkify_finish (struct constant_pool
*pool_list
)
6864 struct constant_pool
*curr_pool
= NULL
;
6868 /* Replace all literal pool references. */
6870 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6873 replace_ltrel_base (&PATTERN (insn
));
6875 curr_pool
= s390_find_pool (pool_list
, insn
);
6879 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6881 rtx addr
, pool_ref
= NULL_RTX
;
6882 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6885 if (s390_execute_label (insn
))
6886 addr
= s390_find_execute (curr_pool
, insn
);
6888 addr
= s390_find_constant (curr_pool
,
6889 get_pool_constant (pool_ref
),
6890 get_pool_mode (pool_ref
));
6892 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6893 INSN_CODE (insn
) = -1;
6898 /* Dump out all literal pools. */
6900 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6901 s390_dump_pool (curr_pool
, 0);
6903 /* Free pool list. */
6907 struct constant_pool
*next
= pool_list
->next
;
6908 s390_free_pool (pool_list
);
6913 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6914 We have decided we cannot use this list, so revert all changes
6915 to the current function that were done by s390_chunkify_start. */
6918 s390_chunkify_cancel (struct constant_pool
*pool_list
)
6920 struct constant_pool
*curr_pool
= NULL
;
6923 /* Remove all pool placeholder insns. */
6925 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6927 /* Did we insert an extra barrier? Remove it. */
6928 rtx barrier
= PREV_INSN (curr_pool
->pool_insn
);
6929 rtx jump
= barrier
? PREV_INSN (barrier
) : NULL_RTX
;
6930 rtx label
= NEXT_INSN (curr_pool
->pool_insn
);
6932 if (jump
&& GET_CODE (jump
) == JUMP_INSN
6933 && barrier
&& GET_CODE (barrier
) == BARRIER
6934 && label
&& GET_CODE (label
) == CODE_LABEL
6935 && GET_CODE (PATTERN (jump
)) == SET
6936 && SET_DEST (PATTERN (jump
)) == pc_rtx
6937 && GET_CODE (SET_SRC (PATTERN (jump
))) == LABEL_REF
6938 && XEXP (SET_SRC (PATTERN (jump
)), 0) == label
)
6941 remove_insn (barrier
);
6942 remove_insn (label
);
6945 remove_insn (curr_pool
->pool_insn
);
6948 /* Remove all base register reload insns. */
6950 for (insn
= get_insns (); insn
; )
6952 rtx next_insn
= NEXT_INSN (insn
);
6954 if (GET_CODE (insn
) == INSN
6955 && GET_CODE (PATTERN (insn
)) == SET
6956 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC
6957 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPEC_RELOAD_BASE
)
6963 /* Free pool list. */
6967 struct constant_pool
*next
= pool_list
->next
;
6968 s390_free_pool (pool_list
);
6973 /* Output the constant pool entry EXP in mode MODE with alignment ALIGN. */
6976 s390_output_pool_entry (rtx exp
, enum machine_mode mode
, unsigned int align
)
6980 switch (GET_MODE_CLASS (mode
))
6983 case MODE_DECIMAL_FLOAT
:
6984 gcc_assert (GET_CODE (exp
) == CONST_DOUBLE
);
6986 REAL_VALUE_FROM_CONST_DOUBLE (r
, exp
);
6987 assemble_real (r
, mode
, align
);
6991 assemble_integer (exp
, GET_MODE_SIZE (mode
), align
, 1);
6992 mark_symbol_refs_as_used (exp
);
7001 /* Return an RTL expression representing the value of the return address
7002 for the frame COUNT steps up from the current frame. FRAME is the
7003 frame pointer of that frame. */
7006 s390_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
7011 /* Without backchain, we fail for all but the current frame. */
7013 if (!TARGET_BACKCHAIN
&& count
> 0)
7016 /* For the current frame, we need to make sure the initial
7017 value of RETURN_REGNUM is actually saved. */
7021 /* On non-z architectures branch splitting could overwrite r14. */
7022 if (TARGET_CPU_ZARCH
)
7023 return get_hard_reg_initial_val (Pmode
, RETURN_REGNUM
);
7026 cfun_frame_layout
.save_return_addr_p
= true;
7027 return gen_rtx_MEM (Pmode
, return_address_pointer_rtx
);
7031 if (TARGET_PACKED_STACK
)
7032 offset
= -2 * UNITS_PER_LONG
;
7034 offset
= RETURN_REGNUM
* UNITS_PER_LONG
;
7036 addr
= plus_constant (frame
, offset
);
7037 addr
= memory_address (Pmode
, addr
);
7038 return gen_rtx_MEM (Pmode
, addr
);
7041 /* Return an RTL expression representing the back chain stored in
7042 the current stack frame. */
7045 s390_back_chain_rtx (void)
7049 gcc_assert (TARGET_BACKCHAIN
);
7051 if (TARGET_PACKED_STACK
)
7052 chain
= plus_constant (stack_pointer_rtx
,
7053 STACK_POINTER_OFFSET
- UNITS_PER_LONG
);
7055 chain
= stack_pointer_rtx
;
7057 chain
= gen_rtx_MEM (Pmode
, chain
);
7061 /* Find first call clobbered register unused in a function.
7062 This could be used as base register in a leaf function
7063 or for holding the return address before epilogue. */
7066 find_unused_clobbered_reg (void)
7069 for (i
= 0; i
< 6; i
++)
7070 if (!df_regs_ever_live_p (i
))
7076 /* Helper function for s390_regs_ever_clobbered. Sets the fields in DATA for all
7077 clobbered hard regs in SETREG. */
7080 s390_reg_clobbered_rtx (rtx setreg
, const_rtx set_insn ATTRIBUTE_UNUSED
, void *data
)
7082 int *regs_ever_clobbered
= (int *)data
;
7083 unsigned int i
, regno
;
7084 enum machine_mode mode
= GET_MODE (setreg
);
7086 if (GET_CODE (setreg
) == SUBREG
)
7088 rtx inner
= SUBREG_REG (setreg
);
7089 if (!GENERAL_REG_P (inner
))
7091 regno
= subreg_regno (setreg
);
7093 else if (GENERAL_REG_P (setreg
))
7094 regno
= REGNO (setreg
);
7099 i
< regno
+ HARD_REGNO_NREGS (regno
, mode
);
7101 regs_ever_clobbered
[i
] = 1;
7104 /* Walks through all basic blocks of the current function looking
7105 for clobbered hard regs using s390_reg_clobbered_rtx. The fields
7106 of the passed integer array REGS_EVER_CLOBBERED are set to one for
7107 each of those regs. */
7110 s390_regs_ever_clobbered (int *regs_ever_clobbered
)
7116 memset (regs_ever_clobbered
, 0, 16 * sizeof (int));
7118 /* For non-leaf functions we have to consider all call clobbered regs to be
7120 if (!current_function_is_leaf
)
7122 for (i
= 0; i
< 16; i
++)
7123 regs_ever_clobbered
[i
] = call_really_used_regs
[i
];
7126 /* Make the "magic" eh_return registers live if necessary. For regs_ever_live
7127 this work is done by liveness analysis (mark_regs_live_at_end).
7128 Special care is needed for functions containing landing pads. Landing pads
7129 may use the eh registers, but the code which sets these registers is not
7130 contained in that function. Hence s390_regs_ever_clobbered is not able to
7131 deal with this automatically. */
7132 if (crtl
->calls_eh_return
|| cfun
->machine
->has_landing_pad_p
)
7133 for (i
= 0; EH_RETURN_DATA_REGNO (i
) != INVALID_REGNUM
; i
++)
7134 if (crtl
->calls_eh_return
7135 || (cfun
->machine
->has_landing_pad_p
7136 && df_regs_ever_live_p (EH_RETURN_DATA_REGNO (i
))))
7137 regs_ever_clobbered
[EH_RETURN_DATA_REGNO (i
)] = 1;
7139 /* For nonlocal gotos all call-saved registers have to be saved.
7140 This flag is also set for the unwinding code in libgcc.
7141 See expand_builtin_unwind_init. For regs_ever_live this is done by
7143 if (cfun
->has_nonlocal_label
)
7144 for (i
= 0; i
< 16; i
++)
7145 if (!call_really_used_regs
[i
])
7146 regs_ever_clobbered
[i
] = 1;
7148 FOR_EACH_BB (cur_bb
)
7150 FOR_BB_INSNS (cur_bb
, cur_insn
)
7152 if (INSN_P (cur_insn
))
7153 note_stores (PATTERN (cur_insn
),
7154 s390_reg_clobbered_rtx
,
7155 regs_ever_clobbered
);
7160 /* Determine the frame area which actually has to be accessed
7161 in the function epilogue. The values are stored at the
7162 given pointers AREA_BOTTOM (address of the lowest used stack
7163 address) and AREA_TOP (address of the first item which does
7164 not belong to the stack frame). */
7167 s390_frame_area (int *area_bottom
, int *area_top
)
7175 if (cfun_frame_layout
.first_restore_gpr
!= -1)
7177 b
= (cfun_frame_layout
.gprs_offset
7178 + cfun_frame_layout
.first_restore_gpr
* UNITS_PER_LONG
);
7179 t
= b
+ (cfun_frame_layout
.last_restore_gpr
7180 - cfun_frame_layout
.first_restore_gpr
+ 1) * UNITS_PER_LONG
;
7183 if (TARGET_64BIT
&& cfun_save_high_fprs_p
)
7185 b
= MIN (b
, cfun_frame_layout
.f8_offset
);
7186 t
= MAX (t
, (cfun_frame_layout
.f8_offset
7187 + cfun_frame_layout
.high_fprs
* 8));
7191 for (i
= 2; i
< 4; i
++)
7192 if (cfun_fpr_bit_p (i
))
7194 b
= MIN (b
, cfun_frame_layout
.f4_offset
+ (i
- 2) * 8);
7195 t
= MAX (t
, cfun_frame_layout
.f4_offset
+ (i
- 1) * 8);
7202 /* Fill cfun->machine with info about register usage of current function.
7203 Return in CLOBBERED_REGS which GPRs are currently considered set. */
7206 s390_register_info (int clobbered_regs
[])
7210 /* fprs 8 - 15 are call saved for 64 Bit ABI. */
7211 cfun_frame_layout
.fpr_bitmap
= 0;
7212 cfun_frame_layout
.high_fprs
= 0;
7214 for (i
= 24; i
< 32; i
++)
7215 if (df_regs_ever_live_p (i
) && !global_regs
[i
])
7217 cfun_set_fpr_bit (i
- 16);
7218 cfun_frame_layout
.high_fprs
++;
7221 /* Find first and last gpr to be saved. We trust regs_ever_live
7222 data, except that we don't save and restore global registers.
7224 Also, all registers with special meaning to the compiler need
7225 to be handled extra. */
7227 s390_regs_ever_clobbered (clobbered_regs
);
7229 for (i
= 0; i
< 16; i
++)
7230 clobbered_regs
[i
] = clobbered_regs
[i
] && !global_regs
[i
] && !fixed_regs
[i
];
7232 if (frame_pointer_needed
)
7233 clobbered_regs
[HARD_FRAME_POINTER_REGNUM
] = 1;
7236 clobbered_regs
[PIC_OFFSET_TABLE_REGNUM
]
7237 |= df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
);
7239 clobbered_regs
[BASE_REGNUM
]
7240 |= (cfun
->machine
->base_reg
7241 && REGNO (cfun
->machine
->base_reg
) == BASE_REGNUM
);
7243 clobbered_regs
[RETURN_REGNUM
]
7244 |= (!current_function_is_leaf
7245 || TARGET_TPF_PROFILING
7246 || cfun
->machine
->split_branches_pending_p
7247 || cfun_frame_layout
.save_return_addr_p
7248 || crtl
->calls_eh_return
7251 clobbered_regs
[STACK_POINTER_REGNUM
]
7252 |= (!current_function_is_leaf
7253 || TARGET_TPF_PROFILING
7254 || cfun_save_high_fprs_p
7255 || get_frame_size () > 0
7256 || cfun
->calls_alloca
7259 for (i
= 6; i
< 16; i
++)
7260 if (df_regs_ever_live_p (i
) || clobbered_regs
[i
])
7262 for (j
= 15; j
> i
; j
--)
7263 if (df_regs_ever_live_p (j
) || clobbered_regs
[j
])
7268 /* Nothing to save/restore. */
7269 cfun_frame_layout
.first_save_gpr_slot
= -1;
7270 cfun_frame_layout
.last_save_gpr_slot
= -1;
7271 cfun_frame_layout
.first_save_gpr
= -1;
7272 cfun_frame_layout
.first_restore_gpr
= -1;
7273 cfun_frame_layout
.last_save_gpr
= -1;
7274 cfun_frame_layout
.last_restore_gpr
= -1;
7278 /* Save slots for gprs from i to j. */
7279 cfun_frame_layout
.first_save_gpr_slot
= i
;
7280 cfun_frame_layout
.last_save_gpr_slot
= j
;
7282 for (i
= cfun_frame_layout
.first_save_gpr_slot
;
7283 i
< cfun_frame_layout
.last_save_gpr_slot
+ 1;
7285 if (clobbered_regs
[i
])
7288 for (j
= cfun_frame_layout
.last_save_gpr_slot
; j
> i
; j
--)
7289 if (clobbered_regs
[j
])
7292 if (i
== cfun_frame_layout
.last_save_gpr_slot
+ 1)
7294 /* Nothing to save/restore. */
7295 cfun_frame_layout
.first_save_gpr
= -1;
7296 cfun_frame_layout
.first_restore_gpr
= -1;
7297 cfun_frame_layout
.last_save_gpr
= -1;
7298 cfun_frame_layout
.last_restore_gpr
= -1;
7302 /* Save / Restore from gpr i to j. */
7303 cfun_frame_layout
.first_save_gpr
= i
;
7304 cfun_frame_layout
.first_restore_gpr
= i
;
7305 cfun_frame_layout
.last_save_gpr
= j
;
7306 cfun_frame_layout
.last_restore_gpr
= j
;
7312 /* Varargs functions need to save gprs 2 to 6. */
7313 if (cfun
->va_list_gpr_size
7314 && crtl
->args
.info
.gprs
< GP_ARG_NUM_REG
)
7316 int min_gpr
= crtl
->args
.info
.gprs
;
7317 int max_gpr
= min_gpr
+ cfun
->va_list_gpr_size
;
7318 if (max_gpr
> GP_ARG_NUM_REG
)
7319 max_gpr
= GP_ARG_NUM_REG
;
7321 if (cfun_frame_layout
.first_save_gpr
== -1
7322 || cfun_frame_layout
.first_save_gpr
> 2 + min_gpr
)
7324 cfun_frame_layout
.first_save_gpr
= 2 + min_gpr
;
7325 cfun_frame_layout
.first_save_gpr_slot
= 2 + min_gpr
;
7328 if (cfun_frame_layout
.last_save_gpr
== -1
7329 || cfun_frame_layout
.last_save_gpr
< 2 + max_gpr
- 1)
7331 cfun_frame_layout
.last_save_gpr
= 2 + max_gpr
- 1;
7332 cfun_frame_layout
.last_save_gpr_slot
= 2 + max_gpr
- 1;
7336 /* Mark f0, f2 for 31 bit and f0-f4 for 64 bit to be saved. */
7337 if (TARGET_HARD_FLOAT
&& cfun
->va_list_fpr_size
7338 && crtl
->args
.info
.fprs
< FP_ARG_NUM_REG
)
7340 int min_fpr
= crtl
->args
.info
.fprs
;
7341 int max_fpr
