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 (Pmode
, 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 (Pmode
, 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 (Pmode
, new_rtx
,
3811 INTVAL (XEXP (XEXP (addr
, 0), 1)));
3812 new_rtx
= force_operand (new_rtx
, 0);
3816 gcc_unreachable (); /* for now ... */
3821 /* Emit insns making the address in operands[1] valid for a standard
3822 move to operands[0]. operands[1] is replaced by an address which
3823 should be used instead of the former RTX to emit the move
3827 emit_symbolic_move (rtx
*operands
)
3829 rtx temp
= !can_create_pseudo_p () ? operands
[0] : gen_reg_rtx (Pmode
);
3831 if (GET_CODE (operands
[0]) == MEM
)
3832 operands
[1] = force_reg (Pmode
, operands
[1]);
3833 else if (TLS_SYMBOLIC_CONST (operands
[1]))
3834 operands
[1] = legitimize_tls_address (operands
[1], temp
);
3836 operands
[1] = legitimize_pic_address (operands
[1], temp
);
3839 /* Try machine-dependent ways of modifying an illegitimate address X
3840 to be legitimate. If we find one, return the new, valid address.
3842 OLDX is the address as it was before break_out_memory_refs was called.
3843 In some cases it is useful to look at this to decide what needs to be done.
3845 MODE is the mode of the operand pointed to by X.
3847 When -fpic is used, special handling is needed for symbolic references.
3848 See comments by legitimize_pic_address for details. */
3851 s390_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
3852 enum machine_mode mode ATTRIBUTE_UNUSED
)
3854 rtx constant_term
= const0_rtx
;
3856 if (TLS_SYMBOLIC_CONST (x
))
3858 x
= legitimize_tls_address (x
, 0);
3860 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3863 else if (GET_CODE (x
) == PLUS
3864 && (TLS_SYMBOLIC_CONST (XEXP (x
, 0))
3865 || TLS_SYMBOLIC_CONST (XEXP (x
, 1))))
3871 if (SYMBOLIC_CONST (x
)
3872 || (GET_CODE (x
) == PLUS
3873 && (SYMBOLIC_CONST (XEXP (x
, 0))
3874 || SYMBOLIC_CONST (XEXP (x
, 1)))))
3875 x
= legitimize_pic_address (x
, 0);
3877 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3881 x
= eliminate_constant_term (x
, &constant_term
);
3883 /* Optimize loading of large displacements by splitting them
3884 into the multiple of 4K and the rest; this allows the
3885 former to be CSE'd if possible.
3887 Don't do this if the displacement is added to a register
3888 pointing into the stack frame, as the offsets will
3889 change later anyway. */
3891 if (GET_CODE (constant_term
) == CONST_INT
3892 && !TARGET_LONG_DISPLACEMENT
3893 && !DISP_IN_RANGE (INTVAL (constant_term
))
3894 && !(REG_P (x
) && REGNO_PTR_FRAME_P (REGNO (x
))))
3896 HOST_WIDE_INT lower
= INTVAL (constant_term
) & 0xfff;
3897 HOST_WIDE_INT upper
= INTVAL (constant_term
) ^ lower
;
3899 rtx temp
= gen_reg_rtx (Pmode
);
3900 rtx val
= force_operand (GEN_INT (upper
), temp
);
3902 emit_move_insn (temp
, val
);
3904 x
= gen_rtx_PLUS (Pmode
, x
, temp
);
3905 constant_term
= GEN_INT (lower
);
3908 if (GET_CODE (x
) == PLUS
)
3910 if (GET_CODE (XEXP (x
, 0)) == REG
)
3912 rtx temp
= gen_reg_rtx (Pmode
);
3913 rtx val
= force_operand (XEXP (x
, 1), temp
);
3915 emit_move_insn (temp
, val
);
3917 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0), temp
);
3920 else if (GET_CODE (XEXP (x
, 1)) == REG
)
3922 rtx temp
= gen_reg_rtx (Pmode
);
3923 rtx val
= force_operand (XEXP (x
, 0), temp
);
3925 emit_move_insn (temp
, val
);
3927 x
= gen_rtx_PLUS (Pmode
, temp
, XEXP (x
, 1));
3931 if (constant_term
!= const0_rtx
)
3932 x
= gen_rtx_PLUS (Pmode
, x
, constant_term
);
3937 /* Try a machine-dependent way of reloading an illegitimate address AD
3938 operand. If we find one, push the reload and return the new address.
3940 MODE is the mode of the enclosing MEM. OPNUM is the operand number
3941 and TYPE is the reload type of the current reload. */
3944 legitimize_reload_address (rtx ad
, enum machine_mode mode ATTRIBUTE_UNUSED
,
3945 int opnum
, int type
)
3947 if (!optimize
|| TARGET_LONG_DISPLACEMENT
)
3950 if (GET_CODE (ad
) == PLUS
)
3952 rtx tem
= simplify_binary_operation (PLUS
, Pmode
,
3953 XEXP (ad
, 0), XEXP (ad
, 1));
3958 if (GET_CODE (ad
) == PLUS
3959 && GET_CODE (XEXP (ad
, 0)) == REG
3960 && GET_CODE (XEXP (ad
, 1)) == CONST_INT
3961 && !DISP_IN_RANGE (INTVAL (XEXP (ad
, 1))))
3963 HOST_WIDE_INT lower
= INTVAL (XEXP (ad
, 1)) & 0xfff;
3964 HOST_WIDE_INT upper
= INTVAL (XEXP (ad
, 1)) ^ lower
;
3965 rtx cst
, tem
, new_rtx
;
3967 cst
= GEN_INT (upper
);
3968 if (!legitimate_reload_constant_p (cst
))
3969 cst
= force_const_mem (Pmode
, cst
);
3971 tem
= gen_rtx_PLUS (Pmode
, XEXP (ad
, 0), cst
);
3972 new_rtx
= gen_rtx_PLUS (Pmode
, tem
, GEN_INT (lower
));
3974 push_reload (XEXP (tem
, 1), 0, &XEXP (tem
, 1), 0,
3975 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0,
3976 opnum
, (enum reload_type
) type
);
3983 /* Emit code to move LEN bytes from DST to SRC. */
3986 s390_expand_movmem (rtx dst
, rtx src
, rtx len
)
3988 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
3990 if (INTVAL (len
) > 0)
3991 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (INTVAL (len
) - 1)));
3994 else if (TARGET_MVCLE
)
3996 emit_insn (gen_movmem_long (dst
, src
, convert_to_mode (Pmode
, len
, 1)));
4001 rtx dst_addr
, src_addr
, count
, blocks
, temp
;
4002 rtx loop_start_label
= gen_label_rtx ();
4003 rtx loop_end_label
= gen_label_rtx ();
4004 rtx end_label
= gen_label_rtx ();
4005 enum machine_mode mode
;
4007 mode
= GET_MODE (len
);
4008 if (mode
== VOIDmode
)
4011 dst_addr
= gen_reg_rtx (Pmode
);
4012 src_addr
= gen_reg_rtx (Pmode
);
4013 count
= gen_reg_rtx (mode
);
4014 blocks
= gen_reg_rtx (mode
);
4016 convert_move (count
, len
, 1);
4017 emit_cmp_and_jump_insns (count
, const0_rtx
,
4018 EQ
, NULL_RTX
, mode
, 1, end_label
);
4020 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4021 emit_move_insn (src_addr
, force_operand (XEXP (src
, 0), NULL_RTX
));
4022 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4023 src
= change_address (src
, VOIDmode
, src_addr
);
4025 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4028 emit_move_insn (count
, temp
);
4030 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4033 emit_move_insn (blocks
, temp
);
4035 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4036 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4038 emit_label (loop_start_label
);
4041 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 768))
4045 /* Issue a read prefetch for the +3 cache line. */
4046 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (768)),
4047 const0_rtx
, const0_rtx
);
4048 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4049 emit_insn (prefetch
);
4051 /* Issue a write prefetch for the +3 cache line. */
4052 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (768)),
4053 const1_rtx
, const0_rtx
);
4054 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4055 emit_insn (prefetch
);
4058 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (255)));
4059 s390_load_address (dst_addr
,
4060 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4061 s390_load_address (src_addr
,
4062 gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (256)));
4064 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4067 emit_move_insn (blocks
, temp
);
4069 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4070 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4072 emit_jump (loop_start_label
);
4073 emit_label (loop_end_label
);
4075 emit_insn (gen_movmem_short (dst
, src
,
4076 convert_to_mode (Pmode
, count
, 1)));
4077 emit_label (end_label
);
4081 /* Emit code to set LEN bytes at DST to VAL.
4082 Make use of clrmem if VAL is zero. */
4085 s390_expand_setmem (rtx dst
, rtx len
, rtx val
)
4087 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) == 0)
4090 gcc_assert (GET_CODE (val
) == CONST_INT
|| GET_MODE (val
) == QImode
);
4092 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) > 0 && INTVAL (len
) <= 257)
4094 if (val
== const0_rtx
&& INTVAL (len
) <= 256)
4095 emit_insn (gen_clrmem_short (dst
, GEN_INT (INTVAL (len
) - 1)));
4098 /* Initialize memory by storing the first byte. */
4099 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4101 if (INTVAL (len
) > 1)
4103 /* Initiate 1 byte overlap move.
4104 The first byte of DST is propagated through DSTP1.
4105 Prepare a movmem for: DST+1 = DST (length = LEN - 1).
4106 DST is set to size 1 so the rest of the memory location
4107 does not count as source operand. */
4108 rtx dstp1
= adjust_address (dst
, VOIDmode
, 1);
4109 set_mem_size (dst
, 1);
4111 emit_insn (gen_movmem_short (dstp1
, dst
,
4112 GEN_INT (INTVAL (len
) - 2)));
4117 else if (TARGET_MVCLE
)
4119 val
= force_not_mem (convert_modes (Pmode
, QImode
, val
, 1));
4120 emit_insn (gen_setmem_long (dst
, convert_to_mode (Pmode
, len
, 1), val
));
4125 rtx dst_addr
, count
, blocks
, temp
, dstp1
= NULL_RTX
;
4126 rtx loop_start_label
= gen_label_rtx ();
4127 rtx loop_end_label
= gen_label_rtx ();
4128 rtx end_label
= gen_label_rtx ();
4129 enum machine_mode mode
;
4131 mode
= GET_MODE (len
);
4132 if (mode
== VOIDmode
)
4135 dst_addr
= gen_reg_rtx (Pmode
);
4136 count
= gen_reg_rtx (mode
);
4137 blocks
= gen_reg_rtx (mode
);
4139 convert_move (count
, len
, 1);
4140 emit_cmp_and_jump_insns (count
, const0_rtx
,
4141 EQ
, NULL_RTX
, mode
, 1, end_label
);
4143 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4144 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4146 if (val
== const0_rtx
)
4147 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4151 dstp1
= adjust_address (dst
, VOIDmode
, 1);
4152 set_mem_size (dst
, 1);
4154 /* Initialize memory by storing the first byte. */
4155 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4157 /* If count is 1 we are done. */
4158 emit_cmp_and_jump_insns (count
, const1_rtx
,
4159 EQ
, NULL_RTX
, mode
, 1, end_label
);
4161 temp
= expand_binop (mode
, add_optab
, count
, GEN_INT (-2), count
, 1,
4165 emit_move_insn (count
, temp
);
4167 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4170 emit_move_insn (blocks
, temp
);
4172 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4173 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4175 emit_label (loop_start_label
);
4178 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 1024))
4180 /* Issue a write prefetch for the +4 cache line. */
4181 rtx prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
,
4183 const1_rtx
, const0_rtx
);
4184 emit_insn (prefetch
);
4185 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4188 if (val
== const0_rtx
)
4189 emit_insn (gen_clrmem_short (dst
, GEN_INT (255)));
4191 emit_insn (gen_movmem_short (dstp1
, dst
, GEN_INT (255)));
4192 s390_load_address (dst_addr
,
4193 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4195 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4198 emit_move_insn (blocks
, temp
);
4200 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4201 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4203 emit_jump (loop_start_label
);
4204 emit_label (loop_end_label
);
4206 if (val
== const0_rtx
)
4207 emit_insn (gen_clrmem_short (dst
, convert_to_mode (Pmode
, count
, 1)));
4209 emit_insn (gen_movmem_short (dstp1
, dst
, convert_to_mode (Pmode
, count
, 1)));
4210 emit_label (end_label
);
4214 /* Emit code to compare LEN bytes at OP0 with those at OP1,
4215 and return the result in TARGET. */
4218 s390_expand_cmpmem (rtx target
, rtx op0
, rtx op1
, rtx len
)
4220 rtx ccreg
= gen_rtx_REG (CCUmode
, CC_REGNUM
);
4223 /* As the result of CMPINT is inverted compared to what we need,
4224 we have to swap the operands. */
4225 tmp
= op0
; op0
= op1
; op1
= tmp
;
4227 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4229 if (INTVAL (len
) > 0)
4231 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (INTVAL (len
) - 1)));
4232 emit_insn (gen_cmpint (target
, ccreg
));
4235 emit_move_insn (target
, const0_rtx
);
4237 else if (TARGET_MVCLE
)
4239 emit_insn (gen_cmpmem_long (op0
, op1
, convert_to_mode (Pmode
, len
, 1)));
4240 emit_insn (gen_cmpint (target
, ccreg
));
4244 rtx addr0
, addr1
, count
, blocks
, temp
;
4245 rtx loop_start_label
= gen_label_rtx ();
4246 rtx loop_end_label
= gen_label_rtx ();
4247 rtx end_label
= gen_label_rtx ();
4248 enum machine_mode mode
;
4250 mode
= GET_MODE (len
);
4251 if (mode
== VOIDmode
)
4254 addr0
= gen_reg_rtx (Pmode
);
4255 addr1
= gen_reg_rtx (Pmode
);
4256 count
= gen_reg_rtx (mode
);
4257 blocks
= gen_reg_rtx (mode
);
4259 convert_move (count
, len
, 1);
4260 emit_cmp_and_jump_insns (count
, const0_rtx
,
4261 EQ
, NULL_RTX
, mode
, 1, end_label
);
4263 emit_move_insn (addr0
, force_operand (XEXP (op0
, 0), NULL_RTX
));
4264 emit_move_insn (addr1
, force_operand (XEXP (op1
, 0), NULL_RTX
));
4265 op0
= change_address (op0
, VOIDmode
, addr0
);
4266 op1
= change_address (op1
, VOIDmode
, addr1
);
4268 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4271 emit_move_insn (count
, temp
);
4273 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4276 emit_move_insn (blocks
, temp
);
4278 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4279 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4281 emit_label (loop_start_label
);
4284 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 512))
4288 /* Issue a read prefetch for the +2 cache line of operand 1. */
4289 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (512)),
4290 const0_rtx
, const0_rtx
);
4291 emit_insn (prefetch
);
4292 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4294 /* Issue a read prefetch for the +2 cache line of operand 2. */
4295 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (512)),
4296 const0_rtx
, const0_rtx
);
4297 emit_insn (prefetch
);
4298 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4301 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (255)));
4302 temp
= gen_rtx_NE (VOIDmode
, ccreg
, const0_rtx
);
4303 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
4304 gen_rtx_LABEL_REF (VOIDmode
, end_label
), pc_rtx
);
4305 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
4306 emit_jump_insn (temp
);
4308 s390_load_address (addr0
,
4309 gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (256)));
4310 s390_load_address (addr1
,
4311 gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (256)));
4313 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4316 emit_move_insn (blocks
, temp
);
4318 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4319 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4321 emit_jump (loop_start_label
);
4322 emit_label (loop_end_label
);
4324 emit_insn (gen_cmpmem_short (op0
, op1
,
4325 convert_to_mode (Pmode
, count
, 1)));
4326 emit_label (end_label
);
4328 emit_insn (gen_cmpint (target
, ccreg
));
4333 /* Expand conditional increment or decrement using alc/slb instructions.
4334 Should generate code setting DST to either SRC or SRC + INCREMENT,
4335 depending on the result of the comparison CMP_OP0 CMP_CODE CMP_OP1.
4336 Returns true if successful, false otherwise.
4338 That makes it possible to implement some if-constructs without jumps e.g.:
4339 (borrow = CC0 | CC1 and carry = CC2 | CC3)
4340 unsigned int a, b, c;
4341 if (a < b) c++; -> CCU b > a -> CC2; c += carry;
4342 if (a < b) c--; -> CCL3 a - b -> borrow; c -= borrow;
4343 if (a <= b) c++; -> CCL3 b - a -> borrow; c += carry;
4344 if (a <= b) c--; -> CCU a <= b -> borrow; c -= borrow;
4346 Checks for EQ and NE with a nonzero value need an additional xor e.g.:
4347 if (a == b) c++; -> CCL3 a ^= b; 0 - a -> borrow; c += carry;
4348 if (a == b) c--; -> CCU a ^= b; a <= 0 -> CC0 | CC1; c -= borrow;
4349 if (a != b) c++; -> CCU a ^= b; a > 0 -> CC2; c += carry;
4350 if (a != b) c--; -> CCL3 a ^= b; 0 - a -> borrow; c -= borrow; */
4353 s390_expand_addcc (enum rtx_code cmp_code
, rtx cmp_op0
, rtx cmp_op1
,
4354 rtx dst
, rtx src
, rtx increment
)
4356 enum machine_mode cmp_mode
;
4357 enum machine_mode cc_mode
;
4363 if ((GET_MODE (cmp_op0
) == SImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4364 && (GET_MODE (cmp_op1
) == SImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4366 else if ((GET_MODE (cmp_op0
) == DImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4367 && (GET_MODE (cmp_op1
) == DImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4372 /* Try ADD LOGICAL WITH CARRY. */
4373 if (increment
== const1_rtx
)
4375 /* Determine CC mode to use. */
4376 if (cmp_code
== EQ
|| cmp_code
== NE
)
4378 if (cmp_op1
!= const0_rtx
)
4380 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4381 NULL_RTX
, 0, OPTAB_WIDEN
);
4382 cmp_op1
= const0_rtx
;
4385 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4388 if (cmp_code
== LTU
|| cmp_code
== LEU
)
4393 cmp_code
= swap_condition (cmp_code
);
4410 /* Emit comparison instruction pattern. */
4411 if (!register_operand (cmp_op0
, cmp_mode
))
4412 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4414 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4415 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4416 /* We use insn_invalid_p here to add clobbers if required. */
4417 ret
= insn_invalid_p (emit_insn (insn
), false);
4420 /* Emit ALC instruction pattern. */
4421 op_res
= gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4422 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4425 if (src
!= const0_rtx
)
4427 if (!register_operand (src
, GET_MODE (dst
)))
4428 src
= force_reg (GET_MODE (dst
), src
);
4430 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, src
);
4431 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, const0_rtx
);
4434 p
= rtvec_alloc (2);
4436 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4438 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4439 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4444 /* Try SUBTRACT LOGICAL WITH BORROW. */
4445 if (increment
== constm1_rtx
)
4447 /* Determine CC mode to use. */
4448 if (cmp_code
== EQ
|| cmp_code
== NE
)
4450 if (cmp_op1
!= const0_rtx
)
4452 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4453 NULL_RTX
, 0, OPTAB_WIDEN
);
4454 cmp_op1
= const0_rtx
;
4457 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4460 if (cmp_code
== GTU
|| cmp_code
== GEU
)
4465 cmp_code
= swap_condition (cmp_code
);
4482 /* Emit comparison instruction pattern. */
4483 if (!register_operand (cmp_op0
, cmp_mode
))
4484 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4486 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4487 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4488 /* We use insn_invalid_p here to add clobbers if required. */
4489 ret
= insn_invalid_p (emit_insn (insn
), false);
4492 /* Emit SLB instruction pattern. */
4493 if (!register_operand (src
, GET_MODE (dst
)))
4494 src
= force_reg (GET_MODE (dst
), src
);
4496 op_res
= gen_rtx_MINUS (GET_MODE (dst
),
4497 gen_rtx_MINUS (GET_MODE (dst
), src
, const0_rtx
),
4498 gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4499 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4501 p
= rtvec_alloc (2);
4503 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4505 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4506 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4514 /* Expand code for the insv template. Return true if successful. */
4517 s390_expand_insv (rtx dest
, rtx op1
, rtx op2
, rtx src
)
4519 int bitsize
= INTVAL (op1
);
4520 int bitpos
= INTVAL (op2
);
4522 /* On z10 we can use the risbg instruction to implement insv. */
4524 && ((GET_MODE (dest
) == DImode
&& GET_MODE (src
) == DImode
)
4525 || (GET_MODE (dest
) == SImode
&& GET_MODE (src
) == SImode
)))
4530 op
= gen_rtx_SET (GET_MODE(src
),
4531 gen_rtx_ZERO_EXTRACT (GET_MODE (dest
), dest
, op1
, op2
),
4533 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4534 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4539 /* We need byte alignment. */
4540 if (bitsize
% BITS_PER_UNIT
)
4544 && memory_operand (dest
, VOIDmode
)
4545 && (register_operand (src
, word_mode
)
4546 || const_int_operand (src
, VOIDmode
)))
4548 /* Emit standard pattern if possible. */
4549 enum machine_mode mode
= smallest_mode_for_size (bitsize
, MODE_INT
);
4550 if (GET_MODE_BITSIZE (mode
) == bitsize
)
4551 emit_move_insn (adjust_address (dest
, mode
, 0), gen_lowpart (mode
, src
));
4553 /* (set (ze (mem)) (const_int)). */
4554 else if (const_int_operand (src
, VOIDmode
))
4556 int size
= bitsize
/ BITS_PER_UNIT
;
4557 rtx src_mem
= adjust_address (force_const_mem (word_mode
, src
), BLKmode
,
4558 GET_MODE_SIZE (word_mode
) - size
);
4560 dest
= adjust_address (dest
, BLKmode
, 0);
4561 set_mem_size (dest
, size
);
4562 s390_expand_movmem (dest
, src_mem
, GEN_INT (size
));
4565 /* (set (ze (mem)) (reg)). */
4566 else if (register_operand (src
, word_mode
))
4568 if (bitsize
<= GET_MODE_BITSIZE (SImode
))
4569 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, op1
,
4573 /* Emit st,stcmh sequence. */
4574 int stcmh_width
= bitsize
- GET_MODE_BITSIZE (SImode
);
4575 int size
= stcmh_width
/ BITS_PER_UNIT
;
4577 emit_move_insn (adjust_address (dest
, SImode
, size
),
4578 gen_lowpart (SImode
, src
));
4579 set_mem_size (dest
, size
);
4580 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, GEN_INT
4581 (stcmh_width
), const0_rtx
),
4582 gen_rtx_LSHIFTRT (word_mode
, src
, GEN_INT
4583 (GET_MODE_BITSIZE (SImode
))));
4592 /* (set (ze (reg)) (const_int)). */
4594 && register_operand (dest
, word_mode
)
4595 && (bitpos
% 16) == 0
4596 && (bitsize
% 16) == 0
4597 && const_int_operand (src
, VOIDmode
))
4599 HOST_WIDE_INT val
= INTVAL (src
);
4600 int regpos
= bitpos
+ bitsize
;
4602 while (regpos
> bitpos
)
4604 enum machine_mode putmode
;
4607 if (TARGET_EXTIMM
&& (regpos
% 32 == 0) && (regpos
>= bitpos
+ 32))
4612 putsize
= GET_MODE_BITSIZE (putmode
);
4614 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4617 gen_int_mode (val
, putmode
));
4620 gcc_assert (regpos
== bitpos
);
4627 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic which returns a
4628 register that holds VAL of mode MODE shifted by COUNT bits. */
4631 s390_expand_mask_and_shift (rtx val
, enum machine_mode mode
, rtx count
)
4633 val
= expand_simple_binop (SImode
, AND
, val
, GEN_INT (GET_MODE_MASK (mode
)),
4634 NULL_RTX
, 1, OPTAB_DIRECT
);
4635 return expand_simple_binop (SImode
, ASHIFT
, val
, count
,
4636 NULL_RTX
, 1, OPTAB_DIRECT
);
4639 /* Structure to hold the initial parameters for a compare_and_swap operation
4640 in HImode and QImode. */
4642 struct alignment_context
4644 rtx memsi
; /* SI aligned memory location. */
4645 rtx shift
; /* Bit offset with regard to lsb. */
4646 rtx modemask
; /* Mask of the HQImode shifted by SHIFT bits. */
4647 rtx modemaski
; /* ~modemask */
4648 bool aligned
; /* True if memory is aligned, false else. */
4651 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic to initialize
4652 structure AC for transparent simplifying, if the memory alignment is known
4653 to be at least 32bit. MEM is the memory location for the actual operation
4654 and MODE its mode. */
4657 init_alignment_context (struct alignment_context
*ac
, rtx mem
,
4658 enum machine_mode mode
)
4660 ac
->shift
= GEN_INT (GET_MODE_SIZE (SImode
) - GET_MODE_SIZE (mode
));
4661 ac
->aligned
= (MEM_ALIGN (mem
) >= GET_MODE_BITSIZE (SImode
));
4664 ac
->memsi
= adjust_address (mem
, SImode
, 0); /* Memory is aligned. */
4667 /* Alignment is unknown. */
4668 rtx byteoffset
, addr
, align
;
4670 /* Force the address into a register. */
4671 addr
= force_reg (Pmode
, XEXP (mem
, 0));
4673 /* Align it to SImode. */
4674 align
= expand_simple_binop (Pmode
, AND
, addr
,
4675 GEN_INT (-GET_MODE_SIZE (SImode
)),
4676 NULL_RTX
, 1, OPTAB_DIRECT
);
4678 ac
->memsi
= gen_rtx_MEM (SImode
, align
);
4679 MEM_VOLATILE_P (ac
->memsi
) = MEM_VOLATILE_P (mem
);
4680 set_mem_alias_set (ac
->memsi
, ALIAS_SET_MEMORY_BARRIER
);
4681 set_mem_align (ac
->memsi
, GET_MODE_BITSIZE (SImode
));
4683 /* Calculate shiftcount. */
4684 byteoffset
= expand_simple_binop (Pmode
, AND
, addr
,
4685 GEN_INT (GET_MODE_SIZE (SImode
) - 1),
4686 NULL_RTX
, 1, OPTAB_DIRECT
);
4687 /* As we already have some offset, evaluate the remaining distance. */
4688 ac
->shift
= expand_simple_binop (SImode
, MINUS
, ac
->shift
, byteoffset
,
4689 NULL_RTX
, 1, OPTAB_DIRECT
);
4692 /* Shift is the byte count, but we need the bitcount. */
4693 ac
->shift
= expand_simple_binop (SImode
, MULT
, ac
->shift
, GEN_INT (BITS_PER_UNIT
),
4694 NULL_RTX
, 1, OPTAB_DIRECT
);
4695 /* Calculate masks. */
4696 ac
->modemask
= expand_simple_binop (SImode
, ASHIFT
,
4697 GEN_INT (GET_MODE_MASK (mode
)), ac
->shift
,
4698 NULL_RTX
, 1, OPTAB_DIRECT
);
4699 ac
->modemaski
= expand_simple_unop (SImode
, NOT
, ac
->modemask
, NULL_RTX
, 1);
4702 /* Expand an atomic compare and swap operation for HImode and QImode. MEM is
4703 the memory location, CMP the old value to compare MEM with and NEW_RTX the value
4704 to set if CMP == MEM.
