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"
47 #include "target-def.h"
49 #include "langhooks.h"
57 /* Define the specific costs for a given cpu. */
59 struct processor_costs
62 const int m
; /* cost of an M instruction. */
63 const int mghi
; /* cost of an MGHI instruction. */
64 const int mh
; /* cost of an MH instruction. */
65 const int mhi
; /* cost of an MHI instruction. */
66 const int ml
; /* cost of an ML instruction. */
67 const int mr
; /* cost of an MR instruction. */
68 const int ms
; /* cost of an MS instruction. */
69 const int msg
; /* cost of an MSG instruction. */
70 const int msgf
; /* cost of an MSGF instruction. */
71 const int msgfr
; /* cost of an MSGFR instruction. */
72 const int msgr
; /* cost of an MSGR instruction. */
73 const int msr
; /* cost of an MSR instruction. */
74 const int mult_df
; /* cost of multiplication in DFmode. */
77 const int sqxbr
; /* cost of square root in TFmode. */
78 const int sqdbr
; /* cost of square root in DFmode. */
79 const int sqebr
; /* cost of square root in SFmode. */
80 /* multiply and add */
81 const int madbr
; /* cost of multiply and add in DFmode. */
82 const int maebr
; /* cost of multiply and add in SFmode. */
94 const struct processor_costs
*s390_cost
;
97 struct processor_costs z900_cost
=
99 COSTS_N_INSNS (5), /* M */
100 COSTS_N_INSNS (10), /* MGHI */
101 COSTS_N_INSNS (5), /* MH */
102 COSTS_N_INSNS (4), /* MHI */
103 COSTS_N_INSNS (5), /* ML */
104 COSTS_N_INSNS (5), /* MR */
105 COSTS_N_INSNS (4), /* MS */
106 COSTS_N_INSNS (15), /* MSG */
107 COSTS_N_INSNS (7), /* MSGF */
108 COSTS_N_INSNS (7), /* MSGFR */
109 COSTS_N_INSNS (10), /* MSGR */
110 COSTS_N_INSNS (4), /* MSR */
111 COSTS_N_INSNS (7), /* multiplication in DFmode */
112 COSTS_N_INSNS (13), /* MXBR */
113 COSTS_N_INSNS (136), /* SQXBR */
114 COSTS_N_INSNS (44), /* SQDBR */
115 COSTS_N_INSNS (35), /* SQEBR */
116 COSTS_N_INSNS (18), /* MADBR */
117 COSTS_N_INSNS (13), /* MAEBR */
118 COSTS_N_INSNS (134), /* DXBR */
119 COSTS_N_INSNS (30), /* DDBR */
120 COSTS_N_INSNS (27), /* DEBR */
121 COSTS_N_INSNS (220), /* DLGR */
122 COSTS_N_INSNS (34), /* DLR */
123 COSTS_N_INSNS (34), /* DR */
124 COSTS_N_INSNS (32), /* DSGFR */
125 COSTS_N_INSNS (32), /* DSGR */
129 struct processor_costs z990_cost
=
131 COSTS_N_INSNS (4), /* M */
132 COSTS_N_INSNS (2), /* MGHI */
133 COSTS_N_INSNS (2), /* MH */
134 COSTS_N_INSNS (2), /* MHI */
135 COSTS_N_INSNS (4), /* ML */
136 COSTS_N_INSNS (4), /* MR */
137 COSTS_N_INSNS (5), /* MS */
138 COSTS_N_INSNS (6), /* MSG */
139 COSTS_N_INSNS (4), /* MSGF */
140 COSTS_N_INSNS (4), /* MSGFR */
141 COSTS_N_INSNS (4), /* MSGR */
142 COSTS_N_INSNS (4), /* MSR */
143 COSTS_N_INSNS (1), /* multiplication in DFmode */
144 COSTS_N_INSNS (28), /* MXBR */
145 COSTS_N_INSNS (130), /* SQXBR */
146 COSTS_N_INSNS (66), /* SQDBR */
147 COSTS_N_INSNS (38), /* SQEBR */
148 COSTS_N_INSNS (1), /* MADBR */
149 COSTS_N_INSNS (1), /* MAEBR */
150 COSTS_N_INSNS (60), /* DXBR */
151 COSTS_N_INSNS (40), /* DDBR */
152 COSTS_N_INSNS (26), /* DEBR */
153 COSTS_N_INSNS (176), /* DLGR */
154 COSTS_N_INSNS (31), /* DLR */
155 COSTS_N_INSNS (31), /* DR */
156 COSTS_N_INSNS (31), /* DSGFR */
157 COSTS_N_INSNS (31), /* DSGR */
161 struct processor_costs z9_109_cost
=
163 COSTS_N_INSNS (4), /* M */
164 COSTS_N_INSNS (2), /* MGHI */
165 COSTS_N_INSNS (2), /* MH */
166 COSTS_N_INSNS (2), /* MHI */
167 COSTS_N_INSNS (4), /* ML */
168 COSTS_N_INSNS (4), /* MR */
169 COSTS_N_INSNS (5), /* MS */
170 COSTS_N_INSNS (6), /* MSG */
171 COSTS_N_INSNS (4), /* MSGF */
172 COSTS_N_INSNS (4), /* MSGFR */
173 COSTS_N_INSNS (4), /* MSGR */
174 COSTS_N_INSNS (4), /* MSR */
175 COSTS_N_INSNS (1), /* multiplication in DFmode */
176 COSTS_N_INSNS (28), /* MXBR */
177 COSTS_N_INSNS (130), /* SQXBR */
178 COSTS_N_INSNS (66), /* SQDBR */
179 COSTS_N_INSNS (38), /* SQEBR */
180 COSTS_N_INSNS (1), /* MADBR */
181 COSTS_N_INSNS (1), /* MAEBR */
182 COSTS_N_INSNS (60), /* DXBR */
183 COSTS_N_INSNS (40), /* DDBR */
184 COSTS_N_INSNS (26), /* DEBR */
185 COSTS_N_INSNS (30), /* DLGR */
186 COSTS_N_INSNS (23), /* DLR */
187 COSTS_N_INSNS (23), /* DR */
188 COSTS_N_INSNS (24), /* DSGFR */
189 COSTS_N_INSNS (24), /* DSGR */
193 struct processor_costs z10_cost
=
195 COSTS_N_INSNS (10), /* M */
196 COSTS_N_INSNS (10), /* MGHI */
197 COSTS_N_INSNS (10), /* MH */
198 COSTS_N_INSNS (10), /* MHI */
199 COSTS_N_INSNS (10), /* ML */
200 COSTS_N_INSNS (10), /* MR */
201 COSTS_N_INSNS (10), /* MS */
202 COSTS_N_INSNS (10), /* MSG */
203 COSTS_N_INSNS (10), /* MSGF */
204 COSTS_N_INSNS (10), /* MSGFR */
205 COSTS_N_INSNS (10), /* MSGR */
206 COSTS_N_INSNS (10), /* MSR */
207 COSTS_N_INSNS (1) , /* multiplication in DFmode */
208 COSTS_N_INSNS (50), /* MXBR */
209 COSTS_N_INSNS (120), /* SQXBR */
210 COSTS_N_INSNS (52), /* SQDBR */
211 COSTS_N_INSNS (38), /* SQEBR */
212 COSTS_N_INSNS (1), /* MADBR */
213 COSTS_N_INSNS (1), /* MAEBR */
214 COSTS_N_INSNS (111), /* DXBR */
215 COSTS_N_INSNS (39), /* DDBR */
216 COSTS_N_INSNS (32), /* DEBR */
217 COSTS_N_INSNS (160), /* DLGR */
218 COSTS_N_INSNS (71), /* DLR */
219 COSTS_N_INSNS (71), /* DR */
220 COSTS_N_INSNS (71), /* DSGFR */
221 COSTS_N_INSNS (71), /* DSGR */
225 struct processor_costs z196_cost
=
227 COSTS_N_INSNS (7), /* M */
228 COSTS_N_INSNS (5), /* MGHI */
229 COSTS_N_INSNS (5), /* MH */
230 COSTS_N_INSNS (5), /* MHI */
231 COSTS_N_INSNS (7), /* ML */
232 COSTS_N_INSNS (7), /* MR */
233 COSTS_N_INSNS (6), /* MS */
234 COSTS_N_INSNS (8), /* MSG */
235 COSTS_N_INSNS (6), /* MSGF */
236 COSTS_N_INSNS (6), /* MSGFR */
237 COSTS_N_INSNS (8), /* MSGR */
238 COSTS_N_INSNS (6), /* MSR */
239 COSTS_N_INSNS (1) , /* multiplication in DFmode */
240 COSTS_N_INSNS (40), /* MXBR B+40 */
241 COSTS_N_INSNS (100), /* SQXBR B+100 */
242 COSTS_N_INSNS (42), /* SQDBR B+42 */
243 COSTS_N_INSNS (28), /* SQEBR B+28 */
244 COSTS_N_INSNS (1), /* MADBR B */
245 COSTS_N_INSNS (1), /* MAEBR B */
246 COSTS_N_INSNS (101), /* DXBR B+101 */
247 COSTS_N_INSNS (29), /* DDBR */
248 COSTS_N_INSNS (22), /* DEBR */
249 COSTS_N_INSNS (160), /* DLGR cracked */
250 COSTS_N_INSNS (160), /* DLR cracked */
251 COSTS_N_INSNS (160), /* DR expanded */
252 COSTS_N_INSNS (160), /* DSGFR cracked */
253 COSTS_N_INSNS (160), /* DSGR cracked */
256 extern int reload_completed
;
258 /* Kept up to date using the SCHED_VARIABLE_ISSUE hook. */
259 static rtx last_scheduled_insn
;
261 /* Structure used to hold the components of a S/390 memory
262 address. A legitimate address on S/390 is of the general
264 base + index + displacement
265 where any of the components is optional.
267 base and index are registers of the class ADDR_REGS,
268 displacement is an unsigned 12-bit immediate constant. */
279 /* The following structure is embedded in the machine
280 specific part of struct function. */
282 struct GTY (()) s390_frame_layout
284 /* Offset within stack frame. */
285 HOST_WIDE_INT gprs_offset
;
286 HOST_WIDE_INT f0_offset
;
287 HOST_WIDE_INT f4_offset
;
288 HOST_WIDE_INT f8_offset
;
289 HOST_WIDE_INT backchain_offset
;
291 /* Number of first and last gpr where slots in the register
292 save area are reserved for. */
293 int first_save_gpr_slot
;
294 int last_save_gpr_slot
;
296 /* Number of first and last gpr to be saved, restored. */
298 int first_restore_gpr
;
300 int last_restore_gpr
;
302 /* Bits standing for floating point registers. Set, if the
303 respective register has to be saved. Starting with reg 16 (f0)
304 at the rightmost bit.
305 Bit 15 - 8 7 6 5 4 3 2 1 0
306 fpr 15 - 8 7 5 3 1 6 4 2 0
307 reg 31 - 24 23 22 21 20 19 18 17 16 */
308 unsigned int fpr_bitmap
;
310 /* Number of floating point registers f8-f15 which must be saved. */
313 /* Set if return address needs to be saved.
314 This flag is set by s390_return_addr_rtx if it could not use
315 the initial value of r14 and therefore depends on r14 saved
317 bool save_return_addr_p
;
319 /* Size of stack frame. */
320 HOST_WIDE_INT frame_size
;
323 /* Define the structure for the machine field in struct function. */
325 struct GTY(()) machine_function
327 struct s390_frame_layout frame_layout
;
329 /* Literal pool base register. */
332 /* True if we may need to perform branch splitting. */
333 bool split_branches_pending_p
;
335 /* Some local-dynamic TLS symbol name. */
336 const char *some_ld_name
;
338 bool has_landing_pad_p
;
341 /* Few accessor macros for struct cfun->machine->s390_frame_layout. */
343 #define cfun_frame_layout (cfun->machine->frame_layout)
344 #define cfun_save_high_fprs_p (!!cfun_frame_layout.high_fprs)
345 #define cfun_gprs_save_area_size ((cfun_frame_layout.last_save_gpr_slot - \
346 cfun_frame_layout.first_save_gpr_slot + 1) * UNITS_PER_LONG)
347 #define cfun_set_fpr_bit(BITNUM) (cfun->machine->frame_layout.fpr_bitmap |= \
349 #define cfun_fpr_bit_p(BITNUM) (!!(cfun->machine->frame_layout.fpr_bitmap & \
352 /* Number of GPRs and FPRs used for argument passing. */
353 #define GP_ARG_NUM_REG 5
354 #define FP_ARG_NUM_REG (TARGET_64BIT? 4 : 2)
356 /* A couple of shortcuts. */
357 #define CONST_OK_FOR_J(x) \
358 CONST_OK_FOR_CONSTRAINT_P((x), 'J', "J")
359 #define CONST_OK_FOR_K(x) \
360 CONST_OK_FOR_CONSTRAINT_P((x), 'K', "K")
361 #define CONST_OK_FOR_Os(x) \
362 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Os")
363 #define CONST_OK_FOR_Op(x) \
364 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Op")
365 #define CONST_OK_FOR_On(x) \
366 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "On")
368 #define REGNO_PAIR_OK(REGNO, MODE) \
369 (HARD_REGNO_NREGS ((REGNO), (MODE)) == 1 || !((REGNO) & 1))
371 /* That's the read ahead of the dynamic branch prediction unit in
372 bytes on a z10 (or higher) CPU. */
373 #define PREDICT_DISTANCE (TARGET_Z10 ? 384 : 2048)
375 /* Return the alignment for LABEL. We default to the -falign-labels
376 value except for the literal pool base label. */
378 s390_label_align (rtx label
)
380 rtx prev_insn
= prev_active_insn (label
);
382 if (prev_insn
== NULL_RTX
)
385 prev_insn
= single_set (prev_insn
);
387 if (prev_insn
== NULL_RTX
)
390 prev_insn
= SET_SRC (prev_insn
);
392 /* Don't align literal pool base labels. */
393 if (GET_CODE (prev_insn
) == UNSPEC
394 && XINT (prev_insn
, 1) == UNSPEC_MAIN_BASE
)
398 return align_labels_log
;
401 static enum machine_mode
402 s390_libgcc_cmp_return_mode (void)
404 return TARGET_64BIT
? DImode
: SImode
;
407 static enum machine_mode
408 s390_libgcc_shift_count_mode (void)
410 return TARGET_64BIT
? DImode
: SImode
;
413 static enum machine_mode
414 s390_unwind_word_mode (void)
416 return TARGET_64BIT
? DImode
: SImode
;
419 /* Return true if the back end supports mode MODE. */
421 s390_scalar_mode_supported_p (enum machine_mode mode
)
423 /* In contrast to the default implementation reject TImode constants on 31bit
424 TARGET_ZARCH for ABI compliance. */
425 if (!TARGET_64BIT
&& TARGET_ZARCH
&& mode
== TImode
)
428 if (DECIMAL_FLOAT_MODE_P (mode
))
429 return default_decimal_float_supported_p ();
431 return default_scalar_mode_supported_p (mode
);
434 /* Set the has_landing_pad_p flag in struct machine_function to VALUE. */
437 s390_set_has_landing_pad_p (bool value
)
439 cfun
->machine
->has_landing_pad_p
= value
;
442 /* If two condition code modes are compatible, return a condition code
443 mode which is compatible with both. Otherwise, return
446 static enum machine_mode
447 s390_cc_modes_compatible (enum machine_mode m1
, enum machine_mode m2
)
455 if (m2
== CCUmode
|| m2
== CCTmode
|| m2
== CCZ1mode
456 || m2
== CCSmode
|| m2
== CCSRmode
|| m2
== CCURmode
)
477 /* Return true if SET either doesn't set the CC register, or else
478 the source and destination have matching CC modes and that
479 CC mode is at least as constrained as REQ_MODE. */
482 s390_match_ccmode_set (rtx set
, enum machine_mode req_mode
)
484 enum machine_mode set_mode
;
486 gcc_assert (GET_CODE (set
) == SET
);
488 if (GET_CODE (SET_DEST (set
)) != REG
|| !CC_REGNO_P (REGNO (SET_DEST (set
))))
491 set_mode
= GET_MODE (SET_DEST (set
));
505 if (req_mode
!= set_mode
)
510 if (req_mode
!= CCSmode
&& req_mode
!= CCUmode
&& req_mode
!= CCTmode
511 && req_mode
!= CCSRmode
&& req_mode
!= CCURmode
)
517 if (req_mode
!= CCAmode
)
525 return (GET_MODE (SET_SRC (set
)) == set_mode
);
528 /* Return true if every SET in INSN that sets the CC register
529 has source and destination with matching CC modes and that
530 CC mode is at least as constrained as REQ_MODE.
531 If REQ_MODE is VOIDmode, always return false. */
534 s390_match_ccmode (rtx insn
, enum machine_mode req_mode
)
538 /* s390_tm_ccmode returns VOIDmode to indicate failure. */
539 if (req_mode
== VOIDmode
)
542 if (GET_CODE (PATTERN (insn
)) == SET
)
543 return s390_match_ccmode_set (PATTERN (insn
), req_mode
);
545 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
546 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
548 rtx set
= XVECEXP (PATTERN (insn
), 0, i
);
549 if (GET_CODE (set
) == SET
)
550 if (!s390_match_ccmode_set (set
, req_mode
))
557 /* If a test-under-mask instruction can be used to implement
558 (compare (and ... OP1) OP2), return the CC mode required
559 to do that. Otherwise, return VOIDmode.
560 MIXED is true if the instruction can distinguish between
561 CC1 and CC2 for mixed selected bits (TMxx), it is false
562 if the instruction cannot (TM). */
565 s390_tm_ccmode (rtx op1
, rtx op2
, bool mixed
)
569 /* ??? Fixme: should work on CONST_DOUBLE as well. */
570 if (GET_CODE (op1
) != CONST_INT
|| GET_CODE (op2
) != CONST_INT
)
573 /* Selected bits all zero: CC0.
574 e.g.: int a; if ((a & (16 + 128)) == 0) */
575 if (INTVAL (op2
) == 0)
578 /* Selected bits all one: CC3.
579 e.g.: int a; if ((a & (16 + 128)) == 16 + 128) */
580 if (INTVAL (op2
) == INTVAL (op1
))
583 /* Exactly two bits selected, mixed zeroes and ones: CC1 or CC2. e.g.:
585 if ((a & (16 + 128)) == 16) -> CCT1
586 if ((a & (16 + 128)) == 128) -> CCT2 */
589 bit1
= exact_log2 (INTVAL (op2
));
590 bit0
= exact_log2 (INTVAL (op1
) ^ INTVAL (op2
));
591 if (bit0
!= -1 && bit1
!= -1)
592 return bit0
> bit1
? CCT1mode
: CCT2mode
;
598 /* Given a comparison code OP (EQ, NE, etc.) and the operands
599 OP0 and OP1 of a COMPARE, return the mode to be used for the
603 s390_select_ccmode (enum rtx_code code
, rtx op0
, rtx op1
)
609 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
610 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
612 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
613 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
615 if ((GET_CODE (op0
) == PLUS
|| GET_CODE (op0
) == MINUS
616 || GET_CODE (op1
) == NEG
)
617 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
620 if (GET_CODE (op0
) == AND
)
622 /* Check whether we can potentially do it via TM. */
623 enum machine_mode ccmode
;
624 ccmode
= s390_tm_ccmode (XEXP (op0
, 1), op1
, 1);
625 if (ccmode
!= VOIDmode
)
627 /* Relax CCTmode to CCZmode to allow fall-back to AND
628 if that turns out to be beneficial. */
629 return ccmode
== CCTmode
? CCZmode
: ccmode
;
633 if (register_operand (op0
, HImode
)
634 && GET_CODE (op1
) == CONST_INT
635 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 65535))
637 if (register_operand (op0
, QImode
)
638 && GET_CODE (op1
) == CONST_INT
639 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 255))
648 /* The only overflow condition of NEG and ABS happens when
649 -INT_MAX is used as parameter, which stays negative. So
650 we have an overflow from a positive value to a negative.
651 Using CCAP mode the resulting cc can be used for comparisons. */
652 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
653 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
656 /* If constants are involved in an add instruction it is possible to use
657 the resulting cc for comparisons with zero. Knowing the sign of the
658 constant the overflow behavior gets predictable. e.g.:
659 int a, b; if ((b = a + c) > 0)
660 with c as a constant value: c < 0 -> CCAN and c >= 0 -> CCAP */
661 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
662 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
664 if (INTVAL (XEXP((op0
), 1)) < 0)
678 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
679 && GET_CODE (op1
) != CONST_INT
)
685 if (GET_CODE (op0
) == PLUS
686 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
689 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
690 && GET_CODE (op1
) != CONST_INT
)
696 if (GET_CODE (op0
) == MINUS
697 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
700 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
701 && GET_CODE (op1
) != CONST_INT
)
710 /* Replace the comparison OP0 CODE OP1 by a semantically equivalent one
711 that we can implement more efficiently. */
714 s390_canonicalize_comparison (enum rtx_code
*code
, rtx
*op0
, rtx
*op1
)
716 /* Convert ZERO_EXTRACT back to AND to enable TM patterns. */
717 if ((*code
== EQ
|| *code
== NE
)
718 && *op1
== const0_rtx
719 && GET_CODE (*op0
) == ZERO_EXTRACT
720 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
721 && GET_CODE (XEXP (*op0
, 2)) == CONST_INT
722 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
724 rtx inner
= XEXP (*op0
, 0);
725 HOST_WIDE_INT modesize
= GET_MODE_BITSIZE (GET_MODE (inner
));
726 HOST_WIDE_INT len
= INTVAL (XEXP (*op0
, 1));
727 HOST_WIDE_INT pos
= INTVAL (XEXP (*op0
, 2));
729 if (len
> 0 && len
< modesize
730 && pos
>= 0 && pos
+ len
<= modesize
731 && modesize
<= HOST_BITS_PER_WIDE_INT
)
733 unsigned HOST_WIDE_INT block
;
734 block
= ((unsigned HOST_WIDE_INT
) 1 << len
) - 1;
735 block
<<= modesize
- pos
- len
;
737 *op0
= gen_rtx_AND (GET_MODE (inner
), inner
,
738 gen_int_mode (block
, GET_MODE (inner
)));
742 /* Narrow AND of memory against immediate to enable TM. */
743 if ((*code
== EQ
|| *code
== NE
)
744 && *op1
== const0_rtx
745 && GET_CODE (*op0
) == AND
746 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
747 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
749 rtx inner
= XEXP (*op0
, 0);
750 rtx mask
= XEXP (*op0
, 1);
752 /* Ignore paradoxical SUBREGs if all extra bits are masked out. */
753 if (GET_CODE (inner
) == SUBREG
754 && SCALAR_INT_MODE_P (GET_MODE (SUBREG_REG (inner
)))
755 && (GET_MODE_SIZE (GET_MODE (inner
))
756 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner
))))
758 & GET_MODE_MASK (GET_MODE (inner
))
759 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (inner
))))
761 inner
= SUBREG_REG (inner
);
763 /* Do not change volatile MEMs. */
764 if (MEM_P (inner
) && !MEM_VOLATILE_P (inner
))
766 int part
= s390_single_part (XEXP (*op0
, 1),
767 GET_MODE (inner
), QImode
, 0);
770 mask
= gen_int_mode (s390_extract_part (mask
, QImode
, 0), QImode
);
771 inner
= adjust_address_nv (inner
, QImode
, part
);
772 *op0
= gen_rtx_AND (QImode
, inner
, mask
);
777 /* Narrow comparisons against 0xffff to HImode if possible. */
778 if ((*code
== EQ
|| *code
== NE
)
779 && GET_CODE (*op1
) == CONST_INT
780 && INTVAL (*op1
) == 0xffff
781 && SCALAR_INT_MODE_P (GET_MODE (*op0
))
782 && (nonzero_bits (*op0
, GET_MODE (*op0
))
783 & ~(unsigned HOST_WIDE_INT
) 0xffff) == 0)
785 *op0
= gen_lowpart (HImode
, *op0
);
789 /* Remove redundant UNSPEC_CCU_TO_INT conversions if possible. */
790 if (GET_CODE (*op0
) == UNSPEC
791 && XINT (*op0
, 1) == UNSPEC_CCU_TO_INT
792 && XVECLEN (*op0
, 0) == 1
793 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCUmode
794 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
795 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
796 && *op1
== const0_rtx
)
798 enum rtx_code new_code
= UNKNOWN
;
801 case EQ
: new_code
= EQ
; break;
802 case NE
: new_code
= NE
; break;
803 case LT
: new_code
= GTU
; break;
804 case GT
: new_code
= LTU
; break;
805 case LE
: new_code
= GEU
; break;
806 case GE
: new_code
= LEU
; break;
810 if (new_code
!= UNKNOWN
)
812 *op0
= XVECEXP (*op0
, 0, 0);
817 /* Remove redundant UNSPEC_CCZ_TO_INT conversions if possible. */
818 if (GET_CODE (*op0
) == UNSPEC
819 && XINT (*op0
, 1) == UNSPEC_CCZ_TO_INT
820 && XVECLEN (*op0
, 0) == 1
821 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCZmode
822 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
823 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
824 && *op1
== const0_rtx
)
826 enum rtx_code new_code
= UNKNOWN
;
829 case EQ
: new_code
= EQ
; break;
830 case NE
: new_code
= NE
; break;
834 if (new_code
!= UNKNOWN
)
836 *op0
= XVECEXP (*op0
, 0, 0);
841 /* Simplify cascaded EQ, NE with const0_rtx. */
842 if ((*code
== NE
|| *code
== EQ
)
843 && (GET_CODE (*op0
) == EQ
|| GET_CODE (*op0
) == NE
)
844 && GET_MODE (*op0
) == SImode
845 && GET_MODE (XEXP (*op0
, 0)) == CCZ1mode
846 && REG_P (XEXP (*op0
, 0))
847 && XEXP (*op0
, 1) == const0_rtx
848 && *op1
== const0_rtx
)
850 if ((*code
== EQ
&& GET_CODE (*op0
) == NE
)
851 || (*code
== NE
&& GET_CODE (*op0
) == EQ
))
855 *op0
= XEXP (*op0
, 0);
858 /* Prefer register over memory as first operand. */
859 if (MEM_P (*op0
) && REG_P (*op1
))
861 rtx tem
= *op0
; *op0
= *op1
; *op1
= tem
;
862 *code
= swap_condition (*code
);
866 /* Emit a compare instruction suitable to implement the comparison
867 OP0 CODE OP1. Return the correct condition RTL to be placed in
868 the IF_THEN_ELSE of the conditional branch testing the result. */
871 s390_emit_compare (enum rtx_code code
, rtx op0
, rtx op1
)
873 enum machine_mode mode
= s390_select_ccmode (code
, op0
, op1
);
876 /* Do not output a redundant compare instruction if a compare_and_swap
877 pattern already computed the result and the machine modes are compatible. */
878 if (GET_MODE_CLASS (GET_MODE (op0
)) == MODE_CC
)
880 gcc_assert (s390_cc_modes_compatible (GET_MODE (op0
), mode
)
886 cc
= gen_rtx_REG (mode
, CC_REGNUM
);
887 emit_insn (gen_rtx_SET (VOIDmode
, cc
, gen_rtx_COMPARE (mode
, op0
, op1
)));
890 return gen_rtx_fmt_ee (code
, VOIDmode
, cc
, const0_rtx
);
893 /* Emit a SImode compare and swap instruction setting MEM to NEW_RTX if OLD
895 Return the correct condition RTL to be placed in the IF_THEN_ELSE of the
896 conditional branch testing the result. */
899 s390_emit_compare_and_swap (enum rtx_code code
, rtx old
, rtx mem
,
900 rtx cmp
, rtx new_rtx
)
902 emit_insn (gen_atomic_compare_and_swapsi_internal (old
, mem
, cmp
, new_rtx
));
903 return s390_emit_compare (code
, gen_rtx_REG (CCZ1mode
, CC_REGNUM
),
907 /* Emit a jump instruction to TARGET. If COND is NULL_RTX, emit an
908 unconditional jump, else a conditional jump under condition COND. */
911 s390_emit_jump (rtx target
, rtx cond
)
915 target
= gen_rtx_LABEL_REF (VOIDmode
, target
);
917 target
= gen_rtx_IF_THEN_ELSE (VOIDmode
, cond
, target
, pc_rtx
);
919 insn
= gen_rtx_SET (VOIDmode
, pc_rtx
, target
);
920 emit_jump_insn (insn
);
923 /* Return branch condition mask to implement a branch
924 specified by CODE. Return -1 for invalid comparisons. */
927 s390_branch_condition_mask (rtx code
)
929 const int CC0
= 1 << 3;
930 const int CC1
= 1 << 2;
931 const int CC2
= 1 << 1;
932 const int CC3
= 1 << 0;
934 gcc_assert (GET_CODE (XEXP (code
, 0)) == REG
);
935 gcc_assert (REGNO (XEXP (code
, 0)) == CC_REGNUM
);
936 gcc_assert (XEXP (code
, 1) == const0_rtx
);
938 switch (GET_MODE (XEXP (code
, 0)))
942 switch (GET_CODE (code
))
945 case NE
: return CC1
| CC2
| CC3
;
951 switch (GET_CODE (code
))
954 case NE
: return CC0
| CC2
| CC3
;
960 switch (GET_CODE (code
))
963 case NE
: return CC0
| CC1
| CC3
;
969 switch (GET_CODE (code
))
972 case NE
: return CC0
| CC1
| CC2
;
978 switch (GET_CODE (code
))
980 case EQ
: return CC0
| CC2
;
981 case NE
: return CC1
| CC3
;
987 switch (GET_CODE (code
))
989 case LTU
: return CC2
| CC3
; /* carry */
990 case GEU
: return CC0
| CC1
; /* no carry */
996 switch (GET_CODE (code
))
998 case GTU
: return CC0
| CC1
; /* borrow */
999 case LEU
: return CC2
| CC3
; /* no borrow */
1005 switch (GET_CODE (code
))
1007 case EQ
: return CC0
| CC2
;
1008 case NE
: return CC1
| CC3
;
1009 case LTU
: return CC1
;
1010 case GTU
: return CC3
;
1011 case LEU
: return CC1
| CC2
;
1012 case GEU
: return CC2
| CC3
;
1017 switch (GET_CODE (code
))
1019 case EQ
: return CC0
;
1020 case NE
: return CC1
| CC2
| CC3
;
1021 case LTU
: return CC1
;
1022 case GTU
: return CC2
;
1023 case LEU
: return CC0
| CC1
;
1024 case GEU
: return CC0
| CC2
;
1030 switch (GET_CODE (code
))
1032 case EQ
: return CC0
;
1033 case NE
: return CC2
| CC1
| CC3
;
1034 case LTU
: return CC2
;
1035 case GTU
: return CC1
;
1036 case LEU
: return CC0
| CC2
;
1037 case GEU
: return CC0
| CC1
;
1043 switch (GET_CODE (code
))
1045 case EQ
: return CC0
;
1046 case NE
: return CC1
| CC2
| CC3
;
1047 case LT
: return CC1
| CC3
;
1048 case GT
: return CC2
;
1049 case LE
: return CC0
| CC1
| CC3
;
1050 case GE
: return CC0
| CC2
;
1056 switch (GET_CODE (code
))
1058 case EQ
: return CC0
;
1059 case NE
: return CC1
| CC2
| CC3
;
1060 case LT
: return CC1
;
1061 case GT
: return CC2
| CC3
;
1062 case LE
: return CC0
| CC1
;
1063 case GE
: return CC0
| CC2
| CC3
;
1069 switch (GET_CODE (code
))
1071 case EQ
: return CC0
;
1072 case NE
: return CC1
| CC2
| CC3
;
1073 case LT
: return CC1
;
1074 case GT
: return CC2
;
1075 case LE
: return CC0
| CC1
;
1076 case GE
: return CC0
| CC2
;
1077 case UNORDERED
: return CC3
;
1078 case ORDERED
: return CC0
| CC1
| CC2
;
1079 case UNEQ
: return CC0
| CC3
;
1080 case UNLT
: return CC1
| CC3
;
1081 case UNGT
: return CC2
| CC3
;
1082 case UNLE
: return CC0
| CC1
| CC3
;
1083 case UNGE
: return CC0
| CC2
| CC3
;
1084 case LTGT
: return CC1
| CC2
;
1090 switch (GET_CODE (code
))
1092 case EQ
: return CC0
;
1093 case NE
: return CC2
| CC1
| CC3
;
1094 case LT
: return CC2
;
1095 case GT
: return CC1
;
1096 case LE
: return CC0
| CC2
;
1097 case GE
: return CC0
| CC1
;
1098 case UNORDERED
: return CC3
;
1099 case ORDERED
: return CC0
| CC2
| CC1
;
1100 case UNEQ
: return CC0
| CC3
;
1101 case UNLT
: return CC2
| CC3
;
1102 case UNGT
: return CC1
| CC3
;
1103 case UNLE
: return CC0
| CC2
| CC3
;
1104 case UNGE
: return CC0
| CC1
| CC3
;
1105 case LTGT
: return CC2
| CC1
;
1116 /* Return branch condition mask to implement a compare and branch
1117 specified by CODE. Return -1 for invalid comparisons. */
1120 s390_compare_and_branch_condition_mask (rtx code
)
1122 const int CC0
= 1 << 3;
1123 const int CC1
= 1 << 2;
1124 const int CC2
= 1 << 1;
1126 switch (GET_CODE (code
))
1150 /* If INV is false, return assembler mnemonic string to implement
1151 a branch specified by CODE. If INV is true, return mnemonic
1152 for the corresponding inverted branch. */
1155 s390_branch_condition_mnemonic (rtx code
, int inv
)
1159 static const char *const mnemonic
[16] =
1161 NULL
, "o", "h", "nle",
1162 "l", "nhe", "lh", "ne",
1163 "e", "nlh", "he", "nl",
1164 "le", "nh", "no", NULL
1167 if (GET_CODE (XEXP (code
, 0)) == REG
1168 && REGNO (XEXP (code
, 0)) == CC_REGNUM
1169 && XEXP (code
, 1) == const0_rtx
)
1170 mask
= s390_branch_condition_mask (code
);
1172 mask
= s390_compare_and_branch_condition_mask (code
);
1174 gcc_assert (mask
>= 0);
1179 gcc_assert (mask
>= 1 && mask
<= 14);
1181 return mnemonic
[mask
];
1184 /* Return the part of op which has a value different from def.
1185 The size of the part is determined by mode.
1186 Use this function only if you already know that op really
1187 contains such a part. */
1189 unsigned HOST_WIDE_INT
1190 s390_extract_part (rtx op
, enum machine_mode mode
, int def
)
1192 unsigned HOST_WIDE_INT value
= 0;
1193 int max_parts
= HOST_BITS_PER_WIDE_INT
/ GET_MODE_BITSIZE (mode
);
1194 int part_bits
= GET_MODE_BITSIZE (mode
);
1195 unsigned HOST_WIDE_INT part_mask
1196 = ((unsigned HOST_WIDE_INT
)1 << part_bits
) - 1;
1199 for (i
= 0; i
< max_parts
; i
++)
1202 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1204 value
>>= part_bits
;
1206 if ((value
& part_mask
) != (def
& part_mask
))
1207 return value
& part_mask
;
1213 /* If OP is an integer constant of mode MODE with exactly one
1214 part of mode PART_MODE unequal to DEF, return the number of that
1215 part. Otherwise, return -1. */
1218 s390_single_part (rtx op
,
1219 enum machine_mode mode
,
1220 enum machine_mode part_mode
,
1223 unsigned HOST_WIDE_INT value
= 0;
1224 int n_parts
= GET_MODE_SIZE (mode
) / GET_MODE_SIZE (part_mode
);
1225 unsigned HOST_WIDE_INT part_mask
1226 = ((unsigned HOST_WIDE_INT
)1 << GET_MODE_BITSIZE (part_mode
)) - 1;
1229 if (GET_CODE (op
) != CONST_INT
)
1232 for (i
= 0; i
< n_parts
; i
++)
1235 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1237 value
>>= GET_MODE_BITSIZE (part_mode
);
1239 if ((value
& part_mask
) != (def
& part_mask
))
1247 return part
== -1 ? -1 : n_parts
- 1 - part
;
1250 /* Return true if IN contains a contiguous bitfield in the lower SIZE
1251 bits and no other bits are set in IN. POS and LENGTH can be used
1252 to obtain the start position and the length of the bitfield.
1254 POS gives the position of the first bit of the bitfield counting
1255 from the lowest order bit starting with zero. In order to use this
1256 value for S/390 instructions this has to be converted to "bits big
1260 s390_contiguous_bitmask_p (unsigned HOST_WIDE_INT in
, int size
,
1261 int *pos
, int *length
)
1266 unsigned HOST_WIDE_INT mask
= 1ULL;
1267 bool contiguous
= false;
1269 for (i
= 0; i
< size
; mask
<<= 1, i
++)
1293 /* Calculate a mask for all bits beyond the contiguous bits. */
1294 mask
= (-1LL & ~(((1ULL << (tmp_length
+ tmp_pos
- 1)) << 1) - 1));
1299 if (tmp_length
+ tmp_pos
- 1 > size
)
1303 *length
= tmp_length
;
1311 /* Check whether we can (and want to) split a double-word
1312 move in mode MODE from SRC to DST into two single-word
1313 moves, moving the subword FIRST_SUBWORD first. */
1316 s390_split_ok_p (rtx dst
, rtx src
, enum machine_mode mode
, int first_subword
)
1318 /* Floating point registers cannot be split. */
1319 if (FP_REG_P (src
) || FP_REG_P (dst
))
1322 /* We don't need to split if operands are directly accessible. */
1323 if (s_operand (src
, mode
) || s_operand (dst
, mode
))
1326 /* Non-offsettable memory references cannot be split. */
1327 if ((GET_CODE (src
) == MEM
&& !offsettable_memref_p (src
))
1328 || (GET_CODE (dst
) == MEM
&& !offsettable_memref_p (dst
)))
1331 /* Moving the first subword must not clobber a register
1332 needed to move the second subword. */
1333 if (register_operand (dst
, mode
))
1335 rtx subreg
= operand_subword (dst
, first_subword
, 0, mode
);
1336 if (reg_overlap_mentioned_p (subreg
, src
))
1343 /* Return true if it can be proven that [MEM1, MEM1 + SIZE]
1344 and [MEM2, MEM2 + SIZE] do overlap and false
1348 s390_overlap_p (rtx mem1
, rtx mem2
, HOST_WIDE_INT size
)
1350 rtx addr1
, addr2
, addr_delta
;
1351 HOST_WIDE_INT delta
;
1353 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1359 addr1
= XEXP (mem1
, 0);
1360 addr2
= XEXP (mem2
, 0);
1362 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1364 /* This overlapping check is used by peepholes merging memory block operations.