= min_fpr
+ cfun
->va_list_fpr_size
;
7342 if (max_fpr
> FP_ARG_NUM_REG
)
7343 max_fpr
= FP_ARG_NUM_REG
;
7345 /* ??? This is currently required to ensure proper location
7346 of the fpr save slots within the va_list save area. */
7347 if (TARGET_PACKED_STACK
)
7350 for (i
= min_fpr
; i
< max_fpr
; i
++)
7351 cfun_set_fpr_bit (i
);
7356 for (i
= 2; i
< 4; i
++)
7357 if (df_regs_ever_live_p (i
+ 16) && !global_regs
[i
+ 16])
7358 cfun_set_fpr_bit (i
);
7361 /* Fill cfun->machine with info about frame of current function. */
7364 s390_frame_info (void)
7368 cfun_frame_layout
.frame_size
= get_frame_size ();
7369 if (!TARGET_64BIT
&& cfun_frame_layout
.frame_size
> 0x7fff0000)
7370 fatal_error ("total size of local variables exceeds architecture limit");
7372 if (!TARGET_PACKED_STACK
)
7374 cfun_frame_layout
.backchain_offset
= 0;
7375 cfun_frame_layout
.f0_offset
= 16 * UNITS_PER_LONG
;
7376 cfun_frame_layout
.f4_offset
= cfun_frame_layout
.f0_offset
+ 2 * 8;
7377 cfun_frame_layout
.f8_offset
= -cfun_frame_layout
.high_fprs
* 8;
7378 cfun_frame_layout
.gprs_offset
= (cfun_frame_layout
.first_save_gpr_slot
7381 else if (TARGET_BACKCHAIN
) /* kernel stack layout */
7383 cfun_frame_layout
.backchain_offset
= (STACK_POINTER_OFFSET
7385 cfun_frame_layout
.gprs_offset
7386 = (cfun_frame_layout
.backchain_offset
7387 - (STACK_POINTER_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
+ 1)
7392 cfun_frame_layout
.f4_offset
7393 = (cfun_frame_layout
.gprs_offset
7394 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7396 cfun_frame_layout
.f0_offset
7397 = (cfun_frame_layout
.f4_offset
7398 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7402 /* On 31 bit we have to care about alignment of the
7403 floating point regs to provide fastest access. */
7404 cfun_frame_layout
.f0_offset
7405 = ((cfun_frame_layout
.gprs_offset
7406 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1))
7407 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7409 cfun_frame_layout
.f4_offset
7410 = (cfun_frame_layout
.f0_offset
7411 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7414 else /* no backchain */
7416 cfun_frame_layout
.f4_offset
7417 = (STACK_POINTER_OFFSET
7418 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7420 cfun_frame_layout
.f0_offset
7421 = (cfun_frame_layout
.f4_offset
7422 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7424 cfun_frame_layout
.gprs_offset
7425 = cfun_frame_layout
.f0_offset
- cfun_gprs_save_area_size
;
7428 if (current_function_is_leaf
7429 && !TARGET_TPF_PROFILING
7430 && cfun_frame_layout
.frame_size
== 0
7431 && !cfun_save_high_fprs_p
7432 && !cfun
->calls_alloca
7436 if (!TARGET_PACKED_STACK
)
7437 cfun_frame_layout
.frame_size
+= (STACK_POINTER_OFFSET
7438 + crtl
->outgoing_args_size
7439 + cfun_frame_layout
.high_fprs
* 8);
7442 if (TARGET_BACKCHAIN
)
7443 cfun_frame_layout
.frame_size
+= UNITS_PER_LONG
;
7445 /* No alignment trouble here because f8-f15 are only saved under
7447 cfun_frame_layout
.f8_offset
= (MIN (MIN (cfun_frame_layout
.f0_offset
,
7448 cfun_frame_layout
.f4_offset
),
7449 cfun_frame_layout
.gprs_offset
)
7450 - cfun_frame_layout
.high_fprs
* 8);
7452 cfun_frame_layout
.frame_size
+= cfun_frame_layout
.high_fprs
* 8;
7454 for (i
= 0; i
< 8; i
++)
7455 if (cfun_fpr_bit_p (i
))
7456 cfun_frame_layout
.frame_size
+= 8;
7458 cfun_frame_layout
.frame_size
+= cfun_gprs_save_area_size
;
7460 /* If under 31 bit an odd number of gprs has to be saved we have to adjust
7461 the frame size to sustain 8 byte alignment of stack frames. */
7462 cfun_frame_layout
.frame_size
= ((cfun_frame_layout
.frame_size
+
7463 STACK_BOUNDARY
/ BITS_PER_UNIT
- 1)
7464 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1));
7466 cfun_frame_layout
.frame_size
+= crtl
->outgoing_args_size
;
7470 /* Generate frame layout. Fills in register and frame data for the current
7471 function in cfun->machine. This routine can be called multiple times;
7472 it will re-do the complete frame layout every time. */
7475 s390_init_frame_layout (void)
7477 HOST_WIDE_INT frame_size
;
7479 int clobbered_regs
[16];
7481 /* On S/390 machines, we may need to perform branch splitting, which
7482 will require both base and return address register. We have no
7483 choice but to assume we're going to need them until right at the
7484 end of the machine dependent reorg phase. */
7485 if (!TARGET_CPU_ZARCH
)
7486 cfun
->machine
->split_branches_pending_p
= true;
7490 frame_size
= cfun_frame_layout
.frame_size
;
7492 /* Try to predict whether we'll need the base register. */
7493 base_used
= cfun
->machine
->split_branches_pending_p
7494 || crtl
->uses_const_pool
7495 || (!DISP_IN_RANGE (frame_size
)
7496 && !CONST_OK_FOR_K (frame_size
));
7498 /* Decide which register to use as literal pool base. In small
7499 leaf functions, try to use an unused call-clobbered register
7500 as base register to avoid save/restore overhead. */
7502 cfun
->machine
->base_reg
= NULL_RTX
;
7503 else if (current_function_is_leaf
&& !df_regs_ever_live_p (5))
7504 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, 5);
7506 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, BASE_REGNUM
);
7508 s390_register_info (clobbered_regs
);
7511 while (frame_size
!= cfun_frame_layout
.frame_size
);
7514 /* Update frame layout. Recompute actual register save data based on
7515 current info and update regs_ever_live for the special registers.
7516 May be called multiple times, but may never cause *more* registers
7517 to be saved than s390_init_frame_layout allocated room for. */
7520 s390_update_frame_layout (void)
7522 int clobbered_regs
[16];
7524 s390_register_info (clobbered_regs
);
7526 df_set_regs_ever_live (BASE_REGNUM
,
7527 clobbered_regs
[BASE_REGNUM
] ? true : false);
7528 df_set_regs_ever_live (RETURN_REGNUM
,
7529 clobbered_regs
[RETURN_REGNUM
] ? true : false);
7530 df_set_regs_ever_live (STACK_POINTER_REGNUM
,
7531 clobbered_regs
[STACK_POINTER_REGNUM
] ? true : false);
7533 if (cfun
->machine
->base_reg
)
7534 df_set_regs_ever_live (REGNO (cfun
->machine
->base_reg
), true);
7537 /* Return true if it is legal to put a value with MODE into REGNO. */
7540 s390_hard_regno_mode_ok (unsigned int regno
, enum machine_mode mode
)
7542 switch (REGNO_REG_CLASS (regno
))
7545 if (REGNO_PAIR_OK (regno
, mode
))
7547 if (mode
== SImode
|| mode
== DImode
)
7550 if (FLOAT_MODE_P (mode
) && GET_MODE_CLASS (mode
) != MODE_VECTOR_FLOAT
)
7555 if (FRAME_REGNO_P (regno
) && mode
== Pmode
)
7560 if (REGNO_PAIR_OK (regno
, mode
))
7563 || (mode
!= TFmode
&& mode
!= TCmode
&& mode
!= TDmode
))
7568 if (GET_MODE_CLASS (mode
) == MODE_CC
)
7572 if (REGNO_PAIR_OK (regno
, mode
))
7574 if (mode
== SImode
|| mode
== Pmode
)
7585 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
7588 s390_hard_regno_rename_ok (unsigned int old_reg
, unsigned int new_reg
)
7590 /* Once we've decided upon a register to use as base register, it must
7591 no longer be used for any other purpose. */
7592 if (cfun
->machine
->base_reg
)
7593 if (REGNO (cfun
->machine
->base_reg
) == old_reg
7594 || REGNO (cfun
->machine
->base_reg
) == new_reg
)
7600 /* Maximum number of registers to represent a value of mode MODE
7601 in a register of class RCLASS. */
7604 s390_class_max_nregs (enum reg_class rclass
, enum machine_mode mode
)
7609 if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7610 return 2 * ((GET_MODE_SIZE (mode
) / 2 + 8 - 1) / 8);
7612 return (GET_MODE_SIZE (mode
) + 8 - 1) / 8;
7614 return (GET_MODE_SIZE (mode
) + 4 - 1) / 4;
7618 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
7621 /* Return true if register FROM can be eliminated via register TO. */
7624 s390_can_eliminate (const int from
, const int to
)
7626 /* On zSeries machines, we have not marked the base register as fixed.
7627 Instead, we have an elimination rule BASE_REGNUM -> BASE_REGNUM.
7628 If a function requires the base register, we say here that this
7629 elimination cannot be performed. This will cause reload to free
7630 up the base register (as if it were fixed). On the other hand,
7631 if the current function does *not* require the base register, we
7632 say here the elimination succeeds, which in turn allows reload
7633 to allocate the base register for any other purpose. */
7634 if (from
== BASE_REGNUM
&& to
== BASE_REGNUM
)
7636 if (TARGET_CPU_ZARCH
)
7638 s390_init_frame_layout ();
7639 return cfun
->machine
->base_reg
== NULL_RTX
;
7645 /* Everything else must point into the stack frame. */
7646 gcc_assert (to
== STACK_POINTER_REGNUM
7647 || to
== HARD_FRAME_POINTER_REGNUM
);
7649 gcc_assert (from
== FRAME_POINTER_REGNUM
7650 || from
== ARG_POINTER_REGNUM
7651 || from
== RETURN_ADDRESS_POINTER_REGNUM
);
7653 /* Make sure we actually saved the return address. */
7654 if (from
== RETURN_ADDRESS_POINTER_REGNUM
)
7655 if (!crtl
->calls_eh_return
7657 && !cfun_frame_layout
.save_return_addr_p
)
7663 /* Return offset between register FROM and TO initially after prolog. */
7666 s390_initial_elimination_offset (int from
, int to
)
7668 HOST_WIDE_INT offset
;
7671 /* ??? Why are we called for non-eliminable pairs? */
7672 if (!s390_can_eliminate (from
, to
))
7677 case FRAME_POINTER_REGNUM
:
7678 offset
= (get_frame_size()
7679 + STACK_POINTER_OFFSET
7680 + crtl
->outgoing_args_size
);
7683 case ARG_POINTER_REGNUM
:
7684 s390_init_frame_layout ();
7685 offset
= cfun_frame_layout
.frame_size
+ STACK_POINTER_OFFSET
;
7688 case RETURN_ADDRESS_POINTER_REGNUM
:
7689 s390_init_frame_layout ();
7690 index
= RETURN_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
;
7691 gcc_assert (index
>= 0);
7692 offset
= cfun_frame_layout
.frame_size
+ cfun_frame_layout
.gprs_offset
;
7693 offset
+= index
* UNITS_PER_LONG
;
7707 /* Emit insn to save fpr REGNUM at offset OFFSET relative
7708 to register BASE. Return generated insn. */
7711 save_fpr (rtx base
, int offset
, int regnum
)
7714 addr
= gen_rtx_MEM (DFmode
, plus_constant (base
, offset
));
7716 if (regnum
>= 16 && regnum
<= (16 + FP_ARG_NUM_REG
))
7717 set_mem_alias_set (addr
, get_varargs_alias_set ());
7719 set_mem_alias_set (addr
, get_frame_alias_set ());
7721 return emit_move_insn (addr
, gen_rtx_REG (DFmode
, regnum
));
7724 /* Emit insn to restore fpr REGNUM from offset OFFSET relative
7725 to register BASE. Return generated insn. */
7728 restore_fpr (rtx base
, int offset
, int regnum
)
7731 addr
= gen_rtx_MEM (DFmode
, plus_constant (base
, offset
));
7732 set_mem_alias_set (addr
, get_frame_alias_set ());
7734 return emit_move_insn (gen_rtx_REG (DFmode
, regnum
), addr
);
7737 /* Return true if REGNO is a global register, but not one
7738 of the special ones that need to be saved/restored in anyway. */
7741 global_not_special_regno_p (int regno
)
7743 return (global_regs
[regno
]
7744 /* These registers are special and need to be
7745 restored in any case. */
7746 && !(regno
== STACK_POINTER_REGNUM
7747 || regno
== RETURN_REGNUM
7748 || regno
== BASE_REGNUM
7749 || (flag_pic
&& regno
== (int)PIC_OFFSET_TABLE_REGNUM
)));
7752 /* Generate insn to save registers FIRST to LAST into
7753 the register save area located at offset OFFSET
7754 relative to register BASE. */
7757 save_gprs (rtx base
, int offset
, int first
, int last
)
7759 rtx addr
, insn
, note
;
7762 addr
= plus_constant (base
, offset
);
7763 addr
= gen_rtx_MEM (Pmode
, addr
);
7765 set_mem_alias_set (addr
, get_frame_alias_set ());
7767 /* Special-case single register. */
7771 insn
= gen_movdi (addr
, gen_rtx_REG (Pmode
, first
));
7773 insn
= gen_movsi (addr
, gen_rtx_REG (Pmode
, first
));
7775 if (!global_not_special_regno_p (first
))
7776 RTX_FRAME_RELATED_P (insn
) = 1;
7781 insn
= gen_store_multiple (addr
,
7782 gen_rtx_REG (Pmode
, first
),
7783 GEN_INT (last
- first
+ 1));
7785 if (first
<= 6 && cfun
->stdarg
)
7786 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
7788 rtx mem
= XEXP (XVECEXP (PATTERN (insn
), 0, i
), 0);
7791 set_mem_alias_set (mem
, get_varargs_alias_set ());
7794 /* We need to set the FRAME_RELATED flag on all SETs
7795 inside the store-multiple pattern.
7797 However, we must not emit DWARF records for registers 2..5
7798 if they are stored for use by variable arguments ...