4705 CMP is never in memory for compare_and_swap_cc because
4706 expand_bool_compare_and_swap puts it into a register for later compare. */
4709 s390_expand_cs_hqi (enum machine_mode mode
, rtx target
, rtx mem
, rtx cmp
, rtx new_rtx
)
4711 struct alignment_context ac
;
4712 rtx cmpv
, newv
, val
, resv
, cc
;
4713 rtx res
= gen_reg_rtx (SImode
);
4714 rtx csloop
= gen_label_rtx ();
4715 rtx csend
= gen_label_rtx ();
4717 gcc_assert (register_operand (target
, VOIDmode
));
4718 gcc_assert (MEM_P (mem
));
4720 init_alignment_context (&ac
, mem
, mode
);
4722 /* Shift the values to the correct bit positions. */
4723 if (!(ac
.aligned
&& MEM_P (cmp
)))
4724 cmp
= s390_expand_mask_and_shift (cmp
, mode
, ac
.shift
);
4725 if (!(ac
.aligned
&& MEM_P (new_rtx
)))
4726 new_rtx
= s390_expand_mask_and_shift (new_rtx
, mode
, ac
.shift
);
4728 /* Load full word. Subsequent loads are performed by CS. */
4729 val
= expand_simple_binop (SImode
, AND
, ac
.memsi
, ac
.modemaski
,
4730 NULL_RTX
, 1, OPTAB_DIRECT
);
4732 /* Start CS loop. */
4733 emit_label (csloop
);
4734 /* val = "<mem>00..0<mem>"
4735 * cmp = "00..0<cmp>00..0"
4736 * new = "00..0<new>00..0"
4739 /* Patch cmp and new with val at correct position. */
4740 if (ac
.aligned
&& MEM_P (cmp
))
4742 cmpv
= force_reg (SImode
, val
);
4743 store_bit_field (cmpv
, GET_MODE_BITSIZE (mode
), 0,
4747 cmpv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, cmp
, val
,
4748 NULL_RTX
, 1, OPTAB_DIRECT
));
4749 if (ac
.aligned
&& MEM_P (new_rtx
))
4751 newv
= force_reg (SImode
, val
);
4752 store_bit_field (newv
, GET_MODE_BITSIZE (mode
), 0,
4753 0, 0, SImode
, new_rtx
);
4756 newv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4757 NULL_RTX
, 1, OPTAB_DIRECT
));
4759 /* Jump to end if we're done (likely?). */
4760 s390_emit_jump (csend
, s390_emit_compare_and_swap (EQ
, res
, ac
.memsi
,
4763 /* Check for changes outside mode. */
4764 resv
= expand_simple_binop (SImode
, AND
, res
, ac
.modemaski
,
4765 NULL_RTX
, 1, OPTAB_DIRECT
);
4766 cc
= s390_emit_compare (NE
, resv
, val
);
4767 emit_move_insn (val
, resv
);
4768 /* Loop internal if so. */
4769 s390_emit_jump (csloop
, cc
);
4773 /* Return the correct part of the bitfield. */
4774 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
, res
, ac
.shift
,
4775 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4778 /* Expand an atomic operation CODE of mode MODE. MEM is the memory location
4779 and VAL the value to play with. If AFTER is true then store the value
4780 MEM holds after the operation, if AFTER is false then store the value MEM
4781 holds before the operation. If TARGET is zero then discard that value, else
4782 store it to TARGET. */
4785 s390_expand_atomic (enum machine_mode mode
, enum rtx_code code
,
4786 rtx target
, rtx mem
, rtx val
, bool after
)
4788 struct alignment_context ac
;
4790 rtx new_rtx
= gen_reg_rtx (SImode
);
4791 rtx orig
= gen_reg_rtx (SImode
);
4792 rtx csloop
= gen_label_rtx ();
4794 gcc_assert (!target
|| register_operand (target
, VOIDmode
));
4795 gcc_assert (MEM_P (mem
));
4797 init_alignment_context (&ac
, mem
, mode
);
4799 /* Shift val to the correct bit positions.
4800 Preserve "icm", but prevent "ex icm". */
4801 if (!(ac
.aligned
&& code
== SET
&& MEM_P (val
)))
4802 val
= s390_expand_mask_and_shift (val
, mode
, ac
.shift
);
4804 /* Further preparation insns. */
4805 if (code
== PLUS
|| code
== MINUS
)
4806 emit_move_insn (orig
, val
);
4807 else if (code
== MULT
|| code
== AND
) /* val = "11..1<val>11..1" */
4808 val
= expand_simple_binop (SImode
, XOR
, val
, ac
.modemaski
,
4809 NULL_RTX
, 1, OPTAB_DIRECT
);
4811 /* Load full word. Subsequent loads are performed by CS. */
4812 cmp
= force_reg (SImode
, ac
.memsi
);
4814 /* Start CS loop. */
4815 emit_label (csloop
);
4816 emit_move_insn (new_rtx
, cmp
);
4818 /* Patch new with val at correct position. */
4823 val
= expand_simple_binop (SImode
, code
, new_rtx
, orig
,
4824 NULL_RTX
, 1, OPTAB_DIRECT
);
4825 val
= expand_simple_binop (SImode
, AND
, val
, ac
.modemask
,
4826 NULL_RTX
, 1, OPTAB_DIRECT
);
4829 if (ac
.aligned
&& MEM_P (val
))
4830 store_bit_field (new_rtx
, GET_MODE_BITSIZE (mode
), 0,
4834 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, ac
.modemaski
,
4835 NULL_RTX
, 1, OPTAB_DIRECT
);
4836 new_rtx
= expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4837 NULL_RTX
, 1, OPTAB_DIRECT
);
4843 new_rtx
= expand_simple_binop (SImode
, code
, new_rtx
, val
,
4844 NULL_RTX
, 1, OPTAB_DIRECT
);
4846 case MULT
: /* NAND */
4847 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, val
,
4848 NULL_RTX
, 1, OPTAB_DIRECT
);
4849 new_rtx
= expand_simple_binop (SImode
, XOR
, new_rtx
, ac
.modemask
,
4850 NULL_RTX
, 1, OPTAB_DIRECT
);
4856 s390_emit_jump (csloop
, s390_emit_compare_and_swap (NE
, cmp
,
4857 ac
.memsi
, cmp
, new_rtx
));
4859 /* Return the correct part of the bitfield. */
4861 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
,
4862 after
? new_rtx
: cmp
, ac
.shift
,
4863 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4866 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
4867 We need to emit DTP-relative relocations. */
4869 static void s390_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
4872 s390_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
4877 fputs ("\t.long\t", file
);
4880 fputs ("\t.quad\t", file
);
4885 output_addr_const (file
, x
);
4886 fputs ("@DTPOFF", file
);
4889 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
4890 /* Implement TARGET_MANGLE_TYPE. */
4893 s390_mangle_type (const_tree type
)
4895 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
4896 && TARGET_LONG_DOUBLE_128
)
4899 /* For all other types, use normal C++ mangling. */
4904 /* In the name of slightly smaller debug output, and to cater to
4905 general assembler lossage, recognize various UNSPEC sequences
4906 and turn them back into a direct symbol reference. */
4909 s390_delegitimize_address (rtx orig_x
)
4913 orig_x
= delegitimize_mem_from_attrs (orig_x
);
4916 /* Extract the symbol ref from:
4917 (plus:SI (reg:SI 12 %r12)
4918 (const:SI (unspec:SI [(symbol_ref/f:SI ("*.LC0"))]
4919 UNSPEC_GOTOFF/PLTOFF)))
4921 (plus:SI (reg:SI 12 %r12)
4922 (const:SI (plus:SI (unspec:SI [(symbol_ref:SI ("L"))]
4923 UNSPEC_GOTOFF/PLTOFF)
4924 (const_int 4 [0x4])))) */
4925 if (GET_CODE (x
) == PLUS
4926 && REG_P (XEXP (x
, 0))
4927 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
4928 && GET_CODE (XEXP (x
, 1)) == CONST
)
4930 HOST_WIDE_INT offset
= 0;
4932 /* The const operand. */
4933 y
= XEXP (XEXP (x
, 1), 0);
4935 if (GET_CODE (y
) == PLUS
4936 && GET_CODE (XEXP (y
, 1)) == CONST_INT
)
4938 offset
= INTVAL (XEXP (y
, 1));
4942 if (GET_CODE (y
) == UNSPEC
4943 && (XINT (y
, 1) == UNSPEC_GOTOFF
4944 || XINT (y
, 1) == UNSPEC_PLTOFF
))
4945 return plus_constant (Pmode
, XVECEXP (y
, 0, 0), offset
);
4948 if (GET_CODE (x
) != MEM
)
4952 if (GET_CODE (x
) == PLUS
4953 && GET_CODE (XEXP (x
, 1)) == CONST
4954 && GET_CODE (XEXP (x
, 0)) == REG
4955 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
4957 y
= XEXP (XEXP (x
, 1), 0);
4958 if (GET_CODE (y
) == UNSPEC
4959 && XINT (y
, 1) == UNSPEC_GOT
)
4960 y
= XVECEXP (y
, 0, 0);
4964 else if (GET_CODE (x
) == CONST
)
4966 /* Extract the symbol ref from:
4967 (mem:QI (const:DI (unspec:DI [(symbol_ref:DI ("foo"))]
4968 UNSPEC_PLT/GOTENT))) */
4971 if (GET_CODE (y
) == UNSPEC
4972 && (XINT (y
, 1) == UNSPEC_GOTENT
4973 || XINT (y
, 1) == UNSPEC_PLT
))
4974 y
= XVECEXP (y
, 0, 0);
4981 if (GET_MODE (orig_x
) != Pmode
)
4983 if (GET_MODE (orig_x
) == BLKmode
)
4985 y
= lowpart_subreg (GET_MODE (orig_x
), y
, Pmode
);
4992 /* Output operand OP to stdio stream FILE.
4993 OP is an address (register + offset) which is not used to address data;
4994 instead the rightmost bits are interpreted as the value. */
4997 print_shift_count_operand (FILE *file
, rtx op
)
4999 HOST_WIDE_INT offset
;
5002 /* Extract base register and offset. */
5003 if (!s390_decompose_shift_count (op
, &base
, &offset
))
5009 gcc_assert (GET_CODE (base
) == REG
);
5010 gcc_assert (REGNO (base
) < FIRST_PSEUDO_REGISTER
);
5011 gcc_assert (REGNO_REG_CLASS (REGNO (base
)) == ADDR_REGS
);
5014 /* Offsets are constricted to twelve bits. */
5015 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, offset
& ((1 << 12) - 1));
5017 fprintf (file
, "(%s)", reg_names
[REGNO (base
)]);
5020 /* See 'get_some_local_dynamic_name'. */
5023 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
5027 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
5029 x
= get_pool_constant (x
);
5030 return for_each_rtx (&x
, get_some_local_dynamic_name_1
, 0);
5033 if (GET_CODE (x
) == SYMBOL_REF
5034 && tls_symbolic_operand (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
5036 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
5043 /* Locate some local-dynamic symbol still in use by this function
5044 so that we can print its name in local-dynamic base patterns. */
5047 get_some_local_dynamic_name (void)
5051 if (cfun
->machine
->some_ld_name
)
5052 return cfun
->machine
->some_ld_name
;
5054 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5056 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
5057 return cfun
->machine
->some_ld_name
;
5062 /* Output machine-dependent UNSPECs occurring in address constant X
5063 in assembler syntax to stdio stream FILE. Returns true if the
5064 constant X could be recognized, false otherwise. */
5067 s390_output_addr_const_extra (FILE *file
, rtx x
)
5069 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 1)
5070 switch (XINT (x
, 1))
5073 output_addr_const (file
, XVECEXP (x
, 0, 0));
5074 fprintf (file
, "@GOTENT");
5077 output_addr_const (file
, XVECEXP (x
, 0, 0));
5078 fprintf (file
, "@GOT");
5081 output_addr_const (file
, XVECEXP (x
, 0, 0));
5082 fprintf (file
, "@GOTOFF");
5085 output_addr_const (file
, XVECEXP (x
, 0, 0));
5086 fprintf (file
, "@PLT");
5089 output_addr_const (file
, XVECEXP (x
, 0, 0));
5090 fprintf (file
, "@PLTOFF");
5093 output_addr_const (file
, XVECEXP (x
, 0, 0));
5094 fprintf (file
, "@TLSGD");
5097 assemble_name (file
, get_some_local_dynamic_name ());
5098 fprintf (file
, "@TLSLDM");
5101 output_addr_const (file
, XVECEXP (x
, 0, 0));
5102 fprintf (file
, "@DTPOFF");
5105 output_addr_const (file
, XVECEXP (x
, 0, 0));
5106 fprintf (file
, "@NTPOFF");
5108 case UNSPEC_GOTNTPOFF
:
5109 output_addr_const (file
, XVECEXP (x
, 0, 0));
5110 fprintf (file
, "@GOTNTPOFF");
5112 case UNSPEC_INDNTPOFF
:
5113 output_addr_const (file
, XVECEXP (x
, 0, 0));
5114 fprintf (file
, "@INDNTPOFF");
5118 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 2)
5119 switch (XINT (x
, 1))
5121 case UNSPEC_POOL_OFFSET
:
5122 x
= gen_rtx_MINUS (GET_MODE (x
), XVECEXP (x
, 0, 0), XVECEXP (x
, 0, 1));
5123 output_addr_const (file
, x
);
5129 /* Output address operand ADDR in assembler syntax to
5130 stdio stream FILE. */
5133 print_operand_address (FILE *file
, rtx addr
)
5135 struct s390_address ad
;
5137 if (s390_symref_operand_p (addr
, NULL
, NULL
))
5141 output_operand_lossage ("symbolic memory references are "
5142 "only supported on z10 or later");
5145 output_addr_const (file
, addr
);
5149 if (!s390_decompose_address (addr
, &ad
)
5150 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5151 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
5152 output_operand_lossage ("cannot decompose address");
5155 output_addr_const (file
, ad
.disp
);
5157 fprintf (file
, "0");
5159 if (ad
.base
&& ad
.indx
)
5160 fprintf (file
, "(%s,%s)", reg_names
[REGNO (ad
.indx
)],
5161 reg_names
[REGNO (ad
.base
)]);
5163 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5166 /* Output operand X in assembler syntax to stdio stream FILE.
5167 CODE specified the format flag. The following format flags
5170 'C': print opcode suffix for branch condition.
5171 'D': print opcode suffix for inverse branch condition.
5172 'E': print opcode suffix for branch on index instruction.
5173 'J': print tls_load/tls_gdcall/tls_ldcall suffix
5174 'G': print the size of the operand in bytes.
5175 'O': print only the displacement of a memory reference.
5176 'R': print only the base register of a memory reference.
5177 'S': print S-type memory reference (base+displacement).
5178 'N': print the second word of a DImode operand.
5179 'M': print the second word of a TImode operand.
5180 'Y': print shift count operand.
5182 'b': print integer X as if it's an unsigned byte.
5183 'c': print integer X as if it's an signed byte.
5184 'x': print integer X as if it's an unsigned halfword.
5185 'h': print integer X as if it's a signed halfword.
5186 'i': print the first nonzero HImode part of X.
5187 'j': print the first HImode part unequal to -1 of X.
5188 'k': print the first nonzero SImode part of X.
5189 'm': print the first SImode part unequal to -1 of X.
5190 'o': print integer X as if it's an unsigned 32bit word. */
5193 print_operand (FILE *file
, rtx x
, int code
)
5198 fprintf (file
, s390_branch_condition_mnemonic (x
, FALSE
));
5202 fprintf (file
, s390_branch_condition_mnemonic (x
, TRUE
));
5206 if (GET_CODE (x
) == LE
)
5207 fprintf (file
, "l");
5208 else if (GET_CODE (x
) == GT
)
5209 fprintf (file
, "h");
5211 output_operand_lossage ("invalid comparison operator "
5212 "for 'E' output modifier");
5216 if (GET_CODE (x
) == SYMBOL_REF
)
5218 fprintf (file
, "%s", ":tls_load:");
5219 output_addr_const (file
, x
);
5221 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD
)
5223 fprintf (file
, "%s", ":tls_gdcall:");
5224 output_addr_const (file
, XVECEXP (x
, 0, 0));
5226 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM
)
5228 fprintf (file
, "%s", ":tls_ldcall:");
5229 assemble_name (file
, get_some_local_dynamic_name ());
5232 output_operand_lossage ("invalid reference for 'J' output modifier");
5236 fprintf (file
, "%u", GET_MODE_SIZE (GET_MODE (x
)));
5241 struct s390_address ad
;
5246 output_operand_lossage ("memory reference expected for "
5247 "'O' output modifier");
5251 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5254 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5257 output_operand_lossage ("invalid address for 'O' output modifier");
5262 output_addr_const (file
, ad
.disp
);
5264 fprintf (file
, "0");
5270 struct s390_address ad
;
5275 output_operand_lossage ("memory reference expected for "
5276 "'R' output modifier");
5280 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5283 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5286 output_operand_lossage ("invalid address for 'R' output modifier");
5291 fprintf (file
, "%s", reg_names
[REGNO (ad
.base
)]);
5293 fprintf (file
, "0");
5299 struct s390_address ad
;
5304 output_operand_lossage ("memory reference expected for "
5305 "'S' output modifier");
5308 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5311 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5314 output_operand_lossage ("invalid address for 'S' output modifier");
5319 output_addr_const (file
, ad
.disp
);
5321 fprintf (file
, "0");
5324 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5329 if (GET_CODE (x
) == REG
)
5330 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5331 else if (GET_CODE (x
) == MEM
)
5332 x
= change_address (x
, VOIDmode
,
5333 plus_constant (Pmode
, XEXP (x
, 0), 4));
5335 output_operand_lossage ("register or memory expression expected "
5336 "for 'N' output modifier");
5340 if (GET_CODE (x
) == REG
)
5341 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5342 else if (GET_CODE (x
) == MEM
)
5343 x
= change_address (x
, VOIDmode
,
5344 plus_constant (Pmode
, XEXP (x
, 0), 8));
5346 output_operand_lossage ("register or memory expression expected "
5347 "for 'M' output modifier");
5351 print_shift_count_operand (file
, x
);
5355 switch (GET_CODE (x
))
5358 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5362 output_address (XEXP (x
, 0));
5369 output_addr_const (file
, x
);
5374 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xff);
5375 else if (code
== 'c')
5376 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xff) ^ 0x80) - 0x80);
5377 else if (code
== 'x')
5378 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffff);
5379 else if (code
== 'h')
5380 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xffff) ^ 0x8000) - 0x8000);
5381 else if (code
== 'i')
5382 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5383 s390_extract_part (x
, HImode
, 0));
5384 else if (code
== 'j')
5385 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5386 s390_extract_part (x
, HImode
, -1));
5387 else if (code
== 'k')
5388 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5389 s390_extract_part (x
, SImode
, 0));
5390 else if (code
== 'm')
5391 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5392 s390_extract_part (x
, SImode
, -1));
5393 else if (code
== 'o')
5394 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffffffff);
5396 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5400 gcc_assert (GET_MODE (x
) == VOIDmode
);
5402 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xff);
5403 else if (code
== 'x')
5404 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xffff);
5405 else if (code
== 'h')
5406 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5407 ((CONST_DOUBLE_LOW (x
) & 0xffff) ^ 0x8000) - 0x8000);
5411 output_operand_lossage ("invalid constant - try using "
5412 "an output modifier");
5414 output_operand_lossage ("invalid constant for output modifier '%c'",
5421 output_operand_lossage ("invalid expression - try using "
5422 "an output modifier");
5424 output_operand_lossage ("invalid expression for output "
5425 "modifier '%c'", code
);
5430 /* Target hook for assembling integer objects. We need to define it
5431 here to work a round a bug in some versions of GAS, which couldn't
5432 handle values smaller than INT_MIN when printed in decimal. */
5435 s390_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
5437 if (size
== 8 && aligned_p
5438 && GET_CODE (x
) == CONST_INT
&& INTVAL (x
) < INT_MIN
)
5440 fprintf (asm_out_file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n",
5444 return default_assemble_integer (x
, size
, aligned_p
);
5447 /* Returns true if register REGNO is used for forming
5448 a memory address in expression X. */
5451 reg_used_in_mem_p (int regno
, rtx x
)
5453 enum rtx_code code
= GET_CODE (x
);
5459 if (refers_to_regno_p (regno
, regno
+1,
5463 else if (code
== SET
5464 && GET_CODE (SET_DEST (x
)) == PC
)
5466 if (refers_to_regno_p (regno
, regno
+1,
5471 fmt
= GET_RTX_FORMAT (code
);
5472 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
5475 && reg_used_in_mem_p (regno
, XEXP (x
, i
)))
5478 else if (fmt
[i
] == 'E')
5479 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5480 if (reg_used_in_mem_p (regno
, XVECEXP (x
, i
, j
)))
5486 /* Returns true if expression DEP_RTX sets an address register
5487 used by instruction INSN to address memory. */
5490 addr_generation_dependency_p (rtx dep_rtx
, rtx insn
)
5494 if (GET_CODE (dep_rtx
) == INSN
)
5495 dep_rtx
= PATTERN (dep_rtx
);
5497 if (GET_CODE (dep_rtx
) == SET
)
5499 target
= SET_DEST (dep_rtx
);
5500 if (GET_CODE (target
) == STRICT_LOW_PART
)
5501 target
= XEXP (target
, 0);
5502 while (GET_CODE (target
) == SUBREG
)
5503 target
= SUBREG_REG (target
);
5505 if (GET_CODE (target
) == REG
)
5507 int regno
= REGNO (target
);
5509 if (s390_safe_attr_type (insn
) == TYPE_LA
)
5511 pat
= PATTERN (insn
);
5512 if (GET_CODE (pat
) == PARALLEL
)
5514 gcc_assert (XVECLEN (pat
, 0) == 2);
5515 pat
= XVECEXP (pat
, 0, 0);
5517 gcc_assert (GET_CODE (pat
) == SET
);
5518 return refers_to_regno_p (regno
, regno
+1, SET_SRC (pat
), 0);
5520 else if (get_attr_atype (insn
) == ATYPE_AGEN
)
5521 return reg_used_in_mem_p (regno
, PATTERN (insn
));
5527 /* Return 1, if dep_insn sets register used in insn in the agen unit. */
5530 s390_agen_dep_p (rtx dep_insn
, rtx insn
)
5532 rtx dep_rtx
= PATTERN (dep_insn
);
5535 if (GET_CODE (dep_rtx
) == SET
5536 && addr_generation_dependency_p (dep_rtx
, insn
))
5538 else if (GET_CODE (dep_rtx
) == PARALLEL
)
5540 for (i
= 0; i
< XVECLEN (dep_rtx
, 0); i
++)
5542 if (addr_generation_dependency_p (XVECEXP (dep_rtx
, 0, i
), insn
))
5550 /* A C statement (sans semicolon) to update the integer scheduling priority
5551 INSN_PRIORITY (INSN). Increase the priority to execute the INSN earlier,
5552 reduce the priority to execute INSN later. Do not define this macro if
5553 you do not need to adjust the scheduling priorities of insns.
5555 A STD instruction should be scheduled earlier,
5556 in order to use the bypass. */
5558 s390_adjust_priority (rtx insn ATTRIBUTE_UNUSED
, int priority
)
5560 if (! INSN_P (insn
))
5563 if (s390_tune
!= PROCESSOR_2084_Z990
5564 && s390_tune
!= PROCESSOR_2094_Z9_109
5565 && s390_tune
!= PROCESSOR_2097_Z10
5566 && s390_tune
!= PROCESSOR_2817_Z196
)
5569 switch (s390_safe_attr_type (insn
))
5573 priority
= priority
<< 3;
5577 priority
= priority
<< 1;
5586 /* The number of instructions that can be issued per cycle. */
5589 s390_issue_rate (void)
5593 case PROCESSOR_2084_Z990
:
5594 case PROCESSOR_2094_Z9_109
:
5595 case PROCESSOR_2817_Z196
:
5597 case PROCESSOR_2097_Z10
:
5605 s390_first_cycle_multipass_dfa_lookahead (void)
5610 /* Annotate every literal pool reference in X by an UNSPEC_LTREF expression.