1365 Overlapping operations would otherwise be recognized by the S/390 hardware
1366 and would fall back to a slower implementation. Allowing overlapping
1367 operations would lead to slow code but not to wrong code. Therefore we are
1368 somewhat optimistic if we cannot prove that the memory blocks are
1370 That's why we return false here although this may accept operations on
1371 overlapping memory areas. */
1372 if (!addr_delta
|| GET_CODE (addr_delta
) != CONST_INT
)
1375 delta
= INTVAL (addr_delta
);
1378 || (delta
> 0 && delta
< size
)
1379 || (delta
< 0 && -delta
< size
))
1385 /* Check whether the address of memory reference MEM2 equals exactly
1386 the address of memory reference MEM1 plus DELTA. Return true if
1387 we can prove this to be the case, false otherwise. */
1390 s390_offset_p (rtx mem1
, rtx mem2
, rtx delta
)
1392 rtx addr1
, addr2
, addr_delta
;
1394 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1397 addr1
= XEXP (mem1
, 0);
1398 addr2
= XEXP (mem2
, 0);
1400 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1401 if (!addr_delta
|| !rtx_equal_p (addr_delta
, delta
))
1407 /* Expand logical operator CODE in mode MODE with operands OPERANDS. */
1410 s390_expand_logical_operator (enum rtx_code code
, enum machine_mode mode
,
1413 enum machine_mode wmode
= mode
;
1414 rtx dst
= operands
[0];
1415 rtx src1
= operands
[1];
1416 rtx src2
= operands
[2];
1419 /* If we cannot handle the operation directly, use a temp register. */
1420 if (!s390_logical_operator_ok_p (operands
))
1421 dst
= gen_reg_rtx (mode
);
1423 /* QImode and HImode patterns make sense only if we have a destination
1424 in memory. Otherwise perform the operation in SImode. */
1425 if ((mode
== QImode
|| mode
== HImode
) && GET_CODE (dst
) != MEM
)
1428 /* Widen operands if required. */
1431 if (GET_CODE (dst
) == SUBREG
1432 && (tem
= simplify_subreg (wmode
, dst
, mode
, 0)) != 0)
1434 else if (REG_P (dst
))
1435 dst
= gen_rtx_SUBREG (wmode
, dst
, 0);
1437 dst
= gen_reg_rtx (wmode
);
1439 if (GET_CODE (src1
) == SUBREG
1440 && (tem
= simplify_subreg (wmode
, src1
, mode
, 0)) != 0)
1442 else if (GET_MODE (src1
) != VOIDmode
)
1443 src1
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src1
), 0);
1445 if (GET_CODE (src2
) == SUBREG
1446 && (tem
= simplify_subreg (wmode
, src2
, mode
, 0)) != 0)
1448 else if (GET_MODE (src2
) != VOIDmode
)
1449 src2
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src2
), 0);
1452 /* Emit the instruction. */
1453 op
= gen_rtx_SET (VOIDmode
, dst
, gen_rtx_fmt_ee (code
, wmode
, src1
, src2
));
1454 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
1455 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clob
)));
1457 /* Fix up the destination if needed. */
1458 if (dst
!= operands
[0])
1459 emit_move_insn (operands
[0], gen_lowpart (mode
, dst
));
1462 /* Check whether OPERANDS are OK for a logical operation (AND, IOR, XOR). */
1465 s390_logical_operator_ok_p (rtx
*operands
)
1467 /* If the destination operand is in memory, it needs to coincide
1468 with one of the source operands. After reload, it has to be
1469 the first source operand. */
1470 if (GET_CODE (operands
[0]) == MEM
)
1471 return rtx_equal_p (operands
[0], operands
[1])
1472 || (!reload_completed
&& rtx_equal_p (operands
[0], operands
[2]));
1477 /* Narrow logical operation CODE of memory operand MEMOP with immediate
1478 operand IMMOP to switch from SS to SI type instructions. */
1481 s390_narrow_logical_operator (enum rtx_code code
, rtx
*memop
, rtx
*immop
)
1483 int def
= code
== AND
? -1 : 0;
1487 gcc_assert (GET_CODE (*memop
) == MEM
);
1488 gcc_assert (!MEM_VOLATILE_P (*memop
));
1490 mask
= s390_extract_part (*immop
, QImode
, def
);
1491 part
= s390_single_part (*immop
, GET_MODE (*memop
), QImode
, def
);
1492 gcc_assert (part
>= 0);
1494 *memop
= adjust_address (*memop
, QImode
, part
);
1495 *immop
= gen_int_mode (mask
, QImode
);
1499 /* How to allocate a 'struct machine_function'. */
1501 static struct machine_function
*
1502 s390_init_machine_status (void)
1504 return ggc_alloc_cleared_machine_function ();
1508 s390_option_override (void)
1510 /* Set up function hooks. */
1511 init_machine_status
= s390_init_machine_status
;
1513 /* Architecture mode defaults according to ABI. */
1514 if (!(target_flags_explicit
& MASK_ZARCH
))
1517 target_flags
|= MASK_ZARCH
;
1519 target_flags
&= ~MASK_ZARCH
;
1522 /* Set the march default in case it hasn't been specified on
1524 if (s390_arch
== PROCESSOR_max
)
1526 s390_arch_string
= TARGET_ZARCH
? "z900" : "g5";
1527 s390_arch
= TARGET_ZARCH
? PROCESSOR_2064_Z900
: PROCESSOR_9672_G5
;
1528 s390_arch_flags
= processor_flags_table
[(int)s390_arch
];
1531 /* Determine processor to tune for. */
1532 if (s390_tune
== PROCESSOR_max
)
1534 s390_tune
= s390_arch
;
1535 s390_tune_flags
= s390_arch_flags
;
1538 /* Sanity checks. */
1539 if (TARGET_ZARCH
&& !TARGET_CPU_ZARCH
)
1540 error ("z/Architecture mode not supported on %s", s390_arch_string
);
1541 if (TARGET_64BIT
&& !TARGET_ZARCH
)
1542 error ("64-bit ABI not supported in ESA/390 mode");
1544 /* Use hardware DFP if available and not explicitly disabled by
1545 user. E.g. with -m31 -march=z10 -mzarch */
1546 if (!(target_flags_explicit
& MASK_HARD_DFP
) && TARGET_DFP
)
1547 target_flags
|= MASK_HARD_DFP
;
1549 if (TARGET_HARD_DFP
&& !TARGET_DFP
)
1551 if (target_flags_explicit
& MASK_HARD_DFP
)
1553 if (!TARGET_CPU_DFP
)
1554 error ("hardware decimal floating point instructions"
1555 " not available on %s", s390_arch_string
);
1557 error ("hardware decimal floating point instructions"
1558 " not available in ESA/390 mode");
1561 target_flags
&= ~MASK_HARD_DFP
;
1564 if ((target_flags_explicit
& MASK_SOFT_FLOAT
) && TARGET_SOFT_FLOAT
)
1566 if ((target_flags_explicit
& MASK_HARD_DFP
) && TARGET_HARD_DFP
)
1567 error ("-mhard-dfp can%'t be used in conjunction with -msoft-float");
1569 target_flags
&= ~MASK_HARD_DFP
;
1572 /* Set processor cost function. */
1575 case PROCESSOR_2084_Z990
:
1576 s390_cost
= &z990_cost
;
1578 case PROCESSOR_2094_Z9_109
:
1579 s390_cost
= &z9_109_cost
;
1581 case PROCESSOR_2097_Z10
:
1582 s390_cost
= &z10_cost
;
1583 case PROCESSOR_2817_Z196
:
1584 s390_cost
= &z196_cost
;
1587 s390_cost
= &z900_cost
;
1590 if (TARGET_BACKCHAIN
&& TARGET_PACKED_STACK
&& TARGET_HARD_FLOAT
)
1591 error ("-mbackchain -mpacked-stack -mhard-float are not supported "
1594 if (s390_stack_size
)
1596 if (s390_stack_guard
>= s390_stack_size
)
1597 error ("stack size must be greater than the stack guard value");
1598 else if (s390_stack_size
> 1 << 16)
1599 error ("stack size must not be greater than 64k");
1601 else if (s390_stack_guard
)
1602 error ("-mstack-guard implies use of -mstack-size");
1604 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1605 if (!(target_flags_explicit
& MASK_LONG_DOUBLE_128
))
1606 target_flags
|= MASK_LONG_DOUBLE_128
;
1609 if (s390_tune
== PROCESSOR_2097_Z10
1610 || s390_tune
== PROCESSOR_2817_Z196
)
1612 maybe_set_param_value (PARAM_MAX_UNROLLED_INSNS
, 100,
1613 global_options
.x_param_values
,
1614 global_options_set
.x_param_values
);
1615 maybe_set_param_value (PARAM_MAX_UNROLL_TIMES
, 32,
1616 global_options
.x_param_values
,
1617 global_options_set
.x_param_values
);
1618 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEELED_INSNS
, 2000,
1619 global_options
.x_param_values
,
1620 global_options_set
.x_param_values
);
1621 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEEL_TIMES
, 64,
1622 global_options
.x_param_values
,
1623 global_options_set
.x_param_values
);
1626 maybe_set_param_value (PARAM_MAX_PENDING_LIST_LENGTH
, 256,
1627 global_options
.x_param_values
,
1628 global_options_set
.x_param_values
);
1629 /* values for loop prefetching */
1630 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE
, 256,
1631 global_options
.x_param_values
,
1632 global_options_set
.x_param_values
);
1633 maybe_set_param_value (PARAM_L1_CACHE_SIZE
, 128,
1634 global_options
.x_param_values
,
1635 global_options_set
.x_param_values
);
1636 /* s390 has more than 2 levels and the size is much larger. Since
1637 we are always running virtualized assume that we only get a small
1638 part of the caches above l1. */
1639 maybe_set_param_value (PARAM_L2_CACHE_SIZE
, 1500,
1640 global_options
.x_param_values
,
1641 global_options_set
.x_param_values
);
1642 maybe_set_param_value (PARAM_PREFETCH_MIN_INSN_TO_MEM_RATIO
, 2,
1643 global_options
.x_param_values
,
1644 global_options_set
.x_param_values
);
1645 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
, 6,
1646 global_options
.x_param_values
,
1647 global_options_set
.x_param_values
);
1649 /* This cannot reside in s390_option_optimization_table since HAVE_prefetch
1650 requires the arch flags to be evaluated already. Since prefetching
1651 is beneficial on s390, we enable it if available. */
1652 if (flag_prefetch_loop_arrays
< 0 && HAVE_prefetch
&& optimize
>= 3)
1653 flag_prefetch_loop_arrays
= 1;
1655 /* Use the alternative scheduling-pressure algorithm by default. */
1656 maybe_set_param_value (PARAM_SCHED_PRESSURE_ALGORITHM
, 2,
1657 global_options
.x_param_values
,
1658 global_options_set
.x_param_values
);
1662 /* Don't emit DWARF3/4 unless specifically selected. The TPF
1663 debuggers do not yet support DWARF 3/4. */
1664 if (!global_options_set
.x_dwarf_strict
)
1666 if (!global_options_set
.x_dwarf_version
)
1671 /* Map for smallest class containing reg regno. */
1673 const enum reg_class regclass_map
[FIRST_PSEUDO_REGISTER
] =
1674 { GENERAL_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1675 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1676 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1677 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1678 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1679 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1680 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1681 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1682 ADDR_REGS
, CC_REGS
, ADDR_REGS
, ADDR_REGS
,
1683 ACCESS_REGS
, ACCESS_REGS
1686 /* Return attribute type of insn. */
1688 static enum attr_type
1689 s390_safe_attr_type (rtx insn
)
1691 if (recog_memoized (insn
) >= 0)
1692 return get_attr_type (insn
);
1697 /* Return true if DISP is a valid short displacement. */
1700 s390_short_displacement (rtx disp
)
1702 /* No displacement is OK. */
1706 /* Without the long displacement facility we don't need to
1707 distingiush between long and short displacement. */
1708 if (!TARGET_LONG_DISPLACEMENT
)
1711 /* Integer displacement in range. */
1712 if (GET_CODE (disp
) == CONST_INT
)
1713 return INTVAL (disp
) >= 0 && INTVAL (disp
) < 4096;
1715 /* GOT offset is not OK, the GOT can be large. */
1716 if (GET_CODE (disp
) == CONST
1717 && GET_CODE (XEXP (disp
, 0)) == UNSPEC
1718 && (XINT (XEXP (disp
, 0), 1) == UNSPEC_GOT
1719 || XINT (XEXP (disp
, 0), 1) == UNSPEC_GOTNTPOFF
))
1722 /* All other symbolic constants are literal pool references,
1723 which are OK as the literal pool must be small. */
1724 if (GET_CODE (disp
) == CONST
)
1730 /* Decompose a RTL expression ADDR for a memory address into
1731 its components, returned in OUT.
1733 Returns false if ADDR is not a valid memory address, true
1734 otherwise. If OUT is NULL, don't return the components,
1735 but check for validity only.
1737 Note: Only addresses in canonical form are recognized.
1738 LEGITIMIZE_ADDRESS should convert non-canonical forms to the
1739 canonical form so that they will be recognized. */
1742 s390_decompose_address (rtx addr
, struct s390_address
*out
)
1744 HOST_WIDE_INT offset
= 0;
1745 rtx base
= NULL_RTX
;
1746 rtx indx
= NULL_RTX
;
1747 rtx disp
= NULL_RTX
;
1749 bool pointer
= false;
1750 bool base_ptr
= false;
1751 bool indx_ptr
= false;
1752 bool literal_pool
= false;
1754 /* We may need to substitute the literal pool base register into the address
1755 below. However, at this point we do not know which register is going to
1756 be used as base, so we substitute the arg pointer register. This is going
1757 to be treated as holding a pointer below -- it shouldn't be used for any
1759 rtx fake_pool_base
= gen_rtx_REG (Pmode
, ARG_POINTER_REGNUM
);
1761 /* Decompose address into base + index + displacement. */
1763 if (GET_CODE (addr
) == REG
|| GET_CODE (addr
) == UNSPEC
)
1766 else if (GET_CODE (addr
) == PLUS
)
1768 rtx op0
= XEXP (addr
, 0);
1769 rtx op1
= XEXP (addr
, 1);
1770 enum rtx_code code0
= GET_CODE (op0
);
1771 enum rtx_code code1
= GET_CODE (op1
);
1773 if (code0
== REG
|| code0
== UNSPEC
)
1775 if (code1
== REG
|| code1
== UNSPEC
)
1777 indx
= op0
; /* index + base */
1783 base
= op0
; /* base + displacement */
1788 else if (code0
== PLUS
)
1790 indx
= XEXP (op0
, 0); /* index + base + disp */
1791 base
= XEXP (op0
, 1);
1802 disp
= addr
; /* displacement */
1804 /* Extract integer part of displacement. */
1808 if (GET_CODE (disp
) == CONST_INT
)
1810 offset
= INTVAL (disp
);
1813 else if (GET_CODE (disp
) == CONST
1814 && GET_CODE (XEXP (disp
, 0)) == PLUS
1815 && GET_CODE (XEXP (XEXP (disp
, 0), 1)) == CONST_INT
)
1817 offset
= INTVAL (XEXP (XEXP (disp
, 0), 1));
1818 disp
= XEXP (XEXP (disp
, 0), 0);
1822 /* Strip off CONST here to avoid special case tests later. */
1823 if (disp
&& GET_CODE (disp
) == CONST
)
1824 disp
= XEXP (disp
, 0);
1826 /* We can convert literal pool addresses to
1827 displacements by basing them off the base register. */
1828 if (disp
&& GET_CODE (disp
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (disp
))
1830 /* Either base or index must be free to hold the base register. */
1832 base
= fake_pool_base
, literal_pool
= true;
1834 indx
= fake_pool_base
, literal_pool
= true;
1838 /* Mark up the displacement. */
1839 disp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, disp
),
1840 UNSPEC_LTREL_OFFSET
);
1843 /* Validate base register. */
1846 if (GET_CODE (base
) == UNSPEC
)
1847 switch (XINT (base
, 1))
1851 disp
= gen_rtx_UNSPEC (Pmode
,
1852 gen_rtvec (1, XVECEXP (base
, 0, 0)),
1853 UNSPEC_LTREL_OFFSET
);
1857 base
= XVECEXP (base
, 0, 1);
1860 case UNSPEC_LTREL_BASE
:
1861 if (XVECLEN (base
, 0) == 1)
1862 base
= fake_pool_base
, literal_pool
= true;
1864 base
= XVECEXP (base
, 0, 1);
1872 || (GET_MODE (base
) != SImode
1873 && GET_MODE (base
) != Pmode
))
1876 if (REGNO (base
) == STACK_POINTER_REGNUM
1877 || REGNO (base
) == FRAME_POINTER_REGNUM
1878 || ((reload_completed
|| reload_in_progress
)
1879 && frame_pointer_needed
1880 && REGNO (base
) == HARD_FRAME_POINTER_REGNUM
)
1881 || REGNO (base
) == ARG_POINTER_REGNUM
1883 && REGNO (base
) == PIC_OFFSET_TABLE_REGNUM
))
1884 pointer
= base_ptr
= true;
1886 if ((reload_completed
|| reload_in_progress
)
1887 && base
== cfun
->machine
->base_reg
)
1888 pointer
= base_ptr
= literal_pool
= true;
1891 /* Validate index register. */
1894 if (GET_CODE (indx
) == UNSPEC
)
1895 switch (XINT (indx
, 1))
1899 disp
= gen_rtx_UNSPEC (Pmode
,
1900 gen_rtvec (1, XVECEXP (indx
, 0, 0)),
1901 UNSPEC_LTREL_OFFSET
);
1905 indx
= XVECEXP (indx
, 0, 1);
1908 case UNSPEC_LTREL_BASE
:
1909 if (XVECLEN (indx
, 0) == 1)
1910 indx
= fake_pool_base
, literal_pool
= true;
1912 indx
= XVECEXP (indx
, 0, 1);
1920 || (GET_MODE (indx
) != SImode
1921 && GET_MODE (indx
) != Pmode
))
1924 if (REGNO (indx
) == STACK_POINTER_REGNUM
1925 || REGNO (indx
) == FRAME_POINTER_REGNUM
1926 || ((reload_completed
|| reload_in_progress
)
1927 && frame_pointer_needed
1928 && REGNO (indx
) == HARD_FRAME_POINTER_REGNUM
)
1929 || REGNO (indx
) == ARG_POINTER_REGNUM
1931 && REGNO (indx
) == PIC_OFFSET_TABLE_REGNUM
))
1932 pointer
= indx_ptr
= true;
1934 if ((reload_completed
|| reload_in_progress
)
1935 && indx
== cfun
->machine
->base_reg
)
1936 pointer
= indx_ptr
= literal_pool
= true;
1939 /* Prefer to use pointer as base, not index. */
1940 if (base
&& indx
&& !base_ptr
1941 && (indx_ptr
|| (!REG_POINTER (base
) && REG_POINTER (indx
))))
1948 /* Validate displacement. */
1951 /* If virtual registers are involved, the displacement will change later
1952 anyway as the virtual registers get eliminated. This could make a
1953 valid displacement invalid, but it is more likely to make an invalid
1954 displacement valid, because we sometimes access the register save area
1955 via negative offsets to one of those registers.
1956 Thus we don't check the displacement for validity here. If after
1957 elimination the displacement turns out to be invalid after all,
1958 this is fixed up by reload in any case. */
1959 if (base
!= arg_pointer_rtx
1960 && indx
!= arg_pointer_rtx
1961 && base
!= return_address_pointer_rtx
1962 && indx
!= return_address_pointer_rtx
1963 && base
!= frame_pointer_rtx
1964 && indx
!= frame_pointer_rtx
1965 && base
!= virtual_stack_vars_rtx
1966 && indx
!= virtual_stack_vars_rtx
)
1967 if (!DISP_IN_RANGE (offset
))
1972 /* All the special cases are pointers. */
1975 /* In the small-PIC case, the linker converts @GOT
1976 and @GOTNTPOFF offsets to possible displacements. */
1977 if (GET_CODE (disp
) == UNSPEC
1978 && (XINT (disp
, 1) == UNSPEC_GOT
1979 || XINT (disp
, 1) == UNSPEC_GOTNTPOFF
)
1985 /* Accept pool label offsets. */
1986 else if (GET_CODE (disp
) == UNSPEC
1987 && XINT (disp
, 1) == UNSPEC_POOL_OFFSET
)
1990 /* Accept literal pool references. */
1991 else if (GET_CODE (disp
) == UNSPEC
1992 && XINT (disp
, 1) == UNSPEC_LTREL_OFFSET
)
1994 /* In case CSE pulled a non literal pool reference out of
1995 the pool we have to reject the address. This is
1996 especially important when loading the GOT pointer on non
1997 zarch CPUs. In this case the literal pool contains an lt
1998 relative offset to the _GLOBAL_OFFSET_TABLE_ label which
1999 will most likely exceed the displacement. */
2000 if (GET_CODE (XVECEXP (disp
, 0, 0)) != SYMBOL_REF
2001 || !CONSTANT_POOL_ADDRESS_P (XVECEXP (disp
, 0, 0)))
2004 orig_disp
= gen_rtx_CONST (Pmode
, disp
);
2007 /* If we have an offset, make sure it does not
2008 exceed the size of the constant pool entry. */
2009 rtx sym
= XVECEXP (disp
, 0, 0);
2010 if (offset
>= GET_MODE_SIZE (get_pool_mode (sym
)))
2013 orig_disp
= plus_constant (Pmode
, orig_disp
, offset
);
2028 out
->disp
= orig_disp
;
2029 out
->pointer
= pointer
;
2030 out
->literal_pool
= literal_pool
;
2036 /* Decompose a RTL expression OP for a shift count into its components,
2037 and return the base register in BASE and the offset in OFFSET.
2039 Return true if OP is a valid shift count, false if not. */
2042 s390_decompose_shift_count (rtx op
, rtx
*base
, HOST_WIDE_INT
*offset
)
2044 HOST_WIDE_INT off
= 0;
2046 /* We can have an integer constant, an address register,
2047 or a sum of the two. */
2048 if (GET_CODE (op
) == CONST_INT
)
2053 if (op
&& GET_CODE (op
) == PLUS
&& GET_CODE (XEXP (op
, 1)) == CONST_INT
)
2055 off
= INTVAL (XEXP (op
, 1));
2058 while (op
&& GET_CODE (op
) == SUBREG
)
2059 op
= SUBREG_REG (op
);
2061 if (op
&& GET_CODE (op
) != REG
)
2073 /* Return true if CODE is a valid address without index. */
2076 s390_legitimate_address_without_index_p (rtx op
)
2078 struct s390_address addr
;
2080 if (!s390_decompose_address (XEXP (op
, 0), &addr
))
2089 /* Return true if ADDR is of kind symbol_ref or symbol_ref + const_int
2090 and return these parts in SYMREF and ADDEND. You can pass NULL in
2091 SYMREF and/or ADDEND if you are not interested in these values.
2092 Literal pool references are *not* considered symbol references. */
2095 s390_symref_operand_p (rtx addr
, rtx
*symref
, HOST_WIDE_INT
*addend
)
2097 HOST_WIDE_INT tmpaddend
= 0;
2099 if (GET_CODE (addr
) == CONST
)
2100 addr
= XEXP (addr
, 0);
2102 if (GET_CODE (addr
) == PLUS
)
2104 if (GET_CODE (XEXP (addr
, 0)) == SYMBOL_REF
2105 && !CONSTANT_POOL_ADDRESS_P (XEXP (addr
, 0))
2106 && CONST_INT_P (XEXP (addr
, 1)))
2108 tmpaddend
= INTVAL (XEXP (addr
, 1));
2109 addr
= XEXP (addr
, 0);
2115 if (GET_CODE (addr
) != SYMBOL_REF
|| CONSTANT_POOL_ADDRESS_P (addr
))
2121 *addend
= tmpaddend
;
2127 /* Return true if the address in OP is valid for constraint letter C
2128 if wrapped in a MEM rtx. Set LIT_POOL_OK to true if it literal
2129 pool MEMs should be accepted. Only the Q, R, S, T constraint
2130 letters are allowed for C. */
2133 s390_check_qrst_address (char c
, rtx op
, bool lit_pool_ok
)
2135 struct s390_address addr
;
2136 bool decomposed
= false;
2138 /* This check makes sure that no symbolic address (except literal
2139 pool references) are accepted by the R or T constraints. */
2140 if (s390_symref_operand_p (op
, NULL
, NULL
))
2143 /* Ensure literal pool references are only accepted if LIT_POOL_OK. */
2146 if (!s390_decompose_address (op
, &addr
))
2148 if (addr
.literal_pool
)
2155 case 'Q': /* no index short displacement */
2156 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2160 if (!s390_short_displacement (addr
.disp
))
2164 case 'R': /* with index short displacement */
2165 if (TARGET_LONG_DISPLACEMENT
)
2167 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2169 if (!s390_short_displacement (addr
.disp
))
2172 /* Any invalid address here will be fixed up by reload,
2173 so accept it for the most generic constraint. */
2176 case 'S': /* no index long displacement */
2177 if (!TARGET_LONG_DISPLACEMENT
)
2179 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2183 if (s390_short_displacement (addr
.disp
))
2187 case 'T': /* with index long displacement */
2188 if (!TARGET_LONG_DISPLACEMENT
)
2190 /* Any invalid address here will be fixed up by reload,
2191 so accept it for the most generic constraint. */
2192 if ((decomposed
|| s390_decompose_address (op
, &addr
))
2193 && s390_short_displacement (addr
.disp
))
2203 /* Evaluates constraint strings described by the regular expression
2204 ([A|B|Z](Q|R|S|T))|U|W|Y and returns 1 if OP is a valid operand for
2205 the constraint given in STR, or 0 else. */
2208 s390_mem_constraint (const char *str
, rtx op
)
2215 /* Check for offsettable variants of memory constraints. */
2216 if (!MEM_P (op
) || MEM_VOLATILE_P (op
))
2218 if ((reload_completed
|| reload_in_progress
)
2219 ? !offsettable_memref_p (op
) : !offsettable_nonstrict_memref_p (op
))
2221 return s390_check_qrst_address (str
[1], XEXP (op
, 0), true);
2223 /* Check for non-literal-pool variants of memory constraints. */
2226 return s390_check_qrst_address (str
[1], XEXP (op
, 0), false);
2231 if (GET_CODE (op
) != MEM
)
2233 return s390_check_qrst_address (c
, XEXP (op
, 0), true);
2235 return (s390_check_qrst_address ('Q', op
, true)
2236 || s390_check_qrst_address ('R', op
, true));
2238 return (s390_check_qrst_address ('S', op
, true)
2239 || s390_check_qrst_address ('T', op
, true));
2241 /* Simply check for the basic form of a shift count. Reload will
2242 take care of making sure we have a proper base register. */
2243 if (!s390_decompose_shift_count (op
, NULL
, NULL
))
2247 return s390_check_qrst_address (str
[1], op
, true);
2255 /* Evaluates constraint strings starting with letter O. Input
2256 parameter C is the second letter following the "O" in the constraint
2257 string. Returns 1 if VALUE meets the respective constraint and 0
2261 s390_O_constraint_str (const char c
, HOST_WIDE_INT value
)
2269 return trunc_int_for_mode (value
, SImode
) == value
;
2273 || s390_single_part (GEN_INT (value
), DImode
, SImode
, 0) == 1;
2276 return s390_single_part (GEN_INT (value
- 1), DImode
, SImode
, -1) == 1;
2284 /* Evaluates constraint strings starting with letter N. Parameter STR
2285 contains the letters following letter "N" in the constraint string.
2286 Returns true if VALUE matches the constraint. */
2289 s390_N_constraint_str (const char *str
, HOST_WIDE_INT value
)
2291 enum machine_mode mode
, part_mode
;
2293 int part
, part_goal
;
2299 part_goal
= str
[0] - '0';
2343 if (GET_MODE_SIZE (mode
) <= GET_MODE_SIZE (part_mode
))
2346 part
= s390_single_part (GEN_INT (value
), mode
, part_mode
, def
);
2349 if (part_goal
!= -1 && part_goal
!= part
)
2356 /* Returns true if the input parameter VALUE is a float zero. */
2359 s390_float_const_zero_p (rtx value
)
2361 return (GET_MODE_CLASS (GET_MODE (value
)) == MODE_FLOAT
2362 && value
== CONST0_RTX (GET_MODE (value
)));
2365 /* Implement TARGET_REGISTER_MOVE_COST. */
2368 s390_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2369 reg_class_t from
, reg_class_t to
)
2371 /* On s390, copy between fprs and gprs is expensive. */
2372 if ((reg_classes_intersect_p (from
, GENERAL_REGS
)
2373 && reg_classes_intersect_p (to
, FP_REGS
))
2374 || (reg_classes_intersect_p (from
, FP_REGS
)
2375 && reg_classes_intersect_p (to
, GENERAL_REGS
)))
2381 /* Implement TARGET_MEMORY_MOVE_COST. */
2384 s390_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2385 reg_class_t rclass ATTRIBUTE_UNUSED
,
2386 bool in ATTRIBUTE_UNUSED
)
2391 /* Compute a (partial) cost for rtx X. Return true if the complete
2392 cost has been computed, and false if subexpressions should be
2393 scanned. In either case, *TOTAL contains the cost result.
2394 CODE contains GET_CODE (x), OUTER_CODE contains the code
2395 of the superexpression of x. */
2398 s390_rtx_costs (rtx x
, int code
, int outer_code
, int opno ATTRIBUTE_UNUSED
,
2399 int *total
, bool speed ATTRIBUTE_UNUSED
)
2422 *total
= COSTS_N_INSNS (1);
2427 *total
= COSTS_N_INSNS (1);
2431 switch (GET_MODE (x
))
2435 rtx left
= XEXP (x
, 0);
2436 rtx right
= XEXP (x
, 1);
2437 if (GET_CODE (right
) == CONST_INT
2438 && CONST_OK_FOR_K (INTVAL (right
)))
2439 *total
= s390_cost
->mhi
;
2440 else if (GET_CODE (left
) == SIGN_EXTEND
)
2441 *total
= s390_cost
->mh
;
2443 *total
= s390_cost
->ms
; /* msr, ms, msy */
2448 rtx left
= XEXP (x
, 0);
2449 rtx right
= XEXP (x
, 1);
2452 if (GET_CODE (right
) == CONST_INT
2453 && CONST_OK_FOR_K (INTVAL (right
)))
2454 *total
= s390_cost
->mghi
;
2455 else if (GET_CODE (left
) == SIGN_EXTEND
)
2456 *total
= s390_cost
->msgf
;
2458 *total
= s390_cost
->msg
; /* msgr, msg */
2460 else /* TARGET_31BIT */
2462 if (GET_CODE (left
) == SIGN_EXTEND
2463 && GET_CODE (right
) == SIGN_EXTEND
)
2464 /* mulsidi case: mr, m */
2465 *total
= s390_cost
->m
;
2466 else if (GET_CODE (left
) == ZERO_EXTEND
2467 && GET_CODE (right
) == ZERO_EXTEND
2468 && TARGET_CPU_ZARCH
)
2469 /* umulsidi case: ml, mlr */
2470 *total
= s390_cost
->ml
;
2472 /* Complex calculation is required. */
2473 *total
= COSTS_N_INSNS (40);
2479 *total
= s390_cost
->mult_df
;
2482 *total
= s390_cost
->mxbr
;
2490 switch (GET_MODE (x
))
2493 *total
= s390_cost
->madbr
;
2496 *total
= s390_cost
->maebr
;
2501 /* Negate in the third argument is free: FMSUB. */
2502 if (GET_CODE (XEXP (x
, 2)) == NEG
)
2504 *total
+= (rtx_cost (XEXP (x
, 0), FMA
, 0, speed
)
2505 + rtx_cost (XEXP (x
, 1), FMA
, 1, speed
)
2506 + rtx_cost (XEXP (XEXP (x
, 2), 0), FMA
, 2, speed
));
2513 if (GET_MODE (x
) == TImode
) /* 128 bit division */
2514 *total
= s390_cost
->dlgr
;
2515 else if (GET_MODE (x
) == DImode
)
2517 rtx right
= XEXP (x
, 1);
2518 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2519 *total
= s390_cost
->dlr
;
2520 else /* 64 by 64 bit division */
2521 *total
= s390_cost
->dlgr
;
2523 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2524 *total
= s390_cost
->dlr
;
2529 if (GET_MODE (x
) == DImode
)
2531 rtx right
= XEXP (x
, 1);
2532 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2534 *total
= s390_cost
->dsgfr
;
2536 *total
= s390_cost
->dr
;
2537 else /* 64 by 64 bit division */
2538 *total
= s390_cost
->dsgr
;
2540 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2541 *total
= s390_cost
->dlr
;
2542 else if (GET_MODE (x
) == SFmode
)
2544 *total
= s390_cost
->debr
;
2546 else if (GET_MODE (x
) == DFmode
)
2548 *total
= s390_cost
->ddbr
;
2550 else if (GET_MODE (x
) == TFmode
)
2552 *total
= s390_cost
->dxbr
;
2557 if (GET_MODE (x
) == SFmode
)
2558 *total
= s390_cost
->sqebr
;
2559 else if (GET_MODE (x
) == DFmode
)
2560 *total
= s390_cost
->sqdbr
;
2562 *total
= s390_cost
->sqxbr
;
2567 if (outer_code
== MULT
|| outer_code
== DIV
|| outer_code
== MOD
2568 || outer_code
== PLUS
|| outer_code
== MINUS
2569 || outer_code
== COMPARE
)
2574 *total
= COSTS_N_INSNS (1);
2575 if (GET_CODE (XEXP (x
, 0)) == AND
2576 && GET_CODE (XEXP (x
, 1)) == CONST_INT
2577 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
)
2579 rtx op0
= XEXP (XEXP (x
, 0), 0);
2580 rtx op1
= XEXP (XEXP (x
, 0), 1);
2581 rtx op2
= XEXP (x
, 1);
2583 if (memory_operand (op0
, GET_MODE (op0
))
2584 && s390_tm_ccmode (op1
, op2
, 0) != VOIDmode
)
2586 if (register_operand (op0
, GET_MODE (op0
))
2587 && s390_tm_ccmode (op1
, op2
, 1) != VOIDmode
)
2597 /* Return the cost of an address rtx ADDR. */
2600 s390_address_cost (rtx addr
, bool speed ATTRIBUTE_UNUSED
)
2602 struct s390_address ad
;
2603 if (!s390_decompose_address (addr
, &ad
))
2606 return ad
.indx
? COSTS_N_INSNS (1) + 1 : COSTS_N_INSNS (1);
2609 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
2610 otherwise return 0. */
2613 tls_symbolic_operand (rtx op
)
2615 if (GET_CODE (op
) != SYMBOL_REF
)
2617 return SYMBOL_REF_TLS_MODEL (op
);
2620 /* Split DImode access register reference REG (on 64-bit) into its constituent
2621 low and high parts, and store them into LO and HI. Note that gen_lowpart/
2622 gen_highpart cannot be used as they assume all registers are word-sized,
2623 while our access registers have only half that size. */
2626 s390_split_access_reg (rtx reg
, rtx
*lo
, rtx
*hi
)
2628 gcc_assert (TARGET_64BIT
);
2629 gcc_assert (ACCESS_REG_P (reg
));
2630 gcc_assert (GET_MODE (reg
) == DImode
);
2631 gcc_assert (!(REGNO (reg
) & 1));
2633 *lo
= gen_rtx_REG (SImode
, REGNO (reg
) + 1);
2634 *hi
= gen_rtx_REG (SImode
, REGNO (reg
));
2637 /* Return true if OP contains a symbol reference */
2640 symbolic_reference_mentioned_p (rtx op
)
2645 if (GET_CODE (op
) == SYMBOL_REF
|| GET_CODE (op
) == LABEL_REF
)
2648 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2649 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2655 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2656 if (symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2660 else if (fmt
[i
] == 'e' && symbolic_reference_mentioned_p (XEXP (op
, i
)))
2667 /* Return true if OP contains a reference to a thread-local symbol. */
2670 tls_symbolic_reference_mentioned_p (rtx op
)
2675 if (GET_CODE (op
) == SYMBOL_REF
)
2676 return tls_symbolic_operand (op
);
2678 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2679 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2685 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2686 if (tls_symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2690 else if (fmt
[i
] == 'e' && tls_symbolic_reference_mentioned_p (XEXP (op
, i
)))
2698 /* Return true if OP is a legitimate general operand when
2699 generating PIC code. It is given that flag_pic is on
2700 and that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2703 legitimate_pic_operand_p (rtx op
)
2705 /* Accept all non-symbolic constants. */
2706 if (!SYMBOLIC_CONST (op
))
2709 /* Reject everything else; must be handled
2710 via emit_symbolic_move. */
2714 /* Returns true if the constant value OP is a legitimate general operand.
2715 It is given that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2718 s390_legitimate_constant_p (enum machine_mode mode
, rtx op
)
2720 /* Accept all non-symbolic constants. */
2721 if (!SYMBOLIC_CONST (op
))
2724 /* Accept immediate LARL operands. */
2725 if (TARGET_CPU_ZARCH
&& larl_operand (op
, mode
))
2728 /* Thread-local symbols are never legal constants. This is
2729 so that emit_call knows that computing such addresses
2730 might require a function call. */
2731 if (TLS_SYMBOLIC_CONST (op
))
2734 /* In the PIC case, symbolic constants must *not* be
2735 forced into the literal pool. We accept them here,
2736 so that they will be handled by emit_symbolic_move. */
2740 /* All remaining non-PIC symbolic constants are
2741 forced into the literal pool. */
2745 /* Determine if it's legal to put X into the constant pool. This
2746 is not possible if X contains the address of a symbol that is
2747 not constant (TLS) or not known at final link time (PIC). */
2750 s390_cannot_force_const_mem (enum machine_mode mode
, rtx x
)
2752 switch (GET_CODE (x
))
2756 /* Accept all non-symbolic constants. */
2760 /* Labels are OK iff we are non-PIC. */
2761 return flag_pic
!= 0;
2764 /* 'Naked' TLS symbol references are never OK,
2765 non-TLS symbols are OK iff we are non-PIC. */
2766 if (tls_symbolic_operand (x
))
2769 return flag_pic
!= 0;
2772 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0));
2775 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0))
2776 || s390_cannot_force_const_mem (mode
, XEXP (x
, 1));
2779 switch (XINT (x
, 1))
2781 /* Only lt-relative or GOT-relative UNSPECs are OK. */
2782 case UNSPEC_LTREL_OFFSET
:
2790 case UNSPEC_GOTNTPOFF
:
2791 case UNSPEC_INDNTPOFF
:
2794 /* If the literal pool shares the code section, be put
2795 execute template placeholders into the pool as well. */
2797 return TARGET_CPU_ZARCH
;
2809 /* Returns true if the constant value OP is a legitimate general
2810 operand during and after reload. The difference to
2811 legitimate_constant_p is that this function will not accept
2812 a constant that would need to be forced to the literal pool
2813 before it can be used as operand.
2814 This function accepts all constants which can be loaded directly
2818 legitimate_reload_constant_p (rtx op
)
2820 /* Accept la(y) operands. */
2821 if (GET_CODE (op
) == CONST_INT
2822 && DISP_IN_RANGE (INTVAL (op
)))
2825 /* Accept l(g)hi/l(g)fi operands. */
2826 if (GET_CODE (op
) == CONST_INT
2827 && (CONST_OK_FOR_K (INTVAL (op
)) || CONST_OK_FOR_Os (INTVAL (op
))))
2830 /* Accept lliXX operands. */
2832 && GET_CODE (op
) == CONST_INT
2833 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2834 && s390_single_part (op
, word_mode
, HImode
, 0) >= 0)
2838 && GET_CODE (op
) == CONST_INT
2839 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2840 && s390_single_part (op
, word_mode
, SImode
, 0) >= 0)
2843 /* Accept larl operands. */
2844 if (TARGET_CPU_ZARCH
2845 && larl_operand (op
, VOIDmode
))
2848 /* Accept floating-point zero operands that fit into a single GPR. */
2849 if (GET_CODE (op
) == CONST_DOUBLE
2850 && s390_float_const_zero_p (op
)
2851 && GET_MODE_SIZE (GET_MODE (op
)) <= UNITS_PER_WORD
)
2854 /* Accept double-word operands that can be split. */
2855 if (GET_CODE (op
) == CONST_INT
2856 && trunc_int_for_mode (INTVAL (op
), word_mode
) != INTVAL (op
))
2858 enum machine_mode dword_mode
= word_mode
== SImode
? DImode
: TImode
;
2859 rtx hi
= operand_subword (op
, 0, 0, dword_mode
);
2860 rtx lo
= operand_subword (op
, 1, 0, dword_mode
);
2861 return legitimate_reload_constant_p (hi
)
2862 && legitimate_reload_constant_p (lo
);
2865 /* Everything else cannot be handled without reload. */
2869 /* Returns true if the constant value OP is a legitimate fp operand
2870 during and after reload.