7800 ??? Unfortunately, it is not enough to simply not the
7801 FRAME_RELATED flags for those SETs, because the first SET
7802 of the PARALLEL is always treated as if it had the flag
7803 set, even if it does not. Therefore we emit a new pattern
7804 without those registers as REG_FRAME_RELATED_EXPR note. */
7806 if (first
>= 6 && !global_not_special_regno_p (first
))
7808 rtx pat
= PATTERN (insn
);
7810 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
7811 if (GET_CODE (XVECEXP (pat
, 0, i
)) == SET
7812 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (pat
,
7814 RTX_FRAME_RELATED_P (XVECEXP (pat
, 0, i
)) = 1;
7816 RTX_FRAME_RELATED_P (insn
) = 1;
7822 for (start
= first
>= 6 ? first
: 6; start
<= last
; start
++)
7823 if (!global_not_special_regno_p (start
))
7829 addr
= plus_constant (base
, offset
+ (start
- first
) * UNITS_PER_LONG
);
7830 note
= gen_store_multiple (gen_rtx_MEM (Pmode
, addr
),
7831 gen_rtx_REG (Pmode
, start
),
7832 GEN_INT (last
- start
+ 1));
7833 note
= PATTERN (note
);
7835 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, note
);
7837 for (i
= 0; i
< XVECLEN (note
, 0); i
++)
7838 if (GET_CODE (XVECEXP (note
, 0, i
)) == SET
7839 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (note
,
7841 RTX_FRAME_RELATED_P (XVECEXP (note
, 0, i
)) = 1;
7843 RTX_FRAME_RELATED_P (insn
) = 1;
7849 /* Generate insn to restore registers FIRST to LAST from
7850 the register save area located at offset OFFSET
7851 relative to register BASE. */
7854 restore_gprs (rtx base
, int offset
, int first
, int last
)
7858 addr
= plus_constant (base
, offset
);
7859 addr
= gen_rtx_MEM (Pmode
, addr
);
7860 set_mem_alias_set (addr
, get_frame_alias_set ());
7862 /* Special-case single register. */
7866 insn
= gen_movdi (gen_rtx_REG (Pmode
, first
), addr
);
7868 insn
= gen_movsi (gen_rtx_REG (Pmode
, first
), addr
);
7873 insn
= gen_load_multiple (gen_rtx_REG (Pmode
, first
),
7875 GEN_INT (last
- first
+ 1));
7879 /* Return insn sequence to load the GOT register. */
7881 static GTY(()) rtx got_symbol
;
7883 s390_load_got (void)
7887 /* We cannot use pic_offset_table_rtx here since we use this
7888 function also for non-pic if __tls_get_offset is called and in
7889 that case PIC_OFFSET_TABLE_REGNUM as well as pic_offset_table_rtx
7891 rtx got_rtx
= gen_rtx_REG (Pmode
, 12);
7895 got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
7896 SYMBOL_REF_FLAGS (got_symbol
) = SYMBOL_FLAG_LOCAL
;
7901 if (TARGET_CPU_ZARCH
)
7903 emit_move_insn (got_rtx
, got_symbol
);
7909 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, got_symbol
),
7910 UNSPEC_LTREL_OFFSET
);
7911 offset
= gen_rtx_CONST (Pmode
, offset
);
7912 offset
= force_const_mem (Pmode
, offset
);
7914 emit_move_insn (got_rtx
, offset
);
7916 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (offset
, 0)),
7918 offset
= gen_rtx_PLUS (Pmode
, got_rtx
, offset
);
7920 emit_move_insn (got_rtx
, offset
);
7923 insns
= get_insns ();
7928 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
7929 and the change to the stack pointer. */
7932 s390_emit_stack_tie (void)
7934 rtx mem
= gen_frame_mem (BLKmode
,
7935 gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
));
7937 emit_insn (gen_stack_tie (mem
));
7940 /* Expand the prologue into a bunch of separate insns. */
7943 s390_emit_prologue (void)
7951 /* Complete frame layout. */
7953 s390_update_frame_layout ();
7955 /* Annotate all constant pool references to let the scheduler know
7956 they implicitly use the base register. */
7958 push_topmost_sequence ();
7960 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
7963 annotate_constant_pool_refs (&PATTERN (insn
));
7964 df_insn_rescan (insn
);
7967 pop_topmost_sequence ();
7969 /* Choose best register to use for temp use within prologue.
7970 See below for why TPF must use the register 1. */
7972 if (!has_hard_reg_initial_val (Pmode
, RETURN_REGNUM
)
7973 && !current_function_is_leaf
7974 && !TARGET_TPF_PROFILING
)
7975 temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
7977 temp_reg
= gen_rtx_REG (Pmode
, 1);
7979 /* Save call saved gprs. */
7980 if (cfun_frame_layout
.first_save_gpr
!= -1)
7982 insn
= save_gprs (stack_pointer_rtx
,
7983 cfun_frame_layout
.gprs_offset
+
7984 UNITS_PER_LONG
* (cfun_frame_layout
.first_save_gpr
7985 - cfun_frame_layout
.first_save_gpr_slot
),
7986 cfun_frame_layout
.first_save_gpr
,
7987 cfun_frame_layout
.last_save_gpr
);
7991 /* Dummy insn to mark literal pool slot. */
7993 if (cfun
->machine
->base_reg
)
7994 emit_insn (gen_main_pool (cfun
->machine
->base_reg
));
7996 offset
= cfun_frame_layout
.f0_offset
;
7998 /* Save f0 and f2. */
7999 for (i
= 0; i
< 2; i
++)
8001 if (cfun_fpr_bit_p (i
))
8003 save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8006 else if (!TARGET_PACKED_STACK
)
8010 /* Save f4 and f6. */
8011 offset
= cfun_frame_layout
.f4_offset
;
8012 for (i
= 2; i
< 4; i
++)
8014 if (cfun_fpr_bit_p (i
))
8016 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8019 /* If f4 and f6 are call clobbered they are saved due to stdargs and
8020 therefore are not frame related. */
8021 if (!call_really_used_regs
[i
+ 16])
8022 RTX_FRAME_RELATED_P (insn
) = 1;
8024 else if (!TARGET_PACKED_STACK
)
8028 if (TARGET_PACKED_STACK
8029 && cfun_save_high_fprs_p
8030 && cfun_frame_layout
.f8_offset
+ cfun_frame_layout
.high_fprs
* 8 > 0)
8032 offset
= (cfun_frame_layout
.f8_offset
8033 + (cfun_frame_layout
.high_fprs
- 1) * 8);
8035 for (i
= 15; i
> 7 && offset
>= 0; i
--)
8036 if (cfun_fpr_bit_p (i
))
8038 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8040 RTX_FRAME_RELATED_P (insn
) = 1;
8043 if (offset
>= cfun_frame_layout
.f8_offset
)
8047 if (!TARGET_PACKED_STACK
)
8048 next_fpr
= cfun_save_high_fprs_p
? 31 : 0;
8050 if (flag_stack_usage_info
)
8051 current_function_static_stack_size
= cfun_frame_layout
.frame_size
;
8053 /* Decrement stack pointer. */
8055 if (cfun_frame_layout
.frame_size
> 0)
8057 rtx frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8060 if (s390_stack_size
)
8062 HOST_WIDE_INT stack_guard
;
8064 if (s390_stack_guard
)
8065 stack_guard
= s390_stack_guard
;
8068 /* If no value for stack guard is provided the smallest power of 2
8069 larger than the current frame size is chosen. */
8071 while (stack_guard
< cfun_frame_layout
.frame_size
)
8075 if (cfun_frame_layout
.frame_size
>= s390_stack_size
)
8077 warning (0, "frame size of function %qs is %wd"
8078 " bytes exceeding user provided stack limit of "
8080 "An unconditional trap is added.",
8081 current_function_name(), cfun_frame_layout
.frame_size
,
8083 emit_insn (gen_trap ());
8087 /* stack_guard has to be smaller than s390_stack_size.
8088 Otherwise we would emit an AND with zero which would
8089 not match the test under mask pattern. */
8090 if (stack_guard
>= s390_stack_size
)
8092 warning (0, "frame size of function %qs is %wd"
8093 " bytes which is more than half the stack size. "
8094 "The dynamic check would not be reliable. "
8095 "No check emitted for this function.",
8096 current_function_name(),
8097 cfun_frame_layout
.frame_size
);
8101 HOST_WIDE_INT stack_check_mask
= ((s390_stack_size
- 1)
8102 & ~(stack_guard
- 1));
8104 rtx t
= gen_rtx_AND (Pmode
, stack_pointer_rtx
,
8105 GEN_INT (stack_check_mask
));
8107 emit_insn (gen_ctrapdi4 (gen_rtx_EQ (VOIDmode
,
8109 t
, const0_rtx
, const0_rtx
));
8111 emit_insn (gen_ctrapsi4 (gen_rtx_EQ (VOIDmode
,
8113 t
, const0_rtx
, const0_rtx
));
8118 if (s390_warn_framesize
> 0
8119 && cfun_frame_layout
.frame_size
>= s390_warn_framesize
)
8120 warning (0, "frame size of %qs is %wd bytes",
8121 current_function_name (), cfun_frame_layout
.frame_size
);
8123 if (s390_warn_dynamicstack_p
&& cfun
->calls_alloca
)
8124 warning (0, "%qs uses dynamic stack allocation", current_function_name ());
8126 /* Save incoming stack pointer into temp reg. */
8127 if (TARGET_BACKCHAIN
|| next_fpr
)
8128 insn
= emit_insn (gen_move_insn (temp_reg
, stack_pointer_rtx
));
8130 /* Subtract frame size from stack pointer. */
8132 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8134 insn
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8135 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8137 insn
= emit_insn (insn
);
8141 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8142 frame_off
= force_const_mem (Pmode
, frame_off
);
8144 insn
= emit_insn (gen_add2_insn (stack_pointer_rtx
, frame_off
));
8145 annotate_constant_pool_refs (&PATTERN (insn
));
8148 RTX_FRAME_RELATED_P (insn
) = 1;
8149 real_frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8150 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8151 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8152 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8155 /* Set backchain. */
8157 if (TARGET_BACKCHAIN
)
8159 if (cfun_frame_layout
.backchain_offset
)
8160 addr
= gen_rtx_MEM (Pmode
,
8161 plus_constant (stack_pointer_rtx
,
8162 cfun_frame_layout
.backchain_offset
));
8164 addr
= gen_rtx_MEM (Pmode
, stack_pointer_rtx
);
8165 set_mem_alias_set (addr
, get_frame_alias_set ());
8166 insn
= emit_insn (gen_move_insn (addr
, temp_reg
));
8169 /* If we support non-call exceptions (e.g. for Java),
8170 we need to make sure the backchain pointer is set up
8171 before any possibly trapping memory access. */
8172 if (TARGET_BACKCHAIN
&& cfun
->can_throw_non_call_exceptions
)
8174 addr
= gen_rtx_MEM (BLKmode
, gen_rtx_SCRATCH (VOIDmode
));
8175 emit_clobber (addr
);
8179 /* Save fprs 8 - 15 (64 bit ABI). */
8181 if (cfun_save_high_fprs_p
&& next_fpr
)
8183 /* If the stack might be accessed through a different register
8184 we have to make sure that the stack pointer decrement is not
8185 moved below the use of the stack slots. */
8186 s390_emit_stack_tie ();
8188 insn
= emit_insn (gen_add2_insn (temp_reg
,
8189 GEN_INT (cfun_frame_layout
.f8_offset
)));
8193 for (i
= 24; i
<= next_fpr
; i
++)
8194 if (cfun_fpr_bit_p (i
- 16))
8196 rtx addr
= plus_constant (stack_pointer_rtx
,
8197 cfun_frame_layout
.frame_size
8198 + cfun_frame_layout
.f8_offset
8201 insn
= save_fpr (temp_reg
, offset
, i
);
8203 RTX_FRAME_RELATED_P (insn
) = 1;
8204 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8205 gen_rtx_SET (VOIDmode
,
8206 gen_rtx_MEM (DFmode
, addr
),
8207 gen_rtx_REG (DFmode
, i
)));
8211 /* Set frame pointer, if needed. */
8213 if (frame_pointer_needed
)
8215 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
8216 RTX_FRAME_RELATED_P (insn
) = 1;
8219 /* Set up got pointer, if needed. */
8221 if (flag_pic
&& df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
8223 rtx insns
= s390_load_got ();
8225 for (insn
= insns
; insn
; insn
= NEXT_INSN (insn
))
8226 annotate_constant_pool_refs (&PATTERN (insn
));
8231 if (TARGET_TPF_PROFILING
)
8233 /* Generate a BAS instruction to serve as a function
8234 entry intercept to facilitate the use of tracing
8235 algorithms located at the branch target. */
8236 emit_insn (gen_prologue_tpf ());
8238 /* Emit a blockage here so that all code
8239 lies between the profiling mechanisms. */
8240 emit_insn (gen_blockage ());
8244 /* Expand the epilogue into a bunch of separate insns. */
8247 s390_emit_epilogue (bool sibcall
)
8249 rtx frame_pointer
, return_reg
, cfa_restores
= NULL_RTX
;
8250 int area_bottom
, area_top
, offset
= 0;
8255 if (TARGET_TPF_PROFILING
)
8258 /* Generate a BAS instruction to serve as a function
8259 entry intercept to facilitate the use of tracing
8260 algorithms located at the branch target. */
8262 /* Emit a blockage here so that all code
8263 lies between the profiling mechanisms. */
8264 emit_insn (gen_blockage ());
8266 emit_insn (gen_epilogue_tpf ());
8269 /* Check whether to use frame or stack pointer for restore. */
8271 frame_pointer
= (frame_pointer_needed
8272 ? hard_frame_pointer_rtx
: stack_pointer_rtx
);
8274 s390_frame_area (&area_bottom
, &area_top
);
8276 /* Check whether we can access the register save area.