5611 Fix up MEMs as required. */
5614 annotate_constant_pool_refs (rtx
*x
)
5619 gcc_assert (GET_CODE (*x
) != SYMBOL_REF
5620 || !CONSTANT_POOL_ADDRESS_P (*x
));
5622 /* Literal pool references can only occur inside a MEM ... */
5623 if (GET_CODE (*x
) == MEM
)
5625 rtx memref
= XEXP (*x
, 0);
5627 if (GET_CODE (memref
) == SYMBOL_REF
5628 && CONSTANT_POOL_ADDRESS_P (memref
))
5630 rtx base
= cfun
->machine
->base_reg
;
5631 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, memref
, base
),
5634 *x
= replace_equiv_address (*x
, addr
);
5638 if (GET_CODE (memref
) == CONST
5639 && GET_CODE (XEXP (memref
, 0)) == PLUS
5640 && GET_CODE (XEXP (XEXP (memref
, 0), 1)) == CONST_INT
5641 && GET_CODE (XEXP (XEXP (memref
, 0), 0)) == SYMBOL_REF
5642 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (memref
, 0), 0)))
5644 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (memref
, 0), 1));
5645 rtx sym
= XEXP (XEXP (memref
, 0), 0);
5646 rtx base
= cfun
->machine
->base_reg
;
5647 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5650 *x
= replace_equiv_address (*x
, plus_constant (Pmode
, addr
, off
));
5655 /* ... or a load-address type pattern. */
5656 if (GET_CODE (*x
) == SET
)
5658 rtx addrref
= SET_SRC (*x
);
5660 if (GET_CODE (addrref
) == SYMBOL_REF
5661 && CONSTANT_POOL_ADDRESS_P (addrref
))
5663 rtx base
= cfun
->machine
->base_reg
;
5664 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, addrref
, base
),
5667 SET_SRC (*x
) = addr
;
5671 if (GET_CODE (addrref
) == CONST
5672 && GET_CODE (XEXP (addrref
, 0)) == PLUS
5673 && GET_CODE (XEXP (XEXP (addrref
, 0), 1)) == CONST_INT
5674 && GET_CODE (XEXP (XEXP (addrref
, 0), 0)) == SYMBOL_REF
5675 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (addrref
, 0), 0)))
5677 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (addrref
, 0), 1));
5678 rtx sym
= XEXP (XEXP (addrref
, 0), 0);
5679 rtx base
= cfun
->machine
->base_reg
;
5680 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5683 SET_SRC (*x
) = plus_constant (Pmode
, addr
, off
);
5688 /* Annotate LTREL_BASE as well. */
5689 if (GET_CODE (*x
) == UNSPEC
5690 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5692 rtx base
= cfun
->machine
->base_reg
;
5693 *x
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XVECEXP (*x
, 0, 0), base
),
5698 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5699 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5703 annotate_constant_pool_refs (&XEXP (*x
, i
));
5705 else if (fmt
[i
] == 'E')
5707 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5708 annotate_constant_pool_refs (&XVECEXP (*x
, i
, j
));
5713 /* Split all branches that exceed the maximum distance.
5714 Returns true if this created a new literal pool entry. */
5717 s390_split_branches (void)
5719 rtx temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
5720 int new_literal
= 0, ret
;
5721 rtx insn
, pat
, tmp
, target
;
5724 /* We need correct insn addresses. */
5726 shorten_branches (get_insns ());
5728 /* Find all branches that exceed 64KB, and split them. */
5730 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5732 if (GET_CODE (insn
) != JUMP_INSN
)
5735 pat
= PATTERN (insn
);
5736 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
5737 pat
= XVECEXP (pat
, 0, 0);
5738 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
5741 if (GET_CODE (SET_SRC (pat
)) == LABEL_REF
)
5743 label
= &SET_SRC (pat
);
5745 else if (GET_CODE (SET_SRC (pat
)) == IF_THEN_ELSE
)
5747 if (GET_CODE (XEXP (SET_SRC (pat
), 1)) == LABEL_REF
)
5748 label
= &XEXP (SET_SRC (pat
), 1);
5749 else if (GET_CODE (XEXP (SET_SRC (pat
), 2)) == LABEL_REF
)
5750 label
= &XEXP (SET_SRC (pat
), 2);
5757 if (get_attr_length (insn
) <= 4)
5760 /* We are going to use the return register as scratch register,
5761 make sure it will be saved/restored by the prologue/epilogue. */
5762 cfun_frame_layout
.save_return_addr_p
= 1;
5767 tmp
= force_const_mem (Pmode
, *label
);
5768 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, tmp
), insn
);
5769 INSN_ADDRESSES_NEW (tmp
, -1);
5770 annotate_constant_pool_refs (&PATTERN (tmp
));
5777 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, *label
),
5778 UNSPEC_LTREL_OFFSET
);
5779 target
= gen_rtx_CONST (Pmode
, target
);
5780 target
= force_const_mem (Pmode
, target
);
5781 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, target
), insn
);
5782 INSN_ADDRESSES_NEW (tmp
, -1);
5783 annotate_constant_pool_refs (&PATTERN (tmp
));
5785 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XEXP (target
, 0),
5786 cfun
->machine
->base_reg
),
5788 target
= gen_rtx_PLUS (Pmode
, temp_reg
, target
);
5791 ret
= validate_change (insn
, label
, target
, 0);
5799 /* Find an annotated literal pool symbol referenced in RTX X,
5800 and store it at REF. Will abort if X contains references to
5801 more than one such pool symbol; multiple references to the same
5802 symbol are allowed, however.
5804 The rtx pointed to by REF must be initialized to NULL_RTX
5805 by the caller before calling this routine. */
5808 find_constant_pool_ref (rtx x
, rtx
*ref
)
5813 /* Ignore LTREL_BASE references. */
5814 if (GET_CODE (x
) == UNSPEC
5815 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5817 /* Likewise POOL_ENTRY insns. */
5818 if (GET_CODE (x
) == UNSPEC_VOLATILE
5819 && XINT (x
, 1) == UNSPECV_POOL_ENTRY
)
5822 gcc_assert (GET_CODE (x
) != SYMBOL_REF
5823 || !CONSTANT_POOL_ADDRESS_P (x
));
5825 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_LTREF
)
5827 rtx sym
= XVECEXP (x
, 0, 0);
5828 gcc_assert (GET_CODE (sym
) == SYMBOL_REF
5829 && CONSTANT_POOL_ADDRESS_P (sym
));
5831 if (*ref
== NULL_RTX
)
5834 gcc_assert (*ref
== sym
);
5839 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5840 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5844 find_constant_pool_ref (XEXP (x
, i
), ref
);
5846 else if (fmt
[i
] == 'E')
5848 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5849 find_constant_pool_ref (XVECEXP (x
, i
, j
), ref
);
5854 /* Replace every reference to the annotated literal pool
5855 symbol REF in X by its base plus OFFSET. */
5858 replace_constant_pool_ref (rtx
*x
, rtx ref
, rtx offset
)
5863 gcc_assert (*x
!= ref
);
5865 if (GET_CODE (*x
) == UNSPEC
5866 && XINT (*x
, 1) == UNSPEC_LTREF
5867 && XVECEXP (*x
, 0, 0) == ref
)
5869 *x
= gen_rtx_PLUS (Pmode
, XVECEXP (*x
, 0, 1), offset
);
5873 if (GET_CODE (*x
) == PLUS
5874 && GET_CODE (XEXP (*x
, 1)) == CONST_INT
5875 && GET_CODE (XEXP (*x
, 0)) == UNSPEC
5876 && XINT (XEXP (*x
, 0), 1) == UNSPEC_LTREF
5877 && XVECEXP (XEXP (*x
, 0), 0, 0) == ref
)
5879 rtx addr
= gen_rtx_PLUS (Pmode
, XVECEXP (XEXP (*x
, 0), 0, 1), offset
);
5880 *x
= plus_constant (Pmode
, addr
, INTVAL (XEXP (*x
, 1)));
5884 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5885 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5889 replace_constant_pool_ref (&XEXP (*x
, i
), ref
, offset
);
5891 else if (fmt
[i
] == 'E')
5893 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5894 replace_constant_pool_ref (&XVECEXP (*x
, i
, j
), ref
, offset
);
5899 /* Check whether X contains an UNSPEC_LTREL_BASE.
5900 Return its constant pool symbol if found, NULL_RTX otherwise. */
5903 find_ltrel_base (rtx x
)
5908 if (GET_CODE (x
) == UNSPEC
5909 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5910 return XVECEXP (x
, 0, 0);
5912 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5913 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5917 rtx fnd
= find_ltrel_base (XEXP (x
, i
));
5921 else if (fmt
[i
] == 'E')
5923 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5925 rtx fnd
= find_ltrel_base (XVECEXP (x
, i
, j
));
5935 /* Replace any occurrence of UNSPEC_LTREL_BASE in X with its base. */
5938 replace_ltrel_base (rtx
*x
)
5943 if (GET_CODE (*x
) == UNSPEC
5944 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5946 *x
= XVECEXP (*x
, 0, 1);
5950 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5951 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5955 replace_ltrel_base (&XEXP (*x
, i
));
5957 else if (fmt
[i
] == 'E')
5959 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5960 replace_ltrel_base (&XVECEXP (*x
, i
, j
));
5966 /* We keep a list of constants which we have to add to internal
5967 constant tables in the middle of large functions. */
5969 #define NR_C_MODES 11
5970 enum machine_mode constant_modes
[NR_C_MODES
] =
5972 TFmode
, TImode
, TDmode
,
5973 DFmode
, DImode
, DDmode
,
5974 SFmode
, SImode
, SDmode
,
5981 struct constant
*next
;
5986 struct constant_pool
5988 struct constant_pool
*next
;
5992 rtx emit_pool_after
;
5994 struct constant
*constants
[NR_C_MODES
];
5995 struct constant
*execute
;
6000 /* Allocate new constant_pool structure. */
6002 static struct constant_pool
*
6003 s390_alloc_pool (void)
6005 struct constant_pool
*pool
;
6008 pool
= (struct constant_pool
*) xmalloc (sizeof *pool
);
6010 for (i
= 0; i
< NR_C_MODES
; i
++)
6011 pool
->constants
[i
] = NULL
;
6013 pool
->execute
= NULL
;
6014 pool
->label
= gen_label_rtx ();
6015 pool
->first_insn
= NULL_RTX
;
6016 pool
->pool_insn
= NULL_RTX
;
6017 pool
->insns
= BITMAP_ALLOC (NULL
);
6019 pool
->emit_pool_after
= NULL_RTX
;
6024 /* Create new constant pool covering instructions starting at INSN
6025 and chain it to the end of POOL_LIST. */
6027 static struct constant_pool
*
6028 s390_start_pool (struct constant_pool
**pool_list
, rtx insn
)
6030 struct constant_pool
*pool
, **prev
;
6032 pool
= s390_alloc_pool ();
6033 pool
->first_insn
= insn
;
6035 for (prev
= pool_list
; *prev
; prev
= &(*prev
)->next
)
6042 /* End range of instructions covered by POOL at INSN and emit
6043 placeholder insn representing the pool. */
6046 s390_end_pool (struct constant_pool
*pool
, rtx insn
)
6048 rtx pool_size
= GEN_INT (pool
->size
+ 8 /* alignment slop */);
6051 insn
= get_last_insn ();
6053 pool
->pool_insn
= emit_insn_after (gen_pool (pool_size
), insn
);
6054 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6057 /* Add INSN to the list of insns covered by POOL. */
6060 s390_add_pool_insn (struct constant_pool
*pool
, rtx insn
)
6062 bitmap_set_bit (pool
->insns
, INSN_UID (insn
));
6065 /* Return pool out of POOL_LIST that covers INSN. */
6067 static struct constant_pool
*
6068 s390_find_pool (struct constant_pool
*pool_list
, rtx insn
)
6070 struct constant_pool
*pool
;
6072 for (pool
= pool_list
; pool
; pool
= pool
->next
)
6073 if (bitmap_bit_p (pool
->insns
, INSN_UID (insn
)))
6079 /* Add constant VAL of mode MODE to the constant pool POOL. */
6082 s390_add_constant (struct constant_pool
*pool
, rtx val
, enum machine_mode mode
)
6087 for (i
= 0; i
< NR_C_MODES
; i
++)
6088 if (constant_modes
[i
] == mode
)
6090 gcc_assert (i
!= NR_C_MODES
);
6092 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6093 if (rtx_equal_p (val
, c
->value
))
6098 c
= (struct constant
*) xmalloc (sizeof *c
);
6100 c
->label
= gen_label_rtx ();
6101 c
->next
= pool
->constants
[i
];
6102 pool
->constants
[i
] = c
;
6103 pool
->size
+= GET_MODE_SIZE (mode
);
6107 /* Return an rtx that represents the offset of X from the start of
6111 s390_pool_offset (struct constant_pool
*pool
, rtx x
)
6115 label
= gen_rtx_LABEL_REF (GET_MODE (x
), pool
->label
);
6116 x
= gen_rtx_UNSPEC (GET_MODE (x
), gen_rtvec (2, x
, label
),
6117 UNSPEC_POOL_OFFSET
);
6118 return gen_rtx_CONST (GET_MODE (x
), x
);
6121 /* Find constant VAL of mode MODE in the constant pool POOL.
6122 Return an RTX describing the distance from the start of
6123 the pool to the location of the new constant. */
6126 s390_find_constant (struct constant_pool
*pool
, rtx val
,
6127 enum machine_mode mode
)
6132 for (i
= 0; i
< NR_C_MODES
; i
++)
6133 if (constant_modes
[i
] == mode
)
6135 gcc_assert (i
!= NR_C_MODES
);
6137 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6138 if (rtx_equal_p (val
, c
->value
))
6143 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6146 /* Check whether INSN is an execute. Return the label_ref to its
6147 execute target template if so, NULL_RTX otherwise. */
6150 s390_execute_label (rtx insn
)
6152 if (GET_CODE (insn
) == INSN
6153 && GET_CODE (PATTERN (insn
)) == PARALLEL
6154 && GET_CODE (XVECEXP (PATTERN (insn
), 0, 0)) == UNSPEC
6155 && XINT (XVECEXP (PATTERN (insn
), 0, 0), 1) == UNSPEC_EXECUTE
)
6156 return XVECEXP (XVECEXP (PATTERN (insn
), 0, 0), 0, 2);
6161 /* Add execute target for INSN to the constant pool POOL. */
6164 s390_add_execute (struct constant_pool
*pool
, rtx insn
)
6168 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6169 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6174 c
= (struct constant
*) xmalloc (sizeof *c
);
6176 c
->label
= gen_label_rtx ();
6177 c
->next
= pool
->execute
;
6183 /* Find execute target for INSN in the constant pool POOL.
6184 Return an RTX describing the distance from the start of
6185 the pool to the location of the execute target. */
6188 s390_find_execute (struct constant_pool
*pool
, rtx insn
)
6192 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6193 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6198 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6201 /* For an execute INSN, extract the execute target template. */
6204 s390_execute_target (rtx insn
)
6206 rtx pattern
= PATTERN (insn
);
6207 gcc_assert (s390_execute_label (insn
));
6209 if (XVECLEN (pattern
, 0) == 2)
6211 pattern
= copy_rtx (XVECEXP (pattern
, 0, 1));
6215 rtvec vec
= rtvec_alloc (XVECLEN (pattern
, 0) - 1);
6218 for (i
= 0; i
< XVECLEN (pattern
, 0) - 1; i
++)
6219 RTVEC_ELT (vec
, i
) = copy_rtx (XVECEXP (pattern
, 0, i
+ 1));
6221 pattern
= gen_rtx_PARALLEL (VOIDmode
, vec
);
6227 /* Indicate that INSN cannot be duplicated. This is the case for
6228 execute insns that carry a unique label. */
6231 s390_cannot_copy_insn_p (rtx insn
)
6233 rtx label
= s390_execute_label (insn
);
6234 return label
&& label
!= const0_rtx
;
6237 /* Dump out the constants in POOL. If REMOTE_LABEL is true,
6238 do not emit the pool base label. */
6241 s390_dump_pool (struct constant_pool
*pool
, bool remote_label
)
6244 rtx insn
= pool
->pool_insn
;
6247 /* Switch to rodata section. */
6248 if (TARGET_CPU_ZARCH
)
6250 insn
= emit_insn_after (gen_pool_section_start (), insn
);
6251 INSN_ADDRESSES_NEW (insn
, -1);
6254 /* Ensure minimum pool alignment. */
6255 if (TARGET_CPU_ZARCH
)
6256 insn
= emit_insn_after (gen_pool_align (GEN_INT (8)), insn
);
6258 insn
= emit_insn_after (gen_pool_align (GEN_INT (4)), insn
);
6259 INSN_ADDRESSES_NEW (insn
, -1);
6261 /* Emit pool base label. */
6264 insn
= emit_label_after (pool
->label
, insn
);
6265 INSN_ADDRESSES_NEW (insn
, -1);
6268 /* Dump constants in descending alignment requirement order,
6269 ensuring proper alignment for every constant. */
6270 for (i
= 0; i
< NR_C_MODES
; i
++)
6271 for (c
= pool
->constants
[i
]; c
; c
= c
->next
)
6273 /* Convert UNSPEC_LTREL_OFFSET unspecs to pool-relative references. */
6274 rtx value
= copy_rtx (c
->value
);
6275 if (GET_CODE (value
) == CONST
6276 && GET_CODE (XEXP (value
, 0)) == UNSPEC
6277 && XINT (XEXP (value
, 0), 1) == UNSPEC_LTREL_OFFSET
6278 && XVECLEN (XEXP (value
, 0), 0) == 1)
6279 value
= s390_pool_offset (pool
, XVECEXP (XEXP (value
, 0), 0, 0));
6281 insn
= emit_label_after (c
->label
, insn
);
6282 INSN_ADDRESSES_NEW (insn
, -1);
6284 value
= gen_rtx_UNSPEC_VOLATILE (constant_modes
[i
],
6285 gen_rtvec (1, value
),
6286 UNSPECV_POOL_ENTRY
);
6287 insn
= emit_insn_after (value
, insn
);
6288 INSN_ADDRESSES_NEW (insn
, -1);
6291 /* Ensure minimum alignment for instructions. */
6292 insn
= emit_insn_after (gen_pool_align (GEN_INT (2)), insn
);
6293 INSN_ADDRESSES_NEW (insn
, -1);
6295 /* Output in-pool execute template insns. */
6296 for (c
= pool
->execute
; c
; c
= c
->next
)
6298 insn
= emit_label_after (c
->label
, insn
);
6299 INSN_ADDRESSES_NEW (insn
, -1);
6301 insn
= emit_insn_after (s390_execute_target (c
->value
), insn
);
6302 INSN_ADDRESSES_NEW (insn
, -1);
6305 /* Switch back to previous section. */
6306 if (TARGET_CPU_ZARCH
)
6308 insn
= emit_insn_after (gen_pool_section_end (), insn
);
6309 INSN_ADDRESSES_NEW (insn
, -1);
6312 insn
= emit_barrier_after (insn
);
6313 INSN_ADDRESSES_NEW (insn
, -1);
6315 /* Remove placeholder insn. */
6316 remove_insn (pool
->pool_insn
);
6319 /* Free all memory used by POOL. */
6322 s390_free_pool (struct constant_pool
*pool
)
6324 struct constant
*c
, *next
;
6327 for (i
= 0; i
< NR_C_MODES
; i
++)
6328 for (c
= pool
->constants
[i
]; c
; c
= next
)
6334 for (c
= pool
->execute
; c
; c
= next
)
6340 BITMAP_FREE (pool
->insns
);
6345 /* Collect main literal pool. Return NULL on overflow. */
6347 static struct constant_pool
*
6348 s390_mainpool_start (void)
6350 struct constant_pool
*pool
;
6353 pool
= s390_alloc_pool ();
6355 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6357 if (GET_CODE (insn
) == INSN
6358 && GET_CODE (PATTERN (insn
)) == SET
6359 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC_VOLATILE
6360 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPECV_MAIN_POOL
)
6362 gcc_assert (!pool
->pool_insn
);
6363 pool
->pool_insn
= insn
;
6366 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6368 s390_add_execute (pool
, insn
);
6370 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6372 rtx pool_ref
= NULL_RTX
;
6373 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6376 rtx constant
= get_pool_constant (pool_ref
);
6377 enum machine_mode mode
= get_pool_mode (pool_ref
);
6378 s390_add_constant (pool
, constant
, mode
);
6382 /* If hot/cold partitioning is enabled we have to make sure that
6383 the literal pool is emitted in the same section where the
6384 initialization of the literal pool base pointer takes place.
6385 emit_pool_after is only used in the non-overflow case on non
6386 Z cpus where we can emit the literal pool at the end of the
6387 function body within the text section. */
6389 && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
6390 && !pool
->emit_pool_after
)
6391 pool
->emit_pool_after
= PREV_INSN (insn
);
6394 gcc_assert (pool
->pool_insn
|| pool
->size
== 0);
6396 if (pool
->size
>= 4096)
6398 /* We're going to chunkify the pool, so remove the main
6399 pool placeholder insn. */
6400 remove_insn (pool
->pool_insn
);
6402 s390_free_pool (pool
);
6406 /* If the functions ends with the section where the literal pool
6407 should be emitted set the marker to its end. */
6408 if (pool
&& !pool
->emit_pool_after
)
6409 pool
->emit_pool_after
= get_last_insn ();
6414 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6415 Modify the current function to output the pool constants as well as
6416 the pool register setup instruction. */
6419 s390_mainpool_finish (struct constant_pool
*pool
)
6421 rtx base_reg
= cfun
->machine
->base_reg
;
6424 /* If the pool is empty, we're done. */
6425 if (pool
->size
== 0)
6427 /* We don't actually need a base register after all. */
6428 cfun
->machine
->base_reg
= NULL_RTX
;
6430 if (pool
->pool_insn
)
6431 remove_insn (pool
->pool_insn
);
6432 s390_free_pool (pool
);
6436 /* We need correct insn addresses. */
6437 shorten_branches (get_insns ());
6439 /* On zSeries, we use a LARL to load the pool register. The pool is
6440 located in the .rodata section, so we emit it after the function. */
6441 if (TARGET_CPU_ZARCH
)
6443 insn
= gen_main_base_64 (base_reg
, pool
->label
);
6444 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6445 INSN_ADDRESSES_NEW (insn
, -1);
6446 remove_insn (pool
->pool_insn
);
6448 insn
= get_last_insn ();
6449 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6450 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6452 s390_dump_pool (pool
, 0);
6455 /* On S/390, if the total size of the function's code plus literal pool
6456 does not exceed 4096 bytes, we use BASR to set up a function base
6457 pointer, and emit the literal pool at the end of the function. */
6458 else if (INSN_ADDRESSES (INSN_UID (pool
->emit_pool_after
))
6459 + pool
->size
+ 8 /* alignment slop */ < 4096)
6461 insn
= gen_main_base_31_small (base_reg
, pool
->label
);
6462 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6463 INSN_ADDRESSES_NEW (insn
, -1);
6464 remove_insn (pool
->pool_insn
);
6466 insn
= emit_label_after (pool
->label
, insn
);
6467 INSN_ADDRESSES_NEW (insn
, -1);
6469 /* emit_pool_after will be set by s390_mainpool_start to the
6470 last insn of the section where the literal pool should be
6472 insn
= pool
->emit_pool_after
;
6474 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6475 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6477 s390_dump_pool (pool
, 1);
6480 /* Otherwise, we emit an inline literal pool and use BASR to branch
6481 over it, setting up the pool register at the same time. */
6484 rtx pool_end
= gen_label_rtx ();
6486 insn
= gen_main_base_31_large (base_reg
, pool
->label
, pool_end
);
6487 insn
= emit_jump_insn_after (insn
, pool
->pool_insn
);
6488 JUMP_LABEL (insn
) = pool_end
;
6489 INSN_ADDRESSES_NEW (insn
, -1);
6490 remove_insn (pool
->pool_insn
);
6492 insn
= emit_label_after (pool
->label
, insn
);
6493 INSN_ADDRESSES_NEW (insn
, -1);
6495 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6496 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6498 insn
= emit_label_after (pool_end
, pool
->pool_insn
);
6499 INSN_ADDRESSES_NEW (insn
, -1);
6501 s390_dump_pool (pool
, 1);
6505 /* Replace all literal pool references. */
6507 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6510 replace_ltrel_base (&PATTERN (insn
));
6512 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6514 rtx addr
, pool_ref
= NULL_RTX
;
6515 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6518 if (s390_execute_label (insn
))
6519 addr
= s390_find_execute (pool
, insn
);
6521 addr
= s390_find_constant (pool
, get_pool_constant (pool_ref
),
6522 get_pool_mode (pool_ref
));
6524 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6525 INSN_CODE (insn
) = -1;
6531 /* Free the pool. */
6532 s390_free_pool (pool
);
6535 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6536 We have decided we cannot use this pool, so revert all changes
6537 to the current function that were done by s390_mainpool_start. */
6539 s390_mainpool_cancel (struct constant_pool
*pool
)
6541 /* We didn't actually change the instruction stream, so simply
6542 free the pool memory. */
6543 s390_free_pool (pool
);
6547 /* Chunkify the literal pool. */
6549 #define S390_POOL_CHUNK_MIN 0xc00
6550 #define S390_POOL_CHUNK_MAX 0xe00
6552 static struct constant_pool
*
6553 s390_chunkify_start (void)
6555 struct constant_pool
*curr_pool
= NULL
, *pool_list
= NULL
;
6558 rtx pending_ltrel
= NULL_RTX
;
6561 rtx (*gen_reload_base
) (rtx
, rtx
) =
6562 TARGET_CPU_ZARCH
? gen_reload_base_64
: gen_reload_base_31
;
6565 /* We need correct insn addresses. */
6567 shorten_branches (get_insns ());
6569 /* Scan all insns and move literals to pool chunks. */
6571 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6573 bool section_switch_p
= false;
6575 /* Check for pending LTREL_BASE. */
6578 rtx ltrel_base
= find_ltrel_base (PATTERN (insn
));
6581 gcc_assert (ltrel_base
== pending_ltrel
);
6582 pending_ltrel
= NULL_RTX
;
6586 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6589 curr_pool
= s390_start_pool (&pool_list
, insn
);
6591 s390_add_execute (curr_pool
, insn
);
6592 s390_add_pool_insn (curr_pool
, insn
);
6594 else if (GET_CODE (insn
) == INSN
|| CALL_P (insn
))
6596 rtx pool_ref
= NULL_RTX
;
6597 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6600 rtx constant
= get_pool_constant (pool_ref
);
6601 enum machine_mode mode
= get_pool_mode (pool_ref
);
6604 curr_pool
= s390_start_pool (&pool_list
, insn
);
6606 s390_add_constant (curr_pool
, constant
, mode
);
6607 s390_add_pool_insn (curr_pool
, insn
);
6609 /* Don't split the pool chunk between a LTREL_OFFSET load
6610 and the corresponding LTREL_BASE. */
6611 if (GET_CODE (constant
) == CONST
6612 && GET_CODE (XEXP (constant
, 0)) == UNSPEC
6613 && XINT (XEXP (constant
, 0), 1) == UNSPEC_LTREL_OFFSET
)
6615 gcc_assert (!pending_ltrel
);
6616 pending_ltrel
= pool_ref
;
6621 if (GET_CODE (insn
) == JUMP_INSN
|| GET_CODE (insn
) == CODE_LABEL
)
6624 s390_add_pool_insn (curr_pool
, insn
);
6625 /* An LTREL_BASE must follow within the same basic block. */
6626 gcc_assert (!pending_ltrel
);
6630 switch (NOTE_KIND (insn
))
6632 case NOTE_INSN_SWITCH_TEXT_SECTIONS
:
6633 section_switch_p
= true;
6635 case NOTE_INSN_VAR_LOCATION
:
6636 case NOTE_INSN_CALL_ARG_LOCATION
:
6643 || INSN_ADDRESSES_SIZE () <= (size_t) INSN_UID (insn
)
6644 || INSN_ADDRESSES (INSN_UID (insn
)) == -1)
6647 if (TARGET_CPU_ZARCH
)
6649 if (curr_pool
->size
< S390_POOL_CHUNK_MAX
)
6652 s390_end_pool (curr_pool
, NULL_RTX
);
6657 int chunk_size
= INSN_ADDRESSES (INSN_UID (insn
))
6658 - INSN_ADDRESSES (INSN_UID (curr_pool
->first_insn
))
6661 /* We will later have to insert base register reload insns.