2871 This function accepts all constants which can be loaded directly
2875 legitimate_reload_fp_constant_p (rtx op
)
2877 /* Accept floating-point zero operands if the load zero instruction
2880 && GET_CODE (op
) == CONST_DOUBLE
2881 && s390_float_const_zero_p (op
))
2887 /* Given an rtx OP being reloaded into a reg required to be in class RCLASS,
2888 return the class of reg to actually use. */
2891 s390_preferred_reload_class (rtx op
, reg_class_t rclass
)
2893 switch (GET_CODE (op
))
2895 /* Constants we cannot reload into general registers
2896 must be forced into the literal pool. */
2899 if (reg_class_subset_p (GENERAL_REGS
, rclass
)
2900 && legitimate_reload_constant_p (op
))
2901 return GENERAL_REGS
;
2902 else if (reg_class_subset_p (ADDR_REGS
, rclass
)
2903 && legitimate_reload_constant_p (op
))
2905 else if (reg_class_subset_p (FP_REGS
, rclass
)
2906 && legitimate_reload_fp_constant_p (op
))
2910 /* If a symbolic constant or a PLUS is reloaded,
2911 it is most likely being used as an address, so
2912 prefer ADDR_REGS. If 'class' is not a superset
2913 of ADDR_REGS, e.g. FP_REGS, reject this reload. */
2917 if (!legitimate_reload_constant_p (op
))
2921 /* load address will be used. */
2922 if (reg_class_subset_p (ADDR_REGS
, rclass
))
2934 /* Return true if ADDR is SYMBOL_REF + addend with addend being a
2935 multiple of ALIGNMENT and the SYMBOL_REF being naturally
2939 s390_check_symref_alignment (rtx addr
, HOST_WIDE_INT alignment
)
2941 HOST_WIDE_INT addend
;
2944 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2947 return (!SYMBOL_REF_NOT_NATURALLY_ALIGNED_P (symref
)
2948 && !(addend
& (alignment
- 1)));
2951 /* ADDR is moved into REG using larl. If ADDR isn't a valid larl
2952 operand SCRATCH is used to reload the even part of the address and
2956 s390_reload_larl_operand (rtx reg
, rtx addr
, rtx scratch
)
2958 HOST_WIDE_INT addend
;
2961 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2965 /* Easy case. The addend is even so larl will do fine. */
2966 emit_move_insn (reg
, addr
);
2969 /* We can leave the scratch register untouched if the target
2970 register is a valid base register. */
2971 if (REGNO (reg
) < FIRST_PSEUDO_REGISTER
2972 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
)
2975 gcc_assert (REGNO (scratch
) < FIRST_PSEUDO_REGISTER
);
2976 gcc_assert (REGNO_REG_CLASS (REGNO (scratch
)) == ADDR_REGS
);
2979 emit_move_insn (scratch
,
2980 gen_rtx_CONST (Pmode
,
2981 gen_rtx_PLUS (Pmode
, symref
,
2982 GEN_INT (addend
- 1))));
2984 emit_move_insn (scratch
, symref
);
2986 /* Increment the address using la in order to avoid clobbering cc. */
2987 emit_move_insn (reg
, gen_rtx_PLUS (Pmode
, scratch
, const1_rtx
));
2991 /* Generate what is necessary to move between REG and MEM using
2992 SCRATCH. The direction is given by TOMEM. */
2995 s390_reload_symref_address (rtx reg
, rtx mem
, rtx scratch
, bool tomem
)
2997 /* Reload might have pulled a constant out of the literal pool.
2998 Force it back in. */
2999 if (CONST_INT_P (mem
) || GET_CODE (mem
) == CONST_DOUBLE
3000 || GET_CODE (mem
) == CONST
)
3001 mem
= force_const_mem (GET_MODE (reg
), mem
);
3003 gcc_assert (MEM_P (mem
));
3005 /* For a load from memory we can leave the scratch register
3006 untouched if the target register is a valid base register. */
3008 && REGNO (reg
) < FIRST_PSEUDO_REGISTER
3009 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
3010 && GET_MODE (reg
) == GET_MODE (scratch
))
3013 /* Load address into scratch register. Since we can't have a
3014 secondary reload for a secondary reload we have to cover the case
3015 where larl would need a secondary reload here as well. */
3016 s390_reload_larl_operand (scratch
, XEXP (mem
, 0), scratch
);
3018 /* Now we can use a standard load/store to do the move. */
3020 emit_move_insn (replace_equiv_address (mem
, scratch
), reg
);
3022 emit_move_insn (reg
, replace_equiv_address (mem
, scratch
));
3025 /* Inform reload about cases where moving X with a mode MODE to a register in
3026 RCLASS requires an extra scratch or immediate register. Return the class
3027 needed for the immediate register. */
3030 s390_secondary_reload (bool in_p
, rtx x
, reg_class_t rclass_i
,
3031 enum machine_mode mode
, secondary_reload_info
*sri
)
3033 enum reg_class rclass
= (enum reg_class
) rclass_i
;
3035 /* Intermediate register needed. */
3036 if (reg_classes_intersect_p (CC_REGS
, rclass
))
3037 return GENERAL_REGS
;
3041 HOST_WIDE_INT offset
;
3044 /* On z10 several optimizer steps may generate larl operands with
3047 && s390_symref_operand_p (x
, &symref
, &offset
)
3049 && !SYMBOL_REF_ALIGN1_P (symref
)
3050 && (offset
& 1) == 1)
3051 sri
->icode
= ((mode
== DImode
) ? CODE_FOR_reloaddi_larl_odd_addend_z10
3052 : CODE_FOR_reloadsi_larl_odd_addend_z10
);
3054 /* On z10 we need a scratch register when moving QI, TI or floating
3055 point mode values from or to a memory location with a SYMBOL_REF
3056 or if the symref addend of a SI or DI move is not aligned to the
3057 width of the access. */
3059 && s390_symref_operand_p (XEXP (x
, 0), NULL
, NULL
)
3060 && (mode
== QImode
|| mode
== TImode
|| FLOAT_MODE_P (mode
)
3061 || (!TARGET_ZARCH
&& mode
== DImode
)
3062 || ((mode
== HImode
|| mode
== SImode
|| mode
== DImode
)
3063 && (!s390_check_symref_alignment (XEXP (x
, 0),
3064 GET_MODE_SIZE (mode
))))))
3066 #define __SECONDARY_RELOAD_CASE(M,m) \
3069 sri->icode = in_p ? CODE_FOR_reload##m##di_toreg_z10 : \
3070 CODE_FOR_reload##m##di_tomem_z10; \
3072 sri->icode = in_p ? CODE_FOR_reload##m##si_toreg_z10 : \
3073 CODE_FOR_reload##m##si_tomem_z10; \
3076 switch (GET_MODE (x
))
3078 __SECONDARY_RELOAD_CASE (QI
, qi
);
3079 __SECONDARY_RELOAD_CASE (HI
, hi
);
3080 __SECONDARY_RELOAD_CASE (SI
, si
);
3081 __SECONDARY_RELOAD_CASE (DI
, di
);
3082 __SECONDARY_RELOAD_CASE (TI
, ti
);
3083 __SECONDARY_RELOAD_CASE (SF
, sf
);
3084 __SECONDARY_RELOAD_CASE (DF
, df
);
3085 __SECONDARY_RELOAD_CASE (TF
, tf
);
3086 __SECONDARY_RELOAD_CASE (SD
, sd
);
3087 __SECONDARY_RELOAD_CASE (DD
, dd
);
3088 __SECONDARY_RELOAD_CASE (TD
, td
);
3093 #undef __SECONDARY_RELOAD_CASE
3097 /* We need a scratch register when loading a PLUS expression which
3098 is not a legitimate operand of the LOAD ADDRESS instruction. */
3099 if (in_p
&& s390_plus_operand (x
, mode
))
3100 sri
->icode
= (TARGET_64BIT
?
3101 CODE_FOR_reloaddi_plus
: CODE_FOR_reloadsi_plus
);
3103 /* Performing a multiword move from or to memory we have to make sure the
3104 second chunk in memory is addressable without causing a displacement
3105 overflow. If that would be the case we calculate the address in
3106 a scratch register. */
3108 && GET_CODE (XEXP (x
, 0)) == PLUS
3109 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
3110 && !DISP_IN_RANGE (INTVAL (XEXP (XEXP (x
, 0), 1))
3111 + GET_MODE_SIZE (mode
) - 1))
3113 /* For GENERAL_REGS a displacement overflow is no problem if occurring
3114 in a s_operand address since we may fallback to lm/stm. So we only
3115 have to care about overflows in the b+i+d case. */
3116 if ((reg_classes_intersect_p (GENERAL_REGS
, rclass
)
3117 && s390_class_max_nregs (GENERAL_REGS
, mode
) > 1
3118 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == PLUS
)
3119 /* For FP_REGS no lm/stm is available so this check is triggered
3120 for displacement overflows in b+i+d and b+d like addresses. */
3121 || (reg_classes_intersect_p (FP_REGS
, rclass
)
3122 && s390_class_max_nregs (FP_REGS
, mode
) > 1))
3125 sri
->icode
= (TARGET_64BIT
?
3126 CODE_FOR_reloaddi_nonoffmem_in
:
3127 CODE_FOR_reloadsi_nonoffmem_in
);
3129 sri
->icode
= (TARGET_64BIT
?
3130 CODE_FOR_reloaddi_nonoffmem_out
:
3131 CODE_FOR_reloadsi_nonoffmem_out
);
3135 /* A scratch address register is needed when a symbolic constant is
3136 copied to r0 compiling with -fPIC. In other cases the target
3137 register might be used as temporary (see legitimize_pic_address). */
3138 if (in_p
&& SYMBOLIC_CONST (x
) && flag_pic
== 2 && rclass
!= ADDR_REGS
)
3139 sri
->icode
= (TARGET_64BIT
?
3140 CODE_FOR_reloaddi_PIC_addr
:
3141 CODE_FOR_reloadsi_PIC_addr
);
3143 /* Either scratch or no register needed. */
3147 /* Generate code to load SRC, which is PLUS that is not a
3148 legitimate operand for the LA instruction, into TARGET.
3149 SCRATCH may be used as scratch register. */
3152 s390_expand_plus_operand (rtx target
, rtx src
,
3156 struct s390_address ad
;
3158 /* src must be a PLUS; get its two operands. */
3159 gcc_assert (GET_CODE (src
) == PLUS
);
3160 gcc_assert (GET_MODE (src
) == Pmode
);
3162 /* Check if any of the two operands is already scheduled
3163 for replacement by reload. This can happen e.g. when
3164 float registers occur in an address. */
3165 sum1
= find_replacement (&XEXP (src
, 0));
3166 sum2
= find_replacement (&XEXP (src
, 1));
3167 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3169 /* If the address is already strictly valid, there's nothing to do. */
3170 if (!s390_decompose_address (src
, &ad
)
3171 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3172 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
3174 /* Otherwise, one of the operands cannot be an address register;
3175 we reload its value into the scratch register. */
3176 if (true_regnum (sum1
) < 1 || true_regnum (sum1
) > 15)
3178 emit_move_insn (scratch
, sum1
);
3181 if (true_regnum (sum2
) < 1 || true_regnum (sum2
) > 15)
3183 emit_move_insn (scratch
, sum2
);
3187 /* According to the way these invalid addresses are generated
3188 in reload.c, it should never happen (at least on s390) that
3189 *neither* of the PLUS components, after find_replacements
3190 was applied, is an address register. */
3191 if (sum1
== scratch
&& sum2
== scratch
)
3197 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3200 /* Emit the LOAD ADDRESS pattern. Note that reload of PLUS
3201 is only ever performed on addresses, so we can mark the
3202 sum as legitimate for LA in any case. */
3203 s390_load_address (target
, src
);
3207 /* Return true if ADDR is a valid memory address.
3208 STRICT specifies whether strict register checking applies. */
3211 s390_legitimate_address_p (enum machine_mode mode
, rtx addr
, bool strict
)
3213 struct s390_address ad
;
3216 && larl_operand (addr
, VOIDmode
)
3217 && (mode
== VOIDmode
3218 || s390_check_symref_alignment (addr
, GET_MODE_SIZE (mode
))))
3221 if (!s390_decompose_address (addr
, &ad
))
3226 if (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3229 if (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
)))
3235 && !(REGNO (ad
.base
) >= FIRST_PSEUDO_REGISTER
3236 || REGNO_REG_CLASS (REGNO (ad
.base
)) == ADDR_REGS
))
3240 && !(REGNO (ad
.indx
) >= FIRST_PSEUDO_REGISTER
3241 || REGNO_REG_CLASS (REGNO (ad
.indx
)) == ADDR_REGS
))
3247 /* Return true if OP is a valid operand for the LA instruction.
3248 In 31-bit, we need to prove that the result is used as an
3249 address, as LA performs only a 31-bit addition. */
3252 legitimate_la_operand_p (rtx op
)
3254 struct s390_address addr
;
3255 if (!s390_decompose_address (op
, &addr
))
3258 return (TARGET_64BIT
|| addr
.pointer
);
3261 /* Return true if it is valid *and* preferable to use LA to
3262 compute the sum of OP1 and OP2. */
3265 preferred_la_operand_p (rtx op1
, rtx op2
)
3267 struct s390_address addr
;
3269 if (op2
!= const0_rtx
)
3270 op1
= gen_rtx_PLUS (Pmode
, op1
, op2
);
3272 if (!s390_decompose_address (op1
, &addr
))
3274 if (addr
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (addr
.base
)))
3276 if (addr
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (addr
.indx
)))
3279 /* Avoid LA instructions with index register on z196; it is
3280 preferable to use regular add instructions when possible. */
3281 if (addr
.indx
&& s390_tune
== PROCESSOR_2817_Z196
)
3284 if (!TARGET_64BIT
&& !addr
.pointer
)
3290 if ((addr
.base
&& REG_P (addr
.base
) && REG_POINTER (addr
.base
))
3291 || (addr
.indx
&& REG_P (addr
.indx
) && REG_POINTER (addr
.indx
)))
3297 /* Emit a forced load-address operation to load SRC into DST.
3298 This will use the LOAD ADDRESS instruction even in situations
3299 where legitimate_la_operand_p (SRC) returns false. */
3302 s390_load_address (rtx dst
, rtx src
)
3305 emit_move_insn (dst
, src
);
3307 emit_insn (gen_force_la_31 (dst
, src
));
3310 /* Return a legitimate reference for ORIG (an address) using the
3311 register REG. If REG is 0, a new pseudo is generated.
3313 There are two types of references that must be handled:
3315 1. Global data references must load the address from the GOT, via
3316 the PIC reg. An insn is emitted to do this load, and the reg is
3319 2. Static data references, constant pool addresses, and code labels
3320 compute the address as an offset from the GOT, whose base is in
3321 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
3322 differentiate them from global data objects. The returned
3323 address is the PIC reg + an unspec constant.
3325 TARGET_LEGITIMIZE_ADDRESS_P rejects symbolic references unless the PIC
3326 reg also appears in the address. */
3329 legitimize_pic_address (rtx orig
, rtx reg
)
3335 gcc_assert (!TLS_SYMBOLIC_CONST (addr
));
3337 if (GET_CODE (addr
) == LABEL_REF
3338 || (GET_CODE (addr
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (addr
)))
3340 /* This is a local symbol. */
3341 if (TARGET_CPU_ZARCH
&& larl_operand (addr
, VOIDmode
))
3343 /* Access local symbols PC-relative via LARL.
3344 This is the same as in the non-PIC case, so it is
3345 handled automatically ... */
3349 /* Access local symbols relative to the GOT. */
3351 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3353 if (reload_in_progress
|| reload_completed
)
3354 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3356 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
3357 addr
= gen_rtx_CONST (Pmode
, addr
);
3358 addr
= force_const_mem (Pmode
, addr
);
3359 emit_move_insn (temp
, addr
);
3361 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3364 s390_load_address (reg
, new_rtx
);
3369 else if (GET_CODE (addr
) == SYMBOL_REF
)
3372 reg
= gen_reg_rtx (Pmode
);
3376 /* Assume GOT offset < 4k. This is handled the same way
3377 in both 31- and 64-bit code (@GOT). */
3379 if (reload_in_progress
|| reload_completed
)
3380 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3382 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3383 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3384 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3385 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3386 emit_move_insn (reg
, new_rtx
);
3389 else if (TARGET_CPU_ZARCH
)
3391 /* If the GOT offset might be >= 4k, we determine the position
3392 of the GOT entry via a PC-relative LARL (@GOTENT). */
3394 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3396 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3397 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3399 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTENT
);
3400 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3401 emit_move_insn (temp
, new_rtx
);
3403 new_rtx
= gen_const_mem (Pmode
, temp
);
3404 emit_move_insn (reg
, new_rtx
);
3409 /* If the GOT offset might be >= 4k, we have to load it
3410 from the literal pool (@GOT). */
3412 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3414 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3415 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3417 if (reload_in_progress
|| reload_completed
)
3418 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3420 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3421 addr
= gen_rtx_CONST (Pmode
, addr
);
3422 addr
= force_const_mem (Pmode
, addr
);
3423 emit_move_insn (temp
, addr
);
3425 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3426 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3427 emit_move_insn (reg
, new_rtx
);
3433 if (GET_CODE (addr
) == CONST
)
3435 addr
= XEXP (addr
, 0);
3436 if (GET_CODE (addr
) == UNSPEC
)
3438 gcc_assert (XVECLEN (addr
, 0) == 1);
3439 switch (XINT (addr
, 1))
3441 /* If someone moved a GOT-relative UNSPEC
3442 out of the literal pool, force them back in. */
3445 new_rtx
= force_const_mem (Pmode
, orig
);
3448 /* @GOT is OK as is if small. */
3451 new_rtx
= force_const_mem (Pmode
, orig
);
3454 /* @GOTENT is OK as is. */
3458 /* @PLT is OK as is on 64-bit, must be converted to
3459 GOT-relative @PLTOFF on 31-bit. */
3461 if (!TARGET_CPU_ZARCH
)
3463 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3465 if (reload_in_progress
|| reload_completed
)
3466 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3468 addr
= XVECEXP (addr
, 0, 0);
3469 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
),
3471 addr
= gen_rtx_CONST (Pmode
, addr
);
3472 addr
= force_const_mem (Pmode
, addr
);
3473 emit_move_insn (temp
, addr
);
3475 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3478 s390_load_address (reg
, new_rtx
);
3484 /* Everything else cannot happen. */
3490 gcc_assert (GET_CODE (addr
) == PLUS
);
3492 if (GET_CODE (addr
) == PLUS
)
3494 rtx op0
= XEXP (addr
, 0), op1
= XEXP (addr
, 1);
3496 gcc_assert (!TLS_SYMBOLIC_CONST (op0
));
3497 gcc_assert (!TLS_SYMBOLIC_CONST (op1
));
3499 /* Check first to see if this is a constant offset
3500 from a local symbol reference. */
3501 if ((GET_CODE (op0
) == LABEL_REF
3502 || (GET_CODE (op0
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (op0
)))
3503 && GET_CODE (op1
) == CONST_INT
)
3505 if (TARGET_CPU_ZARCH
3506 && larl_operand (op0
, VOIDmode
)
3507 && INTVAL (op1
) < (HOST_WIDE_INT
)1 << 31
3508 && INTVAL (op1
) >= -((HOST_WIDE_INT
)1 << 31))
3510 if (INTVAL (op1
) & 1)
3512 /* LARL can't handle odd offsets, so emit a
3513 pair of LARL and LA. */
3514 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3516 if (!DISP_IN_RANGE (INTVAL (op1
)))
3518 HOST_WIDE_INT even
= INTVAL (op1
) - 1;
3519 op0
= gen_rtx_PLUS (Pmode
, op0
, GEN_INT (even
));
3520 op0
= gen_rtx_CONST (Pmode
, op0
);
3524 emit_move_insn (temp
, op0
);
3525 new_rtx
= gen_rtx_PLUS (Pmode
, temp
, op1
);
3529 s390_load_address (reg
, new_rtx
);
3535 /* If the offset is even, we can just use LARL.
3536 This will happen automatically. */
3541 /* Access local symbols relative to the GOT. */
3543 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3545 if (reload_in_progress
|| reload_completed
)
3546 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3548 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op0
),
3550 addr
= gen_rtx_PLUS (Pmode
, addr
, op1
);
3551 addr
= gen_rtx_CONST (Pmode
, addr
);
3552 addr
= force_const_mem (Pmode
, addr
);
3553 emit_move_insn (temp
, addr
);
3555 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3558 s390_load_address (reg
, new_rtx
);
3564 /* Now, check whether it is a GOT relative symbol plus offset
3565 that was pulled out of the literal pool. Force it back in. */
3567 else if (GET_CODE (op0
) == UNSPEC
3568 && GET_CODE (op1
) == CONST_INT
3569 && XINT (op0
, 1) == UNSPEC_GOTOFF
)
3571 gcc_assert (XVECLEN (op0
, 0) == 1);
3573 new_rtx
= force_const_mem (Pmode
, orig
);
3576 /* Otherwise, compute the sum. */
3579 base
= legitimize_pic_address (XEXP (addr
, 0), reg
);
3580 new_rtx
= legitimize_pic_address (XEXP (addr
, 1),
3581 base
== reg
? NULL_RTX
: reg
);
3582 if (GET_CODE (new_rtx
) == CONST_INT
)
3583 new_rtx
= plus_constant (Pmode
, base
, INTVAL (new_rtx
));
3586 if (GET_CODE (new_rtx
) == PLUS
&& CONSTANT_P (XEXP (new_rtx
, 1)))
3588 base
= gen_rtx_PLUS (Pmode
, base
, XEXP (new_rtx
, 0));
3589 new_rtx
= XEXP (new_rtx
, 1);
3591 new_rtx
= gen_rtx_PLUS (Pmode
, base
, new_rtx
);
3594 if (GET_CODE (new_rtx
) == CONST
)
3595 new_rtx
= XEXP (new_rtx
, 0);
3596 new_rtx
= force_operand (new_rtx
, 0);
3603 /* Load the thread pointer into a register. */
3606 s390_get_thread_pointer (void)
3608 rtx tp
= gen_reg_rtx (Pmode
);
3610 emit_move_insn (tp
, gen_rtx_REG (Pmode
, TP_REGNUM
));
3611 mark_reg_pointer (tp
, BITS_PER_WORD
);
3616 /* Emit a tls call insn. The call target is the SYMBOL_REF stored
3617 in s390_tls_symbol which always refers to __tls_get_offset.
3618 The returned offset is written to RESULT_REG and an USE rtx is
3619 generated for TLS_CALL. */
3621 static GTY(()) rtx s390_tls_symbol
;
3624 s390_emit_tls_call_insn (rtx result_reg
, rtx tls_call
)
3629 emit_insn (s390_load_got ());
3631 if (!s390_tls_symbol
)
3632 s390_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tls_get_offset");
3634 insn
= s390_emit_call (s390_tls_symbol
, tls_call
, result_reg
,
3635 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
3637 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), result_reg
);
3638 RTL_CONST_CALL_P (insn
) = 1;
3641 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3642 this (thread-local) address. REG may be used as temporary. */
3645 legitimize_tls_address (rtx addr
, rtx reg
)
3647 rtx new_rtx
, tls_call
, temp
, base
, r2
, insn
;
3649 if (GET_CODE (addr
) == SYMBOL_REF
)
3650 switch (tls_symbolic_operand (addr
))
3652 case TLS_MODEL_GLOBAL_DYNAMIC
:
3654 r2
= gen_rtx_REG (Pmode
, 2);
3655 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_TLSGD
);
3656 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3657 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3658 emit_move_insn (r2
, new_rtx
);
3659 s390_emit_tls_call_insn (r2
, tls_call
);
3660 insn
= get_insns ();
3663 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3664 temp
= gen_reg_rtx (Pmode
);
3665 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3667 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3670 s390_load_address (reg
, new_rtx
);
3675 case TLS_MODEL_LOCAL_DYNAMIC
:
3677 r2
= gen_rtx_REG (Pmode
, 2);
3678 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM
);
3679 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3680 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3681 emit_move_insn (r2
, new_rtx
);
3682 s390_emit_tls_call_insn (r2
, tls_call
);
3683 insn
= get_insns ();
3686 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM_NTPOFF
);
3687 temp
= gen_reg_rtx (Pmode
);
3688 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3690 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3691 base
= gen_reg_rtx (Pmode
);
3692 s390_load_address (base
, new_rtx
);
3694 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_DTPOFF
);
3695 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3696 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3697 temp
= gen_reg_rtx (Pmode
);
3698 emit_move_insn (temp
, new_rtx
);
3700 new_rtx
= gen_rtx_PLUS (Pmode
, base
, temp
);
3703 s390_load_address (reg
, new_rtx
);
3708 case TLS_MODEL_INITIAL_EXEC
:
3711 /* Assume GOT offset < 4k. This is handled the same way
3712 in both 31- and 64-bit code. */
3714 if (reload_in_progress
|| reload_completed
)
3715 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3717 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3718 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3719 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3720 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3721 temp
= gen_reg_rtx (Pmode
);
3722 emit_move_insn (temp
, new_rtx
);
3724 else if (TARGET_CPU_ZARCH
)
3726 /* If the GOT offset might be >= 4k, we determine the position
3727 of the GOT entry via a PC-relative LARL. */
3729 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3730 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3731 temp
= gen_reg_rtx (Pmode
);
3732 emit_move_insn (temp
, new_rtx
);
3734 new_rtx
= gen_const_mem (Pmode
, temp
);
3735 temp
= gen_reg_rtx (Pmode
);
3736 emit_move_insn (temp
, new_rtx
);
3740 /* If the GOT offset might be >= 4k, we have to load it
3741 from the literal pool. */
3743 if (reload_in_progress
|| reload_completed
)
3744 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3746 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3747 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3748 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3749 temp
= gen_reg_rtx (Pmode
);
3750 emit_move_insn (temp
, new_rtx
);
3752 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3753 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3755 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3756 temp
= gen_reg_rtx (Pmode
);
3757 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3761 /* In position-dependent code, load the absolute address of
3762 the GOT entry from the literal pool. */
3764 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3765 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3766 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3767 temp
= gen_reg_rtx (Pmode
);
3768 emit_move_insn (temp
, new_rtx
);
3771 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3772 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3773 temp
= gen_reg_rtx (Pmode
);
3774 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3777 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3780 s390_load_address (reg
, new_rtx
);
3785 case TLS_MODEL_LOCAL_EXEC
:
3786 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3787 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3788 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3789 temp
= gen_reg_rtx (Pmode
);
3790 emit_move_insn (temp
, new_rtx
);
3792 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3795 s390_load_address (reg
, new_rtx
);
3804 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == UNSPEC
)
3806 switch (XINT (XEXP (addr
, 0), 1))
3808 case UNSPEC_INDNTPOFF
:
3809 gcc_assert (TARGET_CPU_ZARCH
);
3818 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
3819 && GET_CODE (XEXP (XEXP (addr
, 0), 1)) == CONST_INT
)
3821 new_rtx
= XEXP (XEXP (addr
, 0), 0);
3822 if (GET_CODE (new_rtx
) != SYMBOL_REF
)
3823 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3825 new_rtx
= legitimize_tls_address (new_rtx
, reg
);
3826 new_rtx
= plus_constant (Pmode
, new_rtx
,
3827 INTVAL (XEXP (XEXP (addr
, 0), 1)));
3828 new_rtx
= force_operand (new_rtx
, 0);
3832 gcc_unreachable (); /* for now ... */
3837 /* Emit insns making the address in operands[1] valid for a standard
3838 move to operands[0]. operands[1] is replaced by an address which
3839 should be used instead of the former RTX to emit the move
3843 emit_symbolic_move (rtx
*operands
)
3845 rtx temp
= !can_create_pseudo_p () ? operands
[0] : gen_reg_rtx (Pmode
);
3847 if (GET_CODE (operands
[0]) == MEM
)
3848 operands
[1] = force_reg (Pmode
, operands
[1]);
3849 else if (TLS_SYMBOLIC_CONST (operands
[1]))
3850 operands
[1] = legitimize_tls_address (operands
[1], temp
);
3852 operands
[1] = legitimize_pic_address (operands
[1], temp
);
3855 /* Try machine-dependent ways of modifying an illegitimate address X
3856 to be legitimate. If we find one, return the new, valid address.
3858 OLDX is the address as it was before break_out_memory_refs was called.
3859 In some cases it is useful to look at this to decide what needs to be done.
3861 MODE is the mode of the operand pointed to by X.
3863 When -fpic is used, special handling is needed for symbolic references.
3864 See comments by legitimize_pic_address for details. */
3867 s390_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
3868 enum machine_mode mode ATTRIBUTE_UNUSED
)
3870 rtx constant_term
= const0_rtx
;
3872 if (TLS_SYMBOLIC_CONST (x
))
3874 x
= legitimize_tls_address (x
, 0);
3876 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3879 else if (GET_CODE (x
) == PLUS
3880 && (TLS_SYMBOLIC_CONST (XEXP (x
, 0))
3881 || TLS_SYMBOLIC_CONST (XEXP (x
, 1))))
3887 if (SYMBOLIC_CONST (x
)
3888 || (GET_CODE (x
) == PLUS
3889 && (SYMBOLIC_CONST (XEXP (x
, 0))
3890 || SYMBOLIC_CONST (XEXP (x
, 1)))))
3891 x
= legitimize_pic_address (x
, 0);
3893 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3897 x
= eliminate_constant_term (x
, &constant_term
);
3899 /* Optimize loading of large displacements by splitting them
3900 into the multiple of 4K and the rest; this allows the
3901 former to be CSE'd if possible.
3903 Don't do this if the displacement is added to a register
3904 pointing into the stack frame, as the offsets will
3905 change later anyway. */
3907 if (GET_CODE (constant_term
) == CONST_INT
3908 && !TARGET_LONG_DISPLACEMENT
3909 && !DISP_IN_RANGE (INTVAL (constant_term
))
3910 && !(REG_P (x
) && REGNO_PTR_FRAME_P (REGNO (x
))))
3912 HOST_WIDE_INT lower
= INTVAL (constant_term
) & 0xfff;
3913 HOST_WIDE_INT upper
= INTVAL (constant_term
) ^ lower
;
3915 rtx temp
= gen_reg_rtx (Pmode
);
3916 rtx val
= force_operand (GEN_INT (upper
), temp
);
3918 emit_move_insn (temp
, val
);
3920 x
= gen_rtx_PLUS (Pmode
, x
, temp
);
3921 constant_term
= GEN_INT (lower
);
3924 if (GET_CODE (x
) == PLUS
)
3926 if (GET_CODE (XEXP (x
, 0)) == REG
)
3928 rtx temp
= gen_reg_rtx (Pmode
);
3929 rtx val
= force_operand (XEXP (x
, 1), temp
);
3931 emit_move_insn (temp
, val
);
3933 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0), temp
);
3936 else if (GET_CODE (XEXP (x
, 1)) == REG
)
3938 rtx temp
= gen_reg_rtx (Pmode
);
3939 rtx val
= force_operand (XEXP (x
, 0), temp
);
3941 emit_move_insn (temp
, val
);
3943 x
= gen_rtx_PLUS (Pmode
, temp
, XEXP (x
, 1));
3947 if (constant_term
!= const0_rtx
)
3948 x
= gen_rtx_PLUS (Pmode
, x
, constant_term
);
3953 /* Try a machine-dependent way of reloading an illegitimate address AD
3954 operand. If we find one, push the reload and return the new address.
3956 MODE is the mode of the enclosing MEM. OPNUM is the operand number
3957 and TYPE is the reload type of the current reload. */
3960 legitimize_reload_address (rtx ad
, enum machine_mode mode ATTRIBUTE_UNUSED
,
3961 int opnum
, int type
)
3963 if (!optimize
|| TARGET_LONG_DISPLACEMENT
)
3966 if (GET_CODE (ad
) == PLUS
)
3968 rtx tem
= simplify_binary_operation (PLUS
, Pmode
,
3969 XEXP (ad
, 0), XEXP (ad
, 1));
3974 if (GET_CODE (ad
) == PLUS
3975 && GET_CODE (XEXP (ad
, 0)) == REG
3976 && GET_CODE (XEXP (ad
, 1)) == CONST_INT
3977 && !DISP_IN_RANGE (INTVAL (XEXP (ad
, 1))))
3979 HOST_WIDE_INT lower
= INTVAL (XEXP (ad
, 1)) & 0xfff;
3980 HOST_WIDE_INT upper
= INTVAL (XEXP (ad
, 1)) ^ lower
;
3981 rtx cst
, tem
, new_rtx
;
3983 cst
= GEN_INT (upper
);
3984 if (!legitimate_reload_constant_p (cst
))
3985 cst
= force_const_mem (Pmode
, cst
);
3987 tem
= gen_rtx_PLUS (Pmode
, XEXP (ad
, 0), cst
);
3988 new_rtx
= gen_rtx_PLUS (Pmode
, tem
, GEN_INT (lower
));
3990 push_reload (XEXP (tem
, 1), 0, &XEXP (tem
, 1), 0,
3991 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0,
3992 opnum
, (enum reload_type
) type
);
3999 /* Emit code to move LEN bytes from DST to SRC. */
4002 s390_expand_movmem (rtx dst
, rtx src
, rtx len
)
4004 /* When tuning for z10 or higher we rely on the Glibc functions to
4005 do the right thing. Only for constant lengths below 64k we will
4006 generate inline code. */
4007 if (s390_tune
>= PROCESSOR_2097_Z10
4008 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > (1<<16)))
4011 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4013 if (INTVAL (len
) > 0)
4014 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (INTVAL (len
) - 1)));
4017 else if (TARGET_MVCLE
)
4019 emit_insn (gen_movmem_long (dst
, src
, convert_to_mode (Pmode
, len
, 1)));
4024 rtx dst_addr
, src_addr
, count
, blocks
, temp
;
4025 rtx loop_start_label
= gen_label_rtx ();
4026 rtx loop_end_label
= gen_label_rtx ();
4027 rtx end_label
= gen_label_rtx ();
4028 enum machine_mode mode
;
4030 mode
= GET_MODE (len
);
4031 if (mode
== VOIDmode
)
4034 dst_addr
= gen_reg_rtx (Pmode
);
4035 src_addr
= gen_reg_rtx (Pmode
);
4036 count
= gen_reg_rtx (mode
);
4037 blocks
= gen_reg_rtx (mode
);
4039 convert_move (count
, len
, 1);
4040 emit_cmp_and_jump_insns (count
, const0_rtx
,
4041 EQ
, NULL_RTX
, mode
, 1, end_label
);
4043 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4044 emit_move_insn (src_addr
, force_operand (XEXP (src
, 0), NULL_RTX
));
4045 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4046 src
= change_address (src
, VOIDmode
, src_addr
);
4048 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4051 emit_move_insn (count
, temp
);
4053 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4056 emit_move_insn (blocks
, temp
);
4058 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4059 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4061 emit_label (loop_start_label
);
4064 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 768))
4068 /* Issue a read prefetch for the +3 cache line. */
4069 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (768)),
4070 const0_rtx
, const0_rtx
);
4071 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4072 emit_insn (prefetch
);
4074 /* Issue a write prefetch for the +3 cache line. */
4075 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (768)),
4076 const1_rtx
, const0_rtx
);
4077 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4078 emit_insn (prefetch
);
4081 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (255)));
4082 s390_load_address (dst_addr
,
4083 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4084 s390_load_address (src_addr
,
4085 gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (256)));
4087 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4090 emit_move_insn (blocks
, temp
);
4092 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4093 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4095 emit_jump (loop_start_label
);
4096 emit_label (loop_end_label
);
4098 emit_insn (gen_movmem_short (dst
, src
,
4099 convert_to_mode (Pmode
, count
, 1)));
4100 emit_label (end_label
);
4105 /* Emit code to set LEN bytes at DST to VAL.
4106 Make use of clrmem if VAL is zero. */
4109 s390_expand_setmem (rtx dst
, rtx len
, rtx val
)
4111 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) == 0)
4114 gcc_assert (GET_CODE (val
) == CONST_INT
|| GET_MODE (val
) == QImode
);
4116 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) > 0 && INTVAL (len
) <= 257)
4118 if (val
== const0_rtx
&& INTVAL (len
) <= 256)
4119 emit_insn (gen_clrmem_short (dst
, GEN_INT (INTVAL (len
) - 1)));
4122 /* Initialize memory by storing the first byte. */
4123 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4125 if (INTVAL (len
) > 1)
4127 /* Initiate 1 byte overlap move.