8277 If not, increment the frame pointer as required. */
8279 if (area_top
<= area_bottom
)
8281 /* Nothing to restore. */
8283 else if (DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_bottom
)
8284 && DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_top
- 1))
8286 /* Area is in range. */
8287 offset
= cfun_frame_layout
.frame_size
;
8291 rtx insn
, frame_off
, cfa
;
8293 offset
= area_bottom
< 0 ? -area_bottom
: 0;
8294 frame_off
= GEN_INT (cfun_frame_layout
.frame_size
- offset
);
8296 cfa
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8297 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8298 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8300 insn
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8301 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8302 insn
= emit_insn (insn
);
8306 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8307 frame_off
= force_const_mem (Pmode
, frame_off
);
8309 insn
= emit_insn (gen_add2_insn (frame_pointer
, frame_off
));
8310 annotate_constant_pool_refs (&PATTERN (insn
));
8312 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, cfa
);
8313 RTX_FRAME_RELATED_P (insn
) = 1;
8316 /* Restore call saved fprs. */
8320 if (cfun_save_high_fprs_p
)
8322 next_offset
= cfun_frame_layout
.f8_offset
;
8323 for (i
= 24; i
< 32; i
++)
8325 if (cfun_fpr_bit_p (i
- 16))
8327 restore_fpr (frame_pointer
,
8328 offset
+ next_offset
, i
);
8330 = alloc_reg_note (REG_CFA_RESTORE
,
8331 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8340 next_offset
= cfun_frame_layout
.f4_offset
;
8341 for (i
= 18; i
< 20; i
++)
8343 if (cfun_fpr_bit_p (i
- 16))
8345 restore_fpr (frame_pointer
,
8346 offset
+ next_offset
, i
);
8348 = alloc_reg_note (REG_CFA_RESTORE
,
8349 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8352 else if (!TARGET_PACKED_STACK
)
8358 /* Return register. */
8360 return_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8362 /* Restore call saved gprs. */
8364 if (cfun_frame_layout
.first_restore_gpr
!= -1)
8369 /* Check for global register and save them
8370 to stack location from where they get restored. */
8372 for (i
= cfun_frame_layout
.first_restore_gpr
;
8373 i
<= cfun_frame_layout
.last_restore_gpr
;
8376 if (global_not_special_regno_p (i
))
8378 addr
= plus_constant (frame_pointer
,
8379 offset
+ cfun_frame_layout
.gprs_offset
8380 + (i
- cfun_frame_layout
.first_save_gpr_slot
)
8382 addr
= gen_rtx_MEM (Pmode
, addr
);
8383 set_mem_alias_set (addr
, get_frame_alias_set ());
8384 emit_move_insn (addr
, gen_rtx_REG (Pmode
, i
));
8388 = alloc_reg_note (REG_CFA_RESTORE
,
8389 gen_rtx_REG (Pmode
, i
), cfa_restores
);
8394 /* Fetch return address from stack before load multiple,
8395 this will do good for scheduling. */
8397 if (cfun_frame_layout
.save_return_addr_p
8398 || (cfun_frame_layout
.first_restore_gpr
< BASE_REGNUM
8399 && cfun_frame_layout
.last_restore_gpr
> RETURN_REGNUM
))
8401 int return_regnum
= find_unused_clobbered_reg();
8404 return_reg
= gen_rtx_REG (Pmode
, return_regnum
);
8406 addr
= plus_constant (frame_pointer
,
8407 offset
+ cfun_frame_layout
.gprs_offset
8409 - cfun_frame_layout
.first_save_gpr_slot
)
8411 addr
= gen_rtx_MEM (Pmode
, addr
);
8412 set_mem_alias_set (addr
, get_frame_alias_set ());
8413 emit_move_insn (return_reg
, addr
);
8417 insn
= restore_gprs (frame_pointer
,
8418 offset
+ cfun_frame_layout
.gprs_offset
8419 + (cfun_frame_layout
.first_restore_gpr
8420 - cfun_frame_layout
.first_save_gpr_slot
)
8422 cfun_frame_layout
.first_restore_gpr
,
8423 cfun_frame_layout
.last_restore_gpr
);
8424 insn
= emit_insn (insn
);
8425 REG_NOTES (insn
) = cfa_restores
;
8426 add_reg_note (insn
, REG_CFA_DEF_CFA
,
8427 plus_constant (stack_pointer_rtx
, STACK_POINTER_OFFSET
));
8428 RTX_FRAME_RELATED_P (insn
) = 1;
8434 /* Return to caller. */
8436 p
= rtvec_alloc (2);
8438 RTVEC_ELT (p
, 0) = ret_rtx
;
8439 RTVEC_ELT (p
, 1) = gen_rtx_USE (VOIDmode
, return_reg
);
8440 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
8445 /* Return the size in bytes of a function argument of
8446 type TYPE and/or mode MODE. At least one of TYPE or
8447 MODE must be specified. */
8450 s390_function_arg_size (enum machine_mode mode
, const_tree type
)
8453 return int_size_in_bytes (type
);
8455 /* No type info available for some library calls ... */
8456 if (mode
!= BLKmode
)
8457 return GET_MODE_SIZE (mode
);
8459 /* If we have neither type nor mode, abort */
8463 /* Return true if a function argument of type TYPE and mode MODE
8464 is to be passed in a floating-point register, if available. */
8467 s390_function_arg_float (enum machine_mode mode
, const_tree type
)
8469 int size
= s390_function_arg_size (mode
, type
);
8473 /* Soft-float changes the ABI: no floating-point registers are used. */
8474 if (TARGET_SOFT_FLOAT
)
8477 /* No type info available for some library calls ... */
8479 return mode
== SFmode
|| mode
== DFmode
|| mode
== SDmode
|| mode
== DDmode
;
8481 /* The ABI says that record types with a single member are treated
8482 just like that member would be. */
8483 while (TREE_CODE (type
) == RECORD_TYPE
)
8485 tree field
, single
= NULL_TREE
;
8487 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
8489 if (TREE_CODE (field
) != FIELD_DECL
)
8492 if (single
== NULL_TREE
)
8493 single
= TREE_TYPE (field
);
8498 if (single
== NULL_TREE
)
8504 return TREE_CODE (type
) == REAL_TYPE
;
8507 /* Return true if a function argument of type TYPE and mode MODE
8508 is to be passed in an integer register, or a pair of integer
8509 registers, if available. */
8512 s390_function_arg_integer (enum machine_mode mode
, const_tree type
)
8514 int size
= s390_function_arg_size (mode
, type
);
8518 /* No type info available for some library calls ... */
8520 return GET_MODE_CLASS (mode
) == MODE_INT
8521 || (TARGET_SOFT_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
));
8523 /* We accept small integral (and similar) types. */
8524 if (INTEGRAL_TYPE_P (type
)
8525 || POINTER_TYPE_P (type
)
8526 || TREE_CODE (type
) == NULLPTR_TYPE
8527 || TREE_CODE (type
) == OFFSET_TYPE
8528 || (TARGET_SOFT_FLOAT
&& TREE_CODE (type
) == REAL_TYPE
))
8531 /* We also accept structs of size 1, 2, 4, 8 that are not
8532 passed in floating-point registers. */
8533 if (AGGREGATE_TYPE_P (type
)
8534 && exact_log2 (size
) >= 0
8535 && !s390_function_arg_float (mode
, type
))
8541 /* Return 1 if a function argument of type TYPE and mode MODE
8542 is to be passed by reference. The ABI specifies that only
8543 structures of size 1, 2, 4, or 8 bytes are passed by value,
8544 all other structures (and complex numbers) are passed by
8548 s390_pass_by_reference (cumulative_args_t ca ATTRIBUTE_UNUSED
,
8549 enum machine_mode mode
, const_tree type
,
8550 bool named ATTRIBUTE_UNUSED
)
8552 int size
= s390_function_arg_size (mode
, type
);
8558 if (AGGREGATE_TYPE_P (type
) && exact_log2 (size
) < 0)
8561 if (TREE_CODE (type
) == COMPLEX_TYPE
8562 || TREE_CODE (type
) == VECTOR_TYPE
)
8569 /* Update the data in CUM to advance over an argument of mode MODE and
8570 data type TYPE. (TYPE is null for libcalls where that information
8571 may not be available.). The boolean NAMED specifies whether the
8572 argument is a named argument (as opposed to an unnamed argument
8573 matching an ellipsis). */
8576 s390_function_arg_advance (cumulative_args_t cum_v
, enum machine_mode mode
,
8577 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8579 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8581 if (s390_function_arg_float (mode
, type
))
8585 else if (s390_function_arg_integer (mode
, type
))
8587 int size
= s390_function_arg_size (mode
, type
);
8588 cum
->gprs
+= ((size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
);
8594 /* Define where to put the arguments to a function.
8595 Value is zero to push the argument on the stack,
8596 or a hard register in which to store the argument.
8598 MODE is the argument's machine mode.
8599 TYPE is the data type of the argument (as a tree).
8600 This is null for libcalls where that information may
8602 CUM is a variable of type CUMULATIVE_ARGS which gives info about
8603 the preceding args and about the function being called.
8604 NAMED is nonzero if this argument is a named parameter
8605 (otherwise it is an extra parameter matching an ellipsis).
8607 On S/390, we use general purpose registers 2 through 6 to
8608 pass integer, pointer, and certain structure arguments, and
8609 floating point registers 0 and 2 (0, 2, 4, and 6 on 64-bit)
8610 to pass floating point arguments. All remaining arguments
8611 are pushed to the stack. */
8614 s390_function_arg (cumulative_args_t cum_v
, enum machine_mode mode
,
8615 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8617 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8619 if (s390_function_arg_float (mode
, type
))
8621 if (cum
->fprs
+ 1 > FP_ARG_NUM_REG
)
8624 return gen_rtx_REG (mode
, cum
->fprs
+ 16);
8626 else if (s390_function_arg_integer (mode
, type
))
8628 int size
= s390_function_arg_size (mode
, type
);
8629 int n_gprs
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8631 if (cum
->gprs
+ n_gprs
> GP_ARG_NUM_REG
)
8633 else if (n_gprs
== 1 || UNITS_PER_WORD
== UNITS_PER_LONG
)
8634 return gen_rtx_REG (mode
, cum
->gprs
+ 2);
8635 else if (n_gprs
== 2)
8637 rtvec p
= rtvec_alloc (2);
8640 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 2),
8643 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 3),
8646 return gen_rtx_PARALLEL (mode
, p
);
8650 /* After the real arguments, expand_call calls us once again
8651 with a void_type_node type. Whatever we return here is
8652 passed as operand 2 to the call expanders.
8654 We don't need this feature ... */
8655 else if (type
== void_type_node
)
8661 /* Return true if return values of type TYPE should be returned
8662 in a memory buffer whose address is passed by the caller as
8663 hidden first argument. */
8666 s390_return_in_memory (const_tree type
, const_tree fundecl ATTRIBUTE_UNUSED
)
8668 /* We accept small integral (and similar) types. */
8669 if (INTEGRAL_TYPE_P (type
)
8670 || POINTER_TYPE_P (type
)
8671 || TREE_CODE (type
) == OFFSET_TYPE
8672 || TREE_CODE (type
) == REAL_TYPE
)
8673 return int_size_in_bytes (type
) > 8;
8675 /* Aggregates and similar constructs are always returned
8677 if (AGGREGATE_TYPE_P (type
)
8678 || TREE_CODE (type
) == COMPLEX_TYPE
8679 || TREE_CODE (type
) == VECTOR_TYPE
)
8682 /* ??? We get called on all sorts of random stuff from
8683 aggregate_value_p. We can't abort, but it's not clear
8684 what's safe to return. Pretend it's a struct I guess. */
8688 /* Function arguments and return values are promoted to word size. */
8690 static enum machine_mode
8691 s390_promote_function_mode (const_tree type
, enum machine_mode mode
,
8693 const_tree fntype ATTRIBUTE_UNUSED
,
8694 int for_return ATTRIBUTE_UNUSED
)
8696 if (INTEGRAL_MODE_P (mode
)
8697 && GET_MODE_SIZE (mode
) < UNITS_PER_LONG
)
8699 if (type
!= NULL_TREE
&& POINTER_TYPE_P (type
))
8700 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
8707 /* Define where to return a (scalar) value of type RET_TYPE.
8708 If RET_TYPE is null, define where to return a (scalar)
8709 value of mode MODE from a libcall. */
8712 s390_function_and_libcall_value (enum machine_mode mode
,
8713 const_tree ret_type
,
8714 const_tree fntype_or_decl
,
8715 bool outgoing ATTRIBUTE_UNUSED
)
8717 /* For normal functions perform the promotion as
8718 promote_function_mode would do. */
8721 int unsignedp
= TYPE_UNSIGNED (ret_type
);
8722 mode
= promote_function_mode (ret_type
, mode
, &unsignedp
,
8726 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
|| SCALAR_FLOAT_MODE_P (mode
));
8727 gcc_assert (GET_MODE_SIZE (mode
) <= 8);
8729 if (TARGET_HARD_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
))
8730 return gen_rtx_REG (mode
, 16);
8731 else if (GET_MODE_SIZE (mode
) <= UNITS_PER_LONG
8732 || UNITS_PER_LONG
== UNITS_PER_WORD
)
8733 return gen_rtx_REG (mode
, 2);
8734 else if (GET_MODE_SIZE (mode
) == 2 * UNITS_PER_LONG
)
8736 /* This case is triggered when returning a 64 bit value with
8737 -m31 -mzarch. Although the value would fit into a single
8738 register it has to be forced into a 32 bit register pair in
8739 order to match the ABI. */
8740 rtvec p
= rtvec_alloc (2);
8743 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 2), const0_rtx
);
8745 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 3), GEN_INT (4));
8747 return gen_rtx_PARALLEL (mode
, p
);
8753 /* Define where to return a scalar return value of type RET_TYPE. */
8756 s390_function_value (const_tree ret_type
, const_tree fn_decl_or_type
,
8759 return s390_function_and_libcall_value (TYPE_MODE (ret_type
), ret_type
,
8760 fn_decl_or_type
, outgoing
);
8763 /* Define where to return a scalar libcall return value of mode
8767 s390_libcall_value (enum machine_mode mode
, const_rtx fun ATTRIBUTE_UNUSED
)
8769 return s390_function_and_libcall_value (mode
, NULL_TREE
,
8774 /* Create and return the va_list datatype.
8776 On S/390, va_list is an array type equivalent to
8778 typedef struct __va_list_tag
8782 void *__overflow_arg_area;
8783 void *__reg_save_area;
8786 where __gpr and __fpr hold the number of general purpose
8787 or floating point arguments used up to now, respectively,
8788 __overflow_arg_area points to the stack location of the
8789 next argument passed on the stack, and __reg_save_area
8790 always points to the start of the register area in the
8791 call frame of the current function. The function prologue
8792 saves all registers used for argument passing into this
8793 area if the function uses variable arguments. */
8796 s390_build_builtin_va_list (void)
8798 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
8800 record
= lang_hooks
.types
.make_type (RECORD_TYPE
);
8803 build_decl (BUILTINS_LOCATION
,
8804 TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
8806 f_gpr
= build_decl (BUILTINS_LOCATION
,
8807 FIELD_DECL
, get_identifier ("__gpr"),
8808 long_integer_type_node
);
8809 f_fpr
= build_decl (BUILTINS_LOCATION
,
8810 FIELD_DECL
, get_identifier ("__fpr"),
8811 long_integer_type_node
);
8812 f_ovf
= build_decl (BUILTINS_LOCATION
,
8813 FIELD_DECL
, get_identifier ("__overflow_arg_area"),
8815 f_sav
= build_decl (BUILTINS_LOCATION
,
8816 FIELD_DECL
, get_identifier ("__reg_save_area"),
8819 va_list_gpr_counter_field
= f_gpr
;
8820 va_list_fpr_counter_field
= f_fpr
;
8822 DECL_FIELD_CONTEXT (f_gpr
) = record
;
8823 DECL_FIELD_CONTEXT (f_fpr
) = record
;
8824 DECL_FIELD_CONTEXT (f_ovf
) = record
;
8825 DECL_FIELD_CONTEXT (f_sav
) = record
;
8827 TYPE_STUB_DECL (record
) = type_decl
;
8828 TYPE_NAME (record
) = type_decl
;
8829 TYPE_FIELDS (record
) = f_gpr
;
8830 DECL_CHAIN (f_gpr
) = f_fpr
;
8831 DECL_CHAIN (f_fpr
) = f_ovf
;
8832 DECL_CHAIN (f_ovf
) = f_sav
;
8834 layout_type (record
);
8836 /* The correct type is an array type of one element. */
8837 return build_array_type (record
, build_index_type (size_zero_node
));
8840 /* Implement va_start by filling the va_list structure VALIST.
8841 STDARG_P is always true, and ignored.
8842 NEXTARG points to the first anonymous stack argument.
8844 The following global variables are used to initialize
8845 the va_list structure:
8848 holds number of gprs and fprs used for named arguments.