6662 Those will have an effect on code size, which we need to
6663 consider here. This calculation makes rather pessimistic
6664 worst-case assumptions. */
6665 if (GET_CODE (insn
) == CODE_LABEL
)
6668 if (chunk_size
< S390_POOL_CHUNK_MIN
6669 && curr_pool
->size
< S390_POOL_CHUNK_MIN
6670 && !section_switch_p
)
6673 /* Pool chunks can only be inserted after BARRIERs ... */
6674 if (GET_CODE (insn
) == BARRIER
)
6676 s390_end_pool (curr_pool
, insn
);
6681 /* ... so if we don't find one in time, create one. */
6682 else if (chunk_size
> S390_POOL_CHUNK_MAX
6683 || curr_pool
->size
> S390_POOL_CHUNK_MAX
6684 || section_switch_p
)
6686 rtx label
, jump
, barrier
, next
, prev
;
6688 if (!section_switch_p
)
6690 /* We can insert the barrier only after a 'real' insn. */
6691 if (GET_CODE (insn
) != INSN
&& GET_CODE (insn
) != CALL_INSN
)
6693 if (get_attr_length (insn
) == 0)
6695 /* Don't separate LTREL_BASE from the corresponding
6696 LTREL_OFFSET load. */
6703 next
= NEXT_INSN (insn
);
6707 && (NOTE_KIND (next
) == NOTE_INSN_VAR_LOCATION
6708 || NOTE_KIND (next
) == NOTE_INSN_CALL_ARG_LOCATION
));
6712 gcc_assert (!pending_ltrel
);
6714 /* The old pool has to end before the section switch
6715 note in order to make it part of the current
6717 insn
= PREV_INSN (insn
);
6720 label
= gen_label_rtx ();
6722 if (prev
&& NOTE_P (prev
))
6723 prev
= prev_nonnote_insn (prev
);
6725 jump
= emit_jump_insn_after_setloc (gen_jump (label
), insn
,
6726 INSN_LOCATOR (prev
));
6728 jump
= emit_jump_insn_after_noloc (gen_jump (label
), insn
);
6729 barrier
= emit_barrier_after (jump
);
6730 insn
= emit_label_after (label
, barrier
);
6731 JUMP_LABEL (jump
) = label
;
6732 LABEL_NUSES (label
) = 1;
6734 INSN_ADDRESSES_NEW (jump
, -1);
6735 INSN_ADDRESSES_NEW (barrier
, -1);
6736 INSN_ADDRESSES_NEW (insn
, -1);
6738 s390_end_pool (curr_pool
, barrier
);
6746 s390_end_pool (curr_pool
, NULL_RTX
);
6747 gcc_assert (!pending_ltrel
);
6749 /* Find all labels that are branched into
6750 from an insn belonging to a different chunk. */
6752 far_labels
= BITMAP_ALLOC (NULL
);
6754 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6756 /* Labels marked with LABEL_PRESERVE_P can be target
6757 of non-local jumps, so we have to mark them.
6758 The same holds for named labels.
6760 Don't do that, however, if it is the label before
6763 if (GET_CODE (insn
) == CODE_LABEL
6764 && (LABEL_PRESERVE_P (insn
) || LABEL_NAME (insn
)))
6766 rtx vec_insn
= next_real_insn (insn
);
6767 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6768 PATTERN (vec_insn
) : NULL_RTX
;
6770 || !(GET_CODE (vec_pat
) == ADDR_VEC
6771 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6772 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (insn
));
6775 /* If we have a direct jump (conditional or unconditional)
6776 or a casesi jump, check all potential targets. */
6777 else if (GET_CODE (insn
) == JUMP_INSN
)
6779 rtx pat
= PATTERN (insn
);
6780 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
6781 pat
= XVECEXP (pat
, 0, 0);
6783 if (GET_CODE (pat
) == SET
)
6785 rtx label
= JUMP_LABEL (insn
);
6788 if (s390_find_pool (pool_list
, label
)
6789 != s390_find_pool (pool_list
, insn
))
6790 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6793 else if (GET_CODE (pat
) == PARALLEL
6794 && XVECLEN (pat
, 0) == 2
6795 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
6796 && GET_CODE (XVECEXP (pat
, 0, 1)) == USE
6797 && GET_CODE (XEXP (XVECEXP (pat
, 0, 1), 0)) == LABEL_REF
)
6799 /* Find the jump table used by this casesi jump. */
6800 rtx vec_label
= XEXP (XEXP (XVECEXP (pat
, 0, 1), 0), 0);
6801 rtx vec_insn
= next_real_insn (vec_label
);
6802 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6803 PATTERN (vec_insn
) : NULL_RTX
;
6805 && (GET_CODE (vec_pat
) == ADDR_VEC
6806 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6808 int i
, diff_p
= GET_CODE (vec_pat
) == ADDR_DIFF_VEC
;
6810 for (i
= 0; i
< XVECLEN (vec_pat
, diff_p
); i
++)
6812 rtx label
= XEXP (XVECEXP (vec_pat
, diff_p
, i
), 0);
6814 if (s390_find_pool (pool_list
, label
)
6815 != s390_find_pool (pool_list
, insn
))
6816 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6823 /* Insert base register reload insns before every pool. */
6825 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6827 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6829 rtx insn
= curr_pool
->first_insn
;
6830 INSN_ADDRESSES_NEW (emit_insn_before (new_insn
, insn
), -1);
6833 /* Insert base register reload insns at every far label. */
6835 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6836 if (GET_CODE (insn
) == CODE_LABEL
6837 && bitmap_bit_p (far_labels
, CODE_LABEL_NUMBER (insn
)))
6839 struct constant_pool
*pool
= s390_find_pool (pool_list
, insn
);
6842 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6844 INSN_ADDRESSES_NEW (emit_insn_after (new_insn
, insn
), -1);
6849 BITMAP_FREE (far_labels
);
6852 /* Recompute insn addresses. */
6854 init_insn_lengths ();
6855 shorten_branches (get_insns ());
6860 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6861 After we have decided to use this list, finish implementing
6862 all changes to the current function as required. */
6865 s390_chunkify_finish (struct constant_pool
*pool_list
)
6867 struct constant_pool
*curr_pool
= NULL
;
6871 /* Replace all literal pool references. */
6873 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6876 replace_ltrel_base (&PATTERN (insn
));
6878 curr_pool
= s390_find_pool (pool_list
, insn
);
6882 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6884 rtx addr
, pool_ref
= NULL_RTX
;
6885 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6888 if (s390_execute_label (insn
))
6889 addr
= s390_find_execute (curr_pool
, insn
);
6891 addr
= s390_find_constant (curr_pool
,
6892 get_pool_constant (pool_ref
),
6893 get_pool_mode (pool_ref
));
6895 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6896 INSN_CODE (insn
) = -1;
6901 /* Dump out all literal pools. */
6903 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6904 s390_dump_pool (curr_pool
, 0);
6906 /* Free pool list. */
6910 struct constant_pool
*next
= pool_list
->next
;
6911 s390_free_pool (pool_list
);
6916 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6917 We have decided we cannot use this list, so revert all changes
6918 to the current function that were done by s390_chunkify_start. */
6921 s390_chunkify_cancel (struct constant_pool
*pool_list
)
6923 struct constant_pool
*curr_pool
= NULL
;
6926 /* Remove all pool placeholder insns. */
6928 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6930 /* Did we insert an extra barrier? Remove it. */
6931 rtx barrier
= PREV_INSN (curr_pool
->pool_insn
);
6932 rtx jump
= barrier
? PREV_INSN (barrier
) : NULL_RTX
;
6933 rtx label
= NEXT_INSN (curr_pool
->pool_insn
);
6935 if (jump
&& GET_CODE (jump
) == JUMP_INSN
6936 && barrier
&& GET_CODE (barrier
) == BARRIER
6937 && label
&& GET_CODE (label
) == CODE_LABEL
6938 && GET_CODE (PATTERN (jump
)) == SET
6939 && SET_DEST (PATTERN (jump
)) == pc_rtx
6940 && GET_CODE (SET_SRC (PATTERN (jump
))) == LABEL_REF
6941 && XEXP (SET_SRC (PATTERN (jump
)), 0) == label
)
6944 remove_insn (barrier
);
6945 remove_insn (label
);
6948 remove_insn (curr_pool
->pool_insn
);
6951 /* Remove all base register reload insns. */
6953 for (insn
= get_insns (); insn
; )
6955 rtx next_insn
= NEXT_INSN (insn
);
6957 if (GET_CODE (insn
) == INSN
6958 && GET_CODE (PATTERN (insn
)) == SET
6959 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC
6960 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPEC_RELOAD_BASE
)
6966 /* Free pool list. */
6970 struct constant_pool
*next
= pool_list
->next
;
6971 s390_free_pool (pool_list
);
6976 /* Output the constant pool entry EXP in mode MODE with alignment ALIGN. */
6979 s390_output_pool_entry (rtx exp
, enum machine_mode mode
, unsigned int align
)
6983 switch (GET_MODE_CLASS (mode
))
6986 case MODE_DECIMAL_FLOAT
:
6987 gcc_assert (GET_CODE (exp
) == CONST_DOUBLE
);
6989 REAL_VALUE_FROM_CONST_DOUBLE (r
, exp
);
6990 assemble_real (r
, mode
, align
);
6994 assemble_integer (exp
, GET_MODE_SIZE (mode
), align
, 1);
6995 mark_symbol_refs_as_used (exp
);
7004 /* Return an RTL expression representing the value of the return address
7005 for the frame COUNT steps up from the current frame. FRAME is the
7006 frame pointer of that frame. */
7009 s390_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
7014 /* Without backchain, we fail for all but the current frame. */
7016 if (!TARGET_BACKCHAIN
&& count
> 0)
7019 /* For the current frame, we need to make sure the initial
7020 value of RETURN_REGNUM is actually saved. */
7024 /* On non-z architectures branch splitting could overwrite r14. */
7025 if (TARGET_CPU_ZARCH
)
7026 return get_hard_reg_initial_val (Pmode
, RETURN_REGNUM
);
7029 cfun_frame_layout
.save_return_addr_p
= true;
7030 return gen_rtx_MEM (Pmode
, return_address_pointer_rtx
);
7034 if (TARGET_PACKED_STACK
)
7035 offset
= -2 * UNITS_PER_LONG
;
7037 offset
= RETURN_REGNUM
* UNITS_PER_LONG
;
7039 addr
= plus_constant (Pmode
, frame
, offset
);
7040 addr
= memory_address (Pmode
, addr
);
7041 return gen_rtx_MEM (Pmode
, addr
);
7044 /* Return an RTL expression representing the back chain stored in
7045 the current stack frame. */
7048 s390_back_chain_rtx (void)
7052 gcc_assert (TARGET_BACKCHAIN
);
7054 if (TARGET_PACKED_STACK
)
7055 chain
= plus_constant (Pmode
, stack_pointer_rtx
,
7056 STACK_POINTER_OFFSET
- UNITS_PER_LONG
);
7058 chain
= stack_pointer_rtx
;
7060 chain
= gen_rtx_MEM (Pmode
, chain
);
7064 /* Find first call clobbered register unused in a function.
7065 This could be used as base register in a leaf function
7066 or for holding the return address before epilogue. */
7069 find_unused_clobbered_reg (void)
7072 for (i
= 0; i
< 6; i
++)
7073 if (!df_regs_ever_live_p (i
))
7079 /* Helper function for s390_regs_ever_clobbered. Sets the fields in DATA for all
7080 clobbered hard regs in SETREG. */
7083 s390_reg_clobbered_rtx (rtx setreg
, const_rtx set_insn ATTRIBUTE_UNUSED
, void *data
)
7085 int *regs_ever_clobbered
= (int *)data
;
7086 unsigned int i
, regno
;
7087 enum machine_mode mode
= GET_MODE (setreg
);
7089 if (GET_CODE (setreg
) == SUBREG
)
7091 rtx inner
= SUBREG_REG (setreg
);
7092 if (!GENERAL_REG_P (inner
))
7094 regno
= subreg_regno (setreg
);
7096 else if (GENERAL_REG_P (setreg
))
7097 regno
= REGNO (setreg
);
7102 i
< regno
+ HARD_REGNO_NREGS (regno
, mode
);
7104 regs_ever_clobbered
[i
] = 1;
7107 /* Walks through all basic blocks of the current function looking
7108 for clobbered hard regs using s390_reg_clobbered_rtx. The fields
7109 of the passed integer array REGS_EVER_CLOBBERED are set to one for
7110 each of those regs. */
7113 s390_regs_ever_clobbered (int *regs_ever_clobbered
)
7119 memset (regs_ever_clobbered
, 0, 16 * sizeof (int));
7121 /* For non-leaf functions we have to consider all call clobbered regs to be
7123 if (!current_function_is_leaf
)
7125 for (i
= 0; i
< 16; i
++)
7126 regs_ever_clobbered
[i
] = call_really_used_regs
[i
];
7129 /* Make the "magic" eh_return registers live if necessary. For regs_ever_live
7130 this work is done by liveness analysis (mark_regs_live_at_end).
7131 Special care is needed for functions containing landing pads. Landing pads
7132 may use the eh registers, but the code which sets these registers is not
7133 contained in that function. Hence s390_regs_ever_clobbered is not able to
7134 deal with this automatically. */
7135 if (crtl
->calls_eh_return
|| cfun
->machine
->has_landing_pad_p
)
7136 for (i
= 0; EH_RETURN_DATA_REGNO (i
) != INVALID_REGNUM
; i
++)
7137 if (crtl
->calls_eh_return
7138 || (cfun
->machine
->has_landing_pad_p
7139 && df_regs_ever_live_p (EH_RETURN_DATA_REGNO (i
))))
7140 regs_ever_clobbered
[EH_RETURN_DATA_REGNO (i
)] = 1;
7142 /* For nonlocal gotos all call-saved registers have to be saved.
7143 This flag is also set for the unwinding code in libgcc.
7144 See expand_builtin_unwind_init. For regs_ever_live this is done by
7146 if (cfun
->has_nonlocal_label
)
7147 for (i
= 0; i
< 16; i
++)
7148 if (!call_really_used_regs
[i
])
7149 regs_ever_clobbered
[i
] = 1;
7151 FOR_EACH_BB (cur_bb
)
7153 FOR_BB_INSNS (cur_bb
, cur_insn
)
7155 if (INSN_P (cur_insn
))
7156 note_stores (PATTERN (cur_insn
),
7157 s390_reg_clobbered_rtx
,
7158 regs_ever_clobbered
);
7163 /* Determine the frame area which actually has to be accessed
7164 in the function epilogue. The values are stored at the
7165 given pointers AREA_BOTTOM (address of the lowest used stack
7166 address) and AREA_TOP (address of the first item which does
7167 not belong to the stack frame). */
7170 s390_frame_area (int *area_bottom
, int *area_top
)
7178 if (cfun_frame_layout
.first_restore_gpr
!= -1)
7180 b
= (cfun_frame_layout
.gprs_offset
7181 + cfun_frame_layout
.first_restore_gpr
* UNITS_PER_LONG
);
7182 t
= b
+ (cfun_frame_layout
.last_restore_gpr
7183 - cfun_frame_layout
.first_restore_gpr
+ 1) * UNITS_PER_LONG
;
7186 if (TARGET_64BIT
&& cfun_save_high_fprs_p
)
7188 b
= MIN (b
, cfun_frame_layout
.f8_offset
);
7189 t
= MAX (t
, (cfun_frame_layout
.f8_offset
7190 + cfun_frame_layout
.high_fprs
* 8));
7194 for (i
= 2; i
< 4; i
++)
7195 if (cfun_fpr_bit_p (i
))
7197 b
= MIN (b
, cfun_frame_layout
.f4_offset
+ (i
- 2) * 8);
7198 t
= MAX (t
, cfun_frame_layout
.f4_offset
+ (i
- 1) * 8);
7205 /* Fill cfun->machine with info about register usage of current function.
7206 Return in CLOBBERED_REGS which GPRs are currently considered set. */
7209 s390_register_info (int clobbered_regs
[])
7213 /* fprs 8 - 15 are call saved for 64 Bit ABI. */
7214 cfun_frame_layout
.fpr_bitmap
= 0;
7215 cfun_frame_layout
.high_fprs
= 0;
7217 for (i
= 24; i
< 32; i
++)
7218 if (df_regs_ever_live_p (i
) && !global_regs
[i
])
7220 cfun_set_fpr_bit (i
- 16);
7221 cfun_frame_layout
.high_fprs
++;
7224 /* Find first and last gpr to be saved. We trust regs_ever_live
7225 data, except that we don't save and restore global registers.
7227 Also, all registers with special meaning to the compiler need
7228 to be handled extra. */
7230 s390_regs_ever_clobbered (clobbered_regs
);
7232 for (i
= 0; i
< 16; i
++)
7233 clobbered_regs
[i
] = clobbered_regs
[i
] && !global_regs
[i
] && !fixed_regs
[i
];
7235 if (frame_pointer_needed
)
7236 clobbered_regs
[HARD_FRAME_POINTER_REGNUM
] = 1;
7239 clobbered_regs
[PIC_OFFSET_TABLE_REGNUM
]
7240 |= df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
);
7242 clobbered_regs
[BASE_REGNUM
]
7243 |= (cfun
->machine
->base_reg
7244 && REGNO (cfun
->machine
->base_reg
) == BASE_REGNUM
);
7246 clobbered_regs
[RETURN_REGNUM
]
7247 |= (!current_function_is_leaf
7248 || TARGET_TPF_PROFILING
7249 || cfun
->machine
->split_branches_pending_p
7250 || cfun_frame_layout
.save_return_addr_p
7251 || crtl
->calls_eh_return
7254 clobbered_regs
[STACK_POINTER_REGNUM
]
7255 |= (!current_function_is_leaf
7256 || TARGET_TPF_PROFILING
7257 || cfun_save_high_fprs_p
7258 || get_frame_size () > 0
7259 || cfun
->calls_alloca
7262 for (i
= 6; i
< 16; i
++)
7263 if (df_regs_ever_live_p (i
) || clobbered_regs
[i
])
7265 for (j
= 15; j
> i
; j
--)
7266 if (df_regs_ever_live_p (j
) || clobbered_regs
[j
])
7271 /* Nothing to save/restore. */
7272 cfun_frame_layout
.first_save_gpr_slot
= -1;
7273 cfun_frame_layout
.last_save_gpr_slot
= -1;
7274 cfun_frame_layout
.first_save_gpr
= -1;
7275 cfun_frame_layout
.first_restore_gpr
= -1;
7276 cfun_frame_layout
.last_save_gpr
= -1;
7277 cfun_frame_layout
.last_restore_gpr
= -1;
7281 /* Save slots for gprs from i to j. */
7282 cfun_frame_layout
.first_save_gpr_slot
= i
;
7283 cfun_frame_layout
.last_save_gpr_slot
= j
;
7285 for (i
= cfun_frame_layout
.first_save_gpr_slot
;
7286 i
< cfun_frame_layout
.last_save_gpr_slot
+ 1;
7288 if (clobbered_regs
[i
])
7291 for (j
= cfun_frame_layout
.last_save_gpr_slot
; j
> i
; j
--)
7292 if (clobbered_regs
[j
])
7295 if (i
== cfun_frame_layout
.last_save_gpr_slot
+ 1)
7297 /* Nothing to save/restore. */
7298 cfun_frame_layout
.first_save_gpr
= -1;
7299 cfun_frame_layout
.first_restore_gpr
= -1;
7300 cfun_frame_layout
.last_save_gpr
= -1;
7301 cfun_frame_layout
.last_restore_gpr
= -1;
7305 /* Save / Restore from gpr i to j. */
7306 cfun_frame_layout
.first_save_gpr
= i
;
7307 cfun_frame_layout
.first_restore_gpr
= i
;
7308 cfun_frame_layout
.last_save_gpr
= j
;
7309 cfun_frame_layout
.last_restore_gpr
= j
;
7315 /* Varargs functions need to save gprs 2 to 6. */
7316 if (cfun
->va_list_gpr_size
7317 && crtl
->args
.info
.gprs
< GP_ARG_NUM_REG
)
7319 int min_gpr
= crtl
->args
.info
.gprs
;
7320 int max_gpr
= min_gpr
+ cfun
->va_list_gpr_size
;
7321 if (max_gpr
> GP_ARG_NUM_REG
)
7322 max_gpr
= GP_ARG_NUM_REG
;
7324 if (cfun_frame_layout
.first_save_gpr
== -1
7325 || cfun_frame_layout
.first_save_gpr
> 2 + min_gpr
)
7327 cfun_frame_layout
.first_save_gpr
= 2 + min_gpr
;
7328 cfun_frame_layout
.first_save_gpr_slot
= 2 + min_gpr
;
7331 if (cfun_frame_layout
.last_save_gpr
== -1
7332 || cfun_frame_layout
.last_save_gpr
< 2 + max_gpr
- 1)
7334 cfun_frame_layout
.last_save_gpr
= 2 + max_gpr
- 1;
7335 cfun_frame_layout
.last_save_gpr_slot
= 2 + max_gpr
- 1;
7339 /* Mark f0, f2 for 31 bit and f0-f4 for 64 bit to be saved. */
7340 if (TARGET_HARD_FLOAT
&& cfun
->va_list_fpr_size
7341 && crtl
->args
.info
.fprs
< FP_ARG_NUM_REG
)
7343 int min_fpr
= crtl
->args
.info
.fprs
;
7344 int max_fpr
= min_fpr
+ cfun
->va_list_fpr_size
;
7345 if (max_fpr
> FP_ARG_NUM_REG
)
7346 max_fpr
= FP_ARG_NUM_REG
;
7348 /* ??? This is currently required to ensure proper location
7349 of the fpr save slots within the va_list save area. */
7350 if (TARGET_PACKED_STACK
)
7353 for (i
= min_fpr
; i
< max_fpr
; i
++)
7354 cfun_set_fpr_bit (i
);
7359 for (i
= 2; i
< 4; i
++)
7360 if (df_regs_ever_live_p (i
+ 16) && !global_regs
[i
+ 16])
7361 cfun_set_fpr_bit (i
);
7364 /* Fill cfun->machine with info about frame of current function. */
7367 s390_frame_info (void)
7371 cfun_frame_layout
.frame_size
= get_frame_size ();
7372 if (!TARGET_64BIT
&& cfun_frame_layout
.frame_size
> 0x7fff0000)
7373 fatal_error ("total size of local variables exceeds architecture limit");
7375 if (!TARGET_PACKED_STACK
)
7377 cfun_frame_layout
.backchain_offset
= 0;
7378 cfun_frame_layout
.f0_offset
= 16 * UNITS_PER_LONG
;
7379 cfun_frame_layout
.f4_offset
= cfun_frame_layout
.f0_offset
+ 2 * 8;
7380 cfun_frame_layout
.f8_offset
= -cfun_frame_layout
.high_fprs
* 8;
7381 cfun_frame_layout
.gprs_offset
= (cfun_frame_layout
.first_save_gpr_slot
7384 else if (TARGET_BACKCHAIN
) /* kernel stack layout */
7386 cfun_frame_layout
.backchain_offset
= (STACK_POINTER_OFFSET
7388 cfun_frame_layout
.gprs_offset
7389 = (cfun_frame_layout
.backchain_offset
7390 - (STACK_POINTER_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
+ 1)
7395 cfun_frame_layout
.f4_offset
7396 = (cfun_frame_layout
.gprs_offset
7397 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7399 cfun_frame_layout
.f0_offset
7400 = (cfun_frame_layout
.f4_offset
7401 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7405 /* On 31 bit we have to care about alignment of the
7406 floating point regs to provide fastest access. */
7407 cfun_frame_layout
.f0_offset
7408 = ((cfun_frame_layout
.gprs_offset
7409 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1))
7410 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7412 cfun_frame_layout
.f4_offset
7413 = (cfun_frame_layout
.f0_offset
7414 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7417 else /* no backchain */
7419 cfun_frame_layout
.f4_offset
7420 = (STACK_POINTER_OFFSET
7421 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7423 cfun_frame_layout
.f0_offset
7424 = (cfun_frame_layout
.f4_offset
7425 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7427 cfun_frame_layout
.gprs_offset
7428 = cfun_frame_layout
.f0_offset
- cfun_gprs_save_area_size
;
7431 if (current_function_is_leaf
7432 && !TARGET_TPF_PROFILING
7433 && cfun_frame_layout
.frame_size
== 0
7434 && !cfun_save_high_fprs_p
7435 && !cfun
->calls_alloca
7439 if (!TARGET_PACKED_STACK
)
7440 cfun_frame_layout
.frame_size
+= (STACK_POINTER_OFFSET
7441 + crtl
->outgoing_args_size
7442 + cfun_frame_layout
.high_fprs
* 8);
7445 if (TARGET_BACKCHAIN
)
7446 cfun_frame_layout
.frame_size
+= UNITS_PER_LONG
;
7448 /* No alignment trouble here because f8-f15 are only saved under
7450 cfun_frame_layout
.f8_offset
= (MIN (MIN (cfun_frame_layout
.f0_offset
,
7451 cfun_frame_layout
.f4_offset
),
7452 cfun_frame_layout
.gprs_offset
)
7453 - cfun_frame_layout
.high_fprs
* 8);
7455 cfun_frame_layout
.frame_size
+= cfun_frame_layout
.high_fprs
* 8;
7457 for (i
= 0; i
< 8; i
++)
7458 if (cfun_fpr_bit_p (i
))
7459 cfun_frame_layout
.frame_size
+= 8;
7461 cfun_frame_layout
.frame_size
+= cfun_gprs_save_area_size
;
7463 /* If under 31 bit an odd number of gprs has to be saved we have to adjust
7464 the frame size to sustain 8 byte alignment of stack frames. */
7465 cfun_frame_layout
.frame_size
= ((cfun_frame_layout
.frame_size
+
7466 STACK_BOUNDARY
/ BITS_PER_UNIT
- 1)
7467 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1));
7469 cfun_frame_layout
.frame_size
+= crtl
->outgoing_args_size
;
7473 /* Generate frame layout. Fills in register and frame data for the current
7474 function in cfun->machine. This routine can be called multiple times;
7475 it will re-do the complete frame layout every time. */
7478 s390_init_frame_layout (void)
7480 HOST_WIDE_INT frame_size
;
7482 int clobbered_regs
[16];
7484 /* On S/390 machines, we may need to perform branch splitting, which
7485 will require both base and return address register. We have no
7486 choice but to assume we're going to need them until right at the
7487 end of the machine dependent reorg phase. */
7488 if (!TARGET_CPU_ZARCH
)
7489 cfun
->machine
->split_branches_pending_p
= true;
7493 frame_size
= cfun_frame_layout
.frame_size
;
7495 /* Try to predict whether we'll need the base register. */
7496 base_used
= cfun
->machine
->split_branches_pending_p
7497 || crtl
->uses_const_pool
7498 || (!DISP_IN_RANGE (frame_size
)
7499 && !CONST_OK_FOR_K (frame_size
));
7501 /* Decide which register to use as literal pool base. In small
7502 leaf functions, try to use an unused call-clobbered register
7503 as base register to avoid save/restore overhead. */
7505 cfun
->machine
->base_reg
= NULL_RTX
;
7506 else if (current_function_is_leaf
&& !df_regs_ever_live_p (5))
7507 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, 5);
7509 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, BASE_REGNUM
);
7511 s390_register_info (clobbered_regs
);
7514 while (frame_size
!= cfun_frame_layout
.frame_size
);
7517 /* Update frame layout. Recompute actual register save data based on
7518 current info and update regs_ever_live for the special registers.