4128 The first byte of DST is propagated through DSTP1.
4129 Prepare a movmem for: DST+1 = DST (length = LEN - 1).
4130 DST is set to size 1 so the rest of the memory location
4131 does not count as source operand. */
4132 rtx dstp1
= adjust_address (dst
, VOIDmode
, 1);
4133 set_mem_size (dst
, 1);
4135 emit_insn (gen_movmem_short (dstp1
, dst
,
4136 GEN_INT (INTVAL (len
) - 2)));
4141 else if (TARGET_MVCLE
)
4143 val
= force_not_mem (convert_modes (Pmode
, QImode
, val
, 1));
4144 emit_insn (gen_setmem_long (dst
, convert_to_mode (Pmode
, len
, 1), val
));
4149 rtx dst_addr
, count
, blocks
, temp
, dstp1
= NULL_RTX
;
4150 rtx loop_start_label
= gen_label_rtx ();
4151 rtx loop_end_label
= gen_label_rtx ();
4152 rtx end_label
= gen_label_rtx ();
4153 enum machine_mode mode
;
4155 mode
= GET_MODE (len
);
4156 if (mode
== VOIDmode
)
4159 dst_addr
= gen_reg_rtx (Pmode
);
4160 count
= gen_reg_rtx (mode
);
4161 blocks
= gen_reg_rtx (mode
);
4163 convert_move (count
, len
, 1);
4164 emit_cmp_and_jump_insns (count
, const0_rtx
,
4165 EQ
, NULL_RTX
, mode
, 1, end_label
);
4167 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4168 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4170 if (val
== const0_rtx
)
4171 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4175 dstp1
= adjust_address (dst
, VOIDmode
, 1);
4176 set_mem_size (dst
, 1);
4178 /* Initialize memory by storing the first byte. */
4179 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4181 /* If count is 1 we are done. */
4182 emit_cmp_and_jump_insns (count
, const1_rtx
,
4183 EQ
, NULL_RTX
, mode
, 1, end_label
);
4185 temp
= expand_binop (mode
, add_optab
, count
, GEN_INT (-2), count
, 1,
4189 emit_move_insn (count
, temp
);
4191 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4194 emit_move_insn (blocks
, temp
);
4196 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4197 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4199 emit_label (loop_start_label
);
4202 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 1024))
4204 /* Issue a write prefetch for the +4 cache line. */
4205 rtx prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
,
4207 const1_rtx
, const0_rtx
);
4208 emit_insn (prefetch
);
4209 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4212 if (val
== const0_rtx
)
4213 emit_insn (gen_clrmem_short (dst
, GEN_INT (255)));
4215 emit_insn (gen_movmem_short (dstp1
, dst
, GEN_INT (255)));
4216 s390_load_address (dst_addr
,
4217 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4219 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4222 emit_move_insn (blocks
, temp
);
4224 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4225 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4227 emit_jump (loop_start_label
);
4228 emit_label (loop_end_label
);
4230 if (val
== const0_rtx
)
4231 emit_insn (gen_clrmem_short (dst
, convert_to_mode (Pmode
, count
, 1)));
4233 emit_insn (gen_movmem_short (dstp1
, dst
, convert_to_mode (Pmode
, count
, 1)));
4234 emit_label (end_label
);
4238 /* Emit code to compare LEN bytes at OP0 with those at OP1,
4239 and return the result in TARGET. */
4242 s390_expand_cmpmem (rtx target
, rtx op0
, rtx op1
, rtx len
)
4244 rtx ccreg
= gen_rtx_REG (CCUmode
, CC_REGNUM
);
4247 /* When tuning for z10 or higher we rely on the Glibc functions to
4248 do the right thing. Only for constant lengths below 64k we will
4249 generate inline code. */
4250 if (s390_tune
>= PROCESSOR_2097_Z10
4251 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > (1<<16)))
4254 /* As the result of CMPINT is inverted compared to what we need,
4255 we have to swap the operands. */
4256 tmp
= op0
; op0
= op1
; op1
= tmp
;
4258 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4260 if (INTVAL (len
) > 0)
4262 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (INTVAL (len
) - 1)));
4263 emit_insn (gen_cmpint (target
, ccreg
));
4266 emit_move_insn (target
, const0_rtx
);
4268 else if (TARGET_MVCLE
)
4270 emit_insn (gen_cmpmem_long (op0
, op1
, convert_to_mode (Pmode
, len
, 1)));
4271 emit_insn (gen_cmpint (target
, ccreg
));
4275 rtx addr0
, addr1
, count
, blocks
, temp
;
4276 rtx loop_start_label
= gen_label_rtx ();
4277 rtx loop_end_label
= gen_label_rtx ();
4278 rtx end_label
= gen_label_rtx ();
4279 enum machine_mode mode
;
4281 mode
= GET_MODE (len
);
4282 if (mode
== VOIDmode
)
4285 addr0
= gen_reg_rtx (Pmode
);
4286 addr1
= gen_reg_rtx (Pmode
);
4287 count
= gen_reg_rtx (mode
);
4288 blocks
= gen_reg_rtx (mode
);
4290 convert_move (count
, len
, 1);
4291 emit_cmp_and_jump_insns (count
, const0_rtx
,
4292 EQ
, NULL_RTX
, mode
, 1, end_label
);
4294 emit_move_insn (addr0
, force_operand (XEXP (op0
, 0), NULL_RTX
));
4295 emit_move_insn (addr1
, force_operand (XEXP (op1
, 0), NULL_RTX
));
4296 op0
= change_address (op0
, VOIDmode
, addr0
);
4297 op1
= change_address (op1
, VOIDmode
, addr1
);
4299 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4302 emit_move_insn (count
, temp
);
4304 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4307 emit_move_insn (blocks
, temp
);
4309 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4310 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4312 emit_label (loop_start_label
);
4315 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 512))
4319 /* Issue a read prefetch for the +2 cache line of operand 1. */
4320 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (512)),
4321 const0_rtx
, const0_rtx
);
4322 emit_insn (prefetch
);
4323 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4325 /* Issue a read prefetch for the +2 cache line of operand 2. */
4326 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (512)),
4327 const0_rtx
, const0_rtx
);
4328 emit_insn (prefetch
);
4329 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4332 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (255)));
4333 temp
= gen_rtx_NE (VOIDmode
, ccreg
, const0_rtx
);
4334 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
4335 gen_rtx_LABEL_REF (VOIDmode
, end_label
), pc_rtx
);
4336 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
4337 emit_jump_insn (temp
);
4339 s390_load_address (addr0
,
4340 gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (256)));
4341 s390_load_address (addr1
,
4342 gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (256)));
4344 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4347 emit_move_insn (blocks
, temp
);
4349 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4350 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4352 emit_jump (loop_start_label
);
4353 emit_label (loop_end_label
);
4355 emit_insn (gen_cmpmem_short (op0
, op1
,
4356 convert_to_mode (Pmode
, count
, 1)));
4357 emit_label (end_label
);
4359 emit_insn (gen_cmpint (target
, ccreg
));
4365 /* Expand conditional increment or decrement using alc/slb instructions.
4366 Should generate code setting DST to either SRC or SRC + INCREMENT,
4367 depending on the result of the comparison CMP_OP0 CMP_CODE CMP_OP1.
4368 Returns true if successful, false otherwise.
4370 That makes it possible to implement some if-constructs without jumps e.g.:
4371 (borrow = CC0 | CC1 and carry = CC2 | CC3)
4372 unsigned int a, b, c;
4373 if (a < b) c++; -> CCU b > a -> CC2; c += carry;
4374 if (a < b) c--; -> CCL3 a - b -> borrow; c -= borrow;
4375 if (a <= b) c++; -> CCL3 b - a -> borrow; c += carry;
4376 if (a <= b) c--; -> CCU a <= b -> borrow; c -= borrow;
4378 Checks for EQ and NE with a nonzero value need an additional xor e.g.:
4379 if (a == b) c++; -> CCL3 a ^= b; 0 - a -> borrow; c += carry;
4380 if (a == b) c--; -> CCU a ^= b; a <= 0 -> CC0 | CC1; c -= borrow;
4381 if (a != b) c++; -> CCU a ^= b; a > 0 -> CC2; c += carry;
4382 if (a != b) c--; -> CCL3 a ^= b; 0 - a -> borrow; c -= borrow; */
4385 s390_expand_addcc (enum rtx_code cmp_code
, rtx cmp_op0
, rtx cmp_op1
,
4386 rtx dst
, rtx src
, rtx increment
)
4388 enum machine_mode cmp_mode
;
4389 enum machine_mode cc_mode
;
4395 if ((GET_MODE (cmp_op0
) == SImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4396 && (GET_MODE (cmp_op1
) == SImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4398 else if ((GET_MODE (cmp_op0
) == DImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4399 && (GET_MODE (cmp_op1
) == DImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4404 /* Try ADD LOGICAL WITH CARRY. */
4405 if (increment
== const1_rtx
)
4407 /* Determine CC mode to use. */
4408 if (cmp_code
== EQ
|| cmp_code
== NE
)
4410 if (cmp_op1
!= const0_rtx
)
4412 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4413 NULL_RTX
, 0, OPTAB_WIDEN
);
4414 cmp_op1
= const0_rtx
;
4417 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4420 if (cmp_code
== LTU
|| cmp_code
== LEU
)
4425 cmp_code
= swap_condition (cmp_code
);
4442 /* Emit comparison instruction pattern. */
4443 if (!register_operand (cmp_op0
, cmp_mode
))
4444 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4446 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4447 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4448 /* We use insn_invalid_p here to add clobbers if required. */
4449 ret
= insn_invalid_p (emit_insn (insn
), false);
4452 /* Emit ALC instruction pattern. */
4453 op_res
= gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4454 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4457 if (src
!= const0_rtx
)
4459 if (!register_operand (src
, GET_MODE (dst
)))
4460 src
= force_reg (GET_MODE (dst
), src
);
4462 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, src
);
4463 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, const0_rtx
);
4466 p
= rtvec_alloc (2);
4468 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4470 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4471 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4476 /* Try SUBTRACT LOGICAL WITH BORROW. */
4477 if (increment
== constm1_rtx
)
4479 /* Determine CC mode to use. */
4480 if (cmp_code
== EQ
|| cmp_code
== NE
)
4482 if (cmp_op1
!= const0_rtx
)
4484 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4485 NULL_RTX
, 0, OPTAB_WIDEN
);
4486 cmp_op1
= const0_rtx
;
4489 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4492 if (cmp_code
== GTU
|| cmp_code
== GEU
)
4497 cmp_code
= swap_condition (cmp_code
);
4514 /* Emit comparison instruction pattern. */
4515 if (!register_operand (cmp_op0
, cmp_mode
))
4516 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4518 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4519 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4520 /* We use insn_invalid_p here to add clobbers if required. */
4521 ret
= insn_invalid_p (emit_insn (insn
), false);
4524 /* Emit SLB instruction pattern. */
4525 if (!register_operand (src
, GET_MODE (dst
)))
4526 src
= force_reg (GET_MODE (dst
), src
);
4528 op_res
= gen_rtx_MINUS (GET_MODE (dst
),
4529 gen_rtx_MINUS (GET_MODE (dst
), src
, const0_rtx
),
4530 gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4531 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4533 p
= rtvec_alloc (2);
4535 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4537 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4538 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4546 /* Expand code for the insv template. Return true if successful. */
4549 s390_expand_insv (rtx dest
, rtx op1
, rtx op2
, rtx src
)
4551 int bitsize
= INTVAL (op1
);
4552 int bitpos
= INTVAL (op2
);
4553 enum machine_mode mode
= GET_MODE (dest
);
4554 enum machine_mode smode
;
4555 int smode_bsize
, mode_bsize
;
4558 /* Generate INSERT IMMEDIATE (IILL et al). */
4559 /* (set (ze (reg)) (const_int)). */
4561 && register_operand (dest
, word_mode
)
4562 && (bitpos
% 16) == 0
4563 && (bitsize
% 16) == 0
4564 && const_int_operand (src
, VOIDmode
))
4566 HOST_WIDE_INT val
= INTVAL (src
);
4567 int regpos
= bitpos
+ bitsize
;
4569 while (regpos
> bitpos
)
4571 enum machine_mode putmode
;
4574 if (TARGET_EXTIMM
&& (regpos
% 32 == 0) && (regpos
>= bitpos
+ 32))
4579 putsize
= GET_MODE_BITSIZE (putmode
);
4581 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4584 gen_int_mode (val
, putmode
));
4587 gcc_assert (regpos
== bitpos
);
4591 smode
= smallest_mode_for_size (bitsize
, MODE_INT
);
4592 smode_bsize
= GET_MODE_BITSIZE (smode
);
4593 mode_bsize
= GET_MODE_BITSIZE (mode
);
4595 /* Generate STORE CHARACTERS UNDER MASK (STCM et al). */
4597 && (bitsize
% BITS_PER_UNIT
) == 0
4599 && (register_operand (src
, word_mode
)
4600 || const_int_operand (src
, VOIDmode
)))
4602 /* Emit standard pattern if possible. */
4603 if (smode_bsize
== bitsize
)
4605 emit_move_insn (adjust_address (dest
, smode
, 0),
4606 gen_lowpart (smode
, src
));
4610 /* (set (ze (mem)) (const_int)). */
4611 else if (const_int_operand (src
, VOIDmode
))
4613 int size
= bitsize
/ BITS_PER_UNIT
;
4614 rtx src_mem
= adjust_address (force_const_mem (word_mode
, src
),
4616 UNITS_PER_WORD
- size
);
4618 dest
= adjust_address (dest
, BLKmode
, 0);
4619 set_mem_size (dest
, size
);
4620 s390_expand_movmem (dest
, src_mem
, GEN_INT (size
));
4624 /* (set (ze (mem)) (reg)). */
4625 else if (register_operand (src
, word_mode
))
4628 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, op1
,
4632 /* Emit st,stcmh sequence. */
4633 int stcmh_width
= bitsize
- 32;
4634 int size
= stcmh_width
/ BITS_PER_UNIT
;
4636 emit_move_insn (adjust_address (dest
, SImode
, size
),
4637 gen_lowpart (SImode
, src
));
4638 set_mem_size (dest
, size
);
4639 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4640 GEN_INT (stcmh_width
),
4642 gen_rtx_LSHIFTRT (word_mode
, src
, GEN_INT (32)));
4648 /* Generate INSERT CHARACTERS UNDER MASK (IC, ICM et al). */
4649 if ((bitpos
% BITS_PER_UNIT
) == 0
4650 && (bitsize
% BITS_PER_UNIT
) == 0
4651 && (bitpos
& 32) == ((bitpos
+ bitsize
- 1) & 32)
4653 && (mode
== DImode
|| mode
== SImode
)
4654 && register_operand (dest
, mode
))
4656 /* Emit a strict_low_part pattern if possible. */
4657 if (smode_bsize
== bitsize
&& bitpos
== mode_bsize
- smode_bsize
)
4659 op
= gen_rtx_STRICT_LOW_PART (VOIDmode
, gen_lowpart (smode
, dest
));
4660 op
= gen_rtx_SET (VOIDmode
, op
, gen_lowpart (smode
, src
));
4661 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4662 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4666 /* ??? There are more powerful versions of ICM that are not
4667 completely represented in the md file. */
4670 /* For z10, generate ROTATE THEN INSERT SELECTED BITS (RISBG et al). */
4671 if (TARGET_Z10
&& (mode
== DImode
|| mode
== SImode
))
4673 enum machine_mode mode_s
= GET_MODE (src
);
4675 if (mode_s
== VOIDmode
)
4677 /* Assume const_int etc already in the proper mode. */
4678 src
= force_reg (mode
, src
);
4680 else if (mode_s
!= mode
)
4682 gcc_assert (GET_MODE_BITSIZE (mode_s
) >= bitsize
);
4683 src
= force_reg (mode_s
, src
);
4684 src
= gen_lowpart (mode
, src
);
4687 op
= gen_rtx_SET (mode
,
4688 gen_rtx_ZERO_EXTRACT (mode
, dest
, op1
, op2
),
4690 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4691 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4699 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic which returns a
4700 register that holds VAL of mode MODE shifted by COUNT bits. */
4703 s390_expand_mask_and_shift (rtx val
, enum machine_mode mode
, rtx count
)
4705 val
= expand_simple_binop (SImode
, AND
, val
, GEN_INT (GET_MODE_MASK (mode
)),
4706 NULL_RTX
, 1, OPTAB_DIRECT
);
4707 return expand_simple_binop (SImode
, ASHIFT
, val
, count
,
4708 NULL_RTX
, 1, OPTAB_DIRECT
);
4711 /* Structure to hold the initial parameters for a compare_and_swap operation
4712 in HImode and QImode. */
4714 struct alignment_context
4716 rtx memsi
; /* SI aligned memory location. */
4717 rtx shift
; /* Bit offset with regard to lsb. */
4718 rtx modemask
; /* Mask of the HQImode shifted by SHIFT bits. */
4719 rtx modemaski
; /* ~modemask */
4720 bool aligned
; /* True if memory is aligned, false else. */
4723 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic to initialize
4724 structure AC for transparent simplifying, if the memory alignment is known
4725 to be at least 32bit. MEM is the memory location for the actual operation
4726 and MODE its mode. */
4729 init_alignment_context (struct alignment_context
*ac
, rtx mem
,
4730 enum machine_mode mode
)
4732 ac
->shift
= GEN_INT (GET_MODE_SIZE (SImode
) - GET_MODE_SIZE (mode
));
4733 ac
->aligned
= (MEM_ALIGN (mem
) >= GET_MODE_BITSIZE (SImode
));
4736 ac
->memsi
= adjust_address (mem
, SImode
, 0); /* Memory is aligned. */
4739 /* Alignment is unknown. */
4740 rtx byteoffset
, addr
, align
;
4742 /* Force the address into a register. */
4743 addr
= force_reg (Pmode
, XEXP (mem
, 0));
4745 /* Align it to SImode. */
4746 align
= expand_simple_binop (Pmode
, AND
, addr
,
4747 GEN_INT (-GET_MODE_SIZE (SImode
)),
4748 NULL_RTX
, 1, OPTAB_DIRECT
);
4750 ac
->memsi
= gen_rtx_MEM (SImode
, align
);
4751 MEM_VOLATILE_P (ac
->memsi
) = MEM_VOLATILE_P (mem
);
4752 set_mem_alias_set (ac
->memsi
, ALIAS_SET_MEMORY_BARRIER
);
4753 set_mem_align (ac
->memsi
, GET_MODE_BITSIZE (SImode
));
4755 /* Calculate shiftcount. */
4756 byteoffset
= expand_simple_binop (Pmode
, AND
, addr
,
4757 GEN_INT (GET_MODE_SIZE (SImode
) - 1),
4758 NULL_RTX
, 1, OPTAB_DIRECT
);
4759 /* As we already have some offset, evaluate the remaining distance. */
4760 ac
->shift
= expand_simple_binop (SImode
, MINUS
, ac
->shift
, byteoffset
,
4761 NULL_RTX
, 1, OPTAB_DIRECT
);
4764 /* Shift is the byte count, but we need the bitcount. */
4765 ac
->shift
= expand_simple_binop (SImode
, ASHIFT
, ac
->shift
, GEN_INT (3),
4766 NULL_RTX
, 1, OPTAB_DIRECT
);
4768 /* Calculate masks. */
4769 ac
->modemask
= expand_simple_binop (SImode
, ASHIFT
,
4770 GEN_INT (GET_MODE_MASK (mode
)),
4771 ac
->shift
, NULL_RTX
, 1, OPTAB_DIRECT
);
4772 ac
->modemaski
= expand_simple_unop (SImode
, NOT
, ac
->modemask
,
4776 /* A subroutine of s390_expand_cs_hqi. Insert INS into VAL. If possible,
4777 use a single insv insn into SEQ2. Otherwise, put prep insns in SEQ1 and
4778 perform the merge in SEQ2. */
4781 s390_two_part_insv (struct alignment_context
*ac
, rtx
*seq1
, rtx
*seq2
,
4782 enum machine_mode mode
, rtx val
, rtx ins
)
4789 tmp
= copy_to_mode_reg (SImode
, val
);
4790 if (s390_expand_insv (tmp
, GEN_INT (GET_MODE_BITSIZE (mode
)),
4794 *seq2
= get_insns ();
4801 /* Failed to use insv. Generate a two part shift and mask. */
4803 tmp
= s390_expand_mask_and_shift (ins
, mode
, ac
->shift
);
4804 *seq1
= get_insns ();
4808 tmp
= expand_simple_binop (SImode
, IOR
, tmp
, val
, NULL_RTX
, 1, OPTAB_DIRECT
);
4809 *seq2
= get_insns ();
4815 /* Expand an atomic compare and swap operation for HImode and QImode. MEM is
4816 the memory location, CMP the old value to compare MEM with and NEW_RTX the
4817 value to set if CMP == MEM. */
4820 s390_expand_cs_hqi (enum machine_mode mode
, rtx btarget
, rtx vtarget
, rtx mem
,
4821 rtx cmp
, rtx new_rtx
, bool is_weak
)
4823 struct alignment_context ac
;
4824 rtx cmpv
, newv
, val
, cc
, seq0
, seq1
, seq2
, seq3
;
4825 rtx res
= gen_reg_rtx (SImode
);
4826 rtx csloop
= NULL
, csend
= NULL
;
4828 gcc_assert (MEM_P (mem
));
4830 init_alignment_context (&ac
, mem
, mode
);
4832 /* Load full word. Subsequent loads are performed by CS. */
4833 val
= expand_simple_binop (SImode
, AND
, ac
.memsi
, ac
.modemaski
,
4834 NULL_RTX
, 1, OPTAB_DIRECT
);
4836 /* Prepare insertions of cmp and new_rtx into the loaded value. When
4837 possible, we try to use insv to make this happen efficiently. If
4838 that fails we'll generate code both inside and outside the loop. */
4839 cmpv
= s390_two_part_insv (&ac
, &seq0
, &seq2
, mode
, val
, cmp
);
4840 newv
= s390_two_part_insv (&ac
, &seq1
, &seq3
, mode
, val
, new_rtx
);
4847 /* Start CS loop. */
4850 /* Begin assuming success. */
4851 emit_move_insn (btarget
, const1_rtx
);
4853 csloop
= gen_label_rtx ();
4854 csend
= gen_label_rtx ();
4855 emit_label (csloop
);
4858 /* val = "<mem>00..0<mem>"
4859 * cmp = "00..0<cmp>00..0"
4860 * new = "00..0<new>00..0"
4866 cc
= s390_emit_compare_and_swap (EQ
, res
, ac
.memsi
, cmpv
, newv
);
4868 emit_insn (gen_cstorecc4 (btarget
, cc
, XEXP (cc
, 0), XEXP (cc
, 1)));
4873 /* Jump to end if we're done (likely?). */
4874 s390_emit_jump (csend
, cc
);
4876 /* Check for changes outside mode, and loop internal if so.
4877 Arrange the moves so that the compare is adjacent to the
4878 branch so that we can generate CRJ. */
4879 tmp
= copy_to_reg (val
);
4880 force_expand_binop (SImode
, and_optab
, res
, ac
.modemaski
, val
,
4882 cc
= s390_emit_compare (NE
, val
, tmp
);
4883 s390_emit_jump (csloop
, cc
);
4886 emit_move_insn (btarget
, const0_rtx
);
4890 /* Return the correct part of the bitfield. */
4891 convert_move (vtarget
, expand_simple_binop (SImode
, LSHIFTRT
, res
, ac
.shift
,
4892 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4895 /* Expand an atomic operation CODE of mode MODE. MEM is the memory location
4896 and VAL the value to play with. If AFTER is true then store the value
4897 MEM holds after the operation, if AFTER is false then store the value MEM
4898 holds before the operation. If TARGET is zero then discard that value, else
4899 store it to TARGET. */
4902 s390_expand_atomic (enum machine_mode mode
, enum rtx_code code
,
4903 rtx target
, rtx mem
, rtx val
, bool after
)
4905 struct alignment_context ac
;
4907 rtx new_rtx
= gen_reg_rtx (SImode
);
4908 rtx orig
= gen_reg_rtx (SImode
);
4909 rtx csloop
= gen_label_rtx ();
4911 gcc_assert (!target
|| register_operand (target
, VOIDmode
));
4912 gcc_assert (MEM_P (mem
));
4914 init_alignment_context (&ac
, mem
, mode
);
4916 /* Shift val to the correct bit positions.
4917 Preserve "icm", but prevent "ex icm". */
4918 if (!(ac
.aligned
&& code
== SET
&& MEM_P (val
)))
4919 val
= s390_expand_mask_and_shift (val
, mode
, ac
.shift
);
4921 /* Further preparation insns. */
4922 if (code
== PLUS
|| code
== MINUS
)
4923 emit_move_insn (orig
, val
);
4924 else if (code
== MULT
|| code
== AND
) /* val = "11..1<val>11..1" */
4925 val
= expand_simple_binop (SImode
, XOR
, val
, ac
.modemaski
,
4926 NULL_RTX
, 1, OPTAB_DIRECT
);
4928 /* Load full word. Subsequent loads are performed by CS. */
4929 cmp
= force_reg (SImode
, ac
.memsi
);
4931 /* Start CS loop. */
4932 emit_label (csloop
);
4933 emit_move_insn (new_rtx
, cmp
);
4935 /* Patch new with val at correct position. */
4940 val
= expand_simple_binop (SImode
, code
, new_rtx
, orig
,
4941 NULL_RTX
, 1, OPTAB_DIRECT
);
4942 val
= expand_simple_binop (SImode
, AND
, val
, ac
.modemask
,
4943 NULL_RTX
, 1, OPTAB_DIRECT
);
4946 if (ac
.aligned
&& MEM_P (val
))
4947 store_bit_field (new_rtx
, GET_MODE_BITSIZE (mode
), 0,
4951 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, ac
.modemaski
,
4952 NULL_RTX
, 1, OPTAB_DIRECT
);
4953 new_rtx
= expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4954 NULL_RTX
, 1, OPTAB_DIRECT
);
4960 new_rtx
= expand_simple_binop (SImode
, code
, new_rtx
, val
,
4961 NULL_RTX
, 1, OPTAB_DIRECT
);
4963 case MULT
: /* NAND */
4964 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, val
,
4965 NULL_RTX
, 1, OPTAB_DIRECT
);
4966 new_rtx
= expand_simple_binop (SImode
, XOR
, new_rtx
, ac
.modemask
,
4967 NULL_RTX
, 1, OPTAB_DIRECT
);
4973 s390_emit_jump (csloop
, s390_emit_compare_and_swap (NE
, cmp
,
4974 ac
.memsi
, cmp
, new_rtx
));
4976 /* Return the correct part of the bitfield. */
4978 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
,
4979 after
? new_rtx
: cmp
, ac
.shift
,
4980 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4983 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
4984 We need to emit DTP-relative relocations. */
4986 static void s390_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
4989 s390_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
4994 fputs ("\t.long\t", file
);
4997 fputs ("\t.quad\t", file
);
5002 output_addr_const (file
, x
);
5003 fputs ("@DTPOFF", file
);
5006 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
5007 /* Implement TARGET_MANGLE_TYPE. */
5010 s390_mangle_type (const_tree type
)
5012 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
5013 && TARGET_LONG_DOUBLE_128
)
5016 /* For all other types, use normal C++ mangling. */
5021 /* In the name of slightly smaller debug output, and to cater to
5022 general assembler lossage, recognize various UNSPEC sequences
5023 and turn them back into a direct symbol reference. */
5026 s390_delegitimize_address (rtx orig_x
)
5030 orig_x
= delegitimize_mem_from_attrs (orig_x
);
5033 /* Extract the symbol ref from:
5034 (plus:SI (reg:SI 12 %r12)
5035 (const:SI (unspec:SI [(symbol_ref/f:SI ("*.LC0"))]
5036 UNSPEC_GOTOFF/PLTOFF)))
5038 (plus:SI (reg:SI 12 %r12)
5039 (const:SI (plus:SI (unspec:SI [(symbol_ref:SI ("L"))]
5040 UNSPEC_GOTOFF/PLTOFF)
5041 (const_int 4 [0x4])))) */
5042 if (GET_CODE (x
) == PLUS
5043 && REG_P (XEXP (x
, 0))
5044 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
5045 && GET_CODE (XEXP (x
, 1)) == CONST
)
5047 HOST_WIDE_INT offset
= 0;
5049 /* The const operand. */
5050 y
= XEXP (XEXP (x
, 1), 0);
5052 if (GET_CODE (y
) == PLUS
5053 && GET_CODE (XEXP (y
, 1)) == CONST_INT
)
5055 offset
= INTVAL (XEXP (y
, 1));
5059 if (GET_CODE (y
) == UNSPEC
5060 && (XINT (y
, 1) == UNSPEC_GOTOFF
5061 || XINT (y
, 1) == UNSPEC_PLTOFF
))
5062 return plus_constant (Pmode
, XVECEXP (y
, 0, 0), offset
);
5065 if (GET_CODE (x
) != MEM
)
5069 if (GET_CODE (x
) == PLUS
5070 && GET_CODE (XEXP (x
, 1)) == CONST
5071 && GET_CODE (XEXP (x
, 0)) == REG
5072 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
5074 y
= XEXP (XEXP (x
, 1), 0);
5075 if (GET_CODE (y
) == UNSPEC
5076 && XINT (y
, 1) == UNSPEC_GOT
)
5077 y
= XVECEXP (y
, 0, 0);
5081 else if (GET_CODE (x
) == CONST
)
5083 /* Extract the symbol ref from:
5084 (mem:QI (const:DI (unspec:DI [(symbol_ref:DI ("foo"))]
5085 UNSPEC_PLT/GOTENT))) */
5088 if (GET_CODE (y
) == UNSPEC
5089 && (XINT (y
, 1) == UNSPEC_GOTENT
5090 || XINT (y
, 1) == UNSPEC_PLT
))
5091 y
= XVECEXP (y
, 0, 0);
5098 if (GET_MODE (orig_x
) != Pmode
)
5100 if (GET_MODE (orig_x
) == BLKmode
)
5102 y
= lowpart_subreg (GET_MODE (orig_x
), y
, Pmode
);
5109 /* Output operand OP to stdio stream FILE.
5110 OP is an address (register + offset) which is not used to address data;
5111 instead the rightmost bits are interpreted as the value. */
5114 print_shift_count_operand (FILE *file
, rtx op
)
5116 HOST_WIDE_INT offset
;
5119 /* Extract base register and offset. */
5120 if (!s390_decompose_shift_count (op
, &base
, &offset
))
5126 gcc_assert (GET_CODE (base
) == REG
);
5127 gcc_assert (REGNO (base
) < FIRST_PSEUDO_REGISTER
);
5128 gcc_assert (REGNO_REG_CLASS (REGNO (base
)) == ADDR_REGS
);
5131 /* Offsets are constricted to twelve bits. */
5132 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, offset
& ((1 << 12) - 1));
5134 fprintf (file
, "(%s)", reg_names
[REGNO (base
)]);
5137 /* See 'get_some_local_dynamic_name'. */
5140 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
5144 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
5146 x
= get_pool_constant (x
);
5147 return for_each_rtx (&x
, get_some_local_dynamic_name_1
, 0);
5150 if (GET_CODE (x
) == SYMBOL_REF
5151 && tls_symbolic_operand (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
5153 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
5160 /* Locate some local-dynamic symbol still in use by this function
5161 so that we can print its name in local-dynamic base patterns. */
5164 get_some_local_dynamic_name (void)
5168 if (cfun
->machine
->some_ld_name
)
5169 return cfun
->machine
->some_ld_name
;
5171 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5173 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
5174 return cfun
->machine
->some_ld_name
;
5179 /* Output machine-dependent UNSPECs occurring in address constant X
5180 in assembler syntax to stdio stream FILE. Returns true if the
5181 constant X could be recognized, false otherwise. */
5184 s390_output_addr_const_extra (FILE *file
, rtx x
)
5186 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 1)
5187 switch (XINT (x
, 1))
5190 output_addr_const (file
, XVECEXP (x
, 0, 0));
5191 fprintf (file
, "@GOTENT");
5194 output_addr_const (file
, XVECEXP (x
, 0, 0));
5195 fprintf (file
, "@GOT");
5198 output_addr_const (file
, XVECEXP (x
, 0, 0));
5199 fprintf (file
, "@GOTOFF");
5202 output_addr_const (file
, XVECEXP (x
, 0, 0));
5203 fprintf (file
, "@PLT");
5206 output_addr_const (file
, XVECEXP (x
, 0, 0));
5207 fprintf (file
, "@PLTOFF");
5210 output_addr_const (file
, XVECEXP (x
, 0, 0));
5211 fprintf (file
, "@TLSGD");
5214 assemble_name (file
, get_some_local_dynamic_name ());
5215 fprintf (file
, "@TLSLDM");
5218 output_addr_const (file
, XVECEXP (x
, 0, 0));
5219 fprintf (file
, "@DTPOFF");
5222 output_addr_const (file
, XVECEXP (x
, 0, 0));
5223 fprintf (file
, "@NTPOFF");
5225 case UNSPEC_GOTNTPOFF
:
5226 output_addr_const (file
, XVECEXP (x
, 0, 0));
5227 fprintf (file
, "@GOTNTPOFF");
5229 case UNSPEC_INDNTPOFF
:
5230 output_addr_const (file
, XVECEXP (x
, 0, 0));
5231 fprintf (file
, "@INDNTPOFF");
5235 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 2)
5236 switch (XINT (x
, 1))
5238 case UNSPEC_POOL_OFFSET
:
5239 x
= gen_rtx_MINUS (GET_MODE (x
), XVECEXP (x
, 0, 0), XVECEXP (x
, 0, 1));
5240 output_addr_const (file
, x
);
5246 /* Output address operand ADDR in assembler syntax to
5247 stdio stream FILE. */
5250 print_operand_address (FILE *file
, rtx addr
)
5252 struct s390_address ad
;
5254 if (s390_symref_operand_p (addr
, NULL
, NULL
))
5258 output_operand_lossage ("symbolic memory references are "
5259 "only supported on z10 or later");
5262 output_addr_const (file
, addr
);
5266 if (!s390_decompose_address (addr
, &ad
)
5267 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5268 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
5269 output_operand_lossage ("cannot decompose address");
5272 output_addr_const (file
, ad
.disp
);
5274 fprintf (file
, "0");
5276 if (ad
.base
&& ad
.indx
)
5277 fprintf (file
, "(%s,%s)", reg_names
[REGNO (ad
.indx
)],
5278 reg_names
[REGNO (ad
.base
)]);
5280 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5283 /* Output operand X in assembler syntax to stdio stream FILE.
5284 CODE specified the format flag. The following format flags
5287 'C': print opcode suffix for branch condition.
5288 'D': print opcode suffix for inverse branch condition.
5289 'E': print opcode suffix for branch on index instruction.
5290 'J': print tls_load/tls_gdcall/tls_ldcall suffix
5291 'G': print the size of the operand in bytes.
5292 'O': print only the displacement of a memory reference.
5293 'R': print only the base register of a memory reference.
5294 'S': print S-type memory reference (base+displacement).
5295 'N': print the second word of a DImode operand.
5296 'M': print the second word of a TImode operand.
5297 'Y': print shift count operand.
5299 'b': print integer X as if it's an unsigned byte.
5300 'c': print integer X as if it's an signed byte.
5301 'x': print integer X as if it's an unsigned halfword.
5302 'h': print integer X as if it's a signed halfword.
5303 'i': print the first nonzero HImode part of X.
5304 'j': print the first HImode part unequal to -1 of X.
5305 'k': print the first nonzero SImode part of X.
5306 'm': print the first SImode part unequal to -1 of X.
5307 'o': print integer X as if it's an unsigned 32bit word. */
5310 print_operand (FILE *file
, rtx x
, int code
)
5315 fprintf (file
, s390_branch_condition_mnemonic (x
, FALSE
));
5319 fprintf (file
, s390_branch_condition_mnemonic (x
, TRUE
));
5323 if (GET_CODE (x
) == LE
)
5324 fprintf (file
, "l");
5325 else if (GET_CODE (x
) == GT
)
5326 fprintf (file
, "h");
5328 output_operand_lossage ("invalid comparison operator "
5329 "for 'E' output modifier");
5333 if (GET_CODE (x
) == SYMBOL_REF
)
5335 fprintf (file
, "%s", ":tls_load:");
5336 output_addr_const (file
, x
);
5338 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD
)
5340 fprintf (file
, "%s", ":tls_gdcall:");
5341 output_addr_const (file
, XVECEXP (x
, 0, 0));
5343 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM
)
5345 fprintf (file
, "%s", ":tls_ldcall:");
5346 assemble_name (file
, get_some_local_dynamic_name ());
5349 output_operand_lossage ("invalid reference for 'J' output modifier");
5353 fprintf (file
, "%u", GET_MODE_SIZE (GET_MODE (x
)));
5358 struct s390_address ad
;
5363 output_operand_lossage ("memory reference expected for "
5364 "'O' output modifier");
5368 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5371 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5374 output_operand_lossage ("invalid address for 'O' output modifier");
5379 output_addr_const (file
, ad
.disp
);
5381 fprintf (file
, "0");
5387 struct s390_address ad
;
5392 output_operand_lossage ("memory reference expected for "
5393 "'R' output modifier");
5397 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5400 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5403 output_operand_lossage ("invalid address for 'R' output modifier");
5408 fprintf (file
, "%s", reg_names
[REGNO (ad
.base
)]);
5410 fprintf (file
, "0");
5416 struct s390_address ad
;
5421 output_operand_lossage ("memory reference expected for "
5422 "'S' output modifier");
5425 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5428 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5431 output_operand_lossage ("invalid address for 'S' output modifier");
5436 output_addr_const (file
, ad
.disp
);
5438 fprintf (file
, "0");
5441 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5446 if (GET_CODE (x
) == REG
)
5447 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5448 else if (GET_CODE (x
) == MEM
)
5449 x
= change_address (x
, VOIDmode
,
5450 plus_constant (Pmode
, XEXP (x
, 0), 4));
5452 output_operand_lossage ("register or memory expression expected "
5453 "for 'N' output modifier");
5457 if (GET_CODE (x
) == REG
)
5458 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5459 else if (GET_CODE (x
) == MEM
)
5460 x
= change_address (x
, VOIDmode
,
5461 plus_constant (Pmode
, XEXP (x
, 0), 8));
5463 output_operand_lossage ("register or memory expression expected "
5464 "for 'M' output modifier");
5468 print_shift_count_operand (file
, x
);
5472 switch (GET_CODE (x
))
5475 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5479 output_address (XEXP (x
, 0));
5486 output_addr_const (file
, x
);
5491 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xff);
5492 else if (code
== 'c')
5493 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xff) ^ 0x80) - 0x80);
5494 else if (code
== 'x')
5495 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffff);
5496 else if (code
== 'h')
5497 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xffff) ^ 0x8000) - 0x8000);
5498 else if (code
== 'i')
5499 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5500 s390_extract_part (x
, HImode
, 0));
5501 else if (code
== 'j')
5502 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5503 s390_extract_part (x
, HImode
, -1));
5504 else if (code
== 'k')
5505 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5506 s390_extract_part (x
, SImode
, 0));
5507 else if (code
== 'm')
5508 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5509 s390_extract_part (x
, SImode
, -1));
5510 else if (code
== 'o')
5511 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffffffff);
5513 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5517 gcc_assert (GET_MODE (x
) == VOIDmode
);
5519 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xff);
5520 else if (code
== 'x')
5521 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xffff);
5522 else if (code
== 'h')
5523 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5524 ((CONST_DOUBLE_LOW (x
) & 0xffff) ^ 0x8000) - 0x8000);
5528 output_operand_lossage ("invalid constant - try using "
5529 "an output modifier");
5531 output_operand_lossage ("invalid constant for output modifier '%c'",
5538 output_operand_lossage ("invalid expression - try using "
5539 "an output modifier");
5541 output_operand_lossage ("invalid expression for output "
5542 "modifier '%c'", code
);
5547 /* Target hook for assembling integer objects. We need to define it
5548 here to work a round a bug in some versions of GAS, which couldn't
5549 handle values smaller than INT_MIN when printed in decimal. */
5552 s390_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
5554 if (size
== 8 && aligned_p
5555 && GET_CODE (x
) == CONST_INT
&& INTVAL (x
) < INT_MIN
)
5557 fprintf (asm_out_file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n",
5561 return default_assemble_integer (x
, size
, aligned_p
);
5564 /* Returns true if register REGNO is used for forming
5565 a memory address in expression X. */
5568 reg_used_in_mem_p (int regno
, rtx x
)
5570 enum rtx_code code
= GET_CODE (x
);
5576 if (refers_to_regno_p (regno
, regno
+1,
5580 else if (code
== SET
5581 && GET_CODE (SET_DEST (x
)) == PC
)
5583 if (refers_to_regno_p (regno
, regno
+1,
5588 fmt
= GET_RTX_FORMAT (code
);
5589 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
5592 && reg_used_in_mem_p (regno
, XEXP (x
, i
)))
5595 else if (fmt
[i
] == 'E')
5596 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5597 if (reg_used_in_mem_p (regno
, XVECEXP (x
, i
, j
)))
5603 /* Returns true if expression DEP_RTX sets an address register
5604 used by instruction INSN to address memory. */
5607 addr_generation_dependency_p (rtx dep_rtx
, rtx insn
)
5611 if (GET_CODE (dep_rtx
) == INSN
)
5612 dep_rtx
= PATTERN (dep_rtx
);
5614 if (GET_CODE (dep_rtx
) == SET
)
5616 target
= SET_DEST (dep_rtx
);
5617 if (GET_CODE (target
) == STRICT_LOW_PART
)
5618 target
= XEXP (target
, 0);
5619 while (GET_CODE (target
) == SUBREG
)
5620 target
= SUBREG_REG (target
);
5622 if (GET_CODE (target
) == REG
)
5624 int regno
= REGNO (target
);
5626 if (s390_safe_attr_type (insn
) == TYPE_LA
)
5628 pat
= PATTERN (insn
);
5629 if (GET_CODE (pat
) == PARALLEL
)
5631 gcc_assert (XVECLEN (pat
, 0) == 2);
5632 pat
= XVECEXP (pat
, 0, 0);
5634 gcc_assert (GET_CODE (pat
) == SET
);
5635 return refers_to_regno_p (regno
, regno
+1, SET_SRC (pat
), 0);
5637 else if (get_attr_atype (insn
) == ATYPE_AGEN
)
5638 return reg_used_in_mem_p (regno
, PATTERN (insn
));
5644 /* Return 1, if dep_insn sets register used in insn in the agen unit. */
5647 s390_agen_dep_p (rtx dep_insn
, rtx insn
)
5649 rtx dep_rtx
= PATTERN (dep_insn
);
5652 if (GET_CODE (dep_rtx
) == SET
5653 && addr_generation_dependency_p (dep_rtx
, insn
))
5655 else if (GET_CODE (dep_rtx
) == PARALLEL
)
5657 for (i
= 0; i
< XVECLEN (dep_rtx
, 0); i
++)
5659 if (addr_generation_dependency_p (XVECEXP (dep_rtx
, 0, i
), insn
))
5667 /* A C statement (sans semicolon) to update the integer scheduling priority
5668 INSN_PRIORITY (INSN). Increase the priority to execute the INSN earlier,
5669 reduce the priority to execute INSN later. Do not define this macro if
5670 you do not need to adjust the scheduling priorities of insns.