8849 crtl->args.arg_offset_rtx:
8850 holds the offset of the first anonymous stack argument
8851 (relative to the virtual arg pointer). */
8854 s390_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
8856 HOST_WIDE_INT n_gpr
, n_fpr
;
8858 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8859 tree gpr
, fpr
, ovf
, sav
, t
;
8861 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8862 f_fpr
= DECL_CHAIN (f_gpr
);
8863 f_ovf
= DECL_CHAIN (f_fpr
);
8864 f_sav
= DECL_CHAIN (f_ovf
);
8866 valist
= build_simple_mem_ref (valist
);
8867 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8868 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8869 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8870 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8872 /* Count number of gp and fp argument registers used. */
8874 n_gpr
= crtl
->args
.info
.gprs
;
8875 n_fpr
= crtl
->args
.info
.fprs
;
8877 if (cfun
->va_list_gpr_size
)
8879 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gpr
), gpr
,
8880 build_int_cst (NULL_TREE
, n_gpr
));
8881 TREE_SIDE_EFFECTS (t
) = 1;
8882 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8885 if (cfun
->va_list_fpr_size
)
8887 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fpr
), fpr
,
8888 build_int_cst (NULL_TREE
, n_fpr
));
8889 TREE_SIDE_EFFECTS (t
) = 1;
8890 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8893 /* Find the overflow area. */
8894 if (n_gpr
+ cfun
->va_list_gpr_size
> GP_ARG_NUM_REG
8895 || n_fpr
+ cfun
->va_list_fpr_size
> FP_ARG_NUM_REG
)
8897 t
= make_tree (TREE_TYPE (ovf
), virtual_incoming_args_rtx
);
8899 off
= INTVAL (crtl
->args
.arg_offset_rtx
);
8900 off
= off
< 0 ? 0 : off
;
8901 if (TARGET_DEBUG_ARG
)
8902 fprintf (stderr
, "va_start: n_gpr = %d, n_fpr = %d off %d\n",
8903 (int)n_gpr
, (int)n_fpr
, off
);
8905 t
= fold_build_pointer_plus_hwi (t
, off
);
8907 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovf
), ovf
, t
);
8908 TREE_SIDE_EFFECTS (t
) = 1;
8909 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8912 /* Find the register save area. */
8913 if ((cfun
->va_list_gpr_size
&& n_gpr
< GP_ARG_NUM_REG
)
8914 || (cfun
->va_list_fpr_size
&& n_fpr
< FP_ARG_NUM_REG
))
8916 t
= make_tree (TREE_TYPE (sav
), return_address_pointer_rtx
);
8917 t
= fold_build_pointer_plus_hwi (t
, -RETURN_REGNUM
* UNITS_PER_LONG
);
8919 t
= build2 (MODIFY_EXPR
, TREE_TYPE (sav
), sav
, t
);
8920 TREE_SIDE_EFFECTS (t
) = 1;
8921 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8925 /* Implement va_arg by updating the va_list structure
8926 VALIST as required to retrieve an argument of type
8927 TYPE, and returning that argument.
8929 Generates code equivalent to:
8931 if (integral value) {
8932 if (size <= 4 && args.gpr < 5 ||
8933 size > 4 && args.gpr < 4 )
8934 ret = args.reg_save_area[args.gpr+8]
8936 ret = *args.overflow_arg_area++;
8937 } else if (float value) {
8939 ret = args.reg_save_area[args.fpr+64]
8941 ret = *args.overflow_arg_area++;
8942 } else if (aggregate value) {
8944 ret = *args.reg_save_area[args.gpr]
8946 ret = **args.overflow_arg_area++;
8950 s390_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
8951 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
8953 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8954 tree gpr
, fpr
, ovf
, sav
, reg
, t
, u
;
8955 int indirect_p
, size
, n_reg
, sav_ofs
, sav_scale
, max_reg
;
8956 tree lab_false
, lab_over
, addr
;
8958 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8959 f_fpr
= DECL_CHAIN (f_gpr
);
8960 f_ovf
= DECL_CHAIN (f_fpr
);
8961 f_sav
= DECL_CHAIN (f_ovf
);
8963 valist
= build_va_arg_indirect_ref (valist
);
8964 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8965 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8966 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8968 /* The tree for args* cannot be shared between gpr/fpr and ovf since
8969 both appear on a lhs. */
8970 valist
= unshare_expr (valist
);
8971 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8973 size
= int_size_in_bytes (type
);
8975 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
8977 if (TARGET_DEBUG_ARG
)
8979 fprintf (stderr
, "va_arg: aggregate type");
8983 /* Aggregates are passed by reference. */
8988 /* kernel stack layout on 31 bit: It is assumed here that no padding
8989 will be added by s390_frame_info because for va_args always an even
8990 number of gprs has to be saved r15-r2 = 14 regs. */
8991 sav_ofs
= 2 * UNITS_PER_LONG
;
8992 sav_scale
= UNITS_PER_LONG
;
8993 size
= UNITS_PER_LONG
;
8994 max_reg
= GP_ARG_NUM_REG
- n_reg
;
8996 else if (s390_function_arg_float (TYPE_MODE (type
), type
))
8998 if (TARGET_DEBUG_ARG
)
9000 fprintf (stderr
, "va_arg: float type");
9004 /* FP args go in FP registers, if present. */
9008 sav_ofs
= 16 * UNITS_PER_LONG
;
9010 max_reg
= FP_ARG_NUM_REG
- n_reg
;
9014 if (TARGET_DEBUG_ARG
)
9016 fprintf (stderr
, "va_arg: other type");
9020 /* Otherwise into GP registers. */
9023 n_reg
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
9025 /* kernel stack layout on 31 bit: It is assumed here that no padding
9026 will be added by s390_frame_info because for va_args always an even
9027 number of gprs has to be saved r15-r2 = 14 regs. */
9028 sav_ofs
= 2 * UNITS_PER_LONG
;
9030 if (size
< UNITS_PER_LONG
)
9031 sav_ofs
+= UNITS_PER_LONG
- size
;
9033 sav_scale
= UNITS_PER_LONG
;
9034 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9037 /* Pull the value out of the saved registers ... */
9039 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
9040 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
9041 addr
= create_tmp_var (ptr_type_node
, "addr");
9043 t
= fold_convert (TREE_TYPE (reg
), size_int (max_reg
));
9044 t
= build2 (GT_EXPR
, boolean_type_node
, reg
, t
);
9045 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
9046 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
9047 gimplify_and_add (t
, pre_p
);
9049 t
= fold_build_pointer_plus_hwi (sav
, sav_ofs
);
9050 u
= build2 (MULT_EXPR
, TREE_TYPE (reg
), reg
,
9051 fold_convert (TREE_TYPE (reg
), size_int (sav_scale
)));
9052 t
= fold_build_pointer_plus (t
, u
);
9054 gimplify_assign (addr
, t
, pre_p
);
9056 gimple_seq_add_stmt (pre_p
, gimple_build_goto (lab_over
));
9058 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_false
));
9061 /* ... Otherwise out of the overflow area. */
9064 if (size
< UNITS_PER_LONG
)
9065 t
= fold_build_pointer_plus_hwi (t
, UNITS_PER_LONG
- size
);
9067 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
9069 gimplify_assign (addr
, t
, pre_p
);
9071 t
= fold_build_pointer_plus_hwi (t
, size
);
9072 gimplify_assign (ovf
, t
, pre_p
);
9074 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_over
));
9077 /* Increment register save count. */
9079 u
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (reg
), reg
,
9080 fold_convert (TREE_TYPE (reg
), size_int (n_reg
)));
9081 gimplify_and_add (u
, pre_p
);
9085 t
= build_pointer_type_for_mode (build_pointer_type (type
),
9087 addr
= fold_convert (t
, addr
);
9088 addr
= build_va_arg_indirect_ref (addr
);
9092 t
= build_pointer_type_for_mode (type
, ptr_mode
, true);
9093 addr
= fold_convert (t
, addr
);
9096 return build_va_arg_indirect_ref (addr
);
9104 S390_BUILTIN_THREAD_POINTER
,
9105 S390_BUILTIN_SET_THREAD_POINTER
,
9110 static enum insn_code
const code_for_builtin_64
[S390_BUILTIN_max
] = {
9115 static enum insn_code
const code_for_builtin_31
[S390_BUILTIN_max
] = {
9121 s390_init_builtins (void)
9125 ftype
= build_function_type_list (ptr_type_node
, NULL_TREE
);
9126 add_builtin_function ("__builtin_thread_pointer", ftype
,
9127 S390_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
9130 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
9131 add_builtin_function ("__builtin_set_thread_pointer", ftype
,
9132 S390_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
9136 /* Expand an expression EXP that calls a built-in function,
9137 with result going to TARGET if that's convenient
9138 (and in mode MODE if that's convenient).
9139 SUBTARGET may be used as the target for computing one of EXP's operands.
9140 IGNORE is nonzero if the value is to be ignored. */
9143 s390_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
9144 enum machine_mode mode ATTRIBUTE_UNUSED
,
9145 int ignore ATTRIBUTE_UNUSED
)
9149 enum insn_code
const *code_for_builtin
=
9150 TARGET_64BIT
? code_for_builtin_64
: code_for_builtin_31
;
9152 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
9153 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
9154 enum insn_code icode
;
9155 rtx op
[MAX_ARGS
], pat
;
9159 call_expr_arg_iterator iter
;
9161 if (fcode
>= S390_BUILTIN_max
)
9162 internal_error ("bad builtin fcode");
9163 icode
= code_for_builtin
[fcode
];
9165 internal_error ("bad builtin fcode");
9167 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
9170 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
9172 const struct insn_operand_data
*insn_op
;
9174 if (arg
== error_mark_node
)
9176 if (arity
> MAX_ARGS
)
9179 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
9181 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
9183 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
9184 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
9190 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
9192 || GET_MODE (target
) != tmode
9193 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
9194 target
= gen_reg_rtx (tmode
);
9200 pat
= GEN_FCN (icode
) (target
);
9204 pat
= GEN_FCN (icode
) (target
, op
[0]);
9206 pat
= GEN_FCN (icode
) (op
[0]);
9209 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
9225 /* Output assembly code for the trampoline template to
9228 On S/390, we use gpr 1 internally in the trampoline code;
9229 gpr 0 is used to hold the static chain. */
9232 s390_asm_trampoline_template (FILE *file
)
9235 op
[0] = gen_rtx_REG (Pmode
, 0);
9236 op
[1] = gen_rtx_REG (Pmode
, 1);
9240 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9241 output_asm_insn ("lmg\t%0,%1,14(%1)", op
); /* 6 byte */
9242 output_asm_insn ("br\t%1", op
); /* 2 byte */
9243 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 10));
9247 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9248 output_asm_insn ("lm\t%0,%1,6(%1)", op
); /* 4 byte */
9249 output_asm_insn ("br\t%1", op
); /* 2 byte */
9250 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 8));
9254 /* Emit RTL insns to initialize the variable parts of a trampoline.
9255 FNADDR is an RTX for the address of the function's pure code.
9256 CXT is an RTX for the static chain value for the function. */
9259 s390_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9261 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
9264 emit_block_move (m_tramp
, assemble_trampoline_template (),
9265 GEN_INT (2 * UNITS_PER_LONG
), BLOCK_OP_NORMAL
);
9267 mem
= adjust_address (m_tramp
, Pmode
, 2 * UNITS_PER_LONG
);
9268 emit_move_insn (mem
, cxt
);
9269 mem
= adjust_address (m_tramp
, Pmode
, 3 * UNITS_PER_LONG
);
9270 emit_move_insn (mem
, fnaddr
);
9273 /* Output assembler code to FILE to increment profiler label # LABELNO
9274 for profiling a function entry. */
9277 s390_function_profiler (FILE *file
, int labelno
)
9282 ASM_GENERATE_INTERNAL_LABEL (label
, "LP", labelno
);
9284 fprintf (file
, "# function profiler \n");
9286 op
[0] = gen_rtx_REG (Pmode
, RETURN_REGNUM
);
9287 op
[1] = gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
);
9288 op
[1] = gen_rtx_MEM (Pmode
, plus_constant (op
[1], UNITS_PER_LONG
));
9290 op
[2] = gen_rtx_REG (Pmode
, 1);
9291 op
[3] = gen_rtx_SYMBOL_REF (Pmode
, label
);
9292 SYMBOL_REF_FLAGS (op
[3]) = SYMBOL_FLAG_LOCAL
;
9294 op
[4] = gen_rtx_SYMBOL_REF (Pmode
, "_mcount");
9297 op
[4] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[4]), UNSPEC_PLT
);
9298 op
[4] = gen_rtx_CONST (Pmode
, op
[4]);