7519 May be called multiple times, but may never cause *more* registers
7520 to be saved than s390_init_frame_layout allocated room for. */
7523 s390_update_frame_layout (void)
7525 int clobbered_regs
[16];
7527 s390_register_info (clobbered_regs
);
7529 df_set_regs_ever_live (BASE_REGNUM
,
7530 clobbered_regs
[BASE_REGNUM
] ? true : false);
7531 df_set_regs_ever_live (RETURN_REGNUM
,
7532 clobbered_regs
[RETURN_REGNUM
] ? true : false);
7533 df_set_regs_ever_live (STACK_POINTER_REGNUM
,
7534 clobbered_regs
[STACK_POINTER_REGNUM
] ? true : false);
7536 if (cfun
->machine
->base_reg
)
7537 df_set_regs_ever_live (REGNO (cfun
->machine
->base_reg
), true);
7540 /* Return true if it is legal to put a value with MODE into REGNO. */
7543 s390_hard_regno_mode_ok (unsigned int regno
, enum machine_mode mode
)
7545 switch (REGNO_REG_CLASS (regno
))
7548 if (REGNO_PAIR_OK (regno
, mode
))
7550 if (mode
== SImode
|| mode
== DImode
)
7553 if (FLOAT_MODE_P (mode
) && GET_MODE_CLASS (mode
) != MODE_VECTOR_FLOAT
)
7558 if (FRAME_REGNO_P (regno
) && mode
== Pmode
)
7563 if (REGNO_PAIR_OK (regno
, mode
))
7566 || (mode
!= TFmode
&& mode
!= TCmode
&& mode
!= TDmode
))
7571 if (GET_MODE_CLASS (mode
) == MODE_CC
)
7575 if (REGNO_PAIR_OK (regno
, mode
))
7577 if (mode
== SImode
|| mode
== Pmode
)
7588 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
7591 s390_hard_regno_rename_ok (unsigned int old_reg
, unsigned int new_reg
)
7593 /* Once we've decided upon a register to use as base register, it must
7594 no longer be used for any other purpose. */
7595 if (cfun
->machine
->base_reg
)
7596 if (REGNO (cfun
->machine
->base_reg
) == old_reg
7597 || REGNO (cfun
->machine
->base_reg
) == new_reg
)
7603 /* Maximum number of registers to represent a value of mode MODE
7604 in a register of class RCLASS. */
7607 s390_class_max_nregs (enum reg_class rclass
, enum machine_mode mode
)
7612 if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7613 return 2 * ((GET_MODE_SIZE (mode
) / 2 + 8 - 1) / 8);
7615 return (GET_MODE_SIZE (mode
) + 8 - 1) / 8;
7617 return (GET_MODE_SIZE (mode
) + 4 - 1) / 4;
7621 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
7624 /* Return true if register FROM can be eliminated via register TO. */
7627 s390_can_eliminate (const int from
, const int to
)
7629 /* On zSeries machines, we have not marked the base register as fixed.
7630 Instead, we have an elimination rule BASE_REGNUM -> BASE_REGNUM.
7631 If a function requires the base register, we say here that this
7632 elimination cannot be performed. This will cause reload to free
7633 up the base register (as if it were fixed). On the other hand,
7634 if the current function does *not* require the base register, we
7635 say here the elimination succeeds, which in turn allows reload
7636 to allocate the base register for any other purpose. */
7637 if (from
== BASE_REGNUM
&& to
== BASE_REGNUM
)
7639 if (TARGET_CPU_ZARCH
)
7641 s390_init_frame_layout ();
7642 return cfun
->machine
->base_reg
== NULL_RTX
;
7648 /* Everything else must point into the stack frame. */
7649 gcc_assert (to
== STACK_POINTER_REGNUM
7650 || to
== HARD_FRAME_POINTER_REGNUM
);
7652 gcc_assert (from
== FRAME_POINTER_REGNUM
7653 || from
== ARG_POINTER_REGNUM
7654 || from
== RETURN_ADDRESS_POINTER_REGNUM
);
7656 /* Make sure we actually saved the return address. */
7657 if (from
== RETURN_ADDRESS_POINTER_REGNUM
)
7658 if (!crtl
->calls_eh_return
7660 && !cfun_frame_layout
.save_return_addr_p
)
7666 /* Return offset between register FROM and TO initially after prolog. */
7669 s390_initial_elimination_offset (int from
, int to
)
7671 HOST_WIDE_INT offset
;
7674 /* ??? Why are we called for non-eliminable pairs? */
7675 if (!s390_can_eliminate (from
, to
))
7680 case FRAME_POINTER_REGNUM
:
7681 offset
= (get_frame_size()
7682 + STACK_POINTER_OFFSET
7683 + crtl
->outgoing_args_size
);
7686 case ARG_POINTER_REGNUM
:
7687 s390_init_frame_layout ();
7688 offset
= cfun_frame_layout
.frame_size
+ STACK_POINTER_OFFSET
;
7691 case RETURN_ADDRESS_POINTER_REGNUM
:
7692 s390_init_frame_layout ();
7693 index
= RETURN_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
;
7694 gcc_assert (index
>= 0);
7695 offset
= cfun_frame_layout
.frame_size
+ cfun_frame_layout
.gprs_offset
;
7696 offset
+= index
* UNITS_PER_LONG
;
7710 /* Emit insn to save fpr REGNUM at offset OFFSET relative
7711 to register BASE. Return generated insn. */
7714 save_fpr (rtx base
, int offset
, int regnum
)
7717 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7719 if (regnum
>= 16 && regnum
<= (16 + FP_ARG_NUM_REG
))
7720 set_mem_alias_set (addr
, get_varargs_alias_set ());
7722 set_mem_alias_set (addr
, get_frame_alias_set ());
7724 return emit_move_insn (addr
, gen_rtx_REG (DFmode
, regnum
));
7727 /* Emit insn to restore fpr REGNUM from offset OFFSET relative
7728 to register BASE. Return generated insn. */
7731 restore_fpr (rtx base
, int offset
, int regnum
)
7734 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7735 set_mem_alias_set (addr
, get_frame_alias_set ());
7737 return emit_move_insn (gen_rtx_REG (DFmode
, regnum
), addr
);
7740 /* Return true if REGNO is a global register, but not one
7741 of the special ones that need to be saved/restored in anyway. */
7744 global_not_special_regno_p (int regno
)
7746 return (global_regs
[regno
]
7747 /* These registers are special and need to be
7748 restored in any case. */
7749 && !(regno
== STACK_POINTER_REGNUM
7750 || regno
== RETURN_REGNUM
7751 || regno
== BASE_REGNUM
7752 || (flag_pic
&& regno
== (int)PIC_OFFSET_TABLE_REGNUM
)));
7755 /* Generate insn to save registers FIRST to LAST into
7756 the register save area located at offset OFFSET
7757 relative to register BASE. */
7760 save_gprs (rtx base
, int offset
, int first
, int last
)
7762 rtx addr
, insn
, note
;
7765 addr
= plus_constant (Pmode
, base
, offset
);
7766 addr
= gen_rtx_MEM (Pmode
, addr
);
7768 set_mem_alias_set (addr
, get_frame_alias_set ());
7770 /* Special-case single register. */
7774 insn
= gen_movdi (addr
, gen_rtx_REG (Pmode
, first
));
7776 insn
= gen_movsi (addr
, gen_rtx_REG (Pmode
, first
));
7778 if (!global_not_special_regno_p (first
))
7779 RTX_FRAME_RELATED_P (insn
) = 1;
7784 insn
= gen_store_multiple (addr
,
7785 gen_rtx_REG (Pmode
, first
),
7786 GEN_INT (last
- first
+ 1));
7788 if (first
<= 6 && cfun
->stdarg
)
7789 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
7791 rtx mem
= XEXP (XVECEXP (PATTERN (insn
), 0, i
), 0);
7794 set_mem_alias_set (mem
, get_varargs_alias_set ());
7797 /* We need to set the FRAME_RELATED flag on all SETs
7798 inside the store-multiple pattern.
7800 However, we must not emit DWARF records for registers 2..5
7801 if they are stored for use by variable arguments ...
7803 ??? Unfortunately, it is not enough to simply not the
7804 FRAME_RELATED flags for those SETs, because the first SET
7805 of the PARALLEL is always treated as if it had the flag
7806 set, even if it does not. Therefore we emit a new pattern
7807 without those registers as REG_FRAME_RELATED_EXPR note. */
7809 if (first
>= 6 && !global_not_special_regno_p (first
))
7811 rtx pat
= PATTERN (insn
);
7813 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
7814 if (GET_CODE (XVECEXP (pat
, 0, i
)) == SET
7815 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (pat
,
7817 RTX_FRAME_RELATED_P (XVECEXP (pat
, 0, i
)) = 1;
7819 RTX_FRAME_RELATED_P (insn
) = 1;
7825 for (start
= first
>= 6 ? first
: 6; start
<= last
; start
++)
7826 if (!global_not_special_regno_p (start
))
7832 addr
= plus_constant (Pmode
, base
,
7833 offset
+ (start
- first
) * UNITS_PER_LONG
);
7834 note
= gen_store_multiple (gen_rtx_MEM (Pmode
, addr
),
7835 gen_rtx_REG (Pmode
, start
),
7836 GEN_INT (last
- start
+ 1));
7837 note
= PATTERN (note
);
7839 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, note
);
7841 for (i
= 0; i
< XVECLEN (note
, 0); i
++)
7842 if (GET_CODE (XVECEXP (note
, 0, i
)) == SET
7843 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (note
,
7845 RTX_FRAME_RELATED_P (XVECEXP (note
, 0, i
)) = 1;
7847 RTX_FRAME_RELATED_P (insn
) = 1;
7853 /* Generate insn to restore registers FIRST to LAST from
7854 the register save area located at offset OFFSET
7855 relative to register BASE. */
7858 restore_gprs (rtx base
, int offset
, int first
, int last
)
7862 addr
= plus_constant (Pmode
, base
, offset
);
7863 addr
= gen_rtx_MEM (Pmode
, addr
);
7864 set_mem_alias_set (addr
, get_frame_alias_set ());
7866 /* Special-case single register. */
7870 insn
= gen_movdi (gen_rtx_REG (Pmode
, first
), addr
);
7872 insn
= gen_movsi (gen_rtx_REG (Pmode
, first
), addr
);
7877 insn
= gen_load_multiple (gen_rtx_REG (Pmode
, first
),
7879 GEN_INT (last
- first
+ 1));
7883 /* Return insn sequence to load the GOT register. */
7885 static GTY(()) rtx got_symbol
;
7887 s390_load_got (void)
7891 /* We cannot use pic_offset_table_rtx here since we use this
7892 function also for non-pic if __tls_get_offset is called and in
7893 that case PIC_OFFSET_TABLE_REGNUM as well as pic_offset_table_rtx
7895 rtx got_rtx
= gen_rtx_REG (Pmode
, 12);
7899 got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
7900 SYMBOL_REF_FLAGS (got_symbol
) = SYMBOL_FLAG_LOCAL
;
7905 if (TARGET_CPU_ZARCH
)
7907 emit_move_insn (got_rtx
, got_symbol
);
7913 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, got_symbol
),
7914 UNSPEC_LTREL_OFFSET
);
7915 offset
= gen_rtx_CONST (Pmode
, offset
);
7916 offset
= force_const_mem (Pmode
, offset
);
7918 emit_move_insn (got_rtx
, offset
);
7920 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (offset
, 0)),
7922 offset
= gen_rtx_PLUS (Pmode
, got_rtx
, offset
);
7924 emit_move_insn (got_rtx
, offset
);
7927 insns
= get_insns ();
7932 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
7933 and the change to the stack pointer. */
7936 s390_emit_stack_tie (void)
7938 rtx mem
= gen_frame_mem (BLKmode
,
7939 gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
));
7941 emit_insn (gen_stack_tie (mem
));
7944 /* Expand the prologue into a bunch of separate insns. */
7947 s390_emit_prologue (void)
7955 /* Complete frame layout. */
7957 s390_update_frame_layout ();
7959 /* Annotate all constant pool references to let the scheduler know
7960 they implicitly use the base register. */
7962 push_topmost_sequence ();
7964 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
7967 annotate_constant_pool_refs (&PATTERN (insn
));
7968 df_insn_rescan (insn
);
7971 pop_topmost_sequence ();
7973 /* Choose best register to use for temp use within prologue.
7974 See below for why TPF must use the register 1. */
7976 if (!has_hard_reg_initial_val (Pmode
, RETURN_REGNUM
)
7977 && !current_function_is_leaf
7978 && !TARGET_TPF_PROFILING
)
7979 temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
7981 temp_reg
= gen_rtx_REG (Pmode
, 1);
7983 /* Save call saved gprs. */
7984 if (cfun_frame_layout
.first_save_gpr
!= -1)
7986 insn
= save_gprs (stack_pointer_rtx
,
7987 cfun_frame_layout
.gprs_offset
+
7988 UNITS_PER_LONG
* (cfun_frame_layout
.first_save_gpr
7989 - cfun_frame_layout
.first_save_gpr_slot
),
7990 cfun_frame_layout
.first_save_gpr
,
7991 cfun_frame_layout
.last_save_gpr
);
7995 /* Dummy insn to mark literal pool slot. */
7997 if (cfun
->machine
->base_reg
)
7998 emit_insn (gen_main_pool (cfun
->machine
->base_reg
));
8000 offset
= cfun_frame_layout
.f0_offset
;
8002 /* Save f0 and f2. */
8003 for (i
= 0; i
< 2; i
++)
8005 if (cfun_fpr_bit_p (i
))
8007 save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8010 else if (!TARGET_PACKED_STACK
)
8014 /* Save f4 and f6. */
8015 offset
= cfun_frame_layout
.f4_offset
;
8016 for (i
= 2; i
< 4; i
++)
8018 if (cfun_fpr_bit_p (i
))
8020 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8023 /* If f4 and f6 are call clobbered they are saved due to stdargs and
8024 therefore are not frame related. */
8025 if (!call_really_used_regs
[i
+ 16])
8026 RTX_FRAME_RELATED_P (insn
) = 1;
8028 else if (!TARGET_PACKED_STACK
)
8032 if (TARGET_PACKED_STACK
8033 && cfun_save_high_fprs_p
8034 && cfun_frame_layout
.f8_offset
+ cfun_frame_layout
.high_fprs
* 8 > 0)
8036 offset
= (cfun_frame_layout
.f8_offset
8037 + (cfun_frame_layout
.high_fprs
- 1) * 8);
8039 for (i
= 15; i
> 7 && offset
>= 0; i
--)
8040 if (cfun_fpr_bit_p (i
))
8042 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8044 RTX_FRAME_RELATED_P (insn
) = 1;
8047 if (offset
>= cfun_frame_layout
.f8_offset
)
8051 if (!TARGET_PACKED_STACK
)
8052 next_fpr
= cfun_save_high_fprs_p
? 31 : 0;
8054 if (flag_stack_usage_info
)
8055 current_function_static_stack_size
= cfun_frame_layout
.frame_size
;
8057 /* Decrement stack pointer. */
8059 if (cfun_frame_layout
.frame_size
> 0)
8061 rtx frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8064 if (s390_stack_size
)
8066 HOST_WIDE_INT stack_guard
;
8068 if (s390_stack_guard
)
8069 stack_guard
= s390_stack_guard
;
8072 /* If no value for stack guard is provided the smallest power of 2
8073 larger than the current frame size is chosen. */
8075 while (stack_guard
< cfun_frame_layout
.frame_size
)
8079 if (cfun_frame_layout
.frame_size
>= s390_stack_size
)
8081 warning (0, "frame size of function %qs is %wd"
8082 " bytes exceeding user provided stack limit of "
8084 "An unconditional trap is added.",
8085 current_function_name(), cfun_frame_layout
.frame_size
,
8087 emit_insn (gen_trap ());
8091 /* stack_guard has to be smaller than s390_stack_size.
8092 Otherwise we would emit an AND with zero which would
8093 not match the test under mask pattern. */
8094 if (stack_guard
>= s390_stack_size
)
8096 warning (0, "frame size of function %qs is %wd"
8097 " bytes which is more than half the stack size. "
8098 "The dynamic check would not be reliable. "
8099 "No check emitted for this function.",
8100 current_function_name(),
8101 cfun_frame_layout
.frame_size
);
8105 HOST_WIDE_INT stack_check_mask
= ((s390_stack_size
- 1)
8106 & ~(stack_guard
- 1));
8108 rtx t
= gen_rtx_AND (Pmode
, stack_pointer_rtx
,
8109 GEN_INT (stack_check_mask
));
8111 emit_insn (gen_ctrapdi4 (gen_rtx_EQ (VOIDmode
,
8113 t
, const0_rtx
, const0_rtx
));
8115 emit_insn (gen_ctrapsi4 (gen_rtx_EQ (VOIDmode
,
8117 t
, const0_rtx
, const0_rtx
));
8122 if (s390_warn_framesize
> 0
8123 && cfun_frame_layout
.frame_size
>= s390_warn_framesize
)
8124 warning (0, "frame size of %qs is %wd bytes",
8125 current_function_name (), cfun_frame_layout
.frame_size
);
8127 if (s390_warn_dynamicstack_p
&& cfun
->calls_alloca
)
8128 warning (0, "%qs uses dynamic stack allocation", current_function_name ());
8130 /* Save incoming stack pointer into temp reg. */
8131 if (TARGET_BACKCHAIN
|| next_fpr
)
8132 insn
= emit_insn (gen_move_insn (temp_reg
, stack_pointer_rtx
));
8134 /* Subtract frame size from stack pointer. */
8136 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8138 insn
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8139 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8141 insn
= emit_insn (insn
);
8145 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8146 frame_off
= force_const_mem (Pmode
, frame_off
);
8148 insn
= emit_insn (gen_add2_insn (stack_pointer_rtx
, frame_off
));
8149 annotate_constant_pool_refs (&PATTERN (insn
));
8152 RTX_FRAME_RELATED_P (insn
) = 1;
8153 real_frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8154 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8155 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8156 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8159 /* Set backchain. */
8161 if (TARGET_BACKCHAIN
)
8163 if (cfun_frame_layout
.backchain_offset
)
8164 addr
= gen_rtx_MEM (Pmode
,
8165 plus_constant (Pmode
, stack_pointer_rtx
,
8166 cfun_frame_layout
.backchain_offset
));
8168 addr
= gen_rtx_MEM (Pmode
, stack_pointer_rtx
);
8169 set_mem_alias_set (addr
, get_frame_alias_set ());
8170 insn
= emit_insn (gen_move_insn (addr
, temp_reg
));
8173 /* If we support non-call exceptions (e.g. for Java),
8174 we need to make sure the backchain pointer is set up
8175 before any possibly trapping memory access. */
8176 if (TARGET_BACKCHAIN
&& cfun
->can_throw_non_call_exceptions
)
8178 addr
= gen_rtx_MEM (BLKmode
, gen_rtx_SCRATCH (VOIDmode
));
8179 emit_clobber (addr
);
8183 /* Save fprs 8 - 15 (64 bit ABI). */
8185 if (cfun_save_high_fprs_p
&& next_fpr
)
8187 /* If the stack might be accessed through a different register
8188 we have to make sure that the stack pointer decrement is not
8189 moved below the use of the stack slots. */
8190 s390_emit_stack_tie ();
8192 insn
= emit_insn (gen_add2_insn (temp_reg
,
8193 GEN_INT (cfun_frame_layout
.f8_offset
)));
8197 for (i
= 24; i
<= next_fpr
; i
++)
8198 if (cfun_fpr_bit_p (i
- 16))
8200 rtx addr
= plus_constant (Pmode
, stack_pointer_rtx
,
8201 cfun_frame_layout
.frame_size
8202 + cfun_frame_layout
.f8_offset
8205 insn
= save_fpr (temp_reg
, offset
, i
);
8207 RTX_FRAME_RELATED_P (insn
) = 1;
8208 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8209 gen_rtx_SET (VOIDmode
,
8210 gen_rtx_MEM (DFmode
, addr
),
8211 gen_rtx_REG (DFmode
, i
)));
8215 /* Set frame pointer, if needed. */
8217 if (frame_pointer_needed
)
8219 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
8220 RTX_FRAME_RELATED_P (insn
) = 1;
8223 /* Set up got pointer, if needed. */
8225 if (flag_pic
&& df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
8227 rtx insns
= s390_load_got ();
8229 for (insn
= insns
; insn
; insn
= NEXT_INSN (insn
))
8230 annotate_constant_pool_refs (&PATTERN (insn
));
8235 if (TARGET_TPF_PROFILING
)
8237 /* Generate a BAS instruction to serve as a function
8238 entry intercept to facilitate the use of tracing
8239 algorithms located at the branch target. */
8240 emit_insn (gen_prologue_tpf ());
8242 /* Emit a blockage here so that all code
8243 lies between the profiling mechanisms. */
8244 emit_insn (gen_blockage ());
8248 /* Expand the epilogue into a bunch of separate insns. */
8251 s390_emit_epilogue (bool sibcall
)
8253 rtx frame_pointer
, return_reg
, cfa_restores
= NULL_RTX
;
8254 int area_bottom
, area_top
, offset
= 0;
8259 if (TARGET_TPF_PROFILING
)
8262 /* Generate a BAS instruction to serve as a function
8263 entry intercept to facilitate the use of tracing
8264 algorithms located at the branch target. */
8266 /* Emit a blockage here so that all code
8267 lies between the profiling mechanisms. */
8268 emit_insn (gen_blockage ());
8270 emit_insn (gen_epilogue_tpf ());
8273 /* Check whether to use frame or stack pointer for restore. */
8275 frame_pointer
= (frame_pointer_needed
8276 ? hard_frame_pointer_rtx
: stack_pointer_rtx
);
8278 s390_frame_area (&area_bottom
, &area_top
);
8280 /* Check whether we can access the register save area.