5672 A STD instruction should be scheduled earlier,
5673 in order to use the bypass. */
5675 s390_adjust_priority (rtx insn ATTRIBUTE_UNUSED
, int priority
)
5677 if (! INSN_P (insn
))
5680 if (s390_tune
!= PROCESSOR_2084_Z990
5681 && s390_tune
!= PROCESSOR_2094_Z9_109
5682 && s390_tune
!= PROCESSOR_2097_Z10
5683 && s390_tune
!= PROCESSOR_2817_Z196
)
5686 switch (s390_safe_attr_type (insn
))
5690 priority
= priority
<< 3;
5694 priority
= priority
<< 1;
5703 /* The number of instructions that can be issued per cycle. */
5706 s390_issue_rate (void)
5710 case PROCESSOR_2084_Z990
:
5711 case PROCESSOR_2094_Z9_109
:
5712 case PROCESSOR_2817_Z196
:
5714 case PROCESSOR_2097_Z10
:
5722 s390_first_cycle_multipass_dfa_lookahead (void)
5727 /* Annotate every literal pool reference in X by an UNSPEC_LTREF expression.
5728 Fix up MEMs as required. */
5731 annotate_constant_pool_refs (rtx
*x
)
5736 gcc_assert (GET_CODE (*x
) != SYMBOL_REF
5737 || !CONSTANT_POOL_ADDRESS_P (*x
));
5739 /* Literal pool references can only occur inside a MEM ... */
5740 if (GET_CODE (*x
) == MEM
)
5742 rtx memref
= XEXP (*x
, 0);
5744 if (GET_CODE (memref
) == SYMBOL_REF
5745 && CONSTANT_POOL_ADDRESS_P (memref
))
5747 rtx base
= cfun
->machine
->base_reg
;
5748 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, memref
, base
),
5751 *x
= replace_equiv_address (*x
, addr
);
5755 if (GET_CODE (memref
) == CONST
5756 && GET_CODE (XEXP (memref
, 0)) == PLUS
5757 && GET_CODE (XEXP (XEXP (memref
, 0), 1)) == CONST_INT
5758 && GET_CODE (XEXP (XEXP (memref
, 0), 0)) == SYMBOL_REF
5759 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (memref
, 0), 0)))
5761 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (memref
, 0), 1));
5762 rtx sym
= XEXP (XEXP (memref
, 0), 0);
5763 rtx base
= cfun
->machine
->base_reg
;
5764 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5767 *x
= replace_equiv_address (*x
, plus_constant (Pmode
, addr
, off
));
5772 /* ... or a load-address type pattern. */
5773 if (GET_CODE (*x
) == SET
)
5775 rtx addrref
= SET_SRC (*x
);
5777 if (GET_CODE (addrref
) == SYMBOL_REF
5778 && CONSTANT_POOL_ADDRESS_P (addrref
))
5780 rtx base
= cfun
->machine
->base_reg
;
5781 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, addrref
, base
),
5784 SET_SRC (*x
) = addr
;
5788 if (GET_CODE (addrref
) == CONST
5789 && GET_CODE (XEXP (addrref
, 0)) == PLUS
5790 && GET_CODE (XEXP (XEXP (addrref
, 0), 1)) == CONST_INT
5791 && GET_CODE (XEXP (XEXP (addrref
, 0), 0)) == SYMBOL_REF
5792 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (addrref
, 0), 0)))
5794 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (addrref
, 0), 1));
5795 rtx sym
= XEXP (XEXP (addrref
, 0), 0);
5796 rtx base
= cfun
->machine
->base_reg
;
5797 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5800 SET_SRC (*x
) = plus_constant (Pmode
, addr
, off
);
5805 /* Annotate LTREL_BASE as well. */
5806 if (GET_CODE (*x
) == UNSPEC
5807 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5809 rtx base
= cfun
->machine
->base_reg
;
5810 *x
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XVECEXP (*x
, 0, 0), base
),
5815 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5816 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5820 annotate_constant_pool_refs (&XEXP (*x
, i
));
5822 else if (fmt
[i
] == 'E')
5824 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5825 annotate_constant_pool_refs (&XVECEXP (*x
, i
, j
));
5830 /* Split all branches that exceed the maximum distance.
5831 Returns true if this created a new literal pool entry. */
5834 s390_split_branches (void)
5836 rtx temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
5837 int new_literal
= 0, ret
;
5838 rtx insn
, pat
, tmp
, target
;
5841 /* We need correct insn addresses. */
5843 shorten_branches (get_insns ());
5845 /* Find all branches that exceed 64KB, and split them. */
5847 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5849 if (GET_CODE (insn
) != JUMP_INSN
)
5852 pat
= PATTERN (insn
);
5853 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
5854 pat
= XVECEXP (pat
, 0, 0);
5855 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
5858 if (GET_CODE (SET_SRC (pat
)) == LABEL_REF
)
5860 label
= &SET_SRC (pat
);
5862 else if (GET_CODE (SET_SRC (pat
)) == IF_THEN_ELSE
)
5864 if (GET_CODE (XEXP (SET_SRC (pat
), 1)) == LABEL_REF
)
5865 label
= &XEXP (SET_SRC (pat
), 1);
5866 else if (GET_CODE (XEXP (SET_SRC (pat
), 2)) == LABEL_REF
)
5867 label
= &XEXP (SET_SRC (pat
), 2);
5874 if (get_attr_length (insn
) <= 4)
5877 /* We are going to use the return register as scratch register,
5878 make sure it will be saved/restored by the prologue/epilogue. */
5879 cfun_frame_layout
.save_return_addr_p
= 1;
5884 tmp
= force_const_mem (Pmode
, *label
);
5885 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, tmp
), insn
);
5886 INSN_ADDRESSES_NEW (tmp
, -1);
5887 annotate_constant_pool_refs (&PATTERN (tmp
));
5894 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, *label
),
5895 UNSPEC_LTREL_OFFSET
);
5896 target
= gen_rtx_CONST (Pmode
, target
);
5897 target
= force_const_mem (Pmode
, target
);
5898 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, target
), insn
);
5899 INSN_ADDRESSES_NEW (tmp
, -1);
5900 annotate_constant_pool_refs (&PATTERN (tmp
));
5902 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XEXP (target
, 0),
5903 cfun
->machine
->base_reg
),
5905 target
= gen_rtx_PLUS (Pmode
, temp_reg
, target
);
5908 ret
= validate_change (insn
, label
, target
, 0);
5916 /* Find an annotated literal pool symbol referenced in RTX X,
5917 and store it at REF. Will abort if X contains references to
5918 more than one such pool symbol; multiple references to the same
5919 symbol are allowed, however.
5921 The rtx pointed to by REF must be initialized to NULL_RTX
5922 by the caller before calling this routine. */
5925 find_constant_pool_ref (rtx x
, rtx
*ref
)
5930 /* Ignore LTREL_BASE references. */
5931 if (GET_CODE (x
) == UNSPEC
5932 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5934 /* Likewise POOL_ENTRY insns. */
5935 if (GET_CODE (x
) == UNSPEC_VOLATILE
5936 && XINT (x
, 1) == UNSPECV_POOL_ENTRY
)
5939 gcc_assert (GET_CODE (x
) != SYMBOL_REF
5940 || !CONSTANT_POOL_ADDRESS_P (x
));
5942 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_LTREF
)
5944 rtx sym
= XVECEXP (x
, 0, 0);
5945 gcc_assert (GET_CODE (sym
) == SYMBOL_REF
5946 && CONSTANT_POOL_ADDRESS_P (sym
));
5948 if (*ref
== NULL_RTX
)
5951 gcc_assert (*ref
== sym
);
5956 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5957 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5961 find_constant_pool_ref (XEXP (x
, i
), ref
);
5963 else if (fmt
[i
] == 'E')
5965 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5966 find_constant_pool_ref (XVECEXP (x
, i
, j
), ref
);
5971 /* Replace every reference to the annotated literal pool
5972 symbol REF in X by its base plus OFFSET. */
5975 replace_constant_pool_ref (rtx
*x
, rtx ref
, rtx offset
)
5980 gcc_assert (*x
!= ref
);
5982 if (GET_CODE (*x
) == UNSPEC
5983 && XINT (*x
, 1) == UNSPEC_LTREF
5984 && XVECEXP (*x
, 0, 0) == ref
)
5986 *x
= gen_rtx_PLUS (Pmode
, XVECEXP (*x
, 0, 1), offset
);
5990 if (GET_CODE (*x
) == PLUS
5991 && GET_CODE (XEXP (*x
, 1)) == CONST_INT
5992 && GET_CODE (XEXP (*x
, 0)) == UNSPEC
5993 && XINT (XEXP (*x
, 0), 1) == UNSPEC_LTREF
5994 && XVECEXP (XEXP (*x
, 0), 0, 0) == ref
)
5996 rtx addr
= gen_rtx_PLUS (Pmode
, XVECEXP (XEXP (*x
, 0), 0, 1), offset
);
5997 *x
= plus_constant (Pmode
, addr
, INTVAL (XEXP (*x
, 1)));
6001 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
6002 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
6006 replace_constant_pool_ref (&XEXP (*x
, i
), ref
, offset
);
6008 else if (fmt
[i
] == 'E')
6010 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
6011 replace_constant_pool_ref (&XVECEXP (*x
, i
, j
), ref
, offset
);
6016 /* Check whether X contains an UNSPEC_LTREL_BASE.
6017 Return its constant pool symbol if found, NULL_RTX otherwise. */
6020 find_ltrel_base (rtx x
)
6025 if (GET_CODE (x
) == UNSPEC
6026 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
6027 return XVECEXP (x
, 0, 0);
6029 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
6030 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
6034 rtx fnd
= find_ltrel_base (XEXP (x
, i
));
6038 else if (fmt
[i
] == 'E')
6040 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
6042 rtx fnd
= find_ltrel_base (XVECEXP (x
, i
, j
));
6052 /* Replace any occurrence of UNSPEC_LTREL_BASE in X with its base. */
6055 replace_ltrel_base (rtx
*x
)
6060 if (GET_CODE (*x
) == UNSPEC
6061 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
6063 *x
= XVECEXP (*x
, 0, 1);
6067 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
6068 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
6072 replace_ltrel_base (&XEXP (*x
, i
));
6074 else if (fmt
[i
] == 'E')
6076 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
6077 replace_ltrel_base (&XVECEXP (*x
, i
, j
));
6083 /* We keep a list of constants which we have to add to internal
6084 constant tables in the middle of large functions. */
6086 #define NR_C_MODES 11
6087 enum machine_mode constant_modes
[NR_C_MODES
] =
6089 TFmode
, TImode
, TDmode
,
6090 DFmode
, DImode
, DDmode
,
6091 SFmode
, SImode
, SDmode
,
6098 struct constant
*next
;
6103 struct constant_pool
6105 struct constant_pool
*next
;
6109 rtx emit_pool_after
;
6111 struct constant
*constants
[NR_C_MODES
];
6112 struct constant
*execute
;
6117 /* Allocate new constant_pool structure. */
6119 static struct constant_pool
*
6120 s390_alloc_pool (void)
6122 struct constant_pool
*pool
;
6125 pool
= (struct constant_pool
*) xmalloc (sizeof *pool
);
6127 for (i
= 0; i
< NR_C_MODES
; i
++)
6128 pool
->constants
[i
] = NULL
;
6130 pool
->execute
= NULL
;
6131 pool
->label
= gen_label_rtx ();
6132 pool
->first_insn
= NULL_RTX
;
6133 pool
->pool_insn
= NULL_RTX
;
6134 pool
->insns
= BITMAP_ALLOC (NULL
);
6136 pool
->emit_pool_after
= NULL_RTX
;
6141 /* Create new constant pool covering instructions starting at INSN
6142 and chain it to the end of POOL_LIST. */
6144 static struct constant_pool
*
6145 s390_start_pool (struct constant_pool
**pool_list
, rtx insn
)
6147 struct constant_pool
*pool
, **prev
;
6149 pool
= s390_alloc_pool ();
6150 pool
->first_insn
= insn
;
6152 for (prev
= pool_list
; *prev
; prev
= &(*prev
)->next
)
6159 /* End range of instructions covered by POOL at INSN and emit
6160 placeholder insn representing the pool. */
6163 s390_end_pool (struct constant_pool
*pool
, rtx insn
)
6165 rtx pool_size
= GEN_INT (pool
->size
+ 8 /* alignment slop */);
6168 insn
= get_last_insn ();
6170 pool
->pool_insn
= emit_insn_after (gen_pool (pool_size
), insn
);
6171 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6174 /* Add INSN to the list of insns covered by POOL. */
6177 s390_add_pool_insn (struct constant_pool
*pool
, rtx insn
)
6179 bitmap_set_bit (pool
->insns
, INSN_UID (insn
));
6182 /* Return pool out of POOL_LIST that covers INSN. */
6184 static struct constant_pool
*
6185 s390_find_pool (struct constant_pool
*pool_list
, rtx insn
)
6187 struct constant_pool
*pool
;
6189 for (pool
= pool_list
; pool
; pool
= pool
->next
)
6190 if (bitmap_bit_p (pool
->insns
, INSN_UID (insn
)))
6196 /* Add constant VAL of mode MODE to the constant pool POOL. */
6199 s390_add_constant (struct constant_pool
*pool
, rtx val
, enum machine_mode mode
)
6204 for (i
= 0; i
< NR_C_MODES
; i
++)
6205 if (constant_modes
[i
] == mode
)
6207 gcc_assert (i
!= NR_C_MODES
);
6209 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6210 if (rtx_equal_p (val
, c
->value
))
6215 c
= (struct constant
*) xmalloc (sizeof *c
);
6217 c
->label
= gen_label_rtx ();
6218 c
->next
= pool
->constants
[i
];
6219 pool
->constants
[i
] = c
;
6220 pool
->size
+= GET_MODE_SIZE (mode
);
6224 /* Return an rtx that represents the offset of X from the start of
6228 s390_pool_offset (struct constant_pool
*pool
, rtx x
)
6232 label
= gen_rtx_LABEL_REF (GET_MODE (x
), pool
->label
);
6233 x
= gen_rtx_UNSPEC (GET_MODE (x
), gen_rtvec (2, x
, label
),
6234 UNSPEC_POOL_OFFSET
);
6235 return gen_rtx_CONST (GET_MODE (x
), x
);
6238 /* Find constant VAL of mode MODE in the constant pool POOL.
6239 Return an RTX describing the distance from the start of
6240 the pool to the location of the new constant. */
6243 s390_find_constant (struct constant_pool
*pool
, rtx val
,
6244 enum machine_mode mode
)
6249 for (i
= 0; i
< NR_C_MODES
; i
++)
6250 if (constant_modes
[i
] == mode
)
6252 gcc_assert (i
!= NR_C_MODES
);
6254 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6255 if (rtx_equal_p (val
, c
->value
))
6260 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6263 /* Check whether INSN is an execute. Return the label_ref to its
6264 execute target template if so, NULL_RTX otherwise. */
6267 s390_execute_label (rtx insn
)
6269 if (GET_CODE (insn
) == INSN
6270 && GET_CODE (PATTERN (insn
)) == PARALLEL
6271 && GET_CODE (XVECEXP (PATTERN (insn
), 0, 0)) == UNSPEC
6272 && XINT (XVECEXP (PATTERN (insn
), 0, 0), 1) == UNSPEC_EXECUTE
)
6273 return XVECEXP (XVECEXP (PATTERN (insn
), 0, 0), 0, 2);
6278 /* Add execute target for INSN to the constant pool POOL. */
6281 s390_add_execute (struct constant_pool
*pool
, rtx insn
)
6285 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6286 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6291 c
= (struct constant
*) xmalloc (sizeof *c
);
6293 c
->label
= gen_label_rtx ();
6294 c
->next
= pool
->execute
;
6300 /* Find execute target for INSN in the constant pool POOL.
6301 Return an RTX describing the distance from the start of
6302 the pool to the location of the execute target. */
6305 s390_find_execute (struct constant_pool
*pool
, rtx insn
)
6309 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6310 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6315 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6318 /* For an execute INSN, extract the execute target template. */
6321 s390_execute_target (rtx insn
)
6323 rtx pattern
= PATTERN (insn
);
6324 gcc_assert (s390_execute_label (insn
));
6326 if (XVECLEN (pattern
, 0) == 2)
6328 pattern
= copy_rtx (XVECEXP (pattern
, 0, 1));
6332 rtvec vec
= rtvec_alloc (XVECLEN (pattern
, 0) - 1);
6335 for (i
= 0; i
< XVECLEN (pattern
, 0) - 1; i
++)
6336 RTVEC_ELT (vec
, i
) = copy_rtx (XVECEXP (pattern
, 0, i
+ 1));
6338 pattern
= gen_rtx_PARALLEL (VOIDmode
, vec
);
6344 /* Indicate that INSN cannot be duplicated. This is the case for
6345 execute insns that carry a unique label. */
6348 s390_cannot_copy_insn_p (rtx insn
)
6350 rtx label
= s390_execute_label (insn
);
6351 return label
&& label
!= const0_rtx
;
6354 /* Dump out the constants in POOL. If REMOTE_LABEL is true,
6355 do not emit the pool base label. */
6358 s390_dump_pool (struct constant_pool
*pool
, bool remote_label
)
6361 rtx insn
= pool
->pool_insn
;
6364 /* Switch to rodata section. */
6365 if (TARGET_CPU_ZARCH
)
6367 insn
= emit_insn_after (gen_pool_section_start (), insn
);
6368 INSN_ADDRESSES_NEW (insn
, -1);
6371 /* Ensure minimum pool alignment. */
6372 if (TARGET_CPU_ZARCH
)
6373 insn
= emit_insn_after (gen_pool_align (GEN_INT (8)), insn
);
6375 insn
= emit_insn_after (gen_pool_align (GEN_INT (4)), insn
);
6376 INSN_ADDRESSES_NEW (insn
, -1);
6378 /* Emit pool base label. */
6381 insn
= emit_label_after (pool
->label
, insn
);
6382 INSN_ADDRESSES_NEW (insn
, -1);
6385 /* Dump constants in descending alignment requirement order,
6386 ensuring proper alignment for every constant. */
6387 for (i
= 0; i
< NR_C_MODES
; i
++)
6388 for (c
= pool
->constants
[i
]; c
; c
= c
->next
)
6390 /* Convert UNSPEC_LTREL_OFFSET unspecs to pool-relative references. */
6391 rtx value
= copy_rtx (c
->value
);
6392 if (GET_CODE (value
) == CONST
6393 && GET_CODE (XEXP (value
, 0)) == UNSPEC
6394 && XINT (XEXP (value
, 0), 1) == UNSPEC_LTREL_OFFSET
6395 && XVECLEN (XEXP (value
, 0), 0) == 1)
6396 value
= s390_pool_offset (pool
, XVECEXP (XEXP (value
, 0), 0, 0));
6398 insn
= emit_label_after (c
->label
, insn
);
6399 INSN_ADDRESSES_NEW (insn
, -1);
6401 value
= gen_rtx_UNSPEC_VOLATILE (constant_modes
[i
],
6402 gen_rtvec (1, value
),
6403 UNSPECV_POOL_ENTRY
);
6404 insn
= emit_insn_after (value
, insn
);
6405 INSN_ADDRESSES_NEW (insn
, -1);
6408 /* Ensure minimum alignment for instructions. */
6409 insn
= emit_insn_after (gen_pool_align (GEN_INT (2)), insn
);
6410 INSN_ADDRESSES_NEW (insn
, -1);
6412 /* Output in-pool execute template insns. */
6413 for (c
= pool
->execute
; c
; c
= c
->next
)
6415 insn
= emit_label_after (c
->label
, insn
);
6416 INSN_ADDRESSES_NEW (insn
, -1);
6418 insn
= emit_insn_after (s390_execute_target (c
->value
), insn
);
6419 INSN_ADDRESSES_NEW (insn
, -1);
6422 /* Switch back to previous section. */
6423 if (TARGET_CPU_ZARCH
)
6425 insn
= emit_insn_after (gen_pool_section_end (), insn
);
6426 INSN_ADDRESSES_NEW (insn
, -1);
6429 insn
= emit_barrier_after (insn
);
6430 INSN_ADDRESSES_NEW (insn
, -1);
6432 /* Remove placeholder insn. */
6433 remove_insn (pool
->pool_insn
);
6436 /* Free all memory used by POOL. */
6439 s390_free_pool (struct constant_pool
*pool
)
6441 struct constant
*c
, *next
;
6444 for (i
= 0; i
< NR_C_MODES
; i
++)
6445 for (c
= pool
->constants
[i
]; c
; c
= next
)
6451 for (c
= pool
->execute
; c
; c
= next
)
6457 BITMAP_FREE (pool
->insns
);
6462 /* Collect main literal pool. Return NULL on overflow. */
6464 static struct constant_pool
*
6465 s390_mainpool_start (void)
6467 struct constant_pool
*pool
;
6470 pool
= s390_alloc_pool ();
6472 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6474 if (GET_CODE (insn
) == INSN
6475 && GET_CODE (PATTERN (insn
)) == SET
6476 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC_VOLATILE
6477 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPECV_MAIN_POOL
)
6479 gcc_assert (!pool
->pool_insn
);
6480 pool
->pool_insn
= insn
;
6483 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6485 s390_add_execute (pool
, insn
);
6487 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6489 rtx pool_ref
= NULL_RTX
;
6490 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6493 rtx constant
= get_pool_constant (pool_ref
);
6494 enum machine_mode mode
= get_pool_mode (pool_ref
);
6495 s390_add_constant (pool
, constant
, mode
);
6499 /* If hot/cold partitioning is enabled we have to make sure that
6500 the literal pool is emitted in the same section where the
6501 initialization of the literal pool base pointer takes place.
6502 emit_pool_after is only used in the non-overflow case on non
6503 Z cpus where we can emit the literal pool at the end of the
6504 function body within the text section. */
6506 && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
6507 && !pool
->emit_pool_after
)
6508 pool
->emit_pool_after
= PREV_INSN (insn
);
6511 gcc_assert (pool
->pool_insn
|| pool
->size
== 0);
6513 if (pool
->size
>= 4096)
6515 /* We're going to chunkify the pool, so remove the main
6516 pool placeholder insn. */
6517 remove_insn (pool
->pool_insn
);
6519 s390_free_pool (pool
);
6523 /* If the functions ends with the section where the literal pool
6524 should be emitted set the marker to its end. */
6525 if (pool
&& !pool
->emit_pool_after
)
6526 pool
->emit_pool_after
= get_last_insn ();
6531 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6532 Modify the current function to output the pool constants as well as
6533 the pool register setup instruction. */
6536 s390_mainpool_finish (struct constant_pool
*pool
)
6538 rtx base_reg
= cfun
->machine
->base_reg
;
6541 /* If the pool is empty, we're done. */
6542 if (pool
->size
== 0)
6544 /* We don't actually need a base register after all. */
6545 cfun
->machine
->base_reg
= NULL_RTX
;
6547 if (pool
->pool_insn
)
6548 remove_insn (pool
->pool_insn
);
6549 s390_free_pool (pool
);
6553 /* We need correct insn addresses. */
6554 shorten_branches (get_insns ());
6556 /* On zSeries, we use a LARL to load the pool register. The pool is
6557 located in the .rodata section, so we emit it after the function. */
6558 if (TARGET_CPU_ZARCH
)
6560 insn
= gen_main_base_64 (base_reg
, pool
->label
);
6561 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6562 INSN_ADDRESSES_NEW (insn
, -1);
6563 remove_insn (pool
->pool_insn
);
6565 insn
= get_last_insn ();
6566 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6567 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6569 s390_dump_pool (pool
, 0);
6572 /* On S/390, if the total size of the function's code plus literal pool
6573 does not exceed 4096 bytes, we use BASR to set up a function base
6574 pointer, and emit the literal pool at the end of the function. */
6575 else if (INSN_ADDRESSES (INSN_UID (pool
->emit_pool_after
))
6576 + pool
->size
+ 8 /* alignment slop */ < 4096)
6578 insn
= gen_main_base_31_small (base_reg
, pool
->label
);
6579 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6580 INSN_ADDRESSES_NEW (insn
, -1);
6581 remove_insn (pool
->pool_insn
);
6583 insn
= emit_label_after (pool
->label
, insn
);
6584 INSN_ADDRESSES_NEW (insn
, -1);
6586 /* emit_pool_after will be set by s390_mainpool_start to the
6587 last insn of the section where the literal pool should be
6589 insn
= pool
->emit_pool_after
;
6591 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6592 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6594 s390_dump_pool (pool
, 1);
6597 /* Otherwise, we emit an inline literal pool and use BASR to branch
6598 over it, setting up the pool register at the same time. */
6601 rtx pool_end
= gen_label_rtx ();
6603 insn
= gen_main_base_31_large (base_reg
, pool
->label
, pool_end
);
6604 insn
= emit_jump_insn_after (insn
, pool
->pool_insn
);
6605 JUMP_LABEL (insn
) = pool_end
;
6606 INSN_ADDRESSES_NEW (insn
, -1);
6607 remove_insn (pool
->pool_insn
);
6609 insn
= emit_label_after (pool
->label
, insn
);
6610 INSN_ADDRESSES_NEW (insn
, -1);
6612 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6613 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6615 insn
= emit_label_after (pool_end
, pool
->pool_insn
);
6616 INSN_ADDRESSES_NEW (insn
, -1);
6618 s390_dump_pool (pool
, 1);
6622 /* Replace all literal pool references. */
6624 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6627 replace_ltrel_base (&PATTERN (insn
));
6629 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6631 rtx addr
, pool_ref
= NULL_RTX
;
6632 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6635 if (s390_execute_label (insn
))
6636 addr
= s390_find_execute (pool
, insn
);
6638 addr
= s390_find_constant (pool
, get_pool_constant (pool_ref
),
6639 get_pool_mode (pool_ref
));
6641 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6642 INSN_CODE (insn
) = -1;
6648 /* Free the pool. */
6649 s390_free_pool (pool
);
6652 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6653 We have decided we cannot use this pool, so revert all changes
6654 to the current function that were done by s390_mainpool_start. */
6656 s390_mainpool_cancel (struct constant_pool
*pool
)
6658 /* We didn't actually change the instruction stream, so simply
6659 free the pool memory. */
6660 s390_free_pool (pool
);
6664 /* Chunkify the literal pool. */
6666 #define S390_POOL_CHUNK_MIN 0xc00
6667 #define S390_POOL_CHUNK_MAX 0xe00
6669 static struct constant_pool
*
6670 s390_chunkify_start (void)
6672 struct constant_pool
*curr_pool
= NULL
, *pool_list
= NULL
;
6675 rtx pending_ltrel
= NULL_RTX
;
6678 rtx (*gen_reload_base
) (rtx
, rtx
) =
6679 TARGET_CPU_ZARCH
? gen_reload_base_64
: gen_reload_base_31
;
6682 /* We need correct insn addresses. */
6684 shorten_branches (get_insns ());
6686 /* Scan all insns and move literals to pool chunks. */
6688 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6690 bool section_switch_p
= false;
6692 /* Check for pending LTREL_BASE. */
6695 rtx ltrel_base
= find_ltrel_base (PATTERN (insn
));
6698 gcc_assert (ltrel_base
== pending_ltrel
);
6699 pending_ltrel
= NULL_RTX
;
6703 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6706 curr_pool
= s390_start_pool (&pool_list
, insn
);
6708 s390_add_execute (curr_pool
, insn
);
6709 s390_add_pool_insn (curr_pool
, insn
);
6711 else if (GET_CODE (insn
) == INSN
|| CALL_P (insn
))
6713 rtx pool_ref
= NULL_RTX
;
6714 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6717 rtx constant
= get_pool_constant (pool_ref
);
6718 enum machine_mode mode
= get_pool_mode (pool_ref
);
6721 curr_pool
= s390_start_pool (&pool_list
, insn
);
6723 s390_add_constant (curr_pool
, constant
, mode
);
6724 s390_add_pool_insn (curr_pool
, insn
);
6726 /* Don't split the pool chunk between a LTREL_OFFSET load
6727 and the corresponding LTREL_BASE. */
6728 if (GET_CODE (constant
) == CONST
6729 && GET_CODE (XEXP (constant
, 0)) == UNSPEC
6730 && XINT (XEXP (constant
, 0), 1) == UNSPEC_LTREL_OFFSET
)
6732 gcc_assert (!pending_ltrel
);
6733 pending_ltrel
= pool_ref
;
6738 if (GET_CODE (insn
) == JUMP_INSN
|| GET_CODE (insn
) == CODE_LABEL
)
6741 s390_add_pool_insn (curr_pool
, insn
);
6742 /* An LTREL_BASE must follow within the same basic block. */
6743 gcc_assert (!pending_ltrel
);
6747 switch (NOTE_KIND (insn
))
6749 case NOTE_INSN_SWITCH_TEXT_SECTIONS
:
6750 section_switch_p
= true;
6752 case NOTE_INSN_VAR_LOCATION
:
6753 case NOTE_INSN_CALL_ARG_LOCATION
:
6760 || INSN_ADDRESSES_SIZE () <= (size_t) INSN_UID (insn
)
6761 || INSN_ADDRESSES (INSN_UID (insn
)) == -1)
6764 if (TARGET_CPU_ZARCH
)
6766 if (curr_pool
->size
< S390_POOL_CHUNK_MAX
)
6769 s390_end_pool (curr_pool
, NULL_RTX
);
6774 int chunk_size
= INSN_ADDRESSES (INSN_UID (insn
))
6775 - INSN_ADDRESSES (INSN_UID (curr_pool
->first_insn
))
6778 /* We will later have to insert base register reload insns.
6779 Those will have an effect on code size, which we need to
6780 consider here. This calculation makes rather pessimistic
6781 worst-case assumptions. */
6782 if (GET_CODE (insn
) == CODE_LABEL
)
6785 if (chunk_size
< S390_POOL_CHUNK_MIN
6786 && curr_pool
->size
< S390_POOL_CHUNK_MIN
6787 && !section_switch_p
)
6790 /* Pool chunks can only be inserted after BARRIERs ... */
6791 if (GET_CODE (insn
) == BARRIER
)
6793 s390_end_pool (curr_pool
, insn
);
6798 /* ... so if we don't find one in time, create one. */
6799 else if (chunk_size
> S390_POOL_CHUNK_MAX
6800 || curr_pool
->size
> S390_POOL_CHUNK_MAX
6801 || section_switch_p
)
6803 rtx label
, jump
, barrier
, next
, prev
;
6805 if (!section_switch_p
)
6807 /* We can insert the barrier only after a 'real' insn. */
6808 if (GET_CODE (insn
) != INSN
&& GET_CODE (insn
) != CALL_INSN
)
6810 if (get_attr_length (insn
) == 0)
6812 /* Don't separate LTREL_BASE from the corresponding
6813 LTREL_OFFSET load. */
6820 next
= NEXT_INSN (insn
);
6824 && (NOTE_KIND (next
) == NOTE_INSN_VAR_LOCATION
6825 || NOTE_KIND (next
) == NOTE_INSN_CALL_ARG_LOCATION
));
6829 gcc_assert (!pending_ltrel
);
6831 /* The old pool has to end before the section switch
6832 note in order to make it part of the current
6834 insn
= PREV_INSN (insn
);
6837 label
= gen_label_rtx ();
6839 if (prev
&& NOTE_P (prev
))
6840 prev
= prev_nonnote_insn (prev
);
6842 jump
= emit_jump_insn_after_setloc (gen_jump (label
), insn
,
6843 INSN_LOCATOR (prev
));
6845 jump
= emit_jump_insn_after_noloc (gen_jump (label
), insn
);
6846 barrier
= emit_barrier_after (jump
);
6847 insn
= emit_label_after (label
, barrier
);
6848 JUMP_LABEL (jump
) = label
;
6849 LABEL_NUSES (label
) = 1;
6851 INSN_ADDRESSES_NEW (jump
, -1);
6852 INSN_ADDRESSES_NEW (barrier
, -1);
6853 INSN_ADDRESSES_NEW (insn
, -1);
6855 s390_end_pool (curr_pool
, barrier
);
6863 s390_end_pool (curr_pool
, NULL_RTX
);
6864 gcc_assert (!pending_ltrel
);
6866 /* Find all labels that are branched into
6867 from an insn belonging to a different chunk. */
6869 far_labels
= BITMAP_ALLOC (NULL
);
6871 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6873 /* Labels marked with LABEL_PRESERVE_P can be target
6874 of non-local jumps, so we have to mark them.
6875 The same holds for named labels.
6877 Don't do that, however, if it is the label before
6880 if (GET_CODE (insn
) == CODE_LABEL
6881 && (LABEL_PRESERVE_P (insn
) || LABEL_NAME (insn
)))
6883 rtx vec_insn
= next_real_insn (insn
);
6884 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6885 PATTERN (vec_insn
) : NULL_RTX
;
6887 || !(GET_CODE (vec_pat
) == ADDR_VEC
6888 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6889 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (insn
));
6892 /* If we have a direct jump (conditional or unconditional)
6893 or a casesi jump, check all potential targets. */
6894 else if (GET_CODE (insn
) == JUMP_INSN
)
6896 rtx pat
= PATTERN (insn
);
6897 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
6898 pat
= XVECEXP (pat
, 0, 0);
6900 if (GET_CODE (pat
) == SET
)
6902 rtx label
= JUMP_LABEL (insn
);
6905 if (s390_find_pool (pool_list
, label
)
6906 != s390_find_pool (pool_list
, insn
))
6907 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6910 else if (GET_CODE (pat
) == PARALLEL
6911 && XVECLEN (pat
, 0) == 2
6912 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
6913 && GET_CODE (XVECEXP (pat
, 0, 1)) == USE
6914 && GET_CODE (XEXP (XVECEXP (pat
, 0, 1), 0)) == LABEL_REF
)
6916 /* Find the jump table used by this casesi jump. */
6917 rtx vec_label
= XEXP (XEXP (XVECEXP (pat
, 0, 1), 0), 0);
6918 rtx vec_insn
= next_real_insn (vec_label
);
6919 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6920 PATTERN (vec_insn
) : NULL_RTX
;
6922 && (GET_CODE (vec_pat
) == ADDR_VEC
6923 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6925 int i
, diff_p
= GET_CODE (vec_pat
) == ADDR_DIFF_VEC
;
6927 for (i
= 0; i
< XVECLEN (vec_pat
, diff_p
); i
++)
6929 rtx label
= XEXP (XVECEXP (vec_pat
, diff_p
, i
), 0);
6931 if (s390_find_pool (pool_list
, label
)
6932 != s390_find_pool (pool_list
, insn
))
6933 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6940 /* Insert base register reload insns before every pool. */
6942 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6944 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6946 rtx insn
= curr_pool
->first_insn
;
6947 INSN_ADDRESSES_NEW (emit_insn_before (new_insn
, insn
), -1);
6950 /* Insert base register reload insns at every far label. */
6952 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6953 if (GET_CODE (insn
) == CODE_LABEL
6954 && bitmap_bit_p (far_labels
, CODE_LABEL_NUMBER (insn
)))
6956 struct constant_pool
*pool
= s390_find_pool (pool_list
, insn
);
6959 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6961 INSN_ADDRESSES_NEW (emit_insn_after (new_insn
, insn
), -1);
6966 BITMAP_FREE (far_labels
);
6969 /* Recompute insn addresses. */
6971 init_insn_lengths ();
6972 shorten_branches (get_insns ());
6977 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6978 After we have decided to use this list, finish implementing
6979 all changes to the current function as required. */
6982 s390_chunkify_finish (struct constant_pool
*pool_list
)
6984 struct constant_pool
*curr_pool
= NULL
;
6988 /* Replace all literal pool references. */
6990 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6993 replace_ltrel_base (&PATTERN (insn
));
6995 curr_pool
= s390_find_pool (pool_list
, insn
);
6999 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
7001 rtx addr
, pool_ref
= NULL_RTX
;
7002 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
7005 if (s390_execute_label (insn
))
7006 addr
= s390_find_execute (curr_pool
, insn
);
7008 addr
= s390_find_constant (curr_pool
,
7009 get_pool_constant (pool_ref
),
7010 get_pool_mode (pool_ref
));
7012 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
7013 INSN_CODE (insn
) = -1;
7018 /* Dump out all literal pools. */
7020 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
7021 s390_dump_pool (curr_pool
, 0);
7023 /* Free pool list. */
7027 struct constant_pool
*next
= pool_list
->next
;
7028 s390_free_pool (pool_list
);
7033 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
7034 We have decided we cannot use this list, so revert all changes
7035 to the current function that were done by s390_chunkify_start. */
7038 s390_chunkify_cancel (struct constant_pool
*pool_list
)
7040 struct constant_pool
*curr_pool
= NULL
;
7043 /* Remove all pool placeholder insns. */
7045 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
7047 /* Did we insert an extra barrier? Remove it. */
7048 rtx barrier
= PREV_INSN (curr_pool
->pool_insn
);
7049 rtx jump
= barrier
? PREV_INSN (barrier
) : NULL_RTX
;
7050 rtx label
= NEXT_INSN (curr_pool
->pool_insn
);
7052 if (jump
&& GET_CODE (jump
) == JUMP_INSN
7053 && barrier
&& GET_CODE (barrier
) == BARRIER
7054 && label
&& GET_CODE (label
) == CODE_LABEL
7055 && GET_CODE (PATTERN (jump
)) == SET
7056 && SET_DEST (PATTERN (jump
)) == pc_rtx
7057 && GET_CODE (SET_SRC (PATTERN (jump
))) == LABEL_REF
7058 && XEXP (SET_SRC (PATTERN (jump
)), 0) == label
)
7061 remove_insn (barrier
);
7062 remove_insn (label
);
7065 remove_insn (curr_pool
->pool_insn
);
7068 /* Remove all base register reload insns. */
7070 for (insn
= get_insns (); insn
; )
7072 rtx next_insn
= NEXT_INSN (insn
);
7074 if (GET_CODE (insn
) == INSN
7075 && GET_CODE (PATTERN (insn
)) == SET
7076 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC
7077 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPEC_RELOAD_BASE
)
7083 /* Free pool list. */
7087 struct constant_pool
*next
= pool_list
->next
;
7088 s390_free_pool (pool_list
);
7093 /* Output the constant pool entry EXP in mode MODE with alignment ALIGN. */
7096 s390_output_pool_entry (rtx exp
, enum machine_mode mode
, unsigned int align
)
7100 switch (GET_MODE_CLASS (mode
))
7103 case MODE_DECIMAL_FLOAT
:
7104 gcc_assert (GET_CODE (exp
) == CONST_DOUBLE
);
7106 REAL_VALUE_FROM_CONST_DOUBLE (r
, exp
);
7107 assemble_real (r
, mode
, align
);
7111 assemble_integer (exp
, GET_MODE_SIZE (mode
), align
, 1);
7112 mark_symbol_refs_as_used (exp
);
7121 /* Return an RTL expression representing the value of the return address
7122 for the frame COUNT steps up from the current frame. FRAME is the
7123 frame pointer of that frame. */
7126 s390_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
7131 /* Without backchain, we fail for all but the current frame. */
7133 if (!TARGET_BACKCHAIN
&& count
> 0)
7136 /* For the current frame, we need to make sure the initial
7137 value of RETURN_REGNUM is actually saved. */
7141 /* On non-z architectures branch splitting could overwrite r14. */
7142 if (TARGET_CPU_ZARCH
)
7143 return get_hard_reg_initial_val (Pmode
, RETURN_REGNUM
);
7146 cfun_frame_layout
.save_return_addr_p
= true;
7147 return gen_rtx_MEM (Pmode
, return_address_pointer_rtx
);
7151 if (TARGET_PACKED_STACK
)
7152 offset
= -2 * UNITS_PER_LONG
;
7154 offset
= RETURN_REGNUM
* UNITS_PER_LONG
;
7156 addr
= plus_constant (Pmode
, frame
, offset
);
7157 addr
= memory_address (Pmode
, addr
);
7158 return gen_rtx_MEM (Pmode
, addr
);
7161 /* Return an RTL expression representing the back chain stored in
7162 the current stack frame. */
7165 s390_back_chain_rtx (void)
7169 gcc_assert (TARGET_BACKCHAIN
);
7171 if (TARGET_PACKED_STACK
)
7172 chain
= plus_constant (Pmode
, stack_pointer_rtx
,
7173 STACK_POINTER_OFFSET
- UNITS_PER_LONG
);
7175 chain
= stack_pointer_rtx
;
7177 chain
= gen_rtx_MEM (Pmode
, chain
);
7181 /* Find first call clobbered register unused in a function.