9303 output_asm_insn ("stg\t%0,%1", op
);
9304 output_asm_insn ("larl\t%2,%3", op
);
9305 output_asm_insn ("brasl\t%0,%4", op
);
9306 output_asm_insn ("lg\t%0,%1", op
);
9310 op
[6] = gen_label_rtx ();
9312 output_asm_insn ("st\t%0,%1", op
);
9313 output_asm_insn ("bras\t%2,%l6", op
);
9314 output_asm_insn (".long\t%4", op
);
9315 output_asm_insn (".long\t%3", op
);
9316 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9317 output_asm_insn ("l\t%0,0(%2)", op
);
9318 output_asm_insn ("l\t%2,4(%2)", op
);
9319 output_asm_insn ("basr\t%0,%0", op
);
9320 output_asm_insn ("l\t%0,%1", op
);
9324 op
[5] = gen_label_rtx ();
9325 op
[6] = gen_label_rtx ();
9327 output_asm_insn ("st\t%0,%1", op
);
9328 output_asm_insn ("bras\t%2,%l6", op
);
9329 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[5]));
9330 output_asm_insn (".long\t%4-%l5", op
);
9331 output_asm_insn (".long\t%3-%l5", op
);
9332 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9333 output_asm_insn ("lr\t%0,%2", op
);
9334 output_asm_insn ("a\t%0,0(%2)", op
);
9335 output_asm_insn ("a\t%2,4(%2)", op
);
9336 output_asm_insn ("basr\t%0,%0", op
);
9337 output_asm_insn ("l\t%0,%1", op
);
9341 /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF
9342 into its SYMBOL_REF_FLAGS. */
9345 s390_encode_section_info (tree decl
, rtx rtl
, int first
)
9347 default_encode_section_info (decl
, rtl
, first
);
9349 if (TREE_CODE (decl
) == VAR_DECL
)
9351 /* If a variable has a forced alignment to < 2 bytes, mark it
9352 with SYMBOL_FLAG_ALIGN1 to prevent it from being used as LARL
9354 if (DECL_USER_ALIGN (decl
) && DECL_ALIGN (decl
) < 16)
9355 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_ALIGN1
;
9356 if (!DECL_SIZE (decl
)
9357 || !DECL_ALIGN (decl
)
9358 || !host_integerp (DECL_SIZE (decl
), 0)
9359 || (DECL_ALIGN (decl
) <= 64
9360 && DECL_ALIGN (decl
) != tree_low_cst (DECL_SIZE (decl
), 0)))
9361 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9364 /* Literal pool references don't have a decl so they are handled
9365 differently here. We rely on the information in the MEM_ALIGN
9366 entry to decide upon natural alignment. */
9368 && GET_CODE (XEXP (rtl
, 0)) == SYMBOL_REF
9369 && TREE_CONSTANT_POOL_ADDRESS_P (XEXP (rtl
, 0))
9370 && (MEM_ALIGN (rtl
) == 0
9371 || GET_MODE_BITSIZE (GET_MODE (rtl
)) == 0
9372 || MEM_ALIGN (rtl
) < GET_MODE_BITSIZE (GET_MODE (rtl
))))
9373 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9376 /* Output thunk to FILE that implements a C++ virtual function call (with
9377 multiple inheritance) to FUNCTION. The thunk adjusts the this pointer
9378 by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment
9379 stored at VCALL_OFFSET in the vtable whose address is located at offset 0
9380 relative to the resulting this pointer. */
9383 s390_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
9384 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
9390 /* Make sure unwind info is emitted for the thunk if needed. */
9391 final_start_function (emit_barrier (), file
, 1);
9393 /* Operand 0 is the target function. */
9394 op
[0] = XEXP (DECL_RTL (function
), 0);
9395 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (op
[0]))
9398 op
[0] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[0]),
9399 TARGET_64BIT
? UNSPEC_PLT
: UNSPEC_GOT
);
9400 op
[0] = gen_rtx_CONST (Pmode
, op
[0]);
9403 /* Operand 1 is the 'this' pointer. */
9404 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
9405 op
[1] = gen_rtx_REG (Pmode
, 3);
9407 op
[1] = gen_rtx_REG (Pmode
, 2);
9409 /* Operand 2 is the delta. */
9410 op
[2] = GEN_INT (delta
);
9412 /* Operand 3 is the vcall_offset. */
9413 op
[3] = GEN_INT (vcall_offset
);
9415 /* Operand 4 is the temporary register. */
9416 op
[4] = gen_rtx_REG (Pmode
, 1);
9418 /* Operands 5 to 8 can be used as labels. */
9424 /* Operand 9 can be used for temporary register. */
9427 /* Generate code. */
9430 /* Setup literal pool pointer if required. */
9431 if ((!DISP_IN_RANGE (delta
)
9432 && !CONST_OK_FOR_K (delta
)
9433 && !CONST_OK_FOR_Os (delta
))
9434 || (!DISP_IN_RANGE (vcall_offset
)
9435 && !CONST_OK_FOR_K (vcall_offset
)
9436 && !CONST_OK_FOR_Os (vcall_offset
)))
9438 op
[5] = gen_label_rtx ();
9439 output_asm_insn ("larl\t%4,%5", op
);
9442 /* Add DELTA to this pointer. */
9445 if (CONST_OK_FOR_J (delta
))
9446 output_asm_insn ("la\t%1,%2(%1)", op
);
9447 else if (DISP_IN_RANGE (delta
))
9448 output_asm_insn ("lay\t%1,%2(%1)", op
);
9449 else if (CONST_OK_FOR_K (delta
))
9450 output_asm_insn ("aghi\t%1,%2", op
);
9451 else if (CONST_OK_FOR_Os (delta
))
9452 output_asm_insn ("agfi\t%1,%2", op
);
9455 op
[6] = gen_label_rtx ();
9456 output_asm_insn ("agf\t%1,%6-%5(%4)", op
);
9460 /* Perform vcall adjustment. */
9463 if (DISP_IN_RANGE (vcall_offset
))
9465 output_asm_insn ("lg\t%4,0(%1)", op
);
9466 output_asm_insn ("ag\t%1,%3(%4)", op
);
9468 else if (CONST_OK_FOR_K (vcall_offset
))
9470 output_asm_insn ("lghi\t%4,%3", op
);
9471 output_asm_insn ("ag\t%4,0(%1)", op
);
9472 output_asm_insn ("ag\t%1,0(%4)", op
);
9474 else if (CONST_OK_FOR_Os (vcall_offset
))
9476 output_asm_insn ("lgfi\t%4,%3", op
);
9477 output_asm_insn ("ag\t%4,0(%1)", op
);
9478 output_asm_insn ("ag\t%1,0(%4)", op
);
9482 op
[7] = gen_label_rtx ();
9483 output_asm_insn ("llgf\t%4,%7-%5(%4)", op
);
9484 output_asm_insn ("ag\t%4,0(%1)", op
);
9485 output_asm_insn ("ag\t%1,0(%4)", op
);
9489 /* Jump to target. */
9490 output_asm_insn ("jg\t%0", op
);
9492 /* Output literal pool if required. */
9495 output_asm_insn (".align\t4", op
);
9496 targetm
.asm_out
.internal_label (file
, "L",
9497 CODE_LABEL_NUMBER (op
[5]));
9501 targetm
.asm_out
.internal_label (file
, "L",
9502 CODE_LABEL_NUMBER (op
[6]));
9503 output_asm_insn (".long\t%2", op
);
9507 targetm
.asm_out
.internal_label (file
, "L",
9508 CODE_LABEL_NUMBER (op
[7]));
9509 output_asm_insn (".long\t%3", op
);
9514 /* Setup base pointer if required. */
9516 || (!DISP_IN_RANGE (delta
)
9517 && !CONST_OK_FOR_K (delta
)
9518 && !CONST_OK_FOR_Os (delta
))
9519 || (!DISP_IN_RANGE (delta
)
9520 && !CONST_OK_FOR_K (vcall_offset
)
9521 && !CONST_OK_FOR_Os (vcall_offset
)))
9523 op
[5] = gen_label_rtx ();
9524 output_asm_insn ("basr\t%4,0", op
);
9525 targetm
.asm_out
.internal_label (file
, "L",
9526 CODE_LABEL_NUMBER (op
[5]));
9529 /* Add DELTA to this pointer. */
9532 if (CONST_OK_FOR_J (delta
))
9533 output_asm_insn ("la\t%1,%2(%1)", op
);
9534 else if (DISP_IN_RANGE (delta
))
9535 output_asm_insn ("lay\t%1,%2(%1)", op
);
9536 else if (CONST_OK_FOR_K (delta
))
9537 output_asm_insn ("ahi\t%1,%2", op
);
9538 else if (CONST_OK_FOR_Os (delta
))
9539 output_asm_insn ("afi\t%1,%2", op
);
9542 op
[6] = gen_label_rtx ();
9543 output_asm_insn ("a\t%1,%6-%5(%4)", op
);
9547 /* Perform vcall adjustment. */
9550 if (CONST_OK_FOR_J (vcall_offset
))
9552 output_asm_insn ("l\t%4,0(%1)", op
);
9553 output_asm_insn ("a\t%1,%3(%4)", op
);
9555 else if (DISP_IN_RANGE (vcall_offset
))
9557 output_asm_insn ("l\t%4,0(%1)", op
);
9558 output_asm_insn ("ay\t%1,%3(%4)", op
);
9560 else if (CONST_OK_FOR_K (vcall_offset
))
9562 output_asm_insn ("lhi\t%4,%3", op
);
9563 output_asm_insn ("a\t%4,0(%1)", op
);
9564 output_asm_insn ("a\t%1,0(%4)", op
);
9566 else if (CONST_OK_FOR_Os (vcall_offset
))
9568 output_asm_insn ("iilf\t%4,%3", op
);
9569 output_asm_insn ("a\t%4,0(%1)", op
);
9570 output_asm_insn ("a\t%1,0(%4)", op
);
9574 op
[7] = gen_label_rtx ();
9575 output_asm_insn ("l\t%4,%7-%5(%4)", op
);
9576 output_asm_insn ("a\t%4,0(%1)", op
);
9577 output_asm_insn ("a\t%1,0(%4)", op
);
9580 /* We had to clobber the base pointer register.
9581 Re-setup the base pointer (with a different base). */
9582 op
[5] = gen_label_rtx ();
9583 output_asm_insn ("basr\t%4,0", op
);
9584 targetm
.asm_out
.internal_label (file
, "L",
9585 CODE_LABEL_NUMBER (op
[5]));
9588 /* Jump to target. */
9589 op
[8] = gen_label_rtx ();
9592 output_asm_insn ("l\t%4,%8-%5(%4)", op
);
9594 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9595 /* We cannot call through .plt, since .plt requires %r12 loaded. */
9596 else if (flag_pic
== 1)
9598 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9599 output_asm_insn ("l\t%4,%0(%4)", op
);
9601 else if (flag_pic
== 2)
9603 op
[9] = gen_rtx_REG (Pmode
, 0);
9604 output_asm_insn ("l\t%9,%8-4-%5(%4)", op
);
9605 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9606 output_asm_insn ("ar\t%4,%9", op
);
9607 output_asm_insn ("l\t%4,0(%4)", op
);
9610 output_asm_insn ("br\t%4", op
);
9612 /* Output literal pool. */
9613 output_asm_insn (".align\t4", op
);
9615 if (nonlocal
&& flag_pic
== 2)
9616 output_asm_insn (".long\t%0", op
);
9619 op
[0] = gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
9620 SYMBOL_REF_FLAGS (op
[0]) = SYMBOL_FLAG_LOCAL
;
9623 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[8]));
9625 output_asm_insn (".long\t%0", op
);
9627 output_asm_insn (".long\t%0-%5", op
);
9631 targetm
.asm_out
.internal_label (file
, "L",
9632 CODE_LABEL_NUMBER (op
[6]));
9633 output_asm_insn (".long\t%2", op
);
9637 targetm
.asm_out
.internal_label (file
, "L",
9638 CODE_LABEL_NUMBER (op
[7]));
9639 output_asm_insn (".long\t%3", op
);
9642 final_end_function ();
9646 s390_valid_pointer_mode (enum machine_mode mode
)
9648 return (mode
== SImode
|| (TARGET_64BIT
&& mode
== DImode
));
9651 /* Checks whether the given CALL_EXPR would use a caller
9652 saved register. This is used to decide whether sibling call
9653 optimization could be performed on the respective function
9657 s390_call_saved_register_used (tree call_expr
)
9659 CUMULATIVE_ARGS cum_v
;
9660 cumulative_args_t cum
;
9662 enum machine_mode mode
;
9667 INIT_CUMULATIVE_ARGS (cum_v
, NULL
, NULL
, 0, 0);
9668 cum
= pack_cumulative_args (&cum_v
);
9670 for (i
= 0; i
< call_expr_nargs (call_expr
); i
++)
9672 parameter
= CALL_EXPR_ARG (call_expr
, i
);
9673 gcc_assert (parameter
);
9675 /* For an undeclared variable passed as parameter we will get
9676 an ERROR_MARK node here. */
9677 if (TREE_CODE (parameter
) == ERROR_MARK
)
9680 type
= TREE_TYPE (parameter
);
9683 mode
= TYPE_MODE (type
);
9686 if (pass_by_reference (&cum_v
, mode
, type
, true))
9689 type
= build_pointer_type (type
);
9692 parm_rtx
= s390_function_arg (cum
, mode
, type
, 0);
9694 s390_function_arg_advance (cum
, mode
, type
, 0);
9699 if (REG_P (parm_rtx
))
9702 reg
< HARD_REGNO_NREGS (REGNO (parm_rtx
), GET_MODE (parm_rtx
));
9704 if (!call_used_regs
[reg
+ REGNO (parm_rtx
)])
9708 if (GET_CODE (parm_rtx
) == PARALLEL
)
9712 for (i
= 0; i
< XVECLEN (parm_rtx
, 0); i
++)
9714 rtx r
= XEXP (XVECEXP (parm_rtx
, 0, i
), 0);
9716 gcc_assert (REG_P (r
));
9719 reg
< HARD_REGNO_NREGS (REGNO (r
), GET_MODE (r
));
9721 if (!call_used_regs
[reg
+ REGNO (r
)])
9730 /* Return true if the given call expression can be
9731 turned into a sibling call.
9732 DECL holds the declaration of the function to be called whereas
9733 EXP is the call expression itself. */
9736 s390_function_ok_for_sibcall (tree decl
, tree exp
)
9738 /* The TPF epilogue uses register 1. */
9739 if (TARGET_TPF_PROFILING
)
9742 /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
9743 which would have to be restored before the sibcall. */
9744 if (!TARGET_64BIT
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
9747 /* Register 6 on s390 is available as an argument register but unfortunately
9748 "caller saved". This makes functions needing this register for arguments
9749 not suitable for sibcalls. */
9750 return !s390_call_saved_register_used (exp
);
9753 /* Return the fixed registers used for condition codes. */
9756 s390_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
9759 *p2
= INVALID_REGNUM
;
9764 /* This function is used by the call expanders of the machine description.
9765 It emits the call insn itself together with the necessary operations
9766 to adjust the target address and returns the emitted insn.