8281 If not, increment the frame pointer as required. */
8283 if (area_top
<= area_bottom
)
8285 /* Nothing to restore. */
8287 else if (DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_bottom
)
8288 && DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_top
- 1))
8290 /* Area is in range. */
8291 offset
= cfun_frame_layout
.frame_size
;
8295 rtx insn
, frame_off
, cfa
;
8297 offset
= area_bottom
< 0 ? -area_bottom
: 0;
8298 frame_off
= GEN_INT (cfun_frame_layout
.frame_size
- offset
);
8300 cfa
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8301 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8302 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8304 insn
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8305 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8306 insn
= emit_insn (insn
);
8310 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8311 frame_off
= force_const_mem (Pmode
, frame_off
);
8313 insn
= emit_insn (gen_add2_insn (frame_pointer
, frame_off
));
8314 annotate_constant_pool_refs (&PATTERN (insn
));
8316 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, cfa
);
8317 RTX_FRAME_RELATED_P (insn
) = 1;
8320 /* Restore call saved fprs. */
8324 if (cfun_save_high_fprs_p
)
8326 next_offset
= cfun_frame_layout
.f8_offset
;
8327 for (i
= 24; i
< 32; i
++)
8329 if (cfun_fpr_bit_p (i
- 16))
8331 restore_fpr (frame_pointer
,
8332 offset
+ next_offset
, i
);
8334 = alloc_reg_note (REG_CFA_RESTORE
,
8335 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8344 next_offset
= cfun_frame_layout
.f4_offset
;
8345 for (i
= 18; i
< 20; i
++)
8347 if (cfun_fpr_bit_p (i
- 16))
8349 restore_fpr (frame_pointer
,
8350 offset
+ next_offset
, i
);
8352 = alloc_reg_note (REG_CFA_RESTORE
,
8353 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8356 else if (!TARGET_PACKED_STACK
)
8362 /* Return register. */
8364 return_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8366 /* Restore call saved gprs. */
8368 if (cfun_frame_layout
.first_restore_gpr
!= -1)
8373 /* Check for global register and save them
8374 to stack location from where they get restored. */
8376 for (i
= cfun_frame_layout
.first_restore_gpr
;
8377 i
<= cfun_frame_layout
.last_restore_gpr
;
8380 if (global_not_special_regno_p (i
))
8382 addr
= plus_constant (Pmode
, frame_pointer
,
8383 offset
+ cfun_frame_layout
.gprs_offset
8384 + (i
- cfun_frame_layout
.first_save_gpr_slot
)
8386 addr
= gen_rtx_MEM (Pmode
, addr
);
8387 set_mem_alias_set (addr
, get_frame_alias_set ());
8388 emit_move_insn (addr
, gen_rtx_REG (Pmode
, i
));
8392 = alloc_reg_note (REG_CFA_RESTORE
,
8393 gen_rtx_REG (Pmode
, i
), cfa_restores
);
8398 /* Fetch return address from stack before load multiple,
8399 this will do good for scheduling. */
8401 if (cfun_frame_layout
.save_return_addr_p
8402 || (cfun_frame_layout
.first_restore_gpr
< BASE_REGNUM
8403 && cfun_frame_layout
.last_restore_gpr
> RETURN_REGNUM
))
8405 int return_regnum
= find_unused_clobbered_reg();
8408 return_reg
= gen_rtx_REG (Pmode
, return_regnum
);
8410 addr
= plus_constant (Pmode
, frame_pointer
,
8411 offset
+ cfun_frame_layout
.gprs_offset
8413 - cfun_frame_layout
.first_save_gpr_slot
)
8415 addr
= gen_rtx_MEM (Pmode
, addr
);
8416 set_mem_alias_set (addr
, get_frame_alias_set ());
8417 emit_move_insn (return_reg
, addr
);
8421 insn
= restore_gprs (frame_pointer
,
8422 offset
+ cfun_frame_layout
.gprs_offset
8423 + (cfun_frame_layout
.first_restore_gpr
8424 - cfun_frame_layout
.first_save_gpr_slot
)
8426 cfun_frame_layout
.first_restore_gpr
,
8427 cfun_frame_layout
.last_restore_gpr
);
8428 insn
= emit_insn (insn
);
8429 REG_NOTES (insn
) = cfa_restores
;
8430 add_reg_note (insn
, REG_CFA_DEF_CFA
,
8431 plus_constant (Pmode
, stack_pointer_rtx
,
8432 STACK_POINTER_OFFSET
));
8433 RTX_FRAME_RELATED_P (insn
) = 1;
8439 /* Return to caller. */
8441 p
= rtvec_alloc (2);
8443 RTVEC_ELT (p
, 0) = ret_rtx
;
8444 RTVEC_ELT (p
, 1) = gen_rtx_USE (VOIDmode
, return_reg
);
8445 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
8450 /* Return the size in bytes of a function argument of
8451 type TYPE and/or mode MODE. At least one of TYPE or
8452 MODE must be specified. */
8455 s390_function_arg_size (enum machine_mode mode
, const_tree type
)
8458 return int_size_in_bytes (type
);
8460 /* No type info available for some library calls ... */
8461 if (mode
!= BLKmode
)
8462 return GET_MODE_SIZE (mode
);
8464 /* If we have neither type nor mode, abort */
8468 /* Return true if a function argument of type TYPE and mode MODE
8469 is to be passed in a floating-point register, if available. */
8472 s390_function_arg_float (enum machine_mode mode
, const_tree type
)
8474 int size
= s390_function_arg_size (mode
, type
);
8478 /* Soft-float changes the ABI: no floating-point registers are used. */
8479 if (TARGET_SOFT_FLOAT
)
8482 /* No type info available for some library calls ... */
8484 return mode
== SFmode
|| mode
== DFmode
|| mode
== SDmode
|| mode
== DDmode
;
8486 /* The ABI says that record types with a single member are treated
8487 just like that member would be. */
8488 while (TREE_CODE (type
) == RECORD_TYPE
)
8490 tree field
, single
= NULL_TREE
;
8492 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
8494 if (TREE_CODE (field
) != FIELD_DECL
)
8497 if (single
== NULL_TREE
)
8498 single
= TREE_TYPE (field
);
8503 if (single
== NULL_TREE
)
8509 return TREE_CODE (type
) == REAL_TYPE
;
8512 /* Return true if a function argument of type TYPE and mode MODE
8513 is to be passed in an integer register, or a pair of integer
8514 registers, if available. */
8517 s390_function_arg_integer (enum machine_mode mode
, const_tree type
)
8519 int size
= s390_function_arg_size (mode
, type
);
8523 /* No type info available for some library calls ... */
8525 return GET_MODE_CLASS (mode
) == MODE_INT
8526 || (TARGET_SOFT_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
));
8528 /* We accept small integral (and similar) types. */
8529 if (INTEGRAL_TYPE_P (type
)
8530 || POINTER_TYPE_P (type
)
8531 || TREE_CODE (type
) == NULLPTR_TYPE
8532 || TREE_CODE (type
) == OFFSET_TYPE
8533 || (TARGET_SOFT_FLOAT
&& TREE_CODE (type
) == REAL_TYPE
))
8536 /* We also accept structs of size 1, 2, 4, 8 that are not
8537 passed in floating-point registers. */
8538 if (AGGREGATE_TYPE_P (type
)
8539 && exact_log2 (size
) >= 0
8540 && !s390_function_arg_float (mode
, type
))
8546 /* Return 1 if a function argument of type TYPE and mode MODE
8547 is to be passed by reference. The ABI specifies that only
8548 structures of size 1, 2, 4, or 8 bytes are passed by value,
8549 all other structures (and complex numbers) are passed by
8553 s390_pass_by_reference (cumulative_args_t ca ATTRIBUTE_UNUSED
,
8554 enum machine_mode mode
, const_tree type
,
8555 bool named ATTRIBUTE_UNUSED
)
8557 int size
= s390_function_arg_size (mode
, type
);
8563 if (AGGREGATE_TYPE_P (type
) && exact_log2 (size
) < 0)
8566 if (TREE_CODE (type
) == COMPLEX_TYPE
8567 || TREE_CODE (type
) == VECTOR_TYPE
)
8574 /* Update the data in CUM to advance over an argument of mode MODE and
8575 data type TYPE. (TYPE is null for libcalls where that information
8576 may not be available.). The boolean NAMED specifies whether the
8577 argument is a named argument (as opposed to an unnamed argument
8578 matching an ellipsis). */
8581 s390_function_arg_advance (cumulative_args_t cum_v
, enum machine_mode mode
,
8582 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8584 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8586 if (s390_function_arg_float (mode
, type
))
8590 else if (s390_function_arg_integer (mode
, type
))
8592 int size
= s390_function_arg_size (mode
, type
);
8593 cum
->gprs
+= ((size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
);
8599 /* Define where to put the arguments to a function.
8600 Value is zero to push the argument on the stack,
8601 or a hard register in which to store the argument.
8603 MODE is the argument's machine mode.
8604 TYPE is the data type of the argument (as a tree).
8605 This is null for libcalls where that information may
8607 CUM is a variable of type CUMULATIVE_ARGS which gives info about
8608 the preceding args and about the function being called.
8609 NAMED is nonzero if this argument is a named parameter
8610 (otherwise it is an extra parameter matching an ellipsis).
8612 On S/390, we use general purpose registers 2 through 6 to
8613 pass integer, pointer, and certain structure arguments, and
8614 floating point registers 0 and 2 (0, 2, 4, and 6 on 64-bit)
8615 to pass floating point arguments. All remaining arguments
8616 are pushed to the stack. */
8619 s390_function_arg (cumulative_args_t cum_v
, enum machine_mode mode
,
8620 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8622 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8624 if (s390_function_arg_float (mode
, type
))
8626 if (cum
->fprs
+ 1 > FP_ARG_NUM_REG
)
8629 return gen_rtx_REG (mode
, cum
->fprs
+ 16);
8631 else if (s390_function_arg_integer (mode
, type
))
8633 int size
= s390_function_arg_size (mode
, type
);
8634 int n_gprs
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8636 if (cum
->gprs
+ n_gprs
> GP_ARG_NUM_REG
)
8638 else if (n_gprs
== 1 || UNITS_PER_WORD
== UNITS_PER_LONG
)
8639 return gen_rtx_REG (mode
, cum
->gprs
+ 2);
8640 else if (n_gprs
== 2)
8642 rtvec p
= rtvec_alloc (2);
8645 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 2),
8648 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 3),
8651 return gen_rtx_PARALLEL (mode
, p
);
8655 /* After the real arguments, expand_call calls us once again
8656 with a void_type_node type. Whatever we return here is
8657 passed as operand 2 to the call expanders.
8659 We don't need this feature ... */
8660 else if (type
== void_type_node
)
8666 /* Return true if return values of type TYPE should be returned
8667 in a memory buffer whose address is passed by the caller as
8668 hidden first argument. */
8671 s390_return_in_memory (const_tree type
, const_tree fundecl ATTRIBUTE_UNUSED
)
8673 /* We accept small integral (and similar) types. */
8674 if (INTEGRAL_TYPE_P (type
)
8675 || POINTER_TYPE_P (type
)
8676 || TREE_CODE (type
) == OFFSET_TYPE
8677 || TREE_CODE (type
) == REAL_TYPE
)
8678 return int_size_in_bytes (type
) > 8;
8680 /* Aggregates and similar constructs are always returned
8682 if (AGGREGATE_TYPE_P (type
)
8683 || TREE_CODE (type
) == COMPLEX_TYPE
8684 || TREE_CODE (type
) == VECTOR_TYPE
)
8687 /* ??? We get called on all sorts of random stuff from
8688 aggregate_value_p. We can't abort, but it's not clear
8689 what's safe to return. Pretend it's a struct I guess. */
8693 /* Function arguments and return values are promoted to word size. */
8695 static enum machine_mode
8696 s390_promote_function_mode (const_tree type
, enum machine_mode mode
,
8698 const_tree fntype ATTRIBUTE_UNUSED
,
8699 int for_return ATTRIBUTE_UNUSED
)
8701 if (INTEGRAL_MODE_P (mode
)
8702 && GET_MODE_SIZE (mode
) < UNITS_PER_LONG
)
8704 if (type
!= NULL_TREE
&& POINTER_TYPE_P (type
))
8705 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
8712 /* Define where to return a (scalar) value of type RET_TYPE.
8713 If RET_TYPE is null, define where to return a (scalar)
8714 value of mode MODE from a libcall. */
8717 s390_function_and_libcall_value (enum machine_mode mode
,
8718 const_tree ret_type
,
8719 const_tree fntype_or_decl
,
8720 bool outgoing ATTRIBUTE_UNUSED
)
8722 /* For normal functions perform the promotion as
8723 promote_function_mode would do. */
8726 int unsignedp
= TYPE_UNSIGNED (ret_type
);
8727 mode
= promote_function_mode (ret_type
, mode
, &unsignedp
,
8731 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
|| SCALAR_FLOAT_MODE_P (mode
));
8732 gcc_assert (GET_MODE_SIZE (mode
) <= 8);
8734 if (TARGET_HARD_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
))
8735 return gen_rtx_REG (mode
, 16);
8736 else if (GET_MODE_SIZE (mode
) <= UNITS_PER_LONG
8737 || UNITS_PER_LONG
== UNITS_PER_WORD
)
8738 return gen_rtx_REG (mode
, 2);
8739 else if (GET_MODE_SIZE (mode
) == 2 * UNITS_PER_LONG
)
8741 /* This case is triggered when returning a 64 bit value with
8742 -m31 -mzarch. Although the value would fit into a single
8743 register it has to be forced into a 32 bit register pair in
8744 order to match the ABI. */
8745 rtvec p
= rtvec_alloc (2);
8748 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 2), const0_rtx
);
8750 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 3), GEN_INT (4));
8752 return gen_rtx_PARALLEL (mode
, p
);
8758 /* Define where to return a scalar return value of type RET_TYPE. */
8761 s390_function_value (const_tree ret_type
, const_tree fn_decl_or_type
,
8764 return s390_function_and_libcall_value (TYPE_MODE (ret_type
), ret_type
,
8765 fn_decl_or_type
, outgoing
);
8768 /* Define where to return a scalar libcall return value of mode
8772 s390_libcall_value (enum machine_mode mode
, const_rtx fun ATTRIBUTE_UNUSED
)
8774 return s390_function_and_libcall_value (mode
, NULL_TREE
,
8779 /* Create and return the va_list datatype.
8781 On S/390, va_list is an array type equivalent to
8783 typedef struct __va_list_tag
8787 void *__overflow_arg_area;
8788 void *__reg_save_area;
8791 where __gpr and __fpr hold the number of general purpose
8792 or floating point arguments used up to now, respectively,
8793 __overflow_arg_area points to the stack location of the
8794 next argument passed on the stack, and __reg_save_area
8795 always points to the start of the register area in the
8796 call frame of the current function. The function prologue
8797 saves all registers used for argument passing into this
8798 area if the function uses variable arguments. */
8801 s390_build_builtin_va_list (void)
8803 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
8805 record
= lang_hooks
.types
.make_type (RECORD_TYPE
);
8808 build_decl (BUILTINS_LOCATION
,
8809 TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
8811 f_gpr
= build_decl (BUILTINS_LOCATION
,
8812 FIELD_DECL
, get_identifier ("__gpr"),
8813 long_integer_type_node
);
8814 f_fpr
= build_decl (BUILTINS_LOCATION
,
8815 FIELD_DECL
, get_identifier ("__fpr"),
8816 long_integer_type_node
);
8817 f_ovf
= build_decl (BUILTINS_LOCATION
,
8818 FIELD_DECL
, get_identifier ("__overflow_arg_area"),
8820 f_sav
= build_decl (BUILTINS_LOCATION
,
8821 FIELD_DECL
, get_identifier ("__reg_save_area"),
8824 va_list_gpr_counter_field
= f_gpr
;
8825 va_list_fpr_counter_field
= f_fpr
;
8827 DECL_FIELD_CONTEXT (f_gpr
) = record
;
8828 DECL_FIELD_CONTEXT (f_fpr
) = record
;
8829 DECL_FIELD_CONTEXT (f_ovf
) = record
;
8830 DECL_FIELD_CONTEXT (f_sav
) = record
;
8832 TYPE_STUB_DECL (record
) = type_decl
;
8833 TYPE_NAME (record
) = type_decl
;
8834 TYPE_FIELDS (record
) = f_gpr
;
8835 DECL_CHAIN (f_gpr
) = f_fpr
;
8836 DECL_CHAIN (f_fpr
) = f_ovf
;
8837 DECL_CHAIN (f_ovf
) = f_sav
;
8839 layout_type (record
);
8841 /* The correct type is an array type of one element. */
8842 return build_array_type (record
, build_index_type (size_zero_node
));
8845 /* Implement va_start by filling the va_list structure VALIST.
8846 STDARG_P is always true, and ignored.
8847 NEXTARG points to the first anonymous stack argument.
8849 The following global variables are used to initialize
8850 the va_list structure:
8853 holds number of gprs and fprs used for named arguments.
8854 crtl->args.arg_offset_rtx:
8855 holds the offset of the first anonymous stack argument
8856 (relative to the virtual arg pointer). */
8859 s390_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
8861 HOST_WIDE_INT n_gpr
, n_fpr
;
8863 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8864 tree gpr
, fpr
, ovf
, sav
, t
;
8866 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8867 f_fpr
= DECL_CHAIN (f_gpr
);
8868 f_ovf
= DECL_CHAIN (f_fpr
);
8869 f_sav
= DECL_CHAIN (f_ovf
);
8871 valist
= build_simple_mem_ref (valist
);
8872 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8873 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8874 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8875 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8877 /* Count number of gp and fp argument registers used. */
8879 n_gpr
= crtl
->args
.info
.gprs
;
8880 n_fpr
= crtl
->args
.info
.fprs
;
8882 if (cfun
->va_list_gpr_size
)
8884 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gpr
), gpr
,
8885 build_int_cst (NULL_TREE
, n_gpr
));
8886 TREE_SIDE_EFFECTS (t
) = 1;
8887 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8890 if (cfun
->va_list_fpr_size
)
8892 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fpr
), fpr
,
8893 build_int_cst (NULL_TREE
, n_fpr
));
8894 TREE_SIDE_EFFECTS (t
) = 1;
8895 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8898 /* Find the overflow area. */
8899 if (n_gpr
+ cfun
->va_list_gpr_size
> GP_ARG_NUM_REG
8900 || n_fpr
+ cfun
->va_list_fpr_size
> FP_ARG_NUM_REG
)
8902 t
= make_tree (TREE_TYPE (ovf
), virtual_incoming_args_rtx
);
8904 off
= INTVAL (crtl
->args
.arg_offset_rtx
);
8905 off
= off
< 0 ? 0 : off
;
8906 if (TARGET_DEBUG_ARG
)
8907 fprintf (stderr
, "va_start: n_gpr = %d, n_fpr = %d off %d\n",
8908 (int)n_gpr
, (int)n_fpr
, off
);
8910 t
= fold_build_pointer_plus_hwi (t
, off
);
8912 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovf
), ovf
, t
);
8913 TREE_SIDE_EFFECTS (t
) = 1;
8914 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8917 /* Find the register save area. */
8918 if ((cfun
->va_list_gpr_size
&& n_gpr
< GP_ARG_NUM_REG
)
8919 || (cfun
->va_list_fpr_size
&& n_fpr
< FP_ARG_NUM_REG
))
8921 t
= make_tree (TREE_TYPE (sav
), return_address_pointer_rtx
);
8922 t
= fold_build_pointer_plus_hwi (t
, -RETURN_REGNUM
* UNITS_PER_LONG
);
8924 t
= build2 (MODIFY_EXPR
, TREE_TYPE (sav
), sav
, t
);
8925 TREE_SIDE_EFFECTS (t
) = 1;
8926 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8930 /* Implement va_arg by updating the va_list structure
8931 VALIST as required to retrieve an argument of type
8932 TYPE, and returning that argument.
8934 Generates code equivalent to:
8936 if (integral value) {
8937 if (size <= 4 && args.gpr < 5 ||
8938 size > 4 && args.gpr < 4 )
8939 ret = args.reg_save_area[args.gpr+8]
8941 ret = *args.overflow_arg_area++;
8942 } else if (float value) {
8944 ret = args.reg_save_area[args.fpr+64]
8946 ret = *args.overflow_arg_area++;
8947 } else if (aggregate value) {
8949 ret = *args.reg_save_area[args.gpr]
8951 ret = **args.overflow_arg_area++;
8955 s390_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
8956 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
8958 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8959 tree gpr
, fpr
, ovf
, sav
, reg
, t
, u
;
8960 int indirect_p
, size
, n_reg
, sav_ofs
, sav_scale
, max_reg
;
8961 tree lab_false
, lab_over
, addr
;
8963 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8964 f_fpr
= DECL_CHAIN (f_gpr
);
8965 f_ovf
= DECL_CHAIN (f_fpr
);
8966 f_sav
= DECL_CHAIN (f_ovf
);
8968 valist
= build_va_arg_indirect_ref (valist
);
8969 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8970 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8971 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8973 /* The tree for args* cannot be shared between gpr/fpr and ovf since
8974 both appear on a lhs. */
8975 valist
= unshare_expr (valist
);
8976 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8978 size
= int_size_in_bytes (type
);
8980 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
8982 if (TARGET_DEBUG_ARG
)
8984 fprintf (stderr
, "va_arg: aggregate type");
8988 /* Aggregates are passed by reference. */
8993 /* kernel stack layout on 31 bit: It is assumed here that no padding
8994 will be added by s390_frame_info because for va_args always an even
8995 number of gprs has to be saved r15-r2 = 14 regs. */
8996 sav_ofs
= 2 * UNITS_PER_LONG
;
8997 sav_scale
= UNITS_PER_LONG
;
8998 size
= UNITS_PER_LONG
;
8999 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9001 else if (s390_function_arg_float (TYPE_MODE (type
), type
))
9003 if (TARGET_DEBUG_ARG
)
9005 fprintf (stderr
, "va_arg: float type");
9009 /* FP args go in FP registers, if present. */
9013 sav_ofs
= 16 * UNITS_PER_LONG
;
9015 max_reg
= FP_ARG_NUM_REG
- n_reg
;
9019 if (TARGET_DEBUG_ARG
)
9021 fprintf (stderr
, "va_arg: other type");
9025 /* Otherwise into GP registers. */
9028 n_reg
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
9030 /* kernel stack layout on 31 bit: It is assumed here that no padding
9031 will be added by s390_frame_info because for va_args always an even
9032 number of gprs has to be saved r15-r2 = 14 regs. */
9033 sav_ofs
= 2 * UNITS_PER_LONG
;
9035 if (size
< UNITS_PER_LONG
)
9036 sav_ofs
+= UNITS_PER_LONG
- size
;
9038 sav_scale
= UNITS_PER_LONG
;
9039 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9042 /* Pull the value out of the saved registers ... */
9044 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
9045 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
9046 addr
= create_tmp_var (ptr_type_node
, "addr");
9047 mark_sym_for_renaming (addr
);
9049 t
= fold_convert (TREE_TYPE (reg
), size_int (max_reg
));
9050 t
= build2 (GT_EXPR
, boolean_type_node
, reg
, t
);
9051 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
9052 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
9053 gimplify_and_add (t
, pre_p
);
9055 t
= fold_build_pointer_plus_hwi (sav
, sav_ofs
);
9056 u
= build2 (MULT_EXPR
, TREE_TYPE (reg
), reg
,
9057 fold_convert (TREE_TYPE (reg
), size_int (sav_scale
)));
9058 t
= fold_build_pointer_plus (t
, u
);
9060 gimplify_assign (addr
, t
, pre_p
);
9062 gimple_seq_add_stmt (pre_p
, gimple_build_goto (lab_over
));
9064 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_false
));
9067 /* ... Otherwise out of the overflow area. */
9070 if (size
< UNITS_PER_LONG
)
9071 t
= fold_build_pointer_plus_hwi (t
, UNITS_PER_LONG
- size
);
9073 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
9075 gimplify_assign (addr
, t
, pre_p
);
9077 t
= fold_build_pointer_plus_hwi (t
, size
);
9078 gimplify_assign (ovf
, t
, pre_p
);
9080 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_over
));
9083 /* Increment register save count. */
9085 u
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (reg
), reg
,
9086 fold_convert (TREE_TYPE (reg
), size_int (n_reg
)));
9087 gimplify_and_add (u
, pre_p
);
9091 t
= build_pointer_type_for_mode (build_pointer_type (type
),
9093 addr
= fold_convert (t
, addr
);
9094 addr
= build_va_arg_indirect_ref (addr
);
9098 t
= build_pointer_type_for_mode (type
, ptr_mode
, true);
9099 addr
= fold_convert (t
, addr
);
9102 return build_va_arg_indirect_ref (addr
);
9110 S390_BUILTIN_THREAD_POINTER
,
9111 S390_BUILTIN_SET_THREAD_POINTER
,
9116 static enum insn_code
const code_for_builtin_64
[S390_BUILTIN_max
] = {
9121 static enum insn_code
const code_for_builtin_31
[S390_BUILTIN_max
] = {
9127 s390_init_builtins (void)
9131 ftype
= build_function_type_list (ptr_type_node
, NULL_TREE
);
9132 add_builtin_function ("__builtin_thread_pointer", ftype
,
9133 S390_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
9136 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
9137 add_builtin_function ("__builtin_set_thread_pointer", ftype
,
9138 S390_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
9142 /* Expand an expression EXP that calls a built-in function,
9143 with result going to TARGET if that's convenient
9144 (and in mode MODE if that's convenient).
9145 SUBTARGET may be used as the target for computing one of EXP's operands.
9146 IGNORE is nonzero if the value is to be ignored. */
9149 s390_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
9150 enum machine_mode mode ATTRIBUTE_UNUSED
,
9151 int ignore ATTRIBUTE_UNUSED
)
9155 enum insn_code
const *code_for_builtin
=
9156 TARGET_64BIT
? code_for_builtin_64
: code_for_builtin_31
;
9158 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
9159 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
9160 enum insn_code icode
;
9161 rtx op
[MAX_ARGS
], pat
;
9165 call_expr_arg_iterator iter
;
9167 if (fcode
>= S390_BUILTIN_max
)
9168 internal_error ("bad builtin fcode");
9169 icode
= code_for_builtin
[fcode
];
9171 internal_error ("bad builtin fcode");
9173 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
9176 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
9178 const struct insn_operand_data
*insn_op
;
9180 if (arg
== error_mark_node
)
9182 if (arity
> MAX_ARGS
)
9185 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
9187 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
9189 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
9190 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
9196 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
9198 || GET_MODE (target
) != tmode
9199 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
9200 target
= gen_reg_rtx (tmode
);
9206 pat
= GEN_FCN (icode
) (target
);
9210 pat
= GEN_FCN (icode
) (target
, op
[0]);
9212 pat
= GEN_FCN (icode
) (op
[0]);
9215 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
9231 /* Output assembly code for the trampoline template to
9234 On S/390, we use gpr 1 internally in the trampoline code;
9235 gpr 0 is used to hold the static chain. */
9238 s390_asm_trampoline_template (FILE *file
)
9241 op
[0] = gen_rtx_REG (Pmode
, 0);
9242 op
[1] = gen_rtx_REG (Pmode
, 1);
9246 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9247 output_asm_insn ("lmg\t%0,%1,14(%1)", op
); /* 6 byte */
9248 output_asm_insn ("br\t%1", op
); /* 2 byte */
9249 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 10));
9253 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9254 output_asm_insn ("lm\t%0,%1,6(%1)", op
); /* 4 byte */
9255 output_asm_insn ("br\t%1", op
); /* 2 byte */
9256 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 8));
9260 /* Emit RTL insns to initialize the variable parts of a trampoline.