7182 This could be used as base register in a leaf function
7183 or for holding the return address before epilogue. */
7186 find_unused_clobbered_reg (void)
7189 for (i
= 0; i
< 6; i
++)
7190 if (!df_regs_ever_live_p (i
))
7196 /* Helper function for s390_regs_ever_clobbered. Sets the fields in DATA for all
7197 clobbered hard regs in SETREG. */
7200 s390_reg_clobbered_rtx (rtx setreg
, const_rtx set_insn ATTRIBUTE_UNUSED
, void *data
)
7202 int *regs_ever_clobbered
= (int *)data
;
7203 unsigned int i
, regno
;
7204 enum machine_mode mode
= GET_MODE (setreg
);
7206 if (GET_CODE (setreg
) == SUBREG
)
7208 rtx inner
= SUBREG_REG (setreg
);
7209 if (!GENERAL_REG_P (inner
))
7211 regno
= subreg_regno (setreg
);
7213 else if (GENERAL_REG_P (setreg
))
7214 regno
= REGNO (setreg
);
7219 i
< regno
+ HARD_REGNO_NREGS (regno
, mode
);
7221 regs_ever_clobbered
[i
] = 1;
7224 /* Walks through all basic blocks of the current function looking
7225 for clobbered hard regs using s390_reg_clobbered_rtx. The fields
7226 of the passed integer array REGS_EVER_CLOBBERED are set to one for
7227 each of those regs. */
7230 s390_regs_ever_clobbered (int *regs_ever_clobbered
)
7236 memset (regs_ever_clobbered
, 0, 16 * sizeof (int));
7238 /* For non-leaf functions we have to consider all call clobbered regs to be
7242 for (i
= 0; i
< 16; i
++)
7243 regs_ever_clobbered
[i
] = call_really_used_regs
[i
];
7246 /* Make the "magic" eh_return registers live if necessary. For regs_ever_live
7247 this work is done by liveness analysis (mark_regs_live_at_end).
7248 Special care is needed for functions containing landing pads. Landing pads
7249 may use the eh registers, but the code which sets these registers is not
7250 contained in that function. Hence s390_regs_ever_clobbered is not able to
7251 deal with this automatically. */
7252 if (crtl
->calls_eh_return
|| cfun
->machine
->has_landing_pad_p
)
7253 for (i
= 0; EH_RETURN_DATA_REGNO (i
) != INVALID_REGNUM
; i
++)
7254 if (crtl
->calls_eh_return
7255 || (cfun
->machine
->has_landing_pad_p
7256 && df_regs_ever_live_p (EH_RETURN_DATA_REGNO (i
))))
7257 regs_ever_clobbered
[EH_RETURN_DATA_REGNO (i
)] = 1;
7259 /* For nonlocal gotos all call-saved registers have to be saved.
7260 This flag is also set for the unwinding code in libgcc.
7261 See expand_builtin_unwind_init. For regs_ever_live this is done by
7263 if (cfun
->has_nonlocal_label
)
7264 for (i
= 0; i
< 16; i
++)
7265 if (!call_really_used_regs
[i
])
7266 regs_ever_clobbered
[i
] = 1;
7268 FOR_EACH_BB (cur_bb
)
7270 FOR_BB_INSNS (cur_bb
, cur_insn
)
7272 if (INSN_P (cur_insn
))
7273 note_stores (PATTERN (cur_insn
),
7274 s390_reg_clobbered_rtx
,
7275 regs_ever_clobbered
);
7280 /* Determine the frame area which actually has to be accessed
7281 in the function epilogue. The values are stored at the
7282 given pointers AREA_BOTTOM (address of the lowest used stack
7283 address) and AREA_TOP (address of the first item which does
7284 not belong to the stack frame). */
7287 s390_frame_area (int *area_bottom
, int *area_top
)
7295 if (cfun_frame_layout
.first_restore_gpr
!= -1)
7297 b
= (cfun_frame_layout
.gprs_offset
7298 + cfun_frame_layout
.first_restore_gpr
* UNITS_PER_LONG
);
7299 t
= b
+ (cfun_frame_layout
.last_restore_gpr
7300 - cfun_frame_layout
.first_restore_gpr
+ 1) * UNITS_PER_LONG
;
7303 if (TARGET_64BIT
&& cfun_save_high_fprs_p
)
7305 b
= MIN (b
, cfun_frame_layout
.f8_offset
);
7306 t
= MAX (t
, (cfun_frame_layout
.f8_offset
7307 + cfun_frame_layout
.high_fprs
* 8));
7311 for (i
= 2; i
< 4; i
++)
7312 if (cfun_fpr_bit_p (i
))
7314 b
= MIN (b
, cfun_frame_layout
.f4_offset
+ (i
- 2) * 8);
7315 t
= MAX (t
, cfun_frame_layout
.f4_offset
+ (i
- 1) * 8);
7322 /* Fill cfun->machine with info about register usage of current function.
7323 Return in CLOBBERED_REGS which GPRs are currently considered set. */
7326 s390_register_info (int clobbered_regs
[])
7330 /* fprs 8 - 15 are call saved for 64 Bit ABI. */
7331 cfun_frame_layout
.fpr_bitmap
= 0;
7332 cfun_frame_layout
.high_fprs
= 0;
7334 for (i
= 24; i
< 32; i
++)
7335 if (df_regs_ever_live_p (i
) && !global_regs
[i
])
7337 cfun_set_fpr_bit (i
- 16);
7338 cfun_frame_layout
.high_fprs
++;
7341 /* Find first and last gpr to be saved. We trust regs_ever_live
7342 data, except that we don't save and restore global registers.
7344 Also, all registers with special meaning to the compiler need
7345 to be handled extra. */
7347 s390_regs_ever_clobbered (clobbered_regs
);
7349 for (i
= 0; i
< 16; i
++)
7350 clobbered_regs
[i
] = clobbered_regs
[i
] && !global_regs
[i
] && !fixed_regs
[i
];
7352 if (frame_pointer_needed
)
7353 clobbered_regs
[HARD_FRAME_POINTER_REGNUM
] = 1;
7356 clobbered_regs
[PIC_OFFSET_TABLE_REGNUM
]
7357 |= df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
);
7359 clobbered_regs
[BASE_REGNUM
]
7360 |= (cfun
->machine
->base_reg
7361 && REGNO (cfun
->machine
->base_reg
) == BASE_REGNUM
);
7363 clobbered_regs
[RETURN_REGNUM
]
7365 || TARGET_TPF_PROFILING
7366 || cfun
->machine
->split_branches_pending_p
7367 || cfun_frame_layout
.save_return_addr_p
7368 || crtl
->calls_eh_return
7371 clobbered_regs
[STACK_POINTER_REGNUM
]
7373 || TARGET_TPF_PROFILING
7374 || cfun_save_high_fprs_p
7375 || get_frame_size () > 0
7376 || cfun
->calls_alloca
7379 for (i
= 6; i
< 16; i
++)
7380 if (df_regs_ever_live_p (i
) || clobbered_regs
[i
])
7382 for (j
= 15; j
> i
; j
--)
7383 if (df_regs_ever_live_p (j
) || clobbered_regs
[j
])
7388 /* Nothing to save/restore. */
7389 cfun_frame_layout
.first_save_gpr_slot
= -1;
7390 cfun_frame_layout
.last_save_gpr_slot
= -1;
7391 cfun_frame_layout
.first_save_gpr
= -1;
7392 cfun_frame_layout
.first_restore_gpr
= -1;
7393 cfun_frame_layout
.last_save_gpr
= -1;
7394 cfun_frame_layout
.last_restore_gpr
= -1;
7398 /* Save slots for gprs from i to j. */
7399 cfun_frame_layout
.first_save_gpr_slot
= i
;
7400 cfun_frame_layout
.last_save_gpr_slot
= j
;
7402 for (i
= cfun_frame_layout
.first_save_gpr_slot
;
7403 i
< cfun_frame_layout
.last_save_gpr_slot
+ 1;
7405 if (clobbered_regs
[i
])
7408 for (j
= cfun_frame_layout
.last_save_gpr_slot
; j
> i
; j
--)
7409 if (clobbered_regs
[j
])
7412 if (i
== cfun_frame_layout
.last_save_gpr_slot
+ 1)
7414 /* Nothing to save/restore. */
7415 cfun_frame_layout
.first_save_gpr
= -1;
7416 cfun_frame_layout
.first_restore_gpr
= -1;
7417 cfun_frame_layout
.last_save_gpr
= -1;
7418 cfun_frame_layout
.last_restore_gpr
= -1;
7422 /* Save / Restore from gpr i to j. */
7423 cfun_frame_layout
.first_save_gpr
= i
;
7424 cfun_frame_layout
.first_restore_gpr
= i
;
7425 cfun_frame_layout
.last_save_gpr
= j
;
7426 cfun_frame_layout
.last_restore_gpr
= j
;
7432 /* Varargs functions need to save gprs 2 to 6. */
7433 if (cfun
->va_list_gpr_size
7434 && crtl
->args
.info
.gprs
< GP_ARG_NUM_REG
)
7436 int min_gpr
= crtl
->args
.info
.gprs
;
7437 int max_gpr
= min_gpr
+ cfun
->va_list_gpr_size
;
7438 if (max_gpr
> GP_ARG_NUM_REG
)
7439 max_gpr
= GP_ARG_NUM_REG
;
7441 if (cfun_frame_layout
.first_save_gpr
== -1
7442 || cfun_frame_layout
.first_save_gpr
> 2 + min_gpr
)
7444 cfun_frame_layout
.first_save_gpr
= 2 + min_gpr
;
7445 cfun_frame_layout
.first_save_gpr_slot
= 2 + min_gpr
;
7448 if (cfun_frame_layout
.last_save_gpr
== -1
7449 || cfun_frame_layout
.last_save_gpr
< 2 + max_gpr
- 1)
7451 cfun_frame_layout
.last_save_gpr
= 2 + max_gpr
- 1;
7452 cfun_frame_layout
.last_save_gpr_slot
= 2 + max_gpr
- 1;
7456 /* Mark f0, f2 for 31 bit and f0-f4 for 64 bit to be saved. */
7457 if (TARGET_HARD_FLOAT
&& cfun
->va_list_fpr_size
7458 && crtl
->args
.info
.fprs
< FP_ARG_NUM_REG
)
7460 int min_fpr
= crtl
->args
.info
.fprs
;
7461 int max_fpr
= min_fpr
+ cfun
->va_list_fpr_size
;
7462 if (max_fpr
> FP_ARG_NUM_REG
)
7463 max_fpr
= FP_ARG_NUM_REG
;
7465 /* ??? This is currently required to ensure proper location
7466 of the fpr save slots within the va_list save area. */
7467 if (TARGET_PACKED_STACK
)
7470 for (i
= min_fpr
; i
< max_fpr
; i
++)
7471 cfun_set_fpr_bit (i
);
7476 for (i
= 2; i
< 4; i
++)
7477 if (df_regs_ever_live_p (i
+ 16) && !global_regs
[i
+ 16])
7478 cfun_set_fpr_bit (i
);
7481 /* Fill cfun->machine with info about frame of current function. */
7484 s390_frame_info (void)
7488 cfun_frame_layout
.frame_size
= get_frame_size ();
7489 if (!TARGET_64BIT
&& cfun_frame_layout
.frame_size
> 0x7fff0000)
7490 fatal_error ("total size of local variables exceeds architecture limit");
7492 if (!TARGET_PACKED_STACK
)
7494 cfun_frame_layout
.backchain_offset
= 0;
7495 cfun_frame_layout
.f0_offset
= 16 * UNITS_PER_LONG
;
7496 cfun_frame_layout
.f4_offset
= cfun_frame_layout
.f0_offset
+ 2 * 8;
7497 cfun_frame_layout
.f8_offset
= -cfun_frame_layout
.high_fprs
* 8;
7498 cfun_frame_layout
.gprs_offset
= (cfun_frame_layout
.first_save_gpr_slot
7501 else if (TARGET_BACKCHAIN
) /* kernel stack layout */
7503 cfun_frame_layout
.backchain_offset
= (STACK_POINTER_OFFSET
7505 cfun_frame_layout
.gprs_offset
7506 = (cfun_frame_layout
.backchain_offset
7507 - (STACK_POINTER_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
+ 1)
7512 cfun_frame_layout
.f4_offset
7513 = (cfun_frame_layout
.gprs_offset
7514 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7516 cfun_frame_layout
.f0_offset
7517 = (cfun_frame_layout
.f4_offset
7518 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7522 /* On 31 bit we have to care about alignment of the
7523 floating point regs to provide fastest access. */
7524 cfun_frame_layout
.f0_offset
7525 = ((cfun_frame_layout
.gprs_offset
7526 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1))
7527 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7529 cfun_frame_layout
.f4_offset
7530 = (cfun_frame_layout
.f0_offset
7531 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7534 else /* no backchain */
7536 cfun_frame_layout
.f4_offset
7537 = (STACK_POINTER_OFFSET
7538 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7540 cfun_frame_layout
.f0_offset
7541 = (cfun_frame_layout
.f4_offset
7542 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7544 cfun_frame_layout
.gprs_offset
7545 = cfun_frame_layout
.f0_offset
- cfun_gprs_save_area_size
;
7549 && !TARGET_TPF_PROFILING
7550 && cfun_frame_layout
.frame_size
== 0
7551 && !cfun_save_high_fprs_p
7552 && !cfun
->calls_alloca
7556 if (!TARGET_PACKED_STACK
)
7557 cfun_frame_layout
.frame_size
+= (STACK_POINTER_OFFSET
7558 + crtl
->outgoing_args_size
7559 + cfun_frame_layout
.high_fprs
* 8);
7562 if (TARGET_BACKCHAIN
)
7563 cfun_frame_layout
.frame_size
+= UNITS_PER_LONG
;
7565 /* No alignment trouble here because f8-f15 are only saved under
7567 cfun_frame_layout
.f8_offset
= (MIN (MIN (cfun_frame_layout
.f0_offset
,
7568 cfun_frame_layout
.f4_offset
),
7569 cfun_frame_layout
.gprs_offset
)
7570 - cfun_frame_layout
.high_fprs
* 8);
7572 cfun_frame_layout
.frame_size
+= cfun_frame_layout
.high_fprs
* 8;
7574 for (i
= 0; i
< 8; i
++)
7575 if (cfun_fpr_bit_p (i
))
7576 cfun_frame_layout
.frame_size
+= 8;
7578 cfun_frame_layout
.frame_size
+= cfun_gprs_save_area_size
;
7580 /* If under 31 bit an odd number of gprs has to be saved we have to adjust
7581 the frame size to sustain 8 byte alignment of stack frames. */
7582 cfun_frame_layout
.frame_size
= ((cfun_frame_layout
.frame_size
+
7583 STACK_BOUNDARY
/ BITS_PER_UNIT
- 1)
7584 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1));
7586 cfun_frame_layout
.frame_size
+= crtl
->outgoing_args_size
;
7590 /* Generate frame layout. Fills in register and frame data for the current
7591 function in cfun->machine. This routine can be called multiple times;
7592 it will re-do the complete frame layout every time. */
7595 s390_init_frame_layout (void)
7597 HOST_WIDE_INT frame_size
;
7599 int clobbered_regs
[16];
7601 /* On S/390 machines, we may need to perform branch splitting, which
7602 will require both base and return address register. We have no
7603 choice but to assume we're going to need them until right at the
7604 end of the machine dependent reorg phase. */
7605 if (!TARGET_CPU_ZARCH
)
7606 cfun
->machine
->split_branches_pending_p
= true;
7610 frame_size
= cfun_frame_layout
.frame_size
;
7612 /* Try to predict whether we'll need the base register. */
7613 base_used
= cfun
->machine
->split_branches_pending_p
7614 || crtl
->uses_const_pool
7615 || (!DISP_IN_RANGE (frame_size
)
7616 && !CONST_OK_FOR_K (frame_size
));
7618 /* Decide which register to use as literal pool base. In small
7619 leaf functions, try to use an unused call-clobbered register
7620 as base register to avoid save/restore overhead. */
7622 cfun
->machine
->base_reg
= NULL_RTX
;
7623 else if (crtl
->is_leaf
&& !df_regs_ever_live_p (5))
7624 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, 5);
7626 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, BASE_REGNUM
);
7628 s390_register_info (clobbered_regs
);
7631 while (frame_size
!= cfun_frame_layout
.frame_size
);
7634 /* Update frame layout. Recompute actual register save data based on
7635 current info and update regs_ever_live for the special registers.
7636 May be called multiple times, but may never cause *more* registers
7637 to be saved than s390_init_frame_layout allocated room for. */
7640 s390_update_frame_layout (void)
7642 int clobbered_regs
[16];
7644 s390_register_info (clobbered_regs
);
7646 df_set_regs_ever_live (BASE_REGNUM
,
7647 clobbered_regs
[BASE_REGNUM
] ? true : false);
7648 df_set_regs_ever_live (RETURN_REGNUM
,
7649 clobbered_regs
[RETURN_REGNUM
] ? true : false);
7650 df_set_regs_ever_live (STACK_POINTER_REGNUM
,
7651 clobbered_regs
[STACK_POINTER_REGNUM
] ? true : false);
7653 if (cfun
->machine
->base_reg
)
7654 df_set_regs_ever_live (REGNO (cfun
->machine
->base_reg
), true);
7657 /* Return true if it is legal to put a value with MODE into REGNO. */
7660 s390_hard_regno_mode_ok (unsigned int regno
, enum machine_mode mode
)
7662 switch (REGNO_REG_CLASS (regno
))
7665 if (REGNO_PAIR_OK (regno
, mode
))
7667 if (mode
== SImode
|| mode
== DImode
)
7670 if (FLOAT_MODE_P (mode
) && GET_MODE_CLASS (mode
) != MODE_VECTOR_FLOAT
)
7675 if (FRAME_REGNO_P (regno
) && mode
== Pmode
)
7680 if (REGNO_PAIR_OK (regno
, mode
))
7683 || (mode
!= TFmode
&& mode
!= TCmode
&& mode
!= TDmode
))
7688 if (GET_MODE_CLASS (mode
) == MODE_CC
)
7692 if (REGNO_PAIR_OK (regno
, mode
))
7694 if (mode
== SImode
|| mode
== Pmode
)
7705 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
7708 s390_hard_regno_rename_ok (unsigned int old_reg
, unsigned int new_reg
)
7710 /* Once we've decided upon a register to use as base register, it must
7711 no longer be used for any other purpose. */
7712 if (cfun
->machine
->base_reg
)
7713 if (REGNO (cfun
->machine
->base_reg
) == old_reg
7714 || REGNO (cfun
->machine
->base_reg
) == new_reg
)
7720 /* Maximum number of registers to represent a value of mode MODE
7721 in a register of class RCLASS. */
7724 s390_class_max_nregs (enum reg_class rclass
, enum machine_mode mode
)
7729 if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7730 return 2 * ((GET_MODE_SIZE (mode
) / 2 + 8 - 1) / 8);
7732 return (GET_MODE_SIZE (mode
) + 8 - 1) / 8;
7734 return (GET_MODE_SIZE (mode
) + 4 - 1) / 4;
7738 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
7741 /* Return true if register FROM can be eliminated via register TO. */
7744 s390_can_eliminate (const int from
, const int to
)
7746 /* On zSeries machines, we have not marked the base register as fixed.
7747 Instead, we have an elimination rule BASE_REGNUM -> BASE_REGNUM.
7748 If a function requires the base register, we say here that this
7749 elimination cannot be performed. This will cause reload to free
7750 up the base register (as if it were fixed). On the other hand,
7751 if the current function does *not* require the base register, we
7752 say here the elimination succeeds, which in turn allows reload
7753 to allocate the base register for any other purpose. */
7754 if (from
== BASE_REGNUM
&& to
== BASE_REGNUM
)
7756 if (TARGET_CPU_ZARCH
)
7758 s390_init_frame_layout ();
7759 return cfun
->machine
->base_reg
== NULL_RTX
;
7765 /* Everything else must point into the stack frame. */
7766 gcc_assert (to
== STACK_POINTER_REGNUM
7767 || to
== HARD_FRAME_POINTER_REGNUM
);
7769 gcc_assert (from
== FRAME_POINTER_REGNUM
7770 || from
== ARG_POINTER_REGNUM
7771 || from
== RETURN_ADDRESS_POINTER_REGNUM
);
7773 /* Make sure we actually saved the return address. */
7774 if (from
== RETURN_ADDRESS_POINTER_REGNUM
)
7775 if (!crtl
->calls_eh_return
7777 && !cfun_frame_layout
.save_return_addr_p
)
7783 /* Return offset between register FROM and TO initially after prolog. */
7786 s390_initial_elimination_offset (int from
, int to
)
7788 HOST_WIDE_INT offset
;
7791 /* ??? Why are we called for non-eliminable pairs? */
7792 if (!s390_can_eliminate (from
, to
))
7797 case FRAME_POINTER_REGNUM
:
7798 offset
= (get_frame_size()
7799 + STACK_POINTER_OFFSET
7800 + crtl
->outgoing_args_size
);
7803 case ARG_POINTER_REGNUM
:
7804 s390_init_frame_layout ();
7805 offset
= cfun_frame_layout
.frame_size
+ STACK_POINTER_OFFSET
;
7808 case RETURN_ADDRESS_POINTER_REGNUM
:
7809 s390_init_frame_layout ();
7810 index
= RETURN_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
;
7811 gcc_assert (index
>= 0);
7812 offset
= cfun_frame_layout
.frame_size
+ cfun_frame_layout
.gprs_offset
;
7813 offset
+= index
* UNITS_PER_LONG
;
7827 /* Emit insn to save fpr REGNUM at offset OFFSET relative
7828 to register BASE. Return generated insn. */
7831 save_fpr (rtx base
, int offset
, int regnum
)
7834 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7836 if (regnum
>= 16 && regnum
<= (16 + FP_ARG_NUM_REG
))
7837 set_mem_alias_set (addr
, get_varargs_alias_set ());
7839 set_mem_alias_set (addr
, get_frame_alias_set ());
7841 return emit_move_insn (addr
, gen_rtx_REG (DFmode
, regnum
));
7844 /* Emit insn to restore fpr REGNUM from offset OFFSET relative
7845 to register BASE. Return generated insn. */
7848 restore_fpr (rtx base
, int offset
, int regnum
)
7851 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7852 set_mem_alias_set (addr
, get_frame_alias_set ());
7854 return emit_move_insn (gen_rtx_REG (DFmode
, regnum
), addr
);
7857 /* Return true if REGNO is a global register, but not one
7858 of the special ones that need to be saved/restored in anyway. */
7861 global_not_special_regno_p (int regno
)
7863 return (global_regs
[regno
]
7864 /* These registers are special and need to be
7865 restored in any case. */
7866 && !(regno
== STACK_POINTER_REGNUM
7867 || regno
== RETURN_REGNUM
7868 || regno
== BASE_REGNUM
7869 || (flag_pic
&& regno
== (int)PIC_OFFSET_TABLE_REGNUM
)));
7872 /* Generate insn to save registers FIRST to LAST into
7873 the register save area located at offset OFFSET
7874 relative to register BASE. */
7877 save_gprs (rtx base
, int offset
, int first
, int last
)
7879 rtx addr
, insn
, note
;
7882 addr
= plus_constant (Pmode
, base
, offset
);
7883 addr
= gen_rtx_MEM (Pmode
, addr
);
7885 set_mem_alias_set (addr
, get_frame_alias_set ());
7887 /* Special-case single register. */
7891 insn
= gen_movdi (addr
, gen_rtx_REG (Pmode
, first
));
7893 insn
= gen_movsi (addr
, gen_rtx_REG (Pmode
, first
));
7895 if (!global_not_special_regno_p (first
))
7896 RTX_FRAME_RELATED_P (insn
) = 1;
7901 insn
= gen_store_multiple (addr
,
7902 gen_rtx_REG (Pmode
, first
),
7903 GEN_INT (last
- first
+ 1));
7905 if (first
<= 6 && cfun
->stdarg
)
7906 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
7908 rtx mem
= XEXP (XVECEXP (PATTERN (insn
), 0, i
), 0);
7911 set_mem_alias_set (mem
, get_varargs_alias_set ());
7914 /* We need to set the FRAME_RELATED flag on all SETs
7915 inside the store-multiple pattern.
7917 However, we must not emit DWARF records for registers 2..5
7918 if they are stored for use by variable arguments ...
7920 ??? Unfortunately, it is not enough to simply not the
7921 FRAME_RELATED flags for those SETs, because the first SET
7922 of the PARALLEL is always treated as if it had the flag
7923 set, even if it does not. Therefore we emit a new pattern
7924 without those registers as REG_FRAME_RELATED_EXPR note. */
7926 if (first
>= 6 && !global_not_special_regno_p (first
))
7928 rtx pat
= PATTERN (insn
);
7930 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
7931 if (GET_CODE (XVECEXP (pat
, 0, i
)) == SET
7932 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (pat
,
7934 RTX_FRAME_RELATED_P (XVECEXP (pat
, 0, i
)) = 1;
7936 RTX_FRAME_RELATED_P (insn
) = 1;
7942 for (start
= first
>= 6 ? first
: 6; start
<= last
; start
++)
7943 if (!global_not_special_regno_p (start
))
7949 addr
= plus_constant (Pmode
, base
,
7950 offset
+ (start
- first
) * UNITS_PER_LONG
);
7951 note
= gen_store_multiple (gen_rtx_MEM (Pmode
, addr
),
7952 gen_rtx_REG (Pmode
, start
),
7953 GEN_INT (last
- start
+ 1));
7954 note
= PATTERN (note
);
7956 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, note
);
7958 for (i
= 0; i
< XVECLEN (note
, 0); i
++)
7959 if (GET_CODE (XVECEXP (note
, 0, i
)) == SET
7960 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (note
,
7962 RTX_FRAME_RELATED_P (XVECEXP (note
, 0, i
)) = 1;
7964 RTX_FRAME_RELATED_P (insn
) = 1;
7970 /* Generate insn to restore registers FIRST to LAST from
7971 the register save area located at offset OFFSET
7972 relative to register BASE. */
7975 restore_gprs (rtx base
, int offset
, int first
, int last
)
7979 addr
= plus_constant (Pmode
, base
, offset
);
7980 addr
= gen_rtx_MEM (Pmode
, addr
);
7981 set_mem_alias_set (addr
, get_frame_alias_set ());
7983 /* Special-case single register. */
7987 insn
= gen_movdi (gen_rtx_REG (Pmode
, first
), addr
);
7989 insn
= gen_movsi (gen_rtx_REG (Pmode
, first
), addr
);
7994 insn
= gen_load_multiple (gen_rtx_REG (Pmode
, first
),
7996 GEN_INT (last
- first
+ 1));
8000 /* Return insn sequence to load the GOT register. */
8002 static GTY(()) rtx got_symbol
;
8004 s390_load_got (void)
8008 /* We cannot use pic_offset_table_rtx here since we use this
8009 function also for non-pic if __tls_get_offset is called and in
8010 that case PIC_OFFSET_TABLE_REGNUM as well as pic_offset_table_rtx
8012 rtx got_rtx
= gen_rtx_REG (Pmode
, 12);
8016 got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
8017 SYMBOL_REF_FLAGS (got_symbol
) = SYMBOL_FLAG_LOCAL
;
8022 if (TARGET_CPU_ZARCH
)
8024 emit_move_insn (got_rtx
, got_symbol
);
8030 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, got_symbol
),
8031 UNSPEC_LTREL_OFFSET
);
8032 offset
= gen_rtx_CONST (Pmode
, offset
);
8033 offset
= force_const_mem (Pmode
, offset
);
8035 emit_move_insn (got_rtx
, offset
);
8037 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (offset
, 0)),
8039 offset
= gen_rtx_PLUS (Pmode
, got_rtx
, offset
);
8041 emit_move_insn (got_rtx
, offset
);
8044 insns
= get_insns ();
8049 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
8050 and the change to the stack pointer. */
8053 s390_emit_stack_tie (void)
8055 rtx mem
= gen_frame_mem (BLKmode
,
8056 gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
));
8058 emit_insn (gen_stack_tie (mem
));
8061 /* Expand the prologue into a bunch of separate insns. */
8064 s390_emit_prologue (void)
8072 /* Complete frame layout. */
8074 s390_update_frame_layout ();
8076 /* Annotate all constant pool references to let the scheduler know
8077 they implicitly use the base register. */
8079 push_topmost_sequence ();
8081 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
8084 annotate_constant_pool_refs (&PATTERN (insn
));
8085 df_insn_rescan (insn
);
8088 pop_topmost_sequence ();
8090 /* Choose best register to use for temp use within prologue.
8091 See below for why TPF must use the register 1. */
8093 if (!has_hard_reg_initial_val (Pmode
, RETURN_REGNUM
)
8095 && !TARGET_TPF_PROFILING
)
8096 temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8098 temp_reg
= gen_rtx_REG (Pmode
, 1);
8100 /* Save call saved gprs. */
8101 if (cfun_frame_layout
.first_save_gpr
!= -1)
8103 insn
= save_gprs (stack_pointer_rtx
,
8104 cfun_frame_layout
.gprs_offset
+
8105 UNITS_PER_LONG
* (cfun_frame_layout
.first_save_gpr
8106 - cfun_frame_layout
.first_save_gpr_slot
),
8107 cfun_frame_layout
.first_save_gpr
,
8108 cfun_frame_layout
.last_save_gpr
);
8112 /* Dummy insn to mark literal pool slot. */
8114 if (cfun
->machine
->base_reg
)
8115 emit_insn (gen_main_pool (cfun
->machine
->base_reg
));
8117 offset
= cfun_frame_layout
.f0_offset
;
8119 /* Save f0 and f2. */
8120 for (i
= 0; i
< 2; i
++)
8122 if (cfun_fpr_bit_p (i
))
8124 save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8127 else if (!TARGET_PACKED_STACK
)
8131 /* Save f4 and f6. */
8132 offset
= cfun_frame_layout
.f4_offset
;
8133 for (i
= 2; i
< 4; i
++)
8135 if (cfun_fpr_bit_p (i
))
8137 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8140 /* If f4 and f6 are call clobbered they are saved due to stdargs and
8141 therefore are not frame related. */
8142 if (!call_really_used_regs
[i
+ 16])
8143 RTX_FRAME_RELATED_P (insn
) = 1;
8145 else if (!TARGET_PACKED_STACK
)
8149 if (TARGET_PACKED_STACK
8150 && cfun_save_high_fprs_p
8151 && cfun_frame_layout
.f8_offset
+ cfun_frame_layout
.high_fprs
* 8 > 0)
8153 offset
= (cfun_frame_layout
.f8_offset
8154 + (cfun_frame_layout
.high_fprs
- 1) * 8);
8156 for (i
= 15; i
> 7 && offset
>= 0; i
--)
8157 if (cfun_fpr_bit_p (i
))
8159 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8161 RTX_FRAME_RELATED_P (insn
) = 1;
8164 if (offset
>= cfun_frame_layout
.f8_offset
)
8168 if (!TARGET_PACKED_STACK
)
8169 next_fpr
= cfun_save_high_fprs_p
? 31 : 0;
8171 if (flag_stack_usage_info
)
8172 current_function_static_stack_size
= cfun_frame_layout
.frame_size
;
8174 /* Decrement stack pointer. */
8176 if (cfun_frame_layout
.frame_size
> 0)
8178 rtx frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8181 if (s390_stack_size
)
8183 HOST_WIDE_INT stack_guard
;
8185 if (s390_stack_guard
)
8186 stack_guard
= s390_stack_guard
;
8189 /* If no value for stack guard is provided the smallest power of 2
8190 larger than the current frame size is chosen. */
8192 while (stack_guard
< cfun_frame_layout
.frame_size
)
8196 if (cfun_frame_layout
.frame_size
>= s390_stack_size
)
8198 warning (0, "frame size of function %qs is %wd"
8199 " bytes exceeding user provided stack limit of "
8201 "An unconditional trap is added.",
8202 current_function_name(), cfun_frame_layout
.frame_size
,
8204 emit_insn (gen_trap ());
8208 /* stack_guard has to be smaller than s390_stack_size.
8209 Otherwise we would emit an AND with zero which would
8210 not match the test under mask pattern. */
8211 if (stack_guard
>= s390_stack_size
)
8213 warning (0, "frame size of function %qs is %wd"
8214 " bytes which is more than half the stack size. "
8215 "The dynamic check would not be reliable. "
8216 "No check emitted for this function.",
8217 current_function_name(),
8218 cfun_frame_layout
.frame_size
);
8222 HOST_WIDE_INT stack_check_mask
= ((s390_stack_size
- 1)
8223 & ~(stack_guard
- 1));
8225 rtx t
= gen_rtx_AND (Pmode
, stack_pointer_rtx
,
8226 GEN_INT (stack_check_mask
));
8228 emit_insn (gen_ctrapdi4 (gen_rtx_EQ (VOIDmode
,
8230 t
, const0_rtx
, const0_rtx
));
8232 emit_insn (gen_ctrapsi4 (gen_rtx_EQ (VOIDmode
,
8234 t
, const0_rtx
, const0_rtx
));
8239 if (s390_warn_framesize
> 0
8240 && cfun_frame_layout
.frame_size
>= s390_warn_framesize
)
8241 warning (0, "frame size of %qs is %wd bytes",
8242 current_function_name (), cfun_frame_layout
.frame_size
);
8244 if (s390_warn_dynamicstack_p
&& cfun
->calls_alloca
)
8245 warning (0, "%qs uses dynamic stack allocation", current_function_name ());
8247 /* Save incoming stack pointer into temp reg. */
8248 if (TARGET_BACKCHAIN
|| next_fpr
)
8249 insn
= emit_insn (gen_move_insn (temp_reg
, stack_pointer_rtx
));
8251 /* Subtract frame size from stack pointer. */
8253 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8255 insn
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8256 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8258 insn
= emit_insn (insn
);
8262 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8263 frame_off
= force_const_mem (Pmode
, frame_off
);
8265 insn
= emit_insn (gen_add2_insn (stack_pointer_rtx
, frame_off
));
8266 annotate_constant_pool_refs (&PATTERN (insn
));
8269 RTX_FRAME_RELATED_P (insn
) = 1;
8270 real_frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8271 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8272 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8273 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8276 /* Set backchain. */
8278 if (TARGET_BACKCHAIN
)
8280 if (cfun_frame_layout
.backchain_offset
)
8281 addr
= gen_rtx_MEM (Pmode
,
8282 plus_constant (Pmode
, stack_pointer_rtx
,
8283 cfun_frame_layout
.backchain_offset
));
8285 addr
= gen_rtx_MEM (Pmode
, stack_pointer_rtx
);
8286 set_mem_alias_set (addr
, get_frame_alias_set ());
8287 insn
= emit_insn (gen_move_insn (addr
, temp_reg
));
8290 /* If we support non-call exceptions (e.g. for Java),
8291 we need to make sure the backchain pointer is set up
8292 before any possibly trapping memory access. */
8293 if (TARGET_BACKCHAIN
&& cfun
->can_throw_non_call_exceptions
)
8295 addr
= gen_rtx_MEM (BLKmode
, gen_rtx_SCRATCH (VOIDmode
));
8296 emit_clobber (addr
);
8300 /* Save fprs 8 - 15 (64 bit ABI). */
8302 if (cfun_save_high_fprs_p
&& next_fpr
)
8304 /* If the stack might be accessed through a different register
8305 we have to make sure that the stack pointer decrement is not
8306 moved below the use of the stack slots. */
8307 s390_emit_stack_tie ();
8309 insn
= emit_insn (gen_add2_insn (temp_reg
,
8310 GEN_INT (cfun_frame_layout
.f8_offset
)));
8314 for (i
= 24; i
<= next_fpr
; i
++)
8315 if (cfun_fpr_bit_p (i
- 16))
8317 rtx addr
= plus_constant (Pmode
, stack_pointer_rtx
,
8318 cfun_frame_layout
.frame_size
8319 + cfun_frame_layout
.f8_offset
8322 insn
= save_fpr (temp_reg
, offset
, i
);
8324 RTX_FRAME_RELATED_P (insn
) = 1;
8325 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8326 gen_rtx_SET (VOIDmode
,
8327 gen_rtx_MEM (DFmode
, addr
),
8328 gen_rtx_REG (DFmode
, i
)));
8332 /* Set frame pointer, if needed. */
8334 if (frame_pointer_needed
)
8336 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
8337 RTX_FRAME_RELATED_P (insn
) = 1;
8340 /* Set up got pointer, if needed. */
8342 if (flag_pic
&& df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
8344 rtx insns
= s390_load_got ();
8346 for (insn
= insns
; insn
; insn
= NEXT_INSN (insn
))
8347 annotate_constant_pool_refs (&PATTERN (insn
));
8352 if (TARGET_TPF_PROFILING
)
8354 /* Generate a BAS instruction to serve as a function
8355 entry intercept to facilitate the use of tracing
8356 algorithms located at the branch target. */
8357 emit_insn (gen_prologue_tpf ());
8359 /* Emit a blockage here so that all code
8360 lies between the profiling mechanisms. */
8361 emit_insn (gen_blockage ());
8365 /* Expand the epilogue into a bunch of separate insns. */
8368 s390_emit_epilogue (bool sibcall
)
8370 rtx frame_pointer
, return_reg
, cfa_restores
= NULL_RTX
;
8371 int area_bottom
, area_top
, offset
= 0;
8376 if (TARGET_TPF_PROFILING
)
8379 /* Generate a BAS instruction to serve as a function
8380 entry intercept to facilitate the use of tracing
8381 algorithms located at the branch target. */
8383 /* Emit a blockage here so that all code
8384 lies between the profiling mechanisms. */
8385 emit_insn (gen_blockage ());
8387 emit_insn (gen_epilogue_tpf ());
8390 /* Check whether to use frame or stack pointer for restore. */
8392 frame_pointer
= (frame_pointer_needed
8393 ? hard_frame_pointer_rtx
: stack_pointer_rtx
);
8395 s390_frame_area (&area_bottom
, &area_top
);
8397 /* Check whether we can access the register save area.