9767 ADDR_LOCATION is the target address rtx
9768 TLS_CALL the location of the thread-local symbol
9769 RESULT_REG the register where the result of the call should be stored
9770 RETADDR_REG the register where the return address should be stored
9771 If this parameter is NULL_RTX the call is considered
9772 to be a sibling call. */
9775 s390_emit_call (rtx addr_location
, rtx tls_call
, rtx result_reg
,
9778 bool plt_call
= false;
9784 /* Direct function calls need special treatment. */
9785 if (GET_CODE (addr_location
) == SYMBOL_REF
)
9787 /* When calling a global routine in PIC mode, we must
9788 replace the symbol itself with the PLT stub. */
9789 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (addr_location
))
9791 if (retaddr_reg
!= NULL_RTX
)
9793 addr_location
= gen_rtx_UNSPEC (Pmode
,
9794 gen_rtvec (1, addr_location
),
9796 addr_location
= gen_rtx_CONST (Pmode
, addr_location
);
9800 /* For -fpic code the PLT entries might use r12 which is
9801 call-saved. Therefore we cannot do a sibcall when
9802 calling directly using a symbol ref. When reaching
9803 this point we decided (in s390_function_ok_for_sibcall)
9804 to do a sibcall for a function pointer but one of the
9805 optimizers was able to get rid of the function pointer
9806 by propagating the symbol ref into the call. This
9807 optimization is illegal for S/390 so we turn the direct
9808 call into a indirect call again. */
9809 addr_location
= force_reg (Pmode
, addr_location
);
9812 /* Unless we can use the bras(l) insn, force the
9813 routine address into a register. */
9814 if (!TARGET_SMALL_EXEC
&& !TARGET_CPU_ZARCH
)
9817 addr_location
= legitimize_pic_address (addr_location
, 0);
9819 addr_location
= force_reg (Pmode
, addr_location
);
9823 /* If it is already an indirect call or the code above moved the
9824 SYMBOL_REF to somewhere else make sure the address can be found in
9826 if (retaddr_reg
== NULL_RTX
9827 && GET_CODE (addr_location
) != SYMBOL_REF
9830 emit_move_insn (gen_rtx_REG (Pmode
, SIBCALL_REGNUM
), addr_location
);
9831 addr_location
= gen_rtx_REG (Pmode
, SIBCALL_REGNUM
);
9834 addr_location
= gen_rtx_MEM (QImode
, addr_location
);
9835 call
= gen_rtx_CALL (VOIDmode
, addr_location
, const0_rtx
);
9837 if (result_reg
!= NULL_RTX
)
9838 call
= gen_rtx_SET (VOIDmode
, result_reg
, call
);
9840 if (retaddr_reg
!= NULL_RTX
)
9842 clobber
= gen_rtx_CLOBBER (VOIDmode
, retaddr_reg
);
9844 if (tls_call
!= NULL_RTX
)
9845 vec
= gen_rtvec (3, call
, clobber
,
9846 gen_rtx_USE (VOIDmode
, tls_call
));
9848 vec
= gen_rtvec (2, call
, clobber
);
9850 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
9853 insn
= emit_call_insn (call
);
9855 /* 31-bit PLT stubs and tls calls use the GOT register implicitly. */
9856 if ((!TARGET_64BIT
&& plt_call
) || tls_call
!= NULL_RTX
)
9858 /* s390_function_ok_for_sibcall should
9859 have denied sibcalls in this case. */
9860 gcc_assert (retaddr_reg
!= NULL_RTX
);
9861 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), gen_rtx_REG (Pmode
, 12));
9866 /* Implement TARGET_CONDITIONAL_REGISTER_USAGE. */
9869 s390_conditional_register_usage (void)
9875 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9876 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9878 if (TARGET_CPU_ZARCH
)
9880 fixed_regs
[BASE_REGNUM
] = 0;
9881 call_used_regs
[BASE_REGNUM
] = 0;
9882 fixed_regs
[RETURN_REGNUM
] = 0;
9883 call_used_regs
[RETURN_REGNUM
] = 0;
9887 for (i
= 24; i
< 32; i
++)
9888 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9892 for (i
= 18; i
< 20; i
++)
9893 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9896 if (TARGET_SOFT_FLOAT
)
9898 for (i
= 16; i
< 32; i
++)
9899 call_used_regs
[i
] = fixed_regs
[i
] = 1;
9903 /* Corresponding function to eh_return expander. */
9905 static GTY(()) rtx s390_tpf_eh_return_symbol
;
9907 s390_emit_tpf_eh_return (rtx target
)
9911 if (!s390_tpf_eh_return_symbol
)
9912 s390_tpf_eh_return_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tpf_eh_return");
9914 reg
= gen_rtx_REG (Pmode
, 2);
9916 emit_move_insn (reg
, target
);
9917 insn
= s390_emit_call (s390_tpf_eh_return_symbol
, NULL_RTX
, reg
,
9918 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
9919 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), reg
);
9921 emit_move_insn (EH_RETURN_HANDLER_RTX
, reg
);
9924 /* Rework the prologue/epilogue to avoid saving/restoring
9925 registers unnecessarily. */
9928 s390_optimize_prologue (void)
9930 rtx insn
, new_insn
, next_insn
;
9932 /* Do a final recompute of the frame-related data. */
9934 s390_update_frame_layout ();
9936 /* If all special registers are in fact used, there's nothing we
9937 can do, so no point in walking the insn list. */
9939 if (cfun_frame_layout
.first_save_gpr
<= BASE_REGNUM
9940 && cfun_frame_layout
.last_save_gpr
>= BASE_REGNUM
9941 && (TARGET_CPU_ZARCH
9942 || (cfun_frame_layout
.first_save_gpr
<= RETURN_REGNUM
9943 && cfun_frame_layout
.last_save_gpr
>= RETURN_REGNUM
)))
9946 /* Search for prologue/epilogue insns and replace them. */
9948 for (insn
= get_insns (); insn
; insn
= next_insn
)
9950 int first
, last
, off
;
9951 rtx set
, base
, offset
;
9953 next_insn
= NEXT_INSN (insn
);
9955 if (GET_CODE (insn
) != INSN
)
9958 if (GET_CODE (PATTERN (insn
)) == PARALLEL
9959 && store_multiple_operation (PATTERN (insn
), VOIDmode
))
9961 set
= XVECEXP (PATTERN (insn
), 0, 0);
9962 first
= REGNO (SET_SRC (set
));
9963 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
9964 offset
= const0_rtx
;
9965 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
9966 off
= INTVAL (offset
);
9968 if (GET_CODE (base
) != REG
|| off
< 0)
9970 if (cfun_frame_layout
.first_save_gpr
!= -1
9971 && (cfun_frame_layout
.first_save_gpr
< first
9972 || cfun_frame_layout
.last_save_gpr
> last
))
9974 if (REGNO (base
) != STACK_POINTER_REGNUM
9975 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
9977 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
9980 if (cfun_frame_layout
.first_save_gpr
!= -1)
9982 new_insn
= save_gprs (base
,
9983 off
+ (cfun_frame_layout
.first_save_gpr
9984 - first
) * UNITS_PER_LONG
,
9985 cfun_frame_layout
.first_save_gpr
,
9986 cfun_frame_layout
.last_save_gpr
);
9987 new_insn
= emit_insn_before (new_insn
, insn
);
9988 INSN_ADDRESSES_NEW (new_insn
, -1);
9995 if (cfun_frame_layout
.first_save_gpr
== -1
9996 && GET_CODE (PATTERN (insn
)) == SET
9997 && GET_CODE (SET_SRC (PATTERN (insn
))) == REG
9998 && (REGNO (SET_SRC (PATTERN (insn
))) == BASE_REGNUM
9999 || (!TARGET_CPU_ZARCH
10000 && REGNO (SET_SRC (PATTERN (insn
))) == RETURN_REGNUM
))
10001 && GET_CODE (SET_DEST (PATTERN (insn
))) == MEM
)
10003 set
= PATTERN (insn
);
10004 first
= REGNO (SET_SRC (set
));
10005 offset
= const0_rtx
;
10006 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10007 off
= INTVAL (offset
);
10009 if (GET_CODE (base
) != REG
|| off
< 0)
10011 if (REGNO (base
) != STACK_POINTER_REGNUM
10012 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10015 remove_insn (insn
);
10019 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10020 && load_multiple_operation (PATTERN (insn
), VOIDmode
))
10022 set
= XVECEXP (PATTERN (insn
), 0, 0);
10023 first
= REGNO (SET_DEST (set
));
10024 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10025 offset
= const0_rtx
;
10026 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10027 off
= INTVAL (offset
);
10029 if (GET_CODE (base
) != REG
|| off
< 0)
10031 if (cfun_frame_layout
.first_restore_gpr
!= -1
10032 && (cfun_frame_layout
.first_restore_gpr
< first
10033 || cfun_frame_layout
.last_restore_gpr
> last
))
10035 if (REGNO (base
) != STACK_POINTER_REGNUM
10036 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10038 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10041 if (cfun_frame_layout
.first_restore_gpr
!= -1)
10043 new_insn
= restore_gprs (base
,
10044 off
+ (cfun_frame_layout
.first_restore_gpr
10045 - first
) * UNITS_PER_LONG
,
10046 cfun_frame_layout
.first_restore_gpr
,
10047 cfun_frame_layout
.last_restore_gpr
);
10048 new_insn
= emit_insn_before (new_insn
, insn
);
10049 INSN_ADDRESSES_NEW (new_insn
, -1);
10052 remove_insn (insn
);
10056 if (cfun_frame_layout
.first_restore_gpr
== -1
10057 && GET_CODE (PATTERN (insn
)) == SET
10058 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
10059 && (REGNO (SET_DEST (PATTERN (insn
))) == BASE_REGNUM
10060 || (!TARGET_CPU_ZARCH
10061 && REGNO (SET_DEST (PATTERN (insn
))) == RETURN_REGNUM
))
10062 && GET_CODE (SET_SRC (PATTERN (insn
))) == MEM
)
10064 set
= PATTERN (insn
);
10065 first
= REGNO (SET_DEST (set
));
10066 offset
= const0_rtx
;
10067 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10068 off
= INTVAL (offset
);
10070 if (GET_CODE (base
) != REG
|| off
< 0)
10072 if (REGNO (base
) != STACK_POINTER_REGNUM
10073 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10076 remove_insn (insn
);
10082 /* On z10 and later the dynamic branch prediction must see the
10083 backward jump within a certain windows. If not it falls back to
10084 the static prediction. This function rearranges the loop backward
10085 branch in a way which makes the static prediction always correct.
10086 The function returns true if it added an instruction. */
10088 s390_fix_long_loop_prediction (rtx insn
)
10090 rtx set
= single_set (insn
);
10091 rtx code_label
, label_ref
, new_label
;
10097 /* This will exclude branch on count and branch on index patterns
10098 since these are correctly statically predicted. */
10100 || SET_DEST (set
) != pc_rtx
10101 || GET_CODE (SET_SRC(set
)) != IF_THEN_ELSE
)
10104 label_ref
= (GET_CODE (XEXP (SET_SRC (set
), 1)) == LABEL_REF
?
10105 XEXP (SET_SRC (set
), 1) : XEXP (SET_SRC (set
), 2));
10107 gcc_assert (GET_CODE (label_ref
) == LABEL_REF
);
10109 code_label
= XEXP (label_ref
, 0);
10111 if (INSN_ADDRESSES (INSN_UID (code_label
)) == -1
10112 || INSN_ADDRESSES (INSN_UID (insn
)) == -1
10113 || (INSN_ADDRESSES (INSN_UID (insn
))
10114 - INSN_ADDRESSES (INSN_UID (code_label
)) < PREDICT_DISTANCE
))
10117 for (distance
= 0, cur_insn
= PREV_INSN (insn
);
10118 distance
< PREDICT_DISTANCE
- 6;
10119 distance
+= get_attr_length (cur_insn
), cur_insn
= PREV_INSN (cur_insn
))
10120 if (!cur_insn
|| JUMP_P (cur_insn
) || LABEL_P (cur_insn
))
10123 new_label
= gen_label_rtx ();
10124 uncond_jump
= emit_jump_insn_after (
10125 gen_rtx_SET (VOIDmode
, pc_rtx
,
10126 gen_rtx_LABEL_REF (VOIDmode
, code_label
)),
10128 emit_label_after (new_label
, uncond_jump
);
10130 tmp
= XEXP (SET_SRC (set
), 1);
10131 XEXP (SET_SRC (set
), 1) = XEXP (SET_SRC (set
), 2);
10132 XEXP (SET_SRC (set
), 2) = tmp
;
10133 INSN_CODE (insn
) = -1;
10135 XEXP (label_ref
, 0) = new_label
;
10136 JUMP_LABEL (insn
) = new_label
;
10137 JUMP_LABEL (uncond_jump
) = code_label
;
10142 /* Returns 1 if INSN reads the value of REG for purposes not related
10143 to addressing of memory, and 0 otherwise. */
10145 s390_non_addr_reg_read_p (rtx reg
, rtx insn
)
10147 return reg_referenced_p (reg
, PATTERN (insn
))
10148 && !reg_used_in_mem_p (REGNO (reg
), PATTERN (insn
));
10151 /* Starting from INSN find_cond_jump looks downwards in the insn
10152 stream for a single jump insn which is the last user of the
10153 condition code set in INSN. */
10155 find_cond_jump (rtx insn
)
10157 for (; insn
; insn
= NEXT_INSN (insn
))
10161 if (LABEL_P (insn
))
10164 if (!JUMP_P (insn
))
10166 if (reg_mentioned_p (gen_rtx_REG (CCmode
, CC_REGNUM
), insn
))
10171 /* This will be triggered by a return. */
10172 if (GET_CODE (PATTERN (insn
)) != SET
)
10175 gcc_assert (SET_DEST (PATTERN (insn
)) == pc_rtx
);
10176 ite
= SET_SRC (PATTERN (insn
));
10178 if (GET_CODE (ite
) != IF_THEN_ELSE
)
10181 cc
= XEXP (XEXP (ite
, 0), 0);
10182 if (!REG_P (cc
) || !CC_REGNO_P (REGNO (cc
)))
10185 if (find_reg_note (insn
, REG_DEAD
, cc
))
10193 /* Swap the condition in COND and the operands in OP0 and OP1 so that
10194 the semantics does not change. If NULL_RTX is passed as COND the
10195 function tries to find the conditional jump starting with INSN. */
10197 s390_swap_cmp (rtx cond
, rtx
*op0
, rtx
*op1
, rtx insn
)
10201 if (cond
== NULL_RTX
)
10203 rtx jump
= find_cond_jump (NEXT_INSN (insn
));
10204 jump
= jump
? single_set (jump
) : NULL_RTX
;
10206 if (jump
== NULL_RTX
)
10209 cond
= XEXP (XEXP (jump
, 1), 0);
10214 PUT_CODE (cond
, swap_condition (GET_CODE (cond
)));
10217 /* On z10, instructions of the compare-and-branch family have the
10218 property to access the register occurring as second operand with
10219 its bits complemented. If such a compare is grouped with a second
10220 instruction that accesses the same register non-complemented, and
10221 if that register's value is delivered via a bypass, then the
10222 pipeline recycles, thereby causing significant performance decline.
10223 This function locates such situations and exchanges the two
10224 operands of the compare. The function return true whenever it
10227 s390_z10_optimize_cmp (rtx insn
)
10229 rtx prev_insn
, next_insn
;
10230 bool insn_added_p
= false;
10231 rtx cond
, *op0
, *op1
;
10233 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
10235 /* Handle compare and branch and branch on count
10237 rtx pattern
= single_set (insn
);
10240 || SET_DEST (pattern
) != pc_rtx
10241 || GET_CODE (SET_SRC (pattern
)) != IF_THEN_ELSE
)
10244 cond
= XEXP (SET_SRC (pattern
), 0);
10245 op0
= &XEXP (cond
, 0);
10246 op1
= &XEXP (cond
, 1);
10248 else if (GET_CODE (PATTERN (insn
)) == SET
)
10252 /* Handle normal compare instructions. */
10253 src
= SET_SRC (PATTERN (insn
));
10254 dest
= SET_DEST (PATTERN (insn
));
10257 || !CC_REGNO_P (REGNO (dest
))
10258 || GET_CODE (src
) != COMPARE
)
10261 /* s390_swap_cmp will try to find the conditional
10262 jump when passing NULL_RTX as condition. */
10264 op0
= &XEXP (src
, 0);
10265 op1
= &XEXP (src
, 1);
10270 if (!REG_P (*op0
) || !REG_P (*op1
))
10273 if (GET_MODE_CLASS (GET_MODE (*op0
)) != MODE_INT
)
10276 /* Swap the COMPARE arguments and its mask if there is a
10277 conflicting access in the previous insn. */
10278 prev_insn
= prev_active_insn (insn
);
10279 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10280 && reg_referenced_p (*op1
, PATTERN (prev_insn
)))
10281 s390_swap_cmp (cond
, op0
, op1
, insn
);
10283 /* Check if there is a conflict with the next insn. If there
10284 was no conflict with the previous insn, then swap the
10285 COMPARE arguments and its mask. If we already swapped
10286 the operands, or if swapping them would cause a conflict
10287 with the previous insn, issue a NOP after the COMPARE in
10288 order to separate the two instuctions. */
10289 next_insn
= next_active_insn (insn
);
10290 if (next_insn
!= NULL_RTX
&& INSN_P (next_insn
)
10291 && s390_non_addr_reg_read_p (*op1
, next_insn
))
10293 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10294 && s390_non_addr_reg_read_p (*op0
, prev_insn
))
10296 if (REGNO (*op1
) == 0)
10297 emit_insn_after (gen_nop1 (), insn
);
10299 emit_insn_after (gen_nop (), insn
);
10300 insn_added_p
= true;
10303 s390_swap_cmp (cond
, op0
, op1
, insn
);
10305 return insn_added_p
;
10308 /* Perform machine-dependent processing. */
10313 bool pool_overflow
= false;
10315 /* Make sure all splits have been performed; splits after
10316 machine_dependent_reorg might confuse insn length counts. */
10317 split_all_insns_noflow ();
10319 /* Install the main literal pool and the associated base
10320 register load insns.
10322 In addition, there are two problematic situations we need
10325 - the literal pool might be > 4096 bytes in size, so that
10326 some of its elements cannot be directly accessed
10328 - a branch target might be > 64K away from the branch, so that
10329 it is not possible to use a PC-relative instruction.
10331 To fix those, we split the single literal pool into multiple
10332 pool chunks, reloading the pool base register at various
10333 points throughout the function to ensure it always points to
10334 the pool chunk the following code expects, and / or replace
10335 PC-relative branches by absolute branches.
10337 However, the two problems are interdependent: splitting the
10338 literal pool can move a branch further away from its target,
10339 causing the 64K limit to overflow, and on the other hand,
10340 replacing a PC-relative branch by an absolute branch means
10341 we need to put the branch target address into the literal
10342 pool, possibly causing it to overflow.
10344 So, we loop trying to fix up both problems until we manage
10345 to satisfy both conditions at the same time. Note that the
10346 loop is guaranteed to terminate as every pass of the loop
10347 strictly decreases the total number of PC-relative branches
10348 in the function. (This is not completely true as there
10349 might be branch-over-pool insns introduced by chunkify_start.
10350 Those never need to be split however.) */
10354 struct constant_pool
*pool
= NULL
;
10356 /* Collect the literal pool. */
10357 if (!pool_overflow
)
10359 pool
= s390_mainpool_start ();
10361 pool_overflow
= true;
10364 /* If literal pool overflowed, start to chunkify it. */
10366 pool
= s390_chunkify_start ();
10368 /* Split out-of-range branches. If this has created new
10369 literal pool entries, cancel current chunk list and
10370 recompute it. zSeries machines have large branch
10371 instructions, so we never need to split a branch. */
10372 if (!TARGET_CPU_ZARCH
&& s390_split_branches ())
10375 s390_chunkify_cancel (pool
);
10377 s390_mainpool_cancel (pool
);
10382 /* If we made it up to here, both conditions are satisfied.