9261 FNADDR is an RTX for the address of the function's pure code.
9262 CXT is an RTX for the static chain value for the function. */
9265 s390_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9267 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
9270 emit_block_move (m_tramp
, assemble_trampoline_template (),
9271 GEN_INT (2 * UNITS_PER_LONG
), BLOCK_OP_NORMAL
);
9273 mem
= adjust_address (m_tramp
, Pmode
, 2 * UNITS_PER_LONG
);
9274 emit_move_insn (mem
, cxt
);
9275 mem
= adjust_address (m_tramp
, Pmode
, 3 * UNITS_PER_LONG
);
9276 emit_move_insn (mem
, fnaddr
);
9279 /* Output assembler code to FILE to increment profiler label # LABELNO
9280 for profiling a function entry. */
9283 s390_function_profiler (FILE *file
, int labelno
)
9288 ASM_GENERATE_INTERNAL_LABEL (label
, "LP", labelno
);
9290 fprintf (file
, "# function profiler \n");
9292 op
[0] = gen_rtx_REG (Pmode
, RETURN_REGNUM
);
9293 op
[1] = gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
);
9294 op
[1] = gen_rtx_MEM (Pmode
, plus_constant (Pmode
, op
[1], UNITS_PER_LONG
));
9296 op
[2] = gen_rtx_REG (Pmode
, 1);
9297 op
[3] = gen_rtx_SYMBOL_REF (Pmode
, label
);
9298 SYMBOL_REF_FLAGS (op
[3]) = SYMBOL_FLAG_LOCAL
;
9300 op
[4] = gen_rtx_SYMBOL_REF (Pmode
, "_mcount");
9303 op
[4] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[4]), UNSPEC_PLT
);
9304 op
[4] = gen_rtx_CONST (Pmode
, op
[4]);
9309 output_asm_insn ("stg\t%0,%1", op
);
9310 output_asm_insn ("larl\t%2,%3", op
);
9311 output_asm_insn ("brasl\t%0,%4", op
);
9312 output_asm_insn ("lg\t%0,%1", op
);
9316 op
[6] = gen_label_rtx ();
9318 output_asm_insn ("st\t%0,%1", op
);
9319 output_asm_insn ("bras\t%2,%l6", op
);
9320 output_asm_insn (".long\t%4", op
);
9321 output_asm_insn (".long\t%3", op
);
9322 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9323 output_asm_insn ("l\t%0,0(%2)", op
);
9324 output_asm_insn ("l\t%2,4(%2)", op
);
9325 output_asm_insn ("basr\t%0,%0", op
);
9326 output_asm_insn ("l\t%0,%1", op
);
9330 op
[5] = gen_label_rtx ();
9331 op
[6] = gen_label_rtx ();
9333 output_asm_insn ("st\t%0,%1", op
);
9334 output_asm_insn ("bras\t%2,%l6", op
);
9335 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[5]));
9336 output_asm_insn (".long\t%4-%l5", op
);
9337 output_asm_insn (".long\t%3-%l5", op
);
9338 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9339 output_asm_insn ("lr\t%0,%2", op
);
9340 output_asm_insn ("a\t%0,0(%2)", op
);
9341 output_asm_insn ("a\t%2,4(%2)", op
);
9342 output_asm_insn ("basr\t%0,%0", op
);
9343 output_asm_insn ("l\t%0,%1", op
);
9347 /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF
9348 into its SYMBOL_REF_FLAGS. */
9351 s390_encode_section_info (tree decl
, rtx rtl
, int first
)
9353 default_encode_section_info (decl
, rtl
, first
);
9355 if (TREE_CODE (decl
) == VAR_DECL
)
9357 /* If a variable has a forced alignment to < 2 bytes, mark it
9358 with SYMBOL_FLAG_ALIGN1 to prevent it from being used as LARL
9360 if (DECL_USER_ALIGN (decl
) && DECL_ALIGN (decl
) < 16)
9361 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_ALIGN1
;
9362 if (!DECL_SIZE (decl
)
9363 || !DECL_ALIGN (decl
)
9364 || !host_integerp (DECL_SIZE (decl
), 0)
9365 || (DECL_ALIGN (decl
) <= 64
9366 && DECL_ALIGN (decl
) != tree_low_cst (DECL_SIZE (decl
), 0)))
9367 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9370 /* Literal pool references don't have a decl so they are handled
9371 differently here. We rely on the information in the MEM_ALIGN
9372 entry to decide upon natural alignment. */
9374 && GET_CODE (XEXP (rtl
, 0)) == SYMBOL_REF
9375 && TREE_CONSTANT_POOL_ADDRESS_P (XEXP (rtl
, 0))
9376 && (MEM_ALIGN (rtl
) == 0
9377 || GET_MODE_BITSIZE (GET_MODE (rtl
)) == 0
9378 || MEM_ALIGN (rtl
) < GET_MODE_BITSIZE (GET_MODE (rtl
))))
9379 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9382 /* Output thunk to FILE that implements a C++ virtual function call (with
9383 multiple inheritance) to FUNCTION. The thunk adjusts the this pointer
9384 by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment
9385 stored at VCALL_OFFSET in the vtable whose address is located at offset 0
9386 relative to the resulting this pointer. */
9389 s390_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
9390 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
9396 /* Make sure unwind info is emitted for the thunk if needed. */
9397 final_start_function (emit_barrier (), file
, 1);
9399 /* Operand 0 is the target function. */
9400 op
[0] = XEXP (DECL_RTL (function
), 0);
9401 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (op
[0]))
9404 op
[0] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[0]),
9405 TARGET_64BIT
? UNSPEC_PLT
: UNSPEC_GOT
);
9406 op
[0] = gen_rtx_CONST (Pmode
, op
[0]);
9409 /* Operand 1 is the 'this' pointer. */
9410 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
9411 op
[1] = gen_rtx_REG (Pmode
, 3);
9413 op
[1] = gen_rtx_REG (Pmode
, 2);
9415 /* Operand 2 is the delta. */
9416 op
[2] = GEN_INT (delta
);
9418 /* Operand 3 is the vcall_offset. */
9419 op
[3] = GEN_INT (vcall_offset
);
9421 /* Operand 4 is the temporary register. */
9422 op
[4] = gen_rtx_REG (Pmode
, 1);
9424 /* Operands 5 to 8 can be used as labels. */
9430 /* Operand 9 can be used for temporary register. */
9433 /* Generate code. */
9436 /* Setup literal pool pointer if required. */
9437 if ((!DISP_IN_RANGE (delta
)
9438 && !CONST_OK_FOR_K (delta
)
9439 && !CONST_OK_FOR_Os (delta
))
9440 || (!DISP_IN_RANGE (vcall_offset
)
9441 && !CONST_OK_FOR_K (vcall_offset
)
9442 && !CONST_OK_FOR_Os (vcall_offset
)))
9444 op
[5] = gen_label_rtx ();
9445 output_asm_insn ("larl\t%4,%5", op
);
9448 /* Add DELTA to this pointer. */
9451 if (CONST_OK_FOR_J (delta
))
9452 output_asm_insn ("la\t%1,%2(%1)", op
);
9453 else if (DISP_IN_RANGE (delta
))
9454 output_asm_insn ("lay\t%1,%2(%1)", op
);
9455 else if (CONST_OK_FOR_K (delta
))
9456 output_asm_insn ("aghi\t%1,%2", op
);
9457 else if (CONST_OK_FOR_Os (delta
))
9458 output_asm_insn ("agfi\t%1,%2", op
);
9461 op
[6] = gen_label_rtx ();
9462 output_asm_insn ("agf\t%1,%6-%5(%4)", op
);
9466 /* Perform vcall adjustment. */
9469 if (DISP_IN_RANGE (vcall_offset
))
9471 output_asm_insn ("lg\t%4,0(%1)", op
);
9472 output_asm_insn ("ag\t%1,%3(%4)", op
);
9474 else if (CONST_OK_FOR_K (vcall_offset
))
9476 output_asm_insn ("lghi\t%4,%3", op
);
9477 output_asm_insn ("ag\t%4,0(%1)", op
);
9478 output_asm_insn ("ag\t%1,0(%4)", op
);
9480 else if (CONST_OK_FOR_Os (vcall_offset
))
9482 output_asm_insn ("lgfi\t%4,%3", op
);
9483 output_asm_insn ("ag\t%4,0(%1)", op
);
9484 output_asm_insn ("ag\t%1,0(%4)", op
);
9488 op
[7] = gen_label_rtx ();
9489 output_asm_insn ("llgf\t%4,%7-%5(%4)", op
);
9490 output_asm_insn ("ag\t%4,0(%1)", op
);
9491 output_asm_insn ("ag\t%1,0(%4)", op
);
9495 /* Jump to target. */
9496 output_asm_insn ("jg\t%0", op
);
9498 /* Output literal pool if required. */
9501 output_asm_insn (".align\t4", op
);
9502 targetm
.asm_out
.internal_label (file
, "L",
9503 CODE_LABEL_NUMBER (op
[5]));
9507 targetm
.asm_out
.internal_label (file
, "L",
9508 CODE_LABEL_NUMBER (op
[6]));
9509 output_asm_insn (".long\t%2", op
);
9513 targetm
.asm_out
.internal_label (file
, "L",
9514 CODE_LABEL_NUMBER (op
[7]));
9515 output_asm_insn (".long\t%3", op
);
9520 /* Setup base pointer if required. */
9522 || (!DISP_IN_RANGE (delta
)
9523 && !CONST_OK_FOR_K (delta
)
9524 && !CONST_OK_FOR_Os (delta
))
9525 || (!DISP_IN_RANGE (delta
)
9526 && !CONST_OK_FOR_K (vcall_offset
)
9527 && !CONST_OK_FOR_Os (vcall_offset
)))
9529 op
[5] = gen_label_rtx ();
9530 output_asm_insn ("basr\t%4,0", op
);
9531 targetm
.asm_out
.internal_label (file
, "L",
9532 CODE_LABEL_NUMBER (op
[5]));
9535 /* Add DELTA to this pointer. */
9538 if (CONST_OK_FOR_J (delta
))
9539 output_asm_insn ("la\t%1,%2(%1)", op
);
9540 else if (DISP_IN_RANGE (delta
))
9541 output_asm_insn ("lay\t%1,%2(%1)", op
);
9542 else if (CONST_OK_FOR_K (delta
))
9543 output_asm_insn ("ahi\t%1,%2", op
);
9544 else if (CONST_OK_FOR_Os (delta
))
9545 output_asm_insn ("afi\t%1,%2", op
);
9548 op
[6] = gen_label_rtx ();
9549 output_asm_insn ("a\t%1,%6-%5(%4)", op
);
9553 /* Perform vcall adjustment. */
9556 if (CONST_OK_FOR_J (vcall_offset
))
9558 output_asm_insn ("l\t%4,0(%1)", op
);
9559 output_asm_insn ("a\t%1,%3(%4)", op
);
9561 else if (DISP_IN_RANGE (vcall_offset
))
9563 output_asm_insn ("l\t%4,0(%1)", op
);
9564 output_asm_insn ("ay\t%1,%3(%4)", op
);
9566 else if (CONST_OK_FOR_K (vcall_offset
))
9568 output_asm_insn ("lhi\t%4,%3", op
);
9569 output_asm_insn ("a\t%4,0(%1)", op
);
9570 output_asm_insn ("a\t%1,0(%4)", op
);
9572 else if (CONST_OK_FOR_Os (vcall_offset
))
9574 output_asm_insn ("iilf\t%4,%3", op
);
9575 output_asm_insn ("a\t%4,0(%1)", op
);
9576 output_asm_insn ("a\t%1,0(%4)", op
);
9580 op
[7] = gen_label_rtx ();
9581 output_asm_insn ("l\t%4,%7-%5(%4)", op
);
9582 output_asm_insn ("a\t%4,0(%1)", op
);
9583 output_asm_insn ("a\t%1,0(%4)", op
);
9586 /* We had to clobber the base pointer register.
9587 Re-setup the base pointer (with a different base). */
9588 op
[5] = gen_label_rtx ();
9589 output_asm_insn ("basr\t%4,0", op
);
9590 targetm
.asm_out
.internal_label (file
, "L",
9591 CODE_LABEL_NUMBER (op
[5]));
9594 /* Jump to target. */
9595 op
[8] = gen_label_rtx ();
9598 output_asm_insn ("l\t%4,%8-%5(%4)", op
);
9600 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9601 /* We cannot call through .plt, since .plt requires %r12 loaded. */
9602 else if (flag_pic
== 1)
9604 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9605 output_asm_insn ("l\t%4,%0(%4)", op
);
9607 else if (flag_pic
== 2)
9609 op
[9] = gen_rtx_REG (Pmode
, 0);
9610 output_asm_insn ("l\t%9,%8-4-%5(%4)", op
);
9611 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9612 output_asm_insn ("ar\t%4,%9", op
);
9613 output_asm_insn ("l\t%4,0(%4)", op
);
9616 output_asm_insn ("br\t%4", op
);
9618 /* Output literal pool. */
9619 output_asm_insn (".align\t4", op
);
9621 if (nonlocal
&& flag_pic
== 2)
9622 output_asm_insn (".long\t%0", op
);
9625 op
[0] = gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
9626 SYMBOL_REF_FLAGS (op
[0]) = SYMBOL_FLAG_LOCAL
;
9629 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[8]));
9631 output_asm_insn (".long\t%0", op
);
9633 output_asm_insn (".long\t%0-%5", op
);
9637 targetm
.asm_out
.internal_label (file
, "L",
9638 CODE_LABEL_NUMBER (op
[6]));
9639 output_asm_insn (".long\t%2", op
);
9643 targetm
.asm_out
.internal_label (file
, "L",
9644 CODE_LABEL_NUMBER (op
[7]));
9645 output_asm_insn (".long\t%3", op
);
9648 final_end_function ();
9652 s390_valid_pointer_mode (enum machine_mode mode
)
9654 return (mode
== SImode
|| (TARGET_64BIT
&& mode
== DImode
));
9657 /* Checks whether the given CALL_EXPR would use a caller
9658 saved register. This is used to decide whether sibling call
9659 optimization could be performed on the respective function
9663 s390_call_saved_register_used (tree call_expr
)
9665 CUMULATIVE_ARGS cum_v
;
9666 cumulative_args_t cum
;
9668 enum machine_mode mode
;
9673 INIT_CUMULATIVE_ARGS (cum_v
, NULL
, NULL
, 0, 0);
9674 cum
= pack_cumulative_args (&cum_v
);
9676 for (i
= 0; i
< call_expr_nargs (call_expr
); i
++)
9678 parameter
= CALL_EXPR_ARG (call_expr
, i
);
9679 gcc_assert (parameter
);
9681 /* For an undeclared variable passed as parameter we will get
9682 an ERROR_MARK node here. */
9683 if (TREE_CODE (parameter
) == ERROR_MARK
)
9686 type
= TREE_TYPE (parameter
);
9689 mode
= TYPE_MODE (type
);
9692 if (pass_by_reference (&cum_v
, mode
, type
, true))
9695 type
= build_pointer_type (type
);
9698 parm_rtx
= s390_function_arg (cum
, mode
, type
, 0);
9700 s390_function_arg_advance (cum
, mode
, type
, 0);
9705 if (REG_P (parm_rtx
))
9708 reg
< HARD_REGNO_NREGS (REGNO (parm_rtx
), GET_MODE (parm_rtx
));
9710 if (!call_used_regs
[reg
+ REGNO (parm_rtx
)])
9714 if (GET_CODE (parm_rtx
) == PARALLEL
)
9718 for (i
= 0; i
< XVECLEN (parm_rtx
, 0); i
++)
9720 rtx r
= XEXP (XVECEXP (parm_rtx
, 0, i
), 0);
9722 gcc_assert (REG_P (r
));
9725 reg
< HARD_REGNO_NREGS (REGNO (r
), GET_MODE (r
));
9727 if (!call_used_regs
[reg
+ REGNO (r
)])
9736 /* Return true if the given call expression can be
9737 turned into a sibling call.
9738 DECL holds the declaration of the function to be called whereas
9739 EXP is the call expression itself. */
9742 s390_function_ok_for_sibcall (tree decl
, tree exp
)
9744 /* The TPF epilogue uses register 1. */
9745 if (TARGET_TPF_PROFILING
)
9748 /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
9749 which would have to be restored before the sibcall. */
9750 if (!TARGET_64BIT
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
9753 /* Register 6 on s390 is available as an argument register but unfortunately
9754 "caller saved". This makes functions needing this register for arguments
9755 not suitable for sibcalls. */
9756 return !s390_call_saved_register_used (exp
);
9759 /* Return the fixed registers used for condition codes. */
9762 s390_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
9765 *p2
= INVALID_REGNUM
;
9770 /* This function is used by the call expanders of the machine description.
9771 It emits the call insn itself together with the necessary operations
9772 to adjust the target address and returns the emitted insn.
9773 ADDR_LOCATION is the target address rtx
9774 TLS_CALL the location of the thread-local symbol
9775 RESULT_REG the register where the result of the call should be stored
9776 RETADDR_REG the register where the return address should be stored
9777 If this parameter is NULL_RTX the call is considered
9778 to be a sibling call. */
9781 s390_emit_call (rtx addr_location
, rtx tls_call
, rtx result_reg
,
9784 bool plt_call
= false;
9790 /* Direct function calls need special treatment. */
9791 if (GET_CODE (addr_location
) == SYMBOL_REF
)
9793 /* When calling a global routine in PIC mode, we must
9794 replace the symbol itself with the PLT stub. */
9795 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (addr_location
))
9797 if (retaddr_reg
!= NULL_RTX
)
9799 addr_location
= gen_rtx_UNSPEC (Pmode
,
9800 gen_rtvec (1, addr_location
),
9802 addr_location
= gen_rtx_CONST (Pmode
, addr_location
);
9806 /* For -fpic code the PLT entries might use r12 which is
9807 call-saved. Therefore we cannot do a sibcall when
9808 calling directly using a symbol ref. When reaching
9809 this point we decided (in s390_function_ok_for_sibcall)
9810 to do a sibcall for a function pointer but one of the
9811 optimizers was able to get rid of the function pointer
9812 by propagating the symbol ref into the call. This
9813 optimization is illegal for S/390 so we turn the direct
9814 call into a indirect call again. */
9815 addr_location
= force_reg (Pmode
, addr_location
);
9818 /* Unless we can use the bras(l) insn, force the
9819 routine address into a register. */
9820 if (!TARGET_SMALL_EXEC
&& !TARGET_CPU_ZARCH
)
9823 addr_location
= legitimize_pic_address (addr_location
, 0);
9825 addr_location
= force_reg (Pmode
, addr_location
);
9829 /* If it is already an indirect call or the code above moved the
9830 SYMBOL_REF to somewhere else make sure the address can be found in
9832 if (retaddr_reg
== NULL_RTX
9833 && GET_CODE (addr_location
) != SYMBOL_REF
9836 emit_move_insn (gen_rtx_REG (Pmode
, SIBCALL_REGNUM
), addr_location
);
9837 addr_location
= gen_rtx_REG (Pmode
, SIBCALL_REGNUM
);
9840 addr_location
= gen_rtx_MEM (QImode
, addr_location
);
9841 call
= gen_rtx_CALL (VOIDmode
, addr_location
, const0_rtx
);
9843 if (result_reg
!= NULL_RTX
)
9844 call
= gen_rtx_SET (VOIDmode
, result_reg
, call
);
9846 if (retaddr_reg
!= NULL_RTX
)
9848 clobber
= gen_rtx_CLOBBER (VOIDmode
, retaddr_reg
);
9850 if (tls_call
!= NULL_RTX
)
9851 vec
= gen_rtvec (3, call
, clobber
,
9852 gen_rtx_USE (VOIDmode
, tls_call
));
9854 vec
= gen_rtvec (2, call
, clobber
);
9856 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
9859 insn
= emit_call_insn (call
);
9861 /* 31-bit PLT stubs and tls calls use the GOT register implicitly. */
9862 if ((!TARGET_64BIT
&& plt_call
) || tls_call
!= NULL_RTX
)
9864 /* s390_function_ok_for_sibcall should
9865 have denied sibcalls in this case. */
9866 gcc_assert (retaddr_reg
!= NULL_RTX
);
9867 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), gen_rtx_REG (Pmode
, 12));
9872 /* Implement TARGET_CONDITIONAL_REGISTER_USAGE. */
9875 s390_conditional_register_usage (void)
9881 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9882 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9884 if (TARGET_CPU_ZARCH
)
9886 fixed_regs
[BASE_REGNUM
] = 0;
9887 call_used_regs
[BASE_REGNUM
] = 0;
9888 fixed_regs
[RETURN_REGNUM
] = 0;
9889 call_used_regs
[RETURN_REGNUM
] = 0;
9893 for (i
= 24; i
< 32; i
++)
9894 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9898 for (i
= 18; i
< 20; i
++)
9899 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9902 if (TARGET_SOFT_FLOAT
)
9904 for (i
= 16; i
< 32; i
++)
9905 call_used_regs
[i
] = fixed_regs
[i
] = 1;
9909 /* Corresponding function to eh_return expander. */
9911 static GTY(()) rtx s390_tpf_eh_return_symbol
;
9913 s390_emit_tpf_eh_return (rtx target
)
9917 if (!s390_tpf_eh_return_symbol
)
9918 s390_tpf_eh_return_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tpf_eh_return");
9920 reg
= gen_rtx_REG (Pmode
, 2);
9922 emit_move_insn (reg
, target
);
9923 insn
= s390_emit_call (s390_tpf_eh_return_symbol
, NULL_RTX
, reg
,
9924 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
9925 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), reg
);
9927 emit_move_insn (EH_RETURN_HANDLER_RTX
, reg
);
9930 /* Rework the prologue/epilogue to avoid saving/restoring
9931 registers unnecessarily. */
9934 s390_optimize_prologue (void)
9936 rtx insn
, new_insn
, next_insn
;
9938 /* Do a final recompute of the frame-related data. */
9940 s390_update_frame_layout ();
9942 /* If all special registers are in fact used, there's nothing we
9943 can do, so no point in walking the insn list. */
9945 if (cfun_frame_layout
.first_save_gpr
<= BASE_REGNUM
9946 && cfun_frame_layout
.last_save_gpr
>= BASE_REGNUM
9947 && (TARGET_CPU_ZARCH
9948 || (cfun_frame_layout
.first_save_gpr
<= RETURN_REGNUM
9949 && cfun_frame_layout
.last_save_gpr
>= RETURN_REGNUM
)))
9952 /* Search for prologue/epilogue insns and replace them. */
9954 for (insn
= get_insns (); insn
; insn
= next_insn
)
9956 int first
, last
, off
;
9957 rtx set
, base
, offset
;
9959 next_insn
= NEXT_INSN (insn
);
9961 if (GET_CODE (insn
) != INSN
)
9964 if (GET_CODE (PATTERN (insn
)) == PARALLEL
9965 && store_multiple_operation (PATTERN (insn
), VOIDmode
))
9967 set
= XVECEXP (PATTERN (insn
), 0, 0);
9968 first
= REGNO (SET_SRC (set
));
9969 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
9970 offset
= const0_rtx
;
9971 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
9972 off
= INTVAL (offset
);
9974 if (GET_CODE (base
) != REG
|| off
< 0)
9976 if (cfun_frame_layout
.first_save_gpr
!= -1
9977 && (cfun_frame_layout
.first_save_gpr
< first
9978 || cfun_frame_layout
.last_save_gpr
> last
))
9980 if (REGNO (base
) != STACK_POINTER_REGNUM
9981 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
9983 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
9986 if (cfun_frame_layout
.first_save_gpr
!= -1)
9988 new_insn
= save_gprs (base
,
9989 off
+ (cfun_frame_layout
.first_save_gpr
9990 - first
) * UNITS_PER_LONG
,
9991 cfun_frame_layout
.first_save_gpr
,
9992 cfun_frame_layout
.last_save_gpr
);
9993 new_insn
= emit_insn_before (new_insn
, insn
);
9994 INSN_ADDRESSES_NEW (new_insn
, -1);
10001 if (cfun_frame_layout
.first_save_gpr
== -1
10002 && GET_CODE (PATTERN (insn
)) == SET
10003 && GET_CODE (SET_SRC (PATTERN (insn
))) == REG
10004 && (REGNO (SET_SRC (PATTERN (insn
))) == BASE_REGNUM
10005 || (!TARGET_CPU_ZARCH
10006 && REGNO (SET_SRC (PATTERN (insn
))) == RETURN_REGNUM
))
10007 && GET_CODE (SET_DEST (PATTERN (insn
))) == MEM
)
10009 set
= PATTERN (insn
);
10010 first
= REGNO (SET_SRC (set
));
10011 offset
= const0_rtx
;
10012 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10013 off
= INTVAL (offset
);
10015 if (GET_CODE (base
) != REG
|| off
< 0)
10017 if (REGNO (base
) != STACK_POINTER_REGNUM
10018 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10021 remove_insn (insn
);
10025 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10026 && load_multiple_operation (PATTERN (insn
), VOIDmode
))
10028 set
= XVECEXP (PATTERN (insn
), 0, 0);
10029 first
= REGNO (SET_DEST (set
));
10030 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10031 offset
= const0_rtx
;
10032 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10033 off
= INTVAL (offset
);
10035 if (GET_CODE (base
) != REG
|| off
< 0)
10037 if (cfun_frame_layout
.first_restore_gpr
!= -1
10038 && (cfun_frame_layout
.first_restore_gpr
< first
10039 || cfun_frame_layout
.last_restore_gpr
> last
))
10041 if (REGNO (base
) != STACK_POINTER_REGNUM
10042 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10044 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10047 if (cfun_frame_layout
.first_restore_gpr
!= -1)
10049 new_insn
= restore_gprs (base
,
10050 off
+ (cfun_frame_layout
.first_restore_gpr
10051 - first
) * UNITS_PER_LONG
,
10052 cfun_frame_layout
.first_restore_gpr
,
10053 cfun_frame_layout
.last_restore_gpr
);
10054 new_insn
= emit_insn_before (new_insn
, insn
);
10055 INSN_ADDRESSES_NEW (new_insn
, -1);
10058 remove_insn (insn
);
10062 if (cfun_frame_layout
.first_restore_gpr
== -1
10063 && GET_CODE (PATTERN (insn
)) == SET
10064 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
10065 && (REGNO (SET_DEST (PATTERN (insn
))) == BASE_REGNUM
10066 || (!TARGET_CPU_ZARCH
10067 && REGNO (SET_DEST (PATTERN (insn
))) == RETURN_REGNUM
))
10068 && GET_CODE (SET_SRC (PATTERN (insn
))) == MEM
)
10070 set
= PATTERN (insn
);
10071 first
= REGNO (SET_DEST (set
));
10072 offset
= const0_rtx
;
10073 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10074 off
= INTVAL (offset
);
10076 if (GET_CODE (base
) != REG
|| off
< 0)
10078 if (REGNO (base
) != STACK_POINTER_REGNUM
10079 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10082 remove_insn (insn
);
10088 /* On z10 and later the dynamic branch prediction must see the
10089 backward jump within a certain windows. If not it falls back to
10090 the static prediction. This function rearranges the loop backward
10091 branch in a way which makes the static prediction always correct.