8398 If not, increment the frame pointer as required. */
8400 if (area_top
<= area_bottom
)
8402 /* Nothing to restore. */
8404 else if (DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_bottom
)
8405 && DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_top
- 1))
8407 /* Area is in range. */
8408 offset
= cfun_frame_layout
.frame_size
;
8412 rtx insn
, frame_off
, cfa
;
8414 offset
= area_bottom
< 0 ? -area_bottom
: 0;
8415 frame_off
= GEN_INT (cfun_frame_layout
.frame_size
- offset
);
8417 cfa
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8418 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8419 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8421 insn
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8422 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8423 insn
= emit_insn (insn
);
8427 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8428 frame_off
= force_const_mem (Pmode
, frame_off
);
8430 insn
= emit_insn (gen_add2_insn (frame_pointer
, frame_off
));
8431 annotate_constant_pool_refs (&PATTERN (insn
));
8433 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, cfa
);
8434 RTX_FRAME_RELATED_P (insn
) = 1;
8437 /* Restore call saved fprs. */
8441 if (cfun_save_high_fprs_p
)
8443 next_offset
= cfun_frame_layout
.f8_offset
;
8444 for (i
= 24; i
< 32; i
++)
8446 if (cfun_fpr_bit_p (i
- 16))
8448 restore_fpr (frame_pointer
,
8449 offset
+ next_offset
, i
);
8451 = alloc_reg_note (REG_CFA_RESTORE
,
8452 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8461 next_offset
= cfun_frame_layout
.f4_offset
;
8462 for (i
= 18; i
< 20; i
++)
8464 if (cfun_fpr_bit_p (i
- 16))
8466 restore_fpr (frame_pointer
,
8467 offset
+ next_offset
, i
);
8469 = alloc_reg_note (REG_CFA_RESTORE
,
8470 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8473 else if (!TARGET_PACKED_STACK
)
8479 /* Return register. */
8481 return_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8483 /* Restore call saved gprs. */
8485 if (cfun_frame_layout
.first_restore_gpr
!= -1)
8490 /* Check for global register and save them
8491 to stack location from where they get restored. */
8493 for (i
= cfun_frame_layout
.first_restore_gpr
;
8494 i
<= cfun_frame_layout
.last_restore_gpr
;
8497 if (global_not_special_regno_p (i
))
8499 addr
= plus_constant (Pmode
, frame_pointer
,
8500 offset
+ cfun_frame_layout
.gprs_offset
8501 + (i
- cfun_frame_layout
.first_save_gpr_slot
)
8503 addr
= gen_rtx_MEM (Pmode
, addr
);
8504 set_mem_alias_set (addr
, get_frame_alias_set ());
8505 emit_move_insn (addr
, gen_rtx_REG (Pmode
, i
));
8509 = alloc_reg_note (REG_CFA_RESTORE
,
8510 gen_rtx_REG (Pmode
, i
), cfa_restores
);
8515 /* Fetch return address from stack before load multiple,
8516 this will do good for scheduling. */
8518 if (cfun_frame_layout
.save_return_addr_p
8519 || (cfun_frame_layout
.first_restore_gpr
< BASE_REGNUM
8520 && cfun_frame_layout
.last_restore_gpr
> RETURN_REGNUM
))
8522 int return_regnum
= find_unused_clobbered_reg();
8525 return_reg
= gen_rtx_REG (Pmode
, return_regnum
);
8527 addr
= plus_constant (Pmode
, frame_pointer
,
8528 offset
+ cfun_frame_layout
.gprs_offset
8530 - cfun_frame_layout
.first_save_gpr_slot
)
8532 addr
= gen_rtx_MEM (Pmode
, addr
);
8533 set_mem_alias_set (addr
, get_frame_alias_set ());
8534 emit_move_insn (return_reg
, addr
);
8538 insn
= restore_gprs (frame_pointer
,
8539 offset
+ cfun_frame_layout
.gprs_offset
8540 + (cfun_frame_layout
.first_restore_gpr
8541 - cfun_frame_layout
.first_save_gpr_slot
)
8543 cfun_frame_layout
.first_restore_gpr
,
8544 cfun_frame_layout
.last_restore_gpr
);
8545 insn
= emit_insn (insn
);
8546 REG_NOTES (insn
) = cfa_restores
;
8547 add_reg_note (insn
, REG_CFA_DEF_CFA
,
8548 plus_constant (Pmode
, stack_pointer_rtx
,
8549 STACK_POINTER_OFFSET
));
8550 RTX_FRAME_RELATED_P (insn
) = 1;
8556 /* Return to caller. */
8558 p
= rtvec_alloc (2);
8560 RTVEC_ELT (p
, 0) = ret_rtx
;
8561 RTVEC_ELT (p
, 1) = gen_rtx_USE (VOIDmode
, return_reg
);
8562 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
8567 /* Return the size in bytes of a function argument of
8568 type TYPE and/or mode MODE. At least one of TYPE or
8569 MODE must be specified. */
8572 s390_function_arg_size (enum machine_mode mode
, const_tree type
)
8575 return int_size_in_bytes (type
);
8577 /* No type info available for some library calls ... */
8578 if (mode
!= BLKmode
)
8579 return GET_MODE_SIZE (mode
);
8581 /* If we have neither type nor mode, abort */
8585 /* Return true if a function argument of type TYPE and mode MODE
8586 is to be passed in a floating-point register, if available. */
8589 s390_function_arg_float (enum machine_mode mode
, const_tree type
)
8591 int size
= s390_function_arg_size (mode
, type
);
8595 /* Soft-float changes the ABI: no floating-point registers are used. */
8596 if (TARGET_SOFT_FLOAT
)
8599 /* No type info available for some library calls ... */
8601 return mode
== SFmode
|| mode
== DFmode
|| mode
== SDmode
|| mode
== DDmode
;
8603 /* The ABI says that record types with a single member are treated
8604 just like that member would be. */
8605 while (TREE_CODE (type
) == RECORD_TYPE
)
8607 tree field
, single
= NULL_TREE
;
8609 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
8611 if (TREE_CODE (field
) != FIELD_DECL
)
8614 if (single
== NULL_TREE
)
8615 single
= TREE_TYPE (field
);
8620 if (single
== NULL_TREE
)
8626 return TREE_CODE (type
) == REAL_TYPE
;
8629 /* Return true if a function argument of type TYPE and mode MODE
8630 is to be passed in an integer register, or a pair of integer
8631 registers, if available. */
8634 s390_function_arg_integer (enum machine_mode mode
, const_tree type
)
8636 int size
= s390_function_arg_size (mode
, type
);
8640 /* No type info available for some library calls ... */
8642 return GET_MODE_CLASS (mode
) == MODE_INT
8643 || (TARGET_SOFT_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
));
8645 /* We accept small integral (and similar) types. */
8646 if (INTEGRAL_TYPE_P (type
)
8647 || POINTER_TYPE_P (type
)
8648 || TREE_CODE (type
) == NULLPTR_TYPE
8649 || TREE_CODE (type
) == OFFSET_TYPE
8650 || (TARGET_SOFT_FLOAT
&& TREE_CODE (type
) == REAL_TYPE
))
8653 /* We also accept structs of size 1, 2, 4, 8 that are not
8654 passed in floating-point registers. */
8655 if (AGGREGATE_TYPE_P (type
)
8656 && exact_log2 (size
) >= 0
8657 && !s390_function_arg_float (mode
, type
))
8663 /* Return 1 if a function argument of type TYPE and mode MODE
8664 is to be passed by reference. The ABI specifies that only
8665 structures of size 1, 2, 4, or 8 bytes are passed by value,
8666 all other structures (and complex numbers) are passed by
8670 s390_pass_by_reference (cumulative_args_t ca ATTRIBUTE_UNUSED
,
8671 enum machine_mode mode
, const_tree type
,
8672 bool named ATTRIBUTE_UNUSED
)
8674 int size
= s390_function_arg_size (mode
, type
);
8680 if (AGGREGATE_TYPE_P (type
) && exact_log2 (size
) < 0)
8683 if (TREE_CODE (type
) == COMPLEX_TYPE
8684 || TREE_CODE (type
) == VECTOR_TYPE
)
8691 /* Update the data in CUM to advance over an argument of mode MODE and
8692 data type TYPE. (TYPE is null for libcalls where that information
8693 may not be available.). The boolean NAMED specifies whether the
8694 argument is a named argument (as opposed to an unnamed argument
8695 matching an ellipsis). */
8698 s390_function_arg_advance (cumulative_args_t cum_v
, enum machine_mode mode
,
8699 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8701 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8703 if (s390_function_arg_float (mode
, type
))
8707 else if (s390_function_arg_integer (mode
, type
))
8709 int size
= s390_function_arg_size (mode
, type
);
8710 cum
->gprs
+= ((size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
);
8716 /* Define where to put the arguments to a function.
8717 Value is zero to push the argument on the stack,
8718 or a hard register in which to store the argument.
8720 MODE is the argument's machine mode.
8721 TYPE is the data type of the argument (as a tree).
8722 This is null for libcalls where that information may
8724 CUM is a variable of type CUMULATIVE_ARGS which gives info about
8725 the preceding args and about the function being called.
8726 NAMED is nonzero if this argument is a named parameter
8727 (otherwise it is an extra parameter matching an ellipsis).
8729 On S/390, we use general purpose registers 2 through 6 to
8730 pass integer, pointer, and certain structure arguments, and
8731 floating point registers 0 and 2 (0, 2, 4, and 6 on 64-bit)
8732 to pass floating point arguments. All remaining arguments
8733 are pushed to the stack. */
8736 s390_function_arg (cumulative_args_t cum_v
, enum machine_mode mode
,
8737 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8739 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8741 if (s390_function_arg_float (mode
, type
))
8743 if (cum
->fprs
+ 1 > FP_ARG_NUM_REG
)
8746 return gen_rtx_REG (mode
, cum
->fprs
+ 16);
8748 else if (s390_function_arg_integer (mode
, type
))
8750 int size
= s390_function_arg_size (mode
, type
);
8751 int n_gprs
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8753 if (cum
->gprs
+ n_gprs
> GP_ARG_NUM_REG
)
8755 else if (n_gprs
== 1 || UNITS_PER_WORD
== UNITS_PER_LONG
)
8756 return gen_rtx_REG (mode
, cum
->gprs
+ 2);
8757 else if (n_gprs
== 2)
8759 rtvec p
= rtvec_alloc (2);
8762 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 2),
8765 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 3),
8768 return gen_rtx_PARALLEL (mode
, p
);
8772 /* After the real arguments, expand_call calls us once again
8773 with a void_type_node type. Whatever we return here is
8774 passed as operand 2 to the call expanders.
8776 We don't need this feature ... */
8777 else if (type
== void_type_node
)
8783 /* Return true if return values of type TYPE should be returned
8784 in a memory buffer whose address is passed by the caller as
8785 hidden first argument. */
8788 s390_return_in_memory (const_tree type
, const_tree fundecl ATTRIBUTE_UNUSED
)
8790 /* We accept small integral (and similar) types. */
8791 if (INTEGRAL_TYPE_P (type
)
8792 || POINTER_TYPE_P (type
)
8793 || TREE_CODE (type
) == OFFSET_TYPE
8794 || TREE_CODE (type
) == REAL_TYPE
)
8795 return int_size_in_bytes (type
) > 8;
8797 /* Aggregates and similar constructs are always returned
8799 if (AGGREGATE_TYPE_P (type
)
8800 || TREE_CODE (type
) == COMPLEX_TYPE
8801 || TREE_CODE (type
) == VECTOR_TYPE
)
8804 /* ??? We get called on all sorts of random stuff from
8805 aggregate_value_p. We can't abort, but it's not clear
8806 what's safe to return. Pretend it's a struct I guess. */
8810 /* Function arguments and return values are promoted to word size. */
8812 static enum machine_mode
8813 s390_promote_function_mode (const_tree type
, enum machine_mode mode
,
8815 const_tree fntype ATTRIBUTE_UNUSED
,
8816 int for_return ATTRIBUTE_UNUSED
)
8818 if (INTEGRAL_MODE_P (mode
)
8819 && GET_MODE_SIZE (mode
) < UNITS_PER_LONG
)
8821 if (type
!= NULL_TREE
&& POINTER_TYPE_P (type
))
8822 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
8829 /* Define where to return a (scalar) value of type RET_TYPE.
8830 If RET_TYPE is null, define where to return a (scalar)
8831 value of mode MODE from a libcall. */
8834 s390_function_and_libcall_value (enum machine_mode mode
,
8835 const_tree ret_type
,
8836 const_tree fntype_or_decl
,
8837 bool outgoing ATTRIBUTE_UNUSED
)
8839 /* For normal functions perform the promotion as
8840 promote_function_mode would do. */
8843 int unsignedp
= TYPE_UNSIGNED (ret_type
);
8844 mode
= promote_function_mode (ret_type
, mode
, &unsignedp
,
8848 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
|| SCALAR_FLOAT_MODE_P (mode
));
8849 gcc_assert (GET_MODE_SIZE (mode
) <= 8);
8851 if (TARGET_HARD_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
))
8852 return gen_rtx_REG (mode
, 16);
8853 else if (GET_MODE_SIZE (mode
) <= UNITS_PER_LONG
8854 || UNITS_PER_LONG
== UNITS_PER_WORD
)
8855 return gen_rtx_REG (mode
, 2);
8856 else if (GET_MODE_SIZE (mode
) == 2 * UNITS_PER_LONG
)
8858 /* This case is triggered when returning a 64 bit value with
8859 -m31 -mzarch. Although the value would fit into a single
8860 register it has to be forced into a 32 bit register pair in
8861 order to match the ABI. */
8862 rtvec p
= rtvec_alloc (2);
8865 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 2), const0_rtx
);
8867 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 3), GEN_INT (4));
8869 return gen_rtx_PARALLEL (mode
, p
);
8875 /* Define where to return a scalar return value of type RET_TYPE. */
8878 s390_function_value (const_tree ret_type
, const_tree fn_decl_or_type
,
8881 return s390_function_and_libcall_value (TYPE_MODE (ret_type
), ret_type
,
8882 fn_decl_or_type
, outgoing
);
8885 /* Define where to return a scalar libcall return value of mode
8889 s390_libcall_value (enum machine_mode mode
, const_rtx fun ATTRIBUTE_UNUSED
)
8891 return s390_function_and_libcall_value (mode
, NULL_TREE
,
8896 /* Create and return the va_list datatype.
8898 On S/390, va_list is an array type equivalent to
8900 typedef struct __va_list_tag
8904 void *__overflow_arg_area;
8905 void *__reg_save_area;
8908 where __gpr and __fpr hold the number of general purpose
8909 or floating point arguments used up to now, respectively,
8910 __overflow_arg_area points to the stack location of the
8911 next argument passed on the stack, and __reg_save_area
8912 always points to the start of the register area in the
8913 call frame of the current function. The function prologue
8914 saves all registers used for argument passing into this
8915 area if the function uses variable arguments. */
8918 s390_build_builtin_va_list (void)
8920 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
8922 record
= lang_hooks
.types
.make_type (RECORD_TYPE
);
8925 build_decl (BUILTINS_LOCATION
,
8926 TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
8928 f_gpr
= build_decl (BUILTINS_LOCATION
,
8929 FIELD_DECL
, get_identifier ("__gpr"),
8930 long_integer_type_node
);
8931 f_fpr
= build_decl (BUILTINS_LOCATION
,
8932 FIELD_DECL
, get_identifier ("__fpr"),
8933 long_integer_type_node
);
8934 f_ovf
= build_decl (BUILTINS_LOCATION
,
8935 FIELD_DECL
, get_identifier ("__overflow_arg_area"),
8937 f_sav
= build_decl (BUILTINS_LOCATION
,
8938 FIELD_DECL
, get_identifier ("__reg_save_area"),
8941 va_list_gpr_counter_field
= f_gpr
;
8942 va_list_fpr_counter_field
= f_fpr
;
8944 DECL_FIELD_CONTEXT (f_gpr
) = record
;
8945 DECL_FIELD_CONTEXT (f_fpr
) = record
;
8946 DECL_FIELD_CONTEXT (f_ovf
) = record
;
8947 DECL_FIELD_CONTEXT (f_sav
) = record
;
8949 TYPE_STUB_DECL (record
) = type_decl
;
8950 TYPE_NAME (record
) = type_decl
;
8951 TYPE_FIELDS (record
) = f_gpr
;
8952 DECL_CHAIN (f_gpr
) = f_fpr
;
8953 DECL_CHAIN (f_fpr
) = f_ovf
;
8954 DECL_CHAIN (f_ovf
) = f_sav
;
8956 layout_type (record
);
8958 /* The correct type is an array type of one element. */
8959 return build_array_type (record
, build_index_type (size_zero_node
));
8962 /* Implement va_start by filling the va_list structure VALIST.
8963 STDARG_P is always true, and ignored.
8964 NEXTARG points to the first anonymous stack argument.
8966 The following global variables are used to initialize
8967 the va_list structure:
8970 holds number of gprs and fprs used for named arguments.
8971 crtl->args.arg_offset_rtx:
8972 holds the offset of the first anonymous stack argument
8973 (relative to the virtual arg pointer). */
8976 s390_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
8978 HOST_WIDE_INT n_gpr
, n_fpr
;
8980 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8981 tree gpr
, fpr
, ovf
, sav
, t
;
8983 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8984 f_fpr
= DECL_CHAIN (f_gpr
);
8985 f_ovf
= DECL_CHAIN (f_fpr
);
8986 f_sav
= DECL_CHAIN (f_ovf
);
8988 valist
= build_simple_mem_ref (valist
);
8989 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8990 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8991 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8992 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8994 /* Count number of gp and fp argument registers used. */
8996 n_gpr
= crtl
->args
.info
.gprs
;
8997 n_fpr
= crtl
->args
.info
.fprs
;
8999 if (cfun
->va_list_gpr_size
)
9001 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gpr
), gpr
,
9002 build_int_cst (NULL_TREE
, n_gpr
));
9003 TREE_SIDE_EFFECTS (t
) = 1;
9004 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
9007 if (cfun
->va_list_fpr_size
)
9009 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fpr
), fpr
,
9010 build_int_cst (NULL_TREE
, n_fpr
));
9011 TREE_SIDE_EFFECTS (t
) = 1;
9012 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
9015 /* Find the overflow area. */
9016 if (n_gpr
+ cfun
->va_list_gpr_size
> GP_ARG_NUM_REG
9017 || n_fpr
+ cfun
->va_list_fpr_size
> FP_ARG_NUM_REG
)
9019 t
= make_tree (TREE_TYPE (ovf
), virtual_incoming_args_rtx
);
9021 off
= INTVAL (crtl
->args
.arg_offset_rtx
);
9022 off
= off
< 0 ? 0 : off
;
9023 if (TARGET_DEBUG_ARG
)
9024 fprintf (stderr
, "va_start: n_gpr = %d, n_fpr = %d off %d\n",
9025 (int)n_gpr
, (int)n_fpr
, off
);
9027 t
= fold_build_pointer_plus_hwi (t
, off
);
9029 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovf
), ovf
, t
);
9030 TREE_SIDE_EFFECTS (t
) = 1;
9031 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
9034 /* Find the register save area. */
9035 if ((cfun
->va_list_gpr_size
&& n_gpr
< GP_ARG_NUM_REG
)
9036 || (cfun
->va_list_fpr_size
&& n_fpr
< FP_ARG_NUM_REG
))
9038 t
= make_tree (TREE_TYPE (sav
), return_address_pointer_rtx
);
9039 t
= fold_build_pointer_plus_hwi (t
, -RETURN_REGNUM
* UNITS_PER_LONG
);
9041 t
= build2 (MODIFY_EXPR
, TREE_TYPE (sav
), sav
, t
);
9042 TREE_SIDE_EFFECTS (t
) = 1;
9043 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
9047 /* Implement va_arg by updating the va_list structure
9048 VALIST as required to retrieve an argument of type
9049 TYPE, and returning that argument.
9051 Generates code equivalent to:
9053 if (integral value) {
9054 if (size <= 4 && args.gpr < 5 ||
9055 size > 4 && args.gpr < 4 )
9056 ret = args.reg_save_area[args.gpr+8]
9058 ret = *args.overflow_arg_area++;
9059 } else if (float value) {
9061 ret = args.reg_save_area[args.fpr+64]
9063 ret = *args.overflow_arg_area++;
9064 } else if (aggregate value) {
9066 ret = *args.reg_save_area[args.gpr]
9068 ret = **args.overflow_arg_area++;
9072 s390_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
9073 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
9075 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
9076 tree gpr
, fpr
, ovf
, sav
, reg
, t
, u
;
9077 int indirect_p
, size
, n_reg
, sav_ofs
, sav_scale
, max_reg
;
9078 tree lab_false
, lab_over
, addr
;
9080 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
9081 f_fpr
= DECL_CHAIN (f_gpr
);
9082 f_ovf
= DECL_CHAIN (f_fpr
);
9083 f_sav
= DECL_CHAIN (f_ovf
);
9085 valist
= build_va_arg_indirect_ref (valist
);
9086 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
9087 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
9088 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
9090 /* The tree for args* cannot be shared between gpr/fpr and ovf since
9091 both appear on a lhs. */
9092 valist
= unshare_expr (valist
);
9093 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
9095 size
= int_size_in_bytes (type
);
9097 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
9099 if (TARGET_DEBUG_ARG
)
9101 fprintf (stderr
, "va_arg: aggregate type");
9105 /* Aggregates are passed by reference. */
9110 /* kernel stack layout on 31 bit: It is assumed here that no padding
9111 will be added by s390_frame_info because for va_args always an even
9112 number of gprs has to be saved r15-r2 = 14 regs. */
9113 sav_ofs
= 2 * UNITS_PER_LONG
;
9114 sav_scale
= UNITS_PER_LONG
;
9115 size
= UNITS_PER_LONG
;
9116 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9118 else if (s390_function_arg_float (TYPE_MODE (type
), type
))
9120 if (TARGET_DEBUG_ARG
)
9122 fprintf (stderr
, "va_arg: float type");
9126 /* FP args go in FP registers, if present. */
9130 sav_ofs
= 16 * UNITS_PER_LONG
;
9132 max_reg
= FP_ARG_NUM_REG
- n_reg
;
9136 if (TARGET_DEBUG_ARG
)
9138 fprintf (stderr
, "va_arg: other type");
9142 /* Otherwise into GP registers. */
9145 n_reg
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
9147 /* kernel stack layout on 31 bit: It is assumed here that no padding
9148 will be added by s390_frame_info because for va_args always an even
9149 number of gprs has to be saved r15-r2 = 14 regs. */
9150 sav_ofs
= 2 * UNITS_PER_LONG
;
9152 if (size
< UNITS_PER_LONG
)
9153 sav_ofs
+= UNITS_PER_LONG
- size
;
9155 sav_scale
= UNITS_PER_LONG
;
9156 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9159 /* Pull the value out of the saved registers ... */
9161 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
9162 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
9163 addr
= create_tmp_var (ptr_type_node
, "addr");
9165 t
= fold_convert (TREE_TYPE (reg
), size_int (max_reg
));
9166 t
= build2 (GT_EXPR
, boolean_type_node
, reg
, t
);
9167 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
9168 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
9169 gimplify_and_add (t
, pre_p
);
9171 t
= fold_build_pointer_plus_hwi (sav
, sav_ofs
);
9172 u
= build2 (MULT_EXPR
, TREE_TYPE (reg
), reg
,
9173 fold_convert (TREE_TYPE (reg
), size_int (sav_scale
)));
9174 t
= fold_build_pointer_plus (t
, u
);
9176 gimplify_assign (addr
, t
, pre_p
);
9178 gimple_seq_add_stmt (pre_p
, gimple_build_goto (lab_over
));
9180 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_false
));
9183 /* ... Otherwise out of the overflow area. */
9186 if (size
< UNITS_PER_LONG
)
9187 t
= fold_build_pointer_plus_hwi (t
, UNITS_PER_LONG
- size
);
9189 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
9191 gimplify_assign (addr
, t
, pre_p
);
9193 t
= fold_build_pointer_plus_hwi (t
, size
);
9194 gimplify_assign (ovf
, t
, pre_p
);
9196 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_over
));
9199 /* Increment register save count. */
9201 u
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (reg
), reg
,
9202 fold_convert (TREE_TYPE (reg
), size_int (n_reg
)));
9203 gimplify_and_add (u
, pre_p
);
9207 t
= build_pointer_type_for_mode (build_pointer_type (type
),
9209 addr
= fold_convert (t
, addr
);
9210 addr
= build_va_arg_indirect_ref (addr
);
9214 t
= build_pointer_type_for_mode (type
, ptr_mode
, true);
9215 addr
= fold_convert (t
, addr
);
9218 return build_va_arg_indirect_ref (addr
);
9226 S390_BUILTIN_THREAD_POINTER
,
9227 S390_BUILTIN_SET_THREAD_POINTER
,
9232 static enum insn_code
const code_for_builtin_64
[S390_BUILTIN_max
] = {
9237 static enum insn_code
const code_for_builtin_31
[S390_BUILTIN_max
] = {
9243 s390_init_builtins (void)
9247 ftype
= build_function_type_list (ptr_type_node
, NULL_TREE
);
9248 add_builtin_function ("__builtin_thread_pointer", ftype
,
9249 S390_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
9252 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
9253 add_builtin_function ("__builtin_set_thread_pointer", ftype
,
9254 S390_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
9258 /* Expand an expression EXP that calls a built-in function,
9259 with result going to TARGET if that's convenient
9260 (and in mode MODE if that's convenient).
9261 SUBTARGET may be used as the target for computing one of EXP's operands.
9262 IGNORE is nonzero if the value is to be ignored. */
9265 s390_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
9266 enum machine_mode mode ATTRIBUTE_UNUSED
,
9267 int ignore ATTRIBUTE_UNUSED
)
9271 enum insn_code
const *code_for_builtin
=
9272 TARGET_64BIT
? code_for_builtin_64
: code_for_builtin_31
;
9274 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
9275 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
9276 enum insn_code icode
;
9277 rtx op
[MAX_ARGS
], pat
;
9281 call_expr_arg_iterator iter
;
9283 if (fcode
>= S390_BUILTIN_max
)
9284 internal_error ("bad builtin fcode");
9285 icode
= code_for_builtin
[fcode
];
9287 internal_error ("bad builtin fcode");
9289 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
9292 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
9294 const struct insn_operand_data
*insn_op
;
9296 if (arg
== error_mark_node
)
9298 if (arity
> MAX_ARGS
)
9301 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
9303 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
9305 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
9306 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
9312 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
9314 || GET_MODE (target
) != tmode
9315 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
9316 target
= gen_reg_rtx (tmode
);
9322 pat
= GEN_FCN (icode
) (target
);
9326 pat
= GEN_FCN (icode
) (target
, op
[0]);
9328 pat
= GEN_FCN (icode
) (op
[0]);
9331 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
9347 /* Output assembly code for the trampoline template to
9350 On S/390, we use gpr 1 internally in the trampoline code;
9351 gpr 0 is used to hold the static chain. */
9354 s390_asm_trampoline_template (FILE *file
)
9357 op
[0] = gen_rtx_REG (Pmode
, 0);
9358 op
[1] = gen_rtx_REG (Pmode
, 1);
9362 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9363 output_asm_insn ("lmg\t%0,%1,14(%1)", op
); /* 6 byte */
9364 output_asm_insn ("br\t%1", op
); /* 2 byte */
9365 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 10));
9369 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9370 output_asm_insn ("lm\t%0,%1,6(%1)", op
); /* 4 byte */
9371 output_asm_insn ("br\t%1", op
); /* 2 byte */
9372 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 8));
9376 /* Emit RTL insns to initialize the variable parts of a trampoline.
9377 FNADDR is an RTX for the address of the function's pure code.
9378 CXT is an RTX for the static chain value for the function. */
9381 s390_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9383 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
9386 emit_block_move (m_tramp
, assemble_trampoline_template (),
9387 GEN_INT (2 * UNITS_PER_LONG
), BLOCK_OP_NORMAL
);
9389 mem
= adjust_address (m_tramp
, Pmode
, 2 * UNITS_PER_LONG
);
9390 emit_move_insn (mem
, cxt
);
9391 mem
= adjust_address (m_tramp
, Pmode
, 3 * UNITS_PER_LONG
);
9392 emit_move_insn (mem
, fnaddr
);
9395 /* Output assembler code to FILE to increment profiler label # LABELNO
9396 for profiling a function entry. */
9399 s390_function_profiler (FILE *file
, int labelno
)
9404 ASM_GENERATE_INTERNAL_LABEL (label
, "LP", labelno
);
9406 fprintf (file
, "# function profiler \n");
9408 op
[0] = gen_rtx_REG (Pmode
, RETURN_REGNUM
);
9409 op
[1] = gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
);
9410 op
[1] = gen_rtx_MEM (Pmode
, plus_constant (Pmode
, op
[1], UNITS_PER_LONG
));
9412 op
[2] = gen_rtx_REG (Pmode
, 1);
9413 op
[3] = gen_rtx_SYMBOL_REF (Pmode
, label
);
9414 SYMBOL_REF_FLAGS (op
[3]) = SYMBOL_FLAG_LOCAL
;
9416 op
[4] = gen_rtx_SYMBOL_REF (Pmode
, "_mcount");
9419 op
[4] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[4]), UNSPEC_PLT
);
9420 op
[4] = gen_rtx_CONST (Pmode
, op
[4]);
9425 output_asm_insn ("stg\t%0,%1", op
);
9426 output_asm_insn ("larl\t%2,%3", op
);
9427 output_asm_insn ("brasl\t%0,%4", op
);
9428 output_asm_insn ("lg\t%0,%1", op
);
9432 op
[6] = gen_label_rtx ();
9434 output_asm_insn ("st\t%0,%1", op
);
9435 output_asm_insn ("bras\t%2,%l6", op
);
9436 output_asm_insn (".long\t%4", op
);
9437 output_asm_insn (".long\t%3", op
);
9438 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9439 output_asm_insn ("l\t%0,0(%2)", op
);
9440 output_asm_insn ("l\t%2,4(%2)", op
);
9441 output_asm_insn ("basr\t%0,%0", op
);
9442 output_asm_insn ("l\t%0,%1", op
);
9446 op
[5] = gen_label_rtx ();
9447 op
[6] = gen_label_rtx ();
9449 output_asm_insn ("st\t%0,%1", op
);
9450 output_asm_insn ("bras\t%2,%l6", op
);
9451 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[5]));
9452 output_asm_insn (".long\t%4-%l5", op
);
9453 output_asm_insn (".long\t%3-%l5", op
);
9454 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9455 output_asm_insn ("lr\t%0,%2", op
);
9456 output_asm_insn ("a\t%0,0(%2)", op
);
9457 output_asm_insn ("a\t%2,4(%2)", op
);
9458 output_asm_insn ("basr\t%0,%0", op
);
9459 output_asm_insn ("l\t%0,%1", op
);
9463 /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF
9464 into its SYMBOL_REF_FLAGS. */
9467 s390_encode_section_info (tree decl
, rtx rtl
, int first
)
9469 default_encode_section_info (decl
, rtl
, first
);
9471 if (TREE_CODE (decl
) == VAR_DECL
)
9473 /* If a variable has a forced alignment to < 2 bytes, mark it
9474 with SYMBOL_FLAG_ALIGN1 to prevent it from being used as LARL
9476 if (DECL_USER_ALIGN (decl
) && DECL_ALIGN (decl
) < 16)
9477 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_ALIGN1
;
9478 if (!DECL_SIZE (decl
)
9479 || !DECL_ALIGN (decl
)
9480 || !host_integerp (DECL_SIZE (decl
), 0)
9481 || (DECL_ALIGN (decl
) <= 64
9482 && DECL_ALIGN (decl
) != tree_low_cst (DECL_SIZE (decl
), 0)))
9483 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9486 /* Literal pool references don't have a decl so they are handled
9487 differently here. We rely on the information in the MEM_ALIGN
9488 entry to decide upon natural alignment. */
9490 && GET_CODE (XEXP (rtl
, 0)) == SYMBOL_REF
9491 && TREE_CONSTANT_POOL_ADDRESS_P (XEXP (rtl
, 0))
9492 && (MEM_ALIGN (rtl
) == 0
9493 || GET_MODE_BITSIZE (GET_MODE (rtl
)) == 0
9494 || MEM_ALIGN (rtl
) < GET_MODE_BITSIZE (GET_MODE (rtl
))))
9495 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9498 /* Output thunk to FILE that implements a C++ virtual function call (with
9499 multiple inheritance) to FUNCTION. The thunk adjusts the this pointer
9500 by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment
9501 stored at VCALL_OFFSET in the vtable whose address is located at offset 0
9502 relative to the resulting this pointer. */
9505 s390_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
9506 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
9512 /* Make sure unwind info is emitted for the thunk if needed. */
9513 final_start_function (emit_barrier (), file
, 1);
9515 /* Operand 0 is the target function. */
9516 op
[0] = XEXP (DECL_RTL (function
), 0);
9517 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (op
[0]))
9520 op
[0] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[0]),
9521 TARGET_64BIT
? UNSPEC_PLT
: UNSPEC_GOT
);
9522 op
[0] = gen_rtx_CONST (Pmode
, op
[0]);
9525 /* Operand 1 is the 'this' pointer. */
9526 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
9527 op
[1] = gen_rtx_REG (Pmode
, 3);
9529 op
[1] = gen_rtx_REG (Pmode
, 2);
9531 /* Operand 2 is the delta. */
9532 op
[2] = GEN_INT (delta
);
9534 /* Operand 3 is the vcall_offset. */
9535 op
[3] = GEN_INT (vcall_offset
);
9537 /* Operand 4 is the temporary register. */
9538 op
[4] = gen_rtx_REG (Pmode
, 1);
9540 /* Operands 5 to 8 can be used as labels. */
9546 /* Operand 9 can be used for temporary register. */
9549 /* Generate code. */
9552 /* Setup literal pool pointer if required. */
9553 if ((!DISP_IN_RANGE (delta
)
9554 && !CONST_OK_FOR_K (delta
)
9555 && !CONST_OK_FOR_Os (delta
))
9556 || (!DISP_IN_RANGE (vcall_offset
)
9557 && !CONST_OK_FOR_K (vcall_offset
)
9558 && !CONST_OK_FOR_Os (vcall_offset
)))
9560 op
[5] = gen_label_rtx ();
9561 output_asm_insn ("larl\t%4,%5", op
);
9564 /* Add DELTA to this pointer. */
9567 if (CONST_OK_FOR_J (delta
))
9568 output_asm_insn ("la\t%1,%2(%1)", op
);
9569 else if (DISP_IN_RANGE (delta
))
9570 output_asm_insn ("lay\t%1,%2(%1)", op
);
9571 else if (CONST_OK_FOR_K (delta
))
9572 output_asm_insn ("aghi\t%1,%2", op
);
9573 else if (CONST_OK_FOR_Os (delta
))
9574 output_asm_insn ("agfi\t%1,%2", op
);
9577 op
[6] = gen_label_rtx ();
9578 output_asm_insn ("agf\t%1,%6-%5(%4)", op
);
9582 /* Perform vcall adjustment. */
9585 if (DISP_IN_RANGE (vcall_offset
))
9587 output_asm_insn ("lg\t%4,0(%1)", op
);
9588 output_asm_insn ("ag\t%1,%3(%4)", op
);
9590 else if (CONST_OK_FOR_K (vcall_offset
))
9592 output_asm_insn ("lghi\t%4,%3", op
);
9593 output_asm_insn ("ag\t%4,0(%1)", op
);
9594 output_asm_insn ("ag\t%1,0(%4)", op
);
9596 else if (CONST_OK_FOR_Os (vcall_offset
))
9598 output_asm_insn ("lgfi\t%4,%3", op
);
9599 output_asm_insn ("ag\t%4,0(%1)", op
);
9600 output_asm_insn ("ag\t%1,0(%4)", op
);
9604 op
[7] = gen_label_rtx ();
9605 output_asm_insn ("llgf\t%4,%7-%5(%4)", op
);
9606 output_asm_insn ("ag\t%4,0(%1)", op
);
9607 output_asm_insn ("ag\t%1,0(%4)", op
);
9611 /* Jump to target. */
9612 output_asm_insn ("jg\t%0", op
);
9614 /* Output literal pool if required. */
9617 output_asm_insn (".align\t4", op
);
9618 targetm
.asm_out
.internal_label (file
, "L",
9619 CODE_LABEL_NUMBER (op
[5]));
9623 targetm
.asm_out
.internal_label (file
, "L",
9624 CODE_LABEL_NUMBER (op
[6]));
9625 output_asm_insn (".long\t%2", op
);
9629 targetm
.asm_out
.internal_label (file
, "L",
9630 CODE_LABEL_NUMBER (op
[7]));
9631 output_asm_insn (".long\t%3", op
);
9636 /* Setup base pointer if required. */
9638 || (!DISP_IN_RANGE (delta
)
9639 && !CONST_OK_FOR_K (delta
)
9640 && !CONST_OK_FOR_Os (delta
))
9641 || (!DISP_IN_RANGE (delta
)
9642 && !CONST_OK_FOR_K (vcall_offset
)
9643 && !CONST_OK_FOR_Os (vcall_offset
)))
9645 op
[5] = gen_label_rtx ();
9646 output_asm_insn ("basr\t%4,0", op
);
9647 targetm
.asm_out
.internal_label (file
, "L",
9648 CODE_LABEL_NUMBER (op
[5]));
9651 /* Add DELTA to this pointer. */
9654 if (CONST_OK_FOR_J (delta
))
9655 output_asm_insn ("la\t%1,%2(%1)", op
);
9656 else if (DISP_IN_RANGE (delta
))
9657 output_asm_insn ("lay\t%1,%2(%1)", op
);
9658 else if (CONST_OK_FOR_K (delta
))
9659 output_asm_insn ("ahi\t%1,%2", op
);
9660 else if (CONST_OK_FOR_Os (delta
))
9661 output_asm_insn ("afi\t%1,%2", op
);
9664 op
[6] = gen_label_rtx ();
9665 output_asm_insn ("a\t%1,%6-%5(%4)", op
);
9669 /* Perform vcall adjustment. */
9672 if (CONST_OK_FOR_J (vcall_offset
))
9674 output_asm_insn ("l\t%4,0(%1)", op
);
9675 output_asm_insn ("a\t%1,%3(%4)", op
);
9677 else if (DISP_IN_RANGE (vcall_offset
))
9679 output_asm_insn ("l\t%4,0(%1)", op
);
9680 output_asm_insn ("ay\t%1,%3(%4)", op
);
9682 else if (CONST_OK_FOR_K (vcall_offset
))
9684 output_asm_insn ("lhi\t%4,%3", op
);
9685 output_asm_insn ("a\t%4,0(%1)", op
);
9686 output_asm_insn ("a\t%1,0(%4)", op
);
9688 else if (CONST_OK_FOR_Os (vcall_offset
))
9690 output_asm_insn ("iilf\t%4,%3", op
);
9691 output_asm_insn ("a\t%4,0(%1)", op
);
9692 output_asm_insn ("a\t%1,0(%4)", op
);
9696 op
[7] = gen_label_rtx ();
9697 output_asm_insn ("l\t%4,%7-%5(%4)", op
);
9698 output_asm_insn ("a\t%4,0(%1)", op
);
9699 output_asm_insn ("a\t%1,0(%4)", op
);
9702 /* We had to clobber the base pointer register.