10383 Finish up literal pool related changes. */
10385 s390_chunkify_finish (pool
);
10387 s390_mainpool_finish (pool
);
10389 /* We're done splitting branches. */
10390 cfun
->machine
->split_branches_pending_p
= false;
10394 /* Generate out-of-pool execute target insns. */
10395 if (TARGET_CPU_ZARCH
)
10397 rtx insn
, label
, target
;
10399 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10401 label
= s390_execute_label (insn
);
10405 gcc_assert (label
!= const0_rtx
);
10407 target
= emit_label (XEXP (label
, 0));
10408 INSN_ADDRESSES_NEW (target
, -1);
10410 target
= emit_insn (s390_execute_target (insn
));
10411 INSN_ADDRESSES_NEW (target
, -1);
10415 /* Try to optimize prologue and epilogue further. */
10416 s390_optimize_prologue ();
10418 /* Walk over the insns and do some >=z10 specific changes. */
10419 if (s390_tune
== PROCESSOR_2097_Z10
10420 || s390_tune
== PROCESSOR_2817_Z196
)
10423 bool insn_added_p
= false;
10425 /* The insn lengths and addresses have to be up to date for the
10426 following manipulations. */
10427 shorten_branches (get_insns ());
10429 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10431 if (!INSN_P (insn
) || INSN_CODE (insn
) <= 0)
10435 insn_added_p
|= s390_fix_long_loop_prediction (insn
);
10437 if ((GET_CODE (PATTERN (insn
)) == PARALLEL
10438 || GET_CODE (PATTERN (insn
)) == SET
)
10439 && s390_tune
== PROCESSOR_2097_Z10
)
10440 insn_added_p
|= s390_z10_optimize_cmp (insn
);
10443 /* Adjust branches if we added new instructions. */
10445 shorten_branches (get_insns ());
10449 /* Return true if INSN is a fp load insn writing register REGNO. */
10451 s390_fpload_toreg (rtx insn
, unsigned int regno
)
10454 enum attr_type flag
= s390_safe_attr_type (insn
);
10456 if (flag
!= TYPE_FLOADSF
&& flag
!= TYPE_FLOADDF
)
10459 set
= single_set (insn
);
10461 if (set
== NULL_RTX
)
10464 if (!REG_P (SET_DEST (set
)) || !MEM_P (SET_SRC (set
)))
10467 if (REGNO (SET_DEST (set
)) != regno
)
10473 /* This value describes the distance to be avoided between an
10474 aritmetic fp instruction and an fp load writing the same register.
10475 Z10_EARLYLOAD_DISTANCE - 1 as well as Z10_EARLYLOAD_DISTANCE + 1 is
10476 fine but the exact value has to be avoided. Otherwise the FP
10477 pipeline will throw an exception causing a major penalty. */
10478 #define Z10_EARLYLOAD_DISTANCE 7
10480 /* Rearrange the ready list in order to avoid the situation described
10481 for Z10_EARLYLOAD_DISTANCE. A problematic load instruction is
10482 moved to the very end of the ready list. */
10484 s390_z10_prevent_earlyload_conflicts (rtx
*ready
, int *nready_p
)
10486 unsigned int regno
;
10487 int nready
= *nready_p
;
10492 enum attr_type flag
;
10495 /* Skip DISTANCE - 1 active insns. */
10496 for (insn
= last_scheduled_insn
, distance
= Z10_EARLYLOAD_DISTANCE
- 1;
10497 distance
> 0 && insn
!= NULL_RTX
;
10498 distance
--, insn
= prev_active_insn (insn
))
10499 if (CALL_P (insn
) || JUMP_P (insn
))
10502 if (insn
== NULL_RTX
)
10505 set
= single_set (insn
);
10507 if (set
== NULL_RTX
|| !REG_P (SET_DEST (set
))
10508 || GET_MODE_CLASS (GET_MODE (SET_DEST (set
))) != MODE_FLOAT
)
10511 flag
= s390_safe_attr_type (insn
);
10513 if (flag
== TYPE_FLOADSF
|| flag
== TYPE_FLOADDF
)
10516 regno
= REGNO (SET_DEST (set
));
10519 while (!s390_fpload_toreg (ready
[i
], regno
) && i
> 0)
10526 memmove (&ready
[1], &ready
[0], sizeof (rtx
) * i
);
10530 /* This function is called via hook TARGET_SCHED_REORDER before
10531 issueing one insn from list READY which contains *NREADYP entries.
10532 For target z10 it reorders load instructions to avoid early load
10533 conflicts in the floating point pipeline */
10535 s390_sched_reorder (FILE *file ATTRIBUTE_UNUSED
, int verbose ATTRIBUTE_UNUSED
,
10536 rtx
*ready
, int *nreadyp
, int clock ATTRIBUTE_UNUSED
)
10538 if (s390_tune
== PROCESSOR_2097_Z10
)
10539 if (reload_completed
&& *nreadyp
> 1)
10540 s390_z10_prevent_earlyload_conflicts (ready
, nreadyp
);
10542 return s390_issue_rate ();
10545 /* This function is called via hook TARGET_SCHED_VARIABLE_ISSUE after
10546 the scheduler has issued INSN. It stores the last issued insn into
10547 last_scheduled_insn in order to make it available for
10548 s390_sched_reorder. */
10550 s390_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED
,
10551 int verbose ATTRIBUTE_UNUSED
,
10552 rtx insn
, int more
)
10554 last_scheduled_insn
= insn
;
10556 if (GET_CODE (PATTERN (insn
)) != USE
10557 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
10564 s390_sched_init (FILE *file ATTRIBUTE_UNUSED
,
10565 int verbose ATTRIBUTE_UNUSED
,
10566 int max_ready ATTRIBUTE_UNUSED
)
10568 last_scheduled_insn
= NULL_RTX
;
10571 /* This function checks the whole of insn X for memory references. The
10572 function always returns zero because the framework it is called
10573 from would stop recursively analyzing the insn upon a return value
10574 other than zero. The real result of this function is updating
10575 counter variable MEM_COUNT. */
10577 check_dpu (rtx
*x
, unsigned *mem_count
)
10579 if (*x
!= NULL_RTX
&& MEM_P (*x
))
10584 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates
10585 a new number struct loop *loop should be unrolled if tuned for cpus with
10586 a built-in stride prefetcher.
10587 The loop is analyzed for memory accesses by calling check_dpu for
10588 each rtx of the loop. Depending on the loop_depth and the amount of
10589 memory accesses a new number <=nunroll is returned to improve the
10590 behaviour of the hardware prefetch unit. */
10592 s390_loop_unroll_adjust (unsigned nunroll
, struct loop
*loop
)
10597 unsigned mem_count
= 0;
10599 if (s390_tune
!= PROCESSOR_2097_Z10
&& s390_tune
!= PROCESSOR_2817_Z196
)
10602 /* Count the number of memory references within the loop body. */
10603 bbs
= get_loop_body (loop
);
10604 for (i
= 0; i
< loop
->num_nodes
; i
++)
10606 for (insn
= BB_HEAD (bbs
[i
]); insn
!= BB_END (bbs
[i
]); insn
= NEXT_INSN (insn
))
10607 if (INSN_P (insn
) && INSN_CODE (insn
) != -1)
10608 for_each_rtx (&insn
, (rtx_function
) check_dpu
, &mem_count
);
10612 /* Prevent division by zero, and we do not need to adjust nunroll in this case. */
10613 if (mem_count
== 0)
10616 switch (loop_depth(loop
))
10619 return MIN (nunroll
, 28 / mem_count
);
10621 return MIN (nunroll
, 22 / mem_count
);
10623 return MIN (nunroll
, 16 / mem_count
);
10627 /* Initialize GCC target structure. */
10629 #undef TARGET_ASM_ALIGNED_HI_OP
10630 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10631 #undef TARGET_ASM_ALIGNED_DI_OP
10632 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10633 #undef TARGET_ASM_INTEGER
10634 #define TARGET_ASM_INTEGER s390_assemble_integer
10636 #undef TARGET_ASM_OPEN_PAREN
10637 #define TARGET_ASM_OPEN_PAREN ""
10639 #undef TARGET_ASM_CLOSE_PAREN
10640 #define TARGET_ASM_CLOSE_PAREN ""
10642 #undef TARGET_OPTION_OVERRIDE
10643 #define TARGET_OPTION_OVERRIDE s390_option_override
10645 #undef TARGET_ENCODE_SECTION_INFO
10646 #define TARGET_ENCODE_SECTION_INFO s390_encode_section_info
10648 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10649 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10652 #undef TARGET_HAVE_TLS
10653 #define TARGET_HAVE_TLS true
10655 #undef TARGET_CANNOT_FORCE_CONST_MEM
10656 #define TARGET_CANNOT_FORCE_CONST_MEM s390_cannot_force_const_mem
10658 #undef TARGET_DELEGITIMIZE_ADDRESS
10659 #define TARGET_DELEGITIMIZE_ADDRESS s390_delegitimize_address
10661 #undef TARGET_LEGITIMIZE_ADDRESS
10662 #define TARGET_LEGITIMIZE_ADDRESS s390_legitimize_address
10664 #undef TARGET_RETURN_IN_MEMORY
10665 #define TARGET_RETURN_IN_MEMORY s390_return_in_memory
10667 #undef TARGET_INIT_BUILTINS
10668 #define TARGET_INIT_BUILTINS s390_init_builtins
10669 #undef TARGET_EXPAND_BUILTIN
10670 #define TARGET_EXPAND_BUILTIN s390_expand_builtin
10672 #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
10673 #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA s390_output_addr_const_extra
10675 #undef TARGET_ASM_OUTPUT_MI_THUNK
10676 #define TARGET_ASM_OUTPUT_MI_THUNK s390_output_mi_thunk
10677 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10678 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10680 #undef TARGET_SCHED_ADJUST_PRIORITY
10681 #define TARGET_SCHED_ADJUST_PRIORITY s390_adjust_priority
10682 #undef TARGET_SCHED_ISSUE_RATE
10683 #define TARGET_SCHED_ISSUE_RATE s390_issue_rate
10684 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10685 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD s390_first_cycle_multipass_dfa_lookahead
10687 #undef TARGET_SCHED_VARIABLE_ISSUE
10688 #define TARGET_SCHED_VARIABLE_ISSUE s390_sched_variable_issue
10689 #undef TARGET_SCHED_REORDER
10690 #define TARGET_SCHED_REORDER s390_sched_reorder
10691 #undef TARGET_SCHED_INIT
10692 #define TARGET_SCHED_INIT s390_sched_init
10694 #undef TARGET_CANNOT_COPY_INSN_P
10695 #define TARGET_CANNOT_COPY_INSN_P s390_cannot_copy_insn_p
10696 #undef TARGET_RTX_COSTS
10697 #define TARGET_RTX_COSTS s390_rtx_costs
10698 #undef TARGET_ADDRESS_COST
10699 #define TARGET_ADDRESS_COST s390_address_cost
10700 #undef TARGET_REGISTER_MOVE_COST
10701 #define TARGET_REGISTER_MOVE_COST s390_register_move_cost
10702 #undef TARGET_MEMORY_MOVE_COST
10703 #define TARGET_MEMORY_MOVE_COST s390_memory_move_cost
10705 #undef TARGET_MACHINE_DEPENDENT_REORG
10706 #define TARGET_MACHINE_DEPENDENT_REORG s390_reorg
10708 #undef TARGET_VALID_POINTER_MODE
10709 #define TARGET_VALID_POINTER_MODE s390_valid_pointer_mode
10711 #undef TARGET_BUILD_BUILTIN_VA_LIST
10712 #define TARGET_BUILD_BUILTIN_VA_LIST s390_build_builtin_va_list
10713 #undef TARGET_EXPAND_BUILTIN_VA_START
10714 #define TARGET_EXPAND_BUILTIN_VA_START s390_va_start
10715 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10716 #define TARGET_GIMPLIFY_VA_ARG_EXPR s390_gimplify_va_arg
10718 #undef TARGET_PROMOTE_FUNCTION_MODE
10719 #define TARGET_PROMOTE_FUNCTION_MODE s390_promote_function_mode
10720 #undef TARGET_PASS_BY_REFERENCE
10721 #define TARGET_PASS_BY_REFERENCE s390_pass_by_reference
10723 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10724 #define TARGET_FUNCTION_OK_FOR_SIBCALL s390_function_ok_for_sibcall
10725 #undef TARGET_FUNCTION_ARG
10726 #define TARGET_FUNCTION_ARG s390_function_arg
10727 #undef TARGET_FUNCTION_ARG_ADVANCE
10728 #define TARGET_FUNCTION_ARG_ADVANCE s390_function_arg_advance
10729 #undef TARGET_FUNCTION_VALUE
10730 #define TARGET_FUNCTION_VALUE s390_function_value
10731 #undef TARGET_LIBCALL_VALUE
10732 #define TARGET_LIBCALL_VALUE s390_libcall_value
10734 #undef TARGET_FIXED_CONDITION_CODE_REGS
10735 #define TARGET_FIXED_CONDITION_CODE_REGS s390_fixed_condition_code_regs
10737 #undef TARGET_CC_MODES_COMPATIBLE
10738 #define TARGET_CC_MODES_COMPATIBLE s390_cc_modes_compatible
10740 #undef TARGET_INVALID_WITHIN_DOLOOP
10741 #define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null
10744 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
10745 #define TARGET_ASM_OUTPUT_DWARF_DTPREL s390_output_dwarf_dtprel
10748 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10749 #undef TARGET_MANGLE_TYPE
10750 #define TARGET_MANGLE_TYPE s390_mangle_type
10753 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10754 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10756 #undef TARGET_PREFERRED_RELOAD_CLASS
10757 #define TARGET_PREFERRED_RELOAD_CLASS s390_preferred_reload_class
10759 #undef TARGET_SECONDARY_RELOAD
10760 #define TARGET_SECONDARY_RELOAD s390_secondary_reload
10762 #undef TARGET_LIBGCC_CMP_RETURN_MODE
10763 #define TARGET_LIBGCC_CMP_RETURN_MODE s390_libgcc_cmp_return_mode
10765 #undef TARGET_LIBGCC_SHIFT_COUNT_MODE
10766 #define TARGET_LIBGCC_SHIFT_COUNT_MODE s390_libgcc_shift_count_mode
10768 #undef TARGET_LEGITIMATE_ADDRESS_P
10769 #define TARGET_LEGITIMATE_ADDRESS_P s390_legitimate_address_p
10771 #undef TARGET_LEGITIMATE_CONSTANT_P
10772 #define TARGET_LEGITIMATE_CONSTANT_P s390_legitimate_constant_p
10774 #undef TARGET_CAN_ELIMINATE
10775 #define TARGET_CAN_ELIMINATE s390_can_eliminate
10777 #undef TARGET_CONDITIONAL_REGISTER_USAGE
10778 #define TARGET_CONDITIONAL_REGISTER_USAGE s390_conditional_register_usage
10780 #undef TARGET_LOOP_UNROLL_ADJUST
10781 #define TARGET_LOOP_UNROLL_ADJUST s390_loop_unroll_adjust
10783 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
10784 #define TARGET_ASM_TRAMPOLINE_TEMPLATE s390_asm_trampoline_template
10785 #undef TARGET_TRAMPOLINE_INIT
10786 #define TARGET_TRAMPOLINE_INIT s390_trampoline_init
10788 #undef TARGET_UNWIND_WORD_MODE
10789 #define TARGET_UNWIND_WORD_MODE s390_unwind_word_mode
10791 struct gcc_target targetm
= TARGET_INITIALIZER
;
10793 #include "gt-s390.h"