10092 The function returns true if it added an instruction. */
10094 s390_fix_long_loop_prediction (rtx insn
)
10096 rtx set
= single_set (insn
);
10097 rtx code_label
, label_ref
, new_label
;
10103 /* This will exclude branch on count and branch on index patterns
10104 since these are correctly statically predicted. */
10106 || SET_DEST (set
) != pc_rtx
10107 || GET_CODE (SET_SRC(set
)) != IF_THEN_ELSE
)
10110 label_ref
= (GET_CODE (XEXP (SET_SRC (set
), 1)) == LABEL_REF
?
10111 XEXP (SET_SRC (set
), 1) : XEXP (SET_SRC (set
), 2));
10113 gcc_assert (GET_CODE (label_ref
) == LABEL_REF
);
10115 code_label
= XEXP (label_ref
, 0);
10117 if (INSN_ADDRESSES (INSN_UID (code_label
)) == -1
10118 || INSN_ADDRESSES (INSN_UID (insn
)) == -1
10119 || (INSN_ADDRESSES (INSN_UID (insn
))
10120 - INSN_ADDRESSES (INSN_UID (code_label
)) < PREDICT_DISTANCE
))
10123 for (distance
= 0, cur_insn
= PREV_INSN (insn
);
10124 distance
< PREDICT_DISTANCE
- 6;
10125 distance
+= get_attr_length (cur_insn
), cur_insn
= PREV_INSN (cur_insn
))
10126 if (!cur_insn
|| JUMP_P (cur_insn
) || LABEL_P (cur_insn
))
10129 new_label
= gen_label_rtx ();
10130 uncond_jump
= emit_jump_insn_after (
10131 gen_rtx_SET (VOIDmode
, pc_rtx
,
10132 gen_rtx_LABEL_REF (VOIDmode
, code_label
)),
10134 emit_label_after (new_label
, uncond_jump
);
10136 tmp
= XEXP (SET_SRC (set
), 1);
10137 XEXP (SET_SRC (set
), 1) = XEXP (SET_SRC (set
), 2);
10138 XEXP (SET_SRC (set
), 2) = tmp
;
10139 INSN_CODE (insn
) = -1;
10141 XEXP (label_ref
, 0) = new_label
;
10142 JUMP_LABEL (insn
) = new_label
;
10143 JUMP_LABEL (uncond_jump
) = code_label
;
10148 /* Returns 1 if INSN reads the value of REG for purposes not related
10149 to addressing of memory, and 0 otherwise. */
10151 s390_non_addr_reg_read_p (rtx reg
, rtx insn
)
10153 return reg_referenced_p (reg
, PATTERN (insn
))
10154 && !reg_used_in_mem_p (REGNO (reg
), PATTERN (insn
));
10157 /* Starting from INSN find_cond_jump looks downwards in the insn
10158 stream for a single jump insn which is the last user of the
10159 condition code set in INSN. */
10161 find_cond_jump (rtx insn
)
10163 for (; insn
; insn
= NEXT_INSN (insn
))
10167 if (LABEL_P (insn
))
10170 if (!JUMP_P (insn
))
10172 if (reg_mentioned_p (gen_rtx_REG (CCmode
, CC_REGNUM
), insn
))
10177 /* This will be triggered by a return. */
10178 if (GET_CODE (PATTERN (insn
)) != SET
)
10181 gcc_assert (SET_DEST (PATTERN (insn
)) == pc_rtx
);
10182 ite
= SET_SRC (PATTERN (insn
));
10184 if (GET_CODE (ite
) != IF_THEN_ELSE
)
10187 cc
= XEXP (XEXP (ite
, 0), 0);
10188 if (!REG_P (cc
) || !CC_REGNO_P (REGNO (cc
)))
10191 if (find_reg_note (insn
, REG_DEAD
, cc
))
10199 /* Swap the condition in COND and the operands in OP0 and OP1 so that
10200 the semantics does not change. If NULL_RTX is passed as COND the
10201 function tries to find the conditional jump starting with INSN. */
10203 s390_swap_cmp (rtx cond
, rtx
*op0
, rtx
*op1
, rtx insn
)
10207 if (cond
== NULL_RTX
)
10209 rtx jump
= find_cond_jump (NEXT_INSN (insn
));
10210 jump
= jump
? single_set (jump
) : NULL_RTX
;
10212 if (jump
== NULL_RTX
)
10215 cond
= XEXP (XEXP (jump
, 1), 0);
10220 PUT_CODE (cond
, swap_condition (GET_CODE (cond
)));
10223 /* On z10, instructions of the compare-and-branch family have the
10224 property to access the register occurring as second operand with
10225 its bits complemented. If such a compare is grouped with a second
10226 instruction that accesses the same register non-complemented, and
10227 if that register's value is delivered via a bypass, then the
10228 pipeline recycles, thereby causing significant performance decline.
10229 This function locates such situations and exchanges the two
10230 operands of the compare. The function return true whenever it
10233 s390_z10_optimize_cmp (rtx insn
)
10235 rtx prev_insn
, next_insn
;
10236 bool insn_added_p
= false;
10237 rtx cond
, *op0
, *op1
;
10239 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
10241 /* Handle compare and branch and branch on count
10243 rtx pattern
= single_set (insn
);
10246 || SET_DEST (pattern
) != pc_rtx
10247 || GET_CODE (SET_SRC (pattern
)) != IF_THEN_ELSE
)
10250 cond
= XEXP (SET_SRC (pattern
), 0);
10251 op0
= &XEXP (cond
, 0);
10252 op1
= &XEXP (cond
, 1);
10254 else if (GET_CODE (PATTERN (insn
)) == SET
)
10258 /* Handle normal compare instructions. */
10259 src
= SET_SRC (PATTERN (insn
));
10260 dest
= SET_DEST (PATTERN (insn
));
10263 || !CC_REGNO_P (REGNO (dest
))
10264 || GET_CODE (src
) != COMPARE
)
10267 /* s390_swap_cmp will try to find the conditional
10268 jump when passing NULL_RTX as condition. */
10270 op0
= &XEXP (src
, 0);
10271 op1
= &XEXP (src
, 1);
10276 if (!REG_P (*op0
) || !REG_P (*op1
))
10279 if (GET_MODE_CLASS (GET_MODE (*op0
)) != MODE_INT
)
10282 /* Swap the COMPARE arguments and its mask if there is a
10283 conflicting access in the previous insn. */
10284 prev_insn
= prev_active_insn (insn
);
10285 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10286 && reg_referenced_p (*op1
, PATTERN (prev_insn
)))
10287 s390_swap_cmp (cond
, op0
, op1
, insn
);
10289 /* Check if there is a conflict with the next insn. If there
10290 was no conflict with the previous insn, then swap the
10291 COMPARE arguments and its mask. If we already swapped
10292 the operands, or if swapping them would cause a conflict
10293 with the previous insn, issue a NOP after the COMPARE in
10294 order to separate the two instuctions. */
10295 next_insn
= next_active_insn (insn
);
10296 if (next_insn
!= NULL_RTX
&& INSN_P (next_insn
)
10297 && s390_non_addr_reg_read_p (*op1
, next_insn
))
10299 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10300 && s390_non_addr_reg_read_p (*op0
, prev_insn
))
10302 if (REGNO (*op1
) == 0)
10303 emit_insn_after (gen_nop1 (), insn
);
10305 emit_insn_after (gen_nop (), insn
);
10306 insn_added_p
= true;
10309 s390_swap_cmp (cond
, op0
, op1
, insn
);
10311 return insn_added_p
;
10314 /* Perform machine-dependent processing. */
10319 bool pool_overflow
= false;
10321 /* Make sure all splits have been performed; splits after
10322 machine_dependent_reorg might confuse insn length counts. */
10323 split_all_insns_noflow ();
10325 /* Install the main literal pool and the associated base
10326 register load insns.
10328 In addition, there are two problematic situations we need
10331 - the literal pool might be > 4096 bytes in size, so that
10332 some of its elements cannot be directly accessed
10334 - a branch target might be > 64K away from the branch, so that
10335 it is not possible to use a PC-relative instruction.
10337 To fix those, we split the single literal pool into multiple
10338 pool chunks, reloading the pool base register at various
10339 points throughout the function to ensure it always points to
10340 the pool chunk the following code expects, and / or replace
10341 PC-relative branches by absolute branches.
10343 However, the two problems are interdependent: splitting the
10344 literal pool can move a branch further away from its target,
10345 causing the 64K limit to overflow, and on the other hand,
10346 replacing a PC-relative branch by an absolute branch means
10347 we need to put the branch target address into the literal
10348 pool, possibly causing it to overflow.
10350 So, we loop trying to fix up both problems until we manage
10351 to satisfy both conditions at the same time. Note that the
10352 loop is guaranteed to terminate as every pass of the loop
10353 strictly decreases the total number of PC-relative branches
10354 in the function. (This is not completely true as there
10355 might be branch-over-pool insns introduced by chunkify_start.
10356 Those never need to be split however.) */
10360 struct constant_pool
*pool
= NULL
;
10362 /* Collect the literal pool. */
10363 if (!pool_overflow
)
10365 pool
= s390_mainpool_start ();
10367 pool_overflow
= true;
10370 /* If literal pool overflowed, start to chunkify it. */
10372 pool
= s390_chunkify_start ();
10374 /* Split out-of-range branches. If this has created new
10375 literal pool entries, cancel current chunk list and
10376 recompute it. zSeries machines have large branch
10377 instructions, so we never need to split a branch. */
10378 if (!TARGET_CPU_ZARCH
&& s390_split_branches ())
10381 s390_chunkify_cancel (pool
);
10383 s390_mainpool_cancel (pool
);
10388 /* If we made it up to here, both conditions are satisfied.
10389 Finish up literal pool related changes. */
10391 s390_chunkify_finish (pool
);
10393 s390_mainpool_finish (pool
);
10395 /* We're done splitting branches. */
10396 cfun
->machine
->split_branches_pending_p
= false;
10400 /* Generate out-of-pool execute target insns. */
10401 if (TARGET_CPU_ZARCH
)
10403 rtx insn
, label
, target
;
10405 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10407 label
= s390_execute_label (insn
);
10411 gcc_assert (label
!= const0_rtx
);
10413 target
= emit_label (XEXP (label
, 0));
10414 INSN_ADDRESSES_NEW (target
, -1);
10416 target
= emit_insn (s390_execute_target (insn
));
10417 INSN_ADDRESSES_NEW (target
, -1);
10421 /* Try to optimize prologue and epilogue further. */
10422 s390_optimize_prologue ();
10424 /* Walk over the insns and do some >=z10 specific changes. */
10425 if (s390_tune
== PROCESSOR_2097_Z10
10426 || s390_tune
== PROCESSOR_2817_Z196
)
10429 bool insn_added_p
= false;
10431 /* The insn lengths and addresses have to be up to date for the
10432 following manipulations. */
10433 shorten_branches (get_insns ());
10435 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10437 if (!INSN_P (insn
) || INSN_CODE (insn
) <= 0)
10441 insn_added_p
|= s390_fix_long_loop_prediction (insn
);
10443 if ((GET_CODE (PATTERN (insn
)) == PARALLEL
10444 || GET_CODE (PATTERN (insn
)) == SET
)
10445 && s390_tune
== PROCESSOR_2097_Z10
)
10446 insn_added_p
|= s390_z10_optimize_cmp (insn
);
10449 /* Adjust branches if we added new instructions. */
10451 shorten_branches (get_insns ());
10455 /* Return true if INSN is a fp load insn writing register REGNO. */
10457 s390_fpload_toreg (rtx insn
, unsigned int regno
)
10460 enum attr_type flag
= s390_safe_attr_type (insn
);
10462 if (flag
!= TYPE_FLOADSF
&& flag
!= TYPE_FLOADDF
)
10465 set
= single_set (insn
);
10467 if (set
== NULL_RTX
)
10470 if (!REG_P (SET_DEST (set
)) || !MEM_P (SET_SRC (set
)))
10473 if (REGNO (SET_DEST (set
)) != regno
)
10479 /* This value describes the distance to be avoided between an
10480 aritmetic fp instruction and an fp load writing the same register.
10481 Z10_EARLYLOAD_DISTANCE - 1 as well as Z10_EARLYLOAD_DISTANCE + 1 is
10482 fine but the exact value has to be avoided. Otherwise the FP
10483 pipeline will throw an exception causing a major penalty. */
10484 #define Z10_EARLYLOAD_DISTANCE 7
10486 /* Rearrange the ready list in order to avoid the situation described
10487 for Z10_EARLYLOAD_DISTANCE. A problematic load instruction is
10488 moved to the very end of the ready list. */
10490 s390_z10_prevent_earlyload_conflicts (rtx
*ready
, int *nready_p
)
10492 unsigned int regno
;
10493 int nready
= *nready_p
;
10498 enum attr_type flag
;
10501 /* Skip DISTANCE - 1 active insns. */
10502 for (insn
= last_scheduled_insn
, distance
= Z10_EARLYLOAD_DISTANCE
- 1;
10503 distance
> 0 && insn
!= NULL_RTX
;
10504 distance
--, insn
= prev_active_insn (insn
))
10505 if (CALL_P (insn
) || JUMP_P (insn
))
10508 if (insn
== NULL_RTX
)
10511 set
= single_set (insn
);
10513 if (set
== NULL_RTX
|| !REG_P (SET_DEST (set
))
10514 || GET_MODE_CLASS (GET_MODE (SET_DEST (set
))) != MODE_FLOAT
)
10517 flag
= s390_safe_attr_type (insn
);
10519 if (flag
== TYPE_FLOADSF
|| flag
== TYPE_FLOADDF
)
10522 regno
= REGNO (SET_DEST (set
));
10525 while (!s390_fpload_toreg (ready
[i
], regno
) && i
> 0)
10532 memmove (&ready
[1], &ready
[0], sizeof (rtx
) * i
);
10536 /* This function is called via hook TARGET_SCHED_REORDER before
10537 issuing one insn from list READY which contains *NREADYP entries.
10538 For target z10 it reorders load instructions to avoid early load
10539 conflicts in the floating point pipeline */
10541 s390_sched_reorder (FILE *file ATTRIBUTE_UNUSED
, int verbose ATTRIBUTE_UNUSED
,
10542 rtx
*ready
, int *nreadyp
, int clock ATTRIBUTE_UNUSED
)
10544 if (s390_tune
== PROCESSOR_2097_Z10
)
10545 if (reload_completed
&& *nreadyp
> 1)
10546 s390_z10_prevent_earlyload_conflicts (ready
, nreadyp
);
10548 return s390_issue_rate ();
10551 /* This function is called via hook TARGET_SCHED_VARIABLE_ISSUE after
10552 the scheduler has issued INSN. It stores the last issued insn into
10553 last_scheduled_insn in order to make it available for
10554 s390_sched_reorder. */
10556 s390_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED
,
10557 int verbose ATTRIBUTE_UNUSED
,
10558 rtx insn
, int more
)
10560 last_scheduled_insn
= insn
;
10562 if (GET_CODE (PATTERN (insn
)) != USE
10563 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
10570 s390_sched_init (FILE *file ATTRIBUTE_UNUSED
,
10571 int verbose ATTRIBUTE_UNUSED
,
10572 int max_ready ATTRIBUTE_UNUSED
)
10574 last_scheduled_insn
= NULL_RTX
;
10577 /* This function checks the whole of insn X for memory references. The
10578 function always returns zero because the framework it is called
10579 from would stop recursively analyzing the insn upon a return value
10580 other than zero. The real result of this function is updating
10581 counter variable MEM_COUNT. */
10583 check_dpu (rtx
*x
, unsigned *mem_count
)
10585 if (*x
!= NULL_RTX
&& MEM_P (*x
))
10590 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates
10591 a new number struct loop *loop should be unrolled if tuned for cpus with
10592 a built-in stride prefetcher.
10593 The loop is analyzed for memory accesses by calling check_dpu for
10594 each rtx of the loop. Depending on the loop_depth and the amount of
10595 memory accesses a new number <=nunroll is returned to improve the
10596 behaviour of the hardware prefetch unit. */
10598 s390_loop_unroll_adjust (unsigned nunroll
, struct loop
*loop
)
10603 unsigned mem_count
= 0;
10605 if (s390_tune
!= PROCESSOR_2097_Z10
&& s390_tune
!= PROCESSOR_2817_Z196
)
10608 /* Count the number of memory references within the loop body. */
10609 bbs
= get_loop_body (loop
);
10610 for (i
= 0; i
< loop
->num_nodes
; i
++)
10612 for (insn
= BB_HEAD (bbs
[i
]); insn
!= BB_END (bbs
[i
]); insn
= NEXT_INSN (insn
))
10613 if (INSN_P (insn
) && INSN_CODE (insn
) != -1)
10614 for_each_rtx (&insn
, (rtx_function
) check_dpu
, &mem_count
);
10618 /* Prevent division by zero, and we do not need to adjust nunroll in this case. */
10619 if (mem_count
== 0)
10622 switch (loop_depth(loop
))
10625 return MIN (nunroll
, 28 / mem_count
);
10627 return MIN (nunroll
, 22 / mem_count
);
10629 return MIN (nunroll
, 16 / mem_count
);
10633 /* Initialize GCC target structure. */
10635 #undef TARGET_ASM_ALIGNED_HI_OP
10636 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10637 #undef TARGET_ASM_ALIGNED_DI_OP
10638 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10639 #undef TARGET_ASM_INTEGER
10640 #define TARGET_ASM_INTEGER s390_assemble_integer
10642 #undef TARGET_ASM_OPEN_PAREN
10643 #define TARGET_ASM_OPEN_PAREN ""
10645 #undef TARGET_ASM_CLOSE_PAREN
10646 #define TARGET_ASM_CLOSE_PAREN ""
10648 #undef TARGET_OPTION_OVERRIDE
10649 #define TARGET_OPTION_OVERRIDE s390_option_override
10651 #undef TARGET_ENCODE_SECTION_INFO
10652 #define TARGET_ENCODE_SECTION_INFO s390_encode_section_info
10654 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10655 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10658 #undef TARGET_HAVE_TLS
10659 #define TARGET_HAVE_TLS true
10661 #undef TARGET_CANNOT_FORCE_CONST_MEM
10662 #define TARGET_CANNOT_FORCE_CONST_MEM s390_cannot_force_const_mem
10664 #undef TARGET_DELEGITIMIZE_ADDRESS
10665 #define TARGET_DELEGITIMIZE_ADDRESS s390_delegitimize_address
10667 #undef TARGET_LEGITIMIZE_ADDRESS
10668 #define TARGET_LEGITIMIZE_ADDRESS s390_legitimize_address
10670 #undef TARGET_RETURN_IN_MEMORY
10671 #define TARGET_RETURN_IN_MEMORY s390_return_in_memory
10673 #undef TARGET_INIT_BUILTINS
10674 #define TARGET_INIT_BUILTINS s390_init_builtins
10675 #undef TARGET_EXPAND_BUILTIN
10676 #define TARGET_EXPAND_BUILTIN s390_expand_builtin
10678 #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
10679 #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA s390_output_addr_const_extra
10681 #undef TARGET_ASM_OUTPUT_MI_THUNK
10682 #define TARGET_ASM_OUTPUT_MI_THUNK s390_output_mi_thunk
10683 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10684 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10686 #undef TARGET_SCHED_ADJUST_PRIORITY
10687 #define TARGET_SCHED_ADJUST_PRIORITY s390_adjust_priority
10688 #undef TARGET_SCHED_ISSUE_RATE
10689 #define TARGET_SCHED_ISSUE_RATE s390_issue_rate
10690 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10691 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD s390_first_cycle_multipass_dfa_lookahead
10693 #undef TARGET_SCHED_VARIABLE_ISSUE
10694 #define TARGET_SCHED_VARIABLE_ISSUE s390_sched_variable_issue
10695 #undef TARGET_SCHED_REORDER
10696 #define TARGET_SCHED_REORDER s390_sched_reorder
10697 #undef TARGET_SCHED_INIT
10698 #define TARGET_SCHED_INIT s390_sched_init
10700 #undef TARGET_CANNOT_COPY_INSN_P
10701 #define TARGET_CANNOT_COPY_INSN_P s390_cannot_copy_insn_p
10702 #undef TARGET_RTX_COSTS
10703 #define TARGET_RTX_COSTS s390_rtx_costs
10704 #undef TARGET_ADDRESS_COST
10705 #define TARGET_ADDRESS_COST s390_address_cost
10706 #undef TARGET_REGISTER_MOVE_COST
10707 #define TARGET_REGISTER_MOVE_COST s390_register_move_cost
10708 #undef TARGET_MEMORY_MOVE_COST
10709 #define TARGET_MEMORY_MOVE_COST s390_memory_move_cost
10711 #undef TARGET_MACHINE_DEPENDENT_REORG
10712 #define TARGET_MACHINE_DEPENDENT_REORG s390_reorg
10714 #undef TARGET_VALID_POINTER_MODE
10715 #define TARGET_VALID_POINTER_MODE s390_valid_pointer_mode
10717 #undef TARGET_BUILD_BUILTIN_VA_LIST
10718 #define TARGET_BUILD_BUILTIN_VA_LIST s390_build_builtin_va_list
10719 #undef TARGET_EXPAND_BUILTIN_VA_START
10720 #define TARGET_EXPAND_BUILTIN_VA_START s390_va_start
10721 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10722 #define TARGET_GIMPLIFY_VA_ARG_EXPR s390_gimplify_va_arg
10724 #undef TARGET_PROMOTE_FUNCTION_MODE
10725 #define TARGET_PROMOTE_FUNCTION_MODE s390_promote_function_mode
10726 #undef TARGET_PASS_BY_REFERENCE
10727 #define TARGET_PASS_BY_REFERENCE s390_pass_by_reference
10729 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10730 #define TARGET_FUNCTION_OK_FOR_SIBCALL s390_function_ok_for_sibcall
10731 #undef TARGET_FUNCTION_ARG
10732 #define TARGET_FUNCTION_ARG s390_function_arg
10733 #undef TARGET_FUNCTION_ARG_ADVANCE
10734 #define TARGET_FUNCTION_ARG_ADVANCE s390_function_arg_advance
10735 #undef TARGET_FUNCTION_VALUE
10736 #define TARGET_FUNCTION_VALUE s390_function_value
10737 #undef TARGET_LIBCALL_VALUE
10738 #define TARGET_LIBCALL_VALUE s390_libcall_value
10740 #undef TARGET_FIXED_CONDITION_CODE_REGS
10741 #define TARGET_FIXED_CONDITION_CODE_REGS s390_fixed_condition_code_regs
10743 #undef TARGET_CC_MODES_COMPATIBLE
10744 #define TARGET_CC_MODES_COMPATIBLE s390_cc_modes_compatible
10746 #undef TARGET_INVALID_WITHIN_DOLOOP
10747 #define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null
10750 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
10751 #define TARGET_ASM_OUTPUT_DWARF_DTPREL s390_output_dwarf_dtprel
10754 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10755 #undef TARGET_MANGLE_TYPE
10756 #define TARGET_MANGLE_TYPE s390_mangle_type
10759 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10760 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10762 #undef TARGET_PREFERRED_RELOAD_CLASS
10763 #define TARGET_PREFERRED_RELOAD_CLASS s390_preferred_reload_class
10765 #undef TARGET_SECONDARY_RELOAD
10766 #define TARGET_SECONDARY_RELOAD s390_secondary_reload
10768 #undef TARGET_LIBGCC_CMP_RETURN_MODE
10769 #define TARGET_LIBGCC_CMP_RETURN_MODE s390_libgcc_cmp_return_mode
10771 #undef TARGET_LIBGCC_SHIFT_COUNT_MODE
10772 #define TARGET_LIBGCC_SHIFT_COUNT_MODE s390_libgcc_shift_count_mode
10774 #undef TARGET_LEGITIMATE_ADDRESS_P
10775 #define TARGET_LEGITIMATE_ADDRESS_P s390_legitimate_address_p
10777 #undef TARGET_LEGITIMATE_CONSTANT_P
10778 #define TARGET_LEGITIMATE_CONSTANT_P s390_legitimate_constant_p
10780 #undef TARGET_CAN_ELIMINATE
10781 #define TARGET_CAN_ELIMINATE s390_can_eliminate
10783 #undef TARGET_CONDITIONAL_REGISTER_USAGE
10784 #define TARGET_CONDITIONAL_REGISTER_USAGE s390_conditional_register_usage
10786 #undef TARGET_LOOP_UNROLL_ADJUST
10787 #define TARGET_LOOP_UNROLL_ADJUST s390_loop_unroll_adjust
10789 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
10790 #define TARGET_ASM_TRAMPOLINE_TEMPLATE s390_asm_trampoline_template
10791 #undef TARGET_TRAMPOLINE_INIT
10792 #define TARGET_TRAMPOLINE_INIT s390_trampoline_init
10794 #undef TARGET_UNWIND_WORD_MODE
10795 #define TARGET_UNWIND_WORD_MODE s390_unwind_word_mode
10797 struct gcc_target targetm
= TARGET_INITIALIZER
;
10799 #include "gt-s390.h"