9703 Re-setup the base pointer (with a different base). */
9704 op
[5] = gen_label_rtx ();
9705 output_asm_insn ("basr\t%4,0", op
);
9706 targetm
.asm_out
.internal_label (file
, "L",
9707 CODE_LABEL_NUMBER (op
[5]));
9710 /* Jump to target. */
9711 op
[8] = gen_label_rtx ();
9714 output_asm_insn ("l\t%4,%8-%5(%4)", op
);
9716 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9717 /* We cannot call through .plt, since .plt requires %r12 loaded. */
9718 else if (flag_pic
== 1)
9720 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9721 output_asm_insn ("l\t%4,%0(%4)", op
);
9723 else if (flag_pic
== 2)
9725 op
[9] = gen_rtx_REG (Pmode
, 0);
9726 output_asm_insn ("l\t%9,%8-4-%5(%4)", op
);
9727 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9728 output_asm_insn ("ar\t%4,%9", op
);
9729 output_asm_insn ("l\t%4,0(%4)", op
);
9732 output_asm_insn ("br\t%4", op
);
9734 /* Output literal pool. */
9735 output_asm_insn (".align\t4", op
);
9737 if (nonlocal
&& flag_pic
== 2)
9738 output_asm_insn (".long\t%0", op
);
9741 op
[0] = gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
9742 SYMBOL_REF_FLAGS (op
[0]) = SYMBOL_FLAG_LOCAL
;
9745 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[8]));
9747 output_asm_insn (".long\t%0", op
);
9749 output_asm_insn (".long\t%0-%5", op
);
9753 targetm
.asm_out
.internal_label (file
, "L",
9754 CODE_LABEL_NUMBER (op
[6]));
9755 output_asm_insn (".long\t%2", op
);
9759 targetm
.asm_out
.internal_label (file
, "L",
9760 CODE_LABEL_NUMBER (op
[7]));
9761 output_asm_insn (".long\t%3", op
);
9764 final_end_function ();
9768 s390_valid_pointer_mode (enum machine_mode mode
)
9770 return (mode
== SImode
|| (TARGET_64BIT
&& mode
== DImode
));
9773 /* Checks whether the given CALL_EXPR would use a caller
9774 saved register. This is used to decide whether sibling call
9775 optimization could be performed on the respective function
9779 s390_call_saved_register_used (tree call_expr
)
9781 CUMULATIVE_ARGS cum_v
;
9782 cumulative_args_t cum
;
9784 enum machine_mode mode
;
9789 INIT_CUMULATIVE_ARGS (cum_v
, NULL
, NULL
, 0, 0);
9790 cum
= pack_cumulative_args (&cum_v
);
9792 for (i
= 0; i
< call_expr_nargs (call_expr
); i
++)
9794 parameter
= CALL_EXPR_ARG (call_expr
, i
);
9795 gcc_assert (parameter
);
9797 /* For an undeclared variable passed as parameter we will get
9798 an ERROR_MARK node here. */
9799 if (TREE_CODE (parameter
) == ERROR_MARK
)
9802 type
= TREE_TYPE (parameter
);
9805 mode
= TYPE_MODE (type
);
9808 if (pass_by_reference (&cum_v
, mode
, type
, true))
9811 type
= build_pointer_type (type
);
9814 parm_rtx
= s390_function_arg (cum
, mode
, type
, 0);
9816 s390_function_arg_advance (cum
, mode
, type
, 0);
9821 if (REG_P (parm_rtx
))
9824 reg
< HARD_REGNO_NREGS (REGNO (parm_rtx
), GET_MODE (parm_rtx
));
9826 if (!call_used_regs
[reg
+ REGNO (parm_rtx
)])
9830 if (GET_CODE (parm_rtx
) == PARALLEL
)
9834 for (i
= 0; i
< XVECLEN (parm_rtx
, 0); i
++)
9836 rtx r
= XEXP (XVECEXP (parm_rtx
, 0, i
), 0);
9838 gcc_assert (REG_P (r
));
9841 reg
< HARD_REGNO_NREGS (REGNO (r
), GET_MODE (r
));
9843 if (!call_used_regs
[reg
+ REGNO (r
)])
9852 /* Return true if the given call expression can be
9853 turned into a sibling call.
9854 DECL holds the declaration of the function to be called whereas
9855 EXP is the call expression itself. */
9858 s390_function_ok_for_sibcall (tree decl
, tree exp
)
9860 /* The TPF epilogue uses register 1. */
9861 if (TARGET_TPF_PROFILING
)
9864 /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
9865 which would have to be restored before the sibcall. */
9866 if (!TARGET_64BIT
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
9869 /* Register 6 on s390 is available as an argument register but unfortunately
9870 "caller saved". This makes functions needing this register for arguments
9871 not suitable for sibcalls. */
9872 return !s390_call_saved_register_used (exp
);
9875 /* Return the fixed registers used for condition codes. */
9878 s390_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
9881 *p2
= INVALID_REGNUM
;
9886 /* This function is used by the call expanders of the machine description.
9887 It emits the call insn itself together with the necessary operations
9888 to adjust the target address and returns the emitted insn.
9889 ADDR_LOCATION is the target address rtx
9890 TLS_CALL the location of the thread-local symbol
9891 RESULT_REG the register where the result of the call should be stored
9892 RETADDR_REG the register where the return address should be stored
9893 If this parameter is NULL_RTX the call is considered
9894 to be a sibling call. */
9897 s390_emit_call (rtx addr_location
, rtx tls_call
, rtx result_reg
,
9900 bool plt_call
= false;
9906 /* Direct function calls need special treatment. */
9907 if (GET_CODE (addr_location
) == SYMBOL_REF
)
9909 /* When calling a global routine in PIC mode, we must
9910 replace the symbol itself with the PLT stub. */
9911 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (addr_location
))
9913 if (retaddr_reg
!= NULL_RTX
)
9915 addr_location
= gen_rtx_UNSPEC (Pmode
,
9916 gen_rtvec (1, addr_location
),
9918 addr_location
= gen_rtx_CONST (Pmode
, addr_location
);
9922 /* For -fpic code the PLT entries might use r12 which is
9923 call-saved. Therefore we cannot do a sibcall when
9924 calling directly using a symbol ref. When reaching
9925 this point we decided (in s390_function_ok_for_sibcall)
9926 to do a sibcall for a function pointer but one of the
9927 optimizers was able to get rid of the function pointer
9928 by propagating the symbol ref into the call. This
9929 optimization is illegal for S/390 so we turn the direct
9930 call into a indirect call again. */
9931 addr_location
= force_reg (Pmode
, addr_location
);
9934 /* Unless we can use the bras(l) insn, force the
9935 routine address into a register. */
9936 if (!TARGET_SMALL_EXEC
&& !TARGET_CPU_ZARCH
)
9939 addr_location
= legitimize_pic_address (addr_location
, 0);
9941 addr_location
= force_reg (Pmode
, addr_location
);
9945 /* If it is already an indirect call or the code above moved the
9946 SYMBOL_REF to somewhere else make sure the address can be found in
9948 if (retaddr_reg
== NULL_RTX
9949 && GET_CODE (addr_location
) != SYMBOL_REF
9952 emit_move_insn (gen_rtx_REG (Pmode
, SIBCALL_REGNUM
), addr_location
);
9953 addr_location
= gen_rtx_REG (Pmode
, SIBCALL_REGNUM
);
9956 addr_location
= gen_rtx_MEM (QImode
, addr_location
);
9957 call
= gen_rtx_CALL (VOIDmode
, addr_location
, const0_rtx
);
9959 if (result_reg
!= NULL_RTX
)
9960 call
= gen_rtx_SET (VOIDmode
, result_reg
, call
);
9962 if (retaddr_reg
!= NULL_RTX
)
9964 clobber
= gen_rtx_CLOBBER (VOIDmode
, retaddr_reg
);
9966 if (tls_call
!= NULL_RTX
)
9967 vec
= gen_rtvec (3, call
, clobber
,
9968 gen_rtx_USE (VOIDmode
, tls_call
));
9970 vec
= gen_rtvec (2, call
, clobber
);
9972 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
9975 insn
= emit_call_insn (call
);
9977 /* 31-bit PLT stubs and tls calls use the GOT register implicitly. */
9978 if ((!TARGET_64BIT
&& plt_call
) || tls_call
!= NULL_RTX
)
9980 /* s390_function_ok_for_sibcall should
9981 have denied sibcalls in this case. */
9982 gcc_assert (retaddr_reg
!= NULL_RTX
);
9983 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), gen_rtx_REG (Pmode
, 12));
9988 /* Implement TARGET_CONDITIONAL_REGISTER_USAGE. */
9991 s390_conditional_register_usage (void)
9997 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9998 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
10000 if (TARGET_CPU_ZARCH
)
10002 fixed_regs
[BASE_REGNUM
] = 0;
10003 call_used_regs
[BASE_REGNUM
] = 0;
10004 fixed_regs
[RETURN_REGNUM
] = 0;
10005 call_used_regs
[RETURN_REGNUM
] = 0;
10009 for (i
= 24; i
< 32; i
++)
10010 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
10014 for (i
= 18; i
< 20; i
++)
10015 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
10018 if (TARGET_SOFT_FLOAT
)
10020 for (i
= 16; i
< 32; i
++)
10021 call_used_regs
[i
] = fixed_regs
[i
] = 1;
10025 /* Corresponding function to eh_return expander. */
10027 static GTY(()) rtx s390_tpf_eh_return_symbol
;
10029 s390_emit_tpf_eh_return (rtx target
)
10033 if (!s390_tpf_eh_return_symbol
)
10034 s390_tpf_eh_return_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tpf_eh_return");
10036 reg
= gen_rtx_REG (Pmode
, 2);
10038 emit_move_insn (reg
, target
);
10039 insn
= s390_emit_call (s390_tpf_eh_return_symbol
, NULL_RTX
, reg
,
10040 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
10041 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), reg
);
10043 emit_move_insn (EH_RETURN_HANDLER_RTX
, reg
);
10046 /* Rework the prologue/epilogue to avoid saving/restoring
10047 registers unnecessarily. */
10050 s390_optimize_prologue (void)
10052 rtx insn
, new_insn
, next_insn
;
10054 /* Do a final recompute of the frame-related data. */
10056 s390_update_frame_layout ();
10058 /* If all special registers are in fact used, there's nothing we
10059 can do, so no point in walking the insn list. */
10061 if (cfun_frame_layout
.first_save_gpr
<= BASE_REGNUM
10062 && cfun_frame_layout
.last_save_gpr
>= BASE_REGNUM
10063 && (TARGET_CPU_ZARCH
10064 || (cfun_frame_layout
.first_save_gpr
<= RETURN_REGNUM
10065 && cfun_frame_layout
.last_save_gpr
>= RETURN_REGNUM
)))
10068 /* Search for prologue/epilogue insns and replace them. */
10070 for (insn
= get_insns (); insn
; insn
= next_insn
)
10072 int first
, last
, off
;
10073 rtx set
, base
, offset
;
10075 next_insn
= NEXT_INSN (insn
);
10077 if (GET_CODE (insn
) != INSN
)
10080 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10081 && store_multiple_operation (PATTERN (insn
), VOIDmode
))
10083 set
= XVECEXP (PATTERN (insn
), 0, 0);
10084 first
= REGNO (SET_SRC (set
));
10085 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10086 offset
= const0_rtx
;
10087 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10088 off
= INTVAL (offset
);
10090 if (GET_CODE (base
) != REG
|| off
< 0)
10092 if (cfun_frame_layout
.first_save_gpr
!= -1
10093 && (cfun_frame_layout
.first_save_gpr
< first
10094 || cfun_frame_layout
.last_save_gpr
> last
))
10096 if (REGNO (base
) != STACK_POINTER_REGNUM
10097 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10099 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10102 if (cfun_frame_layout
.first_save_gpr
!= -1)
10104 new_insn
= save_gprs (base
,
10105 off
+ (cfun_frame_layout
.first_save_gpr
10106 - first
) * UNITS_PER_LONG
,
10107 cfun_frame_layout
.first_save_gpr
,
10108 cfun_frame_layout
.last_save_gpr
);
10109 new_insn
= emit_insn_before (new_insn
, insn
);
10110 INSN_ADDRESSES_NEW (new_insn
, -1);
10113 remove_insn (insn
);
10117 if (cfun_frame_layout
.first_save_gpr
== -1
10118 && GET_CODE (PATTERN (insn
)) == SET
10119 && GET_CODE (SET_SRC (PATTERN (insn
))) == REG
10120 && (REGNO (SET_SRC (PATTERN (insn
))) == BASE_REGNUM
10121 || (!TARGET_CPU_ZARCH
10122 && REGNO (SET_SRC (PATTERN (insn
))) == RETURN_REGNUM
))
10123 && GET_CODE (SET_DEST (PATTERN (insn
))) == MEM
)
10125 set
= PATTERN (insn
);
10126 first
= REGNO (SET_SRC (set
));
10127 offset
= const0_rtx
;
10128 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10129 off
= INTVAL (offset
);
10131 if (GET_CODE (base
) != REG
|| off
< 0)
10133 if (REGNO (base
) != STACK_POINTER_REGNUM
10134 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10137 remove_insn (insn
);
10141 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10142 && load_multiple_operation (PATTERN (insn
), VOIDmode
))
10144 set
= XVECEXP (PATTERN (insn
), 0, 0);
10145 first
= REGNO (SET_DEST (set
));
10146 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10147 offset
= const0_rtx
;
10148 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10149 off
= INTVAL (offset
);
10151 if (GET_CODE (base
) != REG
|| off
< 0)
10153 if (cfun_frame_layout
.first_restore_gpr
!= -1
10154 && (cfun_frame_layout
.first_restore_gpr
< first
10155 || cfun_frame_layout
.last_restore_gpr
> last
))
10157 if (REGNO (base
) != STACK_POINTER_REGNUM
10158 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10160 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10163 if (cfun_frame_layout
.first_restore_gpr
!= -1)
10165 new_insn
= restore_gprs (base
,
10166 off
+ (cfun_frame_layout
.first_restore_gpr
10167 - first
) * UNITS_PER_LONG
,
10168 cfun_frame_layout
.first_restore_gpr
,
10169 cfun_frame_layout
.last_restore_gpr
);
10170 new_insn
= emit_insn_before (new_insn
, insn
);
10171 INSN_ADDRESSES_NEW (new_insn
, -1);
10174 remove_insn (insn
);
10178 if (cfun_frame_layout
.first_restore_gpr
== -1
10179 && GET_CODE (PATTERN (insn
)) == SET
10180 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
10181 && (REGNO (SET_DEST (PATTERN (insn
))) == BASE_REGNUM
10182 || (!TARGET_CPU_ZARCH
10183 && REGNO (SET_DEST (PATTERN (insn
))) == RETURN_REGNUM
))
10184 && GET_CODE (SET_SRC (PATTERN (insn
))) == MEM
)
10186 set
= PATTERN (insn
);
10187 first
= REGNO (SET_DEST (set
));
10188 offset
= const0_rtx
;
10189 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10190 off
= INTVAL (offset
);
10192 if (GET_CODE (base
) != REG
|| off
< 0)
10194 if (REGNO (base
) != STACK_POINTER_REGNUM
10195 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10198 remove_insn (insn
);
10204 /* On z10 and later the dynamic branch prediction must see the
10205 backward jump within a certain windows. If not it falls back to
10206 the static prediction. This function rearranges the loop backward
10207 branch in a way which makes the static prediction always correct.
10208 The function returns true if it added an instruction. */
10210 s390_fix_long_loop_prediction (rtx insn
)
10212 rtx set
= single_set (insn
);
10213 rtx code_label
, label_ref
, new_label
;
10219 /* This will exclude branch on count and branch on index patterns
10220 since these are correctly statically predicted. */
10222 || SET_DEST (set
) != pc_rtx
10223 || GET_CODE (SET_SRC(set
)) != IF_THEN_ELSE
)
10226 label_ref
= (GET_CODE (XEXP (SET_SRC (set
), 1)) == LABEL_REF
?
10227 XEXP (SET_SRC (set
), 1) : XEXP (SET_SRC (set
), 2));
10229 gcc_assert (GET_CODE (label_ref
) == LABEL_REF
);
10231 code_label
= XEXP (label_ref
, 0);
10233 if (INSN_ADDRESSES (INSN_UID (code_label
)) == -1
10234 || INSN_ADDRESSES (INSN_UID (insn
)) == -1
10235 || (INSN_ADDRESSES (INSN_UID (insn
))
10236 - INSN_ADDRESSES (INSN_UID (code_label
)) < PREDICT_DISTANCE
))
10239 for (distance
= 0, cur_insn
= PREV_INSN (insn
);
10240 distance
< PREDICT_DISTANCE
- 6;
10241 distance
+= get_attr_length (cur_insn
), cur_insn
= PREV_INSN (cur_insn
))
10242 if (!cur_insn
|| JUMP_P (cur_insn
) || LABEL_P (cur_insn
))
10245 new_label
= gen_label_rtx ();
10246 uncond_jump
= emit_jump_insn_after (
10247 gen_rtx_SET (VOIDmode
, pc_rtx
,
10248 gen_rtx_LABEL_REF (VOIDmode
, code_label
)),
10250 emit_label_after (new_label
, uncond_jump
);
10252 tmp
= XEXP (SET_SRC (set
), 1);
10253 XEXP (SET_SRC (set
), 1) = XEXP (SET_SRC (set
), 2);
10254 XEXP (SET_SRC (set
), 2) = tmp
;
10255 INSN_CODE (insn
) = -1;
10257 XEXP (label_ref
, 0) = new_label
;
10258 JUMP_LABEL (insn
) = new_label
;
10259 JUMP_LABEL (uncond_jump
) = code_label
;
10264 /* Returns 1 if INSN reads the value of REG for purposes not related
10265 to addressing of memory, and 0 otherwise. */
10267 s390_non_addr_reg_read_p (rtx reg
, rtx insn
)
10269 return reg_referenced_p (reg
, PATTERN (insn
))
10270 && !reg_used_in_mem_p (REGNO (reg
), PATTERN (insn
));
10273 /* Starting from INSN find_cond_jump looks downwards in the insn
10274 stream for a single jump insn which is the last user of the
10275 condition code set in INSN. */
10277 find_cond_jump (rtx insn
)
10279 for (; insn
; insn
= NEXT_INSN (insn
))
10283 if (LABEL_P (insn
))
10286 if (!JUMP_P (insn
))
10288 if (reg_mentioned_p (gen_rtx_REG (CCmode
, CC_REGNUM
), insn
))
10293 /* This will be triggered by a return. */
10294 if (GET_CODE (PATTERN (insn
)) != SET
)
10297 gcc_assert (SET_DEST (PATTERN (insn
)) == pc_rtx
);
10298 ite
= SET_SRC (PATTERN (insn
));
10300 if (GET_CODE (ite
) != IF_THEN_ELSE
)
10303 cc
= XEXP (XEXP (ite
, 0), 0);
10304 if (!REG_P (cc
) || !CC_REGNO_P (REGNO (cc
)))
10307 if (find_reg_note (insn
, REG_DEAD
, cc
))
10315 /* Swap the condition in COND and the operands in OP0 and OP1 so that
10316 the semantics does not change. If NULL_RTX is passed as COND the
10317 function tries to find the conditional jump starting with INSN. */
10319 s390_swap_cmp (rtx cond
, rtx
*op0
, rtx
*op1
, rtx insn
)
10323 if (cond
== NULL_RTX
)
10325 rtx jump
= find_cond_jump (NEXT_INSN (insn
));
10326 jump
= jump
? single_set (jump
) : NULL_RTX
;
10328 if (jump
== NULL_RTX
)
10331 cond
= XEXP (XEXP (jump
, 1), 0);
10336 PUT_CODE (cond
, swap_condition (GET_CODE (cond
)));
10339 /* On z10, instructions of the compare-and-branch family have the
10340 property to access the register occurring as second operand with
10341 its bits complemented. If such a compare is grouped with a second
10342 instruction that accesses the same register non-complemented, and
10343 if that register's value is delivered via a bypass, then the
10344 pipeline recycles, thereby causing significant performance decline.
10345 This function locates such situations and exchanges the two
10346 operands of the compare. The function return true whenever it
10349 s390_z10_optimize_cmp (rtx insn
)
10351 rtx prev_insn
, next_insn
;
10352 bool insn_added_p
= false;
10353 rtx cond
, *op0
, *op1
;
10355 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
10357 /* Handle compare and branch and branch on count
10359 rtx pattern
= single_set (insn
);
10362 || SET_DEST (pattern
) != pc_rtx
10363 || GET_CODE (SET_SRC (pattern
)) != IF_THEN_ELSE
)
10366 cond
= XEXP (SET_SRC (pattern
), 0);
10367 op0
= &XEXP (cond
, 0);
10368 op1
= &XEXP (cond
, 1);
10370 else if (GET_CODE (PATTERN (insn
)) == SET
)
10374 /* Handle normal compare instructions. */
10375 src
= SET_SRC (PATTERN (insn
));
10376 dest
= SET_DEST (PATTERN (insn
));
10379 || !CC_REGNO_P (REGNO (dest
))
10380 || GET_CODE (src
) != COMPARE
)
10383 /* s390_swap_cmp will try to find the conditional
10384 jump when passing NULL_RTX as condition. */
10386 op0
= &XEXP (src
, 0);
10387 op1
= &XEXP (src
, 1);
10392 if (!REG_P (*op0
) || !REG_P (*op1
))
10395 if (GET_MODE_CLASS (GET_MODE (*op0
)) != MODE_INT
)
10398 /* Swap the COMPARE arguments and its mask if there is a
10399 conflicting access in the previous insn. */
10400 prev_insn
= prev_active_insn (insn
);
10401 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10402 && reg_referenced_p (*op1
, PATTERN (prev_insn
)))
10403 s390_swap_cmp (cond
, op0
, op1
, insn
);
10405 /* Check if there is a conflict with the next insn. If there
10406 was no conflict with the previous insn, then swap the
10407 COMPARE arguments and its mask. If we already swapped
10408 the operands, or if swapping them would cause a conflict
10409 with the previous insn, issue a NOP after the COMPARE in
10410 order to separate the two instuctions. */
10411 next_insn
= next_active_insn (insn
);
10412 if (next_insn
!= NULL_RTX
&& INSN_P (next_insn
)
10413 && s390_non_addr_reg_read_p (*op1
, next_insn
))
10415 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10416 && s390_non_addr_reg_read_p (*op0
, prev_insn
))
10418 if (REGNO (*op1
) == 0)
10419 emit_insn_after (gen_nop1 (), insn
);
10421 emit_insn_after (gen_nop (), insn
);
10422 insn_added_p
= true;
10425 s390_swap_cmp (cond
, op0
, op1
, insn
);
10427 return insn_added_p
;
10430 /* Perform machine-dependent processing. */
10435 bool pool_overflow
= false;
10437 /* Make sure all splits have been performed; splits after
10438 machine_dependent_reorg might confuse insn length counts. */
10439 split_all_insns_noflow ();
10441 /* Install the main literal pool and the associated base
10442 register load insns.
10444 In addition, there are two problematic situations we need
10447 - the literal pool might be > 4096 bytes in size, so that
10448 some of its elements cannot be directly accessed
10450 - a branch target might be > 64K away from the branch, so that
10451 it is not possible to use a PC-relative instruction.
10453 To fix those, we split the single literal pool into multiple
10454 pool chunks, reloading the pool base register at various
10455 points throughout the function to ensure it always points to
10456 the pool chunk the following code expects, and / or replace
10457 PC-relative branches by absolute branches.
10459 However, the two problems are interdependent: splitting the
10460 literal pool can move a branch further away from its target,
10461 causing the 64K limit to overflow, and on the other hand,
10462 replacing a PC-relative branch by an absolute branch means
10463 we need to put the branch target address into the literal
10464 pool, possibly causing it to overflow.
10466 So, we loop trying to fix up both problems until we manage
10467 to satisfy both conditions at the same time. Note that the
10468 loop is guaranteed to terminate as every pass of the loop
10469 strictly decreases the total number of PC-relative branches
10470 in the function. (This is not completely true as there
10471 might be branch-over-pool insns introduced by chunkify_start.
10472 Those never need to be split however.) */
10476 struct constant_pool
*pool
= NULL
;
10478 /* Collect the literal pool. */
10479 if (!pool_overflow
)
10481 pool
= s390_mainpool_start ();
10483 pool_overflow
= true;
10486 /* If literal pool overflowed, start to chunkify it. */
10488 pool
= s390_chunkify_start ();
10490 /* Split out-of-range branches. If this has created new
10491 literal pool entries, cancel current chunk list and
10492 recompute it. zSeries machines have large branch
10493 instructions, so we never need to split a branch. */
10494 if (!TARGET_CPU_ZARCH
&& s390_split_branches ())
10497 s390_chunkify_cancel (pool
);
10499 s390_mainpool_cancel (pool
);
10504 /* If we made it up to here, both conditions are satisfied.
10505 Finish up literal pool related changes. */
10507 s390_chunkify_finish (pool
);
10509 s390_mainpool_finish (pool
);
10511 /* We're done splitting branches. */
10512 cfun
->machine
->split_branches_pending_p
= false;
10516 /* Generate out-of-pool execute target insns. */
10517 if (TARGET_CPU_ZARCH
)
10519 rtx insn
, label
, target
;
10521 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10523 label
= s390_execute_label (insn
);
10527 gcc_assert (label
!= const0_rtx
);
10529 target
= emit_label (XEXP (label
, 0));
10530 INSN_ADDRESSES_NEW (target
, -1);
10532 target
= emit_insn (s390_execute_target (insn
));
10533 INSN_ADDRESSES_NEW (target
, -1);
10537 /* Try to optimize prologue and epilogue further. */
10538 s390_optimize_prologue ();
10540 /* Walk over the insns and do some >=z10 specific changes. */
10541 if (s390_tune
== PROCESSOR_2097_Z10
10542 || s390_tune
== PROCESSOR_2817_Z196
)
10545 bool insn_added_p
= false;
10547 /* The insn lengths and addresses have to be up to date for the
10548 following manipulations. */
10549 shorten_branches (get_insns ());
10551 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10553 if (!INSN_P (insn
) || INSN_CODE (insn
) <= 0)
10557 insn_added_p
|= s390_fix_long_loop_prediction (insn
);
10559 if ((GET_CODE (PATTERN (insn
)) == PARALLEL
10560 || GET_CODE (PATTERN (insn
)) == SET
)
10561 && s390_tune
== PROCESSOR_2097_Z10
)
10562 insn_added_p
|= s390_z10_optimize_cmp (insn
);
10565 /* Adjust branches if we added new instructions. */
10567 shorten_branches (get_insns ());
10571 /* Return true if INSN is a fp load insn writing register REGNO. */
10573 s390_fpload_toreg (rtx insn
, unsigned int regno
)
10576 enum attr_type flag
= s390_safe_attr_type (insn
);
10578 if (flag
!= TYPE_FLOADSF
&& flag
!= TYPE_FLOADDF
)
10581 set
= single_set (insn
);
10583 if (set
== NULL_RTX
)
10586 if (!REG_P (SET_DEST (set
)) || !MEM_P (SET_SRC (set
)))
10589 if (REGNO (SET_DEST (set
)) != regno
)
10595 /* This value describes the distance to be avoided between an
10596 aritmetic fp instruction and an fp load writing the same register.
10597 Z10_EARLYLOAD_DISTANCE - 1 as well as Z10_EARLYLOAD_DISTANCE + 1 is
10598 fine but the exact value has to be avoided. Otherwise the FP
10599 pipeline will throw an exception causing a major penalty. */
10600 #define Z10_EARLYLOAD_DISTANCE 7
10602 /* Rearrange the ready list in order to avoid the situation described
10603 for Z10_EARLYLOAD_DISTANCE. A problematic load instruction is
10604 moved to the very end of the ready list. */
10606 s390_z10_prevent_earlyload_conflicts (rtx
*ready
, int *nready_p
)
10608 unsigned int regno
;
10609 int nready
= *nready_p
;
10614 enum attr_type flag
;
10617 /* Skip DISTANCE - 1 active insns. */
10618 for (insn
= last_scheduled_insn
, distance
= Z10_EARLYLOAD_DISTANCE
- 1;
10619 distance
> 0 && insn
!= NULL_RTX
;
10620 distance
--, insn
= prev_active_insn (insn
))
10621 if (CALL_P (insn
) || JUMP_P (insn
))
10624 if (insn
== NULL_RTX
)
10627 set
= single_set (insn
);
10629 if (set
== NULL_RTX
|| !REG_P (SET_DEST (set
))
10630 || GET_MODE_CLASS (GET_MODE (SET_DEST (set
))) != MODE_FLOAT
)
10633 flag
= s390_safe_attr_type (insn
);
10635 if (flag
== TYPE_FLOADSF
|| flag
== TYPE_FLOADDF
)
10638 regno
= REGNO (SET_DEST (set
));
10641 while (!s390_fpload_toreg (ready
[i
], regno
) && i
> 0)
10648 memmove (&ready
[1], &ready
[0], sizeof (rtx
) * i
);
10652 /* This function is called via hook TARGET_SCHED_REORDER before
10653 issuing one insn from list READY which contains *NREADYP entries.
10654 For target z10 it reorders load instructions to avoid early load
10655 conflicts in the floating point pipeline */
10657 s390_sched_reorder (FILE *file ATTRIBUTE_UNUSED
, int verbose ATTRIBUTE_UNUSED
,
10658 rtx
*ready
, int *nreadyp
, int clock ATTRIBUTE_UNUSED
)
10660 if (s390_tune
== PROCESSOR_2097_Z10
)
10661 if (reload_completed
&& *nreadyp
> 1)
10662 s390_z10_prevent_earlyload_conflicts (ready
, nreadyp
);
10664 return s390_issue_rate ();
10667 /* This function is called via hook TARGET_SCHED_VARIABLE_ISSUE after
10668 the scheduler has issued INSN. It stores the last issued insn into
10669 last_scheduled_insn in order to make it available for
10670 s390_sched_reorder. */
10672 s390_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED
,
10673 int verbose ATTRIBUTE_UNUSED
,
10674 rtx insn
, int more
)
10676 last_scheduled_insn
= insn
;
10678 if (GET_CODE (PATTERN (insn
)) != USE
10679 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
10686 s390_sched_init (FILE *file ATTRIBUTE_UNUSED
,
10687 int verbose ATTRIBUTE_UNUSED
,
10688 int max_ready ATTRIBUTE_UNUSED
)
10690 last_scheduled_insn
= NULL_RTX
;
10693 /* This function checks the whole of insn X for memory references. The
10694 function always returns zero because the framework it is called
10695 from would stop recursively analyzing the insn upon a return value
10696 other than zero. The real result of this function is updating
10697 counter variable MEM_COUNT. */
10699 check_dpu (rtx
*x
, unsigned *mem_count
)
10701 if (*x
!= NULL_RTX
&& MEM_P (*x
))
10706 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates
10707 a new number struct loop *loop should be unrolled if tuned for cpus with
10708 a built-in stride prefetcher.
10709 The loop is analyzed for memory accesses by calling check_dpu for
10710 each rtx of the loop. Depending on the loop_depth and the amount of
10711 memory accesses a new number <=nunroll is returned to improve the
10712 behaviour of the hardware prefetch unit. */
10714 s390_loop_unroll_adjust (unsigned nunroll
, struct loop
*loop
)
10719 unsigned mem_count
= 0;
10721 if (s390_tune
!= PROCESSOR_2097_Z10
&& s390_tune
!= PROCESSOR_2817_Z196
)
10724 /* Count the number of memory references within the loop body. */
10725 bbs
= get_loop_body (loop
);
10726 for (i
= 0; i
< loop
->num_nodes
; i
++)
10728 for (insn
= BB_HEAD (bbs
[i
]); insn
!= BB_END (bbs
[i
]); insn
= NEXT_INSN (insn
))
10729 if (INSN_P (insn
) && INSN_CODE (insn
) != -1)
10730 for_each_rtx (&insn
, (rtx_function
) check_dpu
, &mem_count
);
10734 /* Prevent division by zero, and we do not need to adjust nunroll in this case. */
10735 if (mem_count
== 0)
10738 switch (loop_depth(loop
))
10741 return MIN (nunroll
, 28 / mem_count
);
10743 return MIN (nunroll
, 22 / mem_count
);
10745 return MIN (nunroll
, 16 / mem_count
);
10749 /* Initialize GCC target structure. */
10751 #undef TARGET_ASM_ALIGNED_HI_OP
10752 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10753 #undef TARGET_ASM_ALIGNED_DI_OP
10754 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10755 #undef TARGET_ASM_INTEGER
10756 #define TARGET_ASM_INTEGER s390_assemble_integer
10758 #undef TARGET_ASM_OPEN_PAREN
10759 #define TARGET_ASM_OPEN_PAREN ""
10761 #undef TARGET_ASM_CLOSE_PAREN
10762 #define TARGET_ASM_CLOSE_PAREN ""
10764 #undef TARGET_OPTION_OVERRIDE
10765 #define TARGET_OPTION_OVERRIDE s390_option_override
10767 #undef TARGET_ENCODE_SECTION_INFO
10768 #define TARGET_ENCODE_SECTION_INFO s390_encode_section_info
10770 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10771 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10774 #undef TARGET_HAVE_TLS
10775 #define TARGET_HAVE_TLS true
10777 #undef TARGET_CANNOT_FORCE_CONST_MEM
10778 #define TARGET_CANNOT_FORCE_CONST_MEM s390_cannot_force_const_mem
10780 #undef TARGET_DELEGITIMIZE_ADDRESS
10781 #define TARGET_DELEGITIMIZE_ADDRESS s390_delegitimize_address
10783 #undef TARGET_LEGITIMIZE_ADDRESS
10784 #define TARGET_LEGITIMIZE_ADDRESS s390_legitimize_address
10786 #undef TARGET_RETURN_IN_MEMORY
10787 #define TARGET_RETURN_IN_MEMORY s390_return_in_memory
10789 #undef TARGET_INIT_BUILTINS
10790 #define TARGET_INIT_BUILTINS s390_init_builtins
10791 #undef TARGET_EXPAND_BUILTIN
10792 #define TARGET_EXPAND_BUILTIN s390_expand_builtin
10794 #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
10795 #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA s390_output_addr_const_extra
10797 #undef TARGET_ASM_OUTPUT_MI_THUNK
10798 #define TARGET_ASM_OUTPUT_MI_THUNK s390_output_mi_thunk
10799 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10800 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10802 #undef TARGET_SCHED_ADJUST_PRIORITY
10803 #define TARGET_SCHED_ADJUST_PRIORITY s390_adjust_priority
10804 #undef TARGET_SCHED_ISSUE_RATE
10805 #define TARGET_SCHED_ISSUE_RATE s390_issue_rate
10806 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10807 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD s390_first_cycle_multipass_dfa_lookahead
10809 #undef TARGET_SCHED_VARIABLE_ISSUE
10810 #define TARGET_SCHED_VARIABLE_ISSUE s390_sched_variable_issue
10811 #undef TARGET_SCHED_REORDER
10812 #define TARGET_SCHED_REORDER s390_sched_reorder
10813 #undef TARGET_SCHED_INIT
10814 #define TARGET_SCHED_INIT s390_sched_init
10816 #undef TARGET_CANNOT_COPY_INSN_P
10817 #define TARGET_CANNOT_COPY_INSN_P s390_cannot_copy_insn_p
10818 #undef TARGET_RTX_COSTS
10819 #define TARGET_RTX_COSTS s390_rtx_costs
10820 #undef TARGET_ADDRESS_COST
10821 #define TARGET_ADDRESS_COST s390_address_cost
10822 #undef TARGET_REGISTER_MOVE_COST
10823 #define TARGET_REGISTER_MOVE_COST s390_register_move_cost
10824 #undef TARGET_MEMORY_MOVE_COST
10825 #define TARGET_MEMORY_MOVE_COST s390_memory_move_cost
10827 #undef TARGET_MACHINE_DEPENDENT_REORG
10828 #define TARGET_MACHINE_DEPENDENT_REORG s390_reorg
10830 #undef TARGET_VALID_POINTER_MODE
10831 #define TARGET_VALID_POINTER_MODE s390_valid_pointer_mode
10833 #undef TARGET_BUILD_BUILTIN_VA_LIST
10834 #define TARGET_BUILD_BUILTIN_VA_LIST s390_build_builtin_va_list
10835 #undef TARGET_EXPAND_BUILTIN_VA_START
10836 #define TARGET_EXPAND_BUILTIN_VA_START s390_va_start
10837 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10838 #define TARGET_GIMPLIFY_VA_ARG_EXPR s390_gimplify_va_arg
10840 #undef TARGET_PROMOTE_FUNCTION_MODE
10841 #define TARGET_PROMOTE_FUNCTION_MODE s390_promote_function_mode
10842 #undef TARGET_PASS_BY_REFERENCE
10843 #define TARGET_PASS_BY_REFERENCE s390_pass_by_reference
10845 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10846 #define TARGET_FUNCTION_OK_FOR_SIBCALL s390_function_ok_for_sibcall
10847 #undef TARGET_FUNCTION_ARG
10848 #define TARGET_FUNCTION_ARG s390_function_arg
10849 #undef TARGET_FUNCTION_ARG_ADVANCE
10850 #define TARGET_FUNCTION_ARG_ADVANCE s390_function_arg_advance
10851 #undef TARGET_FUNCTION_VALUE
10852 #define TARGET_FUNCTION_VALUE s390_function_value
10853 #undef TARGET_LIBCALL_VALUE
10854 #define TARGET_LIBCALL_VALUE s390_libcall_value
10856 #undef TARGET_FIXED_CONDITION_CODE_REGS
10857 #define TARGET_FIXED_CONDITION_CODE_REGS s390_fixed_condition_code_regs
10859 #undef TARGET_CC_MODES_COMPATIBLE
10860 #define TARGET_CC_MODES_COMPATIBLE s390_cc_modes_compatible
10862 #undef TARGET_INVALID_WITHIN_DOLOOP
10863 #define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null
10866 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
10867 #define TARGET_ASM_OUTPUT_DWARF_DTPREL s390_output_dwarf_dtprel
10870 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10871 #undef TARGET_MANGLE_TYPE
10872 #define TARGET_MANGLE_TYPE s390_mangle_type
10875 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10876 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10878 #undef TARGET_PREFERRED_RELOAD_CLASS
10879 #define TARGET_PREFERRED_RELOAD_CLASS s390_preferred_reload_class
10881 #undef TARGET_SECONDARY_RELOAD
10882 #define TARGET_SECONDARY_RELOAD s390_secondary_reload
10884 #undef TARGET_LIBGCC_CMP_RETURN_MODE
10885 #define TARGET_LIBGCC_CMP_RETURN_MODE s390_libgcc_cmp_return_mode
10887 #undef TARGET_LIBGCC_SHIFT_COUNT_MODE
10888 #define TARGET_LIBGCC_SHIFT_COUNT_MODE s390_libgcc_shift_count_mode
10890 #undef TARGET_LEGITIMATE_ADDRESS_P
10891 #define TARGET_LEGITIMATE_ADDRESS_P s390_legitimate_address_p
10893 #undef TARGET_LEGITIMATE_CONSTANT_P
10894 #define TARGET_LEGITIMATE_CONSTANT_P s390_legitimate_constant_p
10896 #undef TARGET_CAN_ELIMINATE
10897 #define TARGET_CAN_ELIMINATE s390_can_eliminate
10899 #undef TARGET_CONDITIONAL_REGISTER_USAGE
10900 #define TARGET_CONDITIONAL_REGISTER_USAGE s390_conditional_register_usage
10902 #undef TARGET_LOOP_UNROLL_ADJUST
10903 #define TARGET_LOOP_UNROLL_ADJUST s390_loop_unroll_adjust
10905 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
10906 #define TARGET_ASM_TRAMPOLINE_TEMPLATE s390_asm_trampoline_template
10907 #undef TARGET_TRAMPOLINE_INIT
10908 #define TARGET_TRAMPOLINE_INIT s390_trampoline_init
10910 #undef TARGET_UNWIND_WORD_MODE
10911 #define TARGET_UNWIND_WORD_MODE s390_unwind_word_mode
10913 struct gcc_target targetm
= TARGET_INITIALIZER
;
10915 #include "gt-s390.h"